On 06/11/2015 05:35 PM, Salvatore Orlando wrote:
I am not able to say whether this works for Nova. Surely works for
Neutron - from a functional perspective at least.
I still don't know however whether this choice is the best way to
proceed, and perhaps you can help me understand better.
Role checks are always expressed through policy.json and can be
enforced in middleware. Does this mean that there is also a
centralized policy.json, or will we keep per-project policy files even
for role checks?
The focus of the centralized policy has always been the role
management. I don't propose changing course on that. We want the role
names to be consistant across all the projects.
I see the starting lapproach being that users get one of two roles by
default; Administrator or Member. While these will still be scoped to
the project, only the role assignment itself would be checked by the
dynamic policy. The scope would be checked by static policy.
Scope checks - ie: application-specific checks - can be enforced in
any way the application developers wish. They can use policy.json, be
hardcoded or, if they wish ask Pythia, the Oracle of Delphi. From an
operator perspective, this means that every project can enforce
policies in a different way. Is this going to be practical and
maintainable? I can't speak for operators, but I would like to
understand a bit better what this implies for them.
If the project already has logic doing the policy check using oslo, and
they want to keep using oslo, they can do so. Relaly,only Keystone has
any deep need to do so. The rules that I saw in Nova were simple enough
that they could be checked in oslo or with simple python code. The
limiting thing here is that the object in the database needs to be
checked against the scope anyway; it doesn't matter if the scope on my
token matches the scope on my roject as specified in the URL, if the
object fetched back does not belong to that project. How this check is
done differes from project to project, it is just essential that the
check be done.
For a project that is not doing the check, and needs to start doing so,
it probably makes sense to use oslo.policy.
Thanks for the questions
Salvatore
On 11 June 2015 at 17:47, Adam Young <ayo...@redhat.com
<mailto:ayo...@redhat.com>> wrote:
Sean had a really good point when he mentioned that the Developers
know what need to be enforced, and I think this is why he
suggested that the base policy implementation be in Python code,
not the policy JSON DSL.
The main thrust of the dynamic policy has been to get the
role-to-api assignment more flexible. However, there is another
side to each policy rule; figureing out where the project (nee'
tenant) id is in the request; is it part of the URL, part of the
request body, or in the object returned from the database. This
part really should be handled by the developer working on the
policy rule, and it should not be changed.
So...what if we say that we split policy into two checks; a role
check, and a scope check. Both checks must pass in order for the
user to get access to the API. The Scope check is not going to be
dynamic; once set, they will pretty much stay set. It might be
done using the policy.json, or done in code, but it will be
separate from the role check.
The Neutron policy checks for things like
|"shared": "field:networks:shared=True", "shared_firewalls":
"field:firewalls:shared=True", "shared_firewall_policies":
"field:firewall_policies:shared=True", "shared_subnetpools":
"field:subnetpools:shared=True",
Would be handled by the dev teams later policy check; anything
that requires actually fetching the object from the database is
postponed to this stage.
|
The role check will come from the policy.json file. This will
allow the operator to fine tune how roles are handled. Any thing
else that can be explicitly checked based on the token will be
fair game, but not API specific values; no database fetch will be
performed at this point. The assumption is that this policy check
could be generic enough to be performed in middleware, and might
even be enforced based on the URL instead of the pseudo random
namespacing we do now.
Does this suggestion work for Nova? I think it will make the
overall policy much easier to maintain in the field.
__________________________________________________________________________
OpenStack Development Mailing List (not for usage questions)
Unsubscribe:
openstack-dev-requ...@lists.openstack.org?subject:unsubscribe
<http://openstack-dev-requ...@lists.openstack.org?subject:unsubscribe>
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
__________________________________________________________________________
OpenStack Development Mailing List (not for usage questions)
Unsubscribe: openstack-dev-requ...@lists.openstack.org?subject:unsubscribe
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
__________________________________________________________________________
OpenStack Development Mailing List (not for usage questions)
Unsubscribe: openstack-dev-requ...@lists.openstack.org?subject:unsubscribe
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev