Synchronization

Every thread has a life of its own. Normally, a thread goes about its business without any regard for what other threads in the app are doing. Threads may be time-sliced, which means they can run in arbitrary spurts and bursts as directed by the operating system. On a multiprocessor system, it is even possible for many different threads to be running simultaneously on different CPUs. This section is about coordinating the activities of two or more threads so that they can work together and not collide in their use of the same variables and methods (coordinating their play on the golf course). Java provides a few simple structures for synchronizing the activities of threads. They are all based on the concept of monitors, a widely used synchronization scheme (developed by C.A.R. Hoare). You don't have to know the details about how monitors work to be able to use them, but it may help you to have a picture in mind. A monitor is essentially a lock. The lock is attached to a resource that many threads may need to access, but that should be accessed by only one thread at a time. It's very much like a restroom with a lock on the door. If the resource is not being used, the thread can acquire the lock and access the resource. That is, if the restroom is unlocked, you can enter and lock the door. When the thread is done, it relinquishes the lock, just as you unlock the door and leave it open for the next person. However, if another thread already has the lock for the resource, all other threads have to wait until the current thread finishes and releases the lock. This is just like when the restroom is occupied when you arrive: you have to wait until the current user is done and unlocks the door. Fortunately, Java makes the process of synchronizing access to resources fairly easy. The language handles setting up and acquiring locks; all you have to do is specify which resources require synchronization.

Serializing Access to Methods

The most common need for synchronization among threads in Java is to serialize their access to some resource (an object)in other words, to make sure that only one thread at a time can manipulate an object or variable.[*] In Java, every object has an associated lock. To be more specific, every class and every instance of a class has its own lock. The synchronized keyword marks places where a thread must acquire the lock before proceeding.

[*] Don't confuse the term "serialize" in this context with Java object serialization, which is a mechanism for making objects persistent. The underlying meaning (to place one thing after another) does apply to both, however. In the case of object serialization, the object's data is laid out, byte for byte, in a certain order.

For example, suppose we implemented a SpeechSynthesizer class that contains a say( ) method. We don't want multiple threads calling say( ) at the same time, or we wouldn't be able to understand anything being said. So we mark the say( ) method as synchronized, which means that a thread has to acquire the lock on the SpeechSynthesizer object before it can speak:

 class SpeechSynthesizer {
 synchronized void say( String words ) {
 // speak
 }
 }


Because say( ) is an instance method, a thread has to acquire the lock on the SpeechSynthesizer instance it is using before it can invoke the say( ) method. When say( ) has completed, it gives up the lock, which allows the next waiting thread to acquire the lock and run the method. It doesn't matter whether the thread is owned by the SpeechSynthesizer itself or some other object; every thread has to acquire the same lock, that of the SpeechSynthesizer instance. If say( ) were a class (static) method instead of an instance method, we could still mark it as synchronized. In this case, because no instance object is involved, the lock is on the class object itself. Often, you want to synchronize multiple methods of the same class so that only one method modifies or examines parts of the class at a time. All static synchronized methods in a class use the same class object lock. By the same token, all instance methods in a class use the same instance object lock. In this way, Java can guarantee that only one of a set of synchronized methods is running at a time. For example, a SpreadSheet class might contain a number of instance variables that represent cell values as well as some methods that manipulate the cells in a row:

 class SpreadSheet {
 int cellA1, cellA2, cellA3;
 synchronized int sumRow( ) {
 return cellA1 + cellA2 + cellA3;
 }
 synchronized void setRow( int a1, int a2, int a3 ) {
 cellA1 = a1;
 cellA2 = a2;
 cellA3 = a3;
 }
 ...
 }


In this example, methods setRow( ) and sumRow( ) both access the cell values. You can see that problems might arise if one thread were changing the values of the variables in setRow( ) at the same moment another thread was reading the values in sumRow( ). To prevent this, we have marked both methods as synchronized. When threads are synchronized, only one runs at a time. If a thread is in the middle of executing setRow( ) when another thread calls sumRow( ), the second thread waits until the first one finishes executing setRow( ) before it runs sumRow( ). This synchronization allows us to preserve the consistency of the SpreadSheet. The best part is that all this locking and waiting is handled by Java; it's transparent to the programmer. In addition to synchronizing entire methods, the synchronized keyword can be used in a special construct to guard arbitrary blocks of code. In this form, it also takes an explicit argument that specifies the object for which it is to acquire a lock:

 synchronized ( myObject ) {
 // Functionality that needs to be synced
 }


This code block can appear in any method. When it is reached, the thread has to acquire the lock on myObject before proceeding. In this way, we can synchronize methods (or parts of methods) in different classes in the same way as methods in the same class. A synchronized instance method is, therefore, equivalent to a method with its statements synchronized on the current object. Thus:

 synchronized void myMethod ( ) {
 ...
}


is equivalent to:

 void myMethod ( ) {
 synchronized ( this ) {
 ...
 }
 }


Accessing variables

In the SpreadSheet example, we guarded access to a set of instance variables with a synchronized method, which we did mainly so that we wouldn't change one of the variables while someone was reading the rest of them. We wanted to keep them coordinated. But what about individual variable types? Do they need to be synchronized? Normally, the answer is no. Almost all operations on primitives and object reference types in Java happen atomically: they are handled by the VM in one step, with no opportunity for two threads to collide. You can't be in the middle of changing a reference and be only partially done when another thread looks at the reference. But watch outwe did say almost. If you read the Java VM specification carefully, you will see that the double and long primitive types are not guaranteed to be handled atomically. Both of these types represent 64-bit values. The problem has to do with how the Java VM's stack handles them. It is possible that this specification will be beefed up in the future. But for now, to be strict, you should synchronize access to your double and long instance variables through accessor methods or use the volatile keyword or an atomic wrapper class, as we'll describe next. Another issue, independent of the atomicity of the values, is the notion of different threads in the VM caching values for periods of time. That is, even though one thread may have changed the value, the Java VM may not be obliged to make that value appear until the VM reaches a certain state known as a "memory barrier." While this should not be a problem in most real-world coding cases, you can address this by declaring the variable with the volatile keyword. This keyword indicates to the VM that the value may be changed by external threads and effectively synchronizes access to it automatically. Finally, Java 5.0 added a new package, java.util.concurrent.atomic that provides synchronized wrapper classes for all primitive types and references. These wrappers provide not only simple set( ) and get( ) operations on the values but also specialized "combo" operations, such as compareAndSet( ), that work atomically and can be used to build higher-level synchronized app components. The classes in this package were designed specifically to map down to hardware-level functionality in many cases and can be very efficient. We'll talk more about them later in this chapter.

Reentrant locking

The locks acquired by Java upon entering a synchronized method or block of code are reentrant, meaning that the thread holding onto the lock may acquire the same lock again any number of times and never blocks waiting for itself. In most cases, this means that the code behaves as you'd expect; a thread can call a synchronized method recursively and can itself call upon other synchronized methods within the same object.

The wait( ) and notify( ) Methods

With the synchronized keyword, we can serialize the execution of methods and blocks of code so that only one thread may be executing a synchronized item at a time. The wait( ) and notify( ) methods of the Object class extend this capability by allowing us to coordinate the waiting and running threads explicitly. Every object in Java is a subclass of Object, so every object inherits these methods. By using wait( ) and notify( ), a thread can effectively give up its hold on a lock at an arbitrary point and then wait for another thread to give it back before continuing. All of the coordinated activity still happens inside synchronized blocks, and still only one thread is executing at a given time. By executing wait( ) from a synchronized block, a thread gives up its hold on the lock and goes to sleep. A thread might do this if it needs to wait for something to happen in another part of the app, as we'll see shortly. Later, when the necessary event happens, a thread that is running calls notify( ) from a block synchronized on the same object. Now, the first thread wakes up and begins trying to acquire the lock again. When the first thread manages to reacquire the lock, it continues from the point it left off. However, the thread that was waiting may not get the lock immediately (or perhaps ever). It depends on when the second thread eventually releases the lock and which thread manages to snag it next. The first thread won't wake up from the wait( ) unless another thread calls notify( ). An overloaded version of wait( ), however, allows us to specify a timeout period. If another thread doesn't call notify( ) in the specified period, the waiting thread automatically wakes up. Let's look at a simple scenario to see what's going on. In the following example, we'll assume there are three threadsone waiting to execute each of the three synchronized methods of the MyThing class. We'll call them the waiter, notifier, and related threads. Here's a code fragment to illustrate:

 class MyThing {
 synchronized void waiterMethod( ) {
 // do some stuff
 wait( ); // now wait for notifier to do something
 // continue where we left off
 }
 synchronized void notifierMethod( ) {
 // do some stuff
 notify( ); // notify waiter that we've done it
 // continue doing stuff
 }
 synchronized void relatedMethod( ) {
 // do some related stuff
 }
 ...
 }


Let's assume waiter gets through the gate first and begins executing waiterMethod( ). The two other threads are initially blocked, trying to acquire the lock for the MyThing object. When waiter executes the wait( ) method, it relinquishes its hold on the lock and goes to sleep. Now, two viable threads are waiting for the lock. Which thread gets it depends on several factors, including chance and the priorities of the threads. (We'll discuss thread scheduling in the next section.) Let's say that notifier is the next thread to acquire the lock, so it begins to run notifierMethod( ). waiter continues to sleep, and related languishes, waiting for its turn. When notifier executes the call to notify( ), the runtime system prods the waiter thread, effectively telling it something has changed. waiter wakes up and rejoins related in vying for the MyThing lock. It doesn't receive the lock automatically; it just changes from saying, "Leave me alone" to "I want the lock." At this point, notifier still owns the lock and continues to hold it until the synchronized notifierMethod( ) returnsor perhaps executes a wait( ) itself. At that point, the other two methods get to fight over the lock. waiter would like to continue executing waiterMethod( ) from the point it left off, while related, which has been patient, would like to get started. We'll let you choose your own ending for the story. For each call to notify( ), the runtime system wakes up just one thread that is asleep in a wait( ) call. The group of threads waiting on a lock is called the wait set. If multiple threads are waiting, Java picks a thread on an arbitrary basis, which may be implementation-dependent. The Object class also provides a notifyAll( ) call to wake up all waiting threads. In most cases, you'll probably want to use notifyAll( ) rather than notify( ). Keep in mind that notify( ) really means, "Hey, something related to this object has changed. The condition you are waiting for may have changed, so check it again." In general, there is no reason to assume only one thread at a time is interested in the change or able to act upon it. Different threads might look upon whatever has changed in different ways.

Wait conditions

In general, our waiter thread is waiting for a particular condition to change, and we will want it to sit in a loop like the following:

 while ( condition != true )
 wait( );


This test is called the wait condition. Other synchronized threads call notify( ) or notifyAll( ) when they have modified the environment so that the condition can be checked again. It's important to use a loop on the wait condition to be sure that the thread has been awakened for the right reason. Threads may also use a timed version of wait( ) to do periodic work while checking the condition in this way. Using wait conditions like this is also an alternative to polling and sleeping, as you'll see in the following section.

Passing Messages

We'll next illustrate a classic interaction between two threads: a Producer and a Consumer. A producer thread creates messages and places them into a queue while a consumer reads and displays them. To be realistic, we'll give the queue a maximum depth. And to make things really interesting, we'll have our consumer thread be lazy and run much more slowly than the producer. This means that Producer occasionally has to stop and wait for Consumer to catch up. In Java 5.0, a BlockingQueue interface provides exactly this kind of functionality. But we'll build it ourselves here using basic synchronization techniques first and then take a look at Queues and all of the collection classes in . Here are the Producer and Consumer classes:

 import java.util.*;
 public class Consumer implements Runnable {
 Producer producer;
 Consumer( Producer producer ) {
 this.producer = producer;
 }
 public void run( ) {
 while ( true ) {
 String message = producer.getMessage( );
 System.out.println("Got message: " + message);
 try {
 Thread.sleep( 2000 );
 } catch ( InterruptedException e ) { }
 }
 }
 public static void main(String args[]) {
 Producer producer = new Producer( );
 new Thread( producer ).start( );
 Consumer consumer = new Consumer( producer );
 new Thread( consumer ).start( );
 }
 }
 public class Producer implements Runnable{
 static final int MAXQUEUE = 5;
 private List messages = new ArrayList( );
 public void run( ) {
 while ( true ) {
 putMessage( );
 try {
 Thread.sleep( 1000 );
 } catch ( InterruptedException e ) { }
 }
 }
 private synchronized void putMessage( )
 {
 while ( messages.size( ) >= MAXQUEUE )
 try {
 wait( );
 } catch( InterruptedException e ) { }
 messages.add( new java.util.Date( ).toString( ) );
 notify( );
 }
 // called by Consumer
 public synchronized String getMessage( )
 {
 while ( messages.size( ) == 0 )
 try {
 notify( );
 wait( );
 } catch( InterruptedException e ) { }
 String message = (String)messages.remove(0);
 notify( );
 return message;
 }
 }


For convenience, we have included a main( ) method in the Consumer class that runs the complete example. It creates a Consumer that is tied to a Producer and starts the two classes. You can run the example as follows:

 % java Consumer


This produces the timestamp messages created by the Producer:

 Got message: Sun Dec 19 03:35:55 CST 2006
 Got message: Sun Dec 19 03:35:56 CST 2006
 Got message: Sun Dec 19 03:35:57 CST 2006
 ...


The timestamps initially show a spacing of one second, although they appear every two seconds. Our Producer runs faster than our Consumer. Producer would like to generate a new message every second, while Consumer gets around to reading and displaying a message only every two seconds. Can you see how long it will take the message queue to fill up? What will happen when it does? Let's look at the code. We are using a few new tools here. Producer and Consumer implement the Runnable interface, and each has a thread associated with it. The Producer and Consumer classes pass messages through an instance of a java.util.List object. We haven't discussed the List class yet, but it is essentially a dynamic array of elements. We use this one as a queue by simply adding and removing elements in first-in, first-out order. The List has no maximum capacity of its own, but we impose one with our own check. The important activity is in the synchronized methods: putMessage( ) and getMessage( ). Although one of the methods is used by the Producer thread and the other by the Consumer thread, they both live in the Producer class so that we can coordinate them simply by declaring them synchronized. Here, they both implicitly use the Producer object's lock. If the queue is empty, the Consumer blocks in a call in the Producer, waiting for another message. Another design option would implement the getMessage( ) method in the Consumer class and use a synchronized code block to synchronize explicitly on the Producer object. In either case, synchronizing on the Producer enables us to have multiple Consumer objects that feed on the same Producer. We'll do that later in this section. putMessage( )'s job is to add a new message to the queue. It can't do this if the queue is already full, so it first checks the number of elements in messages. If there is room, it stuffs in another timestamp message. If the queue is at its limit, however, putMessage( ) has to wait until there's space. In this situation, putMessage( ) executes a wait( ) and relies on the consumer to call notify( ) to wake it up after a message has been read. Here, we have putMessage( ) testing the condition in a loop. In this simple example, the test might not seem necessary; we could assume that when putMessage( ) wakes up, there is a free spot. However, it's important to always test our wait condition in a loop like this when we synchronize threads because there is no other way to be certain why our thread has been awakened. Before it finishes, putMessage( ) calls notify( ) itself to prod any Consumer that might be waiting on an empty queue. getMessage( ) retrieves a message for the Consumer. It enters a loop like that of putMessage( ), waiting for the queue to have at least one element before proceeding. If the queue is empty, it executes a wait( ) and expects the Producer to call notify( ) when more items are available. Notice that getMessage( ) makes its own calls to notify( ). It does this any time the queue is empty, to prod a producer that might be sleeping and also after it consumes a message, to give the producer the go ahead to fill the queue again. These scenarios are more plausible if there are more consumers, as we'll see next. Let's add another consumer to the scenario, just to make things more interesting. Most of the necessary changes are in the Consumer class; here's the code for the modified class, now called NamedConsumer:

 public class NamedConsumer implements Runnable
 {
 Producer producer;
 String name;
 NamedConsumer(String name, Producer producer) {
 this.producer = producer;
 this.name = name;
 }
 public void run( ) {
 while ( true ) {
 String message = producer.getMessage( );
 System.out.println(name + " got message: " + message);
 try {
 Thread.sleep( 2000 );
 } catch ( InterruptedException e ) { }
 }
 }
 public static void main(String args[]) {
 Producer producer = new Producer( );
 new Thread( producer ).start( );
 NamedConsumer consumer = new NamedConsumer( "One", producer );
 new Thread( consumer ).start( );
 consumer = new NamedConsumer( "Two", producer );
 new Thread( consumer ).start( );
 }
 }


The NamedConsumer constructor takes a string name to identify each consumer. The run( ) method uses this name in the call to println( ) to identify which consumer received the message. The only required modification to the Producer code is to change the notify( ) calls to notifyAll( ) calls in putMessage( ) and getMessage( ). (We could have used notifyAll( ) in the first place.) Now, instead of the consumer and producer playing tag with the queue, we can have many players waiting for the condition of the queue to change. We might have a number of consumers waiting for a message, or we might have the producer waiting for a consumer to take a message. Whenever the condition of the queue changes, we prod all the waiting methods to reevaluate the situation by calling notifyAll( ). Here is some sample output when two NamedConsumers are running, as in the main( ) method shown previously:

 One got message: Sat Mar 18 20:00:01 CST 2006
 Two got message: Sat Mar 18 20:00:02 CST 2006
 One got message: Sat Mar 18 20:00:03 CST 2006
 Two got message: Sat Mar 18 20:00:04 CST 2006
 One got message: Sat Mar 18 20:00:05 CST 2006
 Two got message: Sat Mar 18 20:00:06 CST 2006
 One got message: Sat Mar 18 20:00:07 CST 2006
 Two got message: Sat Mar 18 20:00:08 CST 2006
 ...


We see nice, orderly alternation between the two consumers as a result of the calls to sleep( ) in the various methods. Interesting things would happen, however, if we were to remove all calls to sleep( ) and let things run at full speed. The threads would compete, and their behavior would depend on whether the system is using time-slicing. On a time-sliced system, there should be a fairly random distribution between the two consumers while on a non-time-sliced system, a single consumer could monopolize the messages. We'll talk shortly about how threads compete for time when we discuss thread priority and scheduling.

Food for thought

Many things could be improved in this simple example. What we've tried to emphasize is a defensive style of coding with respect to notifications by threads. That is, you need to rely on real-world conditions that you can test when synchronizing threads; it's not robust to simply assume that you'll get the right notifications in the right place at the right time. With that said, our example does generate extraneous notifications that wake up threads at times when there may not be work for them. For example, we generate notifications both when the queue is empty and when it's full. A better design might split these cases and use two different object locks. Fortunately, most programmers won't have to deal with issues at this level, especially since Java 5.0 has introduced real Queues and other high-level synchronization constructs.

ThreadLocal Objects

A common issue that arises is the need to maintain some information or state on a per-thread basis. For example, we might want to carry some context with the current thread as it executes our app. Or we might simply want to have a value that is different for different threads in the same way that each thread "sees" its own local variables in a method. Java supports this through the ThreadLocal class. A ThreadLocal is an object wrapper that automatically maintains a separate value for any thread calling it. For example:

 ThreadLocal userID = new ThreadLocal( );
 userID.set("Pat"); // called by thread 1
 userID.set("Bob"); // called by thread 2
 userID.get( ); // thread 1 gets "Pat"
 userID.get( ); // thread 2 gets "Bob"


You can use an instance of ThreadLocal anywhere you might use a static or instance variable to automatically maintain separate values for each thread. You can also extend ThreadLocal and override its initialValue( ) method. The ThreadLocal will then use this method to initialize its value once, the first time get( ) is called:

 class MyThreadLocalFactory extends ThreadLocal<Factory> {
 protected Factory initialValue( ) { return new MyFactory( ); }
 }


ThreadLocals are implemented using a Map attached to each Thread instance, so their values will disappear when the Thread is no longer used and garbage collected.

Comments