On 12/01/2014 09:09 AM, Doug Hellmann wrote:
On Nov 30, 2014, at 8:51 PM, Jamie Lennox <jamielen...@redhat.com> wrote:
TL;DR: I think we can handle most of oslo.context with some additions to
auth_token middleware and simplify policy enforcement (from a service
perspective) at the same time.
There is currently a push to release oslo.context as a
library, for reference:
https://github.com/openstack/oslo.context/blob/master/oslo_context/context.py
Whilst I love the intent to standardize this
functionality I think that many of the requirements in there
are incorrect and don't apply to all services. It is my
understanding for example that read_only, show_deleted are
essentially nova requirements, and the use of is_admin needs
to be killed off, not standardized.
Currently each service builds a context based on headers
made available from auth_token middleware and some
additional interpretations based on that user
authentication. Each service does this slightly differently
based on its needs/when it copied it from nova.
I propose that auth_token middleware essentially handle the
creation and management of an authentication object that
will be passed and used by all services. This will
standardize so much of the oslo.context library that I'm not
sure it will be still needed. I bring this up now as I am
wanting to push this way and don't want to change things
after everyone has adopted oslo.context.
We put the context class in its own library because both oslo.messaging and
oslo.log [1] need to have input into its API. If the middleware wants to get
involved in adding to the API that’s fine, but context is not only used for
authentication so the middleware can’t own it.
I think oslo context is mixing in the auth/access info with the rest of
the request. In policy, the acces info is on the left, and the target is
on the right. The target should be exposed by the end service, but the
access info comes from the token.
I took a swag at what this should really look like from a keystone
perspective. I started with a core domain model in pure python objects:
Spec is here.
https://review.openstack.org/#/c/135774/
Code is here.
https://review.openstack.org/#/c/137231/
However, I am going to move that code to the keystoneclient repository.
It will be populated from the token JSON and available for consumption
by keystonemiddleware.auth_token, policy (whereever that ends up living)
as well as any other code that wants it.
Policy enforcement should not make use of the Oslo Context code. It
will be a complete, self contained library. Instead, each of the
services will start with the access info from KC and then provide their
own data for the target side of policy.
We have a chain of specs. I provided an overview of where policy
enforcement is going here:
https://adam.younglogic.com/2014/11/dynamic-policy-in-keystone/
[1] https://review.openstack.org/132551
The current release of auth_token middleware creates and
passes to services (via env['keystone.token_auth']) an auth
plugin that can be passed to clients to use the current user
authentication. My intention here is to expand that object
to expose all of the authentication information required for
the services to operate.
There are two components to context that I can see:
- The current authentication information that is retrieved
from auth_token middleware.
- service specific context added based on that user
information eg read_only, show_deleted, is_admin,
resource_id
Regarding the first point of current authentication there
are three places I can see this used:
- communicating with other services as that user
- associating resources with a user/project
- policy enforcement
One of the specific logging requests we’ve had from operators is to have the
logs show the authentication context clearly and consistently (i.e., the same
format whether domains are used in the deployment or not). That’s an aspects of
the spec linked above.
Addressing each of the 'current authentication' needs:
- As mentioned for service to service communication
auth_token middleware already provides an auth_plugin
that can be used with (at this point most) of the
clients. This greatly simplifies reusing an existing
token and correctly using the service catalog as each
client would do this differently. In future this plugin
will be extended to provide support for concepts such as
filling in the X-Service-Token [1] on behalf of the
service, managing the request id, and generally
standardizing service->service communication without
requiring explicit support from every project and client.
- Given that this authentication plugin is built within
auth_token middleware it is a fairly trivial step to
provide public properties on this object to give access
to the current user_id, project_id and other relevant
authentication data that the services can access. This is
fairly well handled today but it means it is done without
the service having to fetch all these objects from
headers.
That sounds like a good source of data to populate the context object.
- With upcoming changes to policy to handle features such
as the X-Service-Token the existing context will need to
gain a bunch of new entries. With the keystone team
looking to wrap policy enforcement into its own
standalone library it makes more sense to provide this
authentication object directly to policy enforcement.
This will allow the keystone team to manipulate policy
data from both auth_token and the enforcement side,
letting us introduce new features to policy transparent
to the services. It will also standardize the naming of
variables within these policy files.
What is left for a context object after this is managing
serialization and deserialization of this auth object and
any additional fields (read_only etc) that are generally
calculated at context creation time. This would be a very
small library.
That’s not all it will do, but it will be small. As I mentioned above, we
isolated it in its own library to control dependencies because several aspects
of the system want to add to the API.
There are still a number of steps to getting there:
- Adding enough data to the existing authentication plugin
to allow policy enforcement and general usage.
- Making the authentication object serializable for
transmitting between services.
- Extracting policy enforcement into a library.
However I think that this approach brings enough benefits to
hold off on releasing and standardizing the use of the
current context objects.
I'd love to hear everyone thoughts on this, and where it
would fall down. I see there could be some issues with how
the context would fit into nova's versioned objects for
example - but I think this would be the same issues that an
oslo.context library would face anyway.
Jamie
[1] This is where service->service communication includes
the service token as well as the user token to allow smarter
policy and resource access. For example, a user can't access
certain neutron functions directly however it should be
allowed when nova calls neutron on behalf of a user, or an
object that a service made on behalf of a user can only be
deleted when the service makes the request on behalf of that
user.
_______________________________________________
OpenStack-dev mailing list
OpenStack-dev@lists.openstack.org
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
_______________________________________________
OpenStack-dev mailing list
OpenStack-dev@lists.openstack.org
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
_______________________________________________
OpenStack-dev mailing list
OpenStack-dev@lists.openstack.org
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev