12:09 p.m.
On 17.7.2014 17:00, Stian Thorgersen wrote:
Just started on this. Quick summary of my attack plan:
1. Remove AccessToken from AccessCode. Instead add List<String> rolesRequested and
create AccessToken in TokenService.accessCodeToToken. Any reasons why this wouldn't
work?
2. Move some info from AccessCode to UserSession, and remove anything redundant. For
example username used to login, auth-method (social or form), remember-me.
3. Add info from AccessCode to UserSessionProvider and send signed key instead as the
code query param.
With regards to the required info here's what I've come up with:
To verify the code:
* Need to know it hasn't been done before - can either be done by having a timestamp
on ClientAssociation that is incremented every time a code is requested/used, or by having
a ClientAssociation per-code. As we'll need to store more specific to a code than just
the fact it has been used or not, I think it's best to just have a ClientAssociation
per-code
* Not expired - no prob
* Session active - no prob
* redirect_uri when retrieving code matches redirect_uri query param when swapping for
token - we don't do this currently I think, but spec requires it. This would require a
ClientAssociation per-code.
* Correct client_id - no prob, just make sure code belongs to the client_id query param
Then to create the token we need:
* List of roles requested - the union of user roles and app/client scope. This would
require a ClientAssociation per-code as the role-mappings/scope could change between user
granted the roles to the client and the code is exchanged. Also in the future once we add
a scope query param this will be required.
Did you mean intersection rather than
union?
* Realm - no prob, user-session is associated with a realm
* Client - no prob, ClientAssociation is associated with a client
* User - no prob, user-session is associated with a user
* Session - no prob, user-session is a session;)
All in all it seems like a bit more work than I initially considered. Or am I attacking
this completely wrong?!?
To me it makes sense and should work.
I wonder about a bit less-time consuming flow to implement, when
AccessCode is generated together with AccessToken after authentication
as it's now (in OAuthFlows.processAccessCode) . Only new thing in
OAuthFlows.processAccessCode will be that ClientAssociation with just
key-value pair of <signed-codeId>:<base64-encoded-AccessCode> will be
associated with UserSession. Signed codeId will be then saved as "code"
parameter .
Then once client will want to exchange code, Keycloak will perform
lookup of correct UserSession based on codeId (new method on
UserSessionProvider like: getSessionbyCode). Then when UserSession is
successfully looked-up, ClientAssociation is immediatelly deleted, so
that it's not possible to exchange same code twice. In
TokenService.accessCodeToToken, we will still have AccessCode with
wrapped AccessToken so we can do all verification is it's done now. Only
new thing might be redirect_uri verification, which is not currently done.
Defacto it will be same as current flow, only new thing will be:
- ClientAssociation with the pair
<signed-codeId>:<base64-encoded-AccessCode>, created when code is
generated and deleted when code is exchanged.
- redirect_uri validation during exchange code to token.
I think that this approach will help us to address 2 main issues, which
we currently have (long "code" parameter and same code is possible to
exchange more times). Only thing is, that base64-encoded-AccessCode will
be quite long and will contain redundant info, but is it big problem? If
we want later add new info to AccessCode, we don't need to do any
changes in UserSessionProvider model as it will be still just base64
encoded string.
Not sure which approach is better, just a though:-)
Marek