Android 消息机制原理研究 (Looper.addFd)

消息都是存放在一个消息队列中去，而消息循环线程就是围绕这个消息队列进入一个无限循环的，直到线程退出。如果队列中有消息，消息循环线程就会把它取出来，并分发给相应的Handler进行处理；如果队列中没有消息，消息循环线程就会进入空闲等待状态，等待下一个消息的到来。在编写Android应用程序时，当程序执行的任务比较繁重时，为了不阻塞UI主线程而导致ANR的发生，我们通常的做法的创建一个子线程来完成特定的任务。在创建子线程时，有两种选择，一种通过创建Thread对象来创建一个无消息循环的子线程；还有一种就是创建一个带有消息循环的子线程，而创建带有消息循环的子线程由于两种实现方法，一种是直接利用Android给我们封装好的HandlerThread类来直接生成一个带有消息循环的线程对象，另一种方法是在实现线程的run()方法内使用以下方式启动一个消息循环：

一、消息机制使用

通常消息都是有一个消息线程和一个Handler组成，下面我们看PowerManagerService中的一个消息Handler：

            mHandlerThread = new ServiceThread(TAG,
                    Process.THREAD_PRIORITY_DISPLAY, false /*allowIo*/);
            mHandlerThread.start();
            mHandler = new PowerManagerHandler(mHandlerThread.getLooper());

这里的ServiceThread就是一个HandlerThread，创建Handler的时候，必须把HandlerThread的looper传进去，否则就是默认当前线程的looper。

而每个handler，大致如下：

        private final class PowerManagerHandler extends Handler {
            public PowerManagerHandler(Looper looper) {
                super(looper, null, true /*async*/);
            }
     
            @Override
            public void handleMessage(Message msg) {
                switch (msg.what) {
                    case MSG_USER_ACTIVITY_TIMEOUT:
                        handleUserActivityTimeout();
                        break;
                    case MSG_SANDMAN:
                        handleSandman();
                        break;
                    case MSG_SCREEN_BRIGHTNESS_BOOST_TIMEOUT:
                        handleScreenBrightnessBoostTimeout();
                        break;
                    case MSG_CHECK_WAKE_LOCK_ACQUIRE_TIMEOUT:
                        checkWakeLockAquireTooLong();
                        Message m = mHandler.obtainMessage(MSG_CHECK_WAKE_LOCK_ACQUIRE_TIMEOUT);
                        m.setAsynchronous(true);
                        mHandler.sendMessageDelayed(m, WAKE_LOCK_ACQUIRE_TOO_LONG_TIMEOUT);
                        break;
                }
            }
        }
 

二、消息机制原理

那我们先来看下HandlerThread的主函数run函数：

        public void run() {
            mTid = Process.myTid();
            Looper.prepare();
            synchronized (this) {
                mLooper = Looper.myLooper();//赋值后notifyall，主要是getLooper函数返回的是mLooper
                notifyAll();
            }
            Process.setThreadPriority(mPriority);
            onLooperPrepared();
            Looper.loop();
            mTid = -1;
        }

再来看看Lopper的prepare函数，最后新建了一个Looper对象，并且放在线程的局部变量中。

        public static void prepare() {
            prepare(true);
        }
     
        private static void prepare(boolean quitAllowed) {
            if (sThreadLocal.get() != null) {
                throw new RuntimeException("Only one Looper may be created per thread");
            }
            sThreadLocal.set(new Looper(quitAllowed));
        }

Looper的构造函数中创建了MessageQueue

        private Looper(boolean quitAllowed) {
            mQueue = new MessageQueue(quitAllowed);
            mThread = Thread.currentThread();
        }

我们再来看下MessageQueue的构造函数，其中nativeInit是一个native方法，并且把返回值保存在mPtr显然是用long型变量保存的指针

        MessageQueue(boolean quitAllowed) {
            mQuitAllowed = quitAllowed;
            mPtr = nativeInit();
        }

native函数中主要创建了NativeMessageQueue对象，并且把指针变量返回了。

    static jlong android_os_MessageQueue_nativeInit(JNIEnv* env, jclass clazz) {
        NativeMessageQueue* nativeMessageQueue = new NativeMessageQueue();
        if (!nativeMessageQueue) {
            jniThrowRuntimeException(env, "Unable to allocate native queue");
            return 0;
        }
     
        nativeMessageQueue->incStrong(env);
        return reinterpret_cast<jlong>(nativeMessageQueue);
    }

NativeMessageQueue构造函数就是获取mLooper，如果没有就是新建一个Looper

    NativeMessageQueue::NativeMessageQueue() :
            mPollEnv(NULL), mPollObj(NULL), mExceptionObj(NULL) {
        mLooper = Looper::getForThread();
        if (mLooper == NULL) {
            mLooper = new Looper(false);
            Looper::setForThread(mLooper);
        }
    }

然后我们再看下Looper的构造函数，显示调用了eventfd创建了一个fd，eventfd它的主要是用于进程或者线程间的通信，我们可以看下这篇博客eventfd介绍

    Looper::Looper(bool allowNonCallbacks) :
            mAllowNonCallbacks(allowNonCallbacks), mSendingMessage(false),
            mPolling(false), mEpollFd(-1), mEpollRebuildRequired(false),
            mNextRequestSeq(0), mResponseIndex(0), mNextMessageUptime(LLONG_MAX) {
        mWakeEventFd = eventfd(0, EFD_NONBLOCK);
        LOG_ALWAYS_FATAL_IF(mWakeEventFd < 0, "Could not make wake event fd.  errno=%d", errno);
     
        AutoMutex _l(mLock);
        rebuildEpollLocked();
    }

2.1 c层创建epoll

我们再来看下rebuildEpollLocked函数，创建了epoll，并且把mWakeEventFd加入epoll，而且把mRequests的fd也加入epoll

    void Looper::rebuildEpollLocked() {
        // Close old epoll instance if we have one.
        if (mEpollFd >= 0) {
    #if DEBUG_CALLBACKS
            ALOGD("%p ~ rebuildEpollLocked - rebuilding epoll set", this);
    #endif
            close(mEpollFd);
        }
     
        // Allocate the new epoll instance and register the wake pipe.
        mEpollFd = epoll_create(EPOLL_SIZE_HINT);
        LOG_ALWAYS_FATAL_IF(mEpollFd < 0, "Could not create epoll instance.  errno=%d", errno);
     
        struct epoll_event eventItem;
        memset(& eventItem, 0, sizeof(epoll_event)); // zero out unused members of data field union
        eventItem.events = EPOLLIN;
        eventItem.data.fd = mWakeEventFd;
        int result = epoll_ctl(mEpollFd, EPOLL_CTL_ADD, mWakeEventFd, & eventItem);
        LOG_ALWAYS_FATAL_IF(result != 0, "Could not add wake event fd to epoll instance.  errno=%d",
                errno);
     
        for (size_t i = 0; i < mRequests.size(); i++) {
            const Request& request = mRequests.valueAt(i);
            struct epoll_event eventItem;
            request.initEventItem(&eventItem);
     
            int epollResult = epoll_ctl(mEpollFd, EPOLL_CTL_ADD, request.fd, & eventItem);
            if (epollResult < 0) {
                ALOGE("Error adding epoll events for fd %d while rebuilding epoll set, errno=%d",
                        request.fd, errno);
            }
        }
    }

继续回到HandlerThread的run函数，我们继续分析Looper的loop函数

        public void run() {
            mTid = Process.myTid();
            Looper.prepare();
            synchronized (this) {
                mLooper = Looper.myLooper();
                notifyAll();
            }
            Process.setThreadPriority(mPriority);
            onLooperPrepared();
            Looper.loop();
            mTid = -1;
        }

我们看看Looper的loop函数：

        public static void loop() {
            final Looper me = myLooper();
            if (me == null) {
                throw new RuntimeException("No Looper; Looper.prepare() wasn't called on this thread.");
            }
            final MessageQueue queue = me.mQueue;//得到Looper的mQueue
     
            // Make sure the identity of this thread is that of the local process,
            // and keep track of what that identity token actually is.
            Binder.clearCallingIdentity();
            final long ident = Binder.clearCallingIdentity();
     
            for (;;) {
                Message msg = queue.next(); // might block这个函数会阻塞，阻塞主要是epoll_wait
                if (msg == null) {
                    // No message indicates that the message queue is quitting.
                    return;
                }
     
                // This must be in a local variable, in case a UI event sets the logger
                Printer logging = me.mLogging;//自己打的打印
                if (logging != null) {
                    logging.println(">>>>> Dispatching to " + msg.target + " " +
                            msg.callback + ": " + msg.what);
                }
     
                msg.target.dispatchMessage(msg);
     
                if (logging != null) {
                    logging.println("<<<<< Finished to " + msg.target + " " + msg.callback);
                }
     
                // Make sure that during the course of dispatching the
                // identity of the thread wasn't corrupted.
                final long newIdent = Binder.clearCallingIdentity();
                if (ident != newIdent) {
                    Log.wtf(TAG, "Thread identity changed from 0x"
                            + Long.toHexString(ident) + " to 0x"
                            + Long.toHexString(newIdent) + " while dispatching to "
                            + msg.target.getClass().getName() + " "
                            + msg.callback + " what=" + msg.what);
                }
     
                msg.recycleUnchecked();
            }
        }

MessageQueue类的next函数主要是调用了nativePollOnce函数，后面就是从消息队列中取出一个Message

        Message next() {
            // Return here if the message loop has already quit and been disposed.
            // This can happen if the application tries to restart a looper after quit
            // which is not supported.
            final long ptr = mPtr;//之前保留的指针
            if (ptr == 0) {
                return null;
            }
     
            int pendingIdleHandlerCount = -1; // -1 only during first iteration
            int nextPollTimeoutMillis = 0;
            for (;;) {
                if (nextPollTimeoutMillis != 0) {
                    Binder.flushPendingCommands();
                }
     
                nativePollOnce(ptr, nextPollTimeoutMillis);

下面我们主要看下nativePollOnce这个native函数，把之前的指针强制转换成NativeMessageQueue，然后调用其pollOnce函数

    static void android_os_MessageQueue_nativePollOnce(JNIEnv* env, jobject obj,
            jlong ptr, jint timeoutMillis) {
        NativeMessageQueue* nativeMessageQueue = reinterpret_cast<NativeMessageQueue*>(ptr);
        nativeMessageQueue->pollOnce(env, obj, timeoutMillis);
    }

2.2 c层epoll_wait阻塞

pollOnce函数，这个函数前面的while一般都没有只是处理了indent大于0的情况，这种情况一般没有，所以我们可以直接看pollInner函数

    int Looper::pollOnce(int timeoutMillis, int* outFd, int* outEvents, void** outData) {
        int result = 0;
        for (;;) {
            while (mResponseIndex < mResponses.size()) {
                const Response& response = mResponses.itemAt(mResponseIndex++);
                int ident = response.request.ident;
                if (ident >= 0) {
                    int fd = response.request.fd;
                    int events = response.events;
                    void* data = response.request.data;
    #if DEBUG_POLL_AND_WAKE
                    ALOGD("%p ~ pollOnce - returning signalled identifier %d: "
                            "fd=%d, events=0x%x, data=%p",
                            this, ident, fd, events, data);
    #endif
                    if (outFd != NULL) *outFd = fd;
                    if (outEvents != NULL) *outEvents = events;
                    if (outData != NULL) *outData = data;
                    return ident;
                }
            }
     
            if (result != 0) {
    #if DEBUG_POLL_AND_WAKE
                ALOGD("%p ~ pollOnce - returning result %d", this, result);
    #endif
                if (outFd != NULL) *outFd = 0;
                if (outEvents != NULL) *outEvents = 0;
                if (outData != NULL) *outData = NULL;
                return result;
            }
     
            result = pollInner(timeoutMillis);
        }
    }

pollInner函数主要就是调用epoll_wait阻塞，并且java层会计算每次阻塞的时间传到c层，等待有mWakeEventFd或者之前addFd的fd有事件过来，才会epoll_wait返回。

    int Looper::pollInner(int timeoutMillis) {
    #if DEBUG_POLL_AND_WAKE
        ALOGD("%p ~ pollOnce - waiting: timeoutMillis=%d", this, timeoutMillis);
    #endif
     
        // Adjust the timeout based on when the next message is due.
        if (timeoutMillis != 0 && mNextMessageUptime != LLONG_MAX) {
            nsecs_t now = systemTime(SYSTEM_TIME_MONOTONIC);
            int messageTimeoutMillis = toMillisecondTimeoutDelay(now, mNextMessageUptime);
            if (messageTimeoutMillis >= 0
                    && (timeoutMillis < 0 || messageTimeoutMillis < timeoutMillis)) {
                timeoutMillis = messageTimeoutMillis;
            }
    #if DEBUG_POLL_AND_WAKE
            ALOGD("%p ~ pollOnce - next message in %" PRId64 "ns, adjusted timeout: timeoutMillis=%d",
                    this, mNextMessageUptime - now, timeoutMillis);
    #endif
        }
     
        // Poll.
        int result = POLL_WAKE;
        mResponses.clear();//清空mResponses
        mResponseIndex = 0;
     
        // We are about to idle.
        mPolling = true;
     
        struct epoll_event eventItems[EPOLL_MAX_EVENTS];
        int eventCount = epoll_wait(mEpollFd, eventItems, EPOLL_MAX_EVENTS, timeoutMillis);//epoll_wait主要线程阻塞在这，这个阻塞的时间也是有java层传过来的
     
        // No longer idling.
        mPolling = false;
     
        // Acquire lock.
        mLock.lock();
     
        // Rebuild epoll set if needed.
        if (mEpollRebuildRequired) {
            mEpollRebuildRequired = false;
            rebuildEpollLocked();
            goto Done;
        }
     
        // Check for poll error.
        if (eventCount < 0) {
            if (errno == EINTR) {
                goto Done;
            }
            ALOGW("Poll failed with an unexpected error, errno=%d", errno);
            result = POLL_ERROR;
            goto Done;
        }
     
        // Check for poll timeout.
        if (eventCount == 0) {
    #if DEBUG_POLL_AND_WAKE
            ALOGD("%p ~ pollOnce - timeout", this);
    #endif
            result = POLL_TIMEOUT;
            goto Done;
        }
     
        // Handle all events.
    #if DEBUG_POLL_AND_WAKE
        ALOGD("%p ~ pollOnce - handling events from %d fds", this, eventCount);
    #endif
     
        for (int i = 0; i < eventCount; i++) {
            int fd = eventItems[i].data.fd;
            uint32_t epollEvents = eventItems[i].events;
            if (fd == mWakeEventFd) {//通知唤醒线程的事件
                if (epollEvents & EPOLLIN) {
                    awoken();
                } else {
                    ALOGW("Ignoring unexpected epoll events 0x%x on wake event fd.", epollEvents);
                }
            } else {
                ssize_t requestIndex = mRequests.indexOfKey(fd);//之前addFd的事件
                if (requestIndex >= 0) {
                    int events = 0;
                    if (epollEvents & EPOLLIN) events |= EVENT_INPUT;
                    if (epollEvents & EPOLLOUT) events |= EVENT_OUTPUT;
                    if (epollEvents & EPOLLERR) events |= EVENT_ERROR;
                    if (epollEvents & EPOLLHUP) events |= EVENT_HANGUP;
                    pushResponse(events, mRequests.valueAt(requestIndex));//放在mResponses中
                } else {
                    ALOGW("Ignoring unexpected epoll events 0x%x on fd %d that is "
                            "no longer registered.", epollEvents, fd);
                }
            }
        }
    Done: ;
     
        // Invoke pending message callbacks.
        mNextMessageUptime = LLONG_MAX;
        while (mMessageEnvelopes.size() != 0) {// 这块主要是c层的消息，java层的消息是自己管理的
            nsecs_t now = systemTime(SYSTEM_TIME_MONOTONIC);
            const MessageEnvelope& messageEnvelope = mMessageEnvelopes.itemAt(0);
            if (messageEnvelope.uptime <= now) {
                // Remove the envelope from the list.
                // We keep a strong reference to the handler until the call to handleMessage
                // finishes.  Then we drop it so that the handler can be deleted *before*
                // we reacquire our lock.
                { // obtain handler
                    sp<MessageHandler> handler = messageEnvelope.handler;
                    Message message = messageEnvelope.message;
                    mMessageEnvelopes.removeAt(0);
                    mSendingMessage = true;
                    mLock.unlock();
     
    #if DEBUG_POLL_AND_WAKE || DEBUG_CALLBACKS
                    ALOGD("%p ~ pollOnce - sending message: handler=%p, what=%d",
                            this, handler.get(), message.what);
    #endif
                    handler->handleMessage(message);
                } // release handler
     
                mLock.lock();
                mSendingMessage = false;
                result = POLL_CALLBACK;
            } else {
                // The last message left at the head of the queue determines the next wakeup time.
                mNextMessageUptime = messageEnvelope.uptime;
                break;
            }
        }
     
        // Release lock.
        mLock.unlock();
     
        // Invoke all response callbacks.
        for (size_t i = 0; i < mResponses.size(); i++) {//这是之前addFd的事件的处理，主要是遍历mResponses，然后调用其回调
            Response& response = mResponses.editItemAt(i);
            if (response.request.ident == POLL_CALLBACK) {
                int fd = response.request.fd;
                int events = response.events;
                void* data = response.request.data;
    #if DEBUG_POLL_AND_WAKE || DEBUG_CALLBACKS
                ALOGD("%p ~ pollOnce - invoking fd event callback %p: fd=%d, events=0x%x, data=%p",
                        this, response.request.callback.get(), fd, events, data);
    #endif
                // Invoke the callback.  Note that the file descriptor may be closed by
                // the callback (and potentially even reused) before the function returns so
                // we need to be a little careful when removing the file descriptor afterwards.
                int callbackResult = response.request.callback->handleEvent(fd, events, data);
                if (callbackResult == 0) {
                    removeFd(fd, response.request.seq);
                }
     
                // Clear the callback reference in the response structure promptly because we
                // will not clear the response vector itself until the next poll.
                response.request.callback.clear();
                result = POLL_CALLBACK;
            }
        }
        return result;
    }

继续分析Looper的loop函数，可以增加自己的打印来调试代码，之前调用Message的target的dispatchMessage来分配消息

            for (;;) {
                Message msg = queue.next(); // might block
                if (msg == null) {
                    // No message indicates that the message queue is quitting.
                    return;
                }
     
                // This must be in a local variable, in case a UI event sets the logger
                Printer logging = me.mLogging;//自己的打印
                if (logging != null) {
                    logging.println(">>>>> Dispatching to " + msg.target + " " +
                            msg.callback + ": " + msg.what);
                }
     
                msg.target.dispatchMessage(msg);
     
                if (logging != null) {
                    logging.println("<<<<< Finished to " + msg.target + " " + msg.callback);
                }
     
                // Make sure that during the course of dispatching the
                // identity of the thread wasn't corrupted.
                final long newIdent = Binder.clearCallingIdentity();
                if (ident != newIdent) {
                    Log.wtf(TAG, "Thread identity changed from 0x"
                            + Long.toHexString(ident) + " to 0x"
                            + Long.toHexString(newIdent) + " while dispatching to "
                            + msg.target.getClass().getName() + " "
                            + msg.callback + " what=" + msg.what);
                }
     
                msg.recycleUnchecked();
            }
        }

2.3 增加调试打印

我们先来看自己添加打印，可以通过Lopper的setMessageLogging函数来打印

        public void setMessageLogging(@Nullable Printer printer) {
            mLogging = printer;
        }

Printer就是一个interface

    public interface Printer {
        /**
         * Write a line of text to the output.  There is no need to terminate
         * the given string with a newline.
         */
        void println(String x);
    }

我们再来看下具体的打印类：

    public class LogPrinter implements Printer {
        private final int mPriority;
        private final String mTag;
        private final int mBuffer;
        
        /**
         * Create a new Printer that sends to the log with the given priority
         * and tag.
         *
         * @param priority The desired log priority:
         * {@link android.util.Log#VERBOSE Log.VERBOSE},
         * {@link android.util.Log#DEBUG Log.DEBUG},
         * {@link android.util.Log#INFO Log.INFO},
         * {@link android.util.Log#WARN Log.WARN}, or
         * {@link android.util.Log#ERROR Log.ERROR}.
         * @param tag A string tag to associate with each printed log statement.
         */
        public LogPrinter(int priority, String tag) {
            mPriority = priority;
            mTag = tag;
            mBuffer = Log.LOG_ID_MAIN;
        }
     
        /**
         * @hide
         * Same as above, but buffer is one of the LOG_ID_ constants from android.util.Log.
         */
        public LogPrinter(int priority, String tag, int buffer) {
            mPriority = priority;
            mTag = tag;
            mBuffer = buffer;
        }
        
        public void println(String x) {
            Log.println_native(mBuffer, mPriority, mTag, x);
        }
    }
 

2.4 java层消息分发处理

再来看消息的分发，先是调用Handler的obtainMessage函数

                    Message msg = mHandler.obtainMessage(MSG_CHECK_WAKE_LOCK_ACQUIRE_TIMEOUT);
                    msg.setAsynchronous(true);
                    mHandler.sendMessageDelayed(msg, WAKE_LOCK_ACQUIRE_TOO_LONG_TIMEOUT);

先看obtainMessage调用了Message的obtain函数

        public final Message obtainMessage(int what)
        {
            return Message.obtain(this, what);
        }

Message的obtain函数就是新建一个Message，然后其target就是设置成其Handler

        public static Message obtain(Handler h, int what) {
            Message m = obtain();//就是新建一个Message
            m.target = h;
            m.what = what;
     
            return m;
        }

我们再联系之前分发消息

msg.target.dispatchMessage(msg);

最后就是调用Handler的dispatchMessage函数，最后在Handler中，最后会根据不同的情况对消息进行处理。

        public void dispatchMessage(Message msg) {
            if (msg.callback != null) {
                handleCallback(msg);//这种就是用post形式发送，带Runnable的
            } else {
                if (mCallback != null) {//这种是handler传参的时候就是传入了mCallback回调了
                    if (mCallback.handleMessage(msg)) {
                        return;
                    }
                }
                handleMessage(msg);//最后就是在自己实现的handleMessage处理
            }
        }

2.3 java层 消息发送

我们再看下java层的消息发送，主要也是调用Handler的sendMessage post之类函数，最终都会调用下面这个函数

        public boolean sendMessageAtTime(Message msg, long uptimeMillis) {
            MessageQueue queue = mQueue;
            if (queue == null) {
                RuntimeException e = new RuntimeException(
                        this + " sendMessageAtTime() called with no mQueue");
                Log.w("Looper", e.getMessage(), e);
                return false;
            }
            return enqueueMessage(queue, msg, uptimeMillis);
        }

我们再来看java层发送消息最终都会调用enqueueMessage函数

        private boolean enqueueMessage(MessageQueue queue, Message msg, long uptimeMillis) {
            msg.target = this;
            if (mAsynchronous) {
                msg.setAsynchronous(true);
            }
            return queue.enqueueMessage(msg, uptimeMillis);
        }

最终在enqueueMessage中，把消息加入消息队列，然后需要的话就调用c层的nativeWake函数

        boolean enqueueMessage(Message msg, long when) {
            if (msg.target == null) {
                throw new IllegalArgumentException("Message must have a target.");
            }
            if (msg.isInUse()) {
                throw new IllegalStateException(msg + " This message is already in use.");
            }
     
            synchronized (this) {
                if (mQuitting) {
                    IllegalStateException e = new IllegalStateException(
                            msg.target + " sending message to a Handler on a dead thread");
                    Log.w(TAG, e.getMessage(), e);
                    msg.recycle();
                    return false;
                }
     
                msg.markInUse();
                msg.when = when;
                Message p = mMessages;
                boolean needWake;
                if (p == null || when == 0 || when < p.when) {
                    // New head, wake up the event queue if blocked.
                    msg.next = p;
                    mMessages = msg;
                    needWake = mBlocked;
                } else {
                    // Inserted within the middle of the queue.  Usually we don't have to wake
                    // up the event queue unless there is a barrier at the head of the queue
                    // and the message is the earliest asynchronous message in the queue.
                    needWake = mBlocked && p.target == null && msg.isAsynchronous();
                    Message prev;
                    for (;;) {
                        prev = p;
                        p = p.next;
                        if (p == null || when < p.when) {
                            break;
                        }
                        if (needWake && p.isAsynchronous()) {
                            needWake = false;
                        }
                    }
                    msg.next = p; // invariant: p == prev.next
                    prev.next = msg;
                }
     
                // We can assume mPtr != 0 because mQuitting is false.
                if (needWake) {
                    nativeWake(mPtr);
                }
            }
            return true;
        }

我们看下这个native方法，最后也是调用了Looper的wake函数

    static void android_os_MessageQueue_nativeWake(JNIEnv* env, jclass clazz, jlong ptr) {
        NativeMessageQueue* nativeMessageQueue = reinterpret_cast<NativeMessageQueue*>(ptr);
        nativeMessageQueue->wake();
    }

    void NativeMessageQueue::wake() {
        mLooper->wake();
    }

Looper类的wake，函数只是往mWakeEventfd中写了一些内容，这个fd只是通知而已，类似pipe，最后会把epoll_wait唤醒，线程就不阻塞了继续先发送c层消息，然后处理之前addFd的事件，然后处理java层的消息。

    void Looper::wake() {
    #if DEBUG_POLL_AND_WAKE
        ALOGD("%p ~ wake", this);
    #endif
     
        uint64_t inc = 1;
        ssize_t nWrite = TEMP_FAILURE_RETRY(write(mWakeEventFd, &inc, sizeof(uint64_t)));
        if (nWrite != sizeof(uint64_t)) {
            if (errno != EAGAIN) {
                ALOGW("Could not write wake signal, errno=%d", errno);
            }
        }
    }

2.4 c层发送消息

在c层也是可以发送消息的，主要是调用Looper的sendMessageAtTime函数，参数有有一个handler是一个回调，我们把消息放在mMessageEnvelopes中。

    void Looper::sendMessageAtTime(nsecs_t uptime, const sp<MessageHandler>& handler,
            const Message& message) {
    #if DEBUG_CALLBACKS
        ALOGD("%p ~ sendMessageAtTime - uptime=%" PRId64 ", handler=%p, what=%d",
                this, uptime, handler.get(), message.what);
    #endif
     
        size_t i = 0;
        { // acquire lock
            AutoMutex _l(mLock);
     
            size_t messageCount = mMessageEnvelopes.size();
            while (i < messageCount && uptime >= mMessageEnvelopes.itemAt(i).uptime) {
                i += 1;
            }
     
            MessageEnvelope messageEnvelope(uptime, handler, message);
            mMessageEnvelopes.insertAt(messageEnvelope, i, 1);
     
            // Optimization: If the Looper is currently sending a message, then we can skip
            // the call to wake() because the next thing the Looper will do after processing
            // messages is to decide when the next wakeup time should be.  In fact, it does
            // not even matter whether this code is running on the Looper thread.
            if (mSendingMessage) {
                return;
            }
        } // release lock
     
        // Wake the poll loop only when we enqueue a new message at the head.
        if (i == 0) {
            wake();
        }
    }

当在pollOnce中，在epoll_wait之后，会遍历mMessageEnvelopes中的消息，然后调用其handler的handleMessage函数

        while (mMessageEnvelopes.size() != 0) {
            nsecs_t now = systemTime(SYSTEM_TIME_MONOTONIC);
            const MessageEnvelope& messageEnvelope = mMessageEnvelopes.itemAt(0);
            if (messageEnvelope.uptime <= now) {
                // Remove the envelope from the list.
                // We keep a strong reference to the handler until the call to handleMessage
                // finishes.  Then we drop it so that the handler can be deleted *before*
                // we reacquire our lock.
                { // obtain handler
                    sp<MessageHandler> handler = messageEnvelope.handler;
                    Message message = messageEnvelope.message;
                    mMessageEnvelopes.removeAt(0);
                    mSendingMessage = true;
                    mLock.unlock();
     
    #if DEBUG_POLL_AND_WAKE || DEBUG_CALLBACKS
                    ALOGD("%p ~ pollOnce - sending message: handler=%p, what=%d",
                            this, handler.get(), message.what);
    #endif
                    handler->handleMessage(message);
                } // release handler
     
                mLock.lock();
                mSendingMessage = false;
                result = POLL_CALLBACK;
            } else {
                // The last message left at the head of the queue determines the next wakeup time.
                mNextMessageUptime = messageEnvelope.uptime;
                break;
            }
        }

有一个Looper_test.cpp文件，里面介绍了很多Looper的使用方法，我们来看下

        sp<StubMessageHandler> handler = new StubMessageHandler();
        mLooper->sendMessageAtTime(now + ms2ns(100), handler, Message(MSG_TEST1));

StubMessageHandler继承MessageHandler就必须实现handleMessage方法

    class StubMessageHandler : public MessageHandler {
    public:
        Vector<Message> messages;
     
        virtual void handleMessage(const Message& message) {
            messages.push(message);
        }
    };

我们再顺便看下Message和MessageHandler类

    struct Message {
        Message() : what(0) { }
        Message(int what) : what(what) { }
     
        /* The message type. (interpretation is left up to the handler) */
        int what;
    };
     
     
    /**
     * Interface for a Looper message handler.
     *
     * The Looper holds a strong reference to the message handler whenever it has
     * a message to deliver to it.  Make sure to call Looper::removeMessages
     * to remove any pending messages destined for the handler so that the handler
     * can be destroyed.
     */
    class MessageHandler : public virtual RefBase {
    protected:
        virtual ~MessageHandler() { }
     
    public:
        /**
         * Handles a message.
         */
        virtual void handleMessage(const Message& message) = 0;
    };

2.5 c层addFd

我们也可以在Looper.cpp的addFd中增加fd放入线程epoll中，当fd有数据来我们也可以处理相应的数据，下面我们先来看下addFd函数，我们注意其中有一个callBack回调

    int Looper::addFd(int fd, int ident, int events, Looper_callbackFunc callback, void* data) {
        return addFd(fd, ident, events, callback ? new SimpleLooperCallback(callback) : NULL, data);
    }
     
    int Looper::addFd(int fd, int ident, int events, const sp<LooperCallback>& callback, void* data) {
    #if DEBUG_CALLBACKS
        ALOGD("%p ~ addFd - fd=%d, ident=%d, events=0x%x, callback=%p, data=%p", this, fd, ident,
                events, callback.get(), data);
    #endif
     
        if (!callback.get()) {
            if (! mAllowNonCallbacks) {
                ALOGE("Invalid attempt to set NULL callback but not allowed for this looper.");
                return -1;
            }
     
            if (ident < 0) {
                ALOGE("Invalid attempt to set NULL callback with ident < 0.");
                return -1;
            }
        } else {
            ident = POLL_CALLBACK;
        }
     
        { // acquire lock
            AutoMutex _l(mLock);
     
            Request request;
            request.fd = fd;
            request.ident = ident;
            request.events = events;
            request.seq = mNextRequestSeq++;
            request.callback = callback;
            request.data = data;
            if (mNextRequestSeq == -1) mNextRequestSeq = 0; // reserve sequence number -1
     
            struct epoll_event eventItem;
            request.initEventItem(&eventItem);
     
            ssize_t requestIndex = mRequests.indexOfKey(fd);
            if (requestIndex < 0) {
                int epollResult = epoll_ctl(mEpollFd, EPOLL_CTL_ADD, fd, & eventItem);//加入epoll
                if (epollResult < 0) {
                    ALOGE("Error adding epoll events for fd %d, errno=%d", fd, errno);
                    return -1;
                }
                mRequests.add(fd, request);//放入mRequests中
            } else {
                int epollResult = epoll_ctl(mEpollFd, EPOLL_CTL_MOD, fd, & eventItem);//更新
                if (epollResult < 0) {
                    if (errno == ENOENT) {
                        // Tolerate ENOENT because it means that an older file descriptor was
                        // closed before its callback was unregistered and meanwhile a new
                        // file descriptor with the same number has been created and is now
                        // being registered for the first time.  This error may occur naturally
                        // when a callback has the side-effect of closing the file descriptor
                        // before returning and unregistering itself.  Callback sequence number
                        // checks further ensure that the race is benign.
                        //
                        // Unfortunately due to kernel limitations we need to rebuild the epoll
                        // set from scratch because it may contain an old file handle that we are
                        // now unable to remove since its file descriptor is no longer valid.
                        // No such problem would have occurred if we were using the poll system
                        // call instead, but that approach carries others disadvantages.
    #if DEBUG_CALLBACKS
                        ALOGD("%p ~ addFd - EPOLL_CTL_MOD failed due to file descriptor "
                                "being recycled, falling back on EPOLL_CTL_ADD, errno=%d",
                                this, errno);
    #endif
                        epollResult = epoll_ctl(mEpollFd, EPOLL_CTL_ADD, fd, & eventItem);
                        if (epollResult < 0) {
                            ALOGE("Error modifying or adding epoll events for fd %d, errno=%d",
                                    fd, errno);
                            return -1;
                        }
                        scheduleEpollRebuildLocked();
                    } else {
                        ALOGE("Error modifying epoll events for fd %d, errno=%d", fd, errno);
                        return -1;
                    }
                }
                mRequests.replaceValueAt(requestIndex, request);
            }
        } // release lock
        return 1;
    }

在pollOnce函数中，我们先寻找mRequests中匹配的fd，然后在pushResponse中新建一个Response，然后把Response和Request匹配起来。

            } else {
                ssize_t requestIndex = mRequests.indexOfKey(fd);
                if (requestIndex >= 0) {
                    int events = 0;
                    if (epollEvents & EPOLLIN) events |= EVENT_INPUT;
                    if (epollEvents & EPOLLOUT) events |= EVENT_OUTPUT;
                    if (epollEvents & EPOLLERR) events |= EVENT_ERROR;
                    if (epollEvents & EPOLLHUP) events |= EVENT_HANGUP;
                    pushResponse(events, mRequests.valueAt(requestIndex));
                } else {
                    ALOGW("Ignoring unexpected epoll events 0x%x on fd %d that is "
                            "no longer registered.", epollEvents, fd);
                }
            }

下面我们就会遍历mResponses中的Response，然后调用其request中的回调

        for (size_t i = 0; i < mResponses.size(); i++) {
            Response& response = mResponses.editItemAt(i);
            if (response.request.ident == POLL_CALLBACK) {
                int fd = response.request.fd;
                int events = response.events;
                void* data = response.request.data;
    #if DEBUG_POLL_AND_WAKE || DEBUG_CALLBACKS
                ALOGD("%p ~ pollOnce - invoking fd event callback %p: fd=%d, events=0x%x, data=%p",
                        this, response.request.callback.get(), fd, events, data);
    #endif
                // Invoke the callback.  Note that the file descriptor may be closed by
                // the callback (and potentially even reused) before the function returns so
                // we need to be a little careful when removing the file descriptor afterwards.
                int callbackResult = response.request.callback->handleEvent(fd, events, data);
                if (callbackResult == 0) {
                    removeFd(fd, response.request.seq);
                }
     
                // Clear the callback reference in the response structure promptly because we
                // will not clear the response vector itself until the next poll.
                response.request.callback.clear();
                result = POLL_CALLBACK;
            }
        }

同样我们再来看看Looper_test.cpp是如何使用的？

        Pipe pipe;
        StubCallbackHandler handler(true);
     
        handler.setCallback(mLooper, pipe.receiveFd, Looper::EVENT_INPUT);

我们看下handler的setCallback函数

    class CallbackHandler {
    public:
        void setCallback(const sp<Looper>& looper, int fd, int events) {
            looper->addFd(fd, 0, events, staticHandler, this);//就是调用了looper的addFd函数，并且回调
        }
     
    protected:
        virtual ~CallbackHandler() { }
     
        virtual int handler(int fd, int events) = 0;
     
    private:
        static int staticHandler(int fd, int events, void* data) {//这个就是回调函数
            return static_cast<CallbackHandler*>(data)->handler(fd, events);
        }
    };
     
    class StubCallbackHandler : public CallbackHandler {
    public:
        int nextResult;
        int callbackCount;
     
        int fd;
        int events;
     
        StubCallbackHandler(int nextResult) : nextResult(nextResult),
                callbackCount(0), fd(-1), events(-1) {
        }
     
    protected:
        virtual int handler(int fd, int events) {//这个是通过回调函数再调到这里的
            callbackCount += 1;
            this->fd = fd;
            this->events = events;
            return nextResult;
        }
    };

我们结合Looper的addFd一起来看，当callback是有的，我们新建一个SimpleLooperCallback

    int Looper::addFd(int fd, int ident, int events, Looper_callbackFunc callback, void* data) {
        return addFd(fd, ident, events, callback ? new SimpleLooperCallback(callback) : NULL, data);
    }

这里的Looper_callbackFunc是一个typedef

typedef int (*Looper_callbackFunc)(int fd, int events, void* data);

我们再来看SimpleLooperCallback

    class SimpleLooperCallback : public LooperCallback {
    protected:
        virtual ~SimpleLooperCallback();
     
    public:
        SimpleLooperCallback(Looper_callbackFunc callback);
        virtual int handleEvent(int fd, int events, void* data);
     
    private:
        Looper_callbackFunc mCallback;
    };

    SimpleLooperCallback::SimpleLooperCallback(Looper_callbackFunc callback) :
            mCallback(callback) {
    }
     
    SimpleLooperCallback::~SimpleLooperCallback() {
    }
     
    int SimpleLooperCallback::handleEvent(int fd, int events, void* data) {
        return mCallback(fd, events, data);
    }

最后我们是调用callback->handleEvent(fd, events, data)，而callback就是SimpleLooperCallback，这里的data，之前传进来的就是CallbackHandler 的this指针
因此最后就是调用了staticHandler，而data->handler，就是this->handler，最后是虚函数就调用到了StubCallbackHandler 的handler函数中了。

当然我们也可以不用这么复杂，直接使用第二个addFd函数，当然callBack我们需要自己定义一个类来实现LooperCallBack类就行了，这样就简单多了。

int addFd(int fd, int ident, int events, const sp<LooperCallback>& callback, void* data);

2.6 java层addFd

一直以为只能在c层的Looper中才能addFd，原来在java层也通过jni做了这个功能。

我们可以在MessageQueue中的addOnFileDescriptorEventListener来实现这个功能

        public void addOnFileDescriptorEventListener(@NonNull FileDescriptor fd,
                @OnFileDescriptorEventListener.Events int events,
                @NonNull OnFileDescriptorEventListener listener) {
            if (fd == null) {
                throw new IllegalArgumentException("fd must not be null");
            }
            if (listener == null) {
                throw new IllegalArgumentException("listener must not be null");
            }
     
            synchronized (this) {
                updateOnFileDescriptorEventListenerLocked(fd, events, listener);
            }
        }

我们再来看看OnFileDescriptorEventListener 这个回调

        public interface OnFileDescriptorEventListener {
            public static final int EVENT_INPUT = 1 << 0;
            public static final int EVENT_OUTPUT = 1 << 1;
            public static final int EVENT_ERROR = 1 << 2;
     
            /** @hide */
            @Retention(RetentionPolicy.SOURCE)
            @IntDef(flag=true, value={EVENT_INPUT, EVENT_OUTPUT, EVENT_ERROR})
            public @interface Events {}
     
     
            @Events int onFileDescriptorEvents(@NonNull FileDescriptor fd, @Events int events);
        }

接着调用了updateOnFileDescriptorEventListenerLocked函数

        private void updateOnFileDescriptorEventListenerLocked(FileDescriptor fd, int events,
                OnFileDescriptorEventListener listener) {
            final int fdNum = fd.getInt$();
     
            int index = -1;
            FileDescriptorRecord record = null;
            if (mFileDescriptorRecords != null) {
                index = mFileDescriptorRecords.indexOfKey(fdNum);
                if (index >= 0) {
                    record = mFileDescriptorRecords.valueAt(index);
                    if (record != null && record.mEvents == events) {
                        return;
                    }
                }
            }
     
            if (events != 0) {
                events |= OnFileDescriptorEventListener.EVENT_ERROR;
                if (record == null) {
                    if (mFileDescriptorRecords == null) {
                        mFileDescriptorRecords = new SparseArray<FileDescriptorRecord>();
                    }
                    record = new FileDescriptorRecord(fd, events, listener);//fd保存在FileDescriptorRecord对象
                    mFileDescriptorRecords.put(fdNum, record);//mFileDescriptorRecords然后保存在
                } else {
                    record.mListener = listener;
                    record.mEvents = events;
                    record.mSeq += 1;
                }
                nativeSetFileDescriptorEvents(mPtr, fdNum, events);//调用native函数
            } else if (record != null) {
                record.mEvents = 0;
                mFileDescriptorRecords.removeAt(index);
            }
        }

native最后调用了NativeMessageQueue的setFileDescriptorEvents函数

    static void android_os_MessageQueue_nativeSetFileDescriptorEvents(JNIEnv* env, jclass clazz,
            jlong ptr, jint fd, jint events) {
        NativeMessageQueue* nativeMessageQueue = reinterpret_cast<NativeMessageQueue*>(ptr);
        nativeMessageQueue->setFileDescriptorEvents(fd, events);
    }

setFileDescriptorEvents函数，这个addFd就是调用的第二个addFd，因此我们可以肯定NativeMessageQueue继承了LooperCallback

    void NativeMessageQueue::setFileDescriptorEvents(int fd, int events) {
        if (events) {
            int looperEvents = 0;
            if (events & CALLBACK_EVENT_INPUT) {
                looperEvents |= Looper::EVENT_INPUT;
            }
            if (events & CALLBACK_EVENT_OUTPUT) {
                looperEvents |= Looper::EVENT_OUTPUT;
            }
            mLooper->addFd(fd, Looper::POLL_CALLBACK, looperEvents, this,
                    reinterpret_cast<void*>(events));
        } else {
            mLooper->removeFd(fd);
        }
    }

果然是，需要实现handleEvent函数

    class NativeMessageQueue : public MessageQueue, public LooperCallback {
    public:
        NativeMessageQueue();
        virtual ~NativeMessageQueue();
     
        virtual void raiseException(JNIEnv* env, const char* msg, jthrowable exceptionObj);
     
        void pollOnce(JNIEnv* env, jobject obj, int timeoutMillis);
        void wake();
        void setFileDescriptorEvents(int fd, int events);
     
        virtual int handleEvent(int fd, int events, void* data);

handleEvent就是在looper中epoll_wait之后，当我们增加的fd有数据就会调用这个函数

    int NativeMessageQueue::handleEvent(int fd, int looperEvents, void* data) {
        int events = 0;
        if (looperEvents & Looper::EVENT_INPUT) {
            events |= CALLBACK_EVENT_INPUT;
        }
        if (looperEvents & Looper::EVENT_OUTPUT) {
            events |= CALLBACK_EVENT_OUTPUT;
        }
        if (looperEvents & (Looper::EVENT_ERROR | Looper::EVENT_HANGUP | Looper::EVENT_INVALID)) {
            events |= CALLBACK_EVENT_ERROR;
        }
        int oldWatchedEvents = reinterpret_cast<intptr_t>(data);
        int newWatchedEvents = mPollEnv->CallIntMethod(mPollObj,
                gMessageQueueClassInfo.dispatchEvents, fd, events); //调用回调
        if (!newWatchedEvents) {
            return 0; // unregister the fd
        }
        if (newWatchedEvents != oldWatchedEvents) {
            setFileDescriptorEvents(fd, newWatchedEvents);
        }
        return 1;
    }

最后在java的MessageQueue中的dispatchEvents就是在jni层反调过来的，然后调用之前注册的回调函数

        // Called from native code.
        private int dispatchEvents(int fd, int events) {
            // Get the file descriptor record and any state that might change.
            final FileDescriptorRecord record;
            final int oldWatchedEvents;
            final OnFileDescriptorEventListener listener;
            final int seq;
            synchronized (this) {
                record = mFileDescriptorRecords.get(fd);//通过fd得到FileDescriptorRecord
                if (record == null) {
                    return 0; // spurious, no listener registered
                }
     
                oldWatchedEvents = record.mEvents;
                events &= oldWatchedEvents; // filter events based on current watched set
                if (events == 0) {
                    return oldWatchedEvents; // spurious, watched events changed
                }
     
                listener = record.mListener;
                seq = record.mSeq;
            }
     
            // Invoke the listener outside of the lock.
            int newWatchedEvents = listener.onFileDescriptorEvents(//listener回调
                    record.mDescriptor, events);
            if (newWatchedEvents != 0) {
                newWatchedEvents |= OnFileDescriptorEventListener.EVENT_ERROR;
            }
     
            // Update the file descriptor record if the listener changed the set of
            // events to watch and the listener itself hasn't been updated since.
            if (newWatchedEvents != oldWatchedEvents) {
                synchronized (this) {
                    int index = mFileDescriptorRecords.indexOfKey(fd);
                    if (index >= 0 && mFileDescriptorRecords.valueAt(index) == record
                            && record.mSeq == seq) {
                        record.mEvents = newWatchedEvents;
                        if (newWatchedEvents == 0) {
                            mFileDescriptorRecords.removeAt(index);
                        }
                    }
                }
            }
     
            // Return the new set of events to watch for native code to take care of.
            return newWatchedEvents;
        }

 
---------------------  
作者：kc专栏  
来源：优快云  
原文：https://blog.youkuaiyun.com/kc58236582/article/details/51883024  
版权声明：本文为博主原创文章，转载请附上博文链接！