RFR: JDK-8220671: Initialization race for non-JavaThread PtrQueues
rkennke at redhat.com
Wed Mar 27 07:43:47 UTC 2019
>>> I've tested it against my stress test and Shenandoah test suite, and all seems fine.
>>> Is the extra locking around assignment of global state needed? I think the locking in threads_do() is sufficient?
>> The locking over "_all_active = active;" looks sane to me. This patch has no locking in threads_do(), or am I missing something?
> Right, there's no locking in (non_java_)threads_do, by design.
I don't see how this is supposed to work. There's code that modifies the
list, and there's code that iterates the list, and in the absense of
other synchronization (e.g. participation in safepoints), they need to
be protected by the same locks, no? What am I missing?
> But looking at this again, the expanded exposure of the
> NonJavaThreadsList_lock is making me regret that it's also being used
> to protect the synchronize() in remove_from_the_list. If an iterator
> were to try to lock that mutex while some NonJavaThread was on its way
> out, we'd deadlock.
> The one new use of the mutex being added by this
> change doesn't introduce such a problem, but still... I think it's
> easy to avoid that problem by adding a new mutex just to protect the
> synchronize call.
Uh, no? Wouldn't that be missing the point above? We can't just 'avoid'
the problem by protecting something by two mutexes that ought to be
under one mutex.
Or what am I missing?
More information about the hotspot-gc-dev