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 1.57 by dl, Wed Jul 7 19:52:31 2010 UTC vs.
Revision 1.90 by jsr166, Mon Nov 29 20:58:06 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 60 | Line 65 | import java.util.concurrent.CountDownLat
65   * Runnable}- or {@code Callable}- based activities as well.  However,
66   * tasks that are already executing in a pool should normally
67   * <em>NOT</em> use these pool execution methods, but instead use the
68 < * within-computation forms listed in the table. To avoid inadvertant
64 < * cyclic task dependencies and to improve performance, task
65 < * submissions to the current pool by an ongoing fork/join
66 < * computations may be implicitly translated to the corresponding
67 < * ForkJoinTask forms.
68 > * within-computation forms listed in the table.
69   *
70   * <table BORDER CELLPADDING=3 CELLSPACING=1>
71   *  <tr>
# Line 73 | 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 88 | Line 89 | import java.util.concurrent.CountDownLat
89   *    <td> {@link ForkJoinTask#fork} (ForkJoinTasks <em>are</em> Futures)</td>
90   *  </tr>
91   * </table>
92 < *
92 > *
93   * <p><b>Sample Usage.</b> Normally a single {@code ForkJoinPool} is
94   * used for all parallel task execution in a program or subsystem.
95   * Otherwise, use would not usually outweigh the construction and
# Line 113 | Line 114 | import java.util.concurrent.CountDownLat
114   * {@code IllegalArgumentException}.
115   *
116   * <p>This implementation rejects submitted tasks (that is, by throwing
117 < * {@link RejectedExecutionException}) only when the pool is shut down.
117 > * {@link RejectedExecutionException}) only when the pool is shut down
118 > * or internal resources have been exhausted.
119   *
120   * @since 1.7
121   * @author Doug Lea
# Line 140 | Line 142 | public class ForkJoinPool extends Abstra
142       * of tasks profit from cache affinities, but others are harmed by
143       * cache pollution effects.)
144       *
145 +     * Beyond work-stealing support and essential bookkeeping, the
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 +     *   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 +     *   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 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
186       * other. We do not want to negate this by creating bottlenecks
187 <     * implementing the management responsibilities of this class. So
188 <     * we use a collection of techniques that avoid, reduce, or cope
189 <     * well with contention. These entail several instances of
190 <     * bit-packing into CASable fields to maintain only the minimally
191 <     * required atomicity. To enable such packing, we restrict maximum
192 <     * parallelism to (1<<15)-1 (enabling twice this to fit into a 16
193 <     * bit field), which is far in excess of normal operating range.
194 <     * Even though updates to some of these bookkeeping fields do
195 <     * sometimes contend with each other, they don't normally
196 <     * cache-contend with updates to others enough to warrant memory
197 <     * padding or isolation. So they are all held as fields of
198 <     * ForkJoinPool objects.  The main capabilities are as follows:
187 >     * implementing other management responsibilities. So we use a
188 >     * collection of techniques that avoid, reduce, or cope well with
189 >     * contention. These entail several instances of bit-packing into
190 >     * CASable fields to maintain only the minimally required
191 >     * atomicity. To enable such packing, we restrict maximum
192 >     * parallelism to (1<<15)-1 (enabling twice this (to accommodate
193 >     * unbalanced increments and decrements) to fit into a 16 bit
194 >     * field, which is far in excess of normal operating range.  Even
195 >     * though updates to some of these bookkeeping fields do sometimes
196 >     * contend with each other, they don't normally cache-contend with
197 >     * updates to others enough to warrant memory padding or
198 >     * isolation. So they are all held as fields of ForkJoinPool
199 >     * objects.  The main capabilities are as follows:
200       *
201       * 1. Creating and removing workers. Workers are recorded in the
202       * "workers" array. This is an array as opposed to some other data
# Line 170 | 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.  To support these decisions,
238 <     * updates to spare counts must be prospective (not
239 <     * retrospective).  For example, the running count is decremented
240 <     * before blocking by a thread about to block as a spare, but
186 <     * incremented by the thread about to unblock it. Updates upon
187 <     * resumption ofr threads blocking in awaitJoin or awaitBlocker
188 <     * cannot usually be prospective, so the running count is in
189 <     * general an upper bound of the number of productively running
190 <     * threads Updates to the workerCounts field sometimes transiently
191 <     * encounter a fair amount of contention when join dependencies
192 <     * are such that many threads block or unblock at about the same
193 <     * time. We alleviate this by sometimes performing an alternative
194 <     * action on contention like releasing waiters or locating spares.
237 >     * to create, resume or suspend.  Note however that the
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       *
242       * 3. Maintaining global run state. The run state of the pool
243       * consists of a runLevel (SHUTDOWN, TERMINATING, etc) similar to
# Line 220 | 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 231 | 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,
240 <     * that only releases idle workers until it detects interference
241 <     * by other threads trying to release, and lets them take
242 <     * over. The net effect is a tree-like diffusion of signals, where
243 <     * released threads (and possibly others) help with unparks.  To
244 <     * further reduce contention effects a bit, failed CASes to
245 <     * 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 (see below). Usually, extra threads are needed
293 <     * for only very short periods, yet join dependencies are such
294 <     * that we sometimes need them in bursts. Rather than create new
295 <     * threads each time this happens, we suspend no-longer-needed
296 <     * extra ones as "spares". For most purposes, we don't distinguish
297 <     * "extra" spare threads from normal "core" threads: On each call
298 <     * to 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 awaitJoin and awaitBlocker look
301 <     * for suspended threads to resume before considering creating a
302 <     * new replacement. We don't need a special data structure to
303 <     * maintain spares; simply scanning the workers array looking for
304 <     * worker.isSuspended() is fine because the calling thread is
305 <     * otherwise not doing anything useful anyway; we are at least as
306 <     * happy if after locating a spare, the caller doesn't actually
307 <     * block because the join is ready before we try to adjust and
308 <     * compensate.  Note that this is intrinsically racy.  One thread
309 <     * may become a spare at about the same time as another is
310 <     * needlessly being created. We counteract this and related slop
311 <     * in part by requiring resumed spares to immediately recheck (in
312 <     * preStep) to see whether they they should re-suspend. The only
313 <     * effective difference between "extra" and "core" threads is that
314 <     * we allow the "extra" ones to time out and die if they are not
315 <     * resumed within a keep-alive interval of a few seconds. This is
316 <     * 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 becomes zero. But
324 <     * because blocked joins are typically dependent, we don't
325 <     * necessarily need or want one-to-one replacement. Instead, we
326 <     * use a combination of heuristics that adds threads only when the
327 <     * pool appears to be approaching starvation.  These effectively
328 <     * reduce churn at the price of systematically undershooting
329 <     * target parallelism when many threads are blocked.  However,
330 <     * biasing toward undeshooting partially compensates for the above
331 <     * mechanics to suspend extra threads, that normally lead to
332 <     * overshoot because we can only suspend workers in-between
333 <     * top-level actions. It also better copes with the fact that some
334 <     * of the methods in this class tend to never become compiled (but
335 <     * are interpreted), so some components of the entire set of
336 <     * controls might execute many times faster than others. And
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       *
# Line 308 | 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 a few
355 <     * other coding oddities that help some methods perform reasonably
356 <     * even when interpreted (not compiled).
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 380 | 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 423 | 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_INDEX_MASK = (1L << EVENT_COUNT_SHIFT)-1L;
499 >    private static final int EVENT_COUNT_SHIFT = 32;
500 >    private static final int WAITER_ID_MASK    = (1 << 16) - 1;
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 461 | 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;
464    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 470 | 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
474 <     * configuration parameters.
553 >     * running state.
554       */
555      private volatile int workerCounts;
556  
# Line 503 | Line 582 | public class ForkJoinPool extends Abstra
582       */
583      private final int poolNumber;
584  
585 <    // utilities for updating fields
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;
593          do {} while (!UNSAFE.compareAndSwapInt(this, workerCountsOffset,
594 <                                               c = workerCounts,
594 >                                               c = workerCounts,
595                                                 c + ONE_RUNNING));
596      }
597 <    
597 >
598 >    /**
599 >     * Tries to increment running count part of workerCounts.
600 >     */
601 >    final boolean tryIncrementRunningCount() {
602 >        int c;
603 >        return UNSAFE.compareAndSwapInt(this, workerCountsOffset,
604 >                                        c = workerCounts,
605 >                                        c + ONE_RUNNING);
606 >    }
607 >
608      /**
609 <     * Tries to decrement running count unless already zero
609 >     * Tries to decrement running count unless already zero.
610       */
611      final boolean tryDecrementRunningCount() {
612          int wc = workerCounts;
# Line 527 | Line 617 | public class ForkJoinPool extends Abstra
617      }
618  
619      /**
620 <     * Tries incrementing active count; fails on contention.
621 <     * Called by workers before executing tasks.
620 >     * Forces decrement of encoded workerCounts, awaiting nonzero if
621 >     * (rarely) necessary when other count updates lag.
622       *
623 <     * @return true on success
623 >     * @param dr -- either zero or ONE_RUNNING
624 >     * @param dt -- either zero or ONE_TOTAL
625       */
626 <    final boolean tryIncrementActiveCount() {
627 <        int c;
628 <        return UNSAFE.compareAndSwapInt(this, runStateOffset,
629 <                                        c = runState, c + ONE_ACTIVE);
626 >    private void decrementWorkerCounts(int dr, int dt) {
627 >        for (;;) {
628 >            int wc = workerCounts;
629 >            if ((wc & RUNNING_COUNT_MASK)  - dr < 0 ||
630 >                (wc >>> TOTAL_COUNT_SHIFT) - dt < 0) {
631 >                if ((runState & TERMINATED) != 0)
632 >                    return; // lagging termination on a backout
633 >                Thread.yield();
634 >            }
635 >            if (UNSAFE.compareAndSwapInt(this, workerCountsOffset,
636 >                                         wc, wc - (dr + dt)))
637 >                return;
638 >        }
639      }
640  
641      /**
# Line 545 | Line 645 | public class ForkJoinPool extends Abstra
645      final boolean tryDecrementActiveCount() {
646          int c;
647          return UNSAFE.compareAndSwapInt(this, runStateOffset,
648 <                                        c = runState, c - ONE_ACTIVE);
648 >                                        c = runState, c - 1);
649      }
650  
651      /**
# Line 574 | Line 674 | public class ForkJoinPool extends Abstra
674          lock.lock();
675          try {
676              ForkJoinWorkerThread[] ws = workers;
677 <            int nws = ws.length;
678 <            if (k < 0 || k >= nws || ws[k] != null) {
679 <                for (k = 0; k < nws && ws[k] != null; ++k)
677 >            int n = ws.length;
678 >            if (k < 0 || k >= n || ws[k] != null) {
679 >                for (k = 0; k < n && ws[k] != null; ++k)
680                      ;
681 <                if (k == nws)
682 <                    ws = Arrays.copyOf(ws, nws << 1);
681 >                if (k == n)
682 >                    ws = workers = Arrays.copyOf(ws, n << 1);
683              }
684              ws[k] = w;
685 <            workers = ws; // volatile array write ensures slot visibility
685 >            int c = eventCount; // advance event count to ensure visibility
686 >            UNSAFE.compareAndSwapInt(this, eventCountOffset, c, c+1);
687          } finally {
688              lock.unlock();
689          }
# Line 590 | Line 691 | public class ForkJoinPool extends Abstra
691      }
692  
693      /**
694 <     * Nulls out record of worker in workers array
694 >     * Nulls out record of worker in workers array.
695       */
696      private void forgetWorker(ForkJoinWorkerThread w) {
697          int idx = w.poolIndex;
698 <        // Locking helps method recordWorker avoid unecessary expansion
698 >        // Locking helps method recordWorker avoid unnecessary expansion
699          final ReentrantLock lock = this.workerLock;
700          lock.lock();
701          try {
# Line 606 | Line 707 | public class ForkJoinPool extends Abstra
707          }
708      }
709  
609    // adding and removing workers
610
710      /**
711 <     * Tries to create and add new worker. Assumes that worker counts
712 <     * are already updated to accommodate the worker, so adjusts on
713 <     * failure.
711 >     * Final callback from terminating worker.  Removes record of
712 >     * worker from array, and adjusts counts. If pool is shutting
713 >     * down, tries to complete termination.
714       *
715 <     * @return new worker or null if creation failed
715 >     * @param w the worker
716       */
717 <    private ForkJoinWorkerThread addWorker() {
718 <        ForkJoinWorkerThread w = null;
719 <        try {
720 <            w = factory.newThread(this);
721 <        } finally { // Adjust on either null or exceptional factory return
722 <            if (w == null) {
624 <                onWorkerCreationFailure();
625 <                return null;
626 <            }
627 <        }
628 <        w.start(recordWorker(w), ueh);
629 <        return w;
717 >    final void workerTerminated(ForkJoinWorkerThread w) {
718 >        forgetWorker(w);
719 >        decrementWorkerCounts(w.isTrimmed() ? 0 : ONE_RUNNING, ONE_TOTAL);
720 >        while (w.stealCount != 0) // collect final count
721 >            tryAccumulateStealCount(w);
722 >        tryTerminate(false);
723      }
724  
725 +    // Waiting for and signalling events
726 +
727      /**
728 <     * Adjusts counts upon failure to create worker
728 >     * Releases workers blocked on a count not equal to current count.
729 >     * Normally called after precheck that eventWaiters isn't zero to
730 >     * avoid wasted array checks. Gives up upon a change in count or
731 >     * upon releasing four workers, letting others take over.
732       */
733 <    private void onWorkerCreationFailure() {
734 <        for (;;) {
735 <            int wc = workerCounts;
736 <            if ((wc >>> TOTAL_COUNT_SHIFT) > 0 &&
737 <                UNSAFE.compareAndSwapInt(this, workerCountsOffset,
738 <                                         wc, wc - (ONE_RUNNING|ONE_TOTAL)))
733 >    private void releaseEventWaiters() {
734 >        ForkJoinWorkerThread[] ws = workers;
735 >        int n = ws.length;
736 >        long h = eventWaiters;
737 >        int ec = eventCount;
738 >        int releases = 4;
739 >        ForkJoinWorkerThread w; int id;
740 >        while ((id = (((int)h) & WAITER_ID_MASK) - 1) >= 0 &&
741 >               (int)(h >>> EVENT_COUNT_SHIFT) != ec &&
742 >               id < n && (w = ws[id]) != null) {
743 >            if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
744 >                                          h,  w.nextWaiter)) {
745 >                LockSupport.unpark(w);
746 >                if (--releases == 0)
747 >                    break;
748 >            }
749 >            if (eventCount != ec)
750                  break;
751 +            h = eventWaiters;
752          }
643        tryTerminate(false); // in case of failure during shutdown
753      }
754  
755      /**
756 <     * Create enough total workers to establish target parallelism,
757 <     * giving up if terminating or addWorker fails
756 >     * Tries to advance eventCount and releases waiters. Called only
757 >     * from workers.
758       */
759 <    private void ensureEnoughTotalWorkers() {
760 <        int wc;
761 <        while (((wc = workerCounts) >>> TOTAL_COUNT_SHIFT) < parallelism &&
762 <               runState < TERMINATING) {
763 <            if ((UNSAFE.compareAndSwapInt(this, workerCountsOffset,
655 <                                          wc, wc + (ONE_RUNNING|ONE_TOTAL)) &&
656 <                 addWorker() == null))
657 <                break;
658 <        }
759 >    final void signalWork() {
760 >        int c; // try to increment event count -- CAS failure OK
761 >        UNSAFE.compareAndSwapInt(this, eventCountOffset, c = eventCount, c+1);
762 >        if (eventWaiters != 0L)
763 >            releaseEventWaiters();
764      }
765  
766      /**
767 <     * Final callback from terminating worker.  Removes record of
768 <     * worker from array, and adjusts counts. If pool is shutting
664 <     * down, tries to complete terminatation, else possibly replaces
665 <     * the worker.
767 >     * Adds the given worker to event queue and blocks until
768 >     * terminating or event count advances from the given value
769       *
770 <     * @param w the worker
770 >     * @param w the calling worker thread
771 >     * @param ec the count
772       */
773 <    final void workerTerminated(ForkJoinWorkerThread w) {
774 <        if (w.active) { // force inactive
775 <            w.active = false;
776 <            do {} while (!tryDecrementActiveCount());
777 <        }
778 <        forgetWorker(w);
779 <
780 <        // Decrement total count, and if was running, running count
781 <        // Spin (waiting for other updates) if either would be negative
782 <        int nr = w.isTrimmed() ? 0 : ONE_RUNNING;
679 <        int unit = ONE_TOTAL + nr;
680 <        for (;;) {
681 <            int wc = workerCounts;
682 <            int rc = wc & RUNNING_COUNT_MASK;
683 <            if (rc - nr < 0 || (wc >>> TOTAL_COUNT_SHIFT) == 0)
684 <                Thread.yield(); // back off if waiting for other updates
685 <            else if (UNSAFE.compareAndSwapInt(this, workerCountsOffset,
686 <                                              wc, wc - unit))
773 >    private void eventSync(ForkJoinWorkerThread w, int ec) {
774 >        long nh = (((long)ec) << EVENT_COUNT_SHIFT) | ((long)(w.poolIndex+1));
775 >        long h;
776 >        while ((runState < SHUTDOWN || !tryTerminate(false)) &&
777 >               (((int)(h = eventWaiters) & WAITER_ID_MASK) == 0 ||
778 >                (int)(h >>> EVENT_COUNT_SHIFT) == ec) &&
779 >               eventCount == ec) {
780 >            if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
781 >                                          w.nextWaiter = h, nh)) {
782 >                awaitEvent(w, ec);
783                  break;
784 +            }
785          }
689
690        accumulateStealCount(w); // collect final count
691        if (!tryTerminate(false))
692            ensureEnoughTotalWorkers();
786      }
787  
695    // Waiting for and signalling events
696
788      /**
789 <     * Releases workers blocked on a count not equal to current count.
789 >     * Blocks the given worker (that has already been entered as an
790 >     * event waiter) until terminating or event count advances from
791 >     * the given value. The oldest (first) waiter uses a timed wait to
792 >     * occasionally one-by-one shrink the number of workers (to a
793 >     * minimum of one) if the pool has not been used for extended
794 >     * periods.
795 >     *
796 >     * @param w the calling worker thread
797 >     * @param ec the count
798       */
799 <    private void releaseWaiters() {
800 <        long top;
801 <        int id;
802 <        while ((id = (int)((top = eventWaiters) & WAITER_INDEX_MASK)) > 0 &&
803 <               (int)(top >>> EVENT_COUNT_SHIFT) != eventCount) {
804 <            ForkJoinWorkerThread[] ws = workers;
805 <            ForkJoinWorkerThread w;
806 <            if (ws.length >= id && (w = ws[id - 1]) != null &&
807 <                UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
808 <                                          top, w.nextWaiter))
809 <                LockSupport.unpark(w);
799 >    private void awaitEvent(ForkJoinWorkerThread w, int ec) {
800 >        while (eventCount == ec) {
801 >            if (tryAccumulateStealCount(w)) { // transfer while idle
802 >                boolean untimed = (w.nextWaiter != 0L ||
803 >                                   (workerCounts & RUNNING_COUNT_MASK) <= 1);
804 >                long startTime = untimed ? 0 : System.nanoTime();
805 >                Thread.interrupted();         // clear/ignore interrupt
806 >                if (w.isTerminating() || eventCount != ec)
807 >                    break;                    // recheck after clear
808 >                if (untimed)
809 >                    LockSupport.park(w);
810 >                else {
811 >                    LockSupport.parkNanos(w, SHRINK_RATE_NANOS);
812 >                    if (eventCount != ec || w.isTerminating())
813 >                        break;
814 >                    if (System.nanoTime() - startTime >= SHRINK_RATE_NANOS)
815 >                        tryShutdownUnusedWorker(ec);
816 >                }
817 >            }
818          }
819      }
820  
821 +    // Maintaining parallelism
822 +
823      /**
824 <     * Ensures eventCount on exit is different (mod 2^32) than on
716 <     * entry and wakes up all waiters
824 >     * Pushes worker onto the spare stack.
825       */
826 <    private void signalEvent() {
827 <        int c;
828 <        do {} while (!UNSAFE.compareAndSwapInt(this, eventCountOffset,
829 <                                               c = eventCount, c+1));
722 <        releaseWaiters();
826 >    final void pushSpare(ForkJoinWorkerThread w) {
827 >        int ns = (++w.spareCount << SPARE_COUNT_SHIFT) | (w.poolIndex + 1);
828 >        do {} while (!UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
829 >                                               w.nextSpare = spareWaiters,ns));
830      }
831  
832      /**
833 <     * Advances eventCount and releases waiters until interference by
834 <     * other releasing threads is detected.
833 >     * Tries (once) to resume a spare if the number of running
834 >     * threads is less than target.
835       */
836 <    final void signalWork() {
837 <        // EventCount CAS failures are OK -- any change in count suffices.
838 <        int ec;
839 <        UNSAFE.compareAndSwapInt(this, eventCountOffset, ec=eventCount, ec+1);
840 <        outer:for (;;) {
841 <            long top = eventWaiters;
842 <            ec = eventCount;
843 <            for (;;) {
844 <                ForkJoinWorkerThread[] ws; ForkJoinWorkerThread w;
845 <                int id = (int)(top & WAITER_INDEX_MASK);
846 <                if (id <= 0 || (int)(top >>> EVENT_COUNT_SHIFT) == ec)
847 <                    return;
848 <                if ((ws = workers).length < id || (w = ws[id - 1]) == null ||
849 <                    !UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
850 <                                               top, top = w.nextWaiter))
851 <                    continue outer;      // possibly stale; reread
836 >    private void tryResumeSpare() {
837 >        int sw, id;
838 >        ForkJoinWorkerThread[] ws = workers;
839 >        int n = ws.length;
840 >        ForkJoinWorkerThread w;
841 >        if ((sw = spareWaiters) != 0 &&
842 >            (id = (sw & SPARE_ID_MASK) - 1) >= 0 &&
843 >            id < n && (w = ws[id]) != null &&
844 >            (runState >= TERMINATING ||
845 >             (workerCounts & RUNNING_COUNT_MASK) < parallelism) &&
846 >            spareWaiters == sw &&
847 >            UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
848 >                                     sw, w.nextSpare)) {
849 >            int c; // increment running count before resume
850 >            do {} while (!UNSAFE.compareAndSwapInt
851 >                         (this, workerCountsOffset,
852 >                          c = workerCounts, c + ONE_RUNNING));
853 >            if (w.tryUnsuspend())
854                  LockSupport.unpark(w);
855 <                if (top != eventWaiters) // let someone else take over
856 <                    return;
748 <            }
855 >            else   // back out if w was shutdown
856 >                decrementWorkerCounts(ONE_RUNNING, 0);
857          }
858      }
859  
860      /**
861 <     * If worker is inactive, blocks until terminating or event count
862 <     * advances from last value held by worker; in any case helps
863 <     * release others.
864 <     *
865 <     * @param w the calling worker thread
861 >     * Tries to increase the number of running workers if below target
862 >     * parallelism: If a spare exists tries to resume it via
863 >     * tryResumeSpare.  Otherwise, if not enough total workers or all
864 >     * existing workers are busy, adds a new worker. In all cases also
865 >     * helps wake up releasable workers waiting for work.
866       */
867 <    private void eventSync(ForkJoinWorkerThread w) {
868 <        if (!w.active) {
869 <            int prev = w.lastEventCount;
870 <            long nextTop = (((long)prev << EVENT_COUNT_SHIFT) |
871 <                            ((long)(w.poolIndex + 1)));
872 <            long top;
873 <            while ((runState < SHUTDOWN || !tryTerminate(false)) &&
874 <                   (((int)(top = eventWaiters) & WAITER_INDEX_MASK) == 0 ||
875 <                    (int)(top >>> EVENT_COUNT_SHIFT) == prev) &&
876 <                   eventCount == prev) {
877 <                if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
878 <                                              w.nextWaiter = top, nextTop)) {
879 <                    accumulateStealCount(w); // transfer steals while idle
880 <                    Thread.interrupted();    // clear/ignore interrupt
881 <                    while (eventCount == prev)
882 <                        w.doPark();
867 >    private void helpMaintainParallelism() {
868 >        int pc = parallelism;
869 >        int wc, rs, tc;
870 >        while (((wc = workerCounts) & RUNNING_COUNT_MASK) < pc &&
871 >               (rs = runState) < TERMINATING) {
872 >            if (spareWaiters != 0)
873 >                tryResumeSpare();
874 >            else if ((tc = wc >>> TOTAL_COUNT_SHIFT) >= MAX_WORKERS ||
875 >                     (tc >= pc && (rs & ACTIVE_COUNT_MASK) != tc))
876 >                break;   // enough total
877 >            else if (runState == rs && workerCounts == wc &&
878 >                     UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
879 >                                              wc + (ONE_RUNNING|ONE_TOTAL))) {
880 >                ForkJoinWorkerThread w = null;
881 >                Throwable fail = null;
882 >                try {
883 >                    w = factory.newThread(this);
884 >                } catch (Throwable ex) {
885 >                    fail = ex;
886 >                }
887 >                if (w == null) { // null or exceptional factory return
888 >                    decrementWorkerCounts(ONE_RUNNING, ONE_TOTAL);
889 >                    tryTerminate(false); // handle failure during shutdown
890 >                    // If originating from an external caller,
891 >                    // propagate exception, else ignore
892 >                    if (fail != null && runState < TERMINATING &&
893 >                        !(Thread.currentThread() instanceof
894 >                          ForkJoinWorkerThread))
895 >                        UNSAFE.throwException(fail);
896                      break;
897                  }
898 +                w.start(recordWorker(w), ueh);
899 +                if ((workerCounts >>> TOTAL_COUNT_SHIFT) >= pc)
900 +                    break; // add at most one unless total below target
901 +            }
902 +        }
903 +        if (eventWaiters != 0L)
904 +            releaseEventWaiters();
905 +    }
906 +
907 +    /**
908 +     * Callback from the oldest waiter in awaitEvent waking up after a
909 +     * period of non-use. If all workers are idle, tries (once) to
910 +     * shutdown an event waiter or a spare, if one exists. Note that
911 +     * we don't need CAS or locks here because the method is called
912 +     * only from one thread occasionally waking (and even misfires are
913 +     * OK). Note that until the shutdown worker fully terminates,
914 +     * workerCounts will overestimate total count, which is tolerable.
915 +     *
916 +     * @param ec the event count waited on by caller (to abort
917 +     * attempt if count has since changed).
918 +     */
919 +    private void tryShutdownUnusedWorker(int ec) {
920 +        if (runState == 0 && eventCount == ec) { // only trigger if all idle
921 +            ForkJoinWorkerThread[] ws = workers;
922 +            int n = ws.length;
923 +            ForkJoinWorkerThread w = null;
924 +            boolean shutdown = false;
925 +            int sw;
926 +            long h;
927 +            if ((sw = spareWaiters) != 0) { // prefer killing spares
928 +                int id = (sw & SPARE_ID_MASK) - 1;
929 +                if (id >= 0 && id < n && (w = ws[id]) != null &&
930 +                    UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
931 +                                             sw, w.nextSpare))
932 +                    shutdown = true;
933 +            }
934 +            else if ((h = eventWaiters) != 0L) {
935 +                long nh;
936 +                int id = (((int)h) & WAITER_ID_MASK) - 1;
937 +                if (id >= 0 && id < n && (w = ws[id]) != null &&
938 +                    (nh = w.nextWaiter) != 0L && // keep at least one worker
939 +                    UNSAFE.compareAndSwapLong(this, eventWaitersOffset, h, nh))
940 +                    shutdown = true;
941 +            }
942 +            if (w != null && shutdown) {
943 +                w.shutdown();
944 +                LockSupport.unpark(w);
945              }
778            w.lastEventCount = eventCount;
946          }
947 <        releaseWaiters();
947 >        releaseEventWaiters(); // in case of interference
948      }
949  
950      /**
951       * Callback from workers invoked upon each top-level action (i.e.,
952 <     * stealing a task or taking a submission and running
953 <     * it). Performs one or both of the following:
952 >     * stealing a task or taking a submission and running it).
953 >     * Performs one or more of the following:
954       *
955 <     * * If the worker cannot find work, updates its active status to
956 <     * inactive and updates activeCount unless there is contention, in
957 <     * which case it may try again (either in this or a subsequent
958 <     * call).  Additionally, awaits the next task event and/or helps
959 <     * wake up other releasable waiters.
960 <     *
961 <     * * If there are too many running threads, suspends this worker
962 <     * (first forcing inactivation if necessary).  If it is not
963 <     * resumed before a keepAlive elapses, the worker may be "trimmed"
964 <     * -- killed while suspended within suspendAsSpare. Otherwise,
965 <     * upon resume it rechecks to make sure that it is still needed.
955 >     * 1. If the worker is active and either did not run a task
956 >     *    or there are too many workers, try to set its active status
957 >     *    to inactive and update activeCount. On contention, we may
958 >     *    try again in this or a subsequent call.
959 >     *
960 >     * 2. If not enough total workers, help create some.
961 >     *
962 >     * 3. If there are too many running workers, suspend this worker
963 >     *    (first forcing inactive if necessary).  If it is not needed,
964 >     *    it may be shutdown while suspended (via
965 >     *    tryShutdownUnusedWorker).  Otherwise, upon resume it
966 >     *    rechecks running thread count and need for event sync.
967 >     *
968 >     * 4. If worker did not run a task, await the next task event via
969 >     *    eventSync if necessary (first forcing inactivation), upon
970 >     *    which the worker may be shutdown via
971 >     *    tryShutdownUnusedWorker.  Otherwise, help release any
972 >     *    existing event waiters that are now releasable,
973       *
974       * @param w the worker
975 <     * @param worked false if the worker scanned for work but didn't
802 <     * find any (in which case it may block waiting for work).
975 >     * @param ran true if worker ran a task since last call to this method
976       */
977 <    final void preStep(ForkJoinWorkerThread w, boolean worked) {
977 >    final void preStep(ForkJoinWorkerThread w, boolean ran) {
978 >        int wec = w.lastEventCount;
979          boolean active = w.active;
980 <        boolean inactivate = !worked & active;
981 <        for (;;) {
982 <            if (inactivate) {
983 <                int rs = runState;
984 <                if (UNSAFE.compareAndSwapInt(this, runStateOffset,
985 <                                             rs, rs - ONE_ACTIVE))
986 <                    inactivate = active = w.active = false;
980 >        boolean inactivate = false;
981 >        int pc = parallelism;
982 >        while (w.runState == 0) {
983 >            int rs = runState;
984 >            if (rs >= TERMINATING) {           // propagate shutdown
985 >                w.shutdown();
986 >                break;
987              }
988 <            int wc = workerCounts;
989 <            if ((wc & RUNNING_COUNT_MASK) <= parallelism) {
990 <                if (!worked)
991 <                    eventSync(w);
992 <                return;
988 >            if ((inactivate || (active && (rs & ACTIVE_COUNT_MASK) >= pc)) &&
989 >                UNSAFE.compareAndSwapInt(this, runStateOffset, rs, --rs)) {
990 >                inactivate = active = w.active = false;
991 >                if (rs == SHUTDOWN) {          // all inactive and shut down
992 >                    tryTerminate(false);
993 >                    continue;
994 >                }
995 >            }
996 >            int wc = workerCounts;             // try to suspend as spare
997 >            if ((wc & RUNNING_COUNT_MASK) > pc) {
998 >                if (!(inactivate |= active) && // must inactivate to suspend
999 >                    workerCounts == wc &&
1000 >                    UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1001 >                                             wc, wc - ONE_RUNNING))
1002 >                    w.suspendAsSpare();
1003 >            }
1004 >            else if ((wc >>> TOTAL_COUNT_SHIFT) < pc)
1005 >                helpMaintainParallelism();     // not enough workers
1006 >            else if (ran)
1007 >                break;
1008 >            else {
1009 >                long h = eventWaiters;
1010 >                int ec = eventCount;
1011 >                if (h != 0L && (int)(h >>> EVENT_COUNT_SHIFT) != ec)
1012 >                    releaseEventWaiters();     // release others before waiting
1013 >                else if (ec != wec) {
1014 >                    w.lastEventCount = ec;     // no need to wait
1015 >                    break;
1016 >                }
1017 >                else if (!(inactivate |= active))
1018 >                    eventSync(w, wec);         // must inactivate before sync
1019              }
820            if (!(inactivate |= active) &&  // must inactivate to suspend
821                UNSAFE.compareAndSwapInt(this, workerCountsOffset,
822                                         wc, wc - ONE_RUNNING) &&
823                !w.suspendAsSpare())        // false if trimmed
824                return;
1020          }
1021      }
1022  
1023      /**
1024 <     * Tries to decrement running count, and if so, possibly creates
1025 <     * or resumes compensating threads before blocking on task joinMe.
831 <     * This code is sprawled out with manual inlining to evade some
832 <     * JIT oddities.
1024 >     * Helps and/or blocks awaiting join of the given task.
1025 >     * See above for explanation.
1026       *
1027       * @param joinMe the task to join
1028 <     * @return task status on exit
1029 <     */
1030 <    final int tryAwaitJoin(ForkJoinTask<?> joinMe) {
1031 <        int cw = workerCounts; // read now to spoil CAS if counts change as ...
1032 <        releaseWaiters();      // ... a byproduct of releaseWaiters
1033 <        int stat = joinMe.status;
1034 <        if (stat >= 0 && // inline variant of tryDecrementRunningCount
1035 <            (cw & RUNNING_COUNT_MASK) > 0 &&
1036 <            UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1037 <                                     cw, cw - ONE_RUNNING)) {
1038 <            int pc = parallelism;
1039 <            int scans = 0;  // to require confirming passes to add threads
1040 <            outer: while ((workerCounts & RUNNING_COUNT_MASK) < pc) {
1041 <                if ((stat = joinMe.status) < 0)
1042 <                    break;
1043 <                ForkJoinWorkerThread spare = null;
1044 <                ForkJoinWorkerThread[] ws = workers;
1045 <                int nws = ws.length;
1046 <                for (int i = 0; i < nws; ++i) {
1047 <                    ForkJoinWorkerThread w = ws[i];
1048 <                    if (w != null && w.isSuspended()) {
1049 <                        spare = w;
1050 <                        break;
1051 <                    }
1028 >     * @param worker the current worker thread
1029 >     * @param timed true if wait should time out
1030 >     * @param nanos timeout value if timed
1031 >     */
1032 >    final void awaitJoin(ForkJoinTask<?> joinMe, ForkJoinWorkerThread worker,
1033 >                         boolean timed, long nanos) {
1034 >        long startTime = timed ? System.nanoTime() : 0L;
1035 >        int retries = 2 + (parallelism >> 2); // #helpJoins before blocking
1036 >        boolean running = true;               // false when count decremented
1037 >        while (joinMe.status >= 0) {
1038 >            if (runState >= TERMINATING) {
1039 >                joinMe.cancelIgnoringExceptions();
1040 >                break;
1041 >            }
1042 >            running = worker.helpJoinTask(joinMe, running);
1043 >            if (joinMe.status < 0)
1044 >                break;
1045 >            if (retries > 0) {
1046 >                --retries;
1047 >                continue;
1048 >            }
1049 >            int wc = workerCounts;
1050 >            if ((wc & RUNNING_COUNT_MASK) != 0) {
1051 >                if (running) {
1052 >                    if (!UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1053 >                                                  wc, wc - ONE_RUNNING))
1054 >                        continue;
1055 >                    running = false;
1056                  }
1057 <                if ((stat = joinMe.status) < 0) // recheck to narrow race
1058 <                    break;
1059 <                int wc = workerCounts;
1060 <                int rc = wc & RUNNING_COUNT_MASK;
1061 <                if (rc >= pc)
1062 <                    break;
1063 <                if (spare != null) {
1064 <                    if (spare.tryUnsuspend()) {
868 <                        int c; // inline incrementRunningCount
869 <                        do {} while (!UNSAFE.compareAndSwapInt
870 <                                     (this, workerCountsOffset,
871 <                                      c = workerCounts, c + ONE_RUNNING));
872 <                        LockSupport.unpark(spare);
873 <                        break;
1057 >                long h = eventWaiters;
1058 >                if (h != 0L && (int)(h >>> EVENT_COUNT_SHIFT) != eventCount)
1059 >                    releaseEventWaiters();
1060 >                if ((workerCounts & RUNNING_COUNT_MASK) != 0) {
1061 >                    long ms; int ns;
1062 >                    if (!timed) {
1063 >                        ms = JOIN_TIMEOUT_MILLIS;
1064 >                        ns = 0;
1065                      }
1066 <                    continue;
1067 <                }
1068 <                int tc = wc >>> TOTAL_COUNT_SHIFT;
1069 <                int sc = tc - pc;
1070 <                if (rc > 0) {
1071 <                    int p = pc;
1072 <                    int s = sc;
1073 <                    while (s-- >= 0) { // try keeping 3/4 live
1074 <                        if (rc > (p -= (p >>> 2) + 1))
1075 <                            break outer;
1066 >                    else { // at most JOIN_TIMEOUT_MILLIS per wait
1067 >                        long nt = nanos - (System.nanoTime() - startTime);
1068 >                        if (nt <= 0L)
1069 >                            break;
1070 >                        ms = nt / 1000000;
1071 >                        if (ms > JOIN_TIMEOUT_MILLIS) {
1072 >                            ms = JOIN_TIMEOUT_MILLIS;
1073 >                            ns = 0;
1074 >                        }
1075 >                        else
1076 >                            ns = (int) (nt % 1000000);
1077                      }
1078 +                    joinMe.internalAwaitDone(ms, ns);
1079                  }
1080 <                if (scans++ > sc && tc < MAX_THREADS &&
888 <                    UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
889 <                                             wc + (ONE_RUNNING|ONE_TOTAL))) {
890 <                    addWorker();
1080 >                if (joinMe.status < 0)
1081                      break;
892                }
1082              }
1083 <            if (stat >= 0)
1084 <                stat = joinMe.internalAwaitDone();
1085 <            int c; // inline incrementRunningCount
1083 >            helpMaintainParallelism();
1084 >        }
1085 >        if (!running) {
1086 >            int c;
1087              do {} while (!UNSAFE.compareAndSwapInt
1088                           (this, workerCountsOffset,
1089                            c = workerCounts, c + ONE_RUNNING));
1090          }
901        return stat;
1091      }
1092  
1093      /**
1094 <     * Same idea as (and mostly pasted from) tryAwaitJoin, but
906 <     * self-contained
1094 >     * Same idea as awaitJoin, but no helping, retries, or timeouts.
1095       */
1096      final void awaitBlocker(ManagedBlocker blocker)
1097          throws InterruptedException {
1098 <        for (;;) {
911 <            if (blocker.isReleasable())
912 <                return;
913 <            int cw = workerCounts;
914 <            releaseWaiters();
915 <            if ((cw & RUNNING_COUNT_MASK) > 0 &&
916 <                UNSAFE.compareAndSwapInt(this, workerCountsOffset,
917 <                                         cw, cw - ONE_RUNNING))
918 <                break;
919 <        }
920 <        boolean done = false;
921 <        int pc = parallelism;
922 <        int scans = 0;
923 <        outer: while ((workerCounts & RUNNING_COUNT_MASK) < pc) {
924 <            if (done = blocker.isReleasable())
925 <                break;
926 <            ForkJoinWorkerThread spare = null;
927 <            ForkJoinWorkerThread[] ws = workers;
928 <            int nws = ws.length;
929 <            for (int i = 0; i < nws; ++i) {
930 <                ForkJoinWorkerThread w = ws[i];
931 <                if (w != null && w.isSuspended()) {
932 <                    spare = w;
933 <                    break;
934 <                }
935 <            }
936 <            if (done = blocker.isReleasable())
937 <                break;
1098 >        while (!blocker.isReleasable()) {
1099              int wc = workerCounts;
1100 <            int rc = wc & RUNNING_COUNT_MASK;
1101 <            if (rc >= pc)
1102 <                break;
1103 <            if (spare != null) {
1104 <                if (spare.tryUnsuspend()) {
1100 >            if ((wc & RUNNING_COUNT_MASK) == 0)
1101 >                helpMaintainParallelism();
1102 >            else if (UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1103 >                                              wc, wc - ONE_RUNNING)) {
1104 >                try {
1105 >                    while (!blocker.isReleasable()) {
1106 >                        long h = eventWaiters;
1107 >                        if (h != 0L &&
1108 >                            (int)(h >>> EVENT_COUNT_SHIFT) != eventCount)
1109 >                            releaseEventWaiters();
1110 >                        else if ((workerCounts & RUNNING_COUNT_MASK) == 0 &&
1111 >                                 runState < TERMINATING)
1112 >                            helpMaintainParallelism();
1113 >                        else if (blocker.block())
1114 >                            break;
1115 >                    }
1116 >                } finally {
1117                      int c;
1118                      do {} while (!UNSAFE.compareAndSwapInt
1119                                   (this, workerCountsOffset,
1120                                    c = workerCounts, c + ONE_RUNNING));
948                    LockSupport.unpark(spare);
949                    break;
950                }
951                continue;
952            }
953            int tc = wc >>> TOTAL_COUNT_SHIFT;
954            int sc = tc - pc;
955            if (rc > 0) {
956                int p = pc;
957                int s = sc;
958                while (s-- >= 0) {
959                    if (rc > (p -= (p >>> 2) + 1))
960                        break outer;
1121                  }
962            }
963            if (scans++ > sc && tc < MAX_THREADS &&
964                UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
965                                         wc + (ONE_RUNNING|ONE_TOTAL))) {
966                addWorker();
1122                  break;
1123              }
1124          }
1125 <        try {
971 <            if (!done)
972 <                do {} while (!blocker.isReleasable() &&
973 <                             !blocker.block());
974 <        } finally {
975 <            int c;
976 <            do {} while (!UNSAFE.compareAndSwapInt
977 <                         (this, workerCountsOffset,
978 <                          c = workerCounts, c + ONE_RUNNING));
979 <        }
980 <    }  
1125 >    }
1126  
1127      /**
1128       * Possibly initiates and/or completes termination.
# Line 1000 | Line 1145 | public class ForkJoinPool extends Abstra
1145          // Finish now if all threads terminated; else in some subsequent call
1146          if ((workerCounts >>> TOTAL_COUNT_SHIFT) == 0) {
1147              advanceRunLevel(TERMINATED);
1148 <            termination.arrive();
1148 >            termination.forceTermination();
1149          }
1150          return true;
1151      }
1152  
1153      /**
1154       * Actions on transition to TERMINATING
1155 +     *
1156 +     * Runs up to four passes through workers: (0) shutting down each
1157 +     * (without waking up if parked) to quickly spread notifications
1158 +     * without unnecessary bouncing around event queues etc (1) wake
1159 +     * up and help cancel tasks (2) interrupt (3) mop up races with
1160 +     * interrupted workers
1161       */
1162      private void startTerminating() {
1163 <        for (int i = 0; i < 2; ++i) { // twice to mop up newly created workers
1164 <            cancelSubmissions();
1165 <            shutdownWorkers();
1166 <            cancelWorkerTasks();
1167 <            signalEvent();
1168 <            interruptWorkers();
1163 >        cancelSubmissions();
1164 >        for (int passes = 0; passes < 4 && workerCounts != 0; ++passes) {
1165 >            int c; // advance event count
1166 >            UNSAFE.compareAndSwapInt(this, eventCountOffset,
1167 >                                     c = eventCount, c+1);
1168 >            eventWaiters = 0L; // clobber lists
1169 >            spareWaiters = 0;
1170 >            for (ForkJoinWorkerThread w : workers) {
1171 >                if (w != null) {
1172 >                    w.shutdown();
1173 >                    if (passes > 0 && !w.isTerminated()) {
1174 >                        w.cancelTasks();
1175 >                        LockSupport.unpark(w);
1176 >                        if (passes > 1 && !w.isInterrupted()) {
1177 >                            try {
1178 >                                w.interrupt();
1179 >                            } catch (SecurityException ignore) {
1180 >                            }
1181 >                        }
1182 >                    }
1183 >                }
1184 >            }
1185          }
1186      }
1187  
1188      /**
1189 <     * Clear out and cancel submissions, ignoring exceptions
1189 >     * Clears out and cancels submissions, ignoring exceptions.
1190       */
1191      private void cancelSubmissions() {
1192          ForkJoinTask<?> task;
# Line 1031 | Line 1198 | public class ForkJoinPool extends Abstra
1198          }
1199      }
1200  
1034    /**
1035     * Sets all worker run states to at least shutdown,
1036     * also resuming suspended workers
1037     */
1038    private void shutdownWorkers() {
1039        ForkJoinWorkerThread[] ws = workers;
1040        int nws = ws.length;
1041        for (int i = 0; i < nws; ++i) {
1042            ForkJoinWorkerThread w = ws[i];
1043            if (w != null)
1044                w.shutdown();
1045        }
1046    }
1047
1048    /**
1049     * Clears out and cancels all locally queued tasks
1050     */
1051    private void cancelWorkerTasks() {
1052        ForkJoinWorkerThread[] ws = workers;
1053        int nws = ws.length;
1054        for (int i = 0; i < nws; ++i) {
1055            ForkJoinWorkerThread w = ws[i];
1056            if (w != null)
1057                w.cancelTasks();
1058        }
1059    }
1060
1061    /**
1062     * Unsticks all workers blocked on joins etc
1063     */
1064    private void interruptWorkers() {
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 && !w.isTerminated()) {
1070                try {
1071                    w.interrupt();
1072                } catch (SecurityException ignore) {
1073                }
1074            }
1075        }
1076    }
1077
1201      // misc support for ForkJoinWorkerThread
1202  
1203      /**
1204 <     * Returns pool number
1204 >     * Returns pool number.
1205       */
1206      final int getPoolNumber() {
1207          return poolNumber;
1208      }
1209  
1210      /**
1211 <     * Accumulates steal count from a worker, clearing
1212 <     * the worker's value
1211 >     * Tries to accumulate steal count from a worker, clearing
1212 >     * the worker's value if successful.
1213 >     *
1214 >     * @return true if worker steal count now zero
1215       */
1216 <    final void accumulateStealCount(ForkJoinWorkerThread w) {
1216 >    final boolean tryAccumulateStealCount(ForkJoinWorkerThread w) {
1217          int sc = w.stealCount;
1218 <        if (sc != 0) {
1219 <            long c;
1220 <            w.stealCount = 0;
1221 <            do {} while (!UNSAFE.compareAndSwapLong(this, stealCountOffset,
1222 <                                                    c = stealCount, c + sc));
1218 >        long c = stealCount;
1219 >        // CAS even if zero, for fence effects
1220 >        if (UNSAFE.compareAndSwapLong(this, stealCountOffset, c, c + sc)) {
1221 >            if (sc != 0)
1222 >                w.stealCount = 0;
1223 >            return true;
1224          }
1225 +        return sc == 0;
1226      }
1227  
1228      /**
# Line 1103 | Line 1230 | public class ForkJoinPool extends Abstra
1230       * active thread.
1231       */
1232      final int idlePerActive() {
1233 <        int pc = parallelism; // use targeted parallelism, not rc
1234 <        int ac = runState;    // no mask -- artifically boosts during shutdown
1233 >        int pc = parallelism; // use parallelism, not rc
1234 >        int ac = runState;    // no mask -- artificially boosts during shutdown
1235          // Use exact results for small values, saturate past 4
1236 <        return pc <= ac? 0 : pc >>> 1 <= ac? 1 : pc >>> 2 <= ac? 3 : pc >>> 3;
1236 >        return ((pc <= ac) ? 0 :
1237 >                (pc >>> 1 <= ac) ? 1 :
1238 >                (pc >>> 2 <= ac) ? 3 :
1239 >                pc >>> 3);
1240      }
1241  
1242      // Public and protected methods
# Line 1154 | Line 1284 | public class ForkJoinPool extends Abstra
1284       * use {@link java.lang.Runtime#availableProcessors}.
1285       * @param factory the factory for creating new threads. For default value,
1286       * use {@link #defaultForkJoinWorkerThreadFactory}.
1287 <     * @param handler the handler for internal worker threads that
1288 <     * terminate due to unrecoverable errors encountered while executing
1289 <     * tasks. For default value, use <code>null</code>.
1290 <     * @param asyncMode if true,
1287 >     * @param handler the handler for internal worker threads that
1288 >     * terminate due to unrecoverable errors encountered while executing
1289 >     * tasks. For default value, use {@code null}.
1290 >     * @param asyncMode if true,
1291       * establishes local first-in-first-out scheduling mode for forked
1292       * tasks that are never joined. This mode may be more appropriate
1293       * than default locally stack-based mode in applications in which
1294       * worker threads only process event-style asynchronous tasks.
1295 <     * For default value, use <code>false</code>.
1295 >     * For default value, use {@code false}.
1296       * @throws IllegalArgumentException if parallelism less than or
1297       *         equal to zero, or greater than implementation limit
1298       * @throws NullPointerException if the factory is null
# Line 1171 | Line 1301 | public class ForkJoinPool extends Abstra
1301       *         because it does not hold {@link
1302       *         java.lang.RuntimePermission}{@code ("modifyThread")}
1303       */
1304 <    public ForkJoinPool(int parallelism,
1304 >    public ForkJoinPool(int parallelism,
1305                          ForkJoinWorkerThreadFactory factory,
1306                          Thread.UncaughtExceptionHandler handler,
1307                          boolean asyncMode) {
1308          checkPermission();
1309          if (factory == null)
1310              throw new NullPointerException();
1311 <        if (parallelism <= 0 || parallelism > MAX_THREADS)
1311 >        if (parallelism <= 0 || parallelism > MAX_WORKERS)
1312              throw new IllegalArgumentException();
1313          this.parallelism = parallelism;
1314          this.factory = factory;
# Line 1197 | Line 1327 | public class ForkJoinPool extends Abstra
1327       * @param pc the initial parallelism level
1328       */
1329      private static int initialArraySizeFor(int pc) {
1330 <        // See Hackers Delight, sec 3.2. We know MAX_THREADS < (1 >>> 16)
1331 <        int size = pc < MAX_THREADS ? pc + 1 : MAX_THREADS;
1330 >        // If possible, initially allocate enough space for one spare
1331 >        int size = pc < MAX_WORKERS ? pc + 1 : MAX_WORKERS;
1332 >        // See Hackers Delight, sec 3.2. We know MAX_WORKERS < (1 >>> 16)
1333          size |= size >>> 1;
1334          size |= size >>> 2;
1335          size |= size >>> 4;
# Line 1209 | Line 1340 | public class ForkJoinPool extends Abstra
1340      // Execution methods
1341  
1342      /**
1343 <     * Common code for execute, invoke and submit
1343 >     * Submits task and creates, starts, or resumes some workers if necessary
1344       */
1345      private <T> void doSubmit(ForkJoinTask<T> task) {
1346 <        if (task == null)
1347 <            throw new NullPointerException();
1348 <        if (runState >= SHUTDOWN)
1349 <            throw new RejectedExecutionException();
1219 <        // Convert submissions to current pool into forks
1220 <        Thread t = Thread.currentThread();
1221 <        ForkJoinWorkerThread w;
1222 <        if ((t instanceof ForkJoinWorkerThread) &&
1223 <            (w = (ForkJoinWorkerThread) t).pool == this)
1224 <            w.pushTask(task);
1225 <        else {
1226 <            submissionQueue.offer(task);
1227 <            signalEvent();
1228 <            ensureEnoughTotalWorkers();
1229 <        }
1346 >        submissionQueue.offer(task);
1347 >        int c; // try to increment event count -- CAS failure OK
1348 >        UNSAFE.compareAndSwapInt(this, eventCountOffset, c = eventCount, c+1);
1349 >        helpMaintainParallelism();
1350      }
1351  
1352      /**
1353       * Performs the given task, returning its result upon completion.
1234     * If the caller is already engaged in a fork/join computation in
1235     * the current pool, this method is equivalent in effect to
1236     * {@link ForkJoinTask#invoke}.
1354       *
1355       * @param task the task
1356       * @return the task's result
# Line 1242 | Line 1359 | public class ForkJoinPool extends Abstra
1359       *         scheduled for execution
1360       */
1361      public <T> T invoke(ForkJoinTask<T> task) {
1362 <        doSubmit(task);
1363 <        return task.join();
1362 >        if (task == null)
1363 >            throw new NullPointerException();
1364 >        if (runState >= SHUTDOWN)
1365 >            throw new RejectedExecutionException();
1366 >        Thread t = Thread.currentThread();
1367 >        if ((t instanceof ForkJoinWorkerThread) &&
1368 >            ((ForkJoinWorkerThread)t).pool == this)
1369 >            return task.invoke();  // bypass submit if in same pool
1370 >        else {
1371 >            doSubmit(task);
1372 >            return task.join();
1373 >        }
1374 >    }
1375 >
1376 >    /**
1377 >     * Unless terminating, forks task if within an ongoing FJ
1378 >     * computation in the current pool, else submits as external task.
1379 >     */
1380 >    private <T> void forkOrSubmit(ForkJoinTask<T> task) {
1381 >        if (runState >= SHUTDOWN)
1382 >            throw new RejectedExecutionException();
1383 >        Thread t = Thread.currentThread();
1384 >        if ((t instanceof ForkJoinWorkerThread) &&
1385 >            ((ForkJoinWorkerThread)t).pool == this)
1386 >            task.fork();
1387 >        else
1388 >            doSubmit(task);
1389      }
1390  
1391      /**
1392       * Arranges for (asynchronous) execution of the given task.
1251     * If the caller is already engaged in a fork/join computation in
1252     * the current pool, this method is equivalent in effect to
1253     * {@link ForkJoinTask#fork}.
1393       *
1394       * @param task the task
1395       * @throws NullPointerException if the task is null
# Line 1258 | Line 1397 | public class ForkJoinPool extends Abstra
1397       *         scheduled for execution
1398       */
1399      public void execute(ForkJoinTask<?> task) {
1400 <        doSubmit(task);
1400 >        if (task == null)
1401 >            throw new NullPointerException();
1402 >        forkOrSubmit(task);
1403      }
1404  
1405      // AbstractExecutorService methods
# Line 1269 | Line 1410 | public class ForkJoinPool extends Abstra
1410       *         scheduled for execution
1411       */
1412      public void execute(Runnable task) {
1413 +        if (task == null)
1414 +            throw new NullPointerException();
1415          ForkJoinTask<?> job;
1416          if (task instanceof ForkJoinTask<?>) // avoid re-wrap
1417              job = (ForkJoinTask<?>) task;
1418          else
1419              job = ForkJoinTask.adapt(task, null);
1420 <        doSubmit(job);
1420 >        forkOrSubmit(job);
1421      }
1422  
1423      /**
1424       * Submits a ForkJoinTask for execution.
1282     * If the caller is already engaged in a fork/join computation in
1283     * the current pool, this method is equivalent in effect to
1284     * {@link ForkJoinTask#fork}.
1425       *
1426       * @param task the task to submit
1427       * @return the task
# Line 1290 | Line 1430 | public class ForkJoinPool extends Abstra
1430       *         scheduled for execution
1431       */
1432      public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) {
1433 <        doSubmit(task);
1433 >        if (task == null)
1434 >            throw new NullPointerException();
1435 >        forkOrSubmit(task);
1436          return task;
1437      }
1438  
# Line 1300 | Line 1442 | public class ForkJoinPool extends Abstra
1442       *         scheduled for execution
1443       */
1444      public <T> ForkJoinTask<T> submit(Callable<T> task) {
1445 +        if (task == null)
1446 +            throw new NullPointerException();
1447          ForkJoinTask<T> job = ForkJoinTask.adapt(task);
1448 <        doSubmit(job);
1448 >        forkOrSubmit(job);
1449          return job;
1450      }
1451  
# Line 1311 | Line 1455 | public class ForkJoinPool extends Abstra
1455       *         scheduled for execution
1456       */
1457      public <T> ForkJoinTask<T> submit(Runnable task, T result) {
1458 +        if (task == null)
1459 +            throw new NullPointerException();
1460          ForkJoinTask<T> job = ForkJoinTask.adapt(task, result);
1461 <        doSubmit(job);
1461 >        forkOrSubmit(job);
1462          return job;
1463      }
1464  
# Line 1322 | Line 1468 | public class ForkJoinPool extends Abstra
1468       *         scheduled for execution
1469       */
1470      public ForkJoinTask<?> submit(Runnable task) {
1471 +        if (task == null)
1472 +            throw new NullPointerException();
1473          ForkJoinTask<?> job;
1474          if (task instanceof ForkJoinTask<?>) // avoid re-wrap
1475              job = (ForkJoinTask<?>) task;
1476          else
1477              job = ForkJoinTask.adapt(task, null);
1478 <        doSubmit(job);
1478 >        forkOrSubmit(job);
1479          return job;
1480      }
1481  
# Line 1387 | Line 1535 | public class ForkJoinPool extends Abstra
1535  
1536      /**
1537       * Returns the number of worker threads that have started but not
1538 <     * yet terminated.  This result returned by this method may differ
1538 >     * yet terminated.  The result returned by this method may differ
1539       * from {@link #getParallelism} when threads are created to
1540       * maintain parallelism when others are cooperatively blocked.
1541       *
# Line 1472 | Line 1620 | public class ForkJoinPool extends Abstra
1620       */
1621      public long getQueuedTaskCount() {
1622          long count = 0;
1623 <        ForkJoinWorkerThread[] ws = workers;
1476 <        int nws = ws.length;
1477 <        for (int i = 0; i < nws; ++i) {
1478 <            ForkJoinWorkerThread w = ws[i];
1623 >        for (ForkJoinWorkerThread w : workers)
1624              if (w != null)
1625                  count += w.getQueueSize();
1481        }
1626          return count;
1627      }
1628  
# Line 1532 | Line 1676 | public class ForkJoinPool extends Abstra
1676       * @return the number of elements transferred
1677       */
1678      protected int drainTasksTo(Collection<? super ForkJoinTask<?>> c) {
1679 <        int n = submissionQueue.drainTo(c);
1680 <        ForkJoinWorkerThread[] ws = workers;
1537 <        int nws = ws.length;
1538 <        for (int i = 0; i < nws; ++i) {
1539 <            ForkJoinWorkerThread w = ws[i];
1540 <            if (w != null)
1541 <                n += w.drainTasksTo(c);
1542 <        }
1543 <        return n;
1544 <    }
1545 <
1546 <    /**
1547 <     * Returns count of total parks by existing workers.
1548 <     * Used during development only since not meaningful to users.
1549 <     */
1550 <    private int collectParkCount() {
1551 <        int count = 0;
1552 <        ForkJoinWorkerThread[] ws = workers;
1553 <        int nws = ws.length;
1554 <        for (int i = 0; i < nws; ++i) {
1555 <            ForkJoinWorkerThread w = ws[i];
1679 >        int count = submissionQueue.drainTo(c);
1680 >        for (ForkJoinWorkerThread w : workers)
1681              if (w != null)
1682 <                count += w.parkCount;
1558 <        }
1682 >                count += w.drainTasksTo(c);
1683          return count;
1684      }
1685  
# Line 1576 | Line 1700 | public class ForkJoinPool extends Abstra
1700          int pc = parallelism;
1701          int rs = runState;
1702          int ac = rs & ACTIVE_COUNT_MASK;
1579        //        int pk = collectParkCount();
1703          return super.toString() +
1704              "[" + runLevelToString(rs) +
1705              ", parallelism = " + pc +
# Line 1586 | Line 1709 | public class ForkJoinPool extends Abstra
1709              ", steals = " + st +
1710              ", tasks = " + qt +
1711              ", submissions = " + qs +
1589            //            ", parks = " + pk +
1712              "]";
1713      }
1714  
# Line 1651 | Line 1773 | public class ForkJoinPool extends Abstra
1773       * commenced but not yet completed.  This method may be useful for
1774       * debugging. A return of {@code true} reported a sufficient
1775       * period after shutdown may indicate that submitted tasks have
1776 <     * ignored or suppressed interruption, causing this executor not
1777 <     * to properly terminate.
1776 >     * ignored or suppressed interruption, or are waiting for IO,
1777 >     * causing this executor not to properly terminate. (See the
1778 >     * advisory notes for class {@link ForkJoinTask} stating that
1779 >     * tasks should not normally entail blocking operations.  But if
1780 >     * they do, they must abort them on interrupt.)
1781       *
1782       * @return {@code true} if terminating but not yet terminated
1783       */
# Line 1661 | Line 1786 | public class ForkJoinPool extends Abstra
1786      }
1787  
1788      /**
1789 +     * Returns true if terminating or terminated. Used by ForkJoinWorkerThread.
1790 +     */
1791 +    final boolean isAtLeastTerminating() {
1792 +        return runState >= TERMINATING;
1793 +    }
1794 +
1795 +    /**
1796       * Returns {@code true} if this pool has been shut down.
1797       *
1798       * @return {@code true} if this pool has been shut down
# Line 1683 | Line 1815 | public class ForkJoinPool extends Abstra
1815      public boolean awaitTermination(long timeout, TimeUnit unit)
1816          throws InterruptedException {
1817          try {
1818 <            return termination.awaitAdvanceInterruptibly(0, timeout, unit) > 0;
1819 <        } catch(TimeoutException ex) {
1818 >            termination.awaitAdvanceInterruptibly(0, timeout, unit);
1819 >        } catch (TimeoutException ex) {
1820              return false;
1821          }
1822 +        return true;
1823      }
1824  
1825      /**
1826       * Interface for extending managed parallelism for tasks running
1827       * in {@link ForkJoinPool}s.
1828       *
1829 <     * <p>A {@code ManagedBlocker} provides two methods.
1830 <     * Method {@code isReleasable} must return {@code true} if
1831 <     * blocking is not necessary. Method {@code block} blocks the
1832 <     * current thread if necessary (perhaps internally invoking
1833 <     * {@code isReleasable} before actually blocking).
1829 >     * <p>A {@code ManagedBlocker} provides two methods.  Method
1830 >     * {@code isReleasable} must return {@code true} if blocking is
1831 >     * not necessary. Method {@code block} blocks the current thread
1832 >     * if necessary (perhaps internally invoking {@code isReleasable}
1833 >     * before actually blocking). The unusual methods in this API
1834 >     * accommodate synchronizers that may, but don't usually, block
1835 >     * for long periods. Similarly, they allow more efficient internal
1836 >     * handling of cases in which additional workers may be, but
1837 >     * usually are not, needed to ensure sufficient parallelism.
1838 >     * Toward this end, implementations of method {@code isReleasable}
1839 >     * must be amenable to repeated invocation.
1840       *
1841       * <p>For example, here is a ManagedBlocker based on a
1842       * ReentrantLock:
# Line 1715 | Line 1854 | public class ForkJoinPool extends Abstra
1854       *     return hasLock || (hasLock = lock.tryLock());
1855       *   }
1856       * }}</pre>
1857 +     *
1858 +     * <p>Here is a class that possibly blocks waiting for an
1859 +     * item on a given queue:
1860 +     *  <pre> {@code
1861 +     * class QueueTaker<E> implements ManagedBlocker {
1862 +     *   final BlockingQueue<E> queue;
1863 +     *   volatile E item = null;
1864 +     *   QueueTaker(BlockingQueue<E> q) { this.queue = q; }
1865 +     *   public boolean block() throws InterruptedException {
1866 +     *     if (item == null)
1867 +     *       item = queue.take();
1868 +     *     return true;
1869 +     *   }
1870 +     *   public boolean isReleasable() {
1871 +     *     return item != null || (item = queue.poll()) != null;
1872 +     *   }
1873 +     *   public E getItem() { // call after pool.managedBlock completes
1874 +     *     return item;
1875 +     *   }
1876 +     * }}</pre>
1877       */
1878      public static interface ManagedBlocker {
1879          /**
# Line 1757 | Line 1916 | public class ForkJoinPool extends Abstra
1916      public static void managedBlock(ManagedBlocker blocker)
1917          throws InterruptedException {
1918          Thread t = Thread.currentThread();
1919 <        if (t instanceof ForkJoinWorkerThread)
1920 <            ((ForkJoinWorkerThread) t).pool.awaitBlocker(blocker);
1919 >        if (t instanceof ForkJoinWorkerThread) {
1920 >            ForkJoinWorkerThread w = (ForkJoinWorkerThread) t;
1921 >            w.pool.awaitBlocker(blocker);
1922 >        }
1923          else {
1924              do {} while (!blocker.isReleasable() && !blocker.block());
1925          }
# Line 1786 | Line 1947 | public class ForkJoinPool extends Abstra
1947      private static final long eventCountOffset =
1948          objectFieldOffset("eventCount", ForkJoinPool.class);
1949      private static final long eventWaitersOffset =
1950 <        objectFieldOffset("eventWaiters",ForkJoinPool.class);
1950 >        objectFieldOffset("eventWaiters", ForkJoinPool.class);
1951      private static final long stealCountOffset =
1952 <        objectFieldOffset("stealCount",ForkJoinPool.class);
1952 >        objectFieldOffset("stealCount", ForkJoinPool.class);
1953 >    private static final long spareWaitersOffset =
1954 >        objectFieldOffset("spareWaiters", ForkJoinPool.class);
1955  
1956      private static long objectFieldOffset(String field, Class<?> klazz) {
1957          try {

Diff Legend

Removed lines
+ Added lines
< Changed lines
> Changed lines