源码阅读系列:EventBus

时间:2021-03-30 16:44:18

title: 源码阅读系列:EventBus

date: 2016-12-22 16:16:47

tags: 源码阅读

EventBus 是人们在日常开发中经常会用到的开源库,即使是不直接用的人,也多少借鉴过事件总线的用法。而且EventBus的代码其实是非常简单的,可以试着阅读一下。

源码阅读系列不采用对功能进行归类的方法进行阅读,而是采用一个刚开始阅读源码的视角,从我们平时的API调用,一步步的去理解设计意图和实现原理。

从这里开始

从这里开始吧,我们最常用的地方就是给一个函数添加上注解,我们先抛开apt生成的table,只看这个运行时版本的订阅设定。

// eventbus/Subscribe
@Documented
@Retention(RetentionPolicy.RUNTIME)
@Target({ElementType.METHOD})
public @interface Subscribe {
ThreadMode threadMode() default ThreadMode.POSTING; /**
* If true, delivers the most recent sticky event (posted with
* {@link EventBus#postSticky(Object)}) to this subscriber (if event available).
*/
boolean sticky() default false; /** Subscriber priority to influence the order of event delivery.
* Within the same delivery thread ({@link ThreadMode}), higher priority subscribers will receive events before
* others with a lower priority. The default priority is 0. Note: the priority does *NOT* affect the order of
* delivery among subscribers with different {@link ThreadMode}s! */
int priority() default 0;
}

这个设定还是非常简单的,而且都是我们熟悉的东西,线程类型(默认的是抛出线程),是否是粘性事件,时间的优先级。经过这个类的出现,我们就可以在类里面写我们经常写的某个函数是订阅函数了。

@Subscribe (...)
public void getMessage(Event event) { ... }

下面的问题是我们改怎么让EventBus找到这些方法呢?通过apt的版本我们知道这里面肯定有一个map或者是table的东西记录了Object和Method之间的订阅关系,而且还是一对多的。这个地方就是从每个我们进行register的地方进行的。

register & unregister

// eventbus/EventBus
/**
* Registers the given subscriber to receive events. Subscribers must call {@link #unregister(Object)} once they
* are no longer interested in receiving events.
* <p/>
* Subscribers have event handling methods that must be annotated by {@link Subscribe}.
* The {@link Subscribe} annotation also allows configuration like {@link
* ThreadMode} and priority.
*/
public void register(Object subscriber) {
Class<?> subscriberClass = subscriber.getClass();
List<SubscriberMethod> subscriberMethods = subscriberMethodFinder.findSubscriberMethods(subscriberClass);
synchronized (this) {
for (SubscriberMethod subscriberMethod : subscriberMethods) {
subscribe(subscriber, subscriberMethod);
}
}
}

我们在Activity/Fragment中都有可能会调用这个方法,如果是Fragment里面我们还会在onDestoryView()中进行unregister(...)。在这段函数里我们发现使用反射从这个Class中找到了所有的订阅者函数了,然后对每个订阅者函数进行注册。

这里我们看看我们的SubribeMethod被包装成了什么样子:

/** Used internally by EventBus and generated subscriber indexes. */
public class SubscriberMethod {
final Method method;
final ThreadMode threadMode;
final Class<?> eventType;
final int priority;
final boolean sticky;
/** Used for efficient comparison */
String methodString; public SubscriberMethod(Method method, Class<?> eventType, ThreadMode threadMode, int priority, boolean sticky) {
this.method = method;
this.threadMode = threadMode;
this.eventType = eventType;
this.priority = priority;
this.sticky = sticky;
} @Override
public boolean equals(Object other) {
if (other == this) {
return true;
} else if (other instanceof SubscriberMethod) {
checkMethodString();
SubscriberMethod otherSubscriberMethod = (SubscriberMethod)other;
otherSubscriberMethod.checkMethodString();
// Don't use method.equals because of http://code.google.com/p/android/issues/detail?id=7811#c6
return methodString.equals(otherSubscriberMethod.methodString);
} else {
return false;
}
} private synchronized void checkMethodString() {
if (methodString == null) {
// Method.toString has more overhead, just take relevant parts of the method
StringBuilder builder = new StringBuilder(64);
builder.append(method.getDeclaringClass().getName());
builder.append('#').append(method.getName());
builder.append('(').append(eventType.getName());
methodString = builder.toString();
}
} @Override
public int hashCode() {
return method.hashCode();
}
}

SubscribeMethod 携带了Method函数原型,还有就是我们在注解类里面提供的所有信息。还有一个Class<?>类型的EventType是指我们的事件类所对应的Class,其余的方法都是为了比较和判断是否相等来做的,equal/checkMethodString都是各种的拼字串来进行存储和判断。

下面我们来看register里面调用的这段subscribe,这段非常的重要涉及了EventBus运行时处理的绝大多数部分,还有就是粘性事件的分发。这段使用了大量的JDK的反射包的API,本身注释也提醒我们了这段代码需要加锁,毕竟里面这一堆并发容器。所以我们最好先明确这段里面用的并发容器到底都是什么,这段代码才好继续看的下去。

    private final Map<Class<?>, CopyOnWriteArrayList<Subscription>> subscriptionsByEventType;
private final Map<Object, List<Class<?>>> typesBySubscriber;
private final Map<Class<?>, Object> stickyEvents;

主要的有这几个:

  1. 第一个Map存储的Key是Class类型,Value是一个并发的ArrayList里面存的是对订阅者和订阅函数的一种绑定类Subscription从名字上也能看出Key是Event的Class对象。
  2. 第二个存储的是订阅者(Activity什么的啊)和Event类型的List。
  3. 第三个Map存储的是粘性事件,Key是Event类型,Value是真实存在的StickyEvent对象。

知道这三个都是什么之后,这段代码就好看了。我们来看前一部分。

// Must be called in synchronized block
private void subscribe(Object subscriber, SubscriberMethod subscriberMethod) {
Class<?> eventType = subscriberMethod.eventType;
Subscription newSubscription = new Subscription(subscriber, subscriberMethod);
// Map<Class<?>, CopyOnWriteArrayList<Subscription>>
CopyOnWriteArrayList<Subscription> subscriptions = subscriptionsByEventType.get(eventType) ;
if (subscriptions == null) {
subscriptions = new CopyOnWriteArrayList<>();
subscriptionsByEventType.put(eventType, subscriptions);
} else {
if (subscriptions.contains(newSubscription)) {
throw new EventBusException("Subscriber " + subscriber.getClass() + " already registered to event "
+ eventType);
}
} int size = subscriptions.size();
for (int i = 0; i <= size; i++) {
if (i == size || subscriberMethod.priority > subscriptions.get(i).subscriberMethod.priority) {
subscriptions.add(i, newSubscription);
break;
}
} List<Class<?>> subscribedEvents = typesBySubscriber.get(subscriber);
if (subscribedEvents == null) {
subscribedEvents = new ArrayList<>();
typesBySubscriber.put(subscriber, subscribedEvents);
}
subscribedEvents.add(eventType);

这段写的虽然有点乱套,但实际上写的挺简单的,而且一堆堆的O(n)遍历,性能也就那样(?)。

首先这里面出现了Subscription:

final class Subscription {
final Object subscriber;
final SubscriberMethod subscriberMethod;
/**
* Becomes false as soon as {@link EventBus#unregister(Object)} is called, which is checked by queued event delivery
* {@link EventBus#invokeSubscriber(PendingPost)} to prevent race conditions.
*/
volatile boolean active; Subscription(Object subscriber, SubscriberMethod subscriberMethod) {
this.subscriber = subscriber;
this.subscriberMethod = subscriberMethod;
active = true;
} @Override
public boolean equals(Object other) {
if (other instanceof Subscription) {
Subscription otherSubscription = (Subscription) other;
return subscriber == otherSubscription.subscriber
&& subscriberMethod.equals(otherSubscription.subscriberMethod);
} else {
return false;
}
} @Override
public int hashCode() {
return subscriber.hashCode() + subscriberMethod.methodString.hashCode();
}
}

我们发现了这是订阅者和订阅方法类的一个契约关系类。

所以说上面subscribe函数主要做了,

  • 创建了订阅者和方法类的绑定,然后存进了subscriptionsByEventType
  • 对每个类型重新排列了一次优先级
  • typesBySubscriber添加了对应的类型

然后我们可以看一下这个函数的下一半,我们会惊奇地发现,StickyEvent的发送时机居然是在register的时候:

...
if (subscriberMethod.sticky) {
if (eventInheritance) {
// Existing sticky events of all subclasses of eventType have to be considered.
// Note: Iterating over all events may be inefficient with lots of sticky events,
// thus data structure should be changed to allow a more efficient lookup
// (e.g. an additional map storing sub classes of super classes: Class -> List<Class>).
Set<Map.Entry<Class<?>, Object>> entries = stickyEvents.entrySet();
for (Map.Entry<Class<?>, Object> entry : entries) {
Class<?> candidateEventType = entry.getKey();
if (eventType.isAssignableFrom(candidateEventType)) {
Object stickyEvent = entry.getValue();
checkPostStickyEventToSubscription(newSubscription, stickyEvent);
}
}
} else {
Object stickyEvent = stickyEvents.get(eventType);
checkPostStickyEventToSubscription(newSubscription, stickyEvent);
}
}

这时候轮了一遍所有的粘性事件。isAssignableFrom类似于使用在Class之间的instance of 就是判断两个类是否有相同的接口关系,也就是说有继承和实现关系的事件类,都会被判断处理。

  private void checkPostStickyEventToSubscription(Subscription newSubscription, Object stickyEvent) {
if (stickyEvent != null) {
// If the subscriber is trying to abort the event, it will fail (event is not tracked in posting state)
// --> Strange corner case, which we don't take care of here.
postToSubscription(newSubscription, stickyEvent, Looper.getMainLooper() == Looper.myLooper());
}
}
 private void postToSubscription(Subscription subscription, Object event, boolean isMainThread) {
switch (subscription.subscriberMethod.threadMode) {
case POSTING:
invokeSubscriber(subscription, event);
break;
case MAIN:
if (isMainThread) {
invokeSubscriber(subscription, event);
} else {
mainThreadPoster.enqueue(subscription, event);
}
break;
case BACKGROUND:
if (isMainThread) {
backgroundPoster.enqueue(subscription, event);
} else {
invokeSubscriber(subscription, event);
}
break;
case ASYNC:
asyncPoster.enqueue(subscription, event);
break;
default:
throw new IllegalStateException("Unknown thread mode: " + subscription.subscriberMethod.threadMode);
}
}

之后就是针对各种的ThreadMode进行了处理,同一线程的直接依赖Java的反射invoke执行了,各种不可以的情况,比如说发到主线程但还没在主线程的时候,都是用队列进行发送到对应线程。

接下来我们看看这里面在各线程之间的发送是怎么实现的。

消息转换线程

我们发现在Subscription和event入队的时候我们把他们封装成了一个PendingPost类:

// HandlePoster
void enqueue(Subscription subscription, Object event) {
PendingPost pendingPost = PendingPost.obtainPendingPost(subscription, event);
synchronized (this) {
queue.enqueue(pendingPost);
if (!handlerActive) {
handlerActive = true;
if (!sendMessage(obtainMessage())) {
throw new EventBusException("Could not send handler message");
}
}
}
}

然后才进行的入队和发送,这个PendingPost就是一个带有回收池的掩饰传送类:

final class PendingPost {
private final static List<PendingPost> pendingPostPool = new ArrayList<PendingPost>(); Object event;
Subscription subscription;
PendingPost next; private PendingPost(Object event, Subscription subscription) {
this.event = event;
this.subscription = subscription;
} static PendingPost obtainPendingPost(Subscription subscription, Object event) {
synchronized (pendingPostPool) {
int size = pendingPostPool.size();
if (size > 0) {
PendingPost pendingPost = pendingPostPool.remove(size - 1);
pendingPost.event = event;
pendingPost.subscription = subscription;
pendingPost.next = null;
return pendingPost;
}
}
return new PendingPost(event, subscription);
} static void releasePendingPost(PendingPost pendingPost) {
pendingPost.event = null;
pendingPost.subscription = null;
pendingPost.next = null;
synchronized (pendingPostPool) {
// Don't let the pool grow indefinitely
if (pendingPostPool.size() < 10000) {
pendingPostPool.add(pendingPost);
}
}
}
}

这里的设计其实挺不错的,一个静态的回收池,初始化靠一个静态方法,优先使用被回收的对象,实现和Message其实很像。另一个release方法就是把用完的对象回收起来。

PendingPostQueue 就是一个PendingPost的队列,里面的操作基本上就是入队出队之类的,有点特殊的是入队和出队都有一把锁。

接着这个队列被用在了好几个Poster类中,实现了向各个线程的消息转换,首先我们来看向主线程发送数据的:

HandlePoster

final class HandlerPoster extends Handler {

    private final PendingPostQueue queue;
private final int maxMillisInsideHandleMessage;
private final EventBus eventBus;
private boolean handlerActive; HandlerPoster(EventBus eventBus, Looper looper, int maxMillisInsideHandleMessage) {
super(looper);
this.eventBus = eventBus;
this.maxMillisInsideHandleMessage = maxMillisInsideHandleMessage;
queue = new PendingPostQueue();
} void enqueue(Subscription subscription, Object event) {
PendingPost pendingPost = PendingPost.obtainPendingPost(subscription, event);
synchronized (this) {
queue.enqueue(pendingPost);
if (!handlerActive) {
handlerActive = true;
if (!sendMessage(obtainMessage())) {
throw new EventBusException("Could not send handler message");
}
}
}
} @Override
public void handleMessage(Message msg) {
boolean rescheduled = false;
try {
long started = SystemClock.uptimeMillis();
while (true) {
PendingPost pendingPost = queue.poll();
if (pendingPost == null) {
synchronized (this) {
// Check again, this time in synchronized
pendingPost = queue.poll();
if (pendingPost == null) {
handlerActive = false;
return;
}
}
}
eventBus.invokeSubscriber(pendingPost);
long timeInMethod = SystemClock.uptimeMillis() - started;
if (timeInMethod >= maxMillisInsideHandleMessage) {
if (!sendMessage(obtainMessage())) {
throw new EventBusException("Could not send handler message");
}
rescheduled = true;
return;
}
}
} finally {
handlerActive = rescheduled;
}
}
}

HandlePoster 继承自 Handler 再加上初始化的时候传进去的是Looper.getMainThread();所以能向主线程发送消息。每次入队之后都会发送一条空消息去通知handleMessage函数处理队列数据,使用handlerActive作为控制标记位。handleMessage是个死循环两段的if判断用来处理多线程的情况,invokeSubscriber的方式和之前类似。之后就是有一个阀值,当时间超过10ms的时候就会发一个消息重入,并且退出这次循环,这是防止时间太长阻塞主线程。

BackgroundPoster

final class BackgroundPoster implements Runnable {

    private final PendingPostQueue queue;
private final EventBus eventBus; private volatile boolean executorRunning; BackgroundPoster(EventBus eventBus) {
this.eventBus = eventBus;
queue = new PendingPostQueue();
} public void enqueue(Subscription subscription, Object event) {
PendingPost pendingPost = PendingPost.obtainPendingPost(subscription, event);
synchronized (this) {
queue.enqueue(pendingPost);
if (!executorRunning) {
executorRunning = true;
eventBus.getExecutorService().execute(this);
}
}
} @Override
public void run() {
try {
try {
while (true) {
PendingPost pendingPost = queue.poll(1000);
if (pendingPost == null) {
synchronized (this) {
// Check again, this time in synchronized
pendingPost = queue.poll();
if (pendingPost == null) {
executorRunning = false;
return;
}
}
}
eventBus.invokeSubscriber(pendingPost);
}
} catch (InterruptedException e) {
Log.w("Event", Thread.currentThread().getName() + " was interruppted", e);
}
} finally {
executorRunning = false;
}
}
}

BackgroundPoster 自身是一个Runnable ,入队之后就调用EventBus携带的一个线程池进行运行,同样也是一个死循环,用了一个生产者 vs 消费者模式 进行了有限等待,这1000ms内入队的消息都会被弹出处理。

    synchronized PendingPost poll(int maxMillisToWait) throws InterruptedException {
if (head == null) {
wait(maxMillisToWait);
}
return poll();
}

PendingPostQueue的poll(int)方法对队列为空的情况进行了等待,唤醒则出现在enqueue:

    synchronized void enqueue(PendingPost pendingPost) {
if (pendingPost == null) {
throw new NullPointerException("null cannot be enqueued");
}
if (tail != null) {
tail.next = pendingPost;
tail = pendingPost;
} else if (head == null) {
head = tail = pendingPost;
} else {
throw new IllegalStateException("Head present, but no tail");
}
notifyAll(); // 在这进行了唤醒
}

AsyncPoster

如果说Background尚且能保证在同一个线程内完成,AsyncPoster就完全进行了异步操作。

class AsyncPoster implements Runnable {
private final PendingPostQueue queue;
private final EventBus eventBus;
AsyncPoster(EventBus eventBus) {
this.eventBus = eventBus;
queue = new PendingPostQueue();
}
public void enqueue(Subscription subscription, Object event) {
PendingPost pendingPost = PendingPost.obtainPendingPost(subscription, event);
queue.enqueue(pendingPost);
eventBus.getExecutorService().execute(this);
}
@Override
public void run() {
PendingPost pendingPost = queue.poll();
if(pendingPost == null) {
throw new IllegalStateException("No pending post available");
}
eventBus.invokeSubscriber(pendingPost);
}
}

这里面基本上什么都不控制,直接就来一个运行一次,也不会有什么问题。。。

到这为止我们不但知道了方法是怎么注册和绑定的,我们甚至还知道了粘性事件是怎么发送的了,接着我们来看方法查找和普通事件的发送是怎么进行的。

方法查找

// package org.greenrobot.eventbus.meta;
/** Base class for generated index classes created by annotation processing. */
public interface SubscriberInfo {
// 获取订阅的类
Class<?> getSubscriberClass();
// 所有的method
SubscriberMethod[] getSubscriberMethods();
// 获取父类的info
SubscriberInfo getSuperSubscriberInfo();
// 是否检查父类
boolean shouldCheckSuperclass();
}

SubscriberInfo 描述了能通过注解类生成的Index的方法(具体功能我加了主食)。

/**
* Interface for generated indexes.
*/
public interface SubscriberInfoIndex {
SubscriberInfo getSubscriberInfo(Class<?> subscriberClass);
}

这个接口是查找info的。

另外可以说这其中的SubscriberMethodInfo存储着SubscriberMethod所需的元信息:

public class SubscriberMethodInfo {
final String methodName;
final ThreadMode threadMode;
final Class<?> eventType;
final int priority;
final boolean sticky;
...

AbstractSubscriberInfo是一个抽象类,主要负责从Info创建出Method,又是一个反射:

    protected SubscriberMethod createSubscriberMethod(String methodName, Class<?> eventType, ThreadMode threadMode,
int priority, boolean sticky) {
try {
Method method = subscriberClass.getDeclaredMethod(methodName, eventType);
return new SubscriberMethod(method, eventType, threadMode, priority, sticky);
} catch (NoSuchMethodException e) {
throw new EventBusException("Could not find subscriber method in " + subscriberClass +
". Maybe a missing ProGuard rule?", e);
}
}

另外还有一个SimpleSubscriberInfo作为他的子类。

接下来的SubscriberMethodFinder也非常重要运行时的方法查找都来自这里:

刚才我们在EventBus.register(...)中调用了这个函数:

    List<SubscriberMethod> findSubscriberMethods(Class<?> subscriberClass) {
List<SubscriberMethod> subscriberMethods = METHOD_CACHE.get(subscriberClass);
if (subscriberMethods != null) {
return subscriberMethods;
} if (ignoreGeneratedIndex) {
subscriberMethods = findUsingReflection(subscriberClass);
} else {
subscriberMethods = findUsingInfo(subscriberClass);
}
if (subscriberMethods.isEmpty()) {
throw new EventBusException("Subscriber " + subscriberClass
+ " and its super classes have no public methods with the @Subscribe annotation");
} else {
METHOD_CACHE.put(subscriberClass, subscriberMethods);
return subscriberMethods;
}
}

其中的METHOD_CACHE是对每个类方法进行缓存,防止多次查找,毕竟运行时查找还是个复杂的操作,根据是否忽略生成Index。

private List<SubscriberMethod> findUsingReflection(Class<?> subscriberClass) {
FindState findState = prepareFindState();
findState.initForSubscriber(subscriberClass);
while (findState.clazz != null) {
findUsingReflectionInSingleClass(findState);
findState.moveToSuperclass();
}
return getMethodsAndRelease(findState);
} private void findUsingReflectionInSingleClass(FindState findState) {
Method[] methods;
try {
// This is faster than getMethods, especially when subscribers are fat classes like Activities
methods = findState.clazz.getDeclaredMethods();
} catch (Throwable th) {
// Workaround for java.lang.NoClassDefFoundError, see https://github.com/greenrobot/EventBus/issues/149
methods = findState.clazz.getMethods();
findState.skipSuperClasses = true;
}
for (Method method : methods) {
int modifiers = method.getModifiers();
if ((modifiers & Modifier.PUBLIC) != 0 && (modifiers & MODIFIERS_IGNORE) == 0) {
Class<?>[] parameterTypes = method.getParameterTypes();
if (parameterTypes.length == 1) {
Subscribe subscribeAnnotation = method.getAnnotation(Subscribe.class);
if (subscribeAnnotation != null) {
Class<?> eventType = parameterTypes[0];
if (findState.checkAdd(method, eventType)) {
ThreadMode threadMode = subscribeAnnotation.threadMode();
findState.subscriberMethods.add(new SubscriberMethod(method, eventType, threadMode,
subscribeAnnotation.priority(), subscribeAnnotation.sticky()));
}
}
} else if (strictMethodVerification && method.isAnnotationPresent(Subscribe.class)) {
String methodName = method.getDeclaringClass().getName() + "." + method.getName();
throw new EventBusException("@Subscribe method " + methodName +
"must have exactly 1 parameter but has " + parameterTypes.length);
}
} else if (strictMethodVerification && method.isAnnotationPresent(Subscribe.class)) {
String methodName = method.getDeclaringClass().getName() + "." + method.getName();
throw new EventBusException(methodName +
" is a illegal @Subscribe method: must be public, non-static, and non-abstract");
}
}
}

findUsingReflectionInSingleClass对反射类进行了处理,这里面通过掩模运算检查了访问权限, 检查了参数个数。

        boolean checkAdd(Method method, Class<?> eventType) {
// 2 level check: 1st level with event type only (fast), 2nd level with complete signature when required.
// Usually a subscriber doesn't have methods listening to the same event type.
Object existing = anyMethodByEventType.put(eventType, method);
if (existing == null) {
return true;
} else {
if (existing instanceof Method) {
if (!checkAddWithMethodSignature((Method) existing, eventType)) {
// Paranoia check
throw new IllegalStateException();
}
// Put any non-Method object to "consume" the existing Method
anyMethodByEventType.put(eventType, this);
}
return checkAddWithMethodSignature(method, eventType);
}
}

其中的checkAdd检查了类型和方法签名,每次轮转完成之后都会进行一次findState.moveToSuperclass();对父类进行处理。

使用索引

因为反射所使用的运行时查找速度缓慢,所以我们也经常会通过apt使用已经创建好的Index。

刚才另一个分支的findUsingInfo就是使用已有的Index:

private List<SubscriberMethod> findUsingInfo(Class<?> subscriberClass) {
FindState findState = prepareFindState();
findState.initForSubscriber(subscriberClass);
while (findState.clazz != null) {
findState.subscriberInfo = getSubscriberInfo(findState);
if (findState.subscriberInfo != null) {
SubscriberMethod[] array = findState.subscriberInfo.getSubscriberMethods();
for (SubscriberMethod subscriberMethod : array) {
if (findState.checkAdd(subscriberMethod.method, subscriberMethod.eventType)) {
findState.subscriberMethods.add(subscriberMethod);
}
}
} else {
findUsingReflectionInSingleClass(findState);
}
findState.moveToSuperclass();
}
return getMethodsAndRelease(findState);
}

这段非常简单,几乎就是刚才的验证而已,如果没拿到数据的话,还会进行正常的反射查找。

 // EventBusAnnotationProcessor 负责生成注解路由表
private void createInfoIndexFile(String index) {
BufferedWriter writer = null;
try {
JavaFileObject sourceFile = processingEnv.getFiler().createSourceFile(index);
int period = index.lastIndexOf('.');
String myPackage = period > 0 ? index.substring(0, period) : null;
String clazz = index.substring(period + 1);
writer = new BufferedWriter(sourceFile.openWriter());
if (myPackage != null) {
writer.write("package " + myPackage + ";\n\n");
}
writer.write("import org.greenrobot.eventbus.meta.SimpleSubscriberInfo;\n");
writer.write("import org.greenrobot.eventbus.meta.SubscriberMethodInfo;\n");
writer.write("import org.greenrobot.eventbus.meta.SubscriberInfo;\n");
writer.write("import org.greenrobot.eventbus.meta.SubscriberInfoIndex;\n\n");
writer.write("import org.greenrobot.eventbus.ThreadMode;\n\n");
writer.write("import java.util.HashMap;\n");
writer.write("import java.util.Map;\n\n");
writer.write("/** This class is generated by EventBus, do not edit. */\n");
writer.write("public class " + clazz + " implements SubscriberInfoIndex {\n");
writer.write(" private static final Map<Class<?>, SubscriberInfo> SUBSCRIBER_INDEX;\n\n");
writer.write(" static {\n");
writer.write(" SUBSCRIBER_INDEX = new HashMap<Class<?>, SubscriberInfo>();\n\n");
writeIndexLines(writer, myPackage);
writer.write(" }\n\n");
writer.write(" private static void putIndex(SubscriberInfo info) {\n");
writer.write(" SUBSCRIBER_INDEX.put(info.getSubscriberClass(), info);\n");
writer.write(" }\n\n");
writer.write(" @Override\n");
writer.write(" public SubscriberInfo getSubscriberInfo(Class<?> subscriberClass) {\n");
writer.write(" SubscriberInfo info = SUBSCRIBER_INDEX.get(subscriberClass);\n");
writer.write(" if (info != null) {\n");
writer.write(" return info;\n");
writer.write(" } else {\n");
writer.write(" return null;\n");
writer.write(" }\n");
writer.write(" }\n");
writer.write("}\n");
} catch (IOException e) {
throw new RuntimeException("Could not write source for " + index, e);
} finally {
if (writer != null) {
try {
writer.close();
} catch (IOException e) {
//Silent
}
}
}
} private void writeIndexLines(BufferedWriter writer, String myPackage) throws IOException {
for (TypeElement subscriberTypeElement : methodsByClass.keySet()) {
if (classesToSkip.contains(subscriberTypeElement)) {
continue;
} String subscriberClass = getClassString(subscriberTypeElement, myPackage);
if (isVisible(myPackage, subscriberTypeElement)) {
writeLine(writer, 2,
"putIndex(new SimpleSubscriberInfo(" + subscriberClass + ".class,",
"true,", "new SubscriberMethodInfo[] {");
List<ExecutableElement> methods = methodsByClass.get(subscriberTypeElement);
writeCreateSubscriberMethods(writer, methods, "new SubscriberMethodInfo", myPackage);
writer.write(" }));\n\n");
} else {
writer.write(" // Subscriber not visible to index: " + subscriberClass + "\n");
}
}
}

有了这两个方法之后我们就知道,平常的index就是通过这种方式拼接出来的。

Post消息

    /** Posts the given event to the event bus. */
public void post(Object event) {
PostingThreadState postingState = currentPostingThreadState.get();
List<Object> eventQueue = postingState.eventQueue;
eventQueue.add(event); if (!postingState.isPosting) {
postingState.isMainThread = Looper.getMainLooper() == Looper.myLooper();
postingState.isPosting = true;
if (postingState.canceled) {
throw new EventBusException("Internal error. Abort state was not reset");
}
try {
while (!eventQueue.isEmpty()) {
postSingleEvent(eventQueue.remove(0), postingState);
}
} finally {
postingState.isPosting = false;
postingState.isMainThread = false;
}
}
}

PostingThreadState是一个存储在ThreadLocal中的对象,包含有以下各种内容,线程信息,是否是主线程,是否取消,还有一个相应的事件队列。

  private void postSingleEvent(Object event, PostingThreadState postingState) throws Error {
Class<?> eventClass = event.getClass();
boolean subscriptionFound = false;
if (eventInheritance) {
/** Looks up all Class objects including super classes and interfaces. Should also work for interfaces. */
List<Class<?>> eventTypes = lookupAllEventTypes(eventClass);
int countTypes = eventTypes.size();
// 对所有的订阅函数,都调用发送数据
for (int h = 0; h < countTypes; h++) {
// 所有的订阅类
Class<?> clazz = eventTypes.get(h);
subscriptionFound |= postSingleEventForEventType(event, postingState, clazz);
}
} else {
// 只发送一次
subscriptionFound = postSingleEventForEventType(event, postingState, eventClass);
}
if (!subscriptionFound) {
if (logNoSubscriberMessages) {
Log.d(TAG, "No subscribers registered for event " + eventClass);
}
if (sendNoSubscriberEvent && eventClass != NoSubscriberEvent.class &&
eventClass != SubscriberExceptionEvent.class) {
// 无订阅者的处理
post(new NoSubscriberEvent(this, event));
}
}
}

之后:

private boolean postSingleEventForEventType(Object event, PostingThreadState postingState, Class<?> eventClass) {
CopyOnWriteArrayList<Subscription> subscriptions;
synchronized (this) {
subscriptions = subscriptionsByEventType.get(eventClass);
}
if (subscriptions != null && !subscriptions.isEmpty()) {
for (Subscription subscription : subscriptions) {
postingState.event = event;
postingState.subscription = subscription;
boolean aborted = false;
try {
postToSubscription(subscription, event, postingState.isMainThread);
aborted = postingState.canceled;
} finally {
postingState.event = null;
postingState.subscription = null;
postingState.canceled = false;
}
if (aborted) {
break;
}
}
return true;
}
return false;
}

之后对所有的订阅类的所有订阅者都发送一次数据,发送数据方法和上文相同。

发送粘性数据就是拿锁然后保存到队列中去,这样就可以在重新发送:

    public void postSticky(Object event) {
synchronized (stickyEvents) {
stickyEvents.put(event.getClass(), event);
}
// Should be posted after it is putted, in case the subscriber wants to remove immediately
post(event);
}

因为我们无法确定什么时候粘性事件应该停止继续传播,这取决于我们应用的需要,所以我们应当手动remove掉Sticky Event :

// 系统提供了如下方法
public <T> T removeStickyEvent(Class<T> eventType) {
synchronized (stickyEvents) {
return eventType.cast(stickyEvents.remove(eventType));
}
}
public boolean removeStickyEvent(Object event) {
synchronized (stickyEvents) {
Class<?> eventType = event.getClass();
Object existingEvent = stickyEvents.get(eventType);
if (event.equals(existingEvent)) {
stickyEvents.remove(eventType);
return true;
} else {
return false;
}
}
}
public void removeAllStickyEvents() {
synchronized (stickyEvents) {
stickyEvents.clear();
}
}

至此我们就分析完了EventBus的基本上所有的代码(处理util包下的错误日志),EventBus本身的实现并不复杂,使用运行时的反射技巧也很简单,单纯的使用注解类可能会拖慢速度,但是通过apt生成的静态表把速降提升到了一个新的高度,apt的生成大家也看到了并不是很复杂,几乎就是类型检查和拼接字串,不过想法决定了EventBus仍然是一个优秀的开源库,希望我们在使用的同时,仍能对实现原理有所了解。