Hi Jaroslav,
we all keep writing finalization code like this... welcome to the club!
I think it would be better :
- never use currentTimeMillis to measure elapsed time; use nanoTime instead
- why use complex Phaser when simple CountDownLatch will do?
I just concidentally wrote this code, which I hope will be a model of best
practice:
/** No guarantees, but effective in practice. */
private static void forceFullGc() {
CountDownLatch finalizeDone = new CountDownLatch(1);
WeakReference<?> ref = new WeakReference<Object>(new Object() {
protected void finalize() { finalizeDone.countDown(); }});
try {
for (int i = 0; i < 10; i++) {
System.gc();
if (finalizeDone.await(1L, SECONDS) && ref.get() == null) {
System.runFinalization(); // try to pick up stragglers
return;
}
}
} catch (InterruptedException unexpected) {
throw new AssertionError("unexpected InterruptedException");
}
throw new AssertionError("failed to do a \"full\" gc");
}
On Thu, Oct 8, 2015 at 12:37 AM, Jaroslav Bachorik <
jaroslav.bacho...@oracle.com> wrote:
> Please, review the following test change
>
> Issue : https://bugs.openjdk.java.net/browse/JDK-8135188
> Webrev: http://cr.openjdk.java.net/~jbachorik/8135188/webrev.00
>
> The problem is described in detail in the issue. This patch follows the
> recommendation by Dan Daugherty and runs the part where finalization is
> being requested in a shutdown handler as a separate process. The test then
> checks the result of this separate process to see whether any errors were
> reported. In order to simplify the logic of the shutdown handler the
> failures are reported by simply printing messages into stderr as opposed to
> throwing 'new Error()'.
>
> The modified test is still passing on all supported platforms.
>
> Thanks,
>
> -JB-
>
