[
https://issues.apache.org/jira/browse/FELIX-712?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12629120#action_12629120
]
Karl Pauls commented on FELIX-712:
----------------------------------
I still would like to see an example of this problem to make sure we fixed it
:-).
Regardless, the hack should not just bootdelegate. Rather, the idea is to find
the first class on the call stack that is not from the class loader that loaded
the Felix classes or is not a class loader or class itself, because we want to
ignore calls to ClassLoader.loadClass() and Class.forName() since we are trying
to find out who instigated the class load.
In case the class determined in this fashion is from the outside _then_ we try
to bootdelegate. In other words, I think we just missed the case of somebody
using Bundle.loadClass. Due to the fact that Bundle is loaded using the
framework classloader and the implementation uses only classes from the same
classloader we do bootdelegate and I agree that we shouldn't in this case.
IMO, we should try to special case for this scenario and fix the magic instead
of moving the responsibility to the user.
> Ability to disable automatic parent classloader delegation
> ----------------------------------------------------------
>
> Key: FELIX-712
> URL: https://issues.apache.org/jira/browse/FELIX-712
> Project: Felix
> Issue Type: Improvement
> Components: Framework
> Affects Versions: felix-1.2.0
> Reporter: Don Brown
>
> In debugging a strange issue why a certain test was passing in IDEA, but
> failing in Maven, we discovered that under certain conditions, Felix will
> delegate to the parent classloader if it detects another class in the stack
> who's classloader differs from the classloader of Felix. Specifically, in
> org.apache.felix.framework.searchpolicy.R4SearchPolicyCore starting in line
> 541, there is the problematic code block that has this comment:
> // At this point, the class/resource could not be found by the bundle's
> // static or dynamic imports, nor its own content. Before we throw
> // an exception, we will try to determine if the instigator of the
> // class/resource load was a class from a bundle or not. This is
> necessary
> // because the specification mandates that classes on the class path
> // should be hidden (except for java.*), but it does allow for these
> // classes/resources to be exposed by the system bundle as an export.
> // However, in some situations classes on the class path make the
> faulty
> // assumption that they can access everything on the class path from
> // every other class loader that they come in contact with. This is
> // not true if the class loader in question is from a bundle. Thus,
> // this code tries to detect that situation. If the class
> // instigating the load request was NOT from a bundle, then we will
> // make the assumption that the caller actually wanted to use the
> // parent class loader and we will delegate to it. If the class was
> // from a bundle, then we will enforce strict class loading rules
> // for the bundle and throw an exception.
> What that means is if if you call Bundle.loadClass() from a non-bundle class
> and you happen to have a class in your stack, which was loaded from a
> different classloader than Felix, Felix will try to fix your problem by
> allowing parent classloader delegation. This was the cause of the breakage
> for us because Maven's Surefire plugin uses its own classloader to load
> several of its classes and Felix found those in the call stack and decided to
> enable parent delegation. In IDEA, all the classes were loaded by the same
> classloader, so our tests passed.
> Specifically, we have code that sits outside Felix and tries to load classes
> from installed bundles. Felix sees that the call, by examining the stack, is
> by code outside a bundle, so it enters this routine that may or may not,
> depending on the execution environment, delegate the class loading to Felix's
> class loader.
> This ticket asks for a configuration setting to be able to disable this bit
> of code.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
