RFR: 8170812: Metaspace corruption caused by incorrect memory size for MethodCounters

[coleen.phillimore at oracle.com]

On 4/6/17 8:50 AM, Ioi Lam wrote:
> On 4/6/17 8:33 PM, coleen.phillimore at oracle.com wrote:
>>
>> I'm confused.  Metaspace is allocated in granularity of 3 pointer 
>> sized words.
>>
> But sizeof(MethodCounters) is pretty big (72 bytes on 
> linux/x86/debug). If it becomes, say 76 bytes, due to some build 
> config options,
>
>           sizeof(MethodCounters) / wordSize
>
> will cause the last 4 bytes to be chopped.

I see.  It isn't that it's too small.   I think all of the allocation 
functions might have this, except the last thing in these classes is 
typically a pointer so they aren't unaligned.

I will sponsor this, Andrew.  Can you check the other ones?

Thanks!
Coleen
>
> - Ioi
>> Coleen
>>
>> On 4/5/17 11:13 AM, Andrew Haley wrote:
>>> If sizeof (MethodCounters) is not a multiple of wordSize, memory
>>> allocator metadata is corrupted, causing the VM to become unstable and
>>> eventually crash.
>>>
>>> The fix is very simple:
>>>
>>> diff -r 85b6ca9458ed src/share/vm/oops/methodCounters.hpp
>>> --- a/src/share/vm/oops/methodCounters.hpp      Wed Mar 29 15:44:34 
>>> 2017 +0000
>>> +++ b/src/share/vm/oops/methodCounters.hpp      Wed Apr 05 15:42:18 
>>> 2017 +0100
>>> @@ -116,7 +116,7 @@
>>>
>>>     AOT_ONLY(Method* method() const { return _method; })
>>>
>>> -  static int size() { return sizeof(MethodCounters) / wordSize; }
>>> +  static int size() { return align_size_up(sizeof(MethodCounters), 
>>> wordSize) / wordSize; }
>>>
>>>     bool is_klass() const { return false; }
>>>
>>> This is very low risk because if the size is already a multiple of
>>> wordSize, this patch will have no effect.  If the size is not a
>>> multiple of wordSize, this patch will prevent an inevitable crash.
>>>
>>> I've applied for a JDK9 fix request.  I'll need a sponsor.
>>>
>>> Andrew.
>>
>