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 deactivate, 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 interrupted 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 interrupted 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, along
 944      * with status and bookkeeping fields written (mostly) by owners,
 945      * that otherwise interfere with reading array and base
 946      * fields. There are other variables commonly contributing to
 947      * false-sharing-related performance issues (including fields of
 948      * class Thread), but we can't do much about this except try to
 949      * 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         // fields otherwise causing more unnecessary false-sharing cache misses
1182         @jdk.internal.vm.annotation.Contended("w")
1183         int top;                   // index of next slot for push
1184         @jdk.internal.vm.annotation.Contended("w")
1185         volatile int phase;        // versioned active status
1186         @jdk.internal.vm.annotation.Contended("w")
1187         int stackPred;             // pool stack (ctl) predecessor link
1188         @jdk.internal.vm.annotation.Contended("w")
1189         volatile int parking;      // nonzero if parked in awaitWork
1190         @jdk.internal.vm.annotation.Contended("w")
1191         volatile int source;       // source queue id (or DROPPED)
1192         @jdk.internal.vm.annotation.Contended("w")
1193         int nsteals;               // number of steals from other queues
1194 
1195         // Support for atomic operations
1196         private static final Unsafe U;
1197         private static final long PHASE;
1198         private static final long BASE;
1199         private static final long TOP;
1200         private static final long ARRAY;
1201 
1202         final void updateBase(int v) {
1203             U.putIntVolatile(this, BASE, v);
1204         }
1205         final void updateTop(int v) {
1206             U.putIntOpaque(this, TOP, v);
1207         }
1208         final void updateArray(ForkJoinTask<?>[] a) {
1209             U.getAndSetReference(this, ARRAY, a);
1210         }
1211         final void unlockPhase() {
1212             U.getAndAddInt(this, PHASE, IDLE);
1213         }
1214         final boolean tryLockPhase() {    // seqlock acquire
1215             int p;
1216             return (((p = phase) & IDLE) != 0 &&
1217                     U.compareAndSetInt(this, PHASE, p, p + IDLE));
1218         }
1219 
1220         /**
1221          * Constructor. For internal queues, most fields are initialized
1222          * upon thread start in pool.registerWorker.
1223          */
1224         WorkQueue(ForkJoinWorkerThread owner, int id, int cfg,
1225                   boolean clearThreadLocals) {
1226             this.config = (clearThreadLocals) ? cfg | CLEAR_TLS : cfg;
1227             if ((this.owner = owner) == null) {
1228                 array = new ForkJoinTask<?>[INITIAL_QUEUE_CAPACITY];
1229                 phase = id | IDLE;
1230             }
1231         }
1232 
1233         /**
1234          * Returns an exportable index (used by ForkJoinWorkerThread).
1235          */
1236         final int getPoolIndex() {
1237             return (phase & 0xffff) >>> 1; // ignore odd/even tag bit
1238         }
1239 
1240         /**
1241          * Returns the approximate number of tasks in the queue.
1242          */
1243         final int queueSize() {
1244             int unused = phase;             // for ordering effect
1245             return Math.max(top - base, 0); // ignore transient negative
1246         }
1247 
1248         /**
1249          * Pushes a task. Called only by owner or if already locked
1250          *
1251          * @param task the task; no-op if null
1252          * @param pool the pool to signal if was previously empty, else null
1253          * @param internal if caller owns this queue
1254          * @throws RejectedExecutionException if array could not be resized
1255          */
1256         final void push(ForkJoinTask<?> task, ForkJoinPool pool, boolean internal) {
1257             int s = top, b = base, m, cap, room; ForkJoinTask<?>[] a, na;
1258             if ((a = array) != null && (cap = a.length) > 0) { // else disabled
1259                 int k = (m = cap - 1) & s;
1260                 if ((room = m - (s - b)) >= 0) {
1261                     top = s + 1;
1262                     long pos = slotOffset(k);
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 && (na = growArray(a, cap, s)) != null)
1268                         k = ((a = na).length - 1) & s;      // resize
1269                 }
1270                 if (!internal)
1271                     unlockPhase();
1272                 if (room < 0)
1273                     throw new RejectedExecutionException("Queue capacity exceeded");
1274                 if (pool != null &&
1275                     (room == 0 ||
1276                      U.getReferenceAcquire(a, slotOffset(m & (s - 1))) == null))
1277                     pool.signalWork(a, k);    // may have appeared empty
1278             }
1279         }
1280 
1281         /**
1282          * Resizes the queue array unless out of memory.
1283          * @param a old array
1284          * @param cap old array capacity
1285          * @param s current top
1286          * @return new array, or null on failure
1287          */
1288         private ForkJoinTask<?>[] growArray(ForkJoinTask<?>[] a, int cap, int s) {
1289             int newCap = (cap >= 1 << 16) ? cap << 1 : cap << 2;
1290             ForkJoinTask<?>[] newArray = null;
1291             if (a != null && a.length == cap && cap > 0 && newCap > 0) {
1292                 try {
1293                     newArray = new ForkJoinTask<?>[newCap];
1294                 } catch (OutOfMemoryError ex) {
1295                 }
1296                 if (newArray != null) {               // else throw on next push
1297                     int mask = cap - 1, newMask = newCap - 1;
1298                     for (int k = s, j = cap; j > 0; --j, --k) {
1299                         ForkJoinTask<?> u;            // poll old, push to new
1300                         if ((u = (ForkJoinTask<?>)U.getAndSetReference(
1301                                  a, slotOffset(k & mask), null)) == null)
1302                             break;                    // lost to pollers
1303                         newArray[k & newMask] = u;
1304                     }
1305                     updateArray(newArray);           // fully fenced
1306                 }
1307             }
1308             return newArray;
1309         }
1310 
1311         /**
1312          * Takes next task, if one exists, in lifo order.
1313          */
1314         private ForkJoinTask<?> localPop() {
1315             ForkJoinTask<?> t = null;
1316             int s = top - 1, cap; long k; ForkJoinTask<?>[] a;
1317             if ((a = array) != null && (cap = a.length) > 0 &&
1318                 U.getReference(a, k = slotOffset((cap - 1) & s)) != null &&
1319                 (t = (ForkJoinTask<?>)U.getAndSetReference(a, k, null)) != null)
1320                 updateTop(s);
1321             return t;
1322         }
1323 
1324         /**
1325          * Takes next task, if one exists, in fifo order.
1326          */
1327         private ForkJoinTask<?> localPoll() {
1328             ForkJoinTask<?> t = null;
1329             int p = top, cap; ForkJoinTask<?>[] a;
1330             if ((a = array) != null && (cap = a.length) > 0) {
1331                 for (int b = base; p - b > 0; ) {
1332                     int nb = b + 1;
1333                     long k = slotOffset((cap - 1) & b);
1334                     if (U.getReference(a, k) == null) {
1335                         if (nb == p)
1336                             break;          // else base is lagging
1337                         while (b == (b = U.getIntAcquire(this, BASE)))
1338                             Thread.onSpinWait(); // spin to reduce memory traffic
1339                     }
1340                     else if ((t = (ForkJoinTask<?>)
1341                               U.getAndSetReference(a, k, null)) != null) {
1342                         updateBase(nb);
1343                         break;
1344                     }
1345                     else
1346                         b = base;
1347                 }
1348             }
1349             return t;
1350         }
1351 
1352         /**
1353          * Takes next task, if one exists, using configured mode.
1354          */
1355         final ForkJoinTask<?> nextLocalTask() {
1356             return (config & FIFO) == 0 ? localPop() : localPoll();
1357         }
1358 
1359         /**
1360          * Pops the given task only if it is at the current top.
1361          * @param task the task. Caller must ensure non-null.
1362          * @param internal if caller owns this queue
1363          */
1364         final boolean tryUnpush(ForkJoinTask<?> task, boolean internal) {
1365             boolean taken = false;
1366             ForkJoinTask<?>[] a = array;
1367             int p = top, s = p - 1, cap; long k;
1368             if (a != null && (cap = a.length) > 0 &&
1369                 U.getReference(a, k = slotOffset((cap - 1) & s)) == task &&
1370                 (internal || tryLockPhase())) {
1371                 if (top == p && U.compareAndSetReference(a, k, task, null)) {
1372                     taken = true;
1373                     updateTop(s);
1374                 }
1375                 if (!internal)
1376                     unlockPhase();
1377             }
1378             return taken;
1379         }
1380 
1381         /**
1382          * Returns next task, if one exists, in order specified by mode.
1383          */
1384         final ForkJoinTask<?> peek() {
1385             ForkJoinTask<?>[] a = array;
1386             int b = base, cfg = config, p = top, cap;
1387             if (p != b && a != null && (cap = a.length) > 0) {
1388                 if ((cfg & FIFO) == 0)
1389                     return a[(cap - 1) & (p - 1)];
1390                 else { // skip over in-progress removals
1391                     ForkJoinTask<?> t;
1392                     for ( ; p - b > 0; ++b) {
1393                         if ((t = a[(cap - 1) & b]) != null)
1394                             return t;
1395                     }
1396                 }
1397             }
1398             return null;
1399         }
1400 
1401         /**
1402          * Polls for a task. Used only by non-owners.
1403          */
1404         final ForkJoinTask<?> poll() {
1405             for (int pb = -1, b; ; pb = b) {       // track progress
1406                 ForkJoinTask<?> t; int cap, nb; long k; ForkJoinTask<?>[] a;
1407                 if ((a = array) == null || (cap = a.length) <= 0)
1408                     break;
1409                 t = (ForkJoinTask<?>)U.getReferenceAcquire(
1410                     a, k = slotOffset((cap - 1) & (b = base)));
1411                 Object u = U.getReference(         // next slot
1412                     a, slotOffset((cap - 1) & (nb = b + 1)));
1413                 if (base != b)                     // inconsistent
1414                     ;
1415                 else if (t == null) {
1416                     if (u == null && top - b <= 0)
1417                         break;                     // empty
1418                     if (pb == b)
1419                         Thread.onSpinWait();       // stalled
1420                 }
1421                 else if (U.compareAndSetReference(a, k, t, null)) {
1422                     updateBase(nb);
1423                     return t;
1424                 }
1425             }
1426             return null;
1427         }
1428 
1429         // specialized execution methods
1430 
1431         /**
1432          * Runs the given task, as well as remaining local tasks.
1433          */
1434         final void topLevelExec(ForkJoinTask<?> task, int cfg) {
1435             int fifo = cfg & FIFO, clr = cfg & CLEAR_TLS;
1436             while (task != null) {
1437                 task.doExec();
1438                 task = null;
1439                 int p = top, cap; ForkJoinTask<?>[] a;
1440                 if ((a = array) == null || (cap = a.length) <= 0)
1441                     break;
1442                 if (fifo == 0) {  // specialized localPop
1443                     int s = p - 1; long k;
1444                     if (U.getReference(
1445                             a, k = slotOffset((cap - 1) & s)) != null &&
1446                         (task = (ForkJoinTask<?>)
1447                          U.getAndSetReference(a, k, null)) != null) {
1448                         top = s;
1449                     }
1450                 } else {         // specialized localPoll
1451                     for (int b = base; p - b > 0; ) {
1452                         int nb = b + 1;
1453                         if ((task = (ForkJoinTask<?>)U.getAndSetReference(
1454                                  a, slotOffset((cap - 1) & b), null)) != null) {
1455                             base = nb;
1456                             break;
1457                         }
1458                         if (nb == p)
1459                             break;
1460                         while (b == (b = U.getIntAcquire(this, BASE)))
1461                             Thread.onSpinWait();
1462                     }
1463                 }
1464             }
1465         }
1466 
1467         /**
1468          * Deep form of tryUnpush: Traverses from top and removes and
1469          * runs task if present.
1470          */
1471         final void tryRemoveAndExec(ForkJoinTask<?> task, boolean internal) {
1472             ForkJoinTask<?>[] a = array;
1473             int b = base, p = top, s = p - 1, d = p - b, cap;
1474             if (a != null && (cap = a.length) > 0) {
1475                 for (int m = cap - 1, i = s; d > 0; --i, --d) {
1476                     long k; boolean taken;
1477                     ForkJoinTask<?> t = (ForkJoinTask<?>)U.getReference(
1478                         a, k = slotOffset(i & m));
1479                     if (t == null)
1480                         break;
1481                     if (t == task) {
1482                         if (!internal && !tryLockPhase())
1483                             break;                  // fail if locked
1484                         if (taken =
1485                             (top == p &&
1486                              U.compareAndSetReference(a, k, task, null))) {
1487                             if (i == s)             // act as pop
1488                                 updateTop(s);
1489                             else if (i == base)     // act as poll
1490                                 updateBase(i + 1);
1491                             else {                  // swap with top
1492                                 U.putReferenceVolatile(
1493                                     a, k, (ForkJoinTask<?>)
1494                                     U.getAndSetReference(
1495                                         a, slotOffset(s & m), null));
1496                                 updateTop(s);
1497                             }
1498                         }
1499                         if (!internal)
1500                             unlockPhase();
1501                         if (taken)
1502                             task.doExec();
1503                         break;
1504                     }
1505                 }
1506             }
1507         }
1508 
1509         /**
1510          * Tries to pop and run tasks within the target's computation
1511          * until done, not found, or limit exceeded.
1512          *
1513          * @param task root of computation
1514          * @param limit max runs, or zero for no limit
1515          * @return task status if known to be done
1516          */
1517         final int helpComplete(ForkJoinTask<?> task, boolean internal, int limit) {
1518             int status = 0;
1519             if (task != null) {
1520                 outer: for (;;) {
1521                     ForkJoinTask<?>[] a; boolean taken; Object o;
1522                     int stat, p, s, cap;
1523                     if ((stat = task.status) < 0) {
1524                         status = stat;
1525                         break;
1526                     }
1527                     if ((a = array) == null || (cap = a.length) <= 0)
1528                         break;
1529                     long k = slotOffset((cap - 1) & (s = (p = top) - 1));
1530                     if (!((o = U.getReference(a, k)) instanceof CountedCompleter))
1531                         break;
1532                     CountedCompleter<?> t = (CountedCompleter<?>)o, f = t;
1533                     for (int steps = cap;;) {       // bound path
1534                         if (f == task)
1535                             break;
1536                         if ((f = f.completer) == null || --steps == 0)
1537                             break outer;
1538                     }
1539                     if (!internal && !tryLockPhase())
1540                         break;
1541                     if (taken =
1542                         (top == p &&
1543                          U.compareAndSetReference(a, k, t, null)))
1544                         updateTop(s);
1545                     if (!internal)
1546                         unlockPhase();
1547                     if (!taken)
1548                         break;
1549                     t.doExec();
1550                     if (limit != 0 && --limit == 0)
1551                         break;
1552                 }
1553             }
1554             return status;
1555         }
1556 
1557         /**
1558          * Tries to poll and run AsynchronousCompletionTasks until
1559          * none found or blocker is released
1560          *
1561          * @param blocker the blocker
1562          */
1563         final void helpAsyncBlocker(ManagedBlocker blocker) {
1564             for (;;) {
1565                 ForkJoinTask<?> t; ForkJoinTask<?>[] a; int b, cap; long k;
1566                 if ((a = array) == null || (cap = a.length) <= 0)
1567                     break;
1568                 t = (ForkJoinTask<?>)U.getReferenceAcquire(
1569                     a, k = slotOffset((cap - 1) & (b = base)));
1570                 if (t == null) {
1571                     if (top - b <= 0)
1572                         break;
1573                 }
1574                 else if (!(t instanceof CompletableFuture
1575                            .AsynchronousCompletionTask))
1576                     break;
1577                 if (blocker != null && blocker.isReleasable())
1578                     break;
1579                 if (base == b && t != null &&
1580                     U.compareAndSetReference(a, k, t, null)) {
1581                     updateBase(b + 1);
1582                     t.doExec();
1583                 }
1584             }
1585         }
1586 
1587         // misc
1588 
1589         /**
1590          * Cancels all local tasks. Called only by owner.
1591          */
1592         final void cancelTasks() {
1593             for (ForkJoinTask<?> t; (t = localPop()) != null; ) {
1594                 try {
1595                     t.cancel(false);
1596                 } catch (Throwable ignore) {
1597                 }
1598             }
1599         }
1600 
1601         /**
1602          * Returns true if internal and not known to be blocked.
1603          */
1604         final boolean isApparentlyUnblocked() {
1605             Thread wt; Thread.State s;
1606             return ((wt = owner) != null && (phase & IDLE) != 0 &&
1607                     (s = wt.getState()) != Thread.State.BLOCKED &&
1608                     s != Thread.State.WAITING &&
1609                     s != Thread.State.TIMED_WAITING);
1610         }
1611 
1612         static {
1613             U = Unsafe.getUnsafe();
1614             Class<WorkQueue> klass = WorkQueue.class;
1615             PHASE = U.objectFieldOffset(klass, "phase");
1616             BASE = U.objectFieldOffset(klass, "base");
1617             TOP = U.objectFieldOffset(klass, "top");
1618             ARRAY = U.objectFieldOffset(klass, "array");
1619         }
1620     }
1621 
1622     // static fields (initialized in static initializer below)
1623 
1624     /**
1625      * Creates a new ForkJoinWorkerThread. This factory is used unless
1626      * overridden in ForkJoinPool constructors.
1627      */
1628     public static final ForkJoinWorkerThreadFactory
1629         defaultForkJoinWorkerThreadFactory;
1630 
1631     /**
1632      * Common (static) pool. Non-null for public use unless a static
1633      * construction exception, but internal usages null-check on use
1634      * to paranoically avoid potential initialization circularities
1635      * as well as to simplify generated code.
1636      */
1637     static final ForkJoinPool common;
1638 
1639     /**
1640      * Sequence number for creating worker names
1641      */
1642     private static volatile int poolIds;
1643 
1644     /**
1645      * For VirtualThread intrinsics
1646      */
1647     private static final JavaLangAccess JLA;
1648 
1649     // fields declared in order of their likely layout on most VMs
1650     volatile CountDownLatch termination; // lazily constructed
1651     final Predicate<? super ForkJoinPool> saturate;
1652     final ForkJoinWorkerThreadFactory factory;
1653     final UncaughtExceptionHandler ueh;  // per-worker UEH
1654     final SharedThreadContainer container;
1655     final String workerNamePrefix;       // null for common pool
1656     final String poolName;
1657     volatile DelayScheduler delayScheduler;  // lazily constructed
1658     WorkQueue[] queues;                  // main registry
1659     volatile long runState;              // versioned, lockable
1660     final long keepAlive;                // milliseconds before dropping if idle
1661     final long config;                   // static configuration bits
1662     volatile long stealCount;            // collects worker nsteals
1663     volatile long threadIds;             // for worker thread names
1664 
1665     @jdk.internal.vm.annotation.Contended("fjpctl") // segregate
1666     volatile long ctl;                   // main pool control
1667     @jdk.internal.vm.annotation.Contended("fjpctl") // colocate
1668     int parallelism;                     // target number of workers
1669 
1670     // Support for atomic operations
1671     private static final Unsafe U;
1672     private static final long CTL;
1673     private static final long RUNSTATE;
1674     private static final long PARALLELISM;
1675     private static final long THREADIDS;
1676     private static final long TERMINATION;
1677     private static final Object POOLIDS_BASE;
1678     private static final long POOLIDS;
1679 
1680     private boolean compareAndSetCtl(long c, long v) {
1681         return U.compareAndSetLong(this, CTL, c, v);
1682     }
1683     private long compareAndExchangeCtl(long c, long v) {
1684         return U.compareAndExchangeLong(this, CTL, c, v);
1685     }
1686     private long getAndAddCtl(long v) {
1687         return U.getAndAddLong(this, CTL, v);
1688     }
1689     private long incrementThreadIds() {
1690         return U.getAndAddLong(this, THREADIDS, 1L);
1691     }
1692     private static int getAndAddPoolIds(int x) {
1693         return U.getAndAddInt(POOLIDS_BASE, POOLIDS, x);
1694     }
1695     private int getAndSetParallelism(int v) {
1696         return U.getAndSetInt(this, PARALLELISM, v);
1697     }
1698     private int getParallelismOpaque() {
1699         return U.getIntOpaque(this, PARALLELISM);
1700     }
1701     private CountDownLatch cmpExTerminationSignal(CountDownLatch x) {
1702         return (CountDownLatch)
1703             U.compareAndExchangeReference(this, TERMINATION, null, x);
1704     }
1705 
1706     // runState operations
1707 
1708     private long getAndBitwiseOrRunState(long v) { // for status bits
1709         return U.getAndBitwiseOrLong(this, RUNSTATE, v);
1710     }
1711     private boolean casRunState(long c, long v) {
1712         return U.compareAndSetLong(this, RUNSTATE, c, v);
1713     }
1714     private void unlockRunState() {              // increment lock bit
1715         U.getAndAddLong(this, RUNSTATE, RS_LOCK);
1716     }
1717     private long lockRunState() {                // lock and return current state
1718         long s, u;                               // locked when RS_LOCK set
1719         if (((s = runState) & RS_LOCK) == 0L && casRunState(s, u = s + RS_LOCK))
1720             return u;
1721         else
1722             return spinLockRunState();
1723     }
1724     private long spinLockRunState() {            // spin/sleep
1725         for (int waits = 0;;) {
1726             long s, u;
1727             if (((s = runState) & RS_LOCK) == 0L) {
1728                 if (casRunState(s, u = s + RS_LOCK))
1729                     return u;
1730                 waits = 0;
1731             } else if (waits < SPIN_WAITS) {
1732                 ++waits;
1733                 Thread.onSpinWait();
1734             } else {
1735                 if (waits < MIN_SLEEP)
1736                     waits = MIN_SLEEP;
1737                 LockSupport.parkNanos(this, (long)waits);
1738                 if (waits < MAX_SLEEP)
1739                     waits <<= 1;
1740             }
1741         }
1742     }
1743 
1744     static boolean poolIsStopping(ForkJoinPool p) { // Used by ForkJoinTask
1745         return p != null && (p.runState & STOP) != 0L;
1746     }
1747 
1748     // Creating, registering, and deregistering workers
1749 
1750     /**
1751      * Tries to construct and start one worker. Assumes that total
1752      * count has already been incremented as a reservation.  Invokes
1753      * deregisterWorker on any failure.
1754      *
1755      * @return true if successful
1756      */
1757     private boolean createWorker() {
1758         ForkJoinWorkerThreadFactory fac = factory;
1759         SharedThreadContainer ctr = container;
1760         Throwable ex = null;
1761         ForkJoinWorkerThread wt = null;
1762         try {
1763             if ((runState & STOP) == 0L &&  // avoid construction if terminating
1764                 fac != null && (wt = fac.newThread(this)) != null) {
1765                 if (ctr != null)
1766                     ctr.start(wt);
1767                 else
1768                     wt.start();
1769                 return true;
1770             }
1771         } catch (Throwable rex) {
1772             ex = rex;
1773         }
1774         deregisterWorker(wt, ex);
1775         return false;
1776     }
1777 
1778     /**
1779      * Provides a name for ForkJoinWorkerThread constructor.
1780      */
1781     final String nextWorkerThreadName() {
1782         String prefix = workerNamePrefix;
1783         long tid = incrementThreadIds() + 1L;
1784         if (prefix == null) // commonPool has no prefix
1785             prefix = "ForkJoinPool.commonPool-worker-";
1786         return prefix.concat(Long.toString(tid));
1787     }
1788 
1789     /**
1790      * Finishes initializing and records internal queue.
1791      *
1792      * @param w caller's WorkQueue
1793      */
1794     final void registerWorker(WorkQueue w) {
1795         if (w != null) {
1796             w.array = new ForkJoinTask<?>[INITIAL_QUEUE_CAPACITY];
1797             ThreadLocalRandom.localInit();
1798             int seed = w.stackPred = ThreadLocalRandom.getProbe();
1799             int phaseSeq = seed & ~((IDLE << 1) - 1); // initial phase tag
1800             int id = ((seed << 1) | 1) & SMASK; // base of linear-probe-like scan
1801             long stop = lockRunState() & STOP;
1802             try {
1803                 WorkQueue[] qs; int n;
1804                 if (stop == 0L && (qs = queues) != null && (n = qs.length) > 0) {
1805                     for (int k = n, m = n - 1;  ; id += 2) {
1806                         if (qs[id &= m] == null)
1807                             break;
1808                         if ((k -= 2) <= 0) {
1809                             id |= n;
1810                             break;
1811                         }
1812                     }
1813                     w.phase = id | phaseSeq;    // now publishable
1814                     if (id < n)
1815                         qs[id] = w;
1816                     else {                      // expand
1817                         int an = n << 1, am = an - 1;
1818                         WorkQueue[] as = new WorkQueue[an];
1819                         as[id & am] = w;
1820                         for (int j = 1; j < n; j += 2)
1821                             as[j] = qs[j];
1822                         for (int j = 0; j < n; j += 2) {
1823                             WorkQueue q;        // shared queues may move
1824                             if ((q = qs[j]) != null)
1825                                 as[q.phase & EXTERNAL_ID_MASK & am] = q;
1826                         }
1827                         U.storeFence();         // fill before publish
1828                         queues = as;
1829                     }
1830                 }
1831             } finally {
1832                 unlockRunState();
1833             }
1834         }
1835     }
1836 
1837     /**
1838      * Final callback from terminating worker, as well as upon failure
1839      * to construct or start a worker.  Removes record of worker from
1840      * array, and adjusts counts. If pool is shutting down, tries to
1841      * complete termination.
1842      *
1843      * @param wt the worker thread, or null if construction failed
1844      * @param ex the exception causing failure, or null if none
1845      */
1846     final void deregisterWorker(ForkJoinWorkerThread wt, Throwable ex) {
1847         WorkQueue w = null;                // null if not created
1848         int phase = 0;                     // 0 if not registered
1849         if (wt != null && (w = wt.workQueue) != null &&
1850             (phase = w.phase) != 0 && (phase & IDLE) != 0)
1851             releaseWaiters();              // ensure released
1852         if (w == null || w.source != DROPPED) {
1853             long c = ctl;                  // decrement counts
1854             do {} while (c != (c = compareAndExchangeCtl(
1855                                    c, ((RC_MASK & (c - RC_UNIT)) |
1856                                        (TC_MASK & (c - TC_UNIT)) |
1857                                        (LMASK & c)))));
1858         }
1859         if (phase != 0 && w != null) {     // remove index unless terminating
1860             long ns = w.nsteals & 0xffffffffL;
1861             if ((runState & STOP) == 0L) {
1862                 WorkQueue[] qs; int n, i;
1863                 if ((lockRunState() & STOP) == 0L &&
1864                     (qs = queues) != null && (n = qs.length) > 0 &&
1865                     qs[i = phase & SMASK & (n - 1)] == w) {
1866                     qs[i] = null;
1867                     stealCount += ns;      // accumulate steals
1868                 }
1869                 unlockRunState();
1870             }
1871         }
1872         if ((tryTerminate(false, false) & STOP) == 0L &&
1873             phase != 0 && w != null && w.source != DROPPED) {
1874             w.cancelTasks();               // clean queue
1875             signalWork(null, 0);           // possibly replace
1876         }
1877         if (ex != null)
1878             ForkJoinTask.rethrow(ex);
1879     }
1880 
1881     /**
1882      * Releases an idle worker, or creates one if not enough exist,
1883      * giving up if array a is nonnull and task at a[k] already taken.
1884      */
1885     final void signalWork(ForkJoinTask<?>[] a, int k) {
1886         int pc = parallelism;
1887         for (long c = ctl;;) {
1888             WorkQueue[] qs = queues;
1889             long ac = (c + RC_UNIT) & RC_MASK, nc;
1890             int sp = (int)c, i = sp & SMASK;
1891             if ((short)(c >>> RC_SHIFT) >= pc)
1892                 break;
1893             if (qs == null)
1894                 break;
1895             if (qs.length <= i)
1896                 break;
1897             WorkQueue w = qs[i], v = null;
1898             if (sp == 0) {
1899                 if ((short)(c >>> TC_SHIFT) >= pc)
1900                     break;
1901                 nc = ((c + TC_UNIT) & TC_MASK) | ac;
1902             }
1903             else if ((v = w) == null)
1904                 break;
1905             else
1906                 nc = (v.stackPred & LMASK) | (c & TC_MASK) | ac;
1907             if (a != null && k < a.length && k >= 0 && a[k] == null)
1908                 break;
1909             if (c == (c = ctl) && c == (c = compareAndExchangeCtl(c, nc))) {
1910                 if (v == null)
1911                     createWorker();
1912                 else {
1913                     v.phase = sp;
1914                     if (v.parking != 0)
1915                         U.unpark(v.owner);
1916                 }
1917                 break;
1918             }
1919         }
1920     }
1921 
1922     /**
1923      * Releases all waiting workers. Called only during shutdown.
1924      */
1925     private void releaseWaiters() {
1926         for (long c = ctl;;) {
1927             WorkQueue[] qs; WorkQueue v; int sp, i;
1928             if ((sp = (int)c) == 0 || (qs = queues) == null ||
1929                 qs.length <= (i = sp & SMASK) || (v = qs[i]) == null)
1930                 break;
1931             if (c == (c = compareAndExchangeCtl(
1932                           c, ((UMASK & (c + RC_UNIT)) | (c & TC_MASK) |
1933                               (v.stackPred & LMASK))))) {
1934                 v.phase = sp;
1935                 if (v.parking != 0)
1936                     U.unpark(v.owner);
1937             }
1938         }
1939     }
1940 
1941     /**
1942      * Internal version of isQuiescent and related functionality.
1943      * @return positive if stopping, nonnegative if terminating or all
1944      * workers are inactive and submission queues are empty and
1945      * unlocked; if so, setting STOP if shutdown is enabled
1946      */
1947     private int quiescent() {
1948         for (;;) {
1949             long phaseSum = 0L;
1950             boolean swept = false;
1951             for (long e, prevRunState = 0L; ; prevRunState = e) {
1952                 DelayScheduler ds;
1953                 long c = ctl;
1954                 if (((e = runState) & STOP) != 0L)
1955                     return 1;                             // terminating
1956                 else if ((c & RC_MASK) > 0L)
1957                     return -1;                            // at least one active
1958                 else if (!swept || e != prevRunState || (e & RS_LOCK) != 0) {
1959                     long sum = c;
1960                     WorkQueue[] qs = queues;
1961                     int n = (qs == null) ? 0 : qs.length;
1962                     for (int i = 0; i < n; ++i) {         // scan queues
1963                         WorkQueue q;
1964                         if ((q = qs[i]) != null) {
1965                             int p = q.phase, s = q.top, b = q.base;
1966                             sum += (p & 0xffffffffL) | ((long)b << 32);
1967                             if ((p & IDLE) == 0 || s - b > 0)
1968                                 return -1;
1969                         }
1970                     }
1971                     swept = (phaseSum == (phaseSum = sum));
1972                 }
1973                 else if ((e & SHUTDOWN) == 0)
1974                     return 0;
1975                 else if ((ds = delayScheduler) != null && !ds.canShutDown())
1976                     return 0;
1977                 else if (compareAndSetCtl(c, c) && casRunState(e, e | STOP))
1978                     return 1;                             // enable termination
1979                 else
1980                     break;                                // restart
1981             }
1982         }
1983     }
1984 
1985     /**
1986      * Top-level runloop for workers, called by ForkJoinWorkerThread.run.
1987      * See above for explanation.
1988      *
1989      * @param w caller's WorkQueue (may be null on failed initialization)
1990      */
1991     final void runWorker(WorkQueue w) {
1992         if (w != null && w.phase != 0) {                  // else unregistered
1993             WorkQueue[] qs;
1994             int cfg = w.config, r = w.stackPred;          // seed from registerWorker
1995             int rescans = 0, inactive = 0, taken = 0, n;
1996             while ((runState & STOP) == 0L && (qs = queues) != null &&
1997                    (n = qs.length) > 0) {
1998                 int i = r, step = (r >>> 16) | 1;
1999                 r ^= r << 13; r ^= r >>> 17; r ^= r << 5; // xorshift
2000                 scan: for (int j = n; j != 0; --j, i += step) {
2001                     WorkQueue q; int qid;
2002                     if ((q = qs[qid = i & (n - 1)]) != null) {
2003                         ForkJoinTask<?>[] a; int cap;     // poll queue
2004                         while ((a = q.array) != null && (cap = a.length) > 0) {
2005                             int b, nb, nk; long bp; ForkJoinTask<?> t;
2006                             t = (ForkJoinTask<?>)U.getReferenceAcquire(
2007                                 a, bp = slotOffset((cap - 1) & (b = q.base)));
2008                             if (q.base != b)
2009                                 continue;                 // inconsistent
2010                             long np = slotOffset(nk = (nb = b + 1) & (cap - 1));
2011                             if (t == null) {
2012                                 if (q.array != a)         // resized
2013                                     continue;
2014                                 if (rescans > 0)          // ran or stalled
2015                                     break scan;
2016                                 if (U.getReference(a, np) != null ||
2017                                     (rescans < 0 &&
2018                                      (U.getReferenceAcquire(a, bp) != null ||
2019                                       q.top != q.base))) {
2020                                     rescans = 1;          // may be stalled
2021                                     continue;
2022                                 }
2023                                 break;                    // probably empty
2024                             }
2025                             else if (inactive != 0) {
2026                                 if ((inactive = tryReactivate(w)) != 0) {
2027                                     rescans = 1;          // can't take yet
2028                                     break scan;
2029                                 }
2030                             }
2031                             else if (U.compareAndSetReference(a, bp, t, null)) {
2032                                 q.base = nb;
2033                                 Object nt = U.getReferenceAcquire(a, np);
2034                                 w.source = qid;
2035                                 rescans = 1;
2036                                 ++taken;
2037                                 if (nt != null &&         // confirm a[nk]
2038                                     U.getReferenceAcquire(a, np) == nt)
2039                                     signalWork(a, nk);    // propagate
2040                                 w.topLevelExec(t, cfg);
2041                             }
2042                         }
2043                     }
2044                 }
2045                 if (rescans >= 0)
2046                     --rescans;
2047                 else if (inactive == 0) {
2048                     if ((inactive = deactivate(w, taken)) != 0)
2049                         taken = 0;
2050                 }
2051                 else if (awaitWork(w) == 0)
2052                     inactive = rescans = 0;
2053                 else
2054                     break;
2055             }
2056         }
2057     }
2058 
2059     /**
2060      * Tries to deactivate worker, keeping active on contention
2061      *
2062      * @param w the work queue
2063      * @param taken nonzero if stole tasks since last deactivation
2064      * @return nonzero if inactive
2065      */
2066     private int deactivate(WorkQueue w, int taken) {
2067         int inactive = 0, phase;
2068         if (w != null && (inactive = (phase = w.phase) & IDLE) == 0) {
2069             long sp = (phase + (IDLE << 1)) & LMASK, pc, c;
2070             w.phase = phase | IDLE;
2071             w.stackPred = (int)(pc = ctl);    // set ctl stack link
2072             if (!compareAndSetCtl(            // try to enqueue
2073                     pc, c = ((pc - RC_UNIT) & UMASK) | sp))
2074                 w.phase = phase;              // back out on contention
2075             else {
2076                 if (taken != 0)
2077                     w.nsteals += taken;
2078                 if (((c & RC_MASK) == 0L && quiescent() > 0) || taken == 0)
2079                     inactive = w.phase & IDLE;    // check quiescent termination
2080                 else {                            // spin for approx 1 scan cost
2081                     int tc = (short)(c >>> TC_SHIFT);
2082                     int spins = Math.max((tc << 1) + tc, SPIN_WAITS);
2083                     while ((inactive = w.phase & IDLE) != 0 && --spins != 0)
2084                         Thread.onSpinWait();
2085                 }
2086             }
2087         }
2088         return inactive;
2089     }
2090 
2091     /**
2092      * Reactivates worker w if it is currently top of ctl stack
2093      *
2094      * @param w the work queue
2095      * @return 0 if now active
2096      */
2097     private int tryReactivate(WorkQueue w) {
2098         int inactive = 0;
2099         if (w != null) {                         // always true; hoist checks
2100             int sp = w.stackPred, phase, activePhase; long c;
2101             if ((inactive = (phase = w.phase) & IDLE) != 0 &&
2102                 (int)(c = ctl) == (activePhase = phase + IDLE) &&
2103                 compareAndSetCtl(c, (sp & LMASK) | ((c + RC_UNIT) & UMASK))) {
2104                 w.phase = activePhase;
2105                 inactive = 0;
2106             }
2107         }
2108         return inactive;
2109     }
2110 
2111     /**
2112      * Awaits signal or termination.
2113      *
2114      * @param w the work queue
2115      * @return 0 if now active
2116      */
2117     private int awaitWork(WorkQueue w) {
2118         int inactive = 0, phase;
2119         if (w != null) {                          // always true; hoist checks
2120             long waitTime = (w.source == INVALID_ID) ? 0L : keepAlive;
2121             if ((inactive = (phase = w.phase) & IDLE) != 0) {
2122                 LockSupport.setCurrentBlocker(this);
2123                 int activePhase = phase + IDLE;
2124                 for (long deadline = 0L;;) {
2125                     Thread.interrupted();         // clear status
2126                     if ((runState & STOP) != 0L)
2127                         break;
2128                     boolean trimmable = false;    // use timed wait if trimmable
2129                     long d = 0L, c;
2130                     if (((c = ctl) & RC_MASK) == 0L && (int)c == activePhase) {
2131                         long now = System.currentTimeMillis();
2132                         if (deadline == 0L)
2133                             deadline = waitTime + now;
2134                         if (deadline - now <= TIMEOUT_SLOP) {
2135                             if (tryTrim(w, c, activePhase))
2136                                 break;
2137                             continue;             // lost race to trim
2138                         }
2139                         d = deadline;
2140                         trimmable = true;
2141                     }
2142                     w.parking = 1;                // enable unpark and recheck
2143                     if ((inactive = w.phase & IDLE) != 0)
2144                         U.park(trimmable, d);
2145                     w.parking = 0;                // close unpark window
2146                     if (inactive == 0 || (inactive = w.phase & IDLE) == 0)
2147                         break;
2148                 }
2149                 LockSupport.setCurrentBlocker(null);
2150             }
2151         }
2152         return inactive;
2153     }
2154 
2155     /**
2156      * Tries to remove and deregister worker after timeout, and release
2157      * another to do the same unless new tasks are found.
2158      */
2159     private boolean tryTrim(WorkQueue w, long c, int activePhase) {
2160         if (w != null) {
2161             int vp, i; WorkQueue[] vs; WorkQueue v;
2162             long nc = ((w.stackPred & LMASK) |
2163                        ((RC_MASK & c) | (TC_MASK & (c - TC_UNIT))));
2164             if (compareAndSetCtl(c, nc)) {
2165                 w.source = DROPPED;
2166                 w.phase = activePhase;
2167                 if ((vp = (int)nc) != 0 && (vs = queues) != null &&
2168                     vs.length > (i = vp & SMASK) && (v = vs[i]) != null &&
2169                     compareAndSetCtl(           // try to wake up next waiter
2170                         nc, ((v.stackPred & LMASK) |
2171                              ((UMASK & (nc + RC_UNIT)) | (nc & TC_MASK))))) {
2172                     v.source = INVALID_ID;      // enable cascaded timeouts
2173                     v.phase = vp;
2174                     U.unpark(v.owner);
2175                 }
2176                 return true;
2177             }
2178         }
2179         return false;
2180     }
2181 
2182     /**
2183      * Scans for and returns a polled task, if available.  Used only
2184      * for untracked polls. Begins scan at a random index to avoid
2185      * systematic unfairness.
2186      *
2187      * @param submissionsOnly if true, only scan submission queues
2188      */
2189     private ForkJoinTask<?> pollScan(boolean submissionsOnly) {
2190         if ((runState & STOP) == 0L) {
2191             WorkQueue[] qs; int n; WorkQueue q; ForkJoinTask<?> t;
2192             int r = ThreadLocalRandom.nextSecondarySeed();
2193             if (submissionsOnly)                 // even indices only
2194                 r &= ~1;
2195             int step = (submissionsOnly) ? 2 : 1;
2196             if ((qs = queues) != null && (n = qs.length) > 0) {
2197                 for (int i = n; i > 0; i -= step, r += step) {
2198                     if ((q = qs[r & (n - 1)]) != null &&
2199                         (t = q.poll()) != null)
2200                         return t;
2201                 }
2202             }
2203         }
2204         return null;
2205     }
2206 
2207     /**
2208      * Tries to decrement counts (sometimes implicitly) and possibly
2209      * arrange for a compensating worker in preparation for
2210      * blocking. May fail due to interference, in which case -1 is
2211      * returned so caller may retry. A zero return value indicates
2212      * that the caller doesn't need to re-adjust counts when later
2213      * unblocked.
2214      *
2215      * @param c incoming ctl value
2216      * @return UNCOMPENSATE: block then adjust, 0: block, -1 : retry
2217      */
2218     private int tryCompensate(long c) {
2219         Predicate<? super ForkJoinPool> sat;
2220         long b = config;
2221         int pc        = parallelism,                    // unpack fields
2222             minActive = (short)(b >>> RC_SHIFT),
2223             maxTotal  = (short)(b >>> TC_SHIFT) + pc,
2224             active    = (short)(c >>> RC_SHIFT),
2225             total     = (short)(c >>> TC_SHIFT),
2226             sp        = (int)c,
2227             stat      = -1;                             // default retry return
2228         if (sp != 0 && active <= pc) {                  // activate idle worker
2229             WorkQueue[] qs; WorkQueue v; int i;
2230             if ((qs = queues) != null && qs.length > (i = sp & SMASK) &&
2231                 (v = qs[i]) != null &&
2232                 compareAndSetCtl(c, (c & UMASK) | (v.stackPred & LMASK))) {
2233                 v.phase = sp;
2234                 if (v.parking != 0)
2235                     U.unpark(v.owner);
2236                 stat = UNCOMPENSATE;
2237             }
2238         }
2239         else if (active > minActive && total >= pc) {   // reduce active workers
2240             if (compareAndSetCtl(c, ((c - RC_UNIT) & RC_MASK) | (c & ~RC_MASK)))
2241                 stat = UNCOMPENSATE;
2242         }
2243         else if (total < maxTotal && total < MAX_CAP) { // try to expand pool
2244             long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK);
2245             if ((runState & STOP) != 0L)                // terminating
2246                 stat = 0;
2247             else if (compareAndSetCtl(c, nc))
2248                 stat = createWorker() ? UNCOMPENSATE : 0;
2249         }
2250         else if (!compareAndSetCtl(c, c))               // validate
2251             ;
2252         else if ((sat = saturate) != null && sat.test(this))
2253             stat = 0;
2254         else
2255             throw new RejectedExecutionException(
2256                 "Thread limit exceeded replacing blocked worker");
2257         return stat;
2258     }
2259 
2260     /**
2261      * Readjusts RC count; called from ForkJoinTask after blocking.
2262      */
2263     final void uncompensate() {
2264         getAndAddCtl(RC_UNIT);
2265     }
2266 
2267     /**
2268      * Helps if possible until the given task is done.  Processes
2269      * compatible local tasks and scans other queues for task produced
2270      * by w's stealers; returning compensated blocking sentinel if
2271      * none are found.
2272      *
2273      * @param task the task
2274      * @param w caller's WorkQueue
2275      * @param internal true if w is owned by a ForkJoinWorkerThread
2276      * @return task status on exit, or UNCOMPENSATE for compensated blocking
2277      */
2278     final int helpJoin(ForkJoinTask<?> task, WorkQueue w, boolean internal) {
2279         if (w != null)
2280             w.tryRemoveAndExec(task, internal);
2281         int s = 0;
2282         if (task != null && (s = task.status) >= 0 && internal && w != null) {
2283             int wid = w.phase & SMASK, r = wid + 2, wsrc = w.source;
2284             long sctl = 0L;                             // track stability
2285             outer: for (boolean rescan = true;;) {
2286                 if ((s = task.status) < 0)
2287                     break;
2288                 if (!rescan) {
2289                     if ((runState & STOP) != 0L)
2290                         break;
2291                     if (sctl == (sctl = ctl) && (s = tryCompensate(sctl)) >= 0)
2292                         break;
2293                 }
2294                 rescan = false;
2295                 WorkQueue[] qs = queues;
2296                 int n = (qs == null) ? 0 : qs.length;
2297                 scan: for (int l = n >>> 1; l > 0; --l, r += 2) {
2298                     int j; WorkQueue q;
2299                     if ((q = qs[j = r & SMASK & (n - 1)]) != null) {
2300                         for (;;) {
2301                             ForkJoinTask<?> t; ForkJoinTask<?>[] a;
2302                             boolean eligible = false;
2303                             int sq = q.source, b, cap; long k;
2304                             if ((a = q.array) == null || (cap = a.length) <= 0)
2305                                 break;
2306                             t = (ForkJoinTask<?>)U.getReferenceAcquire(
2307                                 a, k = slotOffset((cap - 1) & (b = q.base)));
2308                             if (t == task)
2309                                 eligible = true;
2310                             else if (t != null) {       // check steal chain
2311                                 for (int v = sq, d = cap;;) {
2312                                     WorkQueue p;
2313                                     if (v == wid) {
2314                                         eligible = true;
2315                                         break;
2316                                     }
2317                                     if ((v & 1) == 0 || // external or none
2318                                         --d < 0 ||      // bound depth
2319                                         (p = qs[v & (n - 1)]) == null)
2320                                         break;
2321                                     v = p.source;
2322                                 }
2323                             }
2324                             if ((s = task.status) < 0)
2325                                 break outer;            // validate
2326                             if (q.source == sq && q.base == b &&
2327                                 U.getReference(a, k) == t) {
2328                                 if (!eligible) {        // revisit if nonempty
2329                                     if (!rescan && t == null && q.top - b > 0)
2330                                         rescan = true;
2331                                     break;
2332                                 }
2333                                 if (U.compareAndSetReference(a, k, t, null)) {
2334                                     q.base = b + 1;
2335                                     w.source = j;    // volatile write
2336                                     t.doExec();
2337                                     w.source = wsrc;
2338                                     rescan = true;   // restart at index r
2339                                     break scan;
2340                                 }
2341                             }
2342                         }
2343                     }
2344                 }
2345             }
2346         }
2347         return s;
2348     }
2349 
2350     /**
2351      * Version of helpJoin for CountedCompleters.
2352      *
2353      * @param task root of computation (only called when a CountedCompleter)
2354      * @param w caller's WorkQueue
2355      * @param internal true if w is owned by a ForkJoinWorkerThread
2356      * @return task status on exit, or UNCOMPENSATE for compensated blocking
2357      */
2358     final int helpComplete(ForkJoinTask<?> task, WorkQueue w, boolean internal) {
2359         int s = 0;
2360         if (task != null && (s = task.status) >= 0 && w != null) {
2361             int r = w.phase + 1;                          // for indexing
2362             long sctl = 0L;                               // track stability
2363             outer: for (boolean rescan = true, locals = true;;) {
2364                 if (locals && (s = w.helpComplete(task, internal, 0)) < 0)
2365                     break;
2366                 if ((s = task.status) < 0)
2367                     break;
2368                 if (!rescan) {
2369                     if ((runState & STOP) != 0L)
2370                         break;
2371                     if (sctl == (sctl = ctl) &&
2372                         (!internal || (s = tryCompensate(sctl)) >= 0))
2373                         break;
2374                 }
2375                 rescan = locals = false;
2376                 WorkQueue[] qs = queues;
2377                 int n = (qs == null) ? 0 : qs.length;
2378                 scan: for (int l = n; l > 0; --l, ++r) {
2379                     int j; WorkQueue q;
2380                     if ((q = qs[j = r & SMASK & (n - 1)]) != null) {
2381                         for (;;) {
2382                             ForkJoinTask<?> t; ForkJoinTask<?>[] a;
2383                             int b, cap, nb; long k;
2384                             boolean eligible = false;
2385                             if ((a = q.array) == null || (cap = a.length) <= 0)
2386                                 break;
2387                             t = (ForkJoinTask<?>)U.getReferenceAcquire(
2388                                 a, k = slotOffset((cap - 1) & (b = q.base)));
2389                             if (t instanceof CountedCompleter) {
2390                                 CountedCompleter<?> f = (CountedCompleter<?>)t;
2391                                 for (int steps = cap; steps > 0; --steps) {
2392                                     if (f == task) {
2393                                         eligible = true;
2394                                         break;
2395                                     }
2396                                     if ((f = f.completer) == null)
2397                                         break;
2398                                 }
2399                             }
2400                             if ((s = task.status) < 0)    // validate
2401                                 break outer;
2402                             if (q.base == b) {
2403                                 if (eligible) {
2404                                     if (U.compareAndSetReference(
2405                                             a, k, t, null)) {
2406                                         q.updateBase(b + 1);
2407                                         t.doExec();
2408                                         locals = rescan = true;
2409                                         break scan;
2410                                     }
2411                                 }
2412                                 else if (U.getReference(a, k) == t) {
2413                                     if (!rescan && t == null && q.top - b > 0)
2414                                         rescan = true;    // revisit
2415                                     break;
2416                                 }
2417                             }
2418                         }
2419                     }
2420                 }
2421             }
2422         }
2423         return s;
2424      }
2425 
2426     /**
2427      * Runs tasks until all workers are inactive and no tasks are
2428      * found. Rather than blocking when tasks cannot be found, rescans
2429      * until all others cannot find tasks either.
2430      *
2431      * @param nanos max wait time (Long.MAX_VALUE if effectively untimed)
2432      * @param interruptible true if return on interrupt
2433      * @return positive if quiescent, negative if interrupted, else 0
2434      */
2435     private int helpQuiesce(WorkQueue w, long nanos, boolean interruptible) {
2436         int phase; // w.phase inactive bit set when temporarily quiescent
2437         if (w == null || ((phase = w.phase) & IDLE) != 0)
2438             return 0;
2439         int wsrc = w.source;
2440         long startTime = System.nanoTime();
2441         long maxSleep = Math.min(nanos >>> 8, MAX_SLEEP); // approx 1% nanos
2442         long prevSum = 0L;
2443         int activePhase = phase, inactivePhase = phase + IDLE;
2444         int r = phase + 1, waits = 0, returnStatus = 1;
2445         boolean locals = true;
2446         for (long e = runState;;) {
2447             if ((e & STOP) != 0L)
2448                 break;                      // terminating
2449             if (interruptible && Thread.interrupted()) {
2450                 returnStatus = -1;
2451                 break;
2452             }
2453             if (locals) {                   // run local tasks before (re)polling
2454                 locals = false;
2455                 for (ForkJoinTask<?> u; (u = w.nextLocalTask()) != null;)
2456                     u.doExec();
2457             }
2458             WorkQueue[] qs = queues;
2459             int n = (qs == null) ? 0 : qs.length;
2460             long phaseSum = 0L;
2461             boolean rescan = false, busy = false;
2462             scan: for (int l = n; l > 0; --l, ++r) {
2463                 int j; WorkQueue q;
2464                 if ((q = qs[j = r & SMASK & (n - 1)]) != null && q != w) {
2465                     for (;;) {
2466                         ForkJoinTask<?> t; ForkJoinTask<?>[] a;
2467                         int b, cap; long k;
2468                         if ((a = q.array) == null || (cap = a.length) <= 0)
2469                             break;
2470                         t = (ForkJoinTask<?>)U.getReferenceAcquire(
2471                             a, k = slotOffset((cap - 1) & (b = q.base)));
2472                         if (t != null && phase == inactivePhase) // reactivate
2473                             w.phase = phase = activePhase;
2474                         if (q.base == b && U.getReference(a, k) == t) {
2475                             int nb = b + 1;
2476                             if (t == null) {
2477                                 if (!rescan) {
2478                                     int qp = q.phase, mq = qp & (IDLE | 1);
2479                                     phaseSum += qp;
2480                                     if (mq == 0 || q.top - b > 0)
2481                                         rescan = true;
2482                                     else if (mq == 1)
2483                                         busy = true;
2484                                 }
2485                                 break;
2486                             }
2487                             if (U.compareAndSetReference(a, k, t, null)) {
2488                                 q.base = nb;
2489                                 w.source = j; // volatile write
2490                                 t.doExec();
2491                                 w.source = wsrc;
2492                                 rescan = locals = true;
2493                                 break scan;
2494                             }
2495                         }
2496                     }
2497                 }
2498             }
2499             if (e != (e = runState) || prevSum != (prevSum = phaseSum) ||
2500                 rescan || (e & RS_LOCK) != 0L)
2501                 ;                   // inconsistent
2502             else if (!busy)
2503                 break;
2504             else if (phase == activePhase) {
2505                 waits = 0;          // recheck, then sleep
2506                 w.phase = phase = inactivePhase;
2507             }
2508             else if (System.nanoTime() - startTime > nanos) {
2509                 returnStatus = 0;   // timed out
2510                 break;
2511             }
2512             else if (waits == 0)   // same as spinLockRunState except
2513                 waits = MIN_SLEEP; //   with rescan instead of onSpinWait
2514             else {
2515                 LockSupport.parkNanos(this, (long)waits);
2516                 if (waits < maxSleep)
2517                     waits <<= 1;
2518             }
2519         }
2520         w.phase = activePhase;
2521         return returnStatus;
2522     }
2523 
2524     /**
2525      * Helps quiesce from external caller until done, interrupted, or timeout
2526      *
2527      * @param nanos max wait time (Long.MAX_VALUE if effectively untimed)
2528      * @param interruptible true if return on interrupt
2529      * @return positive if quiescent, negative if interrupted, else 0
2530      */
2531     private int externalHelpQuiesce(long nanos, boolean interruptible) {
2532         if (quiescent() < 0) {
2533             long startTime = System.nanoTime();
2534             long maxSleep = Math.min(nanos >>> 8, MAX_SLEEP);
2535             for (int waits = 0;;) {
2536                 ForkJoinTask<?> t;
2537                 if (interruptible && Thread.interrupted())
2538                     return -1;
2539                 else if ((t = pollScan(false)) != null) {
2540                     waits = 0;
2541                     t.doExec();
2542                 }
2543                 else if (quiescent() >= 0)
2544                     break;
2545                 else if (System.nanoTime() - startTime > nanos)
2546                     return 0;
2547                 else if (waits == 0)
2548                     waits = MIN_SLEEP;
2549                 else {
2550                     LockSupport.parkNanos(this, (long)waits);
2551                     if (waits < maxSleep)
2552                         waits <<= 1;
2553                 }
2554             }
2555         }
2556         return 1;
2557     }
2558 
2559     /**
2560      * Helps quiesce from either internal or external caller
2561      *
2562      * @param pool the pool to use, or null if any
2563      * @param nanos max wait time (Long.MAX_VALUE if effectively untimed)
2564      * @param interruptible true if return on interrupt
2565      * @return positive if quiescent, negative if interrupted, else 0
2566      */
2567     static final int helpQuiescePool(ForkJoinPool pool, long nanos,
2568                                      boolean interruptible) {
2569         Thread t; ForkJoinPool p; ForkJoinWorkerThread wt;
2570         if ((t = Thread.currentThread()) instanceof ForkJoinWorkerThread &&
2571             (p = (wt = (ForkJoinWorkerThread)t).pool) != null &&
2572             (p == pool || pool == null))
2573             return p.helpQuiesce(wt.workQueue, nanos, interruptible);
2574         else if ((p = pool) != null || (p = common) != null)
2575             return p.externalHelpQuiesce(nanos, interruptible);
2576         else
2577             return 0;
2578     }
2579 
2580     /**
2581      * Gets and removes a local or stolen task for the given worker.
2582      *
2583      * @return a task, if available
2584      */
2585     final ForkJoinTask<?> nextTaskFor(WorkQueue w) {
2586         ForkJoinTask<?> t;
2587         if (w == null || (t = w.nextLocalTask()) == null)
2588             t = pollScan(false);
2589         return t;
2590     }
2591 
2592     // External operations
2593 
2594     /**
2595      * Finds and locks a WorkQueue for an external submitter, or
2596      * throws RejectedExecutionException if shutdown
2597      * @param rejectOnShutdown true if RejectedExecutionException
2598      *        should be thrown when shutdown
2599      */
2600     final WorkQueue externalSubmissionQueue(boolean rejectOnShutdown) {
2601         int r;
2602         if ((r = ThreadLocalRandom.getProbe()) == 0) {
2603             ThreadLocalRandom.localInit();   // initialize caller's probe
2604             r = ThreadLocalRandom.getProbe();
2605         }
2606         for (;;) {
2607             WorkQueue q; WorkQueue[] qs; int n, id, i;
2608             if ((qs = queues) == null || (n = qs.length) <= 0)
2609                 break;
2610             if ((q = qs[i = (id = r & EXTERNAL_ID_MASK) & (n - 1)]) == null) {
2611                 WorkQueue newq = new WorkQueue(null, id, 0, false);
2612                 lockRunState();
2613                 if (qs[i] == null && queues == qs)
2614                     q = qs[i] = newq;         // else lost race to install
2615                 unlockRunState();
2616             }
2617             if (q != null && q.tryLockPhase()) {
2618                 if (rejectOnShutdown && (runState & SHUTDOWN) != 0L) {
2619                     q.unlockPhase();          // check while q lock held
2620                     break;
2621                 }
2622                 return q;
2623             }
2624             r = ThreadLocalRandom.advanceProbe(r); // move
2625         }
2626         throw new RejectedExecutionException();
2627     }
2628 
2629     private <T> ForkJoinTask<T> poolSubmit(boolean signalIfEmpty, ForkJoinTask<T> task) {
2630         Thread t; ForkJoinWorkerThread wt; WorkQueue q; boolean internal;
2631         if (((t = JLA.currentCarrierThread()) instanceof ForkJoinWorkerThread) &&
2632             (wt = (ForkJoinWorkerThread)t).pool == this) {
2633             internal = true;
2634             q = wt.workQueue;
2635         }
2636         else {                     // find and lock queue
2637             internal = false;
2638             q = externalSubmissionQueue(true);
2639         }
2640         q.push(task, signalIfEmpty ? this : null, internal);
2641         return task;
2642     }
2643 
2644     /**
2645      * Returns queue for an external thread, if one exists that has
2646      * possibly ever submitted to the given pool (nonzero probe), or
2647      * null if none.
2648      */
2649     static WorkQueue externalQueue(ForkJoinPool p) {
2650         WorkQueue[] qs; int n;
2651         int r = ThreadLocalRandom.getProbe();
2652         return (p != null && (qs = p.queues) != null &&
2653                 (n = qs.length) > 0 && r != 0) ?
2654             qs[r & EXTERNAL_ID_MASK & (n - 1)] : null;
2655     }
2656 
2657     /**
2658      * Returns external queue for common pool.
2659      */
2660     static WorkQueue commonQueue() {
2661         return externalQueue(common);
2662     }
2663 
2664     /**
2665      * If the given executor is a ForkJoinPool, poll and execute
2666      * AsynchronousCompletionTasks from worker's queue until none are
2667      * available or blocker is released.
2668      */
2669     static void helpAsyncBlocker(Executor e, ManagedBlocker blocker) {
2670         WorkQueue w = null; Thread t; ForkJoinWorkerThread wt;
2671         if (((t = Thread.currentThread()) instanceof ForkJoinWorkerThread) &&
2672             (wt = (ForkJoinWorkerThread)t).pool == e)
2673             w = wt.workQueue;
2674         else if (e instanceof ForkJoinPool)
2675             w = externalQueue((ForkJoinPool)e);
2676         if (w != null)
2677             w.helpAsyncBlocker(blocker);
2678     }
2679 
2680     /**
2681      * Returns a cheap heuristic guide for task partitioning when
2682      * programmers, frameworks, tools, or languages have little or no
2683      * idea about task granularity.  In essence, by offering this
2684      * method, we ask users only about tradeoffs in overhead vs
2685      * expected throughput and its variance, rather than how finely to
2686      * partition tasks.
2687      *
2688      * In a steady state strict (tree-structured) computation, each
2689      * thread makes available for stealing enough tasks for other
2690      * threads to remain active. Inductively, if all threads play by
2691      * the same rules, each thread should make available only a
2692      * constant number of tasks.
2693      *
2694      * The minimum useful constant is just 1. But using a value of 1
2695      * would require immediate replenishment upon each steal to
2696      * maintain enough tasks, which is infeasible.  Further,
2697      * partitionings/granularities of offered tasks should minimize
2698      * steal rates, which in general means that threads nearer the top
2699      * of computation tree should generate more than those nearer the
2700      * bottom. In perfect steady state, each thread is at
2701      * approximately the same level of computation tree. However,
2702      * producing extra tasks amortizes the uncertainty of progress and
2703      * diffusion assumptions.
2704      *
2705      * So, users will want to use values larger (but not much larger)
2706      * than 1 to both smooth over transient shortages and hedge
2707      * against uneven progress; as traded off against the cost of
2708      * extra task overhead. We leave the user to pick a threshold
2709      * value to compare with the results of this call to guide
2710      * decisions, but recommend values such as 3.
2711      *
2712      * When all threads are active, it is on average OK to estimate
2713      * surplus strictly locally. In steady-state, if one thread is
2714      * maintaining say 2 surplus tasks, then so are others. So we can
2715      * just use estimated queue length.  However, this strategy alone
2716      * leads to serious mis-estimates in some non-steady-state
2717      * conditions (ramp-up, ramp-down, other stalls). We can detect
2718      * many of these by further considering the number of "idle"
2719      * threads, that are known to have zero queued tasks, so
2720      * compensate by a factor of (#idle/#active) threads.
2721      */
2722     static int getSurplusQueuedTaskCount() {
2723         Thread t; ForkJoinWorkerThread wt; ForkJoinPool pool; WorkQueue q;
2724         if (((t = Thread.currentThread()) instanceof ForkJoinWorkerThread) &&
2725             (pool = (wt = (ForkJoinWorkerThread)t).pool) != null &&
2726             (q = wt.workQueue) != null) {
2727             int n = q.top - q.base;
2728             int p = pool.parallelism;
2729             int a = (short)(pool.ctl >>> RC_SHIFT);
2730             return n - (a > (p >>>= 1) ? 0 :
2731                         a > (p >>>= 1) ? 1 :
2732                         a > (p >>>= 1) ? 2 :
2733                         a > (p >>>= 1) ? 4 :
2734                         8);
2735         }
2736         return 0;
2737     }
2738 
2739     // Termination
2740 
2741     /**
2742      * Possibly initiates and/or completes pool termination.
2743      *
2744      * @param now if true, unconditionally terminate, else only
2745      * if no work and no active workers
2746      * @param enable if true, terminate when next possible
2747      * @return runState on exit
2748      */
2749     private long tryTerminate(boolean now, boolean enable) {
2750         long e, isShutdown, ps;
2751         if (((e = runState) & TERMINATED) != 0L)
2752             now = false;
2753         else if ((e & STOP) != 0L)
2754             now = true;
2755         else if (now) {
2756             if (((ps = getAndBitwiseOrRunState(SHUTDOWN|STOP) & STOP)) == 0L) {
2757                 if ((ps & RS_LOCK) != 0L) {
2758                     spinLockRunState(); // ensure queues array stable after stop
2759                     unlockRunState();
2760                 }
2761                 interruptAll();
2762             }
2763         }
2764         else if ((isShutdown = (e & SHUTDOWN)) != 0L || enable) {
2765             long quiet; DelayScheduler ds;
2766             if (isShutdown == 0L)
2767                 getAndBitwiseOrRunState(SHUTDOWN);
2768             if ((quiet = quiescent()) > 0)
2769                 now = true;
2770             else if (quiet == 0 && (ds = delayScheduler) != null)
2771                 ds.signal();
2772         }
2773 
2774         if (now) {
2775             DelayScheduler ds;
2776             releaseWaiters();
2777             if ((ds = delayScheduler) != null)
2778                 ds.signal();
2779             for (;;) {
2780                 if (((e = runState) & CLEANED) == 0L) {
2781                     boolean clean = cleanQueues();
2782                     if (((e = runState) & CLEANED) == 0L && clean)
2783                         e = getAndBitwiseOrRunState(CLEANED) | CLEANED;
2784                 }
2785                 if ((e & TERMINATED) != 0L)
2786                     break;
2787                 if (ctl != 0L) // else loop if didn't finish cleaning
2788                     break;
2789                 if ((ds = delayScheduler) != null && ds.signal() >= 0)
2790                     break;
2791                 if ((e & CLEANED) != 0L) {
2792                     e |= TERMINATED;
2793                     if ((getAndBitwiseOrRunState(TERMINATED) & TERMINATED) == 0L) {
2794                         CountDownLatch done; SharedThreadContainer ctr;
2795                         if ((done = termination) != null)
2796                             done.countDown();
2797                         if ((ctr = container) != null)
2798                             ctr.close();
2799                     }
2800                     break;
2801                 }
2802             }
2803         }
2804         return e;
2805     }
2806 
2807     /**
2808      * Scans queues in a psuedorandom order based on thread id,
2809      * cancelling tasks until empty, or returning early upon
2810      * interference or still-active external queues, in which case
2811      * other calls will finish cancellation.
2812      *
2813      * @return true if all queues empty
2814      */
2815     private boolean cleanQueues() {
2816         int r = (int)Thread.currentThread().threadId();
2817         r ^= r << 13; r ^= r >>> 17; r ^= r << 5; // xorshift
2818         int step = (r >>> 16) | 1;                // randomize traversals
2819         WorkQueue[] qs = queues;
2820         int n = (qs == null) ? 0 : qs.length;
2821         for (int l = n; l > 0; --l, r += step) {
2822             WorkQueue q; ForkJoinTask<?>[] a; int cap;
2823             if ((q = qs[r & (n - 1)]) != null &&
2824                 (a = q.array) != null && (cap = a.length) > 0) {
2825                 for (;;) {
2826                     ForkJoinTask<?> t; int b; long k;
2827                     t = (ForkJoinTask<?>)U.getReferenceAcquire(
2828                         a, k = slotOffset((cap - 1) & (b = q.base)));
2829                     if (q.base == b && t != null &&
2830                         U.compareAndSetReference(a, k, t, null)) {
2831                         q.updateBase(b + 1);
2832                         try {
2833                             t.cancel(false);
2834                         } catch (Throwable ignore) {
2835                         }
2836                     }
2837                     else if ((q.phase & (IDLE|1)) == 0 || // externally locked
2838                              q.top - q.base > 0)
2839                         return false;             // incomplete
2840                     else
2841                         break;
2842                 }
2843             }
2844         }
2845         return true;
2846     }
2847 
2848     /**
2849      * Interrupts all workers
2850      */
2851     private void interruptAll() {
2852         Thread current = Thread.currentThread();
2853         WorkQueue[] qs = queues;
2854         int n = (qs == null) ? 0 : qs.length;
2855         for (int i = 1; i < n; i += 2) {
2856             WorkQueue q; Thread o;
2857             if ((q = qs[i]) != null && (o = q.owner) != null && o != current) {
2858                 try {
2859                     o.interrupt();
2860                 } catch (Throwable ignore) {
2861                 }
2862             }
2863         }
2864     }
2865 
2866     /**
2867      * Returns termination signal, constructing if necessary
2868      */
2869     private CountDownLatch terminationSignal() {
2870         CountDownLatch signal, s, u;
2871         if ((signal = termination) == null)
2872             signal = ((u = cmpExTerminationSignal(
2873                            s = new CountDownLatch(1))) == null) ? s : u;
2874         return signal;
2875     }
2876 
2877     // Exported methods
2878 
2879     // Constructors
2880 
2881     /**
2882      * Creates a {@code ForkJoinPool} with parallelism equal to {@link
2883      * java.lang.Runtime#availableProcessors}, using defaults for all
2884      * other parameters (see {@link #ForkJoinPool(int,
2885      * ForkJoinWorkerThreadFactory, UncaughtExceptionHandler, boolean,
2886      * int, int, int, Predicate, long, TimeUnit)}).
2887      */
2888     public ForkJoinPool() {
2889         this(Math.min(MAX_CAP, Runtime.getRuntime().availableProcessors()),
2890              defaultForkJoinWorkerThreadFactory, null, false,
2891              0, MAX_CAP, 1, null, DEFAULT_KEEPALIVE, TimeUnit.MILLISECONDS);
2892     }
2893 
2894     /**
2895      * Creates a {@code ForkJoinPool} with the indicated parallelism
2896      * level, using defaults for all other parameters (see {@link
2897      * #ForkJoinPool(int, ForkJoinWorkerThreadFactory,
2898      * UncaughtExceptionHandler, boolean, int, int, int, Predicate,
2899      * long, TimeUnit)}).
2900      *
2901      * @param parallelism the parallelism level
2902      * @throws IllegalArgumentException if parallelism less than or
2903      *         equal to zero, or greater than implementation limit
2904      */
2905     public ForkJoinPool(int parallelism) {
2906         this(parallelism, defaultForkJoinWorkerThreadFactory, null, false,
2907              0, MAX_CAP, 1, null, DEFAULT_KEEPALIVE, TimeUnit.MILLISECONDS);
2908     }
2909 
2910     /**
2911      * Creates a {@code ForkJoinPool} with the given parameters (using
2912      * defaults for others -- see {@link #ForkJoinPool(int,
2913      * ForkJoinWorkerThreadFactory, UncaughtExceptionHandler, boolean,
2914      * int, int, int, Predicate, long, TimeUnit)}).
2915      *
2916      * @param parallelism the parallelism level. For default value,
2917      * use {@link java.lang.Runtime#availableProcessors}.
2918      * @param factory the factory for creating new threads. For default value,
2919      * use {@link #defaultForkJoinWorkerThreadFactory}.
2920      * @param handler the handler for internal worker threads that
2921      * terminate due to unrecoverable errors encountered while executing
2922      * tasks. For default value, use {@code null}.
2923      * @param asyncMode if true,
2924      * establishes local first-in-first-out scheduling mode for forked
2925      * tasks that are never joined. This mode may be more appropriate
2926      * than default locally stack-based mode in applications in which
2927      * worker threads only process event-style asynchronous tasks.
2928      * For default value, use {@code false}.
2929      * @throws IllegalArgumentException if parallelism less than or
2930      *         equal to zero, or greater than implementation limit
2931      * @throws NullPointerException if the factory is null
2932      */
2933     public ForkJoinPool(int parallelism,
2934                         ForkJoinWorkerThreadFactory factory,
2935                         UncaughtExceptionHandler handler,
2936                         boolean asyncMode) {
2937         this(parallelism, factory, handler, asyncMode,
2938              0, MAX_CAP, 1, null, DEFAULT_KEEPALIVE, TimeUnit.MILLISECONDS);
2939     }
2940 
2941     /**
2942      * Creates a {@code ForkJoinPool} with the given parameters.
2943      *
2944      * @param parallelism the parallelism level. For default value,
2945      * use {@link java.lang.Runtime#availableProcessors}.
2946      *
2947      * @param factory the factory for creating new threads. For
2948      * default value, use {@link #defaultForkJoinWorkerThreadFactory}.
2949      *
2950      * @param handler the handler for internal worker threads that
2951      * terminate due to unrecoverable errors encountered while
2952      * executing tasks. For default value, use {@code null}.
2953      *
2954      * @param asyncMode if true, establishes local first-in-first-out
2955      * scheduling mode for forked tasks that are never joined. This
2956      * mode may be more appropriate than default locally stack-based
2957      * mode in applications in which worker threads only process
2958      * event-style asynchronous tasks.  For default value, use {@code
2959      * false}.
2960      *
2961      * @param corePoolSize ignored: used in previous releases of this
2962      * class but no longer applicable. Using {@code 0} maintains
2963      * compatibility across releases.
2964      *
2965      * @param maximumPoolSize the maximum number of threads allowed.
2966      * When the maximum is reached, attempts to replace blocked
2967      * threads fail.  (However, because creation and termination of
2968      * different threads may overlap, and may be managed by the given
2969      * thread factory, this value may be transiently exceeded.)  To
2970      * arrange the same value as is used by default for the common
2971      * pool, use {@code 256} plus the {@code parallelism} level. (By
2972      * default, the common pool allows a maximum of 256 spare
2973      * threads.)  Using a value (for example {@code
2974      * Integer.MAX_VALUE}) larger than the implementation's total
2975      * thread limit has the same effect as using this limit (which is
2976      * the default).
2977      *
2978      * @param minimumRunnable the minimum allowed number of core
2979      * threads not blocked by a join or {@link ManagedBlocker}.  To
2980      * ensure progress, when too few unblocked threads exist and
2981      * unexecuted tasks may exist, new threads are constructed, up to
2982      * the given maximumPoolSize.  For the default value, use {@code
2983      * 1}, that ensures liveness.  A larger value might improve
2984      * throughput in the presence of blocked activities, but might
2985      * not, due to increased overhead.  A value of zero may be
2986      * acceptable when submitted tasks cannot have dependencies
2987      * requiring additional threads.
2988      *
2989      * @param saturate if non-null, a predicate invoked upon attempts
2990      * to create more than the maximum total allowed threads.  By
2991      * default, when a thread is about to block on a join or {@link
2992      * ManagedBlocker}, but cannot be replaced because the
2993      * maximumPoolSize would be exceeded, a {@link
2994      * RejectedExecutionException} is thrown.  But if this predicate
2995      * returns {@code true}, then no exception is thrown, so the pool
2996      * continues to operate with fewer than the target number of
2997      * runnable threads, which might not ensure progress.
2998      *
2999      * @param keepAliveTime the elapsed time since last use before
3000      * a thread is terminated (and then later replaced if needed).
3001      * For the default value, use {@code 60, TimeUnit.SECONDS}.
3002      *
3003      * @param unit the time unit for the {@code keepAliveTime} argument
3004      *
3005      * @throws IllegalArgumentException if parallelism is less than or
3006      *         equal to zero, or is greater than implementation limit,
3007      *         or if maximumPoolSize is less than parallelism,
3008      *         of if the keepAliveTime is less than or equal to zero.
3009      * @throws NullPointerException if the factory is null
3010      * @since 9
3011      */
3012     public ForkJoinPool(int parallelism,
3013                         ForkJoinWorkerThreadFactory factory,
3014                         UncaughtExceptionHandler handler,
3015                         boolean asyncMode,
3016                         int corePoolSize,
3017                         int maximumPoolSize,
3018                         int minimumRunnable,
3019                         Predicate<? super ForkJoinPool> saturate,
3020                         long keepAliveTime,
3021                         TimeUnit unit) {
3022         int p = parallelism;
3023         if (p <= 0 || p > MAX_CAP || p > maximumPoolSize || keepAliveTime <= 0L)
3024             throw new IllegalArgumentException();
3025         if (factory == null || unit == null)
3026             throw new NullPointerException();
3027         int size = Math.max(MIN_QUEUES_SIZE,
3028                             1 << (33 - Integer.numberOfLeadingZeros(p - 1)));
3029         this.parallelism = p;
3030         this.factory = factory;
3031         this.ueh = handler;
3032         this.saturate = saturate;
3033         this.keepAlive = Math.max(unit.toMillis(keepAliveTime), TIMEOUT_SLOP);
3034         int maxSpares = Math.clamp(maximumPoolSize - p, 0, MAX_CAP);
3035         int minAvail = Math.clamp(minimumRunnable, 0, MAX_CAP);
3036         this.config = (((asyncMode ? FIFO : 0) & LMASK) |
3037                        (((long)maxSpares) << TC_SHIFT) |
3038                        (((long)minAvail)  << RC_SHIFT));
3039         this.queues = new WorkQueue[size];
3040         String pid = Integer.toString(getAndAddPoolIds(1) + 1);
3041         String name = "ForkJoinPool-" + pid;
3042         this.poolName = name;
3043         this.workerNamePrefix = name + "-worker-";
3044         this.container = SharedThreadContainer.create(name);
3045     }
3046 
3047     /**
3048      * Constructor for common pool using parameters possibly
3049      * overridden by system properties
3050      */
3051     private ForkJoinPool(byte forCommonPoolOnly) {
3052         String name = "ForkJoinPool.commonPool";
3053         ForkJoinWorkerThreadFactory fac = defaultForkJoinWorkerThreadFactory;
3054         UncaughtExceptionHandler handler = null;
3055         int maxSpares = DEFAULT_COMMON_MAX_SPARES;
3056         int pc = 0, preset = 0; // nonzero if size set as property
3057         try {  // ignore exceptions in accessing/parsing properties
3058             String pp = System.getProperty
3059                 ("java.util.concurrent.ForkJoinPool.common.parallelism");
3060             if (pp != null) {
3061                 pc = Math.max(0, Integer.parseInt(pp));
3062                 preset = PRESET_SIZE;
3063             }
3064             String ms = System.getProperty
3065                 ("java.util.concurrent.ForkJoinPool.common.maximumSpares");
3066             if (ms != null)
3067                 maxSpares = Math.clamp(Integer.parseInt(ms), 0, MAX_CAP);
3068             String sf = System.getProperty
3069                 ("java.util.concurrent.ForkJoinPool.common.threadFactory");
3070             String sh = System.getProperty
3071                 ("java.util.concurrent.ForkJoinPool.common.exceptionHandler");
3072             if (sf != null || sh != null) {
3073                 ClassLoader ldr = ClassLoader.getSystemClassLoader();
3074                 if (sf != null)
3075                     fac = (ForkJoinWorkerThreadFactory)
3076                         ldr.loadClass(sf).getConstructor().newInstance();
3077                 if (sh != null)
3078                     handler = (UncaughtExceptionHandler)
3079                         ldr.loadClass(sh).getConstructor().newInstance();
3080             }
3081         } catch (Exception ignore) {
3082         }
3083         if (preset == 0)
3084             pc = Math.max(1, Runtime.getRuntime().availableProcessors() - 1);
3085         int p = Math.min(pc, MAX_CAP);
3086         int size = Math.max(MIN_QUEUES_SIZE,
3087                             (p == 0) ? 1 :
3088                             1 << (33 - Integer.numberOfLeadingZeros(p-1)));
3089         this.parallelism = p;
3090         this.config = ((preset & LMASK) | (((long)maxSpares) << TC_SHIFT) |
3091                        (1L << RC_SHIFT));
3092         this.factory = fac;
3093         this.ueh = handler;
3094         this.keepAlive = DEFAULT_KEEPALIVE;
3095         this.saturate = null;
3096         this.workerNamePrefix = null;
3097         this.poolName = name;
3098         this.queues = new WorkQueue[size];
3099         this.container = SharedThreadContainer.create(name);
3100     }
3101 
3102     /**
3103      * Returns the common pool instance. This pool is statically
3104      * constructed; its run state is unaffected by attempts to {@link
3105      * #shutdown} or {@link #shutdownNow}. However this pool and any
3106      * ongoing processing are automatically terminated upon program
3107      * {@link System#exit}.  Any program that relies on asynchronous
3108      * task processing to complete before program termination should
3109      * invoke {@code commonPool().}{@link #awaitQuiescence awaitQuiescence},
3110      * before exit.
3111      *
3112      * @return the common pool instance
3113      * @since 1.8
3114      */
3115     public static ForkJoinPool commonPool() {
3116         // assert common != null : "static init error";
3117         return common;
3118     }
3119 
3120     /**
3121      * Package-private access to commonPool overriding zero parallelism
3122      */
3123     static ForkJoinPool asyncCommonPool() {
3124         ForkJoinPool cp; int p;
3125         if ((p = (cp = common).parallelism) == 0)
3126             U.compareAndSetInt(cp, PARALLELISM, 0, 2);
3127         return cp;
3128     }
3129 
3130     // Execution methods
3131 
3132     /**
3133      * Performs the given task, returning its result upon completion.
3134      * If the computation encounters an unchecked Exception or Error,
3135      * it is rethrown as the outcome of this invocation.  Rethrown
3136      * exceptions behave in the same way as regular exceptions, but,
3137      * when possible, contain stack traces (as displayed for example
3138      * using {@code ex.printStackTrace()}) of both the current thread
3139      * as well as the thread actually encountering the exception;
3140      * minimally only the latter.
3141      *
3142      * @param task the task
3143      * @param <T> the type of the task's result
3144      * @return the task's result
3145      * @throws NullPointerException if the task is null
3146      * @throws RejectedExecutionException if the task cannot be
3147      *         scheduled for execution
3148      */
3149     public <T> T invoke(ForkJoinTask<T> task) {
3150         poolSubmit(true, Objects.requireNonNull(task));
3151         try {
3152             return task.join();
3153         } catch (RuntimeException | Error unchecked) {
3154             throw unchecked;
3155         } catch (Exception checked) {
3156             throw new RuntimeException(checked);
3157         }
3158     }
3159 
3160     /**
3161      * Arranges for (asynchronous) execution of the given task.
3162      *
3163      * @param task the task
3164      * @throws NullPointerException if the task is null
3165      * @throws RejectedExecutionException if the task cannot be
3166      *         scheduled for execution
3167      */
3168     public void execute(ForkJoinTask<?> task) {
3169         poolSubmit(true,  Objects.requireNonNull(task));
3170     }
3171 
3172     // AbstractExecutorService methods
3173 
3174     /**
3175      * @throws NullPointerException if the task is null
3176      * @throws RejectedExecutionException if the task cannot be
3177      *         scheduled for execution
3178      */
3179     @Override
3180     @SuppressWarnings("unchecked")
3181     public void execute(Runnable task) {
3182         poolSubmit(true, (Objects.requireNonNull(task) instanceof ForkJoinTask<?>)
3183                    ? (ForkJoinTask<Void>) task // avoid re-wrap
3184                    : new ForkJoinTask.RunnableExecuteAction(task));
3185     }
3186 
3187     /**
3188      * Submits a ForkJoinTask for execution.
3189      *
3190      * @implSpec
3191      * This method is equivalent to {@link #externalSubmit(ForkJoinTask)}
3192      * when called from a thread that is not in this pool.
3193      *
3194      * @param task the task to submit
3195      * @param <T> the type of the task's result
3196      * @return the task
3197      * @throws NullPointerException if the task is null
3198      * @throws RejectedExecutionException if the task cannot be
3199      *         scheduled for execution
3200      */
3201     public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) {
3202         return poolSubmit(true,  Objects.requireNonNull(task));
3203     }
3204 
3205     /**
3206      * @throws NullPointerException if the task is null
3207      * @throws RejectedExecutionException if the task cannot be
3208      *         scheduled for execution
3209      */
3210     @Override
3211     public <T> ForkJoinTask<T> submit(Callable<T> task) {
3212         Objects.requireNonNull(task);
3213         return poolSubmit(
3214             true,
3215             (Thread.currentThread() instanceof ForkJoinWorkerThread) ?
3216             new ForkJoinTask.AdaptedCallable<T>(task) :
3217             new ForkJoinTask.AdaptedInterruptibleCallable<T>(task));
3218     }
3219 
3220     /**
3221      * @throws NullPointerException if the task is null
3222      * @throws RejectedExecutionException if the task cannot be
3223      *         scheduled for execution
3224      */
3225     @Override
3226     public <T> ForkJoinTask<T> submit(Runnable task, T result) {
3227         Objects.requireNonNull(task);
3228         return poolSubmit(
3229             true,
3230             (Thread.currentThread() instanceof ForkJoinWorkerThread) ?
3231             new ForkJoinTask.AdaptedRunnable<T>(task, result) :
3232             new ForkJoinTask.AdaptedInterruptibleRunnable<T>(task, result));
3233     }
3234 
3235     /**
3236      * @throws NullPointerException if the task is null
3237      * @throws RejectedExecutionException if the task cannot be
3238      *         scheduled for execution
3239      */
3240     @Override
3241     @SuppressWarnings("unchecked")
3242     public ForkJoinTask<?> submit(Runnable task) {
3243         Objects.requireNonNull(task);
3244         return poolSubmit(
3245             true,
3246             (task instanceof ForkJoinTask<?>) ?
3247             (ForkJoinTask<Void>) task : // avoid re-wrap
3248             ((Thread.currentThread() instanceof ForkJoinWorkerThread) ?
3249              new ForkJoinTask.AdaptedRunnable<Void>(task, null) :
3250              new ForkJoinTask.AdaptedInterruptibleRunnable<Void>(task, null)));
3251     }
3252 
3253     /**
3254      * Submits the given task as if submitted from a non-{@code ForkJoinTask}
3255      * client. The task is added to a scheduling queue for submissions to the
3256      * pool even when called from a thread in the pool.
3257      *
3258      * @implSpec
3259      * This method is equivalent to {@link #submit(ForkJoinTask)} when called
3260      * from a thread that is not in this pool.
3261      *
3262      * @return the task
3263      * @param task the task to submit
3264      * @param <T> the type of the task's result
3265      * @throws NullPointerException if the task is null
3266      * @throws RejectedExecutionException if the task cannot be
3267      *         scheduled for execution
3268      * @since 20
3269      */
3270     public <T> ForkJoinTask<T> externalSubmit(ForkJoinTask<T> task) {
3271         Objects.requireNonNull(task);
3272         externalSubmissionQueue(true).push(task, this, false);
3273         return task;
3274     }
3275 
3276     /**
3277      * Submits the given task without guaranteeing that it will
3278      * eventually execute in the absence of available active threads.
3279      * In some contexts, this method may reduce contention and
3280      * overhead by relying on context-specific knowledge that existing
3281      * threads (possibly including the calling thread if operating in
3282      * this pool) will eventually be available to execute the task.
3283      *
3284      * @param task the task
3285      * @param <T> the type of the task's result
3286      * @return the task
3287      * @throws NullPointerException if the task is null
3288      * @throws RejectedExecutionException if the task cannot be
3289      *         scheduled for execution
3290      * @since 19
3291      */
3292     public <T> ForkJoinTask<T> lazySubmit(ForkJoinTask<T> task) {
3293         return poolSubmit(false,  Objects.requireNonNull(task));
3294     }
3295 
3296     /**
3297      * Changes the target parallelism of this pool, controlling the
3298      * future creation, use, and termination of worker threads.
3299      * Applications include contexts in which the number of available
3300      * processors changes over time.
3301      *
3302      * @implNote This implementation restricts the maximum number of
3303      * running threads to 32767
3304      *
3305      * @param size the target parallelism level
3306      * @return the previous parallelism level.
3307      * @throws IllegalArgumentException if size is less than 1 or
3308      *         greater than the maximum supported by this pool.
3309      * @throws UnsupportedOperationException this is the{@link
3310      *         #commonPool()} and parallelism level was set by System
3311      *         property {@systemProperty
3312      *         java.util.concurrent.ForkJoinPool.common.parallelism}.
3313      * @since 19
3314      */
3315     public int setParallelism(int size) {
3316         int prevSize;
3317         if (size < 1 || size > MAX_CAP)
3318             throw new IllegalArgumentException();
3319         if ((config & PRESET_SIZE) != 0)
3320             throw new UnsupportedOperationException("Cannot override System property");
3321         if ((prevSize = getAndSetParallelism(size)) < size)
3322             signalWork(null, 0); // trigger worker activation
3323         return prevSize;
3324     }
3325 
3326     /**
3327      * Uninterrupible version of {@code invokeAll}. Executes the given
3328      * tasks, returning a list of Futures holding their status and
3329      * results when all complete, ignoring interrupts.  {@link
3330      * Future#isDone} is {@code true} for each element of the returned
3331      * list.  Note that a <em>completed</em> task could have
3332      * terminated either normally or by throwing an exception.  The
3333      * results of this method are undefined if the given collection is
3334      * modified while this operation is in progress.
3335      *
3336      * @apiNote This method supports usages that previously relied on an
3337      * incompatible override of
3338      * {@link ExecutorService#invokeAll(java.util.Collection)}.
3339      *
3340      * @param tasks the collection of tasks
3341      * @param <T> the type of the values returned from the tasks
3342      * @return a list of Futures representing the tasks, in the same
3343      *         sequential order as produced by the iterator for the
3344      *         given task list, each of which has completed
3345      * @throws NullPointerException if tasks or any of its elements are {@code null}
3346      * @throws RejectedExecutionException if any task cannot be
3347      *         scheduled for execution
3348      * @since 22
3349      */
3350     public <T> List<Future<T>> invokeAllUninterruptibly(Collection<? extends Callable<T>> tasks) {
3351         ArrayList<Future<T>> futures = new ArrayList<>(tasks.size());
3352         try {
3353             for (Callable<T> t : tasks) {
3354                 ForkJoinTask<T> f = ForkJoinTask.adapt(t);
3355                 futures.add(f);
3356                 poolSubmit(true, f);
3357             }
3358             for (int i = futures.size() - 1; i >= 0; --i)
3359                 ((ForkJoinTask<?>)futures.get(i)).quietlyJoin();
3360             return futures;
3361         } catch (Throwable t) {
3362             for (Future<T> e : futures)
3363                 e.cancel(true);
3364             throw t;
3365         }
3366     }
3367 
3368     /**
3369      * Common support for timed and untimed invokeAll
3370      */
3371     private  <T> List<Future<T>> invokeAll(Collection<? extends Callable<T>> tasks,
3372                                            long deadline)
3373         throws InterruptedException {
3374         ArrayList<Future<T>> futures = new ArrayList<>(tasks.size());
3375         try {
3376             for (Callable<T> t : tasks) {
3377                 ForkJoinTask<T> f = ForkJoinTask.adaptInterruptible(t);
3378                 futures.add(f);
3379                 poolSubmit(true, f);
3380             }
3381             for (int i = futures.size() - 1; i >= 0; --i)
3382                 ((ForkJoinTask<?>)futures.get(i))
3383                     .quietlyJoinPoolInvokeAllTask(deadline);
3384             return futures;
3385         } catch (Throwable t) {
3386             for (Future<T> e : futures)
3387                 e.cancel(true);
3388             throw t;
3389         }
3390     }
3391 
3392     @Override
3393     public <T> List<Future<T>> invokeAll(Collection<? extends Callable<T>> tasks)
3394         throws InterruptedException {
3395         return invokeAll(tasks, 0L);
3396     }
3397     // for jdk version < 22, replace with
3398     // /**
3399     //  * @throws NullPointerException       {@inheritDoc}
3400     //  * @throws RejectedExecutionException {@inheritDoc}
3401     //  */
3402     // @Override
3403     // public <T> List<Future<T>> invokeAll(Collection<? extends Callable<T>> tasks) {
3404     //     return invokeAllUninterruptibly(tasks);
3405     // }
3406 
3407     @Override
3408     public <T> List<Future<T>> invokeAll(Collection<? extends Callable<T>> tasks,
3409                                          long timeout, TimeUnit unit)
3410         throws InterruptedException {
3411         return invokeAll(tasks, (System.nanoTime() + unit.toNanos(timeout)) | 1L);
3412     }
3413 
3414     @Override
3415     public <T> T invokeAny(Collection<? extends Callable<T>> tasks)
3416         throws InterruptedException, ExecutionException {
3417         try {
3418             return new ForkJoinTask.InvokeAnyRoot<T>()
3419                 .invokeAny(tasks, this, false, 0L);
3420         } catch (TimeoutException cannotHappen) {
3421             assert false;
3422             return null;
3423         }
3424     }
3425 
3426     @Override
3427     public <T> T invokeAny(Collection<? extends Callable<T>> tasks,
3428                            long timeout, TimeUnit unit)
3429         throws InterruptedException, ExecutionException, TimeoutException {
3430         return new ForkJoinTask.InvokeAnyRoot<T>()
3431             .invokeAny(tasks, this, true, unit.toNanos(timeout));
3432     }
3433 
3434     // Support for delayed tasks
3435 
3436     /**
3437      * Returns STOP and SHUTDOWN status (zero if neither), masking or
3438      * truncating out other bits.
3439      */
3440     final int shutdownStatus(DelayScheduler ds) {
3441         return (int)(runState & (SHUTDOWN | STOP));
3442     }
3443 
3444     /**
3445      * Tries to stop and possibly terminate if already enabled, return success.
3446      */
3447     final boolean tryStopIfShutdown(DelayScheduler ds) {
3448         return (tryTerminate(false, false) & STOP) != 0L;
3449     }
3450 
3451     /**
3452      *  Creates and starts DelayScheduler
3453      */
3454     private DelayScheduler startDelayScheduler() {
3455         DelayScheduler ds;
3456         if ((ds = delayScheduler) == null) {
3457             boolean start = false;
3458             String name = poolName + "-delayScheduler";
3459             if (workerNamePrefix == null)
3460                 asyncCommonPool();  // override common parallelism zero
3461             long isShutdown = lockRunState() & SHUTDOWN;
3462             try {
3463                 if (isShutdown == 0L && (ds = delayScheduler) == null) {
3464                     ds = delayScheduler = new DelayScheduler(this, name);
3465                     start = true;
3466                 }
3467             } finally {
3468                 unlockRunState();
3469             }
3470             if (start) { // start outside of lock
3471                 SharedThreadContainer ctr;
3472                 try {
3473                     if ((ctr = container) != null)
3474                         ctr.start(ds);
3475                     else
3476                         ds.start();
3477                 } catch (RuntimeException | Error ex) { // back out
3478                     lockRunState();
3479                     ds = delayScheduler = null;
3480                     unlockRunState();
3481                     tryTerminate(false, false);
3482                     if (ex instanceof Error)
3483                         throw ex;
3484                 }
3485             }
3486         }
3487         return ds;
3488     }
3489 
3490     /**
3491      * Arranges execution of a ScheduledForkJoinTask whose delay has
3492      * elapsed
3493      */
3494     final void executeEnabledScheduledTask(ScheduledForkJoinTask<?> task) {
3495         externalSubmissionQueue(false).push(task, this, false);
3496     }
3497 
3498     /**
3499      * Arranges delayed execution of a ScheduledForkJoinTask via the
3500      * DelayScheduler, creating and starting it if necessary.
3501      * @return the task
3502      */
3503     final <T> ScheduledForkJoinTask<T> scheduleDelayedTask(ScheduledForkJoinTask<T> task) {
3504         DelayScheduler ds;
3505         if (((ds = delayScheduler) == null &&
3506              (ds = startDelayScheduler()) == null) ||
3507             (runState & SHUTDOWN) != 0L)
3508             throw new RejectedExecutionException();
3509         ds.pend(task);
3510         return task;
3511     }
3512 
3513     /**
3514      * Submits a one-shot task that becomes enabled for execution after the given
3515      * delay.  At that point it will execute unless explicitly
3516      * cancelled, or fail to execute (eventually reporting
3517      * cancellation) when encountering resource exhaustion, or the
3518      * pool is {@link #shutdownNow}, or is {@link #shutdown} when
3519      * otherwise quiescent and {@link #cancelDelayedTasksOnShutdown}
3520      * is in effect.
3521      *
3522      * @param command the task to execute
3523      * @param delay the time from now to delay execution
3524      * @param unit the time unit of the delay parameter
3525      * @return a ForkJoinTask implementing the ScheduledFuture
3526      *         interface, whose {@code get()} method will return
3527      *         {@code null} upon normal completion.
3528      * @throws RejectedExecutionException if the pool is shutdown or
3529      *         submission encounters resource exhaustion.
3530      * @throws NullPointerException if command or unit is null
3531      * @since 25
3532      */
3533     public ScheduledFuture<?> schedule(Runnable command,
3534                                        long delay, TimeUnit unit) {
3535         return scheduleDelayedTask(
3536             new ScheduledForkJoinTask<Void>(
3537                 unit.toNanos(delay), 0L, false, // implicit null check of unit
3538                 Objects.requireNonNull(command), null, this));
3539     }
3540 
3541     /**
3542      * Submits a value-returning one-shot task that becomes enabled for execution
3543      * after the given delay. At that point it will execute unless
3544      * explicitly cancelled, or fail to execute (eventually reporting
3545      * cancellation) when encountering resource exhaustion, or the
3546      * pool is {@link #shutdownNow}, or is {@link #shutdown} when
3547      * otherwise quiescent and {@link #cancelDelayedTasksOnShutdown}
3548      * is in effect.
3549      *
3550      * @param callable the function to execute
3551      * @param delay the time from now to delay execution
3552      * @param unit the time unit of the delay parameter
3553      * @param <V> the type of the callable's result
3554      * @return a ForkJoinTask implementing the ScheduledFuture
3555      *         interface, whose {@code get()} method will return the
3556      *         value from the callable upon normal completion.
3557      * @throws RejectedExecutionException if the pool is shutdown or
3558      *         submission encounters resource exhaustion.
3559      * @throws NullPointerException if command or unit is null
3560      * @since 25
3561      */
3562     public <V> ScheduledFuture<V> schedule(Callable<V> callable,
3563                                            long delay, TimeUnit unit) {
3564         return scheduleDelayedTask(
3565             new ScheduledForkJoinTask<V>(
3566                 unit.toNanos(delay), 0L, false, null,  // implicit null check of unit
3567                 Objects.requireNonNull(callable), this));
3568     }
3569 
3570     /**
3571      * Submits a periodic action that becomes enabled for execution first after the
3572      * given initial delay, and subsequently with the given period;
3573      * that is, executions will commence after
3574      * {@code initialDelay}, then {@code initialDelay + period}, then
3575      * {@code initialDelay + 2 * period}, and so on.
3576      *
3577      * <p>The sequence of task executions continues indefinitely until
3578      * one of the following exceptional completions occur:
3579      * <ul>
3580      * <li>The task is {@linkplain Future#cancel explicitly cancelled}
3581      * <li>Method {@link #shutdownNow} is called
3582      * <li>Method {@link #shutdown} is called and the pool is
3583      * otherwise quiescent, in which case existing executions continue
3584      * but subsequent executions do not.
3585      * <li>An execution or the task encounters resource exhaustion.
3586      * <li>An execution of the task throws an exception.  In this case
3587      * calling {@link Future#get() get} on the returned future will throw
3588      * {@link ExecutionException}, holding the exception as its cause.
3589      * </ul>
3590      * Subsequent executions are suppressed.  Subsequent calls to
3591      * {@link Future#isDone isDone()} on the returned future will
3592      * return {@code true}.
3593      *
3594      * <p>If any execution of this task takes longer than its period, then
3595      * subsequent executions may start late, but will not concurrently
3596      * execute.
3597      * @param command the task to execute
3598      * @param initialDelay the time to delay first execution
3599      * @param period the period between successive executions
3600      * @param unit the time unit of the initialDelay and period parameters
3601      * @return a ForkJoinTask implementing the ScheduledFuture
3602      *         interface.  The future's {@link Future#get() get()}
3603      *         method will never return normally, and will throw an
3604      *         exception upon task cancellation or abnormal
3605      *         termination of a task execution.
3606      * @throws RejectedExecutionException if the pool is shutdown or
3607      *         submission encounters resource exhaustion.
3608      * @throws NullPointerException if command or unit is null
3609      * @throws IllegalArgumentException if period less than or equal to zero
3610      * @since 25
3611      */
3612     public ScheduledFuture<?> scheduleAtFixedRate(Runnable command,
3613                                                   long initialDelay,
3614                                                   long period, TimeUnit unit) {
3615         if (period <= 0L)
3616             throw new IllegalArgumentException();
3617         return scheduleDelayedTask(
3618             new ScheduledForkJoinTask<Void>(
3619                 unit.toNanos(initialDelay),  // implicit null check of unit
3620                 unit.toNanos(period), false,
3621                 Objects.requireNonNull(command), null, this));
3622     }
3623 
3624     /**
3625      * Submits a periodic action that becomes enabled for execution first after the
3626      * given initial delay, and subsequently with the given delay
3627      * between the termination of one execution and the commencement of
3628      * the next.
3629      * <p>The sequence of task executions continues indefinitely until
3630      * one of the following exceptional completions occur:
3631      * <ul>
3632      * <li>The task is {@linkplain Future#cancel explicitly cancelled}
3633      * <li>Method {@link #shutdownNow} is called
3634      * <li>Method {@link #shutdown} is called and the pool is
3635      * otherwise quiescent, in which case existing executions continue
3636      * but subsequent executions do not.
3637      * <li>An execution or the task encounters resource exhaustion.
3638      * <li>An execution of the task throws an exception.  In this case
3639      * calling {@link Future#get() get} on the returned future will throw
3640      * {@link ExecutionException}, holding the exception as its cause.
3641      * </ul>
3642      * Subsequent executions are suppressed.  Subsequent calls to
3643      * {@link Future#isDone isDone()} on the returned future will
3644      * return {@code true}.
3645      * @param command the task to execute
3646      * @param initialDelay the time to delay first execution
3647      * @param delay the delay between the termination of one
3648      * execution and the commencement of the next
3649      * @param unit the time unit of the initialDelay and delay parameters
3650      * @return a ForkJoinTask implementing the ScheduledFuture
3651      *         interface.  The future's {@link Future#get() get()}
3652      *         method will never return normally, and will throw an
3653      *         exception upon task cancellation or abnormal
3654      *         termination of a task execution.
3655      * @throws RejectedExecutionException if the pool is shutdown or
3656      *         submission encounters resource exhaustion.
3657      * @throws NullPointerException if command or unit is null
3658      * @throws IllegalArgumentException if delay less than or equal to zero
3659      * @since 25
3660      */
3661     public ScheduledFuture<?> scheduleWithFixedDelay(Runnable command,
3662                                                      long initialDelay,
3663                                                      long delay, TimeUnit unit) {
3664         if (delay <= 0L)
3665             throw new IllegalArgumentException();
3666         return scheduleDelayedTask(
3667             new ScheduledForkJoinTask<Void>(
3668                 unit.toNanos(initialDelay),  // implicit null check of unit
3669                 -unit.toNanos(delay), false, // negative for fixed delay
3670                 Objects.requireNonNull(command), null, this));
3671     }
3672 
3673     /**
3674      * Body of a task performed on timeout of another task
3675      */
3676     static final class TimeoutAction<V> implements Runnable {
3677         // set after construction, nulled after use
3678         ForkJoinTask.CallableWithTimeout<V> task;
3679         Consumer<? super ForkJoinTask<V>> action;
3680         TimeoutAction(Consumer<? super ForkJoinTask<V>> action) {
3681             this.action = action;
3682         }
3683         public void run() {
3684             ForkJoinTask.CallableWithTimeout<V> t = task;
3685             Consumer<? super ForkJoinTask<V>> a = action;
3686             task = null;
3687             action = null;
3688             if (t != null && t.status >= 0) {
3689                 if (a == null)
3690                     t.cancel(true);
3691                 else {
3692                     a.accept(t);
3693                     t.interruptIfRunning(true);
3694                 }
3695             }
3696         }
3697     }
3698 
3699     /**
3700      * Submits a task executing the given function, cancelling the
3701      * task or performing a given timeoutAction if not completed
3702      * within the given timeout period. If the optional {@code
3703      * timeoutAction} is null, the task is cancelled (via {@code
3704      * cancel(true)}.  Otherwise, the action is applied and the task
3705      * may be interrupted if running. Actions may include {@link
3706      * ForkJoinTask#complete} to set a replacement value or {@link
3707      * ForkJoinTask#completeExceptionally} to throw an appropriate
3708      * exception. Note that these can succeed only if the task has
3709      * not already completed when the timeoutAction executes.
3710      *
3711      * @param callable the function to execute
3712      * @param <V> the type of the callable's result
3713      * @param timeout the time to wait before cancelling if not completed
3714      * @param timeoutAction if nonnull, an action to perform on
3715      *        timeout, otherwise the default action is to cancel using
3716      *        {@code cancel(true)}.
3717      * @param unit the time unit of the timeout parameter
3718      * @return a Future that can be used to extract result or cancel
3719      * @throws RejectedExecutionException if the task cannot be
3720      *         scheduled for execution
3721      * @throws NullPointerException if callable or unit is null
3722      * @since 25
3723      */
3724     public <V> ForkJoinTask<V> submitWithTimeout(Callable<V> callable,
3725                                                  long timeout, TimeUnit unit,
3726                                                  Consumer<? super ForkJoinTask<V>> timeoutAction) {
3727         ForkJoinTask.CallableWithTimeout<V> task; TimeoutAction<V> onTimeout;
3728         Objects.requireNonNull(callable);
3729         ScheduledForkJoinTask<Void> timeoutTask =
3730             new ScheduledForkJoinTask<Void>(
3731                 unit.toNanos(timeout), 0L, true,
3732                 onTimeout = new TimeoutAction<V>(timeoutAction), null, this);
3733         onTimeout.task = task =
3734             new ForkJoinTask.CallableWithTimeout<V>(callable, timeoutTask);
3735         scheduleDelayedTask(timeoutTask);
3736         return poolSubmit(true, task);
3737     }
3738 
3739     /**
3740      * Arranges that scheduled tasks that are not executing and have
3741      * not already been enabled for execution will not be executed and
3742      * will be cancelled upon {@link #shutdown} (unless this pool is
3743      * the {@link #commonPool()} which never shuts down). This method
3744      * may be invoked either before {@link #shutdown} to take effect
3745      * upon the next call, or afterwards to cancel such tasks, which
3746      * may then allow termination. Note that subsequent executions of
3747      * periodic tasks are always disabled upon shutdown, so this
3748      * method applies meaningfully only to non-periodic tasks.
3749      * @since 25
3750      */
3751     public void cancelDelayedTasksOnShutdown() {
3752         DelayScheduler ds;
3753         if ((ds = delayScheduler) != null ||
3754             (ds = startDelayScheduler()) != null)
3755             ds.cancelDelayedTasksOnShutdown();
3756     }
3757 
3758     /**
3759      * Returns the factory used for constructing new workers.
3760      *
3761      * @return the factory used for constructing new workers
3762      */
3763     public ForkJoinWorkerThreadFactory getFactory() {
3764         return factory;
3765     }
3766 
3767     /**
3768      * Returns the handler for internal worker threads that terminate
3769      * due to unrecoverable errors encountered while executing tasks.
3770      *
3771      * @return the handler, or {@code null} if none
3772      */
3773     public UncaughtExceptionHandler getUncaughtExceptionHandler() {
3774         return ueh;
3775     }
3776 
3777     /**
3778      * Returns the targeted parallelism level of this pool.
3779      *
3780      * @return the targeted parallelism level of this pool
3781      */
3782     public int getParallelism() {
3783         return Math.max(getParallelismOpaque(), 1);
3784     }
3785 
3786     /**
3787      * Returns the targeted parallelism level of the common pool.
3788      *
3789      * @return the targeted parallelism level of the common pool
3790      * @since 1.8
3791      */
3792     public static int getCommonPoolParallelism() {
3793         return common.getParallelism();
3794     }
3795 
3796     /**
3797      * Returns the number of worker threads that have started but not
3798      * yet terminated.  The result returned by this method may differ
3799      * from {@link #getParallelism} when threads are created to
3800      * maintain parallelism when others are cooperatively blocked.
3801      *
3802      * @return the number of worker threads
3803      */
3804     public int getPoolSize() {
3805         return (short)(ctl >>> TC_SHIFT);
3806     }
3807 
3808     /**
3809      * Returns {@code true} if this pool uses local first-in-first-out
3810      * scheduling mode for forked tasks that are never joined.
3811      *
3812      * @return {@code true} if this pool uses async mode
3813      */
3814     public boolean getAsyncMode() {
3815         return (config & FIFO) != 0;
3816     }
3817 
3818     /**
3819      * Returns an estimate of the number of worker threads that are
3820      * not blocked waiting to join tasks or for other managed
3821      * synchronization. This method may overestimate the
3822      * number of running threads.
3823      *
3824      * @return the number of worker threads
3825      */
3826     public int getRunningThreadCount() {
3827         WorkQueue[] qs; WorkQueue q;
3828         int rc = 0;
3829         if ((runState & TERMINATED) == 0L && (qs = queues) != null) {
3830             for (int i = 1; i < qs.length; i += 2) {
3831                 if ((q = qs[i]) != null && q.isApparentlyUnblocked())
3832                     ++rc;
3833             }
3834         }
3835         return rc;
3836     }
3837 
3838     /**
3839      * Returns an estimate of the number of threads that are currently
3840      * stealing or executing tasks. This method may overestimate the
3841      * number of active threads.
3842      *
3843      * @return the number of active threads
3844      */
3845     public int getActiveThreadCount() {
3846         return Math.max((short)(ctl >>> RC_SHIFT), 0);
3847     }
3848 
3849     /**
3850      * Returns {@code true} if all worker threads are currently idle.
3851      * An idle worker is one that cannot obtain a task to execute
3852      * because none are available to steal from other threads, and
3853      * there are no pending submissions to the pool. This method is
3854      * conservative; it might not return {@code true} immediately upon
3855      * idleness of all threads, but will eventually become true if
3856      * threads remain inactive.
3857      *
3858      * @return {@code true} if all threads are currently idle
3859      */
3860     public boolean isQuiescent() {
3861         return quiescent() >= 0;
3862     }
3863 
3864     /**
3865      * Returns an estimate of the total number of completed tasks that
3866      * were executed by a thread other than their submitter. The
3867      * reported value underestimates the actual total number of steals
3868      * when the pool is not quiescent. This value may be useful for
3869      * monitoring and tuning fork/join programs: in general, steal
3870      * counts should be high enough to keep threads busy, but low
3871      * enough to avoid overhead and contention across threads.
3872      *
3873      * @return the number of steals
3874      */
3875     public long getStealCount() {
3876         long count = stealCount;
3877         WorkQueue[] qs; WorkQueue q;
3878         if ((qs = queues) != null) {
3879             for (int i = 1; i < qs.length; i += 2) {
3880                 if ((q = qs[i]) != null)
3881                      count += (long)q.nsteals & 0xffffffffL;
3882             }
3883         }
3884         return count;
3885     }
3886 
3887     /**
3888      * Returns an estimate of the total number of tasks currently held
3889      * in queues by worker threads (but not including tasks submitted
3890      * to the pool that have not begun executing). This value is only
3891      * an approximation, obtained by iterating across all threads in
3892      * the pool. This method may be useful for tuning task
3893      * granularities.The returned count does not include scheduled
3894      * tasks that are not yet ready to execute, which are reported
3895      * separately by method {@link getDelayedTaskCount}.
3896      *
3897      * @return the number of queued tasks
3898      * @see ForkJoinWorkerThread#getQueuedTaskCount()
3899      */
3900     public long getQueuedTaskCount() {
3901         WorkQueue[] qs; WorkQueue q;
3902         long count = 0;
3903         if ((runState & TERMINATED) == 0L && (qs = queues) != null) {
3904             for (int i = 1; i < qs.length; i += 2) {
3905                 if ((q = qs[i]) != null)
3906                     count += q.queueSize();
3907             }
3908         }
3909         return count;
3910     }
3911 
3912     /**
3913      * Returns an estimate of the number of tasks submitted to this
3914      * pool that have not yet begun executing.  This method may take
3915      * time proportional to the number of submissions.
3916      *
3917      * @return the number of queued submissions
3918      */
3919     public int getQueuedSubmissionCount() {
3920         WorkQueue[] qs; WorkQueue q;
3921         int count = 0;
3922         if ((runState & TERMINATED) == 0L && (qs = queues) != null) {
3923             for (int i = 0; i < qs.length; i += 2) {
3924                 if ((q = qs[i]) != null)
3925                     count += q.queueSize();
3926             }
3927         }
3928         return count;
3929     }
3930 
3931     /**
3932      * Returns an estimate of the number of delayed (including
3933      * periodic) tasks scheduled in this pool that are not yet ready
3934      * to submit for execution. The returned value is inaccurate while
3935      * delayed tasks are being processed.
3936      *
3937      * @return an estimate of the number of delayed tasks
3938      * @since 25
3939      */
3940     public long getDelayedTaskCount() {
3941         DelayScheduler ds;
3942         return ((ds = delayScheduler) == null ? 0 : ds.lastStableSize());
3943     }
3944 
3945     /**
3946      * Returns {@code true} if there are any tasks submitted to this
3947      * pool that have not yet begun executing.
3948      *
3949      * @return {@code true} if there are any queued submissions
3950      */
3951     public boolean hasQueuedSubmissions() {
3952         WorkQueue[] qs; WorkQueue q;
3953         if ((runState & STOP) == 0L && (qs = queues) != null) {
3954             for (int i = 0; i < qs.length; i += 2) {
3955                 if ((q = qs[i]) != null && q.queueSize() > 0)
3956                     return true;
3957             }
3958         }
3959         return false;
3960     }
3961 
3962     /**
3963      * Removes and returns the next unexecuted submission if one is
3964      * available.  This method may be useful in extensions to this
3965      * class that re-assign work in systems with multiple pools.
3966      *
3967      * @return the next submission, or {@code null} if none
3968      */
3969     protected ForkJoinTask<?> pollSubmission() {
3970         return pollScan(true);
3971     }
3972 
3973     /**
3974      * Removes all available unexecuted submitted and forked tasks
3975      * from scheduling queues and adds them to the given collection,
3976      * without altering their execution status. These may include
3977      * artificially generated or wrapped tasks. This method is
3978      * designed to be invoked only when the pool is known to be
3979      * quiescent. Invocations at other times may not remove all
3980      * tasks. A failure encountered while attempting to add elements
3981      * to collection {@code c} may result in elements being in
3982      * neither, either or both collections when the associated
3983      * exception is thrown.  The behavior of this operation is
3984      * undefined if the specified collection is modified while the
3985      * operation is in progress.
3986      *
3987      * @param c the collection to transfer elements into
3988      * @return the number of elements transferred
3989      */
3990     protected int drainTasksTo(Collection<? super ForkJoinTask<?>> c) {
3991         int count = 0;
3992         for (ForkJoinTask<?> t; (t = pollScan(false)) != null; ) {
3993             c.add(t);
3994             ++count;
3995         }
3996         return count;
3997     }
3998 
3999     /**
4000      * Returns a string identifying this pool, as well as its state,
4001      * including indications of run state, parallelism level, and
4002      * worker and task counts.
4003      *
4004      * @return a string identifying this pool, as well as its state
4005      */
4006     public String toString() {
4007         // Use a single pass through queues to collect counts
4008         DelayScheduler ds;
4009         long e = runState;
4010         long st = stealCount;
4011         long qt = 0L, ss = 0L; int rc = 0;
4012         WorkQueue[] qs; WorkQueue q;
4013         if ((qs = queues) != null) {
4014             for (int i = 0; i < qs.length; ++i) {
4015                 if ((q = qs[i]) != null) {
4016                     int size = q.queueSize();
4017                     if ((i & 1) == 0)
4018                         ss += size;
4019                     else {
4020                         qt += size;
4021                         st += (long)q.nsteals & 0xffffffffL;
4022                         if (q.isApparentlyUnblocked())
4023                             ++rc;
4024                     }
4025                 }
4026             }
4027         }
4028         String delayed = ((ds = delayScheduler) == null ? "" :
4029                           ", delayed = " + ds.lastStableSize());
4030         int pc = parallelism;
4031         long c = ctl;
4032         int tc = (short)(c >>> TC_SHIFT);
4033         int ac = (short)(c >>> RC_SHIFT);
4034         if (ac < 0) // ignore transient negative
4035             ac = 0;
4036         String level = ((e & TERMINATED) != 0L ? "Terminated" :
4037                         (e & STOP)       != 0L ? "Terminating" :
4038                         (e & SHUTDOWN)   != 0L ? "Shutting down" :
4039                         "Running");
4040         return super.toString() +
4041             "[" + level +
4042             ", parallelism = " + pc +
4043             ", size = " + tc +
4044             ", active = " + ac +
4045             ", running = " + rc +
4046             ", steals = " + st +
4047             ", tasks = " + qt +
4048             ", submissions = " + ss +
4049             delayed +
4050             "]";
4051     }
4052 
4053     /**
4054      * Possibly initiates an orderly shutdown in which previously
4055      * submitted tasks are executed, but no new tasks will be
4056      * accepted. Invocation has no effect on execution state if this
4057      * is the {@link #commonPool()}, and no additional effect if
4058      * already shut down.  Tasks that are in the process of being
4059      * submitted concurrently during the course of this method may or
4060      * may not be rejected.
4061      */
4062     public void shutdown() {
4063         if (workerNamePrefix != null) // not common pool
4064             tryTerminate(false, true);
4065     }
4066 
4067     /**
4068      * Possibly attempts to cancel and/or stop all tasks, and reject
4069      * all subsequently submitted tasks.  Invocation has no effect on
4070      * execution state if this is the {@link #commonPool()}, and no
4071      * additional effect if already shut down. Otherwise, tasks that
4072      * are in the process of being submitted or executed concurrently
4073      * during the course of this method may or may not be
4074      * rejected. This method cancels both existing and unexecuted
4075      * tasks, in order to permit termination in the presence of task
4076      * dependencies. So the method always returns an empty list
4077      * (unlike the case for some other Executors).
4078      *
4079      * @return an empty list
4080      */
4081     public List<Runnable> shutdownNow() {
4082         if (workerNamePrefix != null) // not common pool
4083             tryTerminate(true, true);
4084         return Collections.emptyList();
4085     }
4086 
4087     /**
4088      * Returns {@code true} if all tasks have completed following shut down.
4089      *
4090      * @return {@code true} if all tasks have completed following shut down
4091      */
4092     public boolean isTerminated() {
4093         return (tryTerminate(false, false) & TERMINATED) != 0;
4094     }
4095 
4096     /**
4097      * Returns {@code true} if the process of termination has
4098      * commenced but not yet completed.  This method may be useful for
4099      * debugging. A return of {@code true} reported a sufficient
4100      * period after shutdown may indicate that submitted tasks have
4101      * ignored or suppressed interruption, or are waiting for I/O,
4102      * causing this executor not to properly terminate. (See the
4103      * advisory notes for class {@link ForkJoinTask} stating that
4104      * tasks should not normally entail blocking operations.  But if
4105      * they do, they must abort them on interrupt.)
4106      *
4107      * @return {@code true} if terminating but not yet terminated
4108      */
4109     public boolean isTerminating() {
4110         return (tryTerminate(false, false) & (STOP | TERMINATED)) == STOP;
4111     }
4112 
4113     /**
4114      * Returns {@code true} if this pool has been shut down.
4115      *
4116      * @return {@code true} if this pool has been shut down
4117      */
4118     public boolean isShutdown() {
4119         return (runState & SHUTDOWN) != 0L;
4120     }
4121 
4122     /**
4123      * Blocks until all tasks have completed execution after a
4124      * shutdown request, or the timeout occurs, or the current thread
4125      * is interrupted, whichever happens first. Because the {@link
4126      * #commonPool()} never terminates until program shutdown, when
4127      * applied to the common pool, this method is equivalent to {@link
4128      * #awaitQuiescence(long, TimeUnit)} but always returns {@code false}.
4129      *
4130      * @param timeout the maximum time to wait
4131      * @param unit the time unit of the timeout argument
4132      * @return {@code true} if this executor terminated and
4133      *         {@code false} if the timeout elapsed before termination
4134      * @throws InterruptedException if interrupted while waiting
4135      */
4136     public boolean awaitTermination(long timeout, TimeUnit unit)
4137         throws InterruptedException {
4138         long nanos = unit.toNanos(timeout);
4139         CountDownLatch done;
4140         if (workerNamePrefix == null) {    // is common pool
4141             if (helpQuiescePool(this, nanos, true) < 0)
4142                 throw new InterruptedException();
4143             return false;
4144         }
4145         else if ((tryTerminate(false, false) & TERMINATED) != 0 ||
4146                  (done = terminationSignal()) == null ||
4147                  (runState & TERMINATED) != 0L)
4148             return true;
4149         else
4150             return done.await(nanos, TimeUnit.NANOSECONDS);
4151     }
4152 
4153     /**
4154      * If called by a ForkJoinTask operating in this pool, equivalent
4155      * in effect to {@link ForkJoinTask#helpQuiesce}. Otherwise,
4156      * waits and/or attempts to assist performing tasks until this
4157      * pool {@link #isQuiescent} or the indicated timeout elapses.
4158      *
4159      * @param timeout the maximum time to wait
4160      * @param unit the time unit of the timeout argument
4161      * @return {@code true} if quiescent; {@code false} if the
4162      * timeout elapsed.
4163      */
4164     public boolean awaitQuiescence(long timeout, TimeUnit unit) {
4165         return (helpQuiescePool(this, unit.toNanos(timeout), false) > 0);
4166     }
4167 
4168     /**
4169      * Unless this is the {@link #commonPool()}, initiates an orderly
4170      * shutdown in which previously submitted tasks are executed, but
4171      * no new tasks will be accepted, and waits until all tasks have
4172      * completed execution and the executor has terminated.
4173      *
4174      * <p> If already terminated, or this is the {@link
4175      * #commonPool()}, this method has no effect on execution, and
4176      * does not wait. Otherwise, if interrupted while waiting, this
4177      * method stops all executing tasks as if by invoking {@link
4178      * #shutdownNow()}. It then continues to wait until all actively
4179      * executing tasks have completed. Tasks that were awaiting
4180      * execution are not executed. The interrupted status will be
4181      * re-asserted before this method returns.
4182      *
4183      * @since 19
4184      */
4185     @Override
4186     public void close() {
4187         if (workerNamePrefix != null) {
4188             CountDownLatch done = null;
4189             boolean interrupted = false;
4190             while ((tryTerminate(interrupted, true) & TERMINATED) == 0) {
4191                 if (done == null)
4192                     done = terminationSignal();
4193                 else {
4194                     try {
4195                         done.await();
4196                         break;
4197                     } catch (InterruptedException ex) {
4198                         interrupted = true;
4199                     }
4200                 }
4201             }
4202             if (interrupted)
4203                 Thread.currentThread().interrupt();
4204         }
4205     }
4206 
4207     /**
4208      * Interface for extending managed parallelism for tasks running
4209      * in {@link ForkJoinPool}s.
4210      *
4211      * <p>A {@code ManagedBlocker} provides two methods.  Method
4212      * {@link #isReleasable} must return {@code true} if blocking is
4213      * not necessary. Method {@link #block} blocks the current thread
4214      * if necessary (perhaps internally invoking {@code isReleasable}
4215      * before actually blocking). These actions are performed by any
4216      * thread invoking {@link
4217      * ForkJoinPool#managedBlock(ManagedBlocker)}.  The unusual
4218      * methods in this API accommodate synchronizers that may, but
4219      * don't usually, block for long periods. Similarly, they allow
4220      * more efficient internal handling of cases in which additional
4221      * workers may be, but usually are not, needed to ensure
4222      * sufficient parallelism.  Toward this end, implementations of
4223      * method {@code isReleasable} must be amenable to repeated
4224      * invocation. Neither method is invoked after a prior invocation
4225      * of {@code isReleasable} or {@code block} returns {@code true}.
4226      *
4227      * <p>For example, here is a ManagedBlocker based on a
4228      * ReentrantLock:
4229      * <pre> {@code
4230      * class ManagedLocker implements ManagedBlocker {
4231      *   final ReentrantLock lock;
4232      *   boolean hasLock = false;
4233      *   ManagedLocker(ReentrantLock lock) { this.lock = lock; }
4234      *   public boolean block() {
4235      *     if (!hasLock)
4236      *       lock.lock();
4237      *     return true;
4238      *   }
4239      *   public boolean isReleasable() {
4240      *     return hasLock || (hasLock = lock.tryLock());
4241      *   }
4242      * }}</pre>
4243      *
4244      * <p>Here is a class that possibly blocks waiting for an
4245      * item on a given queue:
4246      * <pre> {@code
4247      * class QueueTaker<E> implements ManagedBlocker {
4248      *   final BlockingQueue<E> queue;
4249      *   volatile E item = null;
4250      *   QueueTaker(BlockingQueue<E> q) { this.queue = q; }
4251      *   public boolean block() throws InterruptedException {
4252      *     if (item == null)
4253      *       item = queue.take();
4254      *     return true;
4255      *   }
4256      *   public boolean isReleasable() {
4257      *     return item != null || (item = queue.poll()) != null;
4258      *   }
4259      *   public E getItem() { // call after pool.managedBlock completes
4260      *     return item;
4261      *   }
4262      * }}</pre>
4263      */
4264     public static interface ManagedBlocker {
4265         /**
4266          * Possibly blocks the current thread, for example waiting for
4267          * a lock or condition.
4268          *
4269          * @return {@code true} if no additional blocking is necessary
4270          * (i.e., if isReleasable would return true)
4271          * @throws InterruptedException if interrupted while waiting
4272          * (the method is not required to do so, but is allowed to)
4273          */
4274         boolean block() throws InterruptedException;
4275 
4276         /**
4277          * Returns {@code true} if blocking is unnecessary.
4278          * @return {@code true} if blocking is unnecessary
4279          */
4280         boolean isReleasable();
4281     }
4282 
4283     /**
4284      * Runs the given possibly blocking task.  When {@linkplain
4285      * ForkJoinTask#inForkJoinPool() running in a ForkJoinPool}, this
4286      * method possibly arranges for a spare thread to be activated if
4287      * necessary to ensure sufficient parallelism while the current
4288      * thread is blocked in {@link ManagedBlocker#block blocker.block()}.
4289      *
4290      * <p>This method repeatedly calls {@code blocker.isReleasable()} and
4291      * {@code blocker.block()} until either method returns {@code true}.
4292      * Every call to {@code blocker.block()} is preceded by a call to
4293      * {@code blocker.isReleasable()} that returned {@code false}.
4294      *
4295      * <p>If not running in a ForkJoinPool, this method is
4296      * behaviorally equivalent to
4297      * <pre> {@code
4298      * while (!blocker.isReleasable())
4299      *   if (blocker.block())
4300      *     break;}</pre>
4301      *
4302      * If running in a ForkJoinPool, the pool may first be expanded to
4303      * ensure sufficient parallelism available during the call to
4304      * {@code blocker.block()}.
4305      *
4306      * @param blocker the blocker task
4307      * @throws InterruptedException if {@code blocker.block()} did so
4308      */
4309     public static void managedBlock(ManagedBlocker blocker)
4310         throws InterruptedException {
4311         Thread t; ForkJoinPool p;
4312         if ((t = Thread.currentThread()) instanceof ForkJoinWorkerThread &&
4313             (p = ((ForkJoinWorkerThread)t).pool) != null)
4314             p.compensatedBlock(blocker);
4315         else
4316             unmanagedBlock(blocker);
4317     }
4318 
4319     /** ManagedBlock for ForkJoinWorkerThreads */
4320     private void compensatedBlock(ManagedBlocker blocker)
4321         throws InterruptedException {
4322         Objects.requireNonNull(blocker);
4323         for (;;) {
4324             int comp; boolean done;
4325             long c = ctl;
4326             if (blocker.isReleasable())
4327                 break;
4328             if ((runState & STOP) != 0L)
4329                 throw new InterruptedException();
4330             if ((comp = tryCompensate(c)) >= 0) {
4331                 try {
4332                     done = blocker.block();
4333                 } finally {
4334                     if (comp > 0)
4335                         getAndAddCtl(RC_UNIT);
4336                 }
4337                 if (done)
4338                     break;
4339             }
4340         }
4341     }
4342 
4343     /**
4344      * Invokes tryCompensate to create or re-activate a spare thread to
4345      * compensate for a thread that performs a blocking operation. When the
4346      * blocking operation is done then endCompensatedBlock must be invoked
4347      * with the value returned by this method to re-adjust the parallelism.
4348      * @return value to use in endCompensatedBlock
4349      */
4350     final long beginCompensatedBlock() {
4351         int c;
4352         do {} while ((c = tryCompensate(ctl)) < 0);
4353         return (c == 0) ? 0L : RC_UNIT;
4354     }
4355 
4356     /**
4357      * Re-adjusts parallelism after a blocking operation completes.
4358      * @param post value from beginCompensatedBlock
4359      */
4360     void endCompensatedBlock(long post) {
4361         if (post > 0L) {
4362             getAndAddCtl(post);
4363         }
4364     }
4365 
4366     /** ManagedBlock for external threads */
4367     private static void unmanagedBlock(ManagedBlocker blocker)
4368         throws InterruptedException {
4369         Objects.requireNonNull(blocker);
4370         do {} while (!blocker.isReleasable() && !blocker.block());
4371     }
4372 
4373     @Override
4374     protected <T> RunnableFuture<T> newTaskFor(Runnable runnable, T value) {
4375         Objects.requireNonNull(runnable);
4376         return (Thread.currentThread() instanceof ForkJoinWorkerThread) ?
4377             new ForkJoinTask.AdaptedRunnable<T>(runnable, value) :
4378             new ForkJoinTask.AdaptedInterruptibleRunnable<T>(runnable, value);
4379     }
4380 
4381     @Override
4382     protected <T> RunnableFuture<T> newTaskFor(Callable<T> callable) {
4383         Objects.requireNonNull(callable);
4384         return (Thread.currentThread() instanceof ForkJoinWorkerThread) ?
4385             new ForkJoinTask.AdaptedCallable<T>(callable) :
4386             new ForkJoinTask.AdaptedInterruptibleCallable<T>(callable);
4387     }
4388 
4389     static {
4390         U = Unsafe.getUnsafe();
4391         Class<ForkJoinPool> klass = ForkJoinPool.class;
4392         try {
4393             Field poolIdsField = klass.getDeclaredField("poolIds");
4394             POOLIDS_BASE = U.staticFieldBase(poolIdsField);
4395             POOLIDS = U.staticFieldOffset(poolIdsField);
4396         } catch (NoSuchFieldException e) {
4397             throw new ExceptionInInitializerError(e);
4398         }
4399         CTL = U.objectFieldOffset(klass, "ctl");
4400         RUNSTATE = U.objectFieldOffset(klass, "runState");
4401         PARALLELISM =  U.objectFieldOffset(klass, "parallelism");
4402         THREADIDS = U.objectFieldOffset(klass, "threadIds");
4403         TERMINATION = U.objectFieldOffset(klass, "termination");
4404         Class<ForkJoinTask[]> aklass = ForkJoinTask[].class;
4405         ABASE = U.arrayBaseOffset(aklass);
4406         int scale = U.arrayIndexScale(aklass);
4407         ASHIFT = 31 - Integer.numberOfLeadingZeros(scale);
4408         if ((scale & (scale - 1)) != 0)
4409             throw new Error("array index scale not a power of two");
4410 
4411         Class<?> dep = LockSupport.class; // ensure loaded
4412         // allow access to non-public methods
4413         JLA = SharedSecrets.getJavaLangAccess();
4414         SharedSecrets.setJavaUtilConcurrentFJPAccess(
4415             new JavaUtilConcurrentFJPAccess() {
4416                 @Override
4417                 public long beginCompensatedBlock(ForkJoinPool pool) {
4418                     return pool.beginCompensatedBlock();
4419                 }
4420                 public void endCompensatedBlock(ForkJoinPool pool, long post) {
4421                     pool.endCompensatedBlock(post);
4422                 }
4423             });
4424         defaultForkJoinWorkerThreadFactory =
4425             new DefaultForkJoinWorkerThreadFactory();
4426         common = new ForkJoinPool((byte)0);
4427     }
4428 }