Dear David Holmes, Please see inline response,
On Tue, Feb 3, 2015 at 9:38 PM, David Holmes <david.hol...@oracle.com> wrote: > On 4/02/2015 5:00 AM, Carsten Varming wrote: > >> Greetings all, >> >> I was recently introduced to the deadlock described in JDK-8047720 and >> fixed in JDK9. The fix seems a little messy to me, and it looks like it >> left some very short races in the code. So I decided to clean up the >> code. See attached diff. >> >> The change takes a step back and acquires PeriodicTask_lock in >> WatcherThread::stop (like before the fix in JDK9), but this time >> safepoints are allowed to proceed when acquiring PeriodicTask_lock, >> preventing the deadlock. >> > > It isn't obvious that blocking for a safepoint whilst in this method will > always be a safe thing to do. That would need to be examined in detail - > potential interactions can be very subtle. Absolutely true. For reference, the pattern is repeated with the Terminator_lock a few lines below. The pattern is also used in Threads::destroy_vm before and after calling before_exit, and the java shutdown hooks are called right before the call to before_exit. So there is a lot of evidence that safepoints are allowed to happen in this context. Thank you for taking your time to look at this, Carsten > > David > > > Let me know what you think, >> Carsten >> >>
