1 /*
2 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
3 *
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.
9 *
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).
15 *
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.
19 *
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
22 * questions.
23 */
24
25 /*
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
29 * file:
30 *
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/
34 */
35
36 package java.util.concurrent;
37
38 import java.lang.Thread.UncaughtExceptionHandler;
39 import java.lang.reflect.Field;
40 import java.util.ArrayList;
41 import java.util.Collection;
42 import java.util.Collections;
43 import java.util.List;
44 import java.util.Objects;
45 import java.util.function.Consumer;
46 import java.util.function.Predicate;
47 import java.util.concurrent.CountDownLatch;
48 import java.util.concurrent.locks.LockSupport;
49 import jdk.internal.access.JavaLangAccess;
50 import jdk.internal.access.JavaUtilConcurrentFJPAccess;
51 import jdk.internal.access.SharedSecrets;
52 import jdk.internal.misc.Unsafe;
53 import jdk.internal.vm.SharedThreadContainer;
54 import static java.util.concurrent.DelayScheduler.ScheduledForkJoinTask;
55
56 /**
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.
61 *
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 tasks submitted to the pool and/or created by other active
66 * tasks (eventually blocking waiting for work if none exist). This
67 * enables efficient processing when most tasks spawn other subtasks
68 * (as do most {@code ForkJoinTask}s), as well as when many small
69 * tasks are submitted to the pool from external clients. Especially
70 * when setting <em>asyncMode</em> to true in constructors, {@code
71 * ForkJoinPool}s may also be appropriate for use with event-style
72 * tasks that are never joined. All worker threads are initialized
73 * with {@link Thread#isDaemon} set {@code true}.
74 *
75 * <p>A static {@link #commonPool()} is available and appropriate for
76 * most applications. The common pool is used by any ForkJoinTask that
77 * is not explicitly submitted to a specified pool. Using the common
78 * pool normally reduces resource usage (its threads are slowly
79 * reclaimed during periods of non-use, and reinstated upon subsequent
80 * use).
81 *
82 * <p>For applications that require separate or custom pools, a {@code
83 * ForkJoinPool} may be constructed with a given target parallelism
84 * level; by default, equal to the number of available processors.
85 * The pool attempts to maintain enough active (or available) threads
86 * by dynamically adding, suspending, or resuming internal worker
87 * threads, even if some tasks are stalled waiting to join others.
88 * However, no such adjustments are guaranteed in the face of blocked
89 * I/O or other unmanaged synchronization. The nested {@link
90 * ManagedBlocker} interface enables extension of the kinds of
91 * synchronization accommodated. The default policies may be
92 * overridden using a constructor with parameters corresponding to
93 * those documented in class {@link ThreadPoolExecutor}.
94 *
95 * <p>In addition to execution and lifecycle control methods, this
96 * class provides status check methods (for example
97 * {@link #getStealCount}) that are intended to aid in developing,
98 * tuning, and monitoring fork/join applications. Also, method
99 * {@link #toString} returns indications of pool state in a
100 * convenient form for informal monitoring.
101 *
102 * <p>As is the case with other ExecutorServices, there are three
103 * main task execution methods summarized in the following table.
104 * These are designed to be used primarily by clients not already
105 * engaged in fork/join computations in the current pool. The main
106 * forms of these methods accept instances of {@code ForkJoinTask},
107 * but overloaded forms also allow mixed execution of plain {@code
108 * Runnable}- or {@code Callable}- based activities as well. However,
109 * tasks that are already executing in a pool should normally instead
110 * use the within-computation forms listed in the table unless using
111 * async event-style tasks that are not usually joined, in which case
112 * there is little difference among choice of methods.
113 *
114 * <table class="plain">
115 * <caption>Summary of task execution methods</caption>
116 * <tr>
117 * <td></td>
118 * <th scope="col"> Call from non-fork/join clients</th>
119 * <th scope="col"> Call from within fork/join computations</th>
120 * </tr>
121 * <tr>
122 * <th scope="row" style="text-align:left"> Arrange async execution</th>
123 * <td> {@link #execute(ForkJoinTask)}</td>
124 * <td> {@link ForkJoinTask#fork}</td>
125 * </tr>
126 * <tr>
127 * <th scope="row" style="text-align:left"> Await and obtain result</th>
128 * <td> {@link #invoke(ForkJoinTask)}</td>
129 * <td> {@link ForkJoinTask#invoke}</td>
130 * </tr>
131 * <tr>
132 * <th scope="row" style="text-align:left"> Arrange exec and obtain Future</th>
133 * <td> {@link #submit(ForkJoinTask)}</td>
134 * <td> {@link ForkJoinTask#fork} (ForkJoinTasks <em>are</em> Futures)</td>
135 * </tr>
136 * </table>
137 *
138 * <p>Additionally, this class supports {@link
139 * ScheduledExecutorService} methods to delay or periodically execute
140 * tasks, as well as method {@link #submitWithTimeout} to cancel tasks
141 * that take too long. The scheduled functions or actions may create
142 * and invoke other {@linkplain ForkJoinTask ForkJoinTasks}. Delayed
143 * actions become enabled for execution and behave as ordinary submitted
144 * tasks when their delays elapse. Scheduling methods return
145 * {@linkplain ForkJoinTask ForkJoinTasks} that implement the {@link
146 * ScheduledFuture} interface. Resource exhaustion encountered after
147 * initial submission results in task cancellation. When time-based
148 * methods are used, shutdown policies match the default policies of
149 * class {@link ScheduledThreadPoolExecutor}: upon {@link #shutdown},
150 * existing periodic tasks will not re-execute, and the pool
151 * terminates when quiescent and existing delayed tasks
152 * complete. Method {@link #cancelDelayedTasksOnShutdown} may be used
153 * to disable all delayed tasks upon shutdown, and method {@link
154 * #shutdownNow} may be used to instead unconditionally initiate pool
155 * termination. Monitoring methods such as {@link #getQueuedTaskCount}
156 * do not include scheduled tasks that are not yet enabled for execution,
157 * which are reported separately by method {@link
158 * #getDelayedTaskCount}.
159 *
160 * <p>The parameters used to construct the common pool may be controlled by
161 * setting the following {@linkplain System#getProperty system properties}:
162 * <ul>
163 * <li>{@systemProperty java.util.concurrent.ForkJoinPool.common.parallelism}
164 * - the parallelism level, a non-negative integer. Usage is discouraged.
165 * Use {@link #setParallelism} instead.
166 * <li>{@systemProperty java.util.concurrent.ForkJoinPool.common.threadFactory}
167 * - the class name of a {@link ForkJoinWorkerThreadFactory}.
168 * The {@linkplain ClassLoader#getSystemClassLoader() system class loader}
169 * is used to load this class.
170 * <li>{@systemProperty java.util.concurrent.ForkJoinPool.common.exceptionHandler}
171 * - the class name of a {@link UncaughtExceptionHandler}.
172 * The {@linkplain ClassLoader#getSystemClassLoader() system class loader}
173 * is used to load this class.
174 * <li>{@systemProperty java.util.concurrent.ForkJoinPool.common.maximumSpares}
175 * - the maximum number of allowed extra threads to maintain target
176 * parallelism (default 256).
177 * </ul>
178 * If no thread factory is supplied via a system property, then the
179 * common pool uses a factory that uses the system class loader as the
180 * {@linkplain Thread#getContextClassLoader() thread context class loader}.
181 *
182 * Upon any error in establishing these settings, default parameters
183 * are used. It is possible to disable use of threads by using a
184 * factory that may return {@code null}, in which case some tasks may
185 * never execute. While possible, it is strongly discouraged to set
186 * the parallelism property to zero, which may be internally
187 * overridden in the presence of intrinsically async tasks.
188 *
189 * @implNote This implementation restricts the maximum number of
190 * running threads to 32767. Attempts to create pools with greater
191 * than the maximum number result in {@code
192 * IllegalArgumentException}. Also, this implementation rejects
193 * submitted tasks (that is, by throwing {@link
194 * RejectedExecutionException}) only when the pool is shut down or
195 * internal resources have been exhausted.
196 *
197 * @since 1.7
198 * @author Doug Lea
199 */
200 public class ForkJoinPool extends AbstractExecutorService
201 implements ScheduledExecutorService {
202
203 /*
204 * Implementation Overview
205 *
206 * This class and its nested classes provide the main
207 * functionality and control for a set of worker threads. Because
208 * most internal methods and nested classes are interrelated,
209 * their main rationale and descriptions are presented here;
210 * individual methods and nested classes contain only brief
211 * comments about details. Broadly: submissions from non-FJ
212 * threads enter into submission queues. Workers take these tasks
213 * and typically split them into subtasks that may be stolen by
214 * other workers. Work-stealing based on randomized scans
215 * generally leads to better throughput than "work dealing" in
216 * which producers assign tasks to idle threads, in part because
217 * threads that have finished other tasks before the signalled
218 * thread wakes up (which can be a long time) can take the task
219 * instead. Preference rules give first priority to processing
220 * tasks from their own queues (LIFO or FIFO, depending on mode),
221 * then to randomized FIFO steals of tasks in other queues.
222 *
223 * This framework began as vehicle for supporting structured
224 * parallelism using work-stealing, designed to work best when
225 * tasks are dag-structured (wrt completion dependencies), nested
226 * (generated using recursion or completions), of reasonable
227 * granularity, independent (wrt memory and resources) and where
228 * callers participate in task execution. These are properties
229 * that anyone aiming for efficient parallel multicore execution
230 * should design for. Over time, the scalability advantages of
231 * this framework led to extensions to better support more diverse
232 * usage contexts, amounting to weakenings or violations of each
233 * of these properties. Accommodating them may compromise
234 * performance, but mechanics discussed below include tradeoffs
235 * attempting to arrange that no single performance issue dominates.
236 *
237 * Here's a brief history of major revisions, each also with other
238 * minor features and changes.
239 *
240 * 1. Only handle recursively structured computational tasks
241 * 2. Async (FIFO) mode and striped submission queues
242 * 3. Completion-based tasks (mainly CountedCompleters)
243 * 4. CommonPool and parallelStream support
244 * 5. InterruptibleTasks for externally submitted tasks
245 * 6. Support ScheduledExecutorService methods
246 *
247 * Most changes involve adaptions of base algorithms using
248 * combinations of static and dynamic bitwise mode settings (both
249 * here and in ForkJoinTask), and subclassing of ForkJoinTask.
250 * There are a fair number of odd code constructions and design
251 * decisions for components that reside at the edge of Java vs JVM
252 * functionality.
253 *
254 * WorkQueues
255 * ==========
256 *
257 * Most operations occur within work-stealing queues (in nested
258 * class WorkQueue). These are special forms of Deques that
259 * support only three of the four possible end-operations -- push,
260 * pop, and poll (aka steal), under the further constraints that
261 * push and pop are called only from the owning thread (or, as
262 * extended here, under a lock), while poll may be called from
263 * other threads. (If you are unfamiliar with them, you probably
264 * want to read Herlihy and Shavit's book "The Art of
265 * Multiprocessor programming", chapter 16 describing these in
266 * more detail before proceeding.) The main work-stealing queue
267 * design is roughly similar to those in the papers "Dynamic
268 * Circular Work-Stealing Deque" by Chase and Lev, SPAA 2005
269 * (http://research.sun.com/scalable/pubs/index.html) and
270 * "Idempotent work stealing" by Michael, Saraswat, and Vechev,
271 * PPoPP 2009 (http://portal.acm.org/citation.cfm?id=1504186).
272 * The main differences ultimately stem from GC requirements that
273 * we null out taken slots as soon as we can, to maintain as small
274 * a footprint as possible even in programs generating huge
275 * numbers of tasks. To accomplish this, we shift the CAS
276 * arbitrating pop vs poll (steal) from being on the indices
277 * ("base" and "top") to the slots themselves. These provide the
278 * primary required memory ordering -- see "Correct and Efficient
279 * Work-Stealing for Weak Memory Models" by Le, Pop, Cohen, and
280 * Nardelli, PPoPP 2013
281 * (http://www.di.ens.fr/~zappa/readings/ppopp13.pdf) for an
282 * analysis of memory ordering requirements in work-stealing
283 * algorithms similar to the one used here. We use per-operation
284 * ordered writes of various kinds for accesses when required.
285 *
286 * We also support a user mode in which local task processing is
287 * in FIFO, not LIFO order, simply by using a local version of
288 * poll rather than pop. This can be useful in message-passing
289 * frameworks in which tasks are never joined, although with
290 * increased contention among task producers and consumers. Also,
291 * the same data structure (and class) is used for "submission
292 * queues" (described below) holding externally submitted tasks,
293 * that differ only in that a lock (using field "phase"; see below) is
294 * required by external callers to push and pop tasks.
295 *
296 * Adding tasks then takes the form of a classic array push(task)
297 * in a circular buffer:
298 * q.array[q.top++ % length] = task;
299 *
300 * The actual code needs to null-check and size-check the array,
301 * uses masking, not mod, for indexing a power-of-two-sized array,
302 * enforces memory ordering, supports resizing, and possibly
303 * signals waiting workers to start scanning (described below),
304 * which requires stronger forms of order accesses.
305 *
306 * The pop operation (always performed by owner) is of the form:
307 * if ((task = getAndSet(q.array, (q.top-1) % length, null)) != null)
308 * decrement top and return task;
309 * If this fails, the queue is empty. This operation is one part
310 * of the nextLocalTask method, that instead does a local-poll
311 * in FIFO mode.
312 *
313 * The poll operation is, basically:
314 * if (CAS nonnull task t = q.array[k = q.base % length] to null)
315 * increment base and return task;
316 *
317 * However, there are several more cases that must be dealt with.
318 * Some of them are just due to asynchrony; others reflect
319 * contention and stealing policies. Stepping through them
320 * illustrates some of the implementation decisions in this class.
321 *
322 * * Slot k must be read with an acquiring read, which it must
323 * anyway to dereference and run the task if the (acquiring)
324 * CAS succeeds.
325 *
326 * * q.base may change between reading and using its value to
327 * index the slot. To avoid trying to use the wrong t, the
328 * index and slot must be reread (not necessarily immediately)
329 * until consistent, unless this is a local poll by owner, in
330 * which case this form of inconsistency can only appear as t
331 * being null, below.
332 *
333 * * Similarly, q.array may change (due to a resize), unless this
334 * is a local poll by owner. Otherwise, when t is present, this
335 * only needs consideration on CAS failure (since a CAS
336 * confirms the non-resized case.)
337 *
338 * * t may appear null because a previous poll operation has not
339 * yet incremented q.base, so the read is from an already-taken
340 * index. This form of stall reflects the non-lock-freedom of
341 * the poll operation. Stalls can be detected by observing that
342 * q.base doesn't change on repeated reads of null t and when
343 * no other alternatives apply, spin-wait for it to settle. To
344 * reduce producing these kinds of stalls by other stealers, we
345 * encourage timely writes to indices using otherwise
346 * unnecessarily strong writes.
347 *
348 * * The CAS may fail, in which case we may want to retry unless
349 * there is too much contention. One goal is to balance and
350 * spread out the many forms of contention that may be
351 * encountered across polling and other operations to avoid
352 * sustained performance degradations. Across all cases where
353 * alternatives exist, a bounded number of CAS misses or stalls
354 * are tolerated (for slots, ctl, and elsewhere described
355 * below) before taking alternative action. These may move
356 * contention or retries elsewhere, which is still preferable
357 * to single-point bottlenecks.
358 *
359 * * Even though the check "top == base" is quiescently accurate
360 * to determine whether a queue is empty, it is not of much use
361 * when deciding whether to try to poll or repoll after a
362 * failure. Both top and base may move independently, and both
363 * lag updates to the underlying array. To reduce memory
364 * contention, non-owners avoid reading the "top" when
365 * possible, by using one-ahead reads to check whether to
366 * repoll, relying on the fact that a non-empty queue does not
367 * have two null slots in a row, except in cases (resizes and
368 * shifts) that can be detected with a secondary recheck that
369 * is less likely to conflict with owner writes.
370 *
371 * The poll operations in q.poll(), runWorker(), helpJoin(), and
372 * elsewhere differ with respect to whether other queues are
373 * available to try, and the presence or nature of screening steps
374 * when only some kinds of tasks can be taken. When alternatives
375 * (or failing) is an option, they uniformly give up after
376 * bounded numbers of stalls and/or CAS failures, which reduces
377 * contention when too many workers are polling too few tasks.
378 * Overall, in the aggregate, we ensure probabilistic
379 * non-blockingness of work-stealing at least until checking
380 * quiescence (which is intrinsically blocking): If an attempted
381 * steal fails in these ways, a scanning thief chooses a different
382 * target to try next. In contexts where alternatives aren't
383 * available, and when progress conditions can be isolated to
384 * values of a single variable, simple spinloops (using
385 * Thread.onSpinWait) are used to reduce memory traffic.
386 *
387 * WorkQueues are also used in a similar way for tasks submitted
388 * to the pool. We cannot mix these tasks in the same queues used
389 * by workers. Instead, we randomly associate submission queues
390 * with submitting threads (or carriers when using VirtualThreads)
391 * using a form of hashing. The ThreadLocalRandom probe value
392 * serves as a hash code for choosing existing queues, and may be
393 * randomly repositioned upon contention with other submitters.
394 * In essence, submitters act like workers except that they are
395 * restricted to executing local tasks that they submitted (or
396 * when known, subtasks thereof). Insertion of tasks in shared
397 * mode requires a lock. We use only a simple spinlock (as one
398 * role of field "phase") because submitters encountering a busy
399 * queue move to a different position to use or create other
400 * queues. They (spin) block when registering new queues, or
401 * indirectly elsewhere, by revisiting later.
402 *
403 * Management
404 * ==========
405 *
406 * The main throughput advantages of work-stealing stem from
407 * decentralized control -- workers mostly take tasks from
408 * themselves or each other, at rates that can exceed a billion
409 * per second. Most non-atomic control is performed by some form
410 * of scanning across or within queues. The pool itself creates,
411 * activates (enables scanning for and running tasks),
412 * deactivates, blocks, and terminates threads, all with minimal
413 * central information. There are only a few properties that we
414 * can globally track or maintain, so we pack them into a small
415 * number of variables, often maintaining atomicity without
416 * blocking or locking. Nearly all essentially atomic control
417 * state is held in a few variables that are by far most often
418 * read (not written) as status and consistency checks. We pack as
419 * much information into them as we can.
420 *
421 * Field "ctl" contains 64 bits holding information needed to
422 * atomically decide to add, enqueue (on an event queue), and
423 * dequeue and release workers. To enable this packing, we
424 * restrict maximum parallelism to (1<<15)-1 (which is far in
425 * excess of normal operating range) to allow ids, counts, and
426 * their negations (used for thresholding) to fit into 16bit
427 * subfields.
428 *
429 * Field "runState" and per-WorkQueue field "phase" play similar
430 * roles, as lockable, versioned counters. Field runState also
431 * includes monotonic event bits:
432 * * SHUTDOWN: no more external tasks accepted; STOP when quiescent
433 * * STOP: no more tasks run, and deregister all workers
434 * * CLEANED: all unexecuted tasks have been cancelled
435 * * TERMINATED: all workers deregistered and all queues cleaned
436 * The version tags enable detection of state changes (by
437 * comparing two reads) modulo bit wraparound. The bit range in
438 * each case suffices for purposes of determining quiescence,
439 * termination, avoiding ABA-like errors, and signal control, most
440 * of which are ultimately based on at most 15bit ranges (due to
441 * 32767 max total workers). RunState updates do not need to be
442 * atomic with respect to ctl updates, but because they are not,
443 * some care is required to avoid stalls. The seqLock properties
444 * detect changes and conditionally upgrade to coordinate with
445 * updates. It is typically held for less than a dozen
446 * instructions unless the queue array is being resized, during
447 * which contention is rare. To be conservative, lockRunState is
448 * implemented as a spin/sleep loop. Here and elsewhere spin
449 * constants are short enough to apply even on systems with few
450 * available processors. In addition to checking pool status,
451 * reads of runState sometimes serve as acquire fences before
452 * reading other fields.
453 *
454 * Field "parallelism" holds the target parallelism (normally
455 * corresponding to pool size). Users can dynamically reset target
456 * parallelism, but is only accessed when signalling or awaiting
457 * work, so only slowly has an effect in creating threads or
458 * letting them time out and terminate when idle.
459 *
460 * Array "queues" holds references to WorkQueues. It is updated
461 * (only during worker creation and termination) under the
462 * runState lock. It is otherwise concurrently readable but reads
463 * for use in scans (see below) are always prefaced by a volatile
464 * read of runState (or equivalent constructions), ensuring that
465 * its state is current at the point it is used (which is all we
466 * require). To simplify index-based operations, the array size is
467 * always a power of two, and all readers must tolerate null
468 * slots. Worker queues are at odd indices. Worker phase ids
469 * masked with SMASK match their index. Shared (submission) queues
470 * are at even indices. Grouping them together in this way aids in
471 * task scanning: At top-level, both kinds of queues should be
472 * sampled with approximately the same probability, which is
473 * simpler if they are all in the same array. But we also need to
474 * identify what kind they are without looking at them, leading to
475 * this odd/even scheme. One disadvantage is that there are
476 * usually many fewer submission queues, so there can be many
477 * wasted probes (null slots). But this is still cheaper than
478 * alternatives. Other loops over the queues array vary in origin
479 * and stride depending on whether they cover only submission
480 * (even) or worker (odd) queues or both, and whether they require
481 * randomness (in which case cyclically exhaustive strides may be
482 * used).
483 *
484 * All worker thread creation is on-demand, triggered by task
485 * submissions, replacement of terminated workers, and/or
486 * compensation for blocked workers. However, all other support
487 * code is set up to work with other policies. To ensure that we
488 * do not hold on to worker or task references that would prevent
489 * GC, all accesses to workQueues in waiting, signalling, and
490 * control methods are via indices into the queues array (which is
491 * one source of some of the messy code constructions here). In
492 * essence, the queues array serves as a weak reference
493 * mechanism. In particular, the stack top subfield of ctl stores
494 * indices, not references. Operations on queues obtained from
495 * these indices remain valid (with at most some unnecessary extra
496 * work) even if an underlying worker failed and was replaced by
497 * another at the same index. During termination, worker queue
498 * array updates are disabled.
499 *
500 * Queuing Idle Workers. Unlike HPC work-stealing frameworks, we
501 * cannot let workers spin indefinitely scanning for tasks when
502 * none can be found immediately, and we cannot start/resume
503 * workers unless there appear to be tasks available. On the
504 * other hand, we must quickly prod them into action when new
505 * tasks are submitted or generated. These latencies are mainly a
506 * function of JVM park/unpark (and underlying OS) performance,
507 * which can be slow and variable (even though usages are
508 * streamlined as much as possible). In many usages, ramp-up time
509 * is the main limiting factor in overall performance, which is
510 * compounded at program start-up by JIT compilation and
511 * allocation. On the other hand, throughput degrades when too
512 * many threads poll for too few tasks. (See below.)
513 *
514 * The "ctl" field atomically maintains total and "released"
515 * worker counts, plus the head of the available worker queue
516 * (actually stack, represented by the lower 32bit subfield of
517 * ctl). Released workers are those known to be scanning for
518 * and/or running tasks (we cannot accurately determine
519 * which). Unreleased ("available") workers are recorded in the
520 * ctl stack. These workers are made eligible for signalling by
521 * enqueuing in ctl (see method deactivate). This "queue" is a
522 * form of Treiber stack. This is ideal for activating threads in
523 * most-recently used order, and improves performance and
524 * locality, outweighing the disadvantages of being prone to
525 * contention and inability to release a worker unless it is
526 * topmost on stack. The top stack state holds the value of the
527 * "phase" field of the worker: its index and status, plus a
528 * version counter that, in addition to the count subfields (also
529 * serving as version stamps) provide protection against Treiber
530 * stack ABA effects.
531 *
532 * Creating workers. To create a worker, we pre-increment counts
533 * (serving as a reservation), and attempt to construct a
534 * ForkJoinWorkerThread via its factory. On starting, the new
535 * thread first invokes registerWorker, where it is assigned an
536 * index in the queues array (expanding the array if necessary).
537 * Upon any exception across these steps, or null return from
538 * factory, deregisterWorker adjusts counts and records
539 * accordingly. If a null return, the pool continues running with
540 * fewer than the target number workers. If exceptional, the
541 * exception is propagated, generally to some external caller.
542 *
543 * WorkQueue field "phase" encodes the queue array id in lower
544 * bits, and otherwise acts similarly to the pool runState field:
545 * The "IDLE" bit is clear while active (either a released worker
546 * or a locked external queue), with other bits serving as a
547 * version counter to distinguish changes across multiple reads.
548 * Note that phase field updates lag queue CAS releases; seeing a
549 * non-idle phase does not guarantee that the worker is available
550 * (and so is never checked in this way).
551 *
552 * The ctl field also serves as the basis for memory
553 * synchronization surrounding activation. This uses a more
554 * efficient version of a Dekker-like rule that task producers and
555 * consumers sync with each other by both writing/CASing ctl (even
556 * if to its current value). However, rather than CASing ctl to
557 * its current value in the common case where no action is
558 * required, we reduce write contention by ensuring that
559 * signalWork invocations are prefaced with a fully fenced memory
560 * access (which is usually needed anyway).
561 *
562 * Signalling. Signals (in signalWork) cause new or reactivated
563 * workers to scan for tasks. SignalWork is invoked in two cases:
564 * (1) When a task is pushed onto an empty queue, and (2) When a
565 * worker takes a top-level task from a queue that has additional
566 * tasks. Together, these suffice in O(log(#threads)) steps to
567 * fully activate with at least enough workers, and ideally no
568 * more than required. This ideal is unobtainable: Callers do not
569 * know whether another worker will finish its current task and
570 * poll for others without need of a signal (which is otherwise an
571 * advantage of work-stealing vs other schemes), and also must
572 * conservatively estimate the triggering conditions of emptiness
573 * or non-emptiness; all of which usually cause more activations
574 * than necessary (see below). (Method signalWork is also used as
575 * failsafe in case of Thread failures in deregisterWorker, to
576 * activate or create a new worker to replace them).
577 *
578 * Top-Level scheduling
579 * ====================
580 *
581 * Scanning. Method runWorker performs top-level scanning for (and
582 * execution of) tasks by polling a pseudo-random permutation of
583 * the array (by starting at a given index, and using a constant
584 * cyclically exhaustive stride.) It uses the same basic polling
585 * method as WorkQueue.poll(), but restarts with a different
586 * permutation on each rescan. The pseudorandom generator need
587 * not have high-quality statistical properties in the long
588 * term. We use Marsaglia XorShifts, seeded with the Weyl sequence
589 * from ThreadLocalRandom probes, which are cheap and suffice.
590 *
591 * Deactivation. When no tasks are found by a worker in runWorker,
592 * it invokes awaitWork, that first deactivates (to an IDLE
593 * phase). Avoiding missed signals during deactivation requires a
594 * (conservative) rescan, reactivating if there may be tasks to
595 * poll. Because idle workers are often not yet blocked (parked),
596 * we use a WorkQueue field to advertise that a waiter actually
597 * needs unparking upon signal.
598 *
599 * When tasks are constructed as (recursive) DAGs, top-level
600 * scanning is usually infrequent, and doesn't encounter most
601 * of the following problems addressed by runWorker and awaitWork:
602 *
603 * Locality. Polls are organized into "runs", continuing until
604 * empty or contended, while also minimizing interference by
605 * postponing bookeeping to ends of runs. This may reduce
606 * fairness.
607 *
608 * Contention. When many workers try to poll few queues, they
609 * often collide, generating CAS failures and disrupting locality
610 * of workers already running their tasks. This also leads to
611 * stalls when tasks cannot be taken because other workers have
612 * not finished poll operations, which is detected by reading
613 * ahead in queue arrays. In both cases, workers restart scans in a
614 * way that approximates randomized backoff.
615 *
616 * Oversignalling. When many short top-level tasks are present in
617 * a small number of queues, the above signalling strategy may
618 * activate many more workers than needed, worsening locality and
619 * contention problems, while also generating more global
620 * contention (field ctl is CASed on every activation and
621 * deactivation). We filter out (both in runWorker and
622 * signalWork) attempted signals that are surely not needed
623 * because the signalled tasks are already taken.
624 *
625 * Shutdown and Quiescence
626 * =======================
627 *
628 * Quiescence. Workers scan looking for work, giving up when they
629 * don't find any, without being sure that none are available.
630 * However, some required functionality relies on consensus about
631 * quiescence (also termination, discussed below). The count
632 * fields in ctl allow accurate discovery of states in which all
633 * workers are idle. However, because external (asynchronous)
634 * submitters are not part of this vote, these mechanisms
635 * themselves do not guarantee that the pool is in a quiescent
636 * state with respect to methods isQuiescent, shutdown (which
637 * begins termination when quiescent), helpQuiesce, and indirectly
638 * others including tryCompensate. Method quiescent() is used in
639 * all of these contexts. It provides checks that all workers are
640 * idle and there are no submissions that they could poll if they
641 * were not idle, retrying on inconsistent reads of queues and
642 * using the runState seqLock to retry on queue array updates.
643 * (It also reports quiescence if the pool is terminating.) A true
644 * report means only that there was a moment at which quiescence
645 * held. False negatives are inevitable (for example when queues
646 * indices lag updates, as described above), which is accommodated
647 * when (tentatively) idle by scanning for work etc, and then
648 * re-invoking. This includes cases in which the final unparked
649 * thread (in deactivate()) uses quiescent() to check for tasks
650 * that could have been added during a race window that would not
651 * be accompanied by a signal, in which case re-activating itself
652 * (or any other worker) to rescan. Method helpQuiesce acts
653 * similarly but cannot rely on ctl counts to determine that all
654 * workers are inactive because the caller and any others
655 * executing helpQuiesce are not included in counts.
656 *
657 * Termination. Termination is initiated by setting STOP in one of
658 * three ways (via methods tryTerminate and quiescent):
659 * * A call to shutdownNow, in which case all workers are
660 * interrupted, first ensuring that the queues array is stable,
661 * to avoid missing any workers.
662 * * A call to shutdown when quiescent, in which case method
663 * releaseWaiters is used to dequeue them, at which point they notice
664 * STOP state and return from runWorker to deregister();
665 * * The pool becomes quiescent() sometime after shutdown has
666 * been called, in which case releaseWaiters is also used to
667 * propagate as they deregister.
668 * Upon STOP, each worker, as well as external callers to
669 * tryTerminate (via close() etc) race to set CLEANED, indicating
670 * that all tasks have been cancelled. The implementation (method
671 * cleanQueues) balances cases in which there may be many tasks to
672 * cancel (benefitting from parallelism) versus contention and
673 * interference when many threads try to poll remaining queues,
674 * while also avoiding unnecessary rechecks, by using
675 * pseudorandom scans and giving up upon interference. This may be
676 * retried by the same caller only when there are no more
677 * registered workers, using the same criteria as method
678 * quiescent. When CLEANED and all workers have deregistered,
679 * TERMINATED is set, also signalling any caller of
680 * awaitTermination or close. Because shutdownNow-based
681 * termination relies on interrupts, there is no guarantee that
682 * workers will stop if their tasks ignore interrupts. Class
683 * InterruptibleTask (see below) further arranges runState checks
684 * before executing task bodies, and ensures interrupts while
685 * terminating. Even so, there are no guarantees because tasks may
686 * internally enter unbounded loops.
687 *
688 * Trimming workers. To release resources after periods of lack of
689 * use, a worker starting to wait when the pool is quiescent will
690 * time out and terminate if the pool has remained quiescent for
691 * period given by field keepAlive (default 60sec), which applies
692 * to the first timeout of a quiescent pool. Subsequent cases use
693 * minimal delays such that, if still quiescent, all will be
694 * released soon thereafter. This is checked by setting the
695 * "source" field of signallee to an invalid value, that will
696 * remain invalid only if it did not process any tasks.
697 *
698 * Joining Tasks
699 * =============
700 *
701 * The "Join" part of ForkJoinPools consists of a set of
702 * mechanisms that sometimes or always (depending on the kind of
703 * task) avoid context switching or adding worker threads when one
704 * task would otherwise be blocked waiting for completion of
705 * another, basically, just by running that task or one of its
706 * subtasks if not already taken. These mechanics are disabled for
707 * InterruptibleTasks, that guarantee that callers do not execute
708 * submitted tasks.
709 *
710 * The basic structure of joining is an extended spin/block scheme
711 * in which workers check for task completions status between
712 * steps to find other work, until relevant pool state stabilizes
713 * enough to believe that no such tasks are available, at which
714 * point blocking. This is usually a good choice of when to block
715 * that would otherwise be harder to approximate.
716 *
717 * These forms of helping may increase stack space usage, but that
718 * space is bounded in tree/dag structured procedurally parallel
719 * designs to be no more than that if a task were executed only by
720 * the joining thread. This is arranged by associated task
721 * subclasses that also help detect and control the ways in which
722 * this may occur.
723 *
724 * Normally, the first option when joining a task that is not done
725 * is to try to take it from the local queue and run it. Method
726 * tryRemoveAndExec tries to do so. For tasks with any form of
727 * subtasks that must be completed first, we try to locate these
728 * subtasks and run them as well. This is easy when local, but
729 * when stolen, steal-backs are restricted to the same rules as
730 * stealing (polling), which requires additional bookkeeping and
731 * scanning. This cost is still very much worthwhile because of
732 * its impact on task scheduling and resource control.
733 *
734 * The two methods for finding and executing subtasks vary in
735 * details. The algorithm in helpJoin entails a form of "linear
736 * helping". Each worker records (in field "source") the index of
737 * the internal queue from which it last stole a task. (Note:
738 * because chains cannot include even-numbered external queues,
739 * they are ignored, and 0 is an OK default. However, the source
740 * field is set anyway, or eventually to DROPPED, to ensure
741 * volatile memory synchronization effects.) The scan in method
742 * helpJoin uses these markers to try to find a worker to help
743 * (i.e., steal back a task from and execute it) that could make
744 * progress toward completion of the actively joined task. Thus,
745 * the joiner executes a task that would be on its own local deque
746 * if the to-be-joined task had not been stolen. This is a
747 * conservative variant of the approach described in Wagner &
748 * Calder "Leapfrogging: a portable technique for implementing
749 * efficient futures" SIGPLAN Notices, 1993
750 * (http://portal.acm.org/citation.cfm?id=155354). It differs
751 * mainly in that we only record queues, not full dependency
752 * links. This requires a linear scan of the queues to locate
753 * stealers, but isolates cost to when it is needed, rather than
754 * adding to per-task overhead. For CountedCompleters, the
755 * analogous method helpComplete doesn't need stealer-tracking,
756 * but requires a similar (but simpler) check of completion
757 * chains.
758 *
759 * In either case, searches can fail to locate stealers when
760 * stalls delay recording sources or issuing subtasks. We avoid
761 * some of these cases by using snapshotted values of ctl as a
762 * check that the numbers of workers are not changing, along with
763 * rescans to deal with contention and stalls. But even when
764 * accurately identified, stealers might not ever produce a task
765 * that the joiner can in turn help with.
766 *
767 * Related method helpAsyncBlocker does not directly rely on
768 * subtask structure, but instead avoids or postpones blocking of
769 * tagged tasks (CompletableFuture.AsynchronousCompletionTask) by
770 * executing other asyncs that can be processed in any order.
771 * This is currently invoked only in non-join-based blocking
772 * contexts from classes CompletableFuture and
773 * SubmissionPublisher, that could be further generalized.
774 *
775 * When any of the above fail to avoid blocking, we rely on
776 * "compensation" -- an indirect form of context switching that
777 * either activates an existing worker to take the place of the
778 * blocked one, or expands the number of workers.
779 *
780 * Compensation does not by default aim to keep exactly the target
781 * parallelism number of unblocked threads running at any given
782 * time. Some previous versions of this class employed immediate
783 * compensations for any blocked join. However, in practice, the
784 * vast majority of blockages are transient byproducts of GC and
785 * other JVM or OS activities that are made worse by replacement
786 * by causing longer-term oversubscription. These are inevitable
787 * without (unobtainably) perfect information about whether worker
788 * creation is actually necessary. False alarms are common enough
789 * to negatively impact performance, so compensation is by default
790 * attempted only when it appears possible that the pool could
791 * stall due to lack of any unblocked workers. However, we allow
792 * users to override defaults using the long form of the
793 * ForkJoinPool constructor. The compensation mechanism may also
794 * be bounded. Bounds for the commonPool better enable JVMs to
795 * cope with programming errors and abuse before running out of
796 * resources to do so.
797 *
798 * The ManagedBlocker extension API can't use helping so relies
799 * only on compensation in method awaitBlocker. This API was
800 * designed to highlight the uncertainty of compensation decisions
801 * by requiring implementation of method isReleasable to abort
802 * compensation during attempts to obtain a stable snapshot. But
803 * users now rely upon the fact that if isReleasable always
804 * returns false, the API can be used to obtain precautionary
805 * compensation, which is sometimes the only reasonable option
806 * when running unknown code in tasks; which is now supported more
807 * simply (see method beginCompensatedBlock).
808 *
809 * Common Pool
810 * ===========
811 *
812 * The static common pool always exists after static
813 * initialization. Since it (or any other created pool) need
814 * never be used, we minimize initial construction overhead and
815 * footprint to the setup of about a dozen fields, although with
816 * some System property parsing properties are set. The common pool is
817 * distinguished by having a null workerNamePrefix (which is an
818 * odd convention, but avoids the need to decode status in factory
819 * classes). It also has PRESET_SIZE config set if parallelism
820 * was configured by system property.
821 *
822 * When external threads use the common pool, they can perform
823 * subtask processing (see helpComplete and related methods) upon
824 * joins, unless they are submitted using ExecutorService
825 * submission methods, which implicitly disallow this. This
826 * caller-helps policy makes it sensible to set common pool
827 * parallelism level to one (or more) less than the total number
828 * of available cores, or even zero for pure caller-runs. External
829 * threads waiting for joins first check the common pool for their
830 * task, which fails quickly if the caller did not fork to common
831 * pool.
832 *
833 * Guarantees for common pool parallelism zero are limited to
834 * tasks that are joined by their callers in a tree-structured
835 * fashion or use CountedCompleters (as is true for jdk
836 * parallelStreams). Support infiltrates several methods,
837 * including those that retry helping steps until we are sure that
838 * none apply if there are no workers. To deal with conflicting
839 * requirements, uses of the commonPool that require async because
840 * caller-runs need not apply, ensure threads are enabled (by
841 * setting parallelism) via method asyncCommonPool before
842 * proceeding. (In principle, overriding zero parallelism needs to
843 * ensure at least one worker, but due to other backward
844 * compatibility contraints, ensures two.)
845 *
846 * As a more appropriate default in managed environments, unless
847 * overridden by system properties, we use workers of subclass
848 * InnocuousForkJoinWorkerThread for the commonPool. These
849 * workers do not belong to any user-defined ThreadGroup, and
850 * clear all ThreadLocals and reset the ContextClassLoader before
851 * (re)activating to execute top-level tasks. The associated
852 * mechanics may be JVM-dependent and must access particular
853 * Thread class fields to achieve this effect.
854 *
855 * InterruptibleTasks
856 * ====================
857 *
858 * Regular ForkJoinTasks manage task cancellation (method cancel)
859 * independently from the interrupt status of threads running
860 * tasks. Interrupts are issued internally only while
861 * terminating, to wake up workers and cancel queued tasks. By
862 * default, interrupts are cleared only when necessary to ensure
863 * that calls to LockSupport.park do not loop indefinitely (park
864 * returns immediately if the current thread is interrupted).
865 *
866 * To comply with ExecutorService specs, we use subclasses of
867 * abstract class InterruptibleTask for tasks that require
868 * stronger interruption and cancellation guarantees. External
869 * submitters never run these tasks, even if in the common pool
870 * (as indicated by ForkJoinTask.noUserHelp status bit).
871 * InterruptibleTasks include a "runner" field (implemented
872 * similarly to FutureTask) to support cancel(true). Upon pool
873 * shutdown, runners are interrupted so they can cancel. Since
874 * external joining callers never run these tasks, they must await
875 * cancellation by others, which can occur along several different
876 * paths.
877 *
878 * Across these APIs, rules for reporting exceptions for tasks
879 * with results accessed via join() differ from those via get(),
880 * which differ from those invoked using pool submit methods by
881 * non-workers (which comply with Future.get() specs). Internal
882 * usages of ForkJoinTasks ignore interrupt status when executing
883 * or awaiting completion. Otherwise, reporting task results or
884 * exceptions is preferred to throwing InterruptedExceptions,
885 * which are in turn preferred to timeouts. Similarly, completion
886 * status is preferred to reporting cancellation. Cancellation is
887 * reported as an unchecked exception by join(), and by worker
888 * calls to get(), but is otherwise wrapped in a (checked)
889 * ExecutionException.
890 *
891 * Worker Threads cannot be VirtualThreads, as enforced by
892 * requiring ForkJoinWorkerThreads in factories. There are
893 * several constructions relying on this. However as of this
894 * writing, virtual thread bodies are by default run as some form
895 * of InterruptibleTask.
896 *
897 * DelayScheduler
898 * ================
899 *
900 * This class supports ScheduledExecutorService methods by
901 * creating and starting a DelayScheduler on first use of these
902 * methods (via startDelayScheduler). The scheduler operates
903 * independently in its own thread, relaying tasks to the pool to
904 * execute when their delays elapse (see method
905 * executeEnabledScheduledTask). The only other interactions with
906 * the delayScheduler are to control shutdown and maintain
907 * shutdown-related policies in methods quiescent() and
908 * tryTerminate(). In particular, processing must deal with cases
909 * in which tasks are submitted before shutdown, but not enabled
910 * until afterwards, in which case they must bypass some screening
911 * to be allowed to run. Conversely, the DelayScheduler checks
912 * runState status and when enabled, completes termination, using
913 * only methods shutdownStatus and tryStopIfShutdown. All of these
914 * methods are final and have signatures referencing
915 * DelaySchedulers, so cannot conflict with those of any existing
916 * FJP subclasses.
917 *
918 * Memory placement
919 * ================
920 *
921 * Performance is very sensitive to placement of instances of
922 * ForkJoinPool and WorkQueues and their queue arrays, as well as
923 * the placement of their fields. Caches misses and contention due
924 * to false-sharing have been observed to slow down some programs
925 * by more than a factor of four. Effects may vary across initial
926 * memory configuarations, applications, and different garbage
927 * collectors and GC settings, so there is no perfect solution.
928 * Too much isolation may generate more cache misses in common
929 * cases (because some fields snd slots are usually read at the
930 * same time). The @Contended annotation provides only rough
931 * control (for good reason). Similarly for relying on fields
932 * being placed in size-sorted declaration order.
933 *
934 * We isolate the ForkJoinPool.ctl field that otherwise causes the
935 * most false-sharing misses with respect to other fields. Also,
936 * ForkJoinPool fields are ordered such that fields less prone to
937 * contention effects are first, offsetting those that otherwise
938 * would be, while also reducing total footprint vs using
939 * multiple @Contended regions, which tends to slow down
940 * less-contended applications. To help arrange this, some
941 * non-reference fields are declared as "long" even when ints or
942 * shorts would suffice. For class WorkQueue, an
943 * embedded @Contended isolates the very busy top index, and
944 * another segregates status and bookkeeping fields written
945 * (mostly) by owners, that otherwise interfere with reading
946 * array, top, and base fields. There are other variables commonly
947 * contributing to false-sharing-related performance issues
948 * (including fields of class Thread), but we can't do much about
949 * this except try to minimize access.
950 *
951 * Initial sizing and resizing of WorkQueue arrays is an even more
952 * delicate tradeoff because the best strategy systematically
953 * varies across garbage collectors. Small arrays are better for
954 * locality and reduce GC scan time, but large arrays reduce both
955 * direct false-sharing and indirect cases due to GC bookkeeping
956 * (cardmarks etc), and reduce the number of resizes, which are
957 * not especially fast because they require atomic transfers.
958 * Currently, arrays are initialized to be just large enough to
959 * avoid resizing in most tree-structured tasks, but grow rapidly
960 * until large. (Maintenance note: any changes in fields, queues,
961 * or their uses, or JVM layout policies, must be accompanied by
962 * re-evaluation of these placement and sizing decisions.)
963 *
964 * Style notes
965 * ===========
966 *
967 * Memory ordering relies mainly on atomic operations (CAS,
968 * getAndSet, getAndAdd) along with moded accesses. These use
969 * jdk-internal Unsafe for atomics and special memory modes,
970 * rather than VarHandles, to avoid initialization dependencies in
971 * other jdk components that require early parallelism. This can
972 * be awkward and ugly, but also reflects the need to control
973 * outcomes across the unusual cases that arise in very racy code
974 * with very few invariants. All atomic task slot updates use
975 * Unsafe operations requiring offset positions, not indices, as
976 * computed by method slotOffset. All fields are read into locals
977 * before use, and null-checked if they are references, even if
978 * they can never be null under current usages. Usually,
979 * computations (held in local variables) are defined as soon as
980 * logically enabled, sometimes to convince compilers that they
981 * may be performed despite memory ordering constraints. Array
982 * accesses using masked indices include checks (that are always
983 * true) that the array length is non-zero to avoid compilers
984 * inserting more expensive traps. This is usually done in a
985 * "C"-like style of listing declarations at the heads of methods
986 * or blocks, and using inline assignments on first encounter.
987 * Nearly all explicit checks lead to bypass/return, not exception
988 * throws, because they may legitimately arise during shutdown. A
989 * few unusual loop constructions encourage (with varying
990 * effectiveness) JVMs about where (not) to place safepoints. All
991 * public methods screen arguments (mainly null checks) before
992 * creating or executing tasks.
993 *
994 * There is a lot of representation-level coupling among classes
995 * ForkJoinPool, ForkJoinWorkerThread, and ForkJoinTask. The
996 * fields of WorkQueue maintain data structures managed by
997 * ForkJoinPool, so are directly accessed. There is little point
998 * trying to reduce this, since any associated future changes in
999 * representations will need to be accompanied by algorithmic
1000 * changes anyway. Several methods intrinsically sprawl because
1001 * they must accumulate sets of consistent reads of fields held in
1002 * local variables. Some others are artificially broken up to
1003 * reduce producer/consumer imbalances due to dynamic compilation.
1004 * There are also other coding oddities (including several
1005 * unnecessary-looking hoisted null checks) that help some methods
1006 * perform reasonably even when interpreted (not compiled).
1007 *
1008 * The order of declarations in this file is (with a few exceptions):
1009 * (1) Static configuration constants
1010 * (2) Static utility functions
1011 * (3) Nested (static) classes
1012 * (4) Fields, along with constants used when unpacking some of them
1013 * (5) Internal control methods
1014 * (6) Callbacks and other support for ForkJoinTask methods
1015 * (7) Exported methods
1016 * (8) Static block initializing statics in minimally dependent order
1017 *
1018 */
1019
1020 // static configuration constants
1021
1022 /**
1023 * Default idle timeout value (in milliseconds) for idle threads
1024 * to park waiting for new work before terminating.
1025 */
1026 static final long DEFAULT_KEEPALIVE = 60_000L;
1027
1028 /**
1029 * Undershoot tolerance for idle timeouts, also serving as the
1030 * minimum allowed timeout value.
1031 */
1032 static final long TIMEOUT_SLOP = 20L;
1033
1034 /**
1035 * The default value for common pool maxSpares. Overridable using
1036 * the "java.util.concurrent.ForkJoinPool.common.maximumSpares"
1037 * system property. The default value is far in excess of normal
1038 * requirements, but also far short of maximum capacity and typical OS
1039 * thread limits, so allows JVMs to catch misuse/abuse before
1040 * running out of resources needed to do so.
1041 */
1042 static final int DEFAULT_COMMON_MAX_SPARES = 256;
1043
1044 /**
1045 * Initial capacity of work-stealing queue array.
1046 * Must be a power of two, at least 2. See above.
1047 */
1048 static final int INITIAL_QUEUE_CAPACITY = 1 << 6;
1049
1050 // conversions among short, int, long
1051 static final int SMASK = 0xffff; // (unsigned) short bits
1052 static final long LMASK = 0xffffffffL; // lower 32 bits of long
1053 static final long UMASK = ~LMASK; // upper 32 bits
1054
1055 // masks and sentinels for queue indices
1056 static final int MAX_CAP = 0x7fff; // max # workers
1057 static final int EXTERNAL_ID_MASK = 0x3ffe; // max external queue id
1058 static final int INVALID_ID = 0x4000; // unused external queue id
1059
1060 // pool.runState bits
1061 static final long STOP = 1L << 0; // terminating
1062 static final long SHUTDOWN = 1L << 1; // terminate when quiescent
1063 static final long CLEANED = 1L << 2; // stopped and queues cleared
1064 static final long TERMINATED = 1L << 3; // only set if STOP also set
1065 static final long RS_LOCK = 1L << 4; // lowest seqlock bit
1066
1067 // spin/sleep limits for runState locking and elsewhere
1068 static final int SPIN_WAITS = 1 << 7; // max calls to onSpinWait
1069 static final int MIN_SLEEP = 1 << 10; // approx 1 usec as nanos
1070 static final int MAX_SLEEP = 1 << 20; // approx 1 sec as nanos
1071
1072 // {pool, workQueue} config bits
1073 static final int FIFO = 1 << 0; // fifo queue or access mode
1074 static final int CLEAR_TLS = 1 << 1; // set for Innocuous workers
1075 static final int PRESET_SIZE = 1 << 2; // size was set by property
1076
1077 // others
1078 static final int DROPPED = 1 << 16; // removed from ctl counts
1079 static final int UNCOMPENSATE = 1 << 16; // tryCompensate return
1080 static final int IDLE = 1 << 16; // phase seqlock/version count
1081 static final int MIN_QUEUES_SIZE = 1 << 4; // ensure external slots
1082
1083 /*
1084 * Bits and masks for ctl and bounds are packed with 4 16 bit subfields:
1085 * RC: Number of released (unqueued) workers
1086 * TC: Number of total workers
1087 * SS: version count and status of top waiting thread
1088 * ID: poolIndex of top of Treiber stack of waiters
1089 *
1090 * When convenient, we can extract the lower 32 stack top bits
1091 * (including version bits) as sp=(int)ctl. When sp is non-zero,
1092 * there are waiting workers. Count fields may be transiently
1093 * negative during termination because of out-of-order updates.
1094 * To deal with this, we use casts in and out of "short" and/or
1095 * signed shifts to maintain signedness. Because it occupies
1096 * uppermost bits, we can add one release count using getAndAdd of
1097 * RC_UNIT, rather than CAS, when returning from a blocked join.
1098 * Other updates of multiple subfields require CAS.
1099 */
1100
1101 // Release counts
1102 static final int RC_SHIFT = 48;
1103 static final long RC_UNIT = 0x0001L << RC_SHIFT;
1104 static final long RC_MASK = 0xffffL << RC_SHIFT;
1105 // Total counts
1106 static final int TC_SHIFT = 32;
1107 static final long TC_UNIT = 0x0001L << TC_SHIFT;
1108 static final long TC_MASK = 0xffffL << TC_SHIFT;
1109
1110 /*
1111 * All atomic operations on task arrays (queues) use Unsafe
1112 * operations that take array offsets versus indices, based on
1113 * array base and shift constants established during static
1114 * initialization.
1115 */
1116 static final long ABASE;
1117 static final int ASHIFT;
1118
1119 // Static utilities
1120
1121 /**
1122 * Returns the array offset corresponding to the given index for
1123 * Unsafe task queue operations
1124 */
1125 static long slotOffset(int index) {
1126 return ((long)index << ASHIFT) + ABASE;
1127 }
1128
1129 // Nested classes
1130
1131 /**
1132 * Factory for creating new {@link ForkJoinWorkerThread}s.
1133 * A {@code ForkJoinWorkerThreadFactory} must be defined and used
1134 * for {@code ForkJoinWorkerThread} subclasses that extend base
1135 * functionality or initialize threads with different contexts.
1136 */
1137 public static interface ForkJoinWorkerThreadFactory {
1138 /**
1139 * Returns a new worker thread operating in the given pool.
1140 * Returning null or throwing an exception may result in tasks
1141 * never being executed. If this method throws an exception,
1142 * it is relayed to the caller of the method (for example
1143 * {@code execute}) causing attempted thread creation. If this
1144 * method returns null or throws an exception, it is not
1145 * retried until the next attempted creation (for example
1146 * another call to {@code execute}).
1147 *
1148 * @param pool the pool this thread works in
1149 * @return the new worker thread, or {@code null} if the request
1150 * to create a thread is rejected
1151 * @throws NullPointerException if the pool is null
1152 */
1153 public ForkJoinWorkerThread newThread(ForkJoinPool pool);
1154 }
1155
1156 /**
1157 * Default ForkJoinWorkerThreadFactory implementation; creates a
1158 * new ForkJoinWorkerThread using the system class loader as the
1159 * thread context class loader.
1160 */
1161 static final class DefaultForkJoinWorkerThreadFactory
1162 implements ForkJoinWorkerThreadFactory {
1163 public final ForkJoinWorkerThread newThread(ForkJoinPool pool) {
1164 return ((pool.workerNamePrefix == null) ? // is commonPool
1165 new ForkJoinWorkerThread.InnocuousForkJoinWorkerThread(pool) :
1166 new ForkJoinWorkerThread(null, pool, true, false));
1167 }
1168 }
1169
1170 /**
1171 * Queues supporting work-stealing as well as external task
1172 * submission. See above for descriptions and algorithms.
1173 */
1174 static final class WorkQueue {
1175 // fields declared in order of their likely layout on most VMs
1176 final ForkJoinWorkerThread owner; // null if shared
1177 ForkJoinTask<?>[] array; // the queued tasks; power of 2 size
1178 int base; // index of next slot for poll
1179 final int config; // mode bits
1180
1181 @jdk.internal.vm.annotation.Contended("t") // segregate
1182 int top; // index of next slot for push
1183
1184 // fields otherwise causing more unnecessary false-sharing cache misses
1185 @jdk.internal.vm.annotation.Contended("w")
1186 volatile int phase; // versioned active status
1187 @jdk.internal.vm.annotation.Contended("w")
1188 int stackPred; // pool stack (ctl) predecessor link
1189 @jdk.internal.vm.annotation.Contended("w")
1190 volatile int parking; // nonzero if parked in awaitWork
1191 @jdk.internal.vm.annotation.Contended("w")
1192 volatile int source; // source queue id (or DROPPED)
1193 @jdk.internal.vm.annotation.Contended("w")
1194 int nsteals; // number of steals from other queues
1195
1196 // Support for atomic operations
1197 private static final Unsafe U;
1198 private static final long PHASE;
1199 private static final long BASE;
1200 private static final long TOP;
1201 private static final long ARRAY;
1202
1203 final void updateBase(int v) {
1204 U.putIntVolatile(this, BASE, v);
1205 }
1206 final void updateTop(int v) {
1207 U.putIntOpaque(this, TOP, v);
1208 }
1209 final void updateArray(ForkJoinTask<?>[] a) {
1210 U.getAndSetReference(this, ARRAY, a);
1211 }
1212 final void unlockPhase() {
1213 U.getAndAddInt(this, PHASE, IDLE);
1214 }
1215 final boolean tryLockPhase() { // seqlock acquire
1216 int p;
1217 return (((p = phase) & IDLE) != 0 &&
1218 U.compareAndSetInt(this, PHASE, p, p + IDLE));
1219 }
1220
1221 /**
1222 * Constructor. For internal queues, most fields are initialized
1223 * upon thread start in pool.registerWorker.
1224 */
1225 WorkQueue(ForkJoinWorkerThread owner, int id, int cfg,
1226 boolean clearThreadLocals) {
1227 this.config = (clearThreadLocals) ? cfg | CLEAR_TLS : cfg;
1228 if ((this.owner = owner) == null) {
1229 array = new ForkJoinTask<?>[INITIAL_QUEUE_CAPACITY];
1230 phase = id | IDLE;
1231 }
1232 }
1233
1234 /**
1235 * Returns an exportable index (used by ForkJoinWorkerThread).
1236 */
1237 final int getPoolIndex() {
1238 return (phase & 0xffff) >>> 1; // ignore odd/even tag bit
1239 }
1240
1241 /**
1242 * Returns the approximate number of tasks in the queue.
1243 */
1244 final int queueSize() {
1245 int unused = phase; // for ordering effect
1246 return Math.max(top - base, 0); // ignore transient negative
1247 }
1248
1249 /**
1250 * Pushes a task. Called only by owner or if already locked
1251 *
1252 * @param task the task; no-op if null
1253 * @param pool the pool to signal if was previously empty, else null
1254 * @param internal if caller owns this queue
1255 * @throws RejectedExecutionException if array could not be resized
1256 */
1257 final void push(ForkJoinTask<?> task, ForkJoinPool pool, boolean internal) {
1258 int s = top, b = base, m, cap, room; ForkJoinTask<?>[] a;
1259 if ((a = array) != null && (cap = a.length) > 0) { // else disabled
1260 if ((room = (m = cap - 1) - (s - b)) >= 0) {
1261 top = s + 1;
1262 long pos = slotOffset(m & s);
1263 if (!internal)
1264 U.putReference(a, pos, task); // inside lock
1265 else
1266 U.getAndSetReference(a, pos, task); // fully fenced
1267 if (room == 0 && (a = growArray(a, cap, s)) != null)
1268 m = a.length - 1; // resize
1269 }
1270 if (!internal)
1271 unlockPhase();
1272 if (room < 0)
1273 throw new RejectedExecutionException("Queue capacity exceeded");
1274 if (pool != null && a != null &&
1275 U.getReferenceAcquire(a, slotOffset(m & (s - 1))) == null)
1276 pool.signalWork(a, m & s); // may have appeared empty
1277 }
1278 }
1279
1280 /**
1281 * Resizes the queue array unless out of memory.
1282 * @param a old array
1283 * @param cap old array capacity
1284 * @param s current top
1285 * @return new array, or null on failure
1286 */
1287 private ForkJoinTask<?>[] growArray(ForkJoinTask<?>[] a, int cap, int s) {
1288 int newCap = (cap >= 1 << 16) ? cap << 1 : cap << 2;
1289 ForkJoinTask<?>[] newArray = null;
1290 if (a != null && a.length == cap && cap > 0 && newCap > 0) {
1291 try {
1292 newArray = new ForkJoinTask<?>[newCap];
1293 } catch (OutOfMemoryError ex) {
1294 }
1295 if (newArray != null) { // else throw on next push
1296 int mask = cap - 1, newMask = newCap - 1;
1297 for (int k = s, j = cap; j > 0; --j, --k) {
1298 ForkJoinTask<?> u; // poll old, push to new
1299 if ((u = (ForkJoinTask<?>)U.getAndSetReference(
1300 a, slotOffset(k & mask), null)) == null)
1301 break; // lost to pollers
1302 newArray[k & newMask] = u;
1303 }
1304 updateArray(newArray); // fully fenced
1305 }
1306 }
1307 return newArray;
1308 }
1309
1310 /**
1311 * Takes next task, if one exists, in lifo order.
1312 */
1313 private ForkJoinTask<?> localPop() {
1314 ForkJoinTask<?> t = null;
1315 int s = top - 1, cap; long k; ForkJoinTask<?>[] a;
1316 if ((a = array) != null && (cap = a.length) > 0 &&
1317 U.getReference(a, k = slotOffset((cap - 1) & s)) != null &&
1318 (t = (ForkJoinTask<?>)U.getAndSetReference(a, k, null)) != null)
1319 updateTop(s);
1320 return t;
1321 }
1322
1323 /**
1324 * Takes next task, if one exists, in fifo order.
1325 */
1326 private ForkJoinTask<?> localPoll() {
1327 ForkJoinTask<?> t = null;
1328 int p = top, cap; ForkJoinTask<?>[] a;
1329 if ((a = array) != null && (cap = a.length) > 0) {
1330 for (int b = base; p - b > 0; ) {
1331 int nb = b + 1;
1332 long k = slotOffset((cap - 1) & b);
1333 if (U.getReference(a, k) == null) {
1334 if (nb == p)
1335 break; // else base is lagging
1336 while (b == (b = U.getIntAcquire(this, BASE)))
1337 Thread.onSpinWait(); // spin to reduce memory traffic
1338 }
1339 else if ((t = (ForkJoinTask<?>)
1340 U.getAndSetReference(a, k, null)) != null) {
1341 updateBase(nb);
1342 break;
1343 }
1344 else
1345 b = base;
1346 }
1347 }
1348 return t;
1349 }
1350
1351 /**
1352 * Takes next task, if one exists, using configured mode.
1353 */
1354 final ForkJoinTask<?> nextLocalTask() {
1355 return (config & FIFO) == 0 ? localPop() : localPoll();
1356 }
1357
1358 /**
1359 * Pops the given task only if it is at the current top.
1360 * @param task the task. Caller must ensure non-null.
1361 * @param internal if caller owns this queue
1362 */
1363 final boolean tryUnpush(ForkJoinTask<?> task, boolean internal) {
1364 boolean taken = false;
1365 ForkJoinTask<?>[] a = array;
1366 int p = top, s = p - 1, cap; long k;
1367 if (a != null && (cap = a.length) > 0 &&
1368 U.getReference(a, k = slotOffset((cap - 1) & s)) == task &&
1369 (internal || tryLockPhase())) {
1370 if (top == p && U.compareAndSetReference(a, k, task, null)) {
1371 taken = true;
1372 updateTop(s);
1373 }
1374 if (!internal)
1375 unlockPhase();
1376 }
1377 return taken;
1378 }
1379
1380 /**
1381 * Returns next task, if one exists, in order specified by mode.
1382 */
1383 final ForkJoinTask<?> peek() {
1384 ForkJoinTask<?>[] a = array;
1385 int b = base, cfg = config, p = top, cap;
1386 if (p != b && a != null && (cap = a.length) > 0) {
1387 if ((cfg & FIFO) == 0)
1388 return a[(cap - 1) & (p - 1)];
1389 else { // skip over in-progress removals
1390 ForkJoinTask<?> t;
1391 for ( ; p - b > 0; ++b) {
1392 if ((t = a[(cap - 1) & b]) != null)
1393 return t;
1394 }
1395 }
1396 }
1397 return null;
1398 }
1399
1400 /**
1401 * Polls for a task. Used only by non-owners.
1402 */
1403 final ForkJoinTask<?> poll() {
1404 for (int pb = -1, b; ; pb = b) { // track progress
1405 ForkJoinTask<?> t; int cap, nb; long k; ForkJoinTask<?>[] a;
1406 if ((a = array) == null || (cap = a.length) <= 0)
1407 break;
1408 t = (ForkJoinTask<?>)U.getReferenceAcquire(
1409 a, k = slotOffset((cap - 1) & (b = base)));
1410 Object u = U.getReference( // next slot
1411 a, slotOffset((cap - 1) & (nb = b + 1)));
1412 if (base != b) // inconsistent
1413 ;
1414 else if (t == null) {
1415 if (u == null && top - b <= 0)
1416 break; // empty
1417 if (pb == b)
1418 Thread.onSpinWait(); // stalled
1419 }
1420 else if (U.compareAndSetReference(a, k, t, null)) {
1421 updateBase(nb);
1422 return t;
1423 }
1424 }
1425 return null;
1426 }
1427
1428 // specialized execution methods
1429
1430 /*
1431 * Two version (lifo and fifo) of top-level execution, split
1432 * across modes to better isolate task dispatch and local
1433 * processing from top-level scheduling.
1434 */
1435 final void topLevelExecLifo(ForkJoinTask<?> task) {
1436 while (task != null) {
1437 task.doExec();
1438 task = localPop();
1439 }
1440 }
1441
1442 final void topLevelExecFifo(ForkJoinTask<?> task) {
1443 while (task != null) {
1444 task.doExec();
1445 task = localPoll();
1446 }
1447 }
1448
1449 /**
1450 * Deep form of tryUnpush: Traverses from top and removes and
1451 * runs task if present.
1452 */
1453 final void tryRemoveAndExec(ForkJoinTask<?> task, boolean internal) {
1454 ForkJoinTask<?>[] a = array;
1455 int b = base, p = top, s = p - 1, d = p - b, cap;
1456 if (a != null && (cap = a.length) > 0) {
1457 for (int m = cap - 1, i = s; d > 0; --i, --d) {
1458 long k; boolean taken;
1459 ForkJoinTask<?> t = (ForkJoinTask<?>)U.getReference(
1460 a, k = slotOffset(i & m));
1461 if (t == null)
1462 break;
1463 if (t == task) {
1464 if (!internal && !tryLockPhase())
1465 break; // fail if locked
1466 if (taken =
1467 (top == p &&
1468 U.compareAndSetReference(a, k, task, null))) {
1469 if (i == s) // act as pop
1470 updateTop(s);
1471 else if (i == base) // act as poll
1472 updateBase(i + 1);
1473 else { // swap with top
1474 U.putReferenceVolatile(
1475 a, k, (ForkJoinTask<?>)
1476 U.getAndSetReference(
1477 a, slotOffset(s & m), null));
1478 updateTop(s);
1479 }
1480 }
1481 if (!internal)
1482 unlockPhase();
1483 if (taken)
1484 task.doExec();
1485 break;
1486 }
1487 }
1488 }
1489 }
1490
1491 /**
1492 * Tries to pop and run tasks within the target's computation
1493 * until done, not found, or limit exceeded.
1494 *
1495 * @param task root of computation
1496 * @param limit max runs, or zero for no limit
1497 * @return task status if known to be done
1498 */
1499 final int helpComplete(ForkJoinTask<?> task, boolean internal, int limit) {
1500 int status = 0;
1501 if (task != null) {
1502 outer: for (;;) {
1503 ForkJoinTask<?>[] a; boolean taken; Object o;
1504 int stat, p, s, cap;
1505 if ((stat = task.status) < 0) {
1506 status = stat;
1507 break;
1508 }
1509 if ((a = array) == null || (cap = a.length) <= 0)
1510 break;
1511 long k = slotOffset((cap - 1) & (s = (p = top) - 1));
1512 if (!((o = U.getReference(a, k)) instanceof CountedCompleter))
1513 break;
1514 CountedCompleter<?> t = (CountedCompleter<?>)o, f = t;
1515 for (int steps = cap;;) { // bound path
1516 if (f == task)
1517 break;
1518 if ((f = f.completer) == null || --steps == 0)
1519 break outer;
1520 }
1521 if (!internal && !tryLockPhase())
1522 break;
1523 if (taken =
1524 (top == p &&
1525 U.compareAndSetReference(a, k, t, null)))
1526 updateTop(s);
1527 if (!internal)
1528 unlockPhase();
1529 if (!taken)
1530 break;
1531 t.doExec();
1532 if (limit != 0 && --limit == 0)
1533 break;
1534 }
1535 }
1536 return status;
1537 }
1538
1539 /**
1540 * Tries to poll and run AsynchronousCompletionTasks until
1541 * none found or blocker is released
1542 *
1543 * @param blocker the blocker
1544 */
1545 final void helpAsyncBlocker(ManagedBlocker blocker) {
1546 for (;;) {
1547 ForkJoinTask<?> t; ForkJoinTask<?>[] a; int b, cap; long k;
1548 if ((a = array) == null || (cap = a.length) <= 0)
1549 break;
1550 t = (ForkJoinTask<?>)U.getReferenceAcquire(
1551 a, k = slotOffset((cap - 1) & (b = base)));
1552 if (t == null) {
1553 if (top - b <= 0)
1554 break;
1555 }
1556 else if (!(t instanceof CompletableFuture
1557 .AsynchronousCompletionTask))
1558 break;
1559 if (blocker != null && blocker.isReleasable())
1560 break;
1561 if (base == b && t != null &&
1562 U.compareAndSetReference(a, k, t, null)) {
1563 updateBase(b + 1);
1564 t.doExec();
1565 }
1566 }
1567 }
1568
1569 // misc
1570
1571 /**
1572 * Cancels all local tasks. Called only by owner.
1573 */
1574 final void cancelTasks() {
1575 for (ForkJoinTask<?> t; (t = localPop()) != null; ) {
1576 try {
1577 t.cancel(false);
1578 } catch (Throwable ignore) {
1579 }
1580 }
1581 }
1582
1583 /**
1584 * Returns true if internal and not known to be blocked.
1585 */
1586 final boolean isApparentlyUnblocked() {
1587 Thread wt; Thread.State s;
1588 return ((wt = owner) != null && (phase & IDLE) != 0 &&
1589 (s = wt.getState()) != Thread.State.BLOCKED &&
1590 s != Thread.State.WAITING &&
1591 s != Thread.State.TIMED_WAITING);
1592 }
1593
1594 static {
1595 U = Unsafe.getUnsafe();
1596 Class<WorkQueue> klass = WorkQueue.class;
1597 PHASE = U.objectFieldOffset(klass, "phase");
1598 BASE = U.objectFieldOffset(klass, "base");
1599 TOP = U.objectFieldOffset(klass, "top");
1600 ARRAY = U.objectFieldOffset(klass, "array");
1601 }
1602 }
1603
1604 // static fields (initialized in static initializer below)
1605
1606 /**
1607 * Creates a new ForkJoinWorkerThread. This factory is used unless
1608 * overridden in ForkJoinPool constructors.
1609 */
1610 public static final ForkJoinWorkerThreadFactory
1611 defaultForkJoinWorkerThreadFactory;
1612
1613 /**
1614 * Common (static) pool. Non-null for public use unless a static
1615 * construction exception, but internal usages null-check on use
1616 * to paranoically avoid potential initialization circularities
1617 * as well as to simplify generated code.
1618 */
1619 static final ForkJoinPool common;
1620
1621 /**
1622 * Sequence number for creating worker names
1623 */
1624 private static volatile int poolIds;
1625
1626 /**
1627 * For VirtualThread intrinsics
1628 */
1629 private static final JavaLangAccess JLA;
1630
1631 // fields declared in order of their likely layout on most VMs
1632 volatile CountDownLatch termination; // lazily constructed
1633 final Predicate<? super ForkJoinPool> saturate;
1634 final ForkJoinWorkerThreadFactory factory;
1635 final UncaughtExceptionHandler ueh; // per-worker UEH
1636 final SharedThreadContainer container;
1637 final String workerNamePrefix; // null for common pool
1638 final String poolName;
1639 volatile DelayScheduler delayScheduler; // lazily constructed
1640 WorkQueue[] queues; // main registry
1641 volatile long runState; // versioned, lockable
1642 final long keepAlive; // milliseconds before dropping if idle
1643 final long config; // static configuration bits
1644 volatile long stealCount; // collects worker nsteals
1645 volatile long threadIds; // for worker thread names
1646
1647 @jdk.internal.vm.annotation.Contended("fjpctl") // segregate
1648 volatile long ctl; // main pool control
1649 @jdk.internal.vm.annotation.Contended("fjpctl") // colocate
1650 int parallelism; // target number of workers
1651
1652 // Support for atomic operations
1653 private static final Unsafe U;
1654 private static final long CTL;
1655 private static final long RUNSTATE;
1656 private static final long PARALLELISM;
1657 private static final long THREADIDS;
1658 private static final long TERMINATION;
1659 private static final Object POOLIDS_BASE;
1660 private static final long POOLIDS;
1661
1662 private boolean compareAndSetCtl(long c, long v) {
1663 return U.compareAndSetLong(this, CTL, c, v);
1664 }
1665 private long compareAndExchangeCtl(long c, long v) {
1666 return U.compareAndExchangeLong(this, CTL, c, v);
1667 }
1668 private long getAndAddCtl(long v) {
1669 return U.getAndAddLong(this, CTL, v);
1670 }
1671 private long incrementThreadIds() {
1672 return U.getAndAddLong(this, THREADIDS, 1L);
1673 }
1674 private static int getAndAddPoolIds(int x) {
1675 return U.getAndAddInt(POOLIDS_BASE, POOLIDS, x);
1676 }
1677 private int getAndSetParallelism(int v) {
1678 return U.getAndSetInt(this, PARALLELISM, v);
1679 }
1680 private int getParallelismOpaque() {
1681 return U.getIntOpaque(this, PARALLELISM);
1682 }
1683 private CountDownLatch cmpExTerminationSignal(CountDownLatch x) {
1684 return (CountDownLatch)
1685 U.compareAndExchangeReference(this, TERMINATION, null, x);
1686 }
1687
1688 // runState operations
1689
1690 private long getAndBitwiseOrRunState(long v) { // for status bits
1691 return U.getAndBitwiseOrLong(this, RUNSTATE, v);
1692 }
1693 private boolean casRunState(long c, long v) {
1694 return U.compareAndSetLong(this, RUNSTATE, c, v);
1695 }
1696 private void unlockRunState() { // increment lock bit
1697 U.getAndAddLong(this, RUNSTATE, RS_LOCK);
1698 }
1699 private long lockRunState() { // lock and return current state
1700 long s, u; // locked when RS_LOCK set
1701 if (((s = runState) & RS_LOCK) == 0L && casRunState(s, u = s + RS_LOCK))
1702 return u;
1703 else
1704 return spinLockRunState();
1705 }
1706 private long spinLockRunState() { // spin/sleep
1707 for (int waits = 0;;) {
1708 long s, u;
1709 if (((s = runState) & RS_LOCK) == 0L) {
1710 if (casRunState(s, u = s + RS_LOCK))
1711 return u;
1712 waits = 0;
1713 } else if (waits < SPIN_WAITS) {
1714 ++waits;
1715 Thread.onSpinWait();
1716 } else {
1717 if (waits < MIN_SLEEP)
1718 waits = MIN_SLEEP;
1719 LockSupport.parkNanos(this, (long)waits);
1720 if (waits < MAX_SLEEP)
1721 waits <<= 1;
1722 }
1723 }
1724 }
1725
1726 static boolean poolIsStopping(ForkJoinPool p) { // Used by ForkJoinTask
1727 return p != null && (p.runState & STOP) != 0L;
1728 }
1729
1730 // Creating, registering, and deregistering workers
1731
1732 /**
1733 * Tries to construct and start one worker. Assumes that total
1734 * count has already been incremented as a reservation. Invokes
1735 * deregisterWorker on any failure.
1736 *
1737 * @return true if successful
1738 */
1739 private boolean createWorker() {
1740 ForkJoinWorkerThreadFactory fac = factory;
1741 SharedThreadContainer ctr = container;
1742 Throwable ex = null;
1743 ForkJoinWorkerThread wt = null;
1744 try {
1745 if ((runState & STOP) == 0L && // avoid construction if terminating
1746 fac != null && (wt = fac.newThread(this)) != null) {
1747 if (ctr != null)
1748 ctr.start(wt);
1749 else
1750 wt.start();
1751 return true;
1752 }
1753 } catch (Throwable rex) {
1754 ex = rex;
1755 }
1756 deregisterWorker(wt, ex);
1757 return false;
1758 }
1759
1760 /**
1761 * Provides a name for ForkJoinWorkerThread constructor.
1762 */
1763 final String nextWorkerThreadName() {
1764 String prefix = workerNamePrefix;
1765 long tid = incrementThreadIds() + 1L;
1766 if (prefix == null) // commonPool has no prefix
1767 prefix = "ForkJoinPool.commonPool-worker-";
1768 return prefix.concat(Long.toString(tid));
1769 }
1770
1771 /**
1772 * Finishes initializing and records internal queue.
1773 *
1774 * @param w caller's WorkQueue
1775 */
1776 final void registerWorker(WorkQueue w) {
1777 if (w != null) {
1778 w.array = new ForkJoinTask<?>[INITIAL_QUEUE_CAPACITY];
1779 ThreadLocalRandom.localInit();
1780 int seed = w.stackPred = ThreadLocalRandom.getProbe();
1781 int phaseSeq = seed & ~((IDLE << 1) - 1); // initial phase tag
1782 int id = ((seed << 1) | 1) & SMASK; // base of linear-probe-like scan
1783 long stop = lockRunState() & STOP;
1784 try {
1785 WorkQueue[] qs; int n;
1786 if (stop == 0L && (qs = queues) != null && (n = qs.length) > 0) {
1787 for (int k = n, m = n - 1; ; id += 2) {
1788 if (qs[id &= m] == null)
1789 break;
1790 if ((k -= 2) <= 0) {
1791 id |= n;
1792 break;
1793 }
1794 }
1795 w.phase = id | phaseSeq; // now publishable
1796 if (id < n)
1797 qs[id] = w;
1798 else { // expand
1799 int an = n << 1, am = an - 1;
1800 WorkQueue[] as = new WorkQueue[an];
1801 as[id & am] = w;
1802 for (int j = 1; j < n; j += 2)
1803 as[j] = qs[j];
1804 for (int j = 0; j < n; j += 2) {
1805 WorkQueue q; // shared queues may move
1806 if ((q = qs[j]) != null)
1807 as[q.phase & EXTERNAL_ID_MASK & am] = q;
1808 }
1809 U.storeFence(); // fill before publish
1810 queues = as;
1811 }
1812 }
1813 } finally {
1814 unlockRunState();
1815 }
1816 }
1817 }
1818
1819 /**
1820 * Final callback from terminating worker, as well as upon failure
1821 * to construct or start a worker. Removes record of worker from
1822 * array, and adjusts counts. If pool is shutting down, tries to
1823 * complete termination.
1824 *
1825 * @param wt the worker thread, or null if construction failed
1826 * @param ex the exception causing failure, or null if none
1827 */
1828 final void deregisterWorker(ForkJoinWorkerThread wt, Throwable ex) {
1829 WorkQueue w = null; // null if not created
1830 int phase = 0; // 0 if not registered
1831 if (wt != null && (w = wt.workQueue) != null &&
1832 (phase = w.phase) != 0 && (phase & IDLE) != 0)
1833 releaseWaiters(); // ensure released
1834 if (w == null || w.source != DROPPED) {
1835 long c = ctl; // decrement counts
1836 do {} while (c != (c = compareAndExchangeCtl(
1837 c, ((RC_MASK & (c - RC_UNIT)) |
1838 (TC_MASK & (c - TC_UNIT)) |
1839 (LMASK & c)))));
1840 }
1841 if (phase != 0 && w != null) { // remove index unless terminating
1842 long ns = w.nsteals & 0xffffffffL;
1843 if ((runState & STOP) == 0L) {
1844 WorkQueue[] qs; int n, i;
1845 if ((lockRunState() & STOP) == 0L &&
1846 (qs = queues) != null && (n = qs.length) > 0 &&
1847 qs[i = phase & SMASK & (n - 1)] == w) {
1848 qs[i] = null;
1849 stealCount += ns; // accumulate steals
1850 }
1851 unlockRunState();
1852 }
1853 }
1854 if ((tryTerminate(false, false) & STOP) == 0L &&
1855 phase != 0 && w != null && w.source != DROPPED) {
1856 w.cancelTasks(); // clean queue
1857 signalWork(null, 0); // possibly replace
1858 }
1859 if (ex != null)
1860 ForkJoinTask.rethrow(ex);
1861 }
1862
1863 /**
1864 * Releases an idle worker, or creates one if not enough exist,
1865 * giving up if array a is nonnull and task at a[k] already taken.
1866 */
1867 final void signalWork(ForkJoinTask<?>[] a, int k) {
1868 int pc = parallelism;
1869 for (long c = ctl;;) {
1870 WorkQueue[] qs = queues;
1871 long ac = (c + RC_UNIT) & RC_MASK, nc;
1872 int sp = (int)c, i = sp & SMASK;
1873 if ((short)(c >>> RC_SHIFT) >= pc)
1874 break;
1875 if (qs == null)
1876 break;
1877 if (qs.length <= i)
1878 break;
1879 WorkQueue w = qs[i], v = null;
1880 if (sp == 0) {
1881 if ((short)(c >>> TC_SHIFT) >= pc)
1882 break;
1883 nc = ((c + TC_UNIT) & TC_MASK) | ac;
1884 }
1885 else if ((v = w) == null)
1886 break;
1887 else
1888 nc = (v.stackPred & LMASK) | (c & TC_MASK) | ac;
1889 if (a != null && k < a.length && k >= 0 && a[k] == null)
1890 break;
1891 if (c == (c = ctl) && c == (c = compareAndExchangeCtl(c, nc))) {
1892 if (v == null)
1893 createWorker();
1894 else {
1895 v.phase = sp;
1896 if (v.parking != 0)
1897 U.unpark(v.owner);
1898 }
1899 break;
1900 }
1901 }
1902 }
1903
1904 /**
1905 * Releases all waiting workers. Called only during shutdown.
1906 */
1907 private void releaseWaiters() {
1908 for (long c = ctl;;) {
1909 WorkQueue[] qs; WorkQueue v; int sp, i;
1910 if ((sp = (int)c) == 0 || (qs = queues) == null ||
1911 qs.length <= (i = sp & SMASK) || (v = qs[i]) == null)
1912 break;
1913 if (c == (c = compareAndExchangeCtl(
1914 c, ((UMASK & (c + RC_UNIT)) | (c & TC_MASK) |
1915 (v.stackPred & LMASK))))) {
1916 v.phase = sp;
1917 if (v.parking != 0)
1918 U.unpark(v.owner);
1919 }
1920 }
1921 }
1922
1923 /**
1924 * Internal version of isQuiescent and related functionality.
1925 * @return positive if stopping, nonnegative if terminating or all
1926 * workers are inactive and submission queues are empty and
1927 * unlocked; if so, setting STOP if shutdown is enabled
1928 */
1929 private int quiescent() {
1930 for (;;) {
1931 long phaseSum = 0L;
1932 boolean swept = false;
1933 for (long e, prevRunState = 0L; ; prevRunState = e) {
1934 DelayScheduler ds;
1935 long c = ctl;
1936 if (((e = runState) & STOP) != 0L)
1937 return 1; // terminating
1938 else if ((c & RC_MASK) > 0L)
1939 return -1; // at least one active
1940 else if (!swept || e != prevRunState || (e & RS_LOCK) != 0) {
1941 long sum = c;
1942 WorkQueue[] qs = queues;
1943 int n = (qs == null) ? 0 : qs.length;
1944 for (int i = 0; i < n; ++i) { // scan queues
1945 WorkQueue q;
1946 if ((q = qs[i]) != null) {
1947 int p = q.phase, s = q.top, b = q.base;
1948 sum += (p & 0xffffffffL) | ((long)b << 32);
1949 if ((p & IDLE) == 0 || s - b > 0)
1950 return -1;
1951 }
1952 }
1953 swept = (phaseSum == (phaseSum = sum));
1954 }
1955 else if ((e & SHUTDOWN) == 0)
1956 return 0;
1957 else if ((ds = delayScheduler) != null && !ds.canShutDown())
1958 return 0;
1959 else if (compareAndSetCtl(c, c) && casRunState(e, e | STOP))
1960 return 1; // enable termination
1961 else
1962 break; // restart
1963 }
1964 }
1965 }
1966
1967 /**
1968 * Top-level runloop for workers, called by ForkJoinWorkerThread.run.
1969 * See above for explanation.
1970 *
1971 * @param w caller's WorkQueue (may be null on failed initialization)
1972 */
1973 final void runWorker(WorkQueue w) {
1974 if (w != null) {
1975 int r = w.stackPred; // seed from registerWorker
1976 int fifo = (int)config & FIFO;
1977 int nsteals = 0; // shadow w.nsteals
1978 boolean rescan = true;
1979 WorkQueue[] qs; int n;
1980 while ((rescan || deactivate(w) == 0) && (runState & STOP) == 0L &&
1981 (qs = queues) != null && (n = qs.length) > 0) {
1982 rescan = false;
1983 int i = r, step = (r >>> 16) | 1;
1984 r ^= r << 13; r ^= r >>> 17; r ^= r << 5; // xorshift
1985 scan: for (int j = n << 1; j != 0; --j, i += step) { // 2 sweeps
1986 WorkQueue q; int qid;
1987 if ((q = qs[qid = i & (n - 1)]) != null) {
1988 for (;;) { // poll queue q
1989 ForkJoinTask<?>[] a; int cap, b, m, nb, nk;
1990 if ((a = q.array) == null || (cap = a.length) <= 0)
1991 break;
1992 long bp = slotOffset((m = cap - 1) & (b = q.base));
1993 long np = slotOffset(nk = m & (nb = b + 1));
1994 ForkJoinTask<?> t = (ForkJoinTask<?>)
1995 U.getReferenceAcquire(a, bp);
1996 if (q.array != a || q.base != b ||
1997 U.getReference(a, bp) != t)
1998 continue; // inconsistent
1999 if (t == null) {
2000 if (rescan) { // end of run
2001 w.nsteals = nsteals;
2002 break scan;
2003 }
2004 if (U.getReference(a, np) != null) {
2005 rescan = true; // stalled; reorder scan
2006 break scan;
2007 }
2008 break; // probably empty
2009 }
2010 if (U.compareAndSetReference(a, bp, t, null)) {
2011 q.base = nb;
2012 Object nt = U.getReferenceAcquire(a, np);
2013 if (!rescan) { // begin run
2014 rescan = true;
2015 w.source = qid;
2016 }
2017 ++nsteals;
2018 if (nt != null && // confirm a[nk]
2019 U.getReference(a, np) == nt)
2020 signalWork(a, nk); // propagate
2021 if (fifo != 0) // run t & its subtasks
2022 w.topLevelExecFifo(t);
2023 else
2024 w.topLevelExecLifo(t);
2025 }
2026 }
2027 }
2028 }
2029 }
2030 }
2031 }
2032
2033 /**
2034 * Deactivates and awaits signal or termination.
2035 *
2036 * @param w the work queue
2037 * @return zero if now active
2038 */
2039 private int deactivate(WorkQueue w) {
2040 int idle = 1;
2041 if (w != null) { // always true; hoist checks
2042 int p = w.phase, idlePhase = p | IDLE, activePhase = p + (IDLE << 1);
2043 long pc = ctl, qc = (activePhase & LMASK) | ((pc - RC_UNIT) & UMASK);
2044 w.stackPred = (int)pc; // set ctl stack link
2045 w.phase = idlePhase;
2046 if (!compareAndSetCtl(pc, qc)) { // try to enqueue
2047 w.phase = p; // back out on contention
2048 idle = 0;
2049 }
2050 WorkQueue[] qs; int n; long e;
2051 if (idle != 0 && ((e = runState) & STOP) == 0 && // quiescence checks
2052 ((e & SHUTDOWN) == 0L || (qc & RC_MASK) > 0L || quiescent() <= 0) &&
2053 (qs = queues) != null && (n = qs.length) > 1) {
2054 long psp = pc & LMASK; // ctl predecessor prefix
2055 for (int i = 1; i < n; ++i) { // scan; stagger origins
2056 WorkQueue q; long c; // missed signal check
2057 if ((q = qs[(activePhase + i) & (n - 1)]) != null &&
2058 q.top - q.base > 0) {
2059 if ((idle = w.phase - activePhase) != 0 &&
2060 (int)(c = ctl) == activePhase &&
2061 compareAndSetCtl(c, psp | ((c + RC_UNIT) & UMASK))) {
2062 w.phase = activePhase;
2063 idle = 0; // reactivated
2064 } // else ineligible or lost race
2065 break;
2066 }
2067 }
2068 if (idle != 0 && (idle = w.phase - activePhase) != 0)
2069 idle = awaitWork(w, activePhase, n);
2070 }
2071 }
2072 return idle;
2073 }
2074
2075 /**
2076 * Awaits signal or termination.
2077 *
2078 * @param w the work queue
2079 * @param activePhase w's next active phase
2080 * @param qsize current size of queues array
2081 * @return zero if now active
2082 */
2083 private int awaitWork(WorkQueue w, int activePhase, int qsize) {
2084 int idle = 1;
2085 if ((runState & STOP) == 0L && w != null) {
2086 int cfg = w.config;
2087 long waitTime = (w.source == INVALID_ID) ? 0L : keepAlive;
2088 long deadline = waitTime + System.currentTimeMillis();
2089 int spins = (qsize << 1) | (qsize - 1); // approx traversal cost
2090 if ((cfg & CLEAR_TLS) != 0 && // instanceof check always true
2091 Thread.currentThread() instanceof ForkJoinWorkerThread f)
2092 f.resetThreadLocals(); // clear while accessing thread state
2093 LockSupport.setCurrentBlocker(this);
2094 for (;;) {
2095 Thread.interrupted(); // clear status
2096 int s = spins;
2097 while ((idle = w.phase - activePhase) != 0 && --s != 0)
2098 Thread.onSpinWait(); // spin before/between parks
2099 if (idle == 0)
2100 break;
2101 if ((runState & STOP) != 0L)
2102 break;
2103 boolean trimmable = false; // use timed wait if trimmable
2104 long d = 0L, c;
2105 if (((c = ctl) & RC_MASK) == 0L && (int)c == activePhase) {
2106 if (deadline - System.currentTimeMillis() <= TIMEOUT_SLOP) {
2107 if (tryTrim(w, c, activePhase))
2108 break;
2109 continue; // lost race to trim
2110 }
2111 d = deadline;
2112 trimmable = true;
2113 }
2114 w.parking = 1; // enable unpark and recheck
2115 if ((idle = w.phase - activePhase) != 0)
2116 U.park(trimmable, d);
2117 w.parking = 0; // close unpark window
2118 if (idle == 0 || (idle = w.phase - activePhase) == 0)
2119 break;
2120 }
2121 LockSupport.setCurrentBlocker(null);
2122 }
2123 return idle;
2124 }
2125
2126 /**
2127 * Tries to remove and deregister worker after timeout, and release
2128 * another to do the same unless new tasks are found.
2129 */
2130 private boolean tryTrim(WorkQueue w, long c, int activePhase) {
2131 if (w != null) {
2132 int vp, i; WorkQueue[] vs; WorkQueue v;
2133 long nc = ((w.stackPred & LMASK) |
2134 ((RC_MASK & c) | (TC_MASK & (c - TC_UNIT))));
2135 if (compareAndSetCtl(c, nc)) {
2136 w.source = DROPPED;
2137 w.phase = activePhase;
2138 if ((vp = (int)nc) != 0 && (vs = queues) != null &&
2139 vs.length > (i = vp & SMASK) && (v = vs[i]) != null &&
2140 compareAndSetCtl( // try to wake up next waiter
2141 nc, ((v.stackPred & LMASK) |
2142 ((UMASK & (nc + RC_UNIT)) | (nc & TC_MASK))))) {
2143 v.source = INVALID_ID; // enable cascaded timeouts
2144 v.phase = vp;
2145 U.unpark(v.owner);
2146 }
2147 return true;
2148 }
2149 }
2150 return false;
2151 }
2152
2153 /**
2154 * Scans for and returns a polled task, if available. Used only
2155 * for untracked polls. Begins scan at a random index to avoid
2156 * systematic unfairness.
2157 *
2158 * @param submissionsOnly if true, only scan submission queues
2159 */
2160 private ForkJoinTask<?> pollScan(boolean submissionsOnly) {
2161 if ((runState & STOP) == 0L) {
2162 WorkQueue[] qs; int n; WorkQueue q; ForkJoinTask<?> t;
2163 int r = ThreadLocalRandom.nextSecondarySeed();
2164 if (submissionsOnly) // even indices only
2165 r &= ~1;
2166 int step = (submissionsOnly) ? 2 : 1;
2167 if ((qs = queues) != null && (n = qs.length) > 0) {
2168 for (int i = n; i > 0; i -= step, r += step) {
2169 if ((q = qs[r & (n - 1)]) != null &&
2170 (t = q.poll()) != null)
2171 return t;
2172 }
2173 }
2174 }
2175 return null;
2176 }
2177
2178 /**
2179 * Tries to decrement counts (sometimes implicitly) and possibly
2180 * arrange for a compensating worker in preparation for
2181 * blocking. May fail due to interference, in which case -1 is
2182 * returned so caller may retry. A zero return value indicates
2183 * that the caller doesn't need to re-adjust counts when later
2184 * unblocked.
2185 *
2186 * @param c incoming ctl value
2187 * @return UNCOMPENSATE: block then adjust, 0: block, -1 : retry
2188 */
2189 private int tryCompensate(long c) {
2190 Predicate<? super ForkJoinPool> sat;
2191 long b = config;
2192 int pc = parallelism, // unpack fields
2193 minActive = (short)(b >>> RC_SHIFT),
2194 maxTotal = (short)(b >>> TC_SHIFT) + pc,
2195 active = (short)(c >>> RC_SHIFT),
2196 total = (short)(c >>> TC_SHIFT),
2197 sp = (int)c,
2198 stat = -1; // default retry return
2199 if (sp != 0 && active <= pc) { // activate idle worker
2200 WorkQueue[] qs; WorkQueue v; int i;
2201 if ((qs = queues) != null && qs.length > (i = sp & SMASK) &&
2202 (v = qs[i]) != null &&
2203 compareAndSetCtl(c, (c & UMASK) | (v.stackPred & LMASK))) {
2204 v.phase = sp;
2205 if (v.parking != 0)
2206 U.unpark(v.owner);
2207 stat = UNCOMPENSATE;
2208 }
2209 }
2210 else if (active > minActive && total >= pc) { // reduce active workers
2211 if (compareAndSetCtl(c, ((c - RC_UNIT) & RC_MASK) | (c & ~RC_MASK)))
2212 stat = UNCOMPENSATE;
2213 }
2214 else if (total < maxTotal && total < MAX_CAP) { // try to expand pool
2215 long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK);
2216 if ((runState & STOP) != 0L) // terminating
2217 stat = 0;
2218 else if (compareAndSetCtl(c, nc))
2219 stat = createWorker() ? UNCOMPENSATE : 0;
2220 }
2221 else if (!compareAndSetCtl(c, c)) // validate
2222 ;
2223 else if ((sat = saturate) != null && sat.test(this))
2224 stat = 0;
2225 else
2226 throw new RejectedExecutionException(
2227 "Thread limit exceeded replacing blocked worker");
2228 return stat;
2229 }
2230
2231 /**
2232 * Readjusts RC count; called from ForkJoinTask after blocking.
2233 */
2234 final void uncompensate() {
2235 getAndAddCtl(RC_UNIT);
2236 }
2237
2238 /**
2239 * Helps if possible until the given task is done. Processes
2240 * compatible local tasks and scans other queues for task produced
2241 * by w's stealers; returning compensated blocking sentinel if
2242 * none are found.
2243 *
2244 * @param task the task
2245 * @param w caller's WorkQueue
2246 * @param internal true if w is owned by a ForkJoinWorkerThread
2247 * @return task status on exit, or UNCOMPENSATE for compensated blocking
2248 */
2249 final int helpJoin(ForkJoinTask<?> task, WorkQueue w, boolean internal) {
2250 if (w != null)
2251 w.tryRemoveAndExec(task, internal);
2252 int s = 0;
2253 if (task != null && (s = task.status) >= 0 && internal && w != null) {
2254 int wid = w.phase & SMASK, r = wid + 2, wsrc = w.source;
2255 long sctl = 0L; // track stability
2256 outer: for (boolean rescan = true;;) {
2257 if ((s = task.status) < 0)
2258 break;
2259 if (!rescan) {
2260 if ((runState & STOP) != 0L)
2261 break;
2262 if (sctl == (sctl = ctl) && (s = tryCompensate(sctl)) >= 0)
2263 break;
2264 }
2265 rescan = false;
2266 WorkQueue[] qs = queues;
2267 int n = (qs == null) ? 0 : qs.length;
2268 scan: for (int l = n >>> 1; l > 0; --l, r += 2) {
2269 int j; WorkQueue q;
2270 if ((q = qs[j = r & SMASK & (n - 1)]) != null) {
2271 for (;;) {
2272 ForkJoinTask<?> t; ForkJoinTask<?>[] a;
2273 boolean eligible = false;
2274 int sq = q.source, b, cap; long k;
2275 if ((a = q.array) == null || (cap = a.length) <= 0)
2276 break;
2277 t = (ForkJoinTask<?>)U.getReferenceAcquire(
2278 a, k = slotOffset((cap - 1) & (b = q.base)));
2279 if (t == task)
2280 eligible = true;
2281 else if (t != null) { // check steal chain
2282 for (int v = sq, d = cap;;) {
2283 WorkQueue p;
2284 if (v == wid) {
2285 eligible = true;
2286 break;
2287 }
2288 if ((v & 1) == 0 || // external or none
2289 --d < 0 || // bound depth
2290 (p = qs[v & (n - 1)]) == null)
2291 break;
2292 v = p.source;
2293 }
2294 }
2295 if ((s = task.status) < 0)
2296 break outer; // validate
2297 if (q.source == sq && q.base == b &&
2298 U.getReference(a, k) == t) {
2299 if (!eligible) { // revisit if nonempty
2300 if (!rescan && t == null && q.top - b > 0)
2301 rescan = true;
2302 break;
2303 }
2304 if (U.compareAndSetReference(a, k, t, null)) {
2305 q.base = b + 1;
2306 w.source = j; // volatile write
2307 t.doExec();
2308 w.source = wsrc;
2309 rescan = true; // restart at index r
2310 break scan;
2311 }
2312 }
2313 }
2314 }
2315 }
2316 }
2317 }
2318 return s;
2319 }
2320
2321 /**
2322 * Version of helpJoin for CountedCompleters.
2323 *
2324 * @param task root of computation (only called when a CountedCompleter)
2325 * @param w caller's WorkQueue
2326 * @param internal true if w is owned by a ForkJoinWorkerThread
2327 * @return task status on exit, or UNCOMPENSATE for compensated blocking
2328 */
2329 final int helpComplete(ForkJoinTask<?> task, WorkQueue w, boolean internal) {
2330 int s = 0;
2331 if (task != null && (s = task.status) >= 0 && w != null) {
2332 int r = w.phase + 1; // for indexing
2333 long sctl = 0L; // track stability
2334 outer: for (boolean rescan = true, locals = true;;) {
2335 if (locals && (s = w.helpComplete(task, internal, 0)) < 0)
2336 break;
2337 if ((s = task.status) < 0)
2338 break;
2339 if (!rescan) {
2340 if ((runState & STOP) != 0L)
2341 break;
2342 if (sctl == (sctl = ctl) &&
2343 (!internal || (s = tryCompensate(sctl)) >= 0))
2344 break;
2345 }
2346 rescan = locals = false;
2347 WorkQueue[] qs = queues;
2348 int n = (qs == null) ? 0 : qs.length;
2349 scan: for (int l = n; l > 0; --l, ++r) {
2350 int j; WorkQueue q;
2351 if ((q = qs[j = r & SMASK & (n - 1)]) != null) {
2352 for (;;) {
2353 ForkJoinTask<?> t; ForkJoinTask<?>[] a;
2354 int b, cap, nb; long k;
2355 boolean eligible = false;
2356 if ((a = q.array) == null || (cap = a.length) <= 0)
2357 break;
2358 t = (ForkJoinTask<?>)U.getReferenceAcquire(
2359 a, k = slotOffset((cap - 1) & (b = q.base)));
2360 if (t instanceof CountedCompleter) {
2361 CountedCompleter<?> f = (CountedCompleter<?>)t;
2362 for (int steps = cap; steps > 0; --steps) {
2363 if (f == task) {
2364 eligible = true;
2365 break;
2366 }
2367 if ((f = f.completer) == null)
2368 break;
2369 }
2370 }
2371 if ((s = task.status) < 0) // validate
2372 break outer;
2373 if (q.base == b) {
2374 if (eligible) {
2375 if (U.compareAndSetReference(
2376 a, k, t, null)) {
2377 q.updateBase(b + 1);
2378 t.doExec();
2379 locals = rescan = true;
2380 break scan;
2381 }
2382 }
2383 else if (U.getReference(a, k) == t) {
2384 if (!rescan && t == null && q.top - b > 0)
2385 rescan = true; // revisit
2386 break;
2387 }
2388 }
2389 }
2390 }
2391 }
2392 }
2393 }
2394 return s;
2395 }
2396
2397 /**
2398 * Runs tasks until all workers are inactive and no tasks are
2399 * found. Rather than blocking when tasks cannot be found, rescans
2400 * until all others cannot find tasks either.
2401 *
2402 * @param nanos max wait time (Long.MAX_VALUE if effectively untimed)
2403 * @param interruptible true if return on interrupt
2404 * @return positive if quiescent, negative if interrupted, else 0
2405 */
2406 private int helpQuiesce(WorkQueue w, long nanos, boolean interruptible) {
2407 int phase; // w.phase inactive bit set when temporarily quiescent
2408 if (w == null || ((phase = w.phase) & IDLE) != 0)
2409 return 0;
2410 int wsrc = w.source;
2411 long startTime = System.nanoTime();
2412 long maxSleep = Math.min(nanos >>> 8, MAX_SLEEP); // approx 1% nanos
2413 long prevSum = 0L;
2414 int activePhase = phase, inactivePhase = phase + IDLE;
2415 int r = phase + 1, waits = 0, returnStatus = 1;
2416 boolean locals = true;
2417 for (long e = runState;;) {
2418 if ((e & STOP) != 0L)
2419 break; // terminating
2420 if (interruptible && Thread.interrupted()) {
2421 returnStatus = -1;
2422 break;
2423 }
2424 if (locals) { // run local tasks before (re)polling
2425 locals = false;
2426 for (ForkJoinTask<?> u; (u = w.nextLocalTask()) != null;)
2427 u.doExec();
2428 }
2429 WorkQueue[] qs = queues;
2430 int n = (qs == null) ? 0 : qs.length;
2431 long phaseSum = 0L;
2432 boolean rescan = false, busy = false;
2433 scan: for (int l = n; l > 0; --l, ++r) {
2434 int j; WorkQueue q;
2435 if ((q = qs[j = r & SMASK & (n - 1)]) != null && q != w) {
2436 for (;;) {
2437 ForkJoinTask<?> t; ForkJoinTask<?>[] a;
2438 int b, cap; long k;
2439 if ((a = q.array) == null || (cap = a.length) <= 0)
2440 break;
2441 t = (ForkJoinTask<?>)U.getReferenceAcquire(
2442 a, k = slotOffset((cap - 1) & (b = q.base)));
2443 if (t != null && phase == inactivePhase) // reactivate
2444 w.phase = phase = activePhase;
2445 if (q.base == b && U.getReference(a, k) == t) {
2446 int nb = b + 1;
2447 if (t == null) {
2448 if (!rescan) {
2449 int qp = q.phase, mq = qp & (IDLE | 1);
2450 phaseSum += qp;
2451 if (mq == 0 || q.top - b > 0)
2452 rescan = true;
2453 else if (mq == 1)
2454 busy = true;
2455 }
2456 break;
2457 }
2458 if (U.compareAndSetReference(a, k, t, null)) {
2459 q.base = nb;
2460 w.source = j; // volatile write
2461 t.doExec();
2462 w.source = wsrc;
2463 rescan = locals = true;
2464 break scan;
2465 }
2466 }
2467 }
2468 }
2469 }
2470 if (e != (e = runState) || prevSum != (prevSum = phaseSum) ||
2471 rescan || (e & RS_LOCK) != 0L)
2472 ; // inconsistent
2473 else if (!busy)
2474 break;
2475 else if (phase == activePhase) {
2476 waits = 0; // recheck, then sleep
2477 w.phase = phase = inactivePhase;
2478 }
2479 else if (System.nanoTime() - startTime > nanos) {
2480 returnStatus = 0; // timed out
2481 break;
2482 }
2483 else if (waits == 0) // same as spinLockRunState except
2484 waits = MIN_SLEEP; // with rescan instead of onSpinWait
2485 else {
2486 LockSupport.parkNanos(this, (long)waits);
2487 if (waits < maxSleep)
2488 waits <<= 1;
2489 }
2490 }
2491 w.phase = activePhase;
2492 return returnStatus;
2493 }
2494
2495 /**
2496 * Helps quiesce from external caller until done, interrupted, or timeout
2497 *
2498 * @param nanos max wait time (Long.MAX_VALUE if effectively untimed)
2499 * @param interruptible true if return on interrupt
2500 * @return positive if quiescent, negative if interrupted, else 0
2501 */
2502 private int externalHelpQuiesce(long nanos, boolean interruptible) {
2503 if (quiescent() < 0) {
2504 long startTime = System.nanoTime();
2505 long maxSleep = Math.min(nanos >>> 8, MAX_SLEEP);
2506 for (int waits = 0;;) {
2507 ForkJoinTask<?> t;
2508 if (interruptible && Thread.interrupted())
2509 return -1;
2510 else if ((t = pollScan(false)) != null) {
2511 waits = 0;
2512 t.doExec();
2513 }
2514 else if (quiescent() >= 0)
2515 break;
2516 else if (System.nanoTime() - startTime > nanos)
2517 return 0;
2518 else if (waits == 0)
2519 waits = MIN_SLEEP;
2520 else {
2521 LockSupport.parkNanos(this, (long)waits);
2522 if (waits < maxSleep)
2523 waits <<= 1;
2524 }
2525 }
2526 }
2527 return 1;
2528 }
2529
2530 /**
2531 * Helps quiesce from either internal or external caller
2532 *
2533 * @param pool the pool to use, or null if any
2534 * @param nanos max wait time (Long.MAX_VALUE if effectively untimed)
2535 * @param interruptible true if return on interrupt
2536 * @return positive if quiescent, negative if interrupted, else 0
2537 */
2538 static final int helpQuiescePool(ForkJoinPool pool, long nanos,
2539 boolean interruptible) {
2540 Thread t; ForkJoinPool p; ForkJoinWorkerThread wt;
2541 if ((t = Thread.currentThread()) instanceof ForkJoinWorkerThread &&
2542 (p = (wt = (ForkJoinWorkerThread)t).pool) != null &&
2543 (p == pool || pool == null))
2544 return p.helpQuiesce(wt.workQueue, nanos, interruptible);
2545 else if ((p = pool) != null || (p = common) != null)
2546 return p.externalHelpQuiesce(nanos, interruptible);
2547 else
2548 return 0;
2549 }
2550
2551 /**
2552 * Gets and removes a local or stolen task for the given worker.
2553 *
2554 * @return a task, if available
2555 */
2556 final ForkJoinTask<?> nextTaskFor(WorkQueue w) {
2557 ForkJoinTask<?> t;
2558 if (w == null || (t = w.nextLocalTask()) == null)
2559 t = pollScan(false);
2560 return t;
2561 }
2562
2563 // External operations
2564
2565 /**
2566 * Finds and locks a WorkQueue for an external submitter, or
2567 * throws RejectedExecutionException if shutdown
2568 * @param rejectOnShutdown true if RejectedExecutionException
2569 * should be thrown when shutdown
2570 */
2571 final WorkQueue externalSubmissionQueue(boolean rejectOnShutdown) {
2572 int r;
2573 if ((r = ThreadLocalRandom.getProbe()) == 0) {
2574 ThreadLocalRandom.localInit(); // initialize caller's probe
2575 r = ThreadLocalRandom.getProbe();
2576 }
2577 for (;;) {
2578 WorkQueue q; WorkQueue[] qs; int n, id, i;
2579 if ((qs = queues) == null || (n = qs.length) <= 0)
2580 break;
2581 if ((q = qs[i = (id = r & EXTERNAL_ID_MASK) & (n - 1)]) == null) {
2582 WorkQueue newq = new WorkQueue(null, id, 0, false);
2583 lockRunState();
2584 if (qs[i] == null && queues == qs)
2585 q = qs[i] = newq; // else lost race to install
2586 unlockRunState();
2587 }
2588 if (q != null && q.tryLockPhase()) {
2589 if (rejectOnShutdown && (runState & SHUTDOWN) != 0L) {
2590 q.unlockPhase(); // check while q lock held
2591 break;
2592 }
2593 return q;
2594 }
2595 r = ThreadLocalRandom.advanceProbe(r); // move
2596 }
2597 throw new RejectedExecutionException();
2598 }
2599
2600 private <T> ForkJoinTask<T> poolSubmit(boolean signalIfEmpty, ForkJoinTask<T> task) {
2601 Thread t; ForkJoinWorkerThread wt; WorkQueue q; boolean internal;
2602 if (((t = JLA.currentCarrierThread()) instanceof ForkJoinWorkerThread) &&
2603 (wt = (ForkJoinWorkerThread)t).pool == this) {
2604 internal = true;
2605 q = wt.workQueue;
2606 }
2607 else { // find and lock queue
2608 internal = false;
2609 q = externalSubmissionQueue(true);
2610 }
2611 q.push(task, signalIfEmpty ? this : null, internal);
2612 return task;
2613 }
2614
2615 /**
2616 * Returns queue for an external thread, if one exists that has
2617 * possibly ever submitted to the given pool (nonzero probe), or
2618 * null if none.
2619 */
2620 static WorkQueue externalQueue(ForkJoinPool p) {
2621 WorkQueue[] qs; int n;
2622 int r = ThreadLocalRandom.getProbe();
2623 return (p != null && (qs = p.queues) != null &&
2624 (n = qs.length) > 0 && r != 0) ?
2625 qs[r & EXTERNAL_ID_MASK & (n - 1)] : null;
2626 }
2627
2628 /**
2629 * Returns external queue for common pool.
2630 */
2631 static WorkQueue commonQueue() {
2632 return externalQueue(common);
2633 }
2634
2635 /**
2636 * If the given executor is a ForkJoinPool, poll and execute
2637 * AsynchronousCompletionTasks from worker's queue until none are
2638 * available or blocker is released.
2639 */
2640 static void helpAsyncBlocker(Executor e, ManagedBlocker blocker) {
2641 WorkQueue w = null; Thread t; ForkJoinWorkerThread wt;
2642 if (((t = Thread.currentThread()) instanceof ForkJoinWorkerThread) &&
2643 (wt = (ForkJoinWorkerThread)t).pool == e)
2644 w = wt.workQueue;
2645 else if (e instanceof ForkJoinPool)
2646 w = externalQueue((ForkJoinPool)e);
2647 if (w != null)
2648 w.helpAsyncBlocker(blocker);
2649 }
2650
2651 /**
2652 * Returns a cheap heuristic guide for task partitioning when
2653 * programmers, frameworks, tools, or languages have little or no
2654 * idea about task granularity. In essence, by offering this
2655 * method, we ask users only about tradeoffs in overhead vs
2656 * expected throughput and its variance, rather than how finely to
2657 * partition tasks.
2658 *
2659 * In a steady state strict (tree-structured) computation, each
2660 * thread makes available for stealing enough tasks for other
2661 * threads to remain active. Inductively, if all threads play by
2662 * the same rules, each thread should make available only a
2663 * constant number of tasks.
2664 *
2665 * The minimum useful constant is just 1. But using a value of 1
2666 * would require immediate replenishment upon each steal to
2667 * maintain enough tasks, which is infeasible. Further,
2668 * partitionings/granularities of offered tasks should minimize
2669 * steal rates, which in general means that threads nearer the top
2670 * of computation tree should generate more than those nearer the
2671 * bottom. In perfect steady state, each thread is at
2672 * approximately the same level of computation tree. However,
2673 * producing extra tasks amortizes the uncertainty of progress and
2674 * diffusion assumptions.
2675 *
2676 * So, users will want to use values larger (but not much larger)
2677 * than 1 to both smooth over transient shortages and hedge
2678 * against uneven progress; as traded off against the cost of
2679 * extra task overhead. We leave the user to pick a threshold
2680 * value to compare with the results of this call to guide
2681 * decisions, but recommend values such as 3.
2682 *
2683 * When all threads are active, it is on average OK to estimate
2684 * surplus strictly locally. In steady-state, if one thread is
2685 * maintaining say 2 surplus tasks, then so are others. So we can
2686 * just use estimated queue length. However, this strategy alone
2687 * leads to serious mis-estimates in some non-steady-state
2688 * conditions (ramp-up, ramp-down, other stalls). We can detect
2689 * many of these by further considering the number of "idle"
2690 * threads, that are known to have zero queued tasks, so
2691 * compensate by a factor of (#idle/#active) threads.
2692 */
2693 static int getSurplusQueuedTaskCount() {
2694 Thread t; ForkJoinWorkerThread wt; ForkJoinPool pool; WorkQueue q;
2695 if (((t = Thread.currentThread()) instanceof ForkJoinWorkerThread) &&
2696 (pool = (wt = (ForkJoinWorkerThread)t).pool) != null &&
2697 (q = wt.workQueue) != null) {
2698 int n = q.top - q.base;
2699 int p = pool.parallelism;
2700 int a = (short)(pool.ctl >>> RC_SHIFT);
2701 return n - (a > (p >>>= 1) ? 0 :
2702 a > (p >>>= 1) ? 1 :
2703 a > (p >>>= 1) ? 2 :
2704 a > (p >>>= 1) ? 4 :
2705 8);
2706 }
2707 return 0;
2708 }
2709
2710 // Termination
2711
2712 /**
2713 * Possibly initiates and/or completes pool termination.
2714 *
2715 * @param now if true, unconditionally terminate, else only
2716 * if no work and no active workers
2717 * @param enable if true, terminate when next possible
2718 * @return runState on exit
2719 */
2720 private long tryTerminate(boolean now, boolean enable) {
2721 long e, isShutdown, ps;
2722 if (((e = runState) & TERMINATED) != 0L)
2723 now = false;
2724 else if ((e & STOP) != 0L)
2725 now = true;
2726 else if (now) {
2727 if (((ps = getAndBitwiseOrRunState(SHUTDOWN|STOP) & STOP)) == 0L) {
2728 if ((ps & RS_LOCK) != 0L) {
2729 spinLockRunState(); // ensure queues array stable after stop
2730 unlockRunState();
2731 }
2732 interruptAll();
2733 }
2734 }
2735 else if ((isShutdown = (e & SHUTDOWN)) != 0L || enable) {
2736 long quiet; DelayScheduler ds;
2737 if (isShutdown == 0L)
2738 getAndBitwiseOrRunState(SHUTDOWN);
2739 if ((quiet = quiescent()) > 0)
2740 now = true;
2741 else if (quiet == 0 && (ds = delayScheduler) != null)
2742 ds.signal();
2743 }
2744
2745 if (now) {
2746 DelayScheduler ds;
2747 releaseWaiters();
2748 if ((ds = delayScheduler) != null)
2749 ds.signal();
2750 for (;;) {
2751 if (((e = runState) & CLEANED) == 0L) {
2752 boolean clean = cleanQueues();
2753 if (((e = runState) & CLEANED) == 0L && clean)
2754 e = getAndBitwiseOrRunState(CLEANED) | CLEANED;
2755 }
2756 if ((e & TERMINATED) != 0L)
2757 break;
2758 if (ctl != 0L) // else loop if didn't finish cleaning
2759 break;
2760 if ((ds = delayScheduler) != null && ds.signal() >= 0)
2761 break;
2762 if ((e & CLEANED) != 0L) {
2763 e |= TERMINATED;
2764 if ((getAndBitwiseOrRunState(TERMINATED) & TERMINATED) == 0L) {
2765 CountDownLatch done; SharedThreadContainer ctr;
2766 if ((done = termination) != null)
2767 done.countDown();
2768 if ((ctr = container) != null)
2769 ctr.close();
2770 }
2771 break;
2772 }
2773 }
2774 }
2775 return e;
2776 }
2777
2778 /**
2779 * Scans queues in a psuedorandom order based on thread id,
2780 * cancelling tasks until empty, or returning early upon
2781 * interference or still-active external queues, in which case
2782 * other calls will finish cancellation.
2783 *
2784 * @return true if all queues empty
2785 */
2786 private boolean cleanQueues() {
2787 int r = (int)Thread.currentThread().threadId();
2788 r ^= r << 13; r ^= r >>> 17; r ^= r << 5; // xorshift
2789 int step = (r >>> 16) | 1; // randomize traversals
2790 WorkQueue[] qs = queues;
2791 int n = (qs == null) ? 0 : qs.length;
2792 for (int l = n; l > 0; --l, r += step) {
2793 WorkQueue q; ForkJoinTask<?>[] a; int cap;
2794 if ((q = qs[r & (n - 1)]) != null &&
2795 (a = q.array) != null && (cap = a.length) > 0) {
2796 for (;;) {
2797 ForkJoinTask<?> t; int b; long k;
2798 t = (ForkJoinTask<?>)U.getReferenceAcquire(
2799 a, k = slotOffset((cap - 1) & (b = q.base)));
2800 if (q.base == b && t != null &&
2801 U.compareAndSetReference(a, k, t, null)) {
2802 q.updateBase(b + 1);
2803 try {
2804 t.cancel(false);
2805 } catch (Throwable ignore) {
2806 }
2807 }
2808 else if ((q.phase & (IDLE|1)) == 0 || // externally locked
2809 q.top - q.base > 0)
2810 return false; // incomplete
2811 else
2812 break;
2813 }
2814 }
2815 }
2816 return true;
2817 }
2818
2819 /**
2820 * Interrupts all workers
2821 */
2822 private void interruptAll() {
2823 Thread current = Thread.currentThread();
2824 WorkQueue[] qs = queues;
2825 int n = (qs == null) ? 0 : qs.length;
2826 for (int i = 1; i < n; i += 2) {
2827 WorkQueue q; Thread o;
2828 if ((q = qs[i]) != null && (o = q.owner) != null && o != current) {
2829 try {
2830 o.interrupt();
2831 } catch (Throwable ignore) {
2832 }
2833 }
2834 }
2835 }
2836
2837 /**
2838 * Returns termination signal, constructing if necessary
2839 */
2840 private CountDownLatch terminationSignal() {
2841 CountDownLatch signal, s, u;
2842 if ((signal = termination) == null)
2843 signal = ((u = cmpExTerminationSignal(
2844 s = new CountDownLatch(1))) == null) ? s : u;
2845 return signal;
2846 }
2847
2848 // Exported methods
2849
2850 // Constructors
2851
2852 /**
2853 * Creates a {@code ForkJoinPool} with parallelism equal to {@link
2854 * java.lang.Runtime#availableProcessors}, using defaults for all
2855 * other parameters (see {@link #ForkJoinPool(int,
2856 * ForkJoinWorkerThreadFactory, UncaughtExceptionHandler, boolean,
2857 * int, int, int, Predicate, long, TimeUnit)}).
2858 */
2859 public ForkJoinPool() {
2860 this(Math.min(MAX_CAP, Runtime.getRuntime().availableProcessors()),
2861 defaultForkJoinWorkerThreadFactory, null, false,
2862 0, MAX_CAP, 1, null, DEFAULT_KEEPALIVE, TimeUnit.MILLISECONDS);
2863 }
2864
2865 /**
2866 * Creates a {@code ForkJoinPool} with the indicated parallelism
2867 * level, using defaults for all other parameters (see {@link
2868 * #ForkJoinPool(int, ForkJoinWorkerThreadFactory,
2869 * UncaughtExceptionHandler, boolean, int, int, int, Predicate,
2870 * long, TimeUnit)}).
2871 *
2872 * @param parallelism the parallelism level
2873 * @throws IllegalArgumentException if parallelism less than or
2874 * equal to zero, or greater than implementation limit
2875 */
2876 public ForkJoinPool(int parallelism) {
2877 this(parallelism, defaultForkJoinWorkerThreadFactory, null, false,
2878 0, MAX_CAP, 1, null, DEFAULT_KEEPALIVE, TimeUnit.MILLISECONDS);
2879 }
2880
2881 /**
2882 * Creates a {@code ForkJoinPool} with the given parameters (using
2883 * defaults for others -- see {@link #ForkJoinPool(int,
2884 * ForkJoinWorkerThreadFactory, UncaughtExceptionHandler, boolean,
2885 * int, int, int, Predicate, long, TimeUnit)}).
2886 *
2887 * @param parallelism the parallelism level. For default value,
2888 * use {@link java.lang.Runtime#availableProcessors}.
2889 * @param factory the factory for creating new threads. For default value,
2890 * use {@link #defaultForkJoinWorkerThreadFactory}.
2891 * @param handler the handler for internal worker threads that
2892 * terminate due to unrecoverable errors encountered while executing
2893 * tasks. For default value, use {@code null}.
2894 * @param asyncMode if true,
2895 * establishes local first-in-first-out scheduling mode for forked
2896 * tasks that are never joined. This mode may be more appropriate
2897 * than default locally stack-based mode in applications in which
2898 * worker threads only process event-style asynchronous tasks.
2899 * For default value, use {@code false}.
2900 * @throws IllegalArgumentException if parallelism less than or
2901 * equal to zero, or greater than implementation limit
2902 * @throws NullPointerException if the factory is null
2903 */
2904 public ForkJoinPool(int parallelism,
2905 ForkJoinWorkerThreadFactory factory,
2906 UncaughtExceptionHandler handler,
2907 boolean asyncMode) {
2908 this(parallelism, factory, handler, asyncMode,
2909 0, MAX_CAP, 1, null, DEFAULT_KEEPALIVE, TimeUnit.MILLISECONDS);
2910 }
2911
2912 /**
2913 * Creates a {@code ForkJoinPool} with the given parameters.
2914 *
2915 * @param parallelism the parallelism level. For default value,
2916 * use {@link java.lang.Runtime#availableProcessors}.
2917 *
2918 * @param factory the factory for creating new threads. For
2919 * default value, use {@link #defaultForkJoinWorkerThreadFactory}.
2920 *
2921 * @param handler the handler for internal worker threads that
2922 * terminate due to unrecoverable errors encountered while
2923 * executing tasks. For default value, use {@code null}.
2924 *
2925 * @param asyncMode if true, establishes local first-in-first-out
2926 * scheduling mode for forked tasks that are never joined. This
2927 * mode may be more appropriate than default locally stack-based
2928 * mode in applications in which worker threads only process
2929 * event-style asynchronous tasks. For default value, use {@code
2930 * false}.
2931 *
2932 * @param corePoolSize ignored: used in previous releases of this
2933 * class but no longer applicable. Using {@code 0} maintains
2934 * compatibility across releases.
2935 *
2936 * @param maximumPoolSize the maximum number of threads allowed.
2937 * When the maximum is reached, attempts to replace blocked
2938 * threads fail. (However, because creation and termination of
2939 * different threads may overlap, and may be managed by the given
2940 * thread factory, this value may be transiently exceeded.) To
2941 * arrange the same value as is used by default for the common
2942 * pool, use {@code 256} plus the {@code parallelism} level. (By
2943 * default, the common pool allows a maximum of 256 spare
2944 * threads.) Using a value (for example {@code
2945 * Integer.MAX_VALUE}) larger than the implementation's total
2946 * thread limit has the same effect as using this limit (which is
2947 * the default).
2948 *
2949 * @param minimumRunnable the minimum allowed number of core
2950 * threads not blocked by a join or {@link ManagedBlocker}. To
2951 * ensure progress, when too few unblocked threads exist and
2952 * unexecuted tasks may exist, new threads are constructed, up to
2953 * the given maximumPoolSize. For the default value, use {@code
2954 * 1}, that ensures liveness. A larger value might improve
2955 * throughput in the presence of blocked activities, but might
2956 * not, due to increased overhead. A value of zero may be
2957 * acceptable when submitted tasks cannot have dependencies
2958 * requiring additional threads.
2959 *
2960 * @param saturate if non-null, a predicate invoked upon attempts
2961 * to create more than the maximum total allowed threads. By
2962 * default, when a thread is about to block on a join or {@link
2963 * ManagedBlocker}, but cannot be replaced because the
2964 * maximumPoolSize would be exceeded, a {@link
2965 * RejectedExecutionException} is thrown. But if this predicate
2966 * returns {@code true}, then no exception is thrown, so the pool
2967 * continues to operate with fewer than the target number of
2968 * runnable threads, which might not ensure progress.
2969 *
2970 * @param keepAliveTime the elapsed time since last use before
2971 * a thread is terminated (and then later replaced if needed).
2972 * For the default value, use {@code 60, TimeUnit.SECONDS}.
2973 *
2974 * @param unit the time unit for the {@code keepAliveTime} argument
2975 *
2976 * @throws IllegalArgumentException if parallelism is less than or
2977 * equal to zero, or is greater than implementation limit,
2978 * or if maximumPoolSize is less than parallelism,
2979 * of if the keepAliveTime is less than or equal to zero.
2980 * @throws NullPointerException if the factory is null
2981 * @since 9
2982 */
2983 public ForkJoinPool(int parallelism,
2984 ForkJoinWorkerThreadFactory factory,
2985 UncaughtExceptionHandler handler,
2986 boolean asyncMode,
2987 int corePoolSize,
2988 int maximumPoolSize,
2989 int minimumRunnable,
2990 Predicate<? super ForkJoinPool> saturate,
2991 long keepAliveTime,
2992 TimeUnit unit) {
2993 int p = parallelism;
2994 if (p <= 0 || p > MAX_CAP || p > maximumPoolSize || keepAliveTime <= 0L)
2995 throw new IllegalArgumentException();
2996 if (factory == null || unit == null)
2997 throw new NullPointerException();
2998 int size = Math.max(MIN_QUEUES_SIZE,
2999 1 << (33 - Integer.numberOfLeadingZeros(p - 1)));
3000 this.parallelism = p;
3001 this.factory = factory;
3002 this.ueh = handler;
3003 this.saturate = saturate;
3004 this.keepAlive = Math.max(unit.toMillis(keepAliveTime), TIMEOUT_SLOP);
3005 int maxSpares = Math.clamp(maximumPoolSize - p, 0, MAX_CAP);
3006 int minAvail = Math.clamp(minimumRunnable, 0, MAX_CAP);
3007 this.config = (((asyncMode ? FIFO : 0) & LMASK) |
3008 (((long)maxSpares) << TC_SHIFT) |
3009 (((long)minAvail) << RC_SHIFT));
3010 this.queues = new WorkQueue[size];
3011 String pid = Integer.toString(getAndAddPoolIds(1) + 1);
3012 String name = "ForkJoinPool-" + pid;
3013 this.poolName = name;
3014 this.workerNamePrefix = name + "-worker-";
3015 this.container = SharedThreadContainer.create(name);
3016 }
3017
3018 /**
3019 * Constructor for common pool using parameters possibly
3020 * overridden by system properties
3021 */
3022 private ForkJoinPool(byte forCommonPoolOnly) {
3023 String name = "ForkJoinPool.commonPool";
3024 ForkJoinWorkerThreadFactory fac = defaultForkJoinWorkerThreadFactory;
3025 UncaughtExceptionHandler handler = null;
3026 int maxSpares = DEFAULT_COMMON_MAX_SPARES;
3027 int pc = 0, preset = 0; // nonzero if size set as property
3028 try { // ignore exceptions in accessing/parsing properties
3029 String pp = System.getProperty
3030 ("java.util.concurrent.ForkJoinPool.common.parallelism");
3031 if (pp != null) {
3032 pc = Math.max(0, Integer.parseInt(pp));
3033 preset = PRESET_SIZE;
3034 }
3035 String ms = System.getProperty
3036 ("java.util.concurrent.ForkJoinPool.common.maximumSpares");
3037 if (ms != null)
3038 maxSpares = Math.clamp(Integer.parseInt(ms), 0, MAX_CAP);
3039 String sf = System.getProperty
3040 ("java.util.concurrent.ForkJoinPool.common.threadFactory");
3041 String sh = System.getProperty
3042 ("java.util.concurrent.ForkJoinPool.common.exceptionHandler");
3043 if (sf != null || sh != null) {
3044 ClassLoader ldr = ClassLoader.getSystemClassLoader();
3045 if (sf != null)
3046 fac = (ForkJoinWorkerThreadFactory)
3047 ldr.loadClass(sf).getConstructor().newInstance();
3048 if (sh != null)
3049 handler = (UncaughtExceptionHandler)
3050 ldr.loadClass(sh).getConstructor().newInstance();
3051 }
3052 } catch (Exception ignore) {
3053 }
3054 if (preset == 0)
3055 pc = Math.max(1, Runtime.getRuntime().availableProcessors() - 1);
3056 int p = Math.min(pc, MAX_CAP);
3057 int size = Math.max(MIN_QUEUES_SIZE,
3058 (p == 0) ? 1 :
3059 1 << (33 - Integer.numberOfLeadingZeros(p-1)));
3060 this.parallelism = p;
3061 this.config = ((preset & LMASK) | (((long)maxSpares) << TC_SHIFT) |
3062 (1L << RC_SHIFT));
3063 this.factory = fac;
3064 this.ueh = handler;
3065 this.keepAlive = DEFAULT_KEEPALIVE;
3066 this.saturate = null;
3067 this.workerNamePrefix = null;
3068 this.poolName = name;
3069 this.queues = new WorkQueue[size];
3070 this.container = SharedThreadContainer.create(name);
3071 }
3072
3073 /**
3074 * Returns the common pool instance. This pool is statically
3075 * constructed; its run state is unaffected by attempts to {@link
3076 * #shutdown} or {@link #shutdownNow}. However this pool and any
3077 * ongoing processing are automatically terminated upon program
3078 * {@link System#exit}. Any program that relies on asynchronous
3079 * task processing to complete before program termination should
3080 * invoke {@code commonPool().}{@link #awaitQuiescence awaitQuiescence},
3081 * before exit.
3082 *
3083 * @return the common pool instance
3084 * @since 1.8
3085 */
3086 public static ForkJoinPool commonPool() {
3087 // assert common != null : "static init error";
3088 return common;
3089 }
3090
3091 /**
3092 * Package-private access to commonPool overriding zero parallelism
3093 */
3094 static ForkJoinPool asyncCommonPool() {
3095 ForkJoinPool cp; int p;
3096 if ((p = (cp = common).parallelism) == 0)
3097 U.compareAndSetInt(cp, PARALLELISM, 0, 2);
3098 return cp;
3099 }
3100
3101 // Execution methods
3102
3103 /**
3104 * Performs the given task, returning its result upon completion.
3105 * If the computation encounters an unchecked Exception or Error,
3106 * it is rethrown as the outcome of this invocation. Rethrown
3107 * exceptions behave in the same way as regular exceptions, but,
3108 * when possible, contain stack traces (as displayed for example
3109 * using {@code ex.printStackTrace()}) of both the current thread
3110 * as well as the thread actually encountering the exception;
3111 * minimally only the latter.
3112 *
3113 * @param task the task
3114 * @param <T> the type of the task's result
3115 * @return the task's result
3116 * @throws NullPointerException if the task is null
3117 * @throws RejectedExecutionException if the task cannot be
3118 * scheduled for execution
3119 */
3120 public <T> T invoke(ForkJoinTask<T> task) {
3121 poolSubmit(true, Objects.requireNonNull(task));
3122 try {
3123 return task.join();
3124 } catch (RuntimeException | Error unchecked) {
3125 throw unchecked;
3126 } catch (Exception checked) {
3127 throw new RuntimeException(checked);
3128 }
3129 }
3130
3131 /**
3132 * Arranges for (asynchronous) execution of the given task.
3133 *
3134 * @param task the task
3135 * @throws NullPointerException if the task is null
3136 * @throws RejectedExecutionException if the task cannot be
3137 * scheduled for execution
3138 */
3139 public void execute(ForkJoinTask<?> task) {
3140 poolSubmit(true, Objects.requireNonNull(task));
3141 }
3142
3143 // AbstractExecutorService methods
3144
3145 /**
3146 * @throws NullPointerException if the task is null
3147 * @throws RejectedExecutionException if the task cannot be
3148 * scheduled for execution
3149 */
3150 @Override
3151 @SuppressWarnings("unchecked")
3152 public void execute(Runnable task) {
3153 poolSubmit(true, (Objects.requireNonNull(task) instanceof ForkJoinTask<?>)
3154 ? (ForkJoinTask<Void>) task // avoid re-wrap
3155 : new ForkJoinTask.RunnableExecuteAction(task));
3156 }
3157
3158 /**
3159 * Submits a ForkJoinTask for execution.
3160 *
3161 * @implSpec
3162 * This method is equivalent to {@link #externalSubmit(ForkJoinTask)}
3163 * when called from a thread that is not in this pool.
3164 *
3165 * @param task the task to submit
3166 * @param <T> the type of the task's result
3167 * @return the task
3168 * @throws NullPointerException if the task is null
3169 * @throws RejectedExecutionException if the task cannot be
3170 * scheduled for execution
3171 */
3172 public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) {
3173 return poolSubmit(true, Objects.requireNonNull(task));
3174 }
3175
3176 /**
3177 * @throws NullPointerException if the task is null
3178 * @throws RejectedExecutionException if the task cannot be
3179 * scheduled for execution
3180 */
3181 @Override
3182 public <T> ForkJoinTask<T> submit(Callable<T> task) {
3183 Objects.requireNonNull(task);
3184 return poolSubmit(
3185 true,
3186 (Thread.currentThread() instanceof ForkJoinWorkerThread) ?
3187 new ForkJoinTask.AdaptedCallable<T>(task) :
3188 new ForkJoinTask.AdaptedInterruptibleCallable<T>(task));
3189 }
3190
3191 /**
3192 * @throws NullPointerException if the task is null
3193 * @throws RejectedExecutionException if the task cannot be
3194 * scheduled for execution
3195 */
3196 @Override
3197 public <T> ForkJoinTask<T> submit(Runnable task, T result) {
3198 Objects.requireNonNull(task);
3199 return poolSubmit(
3200 true,
3201 (Thread.currentThread() instanceof ForkJoinWorkerThread) ?
3202 new ForkJoinTask.AdaptedRunnable<T>(task, result) :
3203 new ForkJoinTask.AdaptedInterruptibleRunnable<T>(task, result));
3204 }
3205
3206 /**
3207 * @throws NullPointerException if the task is null
3208 * @throws RejectedExecutionException if the task cannot be
3209 * scheduled for execution
3210 */
3211 @Override
3212 @SuppressWarnings("unchecked")
3213 public ForkJoinTask<?> submit(Runnable task) {
3214 Objects.requireNonNull(task);
3215 return poolSubmit(
3216 true,
3217 (task instanceof ForkJoinTask<?>) ?
3218 (ForkJoinTask<Void>) task : // avoid re-wrap
3219 ((Thread.currentThread() instanceof ForkJoinWorkerThread) ?
3220 new ForkJoinTask.AdaptedRunnable<Void>(task, null) :
3221 new ForkJoinTask.AdaptedInterruptibleRunnable<Void>(task, null)));
3222 }
3223
3224 /**
3225 * Submits the given task as if submitted from a non-{@code ForkJoinTask}
3226 * client. The task is added to a scheduling queue for submissions to the
3227 * pool even when called from a thread in the pool.
3228 *
3229 * @implSpec
3230 * This method is equivalent to {@link #submit(ForkJoinTask)} when called
3231 * from a thread that is not in this pool.
3232 *
3233 * @return the task
3234 * @param task the task to submit
3235 * @param <T> the type of the task's result
3236 * @throws NullPointerException if the task is null
3237 * @throws RejectedExecutionException if the task cannot be
3238 * scheduled for execution
3239 * @since 20
3240 */
3241 public <T> ForkJoinTask<T> externalSubmit(ForkJoinTask<T> task) {
3242 Objects.requireNonNull(task);
3243 externalSubmissionQueue(true).push(task, this, false);
3244 return task;
3245 }
3246
3247 /**
3248 * Submits the given task without guaranteeing that it will
3249 * eventually execute in the absence of available active threads.
3250 * In some contexts, this method may reduce contention and
3251 * overhead by relying on context-specific knowledge that existing
3252 * threads (possibly including the calling thread if operating in
3253 * this pool) will eventually be available to execute the task.
3254 *
3255 * @param task the task
3256 * @param <T> the type of the task's result
3257 * @return the task
3258 * @throws NullPointerException if the task is null
3259 * @throws RejectedExecutionException if the task cannot be
3260 * scheduled for execution
3261 * @since 19
3262 */
3263 public <T> ForkJoinTask<T> lazySubmit(ForkJoinTask<T> task) {
3264 return poolSubmit(false, Objects.requireNonNull(task));
3265 }
3266
3267 /**
3268 * Changes the target parallelism of this pool, controlling the
3269 * future creation, use, and termination of worker threads.
3270 * Applications include contexts in which the number of available
3271 * processors changes over time.
3272 *
3273 * @implNote This implementation restricts the maximum number of
3274 * running threads to 32767
3275 *
3276 * @param size the target parallelism level
3277 * @return the previous parallelism level.
3278 * @throws IllegalArgumentException if size is less than 1 or
3279 * greater than the maximum supported by this pool.
3280 * @throws UnsupportedOperationException this is the{@link
3281 * #commonPool()} and parallelism level was set by System
3282 * property {@systemProperty
3283 * java.util.concurrent.ForkJoinPool.common.parallelism}.
3284 * @since 19
3285 */
3286 public int setParallelism(int size) {
3287 if (size < 1 || size > MAX_CAP)
3288 throw new IllegalArgumentException();
3289 if ((config & PRESET_SIZE) != 0)
3290 throw new UnsupportedOperationException("Cannot override System property");
3291 return getAndSetParallelism(size);
3292 }
3293
3294 /**
3295 * Uninterrupible version of {@code invokeAll}. Executes the given
3296 * tasks, returning a list of Futures holding their status and
3297 * results when all complete, ignoring interrupts. {@link
3298 * Future#isDone} is {@code true} for each element of the returned
3299 * list. Note that a <em>completed</em> task could have
3300 * terminated either normally or by throwing an exception. The
3301 * results of this method are undefined if the given collection is
3302 * modified while this operation is in progress.
3303 *
3304 * @apiNote This method supports usages that previously relied on an
3305 * incompatible override of
3306 * {@link ExecutorService#invokeAll(java.util.Collection)}.
3307 *
3308 * @param tasks the collection of tasks
3309 * @param <T> the type of the values returned from the tasks
3310 * @return a list of Futures representing the tasks, in the same
3311 * sequential order as produced by the iterator for the
3312 * given task list, each of which has completed
3313 * @throws NullPointerException if tasks or any of its elements are {@code null}
3314 * @throws RejectedExecutionException if any task cannot be
3315 * scheduled for execution
3316 * @since 22
3317 */
3318 public <T> List<Future<T>> invokeAllUninterruptibly(Collection<? extends Callable<T>> tasks) {
3319 ArrayList<Future<T>> futures = new ArrayList<>(tasks.size());
3320 try {
3321 for (Callable<T> t : tasks) {
3322 ForkJoinTask<T> f = ForkJoinTask.adapt(t);
3323 futures.add(f);
3324 poolSubmit(true, f);
3325 }
3326 for (int i = futures.size() - 1; i >= 0; --i)
3327 ((ForkJoinTask<?>)futures.get(i)).quietlyJoin();
3328 return futures;
3329 } catch (Throwable t) {
3330 for (Future<T> e : futures)
3331 e.cancel(true);
3332 throw t;
3333 }
3334 }
3335
3336 /**
3337 * Common support for timed and untimed invokeAll
3338 */
3339 private <T> List<Future<T>> invokeAll(Collection<? extends Callable<T>> tasks,
3340 long deadline)
3341 throws InterruptedException {
3342 ArrayList<Future<T>> futures = new ArrayList<>(tasks.size());
3343 try {
3344 for (Callable<T> t : tasks) {
3345 ForkJoinTask<T> f = ForkJoinTask.adaptInterruptible(t);
3346 futures.add(f);
3347 poolSubmit(true, f);
3348 }
3349 for (int i = futures.size() - 1; i >= 0; --i)
3350 ((ForkJoinTask<?>)futures.get(i))
3351 .quietlyJoinPoolInvokeAllTask(deadline);
3352 return futures;
3353 } catch (Throwable t) {
3354 for (Future<T> e : futures)
3355 e.cancel(true);
3356 throw t;
3357 }
3358 }
3359
3360 @Override
3361 public <T> List<Future<T>> invokeAll(Collection<? extends Callable<T>> tasks)
3362 throws InterruptedException {
3363 return invokeAll(tasks, 0L);
3364 }
3365 // for jdk version < 22, replace with
3366 // /**
3367 // * @throws NullPointerException {@inheritDoc}
3368 // * @throws RejectedExecutionException {@inheritDoc}
3369 // */
3370 // @Override
3371 // public <T> List<Future<T>> invokeAll(Collection<? extends Callable<T>> tasks) {
3372 // return invokeAllUninterruptibly(tasks);
3373 // }
3374
3375 @Override
3376 public <T> List<Future<T>> invokeAll(Collection<? extends Callable<T>> tasks,
3377 long timeout, TimeUnit unit)
3378 throws InterruptedException {
3379 return invokeAll(tasks, (System.nanoTime() + unit.toNanos(timeout)) | 1L);
3380 }
3381
3382 @Override
3383 public <T> T invokeAny(Collection<? extends Callable<T>> tasks)
3384 throws InterruptedException, ExecutionException {
3385 try {
3386 return new ForkJoinTask.InvokeAnyRoot<T>()
3387 .invokeAny(tasks, this, false, 0L);
3388 } catch (TimeoutException cannotHappen) {
3389 assert false;
3390 return null;
3391 }
3392 }
3393
3394 @Override
3395 public <T> T invokeAny(Collection<? extends Callable<T>> tasks,
3396 long timeout, TimeUnit unit)
3397 throws InterruptedException, ExecutionException, TimeoutException {
3398 return new ForkJoinTask.InvokeAnyRoot<T>()
3399 .invokeAny(tasks, this, true, unit.toNanos(timeout));
3400 }
3401
3402 // Support for delayed tasks
3403
3404 /**
3405 * Returns STOP and SHUTDOWN status (zero if neither), masking or
3406 * truncating out other bits.
3407 */
3408 final int shutdownStatus(DelayScheduler ds) {
3409 return (int)(runState & (SHUTDOWN | STOP));
3410 }
3411
3412 /**
3413 * Tries to stop and possibly terminate if already enabled, return success.
3414 */
3415 final boolean tryStopIfShutdown(DelayScheduler ds) {
3416 return (tryTerminate(false, false) & STOP) != 0L;
3417 }
3418
3419 /**
3420 * Creates and starts DelayScheduler
3421 */
3422 private DelayScheduler startDelayScheduler() {
3423 DelayScheduler ds;
3424 if ((ds = delayScheduler) == null) {
3425 boolean start = false;
3426 String name = poolName + "-delayScheduler";
3427 if (workerNamePrefix == null)
3428 asyncCommonPool(); // override common parallelism zero
3429 long isShutdown = lockRunState() & SHUTDOWN;
3430 try {
3431 if (isShutdown == 0L && (ds = delayScheduler) == null) {
3432 ds = delayScheduler = new DelayScheduler(this, name);
3433 start = true;
3434 }
3435 } finally {
3436 unlockRunState();
3437 }
3438 if (start) { // start outside of lock
3439 SharedThreadContainer ctr;
3440 try {
3441 if ((ctr = container) != null)
3442 ctr.start(ds);
3443 else
3444 ds.start();
3445 } catch (RuntimeException | Error ex) { // back out
3446 lockRunState();
3447 ds = delayScheduler = null;
3448 unlockRunState();
3449 tryTerminate(false, false);
3450 if (ex instanceof Error)
3451 throw ex;
3452 }
3453 }
3454 }
3455 return ds;
3456 }
3457
3458 /**
3459 * Arranges execution of a ScheduledForkJoinTask whose delay has
3460 * elapsed
3461 */
3462 final void executeEnabledScheduledTask(ScheduledForkJoinTask<?> task) {
3463 externalSubmissionQueue(false).push(task, this, false);
3464 }
3465
3466 /**
3467 * Arranges delayed execution of a ScheduledForkJoinTask via the
3468 * DelayScheduler, creating and starting it if necessary.
3469 * @return the task
3470 */
3471 final <T> ScheduledForkJoinTask<T> scheduleDelayedTask(ScheduledForkJoinTask<T> task) {
3472 DelayScheduler ds;
3473 if (((ds = delayScheduler) == null &&
3474 (ds = startDelayScheduler()) == null) ||
3475 (runState & SHUTDOWN) != 0L)
3476 throw new RejectedExecutionException();
3477 ds.pend(task);
3478 return task;
3479 }
3480
3481 /**
3482 * Submits a one-shot task that becomes enabled for execution after the given
3483 * delay. At that point it will execute unless explicitly
3484 * cancelled, or fail to execute (eventually reporting
3485 * cancellation) when encountering resource exhaustion, or the
3486 * pool is {@link #shutdownNow}, or is {@link #shutdown} when
3487 * otherwise quiescent and {@link #cancelDelayedTasksOnShutdown}
3488 * is in effect.
3489 *
3490 * @param command the task to execute
3491 * @param delay the time from now to delay execution
3492 * @param unit the time unit of the delay parameter
3493 * @return a ForkJoinTask implementing the ScheduledFuture
3494 * interface, whose {@code get()} method will return
3495 * {@code null} upon normal completion.
3496 * @throws RejectedExecutionException if the pool is shutdown or
3497 * submission encounters resource exhaustion.
3498 * @throws NullPointerException if command or unit is null
3499 * @since 25
3500 */
3501 public ScheduledFuture<?> schedule(Runnable command,
3502 long delay, TimeUnit unit) {
3503 return scheduleDelayedTask(
3504 new ScheduledForkJoinTask<Void>(
3505 unit.toNanos(delay), 0L, false, // implicit null check of unit
3506 Objects.requireNonNull(command), null, this));
3507 }
3508
3509 /**
3510 * Submits a value-returning one-shot task that becomes enabled for execution
3511 * after the given delay. At that point it will execute unless
3512 * explicitly cancelled, or fail to execute (eventually reporting
3513 * cancellation) when encountering resource exhaustion, or the
3514 * pool is {@link #shutdownNow}, or is {@link #shutdown} when
3515 * otherwise quiescent and {@link #cancelDelayedTasksOnShutdown}
3516 * is in effect.
3517 *
3518 * @param callable the function to execute
3519 * @param delay the time from now to delay execution
3520 * @param unit the time unit of the delay parameter
3521 * @param <V> the type of the callable's result
3522 * @return a ForkJoinTask implementing the ScheduledFuture
3523 * interface, whose {@code get()} method will return the
3524 * value from the callable upon normal completion.
3525 * @throws RejectedExecutionException if the pool is shutdown or
3526 * submission encounters resource exhaustion.
3527 * @throws NullPointerException if command or unit is null
3528 * @since 25
3529 */
3530 public <V> ScheduledFuture<V> schedule(Callable<V> callable,
3531 long delay, TimeUnit unit) {
3532 return scheduleDelayedTask(
3533 new ScheduledForkJoinTask<V>(
3534 unit.toNanos(delay), 0L, false, null, // implicit null check of unit
3535 Objects.requireNonNull(callable), this));
3536 }
3537
3538 /**
3539 * Submits a periodic action that becomes enabled for execution first after the
3540 * given initial delay, and subsequently with the given period;
3541 * that is, executions will commence after
3542 * {@code initialDelay}, then {@code initialDelay + period}, then
3543 * {@code initialDelay + 2 * period}, and so on.
3544 *
3545 * <p>The sequence of task executions continues indefinitely until
3546 * one of the following exceptional completions occur:
3547 * <ul>
3548 * <li>The task is {@linkplain Future#cancel explicitly cancelled}
3549 * <li>Method {@link #shutdownNow} is called
3550 * <li>Method {@link #shutdown} is called and the pool is
3551 * otherwise quiescent, in which case existing executions continue
3552 * but subsequent executions do not.
3553 * <li>An execution or the task encounters resource exhaustion.
3554 * <li>An execution of the task throws an exception. In this case
3555 * calling {@link Future#get() get} on the returned future will throw
3556 * {@link ExecutionException}, holding the exception as its cause.
3557 * </ul>
3558 * Subsequent executions are suppressed. Subsequent calls to
3559 * {@link Future#isDone isDone()} on the returned future will
3560 * return {@code true}.
3561 *
3562 * <p>If any execution of this task takes longer than its period, then
3563 * subsequent executions may start late, but will not concurrently
3564 * execute.
3565 * @param command the task to execute
3566 * @param initialDelay the time to delay first execution
3567 * @param period the period between successive executions
3568 * @param unit the time unit of the initialDelay and period parameters
3569 * @return a ForkJoinTask implementing the ScheduledFuture
3570 * interface. The future's {@link Future#get() get()}
3571 * method will never return normally, and will throw an
3572 * exception upon task cancellation or abnormal
3573 * termination of a task execution.
3574 * @throws RejectedExecutionException if the pool is shutdown or
3575 * submission encounters resource exhaustion.
3576 * @throws NullPointerException if command or unit is null
3577 * @throws IllegalArgumentException if period less than or equal to zero
3578 * @since 25
3579 */
3580 public ScheduledFuture<?> scheduleAtFixedRate(Runnable command,
3581 long initialDelay,
3582 long period, TimeUnit unit) {
3583 if (period <= 0L)
3584 throw new IllegalArgumentException();
3585 return scheduleDelayedTask(
3586 new ScheduledForkJoinTask<Void>(
3587 unit.toNanos(initialDelay), // implicit null check of unit
3588 unit.toNanos(period), false,
3589 Objects.requireNonNull(command), null, this));
3590 }
3591
3592 /**
3593 * Submits a periodic action that becomes enabled for execution first after the
3594 * given initial delay, and subsequently with the given delay
3595 * between the termination of one execution and the commencement of
3596 * the next.
3597 * <p>The sequence of task executions continues indefinitely until
3598 * one of the following exceptional completions occur:
3599 * <ul>
3600 * <li>The task is {@linkplain Future#cancel explicitly cancelled}
3601 * <li>Method {@link #shutdownNow} is called
3602 * <li>Method {@link #shutdown} is called and the pool is
3603 * otherwise quiescent, in which case existing executions continue
3604 * but subsequent executions do not.
3605 * <li>An execution or the task encounters resource exhaustion.
3606 * <li>An execution of the task throws an exception. In this case
3607 * calling {@link Future#get() get} on the returned future will throw
3608 * {@link ExecutionException}, holding the exception as its cause.
3609 * </ul>
3610 * Subsequent executions are suppressed. Subsequent calls to
3611 * {@link Future#isDone isDone()} on the returned future will
3612 * return {@code true}.
3613 * @param command the task to execute
3614 * @param initialDelay the time to delay first execution
3615 * @param delay the delay between the termination of one
3616 * execution and the commencement of the next
3617 * @param unit the time unit of the initialDelay and delay parameters
3618 * @return a ForkJoinTask implementing the ScheduledFuture
3619 * interface. The future's {@link Future#get() get()}
3620 * method will never return normally, and will throw an
3621 * exception upon task cancellation or abnormal
3622 * termination of a task execution.
3623 * @throws RejectedExecutionException if the pool is shutdown or
3624 * submission encounters resource exhaustion.
3625 * @throws NullPointerException if command or unit is null
3626 * @throws IllegalArgumentException if delay less than or equal to zero
3627 * @since 25
3628 */
3629 public ScheduledFuture<?> scheduleWithFixedDelay(Runnable command,
3630 long initialDelay,
3631 long delay, TimeUnit unit) {
3632 if (delay <= 0L)
3633 throw new IllegalArgumentException();
3634 return scheduleDelayedTask(
3635 new ScheduledForkJoinTask<Void>(
3636 unit.toNanos(initialDelay), // implicit null check of unit
3637 -unit.toNanos(delay), false, // negative for fixed delay
3638 Objects.requireNonNull(command), null, this));
3639 }
3640
3641 /**
3642 * Body of a task performed on timeout of another task
3643 */
3644 static final class TimeoutAction<V> implements Runnable {
3645 // set after construction, nulled after use
3646 ForkJoinTask.CallableWithTimeout<V> task;
3647 Consumer<? super ForkJoinTask<V>> action;
3648 TimeoutAction(Consumer<? super ForkJoinTask<V>> action) {
3649 this.action = action;
3650 }
3651 public void run() {
3652 ForkJoinTask.CallableWithTimeout<V> t = task;
3653 Consumer<? super ForkJoinTask<V>> a = action;
3654 task = null;
3655 action = null;
3656 if (t != null && t.status >= 0) {
3657 if (a == null)
3658 t.cancel(true);
3659 else {
3660 a.accept(t);
3661 t.interruptIfRunning(true);
3662 }
3663 }
3664 }
3665 }
3666
3667 /**
3668 * Submits a task executing the given function, cancelling the
3669 * task or performing a given timeoutAction if not completed
3670 * within the given timeout period. If the optional {@code
3671 * timeoutAction} is null, the task is cancelled (via {@code
3672 * cancel(true)}. Otherwise, the action is applied and the task
3673 * may be interrupted if running. Actions may include {@link
3674 * ForkJoinTask#complete} to set a replacement value or {@link
3675 * ForkJoinTask#completeExceptionally} to throw an appropriate
3676 * exception. Note that these can succeed only if the task has
3677 * not already completed when the timeoutAction executes.
3678 *
3679 * @param callable the function to execute
3680 * @param <V> the type of the callable's result
3681 * @param timeout the time to wait before cancelling if not completed
3682 * @param timeoutAction if nonnull, an action to perform on
3683 * timeout, otherwise the default action is to cancel using
3684 * {@code cancel(true)}.
3685 * @param unit the time unit of the timeout parameter
3686 * @return a Future that can be used to extract result or cancel
3687 * @throws RejectedExecutionException if the task cannot be
3688 * scheduled for execution
3689 * @throws NullPointerException if callable or unit is null
3690 * @since 25
3691 */
3692 public <V> ForkJoinTask<V> submitWithTimeout(Callable<V> callable,
3693 long timeout, TimeUnit unit,
3694 Consumer<? super ForkJoinTask<V>> timeoutAction) {
3695 ForkJoinTask.CallableWithTimeout<V> task; TimeoutAction<V> onTimeout;
3696 Objects.requireNonNull(callable);
3697 ScheduledForkJoinTask<Void> timeoutTask =
3698 new ScheduledForkJoinTask<Void>(
3699 unit.toNanos(timeout), 0L, true,
3700 onTimeout = new TimeoutAction<V>(timeoutAction), null, this);
3701 onTimeout.task = task =
3702 new ForkJoinTask.CallableWithTimeout<V>(callable, timeoutTask);
3703 scheduleDelayedTask(timeoutTask);
3704 return poolSubmit(true, task);
3705 }
3706
3707 /**
3708 * Arranges that scheduled tasks that are not executing and have
3709 * not already been enabled for execution will not be executed and
3710 * will be cancelled upon {@link #shutdown} (unless this pool is
3711 * the {@link #commonPool()} which never shuts down). This method
3712 * may be invoked either before {@link #shutdown} to take effect
3713 * upon the next call, or afterwards to cancel such tasks, which
3714 * may then allow termination. Note that subsequent executions of
3715 * periodic tasks are always disabled upon shutdown, so this
3716 * method applies meaningfully only to non-periodic tasks.
3717 * @since 25
3718 */
3719 public void cancelDelayedTasksOnShutdown() {
3720 DelayScheduler ds;
3721 if ((ds = delayScheduler) != null ||
3722 (ds = startDelayScheduler()) != null)
3723 ds.cancelDelayedTasksOnShutdown();
3724 }
3725
3726 /**
3727 * Returns the factory used for constructing new workers.
3728 *
3729 * @return the factory used for constructing new workers
3730 */
3731 public ForkJoinWorkerThreadFactory getFactory() {
3732 return factory;
3733 }
3734
3735 /**
3736 * Returns the handler for internal worker threads that terminate
3737 * due to unrecoverable errors encountered while executing tasks.
3738 *
3739 * @return the handler, or {@code null} if none
3740 */
3741 public UncaughtExceptionHandler getUncaughtExceptionHandler() {
3742 return ueh;
3743 }
3744
3745 /**
3746 * Returns the targeted parallelism level of this pool.
3747 *
3748 * @return the targeted parallelism level of this pool
3749 */
3750 public int getParallelism() {
3751 return Math.max(getParallelismOpaque(), 1);
3752 }
3753
3754 /**
3755 * Returns the targeted parallelism level of the common pool.
3756 *
3757 * @return the targeted parallelism level of the common pool
3758 * @since 1.8
3759 */
3760 public static int getCommonPoolParallelism() {
3761 return common.getParallelism();
3762 }
3763
3764 /**
3765 * Returns the number of worker threads that have started but not
3766 * yet terminated. The result returned by this method may differ
3767 * from {@link #getParallelism} when threads are created to
3768 * maintain parallelism when others are cooperatively blocked.
3769 *
3770 * @return the number of worker threads
3771 */
3772 public int getPoolSize() {
3773 return (short)(ctl >>> TC_SHIFT);
3774 }
3775
3776 /**
3777 * Returns {@code true} if this pool uses local first-in-first-out
3778 * scheduling mode for forked tasks that are never joined.
3779 *
3780 * @return {@code true} if this pool uses async mode
3781 */
3782 public boolean getAsyncMode() {
3783 return (config & FIFO) != 0;
3784 }
3785
3786 /**
3787 * Returns an estimate of the number of worker threads that are
3788 * not blocked waiting to join tasks or for other managed
3789 * synchronization. This method may overestimate the
3790 * number of running threads.
3791 *
3792 * @return the number of worker threads
3793 */
3794 public int getRunningThreadCount() {
3795 WorkQueue[] qs; WorkQueue q;
3796 int rc = 0;
3797 if ((runState & TERMINATED) == 0L && (qs = queues) != null) {
3798 for (int i = 1; i < qs.length; i += 2) {
3799 if ((q = qs[i]) != null && q.isApparentlyUnblocked())
3800 ++rc;
3801 }
3802 }
3803 return rc;
3804 }
3805
3806 /**
3807 * Returns an estimate of the number of threads that are currently
3808 * stealing or executing tasks. This method may overestimate the
3809 * number of active threads.
3810 *
3811 * @return the number of active threads
3812 */
3813 public int getActiveThreadCount() {
3814 return Math.max((short)(ctl >>> RC_SHIFT), 0);
3815 }
3816
3817 /**
3818 * Returns {@code true} if all worker threads are currently idle.
3819 * An idle worker is one that cannot obtain a task to execute
3820 * because none are available to steal from other threads, and
3821 * there are no pending submissions to the pool. This method is
3822 * conservative; it might not return {@code true} immediately upon
3823 * idleness of all threads, but will eventually become true if
3824 * threads remain inactive.
3825 *
3826 * @return {@code true} if all threads are currently idle
3827 */
3828 public boolean isQuiescent() {
3829 return quiescent() >= 0;
3830 }
3831
3832 /**
3833 * Returns an estimate of the total number of completed tasks that
3834 * were executed by a thread other than their submitter. The
3835 * reported value underestimates the actual total number of steals
3836 * when the pool is not quiescent. This value may be useful for
3837 * monitoring and tuning fork/join programs: in general, steal
3838 * counts should be high enough to keep threads busy, but low
3839 * enough to avoid overhead and contention across threads.
3840 *
3841 * @return the number of steals
3842 */
3843 public long getStealCount() {
3844 long count = stealCount;
3845 WorkQueue[] qs; WorkQueue q;
3846 if ((qs = queues) != null) {
3847 for (int i = 1; i < qs.length; i += 2) {
3848 if ((q = qs[i]) != null)
3849 count += (long)q.nsteals & 0xffffffffL;
3850 }
3851 }
3852 return count;
3853 }
3854
3855 /**
3856 * Returns an estimate of the total number of tasks currently held
3857 * in queues by worker threads (but not including tasks submitted
3858 * to the pool that have not begun executing). This value is only
3859 * an approximation, obtained by iterating across all threads in
3860 * the pool. This method may be useful for tuning task
3861 * granularities.The returned count does not include scheduled
3862 * tasks that are not yet ready to execute, which are reported
3863 * separately by method {@link getDelayedTaskCount}.
3864 *
3865 * @return the number of queued tasks
3866 * @see ForkJoinWorkerThread#getQueuedTaskCount()
3867 */
3868 public long getQueuedTaskCount() {
3869 WorkQueue[] qs; WorkQueue q;
3870 long count = 0;
3871 if ((runState & TERMINATED) == 0L && (qs = queues) != null) {
3872 for (int i = 1; i < qs.length; i += 2) {
3873 if ((q = qs[i]) != null)
3874 count += q.queueSize();
3875 }
3876 }
3877 return count;
3878 }
3879
3880 /**
3881 * Returns an estimate of the number of tasks submitted to this
3882 * pool that have not yet begun executing. This method may take
3883 * time proportional to the number of submissions.
3884 *
3885 * @return the number of queued submissions
3886 */
3887 public int getQueuedSubmissionCount() {
3888 WorkQueue[] qs; WorkQueue q;
3889 int count = 0;
3890 if ((runState & TERMINATED) == 0L && (qs = queues) != null) {
3891 for (int i = 0; i < qs.length; i += 2) {
3892 if ((q = qs[i]) != null)
3893 count += q.queueSize();
3894 }
3895 }
3896 return count;
3897 }
3898
3899 /**
3900 * Returns an estimate of the number of delayed (including
3901 * periodic) tasks scheduled in this pool that are not yet ready
3902 * to submit for execution. The returned value is inaccurate while
3903 * delayed tasks are being processed.
3904 *
3905 * @return an estimate of the number of delayed tasks
3906 * @since 25
3907 */
3908 public long getDelayedTaskCount() {
3909 DelayScheduler ds;
3910 return ((ds = delayScheduler) == null ? 0 : ds.lastStableSize());
3911 }
3912
3913 /**
3914 * Returns {@code true} if there are any tasks submitted to this
3915 * pool that have not yet begun executing.
3916 *
3917 * @return {@code true} if there are any queued submissions
3918 */
3919 public boolean hasQueuedSubmissions() {
3920 WorkQueue[] qs; WorkQueue q;
3921 if ((runState & STOP) == 0L && (qs = queues) != null) {
3922 for (int i = 0; i < qs.length; i += 2) {
3923 if ((q = qs[i]) != null && q.queueSize() > 0)
3924 return true;
3925 }
3926 }
3927 return false;
3928 }
3929
3930 /**
3931 * Removes and returns the next unexecuted submission if one is
3932 * available. This method may be useful in extensions to this
3933 * class that re-assign work in systems with multiple pools.
3934 *
3935 * @return the next submission, or {@code null} if none
3936 */
3937 protected ForkJoinTask<?> pollSubmission() {
3938 return pollScan(true);
3939 }
3940
3941 /**
3942 * Removes all available unexecuted submitted and forked tasks
3943 * from scheduling queues and adds them to the given collection,
3944 * without altering their execution status. These may include
3945 * artificially generated or wrapped tasks. This method is
3946 * designed to be invoked only when the pool is known to be
3947 * quiescent. Invocations at other times may not remove all
3948 * tasks. A failure encountered while attempting to add elements
3949 * to collection {@code c} may result in elements being in
3950 * neither, either or both collections when the associated
3951 * exception is thrown. The behavior of this operation is
3952 * undefined if the specified collection is modified while the
3953 * operation is in progress.
3954 *
3955 * @param c the collection to transfer elements into
3956 * @return the number of elements transferred
3957 */
3958 protected int drainTasksTo(Collection<? super ForkJoinTask<?>> c) {
3959 int count = 0;
3960 for (ForkJoinTask<?> t; (t = pollScan(false)) != null; ) {
3961 c.add(t);
3962 ++count;
3963 }
3964 return count;
3965 }
3966
3967 /**
3968 * Returns a string identifying this pool, as well as its state,
3969 * including indications of run state, parallelism level, and
3970 * worker and task counts.
3971 *
3972 * @return a string identifying this pool, as well as its state
3973 */
3974 public String toString() {
3975 // Use a single pass through queues to collect counts
3976 DelayScheduler ds;
3977 long e = runState;
3978 long st = stealCount;
3979 long qt = 0L, ss = 0L; int rc = 0;
3980 WorkQueue[] qs; WorkQueue q;
3981 if ((qs = queues) != null) {
3982 for (int i = 0; i < qs.length; ++i) {
3983 if ((q = qs[i]) != null) {
3984 int size = q.queueSize();
3985 if ((i & 1) == 0)
3986 ss += size;
3987 else {
3988 qt += size;
3989 st += (long)q.nsteals & 0xffffffffL;
3990 if (q.isApparentlyUnblocked())
3991 ++rc;
3992 }
3993 }
3994 }
3995 }
3996 String delayed = ((ds = delayScheduler) == null ? "" :
3997 ", delayed = " + ds.lastStableSize());
3998 int pc = parallelism;
3999 long c = ctl;
4000 int tc = (short)(c >>> TC_SHIFT);
4001 int ac = (short)(c >>> RC_SHIFT);
4002 if (ac < 0) // ignore transient negative
4003 ac = 0;
4004 String level = ((e & TERMINATED) != 0L ? "Terminated" :
4005 (e & STOP) != 0L ? "Terminating" :
4006 (e & SHUTDOWN) != 0L ? "Shutting down" :
4007 "Running");
4008 return super.toString() +
4009 "[" + level +
4010 ", parallelism = " + pc +
4011 ", size = " + tc +
4012 ", active = " + ac +
4013 ", running = " + rc +
4014 ", steals = " + st +
4015 ", tasks = " + qt +
4016 ", submissions = " + ss +
4017 delayed +
4018 "]";
4019 }
4020
4021 /**
4022 * Possibly initiates an orderly shutdown in which previously
4023 * submitted tasks are executed, but no new tasks will be
4024 * accepted. Invocation has no effect on execution state if this
4025 * is the {@link #commonPool()}, and no additional effect if
4026 * already shut down. Tasks that are in the process of being
4027 * submitted concurrently during the course of this method may or
4028 * may not be rejected.
4029 */
4030 public void shutdown() {
4031 if (workerNamePrefix != null) // not common pool
4032 tryTerminate(false, true);
4033 }
4034
4035 /**
4036 * Possibly attempts to cancel and/or stop all tasks, and reject
4037 * all subsequently submitted tasks. Invocation has no effect on
4038 * execution state if this is the {@link #commonPool()}, and no
4039 * additional effect if already shut down. Otherwise, tasks that
4040 * are in the process of being submitted or executed concurrently
4041 * during the course of this method may or may not be
4042 * rejected. This method cancels both existing and unexecuted
4043 * tasks, in order to permit termination in the presence of task
4044 * dependencies. So the method always returns an empty list
4045 * (unlike the case for some other Executors).
4046 *
4047 * @return an empty list
4048 */
4049 public List<Runnable> shutdownNow() {
4050 if (workerNamePrefix != null) // not common pool
4051 tryTerminate(true, true);
4052 return Collections.emptyList();
4053 }
4054
4055 /**
4056 * Returns {@code true} if all tasks have completed following shut down.
4057 *
4058 * @return {@code true} if all tasks have completed following shut down
4059 */
4060 public boolean isTerminated() {
4061 return (tryTerminate(false, false) & TERMINATED) != 0;
4062 }
4063
4064 /**
4065 * Returns {@code true} if the process of termination has
4066 * commenced but not yet completed. This method may be useful for
4067 * debugging. A return of {@code true} reported a sufficient
4068 * period after shutdown may indicate that submitted tasks have
4069 * ignored or suppressed interruption, or are waiting for I/O,
4070 * causing this executor not to properly terminate. (See the
4071 * advisory notes for class {@link ForkJoinTask} stating that
4072 * tasks should not normally entail blocking operations. But if
4073 * they do, they must abort them on interrupt.)
4074 *
4075 * @return {@code true} if terminating but not yet terminated
4076 */
4077 public boolean isTerminating() {
4078 return (tryTerminate(false, false) & (STOP | TERMINATED)) == STOP;
4079 }
4080
4081 /**
4082 * Returns {@code true} if this pool has been shut down.
4083 *
4084 * @return {@code true} if this pool has been shut down
4085 */
4086 public boolean isShutdown() {
4087 return (runState & SHUTDOWN) != 0L;
4088 }
4089
4090 /**
4091 * Blocks until all tasks have completed execution after a
4092 * shutdown request, or the timeout occurs, or the current thread
4093 * is interrupted, whichever happens first. Because the {@link
4094 * #commonPool()} never terminates until program shutdown, when
4095 * applied to the common pool, this method is equivalent to {@link
4096 * #awaitQuiescence(long, TimeUnit)} but always returns {@code false}.
4097 *
4098 * @param timeout the maximum time to wait
4099 * @param unit the time unit of the timeout argument
4100 * @return {@code true} if this executor terminated and
4101 * {@code false} if the timeout elapsed before termination
4102 * @throws InterruptedException if interrupted while waiting
4103 */
4104 public boolean awaitTermination(long timeout, TimeUnit unit)
4105 throws InterruptedException {
4106 long nanos = unit.toNanos(timeout);
4107 CountDownLatch done;
4108 if (workerNamePrefix == null) { // is common pool
4109 if (helpQuiescePool(this, nanos, true) < 0)
4110 throw new InterruptedException();
4111 return false;
4112 }
4113 else if ((tryTerminate(false, false) & TERMINATED) != 0 ||
4114 (done = terminationSignal()) == null ||
4115 (runState & TERMINATED) != 0L)
4116 return true;
4117 else
4118 return done.await(nanos, TimeUnit.NANOSECONDS);
4119 }
4120
4121 /**
4122 * If called by a ForkJoinTask operating in this pool, equivalent
4123 * in effect to {@link ForkJoinTask#helpQuiesce}. Otherwise,
4124 * waits and/or attempts to assist performing tasks until this
4125 * pool {@link #isQuiescent} or the indicated timeout elapses.
4126 *
4127 * @param timeout the maximum time to wait
4128 * @param unit the time unit of the timeout argument
4129 * @return {@code true} if quiescent; {@code false} if the
4130 * timeout elapsed.
4131 */
4132 public boolean awaitQuiescence(long timeout, TimeUnit unit) {
4133 return (helpQuiescePool(this, unit.toNanos(timeout), false) > 0);
4134 }
4135
4136 /**
4137 * Unless this is the {@link #commonPool()}, initiates an orderly
4138 * shutdown in which previously submitted tasks are executed, but
4139 * no new tasks will be accepted, and waits until all tasks have
4140 * completed execution and the executor has terminated.
4141 *
4142 * <p> If already terminated, or this is the {@link
4143 * #commonPool()}, this method has no effect on execution, and
4144 * does not wait. Otherwise, if interrupted while waiting, this
4145 * method stops all executing tasks as if by invoking {@link
4146 * #shutdownNow()}. It then continues to wait until all actively
4147 * executing tasks have completed. Tasks that were awaiting
4148 * execution are not executed. The interrupt status will be
4149 * re-asserted before this method returns.
4150 *
4151 * @since 19
4152 */
4153 @Override
4154 public void close() {
4155 if (workerNamePrefix != null) {
4156 CountDownLatch done = null;
4157 boolean interrupted = false;
4158 while ((tryTerminate(interrupted, true) & TERMINATED) == 0) {
4159 if (done == null)
4160 done = terminationSignal();
4161 else {
4162 try {
4163 done.await();
4164 break;
4165 } catch (InterruptedException ex) {
4166 interrupted = true;
4167 }
4168 }
4169 }
4170 if (interrupted)
4171 Thread.currentThread().interrupt();
4172 }
4173 }
4174
4175 /**
4176 * Interface for extending managed parallelism for tasks running
4177 * in {@link ForkJoinPool}s.
4178 *
4179 * <p>A {@code ManagedBlocker} provides two methods. Method
4180 * {@link #isReleasable} must return {@code true} if blocking is
4181 * not necessary. Method {@link #block} blocks the current thread
4182 * if necessary (perhaps internally invoking {@code isReleasable}
4183 * before actually blocking). These actions are performed by any
4184 * thread invoking {@link
4185 * ForkJoinPool#managedBlock(ManagedBlocker)}. The unusual
4186 * methods in this API accommodate synchronizers that may, but
4187 * don't usually, block for long periods. Similarly, they allow
4188 * more efficient internal handling of cases in which additional
4189 * workers may be, but usually are not, needed to ensure
4190 * sufficient parallelism. Toward this end, implementations of
4191 * method {@code isReleasable} must be amenable to repeated
4192 * invocation. Neither method is invoked after a prior invocation
4193 * of {@code isReleasable} or {@code block} returns {@code true}.
4194 *
4195 * <p>For example, here is a ManagedBlocker based on a
4196 * ReentrantLock:
4197 * <pre> {@code
4198 * class ManagedLocker implements ManagedBlocker {
4199 * final ReentrantLock lock;
4200 * boolean hasLock = false;
4201 * ManagedLocker(ReentrantLock lock) { this.lock = lock; }
4202 * public boolean block() {
4203 * if (!hasLock)
4204 * lock.lock();
4205 * return true;
4206 * }
4207 * public boolean isReleasable() {
4208 * return hasLock || (hasLock = lock.tryLock());
4209 * }
4210 * }}</pre>
4211 *
4212 * <p>Here is a class that possibly blocks waiting for an
4213 * item on a given queue:
4214 * <pre> {@code
4215 * class QueueTaker<E> implements ManagedBlocker {
4216 * final BlockingQueue<E> queue;
4217 * volatile E item = null;
4218 * QueueTaker(BlockingQueue<E> q) { this.queue = q; }
4219 * public boolean block() throws InterruptedException {
4220 * if (item == null)
4221 * item = queue.take();
4222 * return true;
4223 * }
4224 * public boolean isReleasable() {
4225 * return item != null || (item = queue.poll()) != null;
4226 * }
4227 * public E getItem() { // call after pool.managedBlock completes
4228 * return item;
4229 * }
4230 * }}</pre>
4231 */
4232 public static interface ManagedBlocker {
4233 /**
4234 * Possibly blocks the current thread, for example waiting for
4235 * a lock or condition.
4236 *
4237 * @return {@code true} if no additional blocking is necessary
4238 * (i.e., if isReleasable would return true)
4239 * @throws InterruptedException if interrupted while waiting
4240 * (the method is not required to do so, but is allowed to)
4241 */
4242 boolean block() throws InterruptedException;
4243
4244 /**
4245 * Returns {@code true} if blocking is unnecessary.
4246 * @return {@code true} if blocking is unnecessary
4247 */
4248 boolean isReleasable();
4249 }
4250
4251 /**
4252 * Runs the given possibly blocking task. When {@linkplain
4253 * ForkJoinTask#inForkJoinPool() running in a ForkJoinPool}, this
4254 * method possibly arranges for a spare thread to be activated if
4255 * necessary to ensure sufficient parallelism while the current
4256 * thread is blocked in {@link ManagedBlocker#block blocker.block()}.
4257 *
4258 * <p>This method repeatedly calls {@code blocker.isReleasable()} and
4259 * {@code blocker.block()} until either method returns {@code true}.
4260 * Every call to {@code blocker.block()} is preceded by a call to
4261 * {@code blocker.isReleasable()} that returned {@code false}.
4262 *
4263 * <p>If not running in a ForkJoinPool, this method is
4264 * behaviorally equivalent to
4265 * <pre> {@code
4266 * while (!blocker.isReleasable())
4267 * if (blocker.block())
4268 * break;}</pre>
4269 *
4270 * If running in a ForkJoinPool, the pool may first be expanded to
4271 * ensure sufficient parallelism available during the call to
4272 * {@code blocker.block()}.
4273 *
4274 * @param blocker the blocker task
4275 * @throws InterruptedException if {@code blocker.block()} did so
4276 */
4277 public static void managedBlock(ManagedBlocker blocker)
4278 throws InterruptedException {
4279 Thread t; ForkJoinPool p;
4280 if ((t = Thread.currentThread()) instanceof ForkJoinWorkerThread &&
4281 (p = ((ForkJoinWorkerThread)t).pool) != null)
4282 p.compensatedBlock(blocker);
4283 else
4284 unmanagedBlock(blocker);
4285 }
4286
4287 /** ManagedBlock for ForkJoinWorkerThreads */
4288 private void compensatedBlock(ManagedBlocker blocker)
4289 throws InterruptedException {
4290 Objects.requireNonNull(blocker);
4291 for (;;) {
4292 int comp; boolean done;
4293 long c = ctl;
4294 if (blocker.isReleasable())
4295 break;
4296 if ((runState & STOP) != 0L)
4297 throw new InterruptedException();
4298 if ((comp = tryCompensate(c)) >= 0) {
4299 try {
4300 done = blocker.block();
4301 } finally {
4302 if (comp > 0)
4303 getAndAddCtl(RC_UNIT);
4304 }
4305 if (done)
4306 break;
4307 }
4308 }
4309 }
4310
4311 /**
4312 * Invokes tryCompensate to create or re-activate a spare thread to
4313 * compensate for a thread that performs a blocking operation. When the
4314 * blocking operation is done then endCompensatedBlock must be invoked
4315 * with the value returned by this method to re-adjust the parallelism.
4316 * @return value to use in endCompensatedBlock
4317 */
4318 final long beginCompensatedBlock() {
4319 int c;
4320 do {} while ((c = tryCompensate(ctl)) < 0);
4321 return (c == 0) ? 0L : RC_UNIT;
4322 }
4323
4324 /**
4325 * Re-adjusts parallelism after a blocking operation completes.
4326 * @param post value from beginCompensatedBlock
4327 */
4328 void endCompensatedBlock(long post) {
4329 if (post > 0L) {
4330 getAndAddCtl(post);
4331 }
4332 }
4333
4334 /** ManagedBlock for external threads */
4335 private static void unmanagedBlock(ManagedBlocker blocker)
4336 throws InterruptedException {
4337 Objects.requireNonNull(blocker);
4338 do {} while (!blocker.isReleasable() && !blocker.block());
4339 }
4340
4341 @Override
4342 protected <T> RunnableFuture<T> newTaskFor(Runnable runnable, T value) {
4343 Objects.requireNonNull(runnable);
4344 return (Thread.currentThread() instanceof ForkJoinWorkerThread) ?
4345 new ForkJoinTask.AdaptedRunnable<T>(runnable, value) :
4346 new ForkJoinTask.AdaptedInterruptibleRunnable<T>(runnable, value);
4347 }
4348
4349 @Override
4350 protected <T> RunnableFuture<T> newTaskFor(Callable<T> callable) {
4351 Objects.requireNonNull(callable);
4352 return (Thread.currentThread() instanceof ForkJoinWorkerThread) ?
4353 new ForkJoinTask.AdaptedCallable<T>(callable) :
4354 new ForkJoinTask.AdaptedInterruptibleCallable<T>(callable);
4355 }
4356
4357 static {
4358 U = Unsafe.getUnsafe();
4359 Class<ForkJoinPool> klass = ForkJoinPool.class;
4360 try {
4361 Field poolIdsField = klass.getDeclaredField("poolIds");
4362 POOLIDS_BASE = U.staticFieldBase(poolIdsField);
4363 POOLIDS = U.staticFieldOffset(poolIdsField);
4364 } catch (NoSuchFieldException e) {
4365 throw new ExceptionInInitializerError(e);
4366 }
4367 CTL = U.objectFieldOffset(klass, "ctl");
4368 RUNSTATE = U.objectFieldOffset(klass, "runState");
4369 PARALLELISM = U.objectFieldOffset(klass, "parallelism");
4370 THREADIDS = U.objectFieldOffset(klass, "threadIds");
4371 TERMINATION = U.objectFieldOffset(klass, "termination");
4372 Class<ForkJoinTask[]> aklass = ForkJoinTask[].class;
4373 ABASE = U.arrayBaseOffset(aklass);
4374 int scale = U.arrayIndexScale(aklass);
4375 ASHIFT = 31 - Integer.numberOfLeadingZeros(scale);
4376 if ((scale & (scale - 1)) != 0)
4377 throw new Error("array index scale not a power of two");
4378
4379 Class<?> dep = LockSupport.class; // ensure loaded
4380 // allow access to non-public methods
4381 JLA = SharedSecrets.getJavaLangAccess();
4382 SharedSecrets.setJavaUtilConcurrentFJPAccess(
4383 new JavaUtilConcurrentFJPAccess() {
4384 @Override
4385 public long beginCompensatedBlock(ForkJoinPool pool) {
4386 return pool.beginCompensatedBlock();
4387 }
4388 public void endCompensatedBlock(ForkJoinPool pool, long post) {
4389 pool.endCompensatedBlock(post);
4390 }
4391 });
4392 defaultForkJoinWorkerThreadFactory =
4393 new DefaultForkJoinWorkerThreadFactory();
4394 common = new ForkJoinPool((byte)0);
4395 }
4396 }