RFR: 8065634: Crash in InstanceKlass::clean_method_data when _method is NULL

[harold seigel]

Hi Stefan,

Thanks for restoring the 'else' and the 'nullHandle'.  It looks good!

Harold

On 12/5/2014 8:51 AM, Stefan Karlsson wrote:
> Hi Harold,
>
> Thanks for looking at the patch. My first patch used nullHandle and 
> did the freeing was done inside an else statement, but during a 
> pre-review of the patch I got feedback to change it to CHECK_NULL and 
> get rid of the else statement. :)
>
> Here's a new webrev with nullHandle and the else statement:
> http://cr.openjdk.java.net/~stefank/8065634/webrev.02
>
> Thanks,
> StefanK
>
> On 2014-12-03 14:31, harold seigel wrote:
>> Hi Stefan,
>>
>> The change looks good but I have a couple of questions.
>>
>> 1. Why does the call to create_combined_annotations(), in 
>> parseClassFile(), use CHECK_NULL when other calls use 
>> CHECK_(nullHandle) ?
>>
>> 4021     ClassAnnotationCollector parsed_annotations;
>> 4022     parse_classfile_attributes(&parsed_annotations, 
>> CHECK_(nullHandle));
>> 4023
>> 4024     // Finalize the Annotations metadata object,
>> 4025     // now that all annotation arrays have been created.
>> 4026     create_combined_annotations(CHECK_NULL);
>> 4027
>> 4028     // Make sure this is the end of class file stream
>> 4029     guarantee_property(cfs->at_eos(), "Extra bytes at the end of 
>> class file %s", CHECK_(nullHandle));
>>
>>
>> 2. In ~ClassFileParser(), does MetadataFactory::free_array() and 
>> Annotations::free_contents() need to be called if the *_annotation 
>> fields are NULL?
>>
>>
>> Thanks, Harold
>>
>> On 12/3/2014 9:21 AM, Stefan Karlsson wrote:
>>> Hi all,
>>>
>>> Please, review this patch to fix an issue where the GC encounters an 
>>> InstanceKlass before it has come far enough in the class loading 
>>> process.
>>>
>>> Some background to the fix:
>>>
>>> When the JVM loads classes it has to keep track of the allocated 
>>> Metadata, and in the event of an exception deallocate the Metadata. 
>>> The ClassFileParser saves the Metadata pointers so that it knows 
>>> what needs to be deallocated. When the class loading has come far 
>>> enough there's a point were the GC can take over the ownership of 
>>> the Metadata and do deallocation of the both the InstanceKlass and 
>>> the Metadata, if needed. At that point the Metadata pointers in the 
>>> ClassFileParser are cleared and transfered over to the InstanceKlass.
>>>
>>> Most Metadata allocations happen before the InstanceKlass is 
>>> allocated, but the Annotations are allocated after the InstanceKlass 
>>> but before the transferal and clearing of the Metadata pointers in 
>>> the ClassFileParser. If the GC stops when allocating the 
>>> Annotations, it will likely find this InstanceKlass and assume that 
>>> the Metadata are correctly setup. This is exactly what's happening 
>>> in the bug report. The GC finds that the _methods array, 
>>> unexpectedly, points to NULL.
>>>
>>> The proposed solution to this problem is to allocate the Annotations 
>>> before the InstanceKlass is allocated.
>>>
>>> http://cr.openjdk.java.net/~stefank/8065634/webrev.01
>>> https://bugs.openjdk.java.net/browse/JDK-8065634
>>>
>>> Testing:
>>> JPRT, parallel_class_loading testlist, Aurora adhoc run
>>>
>>> Thanks,
>>> StefanK
>>
>