[AMQ-3185] Closing a VMTransport can cause all other VMTransports to be prematurely closed

• Type: Bug
• Status: Resolved
• Priority: Critical
• Resolution: Fixed
• Affects Version/s: 5.4.2
• Fix Version/s: 5.5.0
• Component/s: Transport
• Patch Info: Patch Available
• Assignee: Timothy A. Bish
• Reporter: Stirling Chow
• Created: 20/Feb/11 19:14
• Updated: 24/Feb/11 16:52
• Resolved: 24/Feb/11 16:52

Description

Symptom
=======
We have eight servers running AMQ 5.3.1 connected in a network-of-brokers over HTTP. Each broker maintains local connections to internal consumers using the VM transport. We were noticing that about once every day, all the local VM connections on a broker and the outbound network bridge connections would fail with the following error:

2010-12-10 04:29:11,663 [processBroker-process-pool-thread-4] ERROR - The worker encountered an exception and will pause for 5 seconds before continuing.
javax.jms.JMSException: Peer (vm://broker-mbus-200005#1052452) disposed.
at org.apache.activemq.util.JMSExceptionSupport.create(JMSExceptionSupport.java:62)
at org.apache.activemq.ActiveMQMessageConsumer.dequeue(ActiveMQMessageConsumer.java:453)
at org.apache.activemq.ActiveMQMessageConsumer.receive(ActiveMQMessageConsumer.java:570)
at com.invoqsystems.foundation.component.communication.jms.source.JMSMessageSource.getMessage(JMSMessageSource.java:33)
at com.invoqsystems.foundation.component.communication.jms.source.JMSMessageSource.getMessage(JMSMessageSource.java:95)
at com.invoqsystems.foundation.component.communication.jms.worker.MessageProcessingWorker.getTask(MessageProcessingWorker.java:9)
at com.invoqsystems.foundation.component.communication.jms.worker.AbstractWorker.iterate(AbstractWorker.java:14)
at com.invoqsystems.foundation.component.communication.jms.worker.AbstractWorker.runUntilStop(AbstractWorker.java:17)
at com.invoqsystems.foundation.component.communication.jms.worker.AbstractWorker.run(AbstractWorker.java:41)
at java.lang.Thread.run(Unknown Source)
Caused by: org.apache.activemq.transport.TransportDisposedIOException: Peer (vm://broker-mbus-200005#1052452) disposed.
at org.apache.activemq.transport.vm.VMTransport.stop(VMTransport.java:70)
at org.apache.activemq.transport.TransportFilter.stop(TransportFilter.java:64)
at org.apache.activemq.transport.TransportFilter.stop(TransportFilter.java:64)
at org.apache.activemq.transport.ResponseCorrelator.stop(ResponseCorrelator.java:132)
at org.apache.activemq.broker.TransportConnection.doStop(TransportConnection.java:956)
at org.apache.activemq.broker.TransportConnection$3.run(TransportConnection.java:918)
at java.util.concurrent.ThreadPoolExecutor$Worker.runTask(Unknown Source)
at java.util.concurrent.ThreadPoolExecutor$Worker.run(Unknown Source)
... 1 more

This was quite unexpected since we create the VM transport connection at broker startup and maintain it (the single connection) throughout the life of the application. The connection is only closed when the process terminates and the broker is stopped.

Since we wrote our consumers against the JMS spec, we handle periodic connection failures by creating a new connection. This works fine sometimes; however, because of AMQ-3127, the sudden reregistration of our consumers simultaneously occurring with bridge re-creation causes frequent deadlock that can only be resolved by restarting the systems.

Cause
=====
We were unable to reliably recreate the failure, so it became clear that a timing issue was involved. Eventually, we determined the cause of the VM transport failure was due to the following code in VMTransportServer:

VMTransportServer.java

    public VMTransport connect() throws IOException {
        TransportAcceptListener al;
        synchronized (this) {
            if (disposed) {
                throw new IOException("Server has been disposed.");
            }
            al = acceptListener;
        }
        if (al == null) {
            throw new IOException("Server TransportAcceptListener is null.");
        }

        connectionCount.incrementAndGet();
        VMTransport client = new VMTransport(location) {
            public void stop() throws Exception {
                if (disposed) {
                    return;
                }
                super.stop();
                if (connectionCount.decrementAndGet() == 0 && disposeOnDisconnect) {
                    VMTransportServer.this.stop();
                }
            };
        };

        VMTransport server = new VMTransport(location);
        client.setPeer(server);
        server.setPeer(client);
        al.onAccept(configure(server));
        return client;
    }

At issue is the override VMTransport.stop() method:

VMTransportServer.java

            public void stop() throws Exception {
                if (disposed) {
                    return;
                }
                super.stop();
                if (connectionCount.decrementAndGet() == 0 && disposeOnDisconnect) {
                    VMTransportServer.this.stop();
                }
            };

Note that VMTransport.disposed is used to protect against multiple calls and subsequently multiple decrements of connectionCount. However, in the implementation of super.stop(), the disposed flag is only set after the peer transport is informed of the stop:

VMTransport.java

    public void stop() throws Exception {
        stopping.set(true);
        
        // If stop() is called while being start()ed.. then we can't stop until we return to the start() method.
        if( enqueueValve.isOn() ) {
        	
            // let the peer know that we are disconnecting..
            try {
>>>                peer.transportListener.onCommand(new ShutdownInfo());
            } catch (Exception ignore) {
            }
        	
        	
            TaskRunner tr = null;
            try {
                enqueueValve.turnOff();
                if (!disposed) {
                    started = false;
>>>                    disposed = true;
                    if (taskRunner != null) {
                        tr = taskRunner;
                        taskRunner = null;
                    }
                }
            } finally {
                stopping.set(false);
                enqueueValve.turnOn();
            }
            if (tr != null) {
                tr.shutdown(1000);
            }
            

        }
        
    }

TransportConnection implements of peer.transportListener.onCommand(new ShutdownInfo()) by launching a asynchronous task that eventually calls back to the same transport that initiated the closure. If the timing is right, VMTransportServer's VMTransport.stop() method is called a second time before the disposed flag is set to true. As a result, the connectionCount is decremented TWICE instead of just once.

In other words, the diposed check and decrement as implemented by VMTransport's anonymous VMTransport subclass are not thread-safe. If VMTransportServer miscounts the connections, it can end up stopping itself while there are still live connections. The result is that the live connections see their peer (the server part of the VMTransport) unexpectedly closed.

Solution
========
The attached patch prevents multiple decrements of the connectionCount by preventing reentrant calls to VMTransportServer's VMTransport stop() method.

A patch is included which demonstrates the problem with the existing AMQ trunk code.

Issue Links

• relates to: AMQ-3127 Network bridge causes deadlock on queue/topic when message dispatch and consumer registration overlap.