[Date Prev] | [Thread Prev] | [Thread Next] | [Date Next] -- [Date Index] | [Thread Index] | [List Home]
Subject: Re: Policy Labelling (was Re: Delegation and on-permit-apply-second)
Hi Erik, On 27/11/2012 8:43 PM, Erik Rissanen wrote:
Hi Steven, See inline. On 2012-11-05 05:42, Steven Legg wrote:Hi Erik, On 1/11/2012 6:40 PM, Erik Rissanen wrote:Hi Steven, I do think there is some practical value in being able to collect any types of policies into a single policy set, but I agree that the value probably is not that big. In come cases where someone would hand over a policy set "here are all my access and admin policies", they would have to hand over two separate sets of policies instead. And, yes, when I think about it more, the evaluation logic change is small. And there is actually one additional benefit. The explicit marking of a policy as an admin policy is probably simpler conceptually to users than to use weird attribute category names.I agree completely.So, I would agree with you to these changes.Reflecting on what you say here and what you were saying during the TC meeting makes me think you are assuming more than the minimal change required to dispense with category prefixing, in which case there are some technical choices to be discussed. I'll start by describing the minimal change. The reduction procedure, as currently described by the administration profile, begins when the PDP encounters an applicable, untrusted access policy. Let's call it 'A'. A reduction graph is formed by considering each ordered pair of policies in the policy set that contains 'A'. Let's call that policy set 'B'. To even get to the point of evaluating access policy 'A', policy set 'B' must be an access policy set (which is to say, it has a target that can only be satisfied by an access request). For policy 'A' to be authorized at all there must be at least one sibling policy of 'A' that is an administrative policy. Let's call it policy 'C'. So an access policy set like 'B' must be allowed to have children that are either access policies (like 'A') or administrative policies (like 'C'). In the current scheme we distinguish between access policies and administrative policies by whether or not they have attribute designators referring to prefixed categories. In the new scheme we would tell the difference by the element name of the policy, e.g., either <Policy> or <AdministrativePolicy>, and in the process exclude the unhelpful possibility of hybrid policies. It is also possible for a sibling of policy 'B' to be an administrative policy set, which is to say, a policy set with a target that can only be satisfied by an administrative request. We can likewise distinguish between access policy sets and administrative policy sets by the element name, e.g., either <PolicySet> or <AdministrativePolicySet>. Let's say we have an administrative policy set 'D' that is a sibling of policy 'A'. Both policy 'C' and policy set 'D' are considered when reducing policy 'A'. In which case they are evaluated with respect to an administrative request. Policy set 'D' is only ever considered when processing an administrative request, so any access policy or access policy set children of policy set 'D' will never be evaluated. We might as well disallow such children. In summary, what we have for the minimal change is <PolicySet> elements that may have <Policy>, <AdministrativePolicy>, <PolicySet> and <AdministrativePolicySet> children, and <AdministrativePolicySet> elements that may have only <AdministrativePolicy> and <AdministrativePolicySet> children. The fact that an access policy set can have administrative child policies and policy sets is a non-issue for most policy set combining algorithms. In the current scheme the administrative children would always evaluate to NotApplicable when evaluating an access request. In the new scheme the administrative children would be skipped (or automatically NotApplicable) when evaluating an access request. Either way, the result of the combining algorithm is unaffected. The one exception is the on-permit-apply-second combining algorithm because it limits a policy set to exactly two children. If, for an access policy set, those children are both access policies, then in practice they have to be trusted policies because there is no room for any administrative policies to authorize untrusted policies (note this is also the case with the current administration profile). If either child of an access policy is an administrative policy, then the policy set will never be applicable. To make the on-permit-apply-second combining rule more useful in an administrative setting, the restriction to two children should be relaxed to two access children and any number of administrative children when the combining rule is used in an access policy set. The "first" and "second" children should be taken to mean the first and second access children, or we can insist that any administrative children must come after the two access children. What to do in the case of an administrative policy set with the on-permit-apply-second combining rule is not so obvious. If we limit it to two administrative children, then those children would have to be trusted in practice. We could say that the first two children are considered by the combining algorithm and any subsequent children are only used if reduction of untrusted children is required.I think the above "minimal change" seems good. I don't think it is a big problem that the on-permit-apply-second is meaningful only in context of trusted children, since it is intended to solve the issue of a condition in policies, and for that purpose everything work fine.That covers the minimal change. The more substantial change is a complete separation of access and administrative policies and policy sets. That is, a <PolicySet> can only have <Policy> and <PolicySet> children, and an <AdministrativePolicySet> can only have <AdministrativePolicy> and <AdministrativePolicySet> children. An access request would initially be assessed against the collection of access policies and access policy sets. The administrative policies and policy sets would become involved if there are any untrusted access policies or policy sets to be reduced. A revision of the reduction procedure would be necessary and I see two main possibilities.I don't like the above since this introduces a top level construct in the PDP which is different from any of the regular XACML policy constructs.
As opposed to the current scheme that introduces a mid-level construct (the reduction graph built from pair-wise assessment) which is different from the regular XACML policy set evaluation ? > It is a fact that the initial request to the PDP is an access request,
so the access request must match the "top level policy construct" in the PDP.
It still would. > Also, if the administrative
profile is going to be of any use, this top level policy construct also has to also be able to contain administrative policies.
Only if you assume that everything has to be wrapped up in a single, physical
policy set.
An important aspect of the policy labelling solution is that administrative
policies and policy sets are automatically NotApplicable when evaluating
access requests, and access policies and policy sets are automatically
NotApplicable when evaluating administrative requests. Section 7.17 of the
core specification has this to say about request evaluation:
"the PDP is defined by a policy-combining algorithm and a set of policies
and/or policy sets. The PDP SHALL return a response context as if it had
evaluated a single policy set consisting of this policy-combining algorithm
and the set of policies and/or policy sets".
We could assume that this virtual, single policy set is exempt from the
separation of access policy and administrative policy. The children could be
access policies and/or access policy sets and/or administrative policies
and/or administrative policy sets. Some of these children will be automatically
NotApplicable depending on whether it is an access request or an administrative
request that is being evaluated.
Alternatively, we could (with respect to option 2) assume that 7.17 applies to
access request evaluation and write similar text to cover administrative request
evaluation. That is, the PDP is defined by a policy-combining algorithm and a
set of access policies and/or access policy sets, and a policy-combining algorithm
(maybe the same one) and a set of administrative policies and/or administrative
policy sets.
> Thus the top level policy construct is something different from the above policy
constructs which you listed. That adds more specification and implementation cost, compared with a PDP which simply starts with a selected top level policy set, which is semantically no different than any other policy set in XACML.
In the context of option (2), and given that an implementation chooses to restrict the "set of policies and/or policy sets" in 7.17 to exactly one policy set, is it really that difficult to have a selected top level access policy set as the place for the PDP to start for an access request and also a selected top level administrative policy set as the place for the PDP to start for an administrative request ?
The first possibility is that we keep the idea of a reduction graph, only instead of the nodes in the graph being the siblings of the untrusted access policy being reduced, the nodes are the standalone administrative policies and policy sets. In effect, every access policy has every standalone administrative policy or policy set as an implied sibling when forming the reduction graph. We have the choice when reducing an untrusted administrative policy whether to consider the actual siblings of that administrative policy when forming the reduction graph (call this (1.a)), or to form the graph from the standalone administrative policies and policy sets (call this (1.b)), just as we would for untrusted access policies. The second possibility (call this (2)) is to do away with the reduction graph and simply assess an administrative request against the collection of administrative policies and policy sets just as we assess an access request against the collection of access policies and policy sets. If we find an untrusted policy (in either collection), then we form the necessary administrative request and assess it against the collection of administrative policies and policy sets. An effect of options (1.b) and (2) is that in the process of reducing an administrative policy we may encounter that very same administrative policy. If we do, then we treat it as NotApplicable rather than recursively reducing it. To do otherwise would open the possibility of self-authorizing, untrusted policies. Note that this effect doesn't arise with the delegation graph because policies aren't paired with themselves. Also with options (1.b) and (2) we don't need to do anything about the on-permit-apply-second combining rule. Administrative policies no longer need to be siblings of the untrusted policies they authorize, so the on-permit-apply-second combining rule can still insist on having exactly two children. I would very much like us to choose option (2) because working out where to put an administrative policy under the current scheme is a burden policy writers can do without. Take the simple case where User A delegates all rights to User D. This is a very simple administrative policy that basically says "if the delegate is User D, then Permit". The issue is where to put it. User A needs to place it (by copy or by reference) in every policy set where User D will potentially be creating access policies. Furthermore, as new policy sets are created that User D might put policies into, User A will have to put the administrative policy in there as well. Conversely, if User A isn't thorough or proactive in placing the administrative policy, then User D will be limited in the rights he or she can actually exercise. Compounding the difficulties is the fact that in the general case, because of multi-valued attributes and augmentation of the delegate category by a context handler, practially any untrusted policy can be authorized by any administrative policy. It all depends on the request context. It is only by making assumptions about certain attributes being single-valued or by knowing about correlations between the attributes of an access subject that we can begin to predict which untrusted policies will be authorized by which administrative policies in the absence of any specific request context. These are things that users shouldn't have to be concerned with when creating administrative policies. Options (1.b) and (2) mean that they don't have to be. Option (2) is the conceptually simpler of the two, and can simulate (1.b) if desired.I don't agree that option 2 is simpler conceptually. I would say it's the opposite since it introduces another type of policy combining at the top level.
Option (1.b) calls for forming a reduction graph from the top level administrative policies and/or policy sets, which is quite different from what 7.17 requires for access requests. Option (2) is just like 7.17, except the policies or policy sets will be different ones.
Also, being able to mix in admin policies in nested access policies is useful for creating hierarchical structures, where a delegated policy contains nested delegation itself. This sort of recursion is useful for being able to import a full policy set of a sub-organization which internally might want to use delegation.
I'm sceptical about the utility of that. Importing policy from a sub-organization isn't necessarily going to be easy. For example, a user of the sub-organization may have "read" access to everything, which is alright in the scope of the sub-organization, but may not be appropriate in the scope of the parent organization. Access rights will generally need to be reconsidered, which will involve rewriting and reorganizing policy for both the parent organization and the sub-organization, and the administrative policies won't be immune to that. Even if restricting the scope of the sub-organization's administrative policies to the sub-organization's access policies is useful, importing the policy of a sub-organization is a relatively infrequent event. Option (2) is about making delegation easier to use all the time. Keeping the reduction graph so that importing a policy set from a sub-organization is (maybe) easier is making the wrong trade off. Also, with option (2) the parent organization would be importing an access policy set and a separate administrative policy set. If there is a simple way to distinguish the issuers in the sub-organization from issuers in the parent organization (e.g., by the domain name in their email addresses), then this could be added to the target of the administrative policy set, effectively restricting the scope of the sub-organization's administrative policies to the sub-organization's access policies, even though they are separated. Regards, Steven
Best regards, ErikRegards, StevenBest regards, Erik On 2012-11-01 06:14, Steven Legg wrote:Hi Erik, On 31/10/2012 11:35 PM, Erik Rissanen wrote:Hi Steven, Sorry about the slow response to this. As I see it, there are two benefits to the category prefixing: 1. A policy can be mixed admin/access.How is that a particular benefit when any such mixed policy (or policy set) can be rewritten as a pair of a pure access policy and a pure administrative policy, together having exactly the same effect ? A policy that authorizes itself is pointless, and merging an access policy with an administrative policy that authorizes unrelated access policies is weird and confusing, so why do it ?2. We don't need to change the XACML evaluation logic as much. To implement the delegation model, the PDP can use the normal XACML evaluation on admin requests, while building the delegation graph. I think the second is a significant benefit. It would significantly complicate both the administrative spec and the implementation if we need to define additional modes of XACML policy evaluation.The delegation model is already a significant departure from normal XACML evaluation, what with building the delegation graph and forming administrative requests. Those things would still be there. The complexity trade off is between category prefixing (which is not yet complete because it has unresolved issues) and a simple boolean test during request processing to see if the kind of the policy matches the kind of the request (i.e., access versus administrative). That test is simple and efficient compared to category prefixing.What about changing the prefix so it works with all cases?The only thing that comes to mind that would be syntactically valid is a prefix to the scheme name. For example, http://example.com/foobar becomes delegated-http://example.com/foobar. > Or changing the admin category to something elsethan it being based on a prefix?That's difficult without making changes to the core syntax. > By means of some other kind of algorithm, or maybe an explicit table? Labelling the policies would be much easier than a managing a mapping table. Regards, StevenBest regards, Erik On 2012-09-03 07:03, Steven Legg wrote:Hi Erik, On 1/09/2012 1:23 AM, Erik Rissanen wrote:I don't think we change the delegation model in this way since it would mean changing the core schema, which we really don't want to do at this stage.I originally suggested an XML attribute for the label, which would need a change to the core schema, but there is another way to label the policies without changing the core schema and that is substitution groups for Policy and PolicySet. For example: <xs:element name="AdministrativePolicy" type="xacml:PolicyType" substitutionGroup="xacml:Policy"/> <xs:element name="AdministrativePolicySet" type="xacml:PolicySetType" substitutionGroup="xacml:PolicySet"/> In a way this is better than a simple XML attribute because it is more visually distinctive. It would be desirable for AdministrativePolicy and AdministrativePolicySet to be in the same namespace as Policy and PolicySet, but if they have to be in their own namespace so be it. > Also, I think we did consider labeling policies like this in earlystages of the delegation work. I don't remember right now why we did not go that way, but I suspect there was a reason.I've found and reported on a bunch of problems with category prefixing, so there are reasons to reconsider using labelling. If the purpose of the prefixing is just to ensure that administrative policies aren't applicable when evaluating access requests, and vice versa, then labelling is a simpler and cleaner way to achieve the same end. > Probably that the current mechanism is more flexible. It's more flexible in that it allows hybrid policies that are a bit access policy and a bit administrative policy. I can't think of a good reason to want to do that, and in any case, it is always possible to split a hybrid policy into a pure access policy and a pure administrative policy. I would happily forego this dubious flexibility to get rid of category prefixing and its problems. Regards, Steven
[Date Prev] | [Thread Prev] | [Thread Next] | [Date Next] -- [Date Index] | [Thread Index] | [List Home]