The java.lang.ThreadLocal<T>
class provides thread-local variables. According to the Java API [API 2014]:
These variables differ from their normal counterparts in that each thread that accesses one (via its
get
orset
method) has its own, independently initialized copy of the variable.ThreadLocal
instances are typically private static fields in classes that wish to associate state with a thread (e.g., a user ID or transaction ID).
The use of ThreadLocal
objects requires care in classes whose objects are required to be executed by multiple threads in a thread pool. The technique of thread pooling allows threads to be reused to reduce thread creation overhead or when creating an unbounded number of threads can diminish the reliability of the system. Each task that enters the pool expects to see ThreadLocal
objects in their initial, default state. However, when ThreadLocal
objects are modified on a thread that is subsequently made available for reuse, the next task executing on the reused thread sees the state of the ThreadLocal
objects as modified by the previous task that executed on that thread [JPL 2006].
Programs must ensure that each task that executes on a thread from a thread pool sees only correctly initialized instances of ThreadLocal
objects.
Noncompliant Code Example
This noncompliant code example consists of an enumeration of days (Day
) and two classes (Diary
and DiaryPool
). The Diary
class uses a ThreadLocal
variable to store thread-specific information, such as each task's current day. The initial value of the current day is Monday; it can be changed later by invoking the setDay()
method. The class also contains a threadSpecificTask()
instance method that performs a thread-specific task.
The DiaryPool
class consists of the doSomething1()
and doSomething2()
methods that each start a thread. The doSomething1()
method changes the initial (default) value of the day to Friday and invokes threadSpecificTask()
. However, doSomething2()
relies on the initial value of the day (Monday) and invokes threadSpecificTask()
. The main()
method creates one thread using doSomething1()
and two more using doSomething2()
.
public enum Day { MONDAY, TUESDAY, WEDNESDAY, THURSDAY, FRIDAY, SATURDAY, SUNDAY; } public final class Diary { private static final ThreadLocal<Day> days = new ThreadLocal<Day>() { // Initialize to Monday protected Day initialValue() { return Day.MONDAY; } }; private static Day currentDay() { return days.get(); } public static void setDay(Day newDay) { days.set(newDay); } // Performs some thread-specific task public void threadSpecificTask() { // Do task ... } } public final class DiaryPool { final int numOfThreads = 2; // Maximum number of threads allowed in pool final Executor exec; final Diary diary; DiaryPool() { exec = (Executor) Executors.newFixedThreadPool(numOfThreads); diary = new Diary(); } public void doSomething1() { exec.execute(new Runnable() { @Override public void run() { diary.setDay(Day.FRIDAY); diary.threadSpecificTask(); } }); } public void doSomething2() { exec.execute(new Runnable() { @Override public void run() { diary.threadSpecificTask(); } }); } public static void main(String[] args) { DiaryPool dp = new DiaryPool(); dp.doSomething1(); // Thread 1, requires current day as Friday dp.doSomething2(); // Thread 2, requires current day as Monday dp.doSomething2(); // Thread 3, requires current day as Monday } }
The DiaryPool
class creates a thread pool that reuses a fixed number of threads operating off a shared, unbounded queue. At any point, no more than numOfThreads
threads are actively processing tasks. If additional tasks are submitted when all threads are active, they wait in the queue until a thread is available. The thread-local state of the thread persists when a thread is recycled.
The following table shows a possible execution order:
Time | Task | Pool Thread | Submitted by Method | Day |
---|---|---|---|---|
1 | t1 | 1 |
| Friday |
2 | t2 | 2 |
| Monday |
3 | t3 | 1 |
| Friday |
In this execution order, it is expected that the two tasks (t2 and t3) started from doSomething2()
would observe the current day as Monday. However, because pool thread 1 is reused, t3 observes the day to be Friday.
Noncompliant Code Example (Increase Thread Pool Size)
This noncompliant code example increases the size of the thread pool from two to three in an attempt to mitigate the issue:
public final class DiaryPool { final int numOfthreads = 3; // ... }
Although increasing the size of the thread pool resolves the problem for this example, it fails to scale because changing the thread pool size is insufficient if additional tasks can be submitted to the pool.
Compliant Solution (try-finally
Clause)
This compliant solution adds the removeDay()
method to the Diary
class and wraps the statements in the doSomething1()
method of class DiaryPool
in a try-finally
block. The finally
block restores the initial state of the thread-local days
object by removing the current thread's value from it.
public final class Diary { // ... public static void removeDay() { days.remove(); } } public final class DiaryPool { // ... public void doSomething1() { exec.execute(new Runnable() { @Override public void run() { try { Diary.setDay(Day.FRIDAY); diary.threadSpecificTask(); } finally { Diary.removeDay(); // Diary.setDay(Day.MONDAY) // can also be used } } }); } // ... }
If the thread-local variable is read by the same thread again, it is reinitialized using the initialValue()
method unless the task has already set the variable's value explicitly [API 2014]. This solution transfers the responsibility for maintenance to the client (DiaryPool
) but is a good option when the Diary
class cannot be modified.
Compliant Solution (beforeExecute()
)
This compliant solution uses a custom ThreadPoolExecutor
that extends ThreadPoolExecutor
and overrides the beforeExecute()
method. The beforeExecute()
method is invoked before the Runnable
task is executed in the specified thread. The method reinitializes the thread-local variable before task r
is executed by thread t
.
class CustomThreadPoolExecutor extends ThreadPoolExecutor { public CustomThreadPoolExecutor(int corePoolSize, int maximumPoolSize, long keepAliveTime, TimeUnit unit, BlockingQueue<Runnable> workQueue) { super(corePoolSize, maximumPoolSize, keepAliveTime, unit, workQueue); } @Override public void beforeExecute(Thread t, Runnable r) { if (t == null || r == null) { throw new NullPointerException(); } Diary.setDay(Day.MONDAY); super.beforeExecute(t, r); } } public final class DiaryPool { // ... DiaryPool() { exec = new CustomThreadPoolExecutor(NumOfthreads, NumOfthreads, 10, TimeUnit.SECONDS, new ArrayBlockingQueue<Runnable>(10)); diary = new Diary(); } // ... }
Exceptions
TPS04-J-EX0: It is unnecessary to reinitialize a ThreadLocal
object that does not change state after initialization. For example, there may be only one type of database connection represented by the initial value of the ThreadLocal
object.
Risk Assessment
Objects using ThreadLocal
data and executed by different tasks in a thread pool without reinitialization might be in an unexpected state when reused.
Rule | Severity | Likelihood | Remediation Cost | Priority | Level |
---|---|---|---|---|---|
TPS04-J | Medium | Probable | High | P4 | L3 |
Bibliography
4 Comments
David Svoboda
Two problems:
First, this rule seems a little like developer carelessness. I think using ThreadLocal objects in thread pools are perfectly safe if you do any of the following:
Presumably we could 'fix' the NCCE with three additional CS's, one for each solution I outline above.
Second, I think there are race conditions in the NCCE. Specifically I'm not convinced that the output arose in the same order the tasks were submitted. This could be fixed by adding a number to each task. It's not the order the output lines appear in that matters, but rather that they indicate exactly which task produced which output.
Dhruv Mohindra
Adding task numbers might not help either because on every run different task numbers will be associated with different days. Basically, we cannot control which task will go next into the thread pool. So, the text would still be incorrect because there are race conditions that might produce a different order. Doing what I've done helped me avoid the mess.
Robert Seacord (Manager)
It seems like there should be a compliant solution based on the
beforeExecute()
method:See http://java.sun.com/javase/6/docs/api/java/util/concurrent/ThreadPoolExecutor.html#beforeExecute%28java.lang.Thread,%20java.lang.Runnable%29
Also see http://java.sun.com/javase/6/docs/api/java/util/concurrent/Executors.html
privilegedThreadFactory
public static ThreadFactory privilegedThreadFactory()
Returns a thread factory used to create new threads that have the same permissions as the current thread. This factory creates threads with the same settings as defaultThreadFactory(), additionally setting the AccessControlContext and contextClassLoader of new threads to be the same as the thread invoking this privilegedThreadFactory method. A new privilegedThreadFactory can be created within an AccessController.doPrivileged(java.security.PrivilegedAction) action setting the current thread's access control context to create threads with the selected permission settings holding within that action.
Note that while tasks running within such threads will have the same access control and class loader settings as the current thread, they need not have the same ThreadLocal or InheritableThreadLocal values. If necessary, particular values of thread locals can be set or reset before any task runs in ThreadPoolExecutor subclasses using ThreadPoolExecutor.beforeExecute(java.lang.Thread, java.lang.Runnable). Also, if it is necessary to initialize worker threads to have the same InheritableThreadLocal settings as some other designated thread, you can create a custom ThreadFactory in which that thread waits for and services requests to create others that will inherit its values.
Returns:
a thread factory
Throws:
AccessControlException - if the current access control context does not have permission to both get and set context class loader.
Dhruv Mohindra
Fixed, please review. I think the following advice that you cited is good but currently beyond the scope of this guideline:
This guideline isn't getting into inheriting the state from another thread. For example, from
InheritableThreadLocal
's API docs:Also I think a CS or exception based on this would be tricky because this article (Goetz) says: