有条件的公平ReadWriteLock

Question

首先，我检查了之前关于这个主题的问题，但是没有一个适合我的具体问题。

我得到了下面的代码，它说明了带有时间戳的传感器和存储在双精度数组中的数据，以及我实现的FairRWLock的一个实例。

class LockedSensors implements Sensors {
    long time = 0;
    double data[];
    FairRWLock lock = new FairRWLock();

    LockedSensors() {
        time = 0;
    }

    // store data and timestamp
    // if and only if data stored previously is older (lower timestamp)
    public void update(long timestamp, double[] data) {
        lock.writeAcquire();
            if (timestamp > time) {
                if (this.data == null)
                    this.data = new double[data.length];
                time = timestamp;
                for (int i = 0; i < data.length; ++i)
                    this.data[i] = data[i];
            }
        lock.writeRelease();
    }

    // pre: val != null
    // pre: val.length matches length of data written via update
    // if no data has been written previously, return 0
    // otherwise return current timestamp and fill current data to array passed
    // as val
    public long get(double val[]) {
        try{
            lock.readAcquire();
                if (time == 0) return 0;
                for (int i = 0; i < data.length; ++i)
                    val[i] = data[i];
                return time;
        } finally{lock.readRelease();}
    }
}

它支持update和get，前者取决于接收新数据的时间，后者提取存储在特定传感器中的数据。

这是我的FairRWLock实现：

class FairRWLock{
    private int readers = 0, writers = 0, readersWaiting = 0, writersWaiting = 0, writersWait = 0;
    private static final int ReaderPriority = 30;
    private Lock lock = new ReentrantLock();
    private Condition readerPass = lock.newCondition();
    private Condition writerPass = lock.newCondition();
    /*
     * readers denotes the number of current readers, writers equivalent, readersWaiting denotes the number of readers
     * awaiting their signal, writersWaiting equivalent. writersWait denotes the number of readers the writers have to
     * let pass before they can proceed, this amount is controlled by the ReaderPriority (reset occurs when writer releases)
     */


    /*
     * increment the number of waiting readers, check if there are any currently working writers OR writers are waiting
     * whilst they don't have to let any readers pass. When signaled, decrement readersWaiting, decrement the number of
     * readers the writers have to let pass and increment the number of current readers.
     */
    public void readAcquire(){
        lock.lock();
            readersWaiting++;
            while(writers > 0 || (writersWaiting > 0 && writersWait <= 0)){
                try {
                    readerPass.await();
                } catch (InterruptedException e) {}
            }
            readersWaiting--;
            writersWait--;
            readers++;
        lock.unlock();
    }

    /*
     * simply decrement number of readers and signal the threads that have to be signaled
     */
    public void readRelease(){
        lock.lock();
            readers--;
            signaling();
        lock.unlock();
    }

    /*
     * increment number of waiting writers, check if there are currently working writers OR readers OR readers currently
     * have priority over the writers. When signaled decrement writersWaiting, increment number of writers
     */
    public void writeAcquire(){
        lock.lock();
            writersWaiting++;
            while(writers > 0 || readers > 0 || (readersWaiting > 0 && writersWait > 0)){
                try{
                    writerPass.await();
                } catch(InterruptedException e) {}
            }
            writersWaiting--;
            writers++;
        lock.unlock();
    }

    /*
     * simply decrement number of current writers, reset the number of readers the writers have to let pass before
     * another writer may pass. signal the ones that should be
     */
    public void writeRelease(){
        lock.lock();
            writers--;
            writersWait = ReaderPriority;
            signaling();
        lock.unlock();
    }

    /*
     * check first if readers currently got priority over the writers. if so (readersWaiting??) ? signal them : signalAll,
     * if not (writersWaiting??) ? signal them : signalAll
     */
    private void signaling(){
        if(writersWait > 0){
            if(readersWaiting > 0) readerPass.signalAll();
            else writerPass.signal();
        } else{
            if(writersWaiting > 0) writerPass.signal();
            else readerPass.signalAll();
        }
    }
}

我对条件锁定不是很熟悉，我的代码似乎不是饿死就是死锁。但是，我找不到问题所在(很可能是在FairRWLock实现中的某个地方)。

Answer 1

试图在不公平的锁上构建公平的锁是没有意义的。当线程进入readAcquire()或writeAcquire()时，它们正在调用lock.lock()，如果不能立即成功，它们可能会进入等待状态，并被任意数量的线程超过，然后才能继续。

在这一点上，无论你以后做什么，都不可能重新建立公平。但值得注意的是，您也忽略了await()的含义。此操作将临时释放锁，因为只有这样才能让其他线程有机会满足您正在等待的条件。当线程获得signal()ed时，它必须重新获取锁，这也不是一个公平的操作。任意数量的线程可能会发出新的锁定请求，从而在很久以前调用了await()的线程继续执行之前完全改变了情况。

最后，你不会想要公平的。update操作的目的是忽略过时的更新，因此如果新的update请求能够更快地进行，那么实际上这将是一种胜利，因为挂起的旧请求将变成无操作。对于并发的get请求，您实际上根本不希望阻塞，所有的读请求都应该能够并发运行，但是，当然，您需要一致性(线程安全)，并且这里没有写入器饥饿。

最好的解决方案是完全不加锁，实现整个操作的无锁：

class LockedSensors implements Sensors {
    private static final class State {
        final long time;
        final double[] data;
        State(long t, double[] in) {
            time = t;
            data = in.clone();
        }
    }
    final AtomicReference<State> current = new AtomicReference<>();
    LockedSensors() {}

    // store data and timestamp
    // if and only if data stored previously is older (lower timestamp)
    public void update(long timestamp, double[] data) {
        State newState = null;
        for(;;) {
            State old = current.get();
            if(old != null && old.time > timestamp) return;
            if(newState == null) newState = new State(timestamp, data);
            if(current.compareAndSet(old, newState)) return;
        }
    }

    // pre: val != null
    // pre: val.length matches length of data written via update
    // if no data has been written previously, return 0
    // otherwise return current timestamp and fill current data to array passed as val
    public long get(double[] val) {
        State actual = current.get();
        if(actual == null) return 0;
        if(actual.data.length != val.length)
            throw new IllegalArgumentException();
        System.arraycopy(actual.data, 0, val, 0, actual.data.length);
        return actual.time;
    }
}

在这里，读者可以继续返回上次完成的更新的结果，而不会阻塞任何写入者。即使是编写器也不会互相阻止，但如果中间发生了另一次更新，则可能不得不旋转，但由于每个编写器都会在遇到较新的时间戳时立即返回，并且始终会有至少一个编写器取得进展，所以这里没有问题。