Wednesday, March 24, 2010

Singletons, Threads & Double Checked Locking

This blog entry extends on the first entry I wrote about Singletons. Today I try to explain the usage of Singletons in a multi-threaded environment.

A typical Singleton would look like the following:


This Singleton class would work perfectly fine in a single-threaded environment. But, how many Java based applications that you have worked on, are single-threaded? None, perhaps.

Can you see what’s terribly wrong with it when your application has multiple threads running simultaneously ??

Well, the answer lies in the way the instance is lazy initialized. The null-checking and initializing the instance need to be an atomic operation. Lets see what would happen if following sequence of events happen:

  • Thead A calls getInstance() for the first time. And checks whether instance is null or not. It finds it true and goes to execute line 11.
  • At this point, it is pre-empted by another thread: Thread B, which also calls getInstance(), checks for the instance null check, since it is not yet initialized, the private constructor is called and instance is now initialized with a state.
  • Thread A resumes control back and starts from where it left i.e line 11. Now the real problem occurs. Since it checked for the null-check on line 10 when it last run and found it to be true, so when it resumes execution, it will execute from line 11 and again initialize the instance and hence two objects would be created and the whole point of having a single object is violated.

Thus, we really need to do something about the situation. No problems, Synchronization comes to the rescue. Most of us (naive programmers) would tend to get rid of this problem by writing code like below:


The only change that occurs here is that we have made our static factory as synchronized method, which makes sure that no possible combination of interleaving of threads could create two objects like it did in the previous case. Although, this works fine but on close inspection, we find that synchronization is necessary only for the first time, and not needed later on (since instance will always be non-null later on) And synchronizing the whole method is thus uncalled for and pretty expensive. We must think of other ways of accomplishing this.

1. Do Not Lazy Initialize

If we do NOT need lazy initialization of the resource, then a perfect method to do away with synchronization would be:


Using this, we depend on the JVM’s capability to execute the static block as soon as the class is loaded first (when it is accessed or referred in code).

Thus, the VM makes sure that the instance is well initialized, before any thread tries to call getInstance() method. And since static blocks are executed only once, so no chance of creating more objects. One scenario that I could fail this approach is when this class is loaded again (perhaps by a different classloader in a different namespace)

Another way to write the above code would be:


However, I prefer the first way (with static blocks) that gives us more flexibility by allowing us to write a block of code instead of just calling the constructor.

2. Use Double Checked Locking

A very common, popular and misunderstood programming idiom that looks like the following and works with Java 1.5 and above only.


See how cleverly, we have limited the scope of synchronization here. The synchronization occurs only for the first time (when the instance is null), after that the synchronization never occurs as the instance is not null. This is what we needed. But you must make this variable as volatile to work it correctly (due to Java Memory Model implementation)

To better understand this programming idiom, I would suggest reading this and this.

I think its enough for one blog post. Any code corrections, new ideas and comments are welcome.

No comments: