Facundo Velazquez created AMQ-6626:
--------------------------------------
Summary: Deadlock caused by ActiveMQMessageConsumer.dispatch() due
to onEvent invocation
Key: AMQ-6626
URL: https://issues.apache.org/jira/browse/AMQ-6626
Project: ActiveMQ
Issue Type: Bug
Components: JMS client
Reporter: Facundo Velazquez
Hi.
In Mule ESB,
I have detected a deadlock caused by the interaction between Spring and
ActiveMQ.
Basically, there are two execution chains that are causing the deadlock:
First Chain: when a Message arrives to the JMS Queue:
1) ActiveMQMessageConsumer.dispatch() gets the lock of
FifoMessageDispatchChannel.mutex.
2) ActiveMQMessageConsumer.dispatch() calls the onMessage() method without
releasing the lock.
3) onMessage Mule Implementation calls
org.springframework.SingleConnectionFactory$SharedConnectionInvocationHandler.invoke()
because It needs a toString of the used resource (in this case the connection)
to logging purposes.
3) in ToString() method,
org.springframework.SingleConnectionFactory.getConnection() method is called
and It takes the org.springframework.SingleConnectionFactory.connectionMonitor
lock to access to connection (that's a shared resource).
Lock ordering in this chain:
FifoMessageDispatchChannel.mutex lock ->
org.springframework.SingleConnectionFactory.connectionMonitor lock
Second Chain: when A reconnection-strategy is applied:
1) Mule stops the current connection calling
org.springframework.SingleConnectionFactory.localStop()
2) LocalStop takes the
org.springframework.SingleConnectionFactory.connectionMonitor lock.
3) After a chain of calls: FifoMessageDispatchChannel.stop() is called.
4) FifoMessageDispatchChannel.stop() takes the FifoMessageDispatchChannel.mutex
lock.
Lock Ordering in this case:
org.springframework.SingleConnectionFactory.connectionMonitor lock ->
FifoMessageDispatchChannel.mutex lock
As you can see, The chains don't take care about lock ordering, and It's
causing a deadlock.
I've sent a mail with same subject to the user list.
Although I agree the best and faster solution is to not use the string() method
of the connection in log (because It's taking a lock), and I'll remove it in
Mule code, I think would be desirable to not call onEvent method without
realising the adquired lock.
Thanks.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
