Start line:  
End line:  

Snippet Preview

Snippet HTML Code

Stack Overflow Questions
   // Revision 1.57
   
   /*
    * Written by Doug Lea with assistance from members of JCP JSR-166
    * Expert Group and released to the public domain, as explained at
    * http://creativecommons.org/publicdomain/zero/1.0/
    */
   
   package org.infinispan.commons.util.concurrent.jdk8backported;
  
  import java.util.Arrays;
  import java.util.List;
An java.util.concurrent.ExecutorService for running ForkJoinTasks. A ForkJoinPool provides the entry point for submissions from non-ForkJoinTask clients, as well as management and monitoring operations.

A ForkJoinPool differs from other kinds of java.util.concurrent.ExecutorService mainly by virtue of employing work-stealing: all threads in the pool attempt to find and execute tasks submitted to the pool and/or created by other active tasks (eventually blocking waiting for work if none exist). This enables efficient processing when most tasks spawn other subtasks (as do most ForkJoinTasks), as well as when many small tasks are submitted to the pool from external clients. Especially when setting asyncMode to true in constructors, ForkJoinPools may also be appropriate for use with event-style tasks that are never joined.

A static commonPool() is available and appropriate for most applications. The common pool is used by any ForkJoinTask that is not explicitly submitted to a specified pool. Using the common pool normally reduces resource usage (its threads are slowly reclaimed during periods of non-use, and reinstated upon subsequent use).

For applications that require separate or custom pools, a ForkJoinPool may be constructed with a given target parallelism level; by default, equal to the number of available processors. The pool attempts to maintain enough active (or available) threads by dynamically adding, suspending, or resuming internal worker threads, even if some tasks are stalled waiting to join others. However, no such adjustments are guaranteed in the face of blocked I/O or other unmanaged synchronization. The nested ForkJoinPool.ManagedBlocker interface enables extension of the kinds of synchronization accommodated.

In addition to execution and lifecycle control methods, this class provides status check methods (for example getStealCount()) that are intended to aid in developing, tuning, and monitoring fork/join applications. Also, method toString() returns indications of pool state in a convenient form for informal monitoring.

As is the case with other ExecutorServices, there are three main task execution methods summarized in the following table. These are designed to be used primarily by clients not already engaged in fork/join computations in the current pool. The main forms of these methods accept instances of ForkJoinTask, but overloaded forms also allow mixed execution of plain Runnable- or Callable- based activities as well. However, tasks that are already executing in a pool should normally instead use the within-computation forms listed in the table unless using async event-style tasks that are not usually joined, in which case there is little difference among choice of methods.

Call from non-fork/join clients Call from within fork/join computations
Arrange async execution execute(org.infinispan.commons.util.concurrent.jdk8backported.ForkJoinTask) ForkJoinTask.fork()
Await and obtain result invoke(org.infinispan.commons.util.concurrent.jdk8backported.ForkJoinTask) ForkJoinTask.invoke()
Arrange exec and obtain Future submit(org.infinispan.commons.util.concurrent.jdk8backported.ForkJoinTask) ForkJoinTask.fork() (ForkJoinTasks are Futures)

The common pool is by default constructed with default parameters, but these may be controlled by setting three java.lang.System.getProperty(java.lang.String) with prefix java.util.concurrent.ForkJoinPool.common: parallelism -- an integer greater than zero, threadFactory -- the class name of a ForkJoinPool.ForkJoinWorkerThreadFactory, and exceptionHandler -- the class name of a java.lang.Thread.UncaughtExceptionHandler. Upon any error in establishing these settings, default parameters are used.

Implementation notes: This implementation restricts the maximum number of running threads to 32767. Attempts to create pools with greater than the maximum number result in IllegalArgumentException.

This implementation rejects submitted tasks (that is, by throwing java.util.concurrent.RejectedExecutionException) only when the pool is shut down or internal resources have been exhausted.

Author(s):
Doug Lea
Since:
1.7
 
 public class ForkJoinPool extends AbstractExecutorService {
 
     /*
      * Implementation Overview
      *
      * This class and its nested classes provide the main
      * functionality and control for a set of worker threads:
      * Submissions from non-FJ threads enter into submission queues.
      * Workers take these tasks and typically split them into subtasks
      * that may be stolen by other workers.  Preference rules give
      * first priority to processing tasks from their own queues (LIFO
      * or FIFO, depending on mode), then to randomized FIFO steals of
      * tasks in other queues.
      *
      * WorkQueues
      * ==========
      *
      * Most operations occur within work-stealing queues (in nested
      * class WorkQueue).  These are special forms of Deques that
      * support only three of the four possible end-operations -- push,
      * pop, and poll (aka steal), under the further constraints that
      * push and pop are called only from the owning thread (or, as
      * extended here, under a lock), while poll may be called from
      * other threads.  (If you are unfamiliar with them, you probably
      * want to read Herlihy and Shavit's book "The Art of
      * Multiprocessor programming", chapter 16 describing these in
      * more detail before proceeding.)  The main work-stealing queue
      * design is roughly similar to those in the papers "Dynamic
      * Circular Work-Stealing Deque" by Chase and Lev, SPAA 2005
      * (http://research.sun.com/scalable/pubs/index.html) and
      * "Idempotent work stealing" by Michael, Saraswat, and Vechev,
      * PPoPP 2009 (http://portal.acm.org/citation.cfm?id=1504186).
      * The main differences ultimately stem from GC requirements that
      * we null out taken slots as soon as we can, to maintain as small
      * a footprint as possible even in programs generating huge
      * numbers of tasks. To accomplish this, we shift the CAS
      * arbitrating pop vs poll (steal) from being on the indices
      * ("base" and "top") to the slots themselves.  So, both a
      * successful pop and poll mainly entail a CAS of a slot from
      * non-null to null.  Because we rely on CASes of references, we
      * do not need tag bits on base or top.  They are simple ints as
      * used in any circular array-based queue (see for example
      * ArrayDeque).  Updates to the indices must still be ordered in a
      * way that guarantees that top == base means the queue is empty,
      * but otherwise may err on the side of possibly making the queue
      * appear nonempty when a push, pop, or poll have not fully
      * committed. Note that this means that the poll operation,
      * considered individually, is not wait-free. One thief cannot
      * successfully continue until another in-progress one (or, if
      * previously empty, a push) completes.  However, in the
      * aggregate, we ensure at least probabilistic non-blockingness.
      * If an attempted steal fails, a thief always chooses a different
      * random victim target to try next. So, in order for one thief to
      * progress, it suffices for any in-progress poll or new push on
      * any empty queue to complete. (This is why we normally use
      * method pollAt and its variants that try once at the apparent
      * base index, else consider alternative actions, rather than
      * method poll.)
      *
      * This approach also enables support of a user mode in which local
      * task processing is in FIFO, not LIFO order, simply by using
      * poll rather than pop.  This can be useful in message-passing
      * frameworks in which tasks are never joined.  However neither
      * mode considers affinities, loads, cache localities, etc, so
      * rarely provide the best possible performance on a given
      * machine, but portably provide good throughput by averaging over
      * these factors.  (Further, even if we did try to use such
      * information, we do not usually have a basis for exploiting it.
      * For example, some sets of tasks profit from cache affinities,
      * but others are harmed by cache pollution effects.)
      *
      * WorkQueues are also used in a similar way for tasks submitted
      * to the pool. We cannot mix these tasks in the same queues used
      * for work-stealing (this would contaminate lifo/fifo
      * processing). Instead, we randomly associate submission queues
      * with submitting threads, using a form of hashing.  The
      * ThreadLocal Submitter class contains a value initially used as
      * a hash code for choosing existing queues, but may be randomly
      * repositioned upon contention with other submitters.  In
      * essence, submitters act like workers except that they are
      * restricted to executing local tasks that they submitted (or in
      * the case of CountedCompleters, others with the same root task).
      * However, because most shared/external queue operations are more
      * expensive than internal, and because, at steady state, external
      * submitters will compete for CPU with workers, ForkJoinTask.join
      * and related methods disable them from repeatedly helping to
      * process tasks if all workers are active.  Insertion of tasks in
      * shared mode requires a lock (mainly to protect in the case of
      * resizing) but we use only a simple spinlock (using bits in
      * field qlock), because submitters encountering a busy queue move
      * on to try or create other queues -- they block only when
      * creating and registering new queues.
      *
      * Management
      * ==========
      *
      * The main throughput advantages of work-stealing stem from
      * decentralized control -- workers mostly take tasks from
      * themselves or each other. We cannot negate this in the
      * implementation of other management responsibilities. The main
      * tactic for avoiding bottlenecks is packing nearly all
      * essentially atomic control state into two volatile variables
      * that are by far most often read (not written) as status and
      * consistency checks.
      *
      * Field "ctl" contains 64 bits holding all the information needed
      * to atomically decide to add, inactivate, enqueue (on an event
      * queue), dequeue, and/or re-activate workers.  To enable this
      * packing, we restrict maximum parallelism to (1<<15)-1 (which is
      * far in excess of normal operating range) to allow ids, counts,
      * and their negations (used for thresholding) to fit into 16bit
      * fields.
      *
      * Field "plock" is a form of sequence lock with a saturating
      * shutdown bit (similarly for per-queue "qlocks"), mainly
      * protecting updates to the workQueues array, as well as to
      * enable shutdown.  When used as a lock, it is normally only very
      * briefly held, so is nearly always available after at most a
      * brief spin, but we use a monitor-based backup strategy to
      * block when needed.
      *
      * Recording WorkQueues.  WorkQueues are recorded in the
      * "workQueues" array that is created upon first use and expanded
      * if necessary.  Updates to the array while recording new workers
      * and unrecording terminated ones are protected from each other
      * by a lock but the array is otherwise concurrently readable, and
      * accessed directly.  To simplify index-based operations, the
      * array size is always a power of two, and all readers must
      * tolerate null slots. Worker queues are at odd indices. Shared
      * (submission) queues are at even indices, up to a maximum of 64
      * slots, to limit growth even if array needs to expand to add
      * more workers. Grouping them together in this way simplifies and
      * speeds up task scanning.
      *
      * All worker thread creation is on-demand, triggered by task
      * submissions, replacement of terminated workers, and/or
      * compensation for blocked workers. However, all other support
      * code is set up to work with other policies.  To ensure that we
      * do not hold on to worker references that would prevent GC, ALL
      * accesses to workQueues are via indices into the workQueues
      * array (which is one source of some of the messy code
      * constructions here). In essence, the workQueues array serves as
      * a weak reference mechanism. Thus for example the wait queue
      * field of ctl stores indices, not references.  Access to the
      * workQueues in associated methods (for example signalWork) must
      * both index-check and null-check the IDs. All such accesses
      * ignore bad IDs by returning out early from what they are doing,
      * since this can only be associated with termination, in which
      * case it is OK to give up.  All uses of the workQueues array
      * also check that it is non-null (even if previously
      * non-null). This allows nulling during termination, which is
      * currently not necessary, but remains an option for
      * resource-revocation-based shutdown schemes. It also helps
      * reduce JIT issuance of uncommon-trap code, which tends to
      * unnecessarily complicate control flow in some methods.
      *
      * Event Queuing. Unlike HPC work-stealing frameworks, we cannot
      * let workers spin indefinitely scanning for tasks when none can
      * be found immediately, and we cannot start/resume workers unless
      * there appear to be tasks available.  On the other hand, we must
      * quickly prod them into action when new tasks are submitted or
      * generated. In many usages, ramp-up time to activate workers is
      * the main limiting factor in overall performance (this is
      * compounded at program start-up by JIT compilation and
      * allocation). So we try to streamline this as much as possible.
      * We park/unpark workers after placing in an event wait queue
      * when they cannot find work. This "queue" is actually a simple
      * Treiber stack, headed by the "id" field of ctl, plus a 15bit
      * counter value (that reflects the number of times a worker has
      * been inactivated) to avoid ABA effects (we need only as many
      * version numbers as worker threads). Successors are held in
      * field WorkQueue.nextWait.  Queuing deals with several intrinsic
      * races, mainly that a task-producing thread can miss seeing (and
      * signalling) another thread that gave up looking for work but
      * has not yet entered the wait queue. We solve this by requiring
      * a full sweep of all workers (via repeated calls to method
      * scan()) both before and after a newly waiting worker is added
      * to the wait queue. During a rescan, the worker might release
      * some other queued worker rather than itself, which has the same
      * net effect. Because enqueued workers may actually be rescanning
      * rather than waiting, we set and clear the "parker" field of
      * WorkQueues to reduce unnecessary calls to unpark.  (This
      * requires a secondary recheck to avoid missed signals.)  Note
      * the unusual conventions about Thread.interrupts surrounding
      * parking and other blocking: Because interrupts are used solely
      * to alert threads to check termination, which is checked anyway
      * upon blocking, we clear status (using Thread.interrupted)
      * before any call to park, so that park does not immediately
      * return due to status being set via some other unrelated call to
      * interrupt in user code.
      *
      * Signalling.  We create or wake up workers only when there
      * appears to be at least one task they might be able to find and
      * execute. However, many other threads may notice the same task
      * and each signal to wake up a thread that might take it. So in
      * general, pools will be over-signalled.  When a submission is
      * added or another worker adds a task to a queue that has fewer
      * than two tasks, they signal waiting workers (or trigger
      * creation of new ones if fewer than the given parallelism level
      * -- signalWork), and may leave a hint to the unparked worker to
      * help signal others upon wakeup).  These primary signals are
      * buttressed by others (see method helpSignal) whenever other
      * threads scan for work or do not have a task to process.  On
      * most platforms, signalling (unpark) overhead time is noticeably
      * long, and the time between signalling a thread and it actually
      * making progress can be very noticeably long, so it is worth
      * offloading these delays from critical paths as much as
      * possible.
      *
      * Trimming workers. To release resources after periods of lack of
      * use, a worker starting to wait when the pool is quiescent will
      * time out and terminate if the pool has remained quiescent for a
      * given period -- a short period if there are more threads than
      * parallelism, longer as the number of threads decreases. This
      * will slowly propagate, eventually terminating all workers after
      * periods of non-use.
      *
      * Shutdown and Termination. A call to shutdownNow atomically sets
      * a plock bit and then (non-atomically) sets each worker's
      * qlock status, cancels all unprocessed tasks, and wakes up
      * all waiting workers.  Detecting whether termination should
      * commence after a non-abrupt shutdown() call requires more work
      * and bookkeeping. We need consensus about quiescence (i.e., that
      * there is no more work). The active count provides a primary
      * indication but non-abrupt shutdown still requires a rechecking
      * scan for any workers that are inactive but not queued.
      *
      * Joining Tasks
      * =============
      *
      * Any of several actions may be taken when one worker is waiting
      * to join a task stolen (or always held) by another.  Because we
      * are multiplexing many tasks on to a pool of workers, we can't
      * just let them block (as in Thread.join).  We also cannot just
      * reassign the joiner's run-time stack with another and replace
      * it later, which would be a form of "continuation", that even if
      * possible is not necessarily a good idea since we sometimes need
      * both an unblocked task and its continuation to progress.
      * Instead we combine two tactics:
      *
      *   Helping: Arranging for the joiner to execute some task that it
      *      would be running if the steal had not occurred.
      *
      *   Compensating: Unless there are already enough live threads,
      *      method tryCompensate() may create or re-activate a spare
      *      thread to compensate for blocked joiners until they unblock.
      *
      * A third form (implemented in tryRemoveAndExec) amounts to
      * helping a hypothetical compensator: If we can readily tell that
      * a possible action of a compensator is to steal and execute the
      * task being joined, the joining thread can do so directly,
      * without the need for a compensation thread (although at the
      * expense of larger run-time stacks, but the tradeoff is
      * typically worthwhile).
      *
      * The ManagedBlocker extension API can't use helping so relies
      * only on compensation in method awaitBlocker.
      *
      * The algorithm in tryHelpStealer entails a form of "linear"
      * helping: Each worker records (in field currentSteal) the most
      * recent task it stole from some other worker. Plus, it records
      * (in field currentJoin) the task it is currently actively
      * joining. Method tryHelpStealer uses these markers to try to
      * find a worker to help (i.e., steal back a task from and execute
      * it) that could hasten completion of the actively joined task.
      * In essence, the joiner executes a task that would be on its own
      * local deque had the to-be-joined task not been stolen. This may
      * be seen as a conservative variant of the approach in Wagner &
      * Calder "Leapfrogging: a portable technique for implementing
      * efficient futures" SIGPLAN Notices, 1993
      * (http://portal.acm.org/citation.cfm?id=155354). It differs in
      * that: (1) We only maintain dependency links across workers upon
      * steals, rather than use per-task bookkeeping.  This sometimes
      * requires a linear scan of workQueues array to locate stealers,
      * but often doesn't because stealers leave hints (that may become
      * stale/wrong) of where to locate them.  It is only a hint
      * because a worker might have had multiple steals and the hint
      * records only one of them (usually the most current).  Hinting
      * isolates cost to when it is needed, rather than adding to
      * per-task overhead.  (2) It is "shallow", ignoring nesting and
      * potentially cyclic mutual steals.  (3) It is intentionally
      * racy: field currentJoin is updated only while actively joining,
      * which means that we miss links in the chain during long-lived
      * tasks, GC stalls etc (which is OK since blocking in such cases
      * is usually a good idea).  (4) We bound the number of attempts
      * to find work (see MAX_HELP) and fall back to suspending the
      * worker and if necessary replacing it with another.
      *
      * Helping actions for CountedCompleters are much simpler: Method
      * helpComplete can take and execute any task with the same root
      * as the task being waited on. However, this still entails some
      * traversal of completer chains, so is less efficient than using
      * CountedCompleters without explicit joins.
      *
      * It is impossible to keep exactly the target parallelism number
      * of threads running at any given time.  Determining the
      * existence of conservatively safe helping targets, the
      * availability of already-created spares, and the apparent need
      * to create new spares are all racy, so we rely on multiple
      * retries of each.  Compensation in the apparent absence of
      * helping opportunities is challenging to control on JVMs, where
      * GC and other activities can stall progress of tasks that in
      * turn stall out many other dependent tasks, without us being
      * able to determine whether they will ever require compensation.
      * Even though work-stealing otherwise encounters little
      * degradation in the presence of more threads than cores,
      * aggressively adding new threads in such cases entails risk of
      * unwanted positive feedback control loops in which more threads
      * cause more dependent stalls (as well as delayed progress of
      * unblocked threads to the point that we know they are available)
      * leading to more situations requiring more threads, and so
      * on. This aspect of control can be seen as an (analytically
      * intractable) game with an opponent that may choose the worst
      * (for us) active thread to stall at any time.  We take several
      * precautions to bound losses (and thus bound gains), mainly in
      * methods tryCompensate and awaitJoin.
      *
      * Common Pool
      * ===========
      *
      * The static common Pool always exists after static
      * initialization.  Since it (or any other created pool) need
      * never be used, we minimize initial construction overhead and
      * footprint to the setup of about a dozen fields, with no nested
      * allocation. Most bootstrapping occurs within method
      * fullExternalPush during the first submission to the pool.
      *
      * When external threads submit to the common pool, they can
      * perform some subtask processing (see externalHelpJoin and
      * related methods).  We do not need to record whether these
      * submissions are to the common pool -- if not, externalHelpJoin
      * returns quickly (at the most helping to signal some common pool
      * workers). These submitters would otherwise be blocked waiting
      * for completion, so the extra effort (with liberally sprinkled
      * task status checks) in inapplicable cases amounts to an odd
      * form of limited spin-wait before blocking in ForkJoinTask.join.
      *
      * Style notes
      * ===========
      *
      * There is a lot of representation-level coupling among classes
      * ForkJoinPool, ForkJoinWorkerThread, and ForkJoinTask.  The
      * fields of WorkQueue maintain data structures managed by
      * ForkJoinPool, so are directly accessed.  There is little point
      * trying to reduce this, since any associated future changes in
      * representations will need to be accompanied by algorithmic
      * changes anyway. Several methods intrinsically sprawl because
      * they must accumulate sets of consistent reads of volatiles held
      * in local variables.  Methods signalWork() and scan() are the
      * main bottlenecks, so are especially heavily
      * micro-optimized/mangled.  There are lots of inline assignments
      * (of form "while ((local = field) != 0)") which are usually the
      * simplest way to ensure the required read orderings (which are
      * sometimes critical). This leads to a "C"-like style of listing
      * declarations of these locals at the heads of methods or blocks.
      * There are several occurrences of the unusual "do {} while
      * (!cas...)"  which is the simplest way to force an update of a
      * CAS'ed variable. There are also other coding oddities (including
      * several unnecessary-looking hoisted null checks) that help
      * some methods perform reasonably even when interpreted (not
      * compiled).
      *
      * The order of declarations in this file is:
      * (1) Static utility functions
      * (2) Nested (static) classes
      * (3) Static fields
      * (4) Fields, along with constants used when unpacking some of them
      * (5) Internal control methods
      * (6) Callbacks and other support for ForkJoinTask methods
      * (7) Exported methods
      * (8) Static block initializing statics in minimally dependent order
      */
 
    // Static utilities
 
   
If there is a security manager, makes sure caller has permission to modify threads.
 
    private static void checkPermission() {
       SecurityManager security = System.getSecurityManager();
       if (security != null)
          security.checkPermission();
    }
 
    // Nested classes
 
   
Factory for creating new ForkJoinWorkerThreads. A ForkJoinWorkerThreadFactory must be defined and used for ForkJoinWorkerThread subclasses that extend base functionality or initialize threads with different contexts.
 
    public static interface ForkJoinWorkerThreadFactory {
      
Returns a new worker thread operating in the given pool.

Parameters:
pool the pool this thread works in
Throws:
java.lang.NullPointerException if the pool is null
 
       public ForkJoinWorkerThread newThread(ForkJoinPool pool);
    }

   
Default ForkJoinWorkerThreadFactory implementation; creates a new ForkJoinWorkerThread.
 
    static final class DefaultForkJoinWorkerThreadFactory
          implements ForkJoinWorkerThreadFactory {
       public final ForkJoinWorkerThread newThread(ForkJoinPool pool) {
          return new ForkJoinWorkerThread(pool);
       }
    }

   
Per-thread records for threads that submit to pools. Currently holds only pseudo-random seed / index that is used to choose submission queues in method externalPush. In the future, this may also incorporate a means to implement different task rejection and resubmission policies. Seeds for submitters and workers/workQueues work in basically the same way but are initialized and updated using slightly different mechanics. Both are initialized using the same approach as in class ThreadLocal, where successive values are unlikely to collide with previous values. Seeds are then randomly modified upon collisions using xorshifts, which requires a non-zero seed.
 
    static final class Submitter {
       int seed;
       Submitter(int s) {  = s; }
    }

   
Class for artificial tasks that are used to replace the target of local joins if they are removed from an interior queue slot in WorkQueue.tryRemoveAndExec. We don't need the proxy to actually do anything beyond having a unique identity.
 
    static final class EmptyTask extends ForkJoinTask<Void> {
       private static final long serialVersionUID = -7721805057305804111L;
       EmptyTask() {  = .; } // force done
       public final Void getRawResult() { return null; }
       public final void setRawResult(Void x) {}
       public final boolean exec() { return true; }
    }

   
Queues supporting work-stealing as well as external task submission. See above for main rationale and algorithms. Implementation relies heavily on "Unsafe" intrinsics and selective use of "volatile": Field "base" is the index (mod array.length) of the least valid queue slot, which is always the next position to steal (poll) from if nonempty. Reads and writes require volatile orderings but not CAS, because updates are only performed after slot CASes. Field "top" is the index (mod array.length) of the next queue slot to push to or pop from. It is written only by owner thread for push, or under lock for external/shared push, and accessed by other threads only after reading (volatile) base. Both top and base are allowed to wrap around on overflow, but (top - base) (or more commonly -(base - top) to force volatile read of base before top) still estimates size. The lock ("qlock") is forced to -1 on termination, causing all further lock attempts to fail. (Note: we don't need CAS for termination state because upon pool shutdown, all shared-queues will stop being used anyway.) Nearly all lock bodies are set up so that exceptions within lock bodies are "impossible" (modulo JVM errors that would cause failure anyway.) The array slots are read and written using the emulation of volatiles/atomics provided by Unsafe. Insertions must in general use putOrderedObject as a form of releasing store to ensure that all writes to the task object are ordered before its publication in the queue. All removals entail a CAS to null. The array is always a power of two. To ensure safety of Unsafe array operations, all accesses perform explicit null checks and implicit bounds checks via power-of-two masking. In addition to basic queuing support, this class contains fields described elsewhere to control execution. It turns out to work better memory-layout-wise to include them in this class rather than a separate class. Performance on most platforms is very sensitive to placement of instances of both WorkQueues and their arrays -- we absolutely do not want multiple WorkQueue instances or multiple queue arrays sharing cache lines. (It would be best for queue objects and their arrays to share, but there is nothing available to help arrange that). Unfortunately, because they are recorded in a common array, WorkQueue instances are often moved to be adjacent by garbage collectors. To reduce impact, we use field padding that works OK on common platforms; this effectively trades off slightly slower average field access for the sake of avoiding really bad worst-case access. (Until better JVM support is in place, this padding is dependent on transient properties of JVM field layout rules.) We also take care in allocating, sizing and resizing the array. Non-shared queue arrays are initialized by workers before use. Others are allocated on first use.
 
    static final class WorkQueue {
      
Capacity of work-stealing queue array upon initialization. Must be a power of two; at least 4, but should be larger to reduce or eliminate cacheline sharing among queues. Currently, it is much larger, as a partial workaround for the fact that JVMs often place arrays in locations that share GC bookkeeping (especially cardmarks) such that per-write accesses encounter serious memory contention.
 
       static final int INITIAL_QUEUE_CAPACITY = 1 << 13;

      
Maximum size for queue arrays. Must be a power of two less than or equal to 1 << (31 - width of array entry) to ensure lack of wraparound of index calculations, but defined to a value a bit less than this to help users trap runaway programs before saturating systems.
 
       static final int MAXIMUM_QUEUE_CAPACITY = 1 << 26; // 64M
 
       // Heuristic padding to ameliorate unfortunate memory placements
       volatile long pad00pad01pad02pad03pad04pad05pad06;
 
       int seed;                  // for random scanning; initialize nonzero
       volatile int eventCount;   // encoded inactivation count; < 0 if inactive
       int nextWait;              // encoded record of next event waiter
       int hint;                  // steal or signal hint (index)
       int poolIndex;             // index of this queue in pool (or 0)
       final int mode;            // 0: lifo, > 0: fifo, < 0: shared
       int nsteals;               // number of steals
       volatile int qlock;        // 1: locked, -1: terminate; else 0
       volatile int base;         // index of next slot for poll
       int top;                   // index of next slot for push
       ForkJoinTask<?>[] array;   // the elements (initially unallocated)
       final ForkJoinPool pool;   // the containing pool (may be null)
       final ForkJoinWorkerThread owner// owning thread or null if shared
       volatile Thread parker;    // == owner during call to park; else null
       volatile ForkJoinTask<?> currentJoin;  // task being joined in awaitJoin
       ForkJoinTask<?> currentSteal// current non-local task being executed
 
       volatile Object pad10pad11pad12pad13pad14pad15pad16pad17;
       volatile Object pad18pad19pad1apad1bpad1cpad1d;
 
       WorkQueue(ForkJoinPool poolForkJoinWorkerThread ownerint mode,
             int seed) {
          this. = pool;
          this. = owner;
          this. = mode;
          this. = seed;
          // Place indices in the center of array (that is not yet allocated)
           =  =  >>> 1;
       }

      
Returns the approximate number of tasks in the queue.
 
       final int queueSize() {
          int n =  - ;       // non-owner callers must read base first
          return (n >= 0) ? 0 : -n// ignore transient negative
       }

      
Provides a more accurate estimate of whether this queue has any tasks than does queueSize, by checking whether a near-empty queue has at least one unclaimed task.
 
       final boolean isEmpty() {
          ForkJoinTask<?>[] aint ms;
          int n =  - (s = );
          return (n >= 0 ||
                        (n == -1 &&
                               ((a = ) == null ||
                                      (m = a.length - 1) < 0 ||
                                      .getObject
                                            (a, (long)((m & (s - 1)) << ) + ) == null)));
       }

      
Pushes a task. Call only by owner in unshared queues. (The shared-queue version is embedded in method externalPush.)

Parameters:
task the task. Caller must ensure non-null.
Throws:
java.util.concurrent.RejectedExecutionException if array cannot be resized
 
       final void push(ForkJoinTask<?> task) {
          ForkJoinTask<?>[] aForkJoinPool p;
          int s = mn;
          if ((a = ) != null) {    // ignore if queue removed
             int j = (((m = a.length - 1) & s) << ) + ;
             .putOrderedObject(ajtask);
             if ((n = ( = s + 1) - ) <= 2) {
                if ((p = ) != null)
                   p.signalWork(this);
             }
             else if (n >= m)
                growArray();
          }
       }

      
Initializes or doubles the capacity of array. Call either by owner or with lock held -- it is OK for base, but not top, to move while resizings are in progress.
 
       final ForkJoinTask<?>[] growArray() {
          ForkJoinTask<?>[] oldA = ;
          int size = oldA != null ? oldA.length << 1 : ;
          if (size > )
             throw new RejectedExecutionException("Queue capacity exceeded");
          int oldMasktb;
          ForkJoinTask<?>[] a =  = new ForkJoinTask<?>[size];
          if (oldA != null && (oldMask = oldA.length - 1) >= 0 &&
                (t = ) - (b = ) > 0) {
             int mask = size - 1;
             do {
                ForkJoinTask<?> x;
                int oldj = ((b & oldMask) << ) + ;
                int j    = ((b &    mask) << ) + ;
                x = (ForkJoinTask<?>).getObjectVolatile(oldAoldj);
                if (x != null &&
                      .compareAndSwapObject(oldAoldjxnull))
                   .putObjectVolatile(ajx);
             } while (++b != t);
          }
          return a;
       }

      
Takes next task, if one exists, in LIFO order. Call only by owner in unshared queues.
 
       final ForkJoinTask<?> pop() {
          ForkJoinTask<?>[] aForkJoinTask<?> tint m;
          if ((a = ) != null && (m = a.length - 1) >= 0) {
             for (int s; (s =  - 1) -  >= 0;) {
                long j = ((m & s) << ) + ;
                if ((t = (ForkJoinTask<?>).getObject(aj)) == null)
                   break;
                if (.compareAndSwapObject(ajtnull)) {
                    = s;
                   return t;
                }
             }
          }
          return null;
       }

      
Takes a task in FIFO order if b is base of queue and a task can be claimed without contention. Specialized versions appear in ForkJoinPool methods scan and tryHelpStealer.
 
       final ForkJoinTask<?> pollAt(int b) {
          ForkJoinTask<?> t; ForkJoinTask<?>[] a;
          if ((a = ) != null) {
             int j = (((a.length - 1) & b) << ) + ;
             if ((t = (ForkJoinTask<?>).getObjectVolatile(aj)) != null &&
                    == b &&
                   .compareAndSwapObject(ajtnull)) {
                 = b + 1;
                return t;
             }
          }
          return null;
       }

      
Takes next task, if one exists, in FIFO order.
 
       final ForkJoinTask<?> poll() {
          ForkJoinTask<?>[] aint bForkJoinTask<?> t;
          while ((b = ) -  < 0 && (a = ) != null) {
             int j = (((a.length - 1) & b) << ) + ;
             t = (ForkJoinTask<?>).getObjectVolatile(aj);
             if (t != null) {
                if ( == b &&
                      .compareAndSwapObject(ajtnull)) {
                    = b + 1;
                   return t;
                }
             }
             else if ( == b) {
                if (b + 1 == )
                   break;
                Thread.yield(); // wait for lagging update (very rare)
             }
          }
          return null;
       }

      
Takes next task, if one exists, in order specified by mode.
 
       final ForkJoinTask<?> nextLocalTask() {
          return  == 0 ? pop() : poll();
       }

      
Returns next task, if one exists, in order specified by mode.
 
       final ForkJoinTask<?> peek() {
          ForkJoinTask<?>[] a = int m;
          if (a == null || (m = a.length - 1) < 0)
             return null;
          int i =  == 0 ?  - 1 : ;
          int j = ((i & m) << ) + ;
          return (ForkJoinTask<?>).getObjectVolatile(aj);
       }

      
Pops the given task only if it is at the current top. (A shared version is available only via FJP.tryExternalUnpush)
 
       final boolean tryUnpush(ForkJoinTask<?> t) {
          ForkJoinTask<?>[] aint s;
          if ((a = ) != null && (s = ) !=  &&
                .compareAndSwapObject
                      (a, (((a.length - 1) & --s) << ) + tnull)) {
              = s;
             return true;
          }
          return false;
       }

      
Removes and cancels all known tasks, ignoring any exceptions.
 
       final void cancelAll() {
          ForkJoinTask.cancelIgnoringExceptions();
          ForkJoinTask.cancelIgnoringExceptions();
          for (ForkJoinTask<?> t; (t = poll()) != null; )
             ForkJoinTask.cancelIgnoringExceptions(t);
       }

      
Computes next value for random probes. Scans don't require a very high quality generator, but also not a crummy one. Marsaglia xor-shift is cheap and works well enough. Note: This is manually inlined in its usages in ForkJoinPool to avoid writes inside busy scan loops.
 
       final int nextSeed() {
          int r = ;
          r ^= r << 13;
          r ^= r >>> 17;
          return  = r ^= r << 5;
       }
 
       // Specialized execution methods
 
      
Pops and runs tasks until empty.
 
       private void popAndExecAll() {
          // A bit faster than repeated pop calls
          ForkJoinTask<?>[] aint mslong jForkJoinTask<?> t;
          while ((a = ) != null && (m = a.length - 1) >= 0 &&
                (s =  - 1) -  >= 0 &&
                (t = ((ForkJoinTask<?>)
                            .getObject(aj = ((m & s) << ) + )))
                      != null) {
             if (.compareAndSwapObject(ajtnull)) {
                 = s;
                t.doExec();
             }
          }
       }

      
Polls and runs tasks until empty.
 
       private void pollAndExecAll() {
          for (ForkJoinTask<?> t; (t = poll()) != null;)
             t.doExec();
       }

      
If present, removes from queue and executes the given task, or any other cancelled task. Returns (true) on any CAS or consistency check failure so caller can retry.

Returns:
false if no progress can be made, else true
 
       final boolean tryRemoveAndExec(ForkJoinTask<?> task) {
          boolean stat = trueremoved = falseempty = true;
          ForkJoinTask<?>[] aint msbn;
          if ((a = ) != null && (m = a.length - 1) >= 0 &&
                (n = (s = ) - (b = )) > 0) {
             for (ForkJoinTask<?> t;;) {           // traverse from s to b
                int j = ((--s & m) << ) + ;
                t = (ForkJoinTask<?>).getObjectVolatile(aj);
                if (t == null)                    // inconsistent length
                   break;
                else if (t == task) {
                   if (s + 1 == ) {           // pop
                      if (!.compareAndSwapObject(ajtasknull))
                         break;
                       = s;
                      removed = true;
                   }
                   else if ( == b)           // replace with proxy
                      removed = .compareAndSwapObject(ajtask,
                            new EmptyTask());
                   break;
                }
                else if (t.status >= 0)
                   empty = false;
                else if (s + 1 == ) {          // pop and throw away
                   if (.compareAndSwapObject(ajtnull))
                       = s;
                   break;
                }
                if (--n == 0) {
                   if (!empty &&  == b)
                      stat = false;
                   break;
                }
             }
          }
          if (removed)
             task.doExec();
          return stat;
       }

      
Polls for and executes the given task or any other task in its CountedCompleter computation.
 
       final boolean pollAndExecCC(ForkJoinTask<?> root) {
          ForkJoinTask<?>[] aint bObject o;
          outer: while ((b = ) -  < 0 && (a = ) != null) {
             long j = (((a.length - 1) & b) << ) + ;
             if ((o = .getObject(aj)) == null ||
                   !(o instanceof CountedCompleter))
                break;
             for (CountedCompleter<?> t = (CountedCompleter<?>)or = t;;) {
                if (r == root) {
                   if ( == b &&
                         .compareAndSwapObject(ajtnull)) {
                       = b + 1;
                      t.doExec();
                      return true;
                   }
                   else
                      break// restart
                }
                if ((r = r.completer) == null)
                   break outer; // not part of root computation
             }
          }
          return false;
       }

      
Executes a top-level task and any local tasks remaining after execution.
 
       final void runTask(ForkJoinTask<?> t) {
          if (t != null) {
             ( = t).doExec();
              = null;
             ++;
             if ( -  < 0) {       // process remaining local tasks
                if ( == 0)
                   popAndExecAll();
                else
                   pollAndExecAll();
             }
          }
      }

      
Executes a non-top-level (stolen) task.
      final void runSubtask(ForkJoinTask<?> t) {
         if (t != null) {
            ForkJoinTask<?> ps = ;
            ( = t).doExec();
             = ps;
         }
      }

      
Returns true if owned and not known to be blocked.
      final boolean isApparentlyUnblocked() {
         Thread wtThread.State s;
         return ( >= 0 &&
                       (wt = ) != null &&
                       (s = wt.getState()) != .. &&
                       s != .. &&
                       s != ..);
      }
      // Unsafe mechanics
      private static final sun.misc.Unsafe U;
      private static final long QLOCK;
      private static final int ABASE;
      private static final int ASHIFT;
      static {
         try {
             = getUnsafe();
            Class<?> k = WorkQueue.class;
            Class<?> ak = ForkJoinTask[].class;
             = .objectFieldOffset
                  (k.getDeclaredField("qlock"));
             = .arrayBaseOffset(ak);
            int scale = .arrayIndexScale(ak);
            if ((scale & (scale - 1)) != 0)
               throw new Error("data type scale not a power of two");
             = 31 - Integer.numberOfLeadingZeros(scale);
         } catch (Exception e) {
            throw new Error(e);
         }
      }
   }
   // static fields (initialized in static initializer below)

   
Creates a new ForkJoinWorkerThread. This factory is used unless overridden in ForkJoinPool constructors.
   public static final ForkJoinWorkerThreadFactory
Per-thread submission bookkeeping. Shared across all pools to reduce ThreadLocal pollution and because random motion to avoid contention in one pool is likely to hold for others. Lazily initialized on first submission (but null-checked in other contexts to avoid unnecessary initialization).
   static final ThreadLocal<Submittersubmitters;

   
Permission required for callers of methods that may start or kill threads.
   private static final RuntimePermission modifyThreadPermission;

   
Common (static) pool. Non-null for public use unless a static construction exception, but internal usages null-check on use to paranoically avoid potential initialization circularities as well as to simplify generated code.
   static final ForkJoinPool common;

   
Common pool parallelism. Must equal common.parallelism.
   static final int commonParallelism;

   
Sequence number for creating workerNamePrefix.
   private static int poolNumberSequence;

   
Returns the next sequence number. We don't expect this to ever contend, so use simple builtin sync.
   private static final synchronized int nextPoolId() {
      return ++;
   }
   // static constants

   
Initial timeout value (in nanoseconds) for the thread triggering quiescence to park waiting for new work. On timeout, the thread will instead try to shrink the number of workers. The value should be large enough to avoid overly aggressive shrinkage during most transient stalls (long GCs etc).
   private static final long IDLE_TIMEOUT      = 2000L * 1000L * 1000L; // 2sec

   
Timeout value when there are more threads than parallelism level
   private static final long FAST_IDLE_TIMEOUT =  200L * 1000L * 1000L;

   
Tolerance for idle timeouts, to cope with timer undershoots
   private static final long TIMEOUT_SLOP = 2000000L;

   
The maximum stolen->joining link depth allowed in method tryHelpStealer. Must be a power of two. Depths for legitimate chains are unbounded, but we use a fixed constant to avoid (otherwise unchecked) cycles and to bound staleness of traversal parameters at the expense of sometimes blocking when we could be helping.
   private static final int MAX_HELP = 64;

   
Increment for seed generators. See class ThreadLocal for explanation.
   private static final int SEED_INCREMENT = 0x61c88647;
    /*
     * Bits and masks for control variables
     *
     * Field ctl is a long packed with:
     * AC: Number of active running workers minus target parallelism (16 bits)
     * TC: Number of total workers minus target parallelism (16 bits)
     * ST: true if pool is terminating (1 bit)
     * EC: the wait count of top waiting thread (15 bits)
     * ID: poolIndex of top of Treiber stack of waiters (16 bits)
     *
     * When convenient, we can extract the upper 32 bits of counts and
     * the lower 32 bits of queue state, u = (int)(ctl >>> 32) and e =
     * (int)ctl.  The ec field is never accessed alone, but always
     * together with id and st. The offsets of counts by the target
     * parallelism and the positionings of fields makes it possible to
     * perform the most common checks via sign tests of fields: When
     * ac is negative, there are not enough active workers, when tc is
     * negative, there are not enough total workers, and when e is
     * negative, the pool is terminating.  To deal with these possibly
     * negative fields, we use casts in and out of "short" and/or
     * signed shifts to maintain signedness.
     *
     * When a thread is queued (inactivated), its eventCount field is
     * set negative, which is the only way to tell if a worker is
     * prevented from executing tasks, even though it must continue to
     * scan for them to avoid queuing races. Note however that
     * eventCount updates lag releases so usage requires care.
     *
     * Field plock is an int packed with:
     * SHUTDOWN: true if shutdown is enabled (1 bit)
     * SEQ:  a sequence lock, with PL_LOCK bit set if locked (30 bits)
     * SIGNAL: set when threads may be waiting on the lock (1 bit)
     *
     * The sequence number enables simple consistency checks:
     * Staleness of read-only operations on the workQueues array can
     * be checked by comparing plock before vs after the reads.
     */
   // bit positions/shifts for fields
   private static final int  AC_SHIFT   = 48;
   private static final int  TC_SHIFT   = 32;
   private static final int  ST_SHIFT   = 31;
   private static final int  EC_SHIFT   = 16;
   // bounds
   private static final int  SMASK      = 0xffff;  // short bits
   private static final int  MAX_CAP    = 0x7fff;  // max #workers - 1
   private static final int  EVENMASK   = 0xfffe;  // even short bits
   private static final int  SQMASK     = 0x007e;  // max 64 (even) slots
   private static final int  SHORT_SIGN = 1 << 15;
   private static final int  INT_SIGN   = 1 << 31;
   // masks
   private static final long STOP_BIT   = 0x0001L << ;
   private static final long AC_MASK    = ((long)) << ;
   private static final long TC_MASK    = ((long)) << ;
   // units for incrementing and decrementing
   private static final long TC_UNIT    = 1L << ;
   private static final long AC_UNIT    = 1L << ;
   // masks and units for dealing with u = (int)(ctl >>> 32)
   private static final int  UAC_SHIFT  =  - 32;
   private static final int  UTC_SHIFT  =  - 32;
   private static final int  UAC_MASK   =  << ;
   private static final int  UTC_MASK   =  << ;
   private static final int  UAC_UNIT   = 1 << ;
   private static final int  UTC_UNIT   = 1 << ;
   // masks and units for dealing with e = (int)ctl
   private static final int E_MASK      = 0x7fffffff; // no STOP_BIT
   private static final int E_SEQ       = 1 << ;
   // plock bits
   private static final int SHUTDOWN    = 1 << 31;
   private static final int PL_LOCK     = 2;
   private static final int PL_SIGNAL   = 1;
   private static final int PL_SPINS    = 1 << 8;
   // access mode for WorkQueue
   static final int LIFO_QUEUE          =  0;
   static final int FIFO_QUEUE          =  1;
   static final int SHARED_QUEUE        = -1;
   // bounds for #steps in scan loop -- must be power 2 minus 1
   private static final int MIN_SCAN    = 0x1ff;   // cover estimation slop
   private static final int MAX_SCAN    = 0x1ffff; // 4 * max workers
   // Instance fields
    /*
     * Field layout of this class tends to matter more than one would
     * like. Runtime layout order is only loosely related to
     * declaration order and may differ across JVMs, but the following
     * empirically works OK on current JVMs.
     */
   // Heuristic padding to ameliorate unfortunate memory placements
   volatile long pad00pad01pad02pad03pad04pad05pad06;
   volatile long stealCount;                  // collects worker counts
   volatile long ctl;                         // main pool control
   volatile int plock;                        // shutdown status and seqLock
   volatile int indexSeed;                    // worker/submitter index seed
   final int config;                          // mode and parallelism level
   WorkQueue[] workQueues;                    // main registry
   final Thread.UncaughtExceptionHandler ueh// per-worker UEH
   final String workerNamePrefix;             // to create worker name string
   volatile Object pad10pad11pad12pad13pad14pad15pad16pad17;
   volatile Object pad18pad19pad1apad1b;

   
Acquires the plock lock to protect worker array and related updates. This method is called only if an initial CAS on plock fails. This acts as a spinlock for normal cases, but falls back to builtin monitor to block when (rarely) needed. This would be a terrible idea for a highly contended lock, but works fine as a more conservative alternative to a pure spinlock.
   private int acquirePlock() {
      int spins = r = 0, psnps;
      for (;;) {
         if (((ps = ) & ) == 0 &&
               .compareAndSwapInt(thispsnps = ps + ))
            return nps;
         else if (r == 0) { // randomize spins if possible
            Thread t = Thread.currentThread(); WorkQueue wSubmitter z;
            if ((t instanceof ForkJoinWorkerThread) &&
                  (w = ((ForkJoinWorkerThread)t).) != null)
               r = w.seed;
            else if ((z = .get()) != null)
               r = z.seed;
            else
               r = 1;
         }
         else if (spins >= 0) {
            r ^= r << 1; r ^= r >>> 3; r ^= r << 10; // xorshift
            if (r >= 0)
               --spins;
         }
         else if (.compareAndSwapInt(thispsps | )) {
            synchronized (this) {
               if (( & ) != 0) {
                  try {
                     wait();
                  } catch (InterruptedException ie) {
                     try {
                        Thread.currentThread().interrupt();
                     } catch (SecurityException ignore) {
                     }
                  }
               }
               else
                  notifyAll();
            }
         }
      }
   }

   
Unlocks and signals any thread waiting for plock. Called only when CAS of seq value for unlock fails.
   private void releasePlock(int ps) {
       = ps;
      synchronized (this) { notifyAll(); }
   }

   
Tries to create and start one worker if fewer than target parallelism level exist. Adjusts counts etc on failure.
   private void tryAddWorker() {
      long cint u;
      while ((u = (int)((c = ) >>> 32)) < 0 &&
            (u & ) != 0 && (int)c == 0) {
         long nc = (long)(((u + ) & ) |
                                ((u + ) & )) << 32;
         if (.compareAndSwapLong(thiscnc)) {
            ForkJoinWorkerThreadFactory fac;
            Throwable ex = null;
            ForkJoinWorkerThread wt = null;
            try {
               if ((fac = ) != null &&
                     (wt = fac.newThread(this)) != null) {
                  wt.start();
                  break;
               }
            } catch (Throwable e) {
               ex = e;
            }
            deregisterWorker(wtex);
            break;
         }
      }
   }
   //  Registering and deregistering workers

   
Callback from ForkJoinWorkerThread to establish and record its WorkQueue. To avoid scanning bias due to packing entries in front of the workQueues array, we treat the array as a simple power-of-two hash table using per-thread seed as hash, expanding as needed.

Parameters:
wt the worker thread
Returns:
the worker's queue
      Thread.UncaughtExceptionHandler handlerWorkQueue[] wsint sps;
      wt.setDaemon(true);
      if ((handler = ) != null)
         wt.setUncaughtExceptionHandler(handler);
      do {} while (!.compareAndSwapInt(thiss = ,
            s += ) ||
            s == 0); // skip 0
      WorkQueue w = new WorkQueue(thiswt >>> 16, s);
      if (((ps = ) & ) != 0 ||
            !.compareAndSwapInt(thispsps += ))
         ps = acquirePlock();
      int nps = (ps & ) | ((ps + ) & ~);
      try {
         if ((ws = ) != null) {    // skip if shutting down
            int n = ws.lengthm = n - 1;
            int r = (s << 1) | 1;           // use odd-numbered indices
            if (ws[r &= m] != null) {       // collision
               int probes = 0;             // step by approx half size
               int step = (n <= 4) ? 2 : ((n >>> 1) & ) + 2;
               while (ws[r = (r + step) & m] != null) {
                  if (++probes >= n) {
                      = ws = Arrays.copyOf(wsn <<= 1);
                     m = n - 1;
                     probes = 0;
                  }
               }
            }
            w.eventCount = w.poolIndex = r// volatile write orders
            ws[r] = w;
         }
      } finally {
         if (!.compareAndSwapInt(thispsnps))
            releasePlock(nps);
      }
      wt.setName(.concat(Integer.toString(w.poolIndex)));
      return w;
   }

   
Final callback from terminating worker, as well as upon failure to construct or start a worker. Removes record of worker from array, and adjusts counts. If pool is shutting down, tries to complete termination.

Parameters:
wt the worker thread or null if construction failed
ex the exception causing failure, or null if none
   final void deregisterWorker(ForkJoinWorkerThread wtThrowable ex) {
      WorkQueue w = null;
      if (wt != null && (w = wt.workQueue) != null) {
         int ps;
         w.qlock = -1;                // ensure set
         long ns = w.nstealssc;     // collect steal count
         do {} while (!.compareAndSwapLong(this,
               sc = sc + ns));
         if (((ps = ) & ) != 0 ||
               !.compareAndSwapInt(thispsps += ))
            ps = acquirePlock();
         int nps = (ps & ) | ((ps + ) & ~);
         try {
            int idx = w.poolIndex;
            WorkQueue[] ws = ;
            if (ws != null && idx >= 0 && idx < ws.length && ws[idx] == w)
               ws[idx] = null;
         } finally {
            if (!.compareAndSwapInt(thispsnps))