Name: elC94547 Date: 03/20/2000
JDK 1.2, 1.2.2 and 1.3 class java.util.Date uses the following
algorithm to maintain many operations like Date.getDay() etc.:
private final int getField(int field) {
if (cal == null) {
if (staticCal == null)
makeStaticCalendars();
synchronized (staticCal) {
staticCal.setTimeZone(TimeZone.getDefault());
staticCal.setTimeInMillis(fastTime);
return staticCal.get(field);
}
}
else {
TimeZone defaultZone = TimeZone.getDefault();
if (!defaultZone.equals(cal.getTimeZone())) {
long ms = cal.getTimeInMillis();
cal.setTimeZone(TimeZone.getDefault());
cal.setTimeInMillis(ms);
}
return cal.get(field);
}
}
Here we see, that operations over the 'staticCal' are embraced
with synchronized (staticCal) {...} section.
But operations over the 'cal' field are not synchronized; and
I'm afraid, that this may cause problems when the class Date
is used by different threads.
In particular, it looks like poor synchronization in the class
Date is causing failures of the stress-test:
nsk/stress/jck12a/jck12a003
which executes hundreds of JCK 1.2a tests as concurrent threads.
The method Date.getDay() intermittently produces an incorrect
result in such executions.
[ Note, that implementation of the method Date.getDay() is the
same in JDK 1.2, 1.2.2 and 1.3. ]
I have observed failures of the method Date.getDay() using the
following environment:
- Dual Pentium-II computer, 2x350MHz, 512Mb RAM
- Windows NT 4.0 Server, SP3, IE4
To reproduce such failure, please try the script DOIT.BAT in
the directory:
/net/sqesvr/vsn/GammaBase/Bugs/<this_bug_number>
======================================================================
Name: elC94547 Date: 03/22/2000
I'm afraid that the description of this bug report was not enough
clear. That was my fault -- I'm sorry! Please, let me explain the
problem in few more details:
Indeed: if a Date object is shared by multiple threads,
^^^^^^^^^^^^^^^^
then operations like following ought to be blocked:
synchronized (d) {
d.setYear(2000 - 1900);
d.setMonth(2);
d.setDay(21);
t = d.getTime();
}
However, the test nsk/stress/jck12a/jck12a003 involve the
JCK tests:
lang/PKGS/pkgs015/pkgs01502/pkgs01502.html
lang/PKGS/pkgs016/pkgs01601/pkgs01601.html
lang/PKGS/pkgs025/pkgs02501/pkgs02501.html
lang/PKGS/pkgs027/pkgs02701/pkgs02701.html
lang/PKGS/pkgs027/pkgs02702/pkgs02702.html
lang/PKGS/pkgs027/pkgs02703/pkgs02703.html
which intermittently fail on much simpler checks like e.g.:
Date d = new Date (96, 6, 10); // July 10,1996
if ( d.getDay () != 3 ) { // Wednesday
out.println ( "not pass 2" );
return 2;
}
In these checks, a Date object is NOT shared by multiple threads.
^^^^^^^^^^^^^^^^^^^^
I'm afraid, that some side effect of manipulating to static
fileds of the class Date may prevent it from returning correct
results when this class (not class instance!) is referenced by
multiple threads.
So, would you take a look at the failures of the test
nsk/stress/jck12a/jck12a003 demonstrated by the script DOIT.BAT
found in the directory /net/sqesvr/vsn/GammaBase/Bugs/4323273,
please?
The failures reproduce quite regularily -- approximately in
3 of 10 executions. So, I believe that those failures still
need to be investigated.
======================================================================