RFR (S): 8245509: Crash handler itself crashes when reporting Unsafe.putInt(0) crash

[coleen.phillimore at oracle.com]

On 6/1/20 7:13 PM, David Holmes wrote:
> On 2/06/2020 8:45 am, coleen.phillimore at oracle.com wrote:
>> On 6/1/20 6:36 PM, David Holmes wrote:
>>> On 2/06/2020 7:21 am, coleen.phillimore at oracle.com wrote:
>>>> On 6/1/20 3:57 PM, gerard ziemski wrote:
>>>>> hi Coleen,
>>>>>
>>>>> On 6/1/20 1:33 PM, coleen.phillimore at oracle.com wrote:
>>>>>> In rev2, in the second part, can you move the ResourceMark just 
>>>>>> out of the loop? Otherwise, looks good and I don't need to see 
>>>>>> another webrev. 
>>>>>
>>>>> I thought the ResourceMark should go with the operation it's meant 
>>>>> for, to make the code more self documenting.
>>>
>>> Yes but you don't have to put it right before the op especially in 
>>> loops. I only had it in the loop because I was not applying this 
>>> change:
>>>
>>> -     if (_verbose && _context && Universe::is_fully_initialized()) {
>>> +     if (_verbose && _context && _thread && 
>>> Universe::is_fully_initialized()) {
>>>
>>> which I do not agree with. We are now excluding printing this 
>>> information if we don't have an attached thread. It was okay to 
>>> exclude "printing register info" because the next section printed 
>>> similar info for when we do not have a _thread. It isn't clear to me 
>>> that is the case here.
>>>
>>> I would prefer not to print less information if we don't have to 
>>> exclude it.
>>
>> If there's no current thread (which is 100% unlikely in this case) it 
>> doesn't make any sense to print it's stack.
>
> Well we already try to do just that - if we crash while attaching the 
> current thread for example. But in that case a bunch of error 
> reporting steps break anyway, including this one:
>
> Stack slot to memory mapping:
> stack at sp + 0 slots:
> [error occurred during error reporting (inspecting top of stack), id 
> 0xe0000000, Internal Error 
> (/jdk-dev/open/src/hotspot/share/runtime/thread.hpp:855)]
>
> So I guess we are not losing anything. :(

I had a look at this code this morning, and am not sure most of it would 
work with a null _thread and _verbose == true. Thanks for testing it out.

Coleen

>
> David
>
>> Coleen
>>
>>>
>>> David
>>> -----
>>>
>>>>>
>>>>> Why does it need to go outside pf the loop? To minimize number of 
>>>>> allocations?
>>>>
>>>> It sets an allocation marker in the Arena to allocate from in the 
>>>> current thread.  It's a pretty cheap operation, and it's for error 
>>>> handling, so it can go inside the loop.  Your change is fine.
>>>>
>>>> Coleen
>>>>
>>>>>
>>>>>
>>>>> cheers
>>>>
>>