Thanks of all of you.
This problem was introduced by the contribution from ICU project, and
I'm responsible to the specific code.
-Yoshito
On 10/6/2011 5:59 PM, Naoto Sato wrote:
Hi David,
Thank you for the quick look and the fix!
Naoto
On 10/6/11 10:09 AM, David Holmes wrote:
Hi Chris,
Thanks. Here's the bug:
private List<Locale> getJRELocales() {
if (availableJRELocales == null) {
synchronized (LocaleServiceProviderPool.class) {
if (availableJRELocales == null) {
Locale[] allLocales = LocaleData.getAvailableLocales();
availableJRELocales = new ArrayList<Locale>(allLocales.length); <<====
YIKES we just published an empty ArrayList
for (Locale locale : allLocales) {
availableJRELocales.add(getLookupLocale(locale));
}
}
}
}
return availableJRELocales;
}
availableJRELocales is a static variable shared across all pools. But it
is being published before it gets populated. We need to use a temporary
for the new ArrayList and only assign to availableJRELocales after it is
populated.
In addition, as you mentioned, availableJRELocales needs to be volatile
for this DCL pattern to be correct.
Thanks,
David
On 6/10/2011 9:02 PM, Chris Hegarty wrote:
I see several exceptions, similar to the following (numbers vary):
Exception in thread "Thread-23"
java.util.ConcurrentModificationException: Expected: 96, got: 156
at java.util.ArrayList$Itr.checkForComodification(ArrayList.java:819)
at java.util.ArrayList$Itr.next(ArrayList.java:791)
at java.util.AbstractCollection.addAll(AbstractCollection.java:333)
at java.util.HashSet.<init>(HashSet.java:117)
at
sun.util.LocaleServiceProviderPool.getAvailableLocales(LocaleServiceProviderPool.java:206)
at Interrupter$TestThread.run(Interrupter.java:49)
There would appear to be 156 JRE Locales ( at least on this system ),
modCount is incremented for each add(), but when the iterator is
created
( implicitly during the HastSet.addAll ) it sees a different value for
modCount.
I think there is a visibility issue here. availableJRELocales is
volatile, but the List reference returned from getJRELocales is not. In
the case where availableJRELocales is not null there will be no memory
barrier to force a HB relationship. Or maybe I've gotten this wrong?
His
is quite bizarre, or maybe it is just the overly complicated use of
locking/DCL in this class.
-Chris.
On 10/ 5/11 07:01 PM, David Holmes wrote:
This might not be related to the CME problem, but here:
public static LocaleServiceProviderPool getPool(Class<? extends
LocaleServiceProvider> providerClass) {
LocaleServiceProviderPool pool = poolOfPools.get(providerClass);
if (pool == null) {
LocaleServiceProviderPool newPool =
new LocaleServiceProviderPool(providerClass);
pool = poolOfPools.put(providerClass, newPool);
if (pool == null) {
pool = newPool;
}
}
return pool;
}
we should probably be using poolOfPools.putIfAbsent(providerClass,
newPool)
David
On 6/10/2011 3:35 AM, Naoto Sato wrote:
I will look into this. Reopened the original CR.
Naoto
On 10/5/11 9:58 AM, Alan Bateman wrote:
Chris Hegarty wrote:
Alan, Naoto, David
I filed CR 7098100: java/util/Locale/Bug6989440.java fails
intermittently.
If you're ok with it please review the patch (below) and I can
push it
to the tl repo. Job done!
I assume there is also some underlying issue in the Locale code and
this
might get hidden if we fix the test (I"ve no objection to fixing the
test of course, just thinking that we should study the Locale
code to
try to identify the deadlock or hang or whatever it is that is
causing
the threads in this test not to terminate).
-Alan
