Mailing-List: contact commits-help@harmony.apache.org; run by ezmlm
Precedence: bulk
Reply-To: dev@harmony.apache.org
Message-ID: <28713272.1172689197071.JavaMail.jira@brutus>
Date: Wed, 28 Feb 2007 10:59:57 -0800 (PST)
From: "Salikh Zakirov (JIRA)" <jira@apache.org>
To: commits@harmony.apache.org
Subject: [jira] Updated: (HARMONY-3267) [drlvm][thread] Thread.stop()
 implementation is broken
In-Reply-To: <23379805.1172685237090.JavaMail.jira@brutus>
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Transfer-Encoding: 7bit


     [ https://issues.apache.org/jira/browse/HARMONY-3267?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Salikh Zakirov updated HARMONY-3267:
------------------------------------

    Attachment: StopBlockedOnMonitor.java

The test StopBlockedOnMonitor tests how VM can stop a thread blocked on a genuine MONITORENTER instruction (rather than a re-acquire of monitor on wait() exit of StopBlocked test). 

This test shows surprising results:

Bea has some race with 2 possible outcomes: (1) can't stop thread, (2) the stopped thread falls through MONITORENTER instruction as if it managed to enter a monitor currently hold by other thread. Looks like a bug in VM.

Sun, J9: fail

DRLVM: passes.
DRLVM works okay on this test because of "thin lock" optimization used in DRLVM. The first thread to contend for a monitor first spins in a loop, and inflates a lock only after it manages to grab a lock. Since busy loop has a safepoint in it, the thread notices stop_callback requested, and successfully throws an exception. 

A modifcation of test to make sure the lock is inflated before blocking on it makes all VMs fail in a stable manner:

--- /home/sszakiro/Desktop/StopBlockedOnMonitor.java    2007-02-28 21:19:44.809276900 +0300
+++ StopBlockedOnMonitor.java   2007-02-28 21:53:38.807620500 +0300
@@ -5,6 +5,11 @@
 
     public static void main(String[] args) {
         try {
+            synchronized (lock) {
+                // use wait() to inflate a lock
+                lock.wait(1);
+            } 
+
             // grab the monitor and keep it forever
             synchronized (lock) {
                 StopBlockedOnMonitor t = new StopBlockedOnMonitor();


> [drlvm][thread] Thread.stop() implementation is broken
> ------------------------------------------------------
>
>                 Key: HARMONY-3267
>                 URL: https://issues.apache.org/jira/browse/HARMONY-3267
>             Project: Harmony
>          Issue Type: Bug
>          Components: DRLVM
>            Reporter: Salikh Zakirov
>            Priority: Minor
>         Attachments: StopBlocked.java, StopBlockedOnMonitor.java, StopWaiting.java
>
>
> The current implementation of Thread.stop() in DRLVM is completely broken.
> The attached tests StopWaiting.java and StopBlocked.java demonstrate current incorrect behaviour:
> DRLVM can't stop the thread which is blocked either waiting or entering monitor.
> The waiting case can be fixed by checking the thread state, setting some checked status and waking thread up, like the patch below.
> However, the blocking case cannot fixed in this way, as we have no easy way to wake up a thread which is blocked on a pthread_mutex_lock or EnterCriticalSection OS call.
> It looks like a correct implementation of stopping blocked thread will require either pthread_kill() or SuspendThread() and modifying thread context (like throwing an exception or longjmp).
> Both tests pass on competition VMs (Sun, Bea, Ibm).
> ----8<-----
> A limited fix for a waiting case only.
> --- vm/thread/src/thread_java_basic.c
> +++ vm/thread/src/thread_java_basic.c
> @@ -416,7 +416,23 @@ void stop_callback(void) {
>      tm_native_thread->suspend_request = 0;
>      hysem_post(tm_native_thread->resume_event);
>      
> +    // Does not return if the exception could be thrown straight away
>      jthread_throw_exception_object(excn);
> +
> +    // getting here means top stack frame is non-unwindable.
> +
> +    if (tm_native_thread->state &
> +            (TM_THREAD_STATE_SLEEPING | TM_THREAD_STATE_WAITING_WITH_TIMEOUT
> +                | TM_THREAD_STATE_WAITING | TM_THREAD_STATE_IN_MONITOR_WAIT
> +                | TM_THREAD_STATE_WAITING_INDEFINITELY | TM_THREAD_STATE_PARKED)) {
> +        // This is needed for correct stopping of a thread blocked on monitor_wait.
> +        // The thread needs some flag to exit its waiting loop.
> +        // We piggy-back on interrupted status. A correct exception from TLS
> +        // will be thrown because the check of exception status on leaving
> +        // JNI frame comes before checking return status in Object.wait().
> +        // Interrupted status will be cleared by 
> +        hythread_interrupt(tm_native_thread);
> +    }
>  }
>  
>  /**

-- 
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.