1 /*
2  * Written by Doug Lea with assistance from members of JCP JSR-166
3  * Expert Group and released to the public domain, as explained at
4  * http://creativecommons.org/publicdomain/zero/1.0/
5  */
6 
7 package java.util.concurrent;
8 
9 import java.lang.Thread.UncaughtExceptionHandler;
10 import java.util.ArrayList;
11 import java.util.Arrays;
12 import java.util.Collection;
13 import java.util.Collections;
14 import java.util.List;
15 import java.util.concurrent.AbstractExecutorService;
16 import java.util.concurrent.Callable;
17 import java.util.concurrent.ExecutorService;
18 import java.util.concurrent.Future;
19 import java.util.concurrent.RejectedExecutionException;
20 import java.util.concurrent.RunnableFuture;
21 import java.util.concurrent.ThreadLocalRandom;
22 import java.util.concurrent.TimeUnit;
23 
24 /**
25  * An {@link ExecutorService} for running {@link ForkJoinTask}s.
26  * A {@code ForkJoinPool} provides the entry point for submissions
27  * from non-{@code ForkJoinTask} clients, as well as management and
28  * monitoring operations.
29  *
30  * <p>A {@code ForkJoinPool} differs from other kinds of {@link
31  * ExecutorService} mainly by virtue of employing
32  * <em>work-stealing</em>: all threads in the pool attempt to find and
33  * execute tasks submitted to the pool and/or created by other active
34  * tasks (eventually blocking waiting for work if none exist). This
35  * enables efficient processing when most tasks spawn other subtasks
36  * (as do most {@code ForkJoinTask}s), as well as when many small
37  * tasks are submitted to the pool from external clients.  Especially
38  * when setting <em>asyncMode</em> to true in constructors, {@code
39  * ForkJoinPool}s may also be appropriate for use with event-style
40  * tasks that are never joined.
41  *
42  * <p>A static {@code commonPool()} is available and appropriate for
43  * most applications. The common pool is used by any ForkJoinTask that
44  * is not explicitly submitted to a specified pool. Using the common
45  * pool normally reduces resource usage (its threads are slowly
46  * reclaimed during periods of non-use, and reinstated upon subsequent
47  * use).
48  *
49  * <p>For applications that require separate or custom pools, a {@code
50  * ForkJoinPool} may be constructed with a given target parallelism
51  * level; by default, equal to the number of available processors. The
52  * pool attempts to maintain enough active (or available) threads by
53  * dynamically adding, suspending, or resuming internal worker
54  * threads, even if some tasks are stalled waiting to join others.
55  * However, no such adjustments are guaranteed in the face of blocked
56  * I/O or other unmanaged synchronization. The nested {@link
57  * ManagedBlocker} interface enables extension of the kinds of
58  * synchronization accommodated.
59  *
60  * <p>In addition to execution and lifecycle control methods, this
61  * class provides status check methods (for example
62  * {@link #getStealCount}) that are intended to aid in developing,
63  * tuning, and monitoring fork/join applications. Also, method
64  * {@link #toString} returns indications of pool state in a
65  * convenient form for informal monitoring.
66  *
67  * <p>As is the case with other ExecutorServices, there are three
68  * main task execution methods summarized in the following table.
69  * These are designed to be used primarily by clients not already
70  * engaged in fork/join computations in the current pool.  The main
71  * forms of these methods accept instances of {@code ForkJoinTask},
72  * but overloaded forms also allow mixed execution of plain {@code
73  * Runnable}- or {@code Callable}- based activities as well.  However,
74  * tasks that are already executing in a pool should normally instead
75  * use the within-computation forms listed in the table unless using
76  * async event-style tasks that are not usually joined, in which case
77  * there is little difference among choice of methods.
78  *
79  * <table BORDER CELLPADDING=3 CELLSPACING=1>
80  * <caption>Summary of task execution methods</caption>
81  *  <tr>
82  *    <td></td>
83  *    <td ALIGN=CENTER> <b>Call from non-fork/join clients</b></td>
84  *    <td ALIGN=CENTER> <b>Call from within fork/join computations</b></td>
85  *  </tr>
86  *  <tr>
87  *    <td> <b>Arrange async execution</b></td>
88  *    <td> {@link #execute(ForkJoinTask)}</td>
89  *    <td> {@link ForkJoinTask#fork}</td>
90  *  </tr>
91  *  <tr>
92  *    <td> <b>Await and obtain result</b></td>
93  *    <td> {@link #invoke(ForkJoinTask)}</td>
94  *    <td> {@link ForkJoinTask#invoke}</td>
95  *  </tr>
96  *  <tr>
97  *    <td> <b>Arrange exec and obtain Future</b></td>
98  *    <td> {@link #submit(ForkJoinTask)}</td>
99  *    <td> {@link ForkJoinTask#fork} (ForkJoinTasks <em>are</em> Futures)</td>
100  *  </tr>
101  * </table>
102  *
103  * <p>The common pool is by default constructed with default
104  * parameters, but these may be controlled by setting three
105  * {@linkplain System#getProperty system properties}:
106  * <ul>
107  * <li>{@code java.util.concurrent.ForkJoinPool.common.parallelism}
108  * - the parallelism level, a non-negative integer
109  * <li>{@code java.util.concurrent.ForkJoinPool.common.threadFactory}
110  * - the class name of a {@link ForkJoinWorkerThreadFactory}
111  * <li>{@code java.util.concurrent.ForkJoinPool.common.exceptionHandler}
112  * - the class name of a {@link UncaughtExceptionHandler}
113  * </ul>
114  * The system class loader is used to load these classes.
115  * Upon any error in establishing these settings, default parameters
116  * are used. It is possible to disable or limit the use of threads in
117  * the common pool by setting the parallelism property to zero, and/or
118  * using a factory that may return {@code null}.
119  *
120  * <p><b>Implementation notes</b>: This implementation restricts the
121  * maximum number of running threads to 32767. Attempts to create
122  * pools with greater than the maximum number result in
123  * {@code IllegalArgumentException}.
124  *
125  * <p>This implementation rejects submitted tasks (that is, by throwing
126  * {@link RejectedExecutionException}) only when the pool is shut down
127  * or internal resources have been exhausted.
128  *
129  * @since 1.7
130  * @author Doug Lea
131  */
132 public class ForkJoinPool extends AbstractExecutorService {
133 
134     /*
135      * Implementation Overview
136      *
137      * This class and its nested classes provide the main
138      * functionality and control for a set of worker threads:
139      * Submissions from non-FJ threads enter into submission queues.
140      * Workers take these tasks and typically split them into subtasks
141      * that may be stolen by other workers.  Preference rules give
142      * first priority to processing tasks from their own queues (LIFO
143      * or FIFO, depending on mode), then to randomized FIFO steals of
144      * tasks in other queues.
145      *
146      * WorkQueues
147      * ==========
148      *
149      * Most operations occur within work-stealing queues (in nested
150      * class WorkQueue).  These are special forms of Deques that
151      * support only three of the four possible end-operations -- push,
152      * pop, and poll (aka steal), under the further constraints that
153      * push and pop are called only from the owning thread (or, as
154      * extended here, under a lock), while poll may be called from
155      * other threads.  (If you are unfamiliar with them, you probably
156      * want to read Herlihy and Shavit's book "The Art of
157      * Multiprocessor programming", chapter 16 describing these in
158      * more detail before proceeding.)  The main work-stealing queue
159      * design is roughly similar to those in the papers "Dynamic
160      * Circular Work-Stealing Deque" by Chase and Lev, SPAA 2005
161      * (http://research.sun.com/scalable/pubs/index.html) and
162      * "Idempotent work stealing" by Michael, Saraswat, and Vechev,
163      * PPoPP 2009 (http://portal.acm.org/citation.cfm?id=1504186).
164      * See also "Correct and Efficient Work-Stealing for Weak Memory
165      * Models" by Le, Pop, Cohen, and Nardelli, PPoPP 2013
166      * (http://www.di.ens.fr/~zappa/readings/ppopp13.pdf) for an
167      * analysis of memory ordering (atomic, volatile etc) issues.  The
168      * main differences ultimately stem from GC requirements that we
169      * null out taken slots as soon as we can, to maintain as small a
170      * footprint as possible even in programs generating huge numbers
171      * of tasks. To accomplish this, we shift the CAS arbitrating pop
172      * vs poll (steal) from being on the indices ("base" and "top") to
173      * the slots themselves.  So, both a successful pop and poll
174      * mainly entail a CAS of a slot from non-null to null.  Because
175      * we rely on CASes of references, we do not need tag bits on base
176      * or top.  They are simple ints as used in any circular
177      * array-based queue (see for example ArrayDeque).  Updates to the
178      * indices must still be ordered in a way that guarantees that top
179      * == base means the queue is empty, but otherwise may err on the
180      * side of possibly making the queue appear nonempty when a push,
181      * pop, or poll have not fully committed. Note that this means
182      * that the poll operation, considered individually, is not
183      * wait-free. One thief cannot successfully continue until another
184      * in-progress one (or, if previously empty, a push) completes.
185      * However, in the aggregate, we ensure at least probabilistic
186      * non-blockingness.  If an attempted steal fails, a thief always
187      * chooses a different random victim target to try next. So, in
188      * order for one thief to progress, it suffices for any
189      * in-progress poll or new push on any empty queue to
190      * complete. (This is why we normally use method pollAt and its
191      * variants that try once at the apparent base index, else
192      * consider alternative actions, rather than method poll.)
193      *
194      * This approach also enables support of a user mode in which local
195      * task processing is in FIFO, not LIFO order, simply by using
196      * poll rather than pop.  This can be useful in message-passing
197      * frameworks in which tasks are never joined.  However neither
198      * mode considers affinities, loads, cache localities, etc, so
199      * rarely provide the best possible performance on a given
200      * machine, but portably provide good throughput by averaging over
201      * these factors.  (Further, even if we did try to use such
202      * information, we do not usually have a basis for exploiting it.
203      * For example, some sets of tasks profit from cache affinities,
204      * but others are harmed by cache pollution effects.)
205      *
206      * WorkQueues are also used in a similar way for tasks submitted
207      * to the pool. We cannot mix these tasks in the same queues used
208      * for work-stealing (this would contaminate lifo/fifo
209      * processing). Instead, we randomly associate submission queues
210      * with submitting threads, using a form of hashing.  The
211      * Submitter probe value serves as a hash code for
212      * choosing existing queues, and may be randomly repositioned upon
213      * contention with other submitters.  In essence, submitters act
214      * like workers except that they are restricted to executing local
215      * tasks that they submitted. However, because most
216      * shared/external queue operations are more expensive than
217      * internal, and because, at steady state, external submitters
218      * will compete for CPU with workers, ForkJoinTask.join and
219      * related methods disable them from repeatedly helping to process
220      * tasks if all workers are active.  Insertion of tasks in shared
221      * mode requires a lock (mainly to protect in the case of
222      * resizing) but we use only a simple spinlock (using bits in
223      * field qlock), because submitters encountering a busy queue move
224      * on to try or create other queues -- they block only when
225      * creating and registering new queues.
226      *
227      * Management
228      * ==========
229      *
230      * The main throughput advantages of work-stealing stem from
231      * decentralized control -- workers mostly take tasks from
232      * themselves or each other. We cannot negate this in the
233      * implementation of other management responsibilities. The main
234      * tactic for avoiding bottlenecks is packing nearly all
235      * essentially atomic control state into two volatile variables
236      * that are by far most often read (not written) as status and
237      * consistency checks.
238      *
239      * Field "ctl" contains 64 bits holding all the information needed
240      * to atomically decide to add, inactivate, enqueue (on an event
241      * queue), dequeue, and/or re-activate workers.  To enable this
242      * packing, we restrict maximum parallelism to (1<<15)-1 (which is
243      * far in excess of normal operating range) to allow ids, counts,
244      * and their negations (used for thresholding) to fit into 16bit
245      * fields.
246      *
247      * Field "plock" is a form of sequence lock with a saturating
248      * shutdown bit (similarly for per-queue "qlocks"), mainly
249      * protecting updates to the workQueues array, as well as to
250      * enable shutdown.  When used as a lock, it is normally only very
251      * briefly held, so is nearly always available after at most a
252      * brief spin, but we use a monitor-based backup strategy to
253      * block when needed.
254      *
255      * Recording WorkQueues.  WorkQueues are recorded in the
256      * "workQueues" array that is created upon first use and expanded
257      * if necessary.  Updates to the array while recording new workers
258      * and unrecording terminated ones are protected from each other
259      * by a lock but the array is otherwise concurrently readable, and
260      * accessed directly.  To simplify index-based operations, the
261      * array size is always a power of two, and all readers must
262      * tolerate null slots. Worker queues are at odd indices. Shared
263      * (submission) queues are at even indices, up to a maximum of 64
264      * slots, to limit growth even if array needs to expand to add
265      * more workers. Grouping them together in this way simplifies and
266      * speeds up task scanning.
267      *
268      * All worker thread creation is on-demand, triggered by task
269      * submissions, replacement of terminated workers, and/or
270      * compensation for blocked workers. However, all other support
271      * code is set up to work with other policies.  To ensure that we
272      * do not hold on to worker references that would prevent GC, ALL
273      * accesses to workQueues are via indices into the workQueues
274      * array (which is one source of some of the messy code
275      * constructions here). In essence, the workQueues array serves as
276      * a weak reference mechanism. Thus for example the wait queue
277      * field of ctl stores indices, not references.  Access to the
278      * workQueues in associated methods (for example signalWork) must
279      * both index-check and null-check the IDs. All such accesses
280      * ignore bad IDs by returning out early from what they are doing,
281      * since this can only be associated with termination, in which
282      * case it is OK to give up.  All uses of the workQueues array
283      * also check that it is non-null (even if previously
284      * non-null). This allows nulling during termination, which is
285      * currently not necessary, but remains an option for
286      * resource-revocation-based shutdown schemes. It also helps
287      * reduce JIT issuance of uncommon-trap code, which tends to
288      * unnecessarily complicate control flow in some methods.
289      *
290      * Event Queuing. Unlike HPC work-stealing frameworks, we cannot
291      * let workers spin indefinitely scanning for tasks when none can
292      * be found immediately, and we cannot start/resume workers unless
293      * there appear to be tasks available.  On the other hand, we must
294      * quickly prod them into action when new tasks are submitted or
295      * generated. In many usages, ramp-up time to activate workers is
296      * the main limiting factor in overall performance (this is
297      * compounded at program start-up by JIT compilation and
298      * allocation). So we try to streamline this as much as possible.
299      * We park/unpark workers after placing in an event wait queue
300      * when they cannot find work. This "queue" is actually a simple
301      * Treiber stack, headed by the "id" field of ctl, plus a 15bit
302      * counter value (that reflects the number of times a worker has
303      * been inactivated) to avoid ABA effects (we need only as many
304      * version numbers as worker threads). Successors are held in
305      * field WorkQueue.nextWait.  Queuing deals with several intrinsic
306      * races, mainly that a task-producing thread can miss seeing (and
307      * signalling) another thread that gave up looking for work but
308      * has not yet entered the wait queue. We solve this by requiring
309      * a full sweep of all workers (via repeated calls to method
310      * scan()) both before and after a newly waiting worker is added
311      * to the wait queue.  Because enqueued workers may actually be
312      * rescanning rather than waiting, we set and clear the "parker"
313      * field of WorkQueues to reduce unnecessary calls to unpark.
314      * (This requires a secondary recheck to avoid missed signals.)
315      * Note the unusual conventions about Thread.interrupts
316      * surrounding parking and other blocking: Because interrupts are
317      * used solely to alert threads to check termination, which is
318      * checked anyway upon blocking, we clear status (using
319      * Thread.interrupted) before any call to park, so that park does
320      * not immediately return due to status being set via some other
321      * unrelated call to interrupt in user code.
322      *
323      * Signalling.  We create or wake up workers only when there
324      * appears to be at least one task they might be able to find and
325      * execute.  When a submission is added or another worker adds a
326      * task to a queue that has fewer than two tasks, they signal
327      * waiting workers (or trigger creation of new ones if fewer than
328      * the given parallelism level -- signalWork).  These primary
329      * signals are buttressed by others whenever other threads remove
330      * a task from a queue and notice that there are other tasks there
331      * as well.  So in general, pools will be over-signalled. On most
332      * platforms, signalling (unpark) overhead time is noticeably
333      * long, and the time between signalling a thread and it actually
334      * making progress can be very noticeably long, so it is worth
335      * offloading these delays from critical paths as much as
336      * possible. Additionally, workers spin-down gradually, by staying
337      * alive so long as they see the ctl state changing.  Similar
338      * stability-sensing techniques are also used before blocking in
339      * awaitJoin and helpComplete.
340      *
341      * Trimming workers. To release resources after periods of lack of
342      * use, a worker starting to wait when the pool is quiescent will
343      * time out and terminate if the pool has remained quiescent for a
344      * given period -- a short period if there are more threads than
345      * parallelism, longer as the number of threads decreases. This
346      * will slowly propagate, eventually terminating all workers after
347      * periods of non-use.
348      *
349      * Shutdown and Termination. A call to shutdownNow atomically sets
350      * a plock bit and then (non-atomically) sets each worker's
351      * qlock status, cancels all unprocessed tasks, and wakes up
352      * all waiting workers.  Detecting whether termination should
353      * commence after a non-abrupt shutdown() call requires more work
354      * and bookkeeping. We need consensus about quiescence (i.e., that
355      * there is no more work). The active count provides a primary
356      * indication but non-abrupt shutdown still requires a rechecking
357      * scan for any workers that are inactive but not queued.
358      *
359      * Joining Tasks
360      * =============
361      *
362      * Any of several actions may be taken when one worker is waiting
363      * to join a task stolen (or always held) by another.  Because we
364      * are multiplexing many tasks on to a pool of workers, we can't
365      * just let them block (as in Thread.join).  We also cannot just
366      * reassign the joiner's run-time stack with another and replace
367      * it later, which would be a form of "continuation", that even if
368      * possible is not necessarily a good idea since we sometimes need
369      * both an unblocked task and its continuation to progress.
370      * Instead we combine two tactics:
371      *
372      *   Helping: Arranging for the joiner to execute some task that it
373      *      would be running if the steal had not occurred.
374      *
375      *   Compensating: Unless there are already enough live threads,
376      *      method tryCompensate() may create or re-activate a spare
377      *      thread to compensate for blocked joiners until they unblock.
378      *
379      * A third form (implemented in tryRemoveAndExec) amounts to
380      * helping a hypothetical compensator: If we can readily tell that
381      * a possible action of a compensator is to steal and execute the
382      * task being joined, the joining thread can do so directly,
383      * without the need for a compensation thread (although at the
384      * expense of larger run-time stacks, but the tradeoff is
385      * typically worthwhile).
386      *
387      * The ManagedBlocker extension API can't use helping so relies
388      * only on compensation in method awaitBlocker.
389      *
390      * The algorithm in tryHelpStealer entails a form of "linear"
391      * helping: Each worker records (in field currentSteal) the most
392      * recent task it stole from some other worker. Plus, it records
393      * (in field currentJoin) the task it is currently actively
394      * joining. Method tryHelpStealer uses these markers to try to
395      * find a worker to help (i.e., steal back a task from and execute
396      * it) that could hasten completion of the actively joined task.
397      * In essence, the joiner executes a task that would be on its own
398      * local deque had the to-be-joined task not been stolen. This may
399      * be seen as a conservative variant of the approach in Wagner &
400      * Calder "Leapfrogging: a portable technique for implementing
401      * efficient futures" SIGPLAN Notices, 1993
402      * (http://portal.acm.org/citation.cfm?id=155354). It differs in
403      * that: (1) We only maintain dependency links across workers upon
404      * steals, rather than use per-task bookkeeping.  This sometimes
405      * requires a linear scan of workQueues array to locate stealers,
406      * but often doesn't because stealers leave hints (that may become
407      * stale/wrong) of where to locate them.  It is only a hint
408      * because a worker might have had multiple steals and the hint
409      * records only one of them (usually the most current).  Hinting
410      * isolates cost to when it is needed, rather than adding to
411      * per-task overhead.  (2) It is "shallow", ignoring nesting and
412      * potentially cyclic mutual steals.  (3) It is intentionally
413      * racy: field currentJoin is updated only while actively joining,
414      * which means that we miss links in the chain during long-lived
415      * tasks, GC stalls etc (which is OK since blocking in such cases
416      * is usually a good idea).  (4) We bound the number of attempts
417      * to find work (see MAX_HELP) and fall back to suspending the
418      * worker and if necessary replacing it with another.
419      *
420      * It is impossible to keep exactly the target parallelism number
421      * of threads running at any given time.  Determining the
422      * existence of conservatively safe helping targets, the
423      * availability of already-created spares, and the apparent need
424      * to create new spares are all racy, so we rely on multiple
425      * retries of each.  Compensation in the apparent absence of
426      * helping opportunities is challenging to control on JVMs, where
427      * GC and other activities can stall progress of tasks that in
428      * turn stall out many other dependent tasks, without us being
429      * able to determine whether they will ever require compensation.
430      * Even though work-stealing otherwise encounters little
431      * degradation in the presence of more threads than cores,
432      * aggressively adding new threads in such cases entails risk of
433      * unwanted positive feedback control loops in which more threads
434      * cause more dependent stalls (as well as delayed progress of
435      * unblocked threads to the point that we know they are available)
436      * leading to more situations requiring more threads, and so
437      * on. This aspect of control can be seen as an (analytically
438      * intractable) game with an opponent that may choose the worst
439      * (for us) active thread to stall at any time.  We take several
440      * precautions to bound losses (and thus bound gains), mainly in
441      * methods tryCompensate and awaitJoin.
442      *
443      * Common Pool
444      * ===========
445      *
446      * The static common pool always exists after static
447      * initialization.  Since it (or any other created pool) need
448      * never be used, we minimize initial construction overhead and
449      * footprint to the setup of about a dozen fields, with no nested
450      * allocation. Most bootstrapping occurs within method
451      * fullExternalPush during the first submission to the pool.
452      *
453      * When external threads submit to the common pool, they can
454      * perform subtask processing (see externalHelpJoin and related
455      * methods).  This caller-helps policy makes it sensible to set
456      * common pool parallelism level to one (or more) less than the
457      * total number of available cores, or even zero for pure
458      * caller-runs.  We do not need to record whether external
459      * submissions are to the common pool -- if not, externalHelpJoin
460      * returns quickly (at the most helping to signal some common pool
461      * workers). These submitters would otherwise be blocked waiting
462      * for completion, so the extra effort (with liberally sprinkled
463      * task status checks) in inapplicable cases amounts to an odd
464      * form of limited spin-wait before blocking in ForkJoinTask.join.
465      *
466      * Style notes
467      * ===========
468      *
469      * There is a lot of representation-level coupling among classes
470      * ForkJoinPool, ForkJoinWorkerThread, and ForkJoinTask.  The
471      * fields of WorkQueue maintain data structures managed by
472      * ForkJoinPool, so are directly accessed.  There is little point
473      * trying to reduce this, since any associated future changes in
474      * representations will need to be accompanied by algorithmic
475      * changes anyway. Several methods intrinsically sprawl because
476      * they must accumulate sets of consistent reads of volatiles held
477      * in local variables.  Methods signalWork() and scan() are the
478      * main bottlenecks, so are especially heavily
479      * micro-optimized/mangled.  There are lots of inline assignments
480      * (of form "while ((local = field) != 0)") which are usually the
481      * simplest way to ensure the required read orderings (which are
482      * sometimes critical). This leads to a "C"-like style of listing
483      * declarations of these locals at the heads of methods or blocks.
484      * There are several occurrences of the unusual "do {} while
485      * (!cas...)"  which is the simplest way to force an update of a
486      * CAS'ed variable. There are also other coding oddities (including
487      * several unnecessary-looking hoisted null checks) that help
488      * some methods perform reasonably even when interpreted (not
489      * compiled).
490      *
491      * The order of declarations in this file is:
492      * (1) Static utility functions
493      * (2) Nested (static) classes
494      * (3) Static fields
495      * (4) Fields, along with constants used when unpacking some of them
496      * (5) Internal control methods
497      * (6) Callbacks and other support for ForkJoinTask methods
498      * (7) Exported methods
499      * (8) Static block initializing statics in minimally dependent order
500      */
501 
502     // Static utilities
503 
504     /**
505      * If there is a security manager, makes sure caller has
506      * permission to modify threads.
507      */
checkPermission()508     private static void checkPermission() {
509         SecurityManager security = System.getSecurityManager();
510         if (security != null)
511             security.checkPermission(modifyThreadPermission);
512     }
513 
514     // Nested classes
515 
516     /**
517      * Factory for creating new {@link ForkJoinWorkerThread}s.
518      * A {@code ForkJoinWorkerThreadFactory} must be defined and used
519      * for {@code ForkJoinWorkerThread} subclasses that extend base
520      * functionality or initialize threads with different contexts.
521      */
522     public static interface ForkJoinWorkerThreadFactory {
523         /**
524          * Returns a new worker thread operating in the given pool.
525          *
526          * @param pool the pool this thread works in
527          * @throws NullPointerException if the pool is null
528          * @return the new worker thread
529          */
newThread(ForkJoinPool pool)530         public ForkJoinWorkerThread newThread(ForkJoinPool pool);
531     }
532 
533     /**
534      * Default ForkJoinWorkerThreadFactory implementation; creates a
535      * new ForkJoinWorkerThread.
536      */
537     static final class DefaultForkJoinWorkerThreadFactory
538         implements ForkJoinWorkerThreadFactory {
newThread(ForkJoinPool pool)539         public final ForkJoinWorkerThread newThread(ForkJoinPool pool) {
540             return new ForkJoinWorkerThread(pool);
541         }
542     }
543 
544     /**
545      * Class for artificial tasks that are used to replace the target
546      * of local joins if they are removed from an interior queue slot
547      * in WorkQueue.tryRemoveAndExec. We don't need the proxy to
548      * actually do anything beyond having a unique identity.
549      */
550     static final class EmptyTask extends ForkJoinTask<Void> {
551         private static final long serialVersionUID = -7721805057305804111L;
EmptyTask()552         EmptyTask() { status = ForkJoinTask.NORMAL; } // force done
getRawResult()553         public final Void getRawResult() { return null; }
setRawResult(Void x)554         public final void setRawResult(Void x) {}
exec()555         public final boolean exec() { return true; }
556     }
557 
558     /**
559      * Queues supporting work-stealing as well as external task
560      * submission. See above for main rationale and algorithms.
561      * Implementation relies heavily on "Unsafe" intrinsics
562      * and selective use of "volatile":
563      *
564      * Field "base" is the index (mod array.length) of the least valid
565      * queue slot, which is always the next position to steal (poll)
566      * from if nonempty. Reads and writes require volatile orderings
567      * but not CAS, because updates are only performed after slot
568      * CASes.
569      *
570      * Field "top" is the index (mod array.length) of the next queue
571      * slot to push to or pop from. It is written only by owner thread
572      * for push, or under lock for external/shared push, and accessed
573      * by other threads only after reading (volatile) base.  Both top
574      * and base are allowed to wrap around on overflow, but (top -
575      * base) (or more commonly -(base - top) to force volatile read of
576      * base before top) still estimates size. The lock ("qlock") is
577      * forced to -1 on termination, causing all further lock attempts
578      * to fail. (Note: we don't need CAS for termination state because
579      * upon pool shutdown, all shared-queues will stop being used
580      * anyway.)  Nearly all lock bodies are set up so that exceptions
581      * within lock bodies are "impossible" (modulo JVM errors that
582      * would cause failure anyway.)
583      *
584      * The array slots are read and written using the emulation of
585      * volatiles/atomics provided by Unsafe. Insertions must in
586      * general use putOrderedObject as a form of releasing store to
587      * ensure that all writes to the task object are ordered before
588      * its publication in the queue.  All removals entail a CAS to
589      * null.  The array is always a power of two. To ensure safety of
590      * Unsafe array operations, all accesses perform explicit null
591      * checks and implicit bounds checks via power-of-two masking.
592      *
593      * In addition to basic queuing support, this class contains
594      * fields described elsewhere to control execution. It turns out
595      * to work better memory-layout-wise to include them in this class
596      * rather than a separate class.
597      *
598      * Performance on most platforms is very sensitive to placement of
599      * instances of both WorkQueues and their arrays -- we absolutely
600      * do not want multiple WorkQueue instances or multiple queue
601      * arrays sharing cache lines. (It would be best for queue objects
602      * and their arrays to share, but there is nothing available to
603      * help arrange that). The @Contended annotation alerts JVMs to
604      * try to keep instances apart.
605      */
606     static final class WorkQueue {
607         /**
608          * Capacity of work-stealing queue array upon initialization.
609          * Must be a power of two; at least 4, but should be larger to
610          * reduce or eliminate cacheline sharing among queues.
611          * Currently, it is much larger, as a partial workaround for
612          * the fact that JVMs often place arrays in locations that
613          * share GC bookkeeping (especially cardmarks) such that
614          * per-write accesses encounter serious memory contention.
615          */
616         static final int INITIAL_QUEUE_CAPACITY = 1 << 13;
617 
618         /**
619          * Maximum size for queue arrays. Must be a power of two less
620          * than or equal to 1 << (31 - width of array entry) to ensure
621          * lack of wraparound of index calculations, but defined to a
622          * value a bit less than this to help users trap runaway
623          * programs before saturating systems.
624          */
625         static final int MAXIMUM_QUEUE_CAPACITY = 1 << 26; // 64M
626 
627         // Heuristic padding to ameliorate unfortunate memory placements
628         volatile long pad00, pad01, pad02, pad03, pad04, pad05, pad06;
629 
630         volatile int eventCount;   // encoded inactivation count; < 0 if inactive
631         int nextWait;              // encoded record of next event waiter
632         int nsteals;               // number of steals
633         int hint;                  // steal index hint
634         short poolIndex;           // index of this queue in pool
635         final short mode;          // 0: lifo, > 0: fifo, < 0: shared
636         volatile int qlock;        // 1: locked, -1: terminate; else 0
637         volatile int base;         // index of next slot for poll
638         int top;                   // index of next slot for push
639         ForkJoinTask<?>[] array;   // the elements (initially unallocated)
640         final ForkJoinPool pool;   // the containing pool (may be null)
641         final ForkJoinWorkerThread owner; // owning thread or null if shared
642         volatile Thread parker;    // == owner during call to park; else null
643         volatile ForkJoinTask<?> currentJoin;  // task being joined in awaitJoin
644         ForkJoinTask<?> currentSteal; // current non-local task being executed
645 
646         volatile Object pad10, pad11, pad12, pad13, pad14, pad15, pad16, pad17;
647         volatile Object pad18, pad19, pad1a, pad1b, pad1c, pad1d;
648 
WorkQueue(ForkJoinPool pool, ForkJoinWorkerThread owner, int mode, int seed)649         WorkQueue(ForkJoinPool pool, ForkJoinWorkerThread owner, int mode,
650                   int seed) {
651             this.pool = pool;
652             this.owner = owner;
653             this.mode = (short)mode;
654             this.hint = seed; // store initial seed for runWorker
655             // Place indices in the center of array (that is not yet allocated)
656             base = top = INITIAL_QUEUE_CAPACITY >>> 1;
657         }
658 
659         /**
660          * Returns the approximate number of tasks in the queue.
661          */
queueSize()662         final int queueSize() {
663             int n = base - top;       // non-owner callers must read base first
664             return (n >= 0) ? 0 : -n; // ignore transient negative
665         }
666 
667         /**
668          * Provides a more accurate estimate of whether this queue has
669          * any tasks than does queueSize, by checking whether a
670          * near-empty queue has at least one unclaimed task.
671          */
isEmpty()672         final boolean isEmpty() {
673             ForkJoinTask<?>[] a; int m, s;
674             int n = base - (s = top);
675             return (n >= 0 ||
676                     (n == -1 &&
677                      ((a = array) == null ||
678                       (m = a.length - 1) < 0 ||
679                       U.getObject
680                       (a, (long)((m & (s - 1)) << ASHIFT) + ABASE) == null)));
681         }
682 
683         /**
684          * Pushes a task. Call only by owner in unshared queues.  (The
685          * shared-queue version is embedded in method externalPush.)
686          *
687          * @param task the task. Caller must ensure non-null.
688          * @throws RejectedExecutionException if array cannot be resized
689          */
push(ForkJoinTask<?> task)690         final void push(ForkJoinTask<?> task) {
691             ForkJoinTask<?>[] a; ForkJoinPool p;
692             int s = top, n;
693             if ((a = array) != null) {    // ignore if queue removed
694                 int m = a.length - 1;
695                 U.putOrderedObject(a, ((m & s) << ASHIFT) + ABASE, task);
696                 if ((n = (top = s + 1) - base) <= 2)
697                     (p = pool).signalWork(p.workQueues, this);
698                 else if (n >= m)
699                     growArray();
700             }
701         }
702 
703         /**
704          * Initializes or doubles the capacity of array. Call either
705          * by owner or with lock held -- it is OK for base, but not
706          * top, to move while resizings are in progress.
707          */
growArray()708         final ForkJoinTask<?>[] growArray() {
709             ForkJoinTask<?>[] oldA = array;
710             int size = oldA != null ? oldA.length << 1 : INITIAL_QUEUE_CAPACITY;
711             if (size > MAXIMUM_QUEUE_CAPACITY)
712                 throw new RejectedExecutionException("Queue capacity exceeded");
713             int oldMask, t, b;
714             ForkJoinTask<?>[] a = array = new ForkJoinTask<?>[size];
715             if (oldA != null && (oldMask = oldA.length - 1) >= 0 &&
716                 (t = top) - (b = base) > 0) {
717                 int mask = size - 1;
718                 do {
719                     ForkJoinTask<?> x;
720                     int oldj = ((b & oldMask) << ASHIFT) + ABASE;
721                     int j    = ((b &    mask) << ASHIFT) + ABASE;
722                     x = (ForkJoinTask<?>)U.getObjectVolatile(oldA, oldj);
723                     if (x != null &&
724                         U.compareAndSwapObject(oldA, oldj, x, null))
725                         U.putObjectVolatile(a, j, x);
726                 } while (++b != t);
727             }
728             return a;
729         }
730 
731         /**
732          * Takes next task, if one exists, in LIFO order.  Call only
733          * by owner in unshared queues.
734          */
pop()735         final ForkJoinTask<?> pop() {
736             ForkJoinTask<?>[] a; ForkJoinTask<?> t; int m;
737             if ((a = array) != null && (m = a.length - 1) >= 0) {
738                 for (int s; (s = top - 1) - base >= 0;) {
739                     long j = ((m & s) << ASHIFT) + ABASE;
740                     if ((t = (ForkJoinTask<?>)U.getObject(a, j)) == null)
741                         break;
742                     if (U.compareAndSwapObject(a, j, t, null)) {
743                         top = s;
744                         return t;
745                     }
746                 }
747             }
748             return null;
749         }
750 
751         /**
752          * Takes a task in FIFO order if b is base of queue and a task
753          * can be claimed without contention. Specialized versions
754          * appear in ForkJoinPool methods scan and tryHelpStealer.
755          */
pollAt(int b)756         final ForkJoinTask<?> pollAt(int b) {
757             ForkJoinTask<?> t; ForkJoinTask<?>[] a;
758             if ((a = array) != null) {
759                 int j = (((a.length - 1) & b) << ASHIFT) + ABASE;
760                 if ((t = (ForkJoinTask<?>)U.getObjectVolatile(a, j)) != null &&
761                     base == b && U.compareAndSwapObject(a, j, t, null)) {
762                     U.putOrderedInt(this, QBASE, b + 1);
763                     return t;
764                 }
765             }
766             return null;
767         }
768 
769         /**
770          * Takes next task, if one exists, in FIFO order.
771          */
poll()772         final ForkJoinTask<?> poll() {
773             ForkJoinTask<?>[] a; int b; ForkJoinTask<?> t;
774             while ((b = base) - top < 0 && (a = array) != null) {
775                 int j = (((a.length - 1) & b) << ASHIFT) + ABASE;
776                 t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
777                 if (t != null) {
778                     if (U.compareAndSwapObject(a, j, t, null)) {
779                         U.putOrderedInt(this, QBASE, b + 1);
780                         return t;
781                     }
782                 }
783                 else if (base == b) {
784                     if (b + 1 == top)
785                         break;
786                     Thread.yield(); // wait for lagging update (very rare)
787                 }
788             }
789             return null;
790         }
791 
792         /**
793          * Takes next task, if one exists, in order specified by mode.
794          */
nextLocalTask()795         final ForkJoinTask<?> nextLocalTask() {
796             return mode == 0 ? pop() : poll();
797         }
798 
799         /**
800          * Returns next task, if one exists, in order specified by mode.
801          */
peek()802         final ForkJoinTask<?> peek() {
803             ForkJoinTask<?>[] a = array; int m;
804             if (a == null || (m = a.length - 1) < 0)
805                 return null;
806             int i = mode == 0 ? top - 1 : base;
807             int j = ((i & m) << ASHIFT) + ABASE;
808             return (ForkJoinTask<?>)U.getObjectVolatile(a, j);
809         }
810 
811         /**
812          * Pops the given task only if it is at the current top.
813          * (A shared version is available only via FJP.tryExternalUnpush)
814          */
tryUnpush(ForkJoinTask<?> t)815         final boolean tryUnpush(ForkJoinTask<?> t) {
816             ForkJoinTask<?>[] a; int s;
817             if ((a = array) != null && (s = top) != base &&
818                 U.compareAndSwapObject
819                 (a, (((a.length - 1) & --s) << ASHIFT) + ABASE, t, null)) {
820                 top = s;
821                 return true;
822             }
823             return false;
824         }
825 
826         /**
827          * Removes and cancels all known tasks, ignoring any exceptions.
828          */
cancelAll()829         final void cancelAll() {
830             ForkJoinTask.cancelIgnoringExceptions(currentJoin);
831             ForkJoinTask.cancelIgnoringExceptions(currentSteal);
832             for (ForkJoinTask<?> t; (t = poll()) != null; )
833                 ForkJoinTask.cancelIgnoringExceptions(t);
834         }
835 
836         // Specialized execution methods
837 
838         /**
839          * Polls and runs tasks until empty.
840          */
pollAndExecAll()841         final void pollAndExecAll() {
842             for (ForkJoinTask<?> t; (t = poll()) != null;)
843                 t.doExec();
844         }
845 
846         /**
847          * Executes a top-level task and any local tasks remaining
848          * after execution.
849          */
runTask(ForkJoinTask<?> task)850         final void runTask(ForkJoinTask<?> task) {
851             if ((currentSteal = task) != null) {
852                 task.doExec();
853                 ForkJoinTask<?>[] a = array;
854                 int md = mode;
855                 ++nsteals;
856                 currentSteal = null;
857                 if (md != 0)
858                     pollAndExecAll();
859                 else if (a != null) {
860                     int s, m = a.length - 1;
861                     while ((s = top - 1) - base >= 0) {
862                         long i = ((m & s) << ASHIFT) + ABASE;
863                         ForkJoinTask<?> t = (ForkJoinTask<?>)U.getObject(a, i);
864                         if (t == null)
865                             break;
866                         if (U.compareAndSwapObject(a, i, t, null)) {
867                             top = s;
868                             t.doExec();
869                         }
870                     }
871                 }
872             }
873         }
874 
875         /**
876          * If present, removes from queue and executes the given task,
877          * or any other cancelled task. Returns (true) on any CAS
878          * or consistency check failure so caller can retry.
879          *
880          * @return false if no progress can be made, else true
881          */
tryRemoveAndExec(ForkJoinTask<?> task)882         final boolean tryRemoveAndExec(ForkJoinTask<?> task) {
883             boolean stat;
884             ForkJoinTask<?>[] a; int m, s, b, n;
885             if (task != null && (a = array) != null && (m = a.length - 1) >= 0 &&
886                 (n = (s = top) - (b = base)) > 0) {
887                 boolean removed = false, empty = true;
888                 stat = true;
889                 for (ForkJoinTask<?> t;;) {           // traverse from s to b
890                     long j = ((--s & m) << ASHIFT) + ABASE;
891                     t = (ForkJoinTask<?>)U.getObject(a, j);
892                     if (t == null)                    // inconsistent length
893                         break;
894                     else if (t == task) {
895                         if (s + 1 == top) {           // pop
896                             if (!U.compareAndSwapObject(a, j, task, null))
897                                 break;
898                             top = s;
899                             removed = true;
900                         }
901                         else if (base == b)           // replace with proxy
902                             removed = U.compareAndSwapObject(a, j, task,
903                                                              new EmptyTask());
904                         break;
905                     }
906                     else if (t.status >= 0)
907                         empty = false;
908                     else if (s + 1 == top) {          // pop and throw away
909                         if (U.compareAndSwapObject(a, j, t, null))
910                             top = s;
911                         break;
912                     }
913                     if (--n == 0) {
914                         if (!empty && base == b)
915                             stat = false;
916                         break;
917                     }
918                 }
919                 if (removed)
920                     task.doExec();
921             }
922             else
923                 stat = false;
924             return stat;
925         }
926 
927         /**
928          * Tries to poll for and execute the given task or any other
929          * task in its CountedCompleter computation.
930          */
pollAndExecCC(CountedCompleter<?> root)931         final boolean pollAndExecCC(CountedCompleter<?> root) {
932             ForkJoinTask<?>[] a; int b; Object o; CountedCompleter<?> t, r;
933             if ((b = base) - top < 0 && (a = array) != null) {
934                 long j = (((a.length - 1) & b) << ASHIFT) + ABASE;
935                 if ((o = U.getObjectVolatile(a, j)) == null)
936                     return true; // retry
937                 if (o instanceof CountedCompleter) {
938                     for (t = (CountedCompleter<?>)o, r = t;;) {
939                         if (r == root) {
940                             if (base == b &&
941                                 U.compareAndSwapObject(a, j, t, null)) {
942                                 U.putOrderedInt(this, QBASE, b + 1);
943                                 t.doExec();
944                             }
945                             return true;
946                         }
947                         else if ((r = r.completer) == null)
948                             break; // not part of root computation
949                     }
950                 }
951             }
952             return false;
953         }
954 
955         /**
956          * Tries to pop and execute the given task or any other task
957          * in its CountedCompleter computation.
958          */
externalPopAndExecCC(CountedCompleter<?> root)959         final boolean externalPopAndExecCC(CountedCompleter<?> root) {
960             ForkJoinTask<?>[] a; int s; Object o; CountedCompleter<?> t, r;
961             if (base - (s = top) < 0 && (a = array) != null) {
962                 long j = (((a.length - 1) & (s - 1)) << ASHIFT) + ABASE;
963                 if ((o = U.getObject(a, j)) instanceof CountedCompleter) {
964                     for (t = (CountedCompleter<?>)o, r = t;;) {
965                         if (r == root) {
966                             if (U.compareAndSwapInt(this, QLOCK, 0, 1)) {
967                                 if (top == s && array == a &&
968                                     U.compareAndSwapObject(a, j, t, null)) {
969                                     top = s - 1;
970                                     qlock = 0;
971                                     t.doExec();
972                                 }
973                                 else
974                                     qlock = 0;
975                             }
976                             return true;
977                         }
978                         else if ((r = r.completer) == null)
979                             break;
980                     }
981                 }
982             }
983             return false;
984         }
985 
986         /**
987          * Internal version
988          */
internalPopAndExecCC(CountedCompleter<?> root)989         final boolean internalPopAndExecCC(CountedCompleter<?> root) {
990             ForkJoinTask<?>[] a; int s; Object o; CountedCompleter<?> t, r;
991             if (base - (s = top) < 0 && (a = array) != null) {
992                 long j = (((a.length - 1) & (s - 1)) << ASHIFT) + ABASE;
993                 if ((o = U.getObject(a, j)) instanceof CountedCompleter) {
994                     for (t = (CountedCompleter<?>)o, r = t;;) {
995                         if (r == root) {
996                             if (U.compareAndSwapObject(a, j, t, null)) {
997                                 top = s - 1;
998                                 t.doExec();
999                             }
1000                             return true;
1001                         }
1002                         else if ((r = r.completer) == null)
1003                             break;
1004                     }
1005                 }
1006             }
1007             return false;
1008         }
1009 
1010         /**
1011          * Returns true if owned and not known to be blocked.
1012          */
isApparentlyUnblocked()1013         final boolean isApparentlyUnblocked() {
1014             Thread wt; Thread.State s;
1015             return (eventCount >= 0 &&
1016                     (wt = owner) != null &&
1017                     (s = wt.getState()) != Thread.State.BLOCKED &&
1018                     s != Thread.State.WAITING &&
1019                     s != Thread.State.TIMED_WAITING);
1020         }
1021 
1022         // Unsafe mechanics
1023         private static final sun.misc.Unsafe U;
1024         private static final long QBASE;
1025         private static final long QLOCK;
1026         private static final int ABASE;
1027         private static final int ASHIFT;
1028         static {
1029             try {
1030                 U = sun.misc.Unsafe.getUnsafe();
1031                 Class<?> k = WorkQueue.class;
1032                 Class<?> ak = ForkJoinTask[].class;
1033                 QBASE = U.objectFieldOffset
1034                     (k.getDeclaredField("base"));
1035                 QLOCK = U.objectFieldOffset
1036                     (k.getDeclaredField("qlock"));
1037                 ABASE = U.arrayBaseOffset(ak);
1038                 int scale = U.arrayIndexScale(ak);
1039                 if ((scale & (scale - 1)) != 0)
1040                     throw new Error("data type scale not a power of two");
1041                 ASHIFT = 31 - Integer.numberOfLeadingZeros(scale);
1042             } catch (Exception e) {
1043                 throw new Error(e);
1044             }
1045         }
1046     }
1047 
1048     // static fields (initialized in static initializer below)
1049 
1050     /**
1051      * Per-thread submission bookkeeping. Shared across all pools
1052      * to reduce ThreadLocal pollution and because random motion
1053      * to avoid contention in one pool is likely to hold for others.
1054      * Lazily initialized on first submission (but null-checked
1055      * in other contexts to avoid unnecessary initialization).
1056      */
1057     static final ThreadLocal<Submitter> submitters;
1058 
1059     /**
1060      * Creates a new ForkJoinWorkerThread. This factory is used unless
1061      * overridden in ForkJoinPool constructors.
1062      */
1063     public static final ForkJoinWorkerThreadFactory
1064         defaultForkJoinWorkerThreadFactory;
1065 
1066     /**
1067      * Permission required for callers of methods that may start or
1068      * kill threads.
1069      */
1070     private static final RuntimePermission modifyThreadPermission;
1071 
1072     /**
1073      * Common (static) pool. Non-null for public use unless a static
1074      * construction exception, but internal usages null-check on use
1075      * to paranoically avoid potential initialization circularities
1076      * as well as to simplify generated code.
1077      */
1078     static final ForkJoinPool common;
1079 
1080     /**
1081      * Common pool parallelism. To allow simpler use and management
1082      * when common pool threads are disabled, we allow the underlying
1083      * common.parallelism field to be zero, but in that case still report
1084      * parallelism as 1 to reflect resulting caller-runs mechanics.
1085      */
1086     static final int commonParallelism;
1087 
1088     /**
1089      * Sequence number for creating workerNamePrefix.
1090      */
1091     private static int poolNumberSequence;
1092 
1093     /**
1094      * Returns the next sequence number. We don't expect this to
1095      * ever contend, so use simple builtin sync.
1096      */
nextPoolId()1097     private static final synchronized int nextPoolId() {
1098         return ++poolNumberSequence;
1099     }
1100 
1101     // static constants
1102 
1103     /**
1104      * Initial timeout value (in nanoseconds) for the thread
1105      * triggering quiescence to park waiting for new work. On timeout,
1106      * the thread will instead try to shrink the number of
1107      * workers. The value should be large enough to avoid overly
1108      * aggressive shrinkage during most transient stalls (long GCs
1109      * etc).
1110      */
1111     private static final long IDLE_TIMEOUT      = 2000L * 1000L * 1000L; // 2sec
1112 
1113     /**
1114      * Timeout value when there are more threads than parallelism level
1115      */
1116     private static final long FAST_IDLE_TIMEOUT =  200L * 1000L * 1000L;
1117 
1118     /**
1119      * Tolerance for idle timeouts, to cope with timer undershoots
1120      */
1121     private static final long TIMEOUT_SLOP = 2000000L;
1122 
1123     /**
1124      * The maximum stolen->joining link depth allowed in method
1125      * tryHelpStealer.  Must be a power of two.  Depths for legitimate
1126      * chains are unbounded, but we use a fixed constant to avoid
1127      * (otherwise unchecked) cycles and to bound staleness of
1128      * traversal parameters at the expense of sometimes blocking when
1129      * we could be helping.
1130      */
1131     private static final int MAX_HELP = 64;
1132 
1133     /**
1134      * Increment for seed generators. See class ThreadLocal for
1135      * explanation.
1136      */
1137     private static final int SEED_INCREMENT = 0x61c88647;
1138 
1139     /*
1140      * Bits and masks for control variables
1141      *
1142      * Field ctl is a long packed with:
1143      * AC: Number of active running workers minus target parallelism (16 bits)
1144      * TC: Number of total workers minus target parallelism (16 bits)
1145      * ST: true if pool is terminating (1 bit)
1146      * EC: the wait count of top waiting thread (15 bits)
1147      * ID: poolIndex of top of Treiber stack of waiters (16 bits)
1148      *
1149      * When convenient, we can extract the upper 32 bits of counts and
1150      * the lower 32 bits of queue state, u = (int)(ctl >>> 32) and e =
1151      * (int)ctl.  The ec field is never accessed alone, but always
1152      * together with id and st. The offsets of counts by the target
1153      * parallelism and the positionings of fields makes it possible to
1154      * perform the most common checks via sign tests of fields: When
1155      * ac is negative, there are not enough active workers, when tc is
1156      * negative, there are not enough total workers, and when e is
1157      * negative, the pool is terminating.  To deal with these possibly
1158      * negative fields, we use casts in and out of "short" and/or
1159      * signed shifts to maintain signedness.
1160      *
1161      * When a thread is queued (inactivated), its eventCount field is
1162      * set negative, which is the only way to tell if a worker is
1163      * prevented from executing tasks, even though it must continue to
1164      * scan for them to avoid queuing races. Note however that
1165      * eventCount updates lag releases so usage requires care.
1166      *
1167      * Field plock is an int packed with:
1168      * SHUTDOWN: true if shutdown is enabled (1 bit)
1169      * SEQ:  a sequence lock, with PL_LOCK bit set if locked (30 bits)
1170      * SIGNAL: set when threads may be waiting on the lock (1 bit)
1171      *
1172      * The sequence number enables simple consistency checks:
1173      * Staleness of read-only operations on the workQueues array can
1174      * be checked by comparing plock before vs after the reads.
1175      */
1176 
1177     // bit positions/shifts for fields
1178     private static final int  AC_SHIFT   = 48;
1179     private static final int  TC_SHIFT   = 32;
1180     private static final int  ST_SHIFT   = 31;
1181     private static final int  EC_SHIFT   = 16;
1182 
1183     // bounds
1184     private static final int  SMASK      = 0xffff;  // short bits
1185     private static final int  MAX_CAP    = 0x7fff;  // max #workers - 1
1186     private static final int  EVENMASK   = 0xfffe;  // even short bits
1187     private static final int  SQMASK     = 0x007e;  // max 64 (even) slots
1188     private static final int  SHORT_SIGN = 1 << 15;
1189     private static final int  INT_SIGN   = 1 << 31;
1190 
1191     // masks
1192     private static final long STOP_BIT   = 0x0001L << ST_SHIFT;
1193     private static final long AC_MASK    = ((long)SMASK) << AC_SHIFT;
1194     private static final long TC_MASK    = ((long)SMASK) << TC_SHIFT;
1195 
1196     // units for incrementing and decrementing
1197     private static final long TC_UNIT    = 1L << TC_SHIFT;
1198     private static final long AC_UNIT    = 1L << AC_SHIFT;
1199 
1200     // masks and units for dealing with u = (int)(ctl >>> 32)
1201     private static final int  UAC_SHIFT  = AC_SHIFT - 32;
1202     private static final int  UTC_SHIFT  = TC_SHIFT - 32;
1203     private static final int  UAC_MASK   = SMASK << UAC_SHIFT;
1204     private static final int  UTC_MASK   = SMASK << UTC_SHIFT;
1205     private static final int  UAC_UNIT   = 1 << UAC_SHIFT;
1206     private static final int  UTC_UNIT   = 1 << UTC_SHIFT;
1207 
1208     // masks and units for dealing with e = (int)ctl
1209     private static final int E_MASK      = 0x7fffffff; // no STOP_BIT
1210     private static final int E_SEQ       = 1 << EC_SHIFT;
1211 
1212     // plock bits
1213     private static final int SHUTDOWN    = 1 << 31;
1214     private static final int PL_LOCK     = 2;
1215     private static final int PL_SIGNAL   = 1;
1216     private static final int PL_SPINS    = 1 << 8;
1217 
1218     // access mode for WorkQueue
1219     static final int LIFO_QUEUE          =  0;
1220     static final int FIFO_QUEUE          =  1;
1221     static final int SHARED_QUEUE        = -1;
1222 
1223     // Heuristic padding to ameliorate unfortunate memory placements
1224     volatile long pad00, pad01, pad02, pad03, pad04, pad05, pad06;
1225 
1226     // Instance fields
1227     volatile long stealCount;                  // collects worker counts
1228     volatile long ctl;                         // main pool control
1229     volatile int plock;                        // shutdown status and seqLock
1230     volatile int indexSeed;                    // worker/submitter index seed
1231     final short parallelism;                   // parallelism level
1232     final short mode;                          // LIFO/FIFO
1233     WorkQueue[] workQueues;                    // main registry
1234     final ForkJoinWorkerThreadFactory factory;
1235     final UncaughtExceptionHandler ueh;        // per-worker UEH
1236     final String workerNamePrefix;             // to create worker name string
1237 
1238     volatile Object pad10, pad11, pad12, pad13, pad14, pad15, pad16, pad17;
1239     volatile Object pad18, pad19, pad1a, pad1b;
1240 
1241     /**
1242      * Acquires the plock lock to protect worker array and related
1243      * updates. This method is called only if an initial CAS on plock
1244      * fails. This acts as a spinlock for normal cases, but falls back
1245      * to builtin monitor to block when (rarely) needed. This would be
1246      * a terrible idea for a highly contended lock, but works fine as
1247      * a more conservative alternative to a pure spinlock.
1248      */
acquirePlock()1249     private int acquirePlock() {
1250         int spins = PL_SPINS, ps, nps;
1251         for (;;) {
1252             if (((ps = plock) & PL_LOCK) == 0 &&
1253                 U.compareAndSwapInt(this, PLOCK, ps, nps = ps + PL_LOCK))
1254                 return nps;
1255             else if (spins >= 0) {
1256                 if (ThreadLocalRandom.current().nextInt() >= 0)
1257                     --spins;
1258             }
1259             else if (U.compareAndSwapInt(this, PLOCK, ps, ps | PL_SIGNAL)) {
1260                 synchronized (this) {
1261                     if ((plock & PL_SIGNAL) != 0) {
1262                         try {
1263                             wait();
1264                         } catch (InterruptedException ie) {
1265                             try {
1266                                 Thread.currentThread().interrupt();
1267                             } catch (SecurityException ignore) {
1268                             }
1269                         }
1270                     }
1271                     else
1272                         notifyAll();
1273                 }
1274             }
1275         }
1276     }
1277 
1278     /**
1279      * Unlocks and signals any thread waiting for plock. Called only
1280      * when CAS of seq value for unlock fails.
1281      */
releasePlock(int ps)1282     private void releasePlock(int ps) {
1283         plock = ps;
1284         synchronized (this) { notifyAll(); }
1285     }
1286 
1287     /**
1288      * Tries to create and start one worker if fewer than target
1289      * parallelism level exist. Adjusts counts etc on failure.
1290      */
tryAddWorker()1291     private void tryAddWorker() {
1292         long c; int u, e;
1293         while ((u = (int)((c = ctl) >>> 32)) < 0 &&
1294                (u & SHORT_SIGN) != 0 && (e = (int)c) >= 0) {
1295             long nc = ((long)(((u + UTC_UNIT) & UTC_MASK) |
1296                               ((u + UAC_UNIT) & UAC_MASK)) << 32) | (long)e;
1297             if (U.compareAndSwapLong(this, CTL, c, nc)) {
1298                 ForkJoinWorkerThreadFactory fac;
1299                 Throwable ex = null;
1300                 ForkJoinWorkerThread wt = null;
1301                 try {
1302                     if ((fac = factory) != null &&
1303                         (wt = fac.newThread(this)) != null) {
1304                         wt.start();
1305                         break;
1306                     }
1307                 } catch (Throwable rex) {
1308                     ex = rex;
1309                 }
1310                 deregisterWorker(wt, ex);
1311                 break;
1312             }
1313         }
1314     }
1315 
1316     //  Registering and deregistering workers
1317 
1318     /**
1319      * Callback from ForkJoinWorkerThread to establish and record its
1320      * WorkQueue. To avoid scanning bias due to packing entries in
1321      * front of the workQueues array, we treat the array as a simple
1322      * power-of-two hash table using per-thread seed as hash,
1323      * expanding as needed.
1324      *
1325      * @param wt the worker thread
1326      * @return the worker's queue
1327      */
registerWorker(ForkJoinWorkerThread wt)1328     final WorkQueue registerWorker(ForkJoinWorkerThread wt) {
1329         UncaughtExceptionHandler handler; WorkQueue[] ws; int s, ps;
1330         wt.setDaemon(true);
1331         if ((handler = ueh) != null)
1332             wt.setUncaughtExceptionHandler(handler);
1333         do {} while (!U.compareAndSwapInt(this, INDEXSEED, s = indexSeed,
1334                                           s += SEED_INCREMENT) ||
1335                      s == 0); // skip 0
1336         WorkQueue w = new WorkQueue(this, wt, mode, s);
1337         if (((ps = plock) & PL_LOCK) != 0 ||
1338             !U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1339             ps = acquirePlock();
1340         int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1341         try {
1342             if ((ws = workQueues) != null) {    // skip if shutting down
1343                 int n = ws.length, m = n - 1;
1344                 int r = (s << 1) | 1;           // use odd-numbered indices
1345                 if (ws[r &= m] != null) {       // collision
1346                     int probes = 0;             // step by approx half size
1347                     int step = (n <= 4) ? 2 : ((n >>> 1) & EVENMASK) + 2;
1348                     while (ws[r = (r + step) & m] != null) {
1349                         if (++probes >= n) {
1350                             workQueues = ws = Arrays.copyOf(ws, n <<= 1);
1351                             m = n - 1;
1352                             probes = 0;
1353                         }
1354                     }
1355                 }
1356                 w.poolIndex = (short)r;
1357                 w.eventCount = r; // volatile write orders
1358                 ws[r] = w;
1359             }
1360         } finally {
1361             if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1362                 releasePlock(nps);
1363         }
1364         wt.setName(workerNamePrefix.concat(Integer.toString(w.poolIndex >>> 1)));
1365         return w;
1366     }
1367 
1368     /**
1369      * Final callback from terminating worker, as well as upon failure
1370      * to construct or start a worker.  Removes record of worker from
1371      * array, and adjusts counts. If pool is shutting down, tries to
1372      * complete termination.
1373      *
1374      * @param wt the worker thread, or null if construction failed
1375      * @param ex the exception causing failure, or null if none
1376      */
deregisterWorker(ForkJoinWorkerThread wt, Throwable ex)1377     final void deregisterWorker(ForkJoinWorkerThread wt, Throwable ex) {
1378         WorkQueue w = null;
1379         if (wt != null && (w = wt.workQueue) != null) {
1380             int ps; long sc;
1381             w.qlock = -1;                // ensure set
1382             do {} while (!U.compareAndSwapLong(this, STEALCOUNT,
1383                                                sc = stealCount,
1384                                                sc + w.nsteals));
1385             if (((ps = plock) & PL_LOCK) != 0 ||
1386                 !U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1387                 ps = acquirePlock();
1388             int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1389             try {
1390                 int idx = w.poolIndex;
1391                 WorkQueue[] ws = workQueues;
1392                 if (ws != null && idx >= 0 && idx < ws.length && ws[idx] == w)
1393                     ws[idx] = null;
1394             } finally {
1395                 if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1396                     releasePlock(nps);
1397             }
1398         }
1399 
1400         long c;                          // adjust ctl counts
1401         do {} while (!U.compareAndSwapLong
1402                      (this, CTL, c = ctl, (((c - AC_UNIT) & AC_MASK) |
1403                                            ((c - TC_UNIT) & TC_MASK) |
1404                                            (c & ~(AC_MASK|TC_MASK)))));
1405 
1406         if (!tryTerminate(false, false) && w != null && w.array != null) {
1407             w.cancelAll();               // cancel remaining tasks
1408             WorkQueue[] ws; WorkQueue v; Thread p; int u, i, e;
1409             while ((u = (int)((c = ctl) >>> 32)) < 0 && (e = (int)c) >= 0) {
1410                 if (e > 0) {             // activate or create replacement
1411                     if ((ws = workQueues) == null ||
1412                         (i = e & SMASK) >= ws.length ||
1413                         (v = ws[i]) == null)
1414                         break;
1415                     long nc = (((long)(v.nextWait & E_MASK)) |
1416                                ((long)(u + UAC_UNIT) << 32));
1417                     if (v.eventCount != (e | INT_SIGN))
1418                         break;
1419                     if (U.compareAndSwapLong(this, CTL, c, nc)) {
1420                         v.eventCount = (e + E_SEQ) & E_MASK;
1421                         if ((p = v.parker) != null)
1422                             U.unpark(p);
1423                         break;
1424                     }
1425                 }
1426                 else {
1427                     if ((short)u < 0)
1428                         tryAddWorker();
1429                     break;
1430                 }
1431             }
1432         }
1433         if (ex == null)                     // help clean refs on way out
1434             ForkJoinTask.helpExpungeStaleExceptions();
1435         else                                // rethrow
1436             ForkJoinTask.rethrow(ex);
1437     }
1438 
1439     // Submissions
1440 
1441     /**
1442      * Per-thread records for threads that submit to pools. Currently
1443      * holds only pseudo-random seed / index that is used to choose
1444      * submission queues in method externalPush. In the future, this may
1445      * also incorporate a means to implement different task rejection
1446      * and resubmission policies.
1447      *
1448      * Seeds for submitters and workers/workQueues work in basically
1449      * the same way but are initialized and updated using slightly
1450      * different mechanics. Both are initialized using the same
1451      * approach as in class ThreadLocal, where successive values are
1452      * unlikely to collide with previous values. Seeds are then
1453      * randomly modified upon collisions using xorshifts, which
1454      * requires a non-zero seed.
1455      */
1456     static final class Submitter {
1457         int seed;
Submitter(int s)1458         Submitter(int s) { seed = s; }
1459     }
1460 
1461     /**
1462      * Unless shutting down, adds the given task to a submission queue
1463      * at submitter's current queue index (modulo submission
1464      * range). Only the most common path is directly handled in this
1465      * method. All others are relayed to fullExternalPush.
1466      *
1467      * @param task the task. Caller must ensure non-null.
1468      */
externalPush(ForkJoinTask<?> task)1469     final void externalPush(ForkJoinTask<?> task) {
1470         Submitter z = submitters.get();
1471         WorkQueue q; int r, m, s, n, am; ForkJoinTask<?>[] a;
1472         int ps = plock;
1473         WorkQueue[] ws = workQueues;
1474         if (z != null && ps > 0 && ws != null && (m = (ws.length - 1)) >= 0 &&
1475             (q = ws[m & (r = z.seed) & SQMASK]) != null && r != 0 &&
1476             U.compareAndSwapInt(q, QLOCK, 0, 1)) { // lock
1477             if ((a = q.array) != null &&
1478                 (am = a.length - 1) > (n = (s = q.top) - q.base)) {
1479                 int j = ((am & s) << ASHIFT) + ABASE;
1480                 U.putOrderedObject(a, j, task);
1481                 q.top = s + 1;                     // push on to deque
1482                 q.qlock = 0;
1483                 if (n <= 1)
1484                     signalWork(ws, q);
1485                 return;
1486             }
1487             q.qlock = 0;
1488         }
1489         fullExternalPush(task);
1490     }
1491 
1492     /**
1493      * Full version of externalPush. This method is called, among
1494      * other times, upon the first submission of the first task to the
1495      * pool, so must perform secondary initialization.  It also
1496      * detects first submission by an external thread by looking up
1497      * its ThreadLocal, and creates a new shared queue if the one at
1498      * index if empty or contended. The plock lock body must be
1499      * exception-free (so no try/finally) so we optimistically
1500      * allocate new queues outside the lock and throw them away if
1501      * (very rarely) not needed.
1502      *
1503      * Secondary initialization occurs when plock is zero, to create
1504      * workQueue array and set plock to a valid value.  This lock body
1505      * must also be exception-free. Because the plock seq value can
1506      * eventually wrap around zero, this method harmlessly fails to
1507      * reinitialize if workQueues exists, while still advancing plock.
1508      */
fullExternalPush(ForkJoinTask<?> task)1509     private void fullExternalPush(ForkJoinTask<?> task) {
1510         int r = 0; // random index seed
1511         for (Submitter z = submitters.get();;) {
1512             WorkQueue[] ws; WorkQueue q; int ps, m, k;
1513             if (z == null) {
1514                 if (U.compareAndSwapInt(this, INDEXSEED, r = indexSeed,
1515                                         r += SEED_INCREMENT) && r != 0)
1516                     submitters.set(z = new Submitter(r));
1517             }
1518             else if (r == 0) {                  // move to a different index
1519                 r = z.seed;
1520                 r ^= r << 13;                   // same xorshift as WorkQueues
1521                 r ^= r >>> 17;
1522                 z.seed = r ^= (r << 5);
1523             }
1524             if ((ps = plock) < 0)
1525                 throw new RejectedExecutionException();
1526             else if (ps == 0 || (ws = workQueues) == null ||
1527                      (m = ws.length - 1) < 0) { // initialize workQueues
1528                 int p = parallelism;            // find power of two table size
1529                 int n = (p > 1) ? p - 1 : 1;    // ensure at least 2 slots
1530                 n |= n >>> 1; n |= n >>> 2;  n |= n >>> 4;
1531                 n |= n >>> 8; n |= n >>> 16; n = (n + 1) << 1;
1532                 WorkQueue[] nws = ((ws = workQueues) == null || ws.length == 0 ?
1533                                    new WorkQueue[n] : null);
1534                 if (((ps = plock) & PL_LOCK) != 0 ||
1535                     !U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1536                     ps = acquirePlock();
1537                 if (((ws = workQueues) == null || ws.length == 0) && nws != null)
1538                     workQueues = nws;
1539                 int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1540                 if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1541                     releasePlock(nps);
1542             }
1543             else if ((q = ws[k = r & m & SQMASK]) != null) {
1544                 if (q.qlock == 0 && U.compareAndSwapInt(q, QLOCK, 0, 1)) {
1545                     ForkJoinTask<?>[] a = q.array;
1546                     int s = q.top;
1547                     boolean submitted = false;
1548                     try {                      // locked version of push
1549                         if ((a != null && a.length > s + 1 - q.base) ||
1550                             (a = q.growArray()) != null) {   // must presize
1551                             int j = (((a.length - 1) & s) << ASHIFT) + ABASE;
1552                             U.putOrderedObject(a, j, task);
1553                             q.top = s + 1;
1554                             submitted = true;
1555                         }
1556                     } finally {
1557                         q.qlock = 0;  // unlock
1558                     }
1559                     if (submitted) {
1560                         signalWork(ws, q);
1561                         return;
1562                     }
1563                 }
1564                 r = 0; // move on failure
1565             }
1566             else if (((ps = plock) & PL_LOCK) == 0) { // create new queue
1567                 q = new WorkQueue(this, null, SHARED_QUEUE, r);
1568                 q.poolIndex = (short)k;
1569                 if (((ps = plock) & PL_LOCK) != 0 ||
1570                     !U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1571                     ps = acquirePlock();
1572                 if ((ws = workQueues) != null && k < ws.length && ws[k] == null)
1573                     ws[k] = q;
1574                 int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1575                 if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1576                     releasePlock(nps);
1577             }
1578             else
1579                 r = 0;
1580         }
1581     }
1582 
1583     // Maintaining ctl counts
1584 
1585     /**
1586      * Increments active count; mainly called upon return from blocking.
1587      */
incrementActiveCount()1588     final void incrementActiveCount() {
1589         long c;
1590         do {} while (!U.compareAndSwapLong
1591                      (this, CTL, c = ctl, ((c & ~AC_MASK) |
1592                                            ((c & AC_MASK) + AC_UNIT))));
1593     }
1594 
1595     /**
1596      * Tries to create or activate a worker if too few are active.
1597      *
1598      * @param ws the worker array to use to find signallees
1599      * @param q if non-null, the queue holding tasks to be processed
1600      */
signalWork(WorkQueue[] ws, WorkQueue q)1601     final void signalWork(WorkQueue[] ws, WorkQueue q) {
1602         for (;;) {
1603             long c; int e, u, i; WorkQueue w; Thread p;
1604             if ((u = (int)((c = ctl) >>> 32)) >= 0)
1605                 break;
1606             if ((e = (int)c) <= 0) {
1607                 if ((short)u < 0)
1608                     tryAddWorker();
1609                 break;
1610             }
1611             if (ws == null || ws.length <= (i = e & SMASK) ||
1612                 (w = ws[i]) == null)
1613                 break;
1614             long nc = (((long)(w.nextWait & E_MASK)) |
1615                        ((long)(u + UAC_UNIT)) << 32);
1616             int ne = (e + E_SEQ) & E_MASK;
1617             if (w.eventCount == (e | INT_SIGN) &&
1618                 U.compareAndSwapLong(this, CTL, c, nc)) {
1619                 w.eventCount = ne;
1620                 if ((p = w.parker) != null)
1621                     U.unpark(p);
1622                 break;
1623             }
1624             if (q != null && q.base >= q.top)
1625                 break;
1626         }
1627     }
1628 
1629     // Scanning for tasks
1630 
1631     /**
1632      * Top-level runloop for workers, called by ForkJoinWorkerThread.run.
1633      */
runWorker(WorkQueue w)1634     final void runWorker(WorkQueue w) {
1635         w.growArray(); // allocate queue
1636         for (int r = w.hint; scan(w, r) == 0; ) {
1637             r ^= r << 13; r ^= r >>> 17; r ^= r << 5; // xorshift
1638         }
1639     }
1640 
1641     /**
1642      * Scans for and, if found, runs one task, else possibly
1643      * inactivates the worker. This method operates on single reads of
1644      * volatile state and is designed to be re-invoked continuously,
1645      * in part because it returns upon detecting inconsistencies,
1646      * contention, or state changes that indicate possible success on
1647      * re-invocation.
1648      *
1649      * The scan searches for tasks across queues starting at a random
1650      * index, checking each at least twice.  The scan terminates upon
1651      * either finding a non-empty queue, or completing the sweep. If
1652      * the worker is not inactivated, it takes and runs a task from
1653      * this queue. Otherwise, if not activated, it tries to activate
1654      * itself or some other worker by signalling. On failure to find a
1655      * task, returns (for retry) if pool state may have changed during
1656      * an empty scan, or tries to inactivate if active, else possibly
1657      * blocks or terminates via method awaitWork.
1658      *
1659      * @param w the worker (via its WorkQueue)
1660      * @param r a random seed
1661      * @return worker qlock status if would have waited, else 0
1662      */
scan(WorkQueue w, int r)1663     private final int scan(WorkQueue w, int r) {
1664         WorkQueue[] ws; int m;
1665         long c = ctl;                            // for consistency check
1666         if ((ws = workQueues) != null && (m = ws.length - 1) >= 0 && w != null) {
1667             for (int j = m + m + 1, ec = w.eventCount;;) {
1668                 WorkQueue q; int b, e; ForkJoinTask<?>[] a; ForkJoinTask<?> t;
1669                 if ((q = ws[(r - j) & m]) != null &&
1670                     (b = q.base) - q.top < 0 && (a = q.array) != null) {
1671                     long i = (((a.length - 1) & b) << ASHIFT) + ABASE;
1672                     if ((t = ((ForkJoinTask<?>)
1673                               U.getObjectVolatile(a, i))) != null) {
1674                         if (ec < 0)
1675                             helpRelease(c, ws, w, q, b);
1676                         else if (q.base == b &&
1677                                  U.compareAndSwapObject(a, i, t, null)) {
1678                             U.putOrderedInt(q, QBASE, b + 1);
1679                             if ((b + 1) - q.top < 0)
1680                                 signalWork(ws, q);
1681                             w.runTask(t);
1682                         }
1683                     }
1684                     break;
1685                 }
1686                 else if (--j < 0) {
1687                     if ((ec | (e = (int)c)) < 0) // inactive or terminating
1688                         return awaitWork(w, c, ec);
1689                     else if (ctl == c) {         // try to inactivate and enqueue
1690                         long nc = (long)ec | ((c - AC_UNIT) & (AC_MASK|TC_MASK));
1691                         w.nextWait = e;
1692                         w.eventCount = ec | INT_SIGN;
1693                         if (!U.compareAndSwapLong(this, CTL, c, nc))
1694                             w.eventCount = ec;   // back out
1695                     }
1696                     break;
1697                 }
1698             }
1699         }
1700         return 0;
1701     }
1702 
1703     /**
1704      * A continuation of scan(), possibly blocking or terminating
1705      * worker w. Returns without blocking if pool state has apparently
1706      * changed since last invocation.  Also, if inactivating w has
1707      * caused the pool to become quiescent, checks for pool
1708      * termination, and, so long as this is not the only worker, waits
1709      * for event for up to a given duration.  On timeout, if ctl has
1710      * not changed, terminates the worker, which will in turn wake up
1711      * another worker to possibly repeat this process.
1712      *
1713      * @param w the calling worker
1714      * @param c the ctl value on entry to scan
1715      * @param ec the worker's eventCount on entry to scan
1716      */
awaitWork(WorkQueue w, long c, int ec)1717     private final int awaitWork(WorkQueue w, long c, int ec) {
1718         int stat, ns; long parkTime, deadline;
1719         if ((stat = w.qlock) >= 0 && w.eventCount == ec && ctl == c &&
1720             !Thread.interrupted()) {
1721             int e = (int)c;
1722             int u = (int)(c >>> 32);
1723             int d = (u >> UAC_SHIFT) + parallelism; // active count
1724 
1725             if (e < 0 || (d <= 0 && tryTerminate(false, false)))
1726                 stat = w.qlock = -1;          // pool is terminating
1727             else if ((ns = w.nsteals) != 0) { // collect steals and retry
1728                 long sc;
1729                 w.nsteals = 0;
1730                 do {} while (!U.compareAndSwapLong(this, STEALCOUNT,
1731                                                    sc = stealCount, sc + ns));
1732             }
1733             else {
1734                 long pc = ((d > 0 || ec != (e | INT_SIGN)) ? 0L :
1735                            ((long)(w.nextWait & E_MASK)) | // ctl to restore
1736                            ((long)(u + UAC_UNIT)) << 32);
1737                 if (pc != 0L) {               // timed wait if last waiter
1738                     int dc = -(short)(c >>> TC_SHIFT);
1739                     parkTime = (dc < 0 ? FAST_IDLE_TIMEOUT:
1740                                 (dc + 1) * IDLE_TIMEOUT);
1741                     deadline = System.nanoTime() + parkTime - TIMEOUT_SLOP;
1742                 }
1743                 else
1744                     parkTime = deadline = 0L;
1745                 if (w.eventCount == ec && ctl == c) {
1746                     Thread wt = Thread.currentThread();
1747                     U.putObject(wt, PARKBLOCKER, this);
1748                     w.parker = wt;            // emulate LockSupport.park
1749                     if (w.eventCount == ec && ctl == c)
1750                         U.park(false, parkTime);  // must recheck before park
1751                     w.parker = null;
1752                     U.putObject(wt, PARKBLOCKER, null);
1753                     if (parkTime != 0L && ctl == c &&
1754                         deadline - System.nanoTime() <= 0L &&
1755                         U.compareAndSwapLong(this, CTL, c, pc))
1756                         stat = w.qlock = -1;  // shrink pool
1757                 }
1758             }
1759         }
1760         return stat;
1761     }
1762 
1763     /**
1764      * Possibly releases (signals) a worker. Called only from scan()
1765      * when a worker with apparently inactive status finds a non-empty
1766      * queue. This requires revalidating all of the associated state
1767      * from caller.
1768      */
helpRelease(long c, WorkQueue[] ws, WorkQueue w, WorkQueue q, int b)1769     private final void helpRelease(long c, WorkQueue[] ws, WorkQueue w,
1770                                    WorkQueue q, int b) {
1771         WorkQueue v; int e, i; Thread p;
1772         if (w != null && w.eventCount < 0 && (e = (int)c) > 0 &&
1773             ws != null && ws.length > (i = e & SMASK) &&
1774             (v = ws[i]) != null && ctl == c) {
1775             long nc = (((long)(v.nextWait & E_MASK)) |
1776                        ((long)((int)(c >>> 32) + UAC_UNIT)) << 32);
1777             int ne = (e + E_SEQ) & E_MASK;
1778             if (q != null && q.base == b && w.eventCount < 0 &&
1779                 v.eventCount == (e | INT_SIGN) &&
1780                 U.compareAndSwapLong(this, CTL, c, nc)) {
1781                 v.eventCount = ne;
1782                 if ((p = v.parker) != null)
1783                     U.unpark(p);
1784             }
1785         }
1786     }
1787 
1788     /**
1789      * Tries to locate and execute tasks for a stealer of the given
1790      * task, or in turn one of its stealers, Traces currentSteal ->
1791      * currentJoin links looking for a thread working on a descendant
1792      * of the given task and with a non-empty queue to steal back and
1793      * execute tasks from. The first call to this method upon a
1794      * waiting join will often entail scanning/search, (which is OK
1795      * because the joiner has nothing better to do), but this method
1796      * leaves hints in workers to speed up subsequent calls. The
1797      * implementation is very branchy to cope with potential
1798      * inconsistencies or loops encountering chains that are stale,
1799      * unknown, or so long that they are likely cyclic.
1800      *
1801      * @param joiner the joining worker
1802      * @param task the task to join
1803      * @return 0 if no progress can be made, negative if task
1804      * known complete, else positive
1805      */
tryHelpStealer(WorkQueue joiner, ForkJoinTask<?> task)1806     private int tryHelpStealer(WorkQueue joiner, ForkJoinTask<?> task) {
1807         int stat = 0, steps = 0;                    // bound to avoid cycles
1808         if (task != null && joiner != null &&
1809             joiner.base - joiner.top >= 0) {        // hoist checks
1810             restart: for (;;) {
1811                 ForkJoinTask<?> subtask = task;     // current target
1812                 for (WorkQueue j = joiner, v;;) {   // v is stealer of subtask
1813                     WorkQueue[] ws; int m, s, h;
1814                     if ((s = task.status) < 0) {
1815                         stat = s;
1816                         break restart;
1817                     }
1818                     if ((ws = workQueues) == null || (m = ws.length - 1) <= 0)
1819                         break restart;              // shutting down
1820                     if ((v = ws[h = (j.hint | 1) & m]) == null ||
1821                         v.currentSteal != subtask) {
1822                         for (int origin = h;;) {    // find stealer
1823                             if (((h = (h + 2) & m) & 15) == 1 &&
1824                                 (subtask.status < 0 || j.currentJoin != subtask))
1825                                 continue restart;   // occasional staleness check
1826                             if ((v = ws[h]) != null &&
1827                                 v.currentSteal == subtask) {
1828                                 j.hint = h;        // save hint
1829                                 break;
1830                             }
1831                             if (h == origin)
1832                                 break restart;      // cannot find stealer
1833                         }
1834                     }
1835                     for (;;) { // help stealer or descend to its stealer
1836                         ForkJoinTask[] a; int b;
1837                         if (subtask.status < 0)     // surround probes with
1838                             continue restart;       //   consistency checks
1839                         if ((b = v.base) - v.top < 0 && (a = v.array) != null) {
1840                             int i = (((a.length - 1) & b) << ASHIFT) + ABASE;
1841                             ForkJoinTask<?> t =
1842                                 (ForkJoinTask<?>)U.getObjectVolatile(a, i);
1843                             if (subtask.status < 0 || j.currentJoin != subtask ||
1844                                 v.currentSteal != subtask)
1845                                 continue restart;   // stale
1846                             stat = 1;               // apparent progress
1847                             if (v.base == b) {
1848                                 if (t == null)
1849                                     break restart;
1850                                 if (U.compareAndSwapObject(a, i, t, null)) {
1851                                     U.putOrderedInt(v, QBASE, b + 1);
1852                                     ForkJoinTask<?> ps = joiner.currentSteal;
1853                                     int jt = joiner.top;
1854                                     do {
1855                                         joiner.currentSteal = t;
1856                                         t.doExec(); // clear local tasks too
1857                                     } while (task.status >= 0 &&
1858                                              joiner.top != jt &&
1859                                              (t = joiner.pop()) != null);
1860                                     joiner.currentSteal = ps;
1861                                     break restart;
1862                                 }
1863                             }
1864                         }
1865                         else {                      // empty -- try to descend
1866                             ForkJoinTask<?> next = v.currentJoin;
1867                             if (subtask.status < 0 || j.currentJoin != subtask ||
1868                                 v.currentSteal != subtask)
1869                                 continue restart;   // stale
1870                             else if (next == null || ++steps == MAX_HELP)
1871                                 break restart;      // dead-end or maybe cyclic
1872                             else {
1873                                 subtask = next;
1874                                 j = v;
1875                                 break;
1876                             }
1877                         }
1878                     }
1879                 }
1880             }
1881         }
1882         return stat;
1883     }
1884 
1885     /**
1886      * Analog of tryHelpStealer for CountedCompleters. Tries to steal
1887      * and run tasks within the target's computation.
1888      *
1889      * @param task the task to join
1890      */
helpComplete(WorkQueue joiner, CountedCompleter<?> task)1891     private int helpComplete(WorkQueue joiner, CountedCompleter<?> task) {
1892         WorkQueue[] ws; int m;
1893         int s = 0;
1894         if ((ws = workQueues) != null && (m = ws.length - 1) >= 0 &&
1895             joiner != null && task != null) {
1896             int j = joiner.poolIndex;
1897             int scans = m + m + 1;
1898             long c = 0L;              // for stability check
1899             for (int k = scans; ; j += 2) {
1900                 WorkQueue q;
1901                 if ((s = task.status) < 0)
1902                     break;
1903                 else if (joiner.internalPopAndExecCC(task))
1904                     k = scans;
1905                 else if ((s = task.status) < 0)
1906                     break;
1907                 else if ((q = ws[j & m]) != null && q.pollAndExecCC(task))
1908                     k = scans;
1909                 else if (--k < 0) {
1910                     if (c == (c = ctl))
1911                         break;
1912                     k = scans;
1913                 }
1914             }
1915         }
1916         return s;
1917     }
1918 
1919     /**
1920      * Tries to decrement active count (sometimes implicitly) and
1921      * possibly release or create a compensating worker in preparation
1922      * for blocking. Fails on contention or termination. Otherwise,
1923      * adds a new thread if no idle workers are available and pool
1924      * may become starved.
1925      *
1926      * @param c the assumed ctl value
1927      */
tryCompensate(long c)1928     final boolean tryCompensate(long c) {
1929         WorkQueue[] ws = workQueues;
1930         int pc = parallelism, e = (int)c, m, tc;
1931         if (ws != null && (m = ws.length - 1) >= 0 && e >= 0 && ctl == c) {
1932             WorkQueue w = ws[e & m];
1933             if (e != 0 && w != null) {
1934                 Thread p;
1935                 long nc = ((long)(w.nextWait & E_MASK) |
1936                            (c & (AC_MASK|TC_MASK)));
1937                 int ne = (e + E_SEQ) & E_MASK;
1938                 if (w.eventCount == (e | INT_SIGN) &&
1939                     U.compareAndSwapLong(this, CTL, c, nc)) {
1940                     w.eventCount = ne;
1941                     if ((p = w.parker) != null)
1942                         U.unpark(p);
1943                     return true;   // replace with idle worker
1944                 }
1945             }
1946             else if ((tc = (short)(c >>> TC_SHIFT)) >= 0 &&
1947                      (int)(c >> AC_SHIFT) + pc > 1) {
1948                 long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK);
1949                 if (U.compareAndSwapLong(this, CTL, c, nc))
1950                     return true;   // no compensation
1951             }
1952             else if (tc + pc < MAX_CAP) {
1953                 long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK);
1954                 if (U.compareAndSwapLong(this, CTL, c, nc)) {
1955                     ForkJoinWorkerThreadFactory fac;
1956                     Throwable ex = null;
1957                     ForkJoinWorkerThread wt = null;
1958                     try {
1959                         if ((fac = factory) != null &&
1960                             (wt = fac.newThread(this)) != null) {
1961                             wt.start();
1962                             return true;
1963                         }
1964                     } catch (Throwable rex) {
1965                         ex = rex;
1966                     }
1967                     deregisterWorker(wt, ex); // clean up and return false
1968                 }
1969             }
1970         }
1971         return false;
1972     }
1973 
1974     /**
1975      * Helps and/or blocks until the given task is done.
1976      *
1977      * @param joiner the joining worker
1978      * @param task the task
1979      * @return task status on exit
1980      */
awaitJoin(WorkQueue joiner, ForkJoinTask<?> task)1981     final int awaitJoin(WorkQueue joiner, ForkJoinTask<?> task) {
1982         int s = 0;
1983         if (task != null && (s = task.status) >= 0 && joiner != null) {
1984             ForkJoinTask<?> prevJoin = joiner.currentJoin;
1985             joiner.currentJoin = task;
1986             do {} while (joiner.tryRemoveAndExec(task) && // process local tasks
1987                          (s = task.status) >= 0);
1988             if (s >= 0 && (task instanceof CountedCompleter))
1989                 s = helpComplete(joiner, (CountedCompleter<?>)task);
1990             long cc = 0;        // for stability checks
1991             while (s >= 0 && (s = task.status) >= 0) {
1992                 if ((s = tryHelpStealer(joiner, task)) == 0 &&
1993                     (s = task.status) >= 0) {
1994                     if (!tryCompensate(cc))
1995                         cc = ctl;
1996                     else {
1997                         if (task.trySetSignal() && (s = task.status) >= 0) {
1998                             synchronized (task) {
1999                                 if (task.status >= 0) {
2000                                     try {                // see ForkJoinTask
2001                                         task.wait();     //  for explanation
2002                                     } catch (InterruptedException ie) {
2003                                     }
2004                                 }
2005                                 else
2006                                     task.notifyAll();
2007                             }
2008                         }
2009                         long c; // reactivate
2010                         do {} while (!U.compareAndSwapLong
2011                                      (this, CTL, c = ctl,
2012                                       ((c & ~AC_MASK) |
2013                                        ((c & AC_MASK) + AC_UNIT))));
2014                     }
2015                 }
2016             }
2017             joiner.currentJoin = prevJoin;
2018         }
2019         return s;
2020     }
2021 
2022     /**
2023      * Stripped-down variant of awaitJoin used by timed joins. Tries
2024      * to help join only while there is continuous progress. (Caller
2025      * will then enter a timed wait.)
2026      *
2027      * @param joiner the joining worker
2028      * @param task the task
2029      */
helpJoinOnce(WorkQueue joiner, ForkJoinTask<?> task)2030     final void helpJoinOnce(WorkQueue joiner, ForkJoinTask<?> task) {
2031         int s;
2032         if (joiner != null && task != null && (s = task.status) >= 0) {
2033             ForkJoinTask<?> prevJoin = joiner.currentJoin;
2034             joiner.currentJoin = task;
2035             do {} while (joiner.tryRemoveAndExec(task) && // process local tasks
2036                          (s = task.status) >= 0);
2037             if (s >= 0) {
2038                 if (task instanceof CountedCompleter)
2039                     helpComplete(joiner, (CountedCompleter<?>)task);
2040                 do {} while (task.status >= 0 &&
2041                              tryHelpStealer(joiner, task) > 0);
2042             }
2043             joiner.currentJoin = prevJoin;
2044         }
2045     }
2046 
2047     /**
2048      * Returns a (probably) non-empty steal queue, if one is found
2049      * during a scan, else null.  This method must be retried by
2050      * caller if, by the time it tries to use the queue, it is empty.
2051      */
findNonEmptyStealQueue()2052     private WorkQueue findNonEmptyStealQueue() {
2053         int r = ThreadLocalRandom.current().nextInt();
2054         for (;;) {
2055             int ps = plock, m; WorkQueue[] ws; WorkQueue q;
2056             if ((ws = workQueues) != null && (m = ws.length - 1) >= 0) {
2057                 for (int j = (m + 1) << 2; j >= 0; --j) {
2058                     if ((q = ws[(((r - j) << 1) | 1) & m]) != null &&
2059                         q.base - q.top < 0)
2060                         return q;
2061                 }
2062             }
2063             if (plock == ps)
2064                 return null;
2065         }
2066     }
2067 
2068     /**
2069      * Runs tasks until {@code isQuiescent()}. We piggyback on
2070      * active count ctl maintenance, but rather than blocking
2071      * when tasks cannot be found, we rescan until all others cannot
2072      * find tasks either.
2073      */
helpQuiescePool(WorkQueue w)2074     final void helpQuiescePool(WorkQueue w) {
2075         ForkJoinTask<?> ps = w.currentSteal;
2076         for (boolean active = true;;) {
2077             long c; WorkQueue q; ForkJoinTask<?> t; int b;
2078             while ((t = w.nextLocalTask()) != null)
2079                 t.doExec();
2080             if ((q = findNonEmptyStealQueue()) != null) {
2081                 if (!active) {      // re-establish active count
2082                     active = true;
2083                     do {} while (!U.compareAndSwapLong
2084                                  (this, CTL, c = ctl,
2085                                   ((c & ~AC_MASK) |
2086                                    ((c & AC_MASK) + AC_UNIT))));
2087                 }
2088                 if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null) {
2089                     (w.currentSteal = t).doExec();
2090                     w.currentSteal = ps;
2091                 }
2092             }
2093             else if (active) {       // decrement active count without queuing
2094                 long nc = ((c = ctl) & ~AC_MASK) | ((c & AC_MASK) - AC_UNIT);
2095                 if ((int)(nc >> AC_SHIFT) + parallelism == 0)
2096                     break;          // bypass decrement-then-increment
2097                 if (U.compareAndSwapLong(this, CTL, c, nc))
2098                     active = false;
2099             }
2100             else if ((int)((c = ctl) >> AC_SHIFT) + parallelism <= 0 &&
2101                      U.compareAndSwapLong
2102                      (this, CTL, c, ((c & ~AC_MASK) |
2103                                      ((c & AC_MASK) + AC_UNIT))))
2104                 break;
2105         }
2106     }
2107 
2108     /**
2109      * Gets and removes a local or stolen task for the given worker.
2110      *
2111      * @return a task, if available
2112      */
nextTaskFor(WorkQueue w)2113     final ForkJoinTask<?> nextTaskFor(WorkQueue w) {
2114         for (ForkJoinTask<?> t;;) {
2115             WorkQueue q; int b;
2116             if ((t = w.nextLocalTask()) != null)
2117                 return t;
2118             if ((q = findNonEmptyStealQueue()) == null)
2119                 return null;
2120             if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null)
2121                 return t;
2122         }
2123     }
2124 
2125     /**
2126      * Returns a cheap heuristic guide for task partitioning when
2127      * programmers, frameworks, tools, or languages have little or no
2128      * idea about task granularity.  In essence by offering this
2129      * method, we ask users only about tradeoffs in overhead vs
2130      * expected throughput and its variance, rather than how finely to
2131      * partition tasks.
2132      *
2133      * In a steady state strict (tree-structured) computation, each
2134      * thread makes available for stealing enough tasks for other
2135      * threads to remain active. Inductively, if all threads play by
2136      * the same rules, each thread should make available only a
2137      * constant number of tasks.
2138      *
2139      * The minimum useful constant is just 1. But using a value of 1
2140      * would require immediate replenishment upon each steal to
2141      * maintain enough tasks, which is infeasible.  Further,
2142      * partitionings/granularities of offered tasks should minimize
2143      * steal rates, which in general means that threads nearer the top
2144      * of computation tree should generate more than those nearer the
2145      * bottom. In perfect steady state, each thread is at
2146      * approximately the same level of computation tree. However,
2147      * producing extra tasks amortizes the uncertainty of progress and
2148      * diffusion assumptions.
2149      *
2150      * So, users will want to use values larger (but not much larger)
2151      * than 1 to both smooth over transient shortages and hedge
2152      * against uneven progress; as traded off against the cost of
2153      * extra task overhead. We leave the user to pick a threshold
2154      * value to compare with the results of this call to guide
2155      * decisions, but recommend values such as 3.
2156      *
2157      * When all threads are active, it is on average OK to estimate
2158      * surplus strictly locally. In steady-state, if one thread is
2159      * maintaining say 2 surplus tasks, then so are others. So we can
2160      * just use estimated queue length.  However, this strategy alone
2161      * leads to serious mis-estimates in some non-steady-state
2162      * conditions (ramp-up, ramp-down, other stalls). We can detect
2163      * many of these by further considering the number of "idle"
2164      * threads, that are known to have zero queued tasks, so
2165      * compensate by a factor of (#idle/#active) threads.
2166      *
2167      * Note: The approximation of #busy workers as #active workers is
2168      * not very good under current signalling scheme, and should be
2169      * improved.
2170      */
getSurplusQueuedTaskCount()2171     static int getSurplusQueuedTaskCount() {
2172         Thread t; ForkJoinWorkerThread wt; ForkJoinPool pool; WorkQueue q;
2173         if (((t = Thread.currentThread()) instanceof ForkJoinWorkerThread)) {
2174             int p = (pool = (wt = (ForkJoinWorkerThread)t).pool).parallelism;
2175             int n = (q = wt.workQueue).top - q.base;
2176             int a = (int)(pool.ctl >> AC_SHIFT) + p;
2177             return n - (a > (p >>>= 1) ? 0 :
2178                         a > (p >>>= 1) ? 1 :
2179                         a > (p >>>= 1) ? 2 :
2180                         a > (p >>>= 1) ? 4 :
2181                         8);
2182         }
2183         return 0;
2184     }
2185 
2186     //  Termination
2187 
2188     /**
2189      * Possibly initiates and/or completes termination.  The caller
2190      * triggering termination runs three passes through workQueues:
2191      * (0) Setting termination status, followed by wakeups of queued
2192      * workers; (1) cancelling all tasks; (2) interrupting lagging
2193      * threads (likely in external tasks, but possibly also blocked in
2194      * joins).  Each pass repeats previous steps because of potential
2195      * lagging thread creation.
2196      *
2197      * @param now if true, unconditionally terminate, else only
2198      * if no work and no active workers
2199      * @param enable if true, enable shutdown when next possible
2200      * @return true if now terminating or terminated
2201      */
tryTerminate(boolean now, boolean enable)2202     private boolean tryTerminate(boolean now, boolean enable) {
2203         int ps;
2204         if (this == common)                        // cannot shut down
2205             return false;
2206         if ((ps = plock) >= 0) {                   // enable by setting plock
2207             if (!enable)
2208                 return false;
2209             if ((ps & PL_LOCK) != 0 ||
2210                 !U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
2211                 ps = acquirePlock();
2212             int nps = ((ps + PL_LOCK) & ~SHUTDOWN) | SHUTDOWN;
2213             if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
2214                 releasePlock(nps);
2215         }
2216         for (long c;;) {
2217             if (((c = ctl) & STOP_BIT) != 0) {     // already terminating
2218                 if ((short)(c >>> TC_SHIFT) + parallelism <= 0) {
2219                     synchronized (this) {
2220                         notifyAll();               // signal when 0 workers
2221                     }
2222                 }
2223                 return true;
2224             }
2225             if (!now) {                            // check if idle & no tasks
2226                 WorkQueue[] ws; WorkQueue w;
2227                 if ((int)(c >> AC_SHIFT) + parallelism > 0)
2228                     return false;
2229                 if ((ws = workQueues) != null) {
2230                     for (int i = 0; i < ws.length; ++i) {
2231                         if ((w = ws[i]) != null &&
2232                             (!w.isEmpty() ||
2233                              ((i & 1) != 0 && w.eventCount >= 0))) {
2234                             signalWork(ws, w);
2235                             return false;
2236                         }
2237                     }
2238                 }
2239             }
2240             if (U.compareAndSwapLong(this, CTL, c, c | STOP_BIT)) {
2241                 for (int pass = 0; pass < 3; ++pass) {
2242                     WorkQueue[] ws; WorkQueue w; Thread wt;
2243                     if ((ws = workQueues) != null) {
2244                         int n = ws.length;
2245                         for (int i = 0; i < n; ++i) {
2246                             if ((w = ws[i]) != null) {
2247                                 w.qlock = -1;
2248                                 if (pass > 0) {
2249                                     w.cancelAll();
2250                                     if (pass > 1 && (wt = w.owner) != null) {
2251                                         if (!wt.isInterrupted()) {
2252                                             try {
2253                                                 wt.interrupt();
2254                                             } catch (Throwable ignore) {
2255                                             }
2256                                         }
2257                                         U.unpark(wt);
2258                                     }
2259                                 }
2260                             }
2261                         }
2262                         // Wake up workers parked on event queue
2263                         int i, e; long cc; Thread p;
2264                         while ((e = (int)(cc = ctl) & E_MASK) != 0 &&
2265                                (i = e & SMASK) < n && i >= 0 &&
2266                                (w = ws[i]) != null) {
2267                             long nc = ((long)(w.nextWait & E_MASK) |
2268                                        ((cc + AC_UNIT) & AC_MASK) |
2269                                        (cc & (TC_MASK|STOP_BIT)));
2270                             if (w.eventCount == (e | INT_SIGN) &&
2271                                 U.compareAndSwapLong(this, CTL, cc, nc)) {
2272                                 w.eventCount = (e + E_SEQ) & E_MASK;
2273                                 w.qlock = -1;
2274                                 if ((p = w.parker) != null)
2275                                     U.unpark(p);
2276                             }
2277                         }
2278                     }
2279                 }
2280             }
2281         }
2282     }
2283 
2284     // external operations on common pool
2285 
2286     /**
2287      * Returns common pool queue for a thread that has submitted at
2288      * least one task.
2289      */
commonSubmitterQueue()2290     static WorkQueue commonSubmitterQueue() {
2291         Submitter z; ForkJoinPool p; WorkQueue[] ws; int m, r;
2292         return ((z = submitters.get()) != null &&
2293                 (p = common) != null &&
2294                 (ws = p.workQueues) != null &&
2295                 (m = ws.length - 1) >= 0) ?
2296             ws[m & z.seed & SQMASK] : null;
2297     }
2298 
2299     /**
2300      * Tries to pop the given task from submitter's queue in common pool.
2301      */
tryExternalUnpush(ForkJoinTask<?> task)2302     final boolean tryExternalUnpush(ForkJoinTask<?> task) {
2303         WorkQueue joiner; ForkJoinTask<?>[] a; int m, s;
2304         Submitter z = submitters.get();
2305         WorkQueue[] ws = workQueues;
2306         boolean popped = false;
2307         if (z != null && ws != null && (m = ws.length - 1) >= 0 &&
2308             (joiner = ws[z.seed & m & SQMASK]) != null &&
2309             joiner.base != (s = joiner.top) &&
2310             (a = joiner.array) != null) {
2311             long j = (((a.length - 1) & (s - 1)) << ASHIFT) + ABASE;
2312             if (U.getObject(a, j) == task &&
2313                 U.compareAndSwapInt(joiner, QLOCK, 0, 1)) {
2314                 if (joiner.top == s && joiner.array == a &&
2315                     U.compareAndSwapObject(a, j, task, null)) {
2316                     joiner.top = s - 1;
2317                     popped = true;
2318                 }
2319                 joiner.qlock = 0;
2320             }
2321         }
2322         return popped;
2323     }
2324 
externalHelpComplete(CountedCompleter<?> task)2325     final int externalHelpComplete(CountedCompleter<?> task) {
2326         WorkQueue joiner; int m, j;
2327         Submitter z = submitters.get();
2328         WorkQueue[] ws = workQueues;
2329         int s = 0;
2330         if (z != null && ws != null && (m = ws.length - 1) >= 0 &&
2331             (joiner = ws[(j = z.seed) & m & SQMASK]) != null && task != null) {
2332             int scans = m + m + 1;
2333             long c = 0L;             // for stability check
2334             j |= 1;                  // poll odd queues
2335             for (int k = scans; ; j += 2) {
2336                 WorkQueue q;
2337                 if ((s = task.status) < 0)
2338                     break;
2339                 else if (joiner.externalPopAndExecCC(task))
2340                     k = scans;
2341                 else if ((s = task.status) < 0)
2342                     break;
2343                 else if ((q = ws[j & m]) != null && q.pollAndExecCC(task))
2344                     k = scans;
2345                 else if (--k < 0) {
2346                     if (c == (c = ctl))
2347                         break;
2348                     k = scans;
2349                 }
2350             }
2351         }
2352         return s;
2353     }
2354 
2355     // Exported methods
2356 
2357     // Constructors
2358 
2359     /**
2360      * Creates a {@code ForkJoinPool} with parallelism equal to {@link
2361      * java.lang.Runtime#availableProcessors}, using the {@linkplain
2362      * #defaultForkJoinWorkerThreadFactory default thread factory},
2363      * no UncaughtExceptionHandler, and non-async LIFO processing mode.
2364      */
ForkJoinPool()2365     public ForkJoinPool() {
2366         this(Math.min(MAX_CAP, Runtime.getRuntime().availableProcessors()),
2367              defaultForkJoinWorkerThreadFactory, null, false);
2368     }
2369 
2370     /**
2371      * Creates a {@code ForkJoinPool} with the indicated parallelism
2372      * level, the {@linkplain
2373      * #defaultForkJoinWorkerThreadFactory default thread factory},
2374      * no UncaughtExceptionHandler, and non-async LIFO processing mode.
2375      *
2376      * @param parallelism the parallelism level
2377      * @throws IllegalArgumentException if parallelism less than or
2378      *         equal to zero, or greater than implementation limit
2379      */
ForkJoinPool(int parallelism)2380     public ForkJoinPool(int parallelism) {
2381         this(parallelism, defaultForkJoinWorkerThreadFactory, null, false);
2382     }
2383 
2384     /**
2385      * Creates a {@code ForkJoinPool} with the given parameters.
2386      *
2387      * @param parallelism the parallelism level. For default value,
2388      * use {@link java.lang.Runtime#availableProcessors}.
2389      * @param factory the factory for creating new threads. For default value,
2390      * use {@link #defaultForkJoinWorkerThreadFactory}.
2391      * @param handler the handler for internal worker threads that
2392      * terminate due to unrecoverable errors encountered while executing
2393      * tasks. For default value, use {@code null}.
2394      * @param asyncMode if true,
2395      * establishes local first-in-first-out scheduling mode for forked
2396      * tasks that are never joined. This mode may be more appropriate
2397      * than default locally stack-based mode in applications in which
2398      * worker threads only process event-style asynchronous tasks.
2399      * For default value, use {@code false}.
2400      * @throws IllegalArgumentException if parallelism less than or
2401      *         equal to zero, or greater than implementation limit
2402      * @throws NullPointerException if the factory is null
2403      */
ForkJoinPool(int parallelism, ForkJoinWorkerThreadFactory factory, UncaughtExceptionHandler handler, boolean asyncMode)2404     public ForkJoinPool(int parallelism,
2405                         ForkJoinWorkerThreadFactory factory,
2406                         UncaughtExceptionHandler handler,
2407                         boolean asyncMode) {
2408         this(checkParallelism(parallelism),
2409              checkFactory(factory),
2410              handler,
2411              (asyncMode ? FIFO_QUEUE : LIFO_QUEUE),
2412              "ForkJoinPool-" + nextPoolId() + "-worker-");
2413         checkPermission();
2414     }
2415 
checkParallelism(int parallelism)2416     private static int checkParallelism(int parallelism) {
2417         if (parallelism <= 0 || parallelism > MAX_CAP)
2418             throw new IllegalArgumentException();
2419         return parallelism;
2420     }
2421 
checkFactory(ForkJoinWorkerThreadFactory factory)2422     private static ForkJoinWorkerThreadFactory checkFactory
2423         (ForkJoinWorkerThreadFactory factory) {
2424         if (factory == null)
2425             throw new NullPointerException();
2426         return factory;
2427     }
2428 
2429     /**
2430      * Creates a {@code ForkJoinPool} with the given parameters, without
2431      * any security checks or parameter validation.  Invoked directly by
2432      * makeCommonPool.
2433      */
ForkJoinPool(int parallelism, ForkJoinWorkerThreadFactory factory, UncaughtExceptionHandler handler, int mode, String workerNamePrefix)2434     private ForkJoinPool(int parallelism,
2435                          ForkJoinWorkerThreadFactory factory,
2436                          UncaughtExceptionHandler handler,
2437                          int mode,
2438                          String workerNamePrefix) {
2439         this.workerNamePrefix = workerNamePrefix;
2440         this.factory = factory;
2441         this.ueh = handler;
2442         this.mode = (short)mode;
2443         this.parallelism = (short)parallelism;
2444         long np = (long)(-parallelism); // offset ctl counts
2445         this.ctl = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK);
2446     }
2447 
2448     /**
2449      * Returns the common pool instance. This pool is statically
2450      * constructed; its run state is unaffected by attempts to {@link
2451      * #shutdown} or {@link #shutdownNow}. However this pool and any
2452      * ongoing processing are automatically terminated upon program
2453      * {@link System#exit}.  Any program that relies on asynchronous
2454      * task processing to complete before program termination should
2455      * invoke {@code commonPool().}{@link #awaitQuiescence awaitQuiescence},
2456      * before exit.
2457      *
2458      * @return the common pool instance
2459      * @since 1.8
2460      * @hide
2461      */
commonPool()2462     public static ForkJoinPool commonPool() {
2463         // assert common != null : "static init error";
2464         return common;
2465     }
2466 
2467     // Execution methods
2468 
2469     /**
2470      * Performs the given task, returning its result upon completion.
2471      * If the computation encounters an unchecked Exception or Error,
2472      * it is rethrown as the outcome of this invocation.  Rethrown
2473      * exceptions behave in the same way as regular exceptions, but,
2474      * when possible, contain stack traces (as displayed for example
2475      * using {@code ex.printStackTrace()}) of both the current thread
2476      * as well as the thread actually encountering the exception;
2477      * minimally only the latter.
2478      *
2479      * @param task the task
2480      * @return the task's result
2481      * @throws NullPointerException if the task is null
2482      * @throws RejectedExecutionException if the task cannot be
2483      *         scheduled for execution
2484      */
invoke(ForkJoinTask<T> task)2485     public <T> T invoke(ForkJoinTask<T> task) {
2486         if (task == null)
2487             throw new NullPointerException();
2488         externalPush(task);
2489         return task.join();
2490     }
2491 
2492     /**
2493      * Arranges for (asynchronous) execution of the given task.
2494      *
2495      * @param task the task
2496      * @throws NullPointerException if the task is null
2497      * @throws RejectedExecutionException if the task cannot be
2498      *         scheduled for execution
2499      */
execute(ForkJoinTask<?> task)2500     public void execute(ForkJoinTask<?> task) {
2501         if (task == null)
2502             throw new NullPointerException();
2503         externalPush(task);
2504     }
2505 
2506     // AbstractExecutorService methods
2507 
2508     /**
2509      * @throws NullPointerException if the task is null
2510      * @throws RejectedExecutionException if the task cannot be
2511      *         scheduled for execution
2512      */
execute(Runnable task)2513     public void execute(Runnable task) {
2514         if (task == null)
2515             throw new NullPointerException();
2516         ForkJoinTask<?> job;
2517         if (task instanceof ForkJoinTask<?>) // avoid re-wrap
2518             job = (ForkJoinTask<?>) task;
2519         else
2520             job = new ForkJoinTask.RunnableExecuteAction(task);
2521         externalPush(job);
2522     }
2523 
2524     /**
2525      * Submits a ForkJoinTask for execution.
2526      *
2527      * @param task the task to submit
2528      * @return the task
2529      * @throws NullPointerException if the task is null
2530      * @throws RejectedExecutionException if the task cannot be
2531      *         scheduled for execution
2532      */
submit(ForkJoinTask<T> task)2533     public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) {
2534         if (task == null)
2535             throw new NullPointerException();
2536         externalPush(task);
2537         return task;
2538     }
2539 
2540     /**
2541      * @throws NullPointerException if the task is null
2542      * @throws RejectedExecutionException if the task cannot be
2543      *         scheduled for execution
2544      */
submit(Callable<T> task)2545     public <T> ForkJoinTask<T> submit(Callable<T> task) {
2546         ForkJoinTask<T> job = new ForkJoinTask.AdaptedCallable<T>(task);
2547         externalPush(job);
2548         return job;
2549     }
2550 
2551     /**
2552      * @throws NullPointerException if the task is null
2553      * @throws RejectedExecutionException if the task cannot be
2554      *         scheduled for execution
2555      */
submit(Runnable task, T result)2556     public <T> ForkJoinTask<T> submit(Runnable task, T result) {
2557         ForkJoinTask<T> job = new ForkJoinTask.AdaptedRunnable<T>(task, result);
2558         externalPush(job);
2559         return job;
2560     }
2561 
2562     /**
2563      * @throws NullPointerException if the task is null
2564      * @throws RejectedExecutionException if the task cannot be
2565      *         scheduled for execution
2566      */
submit(Runnable task)2567     public ForkJoinTask<?> submit(Runnable task) {
2568         if (task == null)
2569             throw new NullPointerException();
2570         ForkJoinTask<?> job;
2571         if (task instanceof ForkJoinTask<?>) // avoid re-wrap
2572             job = (ForkJoinTask<?>) task;
2573         else
2574             job = new ForkJoinTask.AdaptedRunnableAction(task);
2575         externalPush(job);
2576         return job;
2577     }
2578 
2579     /**
2580      * @throws NullPointerException       {@inheritDoc}
2581      * @throws RejectedExecutionException {@inheritDoc}
2582      */
invokeAll(Collection<? extends Callable<T>> tasks)2583     public <T> List<Future<T>> invokeAll(Collection<? extends Callable<T>> tasks) {
2584         // In previous versions of this class, this method constructed
2585         // a task to run ForkJoinTask.invokeAll, but now external
2586         // invocation of multiple tasks is at least as efficient.
2587         ArrayList<Future<T>> futures = new ArrayList<Future<T>>(tasks.size());
2588 
2589         boolean done = false;
2590         try {
2591             for (Callable<T> t : tasks) {
2592                 ForkJoinTask<T> f = new ForkJoinTask.AdaptedCallable<T>(t);
2593                 futures.add(f);
2594                 externalPush(f);
2595             }
2596             for (int i = 0, size = futures.size(); i < size; i++)
2597                 ((ForkJoinTask<?>)futures.get(i)).quietlyJoin();
2598             done = true;
2599             return futures;
2600         } finally {
2601             if (!done)
2602                 for (int i = 0, size = futures.size(); i < size; i++)
2603                     futures.get(i).cancel(false);
2604         }
2605     }
2606 
2607     /**
2608      * Returns the factory used for constructing new workers.
2609      *
2610      * @return the factory used for constructing new workers
2611      */
getFactory()2612     public ForkJoinWorkerThreadFactory getFactory() {
2613         return factory;
2614     }
2615 
2616     /**
2617      * Returns the handler for internal worker threads that terminate
2618      * due to unrecoverable errors encountered while executing tasks.
2619      *
2620      * @return the handler, or {@code null} if none
2621      */
getUncaughtExceptionHandler()2622     public UncaughtExceptionHandler getUncaughtExceptionHandler() {
2623         return ueh;
2624     }
2625 
2626     /**
2627      * Returns the targeted parallelism level of this pool.
2628      *
2629      * @return the targeted parallelism level of this pool
2630      */
getParallelism()2631     public int getParallelism() {
2632         int par;
2633         return ((par = parallelism) > 0) ? par : 1;
2634     }
2635 
2636     /**
2637      * Returns the targeted parallelism level of the common pool.
2638      *
2639      * @return the targeted parallelism level of the common pool
2640      * @since 1.8
2641      * @hide
2642      */
getCommonPoolParallelism()2643     public static int getCommonPoolParallelism() {
2644         return commonParallelism;
2645     }
2646 
2647     /**
2648      * Returns the number of worker threads that have started but not
2649      * yet terminated.  The result returned by this method may differ
2650      * from {@link #getParallelism} when threads are created to
2651      * maintain parallelism when others are cooperatively blocked.
2652      *
2653      * @return the number of worker threads
2654      */
getPoolSize()2655     public int getPoolSize() {
2656         return parallelism + (short)(ctl >>> TC_SHIFT);
2657     }
2658 
2659     /**
2660      * Returns {@code true} if this pool uses local first-in-first-out
2661      * scheduling mode for forked tasks that are never joined.
2662      *
2663      * @return {@code true} if this pool uses async mode
2664      */
getAsyncMode()2665     public boolean getAsyncMode() {
2666         return mode == FIFO_QUEUE;
2667     }
2668 
2669     /**
2670      * Returns an estimate of the number of worker threads that are
2671      * not blocked waiting to join tasks or for other managed
2672      * synchronization. This method may overestimate the
2673      * number of running threads.
2674      *
2675      * @return the number of worker threads
2676      */
getRunningThreadCount()2677     public int getRunningThreadCount() {
2678         int rc = 0;
2679         WorkQueue[] ws; WorkQueue w;
2680         if ((ws = workQueues) != null) {
2681             for (int i = 1; i < ws.length; i += 2) {
2682                 if ((w = ws[i]) != null && w.isApparentlyUnblocked())
2683                     ++rc;
2684             }
2685         }
2686         return rc;
2687     }
2688 
2689     /**
2690      * Returns an estimate of the number of threads that are currently
2691      * stealing or executing tasks. This method may overestimate the
2692      * number of active threads.
2693      *
2694      * @return the number of active threads
2695      */
getActiveThreadCount()2696     public int getActiveThreadCount() {
2697         int r = parallelism + (int)(ctl >> AC_SHIFT);
2698         return (r <= 0) ? 0 : r; // suppress momentarily negative values
2699     }
2700 
2701     /**
2702      * Returns {@code true} if all worker threads are currently idle.
2703      * An idle worker is one that cannot obtain a task to execute
2704      * because none are available to steal from other threads, and
2705      * there are no pending submissions to the pool. This method is
2706      * conservative; it might not return {@code true} immediately upon
2707      * idleness of all threads, but will eventually become true if
2708      * threads remain inactive.
2709      *
2710      * @return {@code true} if all threads are currently idle
2711      */
isQuiescent()2712     public boolean isQuiescent() {
2713         return parallelism + (int)(ctl >> AC_SHIFT) <= 0;
2714     }
2715 
2716     /**
2717      * Returns an estimate of the total number of tasks stolen from
2718      * one thread's work queue by another. The reported value
2719      * underestimates the actual total number of steals when the pool
2720      * is not quiescent. This value may be useful for monitoring and
2721      * tuning fork/join programs: in general, steal counts should be
2722      * high enough to keep threads busy, but low enough to avoid
2723      * overhead and contention across threads.
2724      *
2725      * @return the number of steals
2726      */
getStealCount()2727     public long getStealCount() {
2728         long count = stealCount;
2729         WorkQueue[] ws; WorkQueue w;
2730         if ((ws = workQueues) != null) {
2731             for (int i = 1; i < ws.length; i += 2) {
2732                 if ((w = ws[i]) != null)
2733                     count += w.nsteals;
2734             }
2735         }
2736         return count;
2737     }
2738 
2739     /**
2740      * Returns an estimate of the total number of tasks currently held
2741      * in queues by worker threads (but not including tasks submitted
2742      * to the pool that have not begun executing). This value is only
2743      * an approximation, obtained by iterating across all threads in
2744      * the pool. This method may be useful for tuning task
2745      * granularities.
2746      *
2747      * @return the number of queued tasks
2748      */
getQueuedTaskCount()2749     public long getQueuedTaskCount() {
2750         long count = 0;
2751         WorkQueue[] ws; WorkQueue w;
2752         if ((ws = workQueues) != null) {
2753             for (int i = 1; i < ws.length; i += 2) {
2754                 if ((w = ws[i]) != null)
2755                     count += w.queueSize();
2756             }
2757         }
2758         return count;
2759     }
2760 
2761     /**
2762      * Returns an estimate of the number of tasks submitted to this
2763      * pool that have not yet begun executing.  This method may take
2764      * time proportional to the number of submissions.
2765      *
2766      * @return the number of queued submissions
2767      */
getQueuedSubmissionCount()2768     public int getQueuedSubmissionCount() {
2769         int count = 0;
2770         WorkQueue[] ws; WorkQueue w;
2771         if ((ws = workQueues) != null) {
2772             for (int i = 0; i < ws.length; i += 2) {
2773                 if ((w = ws[i]) != null)
2774                     count += w.queueSize();
2775             }
2776         }
2777         return count;
2778     }
2779 
2780     /**
2781      * Returns {@code true} if there are any tasks submitted to this
2782      * pool that have not yet begun executing.
2783      *
2784      * @return {@code true} if there are any queued submissions
2785      */
hasQueuedSubmissions()2786     public boolean hasQueuedSubmissions() {
2787         WorkQueue[] ws; WorkQueue w;
2788         if ((ws = workQueues) != null) {
2789             for (int i = 0; i < ws.length; i += 2) {
2790                 if ((w = ws[i]) != null && !w.isEmpty())
2791                     return true;
2792             }
2793         }
2794         return false;
2795     }
2796 
2797     /**
2798      * Removes and returns the next unexecuted submission if one is
2799      * available.  This method may be useful in extensions to this
2800      * class that re-assign work in systems with multiple pools.
2801      *
2802      * @return the next submission, or {@code null} if none
2803      */
pollSubmission()2804     protected ForkJoinTask<?> pollSubmission() {
2805         WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2806         if ((ws = workQueues) != null) {
2807             for (int i = 0; i < ws.length; i += 2) {
2808                 if ((w = ws[i]) != null && (t = w.poll()) != null)
2809                     return t;
2810             }
2811         }
2812         return null;
2813     }
2814 
2815     /**
2816      * Removes all available unexecuted submitted and forked tasks
2817      * from scheduling queues and adds them to the given collection,
2818      * without altering their execution status. These may include
2819      * artificially generated or wrapped tasks. This method is
2820      * designed to be invoked only when the pool is known to be
2821      * quiescent. Invocations at other times may not remove all
2822      * tasks. A failure encountered while attempting to add elements
2823      * to collection {@code c} may result in elements being in
2824      * neither, either or both collections when the associated
2825      * exception is thrown.  The behavior of this operation is
2826      * undefined if the specified collection is modified while the
2827      * operation is in progress.
2828      *
2829      * @param c the collection to transfer elements into
2830      * @return the number of elements transferred
2831      */
drainTasksTo(Collection<? super ForkJoinTask<?>> c)2832     protected int drainTasksTo(Collection<? super ForkJoinTask<?>> c) {
2833         int count = 0;
2834         WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2835         if ((ws = workQueues) != null) {
2836             for (int i = 0; i < ws.length; ++i) {
2837                 if ((w = ws[i]) != null) {
2838                     while ((t = w.poll()) != null) {
2839                         c.add(t);
2840                         ++count;
2841                     }
2842                 }
2843             }
2844         }
2845         return count;
2846     }
2847 
2848     /**
2849      * Returns a string identifying this pool, as well as its state,
2850      * including indications of run state, parallelism level, and
2851      * worker and task counts.
2852      *
2853      * @return a string identifying this pool, as well as its state
2854      */
toString()2855     public String toString() {
2856         // Use a single pass through workQueues to collect counts
2857         long qt = 0L, qs = 0L; int rc = 0;
2858         long st = stealCount;
2859         long c = ctl;
2860         WorkQueue[] ws; WorkQueue w;
2861         if ((ws = workQueues) != null) {
2862             for (int i = 0; i < ws.length; ++i) {
2863                 if ((w = ws[i]) != null) {
2864                     int size = w.queueSize();
2865                     if ((i & 1) == 0)
2866                         qs += size;
2867                     else {
2868                         qt += size;
2869                         st += w.nsteals;
2870                         if (w.isApparentlyUnblocked())
2871                             ++rc;
2872                     }
2873                 }
2874             }
2875         }
2876         int pc = parallelism;
2877         int tc = pc + (short)(c >>> TC_SHIFT);
2878         int ac = pc + (int)(c >> AC_SHIFT);
2879         if (ac < 0) // ignore transient negative
2880             ac = 0;
2881         String level;
2882         if ((c & STOP_BIT) != 0)
2883             level = (tc == 0) ? "Terminated" : "Terminating";
2884         else
2885             level = plock < 0 ? "Shutting down" : "Running";
2886         return super.toString() +
2887             "[" + level +
2888             ", parallelism = " + pc +
2889             ", size = " + tc +
2890             ", active = " + ac +
2891             ", running = " + rc +
2892             ", steals = " + st +
2893             ", tasks = " + qt +
2894             ", submissions = " + qs +
2895             "]";
2896     }
2897 
2898     /**
2899      * Possibly initiates an orderly shutdown in which previously
2900      * submitted tasks are executed, but no new tasks will be
2901      * accepted. Invocation has no effect on execution state if this
2902      * is the {@code commonPool()}, and no additional effect if
2903      * already shut down.  Tasks that are in the process of being
2904      * submitted concurrently during the course of this method may or
2905      * may not be rejected.
2906      */
2907     public void shutdown() {
2908         checkPermission();
2909         tryTerminate(false, true);
2910     }
2911 
2912     /**
2913      * Possibly attempts to cancel and/or stop all tasks, and reject
2914      * all subsequently submitted tasks.  Invocation has no effect on
2915      * execution state if this is the {@code commonPool()}, and no
2916      * additional effect if already shut down. Otherwise, tasks that
2917      * are in the process of being submitted or executed concurrently
2918      * during the course of this method may or may not be
2919      * rejected. This method cancels both existing and unexecuted
2920      * tasks, in order to permit termination in the presence of task
2921      * dependencies. So the method always returns an empty list
2922      * (unlike the case for some other Executors).
2923      *
2924      * @return an empty list
2925      */
2926     public List<Runnable> shutdownNow() {
2927         checkPermission();
2928         tryTerminate(true, true);
2929         return Collections.emptyList();
2930     }
2931 
2932     /**
2933      * Returns {@code true} if all tasks have completed following shut down.
2934      *
2935      * @return {@code true} if all tasks have completed following shut down
2936      */
2937     public boolean isTerminated() {
2938         long c = ctl;
2939         return ((c & STOP_BIT) != 0L &&
2940                 (short)(c >>> TC_SHIFT) + parallelism <= 0);
2941     }
2942 
2943     /**
2944      * Returns {@code true} if the process of termination has
2945      * commenced but not yet completed.  This method may be useful for
2946      * debugging. A return of {@code true} reported a sufficient
2947      * period after shutdown may indicate that submitted tasks have
2948      * ignored or suppressed interruption, or are waiting for I/O,
2949      * causing this executor not to properly terminate. (See the
2950      * advisory notes for class {@link ForkJoinTask} stating that
2951      * tasks should not normally entail blocking operations.  But if
2952      * they do, they must abort them on interrupt.)
2953      *
2954      * @return {@code true} if terminating but not yet terminated
2955      */
isTerminating()2956     public boolean isTerminating() {
2957         long c = ctl;
2958         return ((c & STOP_BIT) != 0L &&
2959                 (short)(c >>> TC_SHIFT) + parallelism > 0);
2960     }
2961 
2962     /**
2963      * Returns {@code true} if this pool has been shut down.
2964      *
2965      * @return {@code true} if this pool has been shut down
2966      */
isShutdown()2967     public boolean isShutdown() {
2968         return plock < 0;
2969     }
2970 
2971     /**
2972      * Blocks until all tasks have completed execution after a
2973      * shutdown request, or the timeout occurs, or the current thread
2974      * is interrupted, whichever happens first. Because the {@code
2975      * commonPool()} never terminates until program shutdown, when
2976      * applied to the common pool, this method is equivalent to {@link
2977      * #awaitQuiescence(long, TimeUnit)} but always returns {@code false}.
2978      *
2979      * @param timeout the maximum time to wait
2980      * @param unit the time unit of the timeout argument
2981      * @return {@code true} if this executor terminated and
2982      *         {@code false} if the timeout elapsed before termination
2983      * @throws InterruptedException if interrupted while waiting
2984      */
awaitTermination(long timeout, TimeUnit unit)2985     public boolean awaitTermination(long timeout, TimeUnit unit)
2986         throws InterruptedException {
2987         if (Thread.interrupted())
2988             throw new InterruptedException();
2989         if (this == common) {
2990             awaitQuiescence(timeout, unit);
2991             return false;
2992         }
2993         long nanos = unit.toNanos(timeout);
2994         if (isTerminated())
2995             return true;
2996         if (nanos <= 0L)
2997             return false;
2998         long deadline = System.nanoTime() + nanos;
2999         synchronized (this) {
3000             for (;;) {
3001                 if (isTerminated())
3002                     return true;
3003                 if (nanos <= 0L)
3004                     return false;
3005                 long millis = TimeUnit.NANOSECONDS.toMillis(nanos);
3006                 wait(millis > 0L ? millis : 1L);
3007                 nanos = deadline - System.nanoTime();
3008             }
3009         }
3010     }
3011 
3012     /**
3013      * If called by a ForkJoinTask operating in this pool, equivalent
3014      * in effect to {@link ForkJoinTask#helpQuiesce}. Otherwise,
3015      * waits and/or attempts to assist performing tasks until this
3016      * pool {@link #isQuiescent} or the indicated timeout elapses.
3017      *
3018      * @param timeout the maximum time to wait
3019      * @param unit the time unit of the timeout argument
3020      * @return {@code true} if quiescent; {@code false} if the
3021      * timeout elapsed.
3022      */
awaitQuiescence(long timeout, TimeUnit unit)3023     public boolean awaitQuiescence(long timeout, TimeUnit unit) {
3024         long nanos = unit.toNanos(timeout);
3025         ForkJoinWorkerThread wt;
3026         Thread thread = Thread.currentThread();
3027         if ((thread instanceof ForkJoinWorkerThread) &&
3028             (wt = (ForkJoinWorkerThread)thread).pool == this) {
3029             helpQuiescePool(wt.workQueue);
3030             return true;
3031         }
3032         long startTime = System.nanoTime();
3033         WorkQueue[] ws;
3034         int r = 0, m;
3035         boolean found = true;
3036         while (!isQuiescent() && (ws = workQueues) != null &&
3037                (m = ws.length - 1) >= 0) {
3038             if (!found) {
3039                 if ((System.nanoTime() - startTime) > nanos)
3040                     return false;
3041                 Thread.yield(); // cannot block
3042             }
3043             found = false;
3044             for (int j = (m + 1) << 2; j >= 0; --j) {
3045                 ForkJoinTask<?> t; WorkQueue q; int b;
3046                 if ((q = ws[r++ & m]) != null && (b = q.base) - q.top < 0) {
3047                     found = true;
3048                     if ((t = q.pollAt(b)) != null)
3049                         t.doExec();
3050                     break;
3051                 }
3052             }
3053         }
3054         return true;
3055     }
3056 
3057     /**
3058      * Waits and/or attempts to assist performing tasks indefinitely
3059      * until the {@code commonPool()} {@link #isQuiescent}.
3060      */
quiesceCommonPool()3061     static void quiesceCommonPool() {
3062         common.awaitQuiescence(Long.MAX_VALUE, TimeUnit.NANOSECONDS);
3063     }
3064 
3065     /**
3066      * Interface for extending managed parallelism for tasks running
3067      * in {@link ForkJoinPool}s.
3068      *
3069      * <p>A {@code ManagedBlocker} provides two methods.  Method
3070      * {@code isReleasable} must return {@code true} if blocking is
3071      * not necessary. Method {@code block} blocks the current thread
3072      * if necessary (perhaps internally invoking {@code isReleasable}
3073      * before actually blocking). These actions are performed by any
3074      * thread invoking {@link ForkJoinPool#managedBlock(ManagedBlocker)}.
3075      * The unusual methods in this API accommodate synchronizers that
3076      * may, but don't usually, block for long periods. Similarly, they
3077      * allow more efficient internal handling of cases in which
3078      * additional workers may be, but usually are not, needed to
3079      * ensure sufficient parallelism.  Toward this end,
3080      * implementations of method {@code isReleasable} must be amenable
3081      * to repeated invocation.
3082      *
3083      * <p>For example, here is a ManagedBlocker based on a
3084      * ReentrantLock:
3085      *  <pre> {@code
3086      * class ManagedLocker implements ManagedBlocker {
3087      *   final ReentrantLock lock;
3088      *   boolean hasLock = false;
3089      *   ManagedLocker(ReentrantLock lock) { this.lock = lock; }
3090      *   public boolean block() {
3091      *     if (!hasLock)
3092      *       lock.lock();
3093      *     return true;
3094      *   }
3095      *   public boolean isReleasable() {
3096      *     return hasLock || (hasLock = lock.tryLock());
3097      *   }
3098      * }}</pre>
3099      *
3100      * <p>Here is a class that possibly blocks waiting for an
3101      * item on a given queue:
3102      *  <pre> {@code
3103      * class QueueTaker<E> implements ManagedBlocker {
3104      *   final BlockingQueue<E> queue;
3105      *   volatile E item = null;
3106      *   QueueTaker(BlockingQueue<E> q) { this.queue = q; }
3107      *   public boolean block() throws InterruptedException {
3108      *     if (item == null)
3109      *       item = queue.take();
3110      *     return true;
3111      *   }
3112      *   public boolean isReleasable() {
3113      *     return item != null || (item = queue.poll()) != null;
3114      *   }
3115      *   public E getItem() { // call after pool.managedBlock completes
3116      *     return item;
3117      *   }
3118      * }}</pre>
3119      */
3120     public static interface ManagedBlocker {
3121         /**
3122          * Possibly blocks the current thread, for example waiting for
3123          * a lock or condition.
3124          *
3125          * @return {@code true} if no additional blocking is necessary
3126          * (i.e., if isReleasable would return true)
3127          * @throws InterruptedException if interrupted while waiting
3128          * (the method is not required to do so, but is allowed to)
3129          */
block()3130         boolean block() throws InterruptedException;
3131 
3132         /**
3133          * Returns {@code true} if blocking is unnecessary.
3134          * @return {@code true} if blocking is unnecessary
3135          */
isReleasable()3136         boolean isReleasable();
3137     }
3138 
3139     /**
3140      * Blocks in accord with the given blocker.  If the current thread
3141      * is a {@link ForkJoinWorkerThread}, this method possibly
3142      * arranges for a spare thread to be activated if necessary to
3143      * ensure sufficient parallelism while the current thread is blocked.
3144      *
3145      * <p>If the caller is not a {@link ForkJoinTask}, this method is
3146      * behaviorally equivalent to
3147      *  <pre> {@code
3148      * while (!blocker.isReleasable())
3149      *   if (blocker.block())
3150      *     return;
3151      * }</pre>
3152      *
3153      * If the caller is a {@code ForkJoinTask}, then the pool may
3154      * first be expanded to ensure parallelism, and later adjusted.
3155      *
3156      * @param blocker the blocker
3157      * @throws InterruptedException if blocker.block did so
3158      */
managedBlock(ManagedBlocker blocker)3159     public static void managedBlock(ManagedBlocker blocker)
3160         throws InterruptedException {
3161         Thread t = Thread.currentThread();
3162         if (t instanceof ForkJoinWorkerThread) {
3163             ForkJoinPool p = ((ForkJoinWorkerThread)t).pool;
3164             while (!blocker.isReleasable()) {
3165                 if (p.tryCompensate(p.ctl)) {
3166                     try {
3167                         do {} while (!blocker.isReleasable() &&
3168                                      !blocker.block());
3169                     } finally {
3170                         p.incrementActiveCount();
3171                     }
3172                     break;
3173                 }
3174             }
3175         }
3176         else {
3177             do {} while (!blocker.isReleasable() &&
3178                          !blocker.block());
3179         }
3180     }
3181 
3182     // AbstractExecutorService overrides.  These rely on undocumented
3183     // fact that ForkJoinTask.adapt returns ForkJoinTasks that also
3184     // implement RunnableFuture.
3185 
newTaskFor(Runnable runnable, T value)3186     protected <T> RunnableFuture<T> newTaskFor(Runnable runnable, T value) {
3187         return new ForkJoinTask.AdaptedRunnable<T>(runnable, value);
3188     }
3189 
newTaskFor(Callable<T> callable)3190     protected <T> RunnableFuture<T> newTaskFor(Callable<T> callable) {
3191         return new ForkJoinTask.AdaptedCallable<T>(callable);
3192     }
3193 
3194     // Unsafe mechanics
3195     private static final sun.misc.Unsafe U;
3196     private static final long CTL;
3197     private static final long PARKBLOCKER;
3198     private static final int ABASE;
3199     private static final int ASHIFT;
3200     private static final long STEALCOUNT;
3201     private static final long PLOCK;
3202     private static final long INDEXSEED;
3203     private static final long QBASE;
3204     private static final long QLOCK;
3205 
3206     static {
3207         // initialize field offsets for CAS etc
3208         try {
3209             U = sun.misc.Unsafe.getUnsafe();
3210             Class<?> k = ForkJoinPool.class;
3211             CTL = U.objectFieldOffset
3212                 (k.getDeclaredField("ctl"));
3213             STEALCOUNT = U.objectFieldOffset
3214                 (k.getDeclaredField("stealCount"));
3215             PLOCK = U.objectFieldOffset
3216                 (k.getDeclaredField("plock"));
3217             INDEXSEED = U.objectFieldOffset
3218                 (k.getDeclaredField("indexSeed"));
3219             Class<?> tk = Thread.class;
3220             PARKBLOCKER = U.objectFieldOffset
3221                 (tk.getDeclaredField("parkBlocker"));
3222             Class<?> wk = WorkQueue.class;
3223             QBASE = U.objectFieldOffset
3224                 (wk.getDeclaredField("base"));
3225             QLOCK = U.objectFieldOffset
3226                 (wk.getDeclaredField("qlock"));
3227             Class<?> ak = ForkJoinTask[].class;
3228             ABASE = U.arrayBaseOffset(ak);
3229             int scale = U.arrayIndexScale(ak);
3230             if ((scale & (scale - 1)) != 0)
3231                 throw new Error("data type scale not a power of two");
3232             ASHIFT = 31 - Integer.numberOfLeadingZeros(scale);
3233         } catch (Exception e) {
3234             throw new Error(e);
3235         }
3236 
3237         submitters = new ThreadLocal<Submitter>();
3238         defaultForkJoinWorkerThreadFactory =
3239             new DefaultForkJoinWorkerThreadFactory();
3240         modifyThreadPermission = new RuntimePermission("modifyThread");
3241 
3242         common = java.security.AccessController.doPrivileged
3243             (new java.security.PrivilegedAction<ForkJoinPool>() {
3244                 public ForkJoinPool run() { return makeCommonPool(); }});
3245         int par = common.parallelism; // report 1 even if threads disabled
3246         commonParallelism = par > 0 ? par : 1;
3247     }
3248 
3249     /**
3250      * Creates and returns the common pool, respecting user settings
3251      * specified via system properties.
3252      */
makeCommonPool()3253     private static ForkJoinPool makeCommonPool() {
3254         int parallelism = -1;
3255         ForkJoinWorkerThreadFactory factory
3256             = defaultForkJoinWorkerThreadFactory;
3257         UncaughtExceptionHandler handler = null;
3258         try {  // ignore exceptions in accessing/parsing properties
3259             String pp = System.getProperty
3260                 ("java.util.concurrent.ForkJoinPool.common.parallelism");
3261             String fp = System.getProperty
3262                 ("java.util.concurrent.ForkJoinPool.common.threadFactory");
3263             String hp = System.getProperty
3264                 ("java.util.concurrent.ForkJoinPool.common.exceptionHandler");
3265             if (pp != null)
3266                 parallelism = Integer.parseInt(pp);
3267             if (fp != null)
3268                 factory = ((ForkJoinWorkerThreadFactory)ClassLoader.
3269                            getSystemClassLoader().loadClass(fp).newInstance());
3270             if (hp != null)
3271                 handler = ((UncaughtExceptionHandler)ClassLoader.
3272                            getSystemClassLoader().loadClass(hp).newInstance());
3273         } catch (Exception ignore) {
3274         }
3275 
3276         if (parallelism < 0 && // default 1 less than #cores
3277             (parallelism = Runtime.getRuntime().availableProcessors() - 1) < 0)
3278             parallelism = 0;
3279         if (parallelism > MAX_CAP)
3280             parallelism = MAX_CAP;
3281         return new ForkJoinPool(parallelism, factory, handler, LIFO_QUEUE,
3282                                 "ForkJoinPool.commonPool-worker-");
3283     }
3284 
3285 }
3286