ReentrantLock的实现不仅可以替代隐式的synchronized关键字,而且能够提供超过关键字本身的多种功能。
这里提到一个锁获取的公平性问题,如果在绝对时间上,先对锁进行获取的请求一定被先满足,那么这个锁是公平的,反之,是不公平的,也就是说等待时间最长的线程最有机会获取锁,也可以说锁的获取是有序的。ReentrantLock这个锁提供了一个构造函数,能够控制这个锁是否是公平的。
而锁的名字也是说明了这个锁具备了重复进入的可能,也就是说能够让当前线程多次的进行对锁的获取操作,这样的最大次数限制是Integer.MAX_VALUE,约21亿次左右。
事实上公平的锁机制往往没有非公平的效率高,因为公平的获取锁没有考虑到操作系统对线程的调度因素,这样造成JVM对于等待中的线程调度次序和操作系统对线程的调度之间的不匹配。对于锁的快速且重复的获取过程中,连续获取的概率是非常高的,而公平锁会压制这种情况,虽然公平性得以保障,但是响应比却下降了,但是并不是任何场景都是以TPS作为唯一指标的,因为公平锁能够减少“饥饿”发生的概率,等待越久的请求越是能够得到优先满足。
在ReentrantLock中,对于公平和非公平的定义是通过对同步器AbstractQueuedSynchronizer的扩展加以实现的,也就是在tryAcquire的实现上做了语义的控制。
01 |
final boolean nonfairTryAcquire( int acquires) { |
02 |
final Thread current = Thread.currentThread(); |
03 |
int c = getState(); |
04 |
if (c == 0 ) { |
05 |
if (compareAndSetState( 0 , acquires)) { |
06 |
setExclusiveOwnerThread(current); |
07 |
return true ; |
08 |
} |
09 |
} else if (current == getExclusiveOwnerThread()) { |
10 |
int nextc = c + acquires; |
11 |
if (nextc < 0 ) // overflow |
12 |
throw new Error( "Maximum lock count exceeded" ); |
13 |
setState(nextc); |
14 |
return true ; |
15 |
} |
16 |
return false ; |
17 |
} |
上述逻辑主要包括:
01 |
protected final boolean tryAcquire( int acquires) { |
02 |
final Thread current = Thread.currentThread(); |
03 |
int c = getState(); |
04 |
if (c == 0 ) { |
05 |
if (!hasQueuedPredecessors() && compareAndSetState( 0 , acquires)) { |
06 |
setExclusiveOwnerThread(current); |
07 |
return true ; |
08 |
} |
09 |
} else if (current == getExclusiveOwnerThread()) { |
10 |
int nextc = c + acquires; |
11 |
if (nextc < 0 ) |
12 |
throw new Error( "Maximum lock count exceeded" ); |
13 |
setState(nextc); |
14 |
return true ; |
15 |
} |
16 |
return false ; |
17 |
} |
上述逻辑相比较非公平的获取,仅加入了当前线程(Node)之前是否有前置节点在等待的判断。hasQueuedPredecessors()方法命名有些歧义,其实应该是currentThreadHasQueuedPredecessors()更为妥帖一些,也就是说当前面没有人排在该节点(Node)前面时候队且能够设置成功状态,才能够获取锁。
01 |
protected final boolean tryRelease( int releases) { |
02 |
int c = getState() - releases; |
03 |
if (Thread.currentThread() != getExclusiveOwnerThread()) |
04 |
throw new IllegalMonitorStateException(); |
05 |
boolean free = false ; |
06 |
if (c == 0 ) { |
07 |
free = true ; |
08 |
setExclusiveOwnerThread( null ); |
09 |
} |
10 |
setState(c); |
11 |
return free; |
12 |
} |
上述逻辑主要主要计算了释放状态后的值,如果为0则完全释放,返回true,反之仅是设置状态,返回false。
下面将主要的笔墨放在公平性和非公平性上,首先看一下二者测试的对比:
测试用例如下:
01 |
public class ReentrantLockTest { |
02 |
private static Lock fairLock = new ReentrantLock( true ); |
03 |
private static Lock unfairLock = new ReentrantLock(); |
04 |
05 |
@Test |
06 |
public void fair() { |
07 |
System.out.println( "fair version" ); |
08 |
for ( int i = 0 ; i < 5 ; i++) { |
09 |
Thread thread = new Thread( new Job(fairLock)); |
10 |
thread.setName( "" + i); |
11 |
thread.start(); |
12 |
} |
13 |
14 |
try { |
15 |
Thread.sleep( 5000 ); |
16 |
} catch (InterruptedException e) { |
17 |
e.printStackTrace(); |
18 |
} |
19 |
} |
20 |
21 |
@Test |
22 |
public void unfair() { |
23 |
System.out.println( "unfair version" ); |
24 |
for ( int i = 0 ; i < 5 ; i++) { |
25 |
Thread thread = new Thread( new Job(unfairLock)); |
26 |
thread.setName( "" + i); |
27 |
thread.start(); |
28 |
} |
29 |
30 |
try { |
31 |
Thread.sleep( 5000 ); |
32 |
} catch (InterruptedException e) { |
33 |
e.printStackTrace(); |
34 |
} |
35 |
} |
36 |
37 |
private static class Job implements Runnable { |
38 |
private Lock lock; |
39 |
public Job(Lock lock) { |
40 |
this .lock = lock; |
41 |
} |
42 |
43 |
@Override |
44 |
public void run() { |
45 |
for ( int i = 0 ; i < 5 ; i++) { |
46 |
lock.lock(); |
47 |
try { |
48 |
System.out.println( "Lock by:" |
49 |
+ Thread.currentThread().getName()); |
50 |
} finally { |
51 |
lock.unlock(); |
52 |
} |
53 |
} |
54 |
} |
55 |
} |
56 |
} |
调用非公平的测试方法,返回结果(部分):
unfair version
Lock by:0
Lock by:0
Lock by:2
Lock by:2
Lock by:2
Lock by:2
Lock by:2
Lock by:0
Lock by:0
Lock by:0
Lock by:1
Lock by:1
Lock by:1
调用公平的测试方法,返回结果:
fair version
Lock by:0
Lock by:1
Lock by:0
Lock by:2
Lock by:3
Lock by:4
Lock by:1
Lock by:0
Lock by:2
Lock by:3
Lock by:4
仔细观察返回的结果(其中每个数字代表一个线程),非公平的结果一个线程连续获取锁的情况非常多,而公平的结果连续获取的情况基本没有。那么在一个线程获取了锁的那一刻,究竟锁的公平性会导致锁有什么样的处理逻辑呢?
通过之前的同步器(AbstractQueuedSynchronizer)的介绍,在锁上是存在一个等待队列,sync队列,我们通过复写ReentrantLock的获取当前锁的sync队列,输出在ReentrantLock被获取时刻,当前的sync队列的状态。
修改测试如下:
01 |
public class ReentrantLockTest { |
02 |
private static Lock fairLock = new ReentrantLock2( true ); |
03 |
private static Lock unfairLock = new ReentrantLock2(); |
04 |
@Test |
05 |
public void fair() { |
06 |
System.out.println( "fair version" ); |
07 |
for ( int i = 0 ; i < 5 ; i++) { |
08 |
Thread thread = new Thread( new Job(fairLock)) { |
09 |
public String toString() { |
10 |
return getName(); |
11 |
} |
12 |
}; |
13 |
thread.setName( "" + i); |
14 |
thread.start(); |
15 |
} |
16 |
// sleep 5000ms |
17 |
} |
18 |
19 |
@Test |
20 |
public void unfair() { |
21 |
System.out.println( "unfair version" ); |
22 |
for ( int i = 0 ; i < 5 ; i++) { |
23 |
Thread thread = new Thread( new Job(unfairLock)) { |
24 |
public String toString() { |
25 |
return getName(); |
26 |
} |
27 |
}; |
28 |
thread.setName( "" + i); |
29 |
thread.start(); |
30 |
} |
31 |
// sleep 5000ms |
32 |
} |
33 |
34 |
private static class Job implements Runnable { |
35 |
private Lock lock; |
36 |
37 |
public Job(Lock lock) { |
38 |
this .lock = lock; |
39 |
} |
40 |
41 |
@Override |
42 |
public void run() { |
43 |
for ( int i = 0 ; i < 5 ; i++) { |
44 |
lock.lock(); |
45 |
try { |
46 |
System.out.println( "Lock by:" |
47 |
+ Thread.currentThread().getName() + " and " |
48 |
+ ((ReentrantLock2) lock).getQueuedThreads() |
49 |
+ " waits." ); |
50 |
} finally { |
51 |
lock.unlock(); |
52 |
} |
53 |
} |
54 |
} |
55 |
} |
56 |
57 |
private static class ReentrantLock2 extends ReentrantLock { |
58 |
// Constructor Override |
59 |
60 |
private static final long serialVersionUID = 1773716895097002072L; |
61 |
62 |
public Collection<Thread> getQueuedThreads() { |
63 |
return super .getQueuedThreads(); |
64 |
} |
65 |
} |
66 |
} |
上述逻辑主要是通过构造ReentrantLock2用来输出在sync队列中的线程内容,而且每个线程的toString方法被重写,这样当一个线程获取到锁时,sync队列里的内容也就可以得知了,运行结果如下:
调用非公平方法,返回结果:
unfair version
Lock by:0 and [] waits.
Lock by:0 and [] waits.
Lock by:3 and [2, 1] waits.
Lock by:3 and [4, 2, 1] waits.
Lock by:3 and [4, 2, 1] waits.
Lock by:3 and [0, 4, 2, 1] waits.
Lock by:3 and [0, 4, 2, 1] waits.
Lock by:1 and [0, 4, 2] waits.
Lock by:1 and [0, 4, 2] waits.
调用公平方法,返回结果:
fair version
Lock by:0 and [] waits.
Lock by:1 and [0, 4, 3, 2] waits.
Lock by:2 and [1, 0, 4, 3] waits.
Lock by:3 and [2, 1, 0, 4] waits.
Lock by:4 and [3, 2, 1, 0] waits.
Lock by:0 and [4, 3, 2, 1] waits.
Lock by:1 and [0, 4, 3, 2] waits.
Lock by:2 and [1, 0, 4, 3] waits.
可以明显看出,在非公平获取的过程中,“插队”现象非常严重,后续获取锁的线程根本不顾及sync队列中等待的线程,而是能获取就获取。反观公平获取的过程,锁的获取就类似线性化的,每次都由sync队列中等待最长的线程(链表的第一个,sync队列是由尾部结点添加,当前输出的sync队列是逆序输出)获取锁。一个 hasQueuedPredecessors方法能够获得公平性的特性,这点实际上是由AbstractQueuedSynchronizer来完成的,看一下acquire方法:
1 |
public final void acquire( int arg) { |
2 |
if (!tryAcquire(arg) && acquireQueued(addWaiter(Node.EXCLUSIVE), arg)) |
3 |
selfInterrupt(); |
4 |
} |
可以看到,如果获取状态和在sync队列中排队是短路的判断,也就是说如果tryAcquire成功,那么是不会进入sync队列的,可以通过下图来深刻的认识公平性和AbstractQueuedSynchronizer的获取过程。
非公平的,或者说默认的获取方式如下图所示:
对于状态的获取,可以快速的通过tryAcquire的成功,也就是黄色的Fast路线,也可以由于tryAcquire的失败,构造节点,进入sync队列中排序后再次获取。因此可以理解为Fast就是一个快速通道,当例子中的线程释放锁之后,快速的通过Fast通道再次获取锁,就算当前sync队列中有排队等待的线程也会被忽略。这种模式,可以保证进入和退出锁的吞吐量,但是sync队列中过早排队的线程会一直处于阻塞状态,造成“饥饿”场景。
而公平性锁,就是在tryAcquire的调用中顾及当前sync队列中的等待节点(废弃了Fast通道),也就是任意请求都需要按照sync队列中既有的顺序进行,先到先得。这样很好的确保了公平性,但是可以从结果中看到,吞吐量就没有非公平的锁高了。