ceph/doc/dev/cephx_protocol.rst

============================================================
A Detailed Description of the Cephx Authentication Protocol
============================================================
Peter Reiher
7/13/12

This document provides deeper detail on the Cephx authorization protocol whose high level flow 
is described in the memo by Yehuda (12/19/09).  Because this memo discusses details of 
routines called and variables used, it represents a snapshot.  The code might be changed 
subsequent to the creation of this document, and the document is not likely to be updated in
lockstep.  With luck, code comments will indicate major changes in the way the protocol is
implemented.

Introduction
-------------

The basic idea of the protocol is based on Kerberos.  A client wishes to obtain something from 
a server.  The server will only offer the requested service to authorized clients.  Rather 
than requiring each server to deal with authentication and authorization issues, the system 
uses an authorization server.  Thus, the client must first communicate with the authorization 
server to authenticate itself and to obtain credentials that will grant it access to the
service it wants.

Authorization is not the same as authentication.  Authentication provides evidence that some 
party is who it claims to be.  Authorization provides evidence that a particular party is
allowed to do something.  Generally, secure authorization implies secure authentication 
(since without authentication, you may authorize something for an imposter), but the reverse 
is not necessarily true.  One can authenticate without authorizing.  The purpose 
of this protocol is to authorize.

The basic approach is to use symmetric cryptography throughout.  Each client C has its own
secret key, known only to itself and the authorization server A.  Each server S has its own
secret key, known only to itself and the authorization server A.  Authorization information 
will be passed in tickets, encrypted with the secret key of the entity that offers the service.
There will be a ticket that A gives to C, which permits C to ask A for other tickets.  This 
ticket will be encrypted with A's key, since A is the one who needs to check it.  There will 
later be tickets that A issues that allow C to communicate with S to ask for service.  These 
tickets will be encrypted with S's key, since S needs to check them.   Since we wish to provide 
security of the communications, as well, session keys are set up along with the tickets.  
Currently, those session keys are only used for authentication purposes during this protocol 
and the handshake between the client C and the server S, when the client provides its service 
ticket.  They could be used for authentication or secrecy throughout, with some changes to 
the system.

Several parties need to prove something to each other if this protocol is to achieve its 
desired security effects.

1.  The client C must prove to the authenticator A that it really is C.  Since everything
is being done via messages, the client must also prove that the message proving authenticity
is fresh, and is not being replayed by an attacker.

2.  The authenticator A must prove to client C that it really is the authenticator.  Again,
proof that replay is not occurring is also required.

3.  A and C must securely share a session key to be used for distribution of later
authorization material between them.  Again, no replay is allowable, and the key must be
known only to A and C.

4.  A must receive evidence from C that allows A to look up C's authorized operations with
server S.  

5.  C must receive a ticket from A that will prove to S that C can perform its authorized
operations.   This ticket must be usable only by C.

6.  C must receive from A a session key to protect the communications between C and S.  The
session key must be fresh and not the result of a replay.

Getting Started With Authorization
-----------------------------------

When the client first needs to get service, it contacts the monitor.  At the moment, it has 
no tickets.  Therefore, it uses the "unknown" protocol to talk to the monitor.  This protocol 
is specified as ``CEPH_AUTH_UNKNOWN``.  The monitor also takes on the authentication server 
role, A.  The remainder of the communications will use the cephx protocol (most of whose code 
will be found in files in ``auth/cephx``).  This protocol is responsible for creating and 
communicating the tickets spoken of above.  

Currently, this document does not follow the pre-cephx protocol flow.  It starts up at the 
point where the client has contacted the server and is ready to start the cephx protocol itself.

Once we are in the cephx protocol, we can get the tickets.  First, C needs a ticket that 
allows secure communications with A.  This ticket can then be used to obtain other tickets. 
This is phase I of the protocol, and consists of a send from C to A and a response from A to C.
Then, C needs a ticket to allow it to talk to S to get services.  This is phase II of the 
protocol, and consists of a send from C to A and a response from A to C.

Phase I:
--------

The client is set up to know that it needs certain things, using a variable called ``need``, 
which is part of the ``AuthClientHandler`` class, which the ``CephxClientHandler`` inherits 
from.  At this point, one thing that's encoded in the ``need`` variable is 
``CEPH_ENTITY_TYPE_AUTH``, indicating that we need to start the authentication protocol 
from scratch.  Since we're always talking to the same authorization server, if we've gone 
through this step of the protocol before (and the resulting ticket/session hasn't timed out), 
we can skip this step and just ask for client tickets.  But it must be done initially, and 
we'll assume that we are in that state.

The message C sends to A in phase I is build in ``CephxClientHandler::build_request()`` (in 
``auth/cephx/CephxClientHandler.cc``).  This routine is used for more than one purpose.  
In this case, we first call ``validate_tickets()`` (from routine 
``CephXTicektManager::validate_tickets()`` which lives in ``auth/cephx/CephxProtocol.h``).  
This code runs through the list of possible tickets to determine what we need, setting values 
in the ``need`` flag as necessary.  Then we call ``ticket.get_handler()``.  This routine 
(in ``CephxProtocol.h``) finds a ticket of the specified type (a ticket to perform 
authorization) in the ticket map, creates a ticket handler object for it,  and puts the 
handler into the right place in the map.  Then we hit specialized code to deal with individual 
cases.  The case here is when we still need to authenticate to A (the 
``if (need & CEPH_ENTITY_TYPE_AUTH)`` branch).

We now create a message of type ``CEPH_AUTH_UNKNOWN``.  We need to authenticate 
this message with C's secret key, so we fetch that from the local key repository.  (It's 
called a key server in the code, but it's not really a separate machine or processing entity.
It's more like the place where locally used keys are kept.)  We create a 
random challenge, whose purpose is to prevent replays.  We encrypt that challenge.  We already 
have a server challenge (a similar set of random bytes, but created by the server and sent to
the client) from our pre-cephx stage.  We take both challenges and our secret key and 
produce a combined encrypted challenge value, which goes into ``req.key``.

If we have an old ticket, we store it in ``req.old_ticket``.  We're about to get a new one.

The entire ``req`` structure, including the old ticket and the cryptographic hash of the two 
challenges, gets put into the message.  Then we return from this function, and the 
message is sent.

We now switch over to the authenticator side, A.  The server receives the message that was 
sent, of type ``CEPH_AUTH_UNKNOWN``.  The message gets handled in ``prep_auth()``, 
in ``mon/AuthMonitor.cc``, which calls ``handle_request()`` is ``CephxServiceHandler.cc`` to 
do most of the work.  This routine, also, handles multiple cases.  

The control flow is determined by the ``request_type`` in the ``cephx_header`` associated 
with the message.  Our case here is ``CEPH_AUTH_UNKNOWN``.  We need the 
secret key A shares with C, so we call ``get_secret()`` from out local key repository to get 
it.  We should have set up a server challenge already with this client, so we make sure 
we really do have one.  (This variable is specific to a ``CephxServiceHandler``, so there 
is a different one for each such structure we create, presumably one per client A is 
dealing with.)  If there is no challenge, we'll need to start over, since we need to 
check the client's crypto hash, which depends on a server challenge, in part.

We now call the same routine the client used to calculate the hash, based on the same values: 
the client challenge (which is in the incoming message), the server challenge (which we saved), 
and the client's key (which we just obtained).  We check to see if the client sent the same 
thing we expected.  If so, we know we're talking to the right client.  We know the session is 
fresh, because it used the challenge we sent it to calculate its crypto hash.  So we can
give it an authentication ticket.

We fetch C's ``eauth`` structure.  This contains an ID, a key, and a set of caps (capabilities).

The client sent us its old ticket in the message, if it had one.  If so, we set a flag,
``should_enc_ticket``, to true and set the global ID to the global ID in that old ticket.  
If the attempt to decode its old ticket fails (most probably because it didn't have one),
``should_enc_ticket`` remains false.  Now we set up the new ticket, filling in timestamps, 
the name of C, the global ID provided in the method call (unless there was an old ticket), and 
his ``auid``, obtained from the ``eauth`` structure obtained above.  We need a new session key 
to help the client communicate securely with us, not using its permanent key.    We set the
service ID to ``CEPH_ENTITY_TYPE_AUTH``, which will tell the client C what to do with the 
message we send it.  We build a cephx response header and call 
``cephx_build_service_ticket_reply()``.

``cephx_build_service_ticket_reply()`` is in ``auth/cephx/CephxProtocol.cc``.  This 
routine will build up the response message.   Much of it copies data from its parameters to 
a message structure.  Part of that information (the session key and the validity period) 
gets encrypted with C's permanent key.  If the ``should_encrypt_ticket`` flag is set, 
encrypt it using the old ticket's key.  Otherwise, there was no old ticket key, so the 
new ticket is not encrypted.  (It is, of course, already encrypted with A's permanent key.)  
Presumably the point of this second encryption is to expose less material encrypted with 
permanent keys.

Then we call the key server's ``get_service_caps()`` routine on the entity name, with a 
flag ``CEPH_ENTITY_TYPE_MON``, and capabilities, which will be filled in by this routine.  
The use of that constant flag means we're going to get the client's caps for A, not for some 
other data server.  The ticket here is to access the authorizer A, not the service S.  The 
result of this call is that the caps variable  (a parameter to the routine we're in) is 
filled in with the monitor capabilities that will allow C to  access A's authorization services.

``handle_request()`` itself does not send the response message.  It builds up the 
``result_bl``, which basically holds that message's contents, and the capabilities structure, 
but it doesn't send the message.  We go back to ``prep_auth()``, in ``mon/AuthMonitor.cc``, 
for that.    This routine does some fiddling around with the caps structure that just got 
filled in.  There's a global ID that comes up as a result of this fiddling that is put into 
the reply message.  The reply message is built here (mostly from the ``response_bl`` buffer) 
and sent off.

This completes Phase I of the protocol.  At this point, C has authenticated itself to A, and A has generated a new session key and ticket allowing C to obtain server tickets from A.

Phase II
--------

This phase starts when C receives the message from A containing a new ticket and session key.
The goal of this phase is to provide C with a session key and ticket allowing it to
communicate with S.

The message A sent to C is dispatched to ``build_request()`` in ``CephxClientHandler.cc``, 
the same routine that was used early in Phase I to build the first message in the protocol.  
This time, when ``validate_tickets()`` is called, the ``need`` variable will not contain 
``CEPH_ENTITY_TYPE_AUTH``, so a different branch through the bulk of the routine will be 
used.  This is the branch indicated by ``if (need)``.  We have a ticket for the authorizer, 
but we still need service tickets.

We must send another message to A to obtain the tickets (and session key) for the server 
S.  We set the ``request_type`` of the message to ``CEPHX_GET_PRINCIPAL_SESSION_KEY`` and 
call ``ticket_handler.build_authorizer()`` to obtain an authorizer.  This routine is in 
``CephxProtocol.cc``.  We set the key for this authorizer to be the session key we just got 
from A,and create a new nonce.  We put the global ID, the service ID, and the ticket into a 
message buffer that is part of the authorizer.  Then we create a new ``CephXAuthorize`` 
structure.  The nonce we just created goes there.  We encrypt this ``CephXAuthorize`` 
structure with the current session key and stuff it into the authorizer's buffer.  We 
return the authorizer.

Back in ``build_request()``, we take the part of the authorizer that was just built (its 
buffer, not the session key or anything else) and shove it into the buffer we're creating 
for the message that will go to A.  Then we delete the authorizer.  We put the requirements 
for what we want in ``req.keys``, and we put ``req`` into the buffer.  Then we return, and 
the message gets sent.

The authorizer A receives this message which is of type ``CEPHX_GET_PRINCIPAL_SESSION_KEY``.
The message gets handled in ``prep_auth()``, in ``mon/AuthMonitor.cc``, which again calls 
``handle_request()`` in ``CephxServiceHandler.cc`` to do most of the work.  

In this case, ``handle_request()`` will take the ``CEPHX_GET_PRINCIPAL_SESSION_KEY`` case. 
It will call ``cephx_verify_authorizer()`` in ``CephxProtocol.cc``.  Here, we will grab 
a bunch of data out of the input buffer, including the global and service IDs and the ticket 
for A.   The ticket contains a ``secret_id``, indicating which key is being used for it.     
If the secret ID pulled out of the ticket was -1, the ticket does not specify which secret 
key A should use.  In this case, A should use the key for the specific entity that C wants
to contact, rather than a rotating key shared by all server entities of the same type.
To get that key, A must consult the key repository to find the right key.   Otherwise, 
there's already a structure obtained from the key repository to hold the necessary secret.  
Server secrets rotate on a time expiration basis (key rotation is not covered in this
document), so run through that structure to find its current secret.  Either way, A now 
knows the secret key used to create this ticket.  Now decrypt the encrypted part of the 
ticket, using this key.  It should be a ticket for A.  

The ticket also contains a session key that C should have used to encrypt other parts of 
this message.  Use that session key to decrypt the rest of the message.  

Create a ``CephXAuthorizeReply`` to hold our reply.  Extract the nonce (which was in the stuff 
we just decrypted), add 1 to it, and put the result in the reply.  Encrypt the reply and 
put it in the buffer provided in the call to ``cephx_verify_authorizer()`` and return 
to ``handle_request()``.  This will be used to prove to C that A (rather than an attacker) 
created this response.

Having verified that the message is valid and from C, now we need to build it a ticket for S.
We need to know what S it wants to communicate with and what services it wants.  Pull the
ticket request that describes those things out of its message.  Now run through the ticket
request to see what it wanted.  (He could potentially be asking for multiple different
services in the same request, but we will assume it's just one, for this discussion.)  Once we 
know which service ID it's after, call ``build_session_auth_info()``.

``build_session_auth_info()`` is in ``CephxKeyServer.cc``.  It checks to see if the 
secret for the ``service_ID`` of S is available and puts it into the subfield of one of 
the parameters, and calls the similarly named ``_build_session_auth_info()``, located in 
the same file.      This routine loads up the new ``auth_info`` structure with the 
ID of S, a ticket, and some timestamps for that ticket.  It generates a new session key 
and puts it in the structure.   It then calls ``get_caps()`` to fill in the 
``info.ticket`` caps field.  ``get_caps()`` is also in ``CephxKeyServer.cc``.  It fills the 
``caps_info`` structure it is provided with caps for S allowed to C.

Once ``build_session_auth_info()`` returns, A has a list of the capabilities allowed to 
C for S.  We put a validity period based on the current TTL for this context into the info 
structure, and put it into the ``info_vec`` structure we are preparing in response to the 
message.  

Now call ``build_cephx_response_header()``, also in ``CephxServiceHandler.cc``.   Fill in 
the ``request_type``, which is ``CEPHX_GET_PRINCIPAL_SESSION_KEY``, a status of 0, 
and the result buffer.  

Now call ``cephx_build_service_ticket_reply()``, which is in ``CephxProtocol.cc``.  The 
same routine was used towards the end of A's handling of its response in phase I.  Here, 
the session key (now a session key to talk to S, not A) and the validity period for that 
key will be encrypted with the existing session key shared between C and A.  
The ``should_encrypt_ticket`` parameter is false here, and no key is provided for that 
encryption.  The ticket in question, destined for S once C sends it there, is already 
encrypted with S's secret.  So, essentially, this routine will put ID information, 
the encrypted session key, and the ticket allowing C to talk to S into the buffer to 
be sent to C.

After this routine returns, we exit from ``handle_request()``, going back to ``prep_auth()`` 
and ultimately to the underlying message send code.  

The client receives this message. The nonce is checked as the message passes through
``Pipe::connect()``, which is in ``msg/SimpleMessager.cc``.  In a lengthy ``while(1)`` loop in
the middle of this routine, it gets an authorizer.  If the get was successful, eventually
it will call ``verify_reply()``, which checks the nonce.  ``connect()`` never explicitly
checks to see if it got an authorizer, which would suggest that failure to provide an
authorizer would allow an attacker to skip checking of the nonce.  However, in many places,
if there is no authorizer, important connection fields will get set to zero, which will
ultimately cause the connection to fail to provide data.  It would be worth testing, but
it looks like failure to provide an authorizer, which contains the nonce, would not be helpful
to an attacker.

The message eventually makes its way through to ``handle_response()``, in 
``CephxClientHandler.cc``.    In this routine, we call ``get_handler()`` to get a ticket 
handler to hold the ticket we have just received.  This routine is embedded in the definition 
for a ``CephXTicketManager`` structure.  It takes a type (``CEPH_ENTITY_TYPE_AUTH``, in 
this case) and looks through the ``tickets_map`` to find that type.  There should be one, and 
it should have the session key of the session between C and A in its entry.  This key will 
be used to decrypt the information provided by A, particularly the new session key allowing 
C to talk to S.

We then call ``verify_service_ticket_reply()``, in ``CephxProtocol.cc``.  This routine 
needs to determine if the ticket is OK and also obtain the session key associated with this 
ticket.  It decrypts the encrypted portion of the message buffer, using the session key 
shared with A.  This ticket was not encrypted (well, not twice - tickets are always encrypted, 
but sometimes double encrypted, which this one isn't).  So it can be stored in a service 
ticket buffer directly.  We now grab the ticket out of that buffer.  

The stuff we decrypted with the session key shared between C and A included the new session 
key.  That's our current session key for this ticket, so set it.  Check validity and 
set the expiration times.  Now return true, if we got this far.  

Back in ``handle_response()``, we now call ``validate_tickets()`` to adjust what we think 
we need, since we now have a ticket we didn't have before.  If we've taken care of 
everything we need, we'll return 0.

This ends phase II of the protocol.  We have now successfully set up a ticket and session key 
for client C to talk to server S.  S will know that C is who it claims to be, since A will
verify it.  C will know it is S it's talking to, again because A verified it.  The only
copies of the session key for C and S to communicate were sent encrypted under the permanent
keys of C and S, respectively, so no other party (excepting A, who is trusted by all) knows
that session key.  The ticket will securely indicate to S what C is allowed to do, attested 
to by A.  The nonces passed back and forth between A and C ensure that they have not been 
subject to a replay attack.  C has not yet actually talked to S, but it is ready to.

Much of the security here falls apart if one of the permanent keys is compromised.  Compromise
of C's key means that the attacker can pose as C and obtain all of C's privileges, and can
eavesdrop on C's legitimate conversations.  He can also pretend to be A, but only in 
conversations with C.  Since it does not (by hypothesis) have keys for any services, he
cannot generate any new tickets for services, though it can replay old tickets and session
keys until S's permanent key is changed or the old tickets time out. 

Compromise of S's key means that the attacker can pose as S to anyone, and can eavesdrop on 
any user's conversation with S.  Unless some client's key is also compromised, the attacker
cannot generate new fake client tickets for S, since doing so requires it to authenticate
himself as A, using the client key it doesn't know.