RFR: JDK-8129626: G1: set_in_progress() and clear_started() needs a barrier on non-TSO platforms

[bill pittore]

Generally when you have a storestore on the write side you need a 
loadload on the read side to prevent the second read from floating above 
the first one. The reading thread could read in_progress as 0 before it 
reads starting. Meanwhile the write thread writes 1 to in_progress, 
issues storestore, clears starting. Reading thread then reads starting 
as 0. I don't know if the CGC mutex somehow eliminates this issue as I'm 
not familiar with the code in detail.

bill

On 6/23/2015 4:25 PM, Bengt Rutisson wrote:
>
> Hi everyone,
>
> Could I have a couple of reviews for this change?
>
> https://bugs.openjdk.java.net/browse/JDK-8129626
> http://cr.openjdk.java.net/~brutisso/8129626/webrev.00/
>
> We need to add a barrier between the calls to set_in_progress() and 
> clear_started() to make sure that other threads sees the correct value 
> when they use  ConcurrentMarkThread::during_cycle().
>
> Thanks to Per and Bertrand for helping out identifying and sorting 
> this out.
>
> From the bug report:
>
> ConcurrentMarkThread::during_cycle() is implemented as:
>
> bool during_cycle() { return started() || in_progress(); }
>
> So, it checks both ConcurrentMarkThread::_started and 
> ConcurrentMarkThread::_in_progress and they are meant to overlap. That 
> is, we should not set _started to false until after we have set 
> _in_progress to true. This is done in sleepBeforeNextCycle():
>
> void ConcurrentMarkThread::sleepBeforeNextCycle() {
>   // We join here because we don't want to do the 
> "shouldConcurrentMark()"
>   // below while the world is otherwise stopped.
>   assert(!in_progress(), "should have been cleared");
>
>   MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
>   while (!started() && !_should_terminate) {
>     CGC_lock->wait(Mutex::_no_safepoint_check_flag);
>   }
>
>   if (started()) {
>     set_in_progress();
>     clear_started();
>   }
> }
>
> On non-TSO platforms there is a risk that the write to _in_progress 
> (from set_in_progress()) is seen by other threads after the write to 
> _started (in clear_started()). In that case there is a window when 
> during_cycle() may return false even though we are in a concurrent cycle.
>
>
> Thanks,
> Bengt

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mail.openjdk.org/pipermail/hotspot-gc-dev/attachments/20150623/022f96bc/attachment.htm>