VNA03-J. Do not assume that a group of calls to independently atomic methods is atomic

A consistent locking policy guarantees that multiple threads cannot simultaneously access or modify shared data. When two or more operations must be performed as a single atomic operation, a consistent locking policy must be implemented using either intrinsic synchronization or java.util.concurrent utilities. In the absence of such a policy, the code is susceptible to race conditions.

When presented with a set of operations, where each is guaranteed to be atomic, it is tempting to assume that a single operation consisting of individually-atomic operations is guaranteed to be collectively atomic without additional locking. Similarly, programmers might incorrectly assume that use of a thread-safe Collection is sufficient to preserve an invariant that involves the collection's elements without additional synchronization. A thread-safe class can only guarantee atomicity of its individual methods. A grouping of calls to such methods requires additional synchronization for the group.

Consider, for example, a scenario where the standard thread-safe API lacks a single method to both find a particular person's record in a Hashtable and also update that person's payroll information. In such cases, the two method invocations must be performed atomically.

Enumerations and iterators also require either explicit synchronization on the collection object (client-side locking) or use of a private final lock object.

Compound operations on shared variables are also non-atomic. See rule VNA02-J. Ensure that compound operations on shared variables are atomic for more information.

Rule VNA04-J. Ensure that calls to chained methods are atomic describes a specialized case of this rule.

Noncompliant Code Example (AtomicReference)

This noncompliant code example wraps references to BigInteger objects within thread-safe AtomicReference objects.

final class Adder {
  private final AtomicReference<BigInteger> first;
  private final AtomicReference<BigInteger> second;

  public Adder(BigInteger f, BigInteger s) {
    first  = new AtomicReference<BigInteger>(f);
    second = new AtomicReference<BigInteger>(s);
  }

  public void update(BigInteger f, BigInteger s) { // Unsafe
    first.set(f);
    second.set(s);
  }

  public BigInteger add() { // Unsafe
    return first.get().add(second.get());
  }
}

AtomicReference is an object reference that can be updated atomically. However, operations that combine more than one atomic reference are non-atomic. In this noncompliant code example, one thread may call update() while a second thread may call add(). This might cause the add() method to add the new value of first to the old value of second, yielding an erroneous result.

Compliant Solution (Method Synchronization)

This compliant solution declares the update() and add() methods synchronized to guarantee atomicity.

final class Adder {
  // ...
  private final AtomicReference<BigInteger> first;
  private final AtomicReference<BigInteger> second;

  public Adder(BigInteger f, BigInteger s) {
    first  = new AtomicReference<BigInteger>(f);
    second = new AtomicReference<BigInteger>(s);
  }



  public synchronized void update(BigInteger f, BigInteger s){
    first.set(f);
    second.set(s);
  }

  public synchronized BigInteger add() {
    return first.get().add(second.get());
  }
}

Noncompliant Code Example (synchronizedList())

This noncompliant code example uses a java.util.ArrayList<E> collection, which is not thread-safe. However, the example uses Collections.synchronizedList as a synchronization wrapper for the ArrayList. It subsequently uses an array, rather than an iterator, to iterate over the ArrayList to avoid a ConcurrentModificationException.

final class IPHolder {
  private final List<InetAddress> ips = 
      Collections.synchronizedList(new ArrayList<InetAddress>());

  public void addAndPrintIPAddresses(InetAddress address) {
    ips.add(address);
    InetAddress[] addressCopy = 
        (InetAddress[]) ips.toArray(new InetAddress[0]);
    // Iterate through array addressCopy ...
  }
}

Individually, the add() and toArray() collection methods are atomic. However, when called in succession (as shown in the addAndPrintIPAddresses() method), there is no guarantee that the combined operation is atomic. The addAndPrintIPAddresses() method contains a race condition that allows one thread to add to the list and a second thread to race in and modify the list before the first thread completes. Consequently, the addressCopy array may contain more IP addresses than expected.

Compliant Solution (Synchronized Block)

The race condition can be eliminated by synchronizing on the underlying list's lock. This compliant solution encapsulates all references to the array list within synchronized blocks.

final class IPHolder {
  private final List<InetAddress> ips = 
      Collections.synchronizedList(new ArrayList<InetAddress>());

  public void addAndPrintIPAddresses(InetAddress address) {
    synchronized (ips) {
      ips.add(address);
      InetAddress[] addressCopy = 
          (InetAddress[]) ips.toArray(new InetAddress[0]);
      // Iterate through array addressCopy ...
    }
  }
}

This technique is also called client-side locking [Goetz 2006] because the class holds a lock on an object that might be accessible to other classes. Client-side locking is not always an appropriate strategy; see rule LCK11-J. Avoid client-side locking when using classes that do not commit to their locking strategy for more information.

This code does not violate rule LCK04-J. Do not synchronize on a collection view if the backing collection is accessible because, while it does synchronize on a collection view (the synchronizedList result), the backing collection is inaccessible and consequently cannot be modified by any code.

Note that this compliant solution does not actually use the synchronization offered by Collections.synchronizedList(). If no other code in this solution used it, it could be eliminated.

Noncompliant Code Example (synchronizedMap())

This noncompliant code example defines the KeyedCounter class that is not thread-safe. Although the HashMap is wrapped in a synchronizedMap(), the overall increment operation is not atomic [Lee 2009].

final class KeyedCounter {
  private final Map<String, Integer> map =
      Collections.synchronizedMap(new HashMap<String, Integer>());

  public void increment(String key) {
    Integer old = map.get(key);
    int oldValue = (old == null) ? 0 : old.intValue();
    if (oldValue == Integer.MAX_VALUE) {
      throw new ArithmeticException("Out of range");
    }
    map.put( key, oldValue + 1);
  }

  public Integer getCount(String key) {
    return map.get(key);
  }
}

Compliant Solution (Synchronization)

This compliant solution ensures atomicity by using an internal private lock object to synchronize the statements of the increment() and getCount() methods.

final class KeyedCounter {
  private final Map<String, Integer> map =
      new HashMap<String, Integer>();
  private final Object lock = new Object();

  public void increment(String key) {
    synchronized (lock) {
      Integer old = map.get(key);
      int oldValue = (old == null) ? 0 : old.intValue();
      if (oldValue == Integer.MAX_VALUE) {
        throw new ArithmeticException("Out of range");
      }
      map.put(key, oldValue + 1);
    }
  }

  public Integer getCount(String key) {
    synchronized (lock) {
      return map.get(key);
    }
  }
}

This compliant solution avoids using Collections.synchronizedMap() because locking on the unsynchronized map provides sufficient thread-safety for this application. Rule LCK04-J. Do not synchronize on a collection view if the backing collection is accessible provides more information about synchronizing on synchronizedMap() objects.

Compliant Solution (ConcurrentHashMap)

The previous compliant solution is safe for multithreaded use but does not scale because of excessive synchronization, which can lead to decreased performance.

The ConcurrentHashMap class used in this compliant solution provides several utility methods for performing atomic operations and is often a good choice for algorithms that must scale [Lee 2009].

Note that this compliant solution still requires synchronization, because without it, the test to prevent overflow and the increment will not happen atomically, so two threads calling increment() can still cause overflow. The synchronization block is smaller, and does not include the lookup or addition of new values, so it has less of an impact on performance as the previous compliant solution.

final class KeyedCounter {
  private final ConcurrentMap<String, AtomicInteger> map =
      new ConcurrentHashMap<String, AtomicInteger>();
  private final Object lock = new Object();

  public void increment(String key) {
    AtomicInteger value = new AtomicInteger();
    AtomicInteger old = map.putIfAbsent(key, value);

    if (old != null) {
      value = old;
    }

    synchronized (lock) {
      if (value.get() == Integer.MAX_VALUE) {
        throw new ArithmeticException("Out of range");
      }
      value.incrementAndGet(); // Increment the value atomically
    }
  }

  public Integer getCount(String key) {
    AtomicInteger value = map.get(key);
    return (value == null) ? null : value.get();
  }

  // Other accessors ...
}

According to Section 5.2.1., "ConcurrentHashMap" of the work of Goetz and colleagues [Goetz 2006]:

ConcurrentHashMap, along with the other concurrent collections, further improve on the synchronized collection classes by providing iterators that do not throw ConcurrentModificationException, as a result eliminating the need to lock the collection during iteration. The iterators returned by ConcurrentHashMap are weakly consistent instead of fail-fast. A weakly consistent iterator can tolerate concurrent modification, traverses elements as they existed when the iterator was constructed, and may (but is not guaranteed to) reflect modifications to the collection after the construction of the iterator.

Note that methods such as ConcurrentHashMap.size() and ConcurrentHashMap.isEmpty() are allowed to return an approximate result for performance reasons. Code should avoid relying on these return values when exact results are required.

Risk Assessment

Failure to ensure the atomicity of two or more operations that must be performed as a single atomic operation can result in race conditions in multithreaded applications.

Related Guidelines

Bibliography

      07. Visibility and Atomicity (VNA)