Hi David/Dain, I finally see what's going on here now--and it really makes a lot
of sense. I'm not so sure it's a bug with the classloading process so much as
it's just the way the current code functions. I can't easily show a sequence
diagram here but can briefly explain. It appears the the "EarContext" for a
deployed EAR file is aggregated across successive calls from
EARConfigBuilder.buildConfiguration() to the installModule() method on first the
WebModuleBuilder class, and then second on the EjbModuleBuilder class. This
explains why ClassFinder was working correctly in EjbRefBuilder (i.e. the
deployed module's EarContext had been fully aggregated) but failed for me in
AbstractWebModuleBuilder (i.e., the deployed module's EarContext had not yet
been fully aggregated).
So, rather than fixing something that's not really broken in
AbstractWebModuleBuilder, the best solution in my view is to push the Annotation
processing out of the installModule processing of the module builder(s) up into
the configuration builder. This would allow us to encapsulate the Annotation
processing for deployed EJB applications, Web applications, App Client
applications, and Connectors (not sure these would be annotated though) into a
single class: EARConfigBuilder. Additionally, it would guarantee that we always
have access to a fully aggregated EarContext, and thus a fully populated
classloader to pass to ClassFinder. And finally, I think it would encapsulate
most of the Geronimo annotation processing except for Web Services. This
approach is somewhat in line with my original proposed plan for Annotation
Processing for Geronimo, it's just the conduit has changed somewhat. Do either
of you (or anyone else) have any thoughts, comments and/or concerns ??
Thanks,
Tim McConnell
Tim McConnell wrote:
Hi Dain, What you suggest does make a lot of sense. But for the
stateless-calculator ear file (i.e.,
calculator-stateless-ear-2[1].0-M2.ear) I would then expect to find the
calculator-stateless-ear-2[1].0-M2.jar file on one of the parent
classloaders for the WAR classloader in AbstractWebModuleBuilder. It's
not, so I suspect there is a bug somewhere as you suggest. I shall
investigate further tomorrow. Thanks for the pointer....
Dain Sundstrom wrote:
On Feb 6, 2007, at 12:49 PM, David Blevins wrote:
On Feb 4, 2007, at 7:19 PM, Tim McConnell wrote:
Hi again Dave, after your recommendation below to do all the
annotation discovery during the installModule phase of deployment
ClassFinder is working much better for me. I do still have another
scenario I'd appreciate some advice on. It seems that when an EJB
EAR file (with embedded JAR and WAR files) gets deployed in
Geronimo, there are two builders invoked: e.g.,
TomcatModuleBuilder/AbstractWebModuleBuilder and EJBModuleBuilder
such that the embedded JAR file is not added to the
ClassPath/ClassLoader when the WAR is deployed (I assume this is the
way it should work since I haven't changed it--yet). So, if there
are annotations in the WAR class files pointing to classes in the
JAR file, we'll still encounter NoClassDefException(s). I can just
add the JAR files in the EAR to the classpath of the WAR, which is
what I've done to get around the problem, but I'm not sure this is
the best alternative. Do you have any thoughts?? Thanks much
Those should be added automatically via the deployment system. Very
puzzling. Dain, did you see anything like this when you did that
hack for @EJB annotation support in Servlets?
Nope. The WAR class loader is a child of the EAR class loader so the
WAR should "see" all of the jars in the ear. If that is not the case,
then there is a bug.
-dain
