'Re: [9] RFR (S): 8059340: ConstantPool::_resolved_references is missing in heap dump'

[prev in list] [next in list] [prev in thread] [next in thread] 

List:       openjdk-serviceability-dev
Subject:    Re: [9] RFR (S): 8059340: ConstantPool::_resolved_references is missing in heap dump
From:       Vladimir Ivanov <vladimir.x.ivanov () oracle ! com>
Date:       2015-05-27 10:00:35
Message-ID: 556595C3.7080800 () oracle ! com
[Download RAW message or body]

Unfortunately, it broke class redefinition (see JDK-8081004 [1]) and I 
don't see a reasonable way to fix that.

The problem is that multiple class versions can be alive at the same 
moment w.r.t. class redefinition. Every class version should have its 
dedicated set of resolved_reference slots for invokedynamic/invokehandle 
instructions. They still have dedicated constant pools (and hence 
constant pool caches), but resolved_references array is shared now.

It's not viable to simply append slots to the array during class 
redefinition since it causes memory leak (array size grows 
indefinitely). Compactification strategies are also complex, since slot 
indices should be either (1) stable (they are recorded in CPCEs) and 
slots are reference counted (in order to free and reuse them on next 
class redefinition); or (2) all class versions are patched with new 
indices on every class redefinition.

IMO both approaches complicate things too much compared to keeping 
dedicated resolved_references per ConstantPool instance.

Serguei, Staffan, what do you think? Do you see any viable solutions to 
the problem?

Otherwise, I'm inclined to backout the change.

Best regards,
Vladimir Ivanov

[1] https://bugs.openjdk.java.net/browse/JDK-8081004

On 5/18/15 8:32 PM, Vladimir Ivanov wrote:
> Here's updated version:
>    http://cr.openjdk.java.net/~vlivanov/8059340/webrev.01
>
> Moved ConstantPool::_resolved_references to mirror class instance.
>
> Fixed a couple of issues in CDS and JVMTI (class redefinition) caused by
> this change.
>
> I had to hard code Class::resolved_references offset since it is used in
> template interpreter which is generated earlier than j.l.Class is loaded
> during VM bootstrap.
>
> Testing: hotspot/test, vm testbase (in progress)
>
> Best regards,
> Vladimir Ivanov
>
> On 4/21/15 9:30 PM, Vladimir Ivanov wrote:
>> Coleen, Chris,
>>
>> I'll proceed with moving ConstantPool::_resolved_references to j.l.Class
>> instance then.
>>
>> Thanks for the feedback.
>>
>> Best regards,
>> Vladimir Ivanov
>>
>> On 4/21/15 3:22 AM, Christian Thalinger wrote:
>>>
>>>> On Apr 20, 2015, at 4:34 PM, Coleen Phillimore
>>>> <coleen.phillimore@oracle.com <mailto:coleen.phillimore@oracle.com>>
>>>> wrote:
>>>>
>>>>
>>>> Vladimir,
>>>>
>>>> I think that changing the format of the heap dump isn't a good idea
>>>> either.
>>>>
>>>> On 4/16/15, 2:12 PM, Vladimir Ivanov wrote:
>>>>> (sorry for really late response; just got enough time to return to
>>>>> the bug)
>>>>
>>>> I'd forgotten about it!
>>>>>
>>>>> Coleen, Staffan,
>>>>>
>>>>> Thanks a lot for the feedback!
>>>>>
>>>>> After thinking about the fix more, I don't think that using reserved
>>>>> oop slot in CLASS DUMP for recording _resolved_references is the best
>>>>> thing to do. IMO the change causes too much work for the users (heap
>>>>> dump analysis tools).
>>>>>
>>>>> It needs specification update and then heap dump analyzers should be
>>>>> updated as well.
>>>>>
>>>>> I have 2 alternative approaches (hacky and not-so-hacky :-)):
>>>>>
>>>>>  - artificial class static field in the dump ("<resolved_references>"
>>>>> + optional id to guarantee unique name);
>>>>>
>>>>>  - add j.l.Class::_resolved_references field;
>>>>>    Not sure how much overhead (mostly reads from bytecode) the move
>>>>> from ConstantPool to j.l.Class adds, so I propose just to duplicate
>>>>> it for now.
>>>>
>>>> I really like this second approach, so much so that I had a prototype
>>>> for moving resolved_references directly to the j.l.Class object about
>>>> a year ago.  I couldn't find any benefit other than consolidating oops
>>>> so the GC would have less work to do.  If the resolved_references are
>>>> moved to j.l.C instance, they can not be jobjects and the
>>>> ClassLoaderData::_handles area wouldn't have to contain them (but
>>>> there are other things that could go there so don't delete the
>>>> _handles field yet).
>>>>
>>>> The change I had was relatively simple.  The only annoying part was
>>>> that getting to the resolved references has to be in macroAssembler
>>>> and do:
>>>>
>>>> go through method->cpCache->constants->instanceKlass->java_mirror()
>>>> rather than
>>>> method->cpCache->constants->resolved_references->jmethod indirection
>>>>
>>>> I think it only affects the interpreter so the extra indirection
>>>> wouldn't affect performance, so don't duplicate it!  You don't want to
>>>> increase space used by j.l.C without taking it out somewhere else!
>>>
>>> I like this approach.  Can we do this?
>>>
>>>>
>>>>>
>>>>> What do you think about that?
>>>>
>>>> Is this bug worth doing this?  I don't know but I'd really like it.
>>>>
>>>> Coleen
>>>>
>>>>>
>>>>> Best regards,
>>>>> Vladimir Ivanov
>>>>>
>>>>> On 10/6/14 11:35 AM, Staffan Larsen wrote:
>>>>>> This looks like a good approach. However, there are a couple of more
>>>>>> places that need to be updated.
>>>>>>
>>>>>> The hprof binary format is described in
>>>>>> jdk/src/jdk.hprof.agent/share/native/libhprof/manual.html and needs
>>>>>> to be updated. It’s also more formally specified in hprof_b_spec.h
>>>>>> in the same directory.
>>>>>>
>>>>>> The hprof JVMTI agent in jdk/src/jdk.hprof.agent code would also
>>>>>> need to be updated to show this field. Since this is a JVMTI agent
>>>>>> it needs to be possible to find the resolved_refrences array via the
>>>>>> JVMTI heap walking API. Perhaps that already works? - I haven’t
>>>>>> looked.
>>>>>>
>>>>>> Finally, the Serviceability Agent implements yet another hprof
>>>>>> binary dumper in
>>>>>> hotspot/agent//src/share/classes/sun/jvm/hotspot/utilities/HeapHprofBinWriter.java
>>>>>>
>>>>>>
>>>>>> which also needs to write this reference.
>>>>>>
>>>>>> Thanks,
>>>>>> /Staffan
>>>>>>
>>>>>> On 29 sep 2014, at 16:51, Vladimir Ivanov
>>>>>> <vladimir.x.ivanov@oracle.com <mailto:vladimir.x.ivanov@oracle.com>>
>>>>>> wrote:
>>>>>>
>>>>>>> http://cr.openjdk.java.net/~vlivanov/8059340/webrev.00/
>>>>>>> https://bugs.openjdk.java.net/browse/JDK-8059340
>>>>>>>
>>>>>>> VM heap dump doesn't contain ConstantPool::_resolved_references for
>>>>>>> classes which have resolved references.
>>>>>>>
>>>>>>> ConstantPool::_resolved_references points to an Object[] holding
>>>>>>> resolved constant pool entries (patches for VM anonymous classes,
>>>>>>> linked CallSite & MethodType for invokedynamic instructions).
>>>>>>>
>>>>>>> I've decided to use reserved slot in HPROF class header format.
>>>>>>> It requires an update in jhat to correctly display new info.
>>>>>>>
>>>>>>> The other approach I tried was to dump the reference as a fake
>>>>>>> static field [1], but storing VM internal
>>>>>>> ConstantPool::_resolved_references among user defined fields looks
>>>>>>> confusing.
>>>>>>>
>>>>>>> Testing: manual (verified that corresponding arrays are properly
>>>>>>> linked in Nashorn heap dump).
>>>>>>>
>>>>>>> Thanks!
>>>>>>>
>>>>>>> Best regards,
>>>>>>> Vladimir Ivanov
>>>>>>>
>>>>>>> [1] http://cr.openjdk.java.net/~vlivanov/8059340/static
>>>
[prev in list] [next in list] [prev in thread] [next in thread]
Configure | About | News | Add a list | Sponsored by KoreLogic