On 1 July 2016 at 14:34, Marek Posolda <mposolda@redhat.com> wrote:
On 01/07/16 09:37, Stian Thorgersen wrote:
I don't like option (a) as it would be very clear what scopes are available and what they result in. You'd have to search through all roles and protocol mappers to find a list of scopes as well as somehow manually build the picture of what a scope implies.

A scope could affect:

* Behavior - for example openid and offline_access. We can hard code those, but what about others and custom ones?
For "openid", I think we can hardcode. I was thinking about adding SPI like ScopeParamMapper or something and drive the behaviour with it, however it won't work well and will require callbacks in various stages (ie. IDToken itself is created at some point, but whole AccessTokenResponse itself and control which tokens go to it is different point.

For offline_access we already has that driven by realm role actually.
* Roles - limit what roles are included. This is probably the simplest case as it's just about removing roles.
Not sure what exactly you mean. For roles, we already have the "scopeParamRequired" flag, so we can both include and remove. For example user is member of roleA, roleB, roleC and they have:
role-a : scopeParamRequired=false
role-b : scopeParamRequired=true
role-c : scopeParamRequired=true

and request will have "scope=role-b" then in token will be included both role-a and role-b . Just role-c won't be (as it requires scopeParamRequired but wasn't included in scopeParam).

We currently use it for offline_access role, which is default role, but have scopeParamRequired, so it's included just if "scope=offline_acces" is used.

I meant that a single scope should be able to change several things, not just enable/disable a single role.
 

* Claims - limit what claims are included in the token. This boils down to protocol mappers. Some protocol mappers would want to be included/excluded based on scope, but I can also see that some protocol mappers would internally look at scope to determine what to include.
* Authorization - could also affect authorization separately? The token contains the scope param separately, so authorization policies could be written directly on scope rather than roles/attributes

Another thing is that the consent screen (and also showing application access in account management console) should probably take about scope rather than individual roles. For example:

* Gallery wants to view your personal details
* Backup application wants to access your photos

A few more points:

* It should be possible to view scopes available for a realm directly, not by scanning through roles and protocol mappers
* It should be possible to set a description on a scope
* It should be possible to define a scope that maps to multiple roles
* It should be possible to define a scope that maps to multiple protocol mappers roles
* It may be useful to be able to have protocol mappers that behave differently depending on the scope. Complexity may outweigh usefulness here though.

With that in mind I don't think option (a) is great. Option (b) would need introducing a whole new concept.

How about we use a combination of (a) and (b), by using composite roles? Something along the lines of:

* scope maps to a role. but the role could be a composite role and hence expand to other roles.
* Protocol mappers could require a specific role to be applied
* We could add a page to view scopes for a realm
  - This would show the corresponding role as well as effective roles if it's a composite role
  - It would also list the protocol mappers included
+1 to this concept and "mapping" scope param values to roles. However there are some things to clarify.

- OIDC specs has definition for scopes like "phone" , "email" and "profile" . So we will need the roles for those. Isn't that confusing a bit? IMO our RoleModel applies well for the scope parameter, just not sure if it's not rather mis-using of role concept and bit confusing for users?

Would be better if we had role namespaces as we could then have a OIDC namespace. That would remove some of the confusion I think.
 

- My understanding is, that scope is per-client thing rather than per-realm thing? For example if you use "scope=photo", it's applicable just for client "photogallery". Also the same scope parameter may have different meaning for different clients ( eg. if you use "scope=admin" in client "product-app" , the token will receive claims/permissions for managing products. However using "scope=admin" in client "customer-app" , the token will manage customers). So with respect to this, scope looks to me more like either per-client thing or per-clientTemplate thing (if some logic is common for more clients).

Not sure I full agree with that. I think it can be a per-realm and per-client thing, at least per group of clients. The scope could affect the app, services it invokes and also authorization policies. I would actually think that most organizations would define a "organization" level scopes rather than per-client scopes.
 

- We will need some way how to map value of scope parameter to particular role.
The simple (but probably not correct) way: On role definition, we can have something like "Scope param value" where you can configure value "photo" and it would mean that the "scope=photo" is related to this paricular role.
However what if there are clashes? For example someone can configure the scope parameter value "photo" for realm role "realm-photo-role" or client role "photogalery/client-photo-role" . Then when you use "scope=photo", which one will be used?


Why not just use the role name as the scope name? A composite role can be used to create an "alias" if required.
 
So after we add roleNamespaces, I can imagine something like this:
- Each client will have reserved namespace, where all roles mapped to "scope" parameter lives. Also each clientTemplate will have such namespace.

- After creation of client, there will be a set of corresponding roles added automatically for the client like:
"prefix:/clients/my-client-id/scope-param/phone"
"prefix:/clients/my-client-id/scope-param/address"
"prefix:/clients/my-client-id/scope-param/profile"
"prefix:/clients/my-client-id/scope-param/offline_access"

You won't need to configure "Scope param name" as that will be based on the roleName

- Those roles won't be default roles and won't be automatically assigned to users. However any user, who uses "scope=phone" can request the scope "phone" with corresponding protocolMappers and children composite roles.

- If role is assigned to user directly, it will be always used even if it's not used in scope parameter. So user assigned to role "phone" will automatically receive "phone" and "phone_verified" protocolMappers.

- Keycloak will browse roles from the "scope-param" namespace of particular client, and then also roles from "scope-param" namespace of clientTemplate. So clientTemplate can have namespace for scope parameters, which can be used if there is some "common" scope applicable to more clients with common set of protocolMappers from this clientTemplate.

- For the scopes like "phone" "address" and "profile" it's pretty straightforward. It seems we will need protocolMappers for every OIDC defined claim (address, phone, gender, birthday etc) and those will have role required. So for example protocolMapper for "birthday" will require role "profile" , protocolMapper for "phone" will require role "phone" etc. Basic existing protocolMappers ( name, "preferred_username" etc)  .

- For the "prefix:/clients/my-client-id/scope-param/offline_access" role, it will have one composite role, which is current realm role "offline_access" . The current "offline_access" realm role is default role with scopeParamRequired=true, so it's applied just if it's requested (directly or indirectly) through scope parameter. If client doesn't need offline tokens and "offline_access" scope, the namespaced role "prefix:/clients/my-client-id/scope-param/offline_access". Also if you remove realm role "offline_access" from user, you disallow him to request offline tokens. It's currently realm role, hence it's global setting (so you can't reject user from offline tokens for "clientA" but allow him request offline tokens in "clientB") .

Reading all of this makes me think it would be cleaner to introduce a separate scope concept ;)

A user doesn't have a scope - a user has roles and attributes. Re-using roles concept for the scope just makes it feel awkward and retrofitted.
 


Marek



One more thing is that maybe when a composite role is used on the consent screen we could have an option if the composite role description should be shown rather than the individual roles?

On 30 June 2016 at 15:56, Marek Posolda <mposolda@redhat.com> wrote:
It seems that for OIDC certification, we will need more proper support
for "scope" parameter. There are few tests from OIDC conformance
testsuite, which end with WARNING because of issues with "scope" parameter.


SUMMARY OF SPECS REQUIREMENTS
-----------------------------

- In OIDC specification, the "scope" parameter is actually REQUIRED. And
you must add the scope value "openid" to all authorization requests.
Hence if you don't use "scope=openid", the request is pure OAuth2
request, but it's not OIDC request.

In https://issues.jboss.org/browse/KEYCLOAK-3147 we discuss the
possibility that we should change our adapters and add "scope=openid" to
all requests, and also the possibility to remove IDToken if it's not
OIDC request (and maybe other things). However it may be potential issue
with backward compatibility with older adapters (which don't add
"scope=openid" at all).


- OIDC also prescribes the "scope=offline_access", which you use if you
want offline token. We actually support this as we have realm role
"offline_access", with scopeParamRequired=true . So this role is applied
just if it's included in scope parameter. This is our only support of
scope param actually. ATM we reference the realm roles by name (role
name must match the value of scope parameter) and clientRoles by
"clientId/roleName" . So it's not very flexible and won't work well in
the future with role namespaces.


- OIDC defines four other scope values, which we don't support, with the
meaning like this:

profile
     OPTIONAL. This scope value requests access to the End-User's
default profile Claims, which are: "name", "family_name", "given_name",
"middle_name", "nickname", "preferred_username", "profile", "picture",
"website", "gender", "birthdate", "zoneinfo", "locale", and "updated_at".

email
     OPTIONAL. This scope value requests access to the "email" and
"email_verified" Claims.

address
     OPTIONAL. This scope value requests access to the "address" Claim.

phone
     OPTIONAL. This scope value requests access to the "phone_number"
and "phone_number_verified" Claims.


- Not directly related to scopes, however OIDC also has one parameter
"claims" described in section
http://openid.net/specs/openid-connect-core-1_0.html#ClaimsParameter .
This allows to define some additional claims, which should be included
in IDToken or UserInfo endpoint in addition to claims specified by
"scope" parameter.



HOW TO IMPLEMENT?
-----------------

My current thinking is, that we will have 2 kinds of protocolMappers and
roles.

1) "Always applied" - Those roles/protocolMappers are always applied to
token even if they are not specified by scope parameter.

2) "Applied on demand" - Those roles/protocolMappers are applied just if
they are specifically requested by scope parameter

For roles, we already have that with "scope param required" flag defined
per roleModel. However for protocolMappers we don't have it yet.

IMO We will also need some more flexible way to specify how the value of
scope parameter will be mapped to roles and protocolMappers. For example
if I use "scope=foo", it can mean that I want realm role "foo1", client
role "client1/foo2" and protocolMapper for "firstName" and "lastName" etc.

I can see 2 possibilities:

a) Configure allowed scope param separately per each role / protocolMapper

If some role has "Scope param required" checked, you will have
possibility to configure list of available values of scope parameter,
which this role will be applied to. This will be configured per-each
role separately.

Example: I have realm role "foo" . I check "scope param required" to
true. Then I will define "scope param values" :  "bar" and "baz". It
means that if someone uses parameter "scope=bar" or
scope=baz", then role "foo" will be applied to token. Otherwise it won't
be applied.

Similarly it will be for protocolMappers. We will add switch "Scope
param required" to protocolMappers and we will use list of available
values of scope parameter, which is configured per each protocolMapper
separately.


b) Configure scope parameter in separate place

We will have another tab "Scope parameter config" (or maybe rather
another sub-tab under existing "Scope" tab). Here you will define the
allowed values of scope parameter. For each allowed value, you will
define protocolMappers and roles to apply. Hence for example for
"profile" scope parameter, you will define all protocolMappers for
corresponding claims ( name, family_name, ...) here.

We will still need "scope param required" switch for protocolMappers in
case (b).

My current thinking is to go with (a). So when you go to some role (or
protocolMapper) in admin console you will see if you need scope
parameter and what are available values of scope parameter to request it.

WDYT? Another ideas?


Marek

_______________________________________________
keycloak-dev mailing list
keycloak-dev@lists.jboss.org
https://lists.jboss.org/mailman/listinfo/keycloak-dev