IssueTracker

Problem
-----------
It is easy to see code written in this form: SomeClass.class.getResourceAsStream(someResource) (see https://docs.oracle.com/javase/8/docs/api/java/lang/Class.html#getResourceAsStream-java.lang.String-). However, this code is problematic when multiple class loaders are involved in a parallel build, resulting in concurrency bugs such as Issue 120277395.


What's wrong with Class.getResourceAsStream()?
-------------------------------------------------------------------
Calling Class.getResourceAsStream() is roughly equivalent to calling Class.getClassLoader().getResourceAsStream().

The implementation of ClassLoader.getResourcesAsStream() is probably safe, but when the class loader is a URLClassLoader, then its overridden version URLClassLoader.getResourceAsStream() has a bug: https://bugs.openjdk.java.net/browse/JDK-8205976. (The bug has some sample code that well explains the issue.)

This bug can manifest when two subprojects are building in parallel in separate class loaders, and both call URLClassLoader.getResourceAsStream(). One subproject is finished first, and the corresponding class loader is closed. However, closing the class loader will also close the underlying jar file that the other project might be using when reading the stream returned by URLClassLoader.getResourceAsStream(). (This is what happened in Issue 120277395.)

Note: Calling Class.getResourceAsStream() is okay if Class.getClassLoader() is not a URLClassLoader, and therefore does not suffer from this bug. However, because we usually don't check the ClassLoader type, it is better to simply avoid calling Class.getResourceAsStream().


What's the alternative?
------------------------------
One alternative is to use the default implementation of ClassLoader.getResourcesAsStream(), which is Class.getResource(someResource).openStream().

I've verified locally that using this alternative, closing the class loader doesn't close the underlying jar file, and it should be thread safe. (It fixes Issue 120277395.)


I'm opening this bug to revise the code in AGP, data binding, and Jetifier with this new understanding.