Hi Steve,
Responses inline.
Rich
Steve Bayliss wrote:
003201cbcd45$f379e2c0$0301010a@asusp4t533"
type="cite">
Hi Rich
So the objective here is to ensure that the
(potentially dynamic) hierarchical information gets through to the PDP
so that hierarchical policy combining algorithms can use this
information.
Yes and no. Yes to the first part: i.e. if you have a dynamic
hierarchy, then you need a mechanism
to let the PDP know the current state in order to apply the ancestors
method. i.e. if you are comparing
the ancestors of the requested resource with ancestors of nodes in a
hierarchy where Policy has
been applied, and the hierarchy has changed, then the ancestor
associated with node where the
Policy has been applied may have changed, then the Policy, itself,
needs to be able to get the
ancestors that are associated with the current state of the hierarchy.
No to the second part because the policy combining algorithms just
combine results of policy
evaluation, and would seem to be relatively far removed from the
dynamics that were introduced
by the attribute finder to enable each policy to be evaluated with
dynamic information. i.e. I see
no connection between the attribute finder mechanism and the policy
combining algorithms, they
are involved at different steps in a well-defined sequence of
processing steps defined by the XACML
processing rules, see ch 7 and possibly the combining algorithms in the
appendix.
003201cbcd45$f379e2c0$0301010a@asusp4t533"
type="cite">
To clarify, would the policy still be written
using the hierarchical ancestor/parent attributes?
In one of your earlier emails, you indicated you thought the ancestor
scheme might be useful
for performance reasons. As a result, my responses have been to try to
apply the ancestor
scheme to the use cases you have been describing. So, yes, I am
envisioning that all your
policies that we are discussing are written using ancestor techniques,
and not relying on
the node names at all to inherently represent hierarchical position.
003201cbcd45$f379e2c0$0301010a@asusp4t533"
type="cite">
You say that the policy would look for the new
"descendents-of-a" attribute - so does this mean that the policy also
has to reference this "descendents-of-a" attribute to trigger the
MissingAttribute behaviour - if so, how? By specifying it in the
Resource section of the policy?
That seems like a reasonable place to put it. It needs to be defined
with MustBePresent="true" in
order to guarantee missing attribute processing to be invoked. Or, in
the case of SunXacml,
this will trigger the context handler to invoke attribute finders to
try to find the attribute
somewhere else, and, if found, to add it to the RequestContext to allow
the policy
evaluation to continue with the attribute now being present.
003201cbcd45$f379e2c0$0301010a@asusp4t533"
type="cite">
As an alternative approach, would the following
seem reasonable:
- the context handler is written to both provide
the resource-id hierarchical URI path attribute and the hierarchical
ancestor/parent attributes for the requested resource
One can construct a "resource-id hierarchical URI path attribute",
however, this might
be stretching the concept of URI a bit, as well as resource-id, as I
think these are
generally thought to be of more static character. However, as a
snapshot at a point
in time of a dynamic hierarchy, it might not be unreasonable. I am sure
there are
people who could argue both positions at length. :)
003201cbcd45$f379e2c0$0301010a@asusp4t533"
type="cite">
- the policy is written in terms of resource
ancestor/parent attributes (for this use case specifying a single
"ancestors: bag{'b'}") (and standard policy matching can therefore
match the policy resource specification against the request context -
using the hierarchical ancestor/parent attributes)
I am not sure I understand this. If the policy is written in terms of
ancestor/parent attributes, then
this sounds like a static representation of the hierarchy. i.e. any
explicit list of ancestors is by defn
static, and if the hierarchy is dynamic then the static list may become
obsolete at any time.
003201cbcd45$f379e2c0$0301010a@asusp4t533"
type="cite">
- the PDP has the hierarchical structure (from
the resource-id in the request context), and can use this to determine
hierarchical precedence when combining policies
I am not following this either. I agree that the RequestContext is
providing the current
hierarchical structure for the requested node, but I don't see how the
policies themselves
can provide anything to compare current hierarchical structure to. At
least not without
the dynamic mechanism I suggested above and prev emails.
I think to resolve this one would need to lay out a specific example of
both approaches
and see what is being compared to what.
Rich
003201cbcd45$f379e2c0$0301010a@asusp4t533"
type="cite">
Thanks
Steve
Hi Steve,
I think that, in general, when we talk about "restructuring the
hierarchy", that we
are, in fact, creating a "new hierarchy", at least in the sense that
any existing
representation of that hierarchy has become obsolete and needs to be
updated.
In particular, if a Policy, such as the ones we have been discussing,
have
static Attributes containing their list of ancestors, etc. then those
static Attributes
are no longer valid.
Therefore, if we want a Policy to be able to "keep up" with external
changes that
are applied to the hierarchy, then some dynamic mechanism needs to be
implemented
in order to give the Policy access to the updated representation of the
hierarchy.
One way to approach this would be to define a dynamic Attribute with
AttributeId =
"descendants-of-a", that the Policy would look for in the
RequestContext. Since,
in general, this Attribute will not be supplied by the PEP, then the
Attribute will be
detected as missing. If the Attribute is declared MustBePresent, then
this will trigger
a MissingAttribute condition for Indeterminate, which is described in
Ch 7 of the
XACML core spec 2.0.
Without getting into detail, implementations such as SunXacml when
encountering
these conditions, the ContextHandler invokes any configured
AttributeFinders, which
use the RequestContext as context, and have the AttributeId of the
Attribute that is
missing, and then the AttributeFinder can do what it needs to find the
missing attributes,
which in this case, would presumably be to locate the current
representation of the
hierarchy, calculate or read the current values of ancestors etc., and
return the
new attributes to the RequestContext.
The above technique should address the dynamic requirements that appear
to be
necessary to solve the problem you described, at least, wrt to my
interpretation
of that problem description.
Thanks,
Rich
Steve Bayliss wrote:
000801cbcd28$30132ca0$0301010a@asusp4t533"
type="cite">
Hi Rich
I can see how this would work where
policies include the full set of ancestor/parent attributes in their
resource specification; thanks for the use cases also. I think for the
two policies for "descendents of 'a'" and "descendents of 'b'" these in
fact also apply to "a" and "b" as well as their descendents (as it is
ancestors-or-self) - however I don't see any issue here, as in
the general case, if two policies "X" and "Y" are being evaluated it
will always be possible to see which is "higher" (closer to root) in
the hierarchy - as the lower policy will include the higher policy as
an ancestor - so I'm pretty sure something workable could be
implemented.
However if one wanted to encode polices
such that the policies themselves were not dependent on their own
position in the hierarchy, then this method wouldn't work I think. The
use-case here is where one wants to specify that a policy applies to
"all decendents of x" irrespective of where "x" is in the hierarchy -
this was the situation we did touch on earlier where one wants to allow
for restructuring of the hierarchy.
One way to solve this would be for the
hierarchical policy combining algorithm to have the ability to query
the hierarchy structure directly - would that be "appropriate" within
the XACML architecture?
I think it is best to leave multiple
hierarchies for now - first one would need a good example of a
situation where one could describe what one actually wanted to happen,
and I'm struggling with that - so I would tend to use the
hierarchical-lowest-wins algorithm for a specific hierarchy; and if
there were multiple hierarchies then combine these using the standard
combining algorithms for now.
Regards
Steve
Hi
Steve,
This is first chance I've had to look at this. I basically agree with
the way you have
described the problem. It sounds like you are saying:
- since both "/a/b/c/d" and "/b/a/c/d" produce the same set
of attributes
in the request context, that these two cases are indistinguishable
I think that is true that the requests are indistinguishable, but maybe
not the
policies.
First, this is no longer a single hierarchy problem: "/a/b/c/d" is one
hierarchy,
where "a" is the root node, and "/b/a/c/d" is another hierarchy, where
"b" is
the root node. I don't think this would be allowed in the DAG, because
/a/b
and /b/a would create a loop which is forbidden, so it would have to be
a forest. However, let's leave that aspect aside for now.
If we look at the policies, btw, Anne did a preliminary document on some
use cases, which may help a bit. See attachment to this email:
http://lists.oasis-open.org/archives/xacml/200406/msg00033.html
Doc is here:
http://lists.oasis-open.org/archives/xacml/200406/pdf00003.pdf
In that doc, check section 4.2 which shows an example policy using
ancestors.
In any event, let's consider how we'd write a Policy that specified
"descendants of 'a'". What we could do is collect the attributes for
"a",
namely, in the case "/a/b/c/d", the attributes are:
- self: "a"
- parent: null
- ancestors: null
- ancestors-or-self: "a"
Similarly for "b"
- self: "b"
- parent: "a"
- ancestors: "a"
- ancestors-or-self" "a","b"
For the node in the request, we have:
- self: "d"
- parent: "c"
- ancestors: "a","b","c"
- ancestors-or-self: "a","b","c","d"
In XACML 2.0 there is a set function in section A.3.11 called:
urn:oasis:names:tc:xacml:1.0:function:type-intersection
where "type" can be any XACML datatype
this function returns a bag that contains the set of nodes common to
both bags
that are input.
So, for Policy descendants-of-a we could compare the ancestors-or-self
bags
of the request node and the policy node, which would produce a bag:
{"a"}.
For Policy descendants-of-b the intersection would produce a bag:
{"a","b"}.
Now there is also a function:
urn:oasis:names:tc:xacml:1.0:function:type-bag-size
this function returns an integer.
Possibly, one could then write a policy-combining algorithm that would
include
this integer in the combining, where the greater integer implies the
"lower" policy,
which would then take precedence?
For the "2nd hierarchy" you'd have 2 similar policies. However, I am
not sure what
kind of precedence could be meaningfully created by whether the node
was higher
in hierarchy 1 than it was in hierarchy 2. So, let's leave that
discussion aside as well
for now until we reach agreement on the single hierarchy.
Thanks,
Rich
Steve Bayliss wrote:
003401cbc8fa$4d5ef8c0$0301010a@asusp4t533"
type="cite">
Hi Rich
Thanks, I think my understanding of how
this can work is nearly there.
There's a point I'd like to clarify to make
sure I haven't misunderstood something.
For clarification, this is using the
ancestor/parent scheme but not using the full path identifiers for the
ancestors/parents (I can see if the full path identifiers are used then
there is no ambiguity):
If my assumption is correct then
the given set of parent/ancestor attributes could also result if the
hierarchy was /b/a/c/d - so if policies were present that specied "all
descendents of b" and "all descendents of a" then it would not be
possible to determine which was the "lower" policy and therefore which
takes precedence?
I'm not sure I follow the logic: in the
case given
descendants of "a" are "c" and "d",
and descendants of "b" are "a", "c", "d".
The
scenario here is as follows:
i) The
request from the PEP is for resource "d"
ii) The
context handler builds the set of parent/ancestor attributes
iii) There
are policies specifying 1) descendents of "a", 2) descendents of "b",
3) descendents of "c)
iv) The
policy combining algorithm is "lowest policy wins"
v) the
policy engine needs to determine the order of precedence of these
policies in order to combine the policies according to the combining
algorithm
So,
if the hierarchy was /a/b/c/d, then in (ii) this would result in the
context handler building the following attributes for the resource "d"
ancestors:
bag{a,b,c}
parent:
c
ancestors
or self: bag{a,b,c,d}
If
the hierarchy was in fact /b/a/c/d, then the context handler would
build:
ancestors:
bag{a,b,c}
parent:
c
ancestors
or self: bag{a,b,c,d}
So,
the same attributes in both cases, so I believe the engine won't be
able to determine the hierarchy of "a" and "b" and therefore the
precedence of policies 1 and 2.
This
comes back to my assumption that the policy engine is working on the
set of resource attributes provided by the context handler (or can
request these resource attributes via an attribute query).
Does
this make sense?
Steve
Hi Steve,
I looked over your comments earlier and think we are converging,
but now I will try to respond to your points inline.
Thanks,
Rich
Steve Bayliss wrote:
006401cbc83a$d3a25280$0301010a@asusp4t533" type="cite">
Hi
Rich
Many thanks for this. I can see that I
maybe need to add some clarity in terms of identifiers, nodes and
hierarchies to avoid us discussing at cross purposes. So I'll try and do that first, then
further comment inline to your responses.
And
for further clarification I'll leave aside any more comments on
multiple hierarchies and assume a single rooted hierarchy for this
example.
The
nodes in the example have identifiers a, b, c, d - these identifiers
are unique within the scope of the system.
The
expression
/a/b/c/d
is
meant to state that these resources a, b, c, d are organised in a
hierarchy, so that b is the child of a, c the child of b etc - this
"path" is not a node id per se; but a URI containing this path could be
generated as part of the hierarchical "path" URI scheme, as a
"resource-id".
So a
resource-id (using this scheme) of eg file:///a/b/c/d refers to the node d (and
also gives information about its ancestors and their relationships
within the hierarchy).
I
think this is the assumption you've made below.
I agree. So,
both "d", and "file:///a/b/c/d" identify the same node,
since "d", by itself is unique, and the other also
shows where "d" is in the "hiearchy" in which these nodes have been
assigned. For concreteness, if a,b,c,d were
some sort of physical device, then a,b,c,d might be their unique serial
numbers, which identifies each device,
but the devices themselves have no inherent relation to each other. The
hierarchical relation is a property that
is added later. This has been generally covered in earlier emails, so
nothing really new, just indicating how
the identity of a node "may" also include its position in the
hierarchy, but this is not required, as there are other
"identities" that may represent the node that are independent of any
hierarchy.
006401cbc83a$d3a25280$0301010a@asusp4t533" type="cite">
Further
comments inline.
Steve
Hi Steve,
You raise some interesting questions. I am finding that it is necessary
to qualify
my answers significantly, because the problem statements appear, at
least to me,
somewhat ambiguous, and rather than try to answer all possible
interpretations
of the problem, I opt to select one definitive interpretation and maybe
only mention
the others for reference. I think this particular problem, the
hierarchical profile, is
particularly prone to this difficulty, because there is so much
inherent flexibility in
hierarchical interpretation and it can be applied in many different
ways. So, bear
with me as I try to provide useful answers inline.
Thanks,
Rich
Steve Bayliss wrote:
Hi Rich
This is most useful, so thank you for your
insight. It's prompted me to read through both the 2.0 and 3.0 HRP
specs again in some detail.
In terms of being more specific, a use-case I am
thinking of is a hierarchy where policies are applied at different
levels in the hierarchy, with the lower-level policies overriding
higher-level ones.
That's
pretty clear, sounds like a reasonable problem to be trying to solve:
define policies
on hierarchical members (nodes) that override policies of ancestor
nodes that apply to all
descendant nodes.
So a made-up example would be a single hierarchy:
/a/b/c/d
Ok,
this can be considered to be the URI " '/' + pathname" part of a
specific node-id.
Or possibly what you are saying is that this is the node-id of a
hierarchy with a single
node, namely the one specified.
I will assume we are talking about a single-root, single-parent
hierarchy, where the
root node is "a" and the "made-up example" refers to a specific node
within this
hierarchy. (Maybe this seems like hair-splitting, but I have found
unless all the
details are explicitly nailed down that the discussion quickly diverges
with each
of the people in the discussion working from a different set of
assumptions, which,
imo, ends up being a waste of everyone's time because effectively no
one really
understands what the others are saying. Don't mean to soapbox, but as
this email
progresses I think you will see a substantial number of potential
divergence points.)
I think we're in agreement here, and do appreciate the point you're
making about the potential distinction.
ok
006401cbc83a$d3a25280$0301010a@asusp4t533" type="cite">
Policy 1 states "allow read access to Subjects
with an attribute value of 'public' to all descendents of 'a'" (so a
policy resource-id regex of /a/.*)
Agreed.
Policy 2 states "deny read access to Subjects
with an attribute value of 'public' and an IP address of 'x.x.x.x' to
all descendents of 'b' (ie regex .*/b/.*)
Agreed
with following qualification: I assume that Policy 2 could also be
stated
as "/a/b/*" instead of ".*/b/.*". Not sure why you chose the latter.
The reason for
this distinction is that the policy is to apply to descendents of b,
rather than descendents of b when b is a child of a. A use case for
this may be reorganisation of the hierarchy so that b is no longer a
child of a.
I think I get
your point. Starting with the collection a,b,c,d, if we are interested
in the descendants
of "b", then that is not determined until someone puts together the
hierarchy. However, once the
hierarchy has been assigned, i.e. at any given point in time, then the
list of descendants can be
determined. Also, at some later discrete time, the hierarchy could have
been rearranged and the
list of descendants will be different, in general, from the list at the
previous time.
006401cbc83a$d3a25280$0301010a@asusp4t533" type="cite">
A request for "d" from a Subject 'public' with
IP address "x.x.x.x" should result in a "deny" based on Policy 2
overriding Policy 1.
I think all
that can be said here is a request for "/a/b/c/d" should result in
- a Permit from Policy 1 and
- a Deny from Policy 2.
The
ambiguity of saying "d" as opposed to "/a/b/c/d" comes from the initial
problem statement,
where I assumed there could be multiple nodes in the hierarchy. If you
also had a node named
"/a/b/d" and then stated a request for "d", I am not sure if want all
nodes with node-name "d"
or if you really want the particular "/a/b/c/d".
Hopefully
the clarification with respect to nodes and identifiers resolves this
ambiguity. The user is requesting resource d, that results in a
resource-id of file://a/b/c/d
using the hierarchical resource profile path URI scheme.
Agree. The fact
that you said "these
identifiers are unique within the scope of the system"
means
there will only be one "d", and, as above, at any point in time it may
show up at different points
in the hierarchy, depending on whether the hierarchy has been
re-arranged.
006401cbc83a$d3a25280$0301010a@asusp4t533" type="cite">
The PDP implements a "lowest policy wins"
algorithm. Using the path URI resource-id profile means the PDP has the
full information, on a request for 'd', to determine that (1) both
Policy 1 and Policy 2 apply and (2) Policy 2 takes precedence (the deny
overrides the allow) as it is at a lower level in the context of that
particular hierarchy.
This sounds
like a worthy objective, and one could possibly use the existing
policy-combining algorithms to achieve it or maybe not. As you indicate
ahead,
you are leaving that TBD for now. Possibly a new combining algorithm
would
be in order, which is interesting, but seems possibly as a separate
discussion
issue.
(I am leaving aside the question of what
algorithm should be used to combine policies across multiple
hierarchies!)
I
think I agree based on my previous comment, but now you have introduced
a phrase "combine policies across multiple hierarchies". I agree
multiple hierarchies
is an interesting discussion area, but I thought the example was for a
"single hierarchy"
as stated above, which I will continue to assume.
Agreed. This is a separate discussion, let's
focus on a single hierarchy.
Yes, if the
rules for a single hierarchy (single root, single parent) can get
nailed down
unambiguously, at least for some problem domain, then more complex
problems can
be systematically introduced and analyzed on that foundation.
006401cbc83a$d3a25280$0301010a@asusp4t533" type="cite">
The distinction I was assuming existed between
the URI path and the parent/ancestor attributes was the loss of some
information of structure of the hierarchy in the latter case.
Again, if
we are talking about a single hierarchy, as in the problem statement
above, I am
not sure why you are making this statement., because in the particular
case of a single root,
single parent hierarchy, there is, in fact, no information loss. For
example, in the case given:
- parent of d is c
- ancestors of d are a, b, c
- ancestor or self of d are a, b, c, d
From
this collection of AttributeValues, the policy engine with access to
the node hierarchy
can determine the URI by looking at the parent nodes of c and b, so the
information is not
really "lost" in this case. i.e. one can unambiguously state that a
policy applied to the descendants
of "a" is intended to apply to "d".
I'm still
not clear on this - particularly the phrase "the policy engine with
access to the node hierarchy". I'm making the assumption that the
policy engine has access to the node hierarchy by having been given
these hierarchical attributes (either directly through the context
handler, or through a request for those attributes). Maybe this
assumption is not correct?
Right, in
general, this is a big assumption to make. There are a number of
possibilities, each of which seems to be dependent on situation-specific
characteristics.
- One case is the URI scheme. If a URI is passed in
then it inherently
contains all the parent and ancestor information. But to use the
ancestor
scheme, one has to have a way to parse the URI to produce the list
of ancestors. The Policy could declare an attribute for the list, and if
it is not there, then maybe an attribute finder could be invoked which
would pull the URI from the RequestContext and parse it and return
the parsed components as the value of this "missing attribute". Note:
this only works because all the info is already provided.
- Case 2: might be that some list of ancestors is
passed in plus the resource-id.
The question might arise, how do we know this list is accurate?
Presumably,
it must have come from the hierarchy itself, or some authoritative
representation
of the hierarchy. For example, the PEP could "go to the authority" and
request
the list of ancestors for a specific resource-id.
- Case 3: no URI, no list provided: again we could do a
missing attribute,
however, in this case, the attribute finder would need to go find the
authority
and get the list and return it.
- Case 4: the navigation case you suggested earlier:
somehow the appl is
keeping track of the navigation path, and when the PEP is invoked,
either the node list or the parseable path can be available to the PEP
to do what it chooses.
Bottom line: it
is easy to "talk about" a list of ancestors, but it may not be obvious
how this list is obtained, and depending on what's involved, it may or
may not
influence a possible choice between the URI scheme vs the ancestor
scheme.
006401cbc83a$d3a25280$0301010a@asusp4t533" type="cite">
If my assumption is correct then the given set
of parent/ancestor attributes could also result if the hierarchy was
/b/a/c/d - so if policies were present that specied "all descendents of
b" and "all descendents of a" then it would not be possible to
determine which was the "lower" policy and therefore which takes
precedence?
I'm not sure I
follow the logic: in the case given
descendants of "a" are "c" and "d",
and descendants of "b" are "a", "c", "d".
So, "a" is lower, since its list is a subset "b"'s list.
So, it seems to me that with a single root single parent hierarchy, no
info is
lost whether you use the URI or the ancestor scheme. I think the issue
only arises when multiple hierarchies are combined in a DAG.
In XACML 2.0, this was the only choice one was presented with, at least
as I read the specs, i.e. you could choose the DAG or the URI scheme.
Upon detailed analysis of the DAG, it became well understood that
information
was lost. Then there was subsequent discussion whether that mattered or
not.
The XACML 3.0 spec attempts to resolve this issue by providing both a
representation where the info is lost for applications where that is
ok, and
a representation (forest) where information is retained (i.e. where a
node
has a list of parents, and each parent is "tagged" as to which
single-rooted,
single-parent hierarchy it belongs to).
The algorithm provides for both including a mix of a forest and
multiple DAGs,
if I remember right.
006401cbc83a$d3a25280$0301010a@asusp4t533" type="cite">
It is when you get to "hierarchies" with multiple roots, as in the case
of a DAG, that such
information as the "descendants of a" becomes uncertain.
Again, this is a separate discussion area, that diverges from the
original problem statement.
So focussing now on a single hierarchy example
/a/b/c/d
What I had assumed:
A request for "d" would result in the
following path URI (as resource-id attribute), eg:
Yes, I
agree. Except what do you mean "would result"? Possibly you are
referring
to the navigation client that is collecting specific data as it goes,
which has collected
a,b,c on route to d, and can reconstruct the URI based on this.
I'm assuming here that the context handler, with
access to the full hierarchy information, would construct a
hierarchical URI resource-id file:///a/b/c/d, when the incoming request for d. So that's the case using
the hierarchical resource-id URI, rather than using parent/ancestor
attributes.
I agree, the
context handler is the place where a lot of these questions can be
answered in a straight-forward manner. The tricky part, imo, is that
the CH itself,
is an ambiguity wrt the spec, and since a lot of the assumptions one
needs to make
about the hierarchical profile depend on the characteristics of the
specific CH.
This does not mean, policies cannot be written, but w/o knowing
something about
the CH, it is difficult to say whether one specific strategy is
superior to the other.
006401cbc83a$d3a25280$0301010a@asusp4t533" type="cite">
And the following
ancestor/parent attribute values
parent-id=bag{c}
ancestor-id=bag{a,b,c}
Agree.
Thus if using the
ancestor/parent attribute method, the PDP doesn't have the information
to determine which of the two policies takes precedence (both a and b
are ancestors, but it's not expressed that a is an ancestor of b).
(That's not to say it couldn't go elsewhere to gain information on the
hierarchical structure of the resources and therefore the policy
precedence - perhaps that should be the approach rather than
determining the hierarchy from the request context resource-id
attributes?)
Right,
based on the attributes alone, without access to the hierarchy, one can
probably
not determine whether Policy 1 or Policy 2 takes precedence.
Agreed, this is what I meant by loss of
information about the full hierarchy when using parent/ancestor
attributes as opposed to a resource-id URI with a path specifying the
hierarcy. Again the key phrase is "without access to the hierarchy",
which relates to my earlier assumption.
However, one could probably say the following:
- Policy 1 applies because using some algorithm,
one can presumably determine that
the bag{a,b,c} "matches" the path "/a/.*".
- If Policy 2 were stated as "/a/b/.*" then that
would also "match" bag{a,b,c}, then
maybe one could infer that Policy 2 took precedence because matching a
and b
is more nodes than just matching a, so that might imply "deeper in
hierarchy".
Agreed.
(Personally, imo, this begs the
question "Why would anyone want to do things this way?",
but that is also a divergence and a digression, which could be
discussed separately.
Note: I recognize that there are social networking and semantic web use
cases where
this type of processing might be useful, but I am not sure whether
applying policy
for access control is a workable use case, but again, that's another
discussion.)
What I think you are suggesting is that the
parent-id and ancestor-id values are values that themselves express
their own position in the hierarchy, ie
parent-id=bag{/a/b/c}
ancestor-id=bag{/a,/a/b,/a/b/c}
Yes,
this is my preferred way of looking at this problem. If you have known
hierarchical paths, then I see no reason not to use those as
resource-id's which
enables one to use the URI mechanisms. However, there may be cases where
the URI presents too much overhead of processing, and maybe ancestors
perform
better, but "who knows", to me that is a performance optimization based
on a
specific hierarchy with known characteristics.
I can see what you are saying here - that if
there are known hierarchical paths then these values can be used as
parent/ancestor values. As you state it is a "who knows", but it
"feels" to me that this is doing the job of using the URI mechanism.
Although the advantage I see in using parent/ancestor attributes in
this way is that it avoids needing to match based on regex.
Specifically if one was using a PolicyFinderModule (in the Sun XACML
implementation) then indexing policies (for discovery by the
PolicyFinderModule) based on these parent/ancestor values in the
policies would be easier than indexing based on some decomposition of
regex values of resource-ids (unless one explicitly limited the
allowable range of regex to simple wildcard matches).
I think in some
sense this is getting to the heart of the matter. imo, if there is a
fixed
hierarchy, typically where governance is done over subtrees, then the
regex approach
is probably desirable. For example in a file system where all the files
for one department
are within a specific subtree, then one could simply specify the path
to that subtree and
feel confident that the controls being applied were what's needed. In
this case, I think
representing this path as a collection of ancestors pretty much
obscures what the policy
is doing, so while it might work technically, it would seem to have
little practical value.
In the navigation case, where the list of ancestors is a function of
how you got to
the node, i.e. you are effectively defining your own hierarchy on the
fly, then there
is no way to know in advance what paths need to be protected, since
there can be
any number of paths to a specific node. I think this implicitly places
you in the
multiple hierarchy problem space, and if you don't care what particular
path, but
do care about the ancestors, then a DAG might be the right model, in
which case
one would probably look to the ancestor method as being strategically
advantageous.
The intent of the XACML 3.0 profile is to allow designers to freely
choose the
scheme to apply to different aspects of their problem domain. I suspect
the notion
of "resource-type" would play a significant role in the sense that a
specific resource-type
might be best represented by a URI, or a DAG, or a forest (which is a
superset of the
DAG and URI schemes). It is also conceivable, I expect, that even a
single resource-type
might have aspects of its use in an organization that would suggest
that it have independent
representations including any or all of the 3 schemes.
Hopefully this discussion can help development of a structural
framework for analyzing
specific problem domains, and that it can provide structure for
partitioning a complex
problem into a more manageable small set of relatively simpler problems.
Rich
006401cbc83a$d3a25280$0301010a@asusp4t533" type="cite">
So in this case the structure is not lost
(and I see the equivalence to the path URI method).
Agree.
I need to go through the pseudo-code
algorithm in some more detail, particularly to understand what the
"output" would look like in terms of part of a XACML request context
(ie the resulting XML).
There are parts of the 3.0 spec that do
seem to suggest that the parent/ancestor method could convey less
information than the path URI approach, particularly
2.3: In this
approach, considerable information is discarded. It is not possible to
determine how many hierarchies there are or which ancestors are in
which hierarchies or the relative position of ancestors other than
immediate parents.
Right, this
statement was included to alert users of the nature of the "ancestor"
or DAG method,
which is described in more detail in sections 1.1 and 1.1.1 of the 3.0
profile.
The specific aspect that needs to be preserved is the specific parent
and specific hierarchy
if one is to retain "chain of control" information.
For
example if the CEO of a company applied a policy to all employees,
it is possible that such a policy could be overridden by someone in the
mailroom
who was captain of the company bowling team of which the CEO was a
member.
This has to do with the "tranistive" property of hierarchical
relations, discussion of
which can be found in semantic web references. With single root single
parent
hierarchies this problem can be avoided, and as explained at the end of
section 1.1.1,
retaining the hierarchical identity produces a superset of the DAG,
which also does
not have the cycling problems of DAGs.
None of this was apparent, in the XACML 2.0 version of the hierarchical
profile,
which is the reason for much of the added detail of the 3.0 version,
which still
preserves all the functional capabilities of 2.0, but just
distinguishes the cases.
4.1 The representations of the identities of these parents,
ancestors, or self will not necessarily indicate the path from the root
of the hierarchy to the respective parent, ancestor, or self unless the
representation recommended in Section 3.2: Nodes in a resource that is
not an XML document is used. [ie the path URI method]
Right, this
is included as well to alert users of the choices inherent
in the approach that is used.
From 4.1 it
would seem that the values used for the ancestor/parent attribute
values could in fact be resource-id values that use the path URI
method. So no information is lost.
Yes, if you
use URIs as the node-names you are effectively retain the information,
but in a redundant manner.
However this
seems to be at odds with the assertion in 2.3 (although noting that the
attribute values "... may be identified using identifiers of any XACML
datatype ..."; so there is nothing to preclude usage of path URIs for
these... in which case no information is lost.
This is a
special case of the ancestor method, which allows you to capitalize
on the URI properties from the node names. But it is not a "general
solution"
because it is dependent on a particular node-naming strategy.
Furthermore,
in order to use the retained info, one must "crack" the node-names
which is
outside the specified scope of the ancestor method.
So overall I
am left a little confused as to the intent of the specification in this
respect. Using path URIs for ancestor and parent attribute values
seems like a merging of the two approaches.
The point
of the comment in 4.1 is not to recommend using URIs as node-names
in the ancestor method, it is to point out that if one wanted to use
the ancestor
method, and also if one became "concerned" about the specific paths to
the nodes,
then this would be a "way out" of the problem. However, once one
realized that
the path information was of significant value to a particular use case,
then one would
probably want to consider re-evaluating the original strategy of using
the ancestor
method.
Again, this is an area the could become a whole discussion topic all by
itself.
Though I
guess the datatypes of the ancestor and parent attributes don't have to
be path URIs in terms of the spec, but could be some other datatype
that expresses the hierarchical path.
Agree.
There is a lot of flexibility inherent in the Hierarchical Profile
because it covers
anything that has mathematical hierarchical properties. The original
spec focussed on
the more general multi-root, multi-parent case, which proved to be
problematic
for chain of control use cases, in particular, because the "chain of
control" was tossed
out in the ancestor method, and one was left with only a fairly brittle
URI method.
The purpose of this spec is to unify all these use cases under one
broad umbrella,
where the specifics of each subcase can be identified by specific
properties of
the hierarchy or hierarchies in the application arena that need to be
considered.
The algorithm in 3.3.1 is intended to formalize this unification and
show how all
these methods can logically and peacefully coexist in a fairly
straight-forward
framework for representing the problem space.
Hi
Steve,
Generally, I think your question needs to get more specific, because
of the inherent flexibility of the URI/hierarchical mechanism, imo, one
needs to understand what one is trying to achieve in order to
meaningfully
define an optimum strategy. I will take an initiial shot w responses in
line.
Thanks,
Rich
Steve Bayliss wrote:
Hi Rich
Many thanks for this, this greatly clarifies my
understanding.
To clarify one thing, "Scenario 1" may not be
two user applications - it could be a single user application which
presents a number of different "browsing paths" to resources. But I
don't think that changes anything, so long as the application's PEP
constructs the browsing path the user navigated through.
Agree.
The important thing is the path itself, not which module navigated the
path.
One further quesion (at this stage!):
Would it be possible to
1) in the Request Context, use the hierarchical
URI profile; and at the same time
2) in policies use the parent/ancestor attribute
profile?
Personally,
I do not see a significant distinction between the "URI profile" and the
"parent/ancestor attribute profile", which I see as effectively being
functionally
equivalent, with the distinction that the URI is a special case of the
parent/ancestor attribute,
where the special case is that the URI has a defined syntax.
So, depending on how one approaches the problem, or specifically how
one goes about
"naming" the resources, steps 1 and 2 as you describe may not be
significantly distinct.
For example, if you start with a collection of resources, and each has
a unique identifier,
such as a serial number, or a social security number, there is only a
"1-level" hierarchy,
the collection, itself, is the parent and each individual resource is a
child of that parent,
with no inherent relation to any of the other members of the collection.
The next step is maybe you want to "structure" the membership of the
collection in
some manner so that the members have relations to each other. At this
point it is
necessary to define the "relation" and how the "relation" is
established between
the members. Once this relation is defined, then it makes sense to
start considering
how policies are applied to the collection.
Let's assume you organize them in a single rooted hierarchy such that
each member
has a single parent and all members have some parent. Now you have the
equivalent
of a file directory, where each file has a name and it is also a member
of a directory
in a hierarchy of directories.
Given this scenario, then what is your question 1 and 2? I would assume
I would
pass the dir-path/filename as the resource-id. The list of ancestors is
nothing more
than a succession of truncated dir-paths. So is the question whether to
pass
"/a/b/filename" as resource-id, vs "/filename" as resource-id plus
"/a/b" and "/a"
as ancestor attributes? Since the latter can be derived from the
former, what is
the difference?
Is the parsing of "/a/b/filename" vs "/a/b/.*" more expensive than
doing bag operations
of ancestors? I honestly have no idea. It may depend on the
implementation of
the engine. It may depend on how large the collections are that are
being processed.
I would look to the engine supplier to provide guidance on how to use
these capabilities,
and possibly suggest situations where one or the other would be optimum.
Personally, I think this is a level of detail that policy designers
should not be concerned
about unless there is some macro characteristic that they can relate to
which will guide
the choice.
I'm thinking here that from a perspective of
indexing/finding policies, this would be much easier to do based on
parent/ancestor attributes rather than for instance decomposing a
resource-id regular expresssion (which may specify a range of
possibilities, such as a path /a/b/.*, specifying "all resources under
b when b is under a", and ".*/b/.*" specifying "all resources under b
irrespective of b's parentage").
The
comments above should address, but not necessarily fully answer the
"much easier" question you are thinking about.
However it may be useful to still include the
resource-id in the request as this specifies the complete ancestor
lineage, which could be useful for instance in a custom policy
combining algorithm which had some rules about policy precedence based
on the position in the hierarchy - for example child policies have a
precedence over parent policies.
Again,
depending on what you are trying to accomplish, the importance of which
hierarchy a node's parent belongs
to may or may not be important. As the algorithm in XACML 3.0 shows,
one actually has to go out of one's way
to effectively remove the specific hierarchy identifier of each parent,
at least, if you represent the original hierarchy
by a single column in the n+1 columns of n hierarchies and 1
resource-id as in the algorithm. And as indicated in
the algorithm, if you start with a DAG, then you must have m columns,
where m is the max number of parents any
single member of the DAG has, and if a new parent is added to an
m-parent node, then just add a new column
to the table and increase the DAG width to m+1.
Regards
Steve
Hi
Steve,
I find your explanation much easier to understand now. It sounds like:
- there is one underlying "physical"
resource, "c", that can be obtained either
- by application "A" by navigating via
/a/b/c
- or by application "X" by navigating
via /x/y/c
- basically "A" presents user with /a/b/c,
and "X" presents user with /x/y/c
where /a/b and /x/y are "steps" the user takes thru the appls
I
think the following Policy elements placed in a Target would make the
Policy
applicable to any resource prefixed with "file:///a/b"
<ResourceMatch
MatchId="urn:oasis:names:tc:xacml:1.0:function:regexp-string-match">
<AttributeValue
DataType="http://www.w3.org/2001/XMLSchema#string"
>file:///a/b/.*</AttributeValue>
<ResourceAttributeDesignator
AttributeId="urn:oasis:names:tc:xacml:1.0:resource:resource-id"
DataType="http://www.w3.org/2001/XMLSchema#string" />
</ResourceMatch>
A
similar match statement could be prepared with "file:///x/y/.*" to match any request
prefixed with "file:///x/y".
These match statements can be combined within a Resource statement to
create an "AND",
or combined within 2 separate Resource statements to create an "OR".
Also, I have added a prefix "file://" in order to make
these proper URIs, but will not
use it below. The point is while the example does not use full URI
syntax, there is no
reason why it can't, and we will just assume it is shortened for
convenience.
I think the rest of my comments can now go in line.
Thanks,
Rich
Steve Bayliss wrote:
Hi Rich
Many thanks for your response. I realise
my questions are not clear, I was struggling myself with how to phrase
them, and my terminology is rather loose in places. Perhaps the best
approach is for me to step back and outline two distinct scenarios;
particularly to address what I was trying to express with
"resources-in-path-context".
In both scenarios there is a resource c
that is present in two hierarchies:
/a/b/c
/x/y/c
And in both scenarios, there is a user
application that allows a user to browse through the hierarchies to
locate the resource. The user could browse by starting at the top of
either hierarchy (a or x) to locate the resource c.
Scenario 1:
The user browses starting at "x", and
locates the resource under the hierarchy /x/y/c
Is it possible to write a policy that
applies based on which hierarchy the user navigated to locate the
resource, ie a policy that applies to their browsing path of /x/y/c, so
that the policy would apply if they had browsed through x->y->c
but would not apply if they had browed to the resource through
a->b->c?
The
first (and only) ResourceMatch element above I think is what such a
Policy would need
to contain to match the "/a/b/c" path.
(This is what I was attempting to express
with "resource-path-in-context", the context being the route a user
took to locate the resource.)
That
is much more clear now, thanks.
The application's PEP would be able to
supply the browsing path in the request context.
Agree,
the general assumption is that the PEP is able to collect information
about the
request and transform that information to an appropriate set of
attributes.
(If additional attributes are required they can be obtained via the
ContextHandler or
by using the MissingAttributesDetail technique.)
Is this something that the Hierarchical
Resource Profile is suited to?
I
believe so. It appears to me to be a perfect fit.
Or should this "navigation path" should be
a separate attribute outside of those specified in the Hierarchical
Resource Profile?
I
don't think it is necessary for an additional "navigation path"
attribute. I think the navigation
path makes for a perfectly good resource-id.
I can see how a policy could be written
using a "navigation path" attribute, my question is whether the
scenario is relevant to/better implemented with the Hierarchical
Resource Profile, using the resource-id attribute to represent the
navigational path used by the user.
The
only distinction I can think of right now is that a "navigation path"
in the sense of tracing everything the user
has been doing may have redundant info. However if the "navigation
path" removes backsteps, then I think it
is effectively equivalent to the URI path, which is a legitimate
resource-id value.
Am I missing something here?
Scenario 2:
The user browses to the resource c. It is
irrelevant from a policy perspective of how they navigated to the
resource.
Presumably,
within this particular example there are only the two paths available.
In any event, even if there
were more paths, the profile should work just as well.
Two separate policies are to be written,
one for all resources that are descendents of "a", and another for all
resources that are descendents of "x". Both policies should be applied
when the user requests resource "c".
This
sounds like it could be done with an "AND" of the two ResourceMatch
elements (one for "/a.*",
and the other for "/x.*") within a single Resource element.
It seems that this is suited to the
Hierarchical Resource Profile.
Agree,
but also think scenario 1 is just as applicable.
It would be the responsibility of the
Context Handler to provide two hierarchical resource-id attributes.
The
algorithm in HRP XACML 3.0 section 3.3.1 shows how resource-ids can be
constructed given a list
of unique resource identifiers (can be a simple integer starting from
zero and incremented for each resource
that is added to the collection). This "list" can be regarded as
"column 0" of the "resource map".
For each hierarchy applied to the collection an additional column is
created for the map. Within the column,
the resource identifier of the parent of the current resource is
assigned to create the hierarchy.
The context handler can use this map to create all the resource-ids
needed for the resource.
Policies could be then written using (URI
regex match) /a/.* and /b/.* (or indeed separate rules as you
suggest). Alternatively the "ancestor scheme" could be used as you
suggest, in which case the Context Handler would provide the
resource-parent, resource-ancestor, resource-ancestor-or-self
attributes (and the policy would be written in terms of these rather
than a resource-id regex match.
I
agree the "ancestor scheme" can be used here. Personally, I think the
URI scheme is sufficient for
the use cases I have considered to date. I think it is just a matter of
how the information is viewed
as to which approach is preferable.
With the DAGs, I think the URI would not work very well, if at all
since there are no definitive
unambiguous URIs that can be constructed. So, if resources are viewed
as DAG, then should
use ancestor approach. Same algorithm in 3.3.1 is applicable.
Further comments:
To help my understanding of the difference
between the URI scheme and the "ancestor scheme":
- If the URI scheme is used,
the resource-id attribute is a URI specifying the hierarchical path to
the resource. Multiple resource-id attributes would be present if the
resource is to be found under more than one path.
Yes,
that is correct, and it is allowed as specified in XACML 2.0 section
6.3 lines 2947-2954,
and XACML 3.0 generally Attributes can contain multiple
AttributeValues. It probably would
be more correct to say something like "the Attribute with AttributeId =
"...resource-id" generally
can contain multiple AttributeValue elements with different values".
- If the "ancestor scheme" is
used, the resource-id does not have to be a URI, and does not have to
contain any information about which hierarchies the resource belongs to.
Agree.
- Instead resource-parent,
resource-ancestor and resource-ancestor-or-self attributes are used to
represent this information
Yes,
although, this probably is only relevant to DAGs, and is superfluous
for any number of single parent hierarchies.
The "relaxation" I thought I had identified
for the resource-id in the URI scheme was due to the 2.0 document
stating (section 2.2) that "The identity of a node in a hierarchcial
resource ... SHALL be represented as a URI ...", whereas the 3.0
document (again section 2.2) has "The identity of a node in a
hierarchical resource ... MAY be represented as a URI...". However on
re-reading the subsequent text in section 2.2 of the 3.0 spec, it now
seems clear to me that the resource-id must indeed be a URI.
I hope this goes someway to clarify the
questions I was attempting to pose in my original post.
Yes,
I think we are in agreement, assuming my comments are consistent with
what your
statements are intended to mean.
Rich
Thanks
Steve
Hi Steve,
Not sure if I fully understand your questions, but will try to respond
inline. Note:
my first few responses are attempting to understand your question, the
later
responses are probably closer to being in the answer space of your
questions.
I do not expect that this will fully answer your questions but
hopefully it will
move the ball down the field a bit.
Thanks,
Rich
Steve Bayliss wrote:
Hi
I'm
seeking to clarify my understanding / the intent of the Hierarchical
Resource Profile. Primarily XACML 2.0, but I understand that 3.0
mainly refines and adds clarity, so anything from 3.0 that's relevant
in this respect I am interested in.
I'm
iinterested in the hierarchical path representations of non-XML nodes,
ie 2.0 spec lines 190 et seq.
I
am having trouble parsing the following paragraph. I will deal with
each segment in sequence:
My
question is if the Hierarchical Resource Profile deals (is intented to
deal with) with resources-in-path-context.
The
sentence above appears to be asking the question:
Does the HRP deal with "resources-in-path-context"?
If that is the question, then what does the term
"resources-in-path-context" mean? For example,
XACML Policy deals with information in the RequestContext. Is this what
you are referring to?
That
is, if a resource belongs to multiple hierarchies/has multiple
parents/ancestors, can a policy be written for this
resource-in-path-context.
The
answer to the first part of the question is "Yes, a policy can be
written for it.".
As above the term "resource-in-path-context" is undefined, at least to
me, so I do not understand the last part of the question.
(to
put it another way, policies dealing with resources when being accessed
as part of a particular collection having a specific policy, rather
than a resource being accessed, and because it is a member of - rather
than being accessed in the context of being a member of - having a
specific policy).
I
do not understand the above statement at all. The first phrase appears
to use the term policy redundantly:
"policies (dealing with resources when being
accessed as part of a particular collection) having a specific policy"
The
remainder I find in more difficult to parse. My point here is not to
criticize your sentence structure, but to
try to understand what the point is you are trying to make, and
possibly by explaining why I find it confusing,
will help to establish a terminology where the issues can be addressed.
That being said, the last phrase appears
to me to be saying:
There is some kind of distinction between:
- a resource being accessed because
it is a member of (a specific collection?)
- a resource being accessed in the
context of being a member of (a specific collection?)
Is this the point, that the above two bullets are distinct in
some manner?
Example
paths for a resource "c" (ignoring the fact these aren't URIs, I gather
this restriction was relaxed in 3.0)
/a/b/c
/x/y/c
Ok,
assuming the above URI pathname portions plus leading "/" are provided
as resource-ids (where multiple resource-ids
for the same resource are allowed as specified in section 6.3), the two
provided above seem like a reasonable pair of example paths.
Note: I do not believe there was intentional "relaxation" in 3.0. Note
there are 2 non-XML node identification schemes in the HRP:
- The URI scheme: section 2.2 (of both
XACML 2.0 and 3.0)
- The "ancestor scheme": section 3.2 of
XACML 2.0, and section 2.3 of XACML 3.0.
If
the URI scheme is used the resource-id must be a URI. If the ancestor
scheme is used, then that is not a necessary condition.
According
to the spec, a request for resource c should result in two separate
resources in the request context, a resource-id with value /a/b/c and a
resource-id with value /x/y/z - is that correct?
There
is only one resource, but it will have 2 resource-id's, so the answer
to above is "no" to the first part and "yes" to the 2nd part.
And
two different policies, one specifying (as URI regex match) /a/b/.* and
the other /x/y/.* would both match?
There
is no requirement for two Policies, as opposed to say two Rules, but
yes, it is likely that both the regex's you
propose would appear.
Is
it also permissible for the incoming request to the PEP to specify the
full path, for instance to state that resource c is being accessed in
the context of collection b which is a member of collection a? Thus
only the first policy would match?
It
is ok to specify only one path in the request, but it depends on the
Policy structure whether that would
be accepted or not. i.e. does the Policy state:
- ( /a/b/.* AND /x/y/.* ), or
- ( /a/b/.* OR /x/y/.* )
That
is, because a full resource-id is already present then the engine
shouldn't go away and build the paths.
Sorry,
I am not sure what you are trying to say here. There is no requirement
on the engine to use any
specific method for processing the above scenario.
And
is it conceivable that both forms could exist together?
I
assume so, however, it sounds like you may have a concern that for some
reason both forms should not
co-exist. Is that a concern?
For
instance, if there was a resource attribute of object-ID, and it is
defined that resource-id is formed of the full path to object-ID then
we could have
Request
one:
specifies
resource object-id = "c"
result:
engine builds all applicable paths to "c" - ie two separate resources
with their own resource-id (so both above policies match)
I'm
afraid you are going to have to be more specific here about what
information is being processed.
My assumption is that all required info is in the request context. If
all that is there is "c" then how does one
derive full paths?
Request
two:
specifies
resource-in-path-context, specifying resource-id directly as "/a/b/c"
result:
engine does not build resource-id and only the first policy matches
Correct,
but see comment above. It depends if Policy requires resource to be
member
of both hierarchies or only one hierarchy.
Any
guidance on this appreciated. My reading so far of the specs suggests
that the Hierarchical Resource Profile isn't intended for
resource-in-path-context (or resource-accessed-as-member-of-collection).
Again,
I am not sure what the term "resource-in-path-context"
precisely means or the alternate phrase:
"resource-accessed-as-member-of-collection",
at least in the context of the HRP. The HRP is intended
to represent single nodes in a hierarchy, and to also consider using
the Multiple Resource Profile
for defined "scopes" of nodes within a hierarchy.
Thanks
Steve
|