> On 19 Jan 2017, at 19:23, Martin Buchholz <[email protected]> wrote:
>
> Seems OK.
Thanks.
> It would be nice to somehow get more confidence we don't have remaining
> circularities.
From my experience usually these kind of circularities show up intermittently
but regularly enough. Tricky to induce locally.
> I agree a specialized concurrent hash table to replace ConcurrentHashMap
> might help.
>
Yes, it should reliably break the cycle.
> I was surprised to see below a change in parentheses that seems to change the
> behavior.
>
> final long nextSeed() {
> Thread t; long r; // read and update per-thread seed
> - SEED.set(t = Thread.currentThread(),
> - (r = (long) SEED.get(t)) + GAMMA);
> + U.putLong(t = Thread.currentThread(), SEED,
> + r = U.getLong(t, SEED) + GAMMA);
> return r;
> }
The update reverted to the previous code:
http://hg.openjdk.java.net/jdk9/dev/jdk/file/7901a13a051c/src/java.base/share/classes/java/util/concurrent/ThreadLocalRandom.java
<http://hg.openjdk.java.net/jdk9/dev/jdk/file/7901a13a051c/src/java.base/share/classes/java/util/concurrent/ThreadLocalRandom.java>
I added the parentheses to bound the cast.
Thanks,
Paul.
>
>
>
> On Thu, Jan 19, 2017 at 6:52 PM, Paul Sandoz <[email protected]
> <mailto:[email protected]>> wrote:
> Hi,
>
> Unfortunately the switch from Unsafe to VarHandle in LockSupport and
> ThreadLocalRandom caused some circularity issues (similar to those for
> AtomicInteger) that have shown up in continuous testing (but not when i
> tested explicitly using the infrastructure).
>
> The main problem is MethodType leverages ConcurrentHashMap which under
> certain conditions uses TLR [*].
>
> We know we want to replace CHM usage in MethodType with an alternative Peter
> created, which i think we can pursue post 9 and break the circularity.
>
> This patch reverts the back to Unsafe usage:
>
>
> http://cr.openjdk.java.net/~psandoz/jdk9/JDK-8173083-vh-to-unsafe-ls-tlr/webrev/index.html
>
> <http://cr.openjdk.java.net/~psandoz/jdk9/JDK-8173083-vh-to-unsafe-ls-tlr/webrev/index.html>
>
> I did not revert the changes to Striped64 and ForkJoinWorkerThread.
>
> Thanks,
> Paul.
>
> [*]
> 1) The clinit of LockSupport creates a VarHandle;
>
> 2) a VarHandle creates MethodTypes;
>
> 3) that are stored in a CHM;
>
> 4) which on occasion uses TLR
>
> 5) that uses a VarHandle whose VarForm is being initalized in 1)
>
>
>
