RFR: 6957241: ClassLoader.getResources() returns only 1 instance when using jar indexing

[Daniel Fuchs]

On Tue, 31 Aug 2021 12:11:46 GMT, wxiang <github.com+53162078+shiyuexw at openjdk.org> wrote:

> Using jarIndex for Hibench, there is an unexpected behavior with the exception "Exception in thread "main" org.apache.hadoop.fs.UnsupportedFileSystemException: No FileSystem for scheme "hdfs"".
> 
> After investigating it, it is related to the usage of ServiceLoader with JarIndex.
> The below stack shows the issue with JDK11:
> 
> getResource:1016, URLClassPath$JarLoader (jdk.internal.loader)
> getResource:937, URLClassPath$JarLoader (jdk.internal.loader)
> findResource:912, URLClassPath$JarLoader (jdk.internal.loader)
> next:341, URLClassPath$1 (jdk.internal.loader)
> hasMoreElements:351, URLClassPath$1 (jdk.internal.loader)
> hasNext:355, BuiltinClassLoader$1 (jdk.internal.loader)
> hasMoreElements:363, BuiltinClassLoader$1 (jdk.internal.loader)
> next:3032, CompoundEnumeration (java.lang)
> hasMoreElements:3041, CompoundEnumeration (java.lang)
> nextProviderClass:1203, ServiceLoader$LazyClassPathLookupIterator (java.util)
> hasNextService:1221, ServiceLoader$LazyClassPathLookupIterator (java.util)
> hasNext:1265, ServiceLoader$LazyClassPathLookupIterator (java.util)
> hasNext:1300, ServiceLoader$2 (java.util)
> hasNext:1385, ServiceLoader$3 (java.util)
> 
> The below API tries to get all the resources with the same name.
> 
> public Enumeration<URL> findResources(final String name,
>                                      final boolean check) 
>  ```
> After using JarIndex, URLClassPath.findResources only returns 1 URL.
> It is the same as the description in JDK-6957241.
> 
> The issue still exists in JDK18.
> 
> Root cause:
> 
> public Enumeration<URL> findResources(final String name,
>                                      final boolean check) {
>         return new Enumeration<>() {
>             private int index = 0;
>             private URL url = null;
> 
>             private boolean next() {
>                 if (url != null) {
>                     return true;
>                 } else {
>                     Loader loader;
>                     while ((loader = getLoader(index++)) != null) {
>                         url = loader.findResource(name, check);
>                         if (url != null) {
>                             return true;
>                         }
>                     }
>                     return false;
>                 }
>             }
> ...
>         };
>     }
> 
> With the JarIndex, there is only one loader which is corresponding to the jar with the index due to the implementation in JarLoader.getResource(final String name, boolean check, Set<String> visited).
> 
> Loaders corresponding to other jar packages will not appear in this while.
> So it only returns 1 instance.
> 
> To solve the issue, I change the implementation "private boolean next()".
> If the loader has index, traverse the index and get all the resource from the loader.

If there is a way to simplify the UCP code by removing the support for legacy JAR index, and if it doesn't cause regressions, I'm all for it. But given the intricacy of that code I suspect it will be quite an undertaking - and reviewing such a change will probably not be trivial. The current PR is not trivial to review and seems to be adding an other layer of complexity though, so I agree that we should have this discussion.

-------------

PR: https://git.openjdk.java.net/jdk/pull/5316