+++ /dev/null
-## Rationale for the TLV abstraction
-
-hartmans: Luke, do you have opinions on what we want to do to the tokens for the target name stuff and SNI?
-lukeh: will we be sending a bunch of arbitrary attributes like the acceptor does? like GSS-Asserted-Acceptor-Service-Name
-lukeh: what's SNI?
-hartmans: I think it means that the first token from initiator to acceptor (which I guess is useless today) probably wants to be a TLV token.
-lukeh: yeah, I followed that on the mailing list but didn't implement as I was busy with other stuff
-lukeh: that sounds reasonable, I guess it would require some reorganisation
-lukeh: why do we want to send the target name?
-hartmans: If you're getting a target name back then you probably want a TLV token for the first acceptor to initiator. But today that also includes an eap identity request, right?
-lukeh: yes
-hartmans: The question is do we want to TLV all the EAP tokens at that point?
-lukeh: hmm.
-lukeh: well either way I guess has merit
-hartmans: The down side of TLVing everything is that then you can have news ways of having mallformed tokens.
-lukeh: but I probably need to understand things better
-lukeh: why does the acceptor send the target name back?
-lukeh: I'm starting from a point of not understanding what is wrong with the existing protocol :)
-hartmans: OK, so there are two different usages.
-lukeh: as my understanding is we get the acceptor name back anyway in the EAP CB attributes
-hartmans: 1) server name indication: the initiator knows the target name, but the acceptor may server multiple names.
-hartmans: Here, the acceptor wants to know what name the initiator is calling.
-***hartmans needs to think through the gss implications of what names the server accepts with Nico
-hartmans: The other usage is null target name: the initiator doesn't know the target name but will look at the target name in the established context and see if it likes it.
-hartmans: However in our mechanism, the initiator needs to learn the target name early so it can channel bind to it in EAP so it actually can believe the result.
-hartmans: No, we don't get the acceptor name back in the eap CB.
-hartmans: We *send* it in the eap cb and get back an indication about whether the server considered it in the CB result.
-lukeh: ah.
-hartmans: We don't get the value back.
-hartmans: The other thing I'm wondering.
-hartmans: Is there any actual information in the eap identity request?
-hartmans: Could we just fake that packet and save a round trip?
-lukeh: hmm, good point
-Josh: we need the realm for routing
-lukeh: that's in the response, not the request.
-hartmans: That's in the identity response, right?
-Josh: oh I mis-read
-lukeh: yeah, Sam, that's an interesting point
-lukeh: it didn't occur to me to do that
-Josh: that would be a violation of the EAP SM, no?
-lukeh: possibly
-lukeh: I mean, the initiator would generate it itself rather than getting it from the acceptor
-lukeh: certainly it's easier to understand in the current protocol
-lukeh: but whether it's not worth optimising away is another question.
-hartmans: Josh, it all depends on where you draw the layers.
-Josh: take your point
-hartmans: Note that on the radius server, the EAP SM never generates that packet... The NAS generates that one, right?
-Josh: right
-hartmans: I'm arguing we're basically doing the same thing. Rather than having the NAs generate that, we're asking the client to generate that and feed it to its own state machine.
-hartmans: That only works if the packet is content free.
-Josh: what's the advantage to that?
-lukeh: save a roundtrip
-lukeh: because currently the initiator sends an empty token
-lukeh: just to poke things into action
-Josh: IIRC, the data field can be used in the EAP-Identity/Request, but rarely is.
-lukeh: hmm, if we are optimising something away that might be useful
-hartmans: If it is used, what's it used for.
-lukeh: then that may be bad.
-hartmans: Sadly, Jari is not online. (He's the obvious eap expert in my buddy list)
-Josh: I think we have previously floated the idea of using that data field for federation selection
-lukeh: heh, can we shove the acceptor name in there
-hartmans: Josh, I think that the changes we're talking about for target name null and for sni will give us the rope we need for federation selection.
-***hartmans will investigate what that's useful for.
-lukeh: please define SNI
-lukeh: server name indication
-lukeh: ?
-Josh: True, and cleaner I believe.
-hartmans: Luke if you're going to have an opinion on what tokens should be TLVed, please do so preferably before IETF.
-hartmans: yes
-
-## Actual Design of resulting TLV implementation
-
-lukeh: OK, so the TLV code runs on a table of ( input token, output token, valid states, flags, callback )
-lukeh: there are still a few too many special cases to be completely in love with it
-lukeh: but hopefully after a few more days of hacking it will get there
-lukeh: I've added some debug tokens in the initial leg to test that that works
-lukeh: the previous extension tokens (GSS CB, reauth) are now just, obviously, TLVs in the last leg
-lukeh: there is now only one *GSS* token for context establishment, TOK_TYPE_ESTABLISH_CONTEXT
-lukeh: and a bunch of "inner" token types
-lukeh: #define ITOK_TYPE_CONTEXT_ERR 0x00000001
- #define ITOK_TYPE_ACCEPTOR_NAME_REQ 0x00000002
- #define ITOK_TYPE_ACCEPTOR_NAME_RESP 0x00000003
- #define ITOK_TYPE_EAP_RESP 0x00000004
- #define ITOK_TYPE_EAP_REQ 0x00000005
- #define ITOK_TYPE_GSS_CHANNEL_BINDINGS 0x00000006
- #define ITOK_TYPE_REAUTH_CREDS 0x00000007
- #define ITOK_TYPE_REAUTH_REQ 0x00000008
- #define ITOK_TYPE_REAUTH_RESP 0x00000009
- #define ITOK_TYPE_VERSION_INFO 0x0000000A
- #define ITOK_TYPE_VENDOR_INFO 0x0000000B
- #define ITOK_FLAG_CRITICAL 0x80000000 /* critical, wire flag */
-
-lukeh: s/last leg/last round trip/
-hartmans: What do you mean by token types in the table?
-hartmans: Do you mean that a given callback can only produce one output token type, or something more than that?
-lukeh: a given callback can only produce one output token type
-lukeh: true, a callback could have returned a set of tokens
-lukeh: perhaps that would be desirable
-lukeh: but as long as the state can be passed through the context or credential object
-lukeh: the same effect can be achieved
-hartmans: I think that simplicity is good for now.
-hartmans: OK, that implementation makes sense.
-lukeh: the trunk is simplicity :-)
-lukeh: although one nice thing is that the state machine walker is now shared between initiator and acceptor.
-lukeh: so I may have reduced the line count actually.
-lukeh: meh, about the same.
-hartmans: So, can a callback return success and no token?
-lukeh: yeah.
-hartmans: good.
-lukeh: (well, CONTINUE, or COMPLETE)
-lukeh: I needed that for a couple of cases.
-lukeh: first, if you transit states and don't emit a token
-hartmans: and anyone returning continue means overall result is continue?
-hartmans: anyone returning error means overall result is error?
-lukeh: (hang on)
-lukeh: e.g. going from completing the EAP stage (if no token emitted) to the extension stage
-hartmans: Does that produce a round-trip or does it simply re-run the packet in the new state combining the output?
-lukeh: (I'm going to answer your questions in order asked)
-hartmans: That (state transition) was the case that was making me think you'd need the no token logic
-lukeh: also, if you have a set of optional tokens (e.g. extensions) you need something to mark the end and transit the state.
-hartmans: Yes, answering in order asked is fine.
-lukeh: The very final callback that marks the context established returns COMPLETE
-lukeh: Others return CONTINUE on success or an error.
-lukeh: Returning an error means the overall result is an error and, on the acceptor side, an error token may be generated (this is done at the state machine layer)
-lukeh: In case of a state transition when there is no output token, then the next set of callbacks are called with no input token
-lukeh: to avoid regurgitating the input token from the previous state
-hartmans: OK that last is non-obvious to me (calling nwith no input token)
-lukeh: so, basically, tokens are emitted to the peer
-hartmans: Can I see the table? It might be more clear there.
-lukeh: when we've run out of callbacks or we transit states and have a token to send
-lukeh: (more or less, there are a couple of exceptions to handle things like sending no tokens on the initial context token in order to poke the acceptor into sending us an EAP request)
-lukeh: (rather than sending an empty EAP response from the initiator, as semantically that did not seem correct)
-lukeh: (although perhaps it is, who knows.)
-lukeh: if you checkout tlv
-lukeh: look for eapGssInitiatorSm in init_sec_context.c
-lukeh: and the corresponding one for the acceptor.
-lukeh: note: the states are bitmasks.
-lukeh: there's some verbose commenting in init_sm.c too.
-hartmans: Ah, that makes so much more sense from the table
-hartmans: This is quite clever.
-lukeh: Hmm, it still doesn't feel *quite* right. Too many exceptions. I suspect if that if I was doing it from scratch rather than refactoring it might look different. But, it seems to work for now. Will revise over coming days.
-lukeh: I think it is ugly because it collapses the state and token dimensions into a single one.
-lukeh: However it does make it easier to have tokens that support multiple states.
-lukeh: Although there are some limitations with that (the dispatch table is not retraversed after a state change so it effectively only works for exception tokens; of course, that's easily fixed)
-hartmans: So, we depend on the EAP machine keeping us in sync between the initiator and acceptor?
-hartmans: Not a problem, just confirming I understand.
-hartmans: What do you mean state and token are combined? They are separate columns in the table as far as I can tell
-lukeh: it's not a 2 dimensional array.
-lukeh: they are separate columns yes
-lukeh: re: depending on the EAP machine
-hartmans: Ah.
-lukeh: yes, I guess we do, what else could one do?
-lukeh: it's a black box
-lukeh: I don't know what happens yet if the EAP machine emits a token on success, I don't think that's possible though
-lukeh: I need to check
-***hartmans has too much of a relational database mindset to think of that as more than an efficiency issue
-lukeh: yeah, I never used relational databases, so I never had that mindset
-hartmans: I think depending on EAP to be consistent is fine.
-hartmans: We could actually echo the eap state in some sort of market token. That would be far far worse.
-lukeh: So, adding a new token that doesn't change the state is fairly easy.
-lukeh: You just need to be careful where you put it in the table :)
-lukeh: It needs to be before the state changing entry.
-lukeh: You can set the critical/required flags as desired.
-lukeh: Some care is likely necessary to avoid colliding with the reauthentication path. But I think that's fairly easy.
-lukeh: In my initial implementation I had overloaded the GSS status codes to mean various things, but in the end that got quite ugly.
-lukeh: So I went with the callbacks performing the state changes themselves plus a flag to indicate a few exceptional things (e.g. get-out now)
-hartmans: Presumably error always means get out now?
-lukeh: yes
-lukeh: if you want to get out now on CONTINUE, you need to either
-lukeh: a) change state and emit a token
-lukeh: b) change state and set FORCE_SEND_TOKEN (this is used to handle the initial case where we poke the acceptor without sending an inner token)
-lukeh: c) set STOP_EVAL - this is not used yet
-lukeh: the above applies to COMPLETE as well as CONTINUE
-lukeh: there are some sanity checks to make sure COMPLETE only happens when state becomes ESTABLISHED
-lukeh: see SM_ASSERT_VALID
-hartmans: OK, and we can remove an exception if we decide it's OK for us to fake the eap request identity
-lukeh: ah yes
-hartmans: This is really cool.