Java并发编程--AQS
概述
抽象队列同步器(AbstractQueuedSynchronizer,简称AQS)是用来构建锁或者其他同步组件的基础框架,它使用一个整型的volatile变量(命名为state)来维护同步状态,通过内置的FIFO队列来完成资源获取线程的排队工作。
volatile变量的读写和CAS是concurrent包得以实现的基础。CAS表示如果当前状态值等于预期值,则以原子方式将同步状态设置为给定的更新值,此操作具有volatile读和写的内存语义。AQS通过volatile的读/写和CAS所具有的volatile读和写的内存语义来实现线程之间的通信。
高层类 | Lock 同步器 阻塞队列 Executor 并发容器 |
基础类 | AQS 非阻塞数据结构 原子变量类 |
volatile变量的读/写 CAS |
concurrent包的实现结构如上图所示,AQS、非阻塞数据结构和原子变量类等基础类都是基于volatile变量的读/写和CAS实现,而像Lock、同步器、阻塞队列、Executor和并发容器等高层类又是基于基础类实现。
AQS的域和方法
域
1 private transient volatile Node head; //同步队列的head节点 2 private transient volatile Node tail; //同步队列的tail节点 3 private volatile int state; //同步状态
方法
AQS提供的可以修改同步状态的3个方法:
1 protected final int getState(); //获取同步状态 2 protected final void setState(int newState); //设置同步状态 3 protected final boolean compareAndSetState(int expect, int update); //CAS设置同步状态
AQS提供的模板方法,主要有以下三类:
1)独占式获取和释放同步状态
1 public final void acquire(int arg) //独占式获取同步状态,如果不成功会进入同步队列等待。 2 public final void acquireInterruptibly(int arg) //与acquire不同的是,能响应中断 3 public final boolean tryAcquireNanos(int arg, long nanosTimeout) //增加超时机制 4 5 public final boolean release(int arg) //独占式释放同步状态,该方法会调用重写的tryRelease(int arg)。
以上三种获取同步状态的方法都会调用自定义的tryAcquire(int arg)方法,acquire的获取失败时的入队等待机制、acquireInterruptibly的响应中断机制、tryAcquireNanos的超时机制等AQS已经实现好,这样开发人员只需要实现自己的获取同步状态机制,就可以大大降低了实现一个可靠自定义同步组件的门槛。
2)共享式获取和释放同步状态
1 public final void acquireShared(int arg) //共享式获取同步状态,如果不成功会进入同步队列等待。与独占式不同的是,同一时刻可以有多个线程获取到同步状态。 2 public final void acquireSharedInterruptibly(int arg) //可响应中断 3 public final boolean tryAcquireSharedNanos(int arg, long nanosTimeout) //超时机制 4 5 public final boolean releaseShared(int arg) //共享式释放同步状态,该方法会调用重写的tryReleaseShared(int arg)。
同样以上三种获取同步状态的方法会调用自定义的tryAcquireShared方法。
3)查询同步队列中的等待线程情
1 publicfinalCollection<Thread>getQueuedThreads() 2 publicfinalbooleanhasQueuedThreads()//返回包含可能正在等待获取的线程列表,需要遍历链表。返回的只是个估计值,且是无序的。这个方法的主要是为子类提供的监视同步队列措施而设计。 3 。。。
AQS提供的自定义方法
以上AQS的方法都为final方法,不能被子类重写,因为它们对于任何自定义同步器应该是不需要更改的,下面为AQS提供的可以重写的方法。开发者需要根据自定义同步组件的特点,重写以下方法。这些方法的实现在内部必须是线程安全的,通常应该很短并且不被阻塞。
1 protected boolean tryAcquire(int arg) //独占式获取同步状态,此方法应该查询是否允许它在独占模式下获取对象状态,如果允许,则获取它。返回值语义:true代表获取成功,false代表获取失败。 2 protected boolean tryRelease(int arg) //独占式释放同步状态 3 4 protected int tryAcquireShared(int arg) //共享式获取同步状态,返回值语义:负数代表获取失败、0代表获取成功但没有剩余资源、正数代表获取成功,还有剩余资源。 5 protected boolean tryReleaseShared(int arg) //共享式释放同步状态 6 7 protected boolean isHeldExclusively() //AQS是否被当前线程所独占
AQS的使用
怎么使用AQS实现自定义同步组件?
自定义同步组件实例一:独占锁(使用独占式的获取与释放)。
Mutex同步组件同一时刻只允许一个线程占用锁,不支持可重入。0表示未锁定状态,1表示锁定状态。
1 class Mutex implements Lock, java.io.Serializable { 2 3 //静态内部类,自定义同步器 4 private static class Sync extends AbstractQueuedSynchronizer { 5 6 // 释放处于占用状态(重写isHeldExclusively)Report whether in locked state 7 protected boolean isHeldExclusively() { 8 return getState() == 1; 9 } 10 11 // 独占式获取锁(重写tryAcquire) Acquire the lock if state is zero 12 public boolean tryAcquire(int acquires) { 13 assert acquires == 1; // Otherwise unused 14 if (compareAndSetState(0, 1)) { //CAS设置状态为1。 15 setExclusiveOwnerThread(Thread.currentThread()); 16 return true; 17 } 18 return false; 19 } 20 21 // 独占式释放锁(重写tryRelease) Release the lock by setting state to zero 22 protected boolean tryRelease(int releases) { 23 assert releases == 1; // Otherwise unused 24 if (getState() == 0) //获取状态 25 throw new IllegalMonitorStateException(); 26 setExclusiveOwnerThread(null); 27 setState(0); //设置状态为0 28 return true; 29 } 30 31 // Provide a Condition 32 //每个Condition都包含一个队列 33 Condition newCondition() { return new ConditionObject(); } 34 35 // Deserialize properly 36 private void readObject(ObjectInputStream s) throws IOException, ClassNotFoundException { 37 s.defaultReadObject(); 38 setState(0); // reset to unlocked state 39 } 40 } 41 42 // The sync object does all the hard work. We just forward to it. 43 private final Sync sync = new Sync(); 44 45 //仅需要将操作代理到sync 46 public void lock() { sync.acquire(1); } //调用AQS的模板方法, 47 public boolean tryLock() { return sync.tryAcquire(1); } 48 public void unlock() { sync.release(1); } 49 public Condition newCondition() { return sync.newCondition(); } 50 public boolean isLocked() { return sync.isHeldExclusively(); } 51 public boolean hasQueuedThreads() { return sync.hasQueuedThreads(); } 52 public void lockInterruptibly() throws InterruptedException { 53 sync.acquireInterruptibly(1); 54 } 55 public boolean tryLock(long timeout, TimeUnit unit) throws InterruptedException { 56 return sync.tryAcquireNanos(1, unit.toNanos(timeout)); 57 } 58 }
自定义同步组件实例二:锁存器(使用共享的获取与释放方法)
BooleanLatch可用在多个线程需要等待某个事件发生才能继续执行的情况中。初始状态state=0, 此时所有线程获取同步状态方法tryAcquireShared返回-1,即获取失败,入等待队列。直到有线程调用tryReleaseShared释放同步状态,被阻塞的状态才会进行执行。
1 class BooleanLatch { 2 3 private static class Sync extends AbstractQueuedSynchronizer { 4 boolean isSignalled() { return getState() != 0; } 5 6 protected int tryAcquireShared(int ignore) { 7 return isSignalled()? 1 : -1; 8 } 9 10 protected boolean tryReleaseShared(int ignore) { 11 setState(1); 12 return true; 13 } 14 } 15 16 private final Sync sync = new Sync(); 17 public boolean isSignalled() { return sync.isSignalled(); } 18 public void signal() { sync.releaseShared(1); } 19 public void await() throws InterruptedException { 20 sync.acquireSharedInterruptibly(1); 21 } 22 }
AQS的实现原理
同步队列
同步队列中的节点Node用来保存获取同步状态失败的线程引用、等待状态、以及前驱和后继节点。
AQS中两个域:head节点和tail节点,组成一个FIFO的双向队列。
private transient volatile Node head;
private transient volatile Node tail;
Node源码如下。
1 static final class Node { 2 /** Marker to indicate a node is waiting in shared mode */ 3 static final Node SHARED = new Node(); //共享方式 4 /** Marker to indicate a node is waiting in exclusive mode */ 5 static final Node EXCLUSIVE = null; //独占方式 6 7 /** waitStatus value to indicate thread has cancelled */ 8 static final int CANCELLED = 1; //waitStatus=1为取消状态 9 /** waitStatus value to indicate successor's thread needs unparking */ 10 static final int SIGNAL = -1; //后继节点的线程处于等待状态,需要被唤醒。 11 /** waitStatus value to indicate thread is waiting on condition */ 12 static final int CONDITION = -2; //当前线程在condition上等待 13 static final int PROPAGATE = -3; //表示下一次共享式同步状态获取将会无条件的被传播下去。 14 15 volatile int waitStatus; //等待状态,0-初始状态 16 volatile Node prev; //前驱节点 17 volatile Node next; //后继节点 18 volatile Thread thread; //获取同步的线程 19 Node nextWaiter; 20 21 final boolean isShared() { 22 return nextWaiter == SHARED; 23 } 24 //返回前驱节点 25 final Node predecessor() throws NullPointerException { 26 Node p = prev; 27 if (p == null) 28 throw new NullPointerException(); 29 else 30 return p; 31 } 32 33 Node() { // Used to establish initial head or SHARED marker 34 } 35 36 Node(Thread thread, Node mode) { // Used by addWaiter 37 this.nextWaiter = mode; 38 this.thread = thread; 39 } 40 41 Node(Thread thread, int waitStatus) { // Used by Condition 42 this.waitStatus = waitStatus; 43 this.thread = thread; 44 } 45 }
独占式获取同步状态
1)调用自定义的tryAcquire方法,该方法要保证线程线程安全的获取同步状态(如Mutex中的tryAcquire使用CAS更新保证原子性),如果获取成功则return。
2)如果获取失败,构造Node,并通过addWaiter方法将节点插入队列尾部。Node.EXCLUSIVE表示节点以独占方式等待。
3)acquireQueued方法中该节点自旋方式尝试获取同步状态。如果获取不到同步状态,则阻塞节点中的线程,被阻塞线程唤醒依靠前驱节点的出队或阻塞线程被终端来实现。
1 //该方法对中断不敏感,也就是由于线程获取同步状态失败后进入同步队列中,后续对线程进行中断操作时,线程不会从同步队列中移出. 2 public final void acquire(int arg) { 3 if (!tryAcquire(arg) && 4 acquireQueued(addWaiter(Node.EXCLUSIVE), arg)) 5 selfInterrupt(); 6 } 7 8 //将节点添加到队尾 9 private Node addWaiter(Node mode) { 10 Node node = new Node(Thread.currentThread(), mode); 11 // Try the fast path of enq; backup to full enq on failure 12 //快速尝试入队,如果失败则需要调用enq(node)方法入队,这样做有什么好处?有一定概率减少一次方法调用 13 //compareAndSetTail保证Node入队是线程安全的 14 Node pred = tail; 15 if (pred != null) { 16 node.prev = pred; 17 if (compareAndSetTail(pred, node)) { 18 pred.next = node; 19 return node; 20 } 21 } 22 enq(node); 23 return node; 24 } 25 26 //初始化或自旋CAS直到入队成功 27 private Node enq(final Node node) { 28 for (;;) { 29 Node t = tail; 30 if (t == null) { // Must initialize 31 if (compareAndSetHead(new Node())) 32 tail = head; 33 } else { 34 node.prev = t; 35 if (compareAndSetTail(t, node)) { 36 t.next = node; 37 return t; 38 } 39 } 40 } 41 }
入同步队列之后怎么获取同步状态?阻塞机制是怎样的?
1 //自旋方式尝试获取同步状态 2 final boolean acquireQueued(final Node node, int arg) { 3 boolean failed = true; 4 try { 5 boolean interrupted = false; 6 for (;;) { 7 final Node p = node.predecessor(); //获取当前节点的前驱节点 8 if (p == head && tryAcquire(arg)) { //如果前驱节点是head节点则尝试获取同步状态 9 setHead(node); 10 p.next = null; // help GC 11 failed = false; 12 return interrupted; 13 } 14 if (shouldParkAfterFailedAcquire(p, node) && 15 parkAndCheckInterrupt()) //判断当前线程是否应该被阻塞,如果是则阻塞直到被唤醒继续循环,如果不是则再次尝试获取同步状态。 16 interrupted = true; 17 } 18 } finally { 19 if (failed) 20 cancelAcquire(node); 21 } 22 } 23 24 //判断当前线程是否应该被阻塞,如果线程应该被阻塞则返回true。检查和更新获取同步状态失败Node的前驱节点的waitStatus。 25 //其中pred为node的前驱节点 26 private static boolean shouldParkAfterFailedAcquire(Node pred, Node node) { 27 int ws = pred.waitStatus; 28 if (ws == Node.SIGNAL) 29 /* 30 * This node has already set status asking a release 31 * to signal it, so it can safely park. 32 */ 33 //前驱节点已经设置为SIGNAL状态,在前驱节点的线程释放同步状态会唤醒当前Node的线程。 34 return true; 35 if (ws > 0) { 36 /* 37 * Predecessor was cancelled. Skip over predecessors and 38 * indicate retry. 39 */ 40 //前驱节点是cancelled状态,跳过被取消的Node,直到向前找到waitStatus > 0的Node作为当前节点的前驱,然后重试获取同步状态。 41 do { 42 node.prev = pred = pred.prev; 43 } while (pred.waitStatus > 0); 44 pred.next = node; 45 } else { 46 /* waitStatus = 0是初始状态。 47 * waitStatus must be 0 or PROPAGATE. Indicate that we 48 * need a signal, but don't park yet. Caller will need to 49 * retry to make sure it cannot acquire before parking. 50 */ 51 //将前驱节点的等待状态改为SIGNAL。 52 compareAndSetWaitStatus(pred, ws, Node.SIGNAL); 53 } 54 return false; 55 } 56 57 //阻塞线程 58 private final boolean parkAndCheckInterrupt() { 59 LockSupport.park(this); //使用LockSupport阻塞当前线程 60 return Thread.interrupted(); 61 }
在获取同步状态时,同步器维护一个同步队列,获取状态失败的线程都会被加入到队列中并在队列中进行自旋;移出队列(或停止自旋)的条件是前驱节点为头节点且成功获取了同步状态。
独占式释放同步状态
在释放同步状态时,同步器调用tryRelease(int arg)方法释放同步状态,然后唤醒头节点的后继节点。
1 public final boolean release(int arg) { 2 if (tryRelease(arg)) { 3 Node h = head; 4 if (h != null && h.waitStatus != 0) 5 unparkSuccessor(h); //唤醒后继节点线程 6 return true; 7 } 8 return false; 9 } 10 11 //唤醒后继节点 12 private void unparkSuccessor(Node node) { 13 /* 14 * If status is negative (i.e., possibly needing signal) try 15 * to clear in anticipation of signalling. It is OK if this 16 * fails or if status is changed by waiting thread. 17 */ 18 //将当前节点的waitStatus改为0-原始状态,目的是什么? 19 int ws = node.waitStatus; 20 if (ws < 0) 21 compareAndSetWaitStatus(node, ws, 0); 22 23 /* 24 * Thread to unpark is held in successor, which is normally 25 * just the next node. But if cancelled or apparently null, 26 * traverse backwards from tail to find the actual 27 * non-cancelled successor. 28 */ 29 //如果后继节点为null或被取消,则从tail向前找到正常的后继节点 30 Node s = node.next; 31 if (s == null || s.waitStatus > 0) { 32 s = null; 33 for (Node t = tail; t != null && t != node; t = t.prev) 34 if (t.waitStatus <= 0) 35 s = t; 36 } 37 if (s != null) 38 LockSupport.unpark(s.thread); //唤醒后继节点 39 }
共享式获取同步状态
1)首先调用自定义方法tryAcquireShared尝试获取同步状态,至少调用一次tryAcquireShared方法,如果返回值>=0,则获取成功,return;否则执行步骤2),
2)当获取失败时,为当前线程以共享方式创建Node并插入同步队列。
3)入队后,以自旋方式尝试获取同步状态,如果前驱节点为head节点,则尝试获取同步状态,获取失败,则阻塞线程。
1 //共享式获取同步状态,忽略异常。 2 //注意:实现自定义方法tryAcquireShared时,要遵循AQS定义的返回值语义,负数代表获取失败、0代表获取成功但没有剩余资源、正数代表获取成功,还有剩余资源。 3 public final void acquireShared(int arg) { 4 if (tryAcquireShared(arg) < 0) // 5 doAcquireShared(arg); 6 } 7 8 private void doAcquireShared(int arg) { 9 final Node node = addWaiter(Node.SHARED); //为当前线程以共享方式创建Node并插入同步队列尾部。 10 boolean failed = true; 11 try { 12 boolean interrupted = false; 13 for (;;) { 14 final Node p = node.predecessor(); 15 if (p == head) { //如果前驱节点为head节点,则尝试获取同步状态 16 int r = tryAcquireShared(arg); 17 if (r >= 0) { //获取成功 18 setHeadAndPropagate(node, r); //设置node为head节点,还有剩余资源则继续唤醒后继的node 19 p.next = null; // help GC 20 if (interrupted) //如果等待过程中被中断过,则中断当前线程 21 selfInterrupt(); //Thread.currentThread().interrupt(); 22 failed = false; 23 return; 24 } 25 } 26 //shouldParkAfterFailedAcquire方法判断当前线程是否应该被阻塞,如果是则调用parkAndCheckInterrupt阻塞当前线程 27 if (shouldParkAfterFailedAcquire(p, node) && 28 parkAndCheckInterrupt()) 29 interrupted = true; 30 } 31 } finally { 32 if (failed) 33 cancelAcquire(node); 34 } 35 } 36 37 //获取成功后,设置node为head节点, 38 private void setHeadAndPropagate(Node node, int propagate) { 39 Node h = head; // Record old head for check below 40 setHead(node); //设置node为head节点,因此node出队 41 //如果还有剩余资源,尝试唤醒node节点的后继节点 42 /* 43 * Try to signal next queued node if: 44 * Propagation was indicated by caller, 45 * or was recorded (as h.waitStatus) by a previous operation 46 * (note: this uses sign-check of waitStatus because 47 * PROPAGATE status may transition to SIGNAL.) 48 * and 49 * The next node is waiting in shared mode, 50 * or we don't know, because it appears null 51 * 52 * The conservatism in both of these checks may cause 53 * unnecessary wake-ups, but only when there are multiple 54 * racing acquires/releases, so most need signals now or soon 55 * anyway. 56 */ 57 if (propagate > 0 || h == null || h.waitStatus < 0) { 58 Node s = node.next; 59 if (s == null || s.isShared()) 60 doReleaseShared(); 61 } 62 } 63 64 //设置node为head节点,因此node出队 65 private void setHead(Node node) { 66 head = node; 67 node.thread = null; 68 node.prev = null; 69 }
共享式释放同步状态
1 public final boolean releaseShared(int arg) { 2 if (tryReleaseShared(arg)) { 3 doReleaseShared(); 4 return true; 5 } 6 return false; 7 } 8 9 //唤醒后继节点线程并确保被传播 10 private void doReleaseShared() { 11 /* 12 * Ensure that a release propagates, even if there are other 13 * in-progress acquires/releases. This proceeds in the usual 14 * way of trying to unparkSuccessor of head if it needs 15 * signal. But if it does not, status is set to PROPAGATE to 16 * ensure that upon release, propagation continues. 17 * Additionally, we must loop in case a new node is added 18 * while we are doing this. Also, unlike other uses of 19 * unparkSuccessor, we need to know if CAS to reset status 20 * fails, if so rechecking. 21 */ 22 for (;;) { 23 Node h = head; 24 if (h != null && h != tail) { 25 int ws = h.waitStatus; 26 if (ws == Node.SIGNAL) { 27 if (!compareAndSetWaitStatus(h, Node.SIGNAL, 0)) 28 continue; // loop to recheck cases 29 unparkSuccessor(h); //唤醒后继节点 30 } 31 else if (ws == 0 && 32 !compareAndSetWaitStatus(h, 0, Node.PROPAGATE)) 33 continue; // loop on failed CAS 34 } 35 if (h == head) // loop if head changed 36 break; 37 } 38 }
独占式可中断获取
可响应中断获取与普通获取的区别:当线程被中断时,会立即返回,并抛出InterruptedException,执行cancelAcquire方法取消获取同步状态;而普通获取只是将中断标志位置为true,但线程依旧会阻塞在等待队列中。
1 public final void acquireInterruptibly(int arg) 2 throws InterruptedException { 3 if (Thread.interrupted()) 4 throw new InterruptedException(); //抛出中断异常 5 if (!tryAcquire(arg)) 6 doAcquireInterruptibly(arg); 7 } 8 9 private void doAcquireInterruptibly(int arg) 10 throws InterruptedException { 11 final Node node = addWaiter(Node.EXCLUSIVE); 12 boolean failed = true; 13 try { 14 for (;;) { 15 final Node p = node.predecessor(); 16 if (p == head && tryAcquire(arg)) { 17 setHead(node); 18 p.next = null; // help GC 19 failed = false; 20 return; 21 } 22 if (shouldParkAfterFailedAcquire(p, node) && 23 parkAndCheckInterrupt()) 24 throw new InterruptedException(); //唯一的区别,抛出中断异常。而普通的获取操作,只是将中断标志位置为true。 25 } 26 } finally { 27 if (failed) 28 cancelAcquire(node); 29 } 30 }
如何取消线程?
1 private void cancelAcquire(Node node) { 2 // Ignore if node doesn't exist 3 if (node == null) 4 return; 5 6 node.thread = null; //将线程置为null 7 8 // Skip cancelled predecessors 9 Node pred = node.prev; 10 while (pred.waitStatus > 0) //即waitStatus=1,为cancelled状态。跳过状态为取消的前驱节点 11 node.prev = pred = pred.prev; 12 13 // predNext is the apparent node to unsplice. CASes below will 14 // fail if not, in which case, we lost race vs another cancel 15 // or signal, so no further action is necessary. 16 Node predNext = pred.next; 17 18 // Can use unconditional write instead of CAS here. 19 // After this atomic step, other Nodes can skip past us. 20 // Before, we are free of interference from other threads. 21 node.waitStatus = Node.CANCELLED; //当前节点置为cancelled状态 22 23 // If we are the tail, remove ourselves. 24 if (node == tail && compareAndSetTail(node, pred)) { //如果当前节点为tail节点,删除该节点 25 compareAndSetNext(pred, predNext, null); 26 } else { 27 // If successor needs signal, try to set pred's next-link 28 // so it will get one. Otherwise wake it up to propagate. 29 // 30 int ws; 31 if (pred != head && 32 ((ws = pred.waitStatus) == Node.SIGNAL || 33 (ws <= 0 && compareAndSetWaitStatus(pred, ws, Node.SIGNAL))) && 34 pred.thread != null) { 35 //如果前驱节点waitStatus为SIGNAL或者CAS更新为SIGNAL成功,则pred释放同步状态时会通知后继节点 36 //并且pred.thread不为null,则cas将pred的后继节点置为node.next, 37 Node next = node.next; 38 if (next != null && next.waitStatus <= 0) 39 compareAndSetNext(pred, predNext, next); 40 } else { 41 unparkSuccessor(node); //唤醒后继节点 42 } 43 44 node.next = node; // help GC next指向自己,node从等待队列中移除。 45 } 46 }
独占式可超时获取
在响应中断的基础上增加了超时机制
1 public final boolean tryAcquireNanos(int arg, long nanosTimeout) 2 throws InterruptedException { 3 if (Thread.interrupted()) 4 throw new InterruptedException(); 5 return tryAcquire(arg) || doAcquireNanos(arg, nanosTimeout); 6 } 7 8 private boolean doAcquireNanos(int arg, long nanosTimeout) 9 throws InterruptedException { 10 long lastTime = System.nanoTime(); 11 final Node node = addWaiter(Node.EXCLUSIVE); 12 boolean failed = true; 13 try { 14 for (;;) { 15 final Node p = node.predecessor(); 16 if (p == head && tryAcquire(arg)) { 17 setHead(node); 18 p.next = null; // help GC 19 failed = false; 20 return true; 21 } 22 if (nanosTimeout <= 0) //判断是否超时,nanosTimeout<=0表示超时,如果超时则return false. 23 return false; 24 if (shouldParkAfterFailedAcquire(p, node) && 25 nanosTimeout > spinForTimeoutThreshold) 26 LockSupport.parkNanos(this, nanosTimeout); //如果还剩余的时间nanosTimeout>阈值(spinForTimeoutThreshold=1000纳秒),则阻塞当前线程nanosTimeout纳秒。当nanosTimeout<1000纳秒时,则不阻塞当前线程,而是进入快速的自旋过程。原因在于,非常短的超时等待无法做到十分精确,如果这时再进行超时等待,相反会让nanosTimeout的超时从整体上表现得反而不精确。因此,在超时非常短的场景下,同步器会进入无条件的快速自旋。 27 long now = System.nanoTime(); //当前时间 28 //lastTime表示上次唤醒时间,now - lastTime表示已经睡眠的时间 29 nanosTimeout -= now - lastTime; //nanosTimeout表示还剩余的时间,nanosTimeout>0表示表示超时时间未到 30 lastTime = now; 31 if (Thread.interrupted()) 32 throw new InterruptedException(); //抛异常 33 } 34 } finally { 35 if (failed) 36 cancelAcquire(node); 37 } 38 }
总结,三种方式获取同步状态方式的对比,主要区别在于获取同步状态失败时的处理逻辑:
acquire方法直接阻塞线程,不响应中断,只是将中断标记置为true,但线程依旧会阻塞在等待队列中。
acquireInterruptibly方法直接阻塞线程,响应中断,当线程被中断时,会立即返回,并抛出InterruptedException。
tryAcquireNanos方法将线程阻塞nanosTimeout秒,如何超时还未获取到同步状态,则返回。同时支持响应中断。
1 public final void acquire(int arg) //独占式获取同步状态,如果不成功会进入同步队列等待。 2 public final void acquireInterruptibly(int arg) //与acquire不同的是,能响应中断 3 public final boolean tryAcquireNanos(int arg, long nanosTimeout) //增加超时机制
LockSupport
在阻塞和唤醒线程时,使用了LockSupport类。LockSupport提供了一系列阻塞和唤醒线程的公共方法。底层使用unsafe提供的方法实现。
1 void park() //阻塞当前线程,只有调用unpark或中断才能从park方法中返回 2 void parkNanos(long nanos) //超时阻塞当前线程,超时则返回 3 void parkUntil(long deadline) //截止时间阻塞当前线程,直到deadline 4 5 void unpark(Thread thread) //唤醒处于阻塞状态的线程 6 7 //jdk1.6中增加了以下方法,blocker表示当前线程要等待的对象。线程dump信息比使用park方法要多,方便问题排查和监控。 8 void park(Object blocker) 9 void parkNanos(Object blocker, long nanos) 10 void parkUntil(Object blocker, long deadline)
参考资料:
《Java并发编程的艺术》
《Java并发之AQS详解》http://www.cnblogs.com/waterystone/p/4920797.html