--- jsr166/src/jsr166y/ForkJoinPool.java 2011/02/22 10:50:51 1.92
+++ jsr166/src/jsr166y/ForkJoinPool.java 2012/10/31 12:49:24 1.139
@@ -1,7 +1,7 @@
/*
* 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/licenses/publicdomain
+ * http://creativecommons.org/publicdomain/zero/1.0/
*/
package jsr166y;
@@ -19,10 +19,9 @@ import java.util.concurrent.Future;
import java.util.concurrent.RejectedExecutionException;
import java.util.concurrent.RunnableFuture;
import java.util.concurrent.TimeUnit;
-import java.util.concurrent.TimeoutException;
import java.util.concurrent.atomic.AtomicInteger;
-import java.util.concurrent.locks.LockSupport;
-import java.util.concurrent.locks.ReentrantLock;
+import java.util.concurrent.atomic.AtomicLong;
+import java.util.concurrent.locks.AbstractQueuedSynchronizer;
import java.util.concurrent.locks.Condition;
/**
@@ -34,21 +33,35 @@ import java.util.concurrent.locks.Condit
* A {@code ForkJoinPool} differs from other kinds of {@link
* ExecutorService} mainly by virtue of employing
* work-stealing: all threads in the pool attempt to find and
- * execute subtasks 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 {@code
- * ForkJoinTask}s). When setting asyncMode to true in
- * constructors, {@code ForkJoinPool}s may also be appropriate for use
- * with event-style tasks that are never joined.
+ * 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 {@code ForkJoinTask}s), as well as when many small
+ * tasks are submitted to the pool from external clients. Especially
+ * when setting asyncMode to true in constructors, {@code
+ * ForkJoinPool}s may also be appropriate for use with event-style
+ * tasks that are never joined.
*
- *
A {@code ForkJoinPool} is 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 IO or other unmanaged synchronization. The nested
- * {@link ManagedBlocker} interface enables extension of the kinds of
+ *
A static {@link #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). The common pool is by default constructed with default
+ * parameters, but these may be controlled by setting any or all of
+ * the three properties {@code
+ * java.util.concurrent.ForkJoinPool.common.{parallelism,
+ * threadFactory, exceptionHandler}}.
+ *
+ *
For applications that require separate or custom pools, a {@code
+ * 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 IO or other unmanaged synchronization. The nested {@link
+ * ManagedBlocker} interface enables extension of the kinds of
* synchronization accommodated.
*
*
In addition to execution and lifecycle control methods, this
@@ -59,15 +72,16 @@ import java.util.concurrent.locks.Condit
* 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 by clients not already engaged
- * in fork/join computations in the current pool. The main forms of
- * these methods accept instances of {@code ForkJoinTask}, but
- * overloaded forms also allow mixed execution of plain {@code
+ * 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 {@code ForkJoinTask},
+ * but overloaded forms also allow mixed execution of plain {@code
* Runnable}- or {@code Callable}- based activities as well. However,
- * tasks that are already executing in a pool should normally
- * NOT use these pool execution methods, but instead use the
- * within-computation forms listed in the table.
+ * 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.
*
*
*
@@ -92,24 +106,6 @@ import java.util.concurrent.locks.Condit
*
*
*
- * Sample Usage. Normally a single {@code ForkJoinPool} is
- * used for all parallel task execution in a program or subsystem.
- * Otherwise, use would not usually outweigh the construction and
- * bookkeeping overhead of creating a large set of threads. For
- * example, a common pool could be used for the {@code SortTasks}
- * illustrated in {@link RecursiveAction}. Because {@code
- * ForkJoinPool} uses threads in {@linkplain java.lang.Thread#isDaemon
- * daemon} mode, there is typically no need to explicitly {@link
- * #shutdown} such a pool upon program exit.
- *
- *
- * static final ForkJoinPool mainPool = new ForkJoinPool();
- * ...
- * public void sort(long[] array) {
- * mainPool.invoke(new SortTask(array, 0, array.length));
- * }
- *
- *
* 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
@@ -127,209 +123,356 @@ public class ForkJoinPool extends Abstra
/*
* Implementation Overview
*
- * This class provides the central bookkeeping and control for a
- * set of worker threads: Submissions from non-FJ threads enter
- * into a submission queue. 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 worker queues, and
- * lastly to new submissions.
+ * 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 loosely 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 never
+ * take tasks, and they are multiplexed on to a finite number of
+ * shared work queues. However, classes are set up so that future
+ * extensions could allow submitters to optionally help perform
+ * tasks as well. 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 runState), 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 a single 64bit volatile
- * variable ("ctl"). This variable is read on the order of 10-100
- * times as often as it is modified (always via CAS). (There is
- * some additional control state, for example variable "shutdown"
- * for which we can cope with uncoordinated updates.) This
- * streamlines synchronization and control at the expense of messy
- * constructions needed to repack status bits upon updates.
- * Updates tend not to contend with each other except during
- * bursts while submitted tasks begin or end. In some cases when
- * they do contend, threads can instead do something else
- * (usually, scan for tesks) until contention subsides.
- *
- * To enable 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.
- *
- * Recording Workers. Workers are recorded in the "workers" array
- * that is created upon pool construction and expanded if (rarely)
- * necessary. This is an array as opposed to some other data
- * structure to support index-based random steals by workers.
- * Updates to the array recording new workers and unrecording
- * terminated ones are protected from each other by a seqLock
- * (scanGuard) but the array is otherwise concurrently readable,
- * and accessed directly by workers. To simplify index-based
- * operations, the array size is always a power of two, and all
- * readers must tolerate null slots. To avoid flailing during
- * start-up, the array is presized to hold twice #parallelism
- * workers (which is unlikely to need further resizing during
- * execution). But to avoid dealing with so many null slots,
- * variable scanGuard includes a mask for the nearest power of two
- * that contains all current workers. 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 workers are
- * via indices into the workers array (which is one source of some
- * of the messy code constructions here). In essence, the workers
- * array serves as a weak reference mechanism. Thus for example
- * the wait queue field of ctl stores worker indices, not worker
- * references. Access to the workers 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 workers array, as well as queue arrays, check
- * that the array 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.
- *
- * Wait Queuing. Unlike HPC work-stealing frameworks, we cannot
- * let workers spin indefinitely scanning for tasks when none are
- * can be immediately found, 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. 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 to both wake up waiters (by
- * advancing their count) and avoid ABA effects. Successors are
- * held in worker field "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 both before (in scan())
- * and after (in awaitWork()) 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.
+ * 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 "runState" contains 32 bits needed to register and
+ * deregister WorkQueues, as well as to enable shutdown. It is
+ * only modified under a lock (normally briefly held, but
+ * occasionally protecting allocations and resizings) but even
+ * when locked remains available to check consistency.
+ *
+ * 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. Shared (submission) queues are at even
+ * indices, worker queues at odd indices. 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. When a submission is added or another worker adds a
- * task to a queue that previously had two or fewer tasks, they
+ * task to a queue that previously had fewer than two tasks, they
* signal waiting workers (or trigger creation of new ones if
* fewer than the given parallelism level -- see signalWork).
- * These primary signals are buttressed by signals during rescans
- * as well as those performed when a worker steals a task and
- * notices that there are more tasks too; together these cover the
- * signals needed in cases when more than two tasks are pushed
- * but untaken.
+ * These primary signals are buttressed by signals during rescans;
+ * together these cover the signals needed in cases when more
+ * tasks are pushed but untaken, and improve performance compared
+ * to having one thread wake up all workers.
*
* 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
- * SHRINK_RATE nanosecs.
+ * 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.
*
- * Submissions. External submissions are maintained in an
- * array-based queue that is structured identically to
- * ForkJoinWorkerThread queues (which see) except for the use of
- * submissionLock in method addSubmission. Unlike worker queues,
- * multiple external threads can add new submissions.
- *
- * Compensation. Beyond work-stealing support and lifecycle
- * control, the main responsibility of this framework is to take
- * actions 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:
+ * Shutdown and Termination. A call to shutdownNow atomically sets
+ * a runState bit and then (non-atomically) sets each worker's
+ * runState 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. Method
- * ForkJoinWorkerThread.joinTask tracks joining->stealing
- * links to try to find such a task.
+ * would be running if the steal had not occurred.
*
* Compensating: Unless there are already enough live threads,
- * method tryPreBlock() may create or re-activate a spare
- * thread to compensate for blocked joiners until they
- * unblock.
+ * method tryCompensate() may create or re-activate a spare
+ * thread to compensate for blocked joiners until they unblock.
+ *
+ * A third form (implemented in tryRemoveAndExec and
+ * tryPollForAndExec) 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. A stealHint 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.
+ *
* 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 and require heuristic
- * guidance, so we rely on multiple retries of each. Currently,
- * in keeping with on-demand signalling policy, we compensate only
- * if blocking would leave less than one active (non-waiting,
- * non-blocked) worker. Additionally, to avoid some false alarms
- * due to GC, lagging counters, system activity, etc, compensated
- * blocking for joins is only attempted after a number of rechecks
- * proportional to the current apparent deficit (where retries are
- * interspersed with Thread.yield, for good citizenship). The
- * variable blockedCount, incremented before blocking and
- * decremented after, is sometimes needed to distinguish cases of
- * waiting for work vs blocking on joins or other managed sync,
- * but both the cases are equivalent for most pool control, so we
- * can update non-atomically. (Additionally, contention on
- * blockedCount alleviates some contention on ctl).
- *
- * Shutdown and Termination. A call to shutdownNow atomically sets
- * the ctl stop bit and then (non-atomically) sets each workers
- * "terminate" 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 quiesence (i.e., that
- * there is no more work) which is reflected in active counts so
- * long as there are no current blockers, as well as possible
- * re-evaluations during independent changes in blocking or
- * quiescing workers.
+ * 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: (1) We only try
+ * compensation after attempting enough helping steps (measured
+ * via counting and timing) that we have already consumed the
+ * estimated cost of creating and activating a new thread. (2) We
+ * allow up to 50% of threads to be blocked before initially
+ * adding any others, and unless completely saturated, check that
+ * some work is available for a new worker before adding. Also, we
+ * create up to only 50% more threads until entering a mode that
+ * only adds a thread if all others are possibly blocked. All
+ * together, this means that we might be half as fast to react,
+ * and create half as many threads as possible in the ideal case,
+ * but present vastly fewer anomalies in all other cases compared
+ * to both more aggressive and more conservative alternatives.
*
* Style notes: There is a lot of representation-level coupling
* among classes ForkJoinPool, ForkJoinWorkerThread, and
- * ForkJoinTask. Most fields of ForkJoinWorkerThread maintain
- * data structures managed by ForkJoinPool, so are directly
- * accessed. Conversely we allow access to "workers" array by
- * workers, and direct access to ForkJoinTask.status by both
- * ForkJoinPool and ForkJoinWorkerThread. There is little point
- * trying to reduce this, since any associated future changes in
- * representations will need to be accompanied by algorithmic
- * changes anyway. All together, these low-level implementation
- * choices produce as much as a factor of 4 performance
- * improvement compared to naive implementations, and enable the
- * processing of billions of tasks per second, at the expense of
- * some ugliness.
- *
- * 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 that help some methods perform reasonably even
- * when interpreted (not compiled).
- *
- * The order of declarations in this file is: (1) declarations of
- * statics (2) fields (along with constants used when unpacking
- * some of them), listed in an order that tends to reduce
- * contention among them a bit under most JVMs. (3) internal
- * control methods (4) callbacks and other support for
- * ForkJoinTask and ForkJoinWorkerThread classes, (5) exported
- * methods (plus a few little helpers). (6) static block
- * initializing all statics in a minimally dependent order.
+ * 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 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(modifyThreadPermission);
+ }
+
+ // Nested classes
+
/**
* Factory for creating new {@link ForkJoinWorkerThread}s.
* A {@code ForkJoinWorkerThreadFactory} must be defined and used
@@ -358,114 +501,744 @@ public class ForkJoinPool extends Abstra
}
/**
- * Creates a new ForkJoinWorkerThread. This factory is used unless
- * overridden in ForkJoinPool constructors.
+ * 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 {
+ EmptyTask() { status = ForkJoinTask.NORMAL; } // 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 trySharedPush, 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 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. (Although we can avoid one case
+ * of this when locked in trySharedPush.) 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 (via method growArray) by workers before
+ * use. Others are allocated on first use.
*/
- public static final ForkJoinWorkerThreadFactory
- defaultForkJoinWorkerThreadFactory;
+ 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;
- /**
- * Permission required for callers of methods that may start or
- * kill threads.
- */
- private static final RuntimePermission modifyThreadPermission;
+ /**
+ * 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
+
+ volatile long totalSteals; // cumulative number of steals
+ 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 rescans; // remaining scans until block
+ int nsteals; // top-level task executions since last idle
+ final int mode; // lifo, fifo, or shared
+ int poolIndex; // index of this queue in pool (or 0)
+ int stealHint; // index of most recent known stealer
+ volatile int runState; // 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
+ // Heuristic padding to ameliorate unfortunate memory placements
+ Object p00, p01, p02, p03, p04, p05, p06, p07;
+ Object p08, p09, p0a, p0b, p0c, p0d, p0e;
+
+ WorkQueue(ForkJoinPool pool, ForkJoinWorkerThread owner, int mode) {
+ this.mode = mode;
+ this.pool = pool;
+ this.owner = owner;
+ // Place indices in the center of array (that is not yet allocated)
+ base = top = INITIAL_QUEUE_CAPACITY >>> 1;
+ }
+
+ /**
+ * Returns the approximate number of tasks in the queue.
+ */
+ final int queueSize() {
+ int n = base - top; // 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[] a; int m, s;
+ int n = base - (s = top);
+ return (n >= 0 ||
+ (n == -1 &&
+ ((a = array) == null ||
+ (m = a.length - 1) < 0 ||
+ U.getObjectVolatile
+ (a, ((m & (s - 1)) << ASHIFT) + ABASE) == null)));
+ }
+
+ /**
+ * Pushes a task. Call only by owner in unshared queues.
+ *
+ * @param task the task. Caller must ensure non-null.
+ * @throw RejectedExecutionException if array cannot be resized
+ */
+ final void push(ForkJoinTask task) {
+ ForkJoinTask[] a; ForkJoinPool p;
+ int s = top, m, n;
+ if ((a = array) != null) { // ignore if queue removed
+ U.putOrderedObject
+ (a, (((m = a.length - 1) & s) << ASHIFT) + ABASE, task);
+ if ((n = (top = s + 1) - base) <= 2) {
+ if ((p = pool) != null)
+ p.signalWork();
+ }
+ else if (n >= m)
+ growArray(true);
+ }
+ }
+
+ /**
+ * Pushes a task if lock is free and array is either big
+ * enough or can be resized to be big enough.
+ *
+ * @param task the task. Caller must ensure non-null.
+ * @return true if submitted
+ */
+ final boolean trySharedPush(ForkJoinTask task) {
+ boolean submitted = false;
+ if (runState == 0 && U.compareAndSwapInt(this, RUNSTATE, 0, 1)) {
+ ForkJoinTask[] a = array;
+ int s = top;
+ try {
+ if ((a != null && a.length > s + 1 - base) ||
+ (a = growArray(false)) != null) { // must presize
+ int j = (((a.length - 1) & s) << ASHIFT) + ABASE;
+ U.putObject(a, (long)j, task); // don't need "ordered"
+ top = s + 1;
+ submitted = true;
+ }
+ } finally {
+ runState = 0; // unlock
+ }
+ }
+ return submitted;
+ }
+
+ /**
+ * Takes next task, if one exists, in LIFO order. Call only
+ * by owner in unshared queues.
+ */
+ final ForkJoinTask pop() {
+ ForkJoinTask[] a; ForkJoinTask t; int m;
+ if ((a = array) != null && (m = a.length - 1) >= 0) {
+ for (int s; (s = top - 1) - base >= 0;) {
+ long j = ((m & s) << ASHIFT) + ABASE;
+ if ((t = (ForkJoinTask)U.getObject(a, j)) == null)
+ break;
+ if (U.compareAndSwapObject(a, j, t, null)) {
+ top = s;
+ return t;
+ }
+ }
+ }
+ return null;
+ }
+
+ final ForkJoinTask sharedPop() {
+ ForkJoinTask task = null;
+ if (runState == 0 && U.compareAndSwapInt(this, RUNSTATE, 0, 1)) {
+ try {
+ ForkJoinTask[] a; int m;
+ if ((a = array) != null && (m = a.length - 1) >= 0) {
+ for (int s; (s = top - 1) - base >= 0;) {
+ long j = ((m & s) << ASHIFT) + ABASE;
+ ForkJoinTask t =
+ (ForkJoinTask)U.getObject(a, j);
+ if (t == null)
+ break;
+ if (U.compareAndSwapObject(a, j, t, null)) {
+ top = s;
+ task = t;
+ break;
+ }
+ }
+ }
+ } finally {
+ runState = 0;
+ }
+ }
+ return task;
+ }
+
+
+ /**
+ * 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 = array) != null) {
+ int j = (((a.length - 1) & b) << ASHIFT) + ABASE;
+ if ((t = (ForkJoinTask)U.getObjectVolatile(a, j)) != null &&
+ base == b &&
+ U.compareAndSwapObject(a, j, t, null)) {
+ base = b + 1;
+ return t;
+ }
+ }
+ return null;
+ }
+
+ /**
+ * Takes next task, if one exists, in FIFO order.
+ */
+ final ForkJoinTask poll() {
+ ForkJoinTask[] a; int b; ForkJoinTask t;
+ while ((b = base) - top < 0 && (a = array) != null) {
+ int j = (((a.length - 1) & b) << ASHIFT) + ABASE;
+ t = (ForkJoinTask)U.getObjectVolatile(a, j);
+ if (t != null) {
+ if (base == b &&
+ U.compareAndSwapObject(a, j, t, null)) {
+ base = b + 1;
+ return t;
+ }
+ }
+ else if (base == b) {
+ if (b + 1 == top)
+ break;
+ Thread.yield(); // wait for lagging update
+ }
+ }
+ return null;
+ }
+
+ /**
+ * Takes next task, if one exists, in order specified by mode.
+ */
+ final ForkJoinTask nextLocalTask() {
+ return mode == 0 ? pop() : poll();
+ }
+
+ /**
+ * Returns next task, if one exists, in order specified by mode.
+ */
+ final ForkJoinTask peek() {
+ ForkJoinTask[] a = array; int m;
+ if (a == null || (m = a.length - 1) < 0)
+ return null;
+ int i = mode == 0 ? top - 1 : base;
+ int j = ((i & m) << ASHIFT) + ABASE;
+ return (ForkJoinTask)U.getObjectVolatile(a, j);
+ }
+
+ /**
+ * Pops the given task only if it is at the current top.
+ */
+ final boolean tryUnpush(ForkJoinTask t) {
+ ForkJoinTask[] a; int s;
+ if ((a = array) != null && (s = top) != base &&
+ U.compareAndSwapObject
+ (a, (((a.length - 1) & --s) << ASHIFT) + ABASE, t, null)) {
+ top = s;
+ return true;
+ }
+ return false;
+ }
+
+ /**
+ * Version of tryUnpush for shared queues; called by non-FJ
+ * submitters after prechecking that task probably exists.
+ */
+ final boolean trySharedUnpush(ForkJoinTask t) {
+ boolean success = false;
+ if (runState == 0 && U.compareAndSwapInt(this, RUNSTATE, 0, 1)) {
+ try {
+ ForkJoinTask[] a; int s;
+ if ((a = array) != null && (s = top) != base &&
+ U.compareAndSwapObject
+ (a, (((a.length - 1) & --s) << ASHIFT) + ABASE, t, null)) {
+ top = s;
+ success = true;
+ }
+ } finally {
+ runState = 0; // unlock
+ }
+ }
+ return success;
+ }
+
+ /**
+ * Polls the given task only if it is at the current base.
+ */
+ final boolean pollFor(ForkJoinTask task) {
+ ForkJoinTask[] a; int b;
+ if ((b = base) - top < 0 && (a = array) != null) {
+ int j = (((a.length - 1) & b) << ASHIFT) + ABASE;
+ if (U.getObjectVolatile(a, j) == task && base == b &&
+ U.compareAndSwapObject(a, j, task, null)) {
+ base = b + 1;
+ return true;
+ }
+ }
+ return false;
+ }
+
+ /**
+ * 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.
+ *
+ * @param rejectOnFailure if true, throw exception if capacity
+ * exceeded (relayed ultimately to user); else return null.
+ */
+ final ForkJoinTask[] growArray(boolean rejectOnFailure) {
+ ForkJoinTask[] oldA = array;
+ int size = oldA != null ? oldA.length << 1 : INITIAL_QUEUE_CAPACITY;
+ if (size <= MAXIMUM_QUEUE_CAPACITY) {
+ int oldMask, t, b;
+ ForkJoinTask[] a = array = new ForkJoinTask[size];
+ if (oldA != null && (oldMask = oldA.length - 1) >= 0 &&
+ (t = top) - (b = base) > 0) {
+ int mask = size - 1;
+ do {
+ ForkJoinTask x;
+ int oldj = ((b & oldMask) << ASHIFT) + ABASE;
+ int j = ((b & mask) << ASHIFT) + ABASE;
+ x = (ForkJoinTask)U.getObjectVolatile(oldA, oldj);
+ if (x != null &&
+ U.compareAndSwapObject(oldA, oldj, x, null))
+ U.putObjectVolatile(a, j, x);
+ } while (++b != t);
+ }
+ return a;
+ }
+ else if (!rejectOnFailure)
+ return null;
+ else
+ throw new RejectedExecutionException("Queue capacity exceeded");
+ }
+
+ /**
+ * Removes and cancels all known tasks, ignoring any exceptions.
+ */
+ final void cancelAll() {
+ ForkJoinTask.cancelIgnoringExceptions(currentJoin);
+ ForkJoinTask.cancelIgnoringExceptions(currentSteal);
+ 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 = seed;
+ r ^= r << 13;
+ r ^= r >>> 17;
+ return seed = r ^= r << 5;
+ }
+
+ // Specialized execution methods
+
+ /**
+ * Pops and runs tasks until empty.
+ */
+ private void popAndExecAll() {
+ // A bit faster than repeated pop calls
+ ForkJoinTask[] a; int m, s; long j; ForkJoinTask t;
+ while ((a = array) != null && (m = a.length - 1) >= 0 &&
+ (s = top - 1) - base >= 0 &&
+ (t = ((ForkJoinTask)
+ U.getObject(a, j = ((m & s) << ASHIFT) + ABASE)))
+ != null) {
+ if (U.compareAndSwapObject(a, j, t, null)) {
+ top = 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) immediately on any CAS
+ * or consistency check failure so caller can retry.
+ *
+ * @return 0 if no progress can be made, else positive
+ * (this unusual convention simplifies use with tryHelpStealer.)
+ */
+ final int tryRemoveAndExec(ForkJoinTask task) {
+ int stat = 1;
+ boolean removed = false, empty = true;
+ ForkJoinTask[] a; int m, s, b, n;
+ if ((a = array) != null && (m = a.length - 1) >= 0 &&
+ (n = (s = top) - (b = base)) > 0) {
+ for (ForkJoinTask t;;) { // traverse from s to b
+ int j = ((--s & m) << ASHIFT) + ABASE;
+ t = (ForkJoinTask)U.getObjectVolatile(a, j);
+ if (t == null) // inconsistent length
+ break;
+ else if (t == task) {
+ if (s + 1 == top) { // pop
+ if (!U.compareAndSwapObject(a, j, task, null))
+ break;
+ top = s;
+ removed = true;
+ }
+ else if (base == b) // replace with proxy
+ removed = U.compareAndSwapObject(a, j, task,
+ new EmptyTask());
+ break;
+ }
+ else if (t.status >= 0)
+ empty = false;
+ else if (s + 1 == top) { // pop and throw away
+ if (U.compareAndSwapObject(a, j, t, null))
+ top = s;
+ break;
+ }
+ if (--n == 0) {
+ if (!empty && base == b)
+ stat = 0;
+ break;
+ }
+ }
+ }
+ if (removed)
+ task.doExec();
+ return stat;
+ }
+
+ /**
+ * Version of shared pop that takes top element only if it
+ * its root is the given CountedCompleter.
+ */
+ final CountedCompleter sharedPopCC(CountedCompleter root) {
+ CountedCompleter task = null;
+ if (runState == 0 && U.compareAndSwapInt(this, RUNSTATE, 0, 1)) {
+ try {
+ ForkJoinTask[] a; int m;
+ if ((a = array) != null && (m = a.length - 1) >= 0) {
+ outer:for (int s; (s = top - 1) - base >= 0;) {
+ long j = ((m & s) << ASHIFT) + ABASE;
+ ForkJoinTask t =
+ (ForkJoinTask)U.getObject(a, j);
+ if (t == null || !(t instanceof CountedCompleter))
+ break;
+ CountedCompleter cc = (CountedCompleter)t;
+ for (CountedCompleter q = cc, p;;) {
+ if (q == root) {
+ if (U.compareAndSwapObject(a, j, cc, null)) {
+ top = s;
+ task = cc;
+ break outer;
+ }
+ break;
+ }
+ if ((p = q.completer) == null)
+ break outer;
+ q = p;
+ }
+ }
+ }
+ } finally {
+ runState = 0;
+ }
+ }
+ return task;
+ }
+
+ /**
+ * Executes a top-level task and any local tasks remaining
+ * after execution.
+ */
+ final void runTask(ForkJoinTask t) {
+ if (t != null) {
+ currentSteal = t;
+ t.doExec();
+ if (top != base) { // process remaining local tasks
+ if (mode == 0)
+ popAndExecAll();
+ else
+ pollAndExecAll();
+ }
+ ++nsteals;
+ currentSteal = null;
+ }
+ }
+
+ /**
+ * Executes a non-top-level (stolen) task.
+ */
+ final void runSubtask(ForkJoinTask t) {
+ if (t != null) {
+ ForkJoinTask ps = currentSteal;
+ currentSteal = t;
+ t.doExec();
+ currentSteal = ps;
+ }
+ }
+
+ /**
+ * Returns true if owned and not known to be blocked.
+ */
+ final boolean isApparentlyUnblocked() {
+ Thread wt; Thread.State s;
+ return (eventCount >= 0 &&
+ (wt = owner) != null &&
+ (s = wt.getState()) != Thread.State.BLOCKED &&
+ s != Thread.State.WAITING &&
+ s != Thread.State.TIMED_WAITING);
+ }
+
+ /**
+ * If this owned and is not already interrupted, try to
+ * interrupt and/or unpark, ignoring exceptions.
+ */
+ final void interruptOwner() {
+ Thread wt, p;
+ if ((wt = owner) != null && !wt.isInterrupted()) {
+ try {
+ wt.interrupt();
+ } catch (SecurityException ignore) {
+ }
+ }
+ if ((p = parker) != null)
+ U.unpark(p);
+ }
+
+ // Unsafe mechanics
+ private static final sun.misc.Unsafe U;
+ private static final long RUNSTATE;
+ private static final int ABASE;
+ private static final int ASHIFT;
+ static {
+ int s;
+ try {
+ U = getUnsafe();
+ Class k = WorkQueue.class;
+ Class ak = ForkJoinTask[].class;
+ RUNSTATE = U.objectFieldOffset
+ (k.getDeclaredField("runState"));
+ ABASE = U.arrayBaseOffset(ak);
+ s = U.arrayIndexScale(ak);
+ } catch (Exception e) {
+ throw new Error(e);
+ }
+ if ((s & (s-1)) != 0)
+ throw new Error("data type scale not a power of two");
+ ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
+ }
+ }
/**
- * If there is a security manager, makes sure caller has
- * permission to modify threads.
+ * 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 doSubmit. 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. This is done during
+ * registration for workers, but requires a separate AtomicInteger
+ * for submitters. Seeds are then randomly modified upon
+ * collisions using xorshifts, which requires a non-zero seed.
*/
- private static void checkPermission() {
- SecurityManager security = System.getSecurityManager();
- if (security != null)
- security.checkPermission(modifyThreadPermission);
+ static final class Submitter {
+ int seed;
+ Submitter() {
+ int s = nextSubmitterSeed.getAndAdd(SEED_INCREMENT);
+ seed = (s == 0) ? 1 : s; // ensure non-zero
+ }
+ }
+
+ /** ThreadLocal class for Submitters */
+ static final class ThreadSubmitter extends ThreadLocal {
+ public Submitter initialValue() { return new Submitter(); }
}
+ // static fields (initialized in static initializer below)
+
/**
- * Generator for assigning sequence numbers as pool names.
+ * Creates a new ForkJoinWorkerThread. This factory is used unless
+ * overridden in ForkJoinPool constructors.
*/
- private static final AtomicInteger poolNumberGenerator;
+ public static final ForkJoinWorkerThreadFactory
+ defaultForkJoinWorkerThreadFactory;
+
+ /** Property prefix for constructing common pool */
+ private static final String propPrefix =
+ "java.util.concurrent.ForkJoinPool.common.";
/**
- * Generator for initial random seeds for worker victim
- * selection. This is used only to create initial seeds. Random
- * steals use a cheaper xorshift generator per steal attempt. We
- * don't expect much contention on seedGenerator, so just use a
- * plain Random.
+ * Common (static) pool. Non-null for public use unless a static
+ * construction exception, but internal usages must null-check on
+ * use.
*/
- static final Random workerSeedGenerator;
+ static final ForkJoinPool commonPool;
/**
- * Array holding all worker threads in the pool. Initialized upon
- * construction. Array size must be a power of two. Updates and
- * replacements are protected by scanGuard, but the array is
- * always kept in a consistent enough state to be randomly
- * accessed without locking by workers performing work-stealing,
- * as well as other traversal-based methods in this class, so long
- * as reads memory-acquire by first reading ctl. All readers must
- * tolerate that some array slots may be null.
+ * Common pool parallelism. Must equal commonPool.parallelism.
*/
- ForkJoinWorkerThread[] workers;
+ static final int commonPoolParallelism;
/**
- * Initial size for submission queue array. Must be a power of
- * two. In many applications, these always stay small so we use a
- * small initial cap.
+ * Generator for assigning sequence numbers as pool names.
*/
- private static final int INITIAL_QUEUE_CAPACITY = 8;
+ private static final AtomicInteger poolNumberGenerator;
/**
- * Maximum size for submission queue array. Must be a power of two
- * less than or equal to 1 << (31 - width of array entry) to
- * ensure lack of index wraparound, but is capped at a lower
- * value to help users trap runaway computations.
+ * Generator for initial hashes/seeds for submitters. Accessed by
+ * Submitter class constructor.
*/
- private static final int MAXIMUM_QUEUE_CAPACITY = 1 << 24; // 16M
+ static final AtomicInteger nextSubmitterSeed;
/**
- * Array serving as submission queue. Initialized upon construction.
+ * Permission required for callers of methods that may start or
+ * kill threads.
*/
- private ForkJoinTask[] submissionQueue;
+ private static final RuntimePermission modifyThreadPermission;
/**
- * Lock protecting submissions array for addSubmission
+ * 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.
*/
- private final ReentrantLock submissionLock;
+ private static final ThreadSubmitter submitters;
+
+ // static constants
/**
- * Condition for awaitTermination, using submissionLock for
- * convenience.
+ * 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.
*/
- private final Condition termination;
+ private static final long IDLE_TIMEOUT = 1000L * 1000L * 1000L; // 1sec
/**
- * Creation factory for worker threads.
+ * Timeout value when there are more threads than parallelism level
*/
- private final ForkJoinWorkerThreadFactory factory;
+ private static final long FAST_IDLE_TIMEOUT = 100L * 1000L * 1000L;
/**
- * The uncaught exception handler used when any worker abruptly
- * terminates.
+ * The maximum stolen->joining link depth allowed in method
+ * tryHelpStealer. Must be a power of two. This value also
+ * controls the maximum number of times to try to help join a task
+ * without any apparent progress or change in pool state before
+ * giving up and blocking (see awaitJoin). 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.
*/
- final Thread.UncaughtExceptionHandler ueh;
+ private static final int MAX_HELP = 64;
/**
- * Prefix for assigning names to worker threads
+ * Secondary time-based bound (in nanosecs) for helping attempts
+ * before trying compensated blocking in awaitJoin. Used in
+ * conjunction with MAX_HELP to reduce variance due to different
+ * polling rates associated with different helping options. The
+ * value should roughly approximate the time required to create
+ * and/or activate a worker thread.
*/
- private final String workerNamePrefix;
+ private static final long COMPENSATION_DELAY = 1L << 18; // ~0.25 millisec
/**
- * Sum of per-thread steal counts, updated only when threads are
- * idle or terminating.
+ * Increment for seed generators. See class ThreadLocal for
+ * explanation.
*/
- private volatile long stealCount;
+ private static final int SEED_INCREMENT = 0x61c88647;
/**
- * Main pool control -- a long packed with:
+ * 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 (16bits)
+ * 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 waiting threads (16 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 =
@@ -474,15 +1247,26 @@ public class ForkJoinPool extends Abstra
* 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, when id is
- * negative, there is at least one waiting worker, and when e 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. Note: AC_SHIFT is
- * redundantly declared in ForkJoinWorkerThread in order to
- * integrate a surplus-threads check.
+ * 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 runState is an int packed with:
+ * SHUTDOWN: true if shutdown is enabled (1 bit)
+ * SEQ: a sequence number updated upon (de)registering workers (30 bits)
+ * INIT: set true after workQueues array construction (1 bit)
+ *
+ * The sequence number enables simple consistency checks:
+ * Staleness of read-only operations on the workQueues array can
+ * be checked by comparing runState before vs after the reads.
*/
- volatile long ctl;
// bit positions/shifts for fields
private static final int AC_SHIFT = 48;
@@ -491,8 +1275,9 @@ public class ForkJoinPool extends Abstra
private static final int EC_SHIFT = 16;
// bounds
- private static final int MAX_ID = 0x7fff; // max poolIndex
- private static final int SMASK = 0xffff; // mask short bits
+ private static final int SMASK = 0xffff; // short bits
+ private static final int MAX_CAP = 0x7fff; // max #workers - 1
+ private static final int SQMASK = 0xfffe; // even short bits
private static final int SHORT_SIGN = 1 << 15;
private static final int INT_SIGN = 1 << 31;
@@ -514,830 +1299,1032 @@ public class ForkJoinPool extends Abstra
private static final int UTC_UNIT = 1 << UTC_SHIFT;
// masks and units for dealing with e = (int)ctl
- private static final int E_MASK = 0x7fffffff; // no STOP_BIT
- private static final int EC_UNIT = 1 << EC_SHIFT;
-
- /**
- * The target parallelism level.
- */
- final int parallelism;
+ private static final int E_MASK = 0x7fffffff; // no STOP_BIT
+ private static final int E_SEQ = 1 << EC_SHIFT;
- /**
- * Index (mod submission queue length) of next element to take
- * from submission queue.
- */
- volatile int queueBase;
-
- /**
- * Index (mod submission queue length) of next element to add
- * in submission queue.
- */
- int queueTop;
+ // runState bits
+ private static final int SHUTDOWN = 1 << 31;
- /**
- * True when shutdown() has been called.
- */
- volatile boolean shutdown;
+ // access mode for WorkQueue
+ static final int LIFO_QUEUE = 0;
+ static final int FIFO_QUEUE = 1;
+ static final int SHARED_QUEUE = -1;
- /**
- * True if use local fifo, not default lifo, for local polling
- * Read by, and replicated by ForkJoinWorkerThreads
- */
- final boolean locallyFifo;
+ // Instance fields
- /**
- * The number of threads in ForkJoinWorkerThreads.helpQuiescePool.
- * When non-zero, suppresses automatic shutdown when active
- * counts become zero.
- */
- volatile int quiescerCount;
-
- /**
- * The number of threads blocked in join.
- */
- volatile int blockedCount;
-
- /**
- * Counter for worker Thread names (unrelated to their poolIndex)
- */
- private volatile int nextWorkerNumber;
-
- /**
- * The index for the next created worker. Accessed under scanGuard.
- */
- private int nextWorkerIndex;
+ /*
+ * Field layout order in 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.
+ */
+
+ volatile long stealCount; // collects worker counts
+ volatile long ctl; // main pool control
+ final int parallelism; // parallelism level
+ final int localMode; // per-worker scheduling mode
+ volatile int nextWorkerNumber; // to create worker name string
+ final int submitMask; // submit queue index bound
+ int nextSeed; // for initializing worker seeds
+ volatile int mainLock; // spinlock for array updates
+ volatile int runState; // shutdown status and seq
+ WorkQueue[] workQueues; // main registry
+ final ForkJoinWorkerThreadFactory factory; // factory for new workers
+ final Thread.UncaughtExceptionHandler ueh; // per-worker UEH
+ final String workerNamePrefix; // to create worker name string
- /**
- * SeqLock and index masking for for updates to workers array.
- * Locked when SG_UNIT is set. Unlocking clears bit by adding
- * SG_UNIT. Staleness of read-only operations can be checked by
- * comparing scanGuard to value before the reads. The low 16 bits
- * (i.e, anding with SMASK) hold (the smallest power of two
- * covering all worker indices, minus one, and is used to avoid
- * dealing with large numbers of null slots when the workers array
- * is overallocated.
- */
- volatile int scanGuard;
+ /*
+ * Mechanics for main lock protecting worker array updates. Uses
+ * the same strategy as ConcurrentHashMap bins -- a spinLock for
+ * normal cases, but falling back to builtin lock when (rarely)
+ * needed. See internal ConcurrentHashMap documentation for
+ * explanation.
+ */
+
+ static final int LOCK_WAITING = 2; // bit to indicate need for signal
+ static final int MAX_LOCK_SPINS = 1 << 8;
+
+ private void tryAwaitMainLock() {
+ int spins = MAX_LOCK_SPINS, r = 0, h;
+ while (((h = mainLock) & 1) != 0) {
+ if (r == 0)
+ r = ThreadLocalRandom.current().nextInt(); // randomize spins
+ else if (spins >= 0) {
+ r ^= r << 1; r ^= r >>> 3; r ^= r << 10; // xorshift
+ if (r >= 0)
+ --spins;
+ }
+ else if (U.compareAndSwapInt(this, MAINLOCK, h, h | LOCK_WAITING)) {
+ synchronized (this) {
+ if ((mainLock & LOCK_WAITING) != 0) {
+ try {
+ wait();
+ } catch (InterruptedException ie) {
+ try {
+ Thread.currentThread().interrupt();
+ } catch (SecurityException ignore) {
+ }
+ }
+ }
+ else
+ notifyAll(); // possibly won race vs signaller
+ }
+ break;
+ }
+ }
+ }
- private static final int SG_UNIT = 1 << 16;
+ // Creating, registering, and deregistering workers
/**
- * The wakeup interval (in nanoseconds) for a worker waiting for a
- * task when the pool is quiescent to instead try to shrink the
- * number of workers. The exact value does not matter too
- * much. It must be short enough to release resources during
- * sustained periods of idleness, but not so short that threads
- * are continually re-created.
+ * Tries to create and start a worker
*/
- private static final long SHRINK_RATE =
- 4L * 1000L * 1000L * 1000L; // 4 seconds
+ private void addWorker() {
+ Throwable ex = null;
+ ForkJoinWorkerThread wt = null;
+ try {
+ if ((wt = factory.newThread(this)) != null) {
+ wt.start();
+ return;
+ }
+ } catch (Throwable e) {
+ ex = e;
+ }
+ deregisterWorker(wt, ex); // adjust counts etc on failure
+ }
/**
- * Top-level loop for worker threads: On each step: if the
- * previous step swept through all queues and found no tasks, or
- * there are excess threads, then possibly blocks. Otherwise,
- * scans for and, if found, executes a task. Returns when pool
- * and/or worker terminate.
- *
- * @param w the worker
+ * Callback from ForkJoinWorkerThread constructor to assign a
+ * public name. This must be separate from registerWorker because
+ * it is called during the "super" constructor call in
+ * ForkJoinWorkerThread.
*/
- final void work(ForkJoinWorkerThread w) {
- boolean swept = false; // true on empty scans
- long c;
- while (!w.terminate && (int)(c = ctl) >= 0) {
- int a; // active count
- if (!swept && (a = (int)(c >> AC_SHIFT)) <= 0)
- swept = scan(w, a);
- else if (tryAwaitWork(w, c))
- swept = false;
- }
+ final String nextWorkerName() {
+ int n;
+ do {} while (!U.compareAndSwapInt(this, NEXTWORKERNUMBER,
+ n = nextWorkerNumber, ++n));
+ return workerNamePrefix.concat(Integer.toString(n));
}
- // Signalling
-
/**
- * Wakes up or creates a worker.
+ * Callback from ForkJoinWorkerThread constructor to establish its
+ * poolIndex 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.
+ *
+ * @param w the worker's queue
*/
- final void signalWork() {
- /*
- * The while condition is true if: (there is are too few total
- * workers OR there is at least one waiter) AND (there are too
- * few active workers OR the pool is terminating). The value
- * of e distinguishes the remaining cases: zero (no waiters)
- * for create, negative if terminating (in which case do
- * nothing), else release a waiter. The secondary checks for
- * release (non-null array etc) can fail if the pool begins
- * terminating after the test, and don't impose any added cost
- * because JVMs must perform null and bounds checks anyway.
- */
- long c; int e, u;
- while ((((e = (int)(c = ctl)) | (u = (int)(c >>> 32))) &
- (INT_SIGN|SHORT_SIGN)) == (INT_SIGN|SHORT_SIGN) && e >= 0) {
- if (e > 0) { // release a waiting worker
- int i; ForkJoinWorkerThread w; ForkJoinWorkerThread[] ws;
- if ((ws = workers) == null ||
- (i = ~e & SMASK) >= ws.length ||
- (w = ws[i]) == null)
- break;
- long nc = (((long)(w.nextWait & E_MASK)) |
- ((long)(u + UAC_UNIT) << 32));
- if (w.eventCount == e &&
- UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
- w.eventCount = (e + EC_UNIT) & E_MASK;
- if (w.parked)
- UNSAFE.unpark(w);
- break;
+ final void registerWorker(WorkQueue w) {
+ while (!U.compareAndSwapInt(this, MAINLOCK, 0, 1))
+ tryAwaitMainLock();
+ try {
+ WorkQueue[] ws;
+ if ((ws = workQueues) == null)
+ ws = workQueues = new WorkQueue[submitMask + 1];
+ if (w != null) {
+ int rs, n = ws.length, m = n - 1;
+ int s = nextSeed += SEED_INCREMENT; // rarely-colliding sequence
+ w.seed = (s == 0) ? 1 : s; // ensure non-zero seed
+ 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) & SQMASK) + 2;
+ while (ws[r = (r + step) & m] != null) {
+ if (++probes >= n) {
+ workQueues = ws = Arrays.copyOf(ws, n <<= 1);
+ m = n - 1;
+ probes = 0;
+ }
+ }
}
+ w.eventCount = w.poolIndex = r; // establish before recording
+ ws[r] = w; // also update seq
+ runState = ((rs = runState) & SHUTDOWN) | ((rs + 2) & ~SHUTDOWN);
}
- else if (UNSAFE.compareAndSwapLong
- (this, ctlOffset, c,
- (long)(((u + UTC_UNIT) & UTC_MASK) |
- ((u + UAC_UNIT) & UAC_MASK)) << 32)) {
- addWorker();
- break;
+ } finally {
+ if (!U.compareAndSwapInt(this, MAINLOCK, 1, 0)) {
+ mainLock = 0;
+ synchronized (this) { notifyAll(); };
}
}
}
/**
- * Variant of signalWork to help release waiters on rescans.
- * Tries once to release a waiter if active count < 0.
+ * Final callback from terminating worker, as well as upon failure
+ * to construct or start a worker in addWorker. Removes record of
+ * worker from array, and adjusts counts. If pool is shutting
+ * down, tries to complete termination.
*
- * @return false if failed due to contention, else true
+ * @param wt the worker thread or null if addWorker failed
+ * @param ex the exception causing failure, or null if none
*/
- private boolean tryReleaseWaiter() {
- long c; int e, i; ForkJoinWorkerThread w; ForkJoinWorkerThread[] ws;
- if ((e = (int)(c = ctl)) > 0 &&
- (int)(c >> AC_SHIFT) < 0 &&
- (ws = workers) != null &&
- (i = ~e & SMASK) < ws.length &&
- (w = ws[i]) != null) {
- long nc = ((long)(w.nextWait & E_MASK) |
- ((c + AC_UNIT) & (AC_MASK|TC_MASK)));
- if (w.eventCount != e ||
- !UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc))
- return false;
- w.eventCount = (e + EC_UNIT) & E_MASK;
- if (w.parked)
- UNSAFE.unpark(w);
+ final void deregisterWorker(ForkJoinWorkerThread wt, Throwable ex) {
+ WorkQueue w = null;
+ if (wt != null && (w = wt.workQueue) != null) {
+ w.runState = -1; // ensure runState is set
+ long steals = w.totalSteals + w.nsteals, sc;
+ do {} while (!U.compareAndSwapLong(this, STEALCOUNT,
+ sc = stealCount, sc + steals));
+ int idx = w.poolIndex;
+ while (!U.compareAndSwapInt(this, MAINLOCK, 0, 1))
+ tryAwaitMainLock();
+ try {
+ WorkQueue[] ws = workQueues;
+ if (ws != null && idx >= 0 && idx < ws.length && ws[idx] == w)
+ ws[idx] = null;
+ } finally {
+ if (!U.compareAndSwapInt(this, MAINLOCK, 1, 0)) {
+ mainLock = 0;
+ synchronized (this) { notifyAll(); };
+ }
+ }
}
- return true;
+
+ long c; // adjust ctl counts
+ do {} while (!U.compareAndSwapLong
+ (this, CTL, c = ctl, (((c - AC_UNIT) & AC_MASK) |
+ ((c - TC_UNIT) & TC_MASK) |
+ (c & ~(AC_MASK|TC_MASK)))));
+
+ if (!tryTerminate(false, false) && w != null) {
+ w.cancelAll(); // cancel remaining tasks
+ if (w.array != null) // suppress signal if never ran
+ signalWork(); // wake up or create replacement
+ if (ex == null) // help clean refs on way out
+ ForkJoinTask.helpExpungeStaleExceptions();
+ }
+
+ if (ex != null) // rethrow
+ ForkJoinTask.rethrow(ex);
}
- // Scanning for tasks
+ // Submissions
/**
- * Scans for and, if found, executes one task. Scans start at a
- * random index of workers array, and randomly select the first
- * (2*#workers)-1 probes, and then, if all empty, resort to 2
- * circular sweeps, which is necessary to check quiescence. and
- * taking a submission only if no stealable tasks were found. The
- * steal code inside the loop is a specialized form of
- * ForkJoinWorkerThread.deqTask, followed bookkeeping to support
- * helpJoinTask and signal propagation. The code for submission
- * queues is almost identical. On each steal, the worker completes
- * not only the task, but also all local tasks that this task may
- * have generated. On detecting staleness or contention when
- * trying to take a task, this method returns without finishing
- * sweep, which allows global state rechecks before retry.
- *
- * @param w the worker
- * @param a the number of active workers
- * @return true if swept all queues without finding a task
- */
- private boolean scan(ForkJoinWorkerThread w, int a) {
- int g = scanGuard; // mask 0 avoids useless scans if only one active
- int m = parallelism == 1 - a? 0 : g & SMASK;
- ForkJoinWorkerThread[] ws = workers;
- if (ws == null || ws.length <= m) // staleness check
- return false;
- for (int r = w.seed, k = r, j = -(m + m); j <= m + m; ++j) {
- ForkJoinTask t; ForkJoinTask[] q; int b, i;
- ForkJoinWorkerThread v = ws[k & m];
- if (v != null && (b = v.queueBase) != v.queueTop &&
- (q = v.queue) != null && (i = (q.length - 1) & b) >= 0) {
- long u = (i << ASHIFT) + ABASE;
- if ((t = q[i]) != null && v.queueBase == b &&
- UNSAFE.compareAndSwapObject(q, u, t, null)) {
- int d = (v.queueBase = b + 1) - v.queueTop;
- v.stealHint = w.poolIndex;
- if (d != 0)
- signalWork(); // propagate if nonempty
- w.execTask(t);
+ * Unless shutting down, adds the given task to a submission queue
+ * at submitter's current queue index (modulo submission
+ * range). If no queue exists at the index, one is created. If
+ * the queue is busy, another index is randomly chosen. The
+ * submitMask bounds the effective number of queues to the
+ * (nearest power of two for) parallelism level.
+ *
+ * @param task the task. Caller must ensure non-null.
+ */
+ private void doSubmit(ForkJoinTask task) {
+ Submitter s = submitters.get();
+ for (int r = s.seed, m = submitMask;;) {
+ WorkQueue[] ws; WorkQueue q;
+ int k = r & m & SQMASK; // use only even indices
+ if (runState < 0)
+ throw new RejectedExecutionException(); // shutting down
+ else if ((ws = workQueues) == null || ws.length <= k) {
+ while (!U.compareAndSwapInt(this, MAINLOCK, 0, 1))
+ tryAwaitMainLock();
+ try {
+ if (workQueues == null)
+ workQueues = new WorkQueue[submitMask + 1];
+ } finally {
+ if (!U.compareAndSwapInt(this, MAINLOCK, 1, 0)) {
+ mainLock = 0;
+ synchronized (this) { notifyAll(); };
+ }
}
- r ^= r << 13; r ^= r >>> 17; w.seed = r ^ (r << 5);
- return false; // store next seed
- }
- else if (j < 0) { // xorshift
- r ^= r << 13; r ^= r >>> 17; k = r ^= r << 5;
}
- else
- ++k;
- }
- if (scanGuard != g) // staleness check
- return false;
- else { // try to take submission
- ForkJoinTask t; ForkJoinTask[] q; int b, i;
- if ((b = queueBase) != queueTop &&
- (q = submissionQueue) != null &&
- (i = (q.length - 1) & b) >= 0) {
- long u = (i << ASHIFT) + ABASE;
- if ((t = q[i]) != null && queueBase == b &&
- UNSAFE.compareAndSwapObject(q, u, t, null)) {
- queueBase = b + 1;
- w.execTask(t);
+ else if ((q = ws[k]) == null) { // create new queue
+ WorkQueue nq = new WorkQueue(this, null, SHARED_QUEUE);
+ while (!U.compareAndSwapInt(this, MAINLOCK, 0, 1))
+ tryAwaitMainLock();
+ try {
+ int rs = runState; // to update seq
+ if (ws == workQueues && ws[k] == null) {
+ ws[k] = nq;
+ runState = ((rs & SHUTDOWN) | ((rs + 2) & ~SHUTDOWN));
+ }
+ } finally {
+ if (!U.compareAndSwapInt(this, MAINLOCK, 1, 0)) {
+ mainLock = 0;
+ synchronized (this) { notifyAll(); };
+ }
}
- return false;
}
- return true; // all queues empty
+ else if (q.trySharedPush(task)) {
+ signalWork();
+ return;
+ }
+ else if (m > 1) { // move to a different index
+ r ^= r << 13; // same xorshift as WorkQueues
+ r ^= r >>> 17;
+ s.seed = r ^= r << 5;
+ }
+ else
+ Thread.yield(); // yield if no alternatives
}
}
/**
- * Tries to enqueue worker in wait queue and await change in
- * worker's eventCount. Before blocking, rescans queues to avoid
- * missed signals. If the pool is quiescent, possibly terminates
- * worker upon exit.
+ * Submits the given (non-null) task to the common pool, if possible.
+ */
+ static void submitToCommonPool(ForkJoinTask task) {
+ ForkJoinPool p;
+ if ((p = commonPool) == null)
+ throw new RejectedExecutionException("Common Pool Unavailable");
+ p.doSubmit(task);
+ }
+
+ /**
+ * Returns true if the given task was submitted to common pool
+ * and has not yet commenced execution, and is available for
+ * removal according to execution policies; if so removing the
+ * submission from the pool.
*
- * @param w the calling worker
- * @param c the ctl value on entry
- * @return true if waited or another thread was released upon enq
+ * @param task the task
+ * @return true if successful
*/
- private boolean tryAwaitWork(ForkJoinWorkerThread w, long c) {
- int v = w.eventCount;
- w.nextWait = (int)c; // w's successor record
- long nc = (long)(v & E_MASK) | ((c - AC_UNIT) & (AC_MASK|TC_MASK));
- if (ctl != c || !UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
- long d = ctl; // return true if lost to a deq, to force rescan
- return (int)d != (int)c && ((d - c) & AC_MASK) >= 0L;
- }
- if (parallelism + (int)(c >> AC_SHIFT) == 1 &&
- blockedCount == 0 && quiescerCount == 0)
- idleAwaitWork(w, v); // quiescent -- maybe shrink
-
- boolean rescanned = false;
- for (int sc;;) {
- if (w.eventCount != v)
- return true;
- if ((sc = w.stealCount) != 0) {
- long s = stealCount; // accumulate stealCount
- if (UNSAFE.compareAndSwapLong(this, stealCountOffset, s, s+sc))
- w.stealCount = 0;
- }
- else if (!rescanned) {
- int g = scanGuard, m = g & SMASK;
- ForkJoinWorkerThread[] ws = workers;
- if (ws != null && m < ws.length) {
- rescanned = true;
- for (int i = 0; i <= m; ++i) {
- ForkJoinWorkerThread u = ws[i];
- if (u != null) {
- if (u.queueBase != u.queueTop &&
- !tryReleaseWaiter())
- rescanned = false; // contended
- if (w.eventCount != v)
- return true;
+ static boolean tryUnsubmitFromCommonPool(ForkJoinTask task) {
+ // If not oversaturating platform, peek, looking for task and
+ // eligibility before using trySharedUnpush to actually take
+ // it under lock
+ ForkJoinPool p; WorkQueue[] ws; WorkQueue w, q;
+ ForkJoinTask[] a; int ac, s, m;
+ if ((p = commonPool) != null && (ws = p.workQueues) != null) {
+ int k = submitters.get().seed & p.submitMask & SQMASK;
+ if ((m = ws.length - 1) >= k && (q = ws[k]) != null &&
+ (ac = (int)(p.ctl >> AC_SHIFT)) <= 0) {
+ if (ac == 0) { // double check if all workers active
+ for (int i = 1; i <= m; i += 2) {
+ if ((w = ws[i]) != null && w.parker != null) {
+ ac = -1;
+ break;
}
}
}
- if (scanGuard != g || // stale
- (queueBase != queueTop && !tryReleaseWaiter()))
- rescanned = false;
- if (!rescanned)
- Thread.yield(); // reduce contention
- else
- Thread.interrupted(); // clear before park
- }
- else {
- w.parked = true; // must recheck
- if (w.eventCount != v) {
- w.parked = false;
- return true;
- }
- LockSupport.park(this);
- rescanned = w.parked = false;
+ return (ac < 0 && (a = q.array) != null &&
+ (s = q.top - 1) - q.base >= 0 &&
+ s >= 0 && s < a.length &&
+ a[s] == task &&
+ q.trySharedUnpush(task));
}
}
+ return false;
}
/**
- * If pool is quiescent, checks for termination, and waits for
- * event signal for up to SHRINK_RATE nanosecs. On timeout, if ctl
- * has not changed, terminates the worker. Upon its termination
- * (see deregisterWorker), it may wake up another worker to
- * possibly repeat this process.
- *
- * @param w the calling worker
- * @param v the eventCount w must wait until changed
+ * Tries to pop and run a task within same computation from common pool
*/
- private void idleAwaitWork(ForkJoinWorkerThread w, int v) {
- ForkJoinTask.helpExpungeStaleExceptions(); // help clean weak refs
- if (shutdown)
- tryTerminate(false);
- long c = ctl;
- long nc = (((c & (AC_MASK|TC_MASK)) + AC_UNIT) |
- (long)(w.nextWait & E_MASK)); // ctl value to release w
- if (w.eventCount == v &&
- parallelism + (int)(c >> AC_SHIFT) == 0 &&
- blockedCount == 0 && quiescerCount == 0) {
- long startTime = System.nanoTime();
- Thread.interrupted();
- if (w.eventCount == v) {
- w.parked = true;
- if (w.eventCount == v)
- LockSupport.parkNanos(this, SHRINK_RATE);
- w.parked = false;
- if (w.eventCount == v && ctl == c &&
- System.nanoTime() - startTime >= SHRINK_RATE &&
- UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
- w.terminate = true;
- w.eventCount = ((int)c + EC_UNIT) & E_MASK;
+ static void popAndExecCCFromCommonPool(CountedCompleter cc) {
+ ForkJoinPool p; WorkQueue[] ws; WorkQueue q, w; int m, ac;
+ CountedCompleter par, task;
+ if ((p = commonPool) != null && (ws = p.workQueues) != null) {
+ while ((par = cc.completer) != null) // find root
+ cc = par;
+ int k = submitters.get().seed & p.submitMask & SQMASK;
+ if ((m = ws.length - 1) >= k && (q = ws[k]) != null &&
+ (ac = (int)(p.ctl >> AC_SHIFT)) <= 0) {
+ if (ac == 0) {
+ for (int i = 1; i <= m; i += 2) {
+ if ((w = ws[i]) != null && w.parker != null) {
+ ac = -1;
+ break;
+ }
+ }
}
+ if (ac < 0 && q.top - q.base > 0 &&
+ (task = q.sharedPopCC(cc)) != null)
+ task.exec();
}
}
}
- // Submissions
+ // Maintaining ctl counts
/**
- * Enqueues the given task in the submissionQueue. Same idea as
- * ForkJoinWorkerThread.pushTask except for use of submissionLock.
- *
- * @param t the task
+ * Increments active count; mainly called upon return from blocking.
*/
- private void addSubmission(ForkJoinTask t) {
- final ReentrantLock lock = this.submissionLock;
- lock.lock();
- try {
- ForkJoinTask[] q; int s, m;
- if ((q = submissionQueue) != null) { // ignore if queue removed
- long u = (((s = queueTop) & (m = q.length-1)) << ASHIFT)+ABASE;
- UNSAFE.putOrderedObject(q, u, t);
- queueTop = s + 1;
- if (s - queueBase == m)
- growSubmissionQueue();
- }
- } finally {
- lock.unlock();
- }
- signalWork();
+ final void incrementActiveCount() {
+ long c;
+ do {} while (!U.compareAndSwapLong(this, CTL, c = ctl, c + AC_UNIT));
}
- // (pollSubmission is defined below with exported methods)
-
/**
- * Creates or doubles submissionQueue array.
- * Basically identical to ForkJoinWorkerThread version
+ * Tries to create one or activate one or more workers if too few are active.
*/
- private void growSubmissionQueue() {
- ForkJoinTask[] oldQ = submissionQueue;
- int size = oldQ != null ? oldQ.length << 1 : INITIAL_QUEUE_CAPACITY;
- if (size > MAXIMUM_QUEUE_CAPACITY)
- throw new RejectedExecutionException("Queue capacity exceeded");
- if (size < INITIAL_QUEUE_CAPACITY)
- size = INITIAL_QUEUE_CAPACITY;
- ForkJoinTask[] q = submissionQueue = new ForkJoinTask[size];
- int mask = size - 1;
- int top = queueTop;
- int oldMask;
- if (oldQ != null && (oldMask = oldQ.length - 1) >= 0) {
- for (int b = queueBase; b != top; ++b) {
- long u = ((b & oldMask) << ASHIFT) + ABASE;
- Object x = UNSAFE.getObjectVolatile(oldQ, u);
- if (x != null && UNSAFE.compareAndSwapObject(oldQ, u, x, null))
- UNSAFE.putObjectVolatile
- (q, ((b & mask) << ASHIFT) + ABASE, x);
+ final void signalWork() {
+ long c; int u;
+ while ((u = (int)((c = ctl) >>> 32)) < 0) { // too few active
+ WorkQueue[] ws = workQueues; int e, i; WorkQueue w; Thread p;
+ if ((e = (int)c) > 0) { // at least one waiting
+ if (ws != null && (i = e & SMASK) < ws.length &&
+ (w = ws[i]) != null && w.eventCount == (e | INT_SIGN)) {
+ long nc = (((long)(w.nextWait & E_MASK)) |
+ ((long)(u + UAC_UNIT) << 32));
+ if (U.compareAndSwapLong(this, CTL, c, nc)) {
+ w.eventCount = (e + E_SEQ) & E_MASK;
+ if ((p = w.parker) != null)
+ U.unpark(p); // activate and release
+ break;
+ }
+ }
+ else
+ break;
}
+ else if (e == 0 && (u & SHORT_SIGN) != 0) { // too few total
+ long nc = (long)(((u + UTC_UNIT) & UTC_MASK) |
+ ((u + UAC_UNIT) & UAC_MASK)) << 32;
+ if (U.compareAndSwapLong(this, CTL, c, nc)) {
+ addWorker();
+ break;
+ }
+ }
+ else
+ break;
}
}
- // Blocking support
+ // Scanning for tasks
/**
- * Tries to increment blockedCount, decrement active count
- * (sometimes implicitly) and possibly release or create a
- * compensating worker in preparation for blocking. Fails
- * on contention or termination.
- *
- * @return true if the caller can block, else should recheck and retry
+ * Top-level runloop for workers, called by ForkJoinWorkerThread.run.
*/
- private boolean tryPreBlock() {
- int b = blockedCount;
- if (UNSAFE.compareAndSwapInt(this, blockedCountOffset, b, b + 1)) {
- int pc = parallelism;
- do {
- ForkJoinWorkerThread[] ws; ForkJoinWorkerThread w;
- int e, ac, tc, rc, i;
- long c = ctl;
- int u = (int)(c >>> 32);
- if ((e = (int)c) < 0) {
- // skip -- terminating
- }
- else if ((ac = (u >> UAC_SHIFT)) <= 0 && e != 0 &&
- (ws = workers) != null &&
- (i = ~e & SMASK) < ws.length &&
- (w = ws[i]) != null) {
- long nc = ((long)(w.nextWait & E_MASK) |
- (c & (AC_MASK|TC_MASK)));
- if (w.eventCount == e &&
- UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
- w.eventCount = (e + EC_UNIT) & E_MASK;
- if (w.parked)
- UNSAFE.unpark(w);
- return true; // release an idle worker
+ final void runWorker(WorkQueue w) {
+ w.growArray(false); // initialize queue array in this thread
+ do { w.runTask(scan(w)); } while (w.runState >= 0);
+ }
+
+ /**
+ * Scans for and, if found, returns one task, else possibly
+ * inactivates the worker. This method operates on single reads of
+ * volatile state and is designed to be re-invoked continuously,
+ * in part because it returns upon detecting inconsistencies,
+ * contention, or state changes that indicate possible success on
+ * re-invocation.
+ *
+ * The scan searches for tasks across a random permutation of
+ * queues (starting at a random index and stepping by a random
+ * relative prime, checking each at least once). The scan
+ * terminates upon either finding a non-empty queue, or completing
+ * the sweep. If the worker is not inactivated, it takes and
+ * returns a task from this queue. On failure to find a task, we
+ * take one of the following actions, after which the caller will
+ * retry calling this method unless terminated.
+ *
+ * * If pool is terminating, terminate the worker.
+ *
+ * * If not a complete sweep, try to release a waiting worker. If
+ * the scan terminated because the worker is inactivated, then the
+ * released worker will often be the calling worker, and it can
+ * succeed obtaining a task on the next call. Or maybe it is
+ * another worker, but with same net effect. Releasing in other
+ * cases as well ensures that we have enough workers running.
+ *
+ * * If not already enqueued, try to inactivate and enqueue the
+ * worker on wait queue. Or, if inactivating has caused the pool
+ * to be quiescent, relay to idleAwaitWork to check for
+ * termination and possibly shrink pool.
+ *
+ * * If already inactive, and the caller has run a task since the
+ * last empty scan, return (to allow rescan) unless others are
+ * also inactivated. Field WorkQueue.rescans counts down on each
+ * scan to ensure eventual inactivation and blocking.
+ *
+ * * If already enqueued and none of the above apply, park
+ * awaiting signal,
+ *
+ * @param w the worker (via its WorkQueue)
+ * @return a task or null if none found
+ */
+ private final ForkJoinTask scan(WorkQueue w) {
+ WorkQueue[] ws; // first update random seed
+ int r = w.seed; r ^= r << 13; r ^= r >>> 17; w.seed = r ^= r << 5;
+ int rs = runState, m; // volatile read order matters
+ if ((ws = workQueues) != null && (m = ws.length - 1) > 0) {
+ int ec = w.eventCount; // ec is negative if inactive
+ int step = (r >>> 16) | 1; // relative prime
+ for (int j = (m + 1) << 2; ; r += step) {
+ WorkQueue q; ForkJoinTask t; ForkJoinTask[] a; int b;
+ if ((q = ws[r & m]) != null && (b = q.base) - q.top < 0 &&
+ (a = q.array) != null) { // probably nonempty
+ int i = (((a.length - 1) & b) << ASHIFT) + ABASE;
+ t = (ForkJoinTask)U.getObjectVolatile(a, i);
+ if (q.base == b && ec >= 0 && t != null &&
+ U.compareAndSwapObject(a, i, t, null)) {
+ if (q.top - (q.base = b + 1) > 0)
+ signalWork(); // help pushes signal
+ return t;
+ }
+ else if (ec < 0 || j <= m) {
+ rs = 0; // mark scan as imcomplete
+ break; // caller can retry after release
}
}
- else if ((tc = (short)(u >>> UTC_SHIFT)) >= 0 && ac + pc > 1) {
- long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK);
- if (UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc))
- return true; // no compensation needed
+ if (--j < 0)
+ break;
+ }
+
+ long c = ctl; int e = (int)c, a = (int)(c >> AC_SHIFT), nr, ns;
+ if (e < 0) // decode ctl on empty scan
+ w.runState = -1; // pool is terminating
+ else if (rs == 0 || rs != runState) { // incomplete scan
+ WorkQueue v; Thread p; // try to release a waiter
+ if (e > 0 && a < 0 && w.eventCount == ec &&
+ (v = ws[e & m]) != null && v.eventCount == (e | INT_SIGN)) {
+ long nc = ((long)(v.nextWait & E_MASK) |
+ ((c + AC_UNIT) & (AC_MASK|TC_MASK)));
+ if (ctl == c && U.compareAndSwapLong(this, CTL, c, nc)) {
+ v.eventCount = (e + E_SEQ) & E_MASK;
+ if ((p = v.parker) != null)
+ U.unpark(p);
+ }
}
- else if (tc + pc < MAX_ID) {
- long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK);
- if (UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
- addWorker();
- return true; // create a replacement
+ }
+ else if (ec >= 0) { // try to enqueue/inactivate
+ long nc = (long)ec | ((c - AC_UNIT) & (AC_MASK|TC_MASK));
+ w.nextWait = e;
+ w.eventCount = ec | INT_SIGN; // mark as inactive
+ if (ctl != c || !U.compareAndSwapLong(this, CTL, c, nc))
+ w.eventCount = ec; // unmark on CAS failure
+ else {
+ if ((ns = w.nsteals) != 0) {
+ w.nsteals = 0; // set rescans if ran task
+ w.rescans = (a > 0) ? 0 : a + parallelism;
+ w.totalSteals += ns;
}
+ if (a == 1 - parallelism) // quiescent
+ idleAwaitWork(w, nc, c);
+ }
+ }
+ else if (w.eventCount < 0) { // already queued
+ int ac = a + parallelism;
+ if ((nr = w.rescans) > 0) // continue rescanning
+ w.rescans = (ac < nr) ? ac : nr - 1;
+ else if (((w.seed >>> 16) & ac) == 0) { // randomize park
+ Thread.interrupted(); // clear status
+ Thread wt = Thread.currentThread();
+ U.putObject(wt, PARKBLOCKER, this);
+ w.parker = wt; // emulate LockSupport.park
+ if (w.eventCount < 0) // recheck
+ U.park(false, 0L);
+ w.parker = null;
+ U.putObject(wt, PARKBLOCKER, null);
}
- // try to back out on any failure and let caller retry
- } while (!UNSAFE.compareAndSwapInt(this, blockedCountOffset,
- b = blockedCount, b - 1));
+ }
}
- return false;
- }
-
- /**
- * Decrements blockedCount and increments active count
- */
- private void postBlock() {
- long c;
- do {} while (!UNSAFE.compareAndSwapLong(this, ctlOffset, // no mask
- c = ctl, c + AC_UNIT));
- int b;
- do {} while(!UNSAFE.compareAndSwapInt(this, blockedCountOffset,
- b = blockedCount, b - 1));
+ return null;
}
/**
- * Possibly blocks waiting for the given task to complete, or
- * cancels the task if terminating. Fails to wait if contended.
+ * If inactivating worker w has caused the pool to become
+ * quiescent, checks for pool termination, and, so long as this is
+ * not the only worker, waits for event for up to a given
+ * duration. On timeout, if ctl has not changed, terminates the
+ * worker, which will in turn wake up another worker to possibly
+ * repeat this process.
*
- * @param joinMe the task
+ * @param w the calling worker
+ * @param currentCtl the ctl value triggering possible quiescence
+ * @param prevCtl the ctl value to restore if thread is terminated
*/
- final void tryAwaitJoin(ForkJoinTask joinMe) {
- int s;
- Thread.interrupted(); // clear interrupts before checking termination
- if (joinMe.status >= 0) {
- if (tryPreBlock()) {
- joinMe.tryAwaitDone(0L);
- postBlock();
+ private void idleAwaitWork(WorkQueue w, long currentCtl, long prevCtl) {
+ if (w.eventCount < 0 && !tryTerminate(false, false) &&
+ (int)prevCtl != 0 && !hasQueuedSubmissions() && ctl == currentCtl) {
+ int dc = -(short)(currentCtl >>> TC_SHIFT);
+ long parkTime = dc < 0 ? FAST_IDLE_TIMEOUT: (dc + 1) * IDLE_TIMEOUT;
+ long deadline = System.nanoTime() + parkTime - 100000L; // 1ms slop
+ Thread wt = Thread.currentThread();
+ while (ctl == currentCtl) {
+ Thread.interrupted(); // timed variant of version in scan()
+ U.putObject(wt, PARKBLOCKER, this);
+ w.parker = wt;
+ if (ctl == currentCtl)
+ U.park(false, parkTime);
+ w.parker = null;
+ U.putObject(wt, PARKBLOCKER, null);
+ if (ctl != currentCtl)
+ break;
+ if (deadline - System.nanoTime() <= 0L &&
+ U.compareAndSwapLong(this, CTL, currentCtl, prevCtl)) {
+ w.eventCount = (w.eventCount + E_SEQ) | E_MASK;
+ w.runState = -1; // shrink
+ break;
+ }
}
- if ((ctl & STOP_BIT) != 0L)
- joinMe.cancelIgnoringExceptions();
}
}
/**
- * Possibly blocks the given worker waiting for joinMe to
- * complete or timeout
- *
- * @param joinMe the task
- * @param millis the wait time for underlying Object.wait
- */
- final void timedAwaitJoin(ForkJoinTask joinMe, long nanos) {
- while (joinMe.status >= 0) {
- Thread.interrupted();
- if ((ctl & STOP_BIT) != 0L) {
- joinMe.cancelIgnoringExceptions();
- break;
- }
- if (tryPreBlock()) {
- long last = System.nanoTime();
- while (joinMe.status >= 0) {
- long millis = TimeUnit.NANOSECONDS.toMillis(nanos);
- if (millis <= 0)
- break;
- joinMe.tryAwaitDone(millis);
- if (joinMe.status < 0)
- break;
- if ((ctl & STOP_BIT) != 0L) {
- joinMe.cancelIgnoringExceptions();
- break;
+ * Tries to locate and execute tasks for a stealer of the given
+ * task, or in turn one of its stealers, Traces currentSteal ->
+ * currentJoin links looking for a thread working on a descendant
+ * of the given task and with a non-empty queue to steal back and
+ * execute tasks from. The first call to this method upon a
+ * waiting join will often entail scanning/search, (which is OK
+ * because the joiner has nothing better to do), but this method
+ * leaves hints in workers to speed up subsequent calls. The
+ * implementation is very branchy to cope with potential
+ * inconsistencies or loops encountering chains that are stale,
+ * unknown, or so long that they are likely cyclic.
+ *
+ * @param joiner the joining worker
+ * @param task the task to join
+ * @return 0 if no progress can be made, negative if task
+ * known complete, else positive
+ */
+ private int tryHelpStealer(WorkQueue joiner, ForkJoinTask task) {
+ int stat = 0, steps = 0; // bound to avoid cycles
+ if (joiner != null && task != null) { // hoist null checks
+ restart: for (;;) {
+ ForkJoinTask subtask = task; // current target
+ for (WorkQueue j = joiner, v;;) { // v is stealer of subtask
+ WorkQueue[] ws; int m, s, h;
+ if ((s = task.status) < 0) {
+ stat = s;
+ break restart;
+ }
+ if ((ws = workQueues) == null || (m = ws.length - 1) <= 0)
+ break restart; // shutting down
+ if ((v = ws[h = (j.stealHint | 1) & m]) == null ||
+ v.currentSteal != subtask) {
+ for (int origin = h;;) { // find stealer
+ if (((h = (h + 2) & m) & 15) == 1 &&
+ (subtask.status < 0 || j.currentJoin != subtask))
+ continue restart; // occasional staleness check
+ if ((v = ws[h]) != null &&
+ v.currentSteal == subtask) {
+ j.stealHint = h; // save hint
+ break;
+ }
+ if (h == origin)
+ break restart; // cannot find stealer
+ }
+ }
+ for (;;) { // help stealer or descend to its stealer
+ ForkJoinTask[] a; int b;
+ if (subtask.status < 0) // surround probes with
+ continue restart; // consistency checks
+ if ((b = v.base) - v.top < 0 && (a = v.array) != null) {
+ int i = (((a.length - 1) & b) << ASHIFT) + ABASE;
+ ForkJoinTask t =
+ (ForkJoinTask)U.getObjectVolatile(a, i);
+ if (subtask.status < 0 || j.currentJoin != subtask ||
+ v.currentSteal != subtask)
+ continue restart; // stale
+ stat = 1; // apparent progress
+ if (t != null && v.base == b &&
+ U.compareAndSwapObject(a, i, t, null)) {
+ v.base = b + 1; // help stealer
+ joiner.runSubtask(t);
+ }
+ else if (v.base == b && ++steps == MAX_HELP)
+ break restart; // v apparently stalled
+ }
+ else { // empty -- try to descend
+ ForkJoinTask next = v.currentJoin;
+ if (subtask.status < 0 || j.currentJoin != subtask ||
+ v.currentSteal != subtask)
+ continue restart; // stale
+ else if (next == null || ++steps == MAX_HELP)
+ break restart; // dead-end or maybe cyclic
+ else {
+ subtask = next;
+ j = v;
+ break;
+ }
+ }
}
- long now = System.nanoTime();
- nanos -= now - last;
- last = now;
}
- postBlock();
- break;
}
}
+ return stat;
}
/**
- * If necessary, compensates for blocker, and blocks
+ * If task is at base of some steal queue, steals and executes it.
+ *
+ * @param joiner the joining worker
+ * @param task the task
*/
- private void awaitBlocker(ManagedBlocker blocker)
- throws InterruptedException {
- while (!blocker.isReleasable()) {
- if (tryPreBlock()) {
- try {
- do {} while (!blocker.isReleasable() && !blocker.block());
- } finally {
- postBlock();
+ private void tryPollForAndExec(WorkQueue joiner, ForkJoinTask task) {
+ WorkQueue[] ws;
+ if ((ws = workQueues) != null) {
+ for (int j = 1; j < ws.length && task.status >= 0; j += 2) {
+ WorkQueue q = ws[j];
+ if (q != null && q.pollFor(task)) {
+ joiner.runSubtask(task);
+ break;
}
- break;
}
}
}
- // Creating, registering and deregistring workers
-
/**
- * Tries to create and start a worker; minimally rolls back counts
- * on failure.
- */
- private void addWorker() {
- Throwable ex = null;
- ForkJoinWorkerThread t = null;
- try {
- t = factory.newThread(this);
- } catch (Throwable e) {
- ex = e;
- }
- if (t == null) { // null or exceptional factory return
- long c; // adjust counts
- do {} while (!UNSAFE.compareAndSwapLong
- (this, ctlOffset, c = ctl,
- (((c - AC_UNIT) & AC_MASK) |
- ((c - TC_UNIT) & TC_MASK) |
- (c & ~(AC_MASK|TC_MASK)))));
- // Propagate exception if originating from an external caller
- if (!tryTerminate(false) && ex != null &&
- !(Thread.currentThread() instanceof ForkJoinWorkerThread))
- UNSAFE.throwException(ex);
- }
- else
- t.start();
- }
-
- /**
- * Callback from ForkJoinWorkerThread constructor to assign a
- * public name
+ * Tries to decrement active count (sometimes implicitly) and
+ * possibly release or create a compensating worker in preparation
+ * for blocking. Fails on contention or termination. Otherwise,
+ * adds a new thread if no idle workers are available and either
+ * pool would become completely starved or: (at least half
+ * starved, and fewer than 50% spares exist, and there is at least
+ * one task apparently available). Even though the availability
+ * check requires a full scan, it is worthwhile in reducing false
+ * alarms.
+ *
+ * @param task if non-null, a task being waited for
+ * @param blocker if non-null, a blocker being waited for
+ * @return true if the caller can block, else should recheck and retry
*/
- final String nextWorkerName() {
- for (int n;;) {
- if (UNSAFE.compareAndSwapInt(this, nextWorkerNumberOffset,
- n = nextWorkerNumber, ++n))
- return workerNamePrefix + n;
- }
- }
-
- /**
- * Callback from ForkJoinWorkerThread constructor to
- * determine its poolIndex and record in workers array.
- *
- * @param w the worker
- * @return the worker's pool index
- */
- final int registerWorker(ForkJoinWorkerThread w) {
- /*
- * In the typical case, a new worker acquires the lock, uses
- * next available index and returns quickly. Since we should
- * not block callers (ultimately from signalWork or
- * tryPreBlock) waiting for the lock needed to do this, we
- * instead help release other workers while waiting for the
- * lock.
- */
- for (int g;;) {
- ForkJoinWorkerThread[] ws;
- if (((g = scanGuard) & SG_UNIT) == 0 &&
- UNSAFE.compareAndSwapInt(this, scanGuardOffset,
- g, g | SG_UNIT)) {
- int k = nextWorkerIndex;
- try {
- if ((ws = workers) != null) { // ignore on shutdown
- int n = ws.length;
- if (k < 0 || k >= n || ws[k] != null) {
- for (k = 0; k < n && ws[k] != null; ++k)
- ;
- if (k == n)
- ws = workers = Arrays.copyOf(ws, n << 1);
+ final boolean tryCompensate(ForkJoinTask task, ManagedBlocker blocker) {
+ int pc = parallelism, e;
+ long c = ctl;
+ WorkQueue[] ws = workQueues;
+ if ((e = (int)c) >= 0 && ws != null) {
+ int u, a, ac, hc;
+ int tc = (short)((u = (int)(c >>> 32)) >>> UTC_SHIFT) + pc;
+ boolean replace = false;
+ if ((a = u >> UAC_SHIFT) <= 0) {
+ if ((ac = a + pc) <= 1)
+ replace = true;
+ else if ((e > 0 || (task != null &&
+ ac <= (hc = pc >>> 1) && tc < pc + hc))) {
+ WorkQueue w;
+ for (int j = 0; j < ws.length; ++j) {
+ if ((w = ws[j]) != null && !w.isEmpty()) {
+ replace = true;
+ break; // in compensation range and tasks available
}
- ws[k] = w;
- nextWorkerIndex = k + 1;
- int m = g & SMASK;
- g = k >= m? ((m << 1) + 1) & SMASK : g + (SG_UNIT<<1);
}
- } finally {
- scanGuard = g;
}
- return k;
}
- else if ((ws = workers) != null) { // help release others
- for (ForkJoinWorkerThread u : ws) {
- if (u != null && u.queueBase != u.queueTop) {
- if (tryReleaseWaiter())
- break;
+ if ((task == null || task.status >= 0) && // recheck need to block
+ (blocker == null || !blocker.isReleasable()) && ctl == c) {
+ if (!replace) { // no compensation
+ long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK);
+ if (U.compareAndSwapLong(this, CTL, c, nc))
+ return true;
+ }
+ else if (e != 0) { // release an idle worker
+ WorkQueue w; Thread p; int i;
+ if ((i = e & SMASK) < ws.length && (w = ws[i]) != null) {
+ long nc = ((long)(w.nextWait & E_MASK) |
+ (c & (AC_MASK|TC_MASK)));
+ if (w.eventCount == (e | INT_SIGN) &&
+ U.compareAndSwapLong(this, CTL, c, nc)) {
+ w.eventCount = (e + E_SEQ) & E_MASK;
+ if ((p = w.parker) != null)
+ U.unpark(p);
+ return true;
+ }
+ }
+ }
+ else if (tc < MAX_CAP) { // create replacement
+ long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK);
+ if (U.compareAndSwapLong(this, CTL, c, nc)) {
+ addWorker();
+ return true;
}
}
}
}
+ return false;
}
/**
- * Final callback from terminating worker. Removes record of
- * worker from array, and adjusts counts. If pool is shutting
- * down, tries to complete termination.
- *
- * @param w the worker
- */
- final void deregisterWorker(ForkJoinWorkerThread w, Throwable ex) {
- int idx = w.poolIndex;
- int sc = w.stealCount;
- int steps = 0;
- // Remove from array, adjust worker counts and collect steal count.
- // We can intermix failed removes or adjusts with steal updates
- do {
- long s, c;
- int g;
- if (steps == 0 && ((g = scanGuard) & SG_UNIT) == 0 &&
- UNSAFE.compareAndSwapInt(this, scanGuardOffset,
- g, g |= SG_UNIT)) {
- ForkJoinWorkerThread[] ws = workers;
- if (ws != null && idx >= 0 &&
- idx < ws.length && ws[idx] == w)
- ws[idx] = null; // verify
- nextWorkerIndex = idx;
- scanGuard = g + SG_UNIT;
- steps = 1;
- }
- if (steps == 1 &&
- UNSAFE.compareAndSwapLong(this, ctlOffset, c = ctl,
- (((c - AC_UNIT) & AC_MASK) |
- ((c - TC_UNIT) & TC_MASK) |
- (c & ~(AC_MASK|TC_MASK)))))
- steps = 2;
- if (sc != 0 &&
- UNSAFE.compareAndSwapLong(this, stealCountOffset,
- s = stealCount, s + sc))
- sc = 0;
- } while (steps != 2 || sc != 0);
- if (!tryTerminate(false)) {
- if (ex != null) // possibly replace if died abnormally
- signalWork();
- else
- tryReleaseWaiter();
- }
- }
-
- // Shutdown and termination
-
- /**
- * Possibly initiates and/or completes termination.
+ * Helps and/or blocks until the given task is done.
*
- * @param now if true, unconditionally terminate, else only
- * if shutdown and empty queue and no active workers
- * @return true if now terminating or terminated
+ * @param joiner the joining worker
+ * @param task the task
+ * @return task status on exit
*/
- private boolean tryTerminate(boolean now) {
- long c;
- while (((c = ctl) & STOP_BIT) == 0) {
- if (!now) {
- if ((int)(c >> AC_SHIFT) != -parallelism)
- return false;
- if (!shutdown || blockedCount != 0 || quiescerCount != 0 ||
- queueTop - queueBase > 0) {
- if (ctl == c) // staleness check
- return false;
- continue;
- }
- }
- if (UNSAFE.compareAndSwapLong(this, ctlOffset, c, c | STOP_BIT))
- startTerminating();
- }
- if ((short)(c >>> TC_SHIFT) == -parallelism) {
- submissionLock.lock();
- termination.signalAll();
- submissionLock.unlock();
- }
- return true;
- }
-
- /**
- * Runs up to three passes through workers: (0) Setting
- * termination status for each worker, followed by wakeups up
- * queued workers (1) helping cancel tasks (2) interrupting
- * lagging threads (likely in external tasks, but possibly also
- * blocked in joins). Each pass repeats previous steps because of
- * potential lagging thread creation.
- */
- private void startTerminating() {
- cancelSubmissions();
- for (int pass = 0; pass < 3; ++pass) {
- ForkJoinWorkerThread[] ws = workers;
- if (ws != null) {
- for (ForkJoinWorkerThread w : ws) {
- if (w != null) {
- w.terminate = true;
- if (pass > 0) {
- w.cancelTasks();
- if (pass > 1 && !w.isInterrupted()) {
- try {
- w.interrupt();
- } catch (SecurityException ignore) {
+ final int awaitJoin(WorkQueue joiner, ForkJoinTask task) {
+ int s;
+ if ((s = task.status) >= 0) {
+ ForkJoinTask prevJoin = joiner.currentJoin;
+ joiner.currentJoin = task;
+ long startTime = 0L;
+ for (int k = 0;;) {
+ if ((s = (joiner.isEmpty() ? // try to help
+ tryHelpStealer(joiner, task) :
+ joiner.tryRemoveAndExec(task))) == 0 &&
+ (s = task.status) >= 0) {
+ if (k == 0) {
+ startTime = System.nanoTime();
+ tryPollForAndExec(joiner, task); // check uncommon case
+ }
+ else if ((k & (MAX_HELP - 1)) == 0 &&
+ System.nanoTime() - startTime >=
+ COMPENSATION_DELAY &&
+ tryCompensate(task, null)) {
+ if (task.trySetSignal()) {
+ synchronized (task) {
+ if (task.status >= 0) {
+ try { // see ForkJoinTask
+ task.wait(); // for explanation
+ } catch (InterruptedException ie) {
+ }
}
+ else
+ task.notifyAll();
}
}
+ long c; // re-activate
+ do {} while (!U.compareAndSwapLong
+ (this, CTL, c = ctl, c + AC_UNIT));
}
}
- terminateWaiters();
+ if (s < 0 || (s = task.status) < 0) {
+ joiner.currentJoin = prevJoin;
+ break;
+ }
+ else if ((k++ & (MAX_HELP - 1)) == MAX_HELP >>> 1)
+ Thread.yield(); // for politeness
}
}
+ return s;
}
/**
- * Polls and cancels all submissions. Called only during termination.
+ * Stripped-down variant of awaitJoin used by timed joins. Tries
+ * to help join only while there is continuous progress. (Caller
+ * will then enter a timed wait.)
+ *
+ * @param joiner the joining worker
+ * @param task the task
+ * @return task status on exit
*/
- private void cancelSubmissions() {
- while (queueBase != queueTop) {
- ForkJoinTask task = pollSubmission();
- if (task != null) {
- try {
- task.cancel(false);
- } catch (Throwable ignore) {
+ final int helpJoinOnce(WorkQueue joiner, ForkJoinTask task) {
+ int s;
+ while ((s = task.status) >= 0 &&
+ (joiner.isEmpty() ?
+ tryHelpStealer(joiner, task) :
+ joiner.tryRemoveAndExec(task)) != 0)
+ ;
+ return s;
+ }
+
+ /**
+ * Returns a (probably) non-empty steal queue, if one is found
+ * during a random, then cyclic scan, else null. This method must
+ * be retried by caller if, by the time it tries to use the queue,
+ * it is empty.
+ */
+ private WorkQueue findNonEmptyStealQueue(WorkQueue w) {
+ // Similar to loop in scan(), but ignoring submissions
+ int r = w.seed; r ^= r << 13; r ^= r >>> 17; w.seed = r ^= r << 5;
+ int step = (r >>> 16) | 1;
+ for (WorkQueue[] ws;;) {
+ int rs = runState, m;
+ if ((ws = workQueues) == null || (m = ws.length - 1) < 1)
+ return null;
+ for (int j = (m + 1) << 2; ; r += step) {
+ WorkQueue q = ws[((r << 1) | 1) & m];
+ if (q != null && !q.isEmpty())
+ return q;
+ else if (--j < 0) {
+ if (runState == rs)
+ return null;
+ break;
}
}
}
}
/**
- * Tries to set the termination status of waiting workers, and
- * then wake them up (after which they will terminate).
- */
- private void terminateWaiters() {
- ForkJoinWorkerThread[] ws = workers;
- if (ws != null) {
- ForkJoinWorkerThread w; long c; int i, e;
- int n = ws.length;
- while ((i = ~(e = (int)(c = ctl)) & SMASK) < n &&
- (w = ws[i]) != null && w.eventCount == (e & E_MASK)) {
- if (UNSAFE.compareAndSwapLong(this, ctlOffset, c,
- (long)(w.nextWait & E_MASK) |
- ((c + AC_UNIT) & AC_MASK) |
- (c & (TC_MASK|STOP_BIT)))) {
- w.terminate = true;
- w.eventCount = e + EC_UNIT;
- if (w.parked)
- UNSAFE.unpark(w);
+ * Runs tasks until {@code isQuiescent()}. We piggyback on
+ * active count ctl maintenance, but rather than blocking
+ * when tasks cannot be found, we rescan until all others cannot
+ * find tasks either.
+ */
+ final void helpQuiescePool(WorkQueue w) {
+ for (boolean active = true;;) {
+ ForkJoinTask localTask; // exhaust local queue
+ while ((localTask = w.nextLocalTask()) != null)
+ localTask.doExec();
+ WorkQueue q = findNonEmptyStealQueue(w);
+ if (q != null) {
+ ForkJoinTask t; int b;
+ if (!active) { // re-establish active count
+ long c;
+ active = true;
+ do {} while (!U.compareAndSwapLong
+ (this, CTL, c = ctl, c + AC_UNIT));
+ }
+ if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null)
+ w.runSubtask(t);
+ }
+ else {
+ long c;
+ if (active) { // decrement active count without queuing
+ active = false;
+ do {} while (!U.compareAndSwapLong
+ (this, CTL, c = ctl, c -= AC_UNIT));
+ }
+ else
+ c = ctl; // re-increment on exit
+ if ((int)(c >> AC_SHIFT) + parallelism == 0) {
+ do {} while (!U.compareAndSwapLong
+ (this, CTL, c = ctl, c + AC_UNIT));
+ break;
}
}
}
}
- // misc ForkJoinWorkerThread support
-
/**
- * Increment or decrement quiescerCount. Needed only to prevent
- * triggering shutdown if a worker is transiently inactive while
- * checking quiescence.
- *
- * @param delta 1 for increment, -1 for decrement
+ * Restricted version of helpQuiescePool for non-FJ callers
*/
- final void addQuiescerCount(int delta) {
- int c;
- do {} while(!UNSAFE.compareAndSwapInt(this, quiescerCountOffset,
- c = quiescerCount, c + delta));
+ static void externalHelpQuiescePool() {
+ ForkJoinPool p; WorkQueue[] ws; WorkQueue q, sq;
+ ForkJoinTask[] a; int b;
+ ForkJoinTask t = null;
+ int k = submitters.get().seed & SQMASK;
+ if ((p = commonPool) != null &&
+ (ws = p.workQueues) != null &&
+ ws.length > (k &= p.submitMask) &&
+ (q = ws[k]) != null) {
+ while (q.top - q.base > 0) {
+ if ((t = q.sharedPop()) != null)
+ break;
+ }
+ if (t == null && (sq = p.findNonEmptyStealQueue(q)) != null &&
+ (b = sq.base) - sq.top < 0)
+ t = sq.pollAt(b);
+ if (t != null)
+ t.doExec();
+ }
}
/**
- * Directly increment or decrement active count without
- * queuing. This method is used to transiently assert inactivation
- * while checking quiescence.
+ * Gets and removes a local or stolen task for the given worker.
*
- * @param delta 1 for increment, -1 for decrement
+ * @return a task, if available
*/
- final void addActiveCount(int delta) {
- long d = delta < 0 ? -AC_UNIT : AC_UNIT;
- long c;
- do {} while (!UNSAFE.compareAndSwapLong(this, ctlOffset, c = ctl,
- ((c + d) & AC_MASK) |
- (c & ~AC_MASK)));
+ final ForkJoinTask nextTaskFor(WorkQueue w) {
+ for (ForkJoinTask t;;) {
+ WorkQueue q; int b;
+ if ((t = w.nextLocalTask()) != null)
+ return t;
+ if ((q = findNonEmptyStealQueue(w)) == null)
+ return null;
+ if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null)
+ return t;
+ }
}
/**
* Returns the approximate (non-atomic) number of idle threads per
- * active thread.
+ * active thread to offset steal queue size for method
+ * ForkJoinTask.getSurplusQueuedTaskCount().
*/
final int idlePerActive() {
// Approximate at powers of two for small values, saturate past 4
int p = parallelism;
- int a = p + (int)(ctl >> AC_SHIFT);
- return (a > (p >>>= 1) ? 0 :
- a > (p >>>= 1) ? 1 :
- a > (p >>>= 1) ? 2 :
- a > (p >>>= 1) ? 4 :
- 8);
+ int a = p + (int)(ctl >> AC_SHIFT);
+ return (a > (p >>>= 1) ? 0 :
+ a > (p >>>= 1) ? 1 :
+ a > (p >>>= 1) ? 2 :
+ a > (p >>>= 1) ? 4 :
+ 8);
+ }
+
+ /**
+ * Returns approximate submission queue length for the given caller
+ */
+ static int getEstimatedSubmitterQueueLength() {
+ ForkJoinPool p; WorkQueue[] ws; WorkQueue q;
+ int k = submitters.get().seed & SQMASK;
+ return ((p = commonPool) != null && (ws = p.workQueues) != null &&
+ ws.length > (k &= p.submitMask) &&
+ (q = ws[k]) != null) ?
+ q.queueSize() : 0;
+ }
+
+ // Termination
+
+ /**
+ * Possibly initiates and/or completes termination. The caller
+ * triggering termination runs three passes through workQueues:
+ * (0) Setting termination status, followed by wakeups of queued
+ * workers; (1) cancelling all tasks; (2) interrupting lagging
+ * threads (likely in external tasks, but possibly also blocked in
+ * joins). Each pass repeats previous steps because of potential
+ * lagging thread creation.
+ *
+ * @param now if true, unconditionally terminate, else only
+ * if no work and no active workers
+ * @param enable if true, enable shutdown when next possible
+ * @return true if now terminating or terminated
+ */
+ private boolean tryTerminate(boolean now, boolean enable) {
+ for (long c;;) {
+ if (((c = ctl) & STOP_BIT) != 0) { // already terminating
+ if ((short)(c >>> TC_SHIFT) == -parallelism) {
+ synchronized (this) {
+ notifyAll(); // signal when 0 workers
+ }
+ }
+ return true;
+ }
+ if (runState >= 0) { // not yet enabled
+ if (!enable)
+ return false;
+ while (!U.compareAndSwapInt(this, MAINLOCK, 0, 1))
+ tryAwaitMainLock();
+ try {
+ runState |= SHUTDOWN;
+ } finally {
+ if (!U.compareAndSwapInt(this, MAINLOCK, 1, 0)) {
+ mainLock = 0;
+ synchronized (this) { notifyAll(); };
+ }
+ }
+ }
+ if (!now) { // check if idle & no tasks
+ if ((int)(c >> AC_SHIFT) != -parallelism ||
+ hasQueuedSubmissions())
+ return false;
+ // Check for unqueued inactive workers. One pass suffices.
+ WorkQueue[] ws = workQueues; WorkQueue w;
+ if (ws != null) {
+ for (int i = 1; i < ws.length; i += 2) {
+ if ((w = ws[i]) != null && w.eventCount >= 0)
+ return false;
+ }
+ }
+ }
+ if (U.compareAndSwapLong(this, CTL, c, c | STOP_BIT)) {
+ for (int pass = 0; pass < 3; ++pass) {
+ WorkQueue[] ws = workQueues;
+ if (ws != null) {
+ WorkQueue w;
+ int n = ws.length;
+ for (int i = 0; i < n; ++i) {
+ if ((w = ws[i]) != null) {
+ w.runState = -1;
+ if (pass > 0) {
+ w.cancelAll();
+ if (pass > 1)
+ w.interruptOwner();
+ }
+ }
+ }
+ // Wake up workers parked on event queue
+ int i, e; long cc; Thread p;
+ while ((e = (int)(cc = ctl) & E_MASK) != 0 &&
+ (i = e & SMASK) < n &&
+ (w = ws[i]) != null) {
+ long nc = ((long)(w.nextWait & E_MASK) |
+ ((cc + AC_UNIT) & AC_MASK) |
+ (cc & (TC_MASK|STOP_BIT)));
+ if (w.eventCount == (e | INT_SIGN) &&
+ U.compareAndSwapLong(this, CTL, cc, nc)) {
+ w.eventCount = (e + E_SEQ) & E_MASK;
+ w.runState = -1;
+ if ((p = w.parker) != null)
+ U.unpark(p);
+ }
+ }
+ }
+ }
+ }
+ }
}
// Exported methods
@@ -1409,29 +2396,54 @@ public class ForkJoinPool extends Abstra
checkPermission();
if (factory == null)
throw new NullPointerException();
- if (parallelism <= 0 || parallelism > MAX_ID)
+ if (parallelism <= 0 || parallelism > MAX_CAP)
throw new IllegalArgumentException();
this.parallelism = parallelism;
this.factory = factory;
this.ueh = handler;
- this.locallyFifo = asyncMode;
+ this.localMode = asyncMode ? FIFO_QUEUE : LIFO_QUEUE;
long np = (long)(-parallelism); // offset ctl counts
this.ctl = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK);
- this.submissionQueue = new ForkJoinTask[INITIAL_QUEUE_CAPACITY];
- // initialize workers array with room for 2*parallelism if possible
- int n = parallelism << 1;
- if (n >= MAX_ID)
- n = MAX_ID;
- else { // See Hackers Delight, sec 3.2, where n < (1 << 16)
- n |= n >>> 1; n |= n >>> 2; n |= n >>> 4; n |= n >>> 8;
- }
- workers = new ForkJoinWorkerThread[n + 1];
- this.submissionLock = new ReentrantLock();
- this.termination = submissionLock.newCondition();
+ // Use nearest power 2 for workQueues size. See Hackers Delight sec 3.2.
+ int n = parallelism - 1;
+ n |= n >>> 1; n |= n >>> 2; n |= n >>> 4; n |= n >>> 8; n |= n >>> 16;
+ this.submitMask = ((n + 1) << 1) - 1;
+ int pn = poolNumberGenerator.incrementAndGet();
StringBuilder sb = new StringBuilder("ForkJoinPool-");
- sb.append(poolNumberGenerator.incrementAndGet());
+ sb.append(Integer.toString(pn));
sb.append("-worker-");
this.workerNamePrefix = sb.toString();
+ this.runState = 1; // set init flag
+ }
+
+ /**
+ * Constructor for common pool, suitable only for static initialization.
+ * Basically the same as above, but uses smallest possible initial footprint.
+ */
+ ForkJoinPool(int parallelism, int submitMask,
+ ForkJoinWorkerThreadFactory factory,
+ Thread.UncaughtExceptionHandler handler) {
+ this.factory = factory;
+ this.ueh = handler;
+ this.submitMask = submitMask;
+ this.parallelism = parallelism;
+ long np = (long)(-parallelism);
+ this.ctl = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK);
+ this.localMode = LIFO_QUEUE;
+ this.workerNamePrefix = "ForkJoinPool.commonPool-worker-";
+ this.runState = 1;
+ }
+
+ /**
+ * Returns the common pool instance.
+ *
+ * @return the common pool instance
+ */
+ public static ForkJoinPool commonPool() {
+ ForkJoinPool p;
+ if ((p = commonPool) == null)
+ throw new Error("Common Pool Unavailable");
+ return p;
}
// Execution methods
@@ -1453,34 +2465,10 @@ public class ForkJoinPool extends Abstra
* scheduled for execution
*/
public T invoke(ForkJoinTask task) {
- Thread t = Thread.currentThread();
if (task == null)
throw new NullPointerException();
- if (shutdown)
- throw new RejectedExecutionException();
- if ((t instanceof ForkJoinWorkerThread) &&
- ((ForkJoinWorkerThread)t).pool == this)
- return task.invoke(); // bypass submit if in same pool
- else {
- addSubmission(task);
- return task.join();
- }
- }
-
- /**
- * Unless terminating, forks task if within an ongoing FJ
- * computation in the current pool, else submits as external task.
- */
- private void forkOrSubmit(ForkJoinTask task) {
- ForkJoinWorkerThread w;
- Thread t = Thread.currentThread();
- if (shutdown)
- throw new RejectedExecutionException();
- if ((t instanceof ForkJoinWorkerThread) &&
- (w = (ForkJoinWorkerThread)t).pool == this)
- w.pushTask(task);
- else
- addSubmission(task);
+ doSubmit(task);
+ return task.join();
}
/**
@@ -1494,7 +2482,7 @@ public class ForkJoinPool extends Abstra
public void execute(ForkJoinTask task) {
if (task == null)
throw new NullPointerException();
- forkOrSubmit(task);
+ doSubmit(task);
}
// AbstractExecutorService methods
@@ -1511,8 +2499,8 @@ public class ForkJoinPool extends Abstra
if (task instanceof ForkJoinTask) // avoid re-wrap
job = (ForkJoinTask) task;
else
- job = ForkJoinTask.adapt(task, null);
- forkOrSubmit(job);
+ job = new ForkJoinTask.AdaptedRunnableAction(task);
+ doSubmit(job);
}
/**
@@ -1527,7 +2515,7 @@ public class ForkJoinPool extends Abstra
public ForkJoinTask submit(ForkJoinTask task) {
if (task == null)
throw new NullPointerException();
- forkOrSubmit(task);
+ doSubmit(task);
return task;
}
@@ -1537,10 +2525,8 @@ public class ForkJoinPool extends Abstra
* scheduled for execution
*/
public ForkJoinTask submit(Callable task) {
- if (task == null)
- throw new NullPointerException();
- ForkJoinTask job = ForkJoinTask.adapt(task);
- forkOrSubmit(job);
+ ForkJoinTask job = new ForkJoinTask.AdaptedCallable(task);
+ doSubmit(job);
return job;
}
@@ -1550,10 +2536,8 @@ public class ForkJoinPool extends Abstra
* scheduled for execution
*/
public ForkJoinTask submit(Runnable task, T result) {
- if (task == null)
- throw new NullPointerException();
- ForkJoinTask job = ForkJoinTask.adapt(task, result);
- forkOrSubmit(job);
+ ForkJoinTask job = new ForkJoinTask.AdaptedRunnable(task, result);
+ doSubmit(job);
return job;
}
@@ -1569,8 +2553,8 @@ public class ForkJoinPool extends Abstra
if (task instanceof ForkJoinTask) // avoid re-wrap
job = (ForkJoinTask) task;
else
- job = ForkJoinTask.adapt(task, null);
- forkOrSubmit(job);
+ job = new ForkJoinTask.AdaptedRunnableAction(task);
+ doSubmit(job);
return job;
}
@@ -1579,25 +2563,31 @@ public class ForkJoinPool extends Abstra
* @throws RejectedExecutionException {@inheritDoc}
*/
public List> invokeAll(Collection> tasks) {
- ArrayList> forkJoinTasks =
- new ArrayList>(tasks.size());
- for (Callable task : tasks)
- forkJoinTasks.add(ForkJoinTask.adapt(task));
- invoke(new InvokeAll(forkJoinTasks));
-
+ // In previous versions of this class, this method constructed
+ // a task to run ForkJoinTask.invokeAll, but now external
+ // invocation of multiple tasks is at least as efficient.
+ List> fs = new ArrayList>(tasks.size());
+ // Workaround needed because method wasn't declared with
+ // wildcards in return type but should have been.
@SuppressWarnings({"unchecked", "rawtypes"})
- List> futures = (List>) (List) forkJoinTasks;
- return futures;
- }
+ List> futures = (List>) (List) fs;
- static final class InvokeAll extends RecursiveAction {
- final ArrayList> tasks;
- InvokeAll(ArrayList> tasks) { this.tasks = tasks; }
- public void compute() {
- try { invokeAll(tasks); }
- catch (Exception ignore) {}
+ boolean done = false;
+ try {
+ for (Callable t : tasks) {
+ ForkJoinTask f = new ForkJoinTask.AdaptedCallable(t);
+ doSubmit(f);
+ fs.add(f);
+ }
+ for (ForkJoinTask f : fs)
+ f.quietlyJoin();
+ done = true;
+ return futures;
+ } finally {
+ if (!done)
+ for (ForkJoinTask f : fs)
+ f.cancel(false);
}
- private static final long serialVersionUID = -7914297376763021607L;
}
/**
@@ -1629,6 +2619,15 @@ public class ForkJoinPool extends Abstra
}
/**
+ * Returns the targeted parallelism level of the common pool.
+ *
+ * @return the targeted parallelism level of the common pool
+ */
+ public static int getCommonPoolParallelism() {
+ return commonPoolParallelism;
+ }
+
+ /**
* Returns the number of worker threads that have started but not
* yet terminated. The result returned by this method may differ
* from {@link #getParallelism} when threads are created to
@@ -1647,7 +2646,7 @@ public class ForkJoinPool extends Abstra
* @return {@code true} if this pool uses async mode
*/
public boolean getAsyncMode() {
- return locallyFifo;
+ return localMode != 0;
}
/**
@@ -1659,8 +2658,15 @@ public class ForkJoinPool extends Abstra
* @return the number of worker threads
*/
public int getRunningThreadCount() {
- int r = parallelism + (int)(ctl >> AC_SHIFT);
- return r <= 0? 0 : r; // suppress momentarily negative values
+ int rc = 0;
+ WorkQueue[] ws; WorkQueue w;
+ if ((ws = workQueues) != null) {
+ for (int i = 1; i < ws.length; i += 2) {
+ if ((w = ws[i]) != null && w.isApparentlyUnblocked())
+ ++rc;
+ }
+ }
+ return rc;
}
/**
@@ -1671,8 +2677,8 @@ public class ForkJoinPool extends Abstra
* @return the number of active threads
*/
public int getActiveThreadCount() {
- int r = parallelism + (int)(ctl >> AC_SHIFT) + blockedCount;
- return r <= 0? 0 : r; // suppress momentarily negative values
+ int r = parallelism + (int)(ctl >> AC_SHIFT);
+ return (r <= 0) ? 0 : r; // suppress momentarily negative values
}
/**
@@ -1687,7 +2693,7 @@ public class ForkJoinPool extends Abstra
* @return {@code true} if all threads are currently idle
*/
public boolean isQuiescent() {
- return parallelism + (int)(ctl >> AC_SHIFT) + blockedCount == 0;
+ return (int)(ctl >> AC_SHIFT) + parallelism == 0;
}
/**
@@ -1702,7 +2708,15 @@ public class ForkJoinPool extends Abstra
* @return the number of steals
*/
public long getStealCount() {
- return stealCount;
+ long count = stealCount;
+ WorkQueue[] ws; WorkQueue w;
+ if ((ws = workQueues) != null) {
+ for (int i = 1; i < ws.length; i += 2) {
+ if ((w = ws[i]) != null)
+ count += w.totalSteals;
+ }
+ }
+ return count;
}
/**
@@ -1717,25 +2731,33 @@ public class ForkJoinPool extends Abstra
*/
public long getQueuedTaskCount() {
long count = 0;
- ForkJoinWorkerThread[] ws;
- if ((short)(ctl >>> TC_SHIFT) > -parallelism &&
- (ws = workers) != null) {
- for (ForkJoinWorkerThread w : ws)
- if (w != null)
- count -= w.queueBase - w.queueTop; // must read base first
+ WorkQueue[] ws; WorkQueue w;
+ if ((ws = workQueues) != null) {
+ for (int i = 1; i < ws.length; i += 2) {
+ if ((w = ws[i]) != null)
+ count += w.queueSize();
+ }
}
return count;
}
/**
* Returns an estimate of the number of tasks submitted to this
- * pool that have not yet begun executing. This meThod may take
+ * pool that have not yet begun executing. This method may take
* time proportional to the number of submissions.
*
* @return the number of queued submissions
*/
public int getQueuedSubmissionCount() {
- return -queueBase + queueTop;
+ int count = 0;
+ WorkQueue[] ws; WorkQueue w;
+ if ((ws = workQueues) != null) {
+ for (int i = 0; i < ws.length; i += 2) {
+ if ((w = ws[i]) != null)
+ count += w.queueSize();
+ }
+ }
+ return count;
}
/**
@@ -1745,7 +2767,14 @@ public class ForkJoinPool extends Abstra
* @return {@code true} if there are any queued submissions
*/
public boolean hasQueuedSubmissions() {
- return queueBase != queueTop;
+ WorkQueue[] ws; WorkQueue w;
+ if ((ws = workQueues) != null) {
+ for (int i = 0; i < ws.length; i += 2) {
+ if ((w = ws[i]) != null && !w.isEmpty())
+ return true;
+ }
+ }
+ return false;
}
/**
@@ -1756,16 +2785,11 @@ public class ForkJoinPool extends Abstra
* @return the next submission, or {@code null} if none
*/
protected ForkJoinTask pollSubmission() {
- ForkJoinTask t; ForkJoinTask[] q; int b, i;
- while ((b = queueBase) != queueTop &&
- (q = submissionQueue) != null &&
- (i = (q.length - 1) & b) >= 0) {
- long u = (i << ASHIFT) + ABASE;
- if ((t = q[i]) != null &&
- queueBase == b &&
- UNSAFE.compareAndSwapObject(q, u, t, null)) {
- queueBase = b + 1;
- return t;
+ WorkQueue[] ws; WorkQueue w; ForkJoinTask t;
+ if ((ws = workQueues) != null) {
+ for (int i = 0; i < ws.length; i += 2) {
+ if ((w = ws[i]) != null && (t = w.poll()) != null)
+ return t;
}
}
return null;
@@ -1790,20 +2814,17 @@ public class ForkJoinPool extends Abstra
*/
protected int drainTasksTo(Collection> c) {
int count = 0;
- while (queueBase != queueTop) {
- ForkJoinTask t = pollSubmission();
- if (t != null) {
- c.add(t);
- ++count;
+ WorkQueue[] ws; WorkQueue w; ForkJoinTask t;
+ if ((ws = workQueues) != null) {
+ for (int i = 0; i < ws.length; ++i) {
+ if ((w = ws[i]) != null) {
+ while ((t = w.poll()) != null) {
+ c.add(t);
+ ++count;
+ }
+ }
}
}
- ForkJoinWorkerThread[] ws;
- if ((short)(ctl >>> TC_SHIFT) > -parallelism &&
- (ws = workers) != null) {
- for (ForkJoinWorkerThread w : ws)
- if (w != null)
- count += w.drainTasksTo(c);
- }
return count;
}
@@ -1815,21 +2836,36 @@ public class ForkJoinPool extends Abstra
* @return a string identifying this pool, as well as its state
*/
public String toString() {
- long st = getStealCount();
- long qt = getQueuedTaskCount();
- long qs = getQueuedSubmissionCount();
- int pc = parallelism;
+ // Use a single pass through workQueues to collect counts
+ long qt = 0L, qs = 0L; int rc = 0;
+ long st = stealCount;
long c = ctl;
+ WorkQueue[] ws; WorkQueue w;
+ if ((ws = workQueues) != null) {
+ for (int i = 0; i < ws.length; ++i) {
+ if ((w = ws[i]) != null) {
+ int size = w.queueSize();
+ if ((i & 1) == 0)
+ qs += size;
+ else {
+ qt += size;
+ st += w.totalSteals;
+ if (w.isApparentlyUnblocked())
+ ++rc;
+ }
+ }
+ }
+ }
+ int pc = parallelism;
int tc = pc + (short)(c >>> TC_SHIFT);
- int rc = pc + (int)(c >> AC_SHIFT);
- if (rc < 0) // ignore transient negative
- rc = 0;
- int ac = rc + blockedCount;
+ int ac = pc + (int)(c >> AC_SHIFT);
+ if (ac < 0) // ignore transient negative
+ ac = 0;
String level;
if ((c & STOP_BIT) != 0)
- level = (tc == 0)? "Terminated" : "Terminating";
+ level = (tc == 0) ? "Terminated" : "Terminating";
else
- level = shutdown? "Shutting down" : "Running";
+ level = runState < 0 ? "Shutting down" : "Running";
return super.toString() +
"[" + level +
", parallelism = " + pc +
@@ -1843,11 +2879,13 @@ public class ForkJoinPool extends Abstra
}
/**
- * Initiates an orderly shutdown in which previously submitted
- * tasks are executed, but no new tasks will be accepted.
- * Invocation has no additional effect if already shut down.
- * Tasks that are in the process of being submitted concurrently
- * during the course of this method may or may not be rejected.
+ * Possibly initiates an orderly shutdown in which previously
+ * submitted tasks are executed, but no new tasks will be
+ * accepted. Invocation has no effect on execution state if this
+ * is the {@link #commonPool}, and no additional effect if
+ * already shut down. Tasks that are in the process of being
+ * submitted concurrently during the course of this method may or
+ * may not be rejected.
*
* @throws SecurityException if a security manager exists and
* the caller is not permitted to modify threads
@@ -1856,19 +2894,21 @@ public class ForkJoinPool extends Abstra
*/
public void shutdown() {
checkPermission();
- shutdown = true;
- tryTerminate(false);
+ if (this != commonPool)
+ tryTerminate(false, true);
}
/**
- * Attempts to cancel and/or stop all tasks, and reject all
- * subsequently submitted tasks. Tasks that are in the process of
- * being submitted or executed concurrently during the course of
- * this method may or may not be rejected. This method cancels
- * both existing and unexecuted tasks, in order to permit
- * termination in the presence of task dependencies. So the method
- * always returns an empty list (unlike the case for some other
- * Executors).
+ * Possibly attempts to cancel and/or stop all tasks, and reject
+ * all subsequently submitted tasks. Invocation has no effect on
+ * execution state if this is the {@link #commonPool}, and no
+ * additional effect if already shut down. Otherwise, tasks that
+ * are in the process of being submitted or executed concurrently
+ * during the course of this method may or may not be
+ * rejected. This method cancels both existing and unexecuted
+ * tasks, in order to permit termination in the presence of task
+ * dependencies. So the method always returns an empty list
+ * (unlike the case for some other Executors).
*
* @return an empty list
* @throws SecurityException if a security manager exists and
@@ -1878,8 +2918,8 @@ public class ForkJoinPool extends Abstra
*/
public List shutdownNow() {
checkPermission();
- shutdown = true;
- tryTerminate(true);
+ if (this != commonPool)
+ tryTerminate(true, true);
return Collections.emptyList();
}
@@ -1914,19 +2954,12 @@ public class ForkJoinPool extends Abstra
}
/**
- * Returns true if terminating or terminated. Used by ForkJoinWorkerThread.
- */
- final boolean isAtLeastTerminating() {
- return (ctl & STOP_BIT) != 0L;
- }
-
- /**
* Returns {@code true} if this pool has been shut down.
*
* @return {@code true} if this pool has been shut down
*/
public boolean isShutdown() {
- return shutdown;
+ return runState < 0;
}
/**
@@ -1943,19 +2976,21 @@ public class ForkJoinPool extends Abstra
public boolean awaitTermination(long timeout, TimeUnit unit)
throws InterruptedException {
long nanos = unit.toNanos(timeout);
- final ReentrantLock lock = this.submissionLock;
- lock.lock();
- try {
- for (;;) {
- if (isTerminated())
- return true;
- if (nanos <= 0)
- return false;
- nanos = termination.awaitNanos(nanos);
+ if (isTerminated())
+ return true;
+ long startTime = System.nanoTime();
+ boolean terminated = false;
+ synchronized (this) {
+ for (long waitTime = nanos, millis = 0L;;) {
+ if (terminated = isTerminated() ||
+ waitTime <= 0L ||
+ (millis = unit.toMillis(waitTime)) <= 0L)
+ break;
+ wait(millis);
+ waitTime = nanos - (System.nanoTime() - startTime);
}
- } finally {
- lock.unlock();
}
+ return terminated;
}
/**
@@ -1966,13 +3001,15 @@ public class ForkJoinPool extends Abstra
* {@code isReleasable} must return {@code true} if blocking is
* not necessary. Method {@code block} blocks the current thread
* if necessary (perhaps internally invoking {@code isReleasable}
- * before actually blocking). The unusual methods in this API
- * accommodate synchronizers that may, but don't usually, block
- * for long periods. Similarly, they allow more efficient internal
- * handling of cases in which additional workers may be, but
- * usually are not, needed to ensure sufficient parallelism.
- * Toward this end, implementations of method {@code isReleasable}
- * must be amenable to repeated invocation.
+ * before actually blocking). These actions are performed by any
+ * thread invoking {@link ForkJoinPool#managedBlock}. The
+ * unusual methods in this API accommodate synchronizers that may,
+ * but don't usually, block for long periods. Similarly, they
+ * allow more efficient internal handling of cases in which
+ * additional workers may be, but usually are not, needed to
+ * ensure sufficient parallelism. Toward this end,
+ * implementations of method {@code isReleasable} must be amenable
+ * to repeated invocation.
*
* For example, here is a ManagedBlocker based on a
* ReentrantLock:
@@ -2052,12 +3089,18 @@ public class ForkJoinPool extends Abstra
public static void managedBlock(ManagedBlocker blocker)
throws InterruptedException {
Thread t = Thread.currentThread();
- if (t instanceof ForkJoinWorkerThread) {
- ForkJoinWorkerThread w = (ForkJoinWorkerThread) t;
- w.pool.awaitBlocker(blocker);
- }
- else {
- do {} while (!blocker.isReleasable() && !blocker.block());
+ ForkJoinPool p = ((t instanceof ForkJoinWorkerThread) ?
+ ((ForkJoinWorkerThread)t).pool : null);
+ while (!blocker.isReleasable()) {
+ if (p == null || p.tryCompensate(null, blocker)) {
+ try {
+ do {} while (!blocker.isReleasable() && !blocker.block());
+ } finally {
+ if (p != null)
+ p.incrementActiveCount();
+ }
+ break;
+ }
}
}
@@ -2066,55 +3109,79 @@ public class ForkJoinPool extends Abstra
// implement RunnableFuture.
protected RunnableFuture newTaskFor(Runnable runnable, T value) {
- return (RunnableFuture) ForkJoinTask.adapt(runnable, value);
+ return new ForkJoinTask.AdaptedRunnable(runnable, value);
}
protected RunnableFuture newTaskFor(Callable callable) {
- return (RunnableFuture) ForkJoinTask.adapt(callable);
+ return new ForkJoinTask.AdaptedCallable(callable);
}
// Unsafe mechanics
- private static final sun.misc.Unsafe UNSAFE;
- private static final long ctlOffset;
- private static final long stealCountOffset;
- private static final long blockedCountOffset;
- private static final long quiescerCountOffset;
- private static final long scanGuardOffset;
- private static final long nextWorkerNumberOffset;
- private static final long ABASE;
+ private static final sun.misc.Unsafe U;
+ private static final long CTL;
+ private static final long PARKBLOCKER;
+ private static final int ABASE;
private static final int ASHIFT;
+ private static final long NEXTWORKERNUMBER;
+ private static final long STEALCOUNT;
+ private static final long MAINLOCK;
static {
poolNumberGenerator = new AtomicInteger();
- workerSeedGenerator = new Random();
+ nextSubmitterSeed = new AtomicInteger(0x55555555);
modifyThreadPermission = new RuntimePermission("modifyThread");
defaultForkJoinWorkerThreadFactory =
new DefaultForkJoinWorkerThreadFactory();
+ submitters = new ThreadSubmitter();
int s;
try {
- UNSAFE = getUnsafe();
- Class k = ForkJoinPool.class;
- ctlOffset = UNSAFE.objectFieldOffset
+ U = getUnsafe();
+ Class k = ForkJoinPool.class;
+ Class ak = ForkJoinTask[].class;
+ CTL = U.objectFieldOffset
(k.getDeclaredField("ctl"));
- stealCountOffset = UNSAFE.objectFieldOffset
- (k.getDeclaredField("stealCount"));
- blockedCountOffset = UNSAFE.objectFieldOffset
- (k.getDeclaredField("blockedCount"));
- quiescerCountOffset = UNSAFE.objectFieldOffset
- (k.getDeclaredField("quiescerCount"));
- scanGuardOffset = UNSAFE.objectFieldOffset
- (k.getDeclaredField("scanGuard"));
- nextWorkerNumberOffset = UNSAFE.objectFieldOffset
+ NEXTWORKERNUMBER = U.objectFieldOffset
(k.getDeclaredField("nextWorkerNumber"));
- Class a = ForkJoinTask[].class;
- ABASE = UNSAFE.arrayBaseOffset(a);
- s = UNSAFE.arrayIndexScale(a);
+ STEALCOUNT = U.objectFieldOffset
+ (k.getDeclaredField("stealCount"));
+ MAINLOCK = U.objectFieldOffset
+ (k.getDeclaredField("mainLock"));
+ Class tk = Thread.class;
+ PARKBLOCKER = U.objectFieldOffset
+ (tk.getDeclaredField("parkBlocker"));
+ ABASE = U.arrayBaseOffset(ak);
+ s = U.arrayIndexScale(ak);
+ ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
} catch (Exception e) {
throw new Error(e);
}
if ((s & (s-1)) != 0)
throw new Error("data type scale not a power of two");
- ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
+ try { // Establish common pool
+ String pp = System.getProperty(propPrefix + "parallelism");
+ String fp = System.getProperty(propPrefix + "threadFactory");
+ String up = System.getProperty(propPrefix + "exceptionHandler");
+ ForkJoinWorkerThreadFactory fac = (fp == null) ?
+ defaultForkJoinWorkerThreadFactory :
+ ((ForkJoinWorkerThreadFactory)ClassLoader.
+ getSystemClassLoader().loadClass(fp).newInstance());
+ Thread.UncaughtExceptionHandler ueh = (up == null) ? null :
+ ((Thread.UncaughtExceptionHandler)ClassLoader.
+ getSystemClassLoader().loadClass(up).newInstance());
+ int par;
+ if ((pp == null || (par = Integer.parseInt(pp)) <= 0))
+ par = Runtime.getRuntime().availableProcessors();
+ if (par > MAX_CAP)
+ par = MAX_CAP;
+ commonPoolParallelism = par;
+ int n = par - 1; // precompute submit mask
+ n |= n >>> 1; n |= n >>> 2; n |= n >>> 4;
+ n |= n >>> 8; n |= n >>> 16;
+ int mask = ((n + 1) << 1) - 1;
+ commonPool = new ForkJoinPool(par, mask, fac, ueh);
+ } catch (Exception e) {
+ throw new Error(e);
+ }
}
/**
@@ -2144,4 +3211,5 @@ public class ForkJoinPool extends Abstra
}
}
}
+
}