Gianny,
Can you provide some more information on how you might accomplish what
you are proposing? GeronimoTldLocationsCache is a specialization of the
Jasper TldLocationsCache functionality and we configure Jasper to use
our version. So, much of the fundamental logic involved in resolving
the tld resources is embedded in Jasper.
See http://svn.apache.org/viewvc?view=rev&rev=517712 for the original
change and the JIRA https://issues.apache.org/jira/browse/GERONIMO-2955 .
So, if we revert back to the jasper TldLocationsCache our classloader
implementations of getResource() will never be invoked because we will
not have traversed the parent classloaders correctly to add them to the
cache.
Also, given the differences between MultiParentClassLoader and
ChildrenConfigurationClassLoader it is not clear to me how we would walk
the appropriate parentage without some type of instance check. We do
honor much of the classloading logic there today in walking the
parentage and eventually when jasper goes about loading the resources.
Joe
Gianny Damour wrote:
Hi Joe,
Thanks for your investigations. This is a very good catch.
Based on a cursory review of GeronimoTldLocationsCache, I believe that
the implementation can be improved to remove such a weird instanceof
identification mechanism, which is fragile and increases coupling to
Geronimo class loading internals. Furthermore, the implementation simply
skips any logic, e.g. class loading logic, which is encapsulated within
a class loader. An integration analogy would be to directly retrieve
stuff from a database instead of going through the application layer
providing relevant business logic to retrieve the stuff you are looking
for.
My understanding is that ClassLoader.getResources() can be used to
implement GeronimoTldLocationsCache.scanJars without requiring any
dependency on MCCL and CCCL.
Thanks,
Gianny
On 21/11/2008, at 6:39 AM, Joe Bohn wrote:
I finally just checked in a fix for this. Turned out it wasn't
related to the resource processing in the classloader ... but rather
to the jasper navigation of the classloader hierarchy that had to be
extended for the new ChildrenConfigurationClassLoader. See
http://svn.apache.org/viewvc?rev=719329&view=rev
Joe
Joe Bohn wrote:
I think I'm getting it narrowed down a bit. It seems that when
ChildrenConfigurationClassLoader.getResource(name) is called it never
actually resulting in invoking
MultiParentClassLoader.getResource(name) (at least not in this case).
The MPCL should have been passed in when the CCCL was constructed.
CCCL.getResource calls super.getResource (super being
SecureClassLoader) which should ultimately result in MPCL.getResource
being invoked - but that doesn't seem to be happening. Perhaps we
should update CCCL.getResource to directly invoke MPCL.getResource
rather than super.getResource?
I guess the other possibility is that the parent isn't what I think
it should be ... still investigating.
Joe
Joe Bohn wrote:
Gianny Damour wrote:
On 19/11/2008, at 5:19 AM, Joe Bohn wrote:
Joe Bohn wrote:
Just a heads up that I *think* there are still some issues with
this change.
It appears that the hiddenResource processing from the
MultiParentClassloader was removed.
Correction ... it was not removed but rather changed and
relocated. I'm still looking to understand why this is causing a
problem.
Hi,
Indeed, I changed the implementation to properly encapsulate class
loading rules. The new implementation is way cleaner this way; when
my frustration coming from reported issues will reduce, I may use
this better encapsulation to add import and export class loading
rules.
I agree that what you added for the ClassLoadingRules is cleaner.
However, I think it might be the ChildrenConfigurationClassLoader
and it's replacement of MultiParentClassLoader in the Configuration
that might be causing us problems. The testsuite failures that I
mentioned are failing as well as nearly all of the jstl tck tests
since this change. Perhaps it is working as designed, but it seems
strange to me that as we process the "parent" classloaders they are
all of type "ChildrenConfigurationClassLoader" when they used to be
"MultiParentClassLoaders".
BTW, I've verified that these tests were not failing prior to this
change and begin to fail with just the change a few other necessary
changes (the reverting of the xsd changes and fixing the dependency
history files).
I think this has resulted in some
testsuite failures involving tld processing. There are failures
in the web-testsuite/test-2.1-jsps and web-testsuite/test-myfaces
tests. I'm looking at what would be necessary to add in the
hiddenResource logic again hoping that will resolve the issue.
This is a really nice feature and it is great to have the
capability. However could you please run the testsuite in the
future to avoid problems like this (especially when introducing
fundamental changes like this)?
If this is indeed a bug related to this change, then let's ensure
that we have a proper unit test to prevent regression. Let's also
ensure that this test is collocated with the proper component to
improve cohesion. I have been observing various changes, many of
them do not have proper unit tests and this causes problems further
down the path as other developers can not safely change code.
I'm fine with that, but I'm sure there are probably a number of more
obscure situations that aren't covered by the unit tests ... they
can only do so much. That's one of the reasons for the testsuite.
Regarding private-classes, the current Geronimo <-> OpenEJB DD
coupling is unfortunate. Does the OpenEJB deployer needs to know
about the Geronimo environment? If it does not need to know about
it, then let's strip from the DD the environment element and pass
to OpenEJB the stripped version. This will reduce a little bit
coupling. Another approach is to transform the Geronimo DD into an
OpenEJB supported DD when it is handed over to OpenEJB (in this
case, we simply remove the private-classes). The creation of a
canonical DD representation between Geronimo and OpenEJB will
reduce coupling.
I let you re-revert the private-classes change.
I can't unrevert these changes since I'm not a committer on OpenEJB
... but perhaps we should wait on that anyway until we resolve the
JSTL problems.
Thanks,
Gianny
BTW, I'd personally like to see the plan changes re-introduced
for private-classes if it turns out that we need an OpenEJB
release anyway (and at this point in time I think that is the
case). I think users are more accustomed to using declarative
plans for this type of thing at the moment and would find this
helpful.
Thanks,
Joe
removed content of original change from this discussion thread.
