Skip to end of metadata
Go to start of metadata

Bounded thread pools allow the programmer to specify an upper limit on the number of threads that can concurrently execute in a thread pool. Programs must not use threads from a bounded thread pool to execute tasks that depend on the completion of other tasks in the pool.

A form of deadlock called thread-starvation deadlock arises when all the threads executing in the pool are blocked on tasks that are waiting on an internal queue for an available thread in which to execute. Thread-starvation deadlock occurs when currently executing tasks submit other tasks to a thread pool and wait for them to complete and the thread pool lacks the capacity to accommodate all the tasks at once.

This problem can be confusing because the program can function correctly when fewer threads are needed. The issue can be mitigated, in some cases, by choosing a larger pool size. However, determining a suitable size may be difficult or even impossible.

Similarly, threads in a thread pool may fail to be recycled when two executing tasks each require the other to complete before they can terminate. A blocking operation within a subtask can also lead to unbounded queue growth [Goetz 2006].

Noncompliant Code Example (Interdependent Subtasks)

This noncompliant code example is vulnerable to thread-starvation deadlock. It consists of the ValidationService class, which performs various input validation tasks such as checking whether a user-supplied field exists in a back-end database.

The fieldAggregator() method accepts a variable number of String arguments and creates a task corresponding to each argument to enable concurrent processing. The task performs input validation using the ValidateInput class.

In turn, the ValidateInput class attempts to sanitize the input by creating a subtask for each request using the SanitizeInput class. All tasks are executed in the same thread pool. The fieldAggregator() method blocks until all the tasks have finished executing and, when all results are available, returns the aggregated results as a StringBuilder object to the caller.

public final class ValidationService {
  private final ExecutorService pool;

  public ValidationService(int poolSize) {
    pool = Executors.newFixedThreadPool(poolSize);
  }

  public void shutdown() {
    pool.shutdown();
  }

  public StringBuilder fieldAggregator(String... inputs)
      throws InterruptedException, ExecutionException {

    StringBuilder sb = new StringBuilder();
    // Stores the results
    Future<String>[] results = new Future[inputs.length]; 

    // Submits the tasks to thread pool
    for (int i = 0; i < inputs.length; i++) { 
      results[i] = pool.submit(
        new ValidateInput<String>(inputs[i], pool));
    }

    for (int i = 0; i < inputs.length; i++) { // Aggregates the results
      sb.append(results[i].get());
    }
    return sb;
  }
}

public final class ValidateInput<V> implements Callable<V> {
  private final V input;
  private final ExecutorService pool;

  ValidateInput(V input, ExecutorService pool) {
    this.input = input;
    this.pool = pool;
  }

  @Override public V call() throws Exception {
    // If validation fails, throw an exception here
    // Subtask
    Future<V> future = pool.submit(new SanitizeInput<V>(input)); 
    return (V) future.get();
  }
}

public final class SanitizeInput<V> implements Callable<V> {
  private final V input;

  SanitizeInput(V input) {
    this.input = input;
  }

  @Override public V call() throws Exception {
    // Sanitize input and return
    return (V) input;
  }
}

Assume, for example, that the pool size is set to 6. The ValidationService.fieldAggregator() method is invoked to validate six arguments; consequently, it submits six tasks to the thread pool. Each task submits a corresponding subtask to sanitize the input. The SanitizeInput subtasks must execute before the original six tasks can return their results. However, this is impossible because all six threads in the thread pool are blocked. Furthermore, the shutdown() method cannot shut down the thread pool when it contains active tasks.

Thread-starvation deadlock can also occur when a single-threaded Executor is used, for example, when the caller creates several subtasks and waits for the results.

Compliant Solution (No Interdependent Tasks)

This compliant solution modifies the ValidateInput<V> class so that the SanitizeInput tasks are executed in the same threads as the ValidateInput tasks rather than in separate threads. Consequently, the ValidateInput and SanitizeInput tasks are independent, which eliminates their need to wait for each other to complete. The SanitizeInput class has also been modified to omit implementation of the Callable interface.

public final class ValidationService {
  // ...
  public StringBuilder fieldAggregator(String... inputs)
      throws InterruptedException, ExecutionException {
    // ...
    for (int i = 0; i < inputs.length; i++) {
      // Don't pass-in thread pool  
      results[i] = pool.submit(new ValidateInput<String>(inputs[i])); 
    }
    // ...
  }
}

// Does not use same thread pool
public final class ValidateInput<V> implements Callable<V> { 
  private final V input;

  ValidateInput(V input) {
    this.input = input;
  }

  @Override public V call() throws Exception {
    // If validation fails, throw an exception here
    return (V) new SanitizeInput().sanitize(input);
  }
}

public final class SanitizeInput<V> {  // No longer a Callable task
  public SanitizeInput() {}

  public V sanitize(V input) {
    // Sanitize input and return
    return input;
  }
}

Thread-starvation issues can be partially mitigated by choosing a large thread pool size. However, an untrusted caller can still overwhelm the system by supplying more inputs (see TPS00-J. Use thread pools to enable graceful degradation of service during traffic bursts).

Note that operations that have further constraints, such as the total number of database connections or total ResultSet objects open at a particular time, impose an upper bound on the usable thread pool size, as each thread continues to block until the resource becomes available.

Private static ThreadLocal variables may be used to maintain local state in each thread. When using thread pools, the lifetime of ThreadLocal variables should be bounded by the corresponding task [Goetz 2006]. Furthermore, programs must not use these variables to communicate between tasks. There are additional constraints in the use of ThreadLocal variables in thread pools (see TPS04-J. Ensure ThreadLocal variables are reinitialized when using thread pools for more information).

Noncompliant Code Example (Subtasks)

This noncompliant code example contains a series of subtasks that execute in a shared thread pool [Gafter 2006]. The BrowserManager class calls perUser(), which starts tasks that invoke perProfile(). The perProfile() method starts tasks that invoke perTab(), and in turn, perTab starts tasks that invoke doSomething(). BrowserManager then waits for the tasks to finish. The threads are allowed to invoke doSomething() in any order, provided that count correctly records the number of methods executed.

public final class BrowserManager {
  private final ExecutorService pool = Executors.newFixedThreadPool(10);
  private final int numberOfTimes;
  private static AtomicInteger count = new AtomicInteger(); // count = 0

  public BrowserManager(int n) {
    numberOfTimes = n;
  }

  public void perUser() {
    methodInvoker(numberOfTimes, "perProfile");
    pool.shutdown();
  }

  public void perProfile() {
    methodInvoker(numberOfTimes, "perTab");
  }

  public void perTab() {
    methodInvoker(numberOfTimes, "doSomething");
  }

  public void doSomething() {
    System.out.println(count.getAndIncrement());
  }

  public void methodInvoker(int n, final String method) {
    final BrowserManager manager = this;
    Callable<Object> callable = new Callable<Object>() {
      @Override public Object call() throws Exception {
        Method meth = manager.getClass().getMethod(method);
        return meth.invoke(manager);
      }
    };

    Collection<Callable<Object>> collection = 
        Collections.nCopies(n, callable);
    try {
      Collection<Future<Object>> futures = pool.invokeAll(collection);
    } catch (InterruptedException e) {
      // Forward to handler
      Thread.currentThread().interrupt(); // Reset interrupted status
    }
    // ...
  }

  public static void main(String[] args) {
    BrowserManager manager = new BrowserManager(5);
    manager.perUser();
  }
}

Unfortunately, this program is susceptible to a thread-starvation deadlock. For example, if each of the five perUser tasks spawns five perProfile tasks, where each perProfile task spawns a perTab task, the thread pool will be exhausted, and perTab() will be unable to allocate any additional threads to invoke the doSomething() method.

Compliant Solution (CallerRunsPolicy)

This compliant solution selects and schedules tasks for execution, avoiding thread-starvation deadlock. It sets the CallerRunsPolicy on a ThreadPoolExecutor and uses a SynchronousQueue [Gafter 2006]. The policy dictates that when the thread pool runs out of available threads, any subsequent tasks will run in the thread that submitted the tasks.

public final class BrowserManager {
  private final static ThreadPoolExecutor pool =
      new ThreadPoolExecutor(0, 10, 60L, TimeUnit.SECONDS,
                             new SynchronousQueue<Runnable>());
  private final int numberOfTimes;
  private static AtomicInteger count = new AtomicInteger(); // count = 0

  static {
    pool.setRejectedExecutionHandler(
    new ThreadPoolExecutor.CallerRunsPolicy());
  }

  // ...
}

According to Goetz and colleagues [Goetz 2006]:

A SynchronousQueue is not really a queue at all, but a mechanism for managing handoffs between threads. In order to put an element on the SynchronousQueue, another thread must already be waiting to accept the handoff. If no thread is waiting, but the current pool size is less than the maximum, ThreadPoolExecutor creates a new thread; otherwise, the task is rejected according to the saturation policy.

According to the Java API [API 2014], the CallerRunsPolicy class is

a handler for rejected tasks that runs the rejected task directly in the calling thread of the execute method, unless the executor has been shut down, in which case, the task is discarded.

In this compliant solution, tasks that have other tasks waiting to accept the hand-off are added to the SynchronousQueue when the thread pool is full. For example, tasks corresponding to perTab() are added to the SynchronousQueue because the tasks corresponding to perProfile() are waiting to receive the hand-off. Once the pool is full, additional tasks are rejected according to the saturation policy in effect. Because the CallerRunsPolicy is used to handle these rejected tasks, all the rejected tasks are executed in the main thread that started the initial tasks. When all the threads corresponding to perTab() have finished executing, the next set of tasks corresponding to perProfile() are added to the SynchronousQueue because the hand-off is subsequently used by perUser() tasks.

The CallerRunsPolicy allows graceful degradation of service when faced with many requests by distributing the workload from the thread pool to the work queue. Because the submitted tasks cannot block for any reason other than waiting for other tasks to complete, the policy guarantees that the current thread can handle multiple tasks sequentially. The policy would fail to prevent thread-starvation deadlock if the tasks were to block for some other reason, such as network I/O. Furthermore, this approach avoids unbounded queue growth because SynchronousQueue avoids storing tasks indefinitely for future execution, and all tasks are handled either by the current thread or by a thread in the thread pool.

This compliant solution is subject to the vagaries of the thread scheduler, which might schedule the tasks suboptimally. However, it avoids thread-starvation deadlock.

Risk Assessment

Executing interdependent tasks in a thread pool can lead to denial of service.

Rule

Severity

Likelihood

Remediation Cost

Priority

Level

TPS01-J

Low

Probable

Medium

P4

L3

Bibliography

[API 2014]

Class CallerRunsPolicy

[Gafter 2006]

A Thread Pool Puzzler

[Goetz 2006]

Section 5.3.3, "Dequeues and Work Stealing"
Section 8.3.2, "Managing Queued Tasks"
Section 8.3.3, "Saturation Policies"

 


2 Comments

  1. While I did a lot of work on this rule, lots more is needed:

    • Strictly speaking, a bounded thread pool is OK as long as you ensure that pool size > max # of threads waiting on completion of other threads. But prohibiting interdependent tasks in bounded thread pools is simpler and enforceable.

    Thread starvation issues can be mitigated by choosing a large pool size, however, the limitations of the application when using this approach should be clearly documented.

    • This seems to be saying "you can violate this rule as long as you document it". If so, it should go as an exception, and not just buried under the CS's.
    • The last CS doesn't seem to address the issue. AFAICT the ThreadPoolExecutor it uses is also bound to 10 threads maximum. I don't see why changing the CallerRunsPolicy prevents deadlock. I believe the available solutions are (as in the previous example):
      1. Each thread gets its own fixed threadpool to complete its work.
      2. Use an unbounded thread pool.
      3. Don't use thread pools at all (go back to a loopNTimes method).

    The text to the last CS is also wrong, but that's a minor nit at this point.

    EDIT: We could just get rid of the 2nd NCCE&CS; as I don't think it adds anything new to the rule.

      • Can be added as an exception
      • Same as above, same point
      • Link can be added. Iam curretly unsure if this text will stay here anyway.
      • Work stealing strategy, that's how it works. I think it's neat
      • We can discuss in person wrt what you think is incorrect in the CS

      Other comments:

      • Generifying the String was a good idea as I'd considered it too but I failed to justify why an Integer or Double type would need to be "sanitized".