RFR (sh/jdk8): Cleanup weak JNI refs when not doing reference processing

[Zhengyu Gu]

On 12/13/19 9:05 AM, Roman Kennke wrote:
> System dict and string table are done while doing unloading in 
> parallelCleaning.hpp. Not sure about JVMTI. JNI weakroots popped up 
> because they are hidden in reference-processing.

Okay.

In JDK8u, string table is handled in parallel_cleaning, and JvmtiExport 
weak root is folded into JNIHandles.

It is good to go.

Thanks,

-Zhengyu

> 
> Roman
> 
>> Hi Roman,
>>
>> On 12/13/19 8:34 AM, Roman Kennke wrote:
>>> We may run into troubles in jdk8 when doing class-unloading but not 
>>> reference-processing: in JDK8, ref-processing also handles JNI 
>>> weak-refs (it doesn't really belong there, which is why it has been 
>>> moved out in later releases). This means that when we have a cycle 
>>> that doesn't do ref-processing, we may skip cleaning JNI refs and end 
>>> up with dangling refs there. This manifests easily when running 
>>> IntelliJ.
>>>
>>> Fix is to clean JNI weak-refs explicitely when skipping ref-processing.
>>>
>>
>> Looks like we missed more than just JNI weak root, but whole 
>> weak_roots_work() equivalency in sh/jdk11u, which includes jvmti and 
>> system dictionary roots, etc.
>>
>>
>> -Zhengyu
>>
>>
>>> Fix:
>>> http://cr.openjdk.java.net/~rkennke/shjdk8-cleanup-weakjni/webrev.00/
>>>
>>> Testing: hotspot_gc_shenandoah
>>> IntelliJ with default and traversal mode (failed before/passes now)
>>>
>>> Good?
>>>
>>> Thanks,
>>> Roman
>>>
>>
>