什么是AQS?
- 全称: Abstract Queued Synchronizer: 抽象队列同步器
- 是 java.util.concurrent.locks包下的一个抽象类
- 其编写者: Doug Lea (并发大佬, 整个j.u.c包都是他写的)
- 是 j.u.c 包的基础组件(核心)
我们先来读一下该类的英文说明注释:
/**
* Provides a framework for implementing blocking locks and related synchronizers
* (semaphores, events, etc) that rely on first-in-first-out (FIFO) wait queues.
* 提供了一个实现阻塞式锁 及相关的 依赖于先进先出等待队列的 同步器(如信号量, 事件...) 的框架。
*
* This class is designed to be a useful basis for most kinds of synchronizers that rely on
* a single atomic {@code int} value to represent state.
* 该类被设计为 所有依赖于单个原子性值 来代表其状态的同步器 的基石。
*
* Subclasses must define the protected methods that change this state, and which
* define what that state means in terms of this object being acquired or released.
* 子类必须 通过定义 protected 方法 来改变 此状态, 并且 定义 定义该状态意味着 当前对象 被 获取了
* 或 被释放了。
*
* Given these, the other methods in this class carryout all queuing and blocking mechanics.
* 基于上述已知, 此类中的其他方法 实行了 所有的 队列等待 和 阻塞 机制。
*
* Subclasses can maintain other state fields, but only the atomically updated {@code int}
* value manipulated using methods {@link #getState}, {@link #setState} and {@link
* #compareAndSetState} is tracked with respect to synchronization.
* 子类可以维持其他状态域, 但是只有 被 getState, setState 和 compareAndSetState(CAS 比较并设置
* 转态) 方法 管理的 且 被原子性 更新地 值 才会被同步化关联
*
* <p>Subclasses should be defined as non-public internal helper classes that are used to
* implement the synchronization properties of their enclosing class.
* 子类 应当被定义为 被用于实现它们的 封闭类 的 同步特性 的 非公开帮助类
*
* Class {@code AbstractQueuedSynchronizer} does not implement any synchronization
* interface.
* AQS 并没有实现任何同步接口。
*
* Instead it defines methods such as {@link #acquireInterruptibly} that can be invoked as
* appropriate by concrete locks and related synchronizers to implement their public
* methods.
* 它定义了诸如 acquireInterruptibly 等 可以 被 具体的锁 和 相关的同步器 请求 来实现 它们的公有方
* 法
*
* <p>This class supports either or both a default <em>exclusive</em> mode and a
* <em>shared</em> mode.
* 该类支持 仅排除(排他)模式, 仅共享模式 以及 两种模式并存。
*
*
* When acquired in exclusive mode, attempted acquires by other threads cannot succeed.
* 当一个线程 在 排他模式中 获取了锁, 其他线程的 获取(该锁) 就不能成功
*
* Shared mode acquires by multiple threads may (but need not) succeed.
* 共享模式下, 多个线程获取 一个锁 可能成功(但是 通常 不需要)
*
* This class does not "understand" these differences except in the mechanical
* sense that when a shared mode acquire succeeds, the next waiting thread (if
* one exists) must also determine whether it can acquire as well.
* 该类只有在 一个线程在共享模式下获取到了锁 才会 去判断 下一个线程是否能够获取该锁
*
* Threads waiting in the different modes share the same FIFO queue.
* 在不同模式下等待的线程 共享 先进先出队列
*
* Usually, implementation subclasses support only one of these modes, but both can come
* into play for example in a {@link ReadWriteLock}.
* 通常, 实现该抽象类的子类 只支持其中一种模式(共享, 排他), 但是 在 读写锁中 两种模式可以并存。
*
* Subclasses that support only exclusive or only shared modes need not define the methods
* supporting the unused mode.
* 仅支持排他 或 仅支持共享 模式的子类 不需要定义 未被使用的 模式的方法
*
* <p>This class defines a nested {@link ConditionObject} class that
* can be used as a {@link Condition} implementation by subclasses
* supporting exclusive mode for which method {@link #isHeldExclusively} reports whether
* synchronization is exclusively held with respect to the current thread, method {@link
* #release} invoked with the current {@link #getState} value fully releases
* this object, and {@link #acquire}, given this saved state value,
* eventually restores this object to its previous acquired state.
* 该类定义了一个可被用作 支持排他模式的子类的状态应用 的 嵌套类。isHeldExclusively 方法 报告了
* 同步对当前线程是否维持排他模式。release 方法 通过 getState请求当前 线程状态值, 并且充分释放该对
* 象。acquire方法 根据已保存的状态值 最终 将该对象恢复到 它 上次被 获取之前的状态。
* 衍生: semaphore(信号量) 中的 acquire, release
*
* No {@code AbstractQueuedSynchronizer} method otherwise creates such a
* condition, so if this constraint cannot be met, do not use it.
* 没有抽象队列同步器 方法 就只好 创建 创建一个 状态, 如果这些约束不满足, 就不使用它
*
* The behavior of {@link ConditionObject} depends of course on the
* semantics of its synchronizer implementation.
* 状态对象 的行为 毫无疑问 取决于 它的 同步器应用 的 语义
*
* <p>This class provides inspection, instrumentation, and monitoring
* methods for the internal queue, as well as similar methods for
* condition objects.
* 该类对内部队列 以及相似的 状态对象 提供了 检查, 仪器化 和 监控 的方法
*
* These can be exported as desired into classes using an {@code AbstractQueuedSynchronizer}
* for their synchronization mechanics.
* 这些方法可以根据需求 通过 AQS 来 导入 到 类中 以实现同步机制
*
* <p>Serialization of this class stores only the underlying atomic
* integer maintaining state, so deserialized objects have empty
* thread queues.
* 此类的序列化仅存储原子性的 整形 维持 状态, 所以
*
* Typical subclasses requiring serializability will define a {@code readObject} method that
* restores this to a known initial state upon deserialization.
* 典型的需要序列化属性的 子类会通过 定义 一个 读取 对象的 方法 来将它 恢复为 取决于 反序列化的 已
* 知的的 初始状态
*
* <h3>Usage</h3>
*
* <p>To use this class as the basis of a synchronizer, redefine the
* following methods, as applicable, by inspecting and/or modifying
* the synchronization state using {@link #getState}, {@link
* #setState} and/or {@link #compareAndSetState}:
* 要使用该类作为同步器的基础, 需要通过 使用 getState, setState, CAS 方法 来 检查 或 修改 重定义以下方法为 可执行的 方法。
*
* <ul>
* <li> {@link #tryAcquire}
* <li> {@link #tryRelease}
* <li> {@link #tryAcquireShared}
* <li> {@link #tryReleaseShared}
* <li> {@link #isHeldExclusively}
* </ul>
*
* Each of these methods by default throws {@link UnsupportedOperationException}.
* 以上每个方法都会默认抛出 不支持的操作选项 异常
*
* Implementations of these methods must be internally thread-safe, and should in general be
* short and not block.
* 这些方法 的实现类 必须是 内部 线程 安全的m 并且普遍应当 量小 且 非阻塞
*
* Defining these methods is the <em>only</em> supported means of using this class.
* 以上这些方法 的 定义 仅用于 支持 该类
*
* All other methods are declared {@code final} because they cannot be independently varied.
* 其他的方法都以 final 声明, 因为它们 不能是 无关变化的
*
* <p>You may also find the inherited methods from {@link AbstractOwnableSynchronizer}
* useful to keep track of the thread owning an exclusive synchronizer.
* 你可能也会发现 从 AQS 继承来的 方法 在追踪 一个 拥有 排他模式 同步器的 线程时 很 有用
*
* You are encouraged to use them -- this enables monitoring and diagnostic tools to assist
* users in determining which threads hold locks.
* 使用他们 使得 负责监控 和 诊断的工具 能更好地 帮助用户确定 哪个线程占用着锁
*
* <p>Even though this class is based on an internal FIFO queue, it does not automatically
* enforce FIFO acquisition policies.
* 尽管该类 基于 一个 内部的 先进先出 队列, 他并没有 自动 强迫 使用 先进先出 的获取策略
*
* The core of exclusive synchronization takes the form:
* 排他同步的核心为以下形式:
* <pre>
* Acquire:
* // 只要没有获取到参数
* while (!tryAcquire(arg)) {
* // 如果线程没有入队, 就让他入队
* <em>enqueue thread if it is not already queued</em>;
* // 也可能是当前线程被阻塞了
* <em>possibly block current thread</em>;
* }
*
* Release:
* // 如果释放参数成功
* if (tryRelease(arg))
* // 就解除队列头部的线程
* <em>unblock the first queued thread</em>;
* </pre>
*
* (Shared mode is similar but may involve cascading signals.)
* 共享模式也相似, 但是可能涉及 层级信号
*
* <p id="barging">Because checks in acquire are invoked before
* enqueuing, a newly acquiring thread may <em>barge</em> ahead of
* others that are blocked and queued.
* 因为在线程入队前需要请求检查, 一个新 获取的线程 可能 会 冲撞 到其他 阻塞 且处于 队列中的 线程(即
* 插队) [非公平锁]
*
* However, you can, if desired, define {@code tryAcquire} and/or {@code tryAcquireShared}
* to disable barging by internally invoking one or more of the inspection
* methods, thereby providing a <em>fair</em> FIFO acquisition order.
* 有需求的话 也可以 在 tryAcquire 和/或 tryAcquireShared中 通过内部请求来 禁用 冲撞参数, 从而
* 提供一个公平的 先进先出 获取顺序 [公平锁]
*
* In particular, most fair synchronizers can define {@code tryAcquire}
* to return {@code false} if {@link #hasQueuedPredecessors} (a method
* specifically designed to be used by fair synchronizers) returns
* {@code true}. Other variations are possible.
* 事实上, 更公平的同步器可以定义 tryAcquire 方法 返回 false 如果队列中有前者(这是一个被特定用于定
* 义公平同步器 的设计), 如果队列中没有前者就返回true
*
* <p>Throughput and scalability are generally highest for the default barging (also known
* as <em>greedy</em>,<em>renouncement</em>, and <em>convoy-avoidance</em>) strategy.
* 尽管默认使用 barging 机制 的 写入 和 可扩展性 (又称贪婪重申 避免护航) 是最高的。
*
* While this is not guaranteed to be fair or starvation-free, earlier
* queued threads are allowed to recontend before later queued
* threads, and each recontention has an unbiased chance to succeed
* against incoming threads.
* 尽管这并没有保证 公平 或者说 无饥饿性, 更早入队的线程 被允许 在 后入队的线程 之后 重申 使用权,
* 每次重申都有 公平的 成功机会。
*
* Also, while acquires do not "spin" in the usual sense, they may perform
* multiple invocations of {@code tryAcquire} interspersed with other
* computations before blocking.
* 尽管 请求 并不会自旋, 通常 它们 会执行许多次 tryAcquire 方法调用,在阻塞之前, 其间夹杂着其他的
* 计算
*
* This gives most of the benefits of spins when exclusive synchronization is only briefly
* held, without most of the liabilities when it isn't.
* 当处于排他同步模式时, 它具有了所有自旋的好处, 而当不处于排他模式时, 就不需要这些义务
*
* If so desired, you can augment this by preceding calls to acquire methods with "fast-
* path" checks, possibly prechecking {@link #hasContended} and/or {@link #hasQueuedThreads}
* to only do so if the synchronizer is likely not to be contended.
*
* <p>This class provides an efficient and scalable basis for
* synchronization in part by specializing its range of use to
* synchronizers that can rely on {@code int} state, acquire, and
* release parameters, and an internal FIFO wait queue. When this does
* not suffice, you can build synchronizers from a lower level using
* {@link java.util.concurrent.atomic atomic} classes, your own custom
* {@link java.util.Queue} classes, and {@link LockSupport} blocking
* support.
* 同步器 <-依赖于- state, acquire,release param, 内部等待队列
* 修改 同步器的使用范围 -> 扩展 和 效用
* 可使用原子类, 队列类 和 LockSupport(提供阻塞, park, unpark) 自己构造同步器
*
* <h3>Usage Examples</h3>
*
* <p>Here is a non-reentrant mutual exclusion lock class that uses the value zero to
* represent the unlocked state, and one to represent the locked state.
* 这是一个不可重入的互斥锁类, 0: 未获得锁 1: 已获得锁
*
* While a non-reentrant lock does not strictly require recording of the current owner
* thread, this class does so anyway to make usage easier to monitor.
* 尽管 一个 非可重入锁, 并不严格要求记录当前线程的所有者, 该类确实这么做了, 以使使用和监控更方便
*
* It also supports conditions and exposes one of the instrumentation methods:
* 它还提供了状况 状态 和 暴露设备方法 的支持
*
* <pre> {@code
* class Mutex implements Lock, java.io.Serializable {
*
* // Our internal helper class 内部帮助类
* private static class Sync extends AbstractQueuedSynchronizer {
* // Reports whether in locked state 报告是否在使用锁的状态
* protected boolean isHeldExclusively() {
* return getState() == 1;
* }
*
* // Acquires the lock if state is zero 如果状态为0就获取锁
* public boolean tryAcquire(int acquires) {
* assert acquires == 1; // Otherwise unused
* // CAS
* if (compareAndSetState(0, 1)) {
* setExclusiveOwnerThread(Thread.currentThread());
* return true;
* }
* return false;
* }
*
* // Releases the lock by setting state to zero 通过将状态设置为 0 来 释放锁
* protected boolean tryRelease(int releases) {
* assert releases == 1; // Otherwise unused
* if (getState() == 0) throw new IllegalMonitorStateException();
* setExclusiveOwnerThread(null);
* setState(0);
* return true;
* }
*
* // Provides a Condition 提供状态
* Condition newCondition() { return new ConditionObject(); }
*
* // Deserializes properly 适当的反序列化
* private void readObject(ObjectInputStream s)
* throws IOException, ClassNotFoundException {
* s.defaultReadObject();
* setState(0); // reset to unlocked state 重设
* }
* }
*
* // The sync object does all the hard work. We just forward to it.
* // 同步对象 做了 所有的 苦工
* private final Sync sync = new Sync();
*
* public void lock() { sync.acquire(1); }
* public boolean tryLock() { return sync.tryAcquire(1); }
* public void unlock() { sync.release(1); }
* public Condition newCondition() { return sync.newCondition(); }
* public boolean isLocked() { return sync.isHeldExclusively(); }
* public boolean hasQueuedThreads() { return sync.hasQueuedThreads(); }
* public void lockInterruptibly() throws InterruptedException {
* sync.acquireInterruptibly(1);
* }
* public boolean tryLock(long timeout, TimeUnit unit)
* throws InterruptedException {
* return sync.tryAcquireNanos(1, unit.toNanos(timeout));
* }
* }}</pre>
*
* <p>Here is a latch class that is like a {@link java.util.concurrent.CountDownLatch
* CountDownLatch} except that it only requires a single {@code signal} to
* fire.
* 类似CountDownLatch, 除了只需要 一个简单的信号 来启动
*
* Because a latch is non-exclusive, it uses the {@code shared}
* acquire and release methods.
* Latch 是非排他模式的, 使用共享模式的 获取 和 释放方法
*
* <pre> {@code
* class BooleanLatch {
*
* private static class Sync extends AbstractQueuedSynchronizer {
* boolean isSignalled() { return getState() != 0; }
*
* protected int tryAcquireShared(int ignore) {
* return isSignalled() ? 1 : -1;
* }
*
* protected boolean tryReleaseShared(int ignore) {
* setState(1);
* return true;
* }
* }
*
* private final Sync sync = new Sync();
* public boolean isSignalled() { return sync.isSignalled(); }
* public void signal() { sync.releaseShared(1); }
* public void await() throws InterruptedException {
* sync.acquireSharedInterruptibly(1);
* }
* }}</pre>
*
* @since 1.5
* @author Doug Lea
*/
AQS三板斧
-
CAS(compareAndSwap) [比较并交换]
-
源码:
/** * Atomically sets synchronization state to the given updated * value if the current state value equals the expected value. * 如果当前状态的值 等于 期望的状态的值, 原子性地设置了同步状态 为获取的更新的值 * This operation has memory semantics of a {@code volatile} read * and write. * 该操作 有 与 内存 语义相关的 由 volatile(可见性, 汇编里面其实还是加了锁) 修饰的 读 * 写操作 * @param expect the expected value * @param update the new value * @return {@code true} if successful. False return indicates that the actual * value was not equal to the expected value. */ protected final boolean compareAndSetState(int expect, int update) { // See below for intrinsics setup to support this return unsafe.compareAndSwapInt(this, stateOffset, expect, update); } // sun.misc包 底层通过cpu晶体管来调度, 保证了原子性 public final native boolean compareAndSwapInt(Object var1, long var2, int var4, int var5);
-
-
Spin (自旋)
-
理解上其实就是 写个 无限循环, 来控制线程, 不让它们跳出逻辑区之外。
-
源码:
/** * Inserts node into queue, initializing if necessary. See picture above. * 将节点插入到队列中如果需要就初始化 * @param node the node to insert * @return node's predecessor 返回该节点的前者 */ private Node enq(final Node node) { // 自旋 for (;;) { // 将 节点t 设置为 内队列的尾部节点 其实理解为 将 t 指针 指向 队列尾部 也可以 Node t = tail; if (t == null) { // Must initialize 如果尾节点为空, 必须初始化 if (compareAndSetHead(new Node())) // 将队列尾部设置为头结点 tail = head; } else { // 如果尾节点不为空, 将当前节点的前一个节点设置为尾节点 node.prev = t; if (compareAndSetTail(t, node)) { // 将尾节点的下一个节点设置为当前节点 t.next = node; return t; } } } }
-
-
LockSupport(提供阻塞)
-
park
* @param blocker the synchronization object responsible for this * thread parking * @since 1.6 */ public static void park(Object blocker) { // 获取当前线程 Thread t = Thread.currentThread(); // 设置线程的阻塞者 setBlocker(t, blocker); // 执行阻塞 false 为 isAbsolute 标签的值, 是否绝对时间, 0L 为绝对时间 UNSAFE.park(false, 0L); setBlocker(t, null); }
-
unpark
* @param thread the thread to unpark, or {@code null}, in which case * this operation has no effect */ public static void unpark(Thread thread) { // 如果线程不为空, 就解除阻塞 if (thread != null) UNSAFE.unpark(thread); }
-
-
AQS具备特性
- 阻塞等待队列
- 共享/独占
- 公平/非公平
- 可重入
- 允许中断
-
基于AQS的同步工具
- ReentrantLock(可重入锁)
- ReentrantReadWriteLock(可重入的读写锁)
- Semaphore(信号量) [基本上所有语言都有]
- CountDownLatch(技术插销) 类比golang waitGroups
- FutureTask(未来任务)
- SynchronousQueues(同步队列集)
-
CLH 队列
- 源码注释:
/** * Wait queue node class. * * <p>The wait queue is a variant of a "CLH" (Craig, Landin, and * Hagersten) lock queue. CLH locks are normally used for * spinlocks. We instead use them for blocking synchronizers, but * use the same basic tactic of holding some of the control * information about a thread in the predecessor of its node. A * "status" field in each node keeps track of whether a thread * should block. A node is signalled when its predecessor * releases. Each node of the queue otherwise serves as a * specific-notification-style monitor holding a single waiting * thread. The status field does NOT control whether threads are * granted locks etc though. A thread may try to acquire if it is * first in the queue. But being first does not guarantee success; * it only gives the right to contend. So the currently released * contender thread may need to rewait. * * <p>To enqueue into a CLH lock, you atomically splice it in as new * tail. To dequeue, you just set the head field. * <pre> * +------+ prev +-----+ +-----+ * head | | <---- | | <---- | | tail * +------+ +-----+ +-----+ * </pre> * * <p>Insertion into a CLH queue requires only a single atomic * operation on "tail", so there is a simple atomic point of * demarcation from unqueued to queued. Similarly, dequeuing * involves only updating the "head". However, it takes a bit * more work for nodes to determine who their successors are, * in part to deal with possible cancellation due to timeouts * and interrupts. * * <p>The "prev" links (not used in original CLH locks), are mainly * needed to handle cancellation. If a node is cancelled, its * successor is (normally) relinked to a non-cancelled * predecessor. For explanation of similar mechanics in the case * of spin locks, see the papers by Scott and Scherer at * http://www.cs.rochester.edu/u/scott/synchronization/ * * <p>We also use "next" links to implement blocking mechanics. * The thread id for each node is kept in its own node, so a * predecessor signals the next node to wake up by traversing * next link to determine which thread it is. Determination of * successor must avoid races with newly queued nodes to set * the "next" fields of their predecessors. This is solved * when necessary by checking backwards from the atomically * updated "tail" when a node's successor appears to be null. * (Or, said differently, the next-links are an optimization * so that we don't usually need a backward scan.) * * <p>Cancellation introduces some conservatism to the basic * algorithms. Since we must poll for cancellation of other * nodes, we can miss noticing whether a cancelled node is * ahead or behind us. This is dealt with by always unparking * successors upon cancellation, allowing them to stabilize on * a new predecessor, unless we can identify an uncancelled * predecessor who will carry this responsibility. * * <p>CLH queues need a dummy header node to get started. But * we don't create them on construction, because it would be wasted * effort if there is never contention. Instead, the node * is constructed and head and tail pointers are set upon first * contention. * * <p>Threads waiting on Conditions use the same nodes, but * use an additional link. Conditions only need to link nodes * in simple (non-concurrent) linked queues because they are * only accessed when exclusively held. Upon await, a node is * inserted into a condition queue. Upon signal, the node is * transferred to the main queue. A special value of status * field is used to mark which queue a node is on. * * <p>Thanks go to Dave Dice, Mark Moir, Victor Luchangco, Bill * Scherer and Michael Scott, along with members of JSR-166 * expert group, for helpful ideas, discussions, and critiques * on the design of this class. */
- 它并不是原始的CLH双端队列, 而是其的一个变种。(未完)