RFR: 8209971: TestOptionsWithRanges.java crashes in CDS mode with G1UpdateBufferSize=1

Jiangli Zhou jiangli.zhou at oracle.com
Thu Sep 6 02:58:20 UTC 2018


Please review the following fix for JDK-8209971, 
TestOptionsWithRanges.java crashes in CDS mode with G1UpdateBufferSize=1.

     webrev: http://cr.openjdk.java.net/~jiangli/8209971/webrev.00/

     bug: https://bugs.openjdk.java.net/browse/JDK-8209971

This is a pre-existing issue that's exposed when archived heap object 
relocation is enabled. At CDS runtime, the archived heap regions are 
mapped back to the runtime java heap. As the mapped heap data may not 
occupy the entire runtime GC region(s), the empty spaces within the 
regions are filled with dummy objects, which is done by a fixup process 
after mapping the data. The fixup was done after 
SystemDictionary::initialize(), because fixup needs to access 
SystemDictionary::Object_klass() (in some cases). However that is too 
late as archived java objects are accessed when 
SystemDictionary::initialize() resolves well-known classes.

The solution is to call MetaspaceShared::fixup_mapped_heap_regions() at 
an earlier time. In the webre, the call is moved into 
SystemDictionary::resolve_preloaded_classes(). It's called after 
resolve_wk_klasses_through(WK_KLASS_ENUM_NAME(Object_klass)...), but 
before any of the archived java objects is accessed. No mirror object is 
access/restored during 
resolve_wk_klasses_through(WK_KLASS_ENUM_NAME(Object_klass)...). 
Archived mirrors are restored after java.lang.Class is loaded.

Tested TestOptionsWithRanges.java locally in CDS mode. Tier1 - tier5 
normal runs and tier1 - tier4 default CDS runs are in progress.

Thanks,

Jiangli




More information about the hotspot-runtime-dev mailing list