JAVA 并发：CLH 锁 与 AbstractQueuedSynchronizer

首先向Doug Lea致敬。

CLH

以下是CLH锁的一个简单实现：

class SimpleCLHLock {
    /**
     * initialized with a dummy node
     */
    private Node dummy = new Node();
    private AtomicReference<Node> tail = new AtomicReference<Node>(dummy);

    /**
     * implicit single linked list node
     */
    private static class Node {
        public volatile boolean locked = false;
    }

    private ThreadLocal<Node> threadLockNode = new ThreadLocal<Node>();

    public void lock() {
        Node last_tail;
        Node new_tail = new Node();
        new_tail.locked = true;

        while (true) {
            last_tail = tail.get();

            if (tail.compareAndSet(last_tail, new_tail)) {
                break;
            }
        }
        // we just keep previous tail node as current node's direct predecessor
        // and waiting it jump into unlocked state, so we form a single linked list implicitly
        while (last_tail.locked) {
            Thread.yield();
        }

        // we need the node reference when we want to unlock
        // some threads may waiting on its state to be unlocked in above while-loop statement.
        threadLockNode.set(new_tail);
    }

    public void unlock() {
        // retrieve the lock state held by current thread
        Node node = threadLockNode.get();
        // if null then the current thread has not own the lock, also can't perform unlock operation
        if (node == null) {
            throw new IllegalMonitorStateException();
        }
        node.locked = false;
        threadLockNode.set(null);
    }
}

测试了不同的同步方式的速度

package lock;

import java.util.List;

import java.util.ArrayList;
import java.util.concurrent.*;
import java.util.concurrent.atomic.AtomicBoolean;
import java.util.concurrent.atomic.AtomicLong;
import java.util.concurrent.atomic.AtomicReference;
import java.util.concurrent.locks.ReentrantLock;

/**
 * Created by hegaofeng on 6/30/15.
 */
public class CLH {
    private static volatile long count;

    private static AtomicLong atomicLong = new AtomicLong();

    private static ExecutorService pool = Executors.newCachedThreadPool();

    public static void main(String[] args) {

        final int deltaPerThreads = 10000;
        final int numsOfThreads = 10;

        Callable<Boolean> raw_counter = new Callable<Boolean>() {
            @Override
            public Boolean call() {
                for (int i=0; i<deltaPerThreads; i++) count++;
                return true;
            }
        };

        Callable<Boolean> syn_counter = new Callable<Boolean>() {
            @Override
            public Boolean call() throws Exception {
                    for (int i=0; i<deltaPerThreads; i++) {
                        synchronized (CLH.class) {
                            count++;
                        }
                    }

                return true;
            }
        };

        Callable<Boolean> clh_counter = new Callable<Boolean>() {
            private SimpleCLHLock lock = new SimpleCLHLock();
            @Override
            public Boolean call() throws Exception {
                for (int i=0; i<deltaPerThreads; i++) {
                    lock.lock();
                        count++;
                    lock.unlock();
                }
                return true;
            }
        };

        final Callable<Boolean> lck_counter = new Callable<Boolean>() {
            private ReentrantLock lock = new ReentrantLock();
            @Override
            public Boolean call() throws Exception {
                for (int i=0; i<deltaPerThreads; i++) {
                    lock.lock();
                        count++;
                    lock.unlock();
                }
                return true;
            }
        };

        final Callable<Boolean> lck_fair_counter = new Callable<Boolean>() {
            private ReentrantLock lock = new ReentrantLock(true);
            @Override
            public Boolean call() throws Exception {
                for (int i=0; i<deltaPerThreads; i++) {
                    lock.lock();
                    count++;
                    lock.unlock();
                }
                return true;
            }
        };

        Callable<Boolean> sem_counter = new Callable<Boolean>() {
            private Semaphore sem = new Semaphore(1);
            @Override
            public Boolean call() throws Exception {
                for(int i=0; i<deltaPerThreads; i++) {
                    sem.acquire();
                    count++;
                    sem.release();
                }
                return true;
            }
        };

        Callable<Boolean> sem_fair_counter = new Callable<Boolean>() {
            private Semaphore sem = new Semaphore(1, true);
            @Override
            public Boolean call() throws Exception {
                for(int i=0; i<deltaPerThreads; i++) {
                    sem.acquire();
                    count++;
                    sem.release();
                }
                return true;
            }
        };

        Callable<Boolean> cas_counter = new Callable<Boolean>() {
            private AtomicBoolean locked = new AtomicBoolean();
            @Override
            public Boolean call() throws Exception {
                for (int i=0; i<deltaPerThreads; i++) {
                    while (true) {
                        if (locked.compareAndSet(false, true)) {
                            count++;
                            locked.getAndSet(false);
                            break;
                        }
                    }
                }
                return true;
            }
        };

        Callable<Boolean> pure_cas = new Callable<Boolean>() {
            @Override
            public Boolean call() throws Exception {
                for (int i=0; i<deltaPerThreads; i++) {
                    atomicLong.incrementAndGet();
                }
                return true;
            }
        };

        test("raw", raw_counter, deltaPerThreads, numsOfThreads);

        test("syn", syn_counter, deltaPerThreads, numsOfThreads);

        test("CLH", clh_counter, deltaPerThreads, numsOfThreads);

        test("lock", lck_counter, deltaPerThreads, numsOfThreads);

        test("lock_fair", lck_fair_counter, deltaPerThreads, numsOfThreads);

        test("sem", sem_counter, deltaPerThreads, numsOfThreads);

        test("sem_fair", sem_fair_counter, deltaPerThreads, numsOfThreads);

        test("CAS", cas_counter, deltaPerThreads, numsOfThreads);

        test("pure_CAS", pure_cas, deltaPerThreads, numsOfThreads);
    }

    public static void test(String name, Callable<Boolean> callable, final int deltaPerThreads, final int numsOfThreads) {
        count = 0;
        atomicLong.set(0);
        pool = Executors.newCachedThreadPool();

        List<Callable<Boolean>> calls = new ArrayList<Callable<Boolean>>();
        for (int i=0; i<numsOfThreads; i++) {
            calls.add(callable);
        }
        long start = System.nanoTime();
        try {
            pool.invokeAll(calls);
        } catch (InterruptedException e) {
            e.printStackTrace();
        }
        pool.shutdown();
        long end = System.nanoTime();
        System.out.println("case   : " + name);
        System.out.println("counted: " + (name.equals("pure_CAS") ? atomicLong.get() : count));
        System.out.println("target : " + numsOfThreads * deltaPerThreads);
        System.out.println("time   : " + (end - start) / 1e6 + "ms");
        System.out.println();
    }
}

结果：

case   : raw
counted: 79270
target : 100000
time   : 11.278ms

case   : syn
counted: 100000
target : 100000
time   : 13.553ms

case   : CLH
counted: 100000
target : 100000
time   : 1037.665ms

case   : lock
counted: 100000
target : 100000
time   : 29.134ms

case   : lock_fair
counted: 100000
target : 100000
time   : 638.944ms

case   : sem
counted: 100000
target : 100000
time   : 45.154ms

case   : sem_fair
counted: 100000
target : 100000
time   : 552.855ms

case   : CAS
counted: 100000
target : 100000
time   : 158.422ms

case   : pure_CAS
counted: 100000
target : 100000
time   : 31.964ms

• raw 表示不进行任何同步，现在没有保护的情况下计数累加产生了错误
• syn 表示使用内置锁同步方式即使用synchronized块，在几种方式里表现最好
• CLH 表示使用自己实现的一个CLH锁（具有排队功能的自旋锁），这里的CLH的实现方式决定了它是一个公平锁
• lock 表示用ReentrantLock进行数据保护，速度仅次于内置锁，lock_fair是它的公平版本，不过速度上有大幅下降，变慢了将近20倍
• sem 表示使用Semaphore也就是信号量进行数据保护，速度也不错，sem_fair是它的公平版本，和ReentrantLock上出现的情况一样，公平版本比非公平版本出现了大幅的速度下降，慢了10倍
• CAS 用CAS操作实现简单的自旋锁，不具有排队功能
• pure_CAS 表示直接用使用AtomicLong类型的count变量进行计数，就不需要锁保护了，速度也是非常快得

调整测试的参数将线程数改为2，每线程增量改为50000即总增量不变的情况下得到的结果如下：

case   : raw
counted: 91460
target : 100000
time   : 4.767ms

case   : syn
counted: 100000
target : 100000
time   : 12.385ms

case   : CLH
counted: 100000
target : 100000
time   : 86.12ms

case   : lock
counted: 100000
target : 100000
time   : 33.91ms

case   : lock_fair
counted: 100000
target : 100000
time   : 108.119ms

case   : sem
counted: 100000
target : 100000
time   : 62.503ms

case   : sem_fair
counted: 100000
target : 100000
time   : 83.035ms

case   : CAS
counted: 100000
target : 100000
time   : 13.528ms

case   : pure_CAS
counted: 100000
target : 100000
time   : 11.377ms

基本锁

ReentrantLock的非公平版本在两次测试中并没有很大的变化，比较稳定。

公平锁

在不必要的情况下不去使用锁或者信号量的公平版本，它们相比非公平版本要慢很多，尤其当竞争非常激烈时。从第二组数据来看公平锁与非公平锁的差距缩小了很多，因为在竞争程度比较低（线程数少）的时候花在维护队列上的时间将大大减少。

CAS

正如Java并发编程中提到的CAS在高竞争的环境下性能不如直接使用锁，因为当竞争非常激烈时CAS花在重试上的时间将会大量增加（因为竞争激烈变量变化越快，CAS操作失败可能性越大）。而在竞争不是非常激烈的情况下CAS的开销更小，比如在第二次此时中的CAS自旋锁时间就比直接用ReentrantLock来得少。

内置锁

虽然说在Java并发编程一书中提到JDK1.6的Lock实现比synchronized要略快，这在JDK1.7中恐怕已经不成立了。

AbstractQueuedSynchronizer

http://javarticles.com/2012/10/abstractqueuedsynchronizer-aqs.html#prettyPhoto