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

[Coleen Phillimore]

On 04/02/2013 01:02 PM, serguei.spitsyn at oracle.com wrote:
> On 4/1/13 3:51 PM, Coleen Phillimore wrote:
>>
>> On 04/01/2013 06:02 PM, serguei.spitsyn at oracle.com wrote:
>>> On 4/1/13 1:09 PM, Coleen Phillimore wrote:
>>>>
>>>> Hi Serguei,
>>>>
>>>> Sorry for the delay in reviewing this.
>>>>
>>>> In instanceKlass.cpp line 2730, can you make that mtClass since 
>>>> it's part of class metadata?
>>>
>>> Fixed.
>>> Thank you for the catch!
>>>>
>>>> In methodHandles.cpp, I think the whole member name table mechanism 
>>>> could be protected by #if INCLUDE_JVMTI since it's not needed 
>>>> unless you have jvmti built in the vm.   And you don't need to call 
>>>> add_method_handle() or just have it do JVMTI_RETURN or something 
>>>> for the minimal vm.
>>>
>>> My understanding was that John is considering to use the
>>> MemberNameTable in the future for the compilation purposes.
>>> However, I can temporarily place it under #if INCLUDE_JVMTI
>>> until the compiler team decides what to do with it.
>>>
>>>>
>>>> Also, I think you should consider a global table rather than one 
>>>> per class.  I think we discussed this and may have decided that 
>>>> speed of adding these is important, but depending on how many are 
>>>> actually used per class, there's one pointer per class.
>>>
>>> I thought we've reached a consensus and decided to add MNT pointer
>>> per class as it is performance sensitive for indy calls.
>>> Anyway, before changing it to a hashtable I'd want to hear from John 
>>> and Christian first.
>>
>> Definitely would like to hear from John and Christian.   We did 
>> discuss it but I wasn't totally convinced this field is more 
>> important than the others that were going to be moved out of Klass* 
>> to make it smaller.   Some statistics for some sample applications 
>> like eclipse would be great.
>
> I'm not sure what statistics do you mean here.
> Do you mean a footprint or performance statistics?
>
> As I understand from our previous discussion there can be a more 
> general solution for the footprint issue.
> As such we could attack the footprint problem separately.

Yes, there are more general solutions to the footprint issue.   One 
would be moving these JVMTI fields and initializing them lazily and 
there are other footprint improvements.   One additional improvement 
would be to remove this field which will be mostly zero and put this in 
a hashtable and not allocate it per class.

> For instance, we could separate all JVMTI fields from InstanceKlass 
> and initialize them lazily.
> But I'd not want to go deep into this discussion.
> You do not want me to get involved into the footprint work, do you? :)
>
> Also, the decision how to represent the MNT depends on its future 
> usage by the compiler team.
> As we agreed, the compiler team is going to adjust the MNT to their needs
> at some point when it is more convenient for them.
> So that, could we make a final decision when the whole picture is ready?
> It would be better to approach it in some steps.
> Currently, this bug blocks other work on the JVMTI support of jsr-292.

I don't know what the jsr 292 team has in store for this field but it's 
still a footprint cost that's for only a special case.  So this is okay 
if you file a bug so that we can remove it and reimplement this table to 
be global or a hashtable.

Coleen

>
> Thanks,
> Serguei
>
>>
>> Coleen
>>
>>>
>>>> Maybe they could be organized as a hashtable instead like the 
>>>> method handle intrinsics?
>>>>
>>>> The code itself looks good everywhere, except for the concern about 
>>>> footprint.
>>>
>>> I very appreciate you found a time to review it!
>>> There are not many people having an expertize in this area.
>>>
>>>
>>> Thanks,
>>> Serguei
>>>
>>>>
>>>> Thanks,
>>>> Coleen
>>>>
>>>> Can you make this mtClass on line
>>>> On 03/26/2013 04:37 AM, serguei.spitsyn at oracle.com wrote:
>>>>>
>>>>> 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<jweek>.
>>>>> 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:
>>>>> *  src/share/vm/prims/methodHandles.cpp*:
>>>>>     799 int MethodHandles::find_MemberNames(Klass* k,
>>>>>     . . .
>>>>>     803 DEBUG_ONLY(No_Safepoint_Verifier nsv);
>>>>>     It is probably Ok, but, please, let me know if it is not.
>>>>>     The assert related to locking is fired if it is not 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
>>>>>
>>>>>
>>>>
>>>
>>
>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.openjdk.java.net/pipermail/serviceability-dev/attachments/20130402/6b04af7c/attachment-0001.html