Turning midlet life cycle states and callbacks into an event-driven execution model

One of the most frustrating areas of J2ME is the callback based, magic-thread-spawning approach that the API adopts for delivery of events and life cycle notifications. Basically a midlet can, and will, be constantly interrupted by VM-spawned threads. If you are not extremely careful with your programming the internal state of your midlet will be corrupted, by your own code.

A frequent mistake newbie J2ME programmers make is ignoring how event delivery and callbacks work, or just not bothering to understand them. This is of course a consequence of the grater problem of not understating how concurrency works in Java and how to work with synchronization primitives. Most of the time you can get away with it too, since most mobile VM support only a single flow of execution and just preempt away the current thread.

When the problems hit usually one of two things happen: either the programmers start adding a ton of useless, line-by-line boolean checks that sometimes catch the race condition. Or they dust off their books on Java concurrency and put a mutex check on every object in their code. Enjoy your deadlocks!

The Event-based model

There is a superior alternative. It has been around since the dawn of UI programming and it is jarring that the designers of J2ME choose to ignore it. It is a very simple idea: instead of callbacks and VM threads, you just have a synchronized queue of events. Your main execution loop consumes events from this queue, and the callbacks from the VM do absolutely nothing but add events to this queue.

The main idea is that callback and life cycle code never touches the internal state of your application. It doesn't set booleans. It doesn't add to a counter. It doesn't call a model method. It only, and only, creates an event object and inserts it into a synchronized queue.

How it works in Peak Notes

Here is a simplified version of the Event class in Peak Notes. It is a very simple class that doesn't reference any other object and it is meant to store the data of the callback that created it.

public class Event {
    // add more types when needed
    public static final int POINTER_PRESSED = 1;
    public static final int KEY_PRESSED = 2;
    public int type;
    public int key;
    public int x;
    public int y;
}

Callback methods just create a new object of this class, fill it with the appropriate event data, and queue it in the controller. A pair of example callback methods in the Canvas-derived class:

public void pointerPressed(int x, int y) {
    Event ev = new Event();
    ev.type = Event.POINTER_PRESSED;
    ev.x = x;
    ev.y = y;
    controller.queueEvent(ev);
}

public void keyPressed(int keyCode) {
    Event ev = new Event();
    ev.type = Event.KEY_PRESSED;
    ev.key = keyCode;
    controller.queueEvent(ev);
}

The methods in the controller class allow access to the event queue in a controlled, synchronized way. Since you are not doing model and logic access from the asynchronous callbacks this reduces the mutex usage to just the event queue.

// created in the controller constructor
private Vector eventQueue;

public void queueEvent(Event e) {
    synchronized (eventQueue) {
        eventQueue.addElement(e);
        eventQueue.notify();
    }
}

private Event[] lockAndWaitForEvents() {
    Event[] postEvents = null;
    synchronized (eventQueue) {
        // lock only when there are no existing events
        if (eventQueue.size() == 0) {
                try { eventQueue.wait(); }
                catch (InterruptedException e) { }
            }
        }
        postEvents = new Event[eventQueue.size()];
        eventQueue.copyInto(postEvents);
        eventQueue.removeAllElements();
    }
    return postEvents;
}

The method lockAndWaitForEvents is what you call from your run loop to consume and process new events. If there are no events to process it will wait and sleep, freeing the CPU. Here is a simplified run loop:

public void run() {
    // init code
    ...
    // run loop
    boolean quit = false;
    while (!quit) {
        Event[] events = lockAndWaitForEvents();
        quit = processEvents(events);
    }
    midletInstance.destroyApp(true);
}

processEvents is the method that will call your logic depending on the events received in the queue. Peak Notes uses a complex custom UI system, but on simple games it could be something like:

private boolean processEvents(Event[] events) {
    // this will be modified by the event logic if needed
    boolean doQuit = false;
    for (int i = 0; i < events.length; i++) {
        switch (events[i].type) {
            case Event.POINTER_PRESSED:
            // resolve the event into high level call(s)
            ...
            break;
            case Event.KEY_PRESSED:
            // resolve the event into high level call(s)
            ...
            break;
        }
    }
    return doQuit;      
}

The pain of paint

Given the ridiculously limited and ugly array of UI elements the standard J2ME API provides virtually every midlet uses custom controls, or a fully custom UI. This leads to a hard problem: the only way to paint on screen is by implementing a callback. Which, like all the other callbacks in J2ME, is effectively asynchronous and should be treated like it.

For the Canvas.paint() callback we cannot use an event system since it is only in the execution flow granted by the callback that the midlet is allowed to paint on screen. It is possible to decouple application state into an Event-like class and "double buffer" it, having always a state dedicated to the application model code and another, "dormant" state awaiting for the next time the paint callback is called. This pattern is often used in game programming for more powerful platforms like video consoles. But on J2ME memory is limited and not always possible. It is also complex to implement and can be a real challenge to convince other programmers to follow it.

For Peak Notes I decided to layer a mutex over the logic of the application, just for the paint callback. Fortunately, since all the code that can modifiy the application state is isolated away in a single call, it is extremely easy to implement:

public void run() {
    // init code
    ...
    // run loop
    boolean quit = false;
    while (!quit) {
        Event[] events = lockAndWaitForEvents();
        synchronized (paintMutex) {
            quit = processEvents(events);
        }
    }
    midletInstance.destroyApp(true);
}

And similarly in the paint callback in the Canvas subclass:

public void paint(Graphics g) {
    synchronized (paintMutex) {
        // paint code goes here. it can access the state
        // of the application with no restrictions or locks
    }
}

Conclusion

Multithreading is real even on handheld, single core systems. If J2ME forces you to deal with it, learn how to use it and how it changes your preconceptions on the API.

blogroll

social