Making java.util.concurrent compilable without references to sun.misc.Unsafe and co.
2 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 * This code is free software; you can redistribute it and/or modify it
5 * under the terms of the GNU General Public License version 2 only, as
6 * published by the Free Software Foundation. Oracle designates this
7 * particular file as subject to the "Classpath" exception as provided
8 * by Oracle in the LICENSE file that accompanied this code.
10 * This code is distributed in the hope that it will be useful, but WITHOUT
11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
13 * version 2 for more details (a copy is included in the LICENSE file that
14 * accompanied this code).
16 * You should have received a copy of the GNU General Public License version
17 * 2 along with this work; if not, write to the Free Software Foundation,
18 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
20 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
21 * or visit www.oracle.com if you need additional information or have any
26 * This file is available under and governed by the GNU General Public
27 * License version 2 only, as published by the Free Software Foundation.
28 * However, the following notice accompanied the original version of this
31 * Written by Doug Lea with assistance from members of JCP JSR-166
32 * Expert Group and released to the public domain, as explained at
33 * http://creativecommons.org/publicdomain/zero/1.0/
36 package java.util.concurrent;
38 import java.util.ArrayList;
39 import java.util.Arrays;
40 import java.util.Collection;
41 import java.util.Collections;
42 import java.util.List;
43 import java.util.Random;
44 import java.util.concurrent.AbstractExecutorService;
45 import java.util.concurrent.Callable;
46 import java.util.concurrent.ExecutorService;
47 import java.util.concurrent.Future;
48 import java.util.concurrent.RejectedExecutionException;
49 import java.util.concurrent.RunnableFuture;
50 import java.util.concurrent.TimeUnit;
51 import java.util.concurrent.atomic.AtomicInteger;
52 import java.util.concurrent.locks.LockSupport;
53 import java.util.concurrent.locks.ReentrantLock;
54 import java.util.concurrent.locks.Condition;
57 * An {@link ExecutorService} for running {@link ForkJoinTask}s.
58 * A {@code ForkJoinPool} provides the entry point for submissions
59 * from non-{@code ForkJoinTask} clients, as well as management and
60 * monitoring operations.
62 * <p>A {@code ForkJoinPool} differs from other kinds of {@link
63 * ExecutorService} mainly by virtue of employing
64 * <em>work-stealing</em>: all threads in the pool attempt to find and
65 * execute subtasks created by other active tasks (eventually blocking
66 * waiting for work if none exist). This enables efficient processing
67 * when most tasks spawn other subtasks (as do most {@code
68 * ForkJoinTask}s). When setting <em>asyncMode</em> to true in
69 * constructors, {@code ForkJoinPool}s may also be appropriate for use
70 * with event-style tasks that are never joined.
72 * <p>A {@code ForkJoinPool} is constructed with a given target
73 * parallelism level; by default, equal to the number of available
74 * processors. The pool attempts to maintain enough active (or
75 * available) threads by dynamically adding, suspending, or resuming
76 * internal worker threads, even if some tasks are stalled waiting to
77 * join others. However, no such adjustments are guaranteed in the
78 * face of blocked IO or other unmanaged synchronization. The nested
79 * {@link ManagedBlocker} interface enables extension of the kinds of
80 * synchronization accommodated.
82 * <p>In addition to execution and lifecycle control methods, this
83 * class provides status check methods (for example
84 * {@link #getStealCount}) that are intended to aid in developing,
85 * tuning, and monitoring fork/join applications. Also, method
86 * {@link #toString} returns indications of pool state in a
87 * convenient form for informal monitoring.
89 * <p> As is the case with other ExecutorServices, there are three
90 * main task execution methods summarized in the following
91 * table. These are designed to be used by clients not already engaged
92 * in fork/join computations in the current pool. The main forms of
93 * these methods accept instances of {@code ForkJoinTask}, but
94 * overloaded forms also allow mixed execution of plain {@code
95 * Runnable}- or {@code Callable}- based activities as well. However,
96 * tasks that are already executing in a pool should normally
97 * <em>NOT</em> use these pool execution methods, but instead use the
98 * within-computation forms listed in the table.
100 * <table BORDER CELLPADDING=3 CELLSPACING=1>
103 * <td ALIGN=CENTER> <b>Call from non-fork/join clients</b></td>
104 * <td ALIGN=CENTER> <b>Call from within fork/join computations</b></td>
107 * <td> <b>Arrange async execution</td>
108 * <td> {@link #execute(ForkJoinTask)}</td>
109 * <td> {@link ForkJoinTask#fork}</td>
112 * <td> <b>Await and obtain result</td>
113 * <td> {@link #invoke(ForkJoinTask)}</td>
114 * <td> {@link ForkJoinTask#invoke}</td>
117 * <td> <b>Arrange exec and obtain Future</td>
118 * <td> {@link #submit(ForkJoinTask)}</td>
119 * <td> {@link ForkJoinTask#fork} (ForkJoinTasks <em>are</em> Futures)</td>
123 * <p><b>Sample Usage.</b> Normally a single {@code ForkJoinPool} is
124 * used for all parallel task execution in a program or subsystem.
125 * Otherwise, use would not usually outweigh the construction and
126 * bookkeeping overhead of creating a large set of threads. For
127 * example, a common pool could be used for the {@code SortTasks}
128 * illustrated in {@link RecursiveAction}. Because {@code
129 * ForkJoinPool} uses threads in {@linkplain java.lang.Thread#isDaemon
130 * daemon} mode, there is typically no need to explicitly {@link
131 * #shutdown} such a pool upon program exit.
134 * static final ForkJoinPool mainPool = new ForkJoinPool();
136 * public void sort(long[] array) {
137 * mainPool.invoke(new SortTask(array, 0, array.length));
141 * <p><b>Implementation notes</b>: This implementation restricts the
142 * maximum number of running threads to 32767. Attempts to create
143 * pools with greater than the maximum number result in
144 * {@code IllegalArgumentException}.
146 * <p>This implementation rejects submitted tasks (that is, by throwing
147 * {@link RejectedExecutionException}) only when the pool is shut down
148 * or internal resources have been exhausted.
153 public class ForkJoinPool extends AbstractExecutorService {
156 * Implementation Overview
158 * This class provides the central bookkeeping and control for a
159 * set of worker threads: Submissions from non-FJ threads enter
160 * into a submission queue. Workers take these tasks and typically
161 * split them into subtasks that may be stolen by other workers.
162 * Preference rules give first priority to processing tasks from
163 * their own queues (LIFO or FIFO, depending on mode), then to
164 * randomized FIFO steals of tasks in other worker queues, and
165 * lastly to new submissions.
167 * The main throughput advantages of work-stealing stem from
168 * decentralized control -- workers mostly take tasks from
169 * themselves or each other. We cannot negate this in the
170 * implementation of other management responsibilities. The main
171 * tactic for avoiding bottlenecks is packing nearly all
172 * essentially atomic control state into a single 64bit volatile
173 * variable ("ctl"). This variable is read on the order of 10-100
174 * times as often as it is modified (always via CAS). (There is
175 * some additional control state, for example variable "shutdown"
176 * for which we can cope with uncoordinated updates.) This
177 * streamlines synchronization and control at the expense of messy
178 * constructions needed to repack status bits upon updates.
179 * Updates tend not to contend with each other except during
180 * bursts while submitted tasks begin or end. In some cases when
181 * they do contend, threads can instead do something else
182 * (usually, scan for tasks) until contention subsides.
184 * To enable packing, we restrict maximum parallelism to (1<<15)-1
185 * (which is far in excess of normal operating range) to allow
186 * ids, counts, and their negations (used for thresholding) to fit
189 * Recording Workers. Workers are recorded in the "workers" array
190 * that is created upon pool construction and expanded if (rarely)
191 * necessary. This is an array as opposed to some other data
192 * structure to support index-based random steals by workers.
193 * Updates to the array recording new workers and unrecording
194 * terminated ones are protected from each other by a seqLock
195 * (scanGuard) but the array is otherwise concurrently readable,
196 * and accessed directly by workers. To simplify index-based
197 * operations, the array size is always a power of two, and all
198 * readers must tolerate null slots. To avoid flailing during
199 * start-up, the array is presized to hold twice #parallelism
200 * workers (which is unlikely to need further resizing during
201 * execution). But to avoid dealing with so many null slots,
202 * variable scanGuard includes a mask for the nearest power of two
203 * that contains all current workers. All worker thread creation
204 * is on-demand, triggered by task submissions, replacement of
205 * terminated workers, and/or compensation for blocked
206 * workers. However, all other support code is set up to work with
207 * other policies. To ensure that we do not hold on to worker
208 * references that would prevent GC, ALL accesses to workers are
209 * via indices into the workers array (which is one source of some
210 * of the messy code constructions here). In essence, the workers
211 * array serves as a weak reference mechanism. Thus for example
212 * the wait queue field of ctl stores worker indices, not worker
213 * references. Access to the workers in associated methods (for
214 * example signalWork) must both index-check and null-check the
215 * IDs. All such accesses ignore bad IDs by returning out early
216 * from what they are doing, since this can only be associated
217 * with termination, in which case it is OK to give up.
219 * All uses of the workers array, as well as queue arrays, check
220 * that the array is non-null (even if previously non-null). This
221 * allows nulling during termination, which is currently not
222 * necessary, but remains an option for resource-revocation-based
225 * Wait Queuing. Unlike HPC work-stealing frameworks, we cannot
226 * let workers spin indefinitely scanning for tasks when none can
227 * be found immediately, and we cannot start/resume workers unless
228 * there appear to be tasks available. On the other hand, we must
229 * quickly prod them into action when new tasks are submitted or
230 * generated. We park/unpark workers after placing in an event
231 * wait queue when they cannot find work. This "queue" is actually
232 * a simple Treiber stack, headed by the "id" field of ctl, plus a
233 * 15bit counter value to both wake up waiters (by advancing their
234 * count) and avoid ABA effects. Successors are held in worker
235 * field "nextWait". Queuing deals with several intrinsic races,
236 * mainly that a task-producing thread can miss seeing (and
237 * signalling) another thread that gave up looking for work but
238 * has not yet entered the wait queue. We solve this by requiring
239 * a full sweep of all workers both before (in scan()) and after
240 * (in tryAwaitWork()) a newly waiting worker is added to the wait
241 * queue. During a rescan, the worker might release some other
242 * queued worker rather than itself, which has the same net
243 * effect. Because enqueued workers may actually be rescanning
244 * rather than waiting, we set and clear the "parked" field of
245 * ForkJoinWorkerThread to reduce unnecessary calls to unpark.
246 * (Use of the parked field requires a secondary recheck to avoid
249 * Signalling. We create or wake up workers only when there
250 * appears to be at least one task they might be able to find and
251 * execute. When a submission is added or another worker adds a
252 * task to a queue that previously had two or fewer tasks, they
253 * signal waiting workers (or trigger creation of new ones if
254 * fewer than the given parallelism level -- see signalWork).
255 * These primary signals are buttressed by signals during rescans
256 * as well as those performed when a worker steals a task and
257 * notices that there are more tasks too; together these cover the
258 * signals needed in cases when more than two tasks are pushed
261 * Trimming workers. To release resources after periods of lack of
262 * use, a worker starting to wait when the pool is quiescent will
263 * time out and terminate if the pool has remained quiescent for
264 * SHRINK_RATE nanosecs. This will slowly propagate, eventually
265 * terminating all workers after long periods of non-use.
267 * Submissions. External submissions are maintained in an
268 * array-based queue that is structured identically to
269 * ForkJoinWorkerThread queues except for the use of
270 * submissionLock in method addSubmission. Unlike the case for
271 * worker queues, multiple external threads can add new
272 * submissions, so adding requires a lock.
274 * Compensation. Beyond work-stealing support and lifecycle
275 * control, the main responsibility of this framework is to take
276 * actions when one worker is waiting to join a task stolen (or
277 * always held by) another. Because we are multiplexing many
278 * tasks on to a pool of workers, we can't just let them block (as
279 * in Thread.join). We also cannot just reassign the joiner's
280 * run-time stack with another and replace it later, which would
281 * be a form of "continuation", that even if possible is not
282 * necessarily a good idea since we sometimes need both an
283 * unblocked task and its continuation to progress. Instead we
284 * combine two tactics:
286 * Helping: Arranging for the joiner to execute some task that it
287 * would be running if the steal had not occurred. Method
288 * ForkJoinWorkerThread.joinTask tracks joining->stealing
289 * links to try to find such a task.
291 * Compensating: Unless there are already enough live threads,
292 * method tryPreBlock() may create or re-activate a spare
293 * thread to compensate for blocked joiners until they
296 * The ManagedBlocker extension API can't use helping so relies
297 * only on compensation in method awaitBlocker.
299 * It is impossible to keep exactly the target parallelism number
300 * of threads running at any given time. Determining the
301 * existence of conservatively safe helping targets, the
302 * availability of already-created spares, and the apparent need
303 * to create new spares are all racy and require heuristic
304 * guidance, so we rely on multiple retries of each. Currently,
305 * in keeping with on-demand signalling policy, we compensate only
306 * if blocking would leave less than one active (non-waiting,
307 * non-blocked) worker. Additionally, to avoid some false alarms
308 * due to GC, lagging counters, system activity, etc, compensated
309 * blocking for joins is only attempted after rechecks stabilize
310 * (retries are interspersed with Thread.yield, for good
311 * citizenship). The variable blockedCount, incremented before
312 * blocking and decremented after, is sometimes needed to
313 * distinguish cases of waiting for work vs blocking on joins or
314 * other managed sync. Both cases are equivalent for most pool
315 * control, so we can update non-atomically. (Additionally,
316 * contention on blockedCount alleviates some contention on ctl).
318 * Shutdown and Termination. A call to shutdownNow atomically sets
319 * the ctl stop bit and then (non-atomically) sets each workers
320 * "terminate" status, cancels all unprocessed tasks, and wakes up
321 * all waiting workers. Detecting whether termination should
322 * commence after a non-abrupt shutdown() call requires more work
323 * and bookkeeping. We need consensus about quiesence (i.e., that
324 * there is no more work) which is reflected in active counts so
325 * long as there are no current blockers, as well as possible
326 * re-evaluations during independent changes in blocking or
329 * Style notes: There is a lot of representation-level coupling
330 * among classes ForkJoinPool, ForkJoinWorkerThread, and
331 * ForkJoinTask. Most fields of ForkJoinWorkerThread maintain
332 * data structures managed by ForkJoinPool, so are directly
333 * accessed. Conversely we allow access to "workers" array by
334 * workers, and direct access to ForkJoinTask.status by both
335 * ForkJoinPool and ForkJoinWorkerThread. There is little point
336 * trying to reduce this, since any associated future changes in
337 * representations will need to be accompanied by algorithmic
338 * changes anyway. All together, these low-level implementation
339 * choices produce as much as a factor of 4 performance
340 * improvement compared to naive implementations, and enable the
341 * processing of billions of tasks per second, at the expense of
344 * Methods signalWork() and scan() are the main bottlenecks so are
345 * especially heavily micro-optimized/mangled. There are lots of
346 * inline assignments (of form "while ((local = field) != 0)")
347 * which are usually the simplest way to ensure the required read
348 * orderings (which are sometimes critical). This leads to a
349 * "C"-like style of listing declarations of these locals at the
350 * heads of methods or blocks. There are several occurrences of
351 * the unusual "do {} while (!cas...)" which is the simplest way
352 * to force an update of a CAS'ed variable. There are also other
353 * coding oddities that help some methods perform reasonably even
354 * when interpreted (not compiled).
356 * The order of declarations in this file is: (1) declarations of
357 * statics (2) fields (along with constants used when unpacking
358 * some of them), listed in an order that tends to reduce
359 * contention among them a bit under most JVMs. (3) internal
360 * control methods (4) callbacks and other support for
361 * ForkJoinTask and ForkJoinWorkerThread classes, (5) exported
362 * methods (plus a few little helpers). (6) static block
363 * initializing all statics in a minimally dependent order.
367 * Factory for creating new {@link ForkJoinWorkerThread}s.
368 * A {@code ForkJoinWorkerThreadFactory} must be defined and used
369 * for {@code ForkJoinWorkerThread} subclasses that extend base
370 * functionality or initialize threads with different contexts.
372 public static interface ForkJoinWorkerThreadFactory {
374 * Returns a new worker thread operating in the given pool.
376 * @param pool the pool this thread works in
377 * @throws NullPointerException if the pool is null
379 public ForkJoinWorkerThread newThread(ForkJoinPool pool);
383 * Default ForkJoinWorkerThreadFactory implementation; creates a
384 * new ForkJoinWorkerThread.
386 static class DefaultForkJoinWorkerThreadFactory
387 implements ForkJoinWorkerThreadFactory {
388 public ForkJoinWorkerThread newThread(ForkJoinPool pool) {
389 return new ForkJoinWorkerThread(pool);
394 * Creates a new ForkJoinWorkerThread. This factory is used unless
395 * overridden in ForkJoinPool constructors.
397 public static final ForkJoinWorkerThreadFactory
398 defaultForkJoinWorkerThreadFactory;
401 * If there is a security manager, makes sure caller has
402 * permission to modify threads.
404 private static void checkPermission() {
405 throw new SecurityException();
409 * Generator for assigning sequence numbers as pool names.
411 private static final AtomicInteger poolNumberGenerator;
414 * Generator for initial random seeds for worker victim
415 * selection. This is used only to create initial seeds. Random
416 * steals use a cheaper xorshift generator per steal attempt. We
417 * don't expect much contention on seedGenerator, so just use a
420 static final Random workerSeedGenerator;
423 * Array holding all worker threads in the pool. Initialized upon
424 * construction. Array size must be a power of two. Updates and
425 * replacements are protected by scanGuard, but the array is
426 * always kept in a consistent enough state to be randomly
427 * accessed without locking by workers performing work-stealing,
428 * as well as other traversal-based methods in this class, so long
429 * as reads memory-acquire by first reading ctl. All readers must
430 * tolerate that some array slots may be null.
432 ForkJoinWorkerThread[] workers;
435 * Initial size for submission queue array. Must be a power of
436 * two. In many applications, these always stay small so we use a
439 private static final int INITIAL_QUEUE_CAPACITY = 8;
442 * Maximum size for submission queue array. Must be a power of two
443 * less than or equal to 1 << (31 - width of array entry) to
444 * ensure lack of index wraparound, but is capped at a lower
445 * value to help users trap runaway computations.
447 private static final int MAXIMUM_QUEUE_CAPACITY = 1 << 24; // 16M
450 * Array serving as submission queue. Initialized upon construction.
452 private ForkJoinTask<?>[] submissionQueue;
455 * Lock protecting submissions array for addSubmission
457 private final ReentrantLock submissionLock;
460 * Condition for awaitTermination, using submissionLock for
463 private final Condition termination;
466 * Creation factory for worker threads.
468 private final ForkJoinWorkerThreadFactory factory;
471 * The uncaught exception handler used when any worker abruptly
474 final Thread.UncaughtExceptionHandler ueh;
477 * Prefix for assigning names to worker threads
479 private final String workerNamePrefix;
482 * Sum of per-thread steal counts, updated only when threads are
483 * idle or terminating.
485 private volatile long stealCount;
488 * Main pool control -- a long packed with:
489 * AC: Number of active running workers minus target parallelism (16 bits)
490 * TC: Number of total workers minus target parallelism (16bits)
491 * ST: true if pool is terminating (1 bit)
492 * EC: the wait count of top waiting thread (15 bits)
493 * ID: ~poolIndex of top of Treiber stack of waiting threads (16 bits)
495 * When convenient, we can extract the upper 32 bits of counts and
496 * the lower 32 bits of queue state, u = (int)(ctl >>> 32) and e =
497 * (int)ctl. The ec field is never accessed alone, but always
498 * together with id and st. The offsets of counts by the target
499 * parallelism and the positionings of fields makes it possible to
500 * perform the most common checks via sign tests of fields: When
501 * ac is negative, there are not enough active workers, when tc is
502 * negative, there are not enough total workers, when id is
503 * negative, there is at least one waiting worker, and when e is
504 * negative, the pool is terminating. To deal with these possibly
505 * negative fields, we use casts in and out of "short" and/or
506 * signed shifts to maintain signedness.
510 // bit positions/shifts for fields
511 private static final int AC_SHIFT = 48;
512 private static final int TC_SHIFT = 32;
513 private static final int ST_SHIFT = 31;
514 private static final int EC_SHIFT = 16;
517 private static final int MAX_ID = 0x7fff; // max poolIndex
518 private static final int SMASK = 0xffff; // mask short bits
519 private static final int SHORT_SIGN = 1 << 15;
520 private static final int INT_SIGN = 1 << 31;
523 private static final long STOP_BIT = 0x0001L << ST_SHIFT;
524 private static final long AC_MASK = ((long)SMASK) << AC_SHIFT;
525 private static final long TC_MASK = ((long)SMASK) << TC_SHIFT;
527 // units for incrementing and decrementing
528 private static final long TC_UNIT = 1L << TC_SHIFT;
529 private static final long AC_UNIT = 1L << AC_SHIFT;
531 // masks and units for dealing with u = (int)(ctl >>> 32)
532 private static final int UAC_SHIFT = AC_SHIFT - 32;
533 private static final int UTC_SHIFT = TC_SHIFT - 32;
534 private static final int UAC_MASK = SMASK << UAC_SHIFT;
535 private static final int UTC_MASK = SMASK << UTC_SHIFT;
536 private static final int UAC_UNIT = 1 << UAC_SHIFT;
537 private static final int UTC_UNIT = 1 << UTC_SHIFT;
539 // masks and units for dealing with e = (int)ctl
540 private static final int E_MASK = 0x7fffffff; // no STOP_BIT
541 private static final int EC_UNIT = 1 << EC_SHIFT;
544 * The target parallelism level.
546 final int parallelism;
549 * Index (mod submission queue length) of next element to take
550 * from submission queue. Usage is identical to that for
551 * per-worker queues -- see ForkJoinWorkerThread internal
554 volatile int queueBase;
557 * Index (mod submission queue length) of next element to add
558 * in submission queue. Usage is identical to that for
559 * per-worker queues -- see ForkJoinWorkerThread internal
565 * True when shutdown() has been called.
567 volatile boolean shutdown;
570 * True if use local fifo, not default lifo, for local polling
571 * Read by, and replicated by ForkJoinWorkerThreads
573 final boolean locallyFifo;
576 * The number of threads in ForkJoinWorkerThreads.helpQuiescePool.
577 * When non-zero, suppresses automatic shutdown when active
578 * counts become zero.
580 volatile int quiescerCount;
583 * The number of threads blocked in join.
585 volatile int blockedCount;
588 * Counter for worker Thread names (unrelated to their poolIndex)
590 private volatile int nextWorkerNumber;
593 * The index for the next created worker. Accessed under scanGuard.
595 private int nextWorkerIndex;
598 * SeqLock and index masking for updates to workers array. Locked
599 * when SG_UNIT is set. Unlocking clears bit by adding
600 * SG_UNIT. Staleness of read-only operations can be checked by
601 * comparing scanGuard to value before the reads. The low 16 bits
602 * (i.e, anding with SMASK) hold (the smallest power of two
603 * covering all worker indices, minus one, and is used to avoid
604 * dealing with large numbers of null slots when the workers array
607 volatile int scanGuard;
609 private static final int SG_UNIT = 1 << 16;
612 * The wakeup interval (in nanoseconds) for a worker waiting for a
613 * task when the pool is quiescent to instead try to shrink the
614 * number of workers. The exact value does not matter too
615 * much. It must be short enough to release resources during
616 * sustained periods of idleness, but not so short that threads
617 * are continually re-created.
619 private static final long SHRINK_RATE =
620 4L * 1000L * 1000L * 1000L; // 4 seconds
623 * Top-level loop for worker threads: On each step: if the
624 * previous step swept through all queues and found no tasks, or
625 * there are excess threads, then possibly blocks. Otherwise,
626 * scans for and, if found, executes a task. Returns when pool
627 * and/or worker terminate.
629 * @param w the worker
631 final void work(ForkJoinWorkerThread w) {
632 boolean swept = false; // true on empty scans
634 while (!w.terminate && (int)(c = ctl) >= 0) {
635 int a; // active count
636 if (!swept && (a = (int)(c >> AC_SHIFT)) <= 0)
638 else if (tryAwaitWork(w, c))
646 * Wakes up or creates a worker.
648 final void signalWork() {
650 * The while condition is true if: (there is are too few total
651 * workers OR there is at least one waiter) AND (there are too
652 * few active workers OR the pool is terminating). The value
653 * of e distinguishes the remaining cases: zero (no waiters)
654 * for create, negative if terminating (in which case do
655 * nothing), else release a waiter. The secondary checks for
656 * release (non-null array etc) can fail if the pool begins
657 * terminating after the test, and don't impose any added cost
658 * because JVMs must perform null and bounds checks anyway.
661 while ((((e = (int)(c = ctl)) | (u = (int)(c >>> 32))) &
662 (INT_SIGN|SHORT_SIGN)) == (INT_SIGN|SHORT_SIGN) && e >= 0) {
663 if (e > 0) { // release a waiting worker
664 int i; ForkJoinWorkerThread w; ForkJoinWorkerThread[] ws;
665 if ((ws = workers) == null ||
666 (i = ~e & SMASK) >= ws.length ||
669 long nc = (((long)(w.nextWait & E_MASK)) |
670 ((long)(u + UAC_UNIT) << 32));
671 if (w.eventCount == e &&
672 UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
673 w.eventCount = (e + EC_UNIT) & E_MASK;
679 else if (UNSAFE.compareAndSwapLong
681 (long)(((u + UTC_UNIT) & UTC_MASK) |
682 ((u + UAC_UNIT) & UAC_MASK)) << 32)) {
690 * Variant of signalWork to help release waiters on rescans.
691 * Tries once to release a waiter if active count < 0.
693 * @return false if failed due to contention, else true
695 private boolean tryReleaseWaiter() {
696 long c; int e, i; ForkJoinWorkerThread w; ForkJoinWorkerThread[] ws;
697 if ((e = (int)(c = ctl)) > 0 &&
698 (int)(c >> AC_SHIFT) < 0 &&
699 (ws = workers) != null &&
700 (i = ~e & SMASK) < ws.length &&
701 (w = ws[i]) != null) {
702 long nc = ((long)(w.nextWait & E_MASK) |
703 ((c + AC_UNIT) & (AC_MASK|TC_MASK)));
704 if (w.eventCount != e ||
705 !UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc))
707 w.eventCount = (e + EC_UNIT) & E_MASK;
714 // Scanning for tasks
717 * Scans for and, if found, executes one task. Scans start at a
718 * random index of workers array, and randomly select the first
719 * (2*#workers)-1 probes, and then, if all empty, resort to 2
720 * circular sweeps, which is necessary to check quiescence. and
721 * taking a submission only if no stealable tasks were found. The
722 * steal code inside the loop is a specialized form of
723 * ForkJoinWorkerThread.deqTask, followed bookkeeping to support
724 * helpJoinTask and signal propagation. The code for submission
725 * queues is almost identical. On each steal, the worker completes
726 * not only the task, but also all local tasks that this task may
727 * have generated. On detecting staleness or contention when
728 * trying to take a task, this method returns without finishing
729 * sweep, which allows global state rechecks before retry.
731 * @param w the worker
732 * @param a the number of active workers
733 * @return true if swept all queues without finding a task
735 private boolean scan(ForkJoinWorkerThread w, int a) {
736 int g = scanGuard; // mask 0 avoids useless scans if only one active
737 int m = (parallelism == 1 - a && blockedCount == 0) ? 0 : g & SMASK;
738 ForkJoinWorkerThread[] ws = workers;
739 if (ws == null || ws.length <= m) // staleness check
741 for (int r = w.seed, k = r, j = -(m + m); j <= m + m; ++j) {
742 ForkJoinTask<?> t; ForkJoinTask<?>[] q; int b, i;
743 ForkJoinWorkerThread v = ws[k & m];
744 if (v != null && (b = v.queueBase) != v.queueTop &&
745 (q = v.queue) != null && (i = (q.length - 1) & b) >= 0) {
746 long u = (i << ASHIFT) + ABASE;
747 if ((t = q[i]) != null && v.queueBase == b &&
748 UNSAFE.compareAndSwapObject(q, u, t, null)) {
749 int d = (v.queueBase = b + 1) - v.queueTop;
750 v.stealHint = w.poolIndex;
752 signalWork(); // propagate if nonempty
755 r ^= r << 13; r ^= r >>> 17; w.seed = r ^ (r << 5);
756 return false; // store next seed
758 else if (j < 0) { // xorshift
759 r ^= r << 13; r ^= r >>> 17; k = r ^= r << 5;
764 if (scanGuard != g) // staleness check
766 else { // try to take submission
767 ForkJoinTask<?> t; ForkJoinTask<?>[] q; int b, i;
768 if ((b = queueBase) != queueTop &&
769 (q = submissionQueue) != null &&
770 (i = (q.length - 1) & b) >= 0) {
771 long u = (i << ASHIFT) + ABASE;
772 if ((t = q[i]) != null && queueBase == b &&
773 UNSAFE.compareAndSwapObject(q, u, t, null)) {
779 return true; // all queues empty
784 * Tries to enqueue worker w in wait queue and await change in
785 * worker's eventCount. If the pool is quiescent and there is
786 * more than one worker, possibly terminates worker upon exit.
787 * Otherwise, before blocking, rescans queues to avoid missed
788 * signals. Upon finding work, releases at least one worker
789 * (which may be the current worker). Rescans restart upon
790 * detected staleness or failure to release due to
791 * contention. Note the unusual conventions about Thread.interrupt
792 * here and elsewhere: Because interrupts are used solely to alert
793 * threads to check termination, which is checked here anyway, we
794 * clear status (using Thread.interrupted) before any call to
795 * park, so that park does not immediately return due to status
796 * being set via some other unrelated call to interrupt in user
799 * @param w the calling worker
800 * @param c the ctl value on entry
801 * @return true if waited or another thread was released upon enq
803 private boolean tryAwaitWork(ForkJoinWorkerThread w, long c) {
804 int v = w.eventCount;
805 w.nextWait = (int)c; // w's successor record
806 long nc = (long)(v & E_MASK) | ((c - AC_UNIT) & (AC_MASK|TC_MASK));
807 if (ctl != c || !UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
808 long d = ctl; // return true if lost to a deq, to force scan
809 return (int)d != (int)c && ((d - c) & AC_MASK) >= 0L;
811 for (int sc = w.stealCount; sc != 0;) { // accumulate stealCount
813 if (UNSAFE.compareAndSwapLong(this, stealCountOffset, s, s + sc))
814 sc = w.stealCount = 0;
815 else if (w.eventCount != v)
816 return true; // update next time
818 if ((!shutdown || !tryTerminate(false)) &&
819 (int)c != 0 && parallelism + (int)(nc >> AC_SHIFT) == 0 &&
820 blockedCount == 0 && quiescerCount == 0)
821 idleAwaitWork(w, nc, c, v); // quiescent
822 for (boolean rescanned = false;;) {
823 if (w.eventCount != v)
826 int g = scanGuard, m = g & SMASK;
827 ForkJoinWorkerThread[] ws = workers;
828 if (ws != null && m < ws.length) {
830 for (int i = 0; i <= m; ++i) {
831 ForkJoinWorkerThread u = ws[i];
833 if (u.queueBase != u.queueTop &&
835 rescanned = false; // contended
836 if (w.eventCount != v)
841 if (scanGuard != g || // stale
842 (queueBase != queueTop && !tryReleaseWaiter()))
845 Thread.yield(); // reduce contention
847 Thread.interrupted(); // clear before park
850 w.parked = true; // must recheck
851 if (w.eventCount != v) {
855 LockSupport.park(this);
856 rescanned = w.parked = false;
862 * If inactivating worker w has caused pool to become
863 * quiescent, check for pool termination, and wait for event
864 * for up to SHRINK_RATE nanosecs (rescans are unnecessary in
865 * this case because quiescence reflects consensus about lack
866 * of work). On timeout, if ctl has not changed, terminate the
867 * worker. Upon its termination (see deregisterWorker), it may
868 * wake up another worker to possibly repeat this process.
870 * @param w the calling worker
871 * @param currentCtl the ctl value after enqueuing w
872 * @param prevCtl the ctl value if w terminated
873 * @param v the eventCount w awaits change
875 private void idleAwaitWork(ForkJoinWorkerThread w, long currentCtl,
876 long prevCtl, int v) {
877 if (w.eventCount == v) {
880 ForkJoinTask.helpExpungeStaleExceptions(); // help clean weak refs
881 while (ctl == currentCtl) {
882 long startTime = System.nanoTime();
884 if (w.eventCount == v) // must recheck
885 LockSupport.parkNanos(this, SHRINK_RATE);
887 if (w.eventCount != v)
889 else if (System.nanoTime() - startTime <
890 SHRINK_RATE - (SHRINK_RATE / 10)) // timing slop
891 Thread.interrupted(); // spurious wakeup
892 else if (UNSAFE.compareAndSwapLong(this, ctlOffset,
893 currentCtl, prevCtl)) {
894 w.terminate = true; // restore previous
895 w.eventCount = ((int)currentCtl + EC_UNIT) & E_MASK;
905 * Enqueues the given task in the submissionQueue. Same idea as
906 * ForkJoinWorkerThread.pushTask except for use of submissionLock.
910 private void addSubmission(ForkJoinTask<?> t) {
911 final ReentrantLock lock = this.submissionLock;
914 ForkJoinTask<?>[] q; int s, m;
915 if ((q = submissionQueue) != null) { // ignore if queue removed
916 long u = (((s = queueTop) & (m = q.length-1)) << ASHIFT)+ABASE;
917 UNSAFE.putOrderedObject(q, u, t);
919 if (s - queueBase == m)
920 growSubmissionQueue();
928 // (pollSubmission is defined below with exported methods)
931 * Creates or doubles submissionQueue array.
932 * Basically identical to ForkJoinWorkerThread version.
934 private void growSubmissionQueue() {
935 ForkJoinTask<?>[] oldQ = submissionQueue;
936 int size = oldQ != null ? oldQ.length << 1 : INITIAL_QUEUE_CAPACITY;
937 if (size > MAXIMUM_QUEUE_CAPACITY)
938 throw new RejectedExecutionException("Queue capacity exceeded");
939 if (size < INITIAL_QUEUE_CAPACITY)
940 size = INITIAL_QUEUE_CAPACITY;
941 ForkJoinTask<?>[] q = submissionQueue = new ForkJoinTask<?>[size];
945 if (oldQ != null && (oldMask = oldQ.length - 1) >= 0) {
946 for (int b = queueBase; b != top; ++b) {
947 long u = ((b & oldMask) << ASHIFT) + ABASE;
948 Object x = UNSAFE.getObjectVolatile(oldQ, u);
949 if (x != null && UNSAFE.compareAndSwapObject(oldQ, u, x, null))
950 UNSAFE.putObjectVolatile
951 (q, ((b & mask) << ASHIFT) + ABASE, x);
959 * Tries to increment blockedCount, decrement active count
960 * (sometimes implicitly) and possibly release or create a
961 * compensating worker in preparation for blocking. Fails
962 * on contention or termination.
964 * @return true if the caller can block, else should recheck and retry
966 private boolean tryPreBlock() {
967 int b = blockedCount;
968 if (UNSAFE.compareAndSwapInt(this, blockedCountOffset, b, b + 1)) {
969 int pc = parallelism;
971 ForkJoinWorkerThread[] ws; ForkJoinWorkerThread w;
972 int e, ac, tc, rc, i;
974 int u = (int)(c >>> 32);
975 if ((e = (int)c) < 0) {
976 // skip -- terminating
978 else if ((ac = (u >> UAC_SHIFT)) <= 0 && e != 0 &&
979 (ws = workers) != null &&
980 (i = ~e & SMASK) < ws.length &&
981 (w = ws[i]) != null) {
982 long nc = ((long)(w.nextWait & E_MASK) |
983 (c & (AC_MASK|TC_MASK)));
984 if (w.eventCount == e &&
985 UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
986 w.eventCount = (e + EC_UNIT) & E_MASK;
989 return true; // release an idle worker
992 else if ((tc = (short)(u >>> UTC_SHIFT)) >= 0 && ac + pc > 1) {
993 long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK);
994 if (UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc))
995 return true; // no compensation needed
997 else if (tc + pc < MAX_ID) {
998 long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK);
999 if (UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
1001 return true; // create a replacement
1004 // try to back out on any failure and let caller retry
1005 } while (!UNSAFE.compareAndSwapInt(this, blockedCountOffset,
1006 b = blockedCount, b - 1));
1012 * Decrements blockedCount and increments active count
1014 private void postBlock() {
1016 do {} while (!UNSAFE.compareAndSwapLong(this, ctlOffset, // no mask
1017 c = ctl, c + AC_UNIT));
1019 do {} while (!UNSAFE.compareAndSwapInt(this, blockedCountOffset,
1020 b = blockedCount, b - 1));
1024 * Possibly blocks waiting for the given task to complete, or
1025 * cancels the task if terminating. Fails to wait if contended.
1027 * @param joinMe the task
1029 final void tryAwaitJoin(ForkJoinTask<?> joinMe) {
1031 Thread.interrupted(); // clear interrupts before checking termination
1032 if (joinMe.status >= 0) {
1033 if (tryPreBlock()) {
1034 joinMe.tryAwaitDone(0L);
1037 else if ((ctl & STOP_BIT) != 0L)
1038 joinMe.cancelIgnoringExceptions();
1043 * Possibly blocks the given worker waiting for joinMe to
1044 * complete or timeout
1046 * @param joinMe the task
1047 * @param millis the wait time for underlying Object.wait
1049 final void timedAwaitJoin(ForkJoinTask<?> joinMe, long nanos) {
1050 while (joinMe.status >= 0) {
1051 Thread.interrupted();
1052 if ((ctl & STOP_BIT) != 0L) {
1053 joinMe.cancelIgnoringExceptions();
1056 if (tryPreBlock()) {
1057 long last = System.nanoTime();
1058 while (joinMe.status >= 0) {
1059 long millis = TimeUnit.NANOSECONDS.toMillis(nanos);
1062 joinMe.tryAwaitDone(millis);
1063 if (joinMe.status < 0)
1065 if ((ctl & STOP_BIT) != 0L) {
1066 joinMe.cancelIgnoringExceptions();
1069 long now = System.nanoTime();
1070 nanos -= now - last;
1080 * If necessary, compensates for blocker, and blocks
1082 private void awaitBlocker(ManagedBlocker blocker)
1083 throws InterruptedException {
1084 while (!blocker.isReleasable()) {
1085 if (tryPreBlock()) {
1087 do {} while (!blocker.isReleasable() && !blocker.block());
1096 // Creating, registering and deregistring workers
1099 * Tries to create and start a worker; minimally rolls back counts
1102 private void addWorker() {
1103 Throwable ex = null;
1104 ForkJoinWorkerThread t = null;
1106 t = factory.newThread(this);
1107 } catch (Throwable e) {
1110 if (t == null) { // null or exceptional factory return
1111 long c; // adjust counts
1112 do {} while (!UNSAFE.compareAndSwapLong
1113 (this, ctlOffset, c = ctl,
1114 (((c - AC_UNIT) & AC_MASK) |
1115 ((c - TC_UNIT) & TC_MASK) |
1116 (c & ~(AC_MASK|TC_MASK)))));
1117 // Propagate exception if originating from an external caller
1118 if (!tryTerminate(false) && ex != null &&
1119 !(Thread.currentThread() instanceof ForkJoinWorkerThread))
1120 UNSAFE.throwException(ex);
1127 * Callback from ForkJoinWorkerThread constructor to assign a
1130 final String nextWorkerName() {
1132 if (UNSAFE.compareAndSwapInt(this, nextWorkerNumberOffset,
1133 n = nextWorkerNumber, ++n))
1134 return workerNamePrefix + n;
1139 * Callback from ForkJoinWorkerThread constructor to
1140 * determine its poolIndex and record in workers array.
1142 * @param w the worker
1143 * @return the worker's pool index
1145 final int registerWorker(ForkJoinWorkerThread w) {
1147 * In the typical case, a new worker acquires the lock, uses
1148 * next available index and returns quickly. Since we should
1149 * not block callers (ultimately from signalWork or
1150 * tryPreBlock) waiting for the lock needed to do this, we
1151 * instead help release other workers while waiting for the
1155 ForkJoinWorkerThread[] ws;
1156 if (((g = scanGuard) & SG_UNIT) == 0 &&
1157 UNSAFE.compareAndSwapInt(this, scanGuardOffset,
1159 int k = nextWorkerIndex;
1161 if ((ws = workers) != null) { // ignore on shutdown
1163 if (k < 0 || k >= n || ws[k] != null) {
1164 for (k = 0; k < n && ws[k] != null; ++k)
1167 ws = workers = Arrays.copyOf(ws, n << 1);
1170 nextWorkerIndex = k + 1;
1172 g = (k > m) ? ((m << 1) + 1) & SMASK : g + (SG_UNIT<<1);
1179 else if ((ws = workers) != null) { // help release others
1180 for (ForkJoinWorkerThread u : ws) {
1181 if (u != null && u.queueBase != u.queueTop) {
1182 if (tryReleaseWaiter())
1191 * Final callback from terminating worker. Removes record of
1192 * worker from array, and adjusts counts. If pool is shutting
1193 * down, tries to complete termination.
1195 * @param w the worker
1197 final void deregisterWorker(ForkJoinWorkerThread w, Throwable ex) {
1198 int idx = w.poolIndex;
1199 int sc = w.stealCount;
1201 // Remove from array, adjust worker counts and collect steal count.
1202 // We can intermix failed removes or adjusts with steal updates
1206 if (steps == 0 && ((g = scanGuard) & SG_UNIT) == 0 &&
1207 UNSAFE.compareAndSwapInt(this, scanGuardOffset,
1209 ForkJoinWorkerThread[] ws = workers;
1210 if (ws != null && idx >= 0 &&
1211 idx < ws.length && ws[idx] == w)
1212 ws[idx] = null; // verify
1213 nextWorkerIndex = idx;
1214 scanGuard = g + SG_UNIT;
1218 UNSAFE.compareAndSwapLong(this, ctlOffset, c = ctl,
1219 (((c - AC_UNIT) & AC_MASK) |
1220 ((c - TC_UNIT) & TC_MASK) |
1221 (c & ~(AC_MASK|TC_MASK)))))
1224 UNSAFE.compareAndSwapLong(this, stealCountOffset,
1225 s = stealCount, s + sc))
1227 } while (steps != 2 || sc != 0);
1228 if (!tryTerminate(false)) {
1229 if (ex != null) // possibly replace if died abnormally
1236 // Shutdown and termination
1239 * Possibly initiates and/or completes termination.
1241 * @param now if true, unconditionally terminate, else only
1242 * if shutdown and empty queue and no active workers
1243 * @return true if now terminating or terminated
1245 private boolean tryTerminate(boolean now) {
1247 while (((c = ctl) & STOP_BIT) == 0) {
1249 if ((int)(c >> AC_SHIFT) != -parallelism)
1251 if (!shutdown || blockedCount != 0 || quiescerCount != 0 ||
1252 queueBase != queueTop) {
1253 if (ctl == c) // staleness check
1258 if (UNSAFE.compareAndSwapLong(this, ctlOffset, c, c | STOP_BIT))
1261 if ((short)(c >>> TC_SHIFT) == -parallelism) { // signal when 0 workers
1262 final ReentrantLock lock = this.submissionLock;
1265 termination.signalAll();
1274 * Runs up to three passes through workers: (0) Setting
1275 * termination status for each worker, followed by wakeups up to
1276 * queued workers; (1) helping cancel tasks; (2) interrupting
1277 * lagging threads (likely in external tasks, but possibly also
1278 * blocked in joins). Each pass repeats previous steps because of
1279 * potential lagging thread creation.
1281 private void startTerminating() {
1282 cancelSubmissions();
1283 for (int pass = 0; pass < 3; ++pass) {
1284 ForkJoinWorkerThread[] ws = workers;
1286 for (ForkJoinWorkerThread w : ws) {
1291 if (pass > 1 && !w.isInterrupted()) {
1294 } catch (SecurityException ignore) {
1306 * Polls and cancels all submissions. Called only during termination.
1308 private void cancelSubmissions() {
1309 while (queueBase != queueTop) {
1310 ForkJoinTask<?> task = pollSubmission();
1314 } catch (Throwable ignore) {
1321 * Tries to set the termination status of waiting workers, and
1322 * then wakes them up (after which they will terminate).
1324 private void terminateWaiters() {
1325 ForkJoinWorkerThread[] ws = workers;
1327 ForkJoinWorkerThread w; long c; int i, e;
1329 while ((i = ~(e = (int)(c = ctl)) & SMASK) < n &&
1330 (w = ws[i]) != null && w.eventCount == (e & E_MASK)) {
1331 if (UNSAFE.compareAndSwapLong(this, ctlOffset, c,
1332 (long)(w.nextWait & E_MASK) |
1333 ((c + AC_UNIT) & AC_MASK) |
1334 (c & (TC_MASK|STOP_BIT)))) {
1336 w.eventCount = e + EC_UNIT;
1344 // misc ForkJoinWorkerThread support
1347 * Increment or decrement quiescerCount. Needed only to prevent
1348 * triggering shutdown if a worker is transiently inactive while
1349 * checking quiescence.
1351 * @param delta 1 for increment, -1 for decrement
1353 final void addQuiescerCount(int delta) {
1355 do {} while (!UNSAFE.compareAndSwapInt(this, quiescerCountOffset,
1356 c = quiescerCount, c + delta));
1360 * Directly increment or decrement active count without
1361 * queuing. This method is used to transiently assert inactivation
1362 * while checking quiescence.
1364 * @param delta 1 for increment, -1 for decrement
1366 final void addActiveCount(int delta) {
1367 long d = delta < 0 ? -AC_UNIT : AC_UNIT;
1369 do {} while (!UNSAFE.compareAndSwapLong(this, ctlOffset, c = ctl,
1370 ((c + d) & AC_MASK) |
1375 * Returns the approximate (non-atomic) number of idle threads per
1378 final int idlePerActive() {
1379 // Approximate at powers of two for small values, saturate past 4
1380 int p = parallelism;
1381 int a = p + (int)(ctl >> AC_SHIFT);
1382 return (a > (p >>>= 1) ? 0 :
1383 a > (p >>>= 1) ? 1 :
1384 a > (p >>>= 1) ? 2 :
1385 a > (p >>>= 1) ? 4 :
1394 * Creates a {@code ForkJoinPool} with parallelism equal to {@link
1395 * java.lang.Runtime#availableProcessors}, using the {@linkplain
1396 * #defaultForkJoinWorkerThreadFactory default thread factory},
1397 * no UncaughtExceptionHandler, and non-async LIFO processing mode.
1399 * @throws SecurityException if a security manager exists and
1400 * the caller is not permitted to modify threads
1401 * because it does not hold {@link
1402 * java.lang.RuntimePermission}{@code ("modifyThread")}
1404 public ForkJoinPool() {
1406 defaultForkJoinWorkerThreadFactory, null, false);
1410 * Creates a {@code ForkJoinPool} with the indicated parallelism
1411 * level, the {@linkplain
1412 * #defaultForkJoinWorkerThreadFactory default thread factory},
1413 * no UncaughtExceptionHandler, and non-async LIFO processing mode.
1415 * @param parallelism the parallelism level
1416 * @throws IllegalArgumentException if parallelism less than or
1417 * equal to zero, or greater than implementation limit
1418 * @throws SecurityException if a security manager exists and
1419 * the caller is not permitted to modify threads
1420 * because it does not hold {@link
1421 * java.lang.RuntimePermission}{@code ("modifyThread")}
1423 public ForkJoinPool(int parallelism) {
1424 this(parallelism, defaultForkJoinWorkerThreadFactory, null, false);
1428 * Creates a {@code ForkJoinPool} with the given parameters.
1430 * @param parallelism the parallelism level. For default value,
1431 * use {@link java.lang.Runtime#availableProcessors}.
1432 * @param factory the factory for creating new threads. For default value,
1433 * use {@link #defaultForkJoinWorkerThreadFactory}.
1434 * @param handler the handler for internal worker threads that
1435 * terminate due to unrecoverable errors encountered while executing
1436 * tasks. For default value, use {@code null}.
1437 * @param asyncMode if true,
1438 * establishes local first-in-first-out scheduling mode for forked
1439 * tasks that are never joined. This mode may be more appropriate
1440 * than default locally stack-based mode in applications in which
1441 * worker threads only process event-style asynchronous tasks.
1442 * For default value, use {@code false}.
1443 * @throws IllegalArgumentException if parallelism less than or
1444 * equal to zero, or greater than implementation limit
1445 * @throws NullPointerException if the factory is null
1446 * @throws SecurityException if a security manager exists and
1447 * the caller is not permitted to modify threads
1448 * because it does not hold {@link
1449 * java.lang.RuntimePermission}{@code ("modifyThread")}
1451 public ForkJoinPool(int parallelism,
1452 ForkJoinWorkerThreadFactory factory,
1453 Thread.UncaughtExceptionHandler handler,
1454 boolean asyncMode) {
1456 if (factory == null)
1457 throw new NullPointerException();
1458 if (parallelism <= 0 || parallelism > MAX_ID)
1459 throw new IllegalArgumentException();
1460 this.parallelism = parallelism;
1461 this.factory = factory;
1463 this.locallyFifo = asyncMode;
1464 long np = (long)(-parallelism); // offset ctl counts
1465 this.ctl = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK);
1466 this.submissionQueue = new ForkJoinTask<?>[INITIAL_QUEUE_CAPACITY];
1467 // initialize workers array with room for 2*parallelism if possible
1468 int n = parallelism << 1;
1471 else { // See Hackers Delight, sec 3.2, where n < (1 << 16)
1472 n |= n >>> 1; n |= n >>> 2; n |= n >>> 4; n |= n >>> 8;
1474 workers = new ForkJoinWorkerThread[n + 1];
1475 this.submissionLock = new ReentrantLock();
1476 this.termination = submissionLock.newCondition();
1477 StringBuilder sb = new StringBuilder("ForkJoinPool-");
1478 sb.append(poolNumberGenerator.incrementAndGet());
1479 sb.append("-worker-");
1480 this.workerNamePrefix = sb.toString();
1483 // Execution methods
1486 * Performs the given task, returning its result upon completion.
1487 * If the computation encounters an unchecked Exception or Error,
1488 * it is rethrown as the outcome of this invocation. Rethrown
1489 * exceptions behave in the same way as regular exceptions, but,
1490 * when possible, contain stack traces (as displayed for example
1491 * using {@code ex.printStackTrace()}) of both the current thread
1492 * as well as the thread actually encountering the exception;
1493 * minimally only the latter.
1495 * @param task the task
1496 * @return the task's result
1497 * @throws NullPointerException if the task is null
1498 * @throws RejectedExecutionException if the task cannot be
1499 * scheduled for execution
1501 public <T> T invoke(ForkJoinTask<T> task) {
1502 Thread t = Thread.currentThread();
1504 throw new NullPointerException();
1506 throw new RejectedExecutionException();
1507 if ((t instanceof ForkJoinWorkerThread) &&
1508 ((ForkJoinWorkerThread)t).pool == this)
1509 return task.invoke(); // bypass submit if in same pool
1511 addSubmission(task);
1517 * Unless terminating, forks task if within an ongoing FJ
1518 * computation in the current pool, else submits as external task.
1520 private <T> void forkOrSubmit(ForkJoinTask<T> task) {
1521 ForkJoinWorkerThread w;
1522 Thread t = Thread.currentThread();
1524 throw new RejectedExecutionException();
1525 if ((t instanceof ForkJoinWorkerThread) &&
1526 (w = (ForkJoinWorkerThread)t).pool == this)
1529 addSubmission(task);
1533 * Arranges for (asynchronous) execution of the given task.
1535 * @param task the task
1536 * @throws NullPointerException if the task is null
1537 * @throws RejectedExecutionException if the task cannot be
1538 * scheduled for execution
1540 public void execute(ForkJoinTask<?> task) {
1542 throw new NullPointerException();
1546 // AbstractExecutorService methods
1549 * @throws NullPointerException if the task is null
1550 * @throws RejectedExecutionException if the task cannot be
1551 * scheduled for execution
1553 public void execute(Runnable task) {
1555 throw new NullPointerException();
1556 ForkJoinTask<?> job;
1557 if (task instanceof ForkJoinTask<?>) // avoid re-wrap
1558 job = (ForkJoinTask<?>) task;
1560 job = ForkJoinTask.adapt(task, null);
1565 * Submits a ForkJoinTask for execution.
1567 * @param task the task to submit
1569 * @throws NullPointerException if the task is null
1570 * @throws RejectedExecutionException if the task cannot be
1571 * scheduled for execution
1573 public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) {
1575 throw new NullPointerException();
1581 * @throws NullPointerException if the task is null
1582 * @throws RejectedExecutionException if the task cannot be
1583 * scheduled for execution
1585 public <T> ForkJoinTask<T> submit(Callable<T> task) {
1587 throw new NullPointerException();
1588 ForkJoinTask<T> job = ForkJoinTask.adapt(task);
1594 * @throws NullPointerException if the task is null
1595 * @throws RejectedExecutionException if the task cannot be
1596 * scheduled for execution
1598 public <T> ForkJoinTask<T> submit(Runnable task, T result) {
1600 throw new NullPointerException();
1601 ForkJoinTask<T> job = ForkJoinTask.adapt(task, result);
1607 * @throws NullPointerException if the task is null
1608 * @throws RejectedExecutionException if the task cannot be
1609 * scheduled for execution
1611 public ForkJoinTask<?> submit(Runnable task) {
1613 throw new NullPointerException();
1614 ForkJoinTask<?> job;
1615 if (task instanceof ForkJoinTask<?>) // avoid re-wrap
1616 job = (ForkJoinTask<?>) task;
1618 job = ForkJoinTask.adapt(task, null);
1624 * @throws NullPointerException {@inheritDoc}
1625 * @throws RejectedExecutionException {@inheritDoc}
1627 public <T> List<Future<T>> invokeAll(Collection<? extends Callable<T>> tasks) {
1628 ArrayList<ForkJoinTask<T>> forkJoinTasks =
1629 new ArrayList<ForkJoinTask<T>>(tasks.size());
1630 for (Callable<T> task : tasks)
1631 forkJoinTasks.add(ForkJoinTask.adapt(task));
1632 invoke(new InvokeAll<T>(forkJoinTasks));
1634 @SuppressWarnings({"unchecked", "rawtypes"})
1635 List<Future<T>> futures = (List<Future<T>>) (List) forkJoinTasks;
1639 static final class InvokeAll<T> extends RecursiveAction {
1640 final ArrayList<ForkJoinTask<T>> tasks;
1641 InvokeAll(ArrayList<ForkJoinTask<T>> tasks) { this.tasks = tasks; }
1642 public void compute() {
1643 try { invokeAll(tasks); }
1644 catch (Exception ignore) {}
1646 private static final long serialVersionUID = -7914297376763021607L;
1650 * Returns the factory used for constructing new workers.
1652 * @return the factory used for constructing new workers
1654 public ForkJoinWorkerThreadFactory getFactory() {
1659 * Returns the handler for internal worker threads that terminate
1660 * due to unrecoverable errors encountered while executing tasks.
1662 * @return the handler, or {@code null} if none
1664 public Thread.UncaughtExceptionHandler getUncaughtExceptionHandler() {
1669 * Returns the targeted parallelism level of this pool.
1671 * @return the targeted parallelism level of this pool
1673 public int getParallelism() {
1678 * Returns the number of worker threads that have started but not
1679 * yet terminated. The result returned by this method may differ
1680 * from {@link #getParallelism} when threads are created to
1681 * maintain parallelism when others are cooperatively blocked.
1683 * @return the number of worker threads
1685 public int getPoolSize() {
1686 return parallelism + (short)(ctl >>> TC_SHIFT);
1690 * Returns {@code true} if this pool uses local first-in-first-out
1691 * scheduling mode for forked tasks that are never joined.
1693 * @return {@code true} if this pool uses async mode
1695 public boolean getAsyncMode() {
1700 * Returns an estimate of the number of worker threads that are
1701 * not blocked waiting to join tasks or for other managed
1702 * synchronization. This method may overestimate the
1703 * number of running threads.
1705 * @return the number of worker threads
1707 public int getRunningThreadCount() {
1708 int r = parallelism + (int)(ctl >> AC_SHIFT);
1709 return (r <= 0) ? 0 : r; // suppress momentarily negative values
1713 * Returns an estimate of the number of threads that are currently
1714 * stealing or executing tasks. This method may overestimate the
1715 * number of active threads.
1717 * @return the number of active threads
1719 public int getActiveThreadCount() {
1720 int r = parallelism + (int)(ctl >> AC_SHIFT) + blockedCount;
1721 return (r <= 0) ? 0 : r; // suppress momentarily negative values
1725 * Returns {@code true} if all worker threads are currently idle.
1726 * An idle worker is one that cannot obtain a task to execute
1727 * because none are available to steal from other threads, and
1728 * there are no pending submissions to the pool. This method is
1729 * conservative; it might not return {@code true} immediately upon
1730 * idleness of all threads, but will eventually become true if
1731 * threads remain inactive.
1733 * @return {@code true} if all threads are currently idle
1735 public boolean isQuiescent() {
1736 return parallelism + (int)(ctl >> AC_SHIFT) + blockedCount == 0;
1740 * Returns an estimate of the total number of tasks stolen from
1741 * one thread's work queue by another. The reported value
1742 * underestimates the actual total number of steals when the pool
1743 * is not quiescent. This value may be useful for monitoring and
1744 * tuning fork/join programs: in general, steal counts should be
1745 * high enough to keep threads busy, but low enough to avoid
1746 * overhead and contention across threads.
1748 * @return the number of steals
1750 public long getStealCount() {
1755 * Returns an estimate of the total number of tasks currently held
1756 * in queues by worker threads (but not including tasks submitted
1757 * to the pool that have not begun executing). This value is only
1758 * an approximation, obtained by iterating across all threads in
1759 * the pool. This method may be useful for tuning task
1762 * @return the number of queued tasks
1764 public long getQueuedTaskCount() {
1766 ForkJoinWorkerThread[] ws;
1767 if ((short)(ctl >>> TC_SHIFT) > -parallelism &&
1768 (ws = workers) != null) {
1769 for (ForkJoinWorkerThread w : ws)
1771 count -= w.queueBase - w.queueTop; // must read base first
1777 * Returns an estimate of the number of tasks submitted to this
1778 * pool that have not yet begun executing. This method may take
1779 * time proportional to the number of submissions.
1781 * @return the number of queued submissions
1783 public int getQueuedSubmissionCount() {
1784 return -queueBase + queueTop;
1788 * Returns {@code true} if there are any tasks submitted to this
1789 * pool that have not yet begun executing.
1791 * @return {@code true} if there are any queued submissions
1793 public boolean hasQueuedSubmissions() {
1794 return queueBase != queueTop;
1798 * Removes and returns the next unexecuted submission if one is
1799 * available. This method may be useful in extensions to this
1800 * class that re-assign work in systems with multiple pools.
1802 * @return the next submission, or {@code null} if none
1804 protected ForkJoinTask<?> pollSubmission() {
1805 ForkJoinTask<?> t; ForkJoinTask<?>[] q; int b, i;
1806 while ((b = queueBase) != queueTop &&
1807 (q = submissionQueue) != null &&
1808 (i = (q.length - 1) & b) >= 0) {
1809 long u = (i << ASHIFT) + ABASE;
1810 if ((t = q[i]) != null &&
1812 UNSAFE.compareAndSwapObject(q, u, t, null)) {
1821 * Removes all available unexecuted submitted and forked tasks
1822 * from scheduling queues and adds them to the given collection,
1823 * without altering their execution status. These may include
1824 * artificially generated or wrapped tasks. This method is
1825 * designed to be invoked only when the pool is known to be
1826 * quiescent. Invocations at other times may not remove all
1827 * tasks. A failure encountered while attempting to add elements
1828 * to collection {@code c} may result in elements being in
1829 * neither, either or both collections when the associated
1830 * exception is thrown. The behavior of this operation is
1831 * undefined if the specified collection is modified while the
1832 * operation is in progress.
1834 * @param c the collection to transfer elements into
1835 * @return the number of elements transferred
1837 protected int drainTasksTo(Collection<? super ForkJoinTask<?>> c) {
1839 while (queueBase != queueTop) {
1840 ForkJoinTask<?> t = pollSubmission();
1846 ForkJoinWorkerThread[] ws;
1847 if ((short)(ctl >>> TC_SHIFT) > -parallelism &&
1848 (ws = workers) != null) {
1849 for (ForkJoinWorkerThread w : ws)
1851 count += w.drainTasksTo(c);
1857 * Returns a string identifying this pool, as well as its state,
1858 * including indications of run state, parallelism level, and
1859 * worker and task counts.
1861 * @return a string identifying this pool, as well as its state
1863 public String toString() {
1864 long st = getStealCount();
1865 long qt = getQueuedTaskCount();
1866 long qs = getQueuedSubmissionCount();
1867 int pc = parallelism;
1869 int tc = pc + (short)(c >>> TC_SHIFT);
1870 int rc = pc + (int)(c >> AC_SHIFT);
1871 if (rc < 0) // ignore transient negative
1873 int ac = rc + blockedCount;
1875 if ((c & STOP_BIT) != 0)
1876 level = (tc == 0) ? "Terminated" : "Terminating";
1878 level = shutdown ? "Shutting down" : "Running";
1879 return super.toString() +
1881 ", parallelism = " + pc +
1883 ", active = " + ac +
1884 ", running = " + rc +
1885 ", steals = " + st +
1887 ", submissions = " + qs +
1892 * Initiates an orderly shutdown in which previously submitted
1893 * tasks are executed, but no new tasks will be accepted.
1894 * Invocation has no additional effect if already shut down.
1895 * Tasks that are in the process of being submitted concurrently
1896 * during the course of this method may or may not be rejected.
1898 * @throws SecurityException if a security manager exists and
1899 * the caller is not permitted to modify threads
1900 * because it does not hold {@link
1901 * java.lang.RuntimePermission}{@code ("modifyThread")}
1903 public void shutdown() {
1906 tryTerminate(false);
1910 * Attempts to cancel and/or stop all tasks, and reject all
1911 * subsequently submitted tasks. Tasks that are in the process of
1912 * being submitted or executed concurrently during the course of
1913 * this method may or may not be rejected. This method cancels
1914 * both existing and unexecuted tasks, in order to permit
1915 * termination in the presence of task dependencies. So the method
1916 * always returns an empty list (unlike the case for some other
1919 * @return an empty list
1920 * @throws SecurityException if a security manager exists and
1921 * the caller is not permitted to modify threads
1922 * because it does not hold {@link
1923 * java.lang.RuntimePermission}{@code ("modifyThread")}
1925 public List<Runnable> shutdownNow() {
1929 return Collections.emptyList();
1933 * Returns {@code true} if all tasks have completed following shut down.
1935 * @return {@code true} if all tasks have completed following shut down
1937 public boolean isTerminated() {
1939 return ((c & STOP_BIT) != 0L &&
1940 (short)(c >>> TC_SHIFT) == -parallelism);
1944 * Returns {@code true} if the process of termination has
1945 * commenced but not yet completed. This method may be useful for
1946 * debugging. A return of {@code true} reported a sufficient
1947 * period after shutdown may indicate that submitted tasks have
1948 * ignored or suppressed interruption, or are waiting for IO,
1949 * causing this executor not to properly terminate. (See the
1950 * advisory notes for class {@link ForkJoinTask} stating that
1951 * tasks should not normally entail blocking operations. But if
1952 * they do, they must abort them on interrupt.)
1954 * @return {@code true} if terminating but not yet terminated
1956 public boolean isTerminating() {
1958 return ((c & STOP_BIT) != 0L &&
1959 (short)(c >>> TC_SHIFT) != -parallelism);
1963 * Returns true if terminating or terminated. Used by ForkJoinWorkerThread.
1965 final boolean isAtLeastTerminating() {
1966 return (ctl & STOP_BIT) != 0L;
1970 * Returns {@code true} if this pool has been shut down.
1972 * @return {@code true} if this pool has been shut down
1974 public boolean isShutdown() {
1979 * Blocks until all tasks have completed execution after a shutdown
1980 * request, or the timeout occurs, or the current thread is
1981 * interrupted, whichever happens first.
1983 * @param timeout the maximum time to wait
1984 * @param unit the time unit of the timeout argument
1985 * @return {@code true} if this executor terminated and
1986 * {@code false} if the timeout elapsed before termination
1987 * @throws InterruptedException if interrupted while waiting
1989 public boolean awaitTermination(long timeout, TimeUnit unit)
1990 throws InterruptedException {
1991 long nanos = unit.toNanos(timeout);
1992 final ReentrantLock lock = this.submissionLock;
2000 nanos = termination.awaitNanos(nanos);
2008 * Interface for extending managed parallelism for tasks running
2009 * in {@link ForkJoinPool}s.
2011 * <p>A {@code ManagedBlocker} provides two methods. Method
2012 * {@code isReleasable} must return {@code true} if blocking is
2013 * not necessary. Method {@code block} blocks the current thread
2014 * if necessary (perhaps internally invoking {@code isReleasable}
2015 * before actually blocking). These actions are performed by any
2016 * thread invoking {@link ForkJoinPool#managedBlock}. The
2017 * unusual methods in this API accommodate synchronizers that may,
2018 * but don't usually, block for long periods. Similarly, they
2019 * allow more efficient internal handling of cases in which
2020 * additional workers may be, but usually are not, needed to
2021 * ensure sufficient parallelism. Toward this end,
2022 * implementations of method {@code isReleasable} must be amenable
2023 * to repeated invocation.
2025 * <p>For example, here is a ManagedBlocker based on a
2028 * class ManagedLocker implements ManagedBlocker {
2029 * final ReentrantLock lock;
2030 * boolean hasLock = false;
2031 * ManagedLocker(ReentrantLock lock) { this.lock = lock; }
2032 * public boolean block() {
2037 * public boolean isReleasable() {
2038 * return hasLock || (hasLock = lock.tryLock());
2042 * <p>Here is a class that possibly blocks waiting for an
2043 * item on a given queue:
2045 * class QueueTaker<E> implements ManagedBlocker {
2046 * final BlockingQueue<E> queue;
2047 * volatile E item = null;
2048 * QueueTaker(BlockingQueue<E> q) { this.queue = q; }
2049 * public boolean block() throws InterruptedException {
2051 * item = queue.take();
2054 * public boolean isReleasable() {
2055 * return item != null || (item = queue.poll()) != null;
2057 * public E getItem() { // call after pool.managedBlock completes
2062 public static interface ManagedBlocker {
2064 * Possibly blocks the current thread, for example waiting for
2065 * a lock or condition.
2067 * @return {@code true} if no additional blocking is necessary
2068 * (i.e., if isReleasable would return true)
2069 * @throws InterruptedException if interrupted while waiting
2070 * (the method is not required to do so, but is allowed to)
2072 boolean block() throws InterruptedException;
2075 * Returns {@code true} if blocking is unnecessary.
2077 boolean isReleasable();
2081 * Blocks in accord with the given blocker. If the current thread
2082 * is a {@link ForkJoinWorkerThread}, this method possibly
2083 * arranges for a spare thread to be activated if necessary to
2084 * ensure sufficient parallelism while the current thread is blocked.
2086 * <p>If the caller is not a {@link ForkJoinTask}, this method is
2087 * behaviorally equivalent to
2089 * while (!blocker.isReleasable())
2090 * if (blocker.block())
2094 * If the caller is a {@code ForkJoinTask}, then the pool may
2095 * first be expanded to ensure parallelism, and later adjusted.
2097 * @param blocker the blocker
2098 * @throws InterruptedException if blocker.block did so
2100 public static void managedBlock(ManagedBlocker blocker)
2101 throws InterruptedException {
2102 Thread t = Thread.currentThread();
2103 if (t instanceof ForkJoinWorkerThread) {
2104 ForkJoinWorkerThread w = (ForkJoinWorkerThread) t;
2105 w.pool.awaitBlocker(blocker);
2108 do {} while (!blocker.isReleasable() && !blocker.block());
2112 // AbstractExecutorService overrides. These rely on undocumented
2113 // fact that ForkJoinTask.adapt returns ForkJoinTasks that also
2114 // implement RunnableFuture.
2116 protected <T> RunnableFuture<T> newTaskFor(Runnable runnable, T value) {
2117 return (RunnableFuture<T>) ForkJoinTask.adapt(runnable, value);
2120 protected <T> RunnableFuture<T> newTaskFor(Callable<T> callable) {
2121 return (RunnableFuture<T>) ForkJoinTask.adapt(callable);
2125 private static final Unsafe UNSAFE;
2126 private static final long ctlOffset;
2127 private static final long stealCountOffset;
2128 private static final long blockedCountOffset;
2129 private static final long quiescerCountOffset;
2130 private static final long scanGuardOffset;
2131 private static final long nextWorkerNumberOffset;
2132 private static final long ABASE;
2133 private static final int ASHIFT;
2136 poolNumberGenerator = new AtomicInteger();
2137 workerSeedGenerator = new Random();
2138 defaultForkJoinWorkerThreadFactory =
2139 new DefaultForkJoinWorkerThreadFactory();
2142 UNSAFE = Unsafe.getUnsafe();
2143 Class k = ForkJoinPool.class;
2144 ctlOffset = UNSAFE.objectFieldOffset
2145 (k.getDeclaredField("ctl"));
2146 stealCountOffset = UNSAFE.objectFieldOffset
2147 (k.getDeclaredField("stealCount"));
2148 blockedCountOffset = UNSAFE.objectFieldOffset
2149 (k.getDeclaredField("blockedCount"));
2150 quiescerCountOffset = UNSAFE.objectFieldOffset
2151 (k.getDeclaredField("quiescerCount"));
2152 scanGuardOffset = UNSAFE.objectFieldOffset
2153 (k.getDeclaredField("scanGuard"));
2154 nextWorkerNumberOffset = UNSAFE.objectFieldOffset
2155 (k.getDeclaredField("nextWorkerNumber"));
2156 Class a = ForkJoinTask[].class;
2157 ABASE = UNSAFE.arrayBaseOffset(a);
2158 s = UNSAFE.arrayIndexScale(a);
2159 } catch (Exception e) {
2162 if ((s & (s-1)) != 0)
2163 throw new Error("data type scale not a power of two");
2164 ASHIFT = 31 - Integer.numberOfLeadingZeros(s);