RFR JDK-8028275: Metaspace ShrinkGrowTest causes fatal error if run with JFR

[David Holmes]

On 7/01/2014 6:48 AM, Zhengyu Gu wrote:
> Hi Coleen,
>
> Thanks for the review.
>
> ExceptionMark's destructor only report pending exception as fatal error after initialize is done (is_init_completed() == true), otherwise, it will call vm_exit_during_initialization(), just as you suggested.  Only after set_init_completed() is called (thread.cpp #3568), pending exception has to be checked.

Right - this subtlety is often overlooked. Though perhaps the simpler 
fix is to reduce the scope of the ExceptionMark so that it is 
"destructed" before is_init_completed() is called ? What you have is 
consistent with other error handling code so I'm okay with it as-is. 
Though a comment re CHECK and the ExceptionMark might be useful after 
set_init_completed().

(All this initialization error handling is an awful mess regardless. It 
should be CHECK_JNIERR not CHECK_0.)

David

>
>
> Thanks,
>
> -Zhengyu
>
>
>
> On Jan 6, 2014, at 3:37 PM, Coleen Phillimore wrote:
>
>>
>> Zhengyu,
>>
>> There are other calls to initialize_class in Threads::create_vm() that can create a pending exception and return with CHECK_0 after the EXCEPTION MARK.
>>
>> It seems to me that the EXCEPTION_MARK should be removed and the caller of this function checks for the pending exception.   Better yet would be to encapsulate the calls to initialize_class into a function(s) and check the return for pending exception and exit the jvm with jvm_exit_during_initialization().
>>
>> Coleen
>>
>> On 01/06/2014 10:55 AM, Zhengyu Gu wrote:
>>> This is another bug related on using EXCEPTION_MARK and CHECK macros. JSR292 initialization is done after init_completed flag is set to true, so any pending exceptions from initializing JSR292 classes can cause CHECK macro to return, and trigger the assertion in ExceptionMark destructor.
>>>
>>>
>>> Bug: https://bugs.openjdk.java.net/browse/JDK-8028275
>>> Webrev: http://cr.openjdk.java.net/~zgu/8028275/webrev.00/
>>>
>>> Test:
>>>    Tested the fix on reported platform.
>>>
>>>
>>> Thanks,
>>>
>>> -Zhengyu
>>
>