hybrid vs pure happens-before   

Hello,
As I understand, MTRAT is a hybrid races detector.
Is there a flag which makes MTRAT act like a pure happens-before detector (i.e. treat Unlock()/Lock() as a happens-before arc)?
Thanks,
--kcc

Are thread start/create and finish/join considered as h-b arcs? (looks like yes)
What about notify/wait? (looks like *NO*)
What are the other event pairs handled as h-b arcs?

>> notify/wait are all supported
Hm. I can't see this support.
Did I do something wrong?

% cat -n NotifyWaitTest.java
1 public class NotifyWaitTest {
2 public static int var;
3 public static boolean done;
4 public static Object lock;
5
6 public static void main (String[] args) {
7 System.out.println("NotifyWaitTest:");
8 lock = new Object();
9 done = false;
10 var = 0;
11 Thread t1 = new Thread() {
12 public void run() {
13 var++;
14 // allow Thread2 to block on 'done'.
15 try { Thread.sleep(3000); } catch (Exception e) {}
16 synchronized (lock) {
17 done = true;
18 System.out.println("setting done to true");
19 lock.notify();
20 }
21 }
22 };
23
24 Thread t2 = new Thread() {
25 public void run() {
26 synchronized (lock) {
27 while (!done) {
28 System.out.println("blocking on done");
29 try { lock.wait(); } catch (Exception e) {}
30 }
31 System.out.println("unblocked");
32 }
33 var++;
34 }
35 };
36 t1.start();
37 t2.start();
38 try { t1.join(); t2.join(); } catch (Exception e) {}
39 System.out.println("var = " + var);
40 }
41 }
% javac NotifyWaitTest.java && ../mtrat/mtrat NotifyWaitTest
NotifyWaitTest:
blocking on done
setting done to true
unblocked
Data Race 1 : 55 : NotifyWaitTest : var
Thread "Thread-1" : Tid 11 : Rid 0 : WRITE
Lock Set : [ ]
Vector Clock : 2
NotifyWaitTest$1 : run : 13
Thread "Thread-2" : Tid 12 : Rid 0 : WRITE
Lock Set : [ ]
Vector Clock : 2
NotifyWaitTest$2 : run : 33

var = 2

Ok. I expect this is not hard to fix.
Few more questions:

1. What are the other events that are handled as h-b?
2. Can you add a flag that will treat Unlock=>Lock as h-b arc (even if this will make the tool terribly slow)?
3. Can you add a pair of annotations that create a custom h-b arc? Something like we have for C++: http://code.google.com/p/data-race-test/wiki/DynamicAnnotations#Message_queue

>> 2. I am sorry I don't quite understand your requirement. What do you really want?

In Threadsanitizer (http://code.google.com/p/data-race-test/wiki/ThreadSanitizer) we have two basic modes:
a) hybrid mode (similar to what MTRAT does) and
b) pure happens-before mode: has much fewer false reports (of course, it is slower, less predictable, and misses some real races)

The difference in the implementation is in two lines:
http://code.google.com/p/data-race-test/source/browse/trunk/tsan/thread_sanitizer.cc?spec=svn1037&r=996#3536
http://code.google.com/p/data-race-test/source/browse/trunk/tsan/thread_sanitizer.cc?spec=svn1037&r=996#3560

This shouldn't be too hard to implement: just handle Unlock as Notify and Lock as Wait (under a special command line flag)

> {quote:title=daniel-laptop wrote:}{quote}
> Could you please tell me how you will use MTRAT after that? Why do you want this flag in MTRAT?

We are using ThreadSanitizer to find races in our C++ code.
It has too basic modes:
1. Conservative (pure-happens-before) which could be used by novice users because it gives very few false positives.
2. Aggressive (hybrid) which is used by the folks who can spend more time de-racing their code.
The hybrid mode has more false positives, but with dynamic annotations they can be defeated.

Ideally, we need a similar thing for Java.
Currently, MTRAT seems unusable because it has too many false reports and no way to suppress them.
These are the things we would need:
a. pure happens-before mode
b. annotations for the hybrid mode
c. support for volatile
d. support for java.util.concurrent (e.g., as I can see, CountDownLatch is not supported)
e. some way to suppress warnings that are not fixable by annotations (e.g. third_party code). Something like valgrind suppressions.

We've assembled few unittests for a Java race detector at http://code.google.com/p/data-race-test/source/browse/#svn/trunk/jtsan
MTRAT has a number of false reports there.

BTW, I am a bit puzzled by the license. It says "You are not authorized to use the Program for productive purposes".
So, I'll probably not be able to use it for testing real software...
??

In the section "1 License", you can say some words, shown below. "internal testing" should be fine.


IBM grants You a limited, nonexclusive, nontransferable
license to download, install, and use the Program during the
evaluation period solely for internal testing and evaluation purposes
and to provide feedback to IBM.

You may make a backup copy of the Program to support such
use. You are not authorized to use the Program for productive
purposes or to distribute the Program or any of its parts. You may
not modify or create derivative works of the Program. The
terms of this license apply to each copy that You make. You must
reproduce all copyright notices and all other legends of ownership on
each copy, or partial copy, of the Program.