本文共 15495 字,大约阅读时间需要 51 分钟。
读写锁维护了一对相关的锁,一个用于只读操作,一个用于写入操作。
只要没有writer,读锁可以由多个reader线程同时保持。写锁是独占的。与互斥锁相比,使用读写锁能否提升性能则取决于读写操作期间读取数据相对于修改数据的频率,以及数据的争用,即在同一时间试图对该数据执行读取或写入操作的线程数
。
读写锁适用于读多写少
的情况。
ReentrantReadWriteLock
也是基于 AbstractQueuedSynchronizer
实现的,具有下面这些属性
从读取锁升级到写入锁是不可能的
。ReentrantLock.newCondition()
提供的 Condition 实现对 ReentrantLock 所做的行为相同。当然,此 Condition 只能用于写锁
。 读锁不支持 Condition,readLock().newCondition() 会抛UnsupportedOperationExceptionAQS以单个 int 类型的原子变量来表示其状态,定义了4个抽象方法(
ReentrantLock,它是可重入的独占锁,内部的 Sync 类实现了 tryAcquire(int)、tryRelease(int) 方法,并用状态的值来表示重入次数,加锁或重入锁时状态加 1,释放锁时状态减 1,状态值等于 0 表示锁空闲。
CountDownLatch,它是一个关卡,在条件满足前阻塞所有等待线程,条件满足后允许所有线程通过。内部类 Sync 把状态初始化为大于 0 的某个值,当状态大于 0 时所有wait线程阻塞,每调用一次 countDown 方法就把状态值减 1,减为 0 时允许所有线程通过。利用了AQS的共享模式。
现在,要用AQS来实现 ReentrantReadWriteLock。
ReentrantLock 里,状态值表示重入计数,现在如何在AQS里表示每个读锁、写锁的重入次数呢?如何实现读锁、写锁的公平性呢?
一个状态是没法既表示读锁,又表示写锁的,不够用啊,那就辦成两份用了,状态的高位部分表示读锁,低位表示写锁,由于写锁只有一个,所以写锁的重入计数也解决了,这也会导致写锁可重入的次数减小。
由于读锁可以同时有多个,肯定不能再用辦成两份用的方法来处理了,但我们有 ThreadLocal
,可以把线程重入读锁的次数作为值存在 ThreadLocal
对于公平性的实现,可以通过AQS的等待队列和它的抽象方法来控制,在状态值的另一半里存储当前持有读锁的线程数。如果读线程申请读锁,当前写锁重入次数不为 0 时,则等待,否则可以马上分配;如果是写线程申请写锁,当前状态为 0 则可以马上分配,否则等待。
AQS 的状态是32位(int 类型)的,
读锁用高16位,表示持有读锁的线程数(sharedCount),写锁低16位,表示写锁的重入次数 (exclusiveCount)。状态值为 0 表示锁空闲,sharedCount不为 0 表示分配了读锁,exclusiveCount 不为 0 表示分配了写锁,sharedCount和exclusiveCount 肯定不会同时不为 0。
/** * Synchronization implementation for ReentrantReadWriteLock. * Subclassed into fair and nonfair versions. */abstract static class Sync extends AbstractQueuedSynchronizer { private static final long serialVersionUID = 6317671515068378041L; /* * Read vs write count extraction constants and functions. * Lock state is logically divided into two unsigned shorts: * The lower one representing the exclusive (writer) lock hold count, * and the upper the shared (reader) hold count. */ static final int SHARED_SHIFT = 16; // 由于读锁用32位 int 的高位部分,所以读锁个数加1,其实是加 2^16数值 static final int SHARED_UNIT = (1 << SHARED_SHIFT); // 写锁的可重入的最大次数、读锁允许的最大数量 static final int MAX_COUNT = (1 << SHARED_SHIFT) - 1; // 写锁的掩码,用于状态的低16位有效值 static final int EXCLUSIVE_MASK = (1 << SHARED_SHIFT) - 1; // 读锁计数,当前持有读锁的线程数 /** Returns the number of shared holds represented in count */ static int sharedCount(int c) { return c >>> SHARED_SHIFT; } // 写锁的计数,也就是它的重入次数 /** Returns the number of exclusive holds represented in count */ static int exclusiveCount(int c) { return c & EXCLUSIVE_MASK; }}
/** * A counter for per-thread read hold counts. * Maintained as a ThreadLocal; cached in cachedHoldCounter * 每个线程特定的 read 持有计数。存放在ThreadLocal,缓存在cachedHoldCounter * 不需要是线程安全的。 */ static final class HoldCounter { int count = 0; // 使用id而不是引用是为了避免保留垃圾。注意这是个常量。 // Use id, not reference, to avoid garbage retention final long tid = Thread.currentThread().getId(); } /** * ThreadLocal subclass. Easiest to explicitly define for sake * of deserialization mechanics. * 采用继承是为了重写 initialValue 方法,这样就不用进行这样的处理: * 如果ThreadLocal没有当前线程的计数,则new一个,再放进ThreadLocal里。 * 可以直接调用 get */ static final class ThreadLocalHoldCounter extends ThreadLocal{ public HoldCounter initialValue() { return new HoldCounter(); } } /** * The number of reentrant read locks held by current thread. * Initialized only in constructor and readObject. * Removed whenever a thread's read hold count drops to 0. * 保存当前线程重入读锁的次数的容器。在读锁重入次数为 0 时移除 */ private transient ThreadLocalHoldCounter readHolds; /** * The hold count of the last thread to successfully acquire * readLock. This saves ThreadLocal lookup in the common case * where the next thread to release is the last one to * acquire. This is non-volatile since it is just used * as a heuristic, and would be great for threads to cache. * * Can outlive the Thread for which it is caching the read * hold count, but avoids garbage retention by not retaining a * reference to the Thread. * *
Accessed via a benign data race; relies on the memory * model's final field and out-of-thin-air guarantees. */ /** * 最近一个成功获取读锁的线程的计数。这省却了ThreadLocal查找, * 通常情况下,下一个要释放的线程是最后一个获取的线程。 * 这不是 volatile 的,因为它仅用于试探的,线程进行缓存也是极好的 * (因为判断是否是当前线程是通过线程id来比较的)。 */ private transient HoldCounter cachedHoldCounter; /** * firstReader is the first thread to have acquired the read lock. * firstReaderHoldCount is firstReader's hold count. * *
More precisely, firstReader is the unique thread that last * changed the shared count from 0 to 1, and has not released the * read lock since then; null if there is no such thread. * *
Cannot cause garbage retention unless the thread terminated * without relinquishing its read locks, since tryReleaseShared * sets it to null. * *
Accessed via a benign data race; relies on the memory * model's out-of-thin-air guarantees for references. * *
This allows tracking of read holds for uncontended read * locks to be very cheap. */ /** * firstReader是第一个获取读锁的线程,更加确切地说是最后一个把 共享计数 从 0 改为 1 的(在锁空闲的时候),而且在那之后还没有释放读锁的独特的线程!如果不存在这样的线程则为null * firstReaderHoldCount 是 firstReader 的重入计数。 * * firstReader 不会导致垃圾存留,因此在 tryReleaseShared 里设置为null, * 除非线程异常终止,没有释放读锁 * * 这使得在跟踪无竞争的读锁计数时代价非常低 * * firstReader及其计数firstReaderHoldCount是不会放入 readHolds 的 */ private transient Thread firstReader = null; private transient int firstReaderHoldCount; Sync() { readHolds = new ThreadLocalHoldCounter(); // ensures visibility of readHolds setState(getState()); // 确保 readHolds 的内存可见性,利用 volatile 写的内存语义 }}
通过 tryAcquire 和 tryRelease 实现,源码里有这么一段说明
tryRelease 和 tryAcquire 可能被 Conditions 调用。
因此可能出现参数里包含在条件等待和用 tryAcquire 重新获取到锁的期间内已经释放的 读和写 计数这说明看起来像是在 tryAcquire 里设置状态时要考虑方法参数(acquires)的高位部分,其实是不需要的。由于写锁是独占的,acquires 表示的只能是写锁的计数,如果当前线程成功获取写锁,只需要简单地把当前状态加上 acquires 的值即可,tryRelease 里直接减去其参数值即可。
protected final boolean tryAcquire(int acquires) { /* * Walkthrough: * 1. If read count nonzero or write count nonzero * and owner is a different thread, fail. * 2. If count would saturate, fail. (This can only * happen if count is already nonzero.) * 3. Otherwise, this thread is eligible for lock if * it is either a reentrant acquire or * queue policy allows it. If so, update state * and set owner. */ Thread current = Thread.currentThread(); int c = getState(); int w = exclusiveCount(c); if (c != 0) { // 状态不为0,表示锁被分配出去了 // (Note: if c != 0 and w == 0 then shared count != 0) // c != 0 且 w == 0 : 分配了读锁 // w != 0 && current != getExclusiveOwnerThread() 表示其他线程获取了写锁。 if (w == 0 || current != getExclusiveOwnerThread()) return false ; // 写锁重入 // 检测是否超过最大重入次数。 if (w + exclusiveCount(acquires) > MAX_COUNT) throw new Error("Maximum lock count exceeded"); // 更新写锁重入次数,写锁在低位,直接加上 acquire 即可。 // Reentrant acquire setState(c + acquires); return true ; } // writerShouldBlock 留给子类实现,用于实现公平性策略 // 如果允许获取写锁,则用 CAS 更新状态 if (writerShouldBlock() || !compareAndSetState(c, c + acquires)) return false ; // 不允许获取锁 或 CAS 失败。 // 获取写锁,设置独占线程 setExclusiveOwnerThread(current); return true;}protected final boolean tryRelease(int releases) { if (!isHeldExclusively()) // 是否是当前线程持有写锁 throw new IllegalMonitorStateException(); // 这里不考虑高16位是因为高16位肯定是 0 int nextc = getState() - releases; boolean free = exclusiveCount(nextc) == 0; if (free) setExclusiveOwnerThread( null); // 写锁完全释放,设置独占线程为null setState(nextc); return free;}
// 参数变为 unused 是因为读锁的重入计数是内部维护的protected final int tryAcquireShared(int unused) {/* * Walkthrough: * 1. If write lock held by another thread, fail. * 2. Otherwise, this thread is eligible for * lock wrt state, so ask if it should block * because of queue policy. If not, try * to grant by CASing state and updating count. * Note that step does not check for reentrant * acquires, which is postponed to full version * to avoid having to check hold count in * the more typical non-reentrant case. * 3. If step 2 fails either because thread * apparently not eligible or CAS fails or count * saturated, chain to version with full retry loop. */ Thread current = Thread.currentThread(); int c = getState(); // 持有写锁的线程可以获取读锁 if (exclusiveCount(c) != 0 && // 已分配了写锁 getExclusiveOwnerThread() != current) // 且当前线程不是持有写锁的线程 return -1; int r = sharedCount(c); // 获取读锁的计数 if (!readerShouldBlock() && // 由子类根据其公平策略决定是否允许获取读锁 r < MAX_COUNT && // 读锁数量尚未达到最大值 // 尝试获取读锁,注意读线程计数的单位是 2^16 compareAndSetState(c, c + SHARED_UNIT)) { // 成功获取读锁 // 注意下面对firstReader的处理:firstReader是不会放到readHolds里的 // 这样,在读锁只有一个的情况下,就避免了查找readHolds if (r == 0) { // 是 firstReader,计数不会放入 readHolds firstReader = current; firstReaderHoldCount = 1; } else if (firstReader == current) { // firstReader 重入 firstReaderHoldCount++; } else { // 非 firstReader 读锁重入计数更新 HoldCounter rh = cachedHoldCounter; // 首先访问缓存 if (rh == null || rh.tid != current.getId()) cachedHoldCounter = rh = readHolds.get(); else if (rh.count == 0) readHolds.set(rh); rh.count++; } return 1; } // 获取读锁失败,放到循环里重试。 return fullTryAcquireShared(current);}/** * Full version of acquire for reads, that handles CAS misses * and reentrant reads not dealt with in tryAcquireShared. */final int fullTryAcquireShared(Thread current) { /* * This code is in part redundant with that in * tryAcquireShared but is simpler overall by not * complicating tryAcquireShared with interactions between * retries and lazily reading hold counts. */ HoldCounter rh = null; for (;;) { int c = getState(); if (exclusiveCount(c) != 0) { if (getExclusiveOwnerThread() != current) // 写锁被分配,非写锁线程获取读锁失败 return -1; // else we hold the exclusive lock; blocking here // would cause deadlock. // 否则,当前线程持有写锁,在这里阻塞将导致死锁 } else if (readerShouldBlock()) { // Make sure we're not acquiring read lock reentrantly // 写锁空闲 且 公平策略决定 线程应当被阻塞 // 如果是已获取读锁的线程重入读锁时, // 即使公平策略指示应当阻塞也不会阻塞。 // 否则也会导致死锁 if (firstReader == current) { // assert firstReaderHoldCount > 0; } else { if (rh == null) { rh = cachedHoldCounter; if (rh == null || rh.tid != current.getId()) { rh = readHolds.get(); if (rh.count == 0) readHolds.remove(); } } // 需要阻塞且是非重入(还未获取读锁的),获取失败。 if (rh.count == 0) return -1; } } // 写锁空闲 且 公平策略决定线程可以获取读锁 if (sharedCount(c) == MAX_COUNT) throw new Error( "Maximum lock count exceeded"); if (compareAndSetState(c, c + SHARED_UNIT)) { // 申请读锁成功,下面的处理跟tryAcquireShared类似 if (sharedCount(c) == 0) { firstReader = current; firstReaderHoldCount = 1; } else if (firstReader == current) { firstReaderHoldCount++; } else { // 设定最后一次获取读锁的缓存 if (rh == null) rh = cachedHoldCounter; if (rh == null || rh.tid != current.getId()) rh = readHolds.get(); else if (rh.count == 0) readHolds.set(rh); rh.count++; cachedHoldCounter = rh; // 缓存起来用于释放 cache for release } return 1; } }}protected final boolean tryReleaseShared(int unused) { Thread current = Thread.currentThread(); // 清理firstReader缓存 或 readHolds里的重入计数 if (firstReader == current) { // assert firstReaderHoldCount > 0; if (firstReaderHoldCount == 1) firstReader = null; else firstReaderHoldCount--; } else { HoldCounter rh = cachedHoldCounter; if (rh == null || rh.tid != current.getId()) rh = readHolds.get(); int count = rh.count; if (count <= 1) { // 完全释放读锁 readHolds.remove(); if (count <= 0) throw unmatchedUnlockException(); } --rh.count; // 主要用于重入退出 } // 循环在CAS更新状态值,主要是把读锁数量减 1 for (;;) { int c = getState(); int nextc = c - SHARED_UNIT; if (compareAndSetState(c, nextc)) // 释放读锁对其他读线程没有任何影响, // 但可以允许等待的写线程继续,如果读锁、写锁都空闲。 return nextc == 0; }}
策略由 Sync 的子类 FairSync 和 NonfairSync 实现
/** * 这个非公平策略的同步器是写锁优先的,申请写锁时总是不阻塞。 */static final class NonfairSync extends Sync { private static final long serialVersionUID = -8159625535654395037L; final boolean writerShouldBlock() { return false; // 写线程总是可以突入 } final boolean readerShouldBlock() { /* 作为一个启发用于避免写线程饥饿,如果线程临时出现在等待队列的头部则阻塞, * 如果存在这样的,则是写线程。 */ return apparentlyFirstQueuedIsExclusive(); }}/** * 公平的 Sync,它的策略是:如果线程准备获取锁时, * 同步队列里有等待线程,则阻塞获取锁,不管是否是重入 * 这也就需要tryAcqire、tryAcquireShared方法进行处理。 */static final class FairSync extends Sync { private static final long serialVersionUID = -2274990926593161451L; final boolean writerShouldBlock() { return hasQueuedPredecessors(); } final boolean readerShouldBlock() { return hasQueuedPredecessors(); }}现在用奇数表示申请读锁的读线程,偶数表示申请写锁的写线程,每个数都表示一个不同的线程,存在下面这样的申请队列,假设开始时锁空闲:1 3 5 0 7 9 2 4读线程1申请读锁时,锁是空闲的,马上分配,读线程3、5申请时,由于已分配读锁,它们也可以马上获取读锁。假设此时有线程11申请读锁,由于它不是读锁重入,只能等待。而线程1再次申请读锁是可以的,因为它的重入。写线程0申请写锁时,由于分配了读锁,只能等待,当读线程1、3、5都释放读锁后,线程0可以获取写锁。线程0释放后,线程7、9获取读锁,它们释放后,线程2获取写锁,此时线程4必须等待线程2释放。线程4在线程2释放写锁后获取写锁,它释放写锁后,锁恢复空闲
转载地址:http://ddjpx.baihongyu.com/