I opened https://issues.apache.org/jira/browse/DIRSERVER-1023 for
this and attached my patch. This patch includes the xbean-spring
patch and the server.xml is for xbean-spring. All the integration
tests pass and the server works fine in the geronimo directory plugin.
Here's how the interceptor configuration now looks (using the xbean-
spring format):
<interceptors>
<normalizationService/>
<authenticationService/>
<referralService/>
<authorizationService/>
<defaultAuthorizationService/>
<exceptionService/>
<operationalAttributeService/>
<!-- Uncomment to enable the password policy service
<passwordPolicyService/>
<keyDerivationService/>
-->
<schemaService/>
<subentryService/>
<collectiveAttributeService/>
<eventService/>
<triggerService/>
<!-- Uncomment to enable replication service
<replicationService >
<replicationConfiguration>
<replicationConfiguration
serverPort="10390"
peerReplicas="[EMAIL PROTECTED]:10392">
<replicaId>
<bean xmlns="http://www.springframework.org/schema/
beans" class="org.apache.directory.mitosis.common.ReplicaId">
<constructor-arg>
<value>instance_a</value>
</constructor-arg>
</bean>
</replicaId>
</bean>
</replicationConfiguration>
</replicationService>
-->
</interceptors>
The only interceptor with any configuration is the
ReplicationService. It now has a setter for ReplicationConfiguration
which spring will call. Interceptors are still initialized by the
InterceptorChain which I am not thrilled about but changing that
would definitely be too big to get into 1.5.1.
I don't know how to test the ReplicationService so if someone who
does could try it out that would be great.
Some other comments on the patch:
- Interceptors now have names. These default to the class name of
the interceptor, which makes it impossible to accidently insert 2 of
the same interceptor into the same chain by for instance giving them
different names. Since the ReplicationService has some
configuration, you can set it's name.
- there are a lot of bypass lists. I changed them so it's easier to
see which interceptors are missing by commenting out the lines for
missing interceptors rather than removing the lines.
- I eliminated the double-checked-locking problem by making the non-
final pointers in Entry volatile. I think it would make a lot more
sense to:
a. have an arraylist of interceptors or entries rather than a double-
linked-list
b. do not change these lists after they are set up but rather create
a new one and swap it with the old one. This would mean there is
only one volatile needed for thread safety
c. consider having different lists of interceptors or entries for
different purposes rather than computing skips on each traversal from
the bypass set.
This is IMO really a fairly small change that considerably simplifies
the server.xml and removes the silliest use of configuration objects,
so I'd like to see it get into 1.5.1. If people like the idea but
want to stick with plain spring server.xml I can come up with a new
patch.
thanks
david jencks
On Aug 18, 2007, at 2:21 PM, David Jencks wrote:
On Aug 18, 2007, at 12:14 PM, Alex Karasulu wrote:
I glad to see you involved and thinking about these issues and
aspects
of the server David. I know we're going to gain a lot from your
knowledge
in these areas.
With respect to the Spring discussions I think we hit a stalemate
a few
weeks back. I know I've spoken to both you and Chris about this
who feel
Spring can do much of what we would like to do. As you know my
stance
is to avoid being dependent on Spring.
However I think we have some idea of how we can remain container
independent
while still leveraging Spring. We still need some discussions on
this but I don't
think they're going to happen in the next 3 days.
So maybe we're in danger of confusing 2 issues:
1. spring vs some other container. On this I don't care very much
but I haven't seen anything proposed that is as easy to use,
flexible, and non-invasive as spring. In any case surely 1.5.1 is
going out with spring as the container. (my xbean-spring proposal
fits under this item). Designing components to be container
agnostic is a really good goal.... leading to
2. reasonable component design. To me this includes not putting
container functionality into the components. Currently at least
the [Mutable]InterceptorConfiguration classes are definitely mostly
doing stuff that should be done by the container, whatever it might
end up being. One result is really ugly spring configuration.
I'm proposing to work on (2) a bit, I think it would make the
server.xml quite a bit more comprehensible both for plain spring
and xbean-spring versions of server.xml
Keep in mind 1.5.1 is a feature release where we're just releasing
to get feature
regardless of stability into the hands of our users. We can
release at any point
in time again thereafter so don't feel pressured to pump your
changes into this
release.
well, I think the geronimo directory plugin will probably get
released against 1.5.1 and the interceptor configuration is really
bugging me and I have a couple extra hours.... so I thought I'd see
what I could do here.
thanks
david jencks
Alex
On 8/18/07, David Jencks <[EMAIL PROTECTED]> wrote:
A while back there was some discussion about just how much container
functionality apacheds needs to duplicate for spring, and IIUC there
was general agreement that [Mutable]InterceptorConfiguration could be
replaced by just letting the container create the interceptors
directly.
I started looking at this issue this morning and think it would be
pretty easy to fix so I'm wondering if there's any chance it could
get into 1.5.1.
I also see what looks to me like double-checked-locking problems with
InterceptorChain setting up the chain of Entry objects. IIUC there
are a bunch of methods to insert,remove, etc interceptors in the
running server but the actual code that traverses the interceptor
chain is not at all synchronized. This is a double-checked-locking
scenario and can lead to the live interceptor chain using
incompletely initialized interceptor objects. The usual fix nowadays
for such problems is to make the variables pointing to the possibly
incompletely initialized objects volatile. IIUC this would be the
InterceptorChain.Entry prevEntry and nextEntry fields.
AFAICT the runtime change-the-chain methods are not called in the
apacheds code base at the moment, so I think an acceptable
alternative would be to remove those methods and rely on no one
starting to use the server until it's fully initialized. This would
involve removing the InterceptorChain addFirst, addLast, addBefore,
addAfter methods.
OK... back to the how-much-extra-wrapping question.
I haven't started to code yet but I think the best solution for now
will be to:
- In each interceptor class that actually uses an interceptor
configuration, replace that with individual attributes (I think only
the replication interceptor has any attributes at the moment)
- remove the InterceptorConfiguration cfg argument from
Interceptor.init
At this point we won't need InterceptorConfiguration any more, the
container (spring) can create the interceptors itself, and
InterceptorChain will get a list of interceptors it needs to
initialize rather than a list of InterceptorConfigurations.
thanks
david jencks
