cxf-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Andrea Smyth <andrea.sm...@iona.com>
Subject Re: Ideas for a WS-Policy Framework for CXF
Date Mon, 15 Jan 2007 16:28:37 GMT
Glynn, Eoghan wrote:

> 
>
>  
>
>>-----Original Message-----
>>From: Andrea Smyth [mailto:andrea.smyth@iona.com] 
>>Sent: 15 January 2007 09:21
>>To: cxf-dev@incubator.apache.org
>>Subject: Ideas for a WS-Policy Framework for CXF
>>
>>I have written up some ideas on a WS-Policy implementation 
>>for CXF on the wiki: 
>>http://cwiki.apache.org/confluence/display/CXF/WS-Policy+Framework
>>Any feedback appreciated.
>>    
>>
>
>Hi Andrea,
>
>A couple of thoughts on this ... 
>
>On the issue of the server-side assertion verification, can we
>distinguish two different classes of assertion:
>
>1. Assertions that are manifest from the gross incoming transport-level
>packets, without requiring any introspection of the actual message
>payload, and may even be enforceable prior to the payload being
>available to the receiver ... e.g. asserting that the socket-level comms
>are encrypted with a key strength of at least 128 bits, which could be
>enforced via a TLS handshake, in advance of any message data being sent.
>
>2. Assertions that are only manifest after introspection of the message
>payload ... e.g. asserting that addressing is used, in the sense that
>WS-A headers are present in the incoming message.
>
>Can we impose the restriction that ONLY assertions in class #2 may be
>set per-operation or per-message on the receiver-side?  
>
>Whereas an assertion of class #1 would only be settable for the entire
>endpoint (as this is the level of granularity of the transport
>listener). In WSDL (1.1) terms, such a policy expression could be
>attached to a portType, but not an operation or message. 
>
>So for example, a security policy that would require a specific type of
>listener to be launched would be limited to endpoint-level granularity
>... e.g. a HTTPS listener with enabled cipher suites satisfying at least
>128 bit encryption could NOT be required by a server-side per-operation
>or per-message policy assertion.
>  
>
Hi Eoghan,

This is an interesting question. I believe class #1 assertions would not 
be considered well designed assertions in the context of WS-Policy as 
they'd exhibit no visible behaviour (behaviour that manifests on the 
wire), see Understanding Web Services Policy -  4. Advanced Concepts II: 
Policy Assertion Design 
<http://msdn2.microsoft.com/en-us/library/ms996497.aspx#understwspol_topic4>
Another problem is that transports are not interceptors (nor interceptor 
providers), so we'd have to find a mechanism that allows transports to 
assert parts of the endpoint vocabulary in the same way as interceptors 
would do. Rather than having each interceptor (and the transport too) 
record the fact that they can assert part of the endpoint's vocabulary 
(this would result in a lot of repetitive code), we could instead, at 
the end of the chain on the server side iterate, back over all 
interceptors *and* the transport, and for each candidate that implements 
the AssertionBuilder API, check if it could satisfy the assertions of 
what now is known to be the effective message/operation policy.
To avoid the instanceOf AssertionBuilder check, transports and 
interceptors can instead do this if they want to participate in the 
assertion process:
List<AssertionBuilder> assertionBuilders = 
Message.get(List<AssertionBuilder>);
if (null == assertionBuilders ) {
    message.put(List<AssertionBuilder>.class, new 
ArrayList<AssertionBuilder>());
}
assertionBuilders.add(this);
Then at the end of the chain (in the PolicyVerificationInterceptor) we 
only check back with those AssertionBuilders that have registered 
themselves.

But back to the distinction of class #1/class #2 assertions: I am not 
sure this can be formalised as an assertion can basically be anything. 
We could assume though that assertions are of class #2 by default unless 
registered otherwise with some sort of registry, e.g. the types used in 
JMS configuration could be registered as class #1 assertions.
We should also switch from
<service name="JMSSOAPServiceAddressing">
        <port binding="tns:HWJMSAddressingBinding" 
name="HWJMSAddressingPort">
            <jms:address ....>
           </jms:address>
        </port>
    </service>

to

<service name="JMSSOAPServiceAddressing">   
        <port binding="tns:HWJMSAddressingBinding" 
name="HWJMSAddressingPort">
            <wsp:policy>
                <jms:address ....>
               </jms:address>
         </wsp:policy>
        </port>
    </service>

> 
>--
>
>On the client-side, for performance reasons I think it would be good to
>avoid traversing the AssertionBuilders for every invocation. Instead, we
>could optimize the process to re-use the policy reconciliation done for
>previous invocations.
>  
>
Yes, we should definitely do that - these are good ideas.

Andrea.

>Off the top of my head, something like the following could work:
>
>1. Maintain flyweighted policy sets, so that these may be compared
>efficiently for equality using the == operator (as opposed to doing a
>deep compare).
>
>2. Cache the interceptor chain resulting from the traversal of the
>AssertBuilders for a particular merged policy set.
>  
>
>3. Maintain a second-level cache (only checked if the cache-lookup from
>step #2 yields nothing) of interceptor chains, which is keyed on
>isCompatible()-style logic as opposed to equality. Dumb example, but
>suppose we consider the RM BaseRetransmissionInterval to be a upper
>bound on the wait for an ACK, i.e. an instruction to resend the message
>if an ACK isn't received at *or before* the interval expiry ... with
>that loose interpretation, an asserted RM BaseRetransmissionInterval of
>500ms could be considered compatible with a previously asserted value of
>250ms for an earlier invocation.
>
>The idea would be to always efficiently determine if there's a
>previously computed interceptor chain available for re-use, rather than
>setting one up from scratch for each invocation.
>
>To allow for the fact that the interceptor chains can be modified
>in-flight (e.g. by removing interceptor(s) from the tail of the chain
>during a message dispatch), I guess we'd need to clone the cached chain
>for each new invocation.
>
>Apologies if I'm showing my ignorance of WS-Policy with any of the above
>suggestions ;) 
>
>Cheers,
>Eoghan
>
>  
>


Mime
View raw message