Re: [xacml] Open issue 66

["Rich.Levinson" <rich.levinson@oracle.com>]

Hi Erik,

Answers inline:

Erik Rissanen wrote:
> Hi Rich,
>
> Yes, to respond to this specific "part 1" only, it is true that in 
> this particular example it works very well. The policy is simple 
> enough that there is a single attribute which can be returned by the 
> PDP. If the PEP would provide the missing attribute, then the response 
> would be a Permit.
>
> (BTW, if you prefer to discuss resource attributes, that is good with 
> me. I agree that they are perhaps more interesting since they might be 
> more difficult to handle in practice.)
>
> If I continue beyond "Part 1" into the discussion which you follow up 
> with.
>
> 1. For the sake of argument, let's assume that the PDP can return one 
> or more attributes as "missing".
>
> 2. Then let's assume that, as you say, the policy author puts an 
> attribute in the policy, and the PEP has not provided the attribute to 
> the PDP. Let's say that the attribute has id 
> "http://www.tts-esss.com/pieka/UUUU/3";.
>
> 3. The PEP will get a response saying that attribute 
> http://www.tts-esss.com/pieka/UUUU/3 is missing.
>
> So your question is what should the PEP do? (If I understand you 
> correctly)
That's not really my question. As was already discussed above and in
my prev email, it appears we agree that 7.15.3 says what a PEP can do:
    either ignore it and just reject the request,
    or if it can get the attr, get it, and resubmit the request.
PEP does not "have to do" either of these in particular, but
also "can do" either it chooses. Possibly there is some 3rd
or other choice, but I don't think we need to go there.
>
> It is quite clear to me that if the PEP does not know about this 
> attribute in some form, either directly by itself, or it can look it 
> up in some kind of discovery service, then there is nothing it can do. 
> One cannot expect systems integration to happen by itself. One cannot 
> expect that a policy author can fill in any attribute in a policy and 
> then it would work all by itself.
Right, if PEP does not know what to do, then it should just
reject the request.
>
> So my question then is, what do you think should happen in this stage? 
> Are you looking for a form of general purpose attribute discovery 
> service? That might be an interesting thing to work on.
Yes, I am looking for at least "a form of" somewhat general purpose 
attribute
gathering or collecting, but not "discovery" in the sense of blindly 
going and trying
to find things somewhere.
The "form" which I mentioned in the last sentence of the original issue
    "mechanisms such as vocabularies should be recommended usage here
     with the PEP being responsible for mapping the vocabulary item to the
     particular resource physical access path such as the xpath."
was a suggestion for xml type resources, such as the xml patient record
shown in the core spec.
Let me preface this by saying that this whole "fine grain authorization" 
appears
to me to be a general attempt to abstract security logic out of applications
and into common policy stores and engines. I mean that is the sort of thing
I "hear" a lot and I assume that is a significant part of what this 
whole effort
is all about.
That being said, the actual act of extracting authorization logic out of 
appls
means that some of the tests that they currently do on internal variables to
come to decisions needs to be removed from the code and made available
to be placed in policies that do similar types of tests to come to 
decisions,
but these tests are conducted under the control of the IT Security 
organization
in some central location with an admin console, etc.
So, and I think this is obvious, but maybe, in general people aren't 
thinking
about it enough for it to be obvious, the logic that appl executed and the
attribute values it used in that eval now must be in the policy engine. 
It is
easy to see how the logic can be put there with all the xacml constructs
we have available, but the question remains how do the attribute "values"
get into the policy evaluation.
Again, the answer to this appears fairly obvious to me. In the process of
removing the authorization code from the appl, the devs will notice which
values from the internal resource representation are used by that logic.
Therefore, they must know that when the logic is put somewhere else,
it will need these values to do the evaluation. Therefore, one fairly simple
solution is that the dev identify those object properties, if you will, in a
global sense that can be extracted to a meta-data file, a fairly common
type thing done these days. Maybe it is done w annotations.
Now when the dev replaces the az logic with a simple call for az,
this is where the details need to emerge. (I think the OASIS SCA-*
activities involve quite a bit of abstraction of previously internal
application-based logic into general container framework services
and suspect that this activity I am describing fits fairly naturally
in that sense, except in this case the container service we are
looking at would be the pep-pdp service path.)
First, I would expect that security admins responsible for creating policies
and such would have the list of attrs that are available for such use from
the metadata and doc provided by the appl vendors in the case of
appls that are sold for use by many customers. The vendors should
have a good sense of what attrs might be needed for policies. For
example in the health record case in the core spec, at least all the
attrs currently used in the policies and maybe more depending on
on the nature of the records.
Generally, it would not make sense to send all the possible attrs on
every call, so, when a call is made and a policy evaluated, the list
of missing attrs can be returned, and then the appl az api would
somehow or other get the attrs from the object that it has locally
in context. The example I gave was an xml doc, where to get
the attrs the api might need xpaths and an attr-id. Then the appl
devs could change the formats of the xml docs on their next
release and their metadata would have new xpaths but the
variable id would remain the same.
>
> Maybe I don't understand the use case really. But as far as I see it 
> right now, it's not going to work because:
>
> 1. In general it is not possible to tell which attribute is missing.
>
> 2. Even if it is possible to tell which attribute is missing, one 
> cannot expect a PEP find the missing attribute by itself if it doesn't 
> already know about it.
>
> There just is no way around the fact that if an attribute appears in a 
> policy, then either the PEP or the context handler has to be aware of 
> the attribute in some form, or it won't work. The expected requests 
> from the PEP and the context handler configuration form a contract 
> which defines the vocabulary available to policy authors.
Previous text should answer these points.
>
> But maybe you are talking about lazy fetching of attributes? I do 
> think that there is much value in lazy fetching of attributes. Let's 
> say that the http://www.tts-esss.com/pieka/UUUU/3 attribute is 
> available in some external source and the PEP or context handler is 
> aware of it, but the attribute is not needed every time and it is 
> expensive to retrieve. In this case it may be desirable to set up the 
> system so that the attribute is retrieved only when it is needed by 
> the policies.
No, again, see above.
The point about the context handler, though, is that in the core spec the
context handler is kind of a pdp-local entity which access attrs from a
pip. What I am describing above is getting attrs from an instantiated
business object in an application context. While such data could be
created and made available in an auxiliary pip, this seems like in general
an unnecessary exercise when the info is already right there in the appl
from which the original request was made.
>
> This use case is supported by the context handler abstraction. The 
> context handler is something which can be used by the PDP to retrieve 
> attributes when they are needed on a particular evaluation path, so 
> that early fetching is avoided. But "missing attributes" to the PEP is 
> not going to work.
>
> Anyway, I have a feeling that I don't understand what you are looking 
> for, so I welcome more discussion. :-)
Ok, I provided "more discussion" above. I am interested in your reaction
and what, if any, alternative approaches might be better. My sense is that
the approach I described is fairly "obvious" and I think if we look at
activities in the WS-SecurityPolicy arena there is all kinds of information
being solicited for use in authentication and authorization. WS-Fed is
incorporating provisions for "claims", which afaict are simply attrs w the
rubber stamp of an STS on them, or at least on the wire not a lot more
complicated than that.

    Thanks,
    Rich
>
> Best regards,
> Erik
>
> Rich.Levinson wrote:
>> Hi Erik,
>>
>> I don't think your comments are addressing the specifics that were
>> raised in the issue, which I am copying below for quick reference.
>>
>> To summarize the issue:
>>
>>  Part 1. is basically saying that we need some context about the example
>>     in order to look at the details in a consistent manner. 
>> Basically, my reading
>>     of the example is that when rule 1 is evaluated, the Target 
>> determines that
>>     the Rule is applicable, and presumably the Rule is evaluated and 
>> it should
>>     produce an Indeterminate result because there is no Subject 
>> "patient-number"
>>     as required on line 1134 when the Condition is evaluated.
>>
>>     Based on this Rule I expect that the Indeterminate result causes 
>> us to
>>     look at the prescribed processing in section 7.15.3, which says that
>>     a missing-attribute status code could/should be returned.
>>
>>     To me, this provides an excellent example of how a PEP is informed
>>     of what attribute needs to be supplied.
>>
>> Before I go further into the details, I'd like to get your response 
>> to that.
>>
>> Also, I'd like to note that while I realize that this issue has been 
>> floating
>> there for a long time, one reason is that no one has really wanted to
>> discuss it in detail. Prior responses were more of the nature of there
>> will be some other mechanism for getting the attribute. In general, when
>> we are considering attributes that need to come from users or from
>> resource attributes, I believe there is a non-trivial issue about how
>> those attributes get brought into the picture. The first problem is
>> knowing that they are needed. The fact that they are needed is
>> driven by the existence of an applicable policy that needs it, which
>> clearly is a state that can come and go as an administrator adds or
>> removes a particular attribute from the policy.
>>
>> So, the basic use case here is: an admin creates or changes a policy
>> and decides to add an attr as part of the policy eval. Given this
>> change, how does this attr become part of the request.
>>
>>  (Note: I'd actually prefer to focus on resource attrs rather than
>>     subject attrs, but since this particular situation w Rule 1
>>     appeared obvious to me as a typical example, it happened
>>     that the missing attr was subject-based (there is also an
>>     assoc resource attr, but the req mysteriously already had
>>     supplied that one)
>>
>> My reading of the spec is that 7.15.3 provides a general purpose
>> mechanism for accomplishing this by returning an identifier specifying
>> what attr is needed, based on which the PEP and appl can do
>> whatever they want to get the attr then resubmit the request with
>> the attribute included.
>>
>> This seems to me to be pretty essential functionality. Let me know
>> if you think I am missing some important point here. I will be happy
>> to help "fix" the spec w the detail I think is needed, but before we
>> get to that point there needs to be some agreement that changes
>> are even appropriate.
>>
>>     Thanks,
>>     Rich
>>
>>
>>
>>
>>
>>       66. Missing attributes may be underspecified
>>
>> I did a somewhat detailed analysis of "Example two" in the core spec 
>> from the point of view of understanding how fine grained 
>> authorization (fga) (applying resource attrs to az decision) was 
>> implemented and came across a number of items that I think need to be 
>> addressed especially in potential interoperability situations. I will 
>> put all in one issue initially, we can decide if it needs to be 
>> broken out later.
>>
>> 1. line 1090-91 describing ResourceContent </xacml/ResourceContent>. 
>> In both the core spec and the sample messages, the ResourceContent 
>> </xacml/ResourceContent> contains the following:
>>
>>    *
>>
>>       <ResourceContent </xacml/ResourceContent>>
>>
>>          o
>>
>>             <md:record xmlns:md="urn:med:example:schemas:record"
>>
>>                 + xsi:schemaLocation="urn:med:example:schemas:record
>>
>>                   http:www.med.example.com/schemas/record.xsd">
>>                   <http:www.med.example.com/schemas/record.xsd%22%3E>
>>
>>                +
>>
>>                   <md:patient>
>>
>>                      #
>>
>>                         <md:patientDoB>1992-03-21</md:patientDoB>
>>                         <md:patient-number>555555</md:patient-number>
>>
>>                   </md:patient>
>>
>>             </md:record>
>>
>>       </ResourceContent </xacml/IssuesList/ResourceContent>>
>>       <Attribute AttributeId
>>       
>> </xacml/AttributeId>="urn:oasis:names:tc:xacml:1.0:resource:resource-id"
>>       DataType </xacml/DataType>="xs:string">
>>
>>          o
>>
>>             <AttributeValue </xacml/AttributeValue>>
>>
>>                 + //med.example.com/records/bart-simpson.xml#
>>
>>                   xmlns(md=http:www.med.example.com/schemas/record.xsd)
>>                   xpointer(/md:record/md:patient/md:patientDoB)
>>
>>             </AttributeValue </xacml/IssuesList/AttributeValue>>
>>
>>       </Attribute>
>>
>> While I recognize that the example itself is not intended to be 
>> perfect, it provides a convenient context for raising the following 
>> questions/issues, especially wrt fga.
>>
>>    1. If this is a first request from a PEP, why is the PEP supplying
>>       patient-number on line 1056? This looks like a required attr to
>>       evaluate Rule 1 (line 1141), if the requestor is the patient,
>>       but this example the requestor is the physician.
>>           * The physician-id is supplied in the request (line 1044),
>>             but the only rule it appears in is rule 3 (line 1522).
>>             This rule only allows "write" access (line 1507), so I
>>             expect this request would probably fail as it is currently
>>             set up. i.e. we would need to add a "read" action to rule
>>             3 or add a physician-id test to rule 1.
>>       b. Assuming the above request fails, let's consider what might
>>       be done. There was a "read" request issued (line 1072), so that
>>       would mean that rule 1 (line 1182), rule 2 (line 1347), or rule
>>       4 (line 1668) could be applied.
>>
>> Rule 1 requires a Subject attribute patient-number (line 1134) to 
>> match the patient-number in the requested resource record (line 
>> 1141). Presumably a <MissingAttributeDetail 
>> </xacml/MissingAttributeDetail>> could be returned, somehow 
>> identifying these 2 attributes to the PEP.
>>
>> Rule 2 requires a patientDoB resource attr (line 1297), a 
>> parent-guardian-id subject attr (line 1363), and a parentGuardianId 
>> resource attr (line 1371). Similarly a <MissingAttributeDetail 
>> </xacml/MissingAttributeDetail>> could be returned requesting these.
>>
>>     * Assuming this to be the case, one question I have is how does the
>>
>>       <MissingAttributesDetail </xacml/MissingAttributesDetail>> tell
>>       the PEP whether the attributes that are missing should be
>>       resubmitted as part of the Subject or as part of the Resource?
>>       This info is provided in the Request from the xml structure,
>>       however, the <MissingAttributeDetail
>>       </xacml/MissingAttributeDetail>> does not have equivalent
>>       structure to make such distinctions.
>>
>> The above is intended just to give an example of questions that occur 
>> for this particular example, but it is my opinion that it is 
>> symptomatic of a general problem of how PEPs are supposed to know how 
>> to construct the proper RequestContext </xacml/RequestContext> 
>> necessary, in general, for complex scenarios that require substantive 
>> fga attrs.
>>
>> In these more complex fga scenarios it is likely that 
>> <MissingAttributeDetail </xacml/MissingAttributeDetail>> will be 
>> typically needed to collect all the required attributes. Therefore, I 
>> believe some more robust mechanisms, possibly using 
>> MissingAttributeDetail </xacml/MissingAttributeDetail> as a good 
>> starting point will be needed to adequately define operation in this 
>> area.
>>
>> In this context as well, it is likely that xpaths are probably not 
>> the way to go since they are only applicable to certain types of 
>> resources (xml-based) and those resource structures are likely to 
>> change in time, and these changes should not percolate into the 
>> enterprise Policy arena. Therefore, mechanisms such as vocabularies 
>> should be recommended usage here with the PEP being responsible for 
>> mapping the vocabulary item to the particular resource physical 
>> access path such as the xpath.
>>
>> Status: *OPEN*
>>
>> CHAMPION: *Rich*
>>
>>
>>
>> Erik Rissanen wrote:
>>> All,
>>>
>>> I propose that issue 66, "Missing attributes may be underspecified", 
>>> to be closed without action. The issue has been up there a long time 
>>> with no proposals for a solution. In addition to that, I believe it 
>>> is technically impossible to provide a solution. The reason someone 
>>> gets a not-applicable from the PEP is because "there is no policy 
>>> which applies". In general there is no way to describe in the form 
>>> of "missing attributes" what the PEP needs to provide for the policy 
>>> to apply. Policies can be much too complex for this. In particular, 
>>> a policy could be NotApplicable because an attribute is present, for 
>>> example. Or it might require that three particular integer 
>>> attributes form a pythagorean tripple. How do you express that as 
>>> "missing attributes"?!
>>>
>>> And we demonstrated in the RSA interop that obligations can be used 
>>> to handle simple use cases where some attributes can be expected to 
>>> be missing. An obligation can be used to mark the part of the policy 
>>> which required an attribute, and the obligation can then be returned 
>>> by the PDP if the attribute is missing.
>>>
>>> Best regards,
>>> Erik
>>>
>>>
>>> ---------------------------------------------------------------------
>>> To unsubscribe from this mail list, you must leave the OASIS TC that
>>> generates this mail.  Follow this link to all your TCs in OASIS at:
>>> https://www.oasis-open.org/apps/org/workgroup/portal/my_workgroups.php
>