Cluster broken when pre-touch enabled

[wojciech-adaptive]

If you run a cluster node with -Daeron.pre.touch.mapped.memory=true, it will fail with:

java.lang.NullPointerException
	at io.aeron.ClientConductor.logBuffers(ClientConductor.java:1194)
	at io.aeron.ClientConductor.onNewPublication(ClientConductor.java:298)
	at io.aeron.DriverEventsAdapter.onMessage(DriverEventsAdapter.java:147)
	at org.agrona.concurrent.broadcast.CopyBroadcastReceiver.receive(CopyBroadcastReceiver.java:116)
	at io.aeron.DriverEventsAdapter.receive(DriverEventsAdapter.java:68)
	at io.aeron.ClientConductor.service(ClientConductor.java:1214)
	at io.aeron.ClientConductor.awaitResponse(ClientConductor.java:1272)
	at io.aeron.ClientConductor.addPublication(ClientConductor.java:454)
	at io.aeron.Aeron.addPublication(Aeron.java:274)
	at io.aeron.cluster.service.ContainerClientSession.connect(ContainerClientSession.java:107)
	at io.aeron.cluster.service.ClusteredServiceAgent.onSessionOpen(ClusteredServiceAgent.java:411)
	at io.aeron.cluster.service.BoundedLogAdapter.onMessage(BoundedLogAdapter.java:186)
	at io.aeron.cluster.service.BoundedLogAdapter.onFragment(BoundedLogAdapter.java:68)
	at io.aeron.Image.boundedControlledPoll(Image.java:530)
	at io.aeron.cluster.service.BoundedLogAdapter.poll(BoundedLogAdapter.java:125)
	at io.aeron.cluster.service.ClusteredServiceAgent.doWork(ClusteredServiceAgent.java:169)
	at org.agrona.concurrent.AgentRunner.doWork(AgentRunner.java:292)
	at org.agrona.concurrent.AgentRunner.run(AgentRunner.java:165)
	at java.base/java.lang.Thread.run(Thread.java:829)

Can be reproduced by running the master version of SingleNodeTest.shouldSendMessagesToCluster() with that property set. Seems to be an issue since 1.35.0.

[vyazelenko]

There are two calls to add publication for the same channel:

• An async call which succeeds:

  DriverProxy#addPublication: channel=aeron:udp?endpoint=localhost:59150|term-length=128k, streamId=102
  java.lang.Exception
    at io.aeron.DriverProxy.addPublication(DriverProxy.java:74)
    at io.aeron.ClientConductor.asyncAddPublication(ClientConductor.java:493)
    at io.aeron.Aeron.asyncAddPublication(Aeron.java:300)
    at io.aeron.cluster.ClusterSession.asyncConnect(ClusterSession.java:132)
    at io.aeron.cluster.ConsensusModuleAgent.onSessionConnect(ConsensusModuleAgent.java:378)
    at io.aeron.cluster.IngressAdapter.onFragment(IngressAdapter.java:109)
    at io.aeron.ControlledFragmentAssembler.onFragment(ControlledFragmentAssembler.java:122)
    at io.aeron.Image.controlledPoll(Image.java:369)
    at io.aeron.Subscription.controlledPoll(Subscription.java:235)
    at io.aeron.cluster.IngressAdapter.poll(IngressAdapter.java:178)
    at io.aeron.cluster.ConsensusModuleAgent.consensusWork(ConsensusModuleAgent.java:2025)
    at io.aeron.cluster.ConsensusModuleAgent.doWork(ConsensusModuleAgent.java:337)
    at org.agrona.concurrent.AgentRunner.doWork(AgentRunner.java:292)
    at org.agrona.concurrent.AgentRunner.run(AgentRunner.java:165)
    at java.lang.Thread.run(Thread.java:748)
  

• A blocking call that fails with an NPE:

  DriverProxy#addPublication: channel=aeron:udp?endpoint=localhost:59150|term-length=128k, streamId=102
  java.lang.Exception
    at io.aeron.DriverProxy.addPublication(DriverProxy.java:74)
    at io.aeron.ClientConductor.addPublication(ClientConductor.java:453)
    at io.aeron.Aeron.addPublication(Aeron.java:274)
    at io.aeron.cluster.service.ContainerClientSession.connect(ContainerClientSession.java:107)
    at io.aeron.cluster.service.ClusteredServiceAgent.onSessionOpen(ClusteredServiceAgent.java:411)
    at io.aeron.cluster.service.BoundedLogAdapter.onMessage(BoundedLogAdapter.java:186)
    at io.aeron.cluster.service.BoundedLogAdapter.onFragment(BoundedLogAdapter.java:68)
    at io.aeron.Image.boundedControlledPoll(Image.java:530)
    at io.aeron.cluster.service.BoundedLogAdapter.poll(BoundedLogAdapter.java:125)
    at io.aeron.cluster.service.ClusteredServiceAgent.doWork(ClusteredServiceAgent.java:169)
    at org.agrona.concurrent.AgentRunner.doWork(AgentRunner.java:292)
    at org.agrona.concurrent.AgentRunner.run(AgentRunner.java:165)
    at java.lang.Thread.run(Thread.java:748)
  

[mjpt777]

Using the pre-touch highlights a race which was a bug introduced with async adding of publications. It should be fixed with this commit. 54e4be4

[vyazelenko]

I modified the test to run with pre-touch on and off and it passes with the 54e4be4.

[mjpt777]

Even if pre-touch is not enabled this can result in Publication.channel() returning null and a memory leak in the client conductor.