java多线程设计模式(2)

思考

Q: 服务器线程的wait条件while(queue.size()==0)能否换成if(queue.size()==0)?

A: 在这个例子中可以，因为服务器线程只有一个。但是，如果服务器线程有多个（例如Web应用程序有多个线程处理并发请求，这非常普遍），就会造成严重问题。

Q: 能否用sleep(1000)代替wait()?

A: 绝对不可以。sleep()不会释放锁，因此sleep期间别的线程根本没有办法调用getRequest()和putRequest()，导致所有相关线程都被阻塞。

Q: (Request)queue.remove(0)可以放到synchronized() {}块外面吗？

A: 不可以。因为while()是测试queue，remove()是使用queue，两者是一个原子操作，不能放在synchronized外面。

总结

多线程设计看似简单，实际上必须非常仔细地考虑各种锁定/同步的条件，稍不小心，就可能出错。并且，当线程较少时，很可能发现不了问题，一旦问题出现又难以调试。

所幸的是，已有一些被验证过的模式可以供我们使用，我们会继续介绍一些常用的多线程设计模式。

前面谈了多线程应用程序能极大地改善用户相应。例如对于一个Web应用程序，每当一个用户请求服务器连接时，服务器就可以启动一个新线程为用户服务。

然而，创建和销毁线程本身就有一定的开销，如果频繁创建和销毁线程，CPU和内存开销就不可忽略，垃圾收集器还必须负担更多的工作。因此，线程池就是为了避免频繁创建和销毁线程。

每当服务器接受了一个新的请求后，服务器就从线程池中挑选一个等待的线程并执行请求处理。处理完毕后，线程并不结束，而是转为阻塞状态再次被放入线程池中。这样就避免了频繁创建和销毁线程。

Worker Pattern实现了类似线程池的功能。首先定义Task接口：

package com.crackj2ee.thread;
public interface Task {
    void execute();
}

线程将负责执行execute()方法。注意到任务是由子类通过实现execute()方法实现的，线程本身并不知道自己执行的任务。它只负责运行一个耗时的execute()方法。

具体任务由子类实现，我们定义了一个CalculateTask和一个TimerTask：

// CalculateTask.java
package com.crackj2ee.thread;
public class CalculateTask implements Task {
    private static int count = 0;
    private int num = count;
    public CalculateTask() {
        count++;
    }
    public void execute() {
        System.out.println("[CalculateTask " + num + "] start...");
        try {
            Thread.sleep(3000);
        }
        catch(InterruptedException ie) {}
        System.out.println("[CalculateTask " + num + "] done.");
    }
}

// TimerTask.java
package com.crackj2ee.thread;
public class TimerTask implements Task {
    private static int count = 0;
    private int num = count;
    public TimerTask() {
        count++;
    }
    public void execute() {
        System.out.println("[TimerTask " + num + "] start...");
        try {
            Thread.sleep(2000);
        }
        catch(InterruptedException ie) {}
        System.out.println("[TimerTask " + num + "] done.");
    }
}

以上任务均简单的sleep若干秒。

TaskQueue实现了一个队列，客户端可以将请求放入队列，服务器线程可以从队列中取出任务：

package com.crackj2ee.thread;
import java.util.*;
public class TaskQueue {
    private List queue = new LinkedList();
    public synchronized Task getTask() {
        while(queue.size()==0) {
            try {
                this.wait();
            }
            catch(InterruptedException ie) {
                return null;
            }
        }
        return (Task)queue.remove(0);
    }
    public synchronized void putTask(Task task) {
        queue.add(task);
        this.notifyAll();
    }
}

终于到了真正的WorkerThread，这是真正执行任务的服务器线程：

package com.crackj2ee.thread;
public class WorkerThread extends Thread {
    private static int count = 0;
    private boolean busy = false;
    private boolean stop = false;
    private TaskQueue queue;
    public WorkerThread(ThreadGroup group, TaskQueue queue) {
        super(group, "worker-" + count);
        count++;
        this.queue = queue;
    }
    public void shutdown() {
        stop = true;
        this.interrupt();
        try {
            this.join();
        }
        catch(InterruptedException ie) {}
    }
    public boolean isIdle() {
        return !busy;
    }
    public void run() {
        System.out.println(getName() + " start.");       
        while(!stop) {
            Task task = queue.getTask();
            if(task!=null) {
                busy = true;
                task.execute();
                busy = false;
            }
        }
        System.out.println(getName() + " end.");
    }
}

前面已经讲过，queue.getTask()是一个阻塞方法，服务器线程可能在此wait()一段时间。此外，WorkerThread还有一个shutdown方法，用于安全结束线程。

最后是ThreadPool，负责管理所有的服务器线程，还可以动态增加和减少线程数：

package com.crackj2ee.thread;
import java.util.*;
public class ThreadPool extends ThreadGroup {
    private List threads = new LinkedList();
    private TaskQueue queue;
    public ThreadPool(TaskQueue queue) {
        super("Thread-Pool");
        this.queue = queue;
    }
    public synchronized void addWorkerThread() {
        Thread t = new WorkerThread(this, queue);
        threads.add(t);
        t.start();
    }
    public synchronized void removeWorkerThread() {
        if(threads.size()>0) {
            WorkerThread t = (WorkerThread)threads.remove(0);
            t.shutdown();
        }
    }
    public synchronized void currentStatus() {
        System.out.println("-----------------------------------------------");
        System.out.println("Thread count = " + threads.size());
        Iterator it = threads.iterator();
        while(it.hasNext()) {
            WorkerThread t = (WorkerThread)it.next();
            System.out.println(t.getName() + ": " + (t.isIdle() ? "idle" : "busy"));
        }
        System.out.println("-----------------------------------------------");
    }
}

currentStatus()方法是为了方便调试，打印出所有线程的当前状态。

最后，Main负责完成main()方法：

package com.crackj2ee.thread;
public class Main {
    public static void main(String[] args) {
        TaskQueue queue = new TaskQueue();
        ThreadPool pool = new ThreadPool(queue);
        for(int i=0; i<10; i++) {
            queue.putTask(new CalculateTask());
            queue.putTask(new TimerTask());
        }
        pool.addWorkerThread();
        pool.addWorkerThread();
        doSleep(8000);
        pool.currentStatus();
        pool.addWorkerThread();
        pool.addWorkerThread();
        pool.addWorkerThread();
        pool.addWorkerThread();
        pool.addWorkerThread();
        doSleep(5000);
        pool.currentStatus();
    }
    private static void doSleep(long ms) {
        try {
            Thread.sleep(ms);
        }
        catch(InterruptedException ie) {}
    }
}

main()一开始放入了20个Task，然后动态添加了一些服务线程，并定期打印线程状态，运行结果如下：

worker-0 start.
[CalculateTask 0] start...
worker-1 start.
[TimerTask 0] start...
[TimerTask 0] done.
[CalculateTask 1] start...
[CalculateTask 0] done.
[TimerTask 1] start...
[CalculateTask 1] done.
[CalculateTask 2] start...
[TimerTask 1] done.
[TimerTask 2] start...
[TimerTask 2] done.
[CalculateTask 3] start...
-----------------------------------------------
Thread count = 2
worker-0: busy
worker-1: busy
-----------------------------------------------
[CalculateTask 2] done.
[TimerTask 3] start...
worker-2 start.
[CalculateTask 4] start...
worker-3 start.
[TimerTask 4] start...
worker-4 start.
[CalculateTask 5] start...
worker-5 start.
[TimerTask 5] start...
worker-6 start.
[CalculateTask 6] start...
[CalculateTask 3] done.
[TimerTask 6] start...
[TimerTask 3] done.
[CalculateTask 7] start...
[TimerTask 4] done.
[TimerTask 7] start...
[TimerTask 5] done.
[CalculateTask 8] start...
[CalculateTask 4] done.
[TimerTask 8] start...
[CalculateTask 5] done.
[CalculateTask 9] start...
[CalculateTask 6] done.
[TimerTask 9] start...
[TimerTask 6] done.
[TimerTask 7] done.
-----------------------------------------------
Thread count = 7
worker-0: idle
worker-1: busy
worker-2: busy
worker-3: idle
worker-4: busy
worker-5: busy
worker-6: busy
-----------------------------------------------
[CalculateTask 7] done.
[CalculateTask 8] done.
[TimerTask 8] done.
[TimerTask 9] done.
[CalculateTask 9] done.

仔细观察：一开始只有两个服务器线程，因此线程状态都是忙，后来线程数增多，6个线程中的两个状态变成idle，说明处于wait()状态。

思考：本例的线程调度算法其实根本没有，因为这个应用是围绕TaskQueue设计的，不是以Thread Pool为中心设计的。因此，Task调度取决于TaskQueue的getTask()方法，你可以改进这个方法，例如使用优先队列，使优先级高的任务先被执行。

如果所有的服务器线程都处于busy状态，则说明任务繁忙，TaskQueue的队列越来越长，最终会导致服务器内存耗尽。因此，可以限制TaskQueue的等待任务数，超过最大长度就拒绝处理。许多Web服务器在用户请求繁忙时就会拒绝用户：HTTP 503 SERVICEUNAVAILABLE

多线程读写同一个对象的数据是很普遍的，通常，要避免读写冲突，必须保证任何时候仅有一个线程在写入，有线程正在读取的时候，写入操作就必须等待。简单说，就是要避免“写-写”冲突和“读-写”冲突。但是同时读是允许的，因为“读-读”不冲突，而且很安全。

要实现以上的ReadWriteLock，简单的使用synchronized就不行，我们必须自己设计一个ReadWriteLock类，在读之前，必须先获得“读锁”，写之前，必须先获得“写锁”。举例说明：

DataHandler对象保存了一个可读写的char[]数组：

package com.crackj2ee.thread;

public class DataHandler {
    // store data:
    private char[] buffer = "AAAAAAAAAA".toCharArray();

    private char[] doRead() {
        char[] ret = new char[buffer.length];
        for(int i=0; i<buffer.length; i++) {
            ret[i] = buffer[i];
            sleep(3);
        }
        return ret;
    }

    private void doWrite(char[] data) {
        if(data!=null) {
            buffer = new char[data.length];
            for(int i=0; i<buffer.length; i++) {
                buffer[i] = data[i];
                sleep(10);
            }
        }
    }

    private void sleep(int ms) {
        try {
            Thread.sleep(ms);
        }
        catch(InterruptedException ie) {}
    }
}

doRead()和doWrite()方法是非线程安全的读写方法。为了演示，加入了sleep()，并设置读的速度大约是写的3倍，这符合通常的情况。

为了让多线程能安全读写，我们设计了一个ReadWriteLock：

package com.crackj2ee.thread;
public class ReadWriteLock {
    private int readingThreads = 0;
    private int writingThreads = 0;
    private int waitingThreads = 0; // waiting for write
    private boolean divferWrite = true;

    public synchronized void readLock() throws InterruptedException {
        while(writingThreads>0 || (divferWrite && waitingThreads>0))
            this.wait();
        readingThreads++;
    }

    public synchronized void readUnlock() {
        readingThreads--;
        divferWrite = true;
        notifyAll();
    }

    public synchronized void writeLock() throws InterruptedException {
        waitingThreads++;
        try {
            while(readingThreads>0 || writingThreads>0)
                this.wait();
        }
        finally {
            waitingThreads--;
        }
        writingThreads++;
    }

    public synchronized void writeUnlock() {
        writingThreads--;
        divferWrite = false;
        notifyAll();
    }
}

readLock()用于获得读锁，readUnlock()释放读锁，writeLock()和writeUnlock()一样。由于锁用完必须释放，因此，必须保证lock和unlock匹配。我们修改DataHandler，加入ReadWriteLock：

package com.crackj2ee.thread;
public class DataHandler {
    // store data:
    private char[] buffer = "AAAAAAAAAA".toCharArray();
    // lock:
    private ReadWriteLock lock = new ReadWriteLock();

    public char[] read(String name) throws InterruptedException {
        System.out.println(name + " waiting for read...");
        lock.readLock();
        try {
            char[] data = doRead();
            System.out.println(name + " reads data: " + new String(data));
            return data;
        }
        finally {
            lock.readUnlock();
        }
    }

    public void write(String name, char[] data) throws InterruptedException {
        System.out.println(name + " waiting for write...");
        lock.writeLock();
        try {
            System.out.println(name + " wrote data: " + new String(data));
            doWrite(data);
        }
        finally {
            lock.writeUnlock();
        }
    }

    private char[] doRead() {
        char[] ret = new char[buffer.length];
        for(int i=0; i<buffer.length; i++) {
            ret[i] = buffer[i];
            sleep(3);
        }
        return ret;
    }
    private void doWrite(char[] data) {
        if(data!=null) {
            buffer = new char[data.length];
            for(int i=0; i<buffer.length; i++) {
                buffer[i] = data[i];
                sleep(10);
            }
        }
    }
    private void sleep(int ms) {
        try {
            Thread.sleep(ms);
        }
        catch(InterruptedException ie) {}
    }
}

public方法read()和write()完全封装了底层的ReadWriteLock，因此，多线程可以安全地调用这两个方法：

// ReadingThread不断读取数据：
package com.crackj2ee.thread;
public class ReadingThread extends Thread {
    private DataHandler handler;
    public ReadingThread(DataHandler handler) {
        this.handler = handler;
    }
    public void run() {
        for(;;) {
            try {
                char[] data = handler.read(getName());
                Thread.sleep((long)(Math.random()*1000+100));
            }
            catch(InterruptedException ie) {
                break;
            }
        }
    }
}

// WritingThread不断写入数据，每次写入的都是10个相同的字符：
package com.crackj2ee.thread;
public class WritingThread extends Thread {
    private DataHandler handler;
    public WritingThread(DataHandler handler) {
        this.handler = handler;
    }
    public void run() {
        char[] data = new char[10];
        for(;;) {
            try {
                fill(data);
                handler.write(getName(), data);
                Thread.sleep((long)(Math.random()*1000+100));
            }
            catch(InterruptedException ie) {
                break;
            }
        }
    }
    // 产生一个A-Z随机字符，填入char[10]:
    private void fill(char[] data) {
        char c = (char)(Math.random()*26+'A');
        for(int i=0; i<data.length; i++)
            data[i] = c;
    }
}

最后Main负责启动这些线程：

package com.crackj2ee.thread;
public class Main {
    public static void main(String[] args) {
        DataHandler handler = new DataHandler();
        Thread[] ts = new Thread[] {
                new ReadingThread(handler),
                new ReadingThread(handler),
                new ReadingThread(handler),
                new ReadingThread(handler),
                new ReadingThread(handler),
                new WritingThread(handler),
                new WritingThread(handler)
        };
        for(int i=0; i<ts.length; i++) {
            ts[i].start();
        }
    }
}

我们启动了5个读线程和2个写线程，运行结果如下：

Thread-0 waiting for read...
Thread-1 waiting for read...
Thread-2 waiting for read...
Thread-3 waiting for read...
Thread-4 waiting for read...
Thread-5 waiting for write...
Thread-6 waiting for write...
Thread-4 reads data: AAAAAAAAAA
Thread-3 reads data: AAAAAAAAAA
Thread-2 reads data: AAAAAAAAAA
Thread-1 reads data: AAAAAAAAAA
Thread-0 reads data: AAAAAAAAAA
Thread-5 wrote data: EEEEEEEEEE
Thread-6 wrote data: MMMMMMMMMM
Thread-1 waiting for read...
Thread-4 waiting for read...
Thread-1 reads data: MMMMMMMMMM
Thread-4 reads data: MMMMMMMMMM
Thread-2 waiting for read...
Thread-2 reads data: MMMMMMMMMM
Thread-0 waiting for read...
Thread-0 reads data: MMMMMMMMMM
Thread-4 waiting for read...
Thread-4 reads data: MMMMMMMMMM
Thread-2 waiting for read...
Thread-5 waiting for write...
Thread-2 reads data: MMMMMMMMMM
Thread-5 wrote data: GGGGGGGGGG
Thread-6 waiting for write...
Thread-6 wrote data: AAAAAAAAAA
Thread-3 waiting for read...
Thread-3 reads data: AAAAAAAAAA
......

可以看到，每次读/写都是完整的原子操作，因为我们每次写入的都是10个相同字符。并且，每次读出的都是最近一次写入的内容。

如果去掉ReadWriteLock：

package com.crackj2ee.thread;
public class DataHandler {

    // store data:
    private char[] buffer = "AAAAAAAAAA".toCharArray();

    public char[] read(String name) throws InterruptedException {
        char[] data = doRead();
        System.out.println(name + " reads data: " + new String(data));
        return data;
    }
    public void write(String name, char[] data) throws InterruptedException {
        System.out.println(name + " wrote data: " + new String(data));
        doWrite(data);
    }

    private char[] doRead() {
        char[] ret = new char[10];
        for(int i=0; i<10; i++) {
            ret[i] = buffer[i];
            sleep(3);
        }
        return ret;
    }
    private void doWrite(char[] data) {
        for(int i=0; i<10; i++) {
            buffer[i] = data[i];
            sleep(10);
        }
    }
    private void sleep(int ms) {
        try {
            Thread.sleep(ms);
        }
        catch(InterruptedException ie) {}
    }
}

运行结果如下：

Thread-5 wrote data: AAAAAAAAAA
Thread-6 wrote data: MMMMMMMMMM
Thread-0 reads data: AAAAAAAAAA
Thread-1 reads data: AAAAAAAAAA
Thread-2 reads data: AAAAAAAAAA
Thread-3 reads data: AAAAAAAAAA
Thread-4 reads data: AAAAAAAAAA
Thread-2 reads data: MAAAAAAAAA
Thread-3 reads data: MAAAAAAAAA
Thread-5 wrote data: CCCCCCCCCC
Thread-1 reads data: MAAAAAAAAA
Thread-0 reads data: MAAAAAAAAA
Thread-4 reads data: MAAAAAAAAA
Thread-6 wrote data: EEEEEEEEEE
Thread-3 reads data: EEEEECCCCC
Thread-4 reads data: EEEEEEEEEC
Thread-1 reads data: EEEEEEEEEE

可以看到在Thread-6写入EEEEEEEEEE的过程中，3个线程读取的内容是不同的。

思考

java的synchronized提供了最底层的物理锁，要在synchronized的基础上，实现自己的逻辑锁，就必须仔细设计ReadWriteLock。

Q: lock.readLock()为什么不放入try{ } 内？
A:因为readLock()会抛出InterruptedException，导致readingThreads++不执行，而readUnlock()在finally{ } 中，导致readingThreads--执行，从而使readingThread状态出错。writeLock()也是类似的。

Q: divferWrite有用吗？
A:如果去掉divferWrite，线程安全不受影响。但是，如果读取线程很多，上一个线程还没有读取完，下一个线程又开始读了，就导致写入线程长时间 无法获得writeLock；如果写入线程等待的很多，一个接一个写，也会导致读取线程长时间无法获得readLock。divferWrite的作用是 让读/写交替执行，避免由于读线程繁忙导致写无法进行和由于写线程繁忙导致读无法进行。

Q: notifyAll()换成notify()行不行？
A:不可以。由于divferWrite的存在，如果一个线程刚读取完毕，此时divferWrite=true，再notify()，若恰好唤醒的是一 个读线程，则while(writingThreads>0 || (divferWrite &&waitingThreads>0))可能为true导致该读线程继续等待，而等待写入的线程也处于wait()中，结果所有 线程都处于wait()状态，谁也无法唤醒谁。因此，notifyAll()比notify()要来得安全。程序验证notify()带来的死锁：

Thread-0 waiting for read...
Thread-1 waiting for read...
Thread-2 waiting for read...
Thread-3 waiting for read...
Thread-4 waiting for read...
Thread-5 waiting for write...
Thread-6 waiting for write...
Thread-0 reads data: AAAAAAAAAA
Thread-4 reads data: AAAAAAAAAA
Thread-3 reads data: AAAAAAAAAA
Thread-2 reads data: AAAAAAAAAA
Thread-1 reads data: AAAAAAAAAA
Thread-5 wrote data: CCCCCCCCCC
Thread-2 waiting for read...
Thread-1 waiting for read...
Thread-3 waiting for read...
Thread-0 waiting for read...
Thread-4 waiting for read...
Thread-6 wrote data: LLLLLLLLLL
Thread-5 waiting for write...
Thread-6 waiting for write...
Thread-2 reads data: LLLLLLLLLL
Thread-2 waiting for read...
（运行到此不动 了）

注意到这种死锁是由于所有线程都在等待别的线程唤醒自己，结果都无法醒过来。这和两个线程希望获得对方已有的锁造成死锁不同。因此多线程设计的难度远远高于单线程应用。