RFR: 8017588: SA: jstack -l throws UnalignedAddressException while attaching to core file for java that was started with CMS GC

[Coleen Phillimore]

On 7/5/2013 9:20 AM, Erik Helin wrote:
> On 2013-07-05, Erik Helin wrote:
>> Hi Coleen,
>>
>> thanks for reviewing!
>>
>> On 2013-07-03, Coleen Phillimore wrote:
>>> Eric,
>>>
>>> I like the cleanup of instanceKlass.cpp a lot.
>>>
>>> In the serviceability agent,  does getHeap().getOopSize() return 32
>>> for compressed oops?
> Sorry, I forgot to answer this question. With hotspot 24 and compressed
> oops, I do not think that oops in C++ objects such as methodOop in
> vtableEntry gets compressed.
>
> methodOop gets "typedeffed" at compile-time in oops/oopsHierarchy.hpp to
> methodOopDesc*, so it should always have sizeof(char*), i.e. oopSize, as
> size.
>
> Am I correct?

Yes, you are right for hs24.
Coleen

>
> Thanks,
> Erik
>
>>> The vtables and itables are Method*, ie
>>> metadata.   Is there a better size for these?
>> This change is for hotspot 24, so we have a methodOop in vtableEntry :)
>>
>> So, ObjectHeap.getOopSize() will return oopSize in globalDefinitions.hpp
>> which is sizeof(char*). VM.getAddressSize in SA is dependent on
>> platform, for x86-64 it will return
>> MachineDescriptionAMD64.getAddressSize(), which is defined as 8.
>>
>> Therefore, it does not matter if the code in
>> InstanceKlass.getObjectSize() uses ObjectHeap.getOopSize() or
>> VM.getAddressSize(), the result will be the same.
>>
>> In order to keep the SA code and the hotspot code as similar as
>> possible, I think it is better to use ObjectHeap.getOopSize(), since the
>> hotspot code does sizeof(klassOop).
>>
>> Please see a new webrev at:
>> http://cr.openjdk.java.net/~ehelin/8017588/webrev.01/
>>
>> I've also done some further testing on, the following testlists were
>> successfully run:
>> - vm.quick.testlist
>> - vm.tmtools.testlist
>> - nsk.sajdi.testlist
>>
>> in addition to the previous testing:
>> - JPRT
>> - Running jstack -l -F successfully on a java process using -XX:+UseConcMarkSweepGC
>>
>> Thanks,
>> Erik
>>
>>> Another set of tests that I think we run with vm.quick.testlist is
>>> the nsk.sajdi.testlist and the vm.tmtools.testlist in ute.
>>>
>>> Thanks,
>>> Coleen
>>>
>>> On 7/3/2013 4:33 AM, Erik Helin wrote:
>>>> Hi all,
>>>>
>>>> this change fixes an issue where we could not run jstack -l -F on a
>>>> java process running with -XX:+UseConcMarkSweepGC.
>>>>
>>>> The problem originated from the following change in hotspot:
>>>>   changeset:   3707:8bafad97cd26
>>>>   parent:      3693:973046802b6f
>>>>   7158552: The instanceKlsss::_host_klass is only needed for anonymous class for JSR 292 support.
>>>>
>>>>   diff --git a/src/share/vm/oops/instanceKlass.hpp b/src/share/vm/oops/instanceKlass.hpp
>>>>   --- a/src/share/vm/oops/instanceKlass.hpp
>>>>   +++ b/src/share/vm/oops/instanceKlass.hpp
>>>>   @@ -717,9 +763,11 @@
>>>>      {
>>>>        return object_size(align_object_offset(vtable_length()) +
>>>>                           align_object_offset(itable_length()) +
>>>>   -                       (is_interface() ?
>>>>   -                        (align_object_offset(nonstatic_oop_map_size()) + (int)sizeof(klassOop)/HeapWordSize) :
>>>>   -                        nonstatic_oop_map_size()));
>>>>   +                       ((is_interface() || is_anonymous()) ?
>>>>   +                         align_object_offset(nonstatic_oop_map_size()) :
>>>>   +                         nonstatic_oop_map_size()) +
>>>>   +                       (is_interface() ? (int)sizeof(klassOop)/HeapWordSize : 0) +
>>>>   +                       (is_anonymous() ? (int)sizeof(klassOop)/HeapWordSize : 0));
>>>>      }
>>>>
>>>> The corresponding code in the serviceability agent was not updated:
>>>>
>>>>   agent/src/share/classes/sun/jvm/hotspot/oops/InstanceKlass.java:
>>>>    public long getObjectSize() {
>>>>    long bodySize =    alignObjectOffset(getVtableLen() * getHeap().getOopSize())
>>>>                     + alignObjectOffset(getItableLen() * getHeap().getOopSize())
>>>>                     + (getNonstaticOopMapSize()) * getHeap().getOopSize();
>>>>    return alignObjectSize(headerSize + bodySize);
>>>>    }
>>>>
>>>> This fix updates the SA code to be like the hotspot code. I've also
>>>> introduced variables in the hotspot code and in the SA code to make it
>>>> easier to compare the two versions.
>>>>
>>>> Webrev: http://cr.openjdk.java.net/~ehelin/8017588/webrev.00/
>>>>
>>>> Testing:
>>>> - JPRT
>>>> - Running jstack -l -F successfully on a java process using -XX:+UseConcMarkSweepGC
>>>>
>>>> Thanks,
>>>> Erik