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 1.40 by jsr166, Mon Aug 3 00:53:15 2009 UTC vs.
Revision 1.74 by jsr166, Tue Sep 7 06:32:45 2010 UTC

# Line 7 | Line 7
7   package jsr166y;
8  
9   import java.util.concurrent.*;
10
10   import java.util.ArrayList;
11   import java.util.Arrays;
12   import java.util.Collection;
13   import java.util.Collections;
14   import java.util.List;
16 import java.util.concurrent.locks.Condition;
15   import java.util.concurrent.locks.LockSupport;
16   import java.util.concurrent.locks.ReentrantLock;
17   import java.util.concurrent.atomic.AtomicInteger;
18 < import java.util.concurrent.atomic.AtomicLong;
18 > import java.util.concurrent.CountDownLatch;
19  
20   /**
21   * An {@link ExecutorService} for running {@link ForkJoinTask}s.
22   * A {@code ForkJoinPool} provides the entry point for submissions
23 < * from non-{@code ForkJoinTask}s, as well as management and
24 < * monitoring operations.  Normally a single {@code ForkJoinPool} is
27 < * used for a large number of submitted tasks. Otherwise, use would
28 < * not usually outweigh the construction and bookkeeping overhead of
29 < * creating a large set of threads.
23 > * from non-{@code ForkJoinTask} clients, as well as management and
24 > * monitoring operations.
25   *
26 < * <p>{@code ForkJoinPool}s differ from other kinds of {@link
27 < * Executor}s mainly in that they provide <em>work-stealing</em>: all
28 < * threads in the pool attempt to find and execute subtasks created by
29 < * other active tasks (eventually blocking if none exist). This makes
30 < * them efficient when most tasks spawn other subtasks (as do most
31 < * {@code ForkJoinTask}s), as well as the mixed execution of some
32 < * plain {@code Runnable}- or {@code Callable}- based activities along
33 < * with {@code ForkJoinTask}s. When setting {@linkplain #setAsyncMode
34 < * async mode}, a {@code ForkJoinPool} may also be appropriate for use
40 < * with fine-grained tasks that are never joined. Otherwise, other
41 < * {@code ExecutorService} implementations are typically more
42 < * appropriate choices.
26 > * <p>A {@code ForkJoinPool} differs from other kinds of {@link
27 > * ExecutorService} mainly by virtue of employing
28 > * <em>work-stealing</em>: all threads in the pool attempt to find and
29 > * execute subtasks created by other active tasks (eventually blocking
30 > * waiting for work if none exist). This enables efficient processing
31 > * when most tasks spawn other subtasks (as do most {@code
32 > * ForkJoinTask}s). When setting <em>asyncMode</em> to true in
33 > * constructors, {@code ForkJoinPool}s may also be appropriate for use
34 > * with event-style tasks that are never joined.
35   *
36 < * <p>A {@code ForkJoinPool} may be constructed with a given
37 < * parallelism level (target pool size), which it attempts to maintain
38 < * by dynamically adding, suspending, or resuming threads, even if
39 < * some tasks are waiting to join others. However, no such adjustments
40 < * are performed in the face of blocked IO or other unmanaged
41 < * synchronization. The nested {@link ManagedBlocker} interface
42 < * enables extension of the kinds of synchronization accommodated.
43 < * The target parallelism level may also be changed dynamically
44 < * ({@link #setParallelism}) and thread construction can be limited
53 < * using methods {@link #setMaximumPoolSize} and/or {@link
54 < * #setMaintainsParallelism}.
36 > * <p>A {@code ForkJoinPool} is constructed with a given target
37 > * parallelism level; by default, equal to the number of available
38 > * processors. The pool attempts to maintain enough active (or
39 > * available) threads by dynamically adding, suspending, or resuming
40 > * internal worker threads, even if some tasks are stalled waiting to
41 > * join others. However, no such adjustments are guaranteed in the
42 > * face of blocked IO or other unmanaged synchronization. The nested
43 > * {@link ManagedBlocker} interface enables extension of the kinds of
44 > * synchronization accommodated.
45   *
46   * <p>In addition to execution and lifecycle control methods, this
47   * class provides status check methods (for example
# Line 60 | Line 50 | import java.util.concurrent.atomic.Atomi
50   * {@link #toString} returns indications of pool state in a
51   * convenient form for informal monitoring.
52   *
53 + * <p> As is the case with other ExecutorServices, there are three
54 + * main task execution methods summarized in the following
55 + * table. These are designed to be used by clients not already engaged
56 + * in fork/join computations in the current pool.  The main forms of
57 + * these methods accept instances of {@code ForkJoinTask}, but
58 + * overloaded forms also allow mixed execution of plain {@code
59 + * Runnable}- or {@code Callable}- based activities as well.  However,
60 + * tasks that are already executing in a pool should normally
61 + * <em>NOT</em> use these pool execution methods, but instead use the
62 + * within-computation forms listed in the table.
63 + *
64 + * <table BORDER CELLPADDING=3 CELLSPACING=1>
65 + *  <tr>
66 + *    <td></td>
67 + *    <td ALIGN=CENTER> <b>Call from non-fork/join clients</b></td>
68 + *    <td ALIGN=CENTER> <b>Call from within fork/join computations</b></td>
69 + *  </tr>
70 + *  <tr>
71 + *    <td> <b>Arrange async execution</td>
72 + *    <td> {@link #execute(ForkJoinTask)}</td>
73 + *    <td> {@link ForkJoinTask#fork}</td>
74 + *  </tr>
75 + *  <tr>
76 + *    <td> <b>Await and obtain result</td>
77 + *    <td> {@link #invoke(ForkJoinTask)}</td>
78 + *    <td> {@link ForkJoinTask#invoke}</td>
79 + *  </tr>
80 + *  <tr>
81 + *    <td> <b>Arrange exec and obtain Future</td>
82 + *    <td> {@link #submit(ForkJoinTask)}</td>
83 + *    <td> {@link ForkJoinTask#fork} (ForkJoinTasks <em>are</em> Futures)</td>
84 + *  </tr>
85 + * </table>
86 + *
87 + * <p><b>Sample Usage.</b> Normally a single {@code ForkJoinPool} is
88 + * used for all parallel task execution in a program or subsystem.
89 + * Otherwise, use would not usually outweigh the construction and
90 + * bookkeeping overhead of creating a large set of threads. For
91 + * example, a common pool could be used for the {@code SortTasks}
92 + * illustrated in {@link RecursiveAction}. Because {@code
93 + * ForkJoinPool} uses threads in {@linkplain java.lang.Thread#isDaemon
94 + * daemon} mode, there is typically no need to explicitly {@link
95 + * #shutdown} such a pool upon program exit.
96 + *
97 + * <pre>
98 + * static final ForkJoinPool mainPool = new ForkJoinPool();
99 + * ...
100 + * public void sort(long[] array) {
101 + *   mainPool.invoke(new SortTask(array, 0, array.length));
102 + * }
103 + * </pre>
104 + *
105   * <p><b>Implementation notes</b>: This implementation restricts the
106   * maximum number of running threads to 32767. Attempts to create
107 < * pools with greater than the maximum result in
107 > * pools with greater than the maximum number result in
108   * {@code IllegalArgumentException}.
109   *
110 + * <p>This implementation rejects submitted tasks (that is, by throwing
111 + * {@link RejectedExecutionException}) only when the pool is shut down
112 + * or internal resources have been exhausted.
113 + *
114   * @since 1.7
115   * @author Doug Lea
116   */
117   public class ForkJoinPool extends AbstractExecutorService {
118  
119      /*
120 <     * See the extended comments interspersed below for design,
121 <     * rationale, and walkthroughs.
120 >     * Implementation Overview
121 >     *
122 >     * This class provides the central bookkeeping and control for a
123 >     * set of worker threads: Submissions from non-FJ threads enter
124 >     * into a submission queue. Workers take these tasks and typically
125 >     * split them into subtasks that may be stolen by other workers.
126 >     * The main work-stealing mechanics implemented in class
127 >     * ForkJoinWorkerThread give first priority to processing tasks
128 >     * from their own queues (LIFO or FIFO, depending on mode), then
129 >     * to randomized FIFO steals of tasks in other worker queues, and
130 >     * lastly to new submissions. These mechanics do not consider
131 >     * affinities, loads, cache localities, etc, so rarely provide the
132 >     * best possible performance on a given machine, but portably
133 >     * provide good throughput by averaging over these factors.
134 >     * (Further, even if we did try to use such information, we do not
135 >     * usually have a basis for exploiting it. For example, some sets
136 >     * of tasks profit from cache affinities, but others are harmed by
137 >     * cache pollution effects.)
138 >     *
139 >     * Beyond work-stealing support and essential bookkeeping, the
140 >     * main responsibility of this framework is to take actions when
141 >     * one worker is waiting to join a task stolen (or always held by)
142 >     * another.  Because we are multiplexing many tasks on to a pool
143 >     * of workers, we can't just let them block (as in Thread.join).
144 >     * We also cannot just reassign the joiner's run-time stack with
145 >     * another and replace it later, which would be a form of
146 >     * "continuation", that even if possible is not necessarily a good
147 >     * idea. Given that the creation costs of most threads on most
148 >     * systems mainly surrounds setting up runtime stacks, thread
149 >     * creation and switching is usually not much more expensive than
150 >     * stack creation and switching, and is more flexible). Instead we
151 >     * combine two tactics:
152 >     *
153 >     *   Helping: Arranging for the joiner to execute some task that it
154 >     *      would be running if the steal had not occurred.  Method
155 >     *      ForkJoinWorkerThread.helpJoinTask tracks joining->stealing
156 >     *      links to try to find such a task.
157 >     *
158 >     *   Compensating: Unless there are already enough live threads,
159 >     *      method helpMaintainParallelism() may create or
160 >     *      re-activate a spare thread to compensate for blocked
161 >     *      joiners until they unblock.
162 >     *
163 >     * It is impossible to keep exactly the target (parallelism)
164 >     * number of threads running at any given time.  Determining
165 >     * existence of conservatively safe helping targets, the
166 >     * availability of already-created spares, and the apparent need
167 >     * to create new spares are all racy and require heuristic
168 >     * guidance, so we rely on multiple retries of each.  Compensation
169 >     * occurs in slow-motion. It is triggered only upon timeouts of
170 >     * Object.wait used for joins. This reduces poor decisions that
171 >     * would otherwise be made when threads are waiting for others
172 >     * that are stalled because of unrelated activities such as
173 >     * garbage collection.
174 >     *
175 >     * The ManagedBlocker extension API can't use helping so relies
176 >     * only on compensation in method awaitBlocker.
177 >     *
178 >     * The main throughput advantages of work-stealing stem from
179 >     * decentralized control -- workers mostly steal tasks from each
180 >     * other. We do not want to negate this by creating bottlenecks
181 >     * implementing other management responsibilities. So we use a
182 >     * collection of techniques that avoid, reduce, or cope well with
183 >     * contention. These entail several instances of bit-packing into
184 >     * CASable fields to maintain only the minimally required
185 >     * atomicity. To enable such packing, we restrict maximum
186 >     * parallelism to (1<<15)-1 (enabling twice this (to accommodate
187 >     * unbalanced increments and decrements) to fit into a 16 bit
188 >     * field, which is far in excess of normal operating range.  Even
189 >     * though updates to some of these bookkeeping fields do sometimes
190 >     * contend with each other, they don't normally cache-contend with
191 >     * updates to others enough to warrant memory padding or
192 >     * isolation. So they are all held as fields of ForkJoinPool
193 >     * objects.  The main capabilities are as follows:
194 >     *
195 >     * 1. Creating and removing workers. Workers are recorded in the
196 >     * "workers" array. This is an array as opposed to some other data
197 >     * structure to support index-based random steals by workers.
198 >     * Updates to the array recording new workers and unrecording
199 >     * terminated ones are protected from each other by a lock
200 >     * (workerLock) but the array is otherwise concurrently readable,
201 >     * and accessed directly by workers. To simplify index-based
202 >     * operations, the array size is always a power of two, and all
203 >     * readers must tolerate null slots. Currently, all worker thread
204 >     * creation is on-demand, triggered by task submissions,
205 >     * replacement of terminated workers, and/or compensation for
206 >     * blocked workers. However, all other support code is set up to
207 >     * work with other policies.
208 >     *
209 >     * To ensure that we do not hold on to worker references that
210 >     * would prevent GC, ALL accesses to workers are via indices into
211 >     * the workers array (which is one source of some of the unusual
212 >     * code constructions here). In essence, the workers array serves
213 >     * as a WeakReference mechanism. Thus for example the event queue
214 >     * stores worker indices, not worker references. Access to the
215 >     * workers in associated methods (for example releaseEventWaiters)
216 >     * must both index-check and null-check the IDs. All such accesses
217 >     * ignore bad IDs by returning out early from what they are doing,
218 >     * since this can only be associated with shutdown, in which case
219 >     * it is OK to give up. On termination, we just clobber these
220 >     * data structures without trying to use them.
221 >     *
222 >     * 2. Bookkeeping for dynamically adding and removing workers. We
223 >     * aim to approximately maintain the given level of parallelism.
224 >     * When some workers are known to be blocked (on joins or via
225 >     * ManagedBlocker), we may create or resume others to take their
226 >     * place until they unblock (see below). Implementing this
227 >     * requires counts of the number of "running" threads (i.e., those
228 >     * that are neither blocked nor artificially suspended) as well as
229 >     * the total number.  These two values are packed into one field,
230 >     * "workerCounts" because we need accurate snapshots when deciding
231 >     * to create, resume or suspend.  Note however that the
232 >     * correspondence of these counts to reality is not guaranteed. In
233 >     * particular updates for unblocked threads may lag until they
234 >     * actually wake up.
235 >     *
236 >     * 3. Maintaining global run state. The run state of the pool
237 >     * consists of a runLevel (SHUTDOWN, TERMINATING, etc) similar to
238 >     * those in other Executor implementations, as well as a count of
239 >     * "active" workers -- those that are, or soon will be, or
240 >     * recently were executing tasks. The runLevel and active count
241 >     * are packed together in order to correctly trigger shutdown and
242 >     * termination. Without care, active counts can be subject to very
243 >     * high contention.  We substantially reduce this contention by
244 >     * relaxing update rules.  A worker must claim active status
245 >     * prospectively, by activating if it sees that a submitted or
246 >     * stealable task exists (it may find after activating that the
247 >     * task no longer exists). It stays active while processing this
248 >     * task (if it exists) and any other local subtasks it produces,
249 >     * until it cannot find any other tasks. It then tries
250 >     * inactivating (see method preStep), but upon update contention
251 >     * instead scans for more tasks, later retrying inactivation if it
252 >     * doesn't find any.
253 >     *
254 >     * 4. Managing idle workers waiting for tasks. We cannot let
255 >     * workers spin indefinitely scanning for tasks when none are
256 >     * available. On the other hand, we must quickly prod them into
257 >     * action when new tasks are submitted or generated.  We
258 >     * park/unpark these idle workers using an event-count scheme.
259 >     * Field eventCount is incremented upon events that may enable
260 >     * workers that previously could not find a task to now find one:
261 >     * Submission of a new task to the pool, or another worker pushing
262 >     * a task onto a previously empty queue.  (We also use this
263 >     * mechanism for configuration and termination actions that
264 >     * require wakeups of idle workers).  Each worker maintains its
265 >     * last known event count, and blocks when a scan for work did not
266 >     * find a task AND its lastEventCount matches the current
267 >     * eventCount. Waiting idle workers are recorded in a variant of
268 >     * Treiber stack headed by field eventWaiters which, when nonzero,
269 >     * encodes the thread index and count awaited for by the worker
270 >     * thread most recently calling eventSync. This thread in turn has
271 >     * a record (field nextEventWaiter) for the next waiting worker.
272 >     * In addition to allowing simpler decisions about need for
273 >     * wakeup, the event count bits in eventWaiters serve the role of
274 >     * tags to avoid ABA errors in Treiber stacks. Upon any wakeup,
275 >     * released threads also try to release at most two others.  The
276 >     * net effect is a tree-like diffusion of signals, where released
277 >     * threads (and possibly others) help with unparks.  To further
278 >     * reduce contention effects a bit, failed CASes to increment
279 >     * field eventCount are tolerated without retries in signalWork.
280 >     * Conceptually they are merged into the same event, which is OK
281 >     * when their only purpose is to enable workers to scan for work.
282 >     *
283 >     * 5. Managing suspension of extra workers. When a worker notices
284 >     * (usually upon timeout of a wait()) that there are too few
285 >     * running threads, we may create a new thread to maintain
286 >     * parallelism level, or at least avoid starvation. Usually, extra
287 >     * threads are needed for only very short periods, yet join
288 >     * dependencies are such that we sometimes need them in
289 >     * bursts. Rather than create new threads each time this happens,
290 >     * we suspend no-longer-needed extra ones as "spares". For most
291 >     * purposes, we don't distinguish "extra" spare threads from
292 >     * normal "core" threads: On each call to preStep (the only point
293 >     * at which we can do this) a worker checks to see if there are
294 >     * now too many running workers, and if so, suspends itself.
295 >     * Method helpMaintainParallelism looks for suspended threads to
296 >     * resume before considering creating a new replacement. The
297 >     * spares themselves are encoded on another variant of a Treiber
298 >     * Stack, headed at field "spareWaiters".  Note that the use of
299 >     * spares is intrinsically racy.  One thread may become a spare at
300 >     * about the same time as another is needlessly being created. We
301 >     * counteract this and related slop in part by requiring resumed
302 >     * spares to immediately recheck (in preStep) to see whether they
303 >     * should re-suspend.
304 >     *
305 >     * 6. Killing off unneeded workers. A timeout mechanism is used to
306 >     * shed unused workers: The oldest (first) event queue waiter uses
307 >     * a timed rather than hard wait. When this wait times out without
308 >     * a normal wakeup, it tries to shutdown any one (for convenience
309 >     * the newest) other spare or event waiter via
310 >     * tryShutdownUnusedWorker. This eventually reduces the number of
311 >     * worker threads to a minimum of one after a long enough period
312 >     * without use.
313 >     *
314 >     * 7. Deciding when to create new workers. The main dynamic
315 >     * control in this class is deciding when to create extra threads
316 >     * in method helpMaintainParallelism. We would like to keep
317 >     * exactly #parallelism threads running, which is an impossible
318 >     * task. We always need to create one when the number of running
319 >     * threads would become zero and all workers are busy. Beyond
320 >     * this, we must rely on heuristics that work well in the
321 >     * presence of transient phenomena such as GC stalls, dynamic
322 >     * compilation, and wake-up lags. These transients are extremely
323 >     * common -- we are normally trying to fully saturate the CPUs on
324 >     * a machine, so almost any activity other than running tasks
325 >     * impedes accuracy. Our main defense is to allow parallelism to
326 >     * lapse for a while during joins, and use a timeout to see if,
327 >     * after the resulting settling, there is still a need for
328 >     * additional workers.  This also better copes with the fact that
329 >     * some of the methods in this class tend to never become compiled
330 >     * (but are interpreted), so some components of the entire set of
331 >     * controls might execute 100 times faster than others. And
332 >     * similarly for cases where the apparent lack of work is just due
333 >     * to GC stalls and other transient system activity.
334 >     *
335 >     * Beware that there is a lot of representation-level coupling
336 >     * among classes ForkJoinPool, ForkJoinWorkerThread, and
337 >     * ForkJoinTask.  For example, direct access to "workers" array by
338 >     * workers, and direct access to ForkJoinTask.status by both
339 >     * ForkJoinPool and ForkJoinWorkerThread.  There is little point
340 >     * trying to reduce this, since any associated future changes in
341 >     * representations will need to be accompanied by algorithmic
342 >     * changes anyway.
343 >     *
344 >     * Style notes: There are lots of inline assignments (of form
345 >     * "while ((local = field) != 0)") which are usually the simplest
346 >     * way to ensure the required read orderings (which are sometimes
347 >     * critical). Also several occurrences of the unusual "do {}
348 >     * while (!cas...)" which is the simplest way to force an update of
349 >     * a CAS'ed variable. There are also other coding oddities that
350 >     * help some methods perform reasonably even when interpreted (not
351 >     * compiled), at the expense of some messy constructions that
352 >     * reduce byte code counts.
353 >     *
354 >     * The order of declarations in this file is: (1) statics (2)
355 >     * fields (along with constants used when unpacking some of them)
356 >     * (3) internal control methods (4) callbacks and other support
357 >     * for ForkJoinTask and ForkJoinWorkerThread classes, (5) exported
358 >     * methods (plus a few little helpers).
359       */
360  
78    /** Mask for packing and unpacking shorts */
79    private static final int  shortMask = 0xffff;
80
81    /** Max pool size -- must be a power of two minus 1 */
82    private static final int MAX_THREADS =  0x7FFF;
83
361      /**
362       * Factory for creating new {@link ForkJoinWorkerThread}s.
363       * A {@code ForkJoinWorkerThreadFactory} must be defined and used
# Line 92 | Line 369 | public class ForkJoinPool extends Abstra
369           * Returns a new worker thread operating in the given pool.
370           *
371           * @param pool the pool this thread works in
372 <         * @throws NullPointerException if pool is null
372 >         * @throws NullPointerException if the pool is null
373           */
374          public ForkJoinWorkerThread newThread(ForkJoinPool pool);
375      }
# Line 101 | Line 378 | public class ForkJoinPool extends Abstra
378       * Default ForkJoinWorkerThreadFactory implementation; creates a
379       * new ForkJoinWorkerThread.
380       */
381 <    static class  DefaultForkJoinWorkerThreadFactory
381 >    static class DefaultForkJoinWorkerThreadFactory
382          implements ForkJoinWorkerThreadFactory {
383          public ForkJoinWorkerThread newThread(ForkJoinPool pool) {
384 <            try {
108 <                return new ForkJoinWorkerThread(pool);
109 <            } catch (OutOfMemoryError oom)  {
110 <                return null;
111 <            }
384 >            return new ForkJoinWorkerThread(pool);
385          }
386      }
387  
# Line 144 | Line 417 | public class ForkJoinPool extends Abstra
417          new AtomicInteger();
418  
419      /**
420 <     * Array holding all worker threads in the pool. Initialized upon
421 <     * first use. Array size must be a power of two.  Updates and
422 <     * replacements are protected by workerLock, but it is always kept
423 <     * in a consistent enough state to be randomly accessed without
424 <     * locking by workers performing work-stealing.
420 >     * The time to block in a join (see awaitJoin) before checking if
421 >     * a new worker should be (re)started to maintain parallelism
422 >     * level. The value should be short enough to maintain global
423 >     * responsiveness and progress but long enough to avoid
424 >     * counterproductive firings during GC stalls or unrelated system
425 >     * activity, and to not bog down systems with continual re-firings
426 >     * on GCs or legitimately long waits.
427       */
428 <    volatile ForkJoinWorkerThread[] workers;
428 >    private static final long JOIN_TIMEOUT_MILLIS = 250L; // 4 per second
429  
430      /**
431 <     * Lock protecting access to workers.
431 >     * The wakeup interval (in nanoseconds) for the oldest worker
432 >     * waiting for an event invokes tryShutdownUnusedWorker to shrink
433 >     * the number of workers.  The exact value does not matter too
434 >     * much, but should be long enough to slowly release resources
435 >     * during long periods without use without disrupting normal use.
436       */
437 <    private final ReentrantLock workerLock;
437 >    private static final long SHRINK_RATE_NANOS =
438 >        30L * 1000L * 1000L * 1000L; // 2 per minute
439  
440      /**
441 <     * Condition for awaitTermination.
441 >     * Absolute bound for parallelism level. Twice this number plus
442 >     * one (i.e., 0xfff) must fit into a 16bit field to enable
443 >     * word-packing for some counts and indices.
444       */
445 <    private final Condition termination;
445 >    private static final int MAX_WORKERS   = 0x7fff;
446  
447      /**
448 <     * The uncaught exception handler used when any worker
449 <     * abruptly terminates
448 >     * Array holding all worker threads in the pool.  Array size must
449 >     * be a power of two.  Updates and replacements are protected by
450 >     * workerLock, but the array is always kept in a consistent enough
451 >     * state to be randomly accessed without locking by workers
452 >     * performing work-stealing, as well as other traversal-based
453 >     * methods in this class. All readers must tolerate that some
454 >     * array slots may be null.
455       */
456 <    private Thread.UncaughtExceptionHandler ueh;
456 >    volatile ForkJoinWorkerThread[] workers;
457  
458      /**
459 <     * Creation factory for worker threads.
459 >     * Queue for external submissions.
460       */
461 <    private final ForkJoinWorkerThreadFactory factory;
461 >    private final LinkedTransferQueue<ForkJoinTask<?>> submissionQueue;
462  
463      /**
464 <     * Head of stack of threads that were created to maintain
178 <     * parallelism when other threads blocked, but have since
179 <     * suspended when the parallelism level rose.
464 >     * Lock protecting updates to workers array.
465       */
466 <    private volatile WaitQueueNode spareStack;
466 >    private final ReentrantLock workerLock;
467  
468      /**
469 <     * Sum of per-thread steal counts, updated only when threads are
185 <     * idle or terminating.
469 >     * Latch released upon termination.
470       */
471 <    private final AtomicLong stealCount;
471 >    private final Phaser termination;
472  
473      /**
474 <     * Queue for external submissions.
474 >     * Creation factory for worker threads.
475       */
476 <    private final LinkedTransferQueue<ForkJoinTask<?>> submissionQueue;
476 >    private final ForkJoinWorkerThreadFactory factory;
477  
478      /**
479 <     * Head of Treiber stack for barrier sync. See below for explanation.
479 >     * Sum of per-thread steal counts, updated only when threads are
480 >     * idle or terminating.
481       */
482 <    private volatile WaitQueueNode syncStack;
482 >    private volatile long stealCount;
483  
484      /**
485 <     * The count for event barrier
485 >     * Encoded record of top of Treiber stack of threads waiting for
486 >     * events. The top 32 bits contain the count being waited for. The
487 >     * bottom 16 bits contains one plus the pool index of waiting
488 >     * worker thread. (Bits 16-31 are unused.)
489       */
490 <    private volatile long eventCount;
490 >    private volatile long eventWaiters;
491  
492 <    /**
493 <     * Pool number, just for assigning useful names to worker threads
206 <     */
207 <    private final int poolNumber;
492 >    private static final int  EVENT_COUNT_SHIFT = 32;
493 >    private static final long WAITER_ID_MASK    = (1L << 16) - 1L;
494  
495      /**
496 <     * The maximum allowed pool size
496 >     * A counter for events that may wake up worker threads:
497 >     *   - Submission of a new task to the pool
498 >     *   - A worker pushing a task on an empty queue
499 >     *   - termination
500       */
501 <    private volatile int maxPoolSize;
501 >    private volatile int eventCount;
502  
503      /**
504 <     * The desired parallelism level, updated only under workerLock.
504 >     * Encoded record of top of Treiber stack of spare threads waiting
505 >     * for resumption. The top 16 bits contain an arbitrary count to
506 >     * avoid ABA effects. The bottom 16bits contains one plus the pool
507 >     * index of waiting worker thread.
508       */
509 <    private volatile int parallelism;
509 >    private volatile int spareWaiters;
510 >
511 >    private static final int SPARE_COUNT_SHIFT = 16;
512 >    private static final int SPARE_ID_MASK     = (1 << 16) - 1;
513  
514      /**
515 <     * True if use local fifo, not default lifo, for local polling
515 >     * Lifecycle control. The low word contains the number of workers
516 >     * that are (probably) executing tasks. This value is atomically
517 >     * incremented before a worker gets a task to run, and decremented
518 >     * when worker has no tasks and cannot find any.  Bits 16-18
519 >     * contain runLevel value. When all are zero, the pool is
520 >     * running. Level transitions are monotonic (running -> shutdown
521 >     * -> terminating -> terminated) so each transition adds a bit.
522 >     * These are bundled together to ensure consistent read for
523 >     * termination checks (i.e., that runLevel is at least SHUTDOWN
524 >     * and active threads is zero).
525 >     *
526 >     * Notes: Most direct CASes are dependent on these bitfield
527 >     * positions.  Also, this field is non-private to enable direct
528 >     * performance-sensitive CASes in ForkJoinWorkerThread.
529       */
530 <    private volatile boolean locallyFifo;
530 >    volatile int runState;
531 >
532 >    // Note: The order among run level values matters.
533 >    private static final int RUNLEVEL_SHIFT     = 16;
534 >    private static final int SHUTDOWN           = 1 << RUNLEVEL_SHIFT;
535 >    private static final int TERMINATING        = 1 << (RUNLEVEL_SHIFT + 1);
536 >    private static final int TERMINATED         = 1 << (RUNLEVEL_SHIFT + 2);
537 >    private static final int ACTIVE_COUNT_MASK  = (1 << RUNLEVEL_SHIFT) - 1;
538  
539      /**
540       * Holds number of total (i.e., created and not yet terminated)
541       * and running (i.e., not blocked on joins or other managed sync)
542 <     * threads, packed into one int to ensure consistent snapshot when
542 >     * threads, packed together to ensure consistent snapshot when
543       * making decisions about creating and suspending spare
544 <     * threads. Updated only by CAS.  Note: CASes in
545 <     * updateRunningCount and preJoin assume that running active count
546 <     * is in low word, so need to be modified if this changes.
544 >     * threads. Updated only by CAS. Note that adding a new worker
545 >     * requires incrementing both counts, since workers start off in
546 >     * running state.
547       */
548      private volatile int workerCounts;
549  
550 <    private static int totalCountOf(int s)           { return s >>> 16;  }
551 <    private static int runningCountOf(int s)         { return s & shortMask; }
552 <    private static int workerCountsFor(int t, int r) { return (t << 16) + r; }
550 >    private static final int TOTAL_COUNT_SHIFT  = 16;
551 >    private static final int RUNNING_COUNT_MASK = (1 << TOTAL_COUNT_SHIFT) - 1;
552 >    private static final int ONE_RUNNING        = 1;
553 >    private static final int ONE_TOTAL          = 1 << TOTAL_COUNT_SHIFT;
554  
555      /**
556 <     * Adds delta (which may be negative) to running count.  This must
557 <     * be called before (with negative arg) and after (with positive)
242 <     * any managed synchronization (i.e., mainly, joins).
243 <     *
244 <     * @param delta the number to add
556 >     * The target parallelism level.
557 >     * Accessed directly by ForkJoinWorkerThreads.
558       */
559 <    final void updateRunningCount(int delta) {
247 <        int s;
248 <        do {} while (!casWorkerCounts(s = workerCounts, s + delta));
249 <    }
559 >    final int parallelism;
560  
561      /**
562 <     * Adds delta (which may be negative) to both total and running
563 <     * count.  This must be called upon creation and termination of
254 <     * worker threads.
255 <     *
256 <     * @param delta the number to add
562 >     * True if use local fifo, not default lifo, for local polling
563 >     * Read by, and replicated by ForkJoinWorkerThreads
564       */
565 <    private void updateWorkerCount(int delta) {
259 <        int d = delta + (delta << 16); // add to both lo and hi parts
260 <        int s;
261 <        do {} while (!casWorkerCounts(s = workerCounts, s + d));
262 <    }
565 >    final boolean locallyFifo;
566  
567      /**
568 <     * Lifecycle control. High word contains runState, low word
569 <     * contains the number of workers that are (probably) executing
267 <     * tasks. This value is atomically incremented before a worker
268 <     * gets a task to run, and decremented when worker has no tasks
269 <     * and cannot find any. These two fields are bundled together to
270 <     * support correct termination triggering.  Note: activeCount
271 <     * CAS'es cheat by assuming active count is in low word, so need
272 <     * to be modified if this changes
568 >     * The uncaught exception handler used when any worker abruptly
569 >     * terminates.
570       */
571 <    private volatile int runControl;
571 >    private final Thread.UncaughtExceptionHandler ueh;
572  
573 <    // RunState values. Order among values matters
574 <    private static final int RUNNING     = 0;
575 <    private static final int SHUTDOWN    = 1;
576 <    private static final int TERMINATING = 2;
280 <    private static final int TERMINATED  = 3;
573 >    /**
574 >     * Pool number, just for assigning useful names to worker threads
575 >     */
576 >    private final int poolNumber;
577  
578 <    private static int runStateOf(int c)             { return c >>> 16; }
579 <    private static int activeCountOf(int c)          { return c & shortMask; }
284 <    private static int runControlFor(int r, int a)   { return (r << 16) + a; }
578 >    // Utilities for CASing fields. Note that most of these
579 >    // are usually manually inlined by callers
580  
581      /**
582 <     * Tries incrementing active count; fails on contention.
288 <     * Called by workers before/during executing tasks.
289 <     *
290 <     * @return true on success
582 >     * Increments running count part of workerCounts
583       */
584 <    final boolean tryIncrementActiveCount() {
585 <        int c = runControl;
586 <        return casRunControl(c, c+1);
584 >    final void incrementRunningCount() {
585 >        int c;
586 >        do {} while (!UNSAFE.compareAndSwapInt(this, workerCountsOffset,
587 >                                               c = workerCounts,
588 >                                               c + ONE_RUNNING));
589      }
590  
591      /**
592 <     * Tries decrementing active count; fails on contention.
299 <     * Possibly triggers termination on success.
300 <     * Called by workers when they can't find tasks.
301 <     *
302 <     * @return true on success
592 >     * Tries to decrement running count unless already zero
593       */
594 <    final boolean tryDecrementActiveCount() {
595 <        int c = runControl;
596 <        int nextc = c - 1;
307 <        if (!casRunControl(c, nextc))
594 >    final boolean tryDecrementRunningCount() {
595 >        int wc = workerCounts;
596 >        if ((wc & RUNNING_COUNT_MASK) == 0)
597              return false;
598 <        if (canTerminateOnShutdown(nextc))
599 <            terminateOnShutdown();
311 <        return true;
598 >        return UNSAFE.compareAndSwapInt(this, workerCountsOffset,
599 >                                        wc, wc - ONE_RUNNING);
600      }
601  
602      /**
603 <     * Returns {@code true} if argument represents zero active count
604 <     * and nonzero runstate, which is the triggering condition for
605 <     * terminating on shutdown.
603 >     * Forces decrement of encoded workerCounts, awaiting nonzero if
604 >     * (rarely) necessary when other count updates lag.
605 >     *
606 >     * @param dr -- either zero or ONE_RUNNING
607 >     * @param dt == either zero or ONE_TOTAL
608       */
609 <    private static boolean canTerminateOnShutdown(int c) {
610 <        // i.e. least bit is nonzero runState bit
611 <        return ((c & -c) >>> 16) != 0;
609 >    private void decrementWorkerCounts(int dr, int dt) {
610 >        for (;;) {
611 >            int wc = workerCounts;
612 >            if ((wc & RUNNING_COUNT_MASK)  - dr < 0 ||
613 >                (wc >>> TOTAL_COUNT_SHIFT) - dt < 0) {
614 >                if ((runState & TERMINATED) != 0)
615 >                    return; // lagging termination on a backout
616 >                Thread.yield();
617 >            }
618 >            if (UNSAFE.compareAndSwapInt(this, workerCountsOffset,
619 >                                         wc, wc - (dr + dt)))
620 >                return;
621 >        }
622 >    }
623 >
624 >    /**
625 >     * Tries decrementing active count; fails on contention.
626 >     * Called when workers cannot find tasks to run.
627 >     */
628 >    final boolean tryDecrementActiveCount() {
629 >        int c;
630 >        return UNSAFE.compareAndSwapInt(this, runStateOffset,
631 >                                        c = runState, c - 1);
632      }
633  
634      /**
635 <     * Transition run state to at least the given state. Return true
636 <     * if not already at least given state.
635 >     * Advances to at least the given level. Returns true if not
636 >     * already in at least the given level.
637       */
638 <    private boolean transitionRunStateTo(int state) {
638 >    private boolean advanceRunLevel(int level) {
639          for (;;) {
640 <            int c = runControl;
641 <            if (runStateOf(c) >= state)
640 >            int s = runState;
641 >            if ((s & level) != 0)
642                  return false;
643 <            if (casRunControl(c, runControlFor(state, activeCountOf(c))))
643 >            if (UNSAFE.compareAndSwapInt(this, runStateOffset, s, s | level))
644                  return true;
645          }
646      }
647  
648 +    // workers array maintenance
649 +
650      /**
651 <     * Controls whether to add spares to maintain parallelism
651 >     * Records and returns a workers array index for new worker.
652       */
653 <    private volatile boolean maintainsParallelism;
653 >    private int recordWorker(ForkJoinWorkerThread w) {
654 >        // Try using slot totalCount-1. If not available, scan and/or resize
655 >        int k = (workerCounts >>> TOTAL_COUNT_SHIFT) - 1;
656 >        final ReentrantLock lock = this.workerLock;
657 >        lock.lock();
658 >        try {
659 >            ForkJoinWorkerThread[] ws = workers;
660 >            int n = ws.length;
661 >            if (k < 0 || k >= n || ws[k] != null) {
662 >                for (k = 0; k < n && ws[k] != null; ++k)
663 >                    ;
664 >                if (k == n)
665 >                    ws = Arrays.copyOf(ws, n << 1);
666 >            }
667 >            ws[k] = w;
668 >            workers = ws; // volatile array write ensures slot visibility
669 >        } finally {
670 >            lock.unlock();
671 >        }
672 >        return k;
673 >    }
674  
675 <    // Constructors
675 >    /**
676 >     * Nulls out record of worker in workers array.
677 >     */
678 >    private void forgetWorker(ForkJoinWorkerThread w) {
679 >        int idx = w.poolIndex;
680 >        // Locking helps method recordWorker avoid unnecessary expansion
681 >        final ReentrantLock lock = this.workerLock;
682 >        lock.lock();
683 >        try {
684 >            ForkJoinWorkerThread[] ws = workers;
685 >            if (idx >= 0 && idx < ws.length && ws[idx] == w) // verify
686 >                ws[idx] = null;
687 >        } finally {
688 >            lock.unlock();
689 >        }
690 >    }
691  
692      /**
693 <     * Creates a ForkJoinPool with a pool size equal to the number of
694 <     * processors available on the system, using the default
695 <     * ForkJoinWorkerThreadFactory.
693 >     * Final callback from terminating worker.  Removes record of
694 >     * worker from array, and adjusts counts. If pool is shutting
695 >     * down, tries to complete termination.
696       *
697 <     * @throws SecurityException if a security manager exists and
351 <     *         the caller is not permitted to modify threads
352 <     *         because it does not hold {@link
353 <     *         java.lang.RuntimePermission}{@code ("modifyThread")}
697 >     * @param w the worker
698       */
699 <    public ForkJoinPool() {
700 <        this(Runtime.getRuntime().availableProcessors(),
701 <             defaultForkJoinWorkerThreadFactory);
699 >    final void workerTerminated(ForkJoinWorkerThread w) {
700 >        forgetWorker(w);
701 >        decrementWorkerCounts(w.isTrimmed()? 0 : ONE_RUNNING, ONE_TOTAL);
702 >        while (w.stealCount != 0) // collect final count
703 >            tryAccumulateStealCount(w);
704 >        tryTerminate(false);
705      }
706  
707 +    // Waiting for and signalling events
708 +
709      /**
710 <     * Creates a ForkJoinPool with the indicated parallelism level
711 <     * threads and using the default ForkJoinWorkerThreadFactory.
712 <     *
713 <     * @param parallelism the number of worker threads
365 <     * @throws IllegalArgumentException if parallelism less than or
366 <     * equal to zero
367 <     * @throws SecurityException if a security manager exists and
368 <     *         the caller is not permitted to modify threads
369 <     *         because it does not hold {@link
370 <     *         java.lang.RuntimePermission}{@code ("modifyThread")}
710 >     * Releases workers blocked on a count not equal to current count.
711 >     * Normally called after precheck that eventWaiters isn't zero to
712 >     * avoid wasted array checks. Gives up upon a change in count or
713 >     * upon releasing two workers, letting others take over.
714       */
715 <    public ForkJoinPool(int parallelism) {
716 <        this(parallelism, defaultForkJoinWorkerThreadFactory);
715 >    private void releaseEventWaiters() {
716 >        ForkJoinWorkerThread[] ws = workers;
717 >        int n = ws.length;
718 >        long h = eventWaiters;
719 >        int ec = eventCount;
720 >        boolean releasedOne = false;
721 >        ForkJoinWorkerThread w; int id;
722 >        while ((id = ((int)(h & WAITER_ID_MASK)) - 1) >= 0 &&
723 >               (int)(h >>> EVENT_COUNT_SHIFT) != ec &&
724 >               id < n && (w = ws[id]) != null) {
725 >            if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
726 >                                          h,  w.nextWaiter)) {
727 >                LockSupport.unpark(w);
728 >                if (releasedOne) // exit on second release
729 >                    break;
730 >                releasedOne = true;
731 >            }
732 >            if (eventCount != ec)
733 >                break;
734 >            h = eventWaiters;
735 >        }
736      }
737  
738      /**
739 <     * Creates a ForkJoinPool with parallelism equal to the number of
740 <     * processors available on the system and using the given
379 <     * ForkJoinWorkerThreadFactory.
380 <     *
381 <     * @param factory the factory for creating new threads
382 <     * @throws NullPointerException if factory is null
383 <     * @throws SecurityException if a security manager exists and
384 <     *         the caller is not permitted to modify threads
385 <     *         because it does not hold {@link
386 <     *         java.lang.RuntimePermission}{@code ("modifyThread")}
739 >     * Tries to advance eventCount and releases waiters. Called only
740 >     * from workers.
741       */
742 <    public ForkJoinPool(ForkJoinWorkerThreadFactory factory) {
743 <        this(Runtime.getRuntime().availableProcessors(), factory);
742 >    final void signalWork() {
743 >        int c; // try to increment event count -- CAS failure OK
744 >        UNSAFE.compareAndSwapInt(this, eventCountOffset, c = eventCount, c+1);
745 >        if (eventWaiters != 0L)
746 >            releaseEventWaiters();
747      }
748  
749      /**
750 <     * Creates a ForkJoinPool with the given parallelism and factory.
750 >     * Adds the given worker to event queue and blocks until
751 >     * terminating or event count advances from the given value
752       *
753 <     * @param parallelism the targeted number of worker threads
754 <     * @param factory the factory for creating new threads
397 <     * @throws IllegalArgumentException if parallelism less than or
398 <     * equal to zero, or greater than implementation limit
399 <     * @throws NullPointerException if factory is null
400 <     * @throws SecurityException if a security manager exists and
401 <     *         the caller is not permitted to modify threads
402 <     *         because it does not hold {@link
403 <     *         java.lang.RuntimePermission}{@code ("modifyThread")}
753 >     * @param w the calling worker thread
754 >     * @param ec the count
755       */
756 <    public ForkJoinPool(int parallelism, ForkJoinWorkerThreadFactory factory) {
757 <        if (parallelism <= 0 || parallelism > MAX_THREADS)
758 <            throw new IllegalArgumentException();
759 <        if (factory == null)
760 <            throw new NullPointerException();
761 <        checkPermission();
762 <        this.factory = factory;
763 <        this.parallelism = parallelism;
764 <        this.maxPoolSize = MAX_THREADS;
765 <        this.maintainsParallelism = true;
766 <        this.poolNumber = poolNumberGenerator.incrementAndGet();
767 <        this.workerLock = new ReentrantLock();
768 <        this.termination = workerLock.newCondition();
418 <        this.stealCount = new AtomicLong();
419 <        this.submissionQueue = new LinkedTransferQueue<ForkJoinTask<?>>();
420 <        // worker array and workers are lazily constructed
756 >    private void eventSync(ForkJoinWorkerThread w, int ec) {
757 >        long nh = (((long)ec) << EVENT_COUNT_SHIFT) | ((long)(w.poolIndex+1));
758 >        long h;
759 >        while ((runState < SHUTDOWN || !tryTerminate(false)) &&
760 >               (((int)((h = eventWaiters) & WAITER_ID_MASK)) == 0 ||
761 >                (int)(h >>> EVENT_COUNT_SHIFT) == ec) &&
762 >               eventCount == ec) {
763 >            if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
764 >                                          w.nextWaiter = h, nh)) {
765 >                awaitEvent(w, ec);
766 >                break;
767 >            }
768 >        }
769      }
770  
771      /**
772 <     * Creates a new worker thread using factory.
772 >     * Blocks the given worker (that has already been entered as an
773 >     * event waiter) until terminating or event count advances from
774 >     * the given value. The oldest (first) waiter uses a timed wait to
775 >     * occasionally one-by-one shrink the number of workers (to a
776 >     * minimum of one) if the pool has not been used for extended
777 >     * periods.
778       *
779 <     * @param index the index to assign worker
780 <     * @return new worker, or null of factory failed
779 >     * @param w the calling worker thread
780 >     * @param ec the count
781       */
782 <    private ForkJoinWorkerThread createWorker(int index) {
783 <        Thread.UncaughtExceptionHandler h = ueh;
784 <        ForkJoinWorkerThread w = factory.newThread(this);
785 <        if (w != null) {
786 <            w.poolIndex = index;
787 <            w.setDaemon(true);
788 <            w.setAsyncMode(locallyFifo);
789 <            w.setName("ForkJoinPool-" + poolNumber + "-worker-" + index);
790 <            if (h != null)
791 <                w.setUncaughtExceptionHandler(h);
782 >    private void awaitEvent(ForkJoinWorkerThread w, int ec) {
783 >        while (eventCount == ec) {
784 >            if (tryAccumulateStealCount(w)) { // transfer while idle
785 >                boolean untimed = (w.nextWaiter != 0L ||
786 >                                   (workerCounts & RUNNING_COUNT_MASK) <= 1);
787 >                long startTime = untimed? 0 : System.nanoTime();
788 >                Thread.interrupted();         // clear/ignore interrupt
789 >                if (eventCount != ec || w.runState != 0 ||
790 >                    runState >= TERMINATING)  // recheck after clear
791 >                    break;
792 >                if (untimed)
793 >                    LockSupport.park(w);
794 >                else {
795 >                    LockSupport.parkNanos(w, SHRINK_RATE_NANOS);
796 >                    if (eventCount != ec || w.runState != 0 ||
797 >                        runState >= TERMINATING)
798 >                        break;
799 >                    if (System.nanoTime() - startTime >= SHRINK_RATE_NANOS)
800 >                        tryShutdownUnusedWorker(ec);
801 >                }
802 >            }
803          }
440        return w;
804      }
805  
806 +    // Maintaining parallelism
807 +
808      /**
809 <     * Returns a good size for worker array given pool size.
445 <     * Currently requires size to be a power of two.
809 >     * Pushes worker onto the spare stack.
810       */
811 <    private static int arraySizeFor(int poolSize) {
812 <        return (poolSize <= 1) ? 1 :
813 <            (1 << (32 - Integer.numberOfLeadingZeros(poolSize-1)));
811 >    final void pushSpare(ForkJoinWorkerThread w) {
812 >        int ns = (++w.spareCount << SPARE_COUNT_SHIFT) | (w.poolIndex + 1);
813 >        do {} while (!UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
814 >                                               w.nextSpare = spareWaiters,ns));
815      }
816  
817      /**
818 <     * Creates or resizes array if necessary to hold newLength.
819 <     * Call only under exclusion.
455 <     *
456 <     * @return the array
818 >     * Tries (once) to resume a spare if the number of running
819 >     * threads is less than target.
820       */
821 <    private ForkJoinWorkerThread[] ensureWorkerArrayCapacity(int newLength) {
821 >    private void tryResumeSpare() {
822 >        int sw, id;
823          ForkJoinWorkerThread[] ws = workers;
824 <        if (ws == null)
825 <            return workers = new ForkJoinWorkerThread[arraySizeFor(newLength)];
826 <        else if (newLength > ws.length)
827 <            return workers = Arrays.copyOf(ws, arraySizeFor(newLength));
828 <        else
829 <            return ws;
824 >        int n = ws.length;
825 >        ForkJoinWorkerThread w;
826 >        if ((sw = spareWaiters) != 0 &&
827 >            (id = (sw & SPARE_ID_MASK) - 1) >= 0 &&
828 >            id < n && (w = ws[id]) != null &&
829 >            (workerCounts & RUNNING_COUNT_MASK) < parallelism &&
830 >            spareWaiters == sw &&
831 >            UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
832 >                                     sw, w.nextSpare)) {
833 >            int c; // increment running count before resume
834 >            do {} while (!UNSAFE.compareAndSwapInt
835 >                         (this, workerCountsOffset,
836 >                          c = workerCounts, c + ONE_RUNNING));
837 >            if (w.tryUnsuspend())
838 >                LockSupport.unpark(w);
839 >            else   // back out if w was shutdown
840 >                decrementWorkerCounts(ONE_RUNNING, 0);
841 >        }
842      }
843  
844      /**
845 <     * Tries to shrink workers into smaller array after one or more terminate.
846 <     */
847 <    private void tryShrinkWorkerArray() {
848 <        ForkJoinWorkerThread[] ws = workers;
849 <        if (ws != null) {
850 <            int len = ws.length;
851 <            int last = len - 1;
852 <            while (last >= 0 && ws[last] == null)
853 <                --last;
854 <            int newLength = arraySizeFor(last+1);
855 <            if (newLength < len)
856 <                workers = Arrays.copyOf(ws, newLength);
845 >     * Tries to increase the number of running workers if below target
846 >     * parallelism: If a spare exists tries to resume it via
847 >     * tryResumeSpare.  Otherwise, if not enough total workers or all
848 >     * existing workers are busy, adds a new worker. In all cases also
849 >     * helps wake up releasable workers waiting for work.
850 >     */
851 >    private void helpMaintainParallelism() {
852 >        int pc = parallelism;
853 >        int wc, rs, tc;
854 >        while (((wc = workerCounts) & RUNNING_COUNT_MASK) < pc &&
855 >               (rs = runState) < TERMINATING) {
856 >            if (spareWaiters != 0)
857 >                tryResumeSpare();
858 >            else if ((tc = wc >>> TOTAL_COUNT_SHIFT) >= MAX_WORKERS ||
859 >                     (tc >= pc && (rs & ACTIVE_COUNT_MASK) != tc))
860 >                break;   // enough total
861 >            else if (runState == rs && workerCounts == wc &&
862 >                     UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
863 >                                              wc + (ONE_RUNNING|ONE_TOTAL))) {
864 >                ForkJoinWorkerThread w = null;
865 >                try {
866 >                    w = factory.newThread(this);
867 >                } finally { // adjust on null or exceptional factory return
868 >                    if (w == null) {
869 >                        decrementWorkerCounts(ONE_RUNNING, ONE_TOTAL);
870 >                        tryTerminate(false); // handle failure during shutdown
871 >                    }
872 >                }
873 >                if (w == null)
874 >                    break;
875 >                w.start(recordWorker(w), ueh);
876 >                if ((workerCounts >>> TOTAL_COUNT_SHIFT) >= pc) {
877 >                    int c; // advance event count
878 >                    UNSAFE.compareAndSwapInt(this, eventCountOffset,
879 >                                             c = eventCount, c+1);
880 >                    break; // add at most one unless total below target
881 >                }
882 >            }
883          }
884 +        if (eventWaiters != 0L)
885 +            releaseEventWaiters();
886      }
887  
888      /**
889 <     * Initializes workers if necessary.
889 >     * Callback from the oldest waiter in awaitEvent waking up after a
890 >     * period of non-use. If all workers are idle, tries (once) to
891 >     * shutdown an event waiter or a spare, if one exists. Note that
892 >     * we don't need CAS or locks here because the method is called
893 >     * only from one thread occasionally waking (and even misfires are
894 >     * OK). Note that until the shutdown worker fully terminates,
895 >     * workerCounts will overestimate total count, which is tolerable.
896 >     *
897 >     * @param ec the event count waited on by caller (to abort
898 >     * attempt if count has since changed).
899       */
900 <    final void ensureWorkerInitialization() {
901 <        ForkJoinWorkerThread[] ws = workers;
902 <        if (ws == null) {
903 <            final ReentrantLock lock = this.workerLock;
904 <            lock.lock();
905 <            try {
906 <                ws = workers;
907 <                if (ws == null) {
908 <                    int ps = parallelism;
909 <                    ws = ensureWorkerArrayCapacity(ps);
910 <                    for (int i = 0; i < ps; ++i) {
911 <                        ForkJoinWorkerThread w = createWorker(i);
912 <                        if (w != null) {
913 <                            ws[i] = w;
914 <                            w.start();
915 <                            updateWorkerCount(1);
916 <                        }
917 <                    }
900 >    private void tryShutdownUnusedWorker(int ec) {
901 >        if (runState == 0 && eventCount == ec) { // only trigger if all idle
902 >            ForkJoinWorkerThread[] ws = workers;
903 >            int n = ws.length;
904 >            ForkJoinWorkerThread w = null;
905 >            boolean shutdown = false;
906 >            int sw;
907 >            long h;
908 >            if ((sw = spareWaiters) != 0) { // prefer killing spares
909 >                int id = (sw & SPARE_ID_MASK) - 1;
910 >                if (id >= 0 && id < n && (w = ws[id]) != null &&
911 >                    UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
912 >                                             sw, w.nextSpare))
913 >                    shutdown = true;
914 >            }
915 >            else if ((h = eventWaiters) != 0L) {
916 >                long nh;
917 >                int id = ((int)(h & WAITER_ID_MASK)) - 1;
918 >                if (id >= 0 && id < n && (w = ws[id]) != null &&
919 >                    (nh = w.nextWaiter) != 0L && // keep at least one worker
920 >                    UNSAFE.compareAndSwapLong(this, eventWaitersOffset, h, nh))
921 >                    shutdown = true;
922 >            }
923 >            if (w != null && shutdown) {
924 >                w.shutdown();
925 >                LockSupport.unpark(w);
926 >            }
927 >        }
928 >        releaseEventWaiters(); // in case of interference
929 >    }
930 >
931 >    /**
932 >     * Callback from workers invoked upon each top-level action (i.e.,
933 >     * stealing a task or taking a submission and running it).
934 >     * Performs one or more of the following:
935 >     *
936 >     * 1. If the worker is active and either did not run a task
937 >     *    or there are too many workers, try to set its active status
938 >     *    to inactive and update activeCount. On contention, we may
939 >     *    try again in this or a subsequent call.
940 >     *
941 >     * 2. If not enough total workers, help create some.
942 >     *
943 >     * 3. If there are too many running workers, suspend this worker
944 >     *    (first forcing inactive if necessary).  If it is not needed,
945 >     *    it may be shutdown while suspended (via
946 >     *    tryShutdownUnusedWorker).  Otherwise, upon resume it
947 >     *    rechecks running thread count and need for event sync.
948 >     *
949 >     * 4. If worker did not run a task, await the next task event via
950 >     *    eventSync if necessary (first forcing inactivation), upon
951 >     *    which the worker may be shutdown via
952 >     *    tryShutdownUnusedWorker.  Otherwise, help release any
953 >     *    existing event waiters that are now releasable,
954 >     *
955 >     * @param w the worker
956 >     * @param ran true if worker ran a task since last call to this method
957 >     */
958 >    final void preStep(ForkJoinWorkerThread w, boolean ran) {
959 >        int wec = w.lastEventCount;
960 >        boolean active = w.active;
961 >        boolean inactivate = false;
962 >        int pc = parallelism;
963 >        int rs;
964 >        while (w.runState == 0 && (rs = runState) < TERMINATING) {
965 >            if ((inactivate || (active && (rs & ACTIVE_COUNT_MASK) >= pc)) &&
966 >                UNSAFE.compareAndSwapInt(this, runStateOffset, rs, rs - 1))
967 >                inactivate = active = w.active = false;
968 >            int wc = workerCounts;
969 >            if ((wc & RUNNING_COUNT_MASK) > pc) {
970 >                if (!(inactivate |= active) && // must inactivate to suspend
971 >                    workerCounts == wc &&      // try to suspend as spare
972 >                    UNSAFE.compareAndSwapInt(this, workerCountsOffset,
973 >                                             wc, wc - ONE_RUNNING))
974 >                    w.suspendAsSpare();
975 >            }
976 >            else if ((wc >>> TOTAL_COUNT_SHIFT) < pc)
977 >                helpMaintainParallelism();     // not enough workers
978 >            else if (!ran) {
979 >                long h = eventWaiters;
980 >                int ec = eventCount;
981 >                if (h != 0L && (int)(h >>> EVENT_COUNT_SHIFT) != ec)
982 >                    releaseEventWaiters();     // release others before waiting
983 >                else if (ec != wec) {
984 >                    w.lastEventCount = ec;     // no need to wait
985 >                    break;
986                  }
987 <            } finally {
988 <                lock.unlock();
987 >                else if (!(inactivate |= active))
988 >                    eventSync(w, wec);         // must inactivate before sync
989              }
990 +            else
991 +                break;
992          }
993      }
994  
995      /**
996 <     * Worker creation and startup for threads added via setParallelism.
996 >     * Helps and/or blocks awaiting join of the given task.
997 >     * See above for explanation.
998 >     *
999 >     * @param joinMe the task to join
1000 >     * @param worker the current worker thread
1001       */
1002 <    private void createAndStartAddedWorkers() {
1003 <        resumeAllSpares();  // Allow spares to convert to nonspare
1004 <        int ps = parallelism;
1005 <        ForkJoinWorkerThread[] ws = ensureWorkerArrayCapacity(ps);
1006 <        int len = ws.length;
1007 <        // Sweep through slots, to keep lowest indices most populated
1008 <        int k = 0;
1009 <        while (k < len) {
1010 <            if (ws[k] != null) {
1011 <                ++k;
1012 <                continue;
1002 >    final void awaitJoin(ForkJoinTask<?> joinMe, ForkJoinWorkerThread worker) {
1003 >        int retries = 2 + (parallelism >> 2); // #helpJoins before blocking
1004 >        while (joinMe.status >= 0) {
1005 >            int wc;
1006 >            worker.helpJoinTask(joinMe);
1007 >            if (joinMe.status < 0)
1008 >                break;
1009 >            else if (retries > 0)
1010 >                --retries;
1011 >            else if (((wc = workerCounts) & RUNNING_COUNT_MASK) != 0 &&
1012 >                     UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1013 >                                              wc, wc - ONE_RUNNING)) {
1014 >                int stat, c; long h;
1015 >                while ((stat = joinMe.status) >= 0 &&
1016 >                       (h = eventWaiters) != 0L && // help release others
1017 >                       (int)(h >>> EVENT_COUNT_SHIFT) != eventCount)
1018 >                    releaseEventWaiters();
1019 >                if (stat >= 0 &&
1020 >                    ((workerCounts & RUNNING_COUNT_MASK) == 0 ||
1021 >                     (stat =
1022 >                      joinMe.internalAwaitDone(JOIN_TIMEOUT_MILLIS)) >= 0))
1023 >                    helpMaintainParallelism(); // timeout or no running workers
1024 >                do {} while (!UNSAFE.compareAndSwapInt
1025 >                             (this, workerCountsOffset,
1026 >                              c = workerCounts, c + ONE_RUNNING));
1027 >                if (stat < 0)
1028 >                    break;   // else restart
1029              }
1030 <            int s = workerCounts;
1031 <            int tc = totalCountOf(s);
1032 <            int rc = runningCountOf(s);
1033 <            if (rc >= ps || tc >= ps)
1030 >        }
1031 >    }
1032 >
1033 >    /**
1034 >     * Same idea as awaitJoin, but no helping, retries, or timeouts.
1035 >     */
1036 >    final void awaitBlocker(ManagedBlocker blocker)
1037 >        throws InterruptedException {
1038 >        while (!blocker.isReleasable()) {
1039 >            int wc = workerCounts;
1040 >            if ((wc & RUNNING_COUNT_MASK) != 0 &&
1041 >                UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1042 >                                         wc, wc - ONE_RUNNING)) {
1043 >                try {
1044 >                    while (!blocker.isReleasable()) {
1045 >                        long h = eventWaiters;
1046 >                        if (h != 0L &&
1047 >                            (int)(h >>> EVENT_COUNT_SHIFT) != eventCount)
1048 >                            releaseEventWaiters();
1049 >                        else if ((workerCounts & RUNNING_COUNT_MASK) == 0 &&
1050 >                                 runState < TERMINATING)
1051 >                            helpMaintainParallelism();
1052 >                        else if (blocker.block())
1053 >                            break;
1054 >                    }
1055 >                } finally {
1056 >                    int c;
1057 >                    do {} while (!UNSAFE.compareAndSwapInt
1058 >                                 (this, workerCountsOffset,
1059 >                                  c = workerCounts, c + ONE_RUNNING));
1060 >                }
1061                  break;
1062 <            if (casWorkerCounts (s, workerCountsFor(tc+1, rc+1))) {
1063 <                ForkJoinWorkerThread w = createWorker(k);
1062 >            }
1063 >        }
1064 >    }
1065 >
1066 >    /**
1067 >     * Possibly initiates and/or completes termination.
1068 >     *
1069 >     * @param now if true, unconditionally terminate, else only
1070 >     * if shutdown and empty queue and no active workers
1071 >     * @return true if now terminating or terminated
1072 >     */
1073 >    private boolean tryTerminate(boolean now) {
1074 >        if (now)
1075 >            advanceRunLevel(SHUTDOWN); // ensure at least SHUTDOWN
1076 >        else if (runState < SHUTDOWN ||
1077 >                 !submissionQueue.isEmpty() ||
1078 >                 (runState & ACTIVE_COUNT_MASK) != 0)
1079 >            return false;
1080 >
1081 >        if (advanceRunLevel(TERMINATING))
1082 >            startTerminating();
1083 >
1084 >        // Finish now if all threads terminated; else in some subsequent call
1085 >        if ((workerCounts >>> TOTAL_COUNT_SHIFT) == 0) {
1086 >            advanceRunLevel(TERMINATED);
1087 >            termination.arrive();
1088 >        }
1089 >        return true;
1090 >    }
1091 >
1092 >    /**
1093 >     * Actions on transition to TERMINATING
1094 >     *
1095 >     * Runs up to four passes through workers: (0) shutting down each
1096 >     * (without waking up if parked) to quickly spread notifications
1097 >     * without unnecessary bouncing around event queues etc (1) wake
1098 >     * up and help cancel tasks (2) interrupt (3) mop up races with
1099 >     * interrupted workers
1100 >     */
1101 >    private void startTerminating() {
1102 >        cancelSubmissions();
1103 >        for (int passes = 0; passes < 4 && workerCounts != 0; ++passes) {
1104 >            int c; // advance event count
1105 >            UNSAFE.compareAndSwapInt(this, eventCountOffset,
1106 >                                     c = eventCount, c+1);
1107 >            eventWaiters = 0L; // clobber lists
1108 >            spareWaiters = 0;
1109 >            for (ForkJoinWorkerThread w : workers) {
1110                  if (w != null) {
1111 <                    ws[k++] = w;
1112 <                    w.start();
1113 <                }
1114 <                else {
1115 <                    updateWorkerCount(-1); // back out on failed creation
1116 <                    break;
1111 >                    w.shutdown();
1112 >                    if (passes > 0 && !w.isTerminated()) {
1113 >                        w.cancelTasks();
1114 >                        LockSupport.unpark(w);
1115 >                        if (passes > 1) {
1116 >                            try {
1117 >                                w.interrupt();
1118 >                            } catch (SecurityException ignore) {
1119 >                            }
1120 >                        }
1121 >                    }
1122                  }
1123              }
1124          }
1125      }
1126  
1127 +    /**
1128 +     * Clears out and cancels submissions, ignoring exceptions.
1129 +     */
1130 +    private void cancelSubmissions() {
1131 +        ForkJoinTask<?> task;
1132 +        while ((task = submissionQueue.poll()) != null) {
1133 +            try {
1134 +                task.cancel(false);
1135 +            } catch (Throwable ignore) {
1136 +            }
1137 +        }
1138 +    }
1139 +
1140 +    // misc support for ForkJoinWorkerThread
1141 +
1142 +    /**
1143 +     * Returns pool number.
1144 +     */
1145 +    final int getPoolNumber() {
1146 +        return poolNumber;
1147 +    }
1148 +
1149 +    /**
1150 +     * Tries to accumulate steal count from a worker, clearing
1151 +     * the worker's value if successful.
1152 +     *
1153 +     * @return true if worker steal count now zero
1154 +     */
1155 +    final boolean tryAccumulateStealCount(ForkJoinWorkerThread w) {
1156 +        int sc = w.stealCount;
1157 +        long c = stealCount;
1158 +        // CAS even if zero, for fence effects
1159 +        if (UNSAFE.compareAndSwapLong(this, stealCountOffset, c, c + sc)) {
1160 +            if (sc != 0)
1161 +                w.stealCount = 0;
1162 +            return true;
1163 +        }
1164 +        return sc == 0;
1165 +    }
1166 +
1167 +    /**
1168 +     * Returns the approximate (non-atomic) number of idle threads per
1169 +     * active thread.
1170 +     */
1171 +    final int idlePerActive() {
1172 +        int pc = parallelism; // use parallelism, not rc
1173 +        int ac = runState;    // no mask -- artificially boosts during shutdown
1174 +        // Use exact results for small values, saturate past 4
1175 +        return ((pc <= ac) ? 0 :
1176 +                (pc >>> 1 <= ac) ? 1 :
1177 +                (pc >>> 2 <= ac) ? 3 :
1178 +                pc >>> 3);
1179 +    }
1180 +
1181 +    // Public and protected methods
1182 +
1183 +    // Constructors
1184 +
1185 +    /**
1186 +     * Creates a {@code ForkJoinPool} with parallelism equal to {@link
1187 +     * java.lang.Runtime#availableProcessors}, using the {@linkplain
1188 +     * #defaultForkJoinWorkerThreadFactory default thread factory},
1189 +     * no UncaughtExceptionHandler, and non-async LIFO processing mode.
1190 +     *
1191 +     * @throws SecurityException if a security manager exists and
1192 +     *         the caller is not permitted to modify threads
1193 +     *         because it does not hold {@link
1194 +     *         java.lang.RuntimePermission}{@code ("modifyThread")}
1195 +     */
1196 +    public ForkJoinPool() {
1197 +        this(Runtime.getRuntime().availableProcessors(),
1198 +             defaultForkJoinWorkerThreadFactory, null, false);
1199 +    }
1200 +
1201 +    /**
1202 +     * Creates a {@code ForkJoinPool} with the indicated parallelism
1203 +     * level, the {@linkplain
1204 +     * #defaultForkJoinWorkerThreadFactory default thread factory},
1205 +     * no UncaughtExceptionHandler, and non-async LIFO processing mode.
1206 +     *
1207 +     * @param parallelism the parallelism level
1208 +     * @throws IllegalArgumentException if parallelism less than or
1209 +     *         equal to zero, or greater than implementation limit
1210 +     * @throws SecurityException if a security manager exists and
1211 +     *         the caller is not permitted to modify threads
1212 +     *         because it does not hold {@link
1213 +     *         java.lang.RuntimePermission}{@code ("modifyThread")}
1214 +     */
1215 +    public ForkJoinPool(int parallelism) {
1216 +        this(parallelism, defaultForkJoinWorkerThreadFactory, null, false);
1217 +    }
1218 +
1219 +    /**
1220 +     * Creates a {@code ForkJoinPool} with the given parameters.
1221 +     *
1222 +     * @param parallelism the parallelism level. For default value,
1223 +     * use {@link java.lang.Runtime#availableProcessors}.
1224 +     * @param factory the factory for creating new threads. For default value,
1225 +     * use {@link #defaultForkJoinWorkerThreadFactory}.
1226 +     * @param handler the handler for internal worker threads that
1227 +     * terminate due to unrecoverable errors encountered while executing
1228 +     * tasks. For default value, use {@code null}.
1229 +     * @param asyncMode if true,
1230 +     * establishes local first-in-first-out scheduling mode for forked
1231 +     * tasks that are never joined. This mode may be more appropriate
1232 +     * than default locally stack-based mode in applications in which
1233 +     * worker threads only process event-style asynchronous tasks.
1234 +     * For default value, use {@code false}.
1235 +     * @throws IllegalArgumentException if parallelism less than or
1236 +     *         equal to zero, or greater than implementation limit
1237 +     * @throws NullPointerException if the factory is null
1238 +     * @throws SecurityException if a security manager exists and
1239 +     *         the caller is not permitted to modify threads
1240 +     *         because it does not hold {@link
1241 +     *         java.lang.RuntimePermission}{@code ("modifyThread")}
1242 +     */
1243 +    public ForkJoinPool(int parallelism,
1244 +                        ForkJoinWorkerThreadFactory factory,
1245 +                        Thread.UncaughtExceptionHandler handler,
1246 +                        boolean asyncMode) {
1247 +        checkPermission();
1248 +        if (factory == null)
1249 +            throw new NullPointerException();
1250 +        if (parallelism <= 0 || parallelism > MAX_WORKERS)
1251 +            throw new IllegalArgumentException();
1252 +        this.parallelism = parallelism;
1253 +        this.factory = factory;
1254 +        this.ueh = handler;
1255 +        this.locallyFifo = asyncMode;
1256 +        int arraySize = initialArraySizeFor(parallelism);
1257 +        this.workers = new ForkJoinWorkerThread[arraySize];
1258 +        this.submissionQueue = new LinkedTransferQueue<ForkJoinTask<?>>();
1259 +        this.workerLock = new ReentrantLock();
1260 +        this.termination = new Phaser(1);
1261 +        this.poolNumber = poolNumberGenerator.incrementAndGet();
1262 +    }
1263 +
1264 +    /**
1265 +     * Returns initial power of two size for workers array.
1266 +     * @param pc the initial parallelism level
1267 +     */
1268 +    private static int initialArraySizeFor(int pc) {
1269 +        // If possible, initially allocate enough space for one spare
1270 +        int size = pc < MAX_WORKERS ? pc + 1 : MAX_WORKERS;
1271 +        // See Hackers Delight, sec 3.2. We know MAX_WORKERS < (1 >>> 16)
1272 +        size |= size >>> 1;
1273 +        size |= size >>> 2;
1274 +        size |= size >>> 4;
1275 +        size |= size >>> 8;
1276 +        return size + 1;
1277 +    }
1278 +
1279      // Execution methods
1280  
1281      /**
# Line 551 | Line 1284 | public class ForkJoinPool extends Abstra
1284      private <T> void doSubmit(ForkJoinTask<T> task) {
1285          if (task == null)
1286              throw new NullPointerException();
1287 <        if (isShutdown())
1287 >        if (runState >= SHUTDOWN)
1288              throw new RejectedExecutionException();
556        if (workers == null)
557            ensureWorkerInitialization();
1289          submissionQueue.offer(task);
1290 <        signalIdleWorkers();
1290 >        int c; // try to increment event count -- CAS failure OK
1291 >        UNSAFE.compareAndSwapInt(this, eventCountOffset, c = eventCount, c+1);
1292 >        helpMaintainParallelism(); // create, start, or resume some workers
1293      }
1294  
1295      /**
# Line 564 | Line 1297 | public class ForkJoinPool extends Abstra
1297       *
1298       * @param task the task
1299       * @return the task's result
1300 <     * @throws NullPointerException if task is null
1301 <     * @throws RejectedExecutionException if pool is shut down
1300 >     * @throws NullPointerException if the task is null
1301 >     * @throws RejectedExecutionException if the task cannot be
1302 >     *         scheduled for execution
1303       */
1304      public <T> T invoke(ForkJoinTask<T> task) {
1305          doSubmit(task);
# Line 576 | Line 1310 | public class ForkJoinPool extends Abstra
1310       * Arranges for (asynchronous) execution of the given task.
1311       *
1312       * @param task the task
1313 <     * @throws NullPointerException if task is null
1314 <     * @throws RejectedExecutionException if pool is shut down
1313 >     * @throws NullPointerException if the task is null
1314 >     * @throws RejectedExecutionException if the task cannot be
1315 >     *         scheduled for execution
1316       */
1317      public void execute(ForkJoinTask<?> task) {
1318          doSubmit(task);
# Line 585 | Line 1320 | public class ForkJoinPool extends Abstra
1320  
1321      // AbstractExecutorService methods
1322  
1323 +    /**
1324 +     * @throws NullPointerException if the task is null
1325 +     * @throws RejectedExecutionException if the task cannot be
1326 +     *         scheduled for execution
1327 +     */
1328      public void execute(Runnable task) {
1329          ForkJoinTask<?> job;
1330          if (task instanceof ForkJoinTask<?>) // avoid re-wrap
# Line 594 | Line 1334 | public class ForkJoinPool extends Abstra
1334          doSubmit(job);
1335      }
1336  
1337 +    /**
1338 +     * Submits a ForkJoinTask for execution.
1339 +     *
1340 +     * @param task the task to submit
1341 +     * @return the task
1342 +     * @throws NullPointerException if the task is null
1343 +     * @throws RejectedExecutionException if the task cannot be
1344 +     *         scheduled for execution
1345 +     */
1346 +    public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) {
1347 +        doSubmit(task);
1348 +        return task;
1349 +    }
1350 +
1351 +    /**
1352 +     * @throws NullPointerException if the task is null
1353 +     * @throws RejectedExecutionException if the task cannot be
1354 +     *         scheduled for execution
1355 +     */
1356      public <T> ForkJoinTask<T> submit(Callable<T> task) {
1357          ForkJoinTask<T> job = ForkJoinTask.adapt(task);
1358          doSubmit(job);
1359          return job;
1360      }
1361  
1362 +    /**
1363 +     * @throws NullPointerException if the task is null
1364 +     * @throws RejectedExecutionException if the task cannot be
1365 +     *         scheduled for execution
1366 +     */
1367      public <T> ForkJoinTask<T> submit(Runnable task, T result) {
1368          ForkJoinTask<T> job = ForkJoinTask.adapt(task, result);
1369          doSubmit(job);
1370          return job;
1371      }
1372  
1373 +    /**
1374 +     * @throws NullPointerException if the task is null
1375 +     * @throws RejectedExecutionException if the task cannot be
1376 +     *         scheduled for execution
1377 +     */
1378      public ForkJoinTask<?> submit(Runnable task) {
1379          ForkJoinTask<?> job;
1380          if (task instanceof ForkJoinTask<?>) // avoid re-wrap
# Line 617 | Line 1386 | public class ForkJoinPool extends Abstra
1386      }
1387  
1388      /**
1389 <     * Submits a ForkJoinTask for execution.
1390 <     *
622 <     * @param task the task to submit
623 <     * @return the task
624 <     * @throws RejectedExecutionException if the task cannot be
625 <     *         scheduled for execution
626 <     * @throws NullPointerException if the task is null
1389 >     * @throws NullPointerException       {@inheritDoc}
1390 >     * @throws RejectedExecutionException {@inheritDoc}
1391       */
628    public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) {
629        doSubmit(task);
630        return task;
631    }
632
633
1392      public <T> List<Future<T>> invokeAll(Collection<? extends Callable<T>> tasks) {
1393          ArrayList<ForkJoinTask<T>> forkJoinTasks =
1394              new ArrayList<ForkJoinTask<T>>(tasks.size());
# Line 639 | Line 1397 | public class ForkJoinPool extends Abstra
1397          invoke(new InvokeAll<T>(forkJoinTasks));
1398  
1399          @SuppressWarnings({"unchecked", "rawtypes"})
1400 <        List<Future<T>> futures = (List<Future<T>>) (List) forkJoinTasks;
1400 >            List<Future<T>> futures = (List<Future<T>>) (List) forkJoinTasks;
1401          return futures;
1402      }
1403  
# Line 653 | Line 1411 | public class ForkJoinPool extends Abstra
1411          private static final long serialVersionUID = -7914297376763021607L;
1412      }
1413  
656    // Configuration and status settings and queries
657
1414      /**
1415       * Returns the factory used for constructing new workers.
1416       *
# Line 671 | Line 1427 | public class ForkJoinPool extends Abstra
1427       * @return the handler, or {@code null} if none
1428       */
1429      public Thread.UncaughtExceptionHandler getUncaughtExceptionHandler() {
1430 <        Thread.UncaughtExceptionHandler h;
675 <        final ReentrantLock lock = this.workerLock;
676 <        lock.lock();
677 <        try {
678 <            h = ueh;
679 <        } finally {
680 <            lock.unlock();
681 <        }
682 <        return h;
683 <    }
684 <
685 <    /**
686 <     * Sets the handler for internal worker threads that terminate due
687 <     * to unrecoverable errors encountered while executing tasks.
688 <     * Unless set, the current default or ThreadGroup handler is used
689 <     * as handler.
690 <     *
691 <     * @param h the new handler
692 <     * @return the old handler, or {@code null} if none
693 <     * @throws SecurityException if a security manager exists and
694 <     *         the caller is not permitted to modify threads
695 <     *         because it does not hold {@link
696 <     *         java.lang.RuntimePermission}{@code ("modifyThread")}
697 <     */
698 <    public Thread.UncaughtExceptionHandler
699 <        setUncaughtExceptionHandler(Thread.UncaughtExceptionHandler h) {
700 <        checkPermission();
701 <        Thread.UncaughtExceptionHandler old = null;
702 <        final ReentrantLock lock = this.workerLock;
703 <        lock.lock();
704 <        try {
705 <            old = ueh;
706 <            ueh = h;
707 <            ForkJoinWorkerThread[] ws = workers;
708 <            if (ws != null) {
709 <                for (int i = 0; i < ws.length; ++i) {
710 <                    ForkJoinWorkerThread w = ws[i];
711 <                    if (w != null)
712 <                        w.setUncaughtExceptionHandler(h);
713 <                }
714 <            }
715 <        } finally {
716 <            lock.unlock();
717 <        }
718 <        return old;
719 <    }
720 <
721 <
722 <    /**
723 <     * Sets the target parallelism level of this pool.
724 <     *
725 <     * @param parallelism the target parallelism
726 <     * @throws IllegalArgumentException if parallelism less than or
727 <     * equal to zero or greater than maximum size bounds
728 <     * @throws SecurityException if a security manager exists and
729 <     *         the caller is not permitted to modify threads
730 <     *         because it does not hold {@link
731 <     *         java.lang.RuntimePermission}{@code ("modifyThread")}
732 <     */
733 <    public void setParallelism(int parallelism) {
734 <        checkPermission();
735 <        if (parallelism <= 0 || parallelism > maxPoolSize)
736 <            throw new IllegalArgumentException();
737 <        final ReentrantLock lock = this.workerLock;
738 <        lock.lock();
739 <        try {
740 <            if (!isTerminating()) {
741 <                int p = this.parallelism;
742 <                this.parallelism = parallelism;
743 <                if (parallelism > p)
744 <                    createAndStartAddedWorkers();
745 <                else
746 <                    trimSpares();
747 <            }
748 <        } finally {
749 <            lock.unlock();
750 <        }
751 <        signalIdleWorkers();
1430 >        return ueh;
1431      }
1432  
1433      /**
1434 <     * Returns the targeted number of worker threads in this pool.
1434 >     * Returns the targeted parallelism level of this pool.
1435       *
1436 <     * @return the targeted number of worker threads in this pool
1436 >     * @return the targeted parallelism level of this pool
1437       */
1438      public int getParallelism() {
1439          return parallelism;
# Line 769 | Line 1448 | public class ForkJoinPool extends Abstra
1448       * @return the number of worker threads
1449       */
1450      public int getPoolSize() {
1451 <        return totalCountOf(workerCounts);
773 <    }
774 <
775 <    /**
776 <     * Returns the maximum number of threads allowed to exist in the
777 <     * pool, even if there are insufficient unblocked running threads.
778 <     *
779 <     * @return the maximum
780 <     */
781 <    public int getMaximumPoolSize() {
782 <        return maxPoolSize;
783 <    }
784 <
785 <    /**
786 <     * Sets the maximum number of threads allowed to exist in the
787 <     * pool, even if there are insufficient unblocked running threads.
788 <     * Setting this value has no effect on current pool size. It
789 <     * controls construction of new threads.
790 <     *
791 <     * @throws IllegalArgumentException if negative or greater than
792 <     * internal implementation limit
793 <     */
794 <    public void setMaximumPoolSize(int newMax) {
795 <        if (newMax < 0 || newMax > MAX_THREADS)
796 <            throw new IllegalArgumentException();
797 <        maxPoolSize = newMax;
798 <    }
799 <
800 <
801 <    /**
802 <     * Returns {@code true} if this pool dynamically maintains its
803 <     * target parallelism level. If false, new threads are added only
804 <     * to avoid possible starvation.  This setting is by default true.
805 <     *
806 <     * @return {@code true} if maintains parallelism
807 <     */
808 <    public boolean getMaintainsParallelism() {
809 <        return maintainsParallelism;
810 <    }
811 <
812 <    /**
813 <     * Sets whether this pool dynamically maintains its target
814 <     * parallelism level. If false, new threads are added only to
815 <     * avoid possible starvation.
816 <     *
817 <     * @param enable {@code true} to maintain parallelism
818 <     */
819 <    public void setMaintainsParallelism(boolean enable) {
820 <        maintainsParallelism = enable;
821 <    }
822 <
823 <    /**
824 <     * Establishes local first-in-first-out scheduling mode for forked
825 <     * tasks that are never joined. This mode may be more appropriate
826 <     * than default locally stack-based mode in applications in which
827 <     * worker threads only process asynchronous tasks.  This method is
828 <     * designed to be invoked only when the pool is quiescent, and
829 <     * typically only before any tasks are submitted. The effects of
830 <     * invocations at other times may be unpredictable.
831 <     *
832 <     * @param async if {@code true}, use locally FIFO scheduling
833 <     * @return the previous mode
834 <     * @see #getAsyncMode
835 <     */
836 <    public boolean setAsyncMode(boolean async) {
837 <        boolean oldMode = locallyFifo;
838 <        locallyFifo = async;
839 <        ForkJoinWorkerThread[] ws = workers;
840 <        if (ws != null) {
841 <            for (int i = 0; i < ws.length; ++i) {
842 <                ForkJoinWorkerThread t = ws[i];
843 <                if (t != null)
844 <                    t.setAsyncMode(async);
845 <            }
846 <        }
847 <        return oldMode;
1451 >        return workerCounts >>> TOTAL_COUNT_SHIFT;
1452      }
1453  
1454      /**
# Line 852 | Line 1456 | public class ForkJoinPool extends Abstra
1456       * scheduling mode for forked tasks that are never joined.
1457       *
1458       * @return {@code true} if this pool uses async mode
855     * @see #setAsyncMode
1459       */
1460      public boolean getAsyncMode() {
1461          return locallyFifo;
# Line 861 | Line 1464 | public class ForkJoinPool extends Abstra
1464      /**
1465       * Returns an estimate of the number of worker threads that are
1466       * not blocked waiting to join tasks or for other managed
1467 <     * synchronization.
1467 >     * synchronization. This method may overestimate the
1468 >     * number of running threads.
1469       *
1470       * @return the number of worker threads
1471       */
1472      public int getRunningThreadCount() {
1473 <        return runningCountOf(workerCounts);
1473 >        return workerCounts & RUNNING_COUNT_MASK;
1474      }
1475  
1476      /**
# Line 877 | Line 1481 | public class ForkJoinPool extends Abstra
1481       * @return the number of active threads
1482       */
1483      public int getActiveThreadCount() {
1484 <        return activeCountOf(runControl);
881 <    }
882 <
883 <    /**
884 <     * Returns an estimate of the number of threads that are currently
885 <     * idle waiting for tasks. This method may underestimate the
886 <     * number of idle threads.
887 <     *
888 <     * @return the number of idle threads
889 <     */
890 <    final int getIdleThreadCount() {
891 <        int c = runningCountOf(workerCounts) - activeCountOf(runControl);
892 <        return (c <= 0) ? 0 : c;
1484 >        return runState & ACTIVE_COUNT_MASK;
1485      }
1486  
1487      /**
# Line 904 | Line 1496 | public class ForkJoinPool extends Abstra
1496       * @return {@code true} if all threads are currently idle
1497       */
1498      public boolean isQuiescent() {
1499 <        return activeCountOf(runControl) == 0;
1499 >        return (runState & ACTIVE_COUNT_MASK) == 0;
1500      }
1501  
1502      /**
# Line 919 | Line 1511 | public class ForkJoinPool extends Abstra
1511       * @return the number of steals
1512       */
1513      public long getStealCount() {
1514 <        return stealCount.get();
923 <    }
924 <
925 <    /**
926 <     * Accumulates steal count from a worker.
927 <     * Call only when worker known to be idle.
928 <     */
929 <    private void updateStealCount(ForkJoinWorkerThread w) {
930 <        int sc = w.getAndClearStealCount();
931 <        if (sc != 0)
932 <            stealCount.addAndGet(sc);
1514 >        return stealCount;
1515      }
1516  
1517      /**
# Line 944 | Line 1526 | public class ForkJoinPool extends Abstra
1526       */
1527      public long getQueuedTaskCount() {
1528          long count = 0;
1529 <        ForkJoinWorkerThread[] ws = workers;
1530 <        if (ws != null) {
1531 <            for (int i = 0; i < ws.length; ++i) {
950 <                ForkJoinWorkerThread t = ws[i];
951 <                if (t != null)
952 <                    count += t.getQueueSize();
953 <            }
954 <        }
1529 >        for (ForkJoinWorkerThread w : workers)
1530 >            if (w != null)
1531 >                count += w.getQueueSize();
1532          return count;
1533      }
1534  
# Line 991 | Line 1568 | public class ForkJoinPool extends Abstra
1568       * Removes all available unexecuted submitted and forked tasks
1569       * from scheduling queues and adds them to the given collection,
1570       * without altering their execution status. These may include
1571 <     * artificially generated or wrapped tasks. This method is designed
1572 <     * to be invoked only when the pool is known to be
1571 >     * artificially generated or wrapped tasks. This method is
1572 >     * designed to be invoked only when the pool is known to be
1573       * quiescent. Invocations at other times may not remove all
1574       * tasks. A failure encountered while attempting to add elements
1575       * to collection {@code c} may result in elements being in
# Line 1005 | Line 1582 | public class ForkJoinPool extends Abstra
1582       * @return the number of elements transferred
1583       */
1584      protected int drainTasksTo(Collection<? super ForkJoinTask<?>> c) {
1585 <        int n = submissionQueue.drainTo(c);
1586 <        ForkJoinWorkerThread[] ws = workers;
1587 <        if (ws != null) {
1588 <            for (int i = 0; i < ws.length; ++i) {
1589 <                ForkJoinWorkerThread w = ws[i];
1013 <                if (w != null)
1014 <                    n += w.drainTasksTo(c);
1015 <            }
1016 <        }
1017 <        return n;
1585 >        int count = submissionQueue.drainTo(c);
1586 >        for (ForkJoinWorkerThread w : workers)
1587 >            if (w != null)
1588 >                count += w.drainTasksTo(c);
1589 >        return count;
1590      }
1591  
1592      /**
# Line 1025 | Line 1597 | public class ForkJoinPool extends Abstra
1597       * @return a string identifying this pool, as well as its state
1598       */
1599      public String toString() {
1028        int ps = parallelism;
1029        int wc = workerCounts;
1030        int rc = runControl;
1600          long st = getStealCount();
1601          long qt = getQueuedTaskCount();
1602          long qs = getQueuedSubmissionCount();
1603 +        int wc = workerCounts;
1604 +        int tc = wc >>> TOTAL_COUNT_SHIFT;
1605 +        int rc = wc & RUNNING_COUNT_MASK;
1606 +        int pc = parallelism;
1607 +        int rs = runState;
1608 +        int ac = rs & ACTIVE_COUNT_MASK;
1609          return super.toString() +
1610 <            "[" + runStateToString(runStateOf(rc)) +
1611 <            ", parallelism = " + ps +
1612 <            ", size = " + totalCountOf(wc) +
1613 <            ", active = " + activeCountOf(rc) +
1614 <            ", running = " + runningCountOf(wc) +
1610 >            "[" + runLevelToString(rs) +
1611 >            ", parallelism = " + pc +
1612 >            ", size = " + tc +
1613 >            ", active = " + ac +
1614 >            ", running = " + rc +
1615              ", steals = " + st +
1616              ", tasks = " + qt +
1617              ", submissions = " + qs +
1618              "]";
1619      }
1620  
1621 <    private static String runStateToString(int rs) {
1622 <        switch(rs) {
1623 <        case RUNNING: return "Running";
1624 <        case SHUTDOWN: return "Shutting down";
1625 <        case TERMINATING: return "Terminating";
1051 <        case TERMINATED: return "Terminated";
1052 <        default: throw new Error("Unknown run state");
1053 <        }
1621 >    private static String runLevelToString(int s) {
1622 >        return ((s & TERMINATED) != 0 ? "Terminated" :
1623 >                ((s & TERMINATING) != 0 ? "Terminating" :
1624 >                 ((s & SHUTDOWN) != 0 ? "Shutting down" :
1625 >                  "Running")));
1626      }
1627  
1056    // lifecycle control
1057
1628      /**
1629       * Initiates an orderly shutdown in which previously submitted
1630       * tasks are executed, but no new tasks will be accepted.
# Line 1069 | Line 1639 | public class ForkJoinPool extends Abstra
1639       */
1640      public void shutdown() {
1641          checkPermission();
1642 <        transitionRunStateTo(SHUTDOWN);
1643 <        if (canTerminateOnShutdown(runControl)) {
1074 <            if (workers == null) { // shutting down before workers created
1075 <                final ReentrantLock lock = this.workerLock;
1076 <                lock.lock();
1077 <                try {
1078 <                    if (workers == null) {
1079 <                        terminate();
1080 <                        transitionRunStateTo(TERMINATED);
1081 <                        termination.signalAll();
1082 <                    }
1083 <                } finally {
1084 <                    lock.unlock();
1085 <                }
1086 <            }
1087 <            terminateOnShutdown();
1088 <        }
1642 >        advanceRunLevel(SHUTDOWN);
1643 >        tryTerminate(false);
1644      }
1645  
1646      /**
1647 <     * Attempts to stop all actively executing tasks, and cancels all
1648 <     * waiting tasks.  Tasks that are in the process of being
1649 <     * submitted or executed concurrently during the course of this
1650 <     * method may or may not be rejected. Unlike some other executors,
1651 <     * this method cancels rather than collects non-executed tasks
1652 <     * upon termination, so always returns an empty list. However, you
1653 <     * can use method {@link #drainTasksTo} before invoking this
1654 <     * method to transfer unexecuted tasks to another collection.
1647 >     * Attempts to cancel and/or stop all tasks, and reject all
1648 >     * subsequently submitted tasks.  Tasks that are in the process of
1649 >     * being submitted or executed concurrently during the course of
1650 >     * this method may or may not be rejected. This method cancels
1651 >     * both existing and unexecuted tasks, in order to permit
1652 >     * termination in the presence of task dependencies. So the method
1653 >     * always returns an empty list (unlike the case for some other
1654 >     * Executors).
1655       *
1656       * @return an empty list
1657       * @throws SecurityException if a security manager exists and
# Line 1106 | Line 1661 | public class ForkJoinPool extends Abstra
1661       */
1662      public List<Runnable> shutdownNow() {
1663          checkPermission();
1664 <        terminate();
1664 >        tryTerminate(true);
1665          return Collections.emptyList();
1666      }
1667  
# Line 1116 | Line 1671 | public class ForkJoinPool extends Abstra
1671       * @return {@code true} if all tasks have completed following shut down
1672       */
1673      public boolean isTerminated() {
1674 <        return runStateOf(runControl) == TERMINATED;
1674 >        return runState >= TERMINATED;
1675      }
1676  
1677      /**
1678       * Returns {@code true} if the process of termination has
1679 <     * commenced but possibly not yet completed.
1679 >     * commenced but not yet completed.  This method may be useful for
1680 >     * debugging. A return of {@code true} reported a sufficient
1681 >     * period after shutdown may indicate that submitted tasks have
1682 >     * ignored or suppressed interruption, causing this executor not
1683 >     * to properly terminate.
1684       *
1685 <     * @return {@code true} if terminating
1685 >     * @return {@code true} if terminating but not yet terminated
1686       */
1687      public boolean isTerminating() {
1688 <        return runStateOf(runControl) >= TERMINATING;
1688 >        return (runState & (TERMINATING|TERMINATED)) == TERMINATING;
1689      }
1690  
1691      /**
# Line 1135 | Line 1694 | public class ForkJoinPool extends Abstra
1694       * @return {@code true} if this pool has been shut down
1695       */
1696      public boolean isShutdown() {
1697 <        return runStateOf(runControl) >= SHUTDOWN;
1697 >        return runState >= SHUTDOWN;
1698      }
1699  
1700      /**
# Line 1151 | Line 1710 | public class ForkJoinPool extends Abstra
1710       */
1711      public boolean awaitTermination(long timeout, TimeUnit unit)
1712          throws InterruptedException {
1154        long nanos = unit.toNanos(timeout);
1155        final ReentrantLock lock = this.workerLock;
1156        lock.lock();
1157        try {
1158            for (;;) {
1159                if (isTerminated())
1160                    return true;
1161                if (nanos <= 0)
1162                    return false;
1163                nanos = termination.awaitNanos(nanos);
1164            }
1165        } finally {
1166            lock.unlock();
1167        }
1168    }
1169
1170    // Shutdown and termination support
1171
1172    /**
1173     * Callback from terminating worker. Nulls out the corresponding
1174     * workers slot, and if terminating, tries to terminate; else
1175     * tries to shrink workers array.
1176     *
1177     * @param w the worker
1178     */
1179    final void workerTerminated(ForkJoinWorkerThread w) {
1180        updateStealCount(w);
1181        updateWorkerCount(-1);
1182        final ReentrantLock lock = this.workerLock;
1183        lock.lock();
1184        try {
1185            ForkJoinWorkerThread[] ws = workers;
1186            if (ws != null) {
1187                int idx = w.poolIndex;
1188                if (idx >= 0 && idx < ws.length && ws[idx] == w)
1189                    ws[idx] = null;
1190                if (totalCountOf(workerCounts) == 0) {
1191                    terminate(); // no-op if already terminating
1192                    transitionRunStateTo(TERMINATED);
1193                    termination.signalAll();
1194                }
1195                else if (!isTerminating()) {
1196                    tryShrinkWorkerArray();
1197                    tryResumeSpare(true); // allow replacement
1198                }
1199            }
1200        } finally {
1201            lock.unlock();
1202        }
1203        signalIdleWorkers();
1204    }
1205
1206    /**
1207     * Initiates termination.
1208     */
1209    private void terminate() {
1210        if (transitionRunStateTo(TERMINATING)) {
1211            stopAllWorkers();
1212            resumeAllSpares();
1213            signalIdleWorkers();
1214            cancelQueuedSubmissions();
1215            cancelQueuedWorkerTasks();
1216            interruptUnterminatedWorkers();
1217            signalIdleWorkers(); // resignal after interrupt
1218        }
1219    }
1220
1221    /**
1222     * Possibly terminates when on shutdown state.
1223     */
1224    private void terminateOnShutdown() {
1225        if (!hasQueuedSubmissions() && canTerminateOnShutdown(runControl))
1226            terminate();
1227    }
1228
1229    /**
1230     * Clears out and cancels submissions.
1231     */
1232    private void cancelQueuedSubmissions() {
1233        ForkJoinTask<?> task;
1234        while ((task = pollSubmission()) != null)
1235            task.cancel(false);
1236    }
1237
1238    /**
1239     * Cleans out worker queues.
1240     */
1241    private void cancelQueuedWorkerTasks() {
1242        final ReentrantLock lock = this.workerLock;
1243        lock.lock();
1713          try {
1714 <            ForkJoinWorkerThread[] ws = workers;
1715 <            if (ws != null) {
1247 <                for (int i = 0; i < ws.length; ++i) {
1248 <                    ForkJoinWorkerThread t = ws[i];
1249 <                    if (t != null)
1250 <                        t.cancelTasks();
1251 <                }
1252 <            }
1253 <        } finally {
1254 <            lock.unlock();
1255 <        }
1256 <    }
1257 <
1258 <    /**
1259 <     * Sets each worker's status to terminating. Requires lock to avoid
1260 <     * conflicts with add/remove.
1261 <     */
1262 <    private void stopAllWorkers() {
1263 <        final ReentrantLock lock = this.workerLock;
1264 <        lock.lock();
1265 <        try {
1266 <            ForkJoinWorkerThread[] ws = workers;
1267 <            if (ws != null) {
1268 <                for (int i = 0; i < ws.length; ++i) {
1269 <                    ForkJoinWorkerThread t = ws[i];
1270 <                    if (t != null)
1271 <                        t.shutdownNow();
1272 <                }
1273 <            }
1274 <        } finally {
1275 <            lock.unlock();
1276 <        }
1277 <    }
1278 <
1279 <    /**
1280 <     * Interrupts all unterminated workers.  This is not required for
1281 <     * sake of internal control, but may help unstick user code during
1282 <     * shutdown.
1283 <     */
1284 <    private void interruptUnterminatedWorkers() {
1285 <        final ReentrantLock lock = this.workerLock;
1286 <        lock.lock();
1287 <        try {
1288 <            ForkJoinWorkerThread[] ws = workers;
1289 <            if (ws != null) {
1290 <                for (int i = 0; i < ws.length; ++i) {
1291 <                    ForkJoinWorkerThread t = ws[i];
1292 <                    if (t != null && !t.isTerminated()) {
1293 <                        try {
1294 <                            t.interrupt();
1295 <                        } catch (SecurityException ignore) {
1296 <                        }
1297 <                    }
1298 <                }
1299 <            }
1300 <        } finally {
1301 <            lock.unlock();
1302 <        }
1303 <    }
1304 <
1305 <
1306 <    /*
1307 <     * Nodes for event barrier to manage idle threads.  Queue nodes
1308 <     * are basic Treiber stack nodes, also used for spare stack.
1309 <     *
1310 <     * The event barrier has an event count and a wait queue (actually
1311 <     * a Treiber stack).  Workers are enabled to look for work when
1312 <     * the eventCount is incremented. If they fail to find work, they
1313 <     * may wait for next count. Upon release, threads help others wake
1314 <     * up.
1315 <     *
1316 <     * Synchronization events occur only in enough contexts to
1317 <     * maintain overall liveness:
1318 <     *
1319 <     *   - Submission of a new task to the pool
1320 <     *   - Resizes or other changes to the workers array
1321 <     *   - pool termination
1322 <     *   - A worker pushing a task on an empty queue
1323 <     *
1324 <     * The case of pushing a task occurs often enough, and is heavy
1325 <     * enough compared to simple stack pushes, to require special
1326 <     * handling: Method signalWork returns without advancing count if
1327 <     * the queue appears to be empty.  This would ordinarily result in
1328 <     * races causing some queued waiters not to be woken up. To avoid
1329 <     * this, the first worker enqueued in method sync (see
1330 <     * syncIsReleasable) rescans for tasks after being enqueued, and
1331 <     * helps signal if any are found. This works well because the
1332 <     * worker has nothing better to do, and so might as well help
1333 <     * alleviate the overhead and contention on the threads actually
1334 <     * doing work.  Also, since event counts increments on task
1335 <     * availability exist to maintain liveness (rather than to force
1336 <     * refreshes etc), it is OK for callers to exit early if
1337 <     * contending with another signaller.
1338 <     */
1339 <    static final class WaitQueueNode {
1340 <        WaitQueueNode next; // only written before enqueued
1341 <        volatile ForkJoinWorkerThread thread; // nulled to cancel wait
1342 <        final long count; // unused for spare stack
1343 <
1344 <        WaitQueueNode(long c, ForkJoinWorkerThread w) {
1345 <            count = c;
1346 <            thread = w;
1347 <        }
1348 <
1349 <        /**
1350 <         * Wakes up waiter, returning false if known to already
1351 <         */
1352 <        boolean signal() {
1353 <            ForkJoinWorkerThread t = thread;
1354 <            if (t == null)
1355 <                return false;
1356 <            thread = null;
1357 <            LockSupport.unpark(t);
1358 <            return true;
1359 <        }
1360 <
1361 <        /**
1362 <         * Awaits release on sync.
1363 <         */
1364 <        void awaitSyncRelease(ForkJoinPool p) {
1365 <            while (thread != null && !p.syncIsReleasable(this))
1366 <                LockSupport.park(this);
1367 <        }
1368 <
1369 <        /**
1370 <         * Awaits resumption as spare.
1371 <         */
1372 <        void awaitSpareRelease() {
1373 <            while (thread != null) {
1374 <                if (!Thread.interrupted())
1375 <                    LockSupport.park(this);
1376 <            }
1377 <        }
1378 <    }
1379 <
1380 <    /**
1381 <     * Ensures that no thread is waiting for count to advance from the
1382 <     * current value of eventCount read on entry to this method, by
1383 <     * releasing waiting threads if necessary.
1384 <     *
1385 <     * @return the count
1386 <     */
1387 <    final long ensureSync() {
1388 <        long c = eventCount;
1389 <        WaitQueueNode q;
1390 <        while ((q = syncStack) != null && q.count < c) {
1391 <            if (casBarrierStack(q, null)) {
1392 <                do {
1393 <                    q.signal();
1394 <                } while ((q = q.next) != null);
1395 <                break;
1396 <            }
1397 <        }
1398 <        return c;
1399 <    }
1400 <
1401 <    /**
1402 <     * Increments event count and releases waiting threads.
1403 <     */
1404 <    private void signalIdleWorkers() {
1405 <        long c;
1406 <        do {} while (!casEventCount(c = eventCount, c+1));
1407 <        ensureSync();
1408 <    }
1409 <
1410 <    /**
1411 <     * Signals threads waiting to poll a task. Because method sync
1412 <     * rechecks availability, it is OK to only proceed if queue
1413 <     * appears to be non-empty, and OK to skip under contention to
1414 <     * increment count (since some other thread succeeded).
1415 <     */
1416 <    final void signalWork() {
1417 <        long c;
1418 <        WaitQueueNode q;
1419 <        if (syncStack != null &&
1420 <            casEventCount(c = eventCount, c+1) &&
1421 <            (((q = syncStack) != null && q.count <= c) &&
1422 <             (!casBarrierStack(q, q.next) || !q.signal())))
1423 <            ensureSync();
1424 <    }
1425 <
1426 <    /**
1427 <     * Waits until event count advances from last value held by
1428 <     * caller, or if excess threads, caller is resumed as spare, or
1429 <     * caller or pool is terminating. Updates caller's event on exit.
1430 <     *
1431 <     * @param w the calling worker thread
1432 <     */
1433 <    final void sync(ForkJoinWorkerThread w) {
1434 <        updateStealCount(w); // Transfer w's count while it is idle
1435 <
1436 <        while (!w.isShutdown() && !isTerminating() && !suspendIfSpare(w)) {
1437 <            long prev = w.lastEventCount;
1438 <            WaitQueueNode node = null;
1439 <            WaitQueueNode h;
1440 <            while (eventCount == prev &&
1441 <                   ((h = syncStack) == null || h.count == prev)) {
1442 <                if (node == null)
1443 <                    node = new WaitQueueNode(prev, w);
1444 <                if (casBarrierStack(node.next = h, node)) {
1445 <                    node.awaitSyncRelease(this);
1446 <                    break;
1447 <                }
1448 <            }
1449 <            long ec = ensureSync();
1450 <            if (ec != prev) {
1451 <                w.lastEventCount = ec;
1452 <                break;
1453 <            }
1454 <        }
1455 <    }
1456 <
1457 <    /**
1458 <     * Returns {@code true} if worker waiting on sync can proceed:
1459 <     *  - on signal (thread == null)
1460 <     *  - on event count advance (winning race to notify vs signaller)
1461 <     *  - on interrupt
1462 <     *  - if the first queued node, we find work available
1463 <     * If node was not signalled and event count not advanced on exit,
1464 <     * then we also help advance event count.
1465 <     *
1466 <     * @return {@code true} if node can be released
1467 <     */
1468 <    final boolean syncIsReleasable(WaitQueueNode node) {
1469 <        long prev = node.count;
1470 <        if (!Thread.interrupted() && node.thread != null &&
1471 <            (node.next != null ||
1472 <             !ForkJoinWorkerThread.hasQueuedTasks(workers)) &&
1473 <            eventCount == prev)
1474 <            return false;
1475 <        if (node.thread != null) {
1476 <            node.thread = null;
1477 <            long ec = eventCount;
1478 <            if (prev <= ec) // help signal
1479 <                casEventCount(ec, ec+1);
1480 <        }
1481 <        return true;
1482 <    }
1483 <
1484 <    /**
1485 <     * Returns {@code true} if a new sync event occurred since last
1486 <     * call to sync or this method, if so, updating caller's count.
1487 <     */
1488 <    final boolean hasNewSyncEvent(ForkJoinWorkerThread w) {
1489 <        long lc = w.lastEventCount;
1490 <        long ec = ensureSync();
1491 <        if (ec == lc)
1714 >            return termination.awaitAdvanceInterruptibly(0, timeout, unit) > 0;
1715 >        } catch (TimeoutException ex) {
1716              return false;
1493        w.lastEventCount = ec;
1494        return true;
1495    }
1496
1497    //  Parallelism maintenance
1498
1499    /**
1500     * Decrements running count; if too low, adds spare.
1501     *
1502     * Conceptually, all we need to do here is add or resume a
1503     * spare thread when one is about to block (and remove or
1504     * suspend it later when unblocked -- see suspendIfSpare).
1505     * However, implementing this idea requires coping with
1506     * several problems: we have imperfect information about the
1507     * states of threads. Some count updates can and usually do
1508     * lag run state changes, despite arrangements to keep them
1509     * accurate (for example, when possible, updating counts
1510     * before signalling or resuming), especially when running on
1511     * dynamic JVMs that don't optimize the infrequent paths that
1512     * update counts. Generating too many threads can make these
1513     * problems become worse, because excess threads are more
1514     * likely to be context-switched with others, slowing them all
1515     * down, especially if there is no work available, so all are
1516     * busy scanning or idling.  Also, excess spare threads can
1517     * only be suspended or removed when they are idle, not
1518     * immediately when they aren't needed. So adding threads will
1519     * raise parallelism level for longer than necessary.  Also,
1520     * FJ applications often encounter highly transient peaks when
1521     * many threads are blocked joining, but for less time than it
1522     * takes to create or resume spares.
1523     *
1524     * @param joinMe if non-null, return early if done
1525     * @param maintainParallelism if true, try to stay within
1526     * target counts, else create only to avoid starvation
1527     * @return true if joinMe known to be done
1528     */
1529    final boolean preJoin(ForkJoinTask<?> joinMe,
1530                          boolean maintainParallelism) {
1531        maintainParallelism &= maintainsParallelism; // overrride
1532        boolean dec = false;  // true when running count decremented
1533        while (spareStack == null || !tryResumeSpare(dec)) {
1534            int counts = workerCounts;
1535            if (dec || (dec = casWorkerCounts(counts, --counts))) {
1536                // CAS cheat
1537                if (!needSpare(counts, maintainParallelism))
1538                    break;
1539                if (joinMe.status < 0)
1540                    return true;
1541                if (tryAddSpare(counts))
1542                    break;
1543            }
1544        }
1545        return false;
1546    }
1547
1548    /**
1549     * Same idea as preJoin
1550     */
1551    final boolean preBlock(ManagedBlocker blocker,
1552                           boolean maintainParallelism) {
1553        maintainParallelism &= maintainsParallelism;
1554        boolean dec = false;
1555        while (spareStack == null || !tryResumeSpare(dec)) {
1556            int counts = workerCounts;
1557            if (dec || (dec = casWorkerCounts(counts, --counts))) {
1558                if (!needSpare(counts, maintainParallelism))
1559                    break;
1560                if (blocker.isReleasable())
1561                    return true;
1562                if (tryAddSpare(counts))
1563                    break;
1564            }
1565        }
1566        return false;
1567    }
1568
1569    /**
1570     * Returns {@code true} if a spare thread appears to be needed.
1571     * If maintaining parallelism, returns true when the deficit in
1572     * running threads is more than the surplus of total threads, and
1573     * there is apparently some work to do.  This self-limiting rule
1574     * means that the more threads that have already been added, the
1575     * less parallelism we will tolerate before adding another.
1576     *
1577     * @param counts current worker counts
1578     * @param maintainParallelism try to maintain parallelism
1579     */
1580    private boolean needSpare(int counts, boolean maintainParallelism) {
1581        int ps = parallelism;
1582        int rc = runningCountOf(counts);
1583        int tc = totalCountOf(counts);
1584        int runningDeficit = ps - rc;
1585        int totalSurplus = tc - ps;
1586        return (tc < maxPoolSize &&
1587                (rc == 0 || totalSurplus < 0 ||
1588                 (maintainParallelism &&
1589                  runningDeficit > totalSurplus &&
1590                  ForkJoinWorkerThread.hasQueuedTasks(workers))));
1591    }
1592
1593    /**
1594     * Adds a spare worker if lock available and no more than the
1595     * expected numbers of threads exist.
1596     *
1597     * @return true if successful
1598     */
1599    private boolean tryAddSpare(int expectedCounts) {
1600        final ReentrantLock lock = this.workerLock;
1601        int expectedRunning = runningCountOf(expectedCounts);
1602        int expectedTotal = totalCountOf(expectedCounts);
1603        boolean success = false;
1604        boolean locked = false;
1605        // confirm counts while locking; CAS after obtaining lock
1606        try {
1607            for (;;) {
1608                int s = workerCounts;
1609                int tc = totalCountOf(s);
1610                int rc = runningCountOf(s);
1611                if (rc > expectedRunning || tc > expectedTotal)
1612                    break;
1613                if (!locked && !(locked = lock.tryLock()))
1614                    break;
1615                if (casWorkerCounts(s, workerCountsFor(tc+1, rc+1))) {
1616                    createAndStartSpare(tc);
1617                    success = true;
1618                    break;
1619                }
1620            }
1621        } finally {
1622            if (locked)
1623                lock.unlock();
1624        }
1625        return success;
1626    }
1627
1628    /**
1629     * Adds the kth spare worker. On entry, pool counts are already
1630     * adjusted to reflect addition.
1631     */
1632    private void createAndStartSpare(int k) {
1633        ForkJoinWorkerThread w = null;
1634        ForkJoinWorkerThread[] ws = ensureWorkerArrayCapacity(k + 1);
1635        int len = ws.length;
1636        // Probably, we can place at slot k. If not, find empty slot
1637        if (k < len && ws[k] != null) {
1638            for (k = 0; k < len && ws[k] != null; ++k)
1639                ;
1640        }
1641        if (k < len && !isTerminating() && (w = createWorker(k)) != null) {
1642            ws[k] = w;
1643            w.start();
1644        }
1645        else
1646            updateWorkerCount(-1); // adjust on failure
1647        signalIdleWorkers();
1648    }
1649
1650    /**
1651     * Suspends calling thread w if there are excess threads.  Called
1652     * only from sync.  Spares are enqueued in a Treiber stack using
1653     * the same WaitQueueNodes as barriers.  They are resumed mainly
1654     * in preJoin, but are also woken on pool events that require all
1655     * threads to check run state.
1656     *
1657     * @param w the caller
1658     */
1659    private boolean suspendIfSpare(ForkJoinWorkerThread w) {
1660        WaitQueueNode node = null;
1661        int s;
1662        while (parallelism < runningCountOf(s = workerCounts)) {
1663            if (node == null)
1664                node = new WaitQueueNode(0, w);
1665            if (casWorkerCounts(s, s-1)) { // representation-dependent
1666                // push onto stack
1667                do {} while (!casSpareStack(node.next = spareStack, node));
1668                // block until released by resumeSpare
1669                node.awaitSpareRelease();
1670                return true;
1671            }
1672        }
1673        return false;
1674    }
1675
1676    /**
1677     * Tries to pop and resume a spare thread.
1678     *
1679     * @param updateCount if true, increment running count on success
1680     * @return true if successful
1681     */
1682    private boolean tryResumeSpare(boolean updateCount) {
1683        WaitQueueNode q;
1684        while ((q = spareStack) != null) {
1685            if (casSpareStack(q, q.next)) {
1686                if (updateCount)
1687                    updateRunningCount(1);
1688                q.signal();
1689                return true;
1690            }
1691        }
1692        return false;
1693    }
1694
1695    /**
1696     * Pops and resumes all spare threads. Same idea as ensureSync.
1697     *
1698     * @return true if any spares released
1699     */
1700    private boolean resumeAllSpares() {
1701        WaitQueueNode q;
1702        while ( (q = spareStack) != null) {
1703            if (casSpareStack(q, null)) {
1704                do {
1705                    updateRunningCount(1);
1706                    q.signal();
1707                } while ((q = q.next) != null);
1708                return true;
1709            }
1710        }
1711        return false;
1712    }
1713
1714    /**
1715     * Pops and shuts down excessive spare threads. Call only while
1716     * holding lock. This is not guaranteed to eliminate all excess
1717     * threads, only those suspended as spares, which are the ones
1718     * unlikely to be needed in the future.
1719     */
1720    private void trimSpares() {
1721        int surplus = totalCountOf(workerCounts) - parallelism;
1722        WaitQueueNode q;
1723        while (surplus > 0 && (q = spareStack) != null) {
1724            if (casSpareStack(q, null)) {
1725                do {
1726                    updateRunningCount(1);
1727                    ForkJoinWorkerThread w = q.thread;
1728                    if (w != null && surplus > 0 &&
1729                        runningCountOf(workerCounts) > 0 && w.shutdown())
1730                        --surplus;
1731                    q.signal();
1732                } while ((q = q.next) != null);
1733            }
1717          }
1718      }
1719  
# Line 1738 | Line 1721 | public class ForkJoinPool extends Abstra
1721       * Interface for extending managed parallelism for tasks running
1722       * in {@link ForkJoinPool}s.
1723       *
1724 <     * <p>A {@code ManagedBlocker} provides two methods.
1725 <     * Method {@code isReleasable} must return {@code true} if
1726 <     * blocking is not necessary. Method {@code block} blocks the
1727 <     * current thread if necessary (perhaps internally invoking
1728 <     * {@code isReleasable} before actually blocking).
1724 >     * <p>A {@code ManagedBlocker} provides two methods.  Method
1725 >     * {@code isReleasable} must return {@code true} if blocking is
1726 >     * not necessary. Method {@code block} blocks the current thread
1727 >     * if necessary (perhaps internally invoking {@code isReleasable}
1728 >     * before actually blocking). The unusual methods in this API
1729 >     * accommodate synchronizers that may, but don't usually, block
1730 >     * for long periods. Similarly, they allow more efficient internal
1731 >     * handling of cases in which additional workers may be, but
1732 >     * usually are not, needed to ensure sufficient parallelism.
1733 >     * Toward this end, implementations of method {@code isReleasable}
1734 >     * must be amenable to repeated invocation.
1735       *
1736       * <p>For example, here is a ManagedBlocker based on a
1737       * ReentrantLock:
# Line 1760 | Line 1749 | public class ForkJoinPool extends Abstra
1749       *     return hasLock || (hasLock = lock.tryLock());
1750       *   }
1751       * }}</pre>
1752 +     *
1753 +     * <p>Here is a class that possibly blocks waiting for an
1754 +     * item on a given queue:
1755 +     *  <pre> {@code
1756 +     * class QueueTaker<E> implements ManagedBlocker {
1757 +     *   final BlockingQueue<E> queue;
1758 +     *   volatile E item = null;
1759 +     *   QueueTaker(BlockingQueue<E> q) { this.queue = q; }
1760 +     *   public boolean block() throws InterruptedException {
1761 +     *     if (item == null)
1762 +     *       item = queue.take();
1763 +     *     return true;
1764 +     *   }
1765 +     *   public boolean isReleasable() {
1766 +     *     return item != null || (item = queue.poll()) != null;
1767 +     *   }
1768 +     *   public E getItem() { // call after pool.managedBlock completes
1769 +     *     return item;
1770 +     *   }
1771 +     * }}</pre>
1772       */
1773      public static interface ManagedBlocker {
1774          /**
# Line 1783 | Line 1792 | public class ForkJoinPool extends Abstra
1792       * Blocks in accord with the given blocker.  If the current thread
1793       * is a {@link ForkJoinWorkerThread}, this method possibly
1794       * arranges for a spare thread to be activated if necessary to
1795 <     * ensure parallelism while the current thread is blocked.
1787 <     *
1788 <     * <p>If {@code maintainParallelism} is {@code true} and the pool
1789 <     * supports it ({@link #getMaintainsParallelism}), this method
1790 <     * attempts to maintain the pool's nominal parallelism. Otherwise
1791 <     * it activates a thread only if necessary to avoid complete
1792 <     * starvation. This option may be preferable when blockages use
1793 <     * timeouts, or are almost always brief.
1795 >     * ensure sufficient parallelism while the current thread is blocked.
1796       *
1797       * <p>If the caller is not a {@link ForkJoinTask}, this method is
1798       * behaviorally equivalent to
# Line 1804 | Line 1806 | public class ForkJoinPool extends Abstra
1806       * first be expanded to ensure parallelism, and later adjusted.
1807       *
1808       * @param blocker the blocker
1807     * @param maintainParallelism if {@code true} and supported by
1808     * this pool, attempt to maintain the pool's nominal parallelism;
1809     * otherwise activate a thread only if necessary to avoid
1810     * complete starvation.
1809       * @throws InterruptedException if blocker.block did so
1810       */
1811 <    public static void managedBlock(ManagedBlocker blocker,
1814 <                                    boolean maintainParallelism)
1811 >    public static void managedBlock(ManagedBlocker blocker)
1812          throws InterruptedException {
1813          Thread t = Thread.currentThread();
1814 <        ForkJoinPool pool = ((t instanceof ForkJoinWorkerThread) ?
1815 <                             ((ForkJoinWorkerThread) t).pool : null);
1816 <        if (!blocker.isReleasable()) {
1817 <            try {
1818 <                if (pool == null ||
1819 <                    !pool.preBlock(blocker, maintainParallelism))
1823 <                    awaitBlocker(blocker);
1824 <            } finally {
1825 <                if (pool != null)
1826 <                    pool.updateRunningCount(1);
1827 <            }
1814 >        if (t instanceof ForkJoinWorkerThread) {
1815 >            ForkJoinWorkerThread w = (ForkJoinWorkerThread) t;
1816 >            w.pool.awaitBlocker(blocker);
1817 >        }
1818 >        else {
1819 >            do {} while (!blocker.isReleasable() && !blocker.block());
1820          }
1829    }
1830
1831    private static void awaitBlocker(ManagedBlocker blocker)
1832        throws InterruptedException {
1833        do {} while (!blocker.isReleasable() && !blocker.block());
1821      }
1822  
1823      // AbstractExecutorService overrides.  These rely on undocumented
# Line 1848 | Line 1835 | public class ForkJoinPool extends Abstra
1835      // Unsafe mechanics
1836  
1837      private static final sun.misc.Unsafe UNSAFE = getUnsafe();
1851    private static final long eventCountOffset =
1852        objectFieldOffset("eventCount", ForkJoinPool.class);
1838      private static final long workerCountsOffset =
1839          objectFieldOffset("workerCounts", ForkJoinPool.class);
1840 <    private static final long runControlOffset =
1841 <        objectFieldOffset("runControl", ForkJoinPool.class);
1842 <    private static final long syncStackOffset =
1843 <        objectFieldOffset("syncStack",ForkJoinPool.class);
1844 <    private static final long spareStackOffset =
1845 <        objectFieldOffset("spareStack", ForkJoinPool.class);
1846 <
1847 <    private boolean casEventCount(long cmp, long val) {
1848 <        return UNSAFE.compareAndSwapLong(this, eventCountOffset, cmp, val);
1849 <    }
1865 <    private boolean casWorkerCounts(int cmp, int val) {
1866 <        return UNSAFE.compareAndSwapInt(this, workerCountsOffset, cmp, val);
1867 <    }
1868 <    private boolean casRunControl(int cmp, int val) {
1869 <        return UNSAFE.compareAndSwapInt(this, runControlOffset, cmp, val);
1870 <    }
1871 <    private boolean casSpareStack(WaitQueueNode cmp, WaitQueueNode val) {
1872 <        return UNSAFE.compareAndSwapObject(this, spareStackOffset, cmp, val);
1873 <    }
1874 <    private boolean casBarrierStack(WaitQueueNode cmp, WaitQueueNode val) {
1875 <        return UNSAFE.compareAndSwapObject(this, syncStackOffset, cmp, val);
1876 <    }
1840 >    private static final long runStateOffset =
1841 >        objectFieldOffset("runState", ForkJoinPool.class);
1842 >    private static final long eventCountOffset =
1843 >        objectFieldOffset("eventCount", ForkJoinPool.class);
1844 >    private static final long eventWaitersOffset =
1845 >        objectFieldOffset("eventWaiters",ForkJoinPool.class);
1846 >    private static final long stealCountOffset =
1847 >        objectFieldOffset("stealCount",ForkJoinPool.class);
1848 >    private static final long spareWaitersOffset =
1849 >        objectFieldOffset("spareWaiters",ForkJoinPool.class);
1850  
1851      private static long objectFieldOffset(String field, Class<?> klazz) {
1852          try {

Diff Legend

Removed lines
+ Added lines
< Changed lines
> Changed lines