Concurrency bug when invoking when:disconnected: and other event handlers

namgk / ambienttalk

Automatically exported from code.google.com/p/ambienttalk

0 stars 0 forks source link

When some event handlers like when:disconnected: are invoked, the scheduling of 
the event 
handlers could happen in the wrong thread (the caller thread instead of the 
thread of the actor 
that owns the event handler).

The following is a manifestation of the bug as submitted by Ben Maene:
Client:

{{{
import /.at.lang.futures;
enableFutures(true);

deftype Server;

def vector := /.at.collections.vector.Vector;

def serverList := vector.new();

whenever: Server discovered: { |server|
    when: server<-getName()@FutureMessage becomes: { |name|
    if: (false == serverList.contains(server))
    then: {
        serverList.add(server);
        system.println("Added server: " + name);

        whenever: server disconnected: {
            system.println("Server offline:" + name);
        };
        whenever: server reconnected: {
            system.println("Server online:" + name);
        };
    };
    };
};

network.online();
}}}

Server:
{{{
deftype Server;

def server := actor: {

    def getName(){
        def username := jlobby.java.lang.System.getProperty("user.name");
        username;
    };

    };

export: server as: Server;
network.online()
}}}

The discovery of the server works, but if the server becomes disconnected and 
the 
when:disconnected: handler is triggered, then the following error occurs:

Thu Oct 22 21:15:20 CEST 2009 INFO at.eventloops.vm - event: 
memberLeft(/192.168.205.1:60723[AmbientTalk]): VM disconnected: 
/192.168.205.1:60723[AmbientTalk]
Thu Oct 22 21:15:20 CEST 2009 DEBUG at.eventloops.remoterefs - 
notifyDisconnected for <far 
ref:behaviour of <actormirror:585824>>
Thu Oct 22 21:15:20 CEST 2009 ERROR at.eventloops.remoterefs - error invoking 
when:disconnected: listener
edu.vub.at.exceptions.XIllegalOperation: Detected illegal invocation of 
base_schedule: sharing via 
Java level of object <obj:12222848{createMessage}@[FuturesModule]>
    at edu.vub.at.objects.coercion.Coercer.invoke(Coercer.java:189)
    at $Proxy1.base_schedule(Unknown Source)
    at edu.vub.at.actors.natives.ELActor.acceptSelfSend(ELActor.java:319)
    at edu.vub.at.eval.Evaluator.trigger(Evaluator.java:494)
    at edu.vub.at.actors.natives.NATFarReference.triggerListener(NATFarReference.java:491)
    at edu.vub.at.actors.natives.NATFarReference.notifyDisconnected(NATFarReference.java:245)
    at edu.vub.at.actors.natives.NATFarReference.disconnected(NATFarReference.java:167)
    at 
edu.vub.at.actors.net.ConnectionListenerManager.notifyDisconnected(ConnectionLis
tenerManage
r.java:138)
    at edu.vub.at.actors.natives.ELVirtualMachine$4.process(ELVirtualMachine.java:237)
    at edu.vub.at.actors.natives.ELVirtualMachine.handle(ELVirtualMachine.java:150)
    at edu.vub.at.actors.eventloops.EventLoop.execute(EventLoop.java:248)
    at edu.vub.at.actors.eventloops.EventLoop$EventProcessor.run(EventLoop.java:269)

Original issue reported on code.google.com by tvcut...@gmail.com on 25 Oct 2009 at 8:34

The problem above is caused by an illegal invocation of the ELVirtualMachine thread (which invokes the when:disconnected: event handler) on the actor mirror. Because the actor mirror in this case is replaced using reflection (in the FuturesModule), the actor mirror is a Coercer and the illegal access was trapped. Cause: the method Evaluator.trigger can be invoked by non-actor threads and may directly call methods like closure.meta_receive or actor.acceptSelfSend, which then calls mirror.base_schedule. Since meta_receive and base_schedule may be overridden using reflection, they can execute arbitrary AmbientTalk code that should be serialized in the owner actor! The Evaluator.trigger method should instead schedule the call to actor.acceptSelfSend to be performed by the actor itself. I still have to figure out what the effect of this change will be on the overall interpreter.

Fixed in interpreter/trunk r1634 The bug was related to the method Evaluator.trigger, which can be invoked by any thread (not only actor threads) to send an 'apply' message to a closure owned by another actor. Unfortunately, Evaluator.trigger directly invoked actor.acceptSelfSend(...), which in turn invokes mirror.base_schedule(...), meaning that an external thread could actually concurrently manipulate another actor's inbox! The solution was to add a new event, event_trigger, to the public interface of an actor: other threads may signal this event if they want the actor to perform the asynchronous self-send. An example where the bug would manifest itself is included in at/unit/bugfixes.at , which now, of course, passes without problems. How could this bug manifest in the first place? I think because in the implementation, non-actor threads like ELFarReference event loops, ELVirtualMachine event loops etc. received direct Java references to the subscribing closures. This enables them to invoke apply on these closures without the interpreter being able to catch the bug. Actually, the bug was discovered by Ben because he had set enableFutures(true), which causes the actor's mirror to be replaced by a custom one. At the implementation-level, this means the actor's mirror is actually a coercer (of type ATActorMirror), and coercers *do* trap illegal method invocations. It was thanks to this trap that I could discover the bug.

namgk / ambienttalk

Concurrency bug when invoking when:disconnected: and other event handlers #35