[I2nsf] 答复: Service Layer Policies - Post 1: Background Information on PCIM

["Xialiang (Frank)" <[e-mail]>]

Hi John,
Since current design principle for the policy model of capability interface is to abstract from most common existing NSF security policy implementations and consolidate them into a general and extensive model, I think imperative logic is more suitable for this goal.

I am very interested in whether “event” can be used in I2NSF capability interface. Do you have some suggestions?

Are the “subject” and “target” two different classes in PCIMe? Do you suggest to include them in I2NSF policy model of capability interface for the goal of authorizing the policy?

B.R.
Frank

发件人: I2nsf [mailto:[e-mail]] 代表 John Strassner
发送时间: 2015年12月12日 7:32
收件人: Linda Dunbar
抄送: [e-mail]
主题: Re: [I2nsf] Service Layer Policies - Post 1: Background Information on PCIM

Hi Linda,

thank you for your kind words. My replies are inline:

Thank you very much for the in-depth analysis of PCIM and its suitability as something that I2NSF should be based on. What you write is very valuable. I think the analysis should be added to the I2NSF gap analysis, (or be a standalone Policy Analysis ID).

<jcs>
If desired, I could summarize the key points for the gap analysis. I would be willing to do a Policy Analysis I-D, but I think that this probably belongs in SUPA. However, feedback from I2NSF is invaluable for such a document.
</jcs>


Note: It is possible to define limited constraints in PCIM as conditions. For example, the VendorPolicyCondition class has two attributes, Constraint and ConstraintEncoding, that do this. This will be discussed more in Post 4.

Note: Context is NOT defined in the PCIM.

[Linda]  Do you agree that “Condition of the PCIM” == “I2NSF subject” & “the Contextual states (or the I2NSF object)”?

<jcs>
PCIM/PCIMe conditions are clearly similar to what is now called "I2NSF Subject". However, the contextual state is trickier. Part of the problem is how we want to define Context. I'll save the bulk for Post 4 (as outlined in my Post 0 note). However, looking at page 6 of draft-merged-i2nsf-framework-04:
"Object-Match values, based on context (e.g., state, direction of the traffic, time, geo-location, etc.)", in which case:
   - state:  likely NOT part of the condition; rather, it applies to the entire PolicyRule as a whole
   - traffic direction:  part of the condition
   - time:  part of the condition
   - geo-location:  part of the condition
</jcs>

...

Critical Problems:

There is no unique ID that can be used to distinguish different
instances of the same PolicyRule, PolicyGroup, PolicyCondition,
or PolicyAction.
[Linda] Can we simply create an ID (string or numeric) for instances of the same PolicyRule? Or it is more complicated?

<jcs>
SUPA and DEN-ng use two attributes to do this:

   - a string that contains the content of the identifier
   - a string that is an enum, which defines how to interpret the content

This allows different IDs to be mixed (e.g., URIs, GUIDs, keys, FQDNs, etc.)

That being said, it is certainly easier to pick one and go with it, but since I2NSF SHOULD interoperate with other models easily, I would recommend a scheme similar to the one I used (or the above, of course).
</jcs>


Relationships are implemented by defining an abstract class
with keys, and then overriding the keys in the subclasses of the
relationship. This is broken. Would you override keys in an RDBMS?
[Linda] Is the “Key” to differentiating different “PolicyRule”?   Can you give an example of “subclasses of the relationship”?
What events can cause “overriding the keys in the subclasses of the relationship”?

<jcs>
No, the key (as in primary and foreign keys, just to be clear) is used as part of how CIM data models are built. I believe that this strategy is fundamentally incorrect (as in, would you override keys in an RDBMS). It also violates a host of object-oriented principles (see Post 2b).

Remember, PCIM and PCIMe do NOT define events. This is a design (compile-time) definition of how you build CIM classes.
</jcs>


•  Most of the classes that are used to implement the relationships
do NOT have attributes, so it is impossible to restrict which
classes are related to which other classes (e.g., by context).
Put another way, the same relationship must always be used
between two types, even if the semantics varies. (The two
exceptions are PolicyConditionInPolicyRule and
PolicyActionInPolicyRule).
[Linda] Does “classes” == “PolicyRule”?

<jcs>
No, "classes" == {association | aggregation | composition}.

This is another artifact of CIM. They implement ALL relationships as classes. In UML, there is a difference between these three relationships and realizing them as a class; the CIM approach eliminates this distinction.
</jcs>



  *   The "ignore this unless you need this" philosophy does not
work. How do you know if other implementations that you are
supposed to interoperate with are using them?
[Linda] Is “Ignore this unless you need this” an new type of “PolicyRule”?

<jcs>
No, sorry, this is one of the key points underlying the design philosophy of CIM, and carried into PCIM and PCIMe. Basically, it means that it is OK to define model elements on another model element if they are optional. Examples:

   - defining an optional attribute, method, or relationship on a class
   - defining an optional constraint

This violates all of the 4 design principles in Post 2b (references section)
</jcs>


There is no concept of the subject or the target of a policy
[Linda] Does your “subject” have the same meaning as the “I2NSF subject” which means bits&bytes extracted from packets?

<jcs>
No. A PolicySubject is the author of a policy. One could also differentiate between author, owner, and executor, but that is more complex. It is vital for deontic logic (i.e., the logic of obligations and permissions, as in MUST, MUST NOT, SHOULD, MAY, etc.). This is one of my concerns with the Sbbject-Object-Action-Function model - one cannot build authorization policies that can be logically proven to be correct or not without a PolicySubject.
</jcs>


Adding a new type of policy (e.g., declarative or functional)
will cause a major rewrite of the entire model

[Linda] Can we say that  I2NSF policies have one type “Subject-ContextualStates- action- function? So in I2NSF environment, there is only the notion of adding a new “PolicyRule”, but no adding a new policy type. Is this correct statement?

<jcs>
No. The meaning of what I wrote above is much broader. If you think of ECA policies as imperative in nature, then the analogy is that PCIM/PCIMe are hard-wired to imperative policies. The schema cannot be used for other forms (declarative, functional, etc.) of policies.

Put another way, ECA (or CA, which is all that the PCIM/PCIMe can do) is imperative - it specifies that, given this particular state, the ONLY state that you are allowed to transition to is this other state. Rationality MUST be compiled into the policy. In contrast, for declarative logic, given a particular state, there are a SET of states that either satisfy the goal, or take you closer to the goal. Hence, rationality is GENERATED by a logic planner/optimizer.
</jcs>


There is no approach to detecting and solving policy conflicts;
this is due to the simplistic notion of just using a PDP and a
PEP for applying policies.

[Linda] in I2NSF, there could be a lot of conflicts, for example, there could be two rules entered at different time:

1.       Virtual Network X can’t access “YouTube”,

2.      10.1.1.1 can access “YouTube” (where 10.1.1.1 is a host within the Virtual Network X).

Do you call the 2 rules above conflict?
<jcs>
Bob already answered this in http://www.ietf.org/mail-archive/web/i2nsf/current/msg00729.html
+1 to his answer (or see my reply)
</jcs>


Given the large amount of problems, I recommend that we not consider PCIM further.
[Linda] should I2NSF reference PCIM?

<jcs>
I don't see any harm in **referencing** PCIM or PCIMe; I just don't want to **use** either to construct our policies.
</jcs>


best regards,
John



On Tue, Dec 8, 2015 at 1:10 PM, Linda Dunbar <[e-mail]<mailto:[e-mail]>> wrote:
John,

Thank you very much for the in-depth analysis of PCIM and its suitability as something that I2NSF should be based on. What you write is very valuable. I think the analysis should be added to the I2NSF gap analysis, (or be a standalone Policy Analysis ID).

I have some questions inserted below:

From: I2nsf [mailto:[e-mail]<mailto:[e-mail]>] On Behalf Of John Strassner
Sent: Sunday, December 06, 2015 7:01 PM
To: [e-mail]<mailto:[e-mail]>
Subject: [I2nsf] Service Layer Policies - Post 1: Background Information on PCIM

<Snipped>

2.  Brief description of PCIM

The PCIM Class Hierarchy is very simple, as this is a CA (condition-action) model. Events are NOT represented. PCIM defines a PolicyRule, a PolicyGroup (group of PolicyRules), PolicyCondition, and PolicyAction. These are all siblings; their common superclass is called Policy, as shown below:

ManagedElement (abstract)
  |
  +---Policy (abstract)
  |     |
  |     +---PolicyGroup
  |     |
  |     +---PolicyRule
  |     |
  |     +---PolicyCondition (abstract)
  |     |      |
  |     |      +---PolicyTimePeriodCondition
  |     |      |
  |     |      +---VendorPolicyCondition
  |     |
  |     +---PolicyAction (abstract)
  |            |
  |            +---VendorPolicyAction

Note: PCIM does have a simple notion of Roles, but implements them as attributes, instead of objects. This is not scalable, and does not fully realize the power of using roles. This will be discussed more in Post 3.

Note: It is possible to define limited constraints in PCIM as conditions. For example, the VendorPolicyCondition class has two attributes, Constraint and ConstraintEncoding, that do this. This will be discussed more in Post 4.

Note: Context is NOT defined in the PCIM.

[Linda]  Do you agree that “Condition of the PCIM” == “I2NSF subject” & “the Contextual states (or the I2NSF object)”?


3.  Summary: main drawbacks of using PCIM

Note that the current CIM Policy Model (2.44.1) is significantly different than either PCIM or PCIMe. PCIM is based on CIM 2.4; PCIMe is based on CIM 2.5. PCIMe is NOT backwards compatible with PCIM in all areas.

Critical Problems:

  *   There is no unique ID that can be used to distinguish different
instances of the same PolicyRule, PolicyGroup, PolicyCondition,
or PolicyAction.
[Linda] Can we simply create an ID (string or numeric) for instances of the same PolicyRule? Or it is more complicated?

  *   Relationships are implemented by defining an abstract class
with keys, and then overriding the keys in the subclasses of the
relationship. This is broken. Would you override keys in an RDBMS?
[Linda] Is the “Key” to differentiating different “PolicyRule”?   Can you give an example of “subclasses of the relationship”?
What events can cause “overriding the keys in the subclasses of the relationship”?

  *   Most of the classes that are used to implement the relationships
do NOT have attributes, so it is impossible to restrict which
classes are related to which other classes (e.g., by context).
Put another way, the same relationship must always be used
between two types, even if the semantics varies. (The two
exceptions are PolicyConditionInPolicyRule and
PolicyActionInPolicyRule).
[Linda] Does “classes” == “PolicyRule”?

  *   The "ignore this unless you need this" philosophy does not
work. How do you know if other implementations that you are
supposed to interoperate with are using them?
[Linda] Is “Ignore this unless you need this” an new type of “PolicyRule”?

  *   There is no concept of error or exception handling
  *   There is no concept of the subject or the target of a policy
[Linda] Does your “subject” have the same meaning as the “I2NSF subject” which means bits&bytes extracted from packets?

  *   Adding a new type of policy (e.g., declarative or functional)
will cause a major rewrite of the entire model

[Linda] Can we say that  I2NSF policies have one type “Subject-ContextualStates- action- function? So in I2NSF environment, there is only the notion of adding a new “PolicyRule”, but no adding a new policy type. Is this correct statement?

  *   There is no approach to detecting and solving policy conflicts;
this is due to the simplistic notion of just using a PDP and a
PEP for applying policies.

[Linda] in I2NSF, there could be a lot of conflicts, for example, there could be two rules entered at different time:

1.       Virtual Network X can’t access “YouTube”,

2.      10.1.1.1 can access “YouTube” (where 10.1.1.1 is a host within the Virtual Network X).

Do you call the 2 rules above conflict?



4. Recommendations

Given the large amount of problems, I recommend that we not consider PCIM further.
[Linda] should I2NSF reference PCIM?

Thank you very much. Linda

5. Origin of PCIM/PCIMe

This is a **theory** section; please skip unless you want to understand the model in more detail.

The PCIM is based on the DMTF CIM, which is actually a **data model** (despite its name) because it uses relational technologies (e.g., keys and weak references) to create its model. PCIM is based on an early version of the CIM (2.4); the current version of the CIM is 2.44.1 as of this writing. Since this is tied to the CIM, this means that:

   1) Policy is subclassed from a more generic class (ManagedElement)
   2) ALL relationships (associations, aggregations, and compositions)
       are realized as **classes**
   3) As a result of 2), relationships are subclassed
   4) Both attributes and associations are **inherited**
   5) The same design approach (as much as possible) was used in
        PCIM/PCIMe as was in CIM. Specifically, it was thought better
        to define a generic, but optional, relationship instead of not
        defining the relationship.
   6) The CIM does NOT have a single rooted hierarchy; this makes
        code generation more complex
   7) The CIM uses a proprietary language and is not strictly UML
        compliant, even though it looks like UML (unfortunately, this
        language is also called MOF)

Depending on the complexity of the implementation, the above have been shown to produce problems in large environments. For example, the spirit of "let's define an attribute or association in case it is useful, and make it optional so you don't have to worry about it" not only violates some well-known software architectural guidelines (e.g., Single Responsibility Principle and the Liskov Substitution Principle), but it makes things difficult to interoperate with (since you never really know if an attribute or relationship should be instantiated or not).


regards,
John



--
regards,
John