JTReg Test Failure: java/lang/ClassLoader/forNameLeak/ClassForNameLeak.java

[adam-thorpe]

Failure link

https://ci.adoptopenjdk.net/view/Test_openjdk/job/Test_openjdk8_j9_sanity.openjdk_x86-64_linux/50/

• test category: openjdk
• test target name: jdk_lang
• OS/architecture: all
• version: jdk8
• JVM: openj9
• public build SHAs:

05:13:59  OpenJ9   - 57ec9be1c
05:13:59  OMR      - ea20a3b7
05:13:59  JCL      - 467bd9a90a based on jdk8u232-b05

Optional info

• intermittent failure: no
• regression or new test: regression

Failure output (captured from console output)

05:18:23  STDOUT:
05:18:23  returning phantom ref: java.lang.ref.PhantomReference@8216f6a9 to LeakedClassLoader
05:18:23  returning phantom ref: java.lang.ref.PhantomReference@8b4f50a6 to LeakedClassLoader
05:18:23  returning phantom ref: java.lang.ref.PhantomReference@51fb443d to LeakedClassLoader
05:18:23  returning phantom ref: java.lang.ref.PhantomReference@9f8283b4 to LeakedClassLoader
05:18:23  returning phantom ref: java.lang.ref.PhantomReference@cf971d0b to LeakedClassLoader
05:18:23  returning phantom ref: java.lang.ref.PhantomReference@72146db to LeakedClassLoader
05:18:23  returning phantom ref: java.lang.ref.PhantomReference@cd3ae80 to LeakedClassLoader
05:18:23  returning phantom ref: java.lang.ref.PhantomReference@cfb3d405 to LeakedClassLoader
05:18:23  returning phantom ref: java.lang.ref.PhantomReference@845390d9 to LeakedClassLoader
05:18:23  returning phantom ref: java.lang.ref.PhantomReference@833a1201 to LeakedClassLoader
05:18:23  STDERR:
05:18:23  java.lang.RuntimeException: ClassLoader was never enqueued!
05:18:23  	at ClassForNameLeak.main(ClassForNameLeak.java:116)
05:18:23  	at sun.reflect.NativeMethodAccessorImpl.invoke0(Native Method)
05:18:23  	at sun.reflect.NativeMethodAccessorImpl.invoke(NativeMethodAccessorImpl.java:62)
05:18:23  	at sun.reflect.DelegatingMethodAccessorImpl.invoke(DelegatingMethodAccessorImpl.java:43)
05:18:23  	at java.lang.reflect.Method.invoke(Method.java:498)
05:18:23  	at com.sun.javatest.regtest.agent.MainWrapper$MainThread.run(MainWrapper.java:127)
05:18:23  	at java.lang.Thread.run(Thread.java:819)

[JasonFengJ9]

The test code in question is at [1]:

        // Give the GC a chance to enqueue the PhantomReferences
        for (int i = 0; i < 10; i++) {
            System.gc();
        }
        // Make sure all PhantomReferences to the leaked classloader are enqueued
        for (int j = 0; j < THREADS; j++) {
            Reference rmRef = rq.remove(TIMEOUT);
            if (rmRef == null) {
                throw new RuntimeException("ClassLoader was never enqueued!");
            } else {
                System.out.println("Enqueued " + rmRef);
            }
}

This appears a test of RI GC specific behaviour, OpenJ9 might have different GC policies.

@dmitripivkine please comment.

[1] https://github.com/ibmruntimes/openj9-openjdk-jdk8/blob/467bd9a90a42b9cd6db1c7b8b9faf818769c64b0/jdk/test/java/lang/ClassLoader/forNameLeak/ClassForNameLeak.java#L108-L116

[dmitripivkine]

If I am reading test correctly it expects Classloader to be unloaded by System GC. Class unloading decision in OpenJ9 based on heuristics (not mandatory) but it can be forced by using -Xalwaysclassgc command line option. Also we can add -verbose:gc to get gc verbose log output to be sure that actual Global GC has been executed

[JasonFengJ9]

@adam-thorpe does RI Java 8 fail this test as well?

[adam-thorpe]

@JasonFengJ9 No, it passes: https://ci.adoptopenjdk.net/view/Test_openjdk/job/Test_openjdk8_hs_sanity.openjdk_x86-64_linux/60/testReport/

[JasonFengJ9]

@adam-thorpe could you try -Xalwaysclassgc and -verbose:gc as per #7122 (comment)?

[adam-thorpe]

https://ci.adoptopenjdk.net/view/Test_grinder/job/Grinder/775/console not sure what I'm looking for here

[JasonFengJ9]

The grinder with Xalwaysclassgc above still failed.
@dmitripivkine is there anything helpful in the console output or the result file?

[dmitripivkine]

We need to investigate why this test does not work as expected. Either test is bogus and expected behaviour rely on particular implementation or our code is misbehaving somehow. Let me check code.

[dmitripivkine]

GC does check for ((J2SE_VERSION(javaVM) & J2SE_VERSION_MASK) <= J2SE_18) and keep Phantom Reference Referent alive if it is TRUE. Otherwise Phantom behaves as Weak. So Java 8 is last version where this test is valid. @JasonFengJ9 Is Java version equal J2SE_18 for this test? Is it rely on Java 8 or 2.9?

[JasonFengJ9]

Is Java version equal J2SE_18 for this test? Is it rely on Java 8 or 2.9?

The JVM in question is OpenJ9 Java 8, 2.9 was IBM Java internal level. J2SE_18 determines if it is a Java 8 JVM.

[dmitripivkine]

@amicic FYI

[JasonFengJ9]

OpenJ9 JDK 11 passes this test at a 10x grinder https://ci.adoptopenjdk.net/job/Grinder/808/
Also OpenJ9 JDK 8 failed as reported by this issue at a 3x grinder https://ci.adoptopenjdk.net/job/Grinder/807

This test uses multiple (10) threads to stress the call to the test method however each test method performs following:

        ClassLoader classLoader = new URLClassLoader(
                new URL[]{jarFilePath.toUri().toURL()}) {
            @Override public String toString() { return "LeakedClassLoader"; }
        };

        Class<?> loadClass = Class.forName("ClassForName", true, classLoader);
        ((Runnable) loadClass.newInstance()).run();

        PhantomReference<ClassLoader> ref = new PhantomReference<>(classLoader, rq);
        System.out.println("returning phantom ref: " + ref + " to " + classLoader);
        return ref;

In addition, there is private static final ReferenceQueue<ClassLoader> rq = new ReferenceQueue<>(); which is expected to hold the PhantomReference of the ClassLoader.
It appears OpenJ9 JDK 11 enqueued those ClassLoader PhantomReferences but JDK 8 doesn't.

Note: RI JDK 8 passes this test.

[JasonFengJ9]

JDK 11 with -Xnoclassgc -verbose:gc 10x grinder is at https://ci.adoptopenjdk.net/job/Grinder/809/
fyi @dmitripivkine

[dmitripivkine]

Thank you. All jobs failed as we expected.