LCK10-J. Do not use incorrect forms of the double-checked locking idiom

Lazy initialization defers the construction of a member field or an object referred to by a member field until an instance is actually required rather than computing the field value or constructing the referenced object in the class's constructor. Lazy initialization helps to break harmful circularities in class and instance initialization . It also enables other optimizations [[Bloch 2005]].

Lazy initialization uses either a class or an instance method, depending on whether the member object is static. The method checks whether the instance has already been created and, if not, creates it. When the instance already exists, the method simply returns the instance:

// Correct single threaded version using lazy initialization
final class Foo {
  private Helper helper = null;

  public Helper getHelper() {
    if (helper == null) {
      helper = new Helper();
    }
    return helper;
  }
  // ...
}

Lazy initialization must be synchronized in multithreaded applications to prevent multiple threads from creating extraneous instances of the member object:

// Correct multithreaded version using synchronization
final class Foo {
  private Helper helper = null;

  public synchronized Helper getHelper() {
    if (helper == null) {
      helper = new Helper();
    }
    return helper;
  }
  // ...
}

The double-checked locking idiom improves performance by limiting synchronization to the rare case of computing the field's value or constructing a new instance for the field to reference and by foregoing synchronization during the common case of retrieving an already-created instance or value.

Incorrect forms of the double-checked locking idiom include those that allow publication of an uninitialized or partially initialized object. Consequently, only those forms of the double-checked locking idiom that correctly establish a happens-before relationship both for the helper reference and for the complete construction of the Helper instance are permitted.

Noncompliant Code Example

The double-checked locking pattern uses block synchronization rather than method synchronization and installs an additional null reference check before attempting synchronization. This noncompliant code example uses an incorrect form of the double-checked locking idiom.

// "Double-Checked Locking" idiom
final class Foo {
  private Helper helper = null;
  public Helper getHelper() {
    if (helper == null) {
      synchronized (this) {
        if (helper == null) {
          helper = new Helper();
        }
      }
    }
    return helper;
  }

  // Other methods and members...
}

According to Pugh [[Pugh 2004]]

Writes that initialize the Helper object and the write to the helper field can be done or perceived out of order. As a result, a thread which invokes getHelper() could see a non-null reference to a helper object, but see the default values for fields of the helper object, rather than the values set in the constructor.

Even if the compiler does not reorder those writes, on a multiprocessor, the processor or the memory system may reorder those writes, as perceived by a thread running on another processor.

This code also violates rule TSM03-J. Do not publish partially initialized objects.

Compliant Solution (Volatile)

This compliant solution declares the helper field volatile.

// Works with acquire/release semantics for volatile
// Broken under JDK 1.4 and earlier
final class Foo {
  private volatile Helper helper = null;

  public Helper getHelper() {
    if (helper == null) {
      synchronized (this) {
        if (helper == null) {
          helper = new Helper();
        }
      }
    }
    return helper;
  }
}

When a thread initializes the Helper object, a [happens-before relationship] is established between this thread and any other thread that retrieves and returns the instance [[Pugh 2004], [Manson 2004]].

Compliant Solution (Static Initialization)

This compliant solution initializes the helper field in the declaration of the static variable [[Manson 2006]].

final class Foo {
  private static final Helper helper = new Helper();

  public static Helper getHelper() {
    return helper;
  }
}

Variables that are declared static and initialized at declaration or from a static initializer are guaranteed to be fully constructed before being made visible to other threads. However, this solution forgoes the benefits of lazy initialization.

Compliant Solution (Initialize-on-Demand, Holder Class Idiom)

This compliant solution uses the initialize-on-demand, holder class idiom that implicitly incorporates lazy initialization by declaring a static variable within a static Holder inner class.

final class Foo {
  // Lazy initialization
  private static class Holder {
    static Helper helper = new Helper();
  }

  public static Helper getInstance() {
    return Holder.helper;
  }
}

Initialization of the static helper field is deferred until the getInstance() method is called. The necessary happens-before relationships are created by the combination of the class loader's actions loading and initializing the Holder instance and the guarantees provided by the Java memory model. This idiom is a better choice than the double-checked locking idiom for lazily initializing static fields [[Bloch 2008]]. However, this idiom cannot be used to lazily initialize instance fields [[Bloch 2001]].

Compliant Solution (ThreadLocal Storage)

This compliant solution (originally suggested by Alexander Terekhov [[Pugh 2004]]) uses a ThreadLocal object to track whether each individual thread has participated in the synchronization that creates the needed happens-before relationships. Each thread stores a non-null value into its thread-local perThreadInstance only inside the synchronized createHelper() method; consequently, any thread that sees a null value must establish the necessary happens-before relationships by invoking createHelper().

final class Foo {
  private final ThreadLocal<Foo> perThreadInstance = 
      new ThreadLocal<Foo>();
  private Helper helper = null;

  public Helper getHelper() {
    if (perThreadInstance.get() == null) {
      createHelper();
    }
    return helper;
  }

  private synchronized void createHelper() {
    if (helper == null) {
      helper = new Helper();
    }
    // Any non-null value can be used as an argument to set()
    perThreadInstance.set(this);
  }
}

Compliant Solution (Immutable)

In this compliant solution, suppose that the Helper class is immutable. The Java Memory Model (JMM) guarantees that immutable objects are fully constructed before they become visible to any other thread. Additionally, the block synchronization in the getHelper() method suffices to ensure that all methods that can see a non-null value of the helper field have a proper happens-before relationship for the update to the helper reference. This synchronization and the aforementioned JMM guarantee combine to ensure that only fully initialized Helper objects are visible to threads that see non-null values. Consequently, this compliant solution correctly creates both of the needed happens-before relationships.

public final class Helper {
  private final int n;

  public Helper(int n) {
    this.n = n;
  }

  // Other fields and methods, all fields are final
}

final class Foo {
  private Helper helper = null;

  public Helper getHelper() {
    if (helper == null) {
      synchronized (this) {
        if (helper == null) {
          helper = new Helper();
        }
      }
    }
    return helper;
  }
}

Exceptions

LCK10-EX0: Use of the noncompliant form of the double-checked locking idiom is permitted for 32-bit primitive values (for example, int or float) [[Pugh 2004]], although this usage is discouraged. The noncompliant form establishes the necessary happens-before relationship between threads that see an initialized version of the primitive value. The second happens-before relationship (for the initialization of the fields of the referent) is of no practical value because unsynchronized reads and writes of primitive values up to 32-bits are guaranteed to be atomic. Consequently, the noncompliant form establishes the only needed happens-before relationship in this case. Note, however, that the noncompliant form fails for long or double because unsynchronized reads or writes of 64-bit primitives lack a guarantee of atomicity and consequently require a second happens-before relationship to guarantee that all threads see only fully assigned 64-bit values (See rule VNA05-J. Ensure atomicity when reading and writing 64-bit values for more information.)

Risk Assessment

Using incorrect forms of the double-checked locking idiom can lead to synchronization problems and can expose partially initialized objects.

Related Guidelines

Bibliography

      08. Locking (LCK)