Mailing-List: contact torque-dev-help@db.apache.org; run by ezmlm
Sender: dlr@finemaltcoding.com
To: "Turbine Torque Developers List" <torque-dev@db.apache.org>
Subject: Re: TORQUE_3_0_1 CVS Tag
References: <3EF25C6A.1090102@backstagetech.com.au>
From: Daniel Rall <dlr@finemaltcoding.com>
Date: 20 Jun 2003 00:43:15 -0700
Message-ID: <m3u1alw5ws.fsf@despot.finemaltcoding.com>
Lines: 117
User-Agent: Gnus/5.0808 (Gnus v5.8.8) XEmacs/21.4 (Common Lisp)
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii

Scott Eade <seade@backstagetech.com.au> writes:

> Shouldn't the TORQUE_3_0_BRANCH have had a TORQUE_3_0_1
> tag added with the release of 3.0.1?
> 
> I guess there is little point if there is not likely to
> be a 3.0.2 (i.e. efforts are now focused on 3.1).

What do people think about putting this to get this deadlock fix into
a 3.0.2?  I ask because I have seen it fatal to an entire web
application.


dlr         2003/06/19 17:41:18

  Modified:    src/java/org/apache/torque/manager AbstractBaseManager.java
                        MethodResultCache.java
  Log:
  Corrected deadly multi-CPU thread deadlock problem discovered by Ed
  Korthof <ed@apache.org> and John McNally <jmcnally@apache.org>.  The
  problem was due to emulation of synchronization using an int counter
  (to improve performance by avoiding Java "synchronized" keyword).
  Post-increment and decrement operators compile to three op codes (with
  Sun's JDK 1.3.1 for Linux), unsafe on a multi-CPU box.
  
  * src/java/org/apache/torque/manager/AbstractBaseManager.java
    lockCache, inGet, cacheGet(), removeInstanceImpl(),
    putInstanceImpl(): Removed use of lockeCache and inGet instance
    fields, replaced by consistent use of Java's "synchronized" keyword
    (on the current instance, "this").
  
    getMethodResultCache(), addCacheListenerImpl(), createSubsetList(),
    readObject(): Added JavaDoc.
  
  * src/java/org/apache/torque/manager/MethodResultCache.java
    lockCache, getImpl(), putImpl(), get(): Removed use of lockeCache
    instance fields, replaced by consistent use of Java's "synchronized"
    keyword (on the current instance, "this").
  
    remove(): Added error messages to several method overloads.

Ed Korthof <edk@collab.net> supplied some test code:

Subject: test code demonstrating the lack of atomicity in increment/decrement
Date: Wed, 18 Jun 2003 19:16:53 -0700

This took a while to fail on a single CPU box ... but it fails pretty
quickly on a multi-cpu box.

thanks --

Ed
-- 
   +=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=
   |   Ed Korthof   |  edk@collab.net  |   650-228-2527   |
   +=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=
[2. text/plain; test.java]...

import java.util.HashMap;
public class test
{
    int i = 0;
    boolean lock = false;
    HashMap map = new HashMap();
    int testThreads = 2;

    public static void main(String argv[]) throws Exception {
        new test().runTest();
    }

    public void runTest() throws Exception { 
        map.put("test", new Long(System.currentTimeMillis()));
        for (int j = 0; j < testThreads; ++j) {
            Runnable worker = new Runnable() {
                long current = 0; 
                public void run() {
                    while(true) {
                        if (lock) {
                            synchronized(this) {
                                Object item = map.get("test");
                            }
                        }  
                        else {
                            i++;
                            Long item = (Long)map.get("test");
                            current += item.longValue() +
                                System.currentTimeMillis();
                            i--;
                        }
                    }
                }
            };
            Thread nextThread = new Thread(worker, "worker " + j);
            nextThread.start();
            map.put("worker" + j, nextThread);
            System.err.println("created worker: " + j + " => " + nextThread);
        }
        while (true) {
            synchronized(this) {
                lock = true;
                Thread.yield();
                while (i > 0) {
                    System.err.println("i > 0: " + i);
                    Thread.sleep(100);
                }
                map.put("test", new Long(System.currentTimeMillis()));
                lock = false;
                if (i < 0) {
                    System.err.println("i < 0: " + i);
                }
                System.err.println("one iteration");
            }
            Thread.sleep(500);
        }
    }
}