Request for reviews (XS): 7011386: race in objArrayKlass::array_klass_impl

[Vladimir Kozlov]

Tom,

Using oop_store() instead of oop_store_without_check()
will generate barriers but it is overkill.
I think marking receiver as volatile should be enough:

   void set_higher_dimension(klassOop k) volatile { oop_store_without_check((oop*) &_higher_dimension, (oop) k); }
   void set_lower_dimension(klassOop k)  volatile { oop_store_without_check((oop*) &_lower_dimension, (oop) k); }

Thanks,
Vladimir

Tom Rodriguez wrote:
> That look good, though I keep wondering if we need a barrier in between or if those fields really should be volatile.  It seems like we're playing a little loose with the locking for these lazy values.
> 
> tom
> 
> On Jan 11, 2011, at 9:54 AM, Vladimir Kozlov wrote:
> 
>> http://cr.openjdk.java.net/~kvn/7011386/webrev
>>
>> Fixed 7011386: race in objArrayKlass::array_klass_impl
>>
>> Other threads may access _lower_dimension field before
>> it is initialized by thread which holds the lock in
>> objArrayKlass::array_klass_impl().
>>
>> Move _lower_dimension field initialization before
>> _higher_dimension.
>