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 1.56 by dl, Thu May 27 16:46:48 2010 UTC vs.
Revision 1.80 by dl, Fri Sep 17 14:24:56 2010 UTC

# Line 6 | Line 6
6  
7   package jsr166y;
8  
9 import java.util.concurrent.*;
10
9   import java.util.ArrayList;
10   import java.util.Arrays;
11   import java.util.Collection;
12   import java.util.Collections;
13   import java.util.List;
14 + import java.util.concurrent.AbstractExecutorService;
15 + import java.util.concurrent.Callable;
16 + import java.util.concurrent.ExecutorService;
17 + import java.util.concurrent.Future;
18 + import java.util.concurrent.RejectedExecutionException;
19 + import java.util.concurrent.RunnableFuture;
20 + import java.util.concurrent.TimeUnit;
21 + import java.util.concurrent.TimeoutException;
22 + import java.util.concurrent.atomic.AtomicInteger;
23   import java.util.concurrent.locks.LockSupport;
24   import java.util.concurrent.locks.ReentrantLock;
18 import java.util.concurrent.atomic.AtomicInteger;
19 import java.util.concurrent.CountDownLatch;
25  
26   /**
27   * An {@link ExecutorService} for running {@link ForkJoinTask}s.
28   * A {@code ForkJoinPool} provides the entry point for submissions
29 < * from non-{@code ForkJoinTask}s, as well as management and
29 > * from non-{@code ForkJoinTask} clients, as well as management and
30   * monitoring operations.
31   *
32   * <p>A {@code ForkJoinPool} differs from other kinds of {@link
# Line 30 | Line 35 | import java.util.concurrent.CountDownLat
35   * execute subtasks created by other active tasks (eventually blocking
36   * waiting for work if none exist). This enables efficient processing
37   * when most tasks spawn other subtasks (as do most {@code
38 < * ForkJoinTask}s). A {@code ForkJoinPool} may also be used for mixed
39 < * execution of some plain {@code Runnable}- or {@code Callable}-
40 < * based activities along with {@code ForkJoinTask}s. When setting
36 < * {@linkplain #setAsyncMode async mode}, a {@code ForkJoinPool} may
37 < * also be appropriate for use with fine-grained tasks of any form
38 < * that are never joined. Otherwise, other {@code ExecutorService}
39 < * implementations are typically more appropriate choices.
38 > * ForkJoinTask}s). When setting <em>asyncMode</em> to true in
39 > * constructors, {@code ForkJoinPool}s may also be appropriate for use
40 > * with event-style tasks that are never joined.
41   *
42   * <p>A {@code ForkJoinPool} is constructed with a given target
43   * parallelism level; by default, equal to the number of available
44 < * processors. Unless configured otherwise via {@link
45 < * #setMaintainsParallelism}, the pool attempts to maintain this
46 < * number of active (or available) threads by dynamically adding,
47 < * suspending, or resuming internal worker threads, even if some tasks
48 < * are stalled waiting to join others. However, no such adjustments
49 < * are performed in the face of blocked IO or other unmanaged
50 < * synchronization. The nested {@link ManagedBlocker} interface
50 < * enables extension of the kinds of synchronization accommodated.
51 < * The target parallelism level may also be changed dynamically
52 < * ({@link #setParallelism}). The total number of threads may be
53 < * limited using method {@link #setMaximumPoolSize}, in which case it
54 < * may become possible for the activities of a pool to stall due to
55 < * the lack of available threads to process new tasks. When the pool
56 < * is executing tasks, these and other configuration setting methods
57 < * may only gradually affect actual pool sizes. It is normally best
58 < * practice to invoke these methods only when the pool is known to be
59 < * quiescent.
44 > * processors. The pool attempts to maintain enough active (or
45 > * available) threads by dynamically adding, suspending, or resuming
46 > * internal worker threads, even if some tasks are stalled waiting to
47 > * join others. However, no such adjustments are guaranteed in the
48 > * face of blocked IO or other unmanaged synchronization. The nested
49 > * {@link ManagedBlocker} interface enables extension of the kinds of
50 > * synchronization accommodated.
51   *
52   * <p>In addition to execution and lifecycle control methods, this
53   * class provides status check methods (for example
# Line 65 | Line 56 | import java.util.concurrent.CountDownLat
56   * {@link #toString} returns indications of pool state in a
57   * convenient form for informal monitoring.
58   *
59 + * <p> As is the case with other ExecutorServices, there are three
60 + * main task execution methods summarized in the following
61 + * table. These are designed to be used by clients not already engaged
62 + * in fork/join computations in the current pool.  The main forms of
63 + * these methods accept instances of {@code ForkJoinTask}, but
64 + * overloaded forms also allow mixed execution of plain {@code
65 + * Runnable}- or {@code Callable}- based activities as well.  However,
66 + * tasks that are already executing in a pool should normally
67 + * <em>NOT</em> use these pool execution methods, but instead use the
68 + * within-computation forms listed in the table.
69 + *
70 + * <table BORDER CELLPADDING=3 CELLSPACING=1>
71 + *  <tr>
72 + *    <td></td>
73 + *    <td ALIGN=CENTER> <b>Call from non-fork/join clients</b></td>
74 + *    <td ALIGN=CENTER> <b>Call from within fork/join computations</b></td>
75 + *  </tr>
76 + *  <tr>
77 + *    <td> <b>Arrange async execution</td>
78 + *    <td> {@link #execute(ForkJoinTask)}</td>
79 + *    <td> {@link ForkJoinTask#fork}</td>
80 + *  </tr>
81 + *  <tr>
82 + *    <td> <b>Await and obtain result</td>
83 + *    <td> {@link #invoke(ForkJoinTask)}</td>
84 + *    <td> {@link ForkJoinTask#invoke}</td>
85 + *  </tr>
86 + *  <tr>
87 + *    <td> <b>Arrange exec and obtain Future</td>
88 + *    <td> {@link #submit(ForkJoinTask)}</td>
89 + *    <td> {@link ForkJoinTask#fork} (ForkJoinTasks <em>are</em> Futures)</td>
90 + *  </tr>
91 + * </table>
92 + *
93   * <p><b>Sample Usage.</b> Normally a single {@code ForkJoinPool} is
94   * used for all parallel task execution in a program or subsystem.
95   * Otherwise, use would not usually outweigh the construction and
# Line 89 | Line 114 | import java.util.concurrent.CountDownLat
114   * {@code IllegalArgumentException}.
115   *
116   * <p>This implementation rejects submitted tasks (that is, by throwing
117 < * {@link RejectedExecutionException}) only when the pool is shut down.
117 > * {@link RejectedExecutionException}) only when the pool is shut down
118 > * or internal resources have been exhausted.
119   *
120   * @since 1.7
121   * @author Doug Lea
# Line 116 | Line 142 | public class ForkJoinPool extends Abstra
142       * of tasks profit from cache affinities, but others are harmed by
143       * cache pollution effects.)
144       *
145 +     * Beyond work-stealing support and essential bookkeeping, the
146 +     * main responsibility of this framework is to take actions when
147 +     * one worker is waiting to join a task stolen (or always held by)
148 +     * another.  Because we are multiplexing many tasks on to a pool
149 +     * of workers, we can't just let them block (as in Thread.join).
150 +     * We also cannot just reassign the joiner's run-time stack with
151 +     * another and replace it later, which would be a form of
152 +     * "continuation", that even if possible is not necessarily a good
153 +     * idea. Given that the creation costs of most threads on most
154 +     * systems mainly surrounds setting up runtime stacks, thread
155 +     * creation and switching is usually not much more expensive than
156 +     * stack creation and switching, and is more flexible). Instead we
157 +     * combine two tactics:
158 +     *
159 +     *   Helping: Arranging for the joiner to execute some task that it
160 +     *      would be running if the steal had not occurred.  Method
161 +     *      ForkJoinWorkerThread.helpJoinTask tracks joining->stealing
162 +     *      links to try to find such a task.
163 +     *
164 +     *   Compensating: Unless there are already enough live threads,
165 +     *      method helpMaintainParallelism() may create or
166 +     *      re-activate a spare thread to compensate for blocked
167 +     *      joiners until they unblock.
168 +     *
169 +     * It is impossible to keep exactly the target (parallelism)
170 +     * number of threads running at any given time.  Determining
171 +     * existence of conservatively safe helping targets, the
172 +     * availability of already-created spares, and the apparent need
173 +     * to create new spares are all racy and require heuristic
174 +     * guidance, so we rely on multiple retries of each.  Compensation
175 +     * occurs in slow-motion. It is triggered only upon timeouts of
176 +     * Object.wait used for joins. This reduces poor decisions that
177 +     * would otherwise be made when threads are waiting for others
178 +     * that are stalled because of unrelated activities such as
179 +     * garbage collection.
180 +     *
181 +     * The ManagedBlocker extension API can't use helping so relies
182 +     * only on compensation in method awaitBlocker.
183 +     *
184       * The main throughput advantages of work-stealing stem from
185       * decentralized control -- workers mostly steal tasks from each
186       * other. We do not want to negate this by creating bottlenecks
187 <     * implementing the management responsibilities of this class. So
188 <     * we use a collection of techniques that avoid, reduce, or cope
189 <     * well with contention. These entail several instances of
190 <     * bit-packing into CASable fields to maintain only the minimally
191 <     * required atomicity. To enable such packing, we restrict maximum
192 <     * parallelism to (1<<15)-1 (enabling twice this to fit into a 16
193 <     * bit field), which is far in excess of normal operating range.
194 <     * Even though updates to some of these bookkeeping fields do
195 <     * sometimes contend with each other, they don't normally
196 <     * cache-contend with updates to others enough to warrant memory
197 <     * padding or isolation. So they are all held as fields of
198 <     * ForkJoinPool objects.  The main capabilities are as follows:
187 >     * implementing other management responsibilities. So we use a
188 >     * collection of techniques that avoid, reduce, or cope well with
189 >     * contention. These entail several instances of bit-packing into
190 >     * CASable fields to maintain only the minimally required
191 >     * atomicity. To enable such packing, we restrict maximum
192 >     * parallelism to (1<<15)-1 (enabling twice this (to accommodate
193 >     * unbalanced increments and decrements) to fit into a 16 bit
194 >     * field, which is far in excess of normal operating range.  Even
195 >     * though updates to some of these bookkeeping fields do sometimes
196 >     * contend with each other, they don't normally cache-contend with
197 >     * updates to others enough to warrant memory padding or
198 >     * isolation. So they are all held as fields of ForkJoinPool
199 >     * objects.  The main capabilities are as follows:
200       *
201       * 1. Creating and removing workers. Workers are recorded in the
202       * "workers" array. This is an array as opposed to some other data
# Line 146 | Line 212 | public class ForkJoinPool extends Abstra
212       * blocked workers. However, all other support code is set up to
213       * work with other policies.
214       *
215 +     * To ensure that we do not hold on to worker references that
216 +     * would prevent GC, ALL accesses to workers are via indices into
217 +     * the workers array (which is one source of some of the unusual
218 +     * code constructions here). In essence, the workers array serves
219 +     * as a WeakReference mechanism. Thus for example the event queue
220 +     * stores worker indices, not worker references. Access to the
221 +     * workers in associated methods (for example releaseEventWaiters)
222 +     * must both index-check and null-check the IDs. All such accesses
223 +     * ignore bad IDs by returning out early from what they are doing,
224 +     * since this can only be associated with shutdown, in which case
225 +     * it is OK to give up. On termination, we just clobber these
226 +     * data structures without trying to use them.
227 +     *
228       * 2. Bookkeeping for dynamically adding and removing workers. We
229 <     * maintain a given level of parallelism (or, if
230 <     * maintainsParallelism is false, at least avoid starvation). When
152 <     * some workers are known to be blocked (on joins or via
229 >     * aim to approximately maintain the given level of parallelism.
230 >     * When some workers are known to be blocked (on joins or via
231       * ManagedBlocker), we may create or resume others to take their
232       * place until they unblock (see below). Implementing this
233       * requires counts of the number of "running" threads (i.e., those
234 <     * that are neither blocked nor artifically suspended) as well as
234 >     * that are neither blocked nor artificially suspended) as well as
235       * the total number.  These two values are packed into one field,
236       * "workerCounts" because we need accurate snapshots when deciding
237 <     * to create, resume or suspend.  To support these decisions,
238 <     * updates to spare counts must be prospective (not
239 <     * retrospective).  For example, the running count is decremented
240 <     * before blocking by a thread about to block as a spare, but
163 <     * incremented by the thread about to unblock it. Updates upon
164 <     * resumption ofr threads blocking in awaitJoin or awaitBlocker
165 <     * cannot usually be prospective, so the running count is in
166 <     * general an upper bound of the number of productively running
167 <     * threads Updates to the workerCounts field sometimes transiently
168 <     * encounter a fair amount of contention when join dependencies
169 <     * are such that many threads block or unblock at about the same
170 <     * time. We alleviate this by sometimes bundling updates (for
171 <     * example blocking one thread on join and resuming a spare cancel
172 <     * each other out), and in most other cases performing an
173 <     * alternative action like releasing waiters or locating spares.
237 >     * to create, resume or suspend.  Note however that the
238 >     * correspondence of these counts to reality is not guaranteed. In
239 >     * particular updates for unblocked threads may lag until they
240 >     * actually wake up.
241       *
242       * 3. Maintaining global run state. The run state of the pool
243       * consists of a runLevel (SHUTDOWN, TERMINATING, etc) similar to
# Line 199 | Line 266 | public class ForkJoinPool extends Abstra
266       * workers that previously could not find a task to now find one:
267       * Submission of a new task to the pool, or another worker pushing
268       * a task onto a previously empty queue.  (We also use this
269 <     * mechanism for termination and reconfiguration actions that
269 >     * mechanism for configuration and termination actions that
270       * require wakeups of idle workers).  Each worker maintains its
271       * last known event count, and blocks when a scan for work did not
272       * find a task AND its lastEventCount matches the current
# Line 210 | Line 277 | public class ForkJoinPool extends Abstra
277       * a record (field nextEventWaiter) for the next waiting worker.
278       * In addition to allowing simpler decisions about need for
279       * wakeup, the event count bits in eventWaiters serve the role of
280 <     * tags to avoid ABA errors in Treiber stacks.  To reduce delays
281 <     * in task diffusion, workers not otherwise occupied may invoke
282 <     * method releaseWaiters, that removes and signals (unparks)
283 <     * workers not waiting on current count. To minimize task
284 <     * production stalls associate with signalling, any worker pushing
285 <     * a task on an empty queue invokes the weaker method signalWork,
219 <     * that only releases idle workers until it detects interference
220 <     * by other threads trying to release, and lets them take
221 <     * over. The net effect is a tree-like diffusion of signals, where
222 <     * released threads (and possibly others) help with unparks.  To
223 <     * further reduce contention effects a bit, failed CASes to
224 <     * increment field eventCount are tolerated without retries.
280 >     * tags to avoid ABA errors in Treiber stacks. Upon any wakeup,
281 >     * released threads also try to release at most two others.  The
282 >     * net effect is a tree-like diffusion of signals, where released
283 >     * threads (and possibly others) help with unparks.  To further
284 >     * reduce contention effects a bit, failed CASes to increment
285 >     * field eventCount are tolerated without retries in signalWork.
286       * Conceptually they are merged into the same event, which is OK
287       * when their only purpose is to enable workers to scan for work.
288       *
289 <     * 5. Managing suspension of extra workers. When a worker is about
290 <     * to block waiting for a join (or via ManagedBlockers), we may
291 <     * create a new thread to maintain parallelism level, or at least
292 <     * avoid starvation (see below). Usually, extra threads are needed
293 <     * for only very short periods, yet join dependencies are such
294 <     * that we sometimes need them in bursts. Rather than create new
295 <     * threads each time this happens, we suspend no-longer-needed
296 <     * extra ones as "spares". For most purposes, we don't distinguish
297 <     * "extra" spare threads from normal "core" threads: On each call
298 <     * to preStep (the only point at which we can do this) a worker
299 <     * checks to see if there are now too many running workers, and if
300 <     * so, suspends itself.  Methods awaitJoin and awaitBlocker look
301 <     * for suspended threads to resume before considering creating a
302 <     * new replacement. We don't need a special data structure to
303 <     * maintain spares; simply scanning the workers array looking for
304 <     * worker.isSuspended() is fine because the calling thread is
305 <     * otherwise not doing anything useful anyway; we are at least as
306 <     * happy if after locating a spare, the caller doesn't actually
307 <     * block because the join is ready before we try to adjust and
308 <     * compensate.  Note that this is intrinsically racy.  One thread
309 <     * may become a spare at about the same time as another is
310 <     * needlessly being created. We counteract this and related slop
311 <     * in part by requiring resumed spares to immediately recheck (in
312 <     * preStep) to see whether they they should re-suspend. The only
313 <     * effective difference between "extra" and "core" threads is that
314 <     * we allow the "extra" ones to time out and die if they are not
315 <     * resumed within a keep-alive interval of a few seconds. This is
316 <     * implemented mainly within ForkJoinWorkerThread, but requires
317 <     * some coordination (isTrimmed() -- meaning killed while
318 <     * suspended) to correctly maintain pool counts.
319 <     *
320 <     * 6. Deciding when to create new workers. The main dynamic
321 <     * control in this class is deciding when to create extra threads,
322 <     * in methods awaitJoin and awaitBlocker. We always
323 <     * need to create one when the number of running threads becomes
324 <     * zero. But because blocked joins are typically dependent, we
325 <     * don't necessarily need or want one-to-one replacement. Using a
326 <     * one-to-one compensation rule often leads to enough useless
327 <     * overhead creating, suspending, resuming, and/or killing threads
328 <     * to signficantly degrade throughput.  We use a rule reflecting
329 <     * the idea that, the more spare threads you already have, the
330 <     * more evidence you need to create another one. The "evidence"
331 <     * here takes two forms: (1) Using a creation threshold expressed
332 <     * in terms of the current deficit -- target minus running
333 <     * threads. To reduce flickering and drift around target values,
334 <     * the relation is quadratic: adding a spare if (dc*dc)>=(sc*pc)
335 <     * (where dc is deficit, sc is number of spare threads and pc is
336 <     * target parallelism.)  (2) Using a form of adaptive
337 <     * spionning. requiring a number of threshold checks proportional
338 <     * to the number of spare threads.  This effectively reduces churn
339 <     * at the price of systematically undershooting target parallelism
279 <     * when many threads are blocked.  However, biasing toward
280 <     * undeshooting partially compensates for the above mechanics to
281 <     * suspend extra threads, that normally lead to overshoot because
282 <     * we can only suspend workers in-between top-level actions. It
283 <     * also better copes with the fact that some of the methods in
284 <     * this class tend to never become compiled (but are interpreted),
285 <     * so some components of the entire set of controls might execute
286 <     * many times faster than others. And similarly for cases where
287 <     * the apparent lack of work is just due to GC stalls and other
288 <     * transient system activity.
289 <     *
290 <     * 7. Maintaining other configuration parameters and monitoring
291 <     * statistics. Updates to fields controlling parallelism level,
292 <     * max size, etc can only meaningfully take effect for individual
293 <     * threads upon their next top-level actions; i.e., between
294 <     * stealing/running tasks/submission, which are separated by calls
295 <     * to preStep.  Memory ordering for these (assumed infrequent)
296 <     * reconfiguration calls is ensured by using reads and writes to
297 <     * volatile field workerCounts (that must be read in preStep anyway)
298 <     * as "fences" -- user-level reads are preceded by reads of
299 <     * workCounts, and writes are followed by no-op CAS to
300 <     * workerCounts. The values reported by other management and
301 <     * monitoring methods are either computed on demand, or are kept
302 <     * in fields that are only updated when threads are otherwise
303 <     * idle.
289 >     * 5. Managing suspension of extra workers. When a worker notices
290 >     * (usually upon timeout of a wait()) that there are too few
291 >     * running threads, we may create a new thread to maintain
292 >     * parallelism level, or at least avoid starvation. Usually, extra
293 >     * threads are needed for only very short periods, yet join
294 >     * dependencies are such that we sometimes need them in
295 >     * bursts. Rather than create new threads each time this happens,
296 >     * we suspend no-longer-needed extra ones as "spares". For most
297 >     * purposes, we don't distinguish "extra" spare threads from
298 >     * normal "core" threads: On each call to preStep (the only point
299 >     * at which we can do this) a worker checks to see if there are
300 >     * now too many running workers, and if so, suspends itself.
301 >     * Method helpMaintainParallelism looks for suspended threads to
302 >     * resume before considering creating a new replacement. The
303 >     * spares themselves are encoded on another variant of a Treiber
304 >     * Stack, headed at field "spareWaiters".  Note that the use of
305 >     * spares is intrinsically racy.  One thread may become a spare at
306 >     * about the same time as another is needlessly being created. We
307 >     * counteract this and related slop in part by requiring resumed
308 >     * spares to immediately recheck (in preStep) to see whether they
309 >     * should re-suspend.
310 >     *
311 >     * 6. Killing off unneeded workers. A timeout mechanism is used to
312 >     * shed unused workers: The oldest (first) event queue waiter uses
313 >     * a timed rather than hard wait. When this wait times out without
314 >     * a normal wakeup, it tries to shutdown any one (for convenience
315 >     * the newest) other spare or event waiter via
316 >     * tryShutdownUnusedWorker. This eventually reduces the number of
317 >     * worker threads to a minimum of one after a long enough period
318 >     * without use.
319 >     *
320 >     * 7. Deciding when to create new workers. The main dynamic
321 >     * control in this class is deciding when to create extra threads
322 >     * in method helpMaintainParallelism. We would like to keep
323 >     * exactly #parallelism threads running, which is an impossible
324 >     * task. We always need to create one when the number of running
325 >     * threads would become zero and all workers are busy. Beyond
326 >     * this, we must rely on heuristics that work well in the
327 >     * presence of transient phenomena such as GC stalls, dynamic
328 >     * compilation, and wake-up lags. These transients are extremely
329 >     * common -- we are normally trying to fully saturate the CPUs on
330 >     * a machine, so almost any activity other than running tasks
331 >     * impedes accuracy. Our main defense is to allow parallelism to
332 >     * lapse for a while during joins, and use a timeout to see if,
333 >     * after the resulting settling, there is still a need for
334 >     * additional workers.  This also better copes with the fact that
335 >     * some of the methods in this class tend to never become compiled
336 >     * (but are interpreted), so some components of the entire set of
337 >     * controls might execute 100 times faster than others. And
338 >     * similarly for cases where the apparent lack of work is just due
339 >     * to GC stalls and other transient system activity.
340       *
341       * Beware that there is a lot of representation-level coupling
342       * among classes ForkJoinPool, ForkJoinWorkerThread, and
# Line 313 | Line 349 | public class ForkJoinPool extends Abstra
349       *
350       * Style notes: There are lots of inline assignments (of form
351       * "while ((local = field) != 0)") which are usually the simplest
352 <     * way to ensure read orderings. Also several occurrences of the
353 <     * unusual "do {} while(!cas...)" which is the simplest way to
354 <     * force an update of a CAS'ed variable. There are also a few
355 <     * other coding oddities that help some methods perform reasonably
356 <     * even when interpreted (not compiled).
352 >     * way to ensure the required read orderings (which are sometimes
353 >     * critical). Also several occurrences of the unusual "do {}
354 >     * while (!cas...)" which is the simplest way to force an update of
355 >     * a CAS'ed variable. There are also other coding oddities that
356 >     * help some methods perform reasonably even when interpreted (not
357 >     * compiled), at the expense of some messy constructions that
358 >     * reduce byte code counts.
359       *
360       * The order of declarations in this file is: (1) statics (2)
361       * fields (along with constants used when unpacking some of them)
# Line 346 | Line 384 | public class ForkJoinPool extends Abstra
384       * Default ForkJoinWorkerThreadFactory implementation; creates a
385       * new ForkJoinWorkerThread.
386       */
387 <    static class  DefaultForkJoinWorkerThreadFactory
387 >    static class DefaultForkJoinWorkerThreadFactory
388          implements ForkJoinWorkerThreadFactory {
389          public ForkJoinWorkerThread newThread(ForkJoinPool pool) {
390              return new ForkJoinWorkerThread(pool);
# Line 385 | Line 423 | public class ForkJoinPool extends Abstra
423          new AtomicInteger();
424  
425      /**
426 <     * Absolute bound for parallelism level. Twice this number must
427 <     * fit into a 16bit field to enable word-packing for some counts.
426 >     * The time to block in a join (see awaitJoin) before checking if
427 >     * a new worker should be (re)started to maintain parallelism
428 >     * level. The value should be short enough to maintain global
429 >     * responsiveness and progress but long enough to avoid
430 >     * counterproductive firings during GC stalls or unrelated system
431 >     * activity, and to not bog down systems with continual re-firings
432 >     * on GCs or legitimately long waits.
433 >     */
434 >    private static final long JOIN_TIMEOUT_MILLIS = 250L; // 4 per second
435 >
436 >    /**
437 >     * The wakeup interval (in nanoseconds) for the oldest worker
438 >     * waiting for an event to invoke tryShutdownUnusedWorker to
439 >     * shrink the number of workers.  The exact value does not matter
440 >     * too much. It must be short enough to release resources during
441 >     * sustained periods of idleness, but not so short that threads
442 >     * are continually re-created.
443       */
444 <    private static final int MAX_THREADS = 0x7fff;
444 >    private static final long SHRINK_RATE_NANOS =
445 >        30L * 1000L * 1000L * 1000L; // 2 per minute
446 >
447 >    /**
448 >     * Absolute bound for parallelism level. Twice this number plus
449 >     * one (i.e., 0xfff) must fit into a 16bit field to enable
450 >     * word-packing for some counts and indices.
451 >     */
452 >    private static final int MAX_WORKERS   = 0x7fff;
453  
454      /**
455       * Array holding all worker threads in the pool.  Array size must
# Line 414 | Line 475 | public class ForkJoinPool extends Abstra
475      /**
476       * Latch released upon termination.
477       */
478 <    private final CountDownLatch terminationLatch;
478 >    private final Phaser termination;
479  
480      /**
481       * Creation factory for worker threads.
# Line 428 | Line 489 | public class ForkJoinPool extends Abstra
489      private volatile long stealCount;
490  
491      /**
492 <     * Encoded record of top of treiber stack of threads waiting for
492 >     * Encoded record of top of Treiber stack of threads waiting for
493       * events. The top 32 bits contain the count being waited for. The
494 <     * bottom word contains one plus the pool index of waiting worker
495 <     * thread.
494 >     * bottom 16 bits contains one plus the pool index of waiting
495 >     * worker thread. (Bits 16-31 are unused.)
496       */
497      private volatile long eventWaiters;
498  
499      private static final int  EVENT_COUNT_SHIFT = 32;
500 <    private static final long WAITER_INDEX_MASK = (1L << EVENT_COUNT_SHIFT)-1L;
500 >    private static final long WAITER_ID_MASK    = (1L << 16) - 1L;
501  
502      /**
503       * A counter for events that may wake up worker threads:
504       *   - Submission of a new task to the pool
505       *   - A worker pushing a task on an empty queue
506 <     *   - termination and reconfiguration
506 >     *   - termination
507       */
508      private volatile int eventCount;
509  
510      /**
511 +     * Encoded record of top of Treiber stack of spare threads waiting
512 +     * for resumption. The top 16 bits contain an arbitrary count to
513 +     * avoid ABA effects. The bottom 16bits contains one plus the pool
514 +     * index of waiting worker thread.
515 +     */
516 +    private volatile int spareWaiters;
517 +
518 +    private static final int SPARE_COUNT_SHIFT = 16;
519 +    private static final int SPARE_ID_MASK     = (1 << 16) - 1;
520 +
521 +    /**
522       * Lifecycle control. The low word contains the number of workers
523       * that are (probably) executing tasks. This value is atomically
524       * incremented before a worker gets a task to run, and decremented
525 <     * when worker has no tasks and cannot find any.  Bits 16-18
525 >     * when a worker has no tasks and cannot find any.  Bits 16-18
526       * contain runLevel value. When all are zero, the pool is
527       * running. Level transitions are monotonic (running -> shutdown
528       * -> terminating -> terminated) so each transition adds a bit.
529       * These are bundled together to ensure consistent read for
530       * termination checks (i.e., that runLevel is at least SHUTDOWN
531       * and active threads is zero).
532 +     *
533 +     * Notes: Most direct CASes are dependent on these bitfield
534 +     * positions.  Also, this field is non-private to enable direct
535 +     * performance-sensitive CASes in ForkJoinWorkerThread.
536       */
537 <    private volatile int runState;
537 >    volatile int runState;
538  
539      // Note: The order among run level values matters.
540      private static final int RUNLEVEL_SHIFT     = 16;
# Line 466 | Line 542 | public class ForkJoinPool extends Abstra
542      private static final int TERMINATING        = 1 << (RUNLEVEL_SHIFT + 1);
543      private static final int TERMINATED         = 1 << (RUNLEVEL_SHIFT + 2);
544      private static final int ACTIVE_COUNT_MASK  = (1 << RUNLEVEL_SHIFT) - 1;
469    private static final int ONE_ACTIVE         = 1; // active update delta
545  
546      /**
547       * Holds number of total (i.e., created and not yet terminated)
# Line 475 | Line 550 | public class ForkJoinPool extends Abstra
550       * making decisions about creating and suspending spare
551       * threads. Updated only by CAS. Note that adding a new worker
552       * requires incrementing both counts, since workers start off in
553 <     * running state.  This field is also used for memory-fencing
479 <     * configuration parameters.
553 >     * running state.
554       */
555      private volatile int workerCounts;
556  
# Line 485 | Line 559 | public class ForkJoinPool extends Abstra
559      private static final int ONE_RUNNING        = 1;
560      private static final int ONE_TOTAL          = 1 << TOTAL_COUNT_SHIFT;
561  
488    /*
489     * Fields parallelism. maxPoolSize, and maintainsParallelism are
490     * non-volatile, but external reads/writes use workerCount fences
491     * to ensure visability.
492     */
493
562      /**
563       * The target parallelism level.
564 +     * Accessed directly by ForkJoinWorkerThreads.
565       */
566 <    private int parallelism;
498 <
499 <    /**
500 <     * The maximum allowed pool size.
501 <     */
502 <    private int maxPoolSize;
566 >    final int parallelism;
567  
568      /**
569       * True if use local fifo, not default lifo, for local polling
570 <     * Replicated by ForkJoinWorkerThreads
507 <     */
508 <    private volatile boolean locallyFifo;
509 <
510 <    /**
511 <     * Controls whether to add spares to maintain parallelism
570 >     * Read by, and replicated by ForkJoinWorkerThreads
571       */
572 <    private boolean maintainsParallelism;
572 >    final boolean locallyFifo;
573  
574      /**
575 <     * The uncaught exception handler used when any worker
576 <     * abruptly terminates
575 >     * The uncaught exception handler used when any worker abruptly
576 >     * terminates.
577       */
578 <    private volatile Thread.UncaughtExceptionHandler ueh;
578 >    private final Thread.UncaughtExceptionHandler ueh;
579  
580      /**
581       * Pool number, just for assigning useful names to worker threads
582       */
583      private final int poolNumber;
584  
585 <    // utilities for updating fields
585 >    // Utilities for CASing fields. Note that most of these
586 >    // are usually manually inlined by callers
587  
588      /**
589 <     * Adds delta to running count.  Used mainly by ForkJoinTask.
589 >     * Increments running count part of workerCounts
590       */
591 <    final void updateRunningCount(int delta) {
592 <        int wc;
591 >    final void incrementRunningCount() {
592 >        int c;
593          do {} while (!UNSAFE.compareAndSwapInt(this, workerCountsOffset,
594 <                                               wc = workerCounts,
595 <                                               wc + delta));
594 >                                               c = workerCounts,
595 >                                               c + ONE_RUNNING));
596      }
597  
598      /**
599 <     * Decrements running count unless already zero
599 >     * Tries to decrement running count unless already zero
600       */
601      final boolean tryDecrementRunningCount() {
602          int wc = workerCounts;
# Line 547 | Line 607 | public class ForkJoinPool extends Abstra
607      }
608  
609      /**
610 <     * Write fence for user modifications of pool parameters
611 <     * (parallelism. etc).  Note that it doesn't matter if CAS fails.
552 <     */
553 <    private void workerCountWriteFence() {
554 <        int wc;
555 <        UNSAFE.compareAndSwapInt(this, workerCountsOffset,
556 <                                 wc = workerCounts, wc);
557 <    }
558 <
559 <    /**
560 <     * Read fence for external reads of pool parameters
561 <     * (parallelism. maxPoolSize, etc).
562 <     */
563 <    private void workerCountReadFence() {
564 <        int ignore = workerCounts;
565 <    }
566 <
567 <    /**
568 <     * Tries incrementing active count; fails on contention.
569 <     * Called by workers before executing tasks.
610 >     * Forces decrement of encoded workerCounts, awaiting nonzero if
611 >     * (rarely) necessary when other count updates lag.
612       *
613 <     * @return true on success
613 >     * @param dr -- either zero or ONE_RUNNING
614 >     * @param dt -- either zero or ONE_TOTAL
615       */
616 <    final boolean tryIncrementActiveCount() {
617 <        int c;
618 <        return UNSAFE.compareAndSwapInt(this, runStateOffset,
619 <                                        c = runState, c + ONE_ACTIVE);
616 >    private void decrementWorkerCounts(int dr, int dt) {
617 >        for (;;) {
618 >            int wc = workerCounts;
619 >            if ((wc & RUNNING_COUNT_MASK)  - dr < 0 ||
620 >                (wc >>> TOTAL_COUNT_SHIFT) - dt < 0) {
621 >                if ((runState & TERMINATED) != 0)
622 >                    return; // lagging termination on a backout
623 >                Thread.yield();
624 >            }
625 >            if (UNSAFE.compareAndSwapInt(this, workerCountsOffset,
626 >                                         wc, wc - (dr + dt)))
627 >                return;
628 >        }
629      }
630  
631      /**
# Line 583 | Line 635 | public class ForkJoinPool extends Abstra
635      final boolean tryDecrementActiveCount() {
636          int c;
637          return UNSAFE.compareAndSwapInt(this, runStateOffset,
638 <                                        c = runState, c - ONE_ACTIVE);
638 >                                        c = runState, c - 1);
639      }
640  
641      /**
# Line 612 | Line 664 | public class ForkJoinPool extends Abstra
664          lock.lock();
665          try {
666              ForkJoinWorkerThread[] ws = workers;
667 <            int nws = ws.length;
668 <            if (k < 0 || k >= nws || ws[k] != null) {
669 <                for (k = 0; k < nws && ws[k] != null; ++k)
667 >            int n = ws.length;
668 >            if (k < 0 || k >= n || ws[k] != null) {
669 >                for (k = 0; k < n && ws[k] != null; ++k)
670                      ;
671 <                if (k == nws)
672 <                    ws = Arrays.copyOf(ws, nws << 1);
671 >                if (k == n)
672 >                    ws = Arrays.copyOf(ws, n << 1);
673              }
674              ws[k] = w;
675              workers = ws; // volatile array write ensures slot visibility
# Line 628 | Line 680 | public class ForkJoinPool extends Abstra
680      }
681  
682      /**
683 <     * Nulls out record of worker in workers array
683 >     * Nulls out record of worker in workers array.
684       */
685      private void forgetWorker(ForkJoinWorkerThread w) {
686          int idx = w.poolIndex;
687 <        // Locking helps method recordWorker avoid unecessary expansion
687 >        // Locking helps method recordWorker avoid unnecessary expansion
688          final ReentrantLock lock = this.workerLock;
689          lock.lock();
690          try {
# Line 644 | Line 696 | public class ForkJoinPool extends Abstra
696          }
697      }
698  
647    // adding and removing workers
648
699      /**
700 <     * Tries to create and add new worker. Assumes that worker counts
701 <     * are already updated to accommodate the worker, so adjusts on
702 <     * failure.
700 >     * Final callback from terminating worker.  Removes record of
701 >     * worker from array, and adjusts counts. If pool is shutting
702 >     * down, tries to complete termination.
703       *
704 <     * @return new worker or null if creation failed
704 >     * @param w the worker
705       */
706 <    private ForkJoinWorkerThread addWorker() {
707 <        ForkJoinWorkerThread w = null;
708 <        try {
709 <            w = factory.newThread(this);
710 <        } finally { // Adjust on either null or exceptional factory return
711 <            if (w == null) {
662 <                onWorkerCreationFailure();
663 <                return null;
664 <            }
665 <        }
666 <        w.start(recordWorker(w), locallyFifo, ueh);
667 <        return w;
706 >    final void workerTerminated(ForkJoinWorkerThread w) {
707 >        forgetWorker(w);
708 >        decrementWorkerCounts(w.isTrimmed()? 0 : ONE_RUNNING, ONE_TOTAL);
709 >        while (w.stealCount != 0) // collect final count
710 >            tryAccumulateStealCount(w);
711 >        tryTerminate(false);
712      }
713  
714 +    // Waiting for and signalling events
715 +
716      /**
717 <     * Adjusts counts upon failure to create worker
717 >     * Releases workers blocked on a count not equal to current count.
718 >     * Normally called after precheck that eventWaiters isn't zero to
719 >     * avoid wasted array checks. Gives up upon a change in count or
720 >     * upon releasing two workers, letting others take over.
721       */
722 <    private void onWorkerCreationFailure() {
723 <        for (;;) {
724 <            int wc = workerCounts;
725 <            if ((wc >>> TOTAL_COUNT_SHIFT) > 0 &&
726 <                UNSAFE.compareAndSwapInt(this, workerCountsOffset,
727 <                                         wc, wc - (ONE_RUNNING|ONE_TOTAL)))
722 >    private void releaseEventWaiters() {
723 >        ForkJoinWorkerThread[] ws = workers;
724 >        int n = ws.length;
725 >        long h = eventWaiters;
726 >        int ec = eventCount;
727 >        boolean releasedOne = false;
728 >        ForkJoinWorkerThread w; int id;
729 >        while ((id = ((int)(h & WAITER_ID_MASK)) - 1) >= 0 &&
730 >               (int)(h >>> EVENT_COUNT_SHIFT) != ec &&
731 >               id < n && (w = ws[id]) != null) {
732 >            if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
733 >                                          h,  w.nextWaiter)) {
734 >                LockSupport.unpark(w);
735 >                if (releasedOne) // exit on second release
736 >                    break;
737 >                releasedOne = true;
738 >            }
739 >            if (eventCount != ec)
740                  break;
741 +            h = eventWaiters;
742          }
681        tryTerminate(false); // in case of failure during shutdown
743      }
744  
745      /**
746 <     * Create enough total workers to establish target parallelism,
747 <     * giving up if terminating or addWorker fails
746 >     * Tries to advance eventCount and releases waiters. Called only
747 >     * from workers.
748       */
749 <    private void ensureEnoughTotalWorkers() {
750 <        int wc;
751 <        while (((wc = workerCounts) >>> TOTAL_COUNT_SHIFT) < parallelism &&
752 <               runState < TERMINATING) {
753 <            if ((UNSAFE.compareAndSwapInt(this, workerCountsOffset,
693 <                                          wc, wc + (ONE_RUNNING|ONE_TOTAL)) &&
694 <                 addWorker() == null))
695 <                break;
696 <        }
749 >    final void signalWork() {
750 >        int c; // try to increment event count -- CAS failure OK
751 >        UNSAFE.compareAndSwapInt(this, eventCountOffset, c = eventCount, c+1);
752 >        if (eventWaiters != 0L)
753 >            releaseEventWaiters();
754      }
755  
756      /**
757 <     * Final callback from terminating worker.  Removes record of
758 <     * worker from array, and adjusts counts. If pool is shutting
702 <     * down, tries to complete terminatation, else possibly replaces
703 <     * the worker.
757 >     * Adds the given worker to event queue and blocks until
758 >     * terminating or event count advances from the given value
759       *
760 <     * @param w the worker
760 >     * @param w the calling worker thread
761 >     * @param ec the count
762       */
763 <    final void workerTerminated(ForkJoinWorkerThread w) {
764 <        if (w.active) { // force inactive
765 <            w.active = false;
766 <            do {} while (!tryDecrementActiveCount());
767 <        }
768 <        forgetWorker(w);
769 <
770 <        // Decrement total count, and if was running, running count
771 <        // Spin (waiting for other updates) if either would be negative
772 <        int nr = w.isTrimmed() ? 0 : ONE_RUNNING;
717 <        int unit = ONE_TOTAL + nr;
718 <        for (;;) {
719 <            int wc = workerCounts;
720 <            int rc = wc & RUNNING_COUNT_MASK;
721 <            if (rc - nr < 0 || (wc >>> TOTAL_COUNT_SHIFT) == 0)
722 <                Thread.yield(); // back off if waiting for other updates
723 <            else if (UNSAFE.compareAndSwapInt(this, workerCountsOffset,
724 <                                              wc, wc - unit))
763 >    private void eventSync(ForkJoinWorkerThread w, int ec) {
764 >        long nh = (((long)ec) << EVENT_COUNT_SHIFT) | ((long)(w.poolIndex+1));
765 >        long h;
766 >        while ((runState < SHUTDOWN || !tryTerminate(false)) &&
767 >               (((int)((h = eventWaiters) & WAITER_ID_MASK)) == 0 ||
768 >                (int)(h >>> EVENT_COUNT_SHIFT) == ec) &&
769 >               eventCount == ec) {
770 >            if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
771 >                                          w.nextWaiter = h, nh)) {
772 >                awaitEvent(w, ec);
773                  break;
774 +            }
775          }
727
728        accumulateStealCount(w); // collect final count
729        if (!tryTerminate(false))
730            ensureEnoughTotalWorkers();
776      }
777  
733    // Waiting for and signalling events
734
778      /**
779 <     * Ensures eventCount on exit is different (mod 2^32) than on
780 <     * entry.  CAS failures are OK -- any change in count suffices.
779 >     * Blocks the given worker (that has already been entered as an
780 >     * event waiter) until terminating or event count advances from
781 >     * the given value. The oldest (first) waiter uses a timed wait to
782 >     * occasionally one-by-one shrink the number of workers (to a
783 >     * minimum of one) if the pool has not been used for extended
784 >     * periods.
785 >     *
786 >     * @param w the calling worker thread
787 >     * @param ec the count
788       */
789 <    private void advanceEventCount() {
790 <        int c;
791 <        UNSAFE.compareAndSwapInt(this, eventCountOffset, c = eventCount, c+1);
789 >    private void awaitEvent(ForkJoinWorkerThread w, int ec) {
790 >        while (eventCount == ec) {
791 >            if (tryAccumulateStealCount(w)) { // transfer while idle
792 >                boolean untimed = (w.nextWaiter != 0L ||
793 >                                   (workerCounts & RUNNING_COUNT_MASK) <= 1);
794 >                long startTime = untimed? 0 : System.nanoTime();
795 >                Thread.interrupted();         // clear/ignore interrupt
796 >                if (eventCount != ec || w.isTerminating())
797 >                    break;                    // recheck after clear
798 >                if (untimed)
799 >                    LockSupport.park(w);
800 >                else {
801 >                    LockSupport.parkNanos(w, SHRINK_RATE_NANOS);
802 >                    if (eventCount != ec || w.isTerminating())
803 >                        break;
804 >                    if (System.nanoTime() - startTime >= SHRINK_RATE_NANOS)
805 >                        tryShutdownUnusedWorker(ec);
806 >                }
807 >            }
808 >        }
809      }
810  
811 +    // Maintaining parallelism
812 +
813      /**
814 <     * Releases workers blocked on a count not equal to current count.
814 >     * Pushes worker onto the spare stack.
815       */
816 <    final void releaseWaiters() {
817 <        long top;
818 <        int id;
819 <        while ((id = (int)((top = eventWaiters) & WAITER_INDEX_MASK)) > 0 &&
751 <               (int)(top >>> EVENT_COUNT_SHIFT) != eventCount) {
752 <            ForkJoinWorkerThread[] ws = workers;
753 <            ForkJoinWorkerThread w;
754 <            if (ws.length >= id && (w = ws[id - 1]) != null &&
755 <                UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
756 <                                          top, w.nextWaiter))
757 <                LockSupport.unpark(w);
758 <        }
816 >    final void pushSpare(ForkJoinWorkerThread w) {
817 >        int ns = (++w.spareCount << SPARE_COUNT_SHIFT) | (w.poolIndex + 1);
818 >        do {} while (!UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
819 >                                               w.nextSpare = spareWaiters,ns));
820      }
821  
822      /**
823 <     * Advances eventCount and releases waiters until interference by
824 <     * other releasing threads is detected.
823 >     * Tries (once) to resume a spare if the number of running
824 >     * threads is less than target.
825       */
826 <    final void signalWork() {
827 <        int ec;
828 <        UNSAFE.compareAndSwapInt(this, eventCountOffset, ec=eventCount, ec+1);
829 <        outer:for (;;) {
830 <            long top = eventWaiters;
831 <            ec = eventCount;
832 <            for (;;) {
833 <                ForkJoinWorkerThread[] ws; ForkJoinWorkerThread w;
834 <                int id = (int)(top & WAITER_INDEX_MASK);
835 <                if (id <= 0 || (int)(top >>> EVENT_COUNT_SHIFT) == ec)
836 <                    return;
837 <                if ((ws = workers).length < id || (w = ws[id - 1]) == null ||
838 <                    !UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
839 <                                               top, top = w.nextWaiter))
840 <                    continue outer;      // possibly stale; reread
826 >    private void tryResumeSpare() {
827 >        int sw, id;
828 >        ForkJoinWorkerThread[] ws = workers;
829 >        int n = ws.length;
830 >        ForkJoinWorkerThread w;
831 >        if ((sw = spareWaiters) != 0 &&
832 >            (id = (sw & SPARE_ID_MASK) - 1) >= 0 &&
833 >            id < n && (w = ws[id]) != null &&
834 >            (workerCounts & RUNNING_COUNT_MASK) < parallelism &&
835 >            spareWaiters == sw &&
836 >            UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
837 >                                     sw, w.nextSpare)) {
838 >            int c; // increment running count before resume
839 >            do {} while (!UNSAFE.compareAndSwapInt
840 >                         (this, workerCountsOffset,
841 >                          c = workerCounts, c + ONE_RUNNING));
842 >            if (w.tryUnsuspend())
843                  LockSupport.unpark(w);
844 <                if (top != eventWaiters) // let someone else take over
845 <                    return;
783 <            }
844 >            else   // back out if w was shutdown
845 >                decrementWorkerCounts(ONE_RUNNING, 0);
846          }
847      }
848  
849      /**
850 <     * If worker is inactive, blocks until terminating or event count
851 <     * advances from last value held by worker; in any case helps
852 <     * release others.
853 <     *
854 <     * @param w the calling worker thread
850 >     * Tries to increase the number of running workers if below target
851 >     * parallelism: If a spare exists tries to resume it via
852 >     * tryResumeSpare.  Otherwise, if not enough total workers or all
853 >     * existing workers are busy, adds a new worker. In all cases also
854 >     * helps wake up releasable workers waiting for work.
855       */
856 <    private void eventSync(ForkJoinWorkerThread w) {
857 <        if (!w.active) {
858 <            int prev = w.lastEventCount;
859 <            long nextTop = (((long)prev << EVENT_COUNT_SHIFT) |
860 <                            ((long)(w.poolIndex + 1)));
861 <            long top;
862 <            while ((runState < SHUTDOWN || !tryTerminate(false)) &&
863 <                   (((int)(top = eventWaiters) & WAITER_INDEX_MASK) == 0 ||
864 <                    (int)(top >>> EVENT_COUNT_SHIFT) == prev) &&
865 <                   eventCount == prev) {
866 <                if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
867 <                                              w.nextWaiter = top, nextTop)) {
868 <                    accumulateStealCount(w); // transfer steals while idle
869 <                    Thread.interrupted();    // clear/ignore interrupt
870 <                    while (eventCount == prev)
871 <                        w.doPark();
856 >    private void helpMaintainParallelism() {
857 >        int pc = parallelism;
858 >        int wc, rs, tc;
859 >        while (((wc = workerCounts) & RUNNING_COUNT_MASK) < pc &&
860 >               (rs = runState) < TERMINATING) {
861 >            if (spareWaiters != 0)
862 >                tryResumeSpare();
863 >            else if ((tc = wc >>> TOTAL_COUNT_SHIFT) >= MAX_WORKERS ||
864 >                     (tc >= pc && (rs & ACTIVE_COUNT_MASK) != tc))
865 >                break;   // enough total
866 >            else if (runState == rs && workerCounts == wc &&
867 >                     UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
868 >                                              wc + (ONE_RUNNING|ONE_TOTAL))) {
869 >                ForkJoinWorkerThread w = null;
870 >                Throwable fail = null;
871 >                try {
872 >                    w = factory.newThread(this);
873 >                } catch (Throwable ex) {
874 >                    fail = ex;
875 >                }
876 >                if (w == null) { // null or exceptional factory return
877 >                    decrementWorkerCounts(ONE_RUNNING, ONE_TOTAL);
878 >                    tryTerminate(false); // handle failure during shutdown
879 >                    // If originating from an external caller,
880 >                    // propagate exception, else ignore
881 >                    if (fail != null && runState < TERMINATING &&
882 >                        !(Thread.currentThread() instanceof
883 >                          ForkJoinWorkerThread))
884 >                        UNSAFE.throwException(fail);
885                      break;
886                  }
887 +                w.start(recordWorker(w), ueh);
888 +                if ((workerCounts >>> TOTAL_COUNT_SHIFT) >= pc) {
889 +                    int c; // advance event count
890 +                    UNSAFE.compareAndSwapInt(this, eventCountOffset,
891 +                                             c = eventCount, c+1);
892 +                    break; // add at most one unless total below target
893 +                }
894              }
813            w.lastEventCount = eventCount;
895          }
896 <        releaseWaiters();
896 >        if (eventWaiters != 0L)
897 >            releaseEventWaiters();
898      }
899  
900      /**
901 <     * Callback from workers invoked upon each top-level action (i.e.,
902 <     * stealing a task or taking a submission and running
903 <     * it). Performs one or both of the following:
904 <     *
905 <     * * If the worker cannot find work, updates its active status to
906 <     * inactive and updates activeCount unless there is contention, in
907 <     * which case it may try again (either in this or a subsequent
826 <     * call).  Additionally, awaits the next task event and/or helps
827 <     * wake up other releasable waiters.
828 <     *
829 <     * * If there are too many running threads, suspends this worker
830 <     * (first forcing inactivation if necessary).  If it is not
831 <     * resumed before a keepAlive elapses, the worker may be "trimmed"
832 <     * -- killed while suspended within suspendAsSpare. Otherwise,
833 <     * upon resume it rechecks to make sure that it is still needed.
901 >     * Callback from the oldest waiter in awaitEvent waking up after a
902 >     * period of non-use. If all workers are idle, tries (once) to
903 >     * shutdown an event waiter or a spare, if one exists. Note that
904 >     * we don't need CAS or locks here because the method is called
905 >     * only from one thread occasionally waking (and even misfires are
906 >     * OK). Note that until the shutdown worker fully terminates,
907 >     * workerCounts will overestimate total count, which is tolerable.
908       *
909 <     * @param w the worker
910 <     * @param worked false if the worker scanned for work but didn't
837 <     * find any (in which case it may block waiting for work).
909 >     * @param ec the event count waited on by caller (to abort
910 >     * attempt if count has since changed).
911       */
912 <    final void preStep(ForkJoinWorkerThread w, boolean worked) {
913 <        boolean active = w.active;
914 <        boolean inactivate = !worked & active;
915 <        for (;;) {
916 <            if (inactivate) {
917 <                int c = runState;
918 <                if (UNSAFE.compareAndSwapInt(this, runStateOffset,
919 <                                             c, c - ONE_ACTIVE))
920 <                    inactivate = active = w.active = false;
912 >    private void tryShutdownUnusedWorker(int ec) {
913 >        if (runState == 0 && eventCount == ec) { // only trigger if all idle
914 >            ForkJoinWorkerThread[] ws = workers;
915 >            int n = ws.length;
916 >            ForkJoinWorkerThread w = null;
917 >            boolean shutdown = false;
918 >            int sw;
919 >            long h;
920 >            if ((sw = spareWaiters) != 0) { // prefer killing spares
921 >                int id = (sw & SPARE_ID_MASK) - 1;
922 >                if (id >= 0 && id < n && (w = ws[id]) != null &&
923 >                    UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
924 >                                             sw, w.nextSpare))
925 >                    shutdown = true;
926              }
927 <            int wc = workerCounts;
928 <            if ((wc & RUNNING_COUNT_MASK) <= parallelism) {
929 <                if (!worked)
930 <                    eventSync(w);
931 <                return;
927 >            else if ((h = eventWaiters) != 0L) {
928 >                long nh;
929 >                int id = ((int)(h & WAITER_ID_MASK)) - 1;
930 >                if (id >= 0 && id < n && (w = ws[id]) != null &&
931 >                    (nh = w.nextWaiter) != 0L && // keep at least one worker
932 >                    UNSAFE.compareAndSwapLong(this, eventWaitersOffset, h, nh))
933 >                    shutdown = true;
934 >            }
935 >            if (w != null && shutdown) {
936 >                w.shutdown();
937 >                LockSupport.unpark(w);
938              }
855            if (!(inactivate |= active) &&  // must inactivate to suspend
856                UNSAFE.compareAndSwapInt(this, workerCountsOffset,
857                                         wc, wc - ONE_RUNNING) &&
858                !w.suspendAsSpare())        // false if trimmed
859                return;
939          }
940 +        releaseEventWaiters(); // in case of interference
941      }
942  
943      /**
944 <     * Adjusts counts and creates or resumes compensating threads for
945 <     * a worker blocking on task joinMe.  First tries resuming an
946 <     * existing spare (which usually also avoids any count
867 <     * adjustment), but must then decrement running count to determine
868 <     * whether a new thread is needed. See above for fuller
869 <     * explanation. This code is sprawled out non-modularly mainly
870 <     * because adaptive spinning works best if the entire method is
871 <     * either interpreted or compiled vs having only some pieces of it
872 <     * compiled.
944 >     * Callback from workers invoked upon each top-level action (i.e.,
945 >     * stealing a task or taking a submission and running it).
946 >     * Performs one or more of the following:
947       *
948 <     * @param joinMe the task to join
949 <     * @return task status on exit (to simplify usage by callers)
948 >     * 1. If the worker is active and either did not run a task
949 >     *    or there are too many workers, try to set its active status
950 >     *    to inactive and update activeCount. On contention, we may
951 >     *    try again in this or a subsequent call.
952 >     *
953 >     * 2. If not enough total workers, help create some.
954 >     *
955 >     * 3. If there are too many running workers, suspend this worker
956 >     *    (first forcing inactive if necessary).  If it is not needed,
957 >     *    it may be shutdown while suspended (via
958 >     *    tryShutdownUnusedWorker).  Otherwise, upon resume it
959 >     *    rechecks running thread count and need for event sync.
960 >     *
961 >     * 4. If worker did not run a task, await the next task event via
962 >     *    eventSync if necessary (first forcing inactivation), upon
963 >     *    which the worker may be shutdown via
964 >     *    tryShutdownUnusedWorker.  Otherwise, help release any
965 >     *    existing event waiters that are now releasable,
966 >     *
967 >     * @param w the worker
968 >     * @param ran true if worker ran a task since last call to this method
969       */
970 <    final int awaitJoin(ForkJoinTask<?> joinMe) {
970 >    final void preStep(ForkJoinWorkerThread w, boolean ran) {
971 >        int wec = w.lastEventCount;
972 >        boolean active = w.active;
973 >        boolean inactivate = false;
974          int pc = parallelism;
975 <        boolean adj = false;        // true when running count adjusted
976 <        int scans = 0;
977 <
978 <        while (joinMe.status >= 0) {
979 <            ForkJoinWorkerThread spare = null;
884 <            if ((workerCounts & RUNNING_COUNT_MASK) < pc) {
885 <                ForkJoinWorkerThread[] ws = workers;
886 <                int nws = ws.length;
887 <                for (int i = 0; i < nws; ++i) {
888 <                    ForkJoinWorkerThread w = ws[i];
889 <                    if (w != null && w.isSuspended()) {
890 <                        spare = w;
891 <                        break;
892 <                    }
893 <                }
894 <                if (joinMe.status < 0)
895 <                    break;
975 >        while (w.runState == 0) {
976 >            int rs = runState;
977 >            if (rs >= TERMINATING) { // propagate shutdown
978 >                w.shutdown();
979 >                break;
980              }
981 +            if ((inactivate || (active && (rs & ACTIVE_COUNT_MASK) >= pc)) &&
982 +                UNSAFE.compareAndSwapInt(this, runStateOffset, rs, rs - 1))
983 +                inactivate = active = w.active = false;
984              int wc = workerCounts;
985 <            int rc = wc & RUNNING_COUNT_MASK;
986 <            int dc = pc - rc;
987 <            if (dc > 0 && spare != null && spare.tryUnsuspend()) {
988 <                if (adj) {
989 <                    int c;
990 <                    do {} while (!UNSAFE.compareAndSwapInt
904 <                                 (this, workerCountsOffset,
905 <                                  c = workerCounts, c + ONE_RUNNING));
906 <                }
907 <                adj = true;
908 <                LockSupport.unpark(spare);
985 >            if ((wc & RUNNING_COUNT_MASK) > pc) {
986 >                if (!(inactivate |= active) && // must inactivate to suspend
987 >                    workerCounts == wc &&      // try to suspend as spare
988 >                    UNSAFE.compareAndSwapInt(this, workerCountsOffset,
989 >                                             wc, wc - ONE_RUNNING))
990 >                    w.suspendAsSpare();
991              }
992 <            else if (adj) {
993 <                if (dc <= 0)
992 >            else if ((wc >>> TOTAL_COUNT_SHIFT) < pc)
993 >                helpMaintainParallelism();     // not enough workers
994 >            else if (!ran) {
995 >                long h = eventWaiters;
996 >                int ec = eventCount;
997 >                if (h != 0L && (int)(h >>> EVENT_COUNT_SHIFT) != ec)
998 >                    releaseEventWaiters();     // release others before waiting
999 >                else if (ec != wec) {
1000 >                    w.lastEventCount = ec;     // no need to wait
1001                      break;
913                int tc = wc >>> TOTAL_COUNT_SHIFT;
914                if (scans > tc) {
915                    int ts = (tc - pc) * pc;
916                    if (rc != 0 &&  (dc * dc < ts || !maintainsParallelism))
917                        break;
918                    if (scans > ts && tc < maxPoolSize &&
919                        UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
920                                                 wc+(ONE_RUNNING|ONE_TOTAL))){
921                        addWorker();
922                        break;
923                    }
1002                  }
1003 +                else if (!(inactivate |= active))
1004 +                    eventSync(w, wec);         // must inactivate before sync
1005              }
926            else if (rc != 0)
927                adj = UNSAFE.compareAndSwapInt (this, workerCountsOffset,
928                                                wc, wc - ONE_RUNNING);
929            if ((scans++ & 1) == 0)
930                releaseWaiters();   // help others progress
1006              else
1007 <                Thread.yield();     // avoid starving productive threads
933 <        }
934 <
935 <        if (adj) {
936 <            joinMe.internalAwaitDone();
937 <            int c;
938 <            do {} while (!UNSAFE.compareAndSwapInt
939 <                         (this, workerCountsOffset,
940 <                          c = workerCounts, c + ONE_RUNNING));
1007 >                break;
1008          }
942        return joinMe.status;
1009      }
1010  
1011      /**
1012 <     * Same idea as awaitJoin
1012 >     * Helps and/or blocks awaiting join of the given task.
1013 >     * See above for explanation.
1014 >     *
1015 >     * @param joinMe the task to join
1016 >     * @param worker the current worker thread
1017       */
1018 <    final void awaitBlocker(ManagedBlocker blocker, boolean maintainPar)
1019 <        throws InterruptedException {
1020 <        maintainPar &= maintainsParallelism;
1021 <        int pc = parallelism;
1022 <        boolean adj = false;        // true when running count adjusted
1023 <        int scans = 0;
954 <        boolean done;
955 <
956 <        for (;;) {
957 <            if (done = blocker.isReleasable())
1018 >    final void awaitJoin(ForkJoinTask<?> joinMe, ForkJoinWorkerThread worker) {
1019 >        int retries = 2 + (parallelism >> 2); // #helpJoins before blocking
1020 >        while (joinMe.status >= 0) {
1021 >            int wc;
1022 >            worker.helpJoinTask(joinMe);
1023 >            if (joinMe.status < 0)
1024                  break;
1025 <            ForkJoinWorkerThread spare = null;
1026 <            if ((workerCounts & RUNNING_COUNT_MASK) < pc) {
1027 <                ForkJoinWorkerThread[] ws = workers;
1028 <                int nws = ws.length;
1029 <                for (int i = 0; i < nws; ++i) {
1030 <                    ForkJoinWorkerThread w = ws[i];
1031 <                    if (w != null && w.isSuspended()) {
1032 <                        spare = w;
1033 <                        break;
1034 <                    }
1035 <                }
1036 <                if (done = blocker.isReleasable())
1037 <                    break;
1038 <            }
1039 <            int wc = workerCounts;
974 <            int rc = wc & RUNNING_COUNT_MASK;
975 <            int dc = pc - rc;
976 <            if (dc > 0 && spare != null && spare.tryUnsuspend()) {
977 <                if (adj) {
978 <                    int c;
979 <                    do {} while (!UNSAFE.compareAndSwapInt
980 <                                 (this, workerCountsOffset,
981 <                                  c = workerCounts, c + ONE_RUNNING));
982 <                }
983 <                adj = true;
984 <                LockSupport.unpark(spare);
985 <            }
986 <            else if (adj) {
987 <                if (dc <= 0)
988 <                    break;
989 <                int tc = wc >>> TOTAL_COUNT_SHIFT;
990 <                if (scans > tc) {
991 <                    int ts = (tc - pc) * pc;
992 <                    if (rc != 0 &&  (dc * dc < ts || !maintainPar))
993 <                        break;
994 <                    if (scans > ts && tc < maxPoolSize &&
995 <                        UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
996 <                                                 wc+(ONE_RUNNING|ONE_TOTAL))){
997 <                        addWorker();
998 <                        break;
999 <                    }
1000 <                }
1001 <            }
1002 <            else if (rc != 0)
1003 <                adj = UNSAFE.compareAndSwapInt (this, workerCountsOffset,
1004 <                                                wc, wc - ONE_RUNNING);
1005 <            if ((++scans & 1) == 0)
1006 <                releaseWaiters();   // help others progress
1007 <            else
1008 <                Thread.yield();     // avoid starving productive threads
1009 <        }
1010 <
1011 <        try {
1012 <            if (!done)
1013 <                do {} while (!blocker.isReleasable() && !blocker.block());
1014 <        } finally {
1015 <            if (adj) {
1016 <                int c;
1025 >            else if (retries > 0)
1026 >                --retries;
1027 >            else if (((wc = workerCounts) & RUNNING_COUNT_MASK) != 0 &&
1028 >                     UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1029 >                                              wc, wc - ONE_RUNNING)) {
1030 >                int stat, c; long h;
1031 >                while ((stat = joinMe.status) >= 0 &&
1032 >                       (h = eventWaiters) != 0L && // help release others
1033 >                       (int)(h >>> EVENT_COUNT_SHIFT) != eventCount)
1034 >                    releaseEventWaiters();
1035 >                if (stat >= 0 &&
1036 >                    ((workerCounts & RUNNING_COUNT_MASK) == 0 ||
1037 >                     (stat =
1038 >                      joinMe.internalAwaitDone(JOIN_TIMEOUT_MILLIS)) >= 0))
1039 >                    helpMaintainParallelism(); // timeout or no running workers
1040                  do {} while (!UNSAFE.compareAndSwapInt
1041                               (this, workerCountsOffset,
1042                                c = workerCounts, c + ONE_RUNNING));
1043 +                if (stat < 0)
1044 +                    break;   // else restart
1045              }
1046          }
1047      }
1048  
1049      /**
1050 <     * Unless there are not enough other running threads, adjusts
1026 <     * counts and blocks a worker performing helpJoin that cannot find
1027 <     * any work.
1028 <     *
1029 <     * @return true if joinMe now done
1050 >     * Same idea as awaitJoin, but no helping, retries, or timeouts.
1051       */
1052 <    final boolean tryAwaitBusyJoin(ForkJoinTask<?> joinMe) {
1053 <        int pc = parallelism;
1054 <        outer:for (;;) {
1055 <            releaseWaiters();
1056 <            if ((workerCounts & RUNNING_COUNT_MASK) < pc) {
1057 <                ForkJoinWorkerThread[] ws = workers;
1058 <                int nws = ws.length;
1059 <                for (int i = 0; i < nws; ++i) {
1060 <                    ForkJoinWorkerThread w = ws[i];
1061 <                    if (w != null && w.isSuspended()) {
1062 <                        if (joinMe.status < 0)
1063 <                            return true;
1064 <                        if ((workerCounts & RUNNING_COUNT_MASK) > pc)
1052 >    final void awaitBlocker(ManagedBlocker blocker)
1053 >        throws InterruptedException {
1054 >        while (!blocker.isReleasable()) {
1055 >            int wc = workerCounts;
1056 >            if ((wc & RUNNING_COUNT_MASK) != 0 &&
1057 >                UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1058 >                                         wc, wc - ONE_RUNNING)) {
1059 >                try {
1060 >                    while (!blocker.isReleasable()) {
1061 >                        long h = eventWaiters;
1062 >                        if (h != 0L &&
1063 >                            (int)(h >>> EVENT_COUNT_SHIFT) != eventCount)
1064 >                            releaseEventWaiters();
1065 >                        else if ((workerCounts & RUNNING_COUNT_MASK) == 0 &&
1066 >                                 runState < TERMINATING)
1067 >                            helpMaintainParallelism();
1068 >                        else if (blocker.block())
1069                              break;
1045                        if (w.tryUnsuspend()) {
1046                            LockSupport.unpark(w);
1047                            break outer;
1048                        }
1049                        continue outer;
1070                      }
1071 +                } finally {
1072 +                    int c;
1073 +                    do {} while (!UNSAFE.compareAndSwapInt
1074 +                                 (this, workerCountsOffset,
1075 +                                  c = workerCounts, c + ONE_RUNNING));
1076                  }
1052            }
1053            if (joinMe.status < 0)
1054                return true;
1055            int wc = workerCounts;
1056            if ((wc & RUNNING_COUNT_MASK) <= 2 ||
1057                (wc >>> TOTAL_COUNT_SHIFT) < pc)
1058                return false;  // keep this thread alive
1059            if (UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1060                                         wc, wc - ONE_RUNNING))
1077                  break;
1078 +            }
1079          }
1063
1064        joinMe.internalAwaitDone();
1065        int c;
1066        do {} while (!UNSAFE.compareAndSwapInt
1067                     (this, workerCountsOffset,
1068                      c = workerCounts, c + ONE_RUNNING));
1069        return true;
1080      }
1081  
1082      /**
# Line 1090 | Line 1100 | public class ForkJoinPool extends Abstra
1100          // Finish now if all threads terminated; else in some subsequent call
1101          if ((workerCounts >>> TOTAL_COUNT_SHIFT) == 0) {
1102              advanceRunLevel(TERMINATED);
1103 <            terminationLatch.countDown();
1103 >            termination.arrive();
1104          }
1105          return true;
1106      }
1107  
1108 +
1109      /**
1110       * Actions on transition to TERMINATING
1111 +     *
1112 +     * Runs up to four passes through workers: (0) shutting down each
1113 +     * (without waking up if parked) to quickly spread notifications
1114 +     * without unnecessary bouncing around event queues etc (1) wake
1115 +     * up and help cancel tasks (2) interrupt (3) mop up races with
1116 +     * interrupted workers
1117       */
1118      private void startTerminating() {
1119 <        for (int i = 0; i < 2; ++i) { // twice to mop up newly created workers
1120 <            cancelSubmissions();
1121 <            shutdownWorkers();
1122 <            cancelWorkerTasks();
1123 <            advanceEventCount();
1124 <            releaseWaiters();
1125 <            interruptWorkers();
1119 >        cancelSubmissions();
1120 >        for (int passes = 0; passes < 4 && workerCounts != 0; ++passes) {
1121 >            int c; // advance event count
1122 >            UNSAFE.compareAndSwapInt(this, eventCountOffset,
1123 >                                     c = eventCount, c+1);
1124 >            eventWaiters = 0L; // clobber lists
1125 >            spareWaiters = 0;
1126 >            for (ForkJoinWorkerThread w : workers) {
1127 >                if (w != null) {
1128 >                    w.shutdown();
1129 >                    if (passes > 0 && !w.isTerminated()) {
1130 >                        w.cancelTasks();
1131 >                        LockSupport.unpark(w);
1132 >                        if (passes > 1 && !w.isInterrupted()) {
1133 >                            try {
1134 >                                w.interrupt();
1135 >                            } catch (SecurityException ignore) {
1136 >                            }
1137 >                        }
1138 >                    }
1139 >                }
1140 >            }
1141          }
1142      }
1143  
1144      /**
1145 <     * Clear out and cancel submissions, ignoring exceptions
1145 >     * Clears out and cancels submissions, ignoring exceptions.
1146       */
1147      private void cancelSubmissions() {
1148          ForkJoinTask<?> task;
# Line 1122 | Line 1154 | public class ForkJoinPool extends Abstra
1154          }
1155      }
1156  
1125    /**
1126     * Sets all worker run states to at least shutdown,
1127     * also resuming suspended workers
1128     */
1129    private void shutdownWorkers() {
1130        ForkJoinWorkerThread[] ws = workers;
1131        int nws = ws.length;
1132        for (int i = 0; i < nws; ++i) {
1133            ForkJoinWorkerThread w = ws[i];
1134            if (w != null)
1135                w.shutdown();
1136        }
1137    }
1138
1139    /**
1140     * Clears out and cancels all locally queued tasks
1141     */
1142    private void cancelWorkerTasks() {
1143        ForkJoinWorkerThread[] ws = workers;
1144        int nws = ws.length;
1145        for (int i = 0; i < nws; ++i) {
1146            ForkJoinWorkerThread w = ws[i];
1147            if (w != null)
1148                w.cancelTasks();
1149        }
1150    }
1151
1152    /**
1153     * Unsticks all workers blocked on joins etc
1154     */
1155    private void interruptWorkers() {
1156        ForkJoinWorkerThread[] ws = workers;
1157        int nws = ws.length;
1158        for (int i = 0; i < nws; ++i) {
1159            ForkJoinWorkerThread w = ws[i];
1160            if (w != null && !w.isTerminated()) {
1161                try {
1162                    w.interrupt();
1163                } catch (SecurityException ignore) {
1164                }
1165            }
1166        }
1167    }
1168
1157      // misc support for ForkJoinWorkerThread
1158  
1159      /**
1160 <     * Returns pool number
1160 >     * Returns pool number.
1161       */
1162      final int getPoolNumber() {
1163          return poolNumber;
1164      }
1165  
1166      /**
1167 <     * Accumulates steal count from a worker, clearing
1168 <     * the worker's value
1167 >     * Tries to accumulate steal count from a worker, clearing
1168 >     * the worker's value if successful.
1169 >     *
1170 >     * @return true if worker steal count now zero
1171       */
1172 <    final void accumulateStealCount(ForkJoinWorkerThread w) {
1172 >    final boolean tryAccumulateStealCount(ForkJoinWorkerThread w) {
1173          int sc = w.stealCount;
1174 <        if (sc != 0) {
1175 <            long c;
1176 <            w.stealCount = 0;
1177 <            do {} while (!UNSAFE.compareAndSwapLong(this, stealCountOffset,
1178 <                                                    c = stealCount, c + sc));
1174 >        long c = stealCount;
1175 >        // CAS even if zero, for fence effects
1176 >        if (UNSAFE.compareAndSwapLong(this, stealCountOffset, c, c + sc)) {
1177 >            if (sc != 0)
1178 >                w.stealCount = 0;
1179 >            return true;
1180          }
1181 +        return sc == 0;
1182      }
1183  
1184      /**
# Line 1194 | Line 1186 | public class ForkJoinPool extends Abstra
1186       * active thread.
1187       */
1188      final int idlePerActive() {
1189 <        int ac = runState;    // no mask -- artifically boosts during shutdown
1190 <        int pc = parallelism; // use targeted parallelism, not rc
1189 >        int pc = parallelism; // use parallelism, not rc
1190 >        int ac = runState;    // no mask -- artificially boosts during shutdown
1191          // Use exact results for small values, saturate past 4
1192 <        return pc <= ac? 0 : pc >>> 1 <= ac? 1 : pc >>> 2 <= ac? 3 : pc >>> 3;
1192 >        return ((pc <= ac) ? 0 :
1193 >                (pc >>> 1 <= ac) ? 1 :
1194 >                (pc >>> 2 <= ac) ? 3 :
1195 >                pc >>> 3);
1196      }
1197  
1198      // Public and protected methods
# Line 1206 | Line 1201 | public class ForkJoinPool extends Abstra
1201  
1202      /**
1203       * Creates a {@code ForkJoinPool} with parallelism equal to {@link
1204 <     * java.lang.Runtime#availableProcessors}, and using the {@linkplain
1205 <     * #defaultForkJoinWorkerThreadFactory default thread factory}.
1204 >     * java.lang.Runtime#availableProcessors}, using the {@linkplain
1205 >     * #defaultForkJoinWorkerThreadFactory default thread factory},
1206 >     * no UncaughtExceptionHandler, and non-async LIFO processing mode.
1207       *
1208       * @throws SecurityException if a security manager exists and
1209       *         the caller is not permitted to modify threads
# Line 1216 | Line 1212 | public class ForkJoinPool extends Abstra
1212       */
1213      public ForkJoinPool() {
1214          this(Runtime.getRuntime().availableProcessors(),
1215 <             defaultForkJoinWorkerThreadFactory);
1215 >             defaultForkJoinWorkerThreadFactory, null, false);
1216      }
1217  
1218      /**
1219       * Creates a {@code ForkJoinPool} with the indicated parallelism
1220 <     * level and using the {@linkplain
1221 <     * #defaultForkJoinWorkerThreadFactory default thread factory}.
1220 >     * level, the {@linkplain
1221 >     * #defaultForkJoinWorkerThreadFactory default thread factory},
1222 >     * no UncaughtExceptionHandler, and non-async LIFO processing mode.
1223       *
1224       * @param parallelism the parallelism level
1225       * @throws IllegalArgumentException if parallelism less than or
# Line 1233 | Line 1230 | public class ForkJoinPool extends Abstra
1230       *         java.lang.RuntimePermission}{@code ("modifyThread")}
1231       */
1232      public ForkJoinPool(int parallelism) {
1233 <        this(parallelism, defaultForkJoinWorkerThreadFactory);
1237 <    }
1238 <
1239 <    /**
1240 <     * Creates a {@code ForkJoinPool} with parallelism equal to {@link
1241 <     * java.lang.Runtime#availableProcessors}, and using the given
1242 <     * thread factory.
1243 <     *
1244 <     * @param factory the factory for creating new threads
1245 <     * @throws NullPointerException if the factory is null
1246 <     * @throws SecurityException if a security manager exists and
1247 <     *         the caller is not permitted to modify threads
1248 <     *         because it does not hold {@link
1249 <     *         java.lang.RuntimePermission}{@code ("modifyThread")}
1250 <     */
1251 <    public ForkJoinPool(ForkJoinWorkerThreadFactory factory) {
1252 <        this(Runtime.getRuntime().availableProcessors(), factory);
1233 >        this(parallelism, defaultForkJoinWorkerThreadFactory, null, false);
1234      }
1235  
1236      /**
1237 <     * Creates a {@code ForkJoinPool} with the given parallelism and
1257 <     * thread factory.
1237 >     * Creates a {@code ForkJoinPool} with the given parameters.
1238       *
1239 <     * @param parallelism the parallelism level
1240 <     * @param factory the factory for creating new threads
1239 >     * @param parallelism the parallelism level. For default value,
1240 >     * use {@link java.lang.Runtime#availableProcessors}.
1241 >     * @param factory the factory for creating new threads. For default value,
1242 >     * use {@link #defaultForkJoinWorkerThreadFactory}.
1243 >     * @param handler the handler for internal worker threads that
1244 >     * terminate due to unrecoverable errors encountered while executing
1245 >     * tasks. For default value, use {@code null}.
1246 >     * @param asyncMode if true,
1247 >     * establishes local first-in-first-out scheduling mode for forked
1248 >     * tasks that are never joined. This mode may be more appropriate
1249 >     * than default locally stack-based mode in applications in which
1250 >     * worker threads only process event-style asynchronous tasks.
1251 >     * For default value, use {@code false}.
1252       * @throws IllegalArgumentException if parallelism less than or
1253       *         equal to zero, or greater than implementation limit
1254       * @throws NullPointerException if the factory is null
# Line 1266 | Line 1257 | public class ForkJoinPool extends Abstra
1257       *         because it does not hold {@link
1258       *         java.lang.RuntimePermission}{@code ("modifyThread")}
1259       */
1260 <    public ForkJoinPool(int parallelism, ForkJoinWorkerThreadFactory factory) {
1260 >    public ForkJoinPool(int parallelism,
1261 >                        ForkJoinWorkerThreadFactory factory,
1262 >                        Thread.UncaughtExceptionHandler handler,
1263 >                        boolean asyncMode) {
1264          checkPermission();
1265          if (factory == null)
1266              throw new NullPointerException();
1267 <        if (parallelism <= 0 || parallelism > MAX_THREADS)
1267 >        if (parallelism <= 0 || parallelism > MAX_WORKERS)
1268              throw new IllegalArgumentException();
1275        this.poolNumber = poolNumberGenerator.incrementAndGet();
1276        int arraySize = initialArraySizeFor(parallelism);
1269          this.parallelism = parallelism;
1270          this.factory = factory;
1271 <        this.maxPoolSize = MAX_THREADS;
1272 <        this.maintainsParallelism = true;
1271 >        this.ueh = handler;
1272 >        this.locallyFifo = asyncMode;
1273 >        int arraySize = initialArraySizeFor(parallelism);
1274          this.workers = new ForkJoinWorkerThread[arraySize];
1275          this.submissionQueue = new LinkedTransferQueue<ForkJoinTask<?>>();
1276          this.workerLock = new ReentrantLock();
1277 <        this.terminationLatch = new CountDownLatch(1);
1277 >        this.termination = new Phaser(1);
1278 >        this.poolNumber = poolNumberGenerator.incrementAndGet();
1279      }
1280  
1281      /**
# Line 1289 | Line 1283 | public class ForkJoinPool extends Abstra
1283       * @param pc the initial parallelism level
1284       */
1285      private static int initialArraySizeFor(int pc) {
1286 <        // See Hackers Delight, sec 3.2. We know MAX_THREADS < (1 >>> 16)
1287 <        int size = pc < MAX_THREADS ? pc + 1 : MAX_THREADS;
1286 >        // If possible, initially allocate enough space for one spare
1287 >        int size = pc < MAX_WORKERS ? pc + 1 : MAX_WORKERS;
1288 >        // See Hackers Delight, sec 3.2. We know MAX_WORKERS < (1 >>> 16)
1289          size |= size >>> 1;
1290          size |= size >>> 2;
1291          size |= size >>> 4;
# Line 1309 | Line 1304 | public class ForkJoinPool extends Abstra
1304          if (runState >= SHUTDOWN)
1305              throw new RejectedExecutionException();
1306          submissionQueue.offer(task);
1307 <        advanceEventCount();
1308 <        releaseWaiters();
1309 <        ensureEnoughTotalWorkers();
1307 >        int c; // try to increment event count -- CAS failure OK
1308 >        UNSAFE.compareAndSwapInt(this, eventCountOffset, c = eventCount, c+1);
1309 >        helpMaintainParallelism(); // create, start, or resume some workers
1310      }
1311  
1312      /**
# Line 1357 | Line 1352 | public class ForkJoinPool extends Abstra
1352      }
1353  
1354      /**
1355 +     * Submits a ForkJoinTask for execution.
1356 +     *
1357 +     * @param task the task to submit
1358 +     * @return the task
1359 +     * @throws NullPointerException if the task is null
1360 +     * @throws RejectedExecutionException if the task cannot be
1361 +     *         scheduled for execution
1362 +     */
1363 +    public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) {
1364 +        doSubmit(task);
1365 +        return task;
1366 +    }
1367 +
1368 +    /**
1369       * @throws NullPointerException if the task is null
1370       * @throws RejectedExecutionException if the task cannot be
1371       *         scheduled for execution
# Line 1394 | Line 1403 | public class ForkJoinPool extends Abstra
1403      }
1404  
1405      /**
1397     * Submits a ForkJoinTask for execution.
1398     *
1399     * @param task the task to submit
1400     * @return the task
1401     * @throws NullPointerException if the task is null
1402     * @throws RejectedExecutionException if the task cannot be
1403     *         scheduled for execution
1404     */
1405    public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) {
1406        doSubmit(task);
1407        return task;
1408    }
1409
1410    /**
1406       * @throws NullPointerException       {@inheritDoc}
1407       * @throws RejectedExecutionException {@inheritDoc}
1408       */
# Line 1449 | Line 1444 | public class ForkJoinPool extends Abstra
1444       * @return the handler, or {@code null} if none
1445       */
1446      public Thread.UncaughtExceptionHandler getUncaughtExceptionHandler() {
1452        workerCountReadFence();
1447          return ueh;
1448      }
1449  
1450      /**
1457     * Sets the handler for internal worker threads that terminate due
1458     * to unrecoverable errors encountered while executing tasks.
1459     * Unless set, the current default or ThreadGroup handler is used
1460     * as handler.
1461     *
1462     * @param h the new handler
1463     * @return the old handler, or {@code null} if none
1464     * @throws SecurityException if a security manager exists and
1465     *         the caller is not permitted to modify threads
1466     *         because it does not hold {@link
1467     *         java.lang.RuntimePermission}{@code ("modifyThread")}
1468     */
1469    public Thread.UncaughtExceptionHandler
1470        setUncaughtExceptionHandler(Thread.UncaughtExceptionHandler h) {
1471        checkPermission();
1472        Thread.UncaughtExceptionHandler old = ueh;
1473        if (h != old) {
1474            ueh = h;
1475            ForkJoinWorkerThread[] ws = workers;
1476            int nws = ws.length;
1477            for (int i = 0; i < nws; ++i) {
1478                ForkJoinWorkerThread w = ws[i];
1479                if (w != null)
1480                    w.setUncaughtExceptionHandler(h);
1481            }
1482        }
1483        return old;
1484    }
1485
1486    /**
1487     * Sets the target parallelism level of this pool.
1488     *
1489     * @param parallelism the target parallelism
1490     * @throws IllegalArgumentException if parallelism less than or
1491     * equal to zero or greater than maximum size bounds
1492     * @throws SecurityException if a security manager exists and
1493     *         the caller is not permitted to modify threads
1494     *         because it does not hold {@link
1495     *         java.lang.RuntimePermission}{@code ("modifyThread")}
1496     */
1497    public void setParallelism(int parallelism) {
1498        checkPermission();
1499        if (parallelism <= 0 || parallelism > maxPoolSize)
1500            throw new IllegalArgumentException();
1501        workerCountReadFence();
1502        int pc = this.parallelism;
1503        if (pc != parallelism) {
1504            this.parallelism = parallelism;
1505            workerCountWriteFence();
1506            // Release spares. If too many, some will die after re-suspend
1507            ForkJoinWorkerThread[] ws = workers;
1508            int nws = ws.length;
1509            for (int i = 0; i < nws; ++i) {
1510                ForkJoinWorkerThread w = ws[i];
1511                if (w != null && w.tryUnsuspend()) {
1512                    int c;
1513                    do {} while (!UNSAFE.compareAndSwapInt
1514                                 (this, workerCountsOffset,
1515                                  c = workerCounts, c + ONE_RUNNING));
1516                    LockSupport.unpark(w);
1517                }
1518            }
1519            ensureEnoughTotalWorkers();
1520            advanceEventCount();
1521            releaseWaiters(); // force config recheck by existing workers
1522        }
1523    }
1524
1525    /**
1451       * Returns the targeted parallelism level of this pool.
1452       *
1453       * @return the targeted parallelism level of this pool
1454       */
1455      public int getParallelism() {
1531        //        workerCountReadFence(); // inlined below
1532        int ignore = workerCounts;
1456          return parallelism;
1457      }
1458  
1459      /**
1460       * Returns the number of worker threads that have started but not
1461 <     * yet terminated.  This result returned by this method may differ
1461 >     * yet terminated.  The result returned by this method may differ
1462       * from {@link #getParallelism} when threads are created to
1463       * maintain parallelism when others are cooperatively blocked.
1464       *
# Line 1546 | Line 1469 | public class ForkJoinPool extends Abstra
1469      }
1470  
1471      /**
1549     * Returns the maximum number of threads allowed to exist in the
1550     * pool. Unless set using {@link #setMaximumPoolSize}, the
1551     * maximum is an implementation-defined value designed only to
1552     * prevent runaway growth.
1553     *
1554     * @return the maximum
1555     */
1556    public int getMaximumPoolSize() {
1557        workerCountReadFence();
1558        return maxPoolSize;
1559    }
1560
1561    /**
1562     * Sets the maximum number of threads allowed to exist in the
1563     * pool. The given value should normally be greater than or equal
1564     * to the {@link #getParallelism parallelism} level. Setting this
1565     * value has no effect on current pool size. It controls
1566     * construction of new threads. The use of this method may cause
1567     * tasks that intrinsically require extra threads for dependent
1568     * computations to indefinitely stall. If you are instead trying
1569     * to minimize internal thread creation, consider setting {@link
1570     * #setMaintainsParallelism} as false.
1571     *
1572     * @throws IllegalArgumentException if negative or greater than
1573     * internal implementation limit
1574     */
1575    public void setMaximumPoolSize(int newMax) {
1576        if (newMax < 0 || newMax > MAX_THREADS)
1577            throw new IllegalArgumentException();
1578        maxPoolSize = newMax;
1579        workerCountWriteFence();
1580    }
1581
1582    /**
1583     * Returns {@code true} if this pool dynamically maintains its
1584     * target parallelism level. If false, new threads are added only
1585     * to avoid possible starvation.  This setting is by default true.
1586     *
1587     * @return {@code true} if maintains parallelism
1588     */
1589    public boolean getMaintainsParallelism() {
1590        workerCountReadFence();
1591        return maintainsParallelism;
1592    }
1593
1594    /**
1595     * Sets whether this pool dynamically maintains its target
1596     * parallelism level. If false, new threads are added only to
1597     * avoid possible starvation.
1598     *
1599     * @param enable {@code true} to maintain parallelism
1600     */
1601    public void setMaintainsParallelism(boolean enable) {
1602        maintainsParallelism = enable;
1603        workerCountWriteFence();
1604    }
1605
1606    /**
1607     * Establishes local first-in-first-out scheduling mode for forked
1608     * tasks that are never joined. This mode may be more appropriate
1609     * than default locally stack-based mode in applications in which
1610     * worker threads only process asynchronous tasks.  This method is
1611     * designed to be invoked only when the pool is quiescent, and
1612     * typically only before any tasks are submitted. The effects of
1613     * invocations at other times may be unpredictable.
1614     *
1615     * @param async if {@code true}, use locally FIFO scheduling
1616     * @return the previous mode
1617     * @see #getAsyncMode
1618     */
1619    public boolean setAsyncMode(boolean async) {
1620        workerCountReadFence();
1621        boolean oldMode = locallyFifo;
1622        if (oldMode != async) {
1623            locallyFifo = async;
1624            workerCountWriteFence();
1625            ForkJoinWorkerThread[] ws = workers;
1626            int nws = ws.length;
1627            for (int i = 0; i < nws; ++i) {
1628                ForkJoinWorkerThread w = ws[i];
1629                if (w != null)
1630                    w.setAsyncMode(async);
1631            }
1632        }
1633        return oldMode;
1634    }
1635
1636    /**
1472       * Returns {@code true} if this pool uses local first-in-first-out
1473       * scheduling mode for forked tasks that are never joined.
1474       *
1475       * @return {@code true} if this pool uses async mode
1641     * @see #setAsyncMode
1476       */
1477      public boolean getAsyncMode() {
1644        workerCountReadFence();
1478          return locallyFifo;
1479      }
1480  
# Line 1710 | Line 1543 | public class ForkJoinPool extends Abstra
1543       */
1544      public long getQueuedTaskCount() {
1545          long count = 0;
1546 <        ForkJoinWorkerThread[] ws = workers;
1714 <        int nws = ws.length;
1715 <        for (int i = 0; i < nws; ++i) {
1716 <            ForkJoinWorkerThread w = ws[i];
1546 >        for (ForkJoinWorkerThread w : workers)
1547              if (w != null)
1548                  count += w.getQueueSize();
1719        }
1549          return count;
1550      }
1551  
# Line 1770 | Line 1599 | public class ForkJoinPool extends Abstra
1599       * @return the number of elements transferred
1600       */
1601      protected int drainTasksTo(Collection<? super ForkJoinTask<?>> c) {
1602 <        int n = submissionQueue.drainTo(c);
1603 <        ForkJoinWorkerThread[] ws = workers;
1775 <        int nws = ws.length;
1776 <        for (int i = 0; i < nws; ++i) {
1777 <            ForkJoinWorkerThread w = ws[i];
1602 >        int count = submissionQueue.drainTo(c);
1603 >        for (ForkJoinWorkerThread w : workers)
1604              if (w != null)
1605 <                n += w.drainTasksTo(c);
1606 <        }
1781 <        return n;
1605 >                count += w.drainTasksTo(c);
1606 >        return count;
1607      }
1608  
1609      /**
# Line 1881 | Line 1706 | public class ForkJoinPool extends Abstra
1706      }
1707  
1708      /**
1709 +     * Returns true if terminating or terminated. Used by ForkJoinWorkerThread.
1710 +     */
1711 +    final boolean isAtLeastTerminating() {
1712 +        return runState >= TERMINATING;
1713 +    }
1714 +        
1715 +    /**
1716       * Returns {@code true} if this pool has been shut down.
1717       *
1718       * @return {@code true} if this pool has been shut down
# Line 1902 | Line 1734 | public class ForkJoinPool extends Abstra
1734       */
1735      public boolean awaitTermination(long timeout, TimeUnit unit)
1736          throws InterruptedException {
1737 <        return terminationLatch.await(timeout, unit);
1737 >        try {
1738 >            return termination.awaitAdvanceInterruptibly(0, timeout, unit) > 0;
1739 >        } catch (TimeoutException ex) {
1740 >            return false;
1741 >        }
1742      }
1743  
1744      /**
1745       * Interface for extending managed parallelism for tasks running
1746       * in {@link ForkJoinPool}s.
1747       *
1748 <     * <p>A {@code ManagedBlocker} provides two methods.
1749 <     * Method {@code isReleasable} must return {@code true} if
1750 <     * blocking is not necessary. Method {@code block} blocks the
1751 <     * current thread if necessary (perhaps internally invoking
1752 <     * {@code isReleasable} before actually blocking).
1748 >     * <p>A {@code ManagedBlocker} provides two methods.  Method
1749 >     * {@code isReleasable} must return {@code true} if blocking is
1750 >     * not necessary. Method {@code block} blocks the current thread
1751 >     * if necessary (perhaps internally invoking {@code isReleasable}
1752 >     * before actually blocking). The unusual methods in this API
1753 >     * accommodate synchronizers that may, but don't usually, block
1754 >     * for long periods. Similarly, they allow more efficient internal
1755 >     * handling of cases in which additional workers may be, but
1756 >     * usually are not, needed to ensure sufficient parallelism.
1757 >     * Toward this end, implementations of method {@code isReleasable}
1758 >     * must be amenable to repeated invocation.
1759       *
1760       * <p>For example, here is a ManagedBlocker based on a
1761       * ReentrantLock:
# Line 1931 | Line 1773 | public class ForkJoinPool extends Abstra
1773       *     return hasLock || (hasLock = lock.tryLock());
1774       *   }
1775       * }}</pre>
1776 +     *
1777 +     * <p>Here is a class that possibly blocks waiting for an
1778 +     * item on a given queue:
1779 +     *  <pre> {@code
1780 +     * class QueueTaker<E> implements ManagedBlocker {
1781 +     *   final BlockingQueue<E> queue;
1782 +     *   volatile E item = null;
1783 +     *   QueueTaker(BlockingQueue<E> q) { this.queue = q; }
1784 +     *   public boolean block() throws InterruptedException {
1785 +     *     if (item == null)
1786 +     *       item = queue.take();
1787 +     *     return true;
1788 +     *   }
1789 +     *   public boolean isReleasable() {
1790 +     *     return item != null || (item = queue.poll()) != null;
1791 +     *   }
1792 +     *   public E getItem() { // call after pool.managedBlock completes
1793 +     *     return item;
1794 +     *   }
1795 +     * }}</pre>
1796       */
1797      public static interface ManagedBlocker {
1798          /**
# Line 1954 | Line 1816 | public class ForkJoinPool extends Abstra
1816       * Blocks in accord with the given blocker.  If the current thread
1817       * is a {@link ForkJoinWorkerThread}, this method possibly
1818       * arranges for a spare thread to be activated if necessary to
1819 <     * ensure parallelism while the current thread is blocked.
1958 <     *
1959 <     * <p>If {@code maintainParallelism} is {@code true} and the pool
1960 <     * supports it ({@link #getMaintainsParallelism}), this method
1961 <     * attempts to maintain the pool's nominal parallelism. Otherwise
1962 <     * it activates a thread only if necessary to avoid complete
1963 <     * starvation. This option may be preferable when blockages use
1964 <     * timeouts, or are almost always brief.
1819 >     * ensure sufficient parallelism while the current thread is blocked.
1820       *
1821       * <p>If the caller is not a {@link ForkJoinTask}, this method is
1822       * behaviorally equivalent to
# Line 1975 | Line 1830 | public class ForkJoinPool extends Abstra
1830       * first be expanded to ensure parallelism, and later adjusted.
1831       *
1832       * @param blocker the blocker
1978     * @param maintainParallelism if {@code true} and supported by
1979     * this pool, attempt to maintain the pool's nominal parallelism;
1980     * otherwise activate a thread only if necessary to avoid
1981     * complete starvation.
1833       * @throws InterruptedException if blocker.block did so
1834       */
1835 <    public static void managedBlock(ManagedBlocker blocker,
1985 <                                    boolean maintainParallelism)
1835 >    public static void managedBlock(ManagedBlocker blocker)
1836          throws InterruptedException {
1837          Thread t = Thread.currentThread();
1838 <        if (t instanceof ForkJoinWorkerThread)
1839 <            ((ForkJoinWorkerThread) t).pool.
1840 <                awaitBlocker(blocker, maintainParallelism);
1841 <        else
1842 <            awaitBlocker(blocker);
1843 <    }
1844 <
1995 <    /**
1996 <     * Performs Non-FJ blocking
1997 <     */
1998 <    private static void awaitBlocker(ManagedBlocker blocker)
1999 <        throws InterruptedException {
2000 <        do {} while (!blocker.isReleasable() && !blocker.block());
1838 >        if (t instanceof ForkJoinWorkerThread) {
1839 >            ForkJoinWorkerThread w = (ForkJoinWorkerThread) t;
1840 >            w.pool.awaitBlocker(blocker);
1841 >        }
1842 >        else {
1843 >            do {} while (!blocker.isReleasable() && !blocker.block());
1844 >        }
1845      }
1846  
1847      // AbstractExecutorService overrides.  These rely on undocumented
# Line 2022 | Line 1866 | public class ForkJoinPool extends Abstra
1866      private static final long eventCountOffset =
1867          objectFieldOffset("eventCount", ForkJoinPool.class);
1868      private static final long eventWaitersOffset =
1869 <        objectFieldOffset("eventWaiters",ForkJoinPool.class);
1869 >        objectFieldOffset("eventWaiters", ForkJoinPool.class);
1870      private static final long stealCountOffset =
1871 <        objectFieldOffset("stealCount",ForkJoinPool.class);
1872 <
1871 >        objectFieldOffset("stealCount", ForkJoinPool.class);
1872 >    private static final long spareWaitersOffset =
1873 >        objectFieldOffset("spareWaiters", ForkJoinPool.class);
1874  
1875      private static long objectFieldOffset(String field, Class<?> klazz) {
1876          try {

Diff Legend

Removed lines
+ Added lines
< Changed lines
> Changed lines