ViewVC Help
View File | Revision Log | Show Annotations | Download File | Root Listing
root/jsr166/jsr166/src/jsr166y/ForkJoinPool.java
(Generate patch)

Comparing jsr166/src/jsr166y/ForkJoinPool.java (file contents):
Revision 1.59 by dl, Fri Jul 23 14:09:17 2010 UTC vs.
Revision 1.82 by dl, Sun Oct 10 11:56:11 2010 UTC

# Line 6 | Line 6
6  
7   package jsr166y;
8  
9 import java.util.concurrent.*;
10
9   import java.util.ArrayList;
10   import java.util.Arrays;
11   import java.util.Collection;
12   import java.util.Collections;
13   import java.util.List;
14 + import java.util.concurrent.AbstractExecutorService;
15 + import java.util.concurrent.Callable;
16 + import java.util.concurrent.ExecutorService;
17 + import java.util.concurrent.Future;
18 + import java.util.concurrent.RejectedExecutionException;
19 + import java.util.concurrent.RunnableFuture;
20 + import java.util.concurrent.TimeUnit;
21 + import java.util.concurrent.TimeoutException;
22 + import java.util.concurrent.atomic.AtomicInteger;
23   import java.util.concurrent.locks.LockSupport;
24   import java.util.concurrent.locks.ReentrantLock;
18 import java.util.concurrent.atomic.AtomicInteger;
19 import java.util.concurrent.CountDownLatch;
25  
26   /**
27   * An {@link ExecutorService} for running {@link ForkJoinTask}s.
# Line 52 | Line 57 | import java.util.concurrent.CountDownLat
57   * convenient form for informal monitoring.
58   *
59   * <p> As is the case with other ExecutorServices, there are three
60 < * main task execution methods summarized in the follwoing
60 > * main task execution methods summarized in the following
61   * table. These are designed to be used by clients not already engaged
62   * in fork/join computations in the current pool.  The main forms of
63   * these methods accept instances of {@code ForkJoinTask}, but
# Line 69 | Line 74 | import java.util.concurrent.CountDownLat
74   *    <td ALIGN=CENTER> <b>Call from within fork/join computations</b></td>
75   *  </tr>
76   *  <tr>
77 < *    <td> <b>Arange async execution</td>
77 > *    <td> <b>Arrange async execution</td>
78   *    <td> {@link #execute(ForkJoinTask)}</td>
79   *    <td> {@link ForkJoinTask#fork}</td>
80   *  </tr>
# Line 110 | Line 115 | import java.util.concurrent.CountDownLat
115   *
116   * <p>This implementation rejects submitted tasks (that is, by throwing
117   * {@link RejectedExecutionException}) only when the pool is shut down
118 < * or internal resources have been exhuasted.
118 > * or internal resources have been exhausted.
119   *
120   * @since 1.7
121   * @author Doug Lea
# Line 138 | Line 143 | public class ForkJoinPool extends Abstra
143       * cache pollution effects.)
144       *
145       * Beyond work-stealing support and essential bookkeeping, the
146 <     * main responsibility of this framework is to arrange tactics for
147 <     * when one worker is waiting to join a task stolen (or always
148 <     * held by) another.  Becauae we are multiplexing many tasks on to
149 <     * a pool of workers, we can't just let them block (as in
150 <     * Thread.join).  We also cannot just reassign the joiner's
151 <     * run-time stack with another and replace it later, which would
152 <     * be a form of "continuation", that even if possible is not
153 <     * necessarily a good idea. Given that the creation costs of most
154 <     * threads on most systems mainly surrounds setting up runtime
155 <     * stacks, thread creation and switching is usually not much more
156 <     * expensive than stack creation and switching, and is more
157 <     * flexible). Instead we combine two tactics:
146 >     * main responsibility of this framework is to take actions when
147 >     * one worker is waiting to join a task stolen (or always held by)
148 >     * another.  Because we are multiplexing many tasks on to a pool
149 >     * of workers, we can't just let them block (as in Thread.join).
150 >     * We also cannot just reassign the joiner's run-time stack with
151 >     * another and replace it later, which would be a form of
152 >     * "continuation", that even if possible is not necessarily a good
153 >     * idea. Given that the creation costs of most threads on most
154 >     * systems mainly surrounds setting up runtime stacks, thread
155 >     * creation and switching is usually not much more expensive than
156 >     * stack creation and switching, and is more flexible). Instead we
157 >     * combine two tactics:
158       *
159 <     *   1. Arranging for the joiner to execute some task that it
159 >     *   Helping: Arranging for the joiner to execute some task that it
160       *      would be running if the steal had not occurred.  Method
161       *      ForkJoinWorkerThread.helpJoinTask tracks joining->stealing
162       *      links to try to find such a task.
163       *
164 <     *   2. Unless there are already enough live threads, creating or
165 <     *      or re-activating a spare thread to compensate for the
166 <     *      (blocked) joiner until it unblocks.  Spares then suspend
167 <     *      at their next opportunity or eventually die if unused for
168 <     *      too long.  See below and the internal documentation
169 <     *      for tryAwaitJoin for more details about compensation
170 <     *      rules.
171 <     *
172 <     * Because the determining existence of conservatively safe
173 <     * helping targets, the availability of already-created spares,
174 <     * and the apparent need to create new spares are all racy and
175 <     * require heuristic guidance, joins (in
176 <     * ForkJoinWorkerThread.joinTask) interleave these options until
177 <     * successful.  Creating a new spare always succeeds, but also
178 <     * increases application footprint, so we try to avoid it, within
179 <     * reason.
164 >     *   Compensating: Unless there are already enough live threads,
165 >     *      method helpMaintainParallelism() may create or
166 >     *      re-activate a spare thread to compensate for blocked
167 >     *      joiners until they unblock.
168 >     *
169 >     * It is impossible to keep exactly the target (parallelism)
170 >     * number of threads running at any given time.  Determining
171 >     * existence of conservatively safe helping targets, the
172 >     * availability of already-created spares, and the apparent need
173 >     * to create new spares are all racy and require heuristic
174 >     * guidance, so we rely on multiple retries of each.  Compensation
175 >     * occurs in slow-motion. It is triggered only upon timeouts of
176 >     * Object.wait used for joins. This reduces poor decisions that
177 >     * would otherwise be made when threads are waiting for others
178 >     * that are stalled because of unrelated activities such as
179 >     * garbage collection.
180       *
181 <     * The ManagedBlocker extension API can't use option (1) so uses a
182 <     * special version of (2) in method awaitBlocker.
181 >     * The ManagedBlocker extension API can't use helping so relies
182 >     * only on compensation in method awaitBlocker.
183       *
184       * The main throughput advantages of work-stealing stem from
185       * decentralized control -- workers mostly steal tasks from each
# Line 207 | Line 212 | public class ForkJoinPool extends Abstra
212       * blocked workers. However, all other support code is set up to
213       * work with other policies.
214       *
215 +     * To ensure that we do not hold on to worker references that
216 +     * would prevent GC, ALL accesses to workers are via indices into
217 +     * the workers array (which is one source of some of the unusual
218 +     * code constructions here). In essence, the workers array serves
219 +     * as a WeakReference mechanism. Thus for example the event queue
220 +     * stores worker indices, not worker references. Access to the
221 +     * workers in associated methods (for example releaseEventWaiters)
222 +     * must both index-check and null-check the IDs. All such accesses
223 +     * ignore bad IDs by returning out early from what they are doing,
224 +     * since this can only be associated with shutdown, in which case
225 +     * it is OK to give up. On termination, we just clobber these
226 +     * data structures without trying to use them.
227 +     *
228       * 2. Bookkeeping for dynamically adding and removing workers. We
229       * aim to approximately maintain the given level of parallelism.
230       * When some workers are known to be blocked (on joins or via
231       * ManagedBlocker), we may create or resume others to take their
232       * place until they unblock (see below). Implementing this
233       * requires counts of the number of "running" threads (i.e., those
234 <     * that are neither blocked nor artifically suspended) as well as
234 >     * that are neither blocked nor artificially suspended) as well as
235       * the total number.  These two values are packed into one field,
236       * "workerCounts" because we need accurate snapshots when deciding
237       * to create, resume or suspend.  Note however that the
238 <     * correspondance of these counts to reality is not guaranteed. In
238 >     * correspondence of these counts to reality is not guaranteed. In
239       * particular updates for unblocked threads may lag until they
240       * actually wake up.
241       *
# Line 248 | Line 266 | public class ForkJoinPool extends Abstra
266       * workers that previously could not find a task to now find one:
267       * Submission of a new task to the pool, or another worker pushing
268       * a task onto a previously empty queue.  (We also use this
269 <     * mechanism for termination and reconfiguration actions that
269 >     * mechanism for configuration and termination actions that
270       * require wakeups of idle workers).  Each worker maintains its
271       * last known event count, and blocks when a scan for work did not
272       * find a task AND its lastEventCount matches the current
# Line 259 | Line 277 | public class ForkJoinPool extends Abstra
277       * a record (field nextEventWaiter) for the next waiting worker.
278       * In addition to allowing simpler decisions about need for
279       * wakeup, the event count bits in eventWaiters serve the role of
280 <     * tags to avoid ABA errors in Treiber stacks.  To reduce delays
281 <     * in task diffusion, workers not otherwise occupied may invoke
282 <     * method releaseWaiters, that removes and signals (unparks)
283 <     * workers not waiting on current count. To minimize task
284 <     * production stalls associate with signalling, any worker pushing
285 <     * a task on an empty queue invokes the weaker method signalWork,
268 <     * that only releases idle workers until it detects interference
269 <     * by other threads trying to release, and lets them take
270 <     * over. The net effect is a tree-like diffusion of signals, where
271 <     * released threads (and possibly others) help with unparks.  To
272 <     * further reduce contention effects a bit, failed CASes to
273 <     * increment field eventCount are tolerated without retries.
280 >     * tags to avoid ABA errors in Treiber stacks. Upon any wakeup,
281 >     * released threads also try to release at most two others.  The
282 >     * net effect is a tree-like diffusion of signals, where released
283 >     * threads (and possibly others) help with unparks.  To further
284 >     * reduce contention effects a bit, failed CASes to increment
285 >     * field eventCount are tolerated without retries in signalWork.
286       * Conceptually they are merged into the same event, which is OK
287       * when their only purpose is to enable workers to scan for work.
288       *
289 <     * 5. Managing suspension of extra workers. When a worker is about
290 <     * to block waiting for a join (or via ManagedBlockers), we may
291 <     * create a new thread to maintain parallelism level, or at least
292 <     * avoid starvation. Usually, extra threads are needed for only
293 <     * very short periods, yet join dependencies are such that we
294 <     * sometimes need them in bursts. Rather than create new threads
295 <     * each time this happens, we suspend no-longer-needed extra ones
296 <     * as "spares". For most purposes, we don't distinguish "extra"
297 <     * spare threads from normal "core" threads: On each call to
298 <     * preStep (the only point at which we can do this) a worker
299 <     * checks to see if there are now too many running workers, and if
300 <     * so, suspends itself.  Methods tryAwaitJoin and awaitBlocker
301 <     * look for suspended threads to resume before considering
302 <     * creating a new replacement. We don't need a special data
303 <     * structure to maintain spares; simply scanning the workers array
304 <     * looking for worker.isSuspended() is fine because the calling
305 <     * thread is otherwise not doing anything useful anyway; we are at
306 <     * least as happy if after locating a spare, the caller doesn't
307 <     * actually block because the join is ready before we try to
308 <     * adjust and compensate.  Note that this is intrinsically racy.
309 <     * One thread may become a spare at about the same time as another
310 <     * is needlessly being created. We counteract this and related
311 <     * slop in part by requiring resumed spares to immediately recheck
312 <     * (in preStep) to see whether they they should re-suspend. The
313 <     * only effective difference between "extra" and "core" threads is
314 <     * that we allow the "extra" ones to time out and die if they are
315 <     * not resumed within a keep-alive interval of a few seconds. This
316 <     * is implemented mainly within ForkJoinWorkerThread, but requires
317 <     * some coordination (isTrimmed() -- meaning killed while
318 <     * suspended) to correctly maintain pool counts.
319 <     *
320 <     * 6. Deciding when to create new workers. The main dynamic
321 <     * control in this class is deciding when to create extra threads,
322 <     * in methods awaitJoin and awaitBlocker. We always need to create
323 <     * one when the number of running threads would become zero and
324 <     * all workers are busy. However, this is not easy to detect
325 <     * reliably in the presence of transients so we use retries and
326 <     * allow slack (in tryAwaitJoin) to reduce false alarms.  These
327 <     * effectively reduce churn at the price of systematically
328 <     * undershooting target parallelism when many threads are blocked.
329 <     * However, biasing toward undeshooting partially compensates for
330 <     * the above mechanics to suspend extra threads, that normally
331 <     * lead to overshoot because we can only suspend workers
332 <     * in-between top-level actions. It also better copes with the
333 <     * fact that some of the methods in this class tend to never
334 <     * become compiled (but are interpreted), so some components of
335 <     * the entire set of controls might execute many times faster than
336 <     * others. And similarly for cases where the apparent lack of work
337 <     * is just due to GC stalls and other transient system activity.
289 >     * 5. Managing suspension of extra workers. When a worker notices
290 >     * (usually upon timeout of a wait()) that there are too few
291 >     * running threads, we may create a new thread to maintain
292 >     * parallelism level, or at least avoid starvation. Usually, extra
293 >     * threads are needed for only very short periods, yet join
294 >     * dependencies are such that we sometimes need them in
295 >     * bursts. Rather than create new threads each time this happens,
296 >     * we suspend no-longer-needed extra ones as "spares". For most
297 >     * purposes, we don't distinguish "extra" spare threads from
298 >     * normal "core" threads: On each call to preStep (the only point
299 >     * at which we can do this) a worker checks to see if there are
300 >     * now too many running workers, and if so, suspends itself.
301 >     * Method helpMaintainParallelism looks for suspended threads to
302 >     * resume before considering creating a new replacement. The
303 >     * spares themselves are encoded on another variant of a Treiber
304 >     * Stack, headed at field "spareWaiters".  Note that the use of
305 >     * spares is intrinsically racy.  One thread may become a spare at
306 >     * about the same time as another is needlessly being created. We
307 >     * counteract this and related slop in part by requiring resumed
308 >     * spares to immediately recheck (in preStep) to see whether they
309 >     * should re-suspend.
310 >     *
311 >     * 6. Killing off unneeded workers. A timeout mechanism is used to
312 >     * shed unused workers: The oldest (first) event queue waiter uses
313 >     * a timed rather than hard wait. When this wait times out without
314 >     * a normal wakeup, it tries to shutdown any one (for convenience
315 >     * the newest) other spare or event waiter via
316 >     * tryShutdownUnusedWorker. This eventually reduces the number of
317 >     * worker threads to a minimum of one after a long enough period
318 >     * without use.
319 >     *
320 >     * 7. Deciding when to create new workers. The main dynamic
321 >     * control in this class is deciding when to create extra threads
322 >     * in method helpMaintainParallelism. We would like to keep
323 >     * exactly #parallelism threads running, which is an impossible
324 >     * task. We always need to create one when the number of running
325 >     * threads would become zero and all workers are busy. Beyond
326 >     * this, we must rely on heuristics that work well in the
327 >     * presence of transient phenomena such as GC stalls, dynamic
328 >     * compilation, and wake-up lags. These transients are extremely
329 >     * common -- we are normally trying to fully saturate the CPUs on
330 >     * a machine, so almost any activity other than running tasks
331 >     * impedes accuracy. Our main defense is to allow parallelism to
332 >     * lapse for a while during joins, and use a timeout to see if,
333 >     * after the resulting settling, there is still a need for
334 >     * additional workers.  This also better copes with the fact that
335 >     * some of the methods in this class tend to never become compiled
336 >     * (but are interpreted), so some components of the entire set of
337 >     * controls might execute 100 times faster than others. And
338 >     * similarly for cases where the apparent lack of work is just due
339 >     * to GC stalls and other transient system activity.
340       *
341       * Beware that there is a lot of representation-level coupling
342       * among classes ForkJoinPool, ForkJoinWorkerThread, and
# Line 335 | Line 349 | public class ForkJoinPool extends Abstra
349       *
350       * Style notes: There are lots of inline assignments (of form
351       * "while ((local = field) != 0)") which are usually the simplest
352 <     * way to ensure read orderings. Also several occurrences of the
353 <     * unusual "do {} while(!cas...)" which is the simplest way to
354 <     * force an update of a CAS'ed variable. There are also other
355 <     * coding oddities that help some methods perform reasonably even
356 <     * when interpreted (not compiled), at the expense of messiness.
352 >     * way to ensure the required read orderings (which are sometimes
353 >     * critical). Also several occurrences of the unusual "do {}
354 >     * while (!cas...)" which is the simplest way to force an update of
355 >     * a CAS'ed variable. There are also other coding oddities that
356 >     * help some methods perform reasonably even when interpreted (not
357 >     * compiled), at the expense of some messy constructions that
358 >     * reduce byte code counts.
359       *
360       * The order of declarations in this file is: (1) statics (2)
361       * fields (along with constants used when unpacking some of them)
# Line 407 | Line 423 | public class ForkJoinPool extends Abstra
423          new AtomicInteger();
424  
425      /**
426 <     * Absolute bound for parallelism level. Twice this number must
427 <     * fit into a 16bit field to enable word-packing for some counts.
426 >     * The time to block in a join (see awaitJoin) before checking if
427 >     * a new worker should be (re)started to maintain parallelism
428 >     * level. The value should be short enough to maintain global
429 >     * responsiveness and progress but long enough to avoid
430 >     * counterproductive firings during GC stalls or unrelated system
431 >     * activity, and to not bog down systems with continual re-firings
432 >     * on GCs or legitimately long waits.
433 >     */
434 >    private static final long JOIN_TIMEOUT_MILLIS = 250L; // 4 per second
435 >
436 >    /**
437 >     * The wakeup interval (in nanoseconds) for the oldest worker
438 >     * waiting for an event to invoke tryShutdownUnusedWorker to
439 >     * shrink the number of workers.  The exact value does not matter
440 >     * too much. It must be short enough to release resources during
441 >     * sustained periods of idleness, but not so short that threads
442 >     * are continually re-created.
443 >     */
444 >    private static final long SHRINK_RATE_NANOS =
445 >        30L * 1000L * 1000L * 1000L; // 2 per minute
446 >
447 >    /**
448 >     * Absolute bound for parallelism level. Twice this number plus
449 >     * one (i.e., 0xfff) must fit into a 16bit field to enable
450 >     * word-packing for some counts and indices.
451       */
452 <    private static final int MAX_THREADS = 0x7fff;
452 >    private static final int MAX_WORKERS   = 0x7fff;
453  
454      /**
455       * Array holding all worker threads in the pool.  Array size must
# Line 450 | Line 489 | public class ForkJoinPool extends Abstra
489      private volatile long stealCount;
490  
491      /**
492 <     * Encoded record of top of treiber stack of threads waiting for
492 >     * Encoded record of top of Treiber stack of threads waiting for
493       * events. The top 32 bits contain the count being waited for. The
494 <     * bottom word contains one plus the pool index of waiting worker
495 <     * thread.
494 >     * bottom 16 bits contains one plus the pool index of waiting
495 >     * worker thread. (Bits 16-31 are unused.)
496       */
497      private volatile long eventWaiters;
498  
499      private static final int  EVENT_COUNT_SHIFT = 32;
500 <    private static final long WAITER_ID_MASK = (1L << EVENT_COUNT_SHIFT)-1L;
500 >    private static final long WAITER_ID_MASK    = (1L << 16) - 1L;
501  
502      /**
503       * A counter for events that may wake up worker threads:
504       *   - Submission of a new task to the pool
505       *   - A worker pushing a task on an empty queue
506 <     *   - termination and reconfiguration
506 >     *   - termination
507       */
508      private volatile int eventCount;
509  
510      /**
511 +     * Encoded record of top of Treiber stack of spare threads waiting
512 +     * for resumption. The top 16 bits contain an arbitrary count to
513 +     * avoid ABA effects. The bottom 16bits contains one plus the pool
514 +     * index of waiting worker thread.
515 +     */
516 +    private volatile int spareWaiters;
517 +
518 +    private static final int SPARE_COUNT_SHIFT = 16;
519 +    private static final int SPARE_ID_MASK     = (1 << 16) - 1;
520 +
521 +    /**
522       * Lifecycle control. The low word contains the number of workers
523       * that are (probably) executing tasks. This value is atomically
524       * incremented before a worker gets a task to run, and decremented
525 <     * when worker has no tasks and cannot find any.  Bits 16-18
525 >     * when a worker has no tasks and cannot find any.  Bits 16-18
526       * contain runLevel value. When all are zero, the pool is
527       * running. Level transitions are monotonic (running -> shutdown
528       * -> terminating -> terminated) so each transition adds a bit.
529       * These are bundled together to ensure consistent read for
530       * termination checks (i.e., that runLevel is at least SHUTDOWN
531       * and active threads is zero).
532 +     *
533 +     * Notes: Most direct CASes are dependent on these bitfield
534 +     * positions.  Also, this field is non-private to enable direct
535 +     * performance-sensitive CASes in ForkJoinWorkerThread.
536       */
537 <    private volatile int runState;
537 >    volatile int runState;
538  
539      // Note: The order among run level values matters.
540      private static final int RUNLEVEL_SHIFT     = 16;
# Line 488 | Line 542 | public class ForkJoinPool extends Abstra
542      private static final int TERMINATING        = 1 << (RUNLEVEL_SHIFT + 1);
543      private static final int TERMINATED         = 1 << (RUNLEVEL_SHIFT + 2);
544      private static final int ACTIVE_COUNT_MASK  = (1 << RUNLEVEL_SHIFT) - 1;
491    private static final int ONE_ACTIVE         = 1; // active update delta
545  
546      /**
547       * Holds number of total (i.e., created and not yet terminated)
# Line 497 | Line 550 | public class ForkJoinPool extends Abstra
550       * making decisions about creating and suspending spare
551       * threads. Updated only by CAS. Note that adding a new worker
552       * requires incrementing both counts, since workers start off in
553 <     * running state.  This field is also used for memory-fencing
501 <     * configuration parameters.
553 >     * running state.
554       */
555      private volatile int workerCounts;
556  
# Line 530 | Line 582 | public class ForkJoinPool extends Abstra
582       */
583      private final int poolNumber;
584  
585 <    // Utilities for CASing fields. Note that several of these
586 <    // are manually inlined by callers
585 >    // Utilities for CASing fields. Note that most of these
586 >    // are usually manually inlined by callers
587  
588      /**
589 <     * Increments running count.  Also used by ForkJoinTask.
589 >     * Increments running count part of workerCounts
590       */
591      final void incrementRunningCount() {
592          int c;
# Line 555 | Line 607 | public class ForkJoinPool extends Abstra
607      }
608  
609      /**
610 <     * Tries to increment running count
611 <     */
560 <    final boolean tryIncrementRunningCount() {
561 <        int wc;
562 <        return UNSAFE.compareAndSwapInt(this, workerCountsOffset,
563 <                                        wc = workerCounts, wc + ONE_RUNNING);
564 <    }
565 <
566 <    /**
567 <     * Tries incrementing active count; fails on contention.
568 <     * Called by workers before executing tasks.
610 >     * Forces decrement of encoded workerCounts, awaiting nonzero if
611 >     * (rarely) necessary when other count updates lag.
612       *
613 <     * @return true on success
613 >     * @param dr -- either zero or ONE_RUNNING
614 >     * @param dt -- either zero or ONE_TOTAL
615       */
616 <    final boolean tryIncrementActiveCount() {
617 <        int c;
618 <        return UNSAFE.compareAndSwapInt(this, runStateOffset,
619 <                                        c = runState, c + ONE_ACTIVE);
616 >    private void decrementWorkerCounts(int dr, int dt) {
617 >        for (;;) {
618 >            int wc = workerCounts;
619 >            if ((wc & RUNNING_COUNT_MASK)  - dr < 0 ||
620 >                (wc >>> TOTAL_COUNT_SHIFT) - dt < 0) {
621 >                if ((runState & TERMINATED) != 0)
622 >                    return; // lagging termination on a backout
623 >                Thread.yield();
624 >            }
625 >            if (UNSAFE.compareAndSwapInt(this, workerCountsOffset,
626 >                                         wc, wc - (dr + dt)))
627 >                return;
628 >        }
629      }
630  
631      /**
# Line 582 | Line 635 | public class ForkJoinPool extends Abstra
635      final boolean tryDecrementActiveCount() {
636          int c;
637          return UNSAFE.compareAndSwapInt(this, runStateOffset,
638 <                                        c = runState, c - ONE_ACTIVE);
638 >                                        c = runState, c - 1);
639      }
640  
641      /**
# Line 611 | Line 664 | public class ForkJoinPool extends Abstra
664          lock.lock();
665          try {
666              ForkJoinWorkerThread[] ws = workers;
667 <            int nws = ws.length;
668 <            if (k < 0 || k >= nws || ws[k] != null) {
669 <                for (k = 0; k < nws && ws[k] != null; ++k)
667 >            int n = ws.length;
668 >            if (k < 0 || k >= n || ws[k] != null) {
669 >                for (k = 0; k < n && ws[k] != null; ++k)
670                      ;
671 <                if (k == nws)
672 <                    ws = Arrays.copyOf(ws, nws << 1);
671 >                if (k == n)
672 >                    ws = Arrays.copyOf(ws, n << 1);
673              }
674              ws[k] = w;
675              workers = ws; // volatile array write ensures slot visibility
# Line 627 | Line 680 | public class ForkJoinPool extends Abstra
680      }
681  
682      /**
683 <     * Nulls out record of worker in workers array
683 >     * Nulls out record of worker in workers array.
684       */
685      private void forgetWorker(ForkJoinWorkerThread w) {
686          int idx = w.poolIndex;
687 <        // Locking helps method recordWorker avoid unecessary expansion
687 >        // Locking helps method recordWorker avoid unnecessary expansion
688          final ReentrantLock lock = this.workerLock;
689          lock.lock();
690          try {
# Line 643 | Line 696 | public class ForkJoinPool extends Abstra
696          }
697      }
698  
646    // adding and removing workers
647
699      /**
700 <     * Tries to create and add new worker. Assumes that worker counts
701 <     * are already updated to accommodate the worker, so adjusts on
702 <     * failure.
700 >     * Final callback from terminating worker.  Removes record of
701 >     * worker from array, and adjusts counts. If pool is shutting
702 >     * down, tries to complete termination.
703       *
704 <     * @return new worker or null if creation failed
704 >     * @param w the worker
705       */
706 <    private ForkJoinWorkerThread addWorker() {
707 <        ForkJoinWorkerThread w = null;
708 <        try {
709 <            w = factory.newThread(this);
710 <        } finally { // Adjust on either null or exceptional factory return
711 <            if (w == null) {
661 <                onWorkerCreationFailure();
662 <                return null;
663 <            }
664 <        }
665 <        w.start(recordWorker(w), ueh);
666 <        return w;
706 >    final void workerTerminated(ForkJoinWorkerThread w) {
707 >        forgetWorker(w);
708 >        decrementWorkerCounts(w.isTrimmed()? 0 : ONE_RUNNING, ONE_TOTAL);
709 >        while (w.stealCount != 0) // collect final count
710 >            tryAccumulateStealCount(w);
711 >        tryTerminate(false);
712      }
713  
714 +    // Waiting for and signalling events
715 +
716      /**
717 <     * Adjusts counts upon failure to create worker
717 >     * Releases workers blocked on a count not equal to current count.
718 >     * Normally called after precheck that eventWaiters isn't zero to
719 >     * avoid wasted array checks. Gives up upon a change in count or
720 >     * upon releasing two workers, letting others take over.
721       */
722 <    private void onWorkerCreationFailure() {
723 <        for (;;) {
724 <            int wc = workerCounts;
725 <            if ((wc >>> TOTAL_COUNT_SHIFT) == 0)
726 <                Thread.yield(); // wait for other counts to settle
727 <            else if (UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
728 <                                              wc - (ONE_RUNNING|ONE_TOTAL)))
722 >    private void releaseEventWaiters() {
723 >        ForkJoinWorkerThread[] ws = workers;
724 >        int n = ws.length;
725 >        long h = eventWaiters;
726 >        int ec = eventCount;
727 >        boolean releasedOne = false;
728 >        ForkJoinWorkerThread w; int id;
729 >        while ((id = ((int)(h & WAITER_ID_MASK)) - 1) >= 0 &&
730 >               (int)(h >>> EVENT_COUNT_SHIFT) != ec &&
731 >               id < n && (w = ws[id]) != null) {
732 >            if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
733 >                                          h,  w.nextWaiter)) {
734 >                LockSupport.unpark(w);
735 >                if (releasedOne) // exit on second release
736 >                    break;
737 >                releasedOne = true;
738 >            }
739 >            if (eventCount != ec)
740                  break;
741 +            h = eventWaiters;
742          }
681        tryTerminate(false); // in case of failure during shutdown
743      }
744  
745      /**
746 <     * Creates and/or resumes enough workers to establish target
747 <     * parallelism, giving up if terminating or addWorker fails
687 <     *
688 <     * TODO: recast this to support lazier creation and automated
689 <     * parallelism maintenance
746 >     * Tries to advance eventCount and releases waiters. Called only
747 >     * from workers.
748       */
749 <    private void ensureEnoughWorkers() {
750 <        while ((runState & TERMINATING) == 0) {
751 <            int pc = parallelism;
752 <            int wc = workerCounts;
753 <            int rc = wc & RUNNING_COUNT_MASK;
696 <            int tc = wc >>> TOTAL_COUNT_SHIFT;
697 <            if (tc < pc) {
698 <                if (UNSAFE.compareAndSwapInt
699 <                    (this, workerCountsOffset,
700 <                     wc, wc + (ONE_RUNNING|ONE_TOTAL)) &&
701 <                    addWorker() == null)
702 <                    break;
703 <            }
704 <            else if (tc > pc && rc < pc &&
705 <                     tc > (runState & ACTIVE_COUNT_MASK)) {
706 <                ForkJoinWorkerThread spare = null;
707 <                ForkJoinWorkerThread[] ws = workers;
708 <                int nws = ws.length;
709 <                for (int i = 0; i < nws; ++i) {
710 <                    ForkJoinWorkerThread w = ws[i];
711 <                    if (w != null && w.isSuspended()) {
712 <                        if ((workerCounts & RUNNING_COUNT_MASK) > pc)
713 <                            return;
714 <                        if (w.tryResumeSpare())
715 <                            incrementRunningCount();
716 <                        break;
717 <                    }
718 <                }
719 <            }
720 <            else
721 <                break;
722 <        }
749 >    final void signalWork() {
750 >        int c; // try to increment event count -- CAS failure OK
751 >        UNSAFE.compareAndSwapInt(this, eventCountOffset, c = eventCount, c+1);
752 >        if (eventWaiters != 0L)
753 >            releaseEventWaiters();
754      }
755  
756      /**
757 <     * Final callback from terminating worker.  Removes record of
758 <     * worker from array, and adjusts counts. If pool is shutting
728 <     * down, tries to complete terminatation, else possibly replaces
729 <     * the worker.
757 >     * Adds the given worker to event queue and blocks until
758 >     * terminating or event count advances from the given value
759       *
760 <     * @param w the worker
760 >     * @param w the calling worker thread
761 >     * @param ec the count
762       */
763 <    final void workerTerminated(ForkJoinWorkerThread w) {
764 <        if (w.active) { // force inactive
765 <            w.active = false;
766 <            do {} while (!tryDecrementActiveCount());
767 <        }
768 <        forgetWorker(w);
769 <
770 <        // Decrement total count, and if was running, running count
771 <        // Spin (waiting for other updates) if either would be negative
772 <        int nr = w.isTrimmed() ? 0 : ONE_RUNNING;
743 <        int unit = ONE_TOTAL + nr;
744 <        for (;;) {
745 <            int wc = workerCounts;
746 <            int rc = wc & RUNNING_COUNT_MASK;
747 <            if (rc - nr < 0 || (wc >>> TOTAL_COUNT_SHIFT) == 0)
748 <                Thread.yield(); // back off if waiting for other updates
749 <            else if (UNSAFE.compareAndSwapInt(this, workerCountsOffset,
750 <                                              wc, wc - unit))
763 >    private void eventSync(ForkJoinWorkerThread w, int ec) {
764 >        long nh = (((long)ec) << EVENT_COUNT_SHIFT) | ((long)(w.poolIndex+1));
765 >        long h;
766 >        while ((runState < SHUTDOWN || !tryTerminate(false)) &&
767 >               (((int)((h = eventWaiters) & WAITER_ID_MASK)) == 0 ||
768 >                (int)(h >>> EVENT_COUNT_SHIFT) == ec) &&
769 >               eventCount == ec) {
770 >            if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
771 >                                          w.nextWaiter = h, nh)) {
772 >                awaitEvent(w, ec);
773                  break;
774 +            }
775          }
753
754        accumulateStealCount(w); // collect final count
755        if (!tryTerminate(false))
756            ensureEnoughWorkers();
776      }
777  
759    // Waiting for and signalling events
760
778      /**
779 <     * Releases workers blocked on a count not equal to current count.
780 <     * @return true if any released
779 >     * Blocks the given worker (that has already been entered as an
780 >     * event waiter) until terminating or event count advances from
781 >     * the given value. The oldest (first) waiter uses a timed wait to
782 >     * occasionally one-by-one shrink the number of workers (to a
783 >     * minimum of one) if the pool has not been used for extended
784 >     * periods.
785 >     *
786 >     * @param w the calling worker thread
787 >     * @param ec the count
788       */
789 <    private void releaseWaiters() {
790 <        long top;
791 <        while ((top = eventWaiters) != 0L) {
792 <            ForkJoinWorkerThread[] ws = workers;
793 <            int n = ws.length;
794 <            for (;;) {
795 <                int i = ((int)(top & WAITER_ID_MASK)) - 1;
796 <                if (i < 0 || (int)(top >>> EVENT_COUNT_SHIFT) == eventCount)
797 <                    return;
798 <                ForkJoinWorkerThread w;
799 <                if (i < n && (w = ws[i]) != null &&
800 <                    UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
801 <                                              top, w.nextWaiter)) {
802 <                    LockSupport.unpark(w);
803 <                    top = eventWaiters;
789 >    private void awaitEvent(ForkJoinWorkerThread w, int ec) {
790 >        while (eventCount == ec) {
791 >            if (tryAccumulateStealCount(w)) { // transfer while idle
792 >                boolean untimed = (w.nextWaiter != 0L ||
793 >                                   (workerCounts & RUNNING_COUNT_MASK) <= 1);
794 >                long startTime = untimed? 0 : System.nanoTime();
795 >                Thread.interrupted();         // clear/ignore interrupt
796 >                if (eventCount != ec || w.isTerminating())
797 >                    break;                    // recheck after clear
798 >                if (untimed)
799 >                    LockSupport.park(w);
800 >                else {
801 >                    LockSupport.parkNanos(w, SHRINK_RATE_NANOS);
802 >                    if (eventCount != ec || w.isTerminating())
803 >                        break;
804 >                    if (System.nanoTime() - startTime >= SHRINK_RATE_NANOS)
805 >                        tryShutdownUnusedWorker(ec);
806                  }
781                else
782                    break;      // possibly stale; reread
807              }
808          }
809      }
810  
811 +    // Maintaining parallelism
812 +
813      /**
814 <     * Ensures eventCount on exit is different (mod 2^32) than on
789 <     * entry and wakes up all waiters
814 >     * Pushes worker onto the spare stack.
815       */
816 <    private void signalEvent() {
817 <        int c;
818 <        do {} while (!UNSAFE.compareAndSwapInt(this, eventCountOffset,
819 <                                               c = eventCount, c+1));
795 <        releaseWaiters();
816 >    final void pushSpare(ForkJoinWorkerThread w) {
817 >        int ns = (++w.spareCount << SPARE_COUNT_SHIFT) | (w.poolIndex + 1);
818 >        do {} while (!UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
819 >                                               w.nextSpare = spareWaiters,ns));
820      }
821  
822      /**
823 <     * Advances eventCount and releases waiters until interference by
824 <     * other releasing threads is detected.
823 >     * Tries (once) to resume a spare if the number of running
824 >     * threads is less than target.
825       */
826 <    final void signalWork() {
827 <        int c;
828 <        UNSAFE.compareAndSwapInt(this, eventCountOffset, c=eventCount, c+1);
829 <        long top;
830 <        while ((top = eventWaiters) != 0L) {
831 <            int ec = eventCount;
832 <            ForkJoinWorkerThread[] ws = workers;
833 <            int n = ws.length;
834 <            for (;;) {
835 <                int i = ((int)(top & WAITER_ID_MASK)) - 1;
836 <                if (i < 0 || (int)(top >>> EVENT_COUNT_SHIFT) == ec)
837 <                    return;
838 <                ForkJoinWorkerThread w;
839 <                if (i < n && (w = ws[i]) != null &&
840 <                    UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
841 <                                              top, top = w.nextWaiter)) {
842 <                    LockSupport.unpark(w);
843 <                    if (top != eventWaiters) // let someone else take over
844 <                        return;
845 <                }
822 <                else
823 <                    break;      // possibly stale; reread
824 <            }
826 >    private void tryResumeSpare() {
827 >        int sw, id;
828 >        ForkJoinWorkerThread[] ws = workers;
829 >        int n = ws.length;
830 >        ForkJoinWorkerThread w;
831 >        if ((sw = spareWaiters) != 0 &&
832 >            (id = (sw & SPARE_ID_MASK) - 1) >= 0 &&
833 >            id < n && (w = ws[id]) != null &&
834 >            (workerCounts & RUNNING_COUNT_MASK) < parallelism &&
835 >            spareWaiters == sw &&
836 >            UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
837 >                                     sw, w.nextSpare)) {
838 >            int c; // increment running count before resume
839 >            do {} while (!UNSAFE.compareAndSwapInt
840 >                         (this, workerCountsOffset,
841 >                          c = workerCounts, c + ONE_RUNNING));
842 >            if (w.tryUnsuspend())
843 >                LockSupport.unpark(w);
844 >            else   // back out if w was shutdown
845 >                decrementWorkerCounts(ONE_RUNNING, 0);
846          }
847      }
848  
849      /**
850 <     * If worker is inactive, blocks until terminating or event count
851 <     * advances from last value held by worker; in any case helps
852 <     * release others.
853 <     *
854 <     * @param w the calling worker thread
834 <     * @param retries the number of scans by caller failing to find work
835 <     * @return false if now too many threads running
850 >     * Tries to increase the number of running workers if below target
851 >     * parallelism: If a spare exists tries to resume it via
852 >     * tryResumeSpare.  Otherwise, if not enough total workers or all
853 >     * existing workers are busy, adds a new worker. In all cases also
854 >     * helps wake up releasable workers waiting for work.
855       */
856 <    private boolean eventSync(ForkJoinWorkerThread w, int retries) {
857 <        int wec = w.lastEventCount;
858 <        if (retries > 1) { // can only block after 2nd miss
859 <            long nextTop = (((long)wec << EVENT_COUNT_SHIFT) |
860 <                            ((long)(w.poolIndex + 1)));
861 <            long top;
862 <            while ((runState < SHUTDOWN || !tryTerminate(false)) &&
863 <                   (((int)(top = eventWaiters) & WAITER_ID_MASK) == 0 ||
864 <                    (int)(top >>> EVENT_COUNT_SHIFT) == wec) &&
865 <                   eventCount == wec) {
866 <                if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
867 <                                              w.nextWaiter = top, nextTop)) {
868 <                    accumulateStealCount(w); // transfer steals while idle
869 <                    Thread.interrupted();    // clear/ignore interrupt
870 <                    while (eventCount == wec)
871 <                        w.doPark();
856 >    private void helpMaintainParallelism() {
857 >        int pc = parallelism;
858 >        int wc, rs, tc;
859 >        while (((wc = workerCounts) & RUNNING_COUNT_MASK) < pc &&
860 >               (rs = runState) < TERMINATING) {
861 >            if (spareWaiters != 0)
862 >                tryResumeSpare();
863 >            else if ((tc = wc >>> TOTAL_COUNT_SHIFT) >= MAX_WORKERS ||
864 >                     (tc >= pc && (rs & ACTIVE_COUNT_MASK) != tc))
865 >                break;   // enough total
866 >            else if (runState == rs && workerCounts == wc &&
867 >                     UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
868 >                                              wc + (ONE_RUNNING|ONE_TOTAL))) {
869 >                ForkJoinWorkerThread w = null;
870 >                Throwable fail = null;
871 >                try {
872 >                    w = factory.newThread(this);
873 >                } catch (Throwable ex) {
874 >                    fail = ex;
875 >                }
876 >                if (w == null) { // null or exceptional factory return
877 >                    decrementWorkerCounts(ONE_RUNNING, ONE_TOTAL);
878 >                    tryTerminate(false); // handle failure during shutdown
879 >                    // If originating from an external caller,
880 >                    // propagate exception, else ignore
881 >                    if (fail != null && runState < TERMINATING &&
882 >                        !(Thread.currentThread() instanceof
883 >                          ForkJoinWorkerThread))
884 >                        UNSAFE.throwException(fail);
885                      break;
886                  }
887 +                w.start(recordWorker(w), ueh);
888 +                if ((workerCounts >>> TOTAL_COUNT_SHIFT) >= pc) {
889 +                    int c; // advance event count
890 +                    UNSAFE.compareAndSwapInt(this, eventCountOffset,
891 +                                             c = eventCount, c+1);
892 +                    break; // add at most one unless total below target
893 +                }
894              }
856            wec = eventCount;
895          }
896 <        releaseWaiters();
897 <        int wc = workerCounts;
898 <        if ((wc & RUNNING_COUNT_MASK) <= parallelism) {
899 <            w.lastEventCount = wec;
900 <            return true;
896 >        if (eventWaiters != 0L)
897 >            releaseEventWaiters();
898 >    }
899 >
900 >    /**
901 >     * Callback from the oldest waiter in awaitEvent waking up after a
902 >     * period of non-use. If all workers are idle, tries (once) to
903 >     * shutdown an event waiter or a spare, if one exists. Note that
904 >     * we don't need CAS or locks here because the method is called
905 >     * only from one thread occasionally waking (and even misfires are
906 >     * OK). Note that until the shutdown worker fully terminates,
907 >     * workerCounts will overestimate total count, which is tolerable.
908 >     *
909 >     * @param ec the event count waited on by caller (to abort
910 >     * attempt if count has since changed).
911 >     */
912 >    private void tryShutdownUnusedWorker(int ec) {
913 >        if (runState == 0 && eventCount == ec) { // only trigger if all idle
914 >            ForkJoinWorkerThread[] ws = workers;
915 >            int n = ws.length;
916 >            ForkJoinWorkerThread w = null;
917 >            boolean shutdown = false;
918 >            int sw;
919 >            long h;
920 >            if ((sw = spareWaiters) != 0) { // prefer killing spares
921 >                int id = (sw & SPARE_ID_MASK) - 1;
922 >                if (id >= 0 && id < n && (w = ws[id]) != null &&
923 >                    UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
924 >                                             sw, w.nextSpare))
925 >                    shutdown = true;
926 >            }
927 >            else if ((h = eventWaiters) != 0L) {
928 >                long nh;
929 >                int id = ((int)(h & WAITER_ID_MASK)) - 1;
930 >                if (id >= 0 && id < n && (w = ws[id]) != null &&
931 >                    (nh = w.nextWaiter) != 0L && // keep at least one worker
932 >                    UNSAFE.compareAndSwapLong(this, eventWaitersOffset, h, nh))
933 >                    shutdown = true;
934 >            }
935 >            if (w != null && shutdown) {
936 >                w.shutdown();
937 >                LockSupport.unpark(w);
938 >            }
939          }
940 <        if (wec != w.lastEventCount) // back up if may re-wait
865 <            w.lastEventCount = wec - (wc >>> TOTAL_COUNT_SHIFT);
866 <        return false;
940 >        releaseEventWaiters(); // in case of interference
941      }
942  
943      /**
944       * Callback from workers invoked upon each top-level action (i.e.,
945 <     * stealing a task or taking a submission and running
946 <     * it). Performs one or both of the following:
945 >     * stealing a task or taking a submission and running it).
946 >     * Performs one or more of the following:
947       *
948 <     * * If the worker cannot find work, updates its active status to
949 <     * inactive and updates activeCount unless there is contention, in
950 <     * which case it may try again (either in this or a subsequent
951 <     * call).  Additionally, awaits the next task event and/or helps
952 <     * wake up other releasable waiters.
953 <     *
954 <     * * If there are too many running threads, suspends this worker
955 <     * (first forcing inactivation if necessary).  If it is not
956 <     * resumed before a keepAlive elapses, the worker may be "trimmed"
957 <     * -- killed while suspended within suspendAsSpare. Otherwise,
958 <     * upon resume it rechecks to make sure that it is still needed.
948 >     * 1. If the worker is active and either did not run a task
949 >     *    or there are too many workers, try to set its active status
950 >     *    to inactive and update activeCount. On contention, we may
951 >     *    try again in this or a subsequent call.
952 >     *
953 >     * 2. If not enough total workers, help create some.
954 >     *
955 >     * 3. If there are too many running workers, suspend this worker
956 >     *    (first forcing inactive if necessary).  If it is not needed,
957 >     *    it may be shutdown while suspended (via
958 >     *    tryShutdownUnusedWorker).  Otherwise, upon resume it
959 >     *    rechecks running thread count and need for event sync.
960 >     *
961 >     * 4. If worker did not run a task, await the next task event via
962 >     *    eventSync if necessary (first forcing inactivation), upon
963 >     *    which the worker may be shutdown via
964 >     *    tryShutdownUnusedWorker.  Otherwise, help release any
965 >     *    existing event waiters that are now releasable,
966       *
967       * @param w the worker
968 <     * @param retries the number of scans by caller failing to find work
888 <     * find any (in which case it may block waiting for work).
968 >     * @param ran true if worker ran a task since last call to this method
969       */
970 <    final void preStep(ForkJoinWorkerThread w, int retries) {
970 >    final void preStep(ForkJoinWorkerThread w, boolean ran) {
971 >        int wec = w.lastEventCount;
972          boolean active = w.active;
973 <        boolean inactivate = active && retries != 0;
974 <        for (;;) {
975 <            int rs, wc;
976 <            if (inactivate &&
977 <                UNSAFE.compareAndSwapInt(this, runStateOffset,
978 <                                         rs = runState, rs - ONE_ACTIVE))
973 >        boolean inactivate = false;
974 >        int pc = parallelism;
975 >        while (w.runState == 0) {
976 >            int rs = runState;
977 >            if (rs >= TERMINATING) { // propagate shutdown
978 >                w.shutdown();
979 >                break;
980 >            }
981 >            if ((inactivate || (active && (rs & ACTIVE_COUNT_MASK) >= pc)) &&
982 >                UNSAFE.compareAndSwapInt(this, runStateOffset, rs, rs - 1))
983                  inactivate = active = w.active = false;
984 <            if (((wc = workerCounts) & RUNNING_COUNT_MASK) <= parallelism) {
985 <                if (active || eventSync(w, retries))
984 >            int wc = workerCounts;
985 >            if ((wc & RUNNING_COUNT_MASK) > pc) {
986 >                if (!(inactivate |= active) && // must inactivate to suspend
987 >                    workerCounts == wc &&      // try to suspend as spare
988 >                    UNSAFE.compareAndSwapInt(this, workerCountsOffset,
989 >                                             wc, wc - ONE_RUNNING))
990 >                    w.suspendAsSpare();
991 >            }
992 >            else if ((wc >>> TOTAL_COUNT_SHIFT) < pc)
993 >                helpMaintainParallelism();     // not enough workers
994 >            else if (!ran) {
995 >                long h = eventWaiters;
996 >                int ec = eventCount;
997 >                if (h != 0L && (int)(h >>> EVENT_COUNT_SHIFT) != ec)
998 >                    releaseEventWaiters();     // release others before waiting
999 >                else if (ec != wec) {
1000 >                    w.lastEventCount = ec;     // no need to wait
1001                      break;
1002 +                }
1003 +                else if (!(inactivate |= active))
1004 +                    eventSync(w, wec);         // must inactivate before sync
1005              }
1006 <            else if (!(inactivate |= active) &&  // must inactivate to suspend
904 <                UNSAFE.compareAndSwapInt(this, workerCountsOffset,
905 <                                         wc, wc - ONE_RUNNING) &&
906 <                !w.suspendAsSpare())             // false if trimmed
1006 >            else
1007                  break;
1008          }
1009      }
1010  
1011      /**
1012 <     * Awaits join of the given task if enough threads, or can resume
1013 <     * or create a spare. Fails (in which case the given task might
914 <     * not be done) upon contention or lack of decision about
915 <     * blocking. Returns void because caller must check
916 <     * task status on return anyway.
917 <     *
918 <     * We allow blocking if:
919 <     *
920 <     * 1. There would still be at least as many running threads as
921 <     *    parallelism level if this thread blocks.
922 <     *
923 <     * 2. A spare is resumed to replace this worker. We tolerate
924 <     *    slop in the decision to replace if a spare is found without
925 <     *    first decrementing run count.  This may release too many,
926 <     *    but if so, the superfluous ones will re-suspend via
927 <     *    preStep().
928 <     *
929 <     * 3. After #spares repeated checks, there are no fewer than #spare
930 <     *    threads not running. We allow this slack to avoid hysteresis
931 <     *    and as a hedge against lag/uncertainty of running count
932 <     *    estimates when signalling or unblocking stalls.
933 <     *
934 <     * 4. All existing workers are busy (as rechecked via repeated
935 <     *    retries by caller) and a new spare is created.
936 <     *
937 <     * If none of the above hold, we try to escape out by
938 <     * re-incrementing count and returning to caller, which can retry
939 <     * later.
1012 >     * Helps and/or blocks awaiting join of the given task.
1013 >     * See above for explanation.
1014       *
1015       * @param joinMe the task to join
1016 <     * @param retries if negative, then serve only as a precheck
943 <     *   that the thread can be replaced by a spare. Otherwise,
944 <     *   the number of repeated calls to this method returning busy
945 <     * @return true if the call must be retried because there
946 <     *   none of the blocking checks hold
1016 >     * @param worker the current worker thread
1017       */
1018 <    final boolean tryAwaitJoin(ForkJoinTask<?> joinMe, int retries) {
1019 <        if (joinMe.status < 0) // precheck for cancellation
1020 <            return false;
1021 <        if ((runState & TERMINATING) != 0) { // shutting down
1022 <            joinMe.cancelIgnoringExceptions();
1023 <            return false;
1024 <        }
955 <
956 <        int pc = parallelism;
957 <        boolean running = true; // false when running count decremented
958 <        outer:for (;;) {
959 <            int wc = workerCounts;
960 <            int rc = wc & RUNNING_COUNT_MASK;
961 <            int tc = wc >>> TOTAL_COUNT_SHIFT;
962 <            if (running) { // replace with spare or decrement count
963 <                if (rc <= pc && tc > pc &&
964 <                    (retries > 0 || tc > (runState & ACTIVE_COUNT_MASK))) {
965 <                    ForkJoinWorkerThread[] ws = workers;
966 <                    int nws = ws.length;
967 <                    for (int i = 0; i < nws; ++i) { // search for spare
968 <                        ForkJoinWorkerThread w = ws[i];
969 <                        if (w != null) {
970 <                            if (joinMe.status < 0)
971 <                                return false;
972 <                            if (w.isSuspended()) {
973 <                                if ((workerCounts & RUNNING_COUNT_MASK)>=pc &&
974 <                                    w.tryResumeSpare()) {
975 <                                    running = false;
976 <                                    break outer;
977 <                                }
978 <                                continue outer; // rescan
979 <                            }
980 <                        }
981 <                    }
982 <                }
983 <                if (retries < 0 || // < 0 means replacement check only
984 <                    rc == 0 || joinMe.status < 0 || workerCounts != wc ||
985 <                    !UNSAFE.compareAndSwapInt(this, workerCountsOffset,
986 <                                              wc, wc - ONE_RUNNING))
987 <                    return false; // done or inconsistent or contended
988 <                running = false;
989 <                if (rc > pc)
990 <                    break;
1018 >    final void awaitJoin(ForkJoinTask<?> joinMe, ForkJoinWorkerThread worker) {
1019 >        int retries = 2 + (parallelism >> 2); // #helpJoins before blocking
1020 >        while (joinMe.status >= 0) {
1021 >            int wc;
1022 >            if (runState >= TERMINATING) {
1023 >                joinMe.cancelIgnoringExceptions();
1024 >                break;
1025              }
1026 <            else { // allow blocking if enough threads
1027 <                if (rc >= pc || joinMe.status < 0)
1028 <                    break;
1029 <                int sc = tc - pc + 1; // = spare threads, plus the one to add
1030 <                if (retries > sc) {
1031 <                    if (rc > 0 && rc >= pc - sc) // allow slack
1032 <                        break;
1033 <                    if (tc < MAX_THREADS &&
1034 <                        tc == (runState & ACTIVE_COUNT_MASK) &&
1035 <                        workerCounts == wc &&
1036 <                        UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
1037 <                                                 wc+(ONE_RUNNING|ONE_TOTAL))) {
1038 <                        addWorker();
1039 <                        break;
1040 <                    }
1041 <                }
1042 <                if (workerCounts == wc &&        // back out to allow rescan
1043 <                    UNSAFE.compareAndSwapInt (this, workerCountsOffset,
1044 <                                              wc, wc + ONE_RUNNING)) {
1045 <                    releaseWaiters();            // help others progress
1046 <                    return true;                 // let caller retry
1047 <                }
1026 >            worker.helpJoinTask(joinMe);
1027 >            if (joinMe.status < 0)
1028 >                break;
1029 >            else if (retries > 0)
1030 >                --retries;
1031 >            else if (((wc = workerCounts) & RUNNING_COUNT_MASK) != 0 &&
1032 >                     UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1033 >                                              wc, wc - ONE_RUNNING)) {
1034 >                int stat, c; long h;
1035 >                while ((stat = joinMe.status) >= 0 &&
1036 >                       (h = eventWaiters) != 0L && // help release others
1037 >                       (int)(h >>> EVENT_COUNT_SHIFT) != eventCount)
1038 >                    releaseEventWaiters();
1039 >                if (stat >= 0 &&
1040 >                    ((workerCounts & RUNNING_COUNT_MASK) == 0 ||
1041 >                     (stat =
1042 >                      joinMe.internalAwaitDone(JOIN_TIMEOUT_MILLIS)) >= 0))
1043 >                    helpMaintainParallelism(); // timeout or no running workers
1044 >                do {} while (!UNSAFE.compareAndSwapInt
1045 >                             (this, workerCountsOffset,
1046 >                              c = workerCounts, c + ONE_RUNNING));
1047 >                if (stat < 0)
1048 >                    break;   // else restart
1049              }
1050          }
1016        // arrive here if can block
1017        joinMe.internalAwaitDone();
1018        int c;                      // to inline incrementRunningCount
1019        do {} while (!UNSAFE.compareAndSwapInt
1020                     (this, workerCountsOffset,
1021                      c = workerCounts, c + ONE_RUNNING));
1022        return false;
1051      }
1052  
1053      /**
1054 <     * Same idea as (and shares many code snippets with) tryAwaitJoin,
1027 <     * but self-contained because there are no caller retries.
1028 <     * TODO: Rework to use simpler API.
1054 >     * Same idea as awaitJoin, but no helping, retries, or timeouts.
1055       */
1056      final void awaitBlocker(ManagedBlocker blocker)
1057          throws InterruptedException {
1058 <        boolean done;
1033 <        if (done = blocker.isReleasable())
1034 <            return;
1035 <        int pc = parallelism;
1036 <        int retries = 0;
1037 <        boolean running = true; // false when running count decremented
1038 <        outer:for (;;) {
1058 >        while (!blocker.isReleasable()) {
1059              int wc = workerCounts;
1060 <            int rc = wc & RUNNING_COUNT_MASK;
1061 <            int tc = wc >>> TOTAL_COUNT_SHIFT;
1062 <            if (running) {
1063 <                if (rc <= pc && tc > pc &&
1064 <                    (retries > 0 || tc > (runState & ACTIVE_COUNT_MASK))) {
1065 <                    ForkJoinWorkerThread[] ws = workers;
1066 <                    int nws = ws.length;
1067 <                    for (int i = 0; i < nws; ++i) {
1068 <                        ForkJoinWorkerThread w = ws[i];
1069 <                        if (w != null) {
1070 <                            if (done = blocker.isReleasable())
1071 <                                return;
1072 <                            if (w.isSuspended()) {
1073 <                                if ((workerCounts & RUNNING_COUNT_MASK)>=pc &&
1054 <                                    w.tryResumeSpare()) {
1055 <                                    running = false;
1056 <                                    break outer;
1057 <                                }
1058 <                                continue outer; // rescan
1059 <                            }
1060 <                        }
1061 <                    }
1062 <                }
1063 <                if (done = blocker.isReleasable())
1064 <                    return;
1065 <                if (rc == 0 || workerCounts != wc ||
1066 <                    !UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1067 <                                              wc, wc - ONE_RUNNING))
1068 <                    continue;
1069 <                running = false;
1070 <                if (rc > pc)
1071 <                    break;
1072 <            }
1073 <            else {
1074 <                if (rc >= pc || (done = blocker.isReleasable()))
1075 <                    break;
1076 <                int sc = tc - pc + 1;
1077 <                if (retries++ > sc) {
1078 <                    if (rc > 0 && rc >= pc - sc)
1079 <                        break;
1080 <                    if (tc < MAX_THREADS &&
1081 <                        tc == (runState & ACTIVE_COUNT_MASK) &&
1082 <                        workerCounts == wc &&
1083 <                        UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
1084 <                                                 wc+(ONE_RUNNING|ONE_TOTAL))) {
1085 <                        addWorker();
1086 <                        break;
1060 >            if ((wc & RUNNING_COUNT_MASK) != 0 &&
1061 >                UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1062 >                                         wc, wc - ONE_RUNNING)) {
1063 >                try {
1064 >                    while (!blocker.isReleasable()) {
1065 >                        long h = eventWaiters;
1066 >                        if (h != 0L &&
1067 >                            (int)(h >>> EVENT_COUNT_SHIFT) != eventCount)
1068 >                            releaseEventWaiters();
1069 >                        else if ((workerCounts & RUNNING_COUNT_MASK) == 0 &&
1070 >                                 runState < TERMINATING)
1071 >                            helpMaintainParallelism();
1072 >                        else if (blocker.block())
1073 >                            break;
1074                      }
1075 +                } finally {
1076 +                    int c;
1077 +                    do {} while (!UNSAFE.compareAndSwapInt
1078 +                                 (this, workerCountsOffset,
1079 +                                  c = workerCounts, c + ONE_RUNNING));
1080                  }
1081 <                Thread.yield();
1090 <            }
1091 <        }
1092 <
1093 <        try {
1094 <            if (!done)
1095 <                do {} while (!blocker.isReleasable() && !blocker.block());
1096 <        } finally {
1097 <            if (!running) {
1098 <                int c;
1099 <                do {} while (!UNSAFE.compareAndSwapInt
1100 <                             (this, workerCountsOffset,
1101 <                              c = workerCounts, c + ONE_RUNNING));
1081 >                break;
1082              }
1083          }
1084      }
# Line 1129 | Line 1109 | public class ForkJoinPool extends Abstra
1109          return true;
1110      }
1111  
1112 +
1113      /**
1114       * Actions on transition to TERMINATING
1115 +     *
1116 +     * Runs up to four passes through workers: (0) shutting down each
1117 +     * (without waking up if parked) to quickly spread notifications
1118 +     * without unnecessary bouncing around event queues etc (1) wake
1119 +     * up and help cancel tasks (2) interrupt (3) mop up races with
1120 +     * interrupted workers
1121       */
1122      private void startTerminating() {
1123 <        for (int i = 0; i < 2; ++i) { // twice to mop up newly created workers
1124 <            cancelSubmissions();
1125 <            shutdownWorkers();
1126 <            cancelWorkerTasks();
1127 <            signalEvent();
1128 <            interruptWorkers();
1123 >        cancelSubmissions();
1124 >        for (int passes = 0; passes < 4 && workerCounts != 0; ++passes) {
1125 >            int c; // advance event count
1126 >            UNSAFE.compareAndSwapInt(this, eventCountOffset,
1127 >                                     c = eventCount, c+1);
1128 >            eventWaiters = 0L; // clobber lists
1129 >            spareWaiters = 0;
1130 >            for (ForkJoinWorkerThread w : workers) {
1131 >                if (w != null) {
1132 >                    w.shutdown();
1133 >                    if (passes > 0 && !w.isTerminated()) {
1134 >                        w.cancelTasks();
1135 >                        LockSupport.unpark(w);
1136 >                        if (passes > 1 && !w.isInterrupted()) {
1137 >                            try {
1138 >                                w.interrupt();
1139 >                            } catch (SecurityException ignore) {
1140 >                            }
1141 >                        }
1142 >                    }
1143 >                }
1144 >            }
1145          }
1146      }
1147  
1148      /**
1149 <     * Clear out and cancel submissions, ignoring exceptions
1149 >     * Clears out and cancels submissions, ignoring exceptions.
1150       */
1151      private void cancelSubmissions() {
1152          ForkJoinTask<?> task;
# Line 1155 | Line 1158 | public class ForkJoinPool extends Abstra
1158          }
1159      }
1160  
1158    /**
1159     * Sets all worker run states to at least shutdown,
1160     * also resuming suspended workers
1161     */
1162    private void shutdownWorkers() {
1163        ForkJoinWorkerThread[] ws = workers;
1164        int nws = ws.length;
1165        for (int i = 0; i < nws; ++i) {
1166            ForkJoinWorkerThread w = ws[i];
1167            if (w != null)
1168                w.shutdown();
1169        }
1170    }
1171
1172    /**
1173     * Clears out and cancels all locally queued tasks
1174     */
1175    private void cancelWorkerTasks() {
1176        ForkJoinWorkerThread[] ws = workers;
1177        int nws = ws.length;
1178        for (int i = 0; i < nws; ++i) {
1179            ForkJoinWorkerThread w = ws[i];
1180            if (w != null)
1181                w.cancelTasks();
1182        }
1183    }
1184
1185    /**
1186     * Unsticks all workers blocked on joins etc
1187     */
1188    private void interruptWorkers() {
1189        ForkJoinWorkerThread[] ws = workers;
1190        int nws = ws.length;
1191        for (int i = 0; i < nws; ++i) {
1192            ForkJoinWorkerThread w = ws[i];
1193            if (w != null && !w.isTerminated()) {
1194                try {
1195                    w.interrupt();
1196                } catch (SecurityException ignore) {
1197                }
1198            }
1199        }
1200    }
1201
1161      // misc support for ForkJoinWorkerThread
1162  
1163      /**
1164 <     * Returns pool number
1164 >     * Returns pool number.
1165       */
1166      final int getPoolNumber() {
1167          return poolNumber;
1168      }
1169  
1170      /**
1171 <     * Accumulates steal count from a worker, clearing
1172 <     * the worker's value
1171 >     * Tries to accumulate steal count from a worker, clearing
1172 >     * the worker's value if successful.
1173 >     *
1174 >     * @return true if worker steal count now zero
1175       */
1176 <    final void accumulateStealCount(ForkJoinWorkerThread w) {
1176 >    final boolean tryAccumulateStealCount(ForkJoinWorkerThread w) {
1177          int sc = w.stealCount;
1178 <        if (sc != 0) {
1179 <            long c;
1180 <            w.stealCount = 0;
1181 <            do {} while (!UNSAFE.compareAndSwapLong(this, stealCountOffset,
1182 <                                                    c = stealCount, c + sc));
1178 >        long c = stealCount;
1179 >        // CAS even if zero, for fence effects
1180 >        if (UNSAFE.compareAndSwapLong(this, stealCountOffset, c, c + sc)) {
1181 >            if (sc != 0)
1182 >                w.stealCount = 0;
1183 >            return true;
1184          }
1185 +        return sc == 0;
1186      }
1187  
1188      /**
# Line 1228 | Line 1191 | public class ForkJoinPool extends Abstra
1191       */
1192      final int idlePerActive() {
1193          int pc = parallelism; // use parallelism, not rc
1194 <        int ac = runState;    // no mask -- artifically boosts during shutdown
1194 >        int ac = runState;    // no mask -- artificially boosts during shutdown
1195          // Use exact results for small values, saturate past 4
1196 <        return pc <= ac? 0 : pc >>> 1 <= ac? 1 : pc >>> 2 <= ac? 3 : pc >>> 3;
1196 >        return ((pc <= ac) ? 0 :
1197 >                (pc >>> 1 <= ac) ? 1 :
1198 >                (pc >>> 2 <= ac) ? 3 :
1199 >                pc >>> 3);
1200      }
1201  
1202      // Public and protected methods
# Line 1280 | Line 1246 | public class ForkJoinPool extends Abstra
1246       * use {@link #defaultForkJoinWorkerThreadFactory}.
1247       * @param handler the handler for internal worker threads that
1248       * terminate due to unrecoverable errors encountered while executing
1249 <     * tasks. For default value, use <code>null</code>.
1249 >     * tasks. For default value, use {@code null}.
1250       * @param asyncMode if true,
1251       * establishes local first-in-first-out scheduling mode for forked
1252       * tasks that are never joined. This mode may be more appropriate
1253       * than default locally stack-based mode in applications in which
1254       * worker threads only process event-style asynchronous tasks.
1255 <     * For default value, use <code>false</code>.
1255 >     * For default value, use {@code false}.
1256       * @throws IllegalArgumentException if parallelism less than or
1257       *         equal to zero, or greater than implementation limit
1258       * @throws NullPointerException if the factory is null
# Line 1302 | Line 1268 | public class ForkJoinPool extends Abstra
1268          checkPermission();
1269          if (factory == null)
1270              throw new NullPointerException();
1271 <        if (parallelism <= 0 || parallelism > MAX_THREADS)
1271 >        if (parallelism <= 0 || parallelism > MAX_WORKERS)
1272              throw new IllegalArgumentException();
1273          this.parallelism = parallelism;
1274          this.factory = factory;
# Line 1321 | Line 1287 | public class ForkJoinPool extends Abstra
1287       * @param pc the initial parallelism level
1288       */
1289      private static int initialArraySizeFor(int pc) {
1290 <        // See Hackers Delight, sec 3.2. We know MAX_THREADS < (1 >>> 16)
1291 <        int size = pc < MAX_THREADS ? pc + 1 : MAX_THREADS;
1290 >        // If possible, initially allocate enough space for one spare
1291 >        int size = pc < MAX_WORKERS ? pc + 1 : MAX_WORKERS;
1292 >        // See Hackers Delight, sec 3.2. We know MAX_WORKERS < (1 >>> 16)
1293          size |= size >>> 1;
1294          size |= size >>> 2;
1295          size |= size >>> 4;
# Line 1333 | Line 1300 | public class ForkJoinPool extends Abstra
1300      // Execution methods
1301  
1302      /**
1303 <     * Common code for execute, invoke and submit
1303 >     * Submits task and creates, starts, or resumes some workers if necessary
1304       */
1305      private <T> void doSubmit(ForkJoinTask<T> task) {
1339        if (task == null)
1340            throw new NullPointerException();
1341        if (runState >= SHUTDOWN)
1342            throw new RejectedExecutionException();
1306          submissionQueue.offer(task);
1307 <        signalEvent();
1308 <        ensureEnoughWorkers();
1307 >        int c; // try to increment event count -- CAS failure OK
1308 >        UNSAFE.compareAndSwapInt(this, eventCountOffset, c = eventCount, c+1);
1309 >        helpMaintainParallelism();
1310      }
1311  
1312      /**
1313       * Performs the given task, returning its result upon completion.
1350     * If the caller is already engaged in a fork/join computation in
1351     * the current pool, this method is equivalent in effect to
1352     * {@link ForkJoinTask#invoke}.
1314       *
1315       * @param task the task
1316       * @return the task's result
# Line 1358 | Line 1319 | public class ForkJoinPool extends Abstra
1319       *         scheduled for execution
1320       */
1321      public <T> T invoke(ForkJoinTask<T> task) {
1322 <        doSubmit(task);
1323 <        return task.join();
1322 >        if (task == null)
1323 >            throw new NullPointerException();
1324 >        if (runState >= SHUTDOWN)
1325 >            throw new RejectedExecutionException();
1326 >        Thread t = Thread.currentThread();
1327 >        if ((t instanceof ForkJoinWorkerThread) &&
1328 >            ((ForkJoinWorkerThread)t).pool == this)
1329 >            return task.invoke();  // bypass submit if in same pool
1330 >        else {
1331 >            doSubmit(task);
1332 >            return task.join();
1333 >        }
1334 >    }
1335 >
1336 >    /**
1337 >     * Unless terminating, forks task if within an ongoing FJ
1338 >     * computation in the current pool, else submits as external task.
1339 >     */
1340 >    private <T> void forkOrSubmit(ForkJoinTask<T> task) {
1341 >        if (runState >= SHUTDOWN)
1342 >            throw new RejectedExecutionException();
1343 >        Thread t = Thread.currentThread();
1344 >        if ((t instanceof ForkJoinWorkerThread) &&
1345 >            ((ForkJoinWorkerThread)t).pool == this)
1346 >            task.fork();
1347 >        else
1348 >            doSubmit(task);
1349      }
1350  
1351      /**
1352       * Arranges for (asynchronous) execution of the given task.
1367     * If the caller is already engaged in a fork/join computation in
1368     * the current pool, this method is equivalent in effect to
1369     * {@link ForkJoinTask#fork}.
1353       *
1354       * @param task the task
1355       * @throws NullPointerException if the task is null
# Line 1374 | Line 1357 | public class ForkJoinPool extends Abstra
1357       *         scheduled for execution
1358       */
1359      public void execute(ForkJoinTask<?> task) {
1360 <        doSubmit(task);
1360 >        if (task == null)
1361 >            throw new NullPointerException();
1362 >        forkOrSubmit(task);
1363      }
1364  
1365      // AbstractExecutorService methods
# Line 1385 | Line 1370 | public class ForkJoinPool extends Abstra
1370       *         scheduled for execution
1371       */
1372      public void execute(Runnable task) {
1373 +        if (task == null)
1374 +            throw new NullPointerException();
1375          ForkJoinTask<?> job;
1376          if (task instanceof ForkJoinTask<?>) // avoid re-wrap
1377              job = (ForkJoinTask<?>) task;
1378          else
1379              job = ForkJoinTask.adapt(task, null);
1380 <        doSubmit(job);
1380 >        forkOrSubmit(job);
1381      }
1382  
1383      /**
1384       * Submits a ForkJoinTask for execution.
1398     * If the caller is already engaged in a fork/join computation in
1399     * the current pool, this method is equivalent in effect to
1400     * {@link ForkJoinTask#fork}.
1385       *
1386       * @param task the task to submit
1387       * @return the task
# Line 1406 | Line 1390 | public class ForkJoinPool extends Abstra
1390       *         scheduled for execution
1391       */
1392      public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) {
1393 <        doSubmit(task);
1393 >        if (task == null)
1394 >            throw new NullPointerException();
1395 >        forkOrSubmit(task);
1396          return task;
1397      }
1398  
# Line 1416 | Line 1402 | public class ForkJoinPool extends Abstra
1402       *         scheduled for execution
1403       */
1404      public <T> ForkJoinTask<T> submit(Callable<T> task) {
1405 +        if (task == null)
1406 +            throw new NullPointerException();
1407          ForkJoinTask<T> job = ForkJoinTask.adapt(task);
1408 <        doSubmit(job);
1408 >        forkOrSubmit(job);
1409          return job;
1410      }
1411  
# Line 1427 | Line 1415 | public class ForkJoinPool extends Abstra
1415       *         scheduled for execution
1416       */
1417      public <T> ForkJoinTask<T> submit(Runnable task, T result) {
1418 +        if (task == null)
1419 +            throw new NullPointerException();
1420          ForkJoinTask<T> job = ForkJoinTask.adapt(task, result);
1421 <        doSubmit(job);
1421 >        forkOrSubmit(job);
1422          return job;
1423      }
1424  
# Line 1438 | Line 1428 | public class ForkJoinPool extends Abstra
1428       *         scheduled for execution
1429       */
1430      public ForkJoinTask<?> submit(Runnable task) {
1431 +        if (task == null)
1432 +            throw new NullPointerException();
1433          ForkJoinTask<?> job;
1434          if (task instanceof ForkJoinTask<?>) // avoid re-wrap
1435              job = (ForkJoinTask<?>) task;
1436          else
1437              job = ForkJoinTask.adapt(task, null);
1438 <        doSubmit(job);
1438 >        forkOrSubmit(job);
1439          return job;
1440      }
1441  
# Line 1503 | Line 1495 | public class ForkJoinPool extends Abstra
1495  
1496      /**
1497       * Returns the number of worker threads that have started but not
1498 <     * yet terminated.  This result returned by this method may differ
1498 >     * yet terminated.  The result returned by this method may differ
1499       * from {@link #getParallelism} when threads are created to
1500       * maintain parallelism when others are cooperatively blocked.
1501       *
# Line 1588 | Line 1580 | public class ForkJoinPool extends Abstra
1580       */
1581      public long getQueuedTaskCount() {
1582          long count = 0;
1583 <        ForkJoinWorkerThread[] ws = workers;
1592 <        int nws = ws.length;
1593 <        for (int i = 0; i < nws; ++i) {
1594 <            ForkJoinWorkerThread w = ws[i];
1583 >        for (ForkJoinWorkerThread w : workers)
1584              if (w != null)
1585                  count += w.getQueueSize();
1597        }
1586          return count;
1587      }
1588  
# Line 1648 | Line 1636 | public class ForkJoinPool extends Abstra
1636       * @return the number of elements transferred
1637       */
1638      protected int drainTasksTo(Collection<? super ForkJoinTask<?>> c) {
1639 <        int n = submissionQueue.drainTo(c);
1640 <        ForkJoinWorkerThread[] ws = workers;
1653 <        int nws = ws.length;
1654 <        for (int i = 0; i < nws; ++i) {
1655 <            ForkJoinWorkerThread w = ws[i];
1639 >        int count = submissionQueue.drainTo(c);
1640 >        for (ForkJoinWorkerThread w : workers)
1641              if (w != null)
1642 <                n += w.drainTasksTo(c);
1658 <        }
1659 <        return n;
1660 <    }
1661 <
1662 <    /**
1663 <     * Returns count of total parks by existing workers.
1664 <     * Used during development only since not meaningful to users.
1665 <     */
1666 <    private int collectParkCount() {
1667 <        int count = 0;
1668 <        ForkJoinWorkerThread[] ws = workers;
1669 <        int nws = ws.length;
1670 <        for (int i = 0; i < nws; ++i) {
1671 <            ForkJoinWorkerThread w = ws[i];
1672 <            if (w != null)
1673 <                count += w.parkCount;
1674 <        }
1642 >                count += w.drainTasksTo(c);
1643          return count;
1644      }
1645  
# Line 1692 | Line 1660 | public class ForkJoinPool extends Abstra
1660          int pc = parallelism;
1661          int rs = runState;
1662          int ac = rs & ACTIVE_COUNT_MASK;
1695        //        int pk = collectParkCount();
1663          return super.toString() +
1664              "[" + runLevelToString(rs) +
1665              ", parallelism = " + pc +
# Line 1702 | Line 1669 | public class ForkJoinPool extends Abstra
1669              ", steals = " + st +
1670              ", tasks = " + qt +
1671              ", submissions = " + qs +
1705            //            ", parks = " + pk +
1672              "]";
1673      }
1674  
# Line 1777 | Line 1743 | public class ForkJoinPool extends Abstra
1743      }
1744  
1745      /**
1746 +     * Returns true if terminating or terminated. Used by ForkJoinWorkerThread.
1747 +     */
1748 +    final boolean isAtLeastTerminating() {
1749 +        return runState >= TERMINATING;
1750 +    }
1751 +
1752 +    /**
1753       * Returns {@code true} if this pool has been shut down.
1754       *
1755       * @return {@code true} if this pool has been shut down
# Line 1800 | Line 1773 | public class ForkJoinPool extends Abstra
1773          throws InterruptedException {
1774          try {
1775              return termination.awaitAdvanceInterruptibly(0, timeout, unit) > 0;
1776 <        } catch(TimeoutException ex) {
1776 >        } catch (TimeoutException ex) {
1777              return false;
1778          }
1779      }
# Line 1809 | Line 1782 | public class ForkJoinPool extends Abstra
1782       * Interface for extending managed parallelism for tasks running
1783       * in {@link ForkJoinPool}s.
1784       *
1785 <     * <p>A {@code ManagedBlocker} provides two methods.
1786 <     * Method {@code isReleasable} must return {@code true} if
1787 <     * blocking is not necessary. Method {@code block} blocks the
1788 <     * current thread if necessary (perhaps internally invoking
1789 <     * {@code isReleasable} before actually blocking).
1785 >     * <p>A {@code ManagedBlocker} provides two methods.  Method
1786 >     * {@code isReleasable} must return {@code true} if blocking is
1787 >     * not necessary. Method {@code block} blocks the current thread
1788 >     * if necessary (perhaps internally invoking {@code isReleasable}
1789 >     * before actually blocking). The unusual methods in this API
1790 >     * accommodate synchronizers that may, but don't usually, block
1791 >     * for long periods. Similarly, they allow more efficient internal
1792 >     * handling of cases in which additional workers may be, but
1793 >     * usually are not, needed to ensure sufficient parallelism.
1794 >     * Toward this end, implementations of method {@code isReleasable}
1795 >     * must be amenable to repeated invocation.
1796       *
1797       * <p>For example, here is a ManagedBlocker based on a
1798       * ReentrantLock:
# Line 1831 | Line 1810 | public class ForkJoinPool extends Abstra
1810       *     return hasLock || (hasLock = lock.tryLock());
1811       *   }
1812       * }}</pre>
1813 +     *
1814 +     * <p>Here is a class that possibly blocks waiting for an
1815 +     * item on a given queue:
1816 +     *  <pre> {@code
1817 +     * class QueueTaker<E> implements ManagedBlocker {
1818 +     *   final BlockingQueue<E> queue;
1819 +     *   volatile E item = null;
1820 +     *   QueueTaker(BlockingQueue<E> q) { this.queue = q; }
1821 +     *   public boolean block() throws InterruptedException {
1822 +     *     if (item == null)
1823 +     *       item = queue.take();
1824 +     *     return true;
1825 +     *   }
1826 +     *   public boolean isReleasable() {
1827 +     *     return item != null || (item = queue.poll()) != null;
1828 +     *   }
1829 +     *   public E getItem() { // call after pool.managedBlock completes
1830 +     *     return item;
1831 +     *   }
1832 +     * }}</pre>
1833       */
1834      public static interface ManagedBlocker {
1835          /**
# Line 1873 | Line 1872 | public class ForkJoinPool extends Abstra
1872      public static void managedBlock(ManagedBlocker blocker)
1873          throws InterruptedException {
1874          Thread t = Thread.currentThread();
1875 <        if (t instanceof ForkJoinWorkerThread)
1876 <            ((ForkJoinWorkerThread) t).pool.awaitBlocker(blocker);
1875 >        if (t instanceof ForkJoinWorkerThread) {
1876 >            ForkJoinWorkerThread w = (ForkJoinWorkerThread) t;
1877 >            w.pool.awaitBlocker(blocker);
1878 >        }
1879          else {
1880              do {} while (!blocker.isReleasable() && !blocker.block());
1881          }
# Line 1902 | Line 1903 | public class ForkJoinPool extends Abstra
1903      private static final long eventCountOffset =
1904          objectFieldOffset("eventCount", ForkJoinPool.class);
1905      private static final long eventWaitersOffset =
1906 <        objectFieldOffset("eventWaiters",ForkJoinPool.class);
1906 >        objectFieldOffset("eventWaiters", ForkJoinPool.class);
1907      private static final long stealCountOffset =
1908 <        objectFieldOffset("stealCount",ForkJoinPool.class);
1908 >        objectFieldOffset("stealCount", ForkJoinPool.class);
1909 >    private static final long spareWaitersOffset =
1910 >        objectFieldOffset("spareWaiters", ForkJoinPool.class);
1911  
1912      private static long objectFieldOffset(String field, Class<?> klazz) {
1913          try {

Diff Legend

Removed lines
+ Added lines
< Changed lines
> Changed lines