hs25 review request: 8008511 JSR 292: MemberName vmtarget refs to methods must be updated at class redefinition (original) (raw)

serguei.spitsyn at oracle.com serguei.spitsyn at oracle.com
Tue Mar 26 01:37:16 PDT 2013

Please, review the following fix. The fix was preliminarily reviewed by Coleen, Christian and John but still a final and open jdk review is needed. So that, everyone is welcome to review the fix!

This is open webrev: http://cr.openjdk.java.net/~sspitsyn/webrevs/2013/hotspot/8008511-JVMTI-JSR292.1/

The CR is: http://bugs.sun.com/bugdatabase/view_bug.do?bug_id=8008511 https://jbs.oracle.com/bugs/browse/JDK-8008511

The problem is that the old version of the bootstrap method is re-invoked after a popframe from the bootstrap method execution. It is because the MemberName keeps a stale reference to the old method version.

The solution (suggested by John Rose) is to lazily create and keep up-to-date a MemberNameTable which plays a role of MemberName cache assosiated with the InstanceKlass. Then, at the Class redefinition, this cache is used to replace the old method references in the MemberName's with the new method references.

The MemberNameTable is based on the GrowableArray. A C_HEAP array is allocated at the first call to InstanceKlass::add_member_name(). It is released in the InstanceKlass::release_C_heap_structures().

A global week reference to member name oop is stored in the MemberNameTable. It allowed to avoid having the oops_do() in the MemberNameTable. Also, the MemberNameTable won't hold member name oops in memory.

The MemberNameTable_lock mutex is added to serialize MemberNameTable's updates.

The following No_Safepoint_Verifier has been disabled:

Test coverage: vm.mlvm, nsk.jvmti, nsk.jdi tests on multiple platforms (32 vs 64-bit too). The testing looks good so far.

Thanks, Serguei

More information about the hotspot-dev mailing list