[keycloak-dev] Better support for "scope" in adapters
Stian Thorgersen
sthorger at redhat.com
Tue Feb 5 03:39:37 EST 2019
Tried to filter out implementation details here, so may have lost some
details. It would be good if we can try to keep discussions at a higher
level at least initially as it makes it much easier to follow the
discussion.
For scopes I can see the most common use-case will be the ability to do
incremental scopes. By that I mean the application doesn't request all
permissions it may need, but rather starts small and asks for more as the
user extends use of the application. This is mostly relevant to
applications that require consent.
Now with regards to the application being able to have different tokens to
invoke different services I'm not convinced this is needed so we should
rather wait for demand here. There's two ways a single app can consume
multiple services:
1) Application directly invokes multiple services - in this case the
application should be able to use scopes or token exchange to obtain tokens
to invoke different services. In fact I'd say token exchange is probably
what is wanted here rather than scopes.
2) Application invokes a backend service that aggregates multiple services
- in this case token exchange comes in as the backend service needs to be
able to obtain tokens to invoke the different services
I would think option 2 is the best approach as it allows implementing the
complex code in server side code and also makes the application more
transparent to API changes.
With regards to incremental scope support we need to be able to do that
without requiring logout. For JS adapter that should already work, but it
has one issue and that is you can't set the scope to use with automated
login. We should probably make the scope configurable in the init function
as well as when invoking login function directly. For servlet we should
probably also have a way to configure the initial scope and expose a method
to obtain additional scopes without requiring logout.
I'm not convinced about client scope inheritance. It has an additional
implementation complexity, but most importantly usability with regards to
understand what is actually included in the token when you have
inheritance. It may also have some strange side-effects like how do you
make sure the order of what is applied is correct. Again, probably
something best left until there is demand for it.
On Thu, 31 Jan 2019 at 17:04, Pedro Igor Silva <psilva at redhat.com> wrote:
> +1
>
> On Thu, Jan 31, 2019 at 11:14 AM Marek Posolda <mposolda at redhat.com>
> wrote:
>
> > On 31/01/2019 14:07, Pedro Igor Silva wrote:
> >
> > This what I like most about client scopes (in addition to all mapping you
> > do to claims in the tokens) :) Would also make sense to do the same thing
> > to client scopes ? So clients requesting "foo" would also get "bar" and
> > "xpto", for instance ?
> >
> > Maybe this could avoid the client to request 10 scopes but just a more
> > coarse-grained scope representing all of them.
> >
> > There is opened JIRA for client scopes inheritance
> > https://issues.jboss.org/browse/KEYCLOAK-6633 . I believe this will
> cover
> > what you have in mind? It's just not yet done...
> >
> > Marek
> >
> >
> >
> > On Thu, Jan 31, 2019 at 10:43 AM Marek Posolda <mposolda at redhat.com>
> > wrote:
> >
> >> On 30/01/2019 14:40, Pedro Igor Silva wrote:
> >>
> >>
> >>
> >> On Wed, Jan 30, 2019 at 5:25 AM Marek Posolda <mposolda at redhat.com>
> >> wrote:
> >>
> >>> On 29/01/2019 19:49, Pedro Igor Silva wrote:
> >>>
> >>> I'm not sure if we need to consider that in our adapters.
> >>>
> >>> Usually, the front-end knows the set of scopes that it needs to
> interact
> >>> with the backend and stay functional.
> >>>
> >>> Maybe. I am personally not sure how people expect to use "scope"
> >>> parameters in their frontend applications. Maybe 90% of frontend
> clients
> >>> don't need to use "scope" parameter at all. And from the remaining,
> they
> >>> will be fine with the current support of the "scope" parameter.
> >>>
> >> I would say so, mainly because I think people are still using RBAC to
> >> enforce access to APIs. Enterprise scenarios don't really use scopes as
> >> they are more related with delegation.
> >>
> >> Yeah, maybe. Just a note that our client scopes support also allows to
> >> limit roles in the token. For example you can define role scope
> mappings of
> >> client scope "service1" to have role "service1.my-service1-role" . So by
> >> requesting "scope=service1", you will also receive this role in the
> token
> >> and hence can be used for RBAC based authorization.
> >>
> >> But anyway, I probably won't create any JIRAs for now. Will wait if
> there
> >> is some more feedback or some more requests for better support of
> "scope"
> >> parameter in the adapters.
> >>
> >> Thanks for the feedback Pedro!
> >>
> >> Marek
> >>
> >> One possibility, where I can see usage of this, is when frontend client
> >>> wants to invoke multiple different services and he wants to use
> different
> >>> access tokens with properly "isolated" audiences. So for example you
> want
> >>> to have:
> >>>
> >>> - access token with "scope=service1", which will have in itself
> audience
> >>> claim "aud: service1" and you will use it to invoke backend service1
> >>> - access token with "scope=service2", which will have in itself
> audience
> >>> claim "aud: service2" and you will use it to invoke backend service2
> >>>
> >>> In this case, having the possibility for adapters to "cache" multiple
> >>> tokens for various scopes can be beneficial IMO, so client can easily
> >>> retrieve proper access token based on the service he wants to invoke.
> >>>
> >>> And the backend by itself is free to exchange tokens to call other
> >>> services (service chaining).
> >>>
> >>> Don't think that brings a lot of complexity to the front-end and,
> >> probably, indicates a bad design?
> >>
> >>> IMO in many cases, you're right. For example when frontend client uses
> >>> access token to invoke backend "service1", this backend may want to
> >>> retrieve some other data from "service11". So service1 backend needs to
> >>> reuse the token or he wants to exchange this token.
> >>>
> >>> But in many cases, you want to avoid this. So in my example above, when
> >>> you have access token with "aud: service1", you want this access token
> to
> >>> be used solely to invoke service1. You don't want to have one huge
> access
> >>> token, which will have all the audiences like:
> >>>
> >>> aud: [ "service1", "service2" ]
> >>>
> >> The access token is also tied with the client, what means "this client
> is
> >> allowed to invoke service1 and service2". I usually don't see a problem
> on
> >> that if you consider that multiple audiences mean that a high degree of
> >> trust between the parties involved. What I think is true for most
> >> enterprise use cases where the front-end is basically accessing internal
> >> services.
> >>
> >> It is also worthy to consider, IMO, that the fact that you have distinct
> >> services, does not mean they are not part of the same API, thus the same
> >> audience.
> >>
> >>> You may want separate access tokens with isolated audiences exactly
> >>> because you don't want service1 and service2 to be able to invoke each
> >>> other. IMO this isolation is one of the main usages of the "aud" claim
> in
> >>> the tokens.
> >>>
> >>>
> >>> One thing that makes sense though is "incremental authorization" and
> >>> allow apps to request additional scopes during an authentication
> session,
> >>> so the app gets what was previously granted and the new scopes
> (depending
> >>> on user consent). But I think we support that already, right ?
> >>>
> >>> We don't support it directly and maybe this is something to improve.
> ATM
> >>> you will need programatically do something like this:
> >>>
> >>> String existingScope = existingAccessToken.getScope();
> >>>
> >>> // I want to use existingScope and add "phone" scope to it
> >>> String newScope = existingScope + " phone";
> >>>
> >>> // Request the login for new scope now
> >>>
> >>>
> >>> The part of "requesting login for new scope" is possible with
> javascript
> >>> adapter, but not easily with the "java" adapter. With java adapter,
> there
> >>> is no easy way to manually "build" URL to sent to OIDC authentication
> >>> endpoint (equivalent of keycloak.js method "createLoginUrl"). That's
> also
> >>> one of the things, which I proposed to improve in this email thread.
> >>>
> >> Marek
> >>>
> >>>
> >>> Regards.
> >>> Pedro Igor
> >>>
> >>> On Tue, Jan 29, 2019 at 9:36 AM Marek Posolda <mposolda at redhat.com>
> >>> wrote:
> >>>
> >>>> During my work on Client Scopes last year, I did not any work on the
> >>>> adapters side. I think there is a space for improvement here. I will
> >>>> try
> >>>> to summary current issues and some initial proposals for improve
> >>>> things.
> >>>> Suggestions welcomed! And sorry for longer email :)
> >>>>
> >>>>
> >>>> Both javascript adapter and servlet adapter has some way for
> requesting
> >>>> the additional "scope" and ensure that that initial OIDC
> authentication
> >>>> request sent to Keycloak will contain some custom "scope" parameter.
> >>>> The
> >>>> javascript adapter has support for "scope" as an option of the "login"
> >>>> method [1]. The servlet adapter has a possibility to inject custom
> >>>> "scope" with parameters forwarding [2]. I am not sure about node.js
> and
> >>>> other adapters TBH. But even for javascript and servlet adapter, the
> >>>> current support is quite limited for few reasons. For example:
> >>>>
> >>>> - The approach of parameters forwarding used in servlet adapters
> >>>> requires to logout before requesting the additional scope. Because
> >>>> when
> >>>> I am already authenticated in the application and I open secured URL
> >>>> like http://localhost/app/secured?scope=some-custom-scope, the
> adapter
> >>>> will just ignore it in case that user is already logged-in and it will
> >>>> automatically forward to the application.
> >>>>
> >>>> - Both servlet and javascript adapters support to have just single
> >>>> "triplet" of tokens per browser session. In this context "triplet"
> >>>> means
> >>>> the single set of 3 tokens (ID token , Access Token , Refresh token).
> >>>> So
> >>>> for example when I want to request the custom scope for being able to
> >>>> invoke "service1", I can use "scope=service1". However after Keycloak
> >>>> redirects me back to the application, the existing triplet of tokens
> is
> >>>> just replaced with the new one for "service1" . Then when I want to
> >>>> later invoke another service like "service2", I need to request the
> >>>> additional scope "scope=service2", which will replace my tokens on the
> >>>> adapter's side with the "service2" tokens . But then later when I want
> >>>> to switch again to "service1", I need to redirect to Keycloak again as
> >>>> the current triplet of tokens for "service1" etc.
> >>>>
> >>>>
> >>>> To improve this limitation, I think that it will be good if adapters
> >>>> easily support the following:
> >>>>
> >>>> - Instead of having single triplet of tokens, it will be good if
> >>>> adapters can contain Map of tokens. Key of the map can be "scope"
> >>>> parameter. So for example, the adapter will have "default" tokens
> >>>> (those, which were used for initial login), the tokens for "service1"
> >>>> and the tokens for "service2" .
> >>>>
> >>>> - It will be nice if there is easy way to ask adapter for "service1"
> >>>> scope. In case that I don't have yet this scope, adapter will redirect
> >>>> me to Keycloak with "scope=service1". If I already have it, adapter
> >>>> will
> >>>> return me an existing token. If existing access token is expired,
> >>>> adapter will refresh the access token for requested scope in the
> >>>> background and return me the "updated" token.
> >>>>
> >>>> - When I want to invoke service1 and I need to use "scope=service1", I
> >>>> usually need just access token and refresh token. I don't need ID
> Token
> >>>> anymore. I also don't need the "profile" and "email" claims to be
> >>>> returned in the new access token. This is related to the JIRA of
> having
> >>>> the server-side support for client scopes like (always, default,
> >>>> optional) instead of current (default, optional) [3]. In other words,
> >>>> the client scopes "profile" and "email" will be default client scopes,
> >>>> which means that if I don't use "scope=openid" in the OIDC initial
> >>>> request, the "profile" and "email" will be ommited from the response
> as
> >>>> well as the ID Token will be ommited from the response.
> >>>>
> >>>>
> >>>> So how to support this on adapters? For Keycloak.js, I can think about
> >>>> some variation of existing "update" method like this:
> >>>>
> >>>>
> >>>> keycloak.updateTokenWithScope('service1',
> >>>> 5).success(function(accessToken, refreshed) {
> >>>> if (refreshed) {
> >>>> alert("I had existing accessToken for scope 'service1',
> >>>> but
> >>>> it needed to be refreshed as it was expired or about to expire in less
> >>>> than 5 seconds");
> >>>> } else {
> >>>> alert('I have accessToken for 'service1', which didn't
> >>>> need to be refreshed');
> >>>> }
> >>>> // I can invoke REST service now with the accessToken
> >>>> ...
> >>>> }).error(function() {
> >>>> alert("Failed to refresh the token OR I don't have yet scope
> >>>> for 'service1' .");
> >>>> // User usually need to call keycloak.login with the
> requested
> >>>> scope now...
> >>>> });
> >>>>
> >>>>
> >>>>
> >>>>
> >>>> For servlet adapter something like:
> >>>>
> >>>> KeycloakSecurityContext ctx = ... // Retrieved from HttpServletRequest
> >>>> or Principal as currently
> >>>>
> >>>> if (ctx.hasScope("service1")) {
> >>>> try {
> >>>> String accessToken = ctx.getScope("service1");
> >>>> // Invoke service1 with accessToken now
> >>>> } catch (TokenRefreshException ex) {
> >>>> log.error("I already had scope for service1, but failed to
> >>>> refresh the token. Need to re-login for the scope service1");
> >>>>
> >>>> // See method below
> >>>> redirectToKeycloakWithService1Scope();
> >>>> }
> >>>> } else {
> >>>> // See method below
> >>>> redirectToKeycloakWithService1Scope();
> >>>> }
> >>>>
> >>>>
> >>>> private void redirectToKeycloakWithService1Scope() {
> >>>> KeycloakRedirectUriBuilder builder = ctx.createLoginUrl();
> >>>> URL url = builder.scope("service1").build();
> >>>> httpServletResponse.sendRedirect(url);
> >>>> }
> >>>>
> >>>>
> >>>> Regarding the class KeycloakRedirectUriBuilder, I was thinking about
> >>>> this class so that servlet adapter are able to easily create login URL
> >>>> with custom values for things like scope, prompt, max_age etc. This
> >>>> capability is currently missing in servlet adapters and the current
> >>>> approach based on parameters forwarding is a bit clunky for few
> >>>> reasons.
> >>>> One reason is usability and the other is, that you need to logout
> first.
> >>>>
> >>>>
> >>>> [1]
> >>>>
> >>>>
> https://www.keycloak.org/docs/latest/securing_apps/index.html#javascript-adapter-reference
> >>>> [2]
> >>>>
> >>>>
> https://www.keycloak.org/docs/latest/securing_apps/index.html#_params_forwarding
> >>>> [3] https://issues.jboss.org/browse/KEYCLOAK-8323
> >>>>
> >>>> Marek
> >>>>
> >>>> _______________________________________________
> >>>> keycloak-dev mailing list
> >>>> keycloak-dev at lists.jboss.org
> >>>> https://lists.jboss.org/mailman/listinfo/keycloak-dev
> >>>
> >>>
> >>>
> >>
> >
> _______________________________________________
> keycloak-dev mailing list
> keycloak-dev at lists.jboss.org
> https://lists.jboss.org/mailman/listinfo/keycloak-dev
>
More information about the keycloak-dev
mailing list