getResources in BaseClassLoader, ClassLoaderDomain, and frie

[ttarhan]

Folks,

BaseClassLoader, ClassLoaderDomain and several other classes in JBossCL have public API methods which take Set as a parameter. We can't control what implementation of Set is passed in to these methods, but HashSet is commonly used, causing the URL.hashCode() performance problem (DNS lookup).

I do not understand enough about the public API of JBossCL to know where these methods might be used from. However, I do know that these methods are called internally from several areas in JBossCL with a HashSet.

I see a two-step solution:

1) Change anywhere in JBossCL where HashSet is used to another Set implementation. This is a quick and non-API-breaking change that will actually take care of the most common scenarios.

An example (from user code in the container): ClassLoader.getResources(String) eventually calls BaseClassLoader.loadResources(String) which creates a HashSet and passes that to ClassLoaderDomain.getResources(...) to be filled in, eventually returning the HashSet to the caller. Changing the internal usage of ClassLoaderDomain and friends to use a Set other than HashSet would fix this scenario.

2) Consider if there is an API change required to prevent misuse of this by callers outside of JBossCL. I'm not sure if any of these methods are intended to be used from outside. Some are public, some are not.

I'm willing to tackle solution #1 and may submit a patch later this evening.

However, I need some help/discussion/advice before I could tackle #2.

Thoughts? Comments?

[ttarhan]

Added to JIRA as well, although maybe this is premature?

https://jira.jboss.org/jira/browse/JBCL-118

[ttarhan]

I went ahead and posted a patch to the JIRA to handle item #1 listed above. I also posted a snapshot of jboss-classloader.jar which is built with the patch against tags/2.0.6.GA, thus making it a drop-in replacement for JBoss 5.1.0.GA.

[ttarhan]

As suggested by Adrian in JBCL-117, a better fix is to modify org.jboss.virtual.protocol.vfsmemory.Handler to override the DNS-resolving behavior of hashCode() and equals(). I've attached a patch to the JIRA which does this by overriding getHostAddress(URL) inside org.jboss.virtual.protocol.vfsmemory.Handler to always return null. This is also semantically correct, since vfsmemory "hosts" are not real hosts at all.