Hi Venkat,
On 11/10/17 13:07, Venkateswara R Chintala wrote:
Hi,
In a multi-threaded environment, when java.util.SimpleTimeZone object
is used to create a default timezone, there can be a race condition
between the methods java.util.Timezone.getDefault() and
java.util.Timezone.getDefaultRef() which can result in inconsistency
of cache that is used to validate a particular time/date in DST.
When a thread is cloning a default timezone object (SimpleTimeZone)
and at the same time if a different thread modifies the time/year
values, then the cache values (cacheYear, cacheStart, cacheEnd) can
become inconsistent which leads to incorrect DST determination.
We considered two approaches to fix the issue.
1)Synchronize access to cloning default timezone object when cache is
being modified.
2)Invalidate the cache while returning the clone.
We preferred the second option as synchronization is more expensive.
Unfortunately, the SimpleTimeZone.invalidateCache() is also
synchronized. So of the same cost, and it may inhibit JIT optimization.
Perhaps it would be best to synchronize cloning and then arrange the
ZoneId.systemDefault() codepath to not involve cloning. I refrained from
doing that in a patch for:
https://bugs.openjdk.java.net/browse/JDK-8074002
simply because it was easier and benchmarks showed that cloning is
optimized away. But now we should reconsider that and use
TimeZone.getDefaultRef() from the ZoneId.systemDefault() (introducing
JavaUtilAccess into SharedSecrets mechanism)...
Regards, Peter
