RFR: 8016074 NMT: assertion failed: assert(thread->thread_state() == from) failed: coming from wrong thread state

[Karen Kinnear]

Looks good

thanks Zhengyu for the quick response and to Christian for the very helpful testing,
Karen

On Jul 3, 2013, at 7:32 AM, Zhengyu Gu wrote:

> Solaris os::yield_all() implementation depends on os::sleep(), which requires JavaThread to be in _thread_in_vm state. As NMT tracking call can originate from any thread in any state, transits to _thread_in_vm can be dangerous, if caller holds any locks, deadlock or lock out of order can result, since state transition code may block on Threads_lock.
> 
> The solution is to use os::NakedYield() which calls to system thr_yield().
> 
> On Linux and BSD, os::yield_all() and os::NakedYield() implementations are the same.
> 
> On Windows, the implementations are different, based on our QA feedback, os::NakedYield() does not perform as well as os::yield_all().
> 
> 
> External bug:http://bugs.sun.com/view_bug.do?bug_id=8016074
> JBS:https://jbs.oracle.com/bugs/browse/JDK-8016074
> Webrev:http://cr.openjdk.java.net/~zgu/8016074/webrev.00/
> 
> 
> Tests:
>  - JPRT with -XX:-AutoShutdownNMT flag
>  - Kitchensink on Solaris 11, with -XX:NativeMemoryTracking=detail -XX:-AutoShutdownNMT -XX:+SafepointALot
>    over 12 hours now.
> 
> 
> Thanks,
> 
> -Zhengyu
>