Request for review: 8001341/8004223: SIGSEGV in methodOopDesc::fast_exception_handler_bci_for(KlassHandle, int, Thread*)+0x3e9

[Jiangli Zhou]

Thanks, Coleen!

Jiangli

On 01/07/2013 02:32 PM, Coleen Phillimore wrote:
>
> Jiangli, This looks good.
>
> On 01/07/2013 05:34 PM, Christian Thalinger wrote:
>> On Jan 7, 2013, at 12:30 PM, Jiangli Zhou <jiangli.zhou at oracle.com> 
>> wrote:
>>
>>> Hi,
>>>
>>> Please review the fix for 8001341/8004223: SIGSEGV in 
>>> methodOopDesc::fast_exception_handler_bci_for(KlassHandle,int,Thread*)+0x3e9.
>>>
>>> hs24 webrev: 
>>> http://cr.openjdk.java.net/~jiangli/8001341/webrev_hs24.00/
>>> jdk8 webrev: 
>>> http://cr.openjdk.java.net/~jiangli/8001341/webrev_jdk8.00/
>>>
>>> The SIGSEGV was caused by an unhandled methodOop in 
>>> methodOopDesc::fast_exception_handler_bci_for() and 
>>> JvmtiExport::post_exception_throw(). It's only reproducible on hs24 
>>> as oops in the perm-gen could be moved by the GC. The bug was not 
>>> reported on jdk8. On jdk8, Method* doesn't move. However the fix is 
>>> still needed on jdk8 since the Method* in 
>>> JvmtiExport::post_exception_throw() could be redefined after the  
>>> methodOopDesc::fast_exception_handler_bci_for() call. The handle 
>>> will keep it from being deallocated.
>> I was just going to say it's not required for HS25 but the 
>> redefinition is a valid point.  Although imperceptible.  Could we add 
>> a comment in:
>>
>> src/share/vm/oops/method.hpp
>>
>> to prevent someone undoing this change?
>
> Yes, this is sort of unfortunate because of redefine classes.   The 
> methodHandles that we used for oops because methodOop could move, are 
> still needed for Method* because they can be deallocated.
>
> thanks,
> Coleen
>
>> -- Chris
>>
>>> Tested with -XX:+CheckUnhandledOops on solaris (x64) using 
>>> vm.quick.testlist. Tested with jprt.
>>>
>>> Thanks,
>>> Jiangli
>