Java 并发编程-再谈 AbstractQueuedSynchronizer 2:共享模式与基于 Condition 的等待 / 通知机制实现

共享模式acquire实现流程

上文我们讲解了AbstractQueuedSynchronizer独占模式的acquire实现流程,本文趁热打铁继续看一下AbstractQueuedSynchronizer共享模式acquire的实现流程。连续两篇文章的学习,也可以对比独占模式acquire和共享模式acquire的区别,加深对于AbstractQueuedSynchronizer的理解。

先看一下共享模式acquire的实现,方法为acquireShared和acquireSharedInterruptibly,两者差别不大,区别就在于后者有中断处理,以acquireShared为例:

1
2
3
4
public final void acquireShared(int arg) {
      if (tryAcquireShared(arg) < 0)
            doAcquireShared(arg);
 }

这里就能看出第一个差别来了:独占模式acquire的时候子类重写的方法tryAcquire返回的是boolean,即是否tryAcquire成功;共享模式acquire的时候,返回的是一个int型变量,判断是否<0。doAcquireShared方法的实现为:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
private void doAcquireShared(int arg) {
    final Node node = addWaiter(Node.SHARED);
    boolean failed = true;
    try {
        boolean interrupted = false;
        for (;;) {
            final Node p = node.predecessor();
            if (p == head) {
                int r = tryAcquireShared(arg);
                if (r >= 0) {
                    setHeadAndPropagate(node, r);
                    p.next = null; // help GC
                    if (interrupted)
                        selfInterrupt();
                    failed = false;
                    return;
                }
            }
            if (shouldParkAfterFailedAcquire(p, node) &&
                parkAndCheckInterrupt())
                interrupted = true;
        }
    } finally {
        if (failed)
            cancelAcquire(node);
    }
}

我们来分析一下这段代码做了什么:

  1. addWaiter,把所有tryAcquireShared<0的线程实例化出一个Node,构建为一个FIFO队列,这和独占锁是一样的
  2. 拿当前节点的前驱节点,只有前驱节点是head的节点才能tryAcquireShared,这和独占锁也是一样的
  3. 前驱节点不是head的,执行”shouldParkAfterFailedAcquire() && parkAndCheckInterrupt()”,for(;;)循环,”shouldParkAfterFailedAcquire()”方法执行2次,当前线程阻塞,这和独占锁也是一样的

确实,共享模式下的acquire和独占模式下的acquire大部分逻辑差不多,最大的差别在于tryAcquireShared成功之后,独占模式的acquire是直接将当前节点设置为head节点即可,共享模式会执行setHeadAndPropagate方法,顾名思义,即在设置head之后多执行了一步propagate操作。setHeadAndPropagate方法源码为:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
private void setHeadAndPropagate(Node node, int propagate) {
    Node h = head; // Record old head for check below
    setHead(node);
    /*
     * Try to signal next queued node if:
     *   Propagation was indicated by caller,
     *     or was recorded (as h.waitStatus) by a previous operation
     *     (note: this uses sign-check of waitStatus because
     *      PROPAGATE status may transition to SIGNAL.)
     * and
     *   The next node is waiting in shared mode,
     *     or we don't know, because it appears null
     *
     * The conservatism in both of these checks may cause
     * unnecessary wake-ups, but only when there are multiple
     * racing acquires/releases, so most need signals now or soon
     * anyway.
     */
    if (propagate > 0 || h == null || h.waitStatus < 0) {
        Node s = node.next;
        if (s == null || s.isShared())
            doReleaseShared();
    }
}

第3行的代码设置重设head,第2行的代码由于第3行的代码要重设head,因此先定义一个Node型变量h获得原head的地址,这两行代码很简单。

第19行~第23行的代码是独占锁和共享锁最不一样的一个地方,我们再看独占锁acquireQueued的代码:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
final boolean acquireQueued(final Node node, int arg) {
    boolean failed = true;
    try {
        boolean interrupted = false;
        for (;;) {
            final Node p = node.predecessor();
            if (p == head && tryAcquire(arg)) {
                setHead(node);
                p.next = null; // help GC
                failed = false;
                return interrupted;
            }
            if (shouldParkAfterFailedAcquire(p, node) &&
                parkAndCheckInterrupt())
                interrupted = true;
        }
    } finally {
        if (failed)
            cancelAcquire(node);
    }
}

这意味着独占锁某个节点被唤醒之后,它只需要将这个节点设置成head就完事了,而共享锁不一样,某个节点被设置为head之后,如果它的后继节点是SHARED状态的,那么将继续通过doReleaseShared方法尝试往后唤醒节点,实现了共享状态的向后传播。

共享模式release实现流程

上面讲了共享模式下acquire是如何实现的,下面再看一下release的实现流程,方法为releaseShared:

1
2
3
4
5
6
7
public final boolean releaseShared(int arg) {
    if (tryReleaseShared(arg)) {
        doReleaseShared();
        return true;
    }
    return false;
}

tryReleaseShared方法是子类实现的,如果tryReleaseShared成功,那么执行doReleaseShared()方法:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
private void doReleaseShared() {
    /*
     * Ensure that a release propagates, even if there are other
     * in-progress acquires/releases.  This proceeds in the usual
     * way of trying to unparkSuccessor of head if it needs
     * signal. But if it does not, status is set to PROPAGATE to
     * ensure that upon release, propagation continues.
     * Additionally, we must loop in case a new node is added
     * while we are doing this. Also, unlike other uses of
     * unparkSuccessor, we need to know if CAS to reset status
     * fails, if so rechecking.
     */
    for (;;) {
        Node h = head;
        if (h != null && h != tail) {
            int ws = h.waitStatus;
            if (ws == Node.SIGNAL) {
                if (!compareAndSetWaitStatus(h, Node.SIGNAL, 0))
                    continue;            // loop to recheck cases
                unparkSuccessor(h);
            }
            else if (ws == 0 &&
                     !compareAndSetWaitStatus(h, 0, Node.PROPAGATE))
                continue;                // loop on failed CAS
        }
        if (h == head)                   // loop if head changed
            break;
    }
}

主要是两层逻辑:

  1. 头结点本身的waitStatus是SIGNAL且能通过CAS算法将头结点的waitStatus从SIGNAL设置为0,唤醒头结点的后继节点
  2. 头结点本身的waitStatus是0的话,尝试将其设置为PROPAGATE状态的,意味着共享状态可以向后传播

Condition的await()方法实现原理—-构建等待队列

我们知道,Condition是用于实现通知/等待机制的,和Object的wait()/notify()一样,由于本文之前描述AbstractQueuedSynchronizer的共享模式的篇幅不是很长,加之Condition也是AbstractQueuedSynchronizer的一部分,因此将Condition也放在这里写了。

Condition分为await()和signal()两部分,前者用于等待、后者用于唤醒,首先看一下await()是如何实现的。Condition本身是一个接口,其在AbstractQueuedSynchronizer中的实现为ConditionObject:

1
2
3
4
5
6
7
8
9
public class ConditionObject implements Condition, java.io.Serializable {
        private static final long serialVersionUID = 1173984872572414699L;
        /** First node of condition queue. */
        private transient Node firstWaiter;
        /** Last node of condition queue. */
        private transient Node lastWaiter;
         
        ...
}

这里贴了一些字段定义,后面都是方法就不贴了,会对重点方法进行分析的。从字段定义我们可以看到,ConditionObject全局性地记录了第一个等待的节点与最后一个等待的节点。

像ReentrantLock每次要使用ConditionObject,直接new一个ConditionObject出来即可。我们关注一下await()方法的实现:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
public final void await() throws InterruptedException {
    if (Thread.interrupted())
        throw new InterruptedException();
    Node node = addConditionWaiter();
    int savedState = fullyRelease(node);
    int interruptMode = 0;
    while (!isOnSyncQueue(node)) {
        LockSupport.park(this);
        if ((interruptMode = checkInterruptWhileWaiting(node)) != 0)
            break;
    }
    if (acquireQueued(node, savedState) && interruptMode != THROW_IE)
        interruptMode = REINTERRUPT;
    if (node.nextWaiter != null) // clean up if cancelled
        unlinkCancelledWaiters();
    if (interruptMode != 0)
        reportInterruptAfterWait(interruptMode);
}

第2行~第3行的代码用于处理中断,第4行代码比较关键,添加Condition的等待者,看一下实现:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
private Node addConditionWaiter() {
    Node t = lastWaiter;
    // If lastWaiter is cancelled, clean out.
    if (t != null && t.waitStatus != Node.CONDITION) {
        unlinkCancelledWaiters();
        t = lastWaiter;
    }
    Node node = new Node(Thread.currentThread(), Node.CONDITION);
    if (t == null)
        firstWaiter = node;
    else
        t.nextWaiter = node;
    lastWaiter = node;
    return node;
}

首先拿到队列(注意数据结构,Condition构建出来的也是一个队列)中最后一个等待者,紧接着第4行的的判断,判断最后一个等待者的waitStatus不是CONDITION的话,执行第5行的代码,解绑取消的等待者,因为通过第8行的代码,我们看到,new出来的Node的状态都是CONDITION的。

那么unlinkCancelledWaiters做了什么?里面的流程就不看了,就是一些指针遍历并判断状态的操作,总结一下就是:从头到尾遍历每一个Node,遇到Node的waitStatus不是CONDITION的就从队列中踢掉,该节点的前后节点相连。

接着第8行的代码前面说过了,new出来了一个Node,存储了当前线程,waitStatus是CONDITION,接着第9行~第13行的操作很好理解:

  1. 如果lastWaiter是null,说明FIFO队列中没有任何Node,firstWaiter=Node
  2. 如果lastWaiter不是null,说明FIFO队列中有Node,原lastWaiter的next指向Node
  3. 无论如何,新加入的Node编程lastWaiter,即新加入的Node一定是在最后面

用一张图表示一下构建的数据结构就是:

对比学习,我们总结一下Condition构建出来的队列和AbstractQueuedSynchronizer构建出来的队列的差别,主要体现在2点上:

  1. AbstractQueuedSynchronizer构建出来的队列,头节点是一个没有Thread的空节点,其标识作用,而Condition构建出来的队列,头节点就是真正等待的节点
  2. AbstractQueuedSynchronizer构建出来的队列,节点之间有next与pred相互标识该节点的前一个节点与后一个节点的地址,而Condition构建出来的队列,只使用了nextWaiter标识下一个等待节点的地址

整个过程中,我们看到没有使用任何CAS操作,firstWaiter和lastWaiter也没有用volatile修饰,其实原因很简单:要await()必然要先lock(),既然lock()了就表示没有竞争,没有竞争自然也没必要使用volatile+CAS的机制去保证什么。

Condition的await()方法实现原理—-线程等待

前面我们看了Condition构建等待队列的过程,接下来我们看一下等待的过程,await()方法的代码比较短,再贴一下:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
public final void await() throws InterruptedException {
    if (Thread.interrupted())
        throw new InterruptedException();
    Node node = addConditionWaiter();
    int savedState = fullyRelease(node);
    int interruptMode = 0;
    while (!isOnSyncQueue(node)) {
        LockSupport.park(this);
        if ((interruptMode = checkInterruptWhileWaiting(node)) != 0)
            break;
    }
    if (acquireQueued(node, savedState) && interruptMode != THROW_IE)
        interruptMode = REINTERRUPT;
    if (node.nextWaiter != null) // clean up if cancelled
        unlinkCancelledWaiters();
    if (interruptMode != 0)
        reportInterruptAfterWait(interruptMode);
}

构建完毕队列之后,执行第5行的fullyRelease方法,顾名思义:fullyRelease方法的作用是完全释放Node的状态。方法实现为:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
final int fullyRelease(Node node) {
    boolean failed = true;
    try {
        int savedState = getState();
        if (release(savedState)) {
            failed = false;
            return savedState;
        } else {
            throw new IllegalMonitorStateException();
        }
    } finally {
        if (failed)
            node.waitStatus = Node.CANCELLED;
    }
}

这里第4行获取state,第5行release的时候将整个state传过去,理由是某线程可能多次调用了lock()方法,比如调用了10次lock,那么此线程就将state加到了10,所以这里要将10传过去,将状态全部释放,这样后面的线程才能重新从state=0开始竞争锁,这也是方法被命名为fullyRelease的原因,因为要完全释放锁,释放锁之后,如果有竞争锁的线程,那么就唤醒第一个,这都是release方法的逻辑了,前面的文章详细讲解过。

接着看await()方法的第7行判断”while(!isOnSyncQueue(node))”:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
final boolean isOnSyncQueue(Node node) {
    if (node.waitStatus == Node.CONDITION || node.prev == null)
        return false;
    if (node.next != null) // If has successor, it must be on queue
        return true;
    /*
     * node.prev can be non-null, but not yet on queue because
     * the CAS to place it on queue can fail. So we have to
     * traverse from tail to make sure it actually made it.  It
     * will always be near the tail in calls to this method, and
     * unless the CAS failed (which is unlikely), it will be
     * there, so we hardly ever traverse much.
     */
    return findNodeFromTail(node);
}

注意这里的判断是Node是否在AbstractQueuedSynchronizer构建的队列中而不是Node是否在Condition构建的队列中,如果Node不在AbstractQueuedSynchronizer构建的队列中,那么调用LockSupport的park方法阻塞。

至此调用await()方法的线程构建Condition等待队列–释放锁–等待的过程已经全部分析完毕。 

Condition的signal()实现原理

上面的代码分析了构建Condition等待队列–释放锁–等待的过程,接着看一下signal()方法通知是如何实现的:

1
2
3
4
5
6
7
public final void signal() {
    if (!isHeldExclusively())
        throw new IllegalMonitorStateException();
    Node first = firstWaiter;
    if (first != null)
        doSignal(first);
}

首先从第2行的代码我们看到,要能signal(),当前线程必须持有独占锁,否则抛出异常IllegalMonitorStateException。

那么真正操作的时候,获取第一个waiter,如果有waiter,调用doSignal方法:

1
2
3
4
5
6
7
8
private void doSignal(Node first) {
    do {
        if ( (firstWaiter = first.nextWaiter) == null)
            lastWaiter = null;
        first.nextWaiter = null;
    } while (!transferForSignal(first) &&
             (first = firstWaiter) != null);
}

第3行~第5行的代码很好理解:

  1. 重新设置firstWaiter,指向第一个waiter的nextWaiter
  2. 如果第一个waiter的nextWaiter为null,说明当前队列中只有一个waiter,lastWaiter置空
  3. 因为firstWaiter是要被signal的,因此它没什么用了,nextWaiter置空

接着执行第6行和第7行的代码,这里重点就是第6行的transferForSignal方法:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
final boolean transferForSignal(Node node) {
    /*
     * If cannot change waitStatus, the node has been cancelled.
     */
    if (!compareAndSetWaitStatus(node, Node.CONDITION, 0))
        return false;
 
    /*
     * Splice onto queue and try to set waitStatus of predecessor to
     * indicate that thread is (probably) waiting. If cancelled or
     * attempt to set waitStatus fails, wake up to resync (in which
     * case the waitStatus can be transiently and harmlessly wrong).
     */
    Node p = enq(node);
    int ws = p.waitStatus;
    if (ws > 0 || !compareAndSetWaitStatus(p, ws, Node.SIGNAL))
        LockSupport.unpark(node.thread);
    return true;
}

方法本意是将一个节点从Condition队列转换为AbstractQueuedSynchronizer队列,总结一下方法的实现:

  1. 尝试将Node的waitStatus从CONDITION置为0,这一步失败直接返回false
  2. 当前节点进入调用enq方法进入AbstractQueuedSynchronizer队列
  3. 当前节点通过CAS机制将waitStatus置为SIGNAL

最后上面的步骤全部成功,返回true,返回true唤醒等待节点成功。从唤醒的代码我们可以得出一个重要结论:某个await()的节点被唤醒之后并不意味着它后面的代码会立即执行,它会被加入到AbstractQueuedSynchronizer队列的尾部,只有前面等待的节点获取锁全部完毕才能轮到它。

代码分析到这里,我想类似的signalAll方法也没有必要再分析了,显然signalAll方法的作用就是将所有Condition队列中等待的节点逐一队列中从移除,由CONDITION状态变为SIGNAL状态并加入AbstractQueuedSynchronizer队列的尾部。

代码示例

可能大家看了我分析半天代码会有点迷糊,这里最后我贴一段我用于验证上面Condition结论的示例代码,首先建立一个Thread,我将之命名为ConditionThread:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
/**
 * @author 五月的仓颉http://www.cnblogs.com/xrq730/p/7067904.html
 */
public class ConditionThread implements Runnable {
 
    private Lock lock;
     
    private Condition condition;
     
    public ConditionThread(Lock lock, Condition condition) {
        this.lock = lock;
        this.condition = condition;
    }
     
    @Override
    public void run() {
         
        if ("线程0".equals(JdkUtil.getThreadName())) {
            thread0Process();
        } else if ("线程1".equals(JdkUtil.getThreadName())) {
            thread1Process();
        } else if ("线程2".equals(JdkUtil.getThreadName())) {
            thread2Process();
        }
         
    }
     
    private void thread0Process() {
        try {
            lock.lock();
            System.out.println("线程0休息5秒");
            JdkUtil.sleep(5000);
            condition.signal();
            System.out.println("线程0唤醒等待线程");
        } finally {
            lock.unlock();
        }
    }
     
    private void thread1Process() {
        try {
            lock.lock();
            System.out.println("线程1阻塞");
            condition.await();
            System.out.println("线程1被唤醒");
        } catch (InterruptedException e) {
             
        } finally {
            lock.unlock();
        }
    }
     
    private void thread2Process() {
        try {
            System.out.println("线程2想要获取锁");
            lock.lock();
            System.out.println("线程2获取锁成功");
        } finally {
            lock.unlock();
        }
    }
     
}

这个类里面的方法就不解释了,反正就三个方法片段,根据线程名判断,每个线层执行的是其中的一个代码片段。写一段测试代码:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
/**
 * @author 五月的仓颉http://www.cnblogs.com/xrq730/p/7067904.html
 */
@Test
public void testCondition() throws Exception {
    Lock lock = new ReentrantLock();
    Condition condition = lock.newCondition();
         
    // 线程0的作用是signal
    Runnable runnable0 = new ConditionThread(lock, condition);
    Thread thread0 = new Thread(runnable0);
    thread0.setName("线程0");
    // 线程1的作用是await
    Runnable runnable1 = new ConditionThread(lock, condition);
    Thread thread1 = new Thread(runnable1);
    thread1.setName("线程1");
    // 线程2的作用是lock
    Runnable runnable2 = new ConditionThread(lock, condition);
    Thread thread2 = new Thread(runnable2);
    thread2.setName("线程2");
         
    thread1.start();
    Thread.sleep(1000);
    thread0.start();
    Thread.sleep(1000);
    thread2.start();
         
    thread1.join();
}

测试代码的意思是:

  1. 线程1先启动,获取锁,调用await()方法等待
  2. 线程0后启动,获取锁,休眠5秒准备signal()
  3. 线程2最后启动,获取锁,由于线程0未使用完毕锁,因此线程2排队,可以此时由于线程0还未signal(),因此线程1在线程0执行signal()后,在AbstractQueuedSynchronizer队列中的顺序是在线程2后面的

代码执行结果为:

1
2
3
4
5
6
1 线程1阻塞
2 线程0休息5
3 线程2想要获取锁
4 线程0唤醒等待线程
5 线程2获取锁成功
6 线程1被唤醒

符合我们的结论:signal()并不意味着被唤醒的线程立即执行。由于线程2先于线程0排队,因此看到第5行打印的内容,线程2先获取锁。

posted @ 2017-08-29 15:38  _1900  阅读(319)  评论(0编辑  收藏  举报