RFR (S) JDK-8162553: Crash in class unloading due to null CLD having a zero _keep_alive value

[David Holmes]

Hi Lois,

On 9/08/2016 5:41 AM, Lois Foltan wrote:
> Hello,
>
> Please review the following fix:
>
> Webrev:
> http://cr.openjdk.java.net/~lfoltan/bug_jdk8162553/webrev/
>
> Bug: Crash in class unloading due to null CLD having a zero _keep_alive
> value
> https://bugs.openjdk.java.net/browse/JDK-8162553
>
> Summary:
> Limit the increment/decrement of the ClassLoaderData _keep_alive field
> to just ClassLoaderData(s) created specifically for an anonymous class.

That sounds reasonable but I wonder whether this should be checked in 
the calling code rather than inside those methods? Otherwise in the 
calling code there is an appearance that all CLDs need a keep-alive 
count, when they don't. A comment in classLoaderData.hpp that those 
methods only apply to anonymous classes would also add some clarity to 
the API.

Thanks,
David

> Due to the uniqueness that no other class shares an anonymous class'
> ClassLoaderData, no other non-GC thread has knowledge of the anonymous
> class while it is being defined.  By limiting the increment/decrement to
> just ClassLoaderData(s) dedicated to anonymous classes, the need to make
> this an atomic operation is avoided.  This bug was exposed since
> java_lang_Class::create_mirror() is multi-threaded once java.lang.Class
> is loaded during JVM start up.
>
> Testing:
> Several runs of hotspot nightly testing and several iterations of
> various tests that have exposed this issue.
>
> Thanks,
> Lois