Why Volatile is behaving weirdly

问题:

I have experience this weird behavior of volatile keyword recently. As far as i know,

1. volatile keyword is applied on to the variable to reflect the changes done on the data of the variable by one thread onto the other thread.

2. volatile keyword prevents caching of the data on the thread.

I did a small test........

1. I used an integer variable named count, and used volatile keyword on it.

2. Then made 2 different threads to increment the variable value to 10000, so the end resultant should be 20000.

3. But thats not the case always, with volatile keyword i am getting not getting 20000 consistently, but 18534, 15000, etc.... and sometimes 20000.

4. But while i used synchronized keyword, it just worked fine, why....??

Can anyone please explain me this behaviour of volatile keyword.

i am posting my code with volatile keyword and as well as the one with synchronzied keyword.

The following code below behaves inconsistently with volatile keyword on variable count

public class SynVsVol implements Runnable{

    volatile int  count = 0;

    public void go(){

        for (int i=0 ; i<10000 ; i++){
             count = count + 1;
         }
    }

    @Override
    public void run() {
        go();
    }

    public static void main(String[] args){

        SynVsVol s = new SynVsVol();
        Thread t1 = new Thread(s);
        Thread t2 = new Thread(s);
        t1.start();
        t2.start();

        try {
            t1.join();
            t2.join();
        } catch (InterruptedException e) {
            // TODO Auto-generated catch block
            e.printStackTrace();
        }
        System.out.println("Total Count Value: "+s.count);
    }
}

The following code behaves perfectly with synchronized keyword on the method go().

public class SynVsVol implements Runnable{

    int  count = 0;

    public synchronized void go(){
        for (int i=0 ; i<10000 ; i++){
             count = count + 1;
         }
    }

    @Override
    public void run() {
        go();
    }

    public static void main(String[] args){

        SynVsVol s = new SynVsVol();
        Thread t1 = new Thread(s);
        Thread t2 = new Thread(s);
        t1.start();
        t2.start();

        try {
            t1.join();
            t2.join();
        } catch (InterruptedException e) {
            // TODO Auto-generated catch block
            e.printStackTrace();
        }
        System.out.println("Total Count Value: "+s.count);
    }
}

回答1:

The volatile keyword is not a synchronization primitive. It merely prevents caching of the value on the thread, but it does not prevent two threads from modifying the same value and writing it back concurrently.

Let's say two threads come to the point when they need to increment the counter, which is now set to 5. Both threads see 5, make 6 out of it, and write it back into the counter. If the counter were not volatile, both threads could have assumed that they know the value is 6, and skip the next read. However, it's volatile, so they both would read 6 back, and continue incrementing. Since the threads are not going in lock-step, you may see a value different from 10000 in the output, but there's virtually no chance that you would see 20000.

回答2:

count = count + 1 is not atomic. It has three steps:

1. read the current value of the variable
2. increment the value
3. write the new value back to the variable

These three steps are getting interwoven, resulting in different execution paths, resulting in an incorrect value. Use AtomicInteger.incrementAndGet() instead if you want to avoid the synchronized keyword.

So although the volatile keyword acts pretty much as you described it, that only applies to each seperate operation, not to all three operations collectively.

回答3:

The fact that a variable is volatile does not mean every operation it's involved in is atomic. For instance, this line in SynVsVol.Go:

count = count + 1;

will first have count read, then incremented, and the result will then be written back. If some other thread will execute it at the same time, the results depend on the interleaving of the commands.

Now, when you add the syncronized, SynVsVol.Go executes atomically. Namely, the increment is done as a whole by a single thread, and the other one can't modify count until it is done.

Lastly, caching of member variables that are only modified within a syncronized block is much easier. The compiler can read their value when the monitor is acquired, cache it in a register, have all changes done on that register, and eventually flush it back to the main memory when the monitor is released. This is also the case when you call wait in a synchronized block, and when some other thread notifys you: cached member variables will be synchronized, and your program will remain coherent. That's guaranteed even if the member variable is not declared as volatile:

Synchronization ensures that memory writes by a thread before or during a synchronized block are made visible in a predictable manner to other threads which synchronize on the same monitor.

回答4:

Your code is broken because it treats the read-and-increment operation on a volatile as atomic, which it is not. The code doesn't contain a data race, but it does contain a race condition on the int.