RFR 8165246: [REDO] InstanceKlass::_previous_version_count goes negative

[George Triantafillou]

Hi Coleen,

The fix looks good.  One nit:

src/share/vm/oops/instanceKlass.cpp:

3370 // Globally, there is at least one previous version of a class to walk
3371 // during class unloading, which is saved because there are old 
methods in the class
3372 // are still running.   Otherwise the previous version list is 
cleaned up.

Change to:

3370 // Globally, there is at least one previous version of a class to walk
3371 // during class unloading, which is saved because old methods in 
the class
3372 // are still running.   Otherwise the previous version list is 
cleaned up.

-George

On 9/6/2016 12:24 PM, Coleen Phillimore wrote:
> Summary: make _has_previous_version a boolean that is set to true when 
> previous version of a class is added or during class unloading call to 
> purge_previous_versions
>
> The failed fix was because InstanceKlass's previous_versions klasses 
> were walked more than once during class unloading, because the scratch 
> classes are also on the CLD::_klasses list.  This change does three 
> things.  1. Only walk the class if has_been_redefined (which is not 
> set for the scratch class), 2. set next link to NULL for the scratch 
> class when removed from the previous version list, and 3. change the 
> counter to a flag because the counter may never go to zero if a CLD 
> has redefined classes with previous versions and is unloaded.  For 
> this latter case we don't need to walk previous_versions.
>
> I tested this with changes 1 and 2 in Kitchensink, to verify the 
> diagnosis, and then added 3 and retested.
> Also tested with the fix for 
> https://bugs.openjdk.java.net/browse/JDK-8156137 which was checked 
> into hs_-omp repository.
>
> Also ran through RBT 4 tiers nightly tests.
>
> open webrev at http://cr.openjdk.java.net/~coleenp/8165246.01/webrev
> bug link https://bugs.openjdk.java.net/browse/JDK-8165246
>
> Thanks,
> Coleen
>
>