ForkJoinPool.java revision a807b4d808d2591894daf13aab179b2e9c46a2f5
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
7package java.util.concurrent;
8
9import java.util.ArrayList;
10import java.util.Arrays;
11import java.util.Collection;
12import java.util.Collections;
13import java.util.List;
14import java.util.Random;
15import java.util.concurrent.AbstractExecutorService;
16import java.util.concurrent.Callable;
17import java.util.concurrent.ExecutorService;
18import java.util.concurrent.Future;
19import java.util.concurrent.RejectedExecutionException;
20import java.util.concurrent.RunnableFuture;
21import java.util.concurrent.TimeUnit;
22import java.util.concurrent.atomic.AtomicInteger;
23import java.util.concurrent.locks.LockSupport;
24import java.util.concurrent.locks.ReentrantLock;
25import java.util.concurrent.locks.Condition;
26import libcore.util.SneakyThrow;
27
28// BEGIN android-note
29// removed security manager docs
30// END android-note
31
32/**
33 * An {@link ExecutorService} for running {@link ForkJoinTask}s.
34 * A {@code ForkJoinPool} provides the entry point for submissions
35 * from non-{@code ForkJoinTask} clients, as well as management and
36 * monitoring operations.
37 *
38 * <p>A {@code ForkJoinPool} differs from other kinds of {@link
39 * ExecutorService} mainly by virtue of employing
40 * <em>work-stealing</em>: all threads in the pool attempt to find and
41 * execute subtasks created by other active tasks (eventually blocking
42 * waiting for work if none exist). This enables efficient processing
43 * when most tasks spawn other subtasks (as do most {@code
44 * ForkJoinTask}s). When setting <em>asyncMode</em> to true in
45 * constructors, {@code ForkJoinPool}s may also be appropriate for use
46 * with event-style tasks that are never joined.
47 *
48 * <p>A {@code ForkJoinPool} is constructed with a given target
49 * parallelism level; by default, equal to the number of available
50 * processors. The pool attempts to maintain enough active (or
51 * available) threads by dynamically adding, suspending, or resuming
52 * internal worker threads, even if some tasks are stalled waiting to
53 * join others. However, no such adjustments are guaranteed in the
54 * face of blocked IO or other unmanaged synchronization. The nested
55 * {@link ManagedBlocker} interface enables extension of the kinds of
56 * synchronization accommodated.
57 *
58 * <p>In addition to execution and lifecycle control methods, this
59 * class provides status check methods (for example
60 * {@link #getStealCount}) that are intended to aid in developing,
61 * tuning, and monitoring fork/join applications. Also, method
62 * {@link #toString} returns indications of pool state in a
63 * convenient form for informal monitoring.
64 *
65 * <p> As is the case with other ExecutorServices, there are three
66 * main task execution methods summarized in the following
67 * table. These are designed to be used by clients not already engaged
68 * in fork/join computations in the current pool.  The main forms of
69 * these methods accept instances of {@code ForkJoinTask}, but
70 * overloaded forms also allow mixed execution of plain {@code
71 * Runnable}- or {@code Callable}- based activities as well.  However,
72 * tasks that are already executing in a pool should normally
73 * <em>NOT</em> use these pool execution methods, but instead use the
74 * within-computation forms listed in the table.
75 *
76 * <table BORDER CELLPADDING=3 CELLSPACING=1>
77 *  <tr>
78 *    <td></td>
79 *    <td ALIGN=CENTER> <b>Call from non-fork/join clients</b></td>
80 *    <td ALIGN=CENTER> <b>Call from within fork/join computations</b></td>
81 *  </tr>
82 *  <tr>
83 *    <td> <b>Arrange async execution</td>
84 *    <td> {@link #execute(ForkJoinTask)}</td>
85 *    <td> {@link ForkJoinTask#fork}</td>
86 *  </tr>
87 *  <tr>
88 *    <td> <b>Await and obtain result</td>
89 *    <td> {@link #invoke(ForkJoinTask)}</td>
90 *    <td> {@link ForkJoinTask#invoke}</td>
91 *  </tr>
92 *  <tr>
93 *    <td> <b>Arrange exec and obtain Future</td>
94 *    <td> {@link #submit(ForkJoinTask)}</td>
95 *    <td> {@link ForkJoinTask#fork} (ForkJoinTasks <em>are</em> Futures)</td>
96 *  </tr>
97 * </table>
98 *
99 * <p><b>Sample Usage.</b> Normally a single {@code ForkJoinPool} is
100 * used for all parallel task execution in a program or subsystem.
101 * Otherwise, use would not usually outweigh the construction and
102 * bookkeeping overhead of creating a large set of threads. For
103 * example, a common pool could be used for the {@code SortTasks}
104 * illustrated in {@link RecursiveAction}. Because {@code
105 * ForkJoinPool} uses threads in {@linkplain java.lang.Thread#isDaemon
106 * daemon} mode, there is typically no need to explicitly {@link
107 * #shutdown} such a pool upon program exit.
108 *
109 *  <pre> {@code
110 * static final ForkJoinPool mainPool = new ForkJoinPool();
111 * ...
112 * public void sort(long[] array) {
113 *   mainPool.invoke(new SortTask(array, 0, array.length));
114 * }}</pre>
115 *
116 * <p><b>Implementation notes</b>: This implementation restricts the
117 * maximum number of running threads to 32767. Attempts to create
118 * pools with greater than the maximum number result in
119 * {@code IllegalArgumentException}.
120 *
121 * <p>This implementation rejects submitted tasks (that is, by throwing
122 * {@link RejectedExecutionException}) only when the pool is shut down
123 * or internal resources have been exhausted.
124 *
125 * @since 1.7
126 * @hide
127 * @author Doug Lea
128 */
129public class ForkJoinPool extends AbstractExecutorService {
130
131    /*
132     * Implementation Overview
133     *
134     * This class provides the central bookkeeping and control for a
135     * set of worker threads: Submissions from non-FJ threads enter
136     * into a submission queue. Workers take these tasks and typically
137     * split them into subtasks that may be stolen by other workers.
138     * Preference rules give first priority to processing tasks from
139     * their own queues (LIFO or FIFO, depending on mode), then to
140     * randomized FIFO steals of tasks in other worker queues, and
141     * lastly to new submissions.
142     *
143     * The main throughput advantages of work-stealing stem from
144     * decentralized control -- workers mostly take tasks from
145     * themselves or each other. We cannot negate this in the
146     * implementation of other management responsibilities. The main
147     * tactic for avoiding bottlenecks is packing nearly all
148     * essentially atomic control state into a single 64bit volatile
149     * variable ("ctl"). This variable is read on the order of 10-100
150     * times as often as it is modified (always via CAS). (There is
151     * some additional control state, for example variable "shutdown"
152     * for which we can cope with uncoordinated updates.)  This
153     * streamlines synchronization and control at the expense of messy
154     * constructions needed to repack status bits upon updates.
155     * Updates tend not to contend with each other except during
156     * bursts while submitted tasks begin or end.  In some cases when
157     * they do contend, threads can instead do something else
158     * (usually, scan for tasks) until contention subsides.
159     *
160     * To enable packing, we restrict maximum parallelism to (1<<15)-1
161     * (which is far in excess of normal operating range) to allow
162     * ids, counts, and their negations (used for thresholding) to fit
163     * into 16bit fields.
164     *
165     * Recording Workers.  Workers are recorded in the "workers" array
166     * that is created upon pool construction and expanded if (rarely)
167     * necessary.  This is an array as opposed to some other data
168     * structure to support index-based random steals by workers.
169     * Updates to the array recording new workers and unrecording
170     * terminated ones are protected from each other by a seqLock
171     * (scanGuard) but the array is otherwise concurrently readable,
172     * and accessed directly by workers. To simplify index-based
173     * operations, the array size is always a power of two, and all
174     * readers must tolerate null slots. To avoid flailing during
175     * start-up, the array is presized to hold twice #parallelism
176     * workers (which is unlikely to need further resizing during
177     * execution). But to avoid dealing with so many null slots,
178     * variable scanGuard includes a mask for the nearest power of two
179     * that contains all current workers.  All worker thread creation
180     * is on-demand, triggered by task submissions, replacement of
181     * terminated workers, and/or compensation for blocked
182     * workers. However, all other support code is set up to work with
183     * other policies.  To ensure that we do not hold on to worker
184     * references that would prevent GC, ALL accesses to workers are
185     * via indices into the workers array (which is one source of some
186     * of the messy code constructions here). In essence, the workers
187     * array serves as a weak reference mechanism. Thus for example
188     * the wait queue field of ctl stores worker indices, not worker
189     * references.  Access to the workers in associated methods (for
190     * example signalWork) must both index-check and null-check the
191     * IDs. All such accesses ignore bad IDs by returning out early
192     * from what they are doing, since this can only be associated
193     * with termination, in which case it is OK to give up.
194     *
195     * All uses of the workers array, as well as queue arrays, check
196     * that the array is non-null (even if previously non-null). This
197     * allows nulling during termination, which is currently not
198     * necessary, but remains an option for resource-revocation-based
199     * shutdown schemes.
200     *
201     * Wait Queuing. Unlike HPC work-stealing frameworks, we cannot
202     * let workers spin indefinitely scanning for tasks when none can
203     * be found immediately, and we cannot start/resume workers unless
204     * there appear to be tasks available.  On the other hand, we must
205     * quickly prod them into action when new tasks are submitted or
206     * generated.  We park/unpark workers after placing in an event
207     * wait queue when they cannot find work. This "queue" is actually
208     * a simple Treiber stack, headed by the "id" field of ctl, plus a
209     * 15bit counter value to both wake up waiters (by advancing their
210     * count) and avoid ABA effects. Successors are held in worker
211     * field "nextWait".  Queuing deals with several intrinsic races,
212     * mainly that a task-producing thread can miss seeing (and
213     * signalling) another thread that gave up looking for work but
214     * has not yet entered the wait queue. We solve this by requiring
215     * a full sweep of all workers both before (in scan()) and after
216     * (in tryAwaitWork()) a newly waiting worker is added to the wait
217     * queue. During a rescan, the worker might release some other
218     * queued worker rather than itself, which has the same net
219     * effect. Because enqueued workers may actually be rescanning
220     * rather than waiting, we set and clear the "parked" field of
221     * ForkJoinWorkerThread to reduce unnecessary calls to unpark.
222     * (Use of the parked field requires a secondary recheck to avoid
223     * missed signals.)
224     *
225     * Signalling.  We create or wake up workers only when there
226     * appears to be at least one task they might be able to find and
227     * execute.  When a submission is added or another worker adds a
228     * task to a queue that previously had two or fewer tasks, they
229     * signal waiting workers (or trigger creation of new ones if
230     * fewer than the given parallelism level -- see signalWork).
231     * These primary signals are buttressed by signals during rescans
232     * as well as those performed when a worker steals a task and
233     * notices that there are more tasks too; together these cover the
234     * signals needed in cases when more than two tasks are pushed
235     * but untaken.
236     *
237     * Trimming workers. To release resources after periods of lack of
238     * use, a worker starting to wait when the pool is quiescent will
239     * time out and terminate if the pool has remained quiescent for
240     * SHRINK_RATE nanosecs. This will slowly propagate, eventually
241     * terminating all workers after long periods of non-use.
242     *
243     * Submissions. External submissions are maintained in an
244     * array-based queue that is structured identically to
245     * ForkJoinWorkerThread queues except for the use of
246     * submissionLock in method addSubmission. Unlike the case for
247     * worker queues, multiple external threads can add new
248     * submissions, so adding requires a lock.
249     *
250     * Compensation. Beyond work-stealing support and lifecycle
251     * control, the main responsibility of this framework is to take
252     * actions when one worker is waiting to join a task stolen (or
253     * always held by) another.  Because we are multiplexing many
254     * tasks on to a pool of workers, we can't just let them block (as
255     * in Thread.join).  We also cannot just reassign the joiner's
256     * run-time stack with another and replace it later, which would
257     * be a form of "continuation", that even if possible is not
258     * necessarily a good idea since we sometimes need both an
259     * unblocked task and its continuation to progress. Instead we
260     * combine two tactics:
261     *
262     *   Helping: Arranging for the joiner to execute some task that it
263     *      would be running if the steal had not occurred.  Method
264     *      ForkJoinWorkerThread.joinTask tracks joining->stealing
265     *      links to try to find such a task.
266     *
267     *   Compensating: Unless there are already enough live threads,
268     *      method tryPreBlock() may create or re-activate a spare
269     *      thread to compensate for blocked joiners until they
270     *      unblock.
271     *
272     * The ManagedBlocker extension API can't use helping so relies
273     * only on compensation in method awaitBlocker.
274     *
275     * It is impossible to keep exactly the target parallelism number
276     * of threads running at any given time.  Determining the
277     * existence of conservatively safe helping targets, the
278     * availability of already-created spares, and the apparent need
279     * to create new spares are all racy and require heuristic
280     * guidance, so we rely on multiple retries of each.  Currently,
281     * in keeping with on-demand signalling policy, we compensate only
282     * if blocking would leave less than one active (non-waiting,
283     * non-blocked) worker. Additionally, to avoid some false alarms
284     * due to GC, lagging counters, system activity, etc, compensated
285     * blocking for joins is only attempted after rechecks stabilize
286     * (retries are interspersed with Thread.yield, for good
287     * citizenship).  The variable blockedCount, incremented before
288     * blocking and decremented after, is sometimes needed to
289     * distinguish cases of waiting for work vs blocking on joins or
290     * other managed sync. Both cases are equivalent for most pool
291     * control, so we can update non-atomically. (Additionally,
292     * contention on blockedCount alleviates some contention on ctl).
293     *
294     * Shutdown and Termination. A call to shutdownNow atomically sets
295     * the ctl stop bit and then (non-atomically) sets each workers
296     * "terminate" status, cancels all unprocessed tasks, and wakes up
297     * all waiting workers.  Detecting whether termination should
298     * commence after a non-abrupt shutdown() call requires more work
299     * and bookkeeping. We need consensus about quiescence (i.e., that
300     * there is no more work) which is reflected in active counts so
301     * long as there are no current blockers, as well as possible
302     * re-evaluations during independent changes in blocking or
303     * quiescing workers.
304     *
305     * Style notes: There is a lot of representation-level coupling
306     * among classes ForkJoinPool, ForkJoinWorkerThread, and
307     * ForkJoinTask.  Most fields of ForkJoinWorkerThread maintain
308     * data structures managed by ForkJoinPool, so are directly
309     * accessed.  Conversely we allow access to "workers" array by
310     * workers, and direct access to ForkJoinTask.status by both
311     * ForkJoinPool and ForkJoinWorkerThread.  There is little point
312     * trying to reduce this, since any associated future changes in
313     * representations will need to be accompanied by algorithmic
314     * changes anyway. All together, these low-level implementation
315     * choices produce as much as a factor of 4 performance
316     * improvement compared to naive implementations, and enable the
317     * processing of billions of tasks per second, at the expense of
318     * some ugliness.
319     *
320     * Methods signalWork() and scan() are the main bottlenecks so are
321     * especially heavily micro-optimized/mangled.  There are lots of
322     * inline assignments (of form "while ((local = field) != 0)")
323     * which are usually the simplest way to ensure the required read
324     * orderings (which are sometimes critical). This leads to a
325     * "C"-like style of listing declarations of these locals at the
326     * heads of methods or blocks.  There are several occurrences of
327     * the unusual "do {} while (!cas...)"  which is the simplest way
328     * to force an update of a CAS'ed variable. There are also other
329     * coding oddities that help some methods perform reasonably even
330     * when interpreted (not compiled).
331     *
332     * The order of declarations in this file is: (1) declarations of
333     * statics (2) fields (along with constants used when unpacking
334     * some of them), listed in an order that tends to reduce
335     * contention among them a bit under most JVMs.  (3) internal
336     * control methods (4) callbacks and other support for
337     * ForkJoinTask and ForkJoinWorkerThread classes, (5) exported
338     * methods (plus a few little helpers). (6) static block
339     * initializing all statics in a minimally dependent order.
340     */
341
342    /**
343     * Factory for creating new {@link ForkJoinWorkerThread}s.
344     * A {@code ForkJoinWorkerThreadFactory} must be defined and used
345     * for {@code ForkJoinWorkerThread} subclasses that extend base
346     * functionality or initialize threads with different contexts.
347     */
348    public static interface ForkJoinWorkerThreadFactory {
349        /**
350         * Returns a new worker thread operating in the given pool.
351         *
352         * @param pool the pool this thread works in
353         * @throws NullPointerException if the pool is null
354         */
355        public ForkJoinWorkerThread newThread(ForkJoinPool pool);
356    }
357
358    /**
359     * Default ForkJoinWorkerThreadFactory implementation; creates a
360     * new ForkJoinWorkerThread.
361     */
362    static class DefaultForkJoinWorkerThreadFactory
363        implements ForkJoinWorkerThreadFactory {
364        public ForkJoinWorkerThread newThread(ForkJoinPool pool) {
365            return new ForkJoinWorkerThread(pool);
366        }
367    }
368
369    /**
370     * Creates a new ForkJoinWorkerThread. This factory is used unless
371     * overridden in ForkJoinPool constructors.
372     */
373    public static final ForkJoinWorkerThreadFactory
374        defaultForkJoinWorkerThreadFactory;
375
376    /**
377     * Permission required for callers of methods that may start or
378     * kill threads.
379     */
380    private static final RuntimePermission modifyThreadPermission;
381
382    /**
383     * If there is a security manager, makes sure caller has
384     * permission to modify threads.
385     */
386    private static void checkPermission() {
387        SecurityManager security = System.getSecurityManager();
388        if (security != null)
389            security.checkPermission(modifyThreadPermission);
390    }
391
392    /**
393     * Generator for assigning sequence numbers as pool names.
394     */
395    private static final AtomicInteger poolNumberGenerator;
396
397    /**
398     * Generator for initial random seeds for worker victim
399     * selection. This is used only to create initial seeds. Random
400     * steals use a cheaper xorshift generator per steal attempt. We
401     * don't expect much contention on seedGenerator, so just use a
402     * plain Random.
403     */
404    static final Random workerSeedGenerator;
405
406    /**
407     * Array holding all worker threads in the pool.  Initialized upon
408     * construction. Array size must be a power of two.  Updates and
409     * replacements are protected by scanGuard, but the array is
410     * always kept in a consistent enough state to be randomly
411     * accessed without locking by workers performing work-stealing,
412     * as well as other traversal-based methods in this class, so long
413     * as reads memory-acquire by first reading ctl. All readers must
414     * tolerate that some array slots may be null.
415     */
416    ForkJoinWorkerThread[] workers;
417
418    /**
419     * Initial size for submission queue array. Must be a power of
420     * two.  In many applications, these always stay small so we use a
421     * small initial cap.
422     */
423    private static final int INITIAL_QUEUE_CAPACITY = 8;
424
425    /**
426     * Maximum size for submission queue array. Must be a power of two
427     * less than or equal to 1 << (31 - width of array entry) to
428     * ensure lack of index wraparound, but is capped at a lower
429     * value to help users trap runaway computations.
430     */
431    private static final int MAXIMUM_QUEUE_CAPACITY = 1 << 24; // 16M
432
433    /**
434     * Array serving as submission queue. Initialized upon construction.
435     */
436    private ForkJoinTask<?>[] submissionQueue;
437
438    /**
439     * Lock protecting submissions array for addSubmission
440     */
441    private final ReentrantLock submissionLock;
442
443    /**
444     * Condition for awaitTermination, using submissionLock for
445     * convenience.
446     */
447    private final Condition termination;
448
449    /**
450     * Creation factory for worker threads.
451     */
452    private final ForkJoinWorkerThreadFactory factory;
453
454    /**
455     * The uncaught exception handler used when any worker abruptly
456     * terminates.
457     */
458    final Thread.UncaughtExceptionHandler ueh;
459
460    /**
461     * Prefix for assigning names to worker threads
462     */
463    private final String workerNamePrefix;
464
465    /**
466     * Sum of per-thread steal counts, updated only when threads are
467     * idle or terminating.
468     */
469    private volatile long stealCount;
470
471    /**
472     * Main pool control -- a long packed with:
473     * AC: Number of active running workers minus target parallelism (16 bits)
474     * TC: Number of total workers minus target parallelism (16 bits)
475     * ST: true if pool is terminating (1 bit)
476     * EC: the wait count of top waiting thread (15 bits)
477     * ID: ~poolIndex of top of Treiber stack of waiting threads (16 bits)
478     *
479     * When convenient, we can extract the upper 32 bits of counts and
480     * the lower 32 bits of queue state, u = (int)(ctl >>> 32) and e =
481     * (int)ctl.  The ec field is never accessed alone, but always
482     * together with id and st. The offsets of counts by the target
483     * parallelism and the positionings of fields makes it possible to
484     * perform the most common checks via sign tests of fields: When
485     * ac is negative, there are not enough active workers, when tc is
486     * negative, there are not enough total workers, when id is
487     * negative, there is at least one waiting worker, and when e is
488     * negative, the pool is terminating.  To deal with these possibly
489     * negative fields, we use casts in and out of "short" and/or
490     * signed shifts to maintain signedness.
491     */
492    volatile long ctl;
493
494    // bit positions/shifts for fields
495    private static final int  AC_SHIFT   = 48;
496    private static final int  TC_SHIFT   = 32;
497    private static final int  ST_SHIFT   = 31;
498    private static final int  EC_SHIFT   = 16;
499
500    // bounds
501    private static final int  MAX_ID     = 0x7fff;  // max poolIndex
502    private static final int  SMASK      = 0xffff;  // mask short bits
503    private static final int  SHORT_SIGN = 1 << 15;
504    private static final int  INT_SIGN   = 1 << 31;
505
506    // masks
507    private static final long STOP_BIT   = 0x0001L << ST_SHIFT;
508    private static final long AC_MASK    = ((long)SMASK) << AC_SHIFT;
509    private static final long TC_MASK    = ((long)SMASK) << TC_SHIFT;
510
511    // units for incrementing and decrementing
512    private static final long TC_UNIT    = 1L << TC_SHIFT;
513    private static final long AC_UNIT    = 1L << AC_SHIFT;
514
515    // masks and units for dealing with u = (int)(ctl >>> 32)
516    private static final int  UAC_SHIFT  = AC_SHIFT - 32;
517    private static final int  UTC_SHIFT  = TC_SHIFT - 32;
518    private static final int  UAC_MASK   = SMASK << UAC_SHIFT;
519    private static final int  UTC_MASK   = SMASK << UTC_SHIFT;
520    private static final int  UAC_UNIT   = 1 << UAC_SHIFT;
521    private static final int  UTC_UNIT   = 1 << UTC_SHIFT;
522
523    // masks and units for dealing with e = (int)ctl
524    private static final int  E_MASK     = 0x7fffffff; // no STOP_BIT
525    private static final int  EC_UNIT    = 1 << EC_SHIFT;
526
527    /**
528     * The target parallelism level.
529     */
530    final int parallelism;
531
532    /**
533     * Index (mod submission queue length) of next element to take
534     * from submission queue. Usage is identical to that for
535     * per-worker queues -- see ForkJoinWorkerThread internal
536     * documentation.
537     */
538    volatile int queueBase;
539
540    /**
541     * Index (mod submission queue length) of next element to add
542     * in submission queue. Usage is identical to that for
543     * per-worker queues -- see ForkJoinWorkerThread internal
544     * documentation.
545     */
546    int queueTop;
547
548    /**
549     * True when shutdown() has been called.
550     */
551    volatile boolean shutdown;
552
553    /**
554     * True if use local fifo, not default lifo, for local polling.
555     * Read by, and replicated by ForkJoinWorkerThreads.
556     */
557    final boolean locallyFifo;
558
559    /**
560     * The number of threads in ForkJoinWorkerThreads.helpQuiescePool.
561     * When non-zero, suppresses automatic shutdown when active
562     * counts become zero.
563     */
564    volatile int quiescerCount;
565
566    /**
567     * The number of threads blocked in join.
568     */
569    volatile int blockedCount;
570
571    /**
572     * Counter for worker Thread names (unrelated to their poolIndex)
573     */
574    private volatile int nextWorkerNumber;
575
576    /**
577     * The index for the next created worker. Accessed under scanGuard.
578     */
579    private int nextWorkerIndex;
580
581    /**
582     * SeqLock and index masking for updates to workers array.  Locked
583     * when SG_UNIT is set. Unlocking clears bit by adding
584     * SG_UNIT. Staleness of read-only operations can be checked by
585     * comparing scanGuard to value before the reads. The low 16 bits
586     * (i.e, anding with SMASK) hold (the smallest power of two
587     * covering all worker indices, minus one, and is used to avoid
588     * dealing with large numbers of null slots when the workers array
589     * is overallocated.
590     */
591    volatile int scanGuard;
592
593    private static final int SG_UNIT = 1 << 16;
594
595    /**
596     * The wakeup interval (in nanoseconds) for a worker waiting for a
597     * task when the pool is quiescent to instead try to shrink the
598     * number of workers.  The exact value does not matter too
599     * much. It must be short enough to release resources during
600     * sustained periods of idleness, but not so short that threads
601     * are continually re-created.
602     */
603    private static final long SHRINK_RATE =
604        4L * 1000L * 1000L * 1000L; // 4 seconds
605
606    /**
607     * Top-level loop for worker threads: On each step: if the
608     * previous step swept through all queues and found no tasks, or
609     * there are excess threads, then possibly blocks. Otherwise,
610     * scans for and, if found, executes a task. Returns when pool
611     * and/or worker terminate.
612     *
613     * @param w the worker
614     */
615    final void work(ForkJoinWorkerThread w) {
616        boolean swept = false;                // true on empty scans
617        long c;
618        while (!w.terminate && (int)(c = ctl) >= 0) {
619            int a;                            // active count
620            if (!swept && (a = (int)(c >> AC_SHIFT)) <= 0)
621                swept = scan(w, a);
622            else if (tryAwaitWork(w, c))
623                swept = false;
624        }
625    }
626
627    // Signalling
628
629    /**
630     * Wakes up or creates a worker.
631     */
632    final void signalWork() {
633        /*
634         * The while condition is true if: (there is are too few total
635         * workers OR there is at least one waiter) AND (there are too
636         * few active workers OR the pool is terminating).  The value
637         * of e distinguishes the remaining cases: zero (no waiters)
638         * for create, negative if terminating (in which case do
639         * nothing), else release a waiter. The secondary checks for
640         * release (non-null array etc) can fail if the pool begins
641         * terminating after the test, and don't impose any added cost
642         * because JVMs must perform null and bounds checks anyway.
643         */
644        long c; int e, u;
645        while ((((e = (int)(c = ctl)) | (u = (int)(c >>> 32))) &
646                (INT_SIGN|SHORT_SIGN)) == (INT_SIGN|SHORT_SIGN) && e >= 0) {
647            if (e > 0) {                         // release a waiting worker
648                int i; ForkJoinWorkerThread w; ForkJoinWorkerThread[] ws;
649                if ((ws = workers) == null ||
650                    (i = ~e & SMASK) >= ws.length ||
651                    (w = ws[i]) == null)
652                    break;
653                long nc = (((long)(w.nextWait & E_MASK)) |
654                           ((long)(u + UAC_UNIT) << 32));
655                if (w.eventCount == e &&
656                    UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
657                    w.eventCount = (e + EC_UNIT) & E_MASK;
658                    if (w.parked)
659                        UNSAFE.unpark(w);
660                    break;
661                }
662            }
663            else if (UNSAFE.compareAndSwapLong
664                     (this, ctlOffset, c,
665                      (long)(((u + UTC_UNIT) & UTC_MASK) |
666                             ((u + UAC_UNIT) & UAC_MASK)) << 32)) {
667                addWorker();
668                break;
669            }
670        }
671    }
672
673    /**
674     * Variant of signalWork to help release waiters on rescans.
675     * Tries once to release a waiter if active count < 0.
676     *
677     * @return false if failed due to contention, else true
678     */
679    private boolean tryReleaseWaiter() {
680        long c; int e, i; ForkJoinWorkerThread w; ForkJoinWorkerThread[] ws;
681        if ((e = (int)(c = ctl)) > 0 &&
682            (int)(c >> AC_SHIFT) < 0 &&
683            (ws = workers) != null &&
684            (i = ~e & SMASK) < ws.length &&
685            (w = ws[i]) != null) {
686            long nc = ((long)(w.nextWait & E_MASK) |
687                       ((c + AC_UNIT) & (AC_MASK|TC_MASK)));
688            if (w.eventCount != e ||
689                !UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc))
690                return false;
691            w.eventCount = (e + EC_UNIT) & E_MASK;
692            if (w.parked)
693                UNSAFE.unpark(w);
694        }
695        return true;
696    }
697
698    // Scanning for tasks
699
700    /**
701     * Scans for and, if found, executes one task. Scans start at a
702     * random index of workers array, and randomly select the first
703     * (2*#workers)-1 probes, and then, if all empty, resort to 2
704     * circular sweeps, which is necessary to check quiescence. and
705     * taking a submission only if no stealable tasks were found.  The
706     * steal code inside the loop is a specialized form of
707     * ForkJoinWorkerThread.deqTask, followed bookkeeping to support
708     * helpJoinTask and signal propagation. The code for submission
709     * queues is almost identical. On each steal, the worker completes
710     * not only the task, but also all local tasks that this task may
711     * have generated. On detecting staleness or contention when
712     * trying to take a task, this method returns without finishing
713     * sweep, which allows global state rechecks before retry.
714     *
715     * @param w the worker
716     * @param a the number of active workers
717     * @return true if swept all queues without finding a task
718     */
719    private boolean scan(ForkJoinWorkerThread w, int a) {
720        int g = scanGuard; // mask 0 avoids useless scans if only one active
721        int m = (parallelism == 1 - a && blockedCount == 0) ? 0 : g & SMASK;
722        ForkJoinWorkerThread[] ws = workers;
723        if (ws == null || ws.length <= m)         // staleness check
724            return false;
725        for (int r = w.seed, k = r, j = -(m + m); j <= m + m; ++j) {
726            ForkJoinTask<?> t; ForkJoinTask<?>[] q; int b, i;
727            ForkJoinWorkerThread v = ws[k & m];
728            if (v != null && (b = v.queueBase) != v.queueTop &&
729                (q = v.queue) != null && (i = (q.length - 1) & b) >= 0) {
730                long u = (i << ASHIFT) + ABASE;
731                if ((t = q[i]) != null && v.queueBase == b &&
732                    UNSAFE.compareAndSwapObject(q, u, t, null)) {
733                    int d = (v.queueBase = b + 1) - v.queueTop;
734                    v.stealHint = w.poolIndex;
735                    if (d != 0)
736                        signalWork();             // propagate if nonempty
737                    w.execTask(t);
738                }
739                r ^= r << 13; r ^= r >>> 17; w.seed = r ^ (r << 5);
740                return false;                     // store next seed
741            }
742            else if (j < 0) {                     // xorshift
743                r ^= r << 13; r ^= r >>> 17; k = r ^= r << 5;
744            }
745            else
746                ++k;
747        }
748        if (scanGuard != g)                       // staleness check
749            return false;
750        else {                                    // try to take submission
751            ForkJoinTask<?> t; ForkJoinTask<?>[] q; int b, i;
752            if ((b = queueBase) != queueTop &&
753                (q = submissionQueue) != null &&
754                (i = (q.length - 1) & b) >= 0) {
755                long u = (i << ASHIFT) + ABASE;
756                if ((t = q[i]) != null && queueBase == b &&
757                    UNSAFE.compareAndSwapObject(q, u, t, null)) {
758                    queueBase = b + 1;
759                    w.execTask(t);
760                }
761                return false;
762            }
763            return true;                         // all queues empty
764        }
765    }
766
767    /**
768     * Tries to enqueue worker w in wait queue and await change in
769     * worker's eventCount.  If the pool is quiescent and there is
770     * more than one worker, possibly terminates worker upon exit.
771     * Otherwise, before blocking, rescans queues to avoid missed
772     * signals.  Upon finding work, releases at least one worker
773     * (which may be the current worker). Rescans restart upon
774     * detected staleness or failure to release due to
775     * contention. Note the unusual conventions about Thread.interrupt
776     * here and elsewhere: Because interrupts are used solely to alert
777     * threads to check termination, which is checked here anyway, we
778     * clear status (using Thread.interrupted) before any call to
779     * park, so that park does not immediately return due to status
780     * being set via some other unrelated call to interrupt in user
781     * code.
782     *
783     * @param w the calling worker
784     * @param c the ctl value on entry
785     * @return true if waited or another thread was released upon enq
786     */
787    private boolean tryAwaitWork(ForkJoinWorkerThread w, long c) {
788        int v = w.eventCount;
789        w.nextWait = (int)c;                      // w's successor record
790        long nc = (long)(v & E_MASK) | ((c - AC_UNIT) & (AC_MASK|TC_MASK));
791        if (ctl != c || !UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
792            long d = ctl; // return true if lost to a deq, to force scan
793            return (int)d != (int)c && (d & AC_MASK) >= (c & AC_MASK);
794        }
795        for (int sc = w.stealCount; sc != 0;) {   // accumulate stealCount
796            long s = stealCount;
797            if (UNSAFE.compareAndSwapLong(this, stealCountOffset, s, s + sc))
798                sc = w.stealCount = 0;
799            else if (w.eventCount != v)
800                return true;                      // update next time
801        }
802        if ((!shutdown || !tryTerminate(false)) &&
803            (int)c != 0 && parallelism + (int)(nc >> AC_SHIFT) == 0 &&
804            blockedCount == 0 && quiescerCount == 0)
805            idleAwaitWork(w, nc, c, v);           // quiescent
806        for (boolean rescanned = false;;) {
807            if (w.eventCount != v)
808                return true;
809            if (!rescanned) {
810                int g = scanGuard, m = g & SMASK;
811                ForkJoinWorkerThread[] ws = workers;
812                if (ws != null && m < ws.length) {
813                    rescanned = true;
814                    for (int i = 0; i <= m; ++i) {
815                        ForkJoinWorkerThread u = ws[i];
816                        if (u != null) {
817                            if (u.queueBase != u.queueTop &&
818                                !tryReleaseWaiter())
819                                rescanned = false; // contended
820                            if (w.eventCount != v)
821                                return true;
822                        }
823                    }
824                }
825                if (scanGuard != g ||              // stale
826                    (queueBase != queueTop && !tryReleaseWaiter()))
827                    rescanned = false;
828                if (!rescanned)
829                    Thread.yield();                // reduce contention
830                else
831                    Thread.interrupted();          // clear before park
832            }
833            else {
834                w.parked = true;                   // must recheck
835                if (w.eventCount != v) {
836                    w.parked = false;
837                    return true;
838                }
839                LockSupport.park(this);
840                rescanned = w.parked = false;
841            }
842        }
843    }
844
845    /**
846     * If inactivating worker w has caused pool to become
847     * quiescent, check for pool termination, and wait for event
848     * for up to SHRINK_RATE nanosecs (rescans are unnecessary in
849     * this case because quiescence reflects consensus about lack
850     * of work). On timeout, if ctl has not changed, terminate the
851     * worker. Upon its termination (see deregisterWorker), it may
852     * wake up another worker to possibly repeat this process.
853     *
854     * @param w the calling worker
855     * @param currentCtl the ctl value after enqueuing w
856     * @param prevCtl the ctl value if w terminated
857     * @param v the eventCount w awaits change
858     */
859    private void idleAwaitWork(ForkJoinWorkerThread w, long currentCtl,
860                               long prevCtl, int v) {
861        if (w.eventCount == v) {
862            if (shutdown)
863                tryTerminate(false);
864            ForkJoinTask.helpExpungeStaleExceptions(); // help clean weak refs
865            while (ctl == currentCtl) {
866                long startTime = System.nanoTime();
867                w.parked = true;
868                if (w.eventCount == v)             // must recheck
869                    LockSupport.parkNanos(this, SHRINK_RATE);
870                w.parked = false;
871                if (w.eventCount != v)
872                    break;
873                else if (System.nanoTime() - startTime <
874                         SHRINK_RATE - (SHRINK_RATE / 10)) // timing slop
875                    Thread.interrupted();          // spurious wakeup
876                else if (UNSAFE.compareAndSwapLong(this, ctlOffset,
877                                                   currentCtl, prevCtl)) {
878                    w.terminate = true;            // restore previous
879                    w.eventCount = ((int)currentCtl + EC_UNIT) & E_MASK;
880                    break;
881                }
882            }
883        }
884    }
885
886    // Submissions
887
888    /**
889     * Enqueues the given task in the submissionQueue.  Same idea as
890     * ForkJoinWorkerThread.pushTask except for use of submissionLock.
891     *
892     * @param t the task
893     */
894    private void addSubmission(ForkJoinTask<?> t) {
895        final ReentrantLock lock = this.submissionLock;
896        lock.lock();
897        try {
898            ForkJoinTask<?>[] q; int s, m;
899            if ((q = submissionQueue) != null) {    // ignore if queue removed
900                long u = (((s = queueTop) & (m = q.length-1)) << ASHIFT)+ABASE;
901                UNSAFE.putOrderedObject(q, u, t);
902                queueTop = s + 1;
903                if (s - queueBase == m)
904                    growSubmissionQueue();
905            }
906        } finally {
907            lock.unlock();
908        }
909        signalWork();
910    }
911
912    //  (pollSubmission is defined below with exported methods)
913
914    /**
915     * Creates or doubles submissionQueue array.
916     * Basically identical to ForkJoinWorkerThread version.
917     */
918    private void growSubmissionQueue() {
919        ForkJoinTask<?>[] oldQ = submissionQueue;
920        int size = oldQ != null ? oldQ.length << 1 : INITIAL_QUEUE_CAPACITY;
921        if (size > MAXIMUM_QUEUE_CAPACITY)
922            throw new RejectedExecutionException("Queue capacity exceeded");
923        if (size < INITIAL_QUEUE_CAPACITY)
924            size = INITIAL_QUEUE_CAPACITY;
925        ForkJoinTask<?>[] q = submissionQueue = new ForkJoinTask<?>[size];
926        int mask = size - 1;
927        int top = queueTop;
928        int oldMask;
929        if (oldQ != null && (oldMask = oldQ.length - 1) >= 0) {
930            for (int b = queueBase; b != top; ++b) {
931                long u = ((b & oldMask) << ASHIFT) + ABASE;
932                Object x = UNSAFE.getObjectVolatile(oldQ, u);
933                if (x != null && UNSAFE.compareAndSwapObject(oldQ, u, x, null))
934                    UNSAFE.putObjectVolatile
935                        (q, ((b & mask) << ASHIFT) + ABASE, x);
936            }
937        }
938    }
939
940    // Blocking support
941
942    /**
943     * Tries to increment blockedCount, decrement active count
944     * (sometimes implicitly) and possibly release or create a
945     * compensating worker in preparation for blocking. Fails
946     * on contention or termination.
947     *
948     * @return true if the caller can block, else should recheck and retry
949     */
950    private boolean tryPreBlock() {
951        int b = blockedCount;
952        if (UNSAFE.compareAndSwapInt(this, blockedCountOffset, b, b + 1)) {
953            int pc = parallelism;
954            do {
955                ForkJoinWorkerThread[] ws; ForkJoinWorkerThread w;
956                int e, ac, tc, i;
957                long c = ctl;
958                int u = (int)(c >>> 32);
959                if ((e = (int)c) < 0) {
960                                                 // skip -- terminating
961                }
962                else if ((ac = (u >> UAC_SHIFT)) <= 0 && e != 0 &&
963                         (ws = workers) != null &&
964                         (i = ~e & SMASK) < ws.length &&
965                         (w = ws[i]) != null) {
966                    long nc = ((long)(w.nextWait & E_MASK) |
967                               (c & (AC_MASK|TC_MASK)));
968                    if (w.eventCount == e &&
969                        UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
970                        w.eventCount = (e + EC_UNIT) & E_MASK;
971                        if (w.parked)
972                            UNSAFE.unpark(w);
973                        return true;             // release an idle worker
974                    }
975                }
976                else if ((tc = (short)(u >>> UTC_SHIFT)) >= 0 && ac + pc > 1) {
977                    long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK);
978                    if (UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc))
979                        return true;             // no compensation needed
980                }
981                else if (tc + pc < MAX_ID) {
982                    long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK);
983                    if (UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
984                        addWorker();
985                        return true;            // create a replacement
986                    }
987                }
988                // try to back out on any failure and let caller retry
989            } while (!UNSAFE.compareAndSwapInt(this, blockedCountOffset,
990                                               b = blockedCount, b - 1));
991        }
992        return false;
993    }
994
995    /**
996     * Decrements blockedCount and increments active count.
997     */
998    private void postBlock() {
999        long c;
1000        do {} while (!UNSAFE.compareAndSwapLong(this, ctlOffset,  // no mask
1001                                                c = ctl, c + AC_UNIT));
1002        int b;
1003        do {} while (!UNSAFE.compareAndSwapInt(this, blockedCountOffset,
1004                                               b = blockedCount, b - 1));
1005    }
1006
1007    /**
1008     * Possibly blocks waiting for the given task to complete, or
1009     * cancels the task if terminating.  Fails to wait if contended.
1010     *
1011     * @param joinMe the task
1012     */
1013    final void tryAwaitJoin(ForkJoinTask<?> joinMe) {
1014        Thread.interrupted(); // clear interrupts before checking termination
1015        if (joinMe.status >= 0) {
1016            if (tryPreBlock()) {
1017                joinMe.tryAwaitDone(0L);
1018                postBlock();
1019            }
1020            else if ((ctl & STOP_BIT) != 0L)
1021                joinMe.cancelIgnoringExceptions();
1022        }
1023    }
1024
1025    /**
1026     * Possibly blocks the given worker waiting for joinMe to
1027     * complete or timeout.
1028     *
1029     * @param joinMe the task
1030     * @param nanos the wait time for underlying Object.wait
1031     */
1032    final void timedAwaitJoin(ForkJoinTask<?> joinMe, long nanos) {
1033        while (joinMe.status >= 0) {
1034            Thread.interrupted();
1035            if ((ctl & STOP_BIT) != 0L) {
1036                joinMe.cancelIgnoringExceptions();
1037                break;
1038            }
1039            if (tryPreBlock()) {
1040                long last = System.nanoTime();
1041                while (joinMe.status >= 0) {
1042                    long millis = TimeUnit.NANOSECONDS.toMillis(nanos);
1043                    if (millis <= 0)
1044                        break;
1045                    joinMe.tryAwaitDone(millis);
1046                    if (joinMe.status < 0)
1047                        break;
1048                    if ((ctl & STOP_BIT) != 0L) {
1049                        joinMe.cancelIgnoringExceptions();
1050                        break;
1051                    }
1052                    long now = System.nanoTime();
1053                    nanos -= now - last;
1054                    last = now;
1055                }
1056                postBlock();
1057                break;
1058            }
1059        }
1060    }
1061
1062    /**
1063     * If necessary, compensates for blocker, and blocks.
1064     */
1065    private void awaitBlocker(ManagedBlocker blocker)
1066        throws InterruptedException {
1067        while (!blocker.isReleasable()) {
1068            if (tryPreBlock()) {
1069                try {
1070                    do {} while (!blocker.isReleasable() && !blocker.block());
1071                } finally {
1072                    postBlock();
1073                }
1074                break;
1075            }
1076        }
1077    }
1078
1079    // Creating, registering and deregistring workers
1080
1081    /**
1082     * Tries to create and start a worker; minimally rolls back counts
1083     * on failure.
1084     */
1085    private void addWorker() {
1086        Throwable ex = null;
1087        ForkJoinWorkerThread t = null;
1088        try {
1089            t = factory.newThread(this);
1090        } catch (Throwable e) {
1091            ex = e;
1092        }
1093        if (t == null) {  // null or exceptional factory return
1094            long c;       // adjust counts
1095            do {} while (!UNSAFE.compareAndSwapLong
1096                         (this, ctlOffset, c = ctl,
1097                          (((c - AC_UNIT) & AC_MASK) |
1098                           ((c - TC_UNIT) & TC_MASK) |
1099                           (c & ~(AC_MASK|TC_MASK)))));
1100            // Propagate exception if originating from an external caller
1101            if (!tryTerminate(false) && ex != null &&
1102                !(Thread.currentThread() instanceof ForkJoinWorkerThread))
1103                SneakyThrow.sneakyThrow(ex); // android-changed
1104        }
1105        else
1106            t.start();
1107    }
1108
1109    /**
1110     * Callback from ForkJoinWorkerThread constructor to assign a
1111     * public name
1112     */
1113    final String nextWorkerName() {
1114        for (int n;;) {
1115            if (UNSAFE.compareAndSwapInt(this, nextWorkerNumberOffset,
1116                                         n = nextWorkerNumber, ++n))
1117                return workerNamePrefix + n;
1118        }
1119    }
1120
1121    /**
1122     * Callback from ForkJoinWorkerThread constructor to
1123     * determine its poolIndex and record in workers array.
1124     *
1125     * @param w the worker
1126     * @return the worker's pool index
1127     */
1128    final int registerWorker(ForkJoinWorkerThread w) {
1129        /*
1130         * In the typical case, a new worker acquires the lock, uses
1131         * next available index and returns quickly.  Since we should
1132         * not block callers (ultimately from signalWork or
1133         * tryPreBlock) waiting for the lock needed to do this, we
1134         * instead help release other workers while waiting for the
1135         * lock.
1136         */
1137        for (int g;;) {
1138            ForkJoinWorkerThread[] ws;
1139            if (((g = scanGuard) & SG_UNIT) == 0 &&
1140                UNSAFE.compareAndSwapInt(this, scanGuardOffset,
1141                                         g, g | SG_UNIT)) {
1142                int k = nextWorkerIndex;
1143                try {
1144                    if ((ws = workers) != null) { // ignore on shutdown
1145                        int n = ws.length;
1146                        if (k < 0 || k >= n || ws[k] != null) {
1147                            for (k = 0; k < n && ws[k] != null; ++k)
1148                                ;
1149                            if (k == n)
1150                                ws = workers = Arrays.copyOf(ws, n << 1);
1151                        }
1152                        ws[k] = w;
1153                        nextWorkerIndex = k + 1;
1154                        int m = g & SMASK;
1155                        g = (k > m) ? ((m << 1) + 1) & SMASK : g + (SG_UNIT<<1);
1156                    }
1157                } finally {
1158                    scanGuard = g;
1159                }
1160                return k;
1161            }
1162            else if ((ws = workers) != null) { // help release others
1163                for (ForkJoinWorkerThread u : ws) {
1164                    if (u != null && u.queueBase != u.queueTop) {
1165                        if (tryReleaseWaiter())
1166                            break;
1167                    }
1168                }
1169            }
1170        }
1171    }
1172
1173    /**
1174     * Final callback from terminating worker.  Removes record of
1175     * worker from array, and adjusts counts. If pool is shutting
1176     * down, tries to complete termination.
1177     *
1178     * @param w the worker
1179     */
1180    final void deregisterWorker(ForkJoinWorkerThread w, Throwable ex) {
1181        int idx = w.poolIndex;
1182        int sc = w.stealCount;
1183        int steps = 0;
1184        // Remove from array, adjust worker counts and collect steal count.
1185        // We can intermix failed removes or adjusts with steal updates
1186        do {
1187            long s, c;
1188            int g;
1189            if (steps == 0 && ((g = scanGuard) & SG_UNIT) == 0 &&
1190                UNSAFE.compareAndSwapInt(this, scanGuardOffset,
1191                                         g, g |= SG_UNIT)) {
1192                ForkJoinWorkerThread[] ws = workers;
1193                if (ws != null && idx >= 0 &&
1194                    idx < ws.length && ws[idx] == w)
1195                    ws[idx] = null;    // verify
1196                nextWorkerIndex = idx;
1197                scanGuard = g + SG_UNIT;
1198                steps = 1;
1199            }
1200            if (steps == 1 &&
1201                UNSAFE.compareAndSwapLong(this, ctlOffset, c = ctl,
1202                                          (((c - AC_UNIT) & AC_MASK) |
1203                                           ((c - TC_UNIT) & TC_MASK) |
1204                                           (c & ~(AC_MASK|TC_MASK)))))
1205                steps = 2;
1206            if (sc != 0 &&
1207                UNSAFE.compareAndSwapLong(this, stealCountOffset,
1208                                          s = stealCount, s + sc))
1209                sc = 0;
1210        } while (steps != 2 || sc != 0);
1211        if (!tryTerminate(false)) {
1212            if (ex != null)   // possibly replace if died abnormally
1213                signalWork();
1214            else
1215                tryReleaseWaiter();
1216        }
1217    }
1218
1219    // Shutdown and termination
1220
1221    /**
1222     * Possibly initiates and/or completes termination.
1223     *
1224     * @param now if true, unconditionally terminate, else only
1225     * if shutdown and empty queue and no active workers
1226     * @return true if now terminating or terminated
1227     */
1228    private boolean tryTerminate(boolean now) {
1229        long c;
1230        while (((c = ctl) & STOP_BIT) == 0) {
1231            if (!now) {
1232                if ((int)(c >> AC_SHIFT) != -parallelism)
1233                    return false;
1234                if (!shutdown || blockedCount != 0 || quiescerCount != 0 ||
1235                    queueBase != queueTop) {
1236                    if (ctl == c) // staleness check
1237                        return false;
1238                    continue;
1239                }
1240            }
1241            if (UNSAFE.compareAndSwapLong(this, ctlOffset, c, c | STOP_BIT))
1242                startTerminating();
1243        }
1244        if ((short)(c >>> TC_SHIFT) == -parallelism) { // signal when 0 workers
1245            final ReentrantLock lock = this.submissionLock;
1246            lock.lock();
1247            try {
1248                termination.signalAll();
1249            } finally {
1250                lock.unlock();
1251            }
1252        }
1253        return true;
1254    }
1255
1256    /**
1257     * Runs up to three passes through workers: (0) Setting
1258     * termination status for each worker, followed by wakeups up to
1259     * queued workers; (1) helping cancel tasks; (2) interrupting
1260     * lagging threads (likely in external tasks, but possibly also
1261     * blocked in joins).  Each pass repeats previous steps because of
1262     * potential lagging thread creation.
1263     */
1264    private void startTerminating() {
1265        cancelSubmissions();
1266        for (int pass = 0; pass < 3; ++pass) {
1267            ForkJoinWorkerThread[] ws = workers;
1268            if (ws != null) {
1269                for (ForkJoinWorkerThread w : ws) {
1270                    if (w != null) {
1271                        w.terminate = true;
1272                        if (pass > 0) {
1273                            w.cancelTasks();
1274                            if (pass > 1 && !w.isInterrupted()) {
1275                                try {
1276                                    w.interrupt();
1277                                } catch (SecurityException ignore) {
1278                                }
1279                            }
1280                        }
1281                    }
1282                }
1283                terminateWaiters();
1284            }
1285        }
1286    }
1287
1288    /**
1289     * Polls and cancels all submissions. Called only during termination.
1290     */
1291    private void cancelSubmissions() {
1292        while (queueBase != queueTop) {
1293            ForkJoinTask<?> task = pollSubmission();
1294            if (task != null) {
1295                try {
1296                    task.cancel(false);
1297                } catch (Throwable ignore) {
1298                }
1299            }
1300        }
1301    }
1302
1303    /**
1304     * Tries to set the termination status of waiting workers, and
1305     * then wakes them up (after which they will terminate).
1306     */
1307    private void terminateWaiters() {
1308        ForkJoinWorkerThread[] ws = workers;
1309        if (ws != null) {
1310            ForkJoinWorkerThread w; long c; int i, e;
1311            int n = ws.length;
1312            while ((i = ~(e = (int)(c = ctl)) & SMASK) < n &&
1313                   (w = ws[i]) != null && w.eventCount == (e & E_MASK)) {
1314                if (UNSAFE.compareAndSwapLong(this, ctlOffset, c,
1315                                              (long)(w.nextWait & E_MASK) |
1316                                              ((c + AC_UNIT) & AC_MASK) |
1317                                              (c & (TC_MASK|STOP_BIT)))) {
1318                    w.terminate = true;
1319                    w.eventCount = e + EC_UNIT;
1320                    if (w.parked)
1321                        UNSAFE.unpark(w);
1322                }
1323            }
1324        }
1325    }
1326
1327    // misc ForkJoinWorkerThread support
1328
1329    /**
1330     * Increments or decrements quiescerCount. Needed only to prevent
1331     * triggering shutdown if a worker is transiently inactive while
1332     * checking quiescence.
1333     *
1334     * @param delta 1 for increment, -1 for decrement
1335     */
1336    final void addQuiescerCount(int delta) {
1337        int c;
1338        do {} while (!UNSAFE.compareAndSwapInt(this, quiescerCountOffset,
1339                                               c = quiescerCount, c + delta));
1340    }
1341
1342    /**
1343     * Directly increments or decrements active count without queuing.
1344     * This method is used to transiently assert inactivation while
1345     * checking quiescence.
1346     *
1347     * @param delta 1 for increment, -1 for decrement
1348     */
1349    final void addActiveCount(int delta) {
1350        long d = (long)delta << AC_SHIFT;
1351        long c;
1352        do {} while (!UNSAFE.compareAndSwapLong(this, ctlOffset,
1353                                                c = ctl, c + d));
1354    }
1355
1356    /**
1357     * Returns the approximate (non-atomic) number of idle threads per
1358     * active thread.
1359     */
1360    final int idlePerActive() {
1361        // Approximate at powers of two for small values, saturate past 4
1362        int p = parallelism;
1363        int a = p + (int)(ctl >> AC_SHIFT);
1364        return (a > (p >>>= 1) ? 0 :
1365                a > (p >>>= 1) ? 1 :
1366                a > (p >>>= 1) ? 2 :
1367                a > (p >>>= 1) ? 4 :
1368                8);
1369    }
1370
1371    // Exported methods
1372
1373    // Constructors
1374
1375    /**
1376     * Creates a {@code ForkJoinPool} with parallelism equal to {@link
1377     * java.lang.Runtime#availableProcessors}, using the {@linkplain
1378     * #defaultForkJoinWorkerThreadFactory default thread factory},
1379     * no UncaughtExceptionHandler, and non-async LIFO processing mode.
1380     */
1381    public ForkJoinPool() {
1382        this(Runtime.getRuntime().availableProcessors(),
1383             defaultForkJoinWorkerThreadFactory, null, false);
1384    }
1385
1386    /**
1387     * Creates a {@code ForkJoinPool} with the indicated parallelism
1388     * level, the {@linkplain
1389     * #defaultForkJoinWorkerThreadFactory default thread factory},
1390     * no UncaughtExceptionHandler, and non-async LIFO processing mode.
1391     *
1392     * @param parallelism the parallelism level
1393     * @throws IllegalArgumentException if parallelism less than or
1394     *         equal to zero, or greater than implementation limit
1395     */
1396    public ForkJoinPool(int parallelism) {
1397        this(parallelism, defaultForkJoinWorkerThreadFactory, null, false);
1398    }
1399
1400    /**
1401     * Creates a {@code ForkJoinPool} with the given parameters.
1402     *
1403     * @param parallelism the parallelism level. For default value,
1404     * use {@link java.lang.Runtime#availableProcessors}.
1405     * @param factory the factory for creating new threads. For default value,
1406     * use {@link #defaultForkJoinWorkerThreadFactory}.
1407     * @param handler the handler for internal worker threads that
1408     * terminate due to unrecoverable errors encountered while executing
1409     * tasks. For default value, use {@code null}.
1410     * @param asyncMode if true,
1411     * establishes local first-in-first-out scheduling mode for forked
1412     * tasks that are never joined. This mode may be more appropriate
1413     * than default locally stack-based mode in applications in which
1414     * worker threads only process event-style asynchronous tasks.
1415     * For default value, use {@code false}.
1416     * @throws IllegalArgumentException if parallelism less than or
1417     *         equal to zero, or greater than implementation limit
1418     * @throws NullPointerException if the factory is null
1419     */
1420    public ForkJoinPool(int parallelism,
1421                        ForkJoinWorkerThreadFactory factory,
1422                        Thread.UncaughtExceptionHandler handler,
1423                        boolean asyncMode) {
1424        checkPermission();
1425        if (factory == null)
1426            throw new NullPointerException();
1427        if (parallelism <= 0 || parallelism > MAX_ID)
1428            throw new IllegalArgumentException();
1429        this.parallelism = parallelism;
1430        this.factory = factory;
1431        this.ueh = handler;
1432        this.locallyFifo = asyncMode;
1433        long np = (long)(-parallelism); // offset ctl counts
1434        this.ctl = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK);
1435        this.submissionQueue = new ForkJoinTask<?>[INITIAL_QUEUE_CAPACITY];
1436        // initialize workers array with room for 2*parallelism if possible
1437        int n = parallelism << 1;
1438        if (n >= MAX_ID)
1439            n = MAX_ID;
1440        else { // See Hackers Delight, sec 3.2, where n < (1 << 16)
1441            n |= n >>> 1; n |= n >>> 2; n |= n >>> 4; n |= n >>> 8;
1442        }
1443        workers = new ForkJoinWorkerThread[n + 1];
1444        this.submissionLock = new ReentrantLock();
1445        this.termination = submissionLock.newCondition();
1446        StringBuilder sb = new StringBuilder("ForkJoinPool-");
1447        sb.append(poolNumberGenerator.incrementAndGet());
1448        sb.append("-worker-");
1449        this.workerNamePrefix = sb.toString();
1450    }
1451
1452    // Execution methods
1453
1454    /**
1455     * Performs the given task, returning its result upon completion.
1456     * If the computation encounters an unchecked Exception or Error,
1457     * it is rethrown as the outcome of this invocation.  Rethrown
1458     * exceptions behave in the same way as regular exceptions, but,
1459     * when possible, contain stack traces (as displayed for example
1460     * using {@code ex.printStackTrace()}) of both the current thread
1461     * as well as the thread actually encountering the exception;
1462     * minimally only the latter.
1463     *
1464     * @param task the task
1465     * @return the task's result
1466     * @throws NullPointerException if the task is null
1467     * @throws RejectedExecutionException if the task cannot be
1468     *         scheduled for execution
1469     */
1470    public <T> T invoke(ForkJoinTask<T> task) {
1471        Thread t = Thread.currentThread();
1472        if (task == null)
1473            throw new NullPointerException();
1474        if (shutdown)
1475            throw new RejectedExecutionException();
1476        if ((t instanceof ForkJoinWorkerThread) &&
1477            ((ForkJoinWorkerThread)t).pool == this)
1478            return task.invoke();  // bypass submit if in same pool
1479        else {
1480            addSubmission(task);
1481            return task.join();
1482        }
1483    }
1484
1485    /**
1486     * Unless terminating, forks task if within an ongoing FJ
1487     * computation in the current pool, else submits as external task.
1488     */
1489    private <T> void forkOrSubmit(ForkJoinTask<T> task) {
1490        ForkJoinWorkerThread w;
1491        Thread t = Thread.currentThread();
1492        if (shutdown)
1493            throw new RejectedExecutionException();
1494        if ((t instanceof ForkJoinWorkerThread) &&
1495            (w = (ForkJoinWorkerThread)t).pool == this)
1496            w.pushTask(task);
1497        else
1498            addSubmission(task);
1499    }
1500
1501    /**
1502     * Arranges for (asynchronous) execution of the given task.
1503     *
1504     * @param task the task
1505     * @throws NullPointerException if the task is null
1506     * @throws RejectedExecutionException if the task cannot be
1507     *         scheduled for execution
1508     */
1509    public void execute(ForkJoinTask<?> task) {
1510        if (task == null)
1511            throw new NullPointerException();
1512        forkOrSubmit(task);
1513    }
1514
1515    // AbstractExecutorService methods
1516
1517    /**
1518     * @throws NullPointerException if the task is null
1519     * @throws RejectedExecutionException if the task cannot be
1520     *         scheduled for execution
1521     */
1522    public void execute(Runnable task) {
1523        if (task == null)
1524            throw new NullPointerException();
1525        ForkJoinTask<?> job;
1526        if (task instanceof ForkJoinTask<?>) // avoid re-wrap
1527            job = (ForkJoinTask<?>) task;
1528        else
1529            job = ForkJoinTask.adapt(task, null);
1530        forkOrSubmit(job);
1531    }
1532
1533    /**
1534     * Submits a ForkJoinTask for execution.
1535     *
1536     * @param task the task to submit
1537     * @return the task
1538     * @throws NullPointerException if the task is null
1539     * @throws RejectedExecutionException if the task cannot be
1540     *         scheduled for execution
1541     */
1542    public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) {
1543        if (task == null)
1544            throw new NullPointerException();
1545        forkOrSubmit(task);
1546        return task;
1547    }
1548
1549    /**
1550     * @throws NullPointerException if the task is null
1551     * @throws RejectedExecutionException if the task cannot be
1552     *         scheduled for execution
1553     */
1554    public <T> ForkJoinTask<T> submit(Callable<T> task) {
1555        if (task == null)
1556            throw new NullPointerException();
1557        ForkJoinTask<T> job = ForkJoinTask.adapt(task);
1558        forkOrSubmit(job);
1559        return job;
1560    }
1561
1562    /**
1563     * @throws NullPointerException if the task is null
1564     * @throws RejectedExecutionException if the task cannot be
1565     *         scheduled for execution
1566     */
1567    public <T> ForkJoinTask<T> submit(Runnable task, T result) {
1568        if (task == null)
1569            throw new NullPointerException();
1570        ForkJoinTask<T> job = ForkJoinTask.adapt(task, result);
1571        forkOrSubmit(job);
1572        return job;
1573    }
1574
1575    /**
1576     * @throws NullPointerException if the task is null
1577     * @throws RejectedExecutionException if the task cannot be
1578     *         scheduled for execution
1579     */
1580    public ForkJoinTask<?> submit(Runnable task) {
1581        if (task == null)
1582            throw new NullPointerException();
1583        ForkJoinTask<?> job;
1584        if (task instanceof ForkJoinTask<?>) // avoid re-wrap
1585            job = (ForkJoinTask<?>) task;
1586        else
1587            job = ForkJoinTask.adapt(task, null);
1588        forkOrSubmit(job);
1589        return job;
1590    }
1591
1592    /**
1593     * @throws NullPointerException       {@inheritDoc}
1594     * @throws RejectedExecutionException {@inheritDoc}
1595     */
1596    public <T> List<Future<T>> invokeAll(Collection<? extends Callable<T>> tasks) {
1597        ArrayList<ForkJoinTask<T>> forkJoinTasks =
1598            new ArrayList<ForkJoinTask<T>>(tasks.size());
1599        for (Callable<T> task : tasks)
1600            forkJoinTasks.add(ForkJoinTask.adapt(task));
1601        invoke(new InvokeAll<T>(forkJoinTasks));
1602
1603        @SuppressWarnings({"unchecked", "rawtypes"})
1604            List<Future<T>> futures = (List<Future<T>>) (List) forkJoinTasks;
1605        return futures;
1606    }
1607
1608    static final class InvokeAll<T> extends RecursiveAction {
1609        final ArrayList<ForkJoinTask<T>> tasks;
1610        InvokeAll(ArrayList<ForkJoinTask<T>> tasks) { this.tasks = tasks; }
1611        public void compute() {
1612            try { invokeAll(tasks); }
1613            catch (Exception ignore) {}
1614        }
1615        private static final long serialVersionUID = -7914297376763021607L;
1616    }
1617
1618    /**
1619     * Returns the factory used for constructing new workers.
1620     *
1621     * @return the factory used for constructing new workers
1622     */
1623    public ForkJoinWorkerThreadFactory getFactory() {
1624        return factory;
1625    }
1626
1627    /**
1628     * Returns the handler for internal worker threads that terminate
1629     * due to unrecoverable errors encountered while executing tasks.
1630     *
1631     * @return the handler, or {@code null} if none
1632     */
1633    public Thread.UncaughtExceptionHandler getUncaughtExceptionHandler() {
1634        return ueh;
1635    }
1636
1637    /**
1638     * Returns the targeted parallelism level of this pool.
1639     *
1640     * @return the targeted parallelism level of this pool
1641     */
1642    public int getParallelism() {
1643        return parallelism;
1644    }
1645
1646    /**
1647     * Returns the number of worker threads that have started but not
1648     * yet terminated.  The result returned by this method may differ
1649     * from {@link #getParallelism} when threads are created to
1650     * maintain parallelism when others are cooperatively blocked.
1651     *
1652     * @return the number of worker threads
1653     */
1654    public int getPoolSize() {
1655        return parallelism + (short)(ctl >>> TC_SHIFT);
1656    }
1657
1658    /**
1659     * Returns {@code true} if this pool uses local first-in-first-out
1660     * scheduling mode for forked tasks that are never joined.
1661     *
1662     * @return {@code true} if this pool uses async mode
1663     */
1664    public boolean getAsyncMode() {
1665        return locallyFifo;
1666    }
1667
1668    /**
1669     * Returns an estimate of the number of worker threads that are
1670     * not blocked waiting to join tasks or for other managed
1671     * synchronization. This method may overestimate the
1672     * number of running threads.
1673     *
1674     * @return the number of worker threads
1675     */
1676    public int getRunningThreadCount() {
1677        int r = parallelism + (int)(ctl >> AC_SHIFT);
1678        return (r <= 0) ? 0 : r; // suppress momentarily negative values
1679    }
1680
1681    /**
1682     * Returns an estimate of the number of threads that are currently
1683     * stealing or executing tasks. This method may overestimate the
1684     * number of active threads.
1685     *
1686     * @return the number of active threads
1687     */
1688    public int getActiveThreadCount() {
1689        int r = parallelism + (int)(ctl >> AC_SHIFT) + blockedCount;
1690        return (r <= 0) ? 0 : r; // suppress momentarily negative values
1691    }
1692
1693    /**
1694     * Returns {@code true} if all worker threads are currently idle.
1695     * An idle worker is one that cannot obtain a task to execute
1696     * because none are available to steal from other threads, and
1697     * there are no pending submissions to the pool. This method is
1698     * conservative; it might not return {@code true} immediately upon
1699     * idleness of all threads, but will eventually become true if
1700     * threads remain inactive.
1701     *
1702     * @return {@code true} if all threads are currently idle
1703     */
1704    public boolean isQuiescent() {
1705        return parallelism + (int)(ctl >> AC_SHIFT) + blockedCount == 0;
1706    }
1707
1708    /**
1709     * Returns an estimate of the total number of tasks stolen from
1710     * one thread's work queue by another. The reported value
1711     * underestimates the actual total number of steals when the pool
1712     * is not quiescent. This value may be useful for monitoring and
1713     * tuning fork/join programs: in general, steal counts should be
1714     * high enough to keep threads busy, but low enough to avoid
1715     * overhead and contention across threads.
1716     *
1717     * @return the number of steals
1718     */
1719    public long getStealCount() {
1720        return stealCount;
1721    }
1722
1723    /**
1724     * Returns an estimate of the total number of tasks currently held
1725     * in queues by worker threads (but not including tasks submitted
1726     * to the pool that have not begun executing). This value is only
1727     * an approximation, obtained by iterating across all threads in
1728     * the pool. This method may be useful for tuning task
1729     * granularities.
1730     *
1731     * @return the number of queued tasks
1732     */
1733    public long getQueuedTaskCount() {
1734        long count = 0;
1735        ForkJoinWorkerThread[] ws;
1736        if ((short)(ctl >>> TC_SHIFT) > -parallelism &&
1737            (ws = workers) != null) {
1738            for (ForkJoinWorkerThread w : ws)
1739                if (w != null)
1740                    count -= w.queueBase - w.queueTop; // must read base first
1741        }
1742        return count;
1743    }
1744
1745    /**
1746     * Returns an estimate of the number of tasks submitted to this
1747     * pool that have not yet begun executing.  This method may take
1748     * time proportional to the number of submissions.
1749     *
1750     * @return the number of queued submissions
1751     */
1752    public int getQueuedSubmissionCount() {
1753        return -queueBase + queueTop;
1754    }
1755
1756    /**
1757     * Returns {@code true} if there are any tasks submitted to this
1758     * pool that have not yet begun executing.
1759     *
1760     * @return {@code true} if there are any queued submissions
1761     */
1762    public boolean hasQueuedSubmissions() {
1763        return queueBase != queueTop;
1764    }
1765
1766    /**
1767     * Removes and returns the next unexecuted submission if one is
1768     * available.  This method may be useful in extensions to this
1769     * class that re-assign work in systems with multiple pools.
1770     *
1771     * @return the next submission, or {@code null} if none
1772     */
1773    protected ForkJoinTask<?> pollSubmission() {
1774        ForkJoinTask<?> t; ForkJoinTask<?>[] q; int b, i;
1775        while ((b = queueBase) != queueTop &&
1776               (q = submissionQueue) != null &&
1777               (i = (q.length - 1) & b) >= 0) {
1778            long u = (i << ASHIFT) + ABASE;
1779            if ((t = q[i]) != null &&
1780                queueBase == b &&
1781                UNSAFE.compareAndSwapObject(q, u, t, null)) {
1782                queueBase = b + 1;
1783                return t;
1784            }
1785        }
1786        return null;
1787    }
1788
1789    /**
1790     * Removes all available unexecuted submitted and forked tasks
1791     * from scheduling queues and adds them to the given collection,
1792     * without altering their execution status. These may include
1793     * artificially generated or wrapped tasks. This method is
1794     * designed to be invoked only when the pool is known to be
1795     * quiescent. Invocations at other times may not remove all
1796     * tasks. A failure encountered while attempting to add elements
1797     * to collection {@code c} may result in elements being in
1798     * neither, either or both collections when the associated
1799     * exception is thrown.  The behavior of this operation is
1800     * undefined if the specified collection is modified while the
1801     * operation is in progress.
1802     *
1803     * @param c the collection to transfer elements into
1804     * @return the number of elements transferred
1805     */
1806    protected int drainTasksTo(Collection<? super ForkJoinTask<?>> c) {
1807        int count = 0;
1808        while (queueBase != queueTop) {
1809            ForkJoinTask<?> t = pollSubmission();
1810            if (t != null) {
1811                c.add(t);
1812                ++count;
1813            }
1814        }
1815        ForkJoinWorkerThread[] ws;
1816        if ((short)(ctl >>> TC_SHIFT) > -parallelism &&
1817            (ws = workers) != null) {
1818            for (ForkJoinWorkerThread w : ws)
1819                if (w != null)
1820                    count += w.drainTasksTo(c);
1821        }
1822        return count;
1823    }
1824
1825    /**
1826     * Returns a string identifying this pool, as well as its state,
1827     * including indications of run state, parallelism level, and
1828     * worker and task counts.
1829     *
1830     * @return a string identifying this pool, as well as its state
1831     */
1832    public String toString() {
1833        long st = getStealCount();
1834        long qt = getQueuedTaskCount();
1835        long qs = getQueuedSubmissionCount();
1836        int pc = parallelism;
1837        long c = ctl;
1838        int tc = pc + (short)(c >>> TC_SHIFT);
1839        int rc = pc + (int)(c >> AC_SHIFT);
1840        if (rc < 0) // ignore transient negative
1841            rc = 0;
1842        int ac = rc + blockedCount;
1843        String level;
1844        if ((c & STOP_BIT) != 0)
1845            level = (tc == 0) ? "Terminated" : "Terminating";
1846        else
1847            level = shutdown ? "Shutting down" : "Running";
1848        return super.toString() +
1849            "[" + level +
1850            ", parallelism = " + pc +
1851            ", size = " + tc +
1852            ", active = " + ac +
1853            ", running = " + rc +
1854            ", steals = " + st +
1855            ", tasks = " + qt +
1856            ", submissions = " + qs +
1857            "]";
1858    }
1859
1860    /**
1861     * Initiates an orderly shutdown in which previously submitted
1862     * tasks are executed, but no new tasks will be accepted.
1863     * Invocation has no additional effect if already shut down.
1864     * Tasks that are in the process of being submitted concurrently
1865     * during the course of this method may or may not be rejected.
1866     */
1867    public void shutdown() {
1868        checkPermission();
1869        shutdown = true;
1870        tryTerminate(false);
1871    }
1872
1873    /**
1874     * Attempts to cancel and/or stop all tasks, and reject all
1875     * subsequently submitted tasks.  Tasks that are in the process of
1876     * being submitted or executed concurrently during the course of
1877     * this method may or may not be rejected. This method cancels
1878     * both existing and unexecuted tasks, in order to permit
1879     * termination in the presence of task dependencies. So the method
1880     * always returns an empty list (unlike the case for some other
1881     * Executors).
1882     *
1883     * @return an empty list
1884     */
1885    public List<Runnable> shutdownNow() {
1886        checkPermission();
1887        shutdown = true;
1888        tryTerminate(true);
1889        return Collections.emptyList();
1890    }
1891
1892    /**
1893     * Returns {@code true} if all tasks have completed following shut down.
1894     *
1895     * @return {@code true} if all tasks have completed following shut down
1896     */
1897    public boolean isTerminated() {
1898        long c = ctl;
1899        return ((c & STOP_BIT) != 0L &&
1900                (short)(c >>> TC_SHIFT) == -parallelism);
1901    }
1902
1903    /**
1904     * Returns {@code true} if the process of termination has
1905     * commenced but not yet completed.  This method may be useful for
1906     * debugging. A return of {@code true} reported a sufficient
1907     * period after shutdown may indicate that submitted tasks have
1908     * ignored or suppressed interruption, or are waiting for IO,
1909     * causing this executor not to properly terminate. (See the
1910     * advisory notes for class {@link ForkJoinTask} stating that
1911     * tasks should not normally entail blocking operations.  But if
1912     * they do, they must abort them on interrupt.)
1913     *
1914     * @return {@code true} if terminating but not yet terminated
1915     */
1916    public boolean isTerminating() {
1917        long c = ctl;
1918        return ((c & STOP_BIT) != 0L &&
1919                (short)(c >>> TC_SHIFT) != -parallelism);
1920    }
1921
1922    /**
1923     * Returns true if terminating or terminated. Used by ForkJoinWorkerThread.
1924     */
1925    final boolean isAtLeastTerminating() {
1926        return (ctl & STOP_BIT) != 0L;
1927    }
1928
1929    /**
1930     * Returns {@code true} if this pool has been shut down.
1931     *
1932     * @return {@code true} if this pool has been shut down
1933     */
1934    public boolean isShutdown() {
1935        return shutdown;
1936    }
1937
1938    /**
1939     * Blocks until all tasks have completed execution after a shutdown
1940     * request, or the timeout occurs, or the current thread is
1941     * interrupted, whichever happens first.
1942     *
1943     * @param timeout the maximum time to wait
1944     * @param unit the time unit of the timeout argument
1945     * @return {@code true} if this executor terminated and
1946     *         {@code false} if the timeout elapsed before termination
1947     * @throws InterruptedException if interrupted while waiting
1948     */
1949    public boolean awaitTermination(long timeout, TimeUnit unit)
1950        throws InterruptedException {
1951        long nanos = unit.toNanos(timeout);
1952        final ReentrantLock lock = this.submissionLock;
1953        lock.lock();
1954        try {
1955            for (;;) {
1956                if (isTerminated())
1957                    return true;
1958                if (nanos <= 0)
1959                    return false;
1960                nanos = termination.awaitNanos(nanos);
1961            }
1962        } finally {
1963            lock.unlock();
1964        }
1965    }
1966
1967    /**
1968     * Interface for extending managed parallelism for tasks running
1969     * in {@link ForkJoinPool}s.
1970     *
1971     * <p>A {@code ManagedBlocker} provides two methods.  Method
1972     * {@code isReleasable} must return {@code true} if blocking is
1973     * not necessary. Method {@code block} blocks the current thread
1974     * if necessary (perhaps internally invoking {@code isReleasable}
1975     * before actually blocking). These actions are performed by any
1976     * thread invoking {@link ForkJoinPool#managedBlock}.  The
1977     * unusual methods in this API accommodate synchronizers that may,
1978     * but don't usually, block for long periods. Similarly, they
1979     * allow more efficient internal handling of cases in which
1980     * additional workers may be, but usually are not, needed to
1981     * ensure sufficient parallelism.  Toward this end,
1982     * implementations of method {@code isReleasable} must be amenable
1983     * to repeated invocation.
1984     *
1985     * <p>For example, here is a ManagedBlocker based on a
1986     * ReentrantLock:
1987     *  <pre> {@code
1988     * class ManagedLocker implements ManagedBlocker {
1989     *   final ReentrantLock lock;
1990     *   boolean hasLock = false;
1991     *   ManagedLocker(ReentrantLock lock) { this.lock = lock; }
1992     *   public boolean block() {
1993     *     if (!hasLock)
1994     *       lock.lock();
1995     *     return true;
1996     *   }
1997     *   public boolean isReleasable() {
1998     *     return hasLock || (hasLock = lock.tryLock());
1999     *   }
2000     * }}</pre>
2001     *
2002     * <p>Here is a class that possibly blocks waiting for an
2003     * item on a given queue:
2004     *  <pre> {@code
2005     * class QueueTaker<E> implements ManagedBlocker {
2006     *   final BlockingQueue<E> queue;
2007     *   volatile E item = null;
2008     *   QueueTaker(BlockingQueue<E> q) { this.queue = q; }
2009     *   public boolean block() throws InterruptedException {
2010     *     if (item == null)
2011     *       item = queue.take();
2012     *     return true;
2013     *   }
2014     *   public boolean isReleasable() {
2015     *     return item != null || (item = queue.poll()) != null;
2016     *   }
2017     *   public E getItem() { // call after pool.managedBlock completes
2018     *     return item;
2019     *   }
2020     * }}</pre>
2021     */
2022    public static interface ManagedBlocker {
2023        /**
2024         * Possibly blocks the current thread, for example waiting for
2025         * a lock or condition.
2026         *
2027         * @return {@code true} if no additional blocking is necessary
2028         * (i.e., if isReleasable would return true)
2029         * @throws InterruptedException if interrupted while waiting
2030         * (the method is not required to do so, but is allowed to)
2031         */
2032        boolean block() throws InterruptedException;
2033
2034        /**
2035         * Returns {@code true} if blocking is unnecessary.
2036         */
2037        boolean isReleasable();
2038    }
2039
2040    /**
2041     * Blocks in accord with the given blocker.  If the current thread
2042     * is a {@link ForkJoinWorkerThread}, this method possibly
2043     * arranges for a spare thread to be activated if necessary to
2044     * ensure sufficient parallelism while the current thread is blocked.
2045     *
2046     * <p>If the caller is not a {@link ForkJoinTask}, this method is
2047     * behaviorally equivalent to
2048     *  <pre> {@code
2049     * while (!blocker.isReleasable())
2050     *   if (blocker.block())
2051     *     return;
2052     * }</pre>
2053     *
2054     * If the caller is a {@code ForkJoinTask}, then the pool may
2055     * first be expanded to ensure parallelism, and later adjusted.
2056     *
2057     * @param blocker the blocker
2058     * @throws InterruptedException if blocker.block did so
2059     */
2060    public static void managedBlock(ManagedBlocker blocker)
2061        throws InterruptedException {
2062        Thread t = Thread.currentThread();
2063        if (t instanceof ForkJoinWorkerThread) {
2064            ForkJoinWorkerThread w = (ForkJoinWorkerThread) t;
2065            w.pool.awaitBlocker(blocker);
2066        }
2067        else {
2068            do {} while (!blocker.isReleasable() && !blocker.block());
2069        }
2070    }
2071
2072    // AbstractExecutorService overrides.  These rely on undocumented
2073    // fact that ForkJoinTask.adapt returns ForkJoinTasks that also
2074    // implement RunnableFuture.
2075
2076    protected <T> RunnableFuture<T> newTaskFor(Runnable runnable, T value) {
2077        return (RunnableFuture<T>) ForkJoinTask.adapt(runnable, value);
2078    }
2079
2080    protected <T> RunnableFuture<T> newTaskFor(Callable<T> callable) {
2081        return (RunnableFuture<T>) ForkJoinTask.adapt(callable);
2082    }
2083
2084    // Unsafe mechanics
2085    private static final sun.misc.Unsafe UNSAFE;
2086    private static final long ctlOffset;
2087    private static final long stealCountOffset;
2088    private static final long blockedCountOffset;
2089    private static final long quiescerCountOffset;
2090    private static final long scanGuardOffset;
2091    private static final long nextWorkerNumberOffset;
2092    private static final long ABASE;
2093    private static final int ASHIFT;
2094
2095    static {
2096        poolNumberGenerator = new AtomicInteger();
2097        workerSeedGenerator = new Random();
2098        modifyThreadPermission = new RuntimePermission("modifyThread");
2099        defaultForkJoinWorkerThreadFactory =
2100            new DefaultForkJoinWorkerThreadFactory();
2101        try {
2102            UNSAFE = sun.misc.Unsafe.getUnsafe();
2103            Class<?> k = ForkJoinPool.class;
2104            ctlOffset = UNSAFE.objectFieldOffset
2105                (k.getDeclaredField("ctl"));
2106            stealCountOffset = UNSAFE.objectFieldOffset
2107                (k.getDeclaredField("stealCount"));
2108            blockedCountOffset = UNSAFE.objectFieldOffset
2109                (k.getDeclaredField("blockedCount"));
2110            quiescerCountOffset = UNSAFE.objectFieldOffset
2111                (k.getDeclaredField("quiescerCount"));
2112            scanGuardOffset = UNSAFE.objectFieldOffset
2113                (k.getDeclaredField("scanGuard"));
2114            nextWorkerNumberOffset = UNSAFE.objectFieldOffset
2115                (k.getDeclaredField("nextWorkerNumber"));
2116        } catch (Exception e) {
2117            throw new Error(e);
2118        }
2119        Class<?> a = ForkJoinTask[].class;
2120        ABASE = UNSAFE.arrayBaseOffset(a);
2121        int s = UNSAFE.arrayIndexScale(a);
2122        if ((s & (s-1)) != 0)
2123            throw new Error("data type scale not a power of two");
2124        ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
2125    }
2126
2127}
2128