RFR(XS) 8191369: NMT: Enhance thread stack tracking

[Zhengyu Gu]

Hi David,

Thanks for the suggestions.

On 11/16/2017 01:58 AM, David Holmes wrote:
> Hi Zhengyu,
> 
> I can't review this in detail as I'm not familiar with NMT workings, but 
> have a couple of comments.
> 
> On 16/11/2017 8:25 AM, Zhengyu Gu wrote:
>> This is Linux only enhancement for now, can be extended for other 
>> platforms. (See bug for details)
> 
> I'm concerned about unnecessary platform skew. I added some comments to 
> the bug. I think the mincore trick can be used on Solaris and AIX as 
> well - but not on OS X / BSD. It may be that VirtualQuery can be used on 
> Windows to get the same information - but I'm unclear if the memory 
> states support what you're looking for. It would be good to extend this 
> to other platforms that will support it.

Windows is definitely doable, I can add Windows support if it helps the 
case. But I don't have access to other platforms.

> 
> And on a minor note can you please correct the existing spelling error 
> in get_stack_commited_bottom. :)
> 
Will do.

Thanks,

-Zhengyu


>> Current implementation, thread stack is tracked when thread is 
>> created, its available stack space is tagged as reserved and committed.
>>
>> However, this is not how stack works. Stack pages are committed on 
>> demand, so this approach overstates its memory usage.
>>
>> This enhancement gathers thread stack usage at NMT query time, so that 
>> it can report more accurate usage.
>>
>> Bug: https://bugs.openjdk.java.net/browse/JDK-8191369
>> Webrev: http://cr.openjdk.java.net/~zgu/8191369/webrev.00/index.html
>>
>>
>> Example:
>>
>> Summary shows thread stacks only use small faction of reserved space.
>>
>> -                    Thread (reserved=41216KB, committed=632KB)
>>                              (thread #41)
>>                              (stack: reserved=41032KB, committed=448KB)
>>                              (malloc=137KB #222)
>>                              (arena=47KB #76)
>>
>>
>> Detail tracking shows stack guards and actually used/committed stack 
>> area.
>>
>> [0x00007f66e95e7000 - 0x00007f66e96e8000] reserved 1028KB for Thread 
>> Stack from
>>      [0x00007f67c0b57d5c] JavaThread::run()+0x2c
>>      [0x00007f67c08bbea2] thread_native_entry(Thread*)+0x112
>>
>>          [0x00007f66e95e7000 - 0x00007f66e95eb000] committed 16KB from
>>              [0x00007f67c08b7319] 
>> os::pd_create_stack_guard_pages(char*, unsigned long)+0x29
>>              [0x00007f67c0b56851] 
>> JavaThread::create_stack_guard_pages() [clone .part.125]+0xa1
>>              [0x00007f67c0b57d6e] JavaThread::run()+0x3e
>>              [0x00007f67c08bbea2] thread_native_entry(Thread*)+0x112
>>
>>          [0x00007f66e96e7000 - 0x00007f66e96e8000] committed 4KB
> 
> Can we capture this in a test so we can tell that the tracking has 
> improved?
> 
> Thanks,
> David
> -----
> 
>> Thanks,
>>
>> -Zhengyu
>>