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 1.55 by dl, Sun Apr 18 13:59:57 2010 UTC vs.
Revision 1.90 by jsr166, Mon Nov 29 20:58:06 2010 UTC

# Line 6 | Line 6
6  
7   package jsr166y;
8  
9 import java.util.concurrent.*;
10
9   import java.util.ArrayList;
10   import java.util.Arrays;
11   import java.util.Collection;
12   import java.util.Collections;
13   import java.util.List;
14 + import java.util.concurrent.AbstractExecutorService;
15 + import java.util.concurrent.Callable;
16 + import java.util.concurrent.ExecutorService;
17 + import java.util.concurrent.Future;
18 + import java.util.concurrent.RejectedExecutionException;
19 + import java.util.concurrent.RunnableFuture;
20 + import java.util.concurrent.TimeUnit;
21 + import java.util.concurrent.TimeoutException;
22 + import java.util.concurrent.atomic.AtomicInteger;
23   import java.util.concurrent.locks.LockSupport;
24   import java.util.concurrent.locks.ReentrantLock;
18 import java.util.concurrent.atomic.AtomicInteger;
19 import java.util.concurrent.CountDownLatch;
25  
26   /**
27   * An {@link ExecutorService} for running {@link ForkJoinTask}s.
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 140 | Line 206 | public class ForkJoinPool extends Abstra
206       * (workerLock) but the array is otherwise concurrently readable,
207       * and accessed directly by workers. To simplify index-based
208       * operations, the array size is always a power of two, and all
209 <     * readers must tolerate null slots. Currently, all but the first
210 <     * worker thread creation is on-demand, triggered by task
211 <     * submissions, replacement of terminated workers, and/or
212 <     * compensation for blocked workers. However, all other support
213 <     * code is set up to work with other policies.
209 >     * readers must tolerate null slots. Currently, all worker thread
210 >     * creation is on-demand, triggered by task submissions,
211 >     * replacement of terminated workers, and/or compensation for
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 must be prospective (not retrospective).  For example,
239 <     * the running count is decremented before blocking by a thread
240 <     * about to block, but incremented by the thread about to unblock
163 <     * it. (In a few cases, these prospective updates may need to be
164 <     * rolled back, for example when deciding to create a new worker
165 <     * but the thread factory fails or returns null. In these cases,
166 <     * we are no worse off wrt other decisions than we would be
167 <     * otherwise.)  Updates to the workerCounts field sometimes
168 <     * transiently encounter a fair amount of contention when join
169 <     * dependencies are such that many threads block or unblock at
170 <     * about the same time. We alleviate this by sometimes bundling
171 <     * updates (for example blocking one thread on join and resuming a
172 <     * spare cancel each other out), and in most other cases
173 <     * performing an alternative action (like releasing waiters and
174 <     * finding spares; see below) as a more productive form of
175 <     * backoff.
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 201 | 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 212 | 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,
221 <     * that only releases idle workers until it detects interference
222 <     * by other threads trying to release, and lets them take
223 <     * over. The net effect is a tree-like diffusion of signals, where
224 <     * released threads and possibly others) help with unparks.  To
225 <     * further reduce contention effects a bit, failed CASes to
226 <     * 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 preJoin and doBlock look for
301 <     * suspended threads to resume before considering creating a new
302 <     * replacement. We don't need a special data structure to maintain
303 <     * 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 preJoin and doBlock. We always need to create one
323 <     * when the number of running threads becomes zero. But because
324 <     * blocked joins are typically dependent, we don't necessarily
325 <     * need or want one-to-one replacement. Using a one-to-one
326 <     * compensation rule often leads to enough useless overhead
327 <     * creating, suspending, resuming, and/or killing threads to
328 <     * signficantly degrade throughput.  We use a rule reflecting the
329 <     * idea that, the more spare threads you already have, the more
330 <     * evidence you need to create another one; where "evidence" is
331 <     * expressed as the current deficit -- target minus running
332 <     * threads. To reduce flickering and drift around target values,
333 <     * the relation is quadratic: adding a spare if (dc*dc)>=(sc*pc)
334 <     * (where dc is deficit, sc is number of spare threads and pc is
335 <     * target parallelism.)  This effectively reduces churn at the
336 <     * price of systematically undershooting target parallelism when
337 <     * many threads are blocked.  However, biasing toward undeshooting
338 <     * partially compensates for the above mechanics to suspend extra
339 <     * threads, that normally lead to overshoot because we can only
281 <     * suspend workers in-between top-level actions. It also better
282 <     * copes with the fact that some of the methods in this class tend
283 <     * to never become compiled (but are interpreted), so some
284 <     * components of the entire set of controls might execute many
285 <     * times faster than others. And similarly for cases where the
286 <     * apparent lack of work is just due to GC stalls and other
287 <     * transient system activity.
288 <     *
289 <     * 7. Maintaining other configuration parameters and monitoring
290 <     * statistics. Updates to fields controlling parallelism level,
291 <     * max size, etc can only meaningfully take effect for individual
292 <     * threads upon their next top-level actions; i.e., between
293 <     * stealing/running tasks/submission, which are separated by calls
294 <     * to preStep.  Memory ordering for these (assumed infrequent)
295 <     * reconfiguration calls is ensured by using reads and writes to
296 <     * volatile field workerCounts (that must be read in preStep anyway)
297 <     * as "fences" -- user-level reads are preceded by reads of
298 <     * workCounts, and writes are followed by no-op CAS to
299 <     * workerCounts. The values reported by other management and
300 <     * monitoring methods are either computed on demand, or are kept
301 <     * in fields that are only updated when threads are otherwise
302 <     * 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 312 | 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 345 | 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 384 | Line 423 | public class ForkJoinPool extends Abstra
423          new AtomicInteger();
424  
425      /**
426 <     * Absolute bound for parallelism level. Twice this number must
427 <     * fit into a 16bit field to enable word-packing for some counts.
426 >     * The time to block in a join (see awaitJoin) before checking if
427 >     * a new worker should be (re)started to maintain parallelism
428 >     * level. The value should be short enough to maintain global
429 >     * responsiveness and progress but long enough to avoid
430 >     * counterproductive firings during GC stalls or unrelated system
431 >     * activity, and to not bog down systems with continual re-firings
432 >     * on GCs or legitimately long waits.
433 >     */
434 >    private static final long JOIN_TIMEOUT_MILLIS = 250L; // 4 per second
435 >
436 >    /**
437 >     * The wakeup interval (in nanoseconds) for the oldest worker
438 >     * waiting for an event to invoke tryShutdownUnusedWorker to
439 >     * shrink the number of workers.  The exact value does not matter
440 >     * too much. It must be short enough to release resources during
441 >     * sustained periods of idleness, but not so short that threads
442 >     * are continually re-created.
443 >     */
444 >    private static final long SHRINK_RATE_NANOS =
445 >        30L * 1000L * 1000L * 1000L; // 2 per minute
446 >
447 >    /**
448 >     * Absolute bound for parallelism level. Twice this number plus
449 >     * one (i.e., 0xfff) must fit into a 16bit field to enable
450 >     * word-packing for some counts and indices.
451       */
452 <    private static final int MAX_THREADS = 0x7fff;
452 >    private static final int MAX_WORKERS   = 0x7fff;
453  
454      /**
455       * Array holding all worker threads in the pool.  Array size must
# Line 413 | 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 427 | Line 489 | public class ForkJoinPool extends Abstra
489      private volatile long stealCount;
490  
491      /**
492 <     * Encoded record of top of treiber stack of threads waiting for
492 >     * Encoded record of top of Treiber stack of threads waiting for
493       * events. The top 32 bits contain the count being waited for. The
494 <     * bottom word contains one plus the pool index of waiting worker
495 <     * thread.
494 >     * bottom 16 bits contains one plus the pool index of waiting
495 >     * worker thread. (Bits 16-31 are unused.)
496       */
497      private volatile long eventWaiters;
498  
499 <    private static final int  EVENT_COUNT_SHIFT = 32;
500 <    private static final long WAITER_INDEX_MASK = (1L << EVENT_COUNT_SHIFT)-1L;
499 >    private static final int EVENT_COUNT_SHIFT = 32;
500 >    private static final int WAITER_ID_MASK    = (1 << 16) - 1;
501  
502      /**
503       * A counter for events that may wake up worker threads:
504       *   - Submission of a new task to the pool
505       *   - A worker pushing a task on an empty queue
506 <     *   - termination and reconfiguration
506 >     *   - termination
507       */
508      private volatile int eventCount;
509  
510      /**
511 +     * Encoded record of top of Treiber stack of spare threads waiting
512 +     * for resumption. The top 16 bits contain an arbitrary count to
513 +     * avoid ABA effects. The bottom 16bits contains one plus the pool
514 +     * index of waiting worker thread.
515 +     */
516 +    private volatile int spareWaiters;
517 +
518 +    private static final int SPARE_COUNT_SHIFT = 16;
519 +    private static final int SPARE_ID_MASK     = (1 << 16) - 1;
520 +
521 +    /**
522       * Lifecycle control. The low word contains the number of workers
523       * that are (probably) executing tasks. This value is atomically
524       * incremented before a worker gets a task to run, and decremented
525 <     * when worker has no tasks and cannot find any.  Bits 16-18
525 >     * when a worker has no tasks and cannot find any.  Bits 16-18
526       * contain runLevel value. When all are zero, the pool is
527       * running. Level transitions are monotonic (running -> shutdown
528       * -> terminating -> terminated) so each transition adds a bit.
529       * These are bundled together to ensure consistent read for
530       * termination checks (i.e., that runLevel is at least SHUTDOWN
531       * and active threads is zero).
532 +     *
533 +     * Notes: Most direct CASes are dependent on these bitfield
534 +     * positions.  Also, this field is non-private to enable direct
535 +     * performance-sensitive CASes in ForkJoinWorkerThread.
536       */
537 <    private volatile int runState;
537 >    volatile int runState;
538  
539      // Note: The order among run level values matters.
540      private static final int RUNLEVEL_SHIFT     = 16;
# Line 465 | 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;
468    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 474 | 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
478 <     * configuration parameters.
553 >     * running state.
554       */
555      private volatile int workerCounts;
556  
# Line 484 | 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  
487    /*
488     * Fields parallelism. maxPoolSize, locallyFifo,
489     * maintainsParallelism, and ueh are non-volatile, but external
490     * reads/writes use workerCount fences to ensure visability.
491     */
492
562      /**
563       * The target parallelism level.
564 +     * Accessed directly by ForkJoinWorkerThreads.
565       */
566 <    private int parallelism;
497 <
498 <    /**
499 <     * The maximum allowed pool size.
500 <     */
501 <    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
506 <     */
507 <    private boolean locallyFifo;
508 <
509 <    /**
510 <     * 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 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.
529 <     *
530 <     * @param delta the number to add
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 <     * Write fence for user modifications of pool parameters
541 <     * (parallelism. etc).  Note that it doesn't matter if CAS fails.
599 >     * Tries to increment running count part of workerCounts.
600       */
601 <    private void workerCountWriteFence() {
602 <        int wc;
603 <        UNSAFE.compareAndSwapInt(this, workerCountsOffset,
604 <                                 wc = workerCounts, wc);
601 >    final boolean tryIncrementRunningCount() {
602 >        int c;
603 >        return UNSAFE.compareAndSwapInt(this, workerCountsOffset,
604 >                                        c = workerCounts,
605 >                                        c + ONE_RUNNING);
606      }
607  
608      /**
609 <     * Read fence for external reads of pool parameters
551 <     * (parallelism. maxPoolSize, etc).
609 >     * Tries to decrement running count unless already zero.
610       */
611 <    private void workerCountReadFence() {
612 <        int ignore = workerCounts;
611 >    final boolean tryDecrementRunningCount() {
612 >        int wc = workerCounts;
613 >        if ((wc & RUNNING_COUNT_MASK) == 0)
614 >            return false;
615 >        return UNSAFE.compareAndSwapInt(this, workerCountsOffset,
616 >                                        wc, wc - ONE_RUNNING);
617      }
618  
619      /**
620 <     * Tries incrementing active count; fails on contention.
621 <     * Called by workers before executing tasks.
620 >     * Forces decrement of encoded workerCounts, awaiting nonzero if
621 >     * (rarely) necessary when other count updates lag.
622       *
623 <     * @return true on success
623 >     * @param dr -- either zero or ONE_RUNNING
624 >     * @param dt -- either zero or ONE_TOTAL
625       */
626 <    final boolean tryIncrementActiveCount() {
627 <        int c;
628 <        return UNSAFE.compareAndSwapInt(this, runStateOffset,
629 <                                        c = runState, c + ONE_ACTIVE);
626 >    private void decrementWorkerCounts(int dr, int dt) {
627 >        for (;;) {
628 >            int wc = workerCounts;
629 >            if ((wc & RUNNING_COUNT_MASK)  - dr < 0 ||
630 >                (wc >>> TOTAL_COUNT_SHIFT) - dt < 0) {
631 >                if ((runState & TERMINATED) != 0)
632 >                    return; // lagging termination on a backout
633 >                Thread.yield();
634 >            }
635 >            if (UNSAFE.compareAndSwapInt(this, workerCountsOffset,
636 >                                         wc, wc - (dr + dt)))
637 >                return;
638 >        }
639      }
640  
641      /**
# Line 573 | Line 645 | public class ForkJoinPool extends Abstra
645      final boolean tryDecrementActiveCount() {
646          int c;
647          return UNSAFE.compareAndSwapInt(this, runStateOffset,
648 <                                        c = runState, c - ONE_ACTIVE);
648 >                                        c = runState, c - 1);
649      }
650  
651      /**
# Line 602 | Line 674 | public class ForkJoinPool extends Abstra
674          lock.lock();
675          try {
676              ForkJoinWorkerThread[] ws = workers;
677 <            int len = ws.length;
678 <            if (k < 0 || k >= len || ws[k] != null) {
679 <                for (k = 0; k < len && ws[k] != null; ++k)
677 >            int n = ws.length;
678 >            if (k < 0 || k >= n || ws[k] != null) {
679 >                for (k = 0; k < n && ws[k] != null; ++k)
680                      ;
681 <                if (k == len)
682 <                    ws = Arrays.copyOf(ws, len << 1);
681 >                if (k == n)
682 >                    ws = workers = Arrays.copyOf(ws, n << 1);
683              }
684              ws[k] = w;
685 <            workers = ws; // volatile array write ensures slot visibility
685 >            int c = eventCount; // advance event count to ensure visibility
686 >            UNSAFE.compareAndSwapInt(this, eventCountOffset, c, c+1);
687          } finally {
688              lock.unlock();
689          }
# Line 618 | Line 691 | public class ForkJoinPool extends Abstra
691      }
692  
693      /**
694 <     * Nulls out record of worker in workers array
694 >     * Nulls out record of worker in workers array.
695       */
696      private void forgetWorker(ForkJoinWorkerThread w) {
697          int idx = w.poolIndex;
698 <        // Locking helps method recordWorker avoid unecessary expansion
698 >        // Locking helps method recordWorker avoid unnecessary expansion
699          final ReentrantLock lock = this.workerLock;
700          lock.lock();
701          try {
# Line 634 | Line 707 | public class ForkJoinPool extends Abstra
707          }
708      }
709  
637    // adding and removing workers
638
710      /**
711 <     * Tries to create and add new worker. Assumes that worker counts
712 <     * are already updated to accommodate the worker, so adjusts on
713 <     * failure.
711 >     * Final callback from terminating worker.  Removes record of
712 >     * worker from array, and adjusts counts. If pool is shutting
713 >     * down, tries to complete termination.
714       *
715 <     * @return new worker or null if creation failed
715 >     * @param w the worker
716       */
717 <    private ForkJoinWorkerThread addWorker() {
718 <        ForkJoinWorkerThread w = null;
719 <        try {
720 <            w = factory.newThread(this);
721 <        } finally { // Adjust on either null or exceptional factory return
722 <            if (w == null) {
723 <                onWorkerCreationFailure();
724 <                return null;
717 >    final void workerTerminated(ForkJoinWorkerThread w) {
718 >        forgetWorker(w);
719 >        decrementWorkerCounts(w.isTrimmed() ? 0 : ONE_RUNNING, ONE_TOTAL);
720 >        while (w.stealCount != 0) // collect final count
721 >            tryAccumulateStealCount(w);
722 >        tryTerminate(false);
723 >    }
724 >
725 >    // Waiting for and signalling events
726 >
727 >    /**
728 >     * Releases workers blocked on a count not equal to current count.
729 >     * Normally called after precheck that eventWaiters isn't zero to
730 >     * avoid wasted array checks. Gives up upon a change in count or
731 >     * upon releasing four workers, letting others take over.
732 >     */
733 >    private void releaseEventWaiters() {
734 >        ForkJoinWorkerThread[] ws = workers;
735 >        int n = ws.length;
736 >        long h = eventWaiters;
737 >        int ec = eventCount;
738 >        int releases = 4;
739 >        ForkJoinWorkerThread w; int id;
740 >        while ((id = (((int)h) & WAITER_ID_MASK) - 1) >= 0 &&
741 >               (int)(h >>> EVENT_COUNT_SHIFT) != ec &&
742 >               id < n && (w = ws[id]) != null) {
743 >            if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
744 >                                          h,  w.nextWaiter)) {
745 >                LockSupport.unpark(w);
746 >                if (--releases == 0)
747 >                    break;
748              }
749 +            if (eventCount != ec)
750 +                break;
751 +            h = eventWaiters;
752          }
656        w.start(recordWorker(w), locallyFifo, ueh);
657        return w;
753      }
754  
755      /**
756 <     * Adjusts counts upon failure to create worker
756 >     * Tries to advance eventCount and releases waiters. Called only
757 >     * from workers.
758       */
759 <    private void onWorkerCreationFailure() {
760 <        int c;
761 <        do {} while (!UNSAFE.compareAndSwapInt(this, workerCountsOffset,
762 <                                               c = workerCounts,
763 <                                               c - (ONE_RUNNING|ONE_TOTAL)));
668 <        tryTerminate(false); // in case of failure during shutdown
759 >    final void signalWork() {
760 >        int c; // try to increment event count -- CAS failure OK
761 >        UNSAFE.compareAndSwapInt(this, eventCountOffset, c = eventCount, c+1);
762 >        if (eventWaiters != 0L)
763 >            releaseEventWaiters();
764      }
765  
766      /**
767 <     * Create enough total workers to establish target parallelism,
768 <     * giving up if terminating or addWorker fails
767 >     * Adds the given worker to event queue and blocks until
768 >     * terminating or event count advances from the given value
769 >     *
770 >     * @param w the calling worker thread
771 >     * @param ec the count
772       */
773 <    private void ensureEnoughTotalWorkers() {
774 <        int wc;
775 <        while (runState < TERMINATING &&
776 <               ((wc = workerCounts) >>> TOTAL_COUNT_SHIFT) < parallelism) {
777 <            if ((UNSAFE.compareAndSwapInt(this, workerCountsOffset,
778 <                                          wc, wc + (ONE_RUNNING|ONE_TOTAL)) &&
779 <                 addWorker() == null))
773 >    private void eventSync(ForkJoinWorkerThread w, int ec) {
774 >        long nh = (((long)ec) << EVENT_COUNT_SHIFT) | ((long)(w.poolIndex+1));
775 >        long h;
776 >        while ((runState < SHUTDOWN || !tryTerminate(false)) &&
777 >               (((int)(h = eventWaiters) & WAITER_ID_MASK) == 0 ||
778 >                (int)(h >>> EVENT_COUNT_SHIFT) == ec) &&
779 >               eventCount == ec) {
780 >            if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
781 >                                          w.nextWaiter = h, nh)) {
782 >                awaitEvent(w, ec);
783                  break;
784 +            }
785          }
786      }
787  
788      /**
789 <     * Final callback from terminating worker.  Removes record of
790 <     * worker from array, and adjusts counts. If pool is shutting
791 <     * down, tries to complete terminatation, else possibly replaces
792 <     * the worker.
789 >     * Blocks the given worker (that has already been entered as an
790 >     * event waiter) until terminating or event count advances from
791 >     * the given value. The oldest (first) waiter uses a timed wait to
792 >     * occasionally one-by-one shrink the number of workers (to a
793 >     * minimum of one) if the pool has not been used for extended
794 >     * periods.
795       *
796 <     * @param w the worker
796 >     * @param w the calling worker thread
797 >     * @param ec the count
798       */
799 <    final void workerTerminated(ForkJoinWorkerThread w) {
800 <        if (w.active) { // force inactive
801 <            w.active = false;
802 <            do {} while (!tryDecrementActiveCount());
799 >    private void awaitEvent(ForkJoinWorkerThread w, int ec) {
800 >        while (eventCount == ec) {
801 >            if (tryAccumulateStealCount(w)) { // transfer while idle
802 >                boolean untimed = (w.nextWaiter != 0L ||
803 >                                   (workerCounts & RUNNING_COUNT_MASK) <= 1);
804 >                long startTime = untimed ? 0 : System.nanoTime();
805 >                Thread.interrupted();         // clear/ignore interrupt
806 >                if (w.isTerminating() || eventCount != ec)
807 >                    break;                    // recheck after clear
808 >                if (untimed)
809 >                    LockSupport.park(w);
810 >                else {
811 >                    LockSupport.parkNanos(w, SHRINK_RATE_NANOS);
812 >                    if (eventCount != ec || w.isTerminating())
813 >                        break;
814 >                    if (System.nanoTime() - startTime >= SHRINK_RATE_NANOS)
815 >                        tryShutdownUnusedWorker(ec);
816 >                }
817 >            }
818          }
699        forgetWorker(w);
700
701        // decrement total count, and if was running, running count
702        int unit = w.isTrimmed()? ONE_TOTAL : (ONE_RUNNING|ONE_TOTAL);
703        int wc;
704        do {} while (!UNSAFE.compareAndSwapInt(this, workerCountsOffset,
705                                               wc = workerCounts, wc - unit));
706
707        accumulateStealCount(w); // collect final count
708        if (!tryTerminate(false))
709            ensureEnoughTotalWorkers();
819      }
820  
821 <    // Waiting for and signalling events
821 >    // Maintaining parallelism
822  
823      /**
824 <     * Ensures eventCount on exit is different (mod 2^32) than on
716 <     * entry.  CAS failures are OK -- any change in count suffices.
824 >     * Pushes worker onto the spare stack.
825       */
826 <    private void advanceEventCount() {
827 <        int c;
828 <        UNSAFE.compareAndSwapInt(this, eventCountOffset, c = eventCount, c+1);
826 >    final void pushSpare(ForkJoinWorkerThread w) {
827 >        int ns = (++w.spareCount << SPARE_COUNT_SHIFT) | (w.poolIndex + 1);
828 >        do {} while (!UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
829 >                                               w.nextSpare = spareWaiters,ns));
830      }
831  
832      /**
833 <     * Releases workers blocked on a count not equal to current count.
834 <     */
835 <    final void releaseWaiters() {
836 <        long top;
837 <        int id;
838 <        while ((id = (int)((top = eventWaiters) & WAITER_INDEX_MASK)) > 0 &&
839 <               (int)(top >>> EVENT_COUNT_SHIFT) != eventCount) {
840 <            ForkJoinWorkerThread[] ws = workers;
841 <            ForkJoinWorkerThread w;
842 <            if (ws.length >= id && (w = ws[id - 1]) != null &&
843 <                UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
844 <                                          top, w.nextWaiter))
833 >     * Tries (once) to resume a spare if the number of running
834 >     * threads is less than target.
835 >     */
836 >    private void tryResumeSpare() {
837 >        int sw, id;
838 >        ForkJoinWorkerThread[] ws = workers;
839 >        int n = ws.length;
840 >        ForkJoinWorkerThread w;
841 >        if ((sw = spareWaiters) != 0 &&
842 >            (id = (sw & SPARE_ID_MASK) - 1) >= 0 &&
843 >            id < n && (w = ws[id]) != null &&
844 >            (runState >= TERMINATING ||
845 >             (workerCounts & RUNNING_COUNT_MASK) < parallelism) &&
846 >            spareWaiters == sw &&
847 >            UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
848 >                                     sw, w.nextSpare)) {
849 >            int c; // increment running count before resume
850 >            do {} while (!UNSAFE.compareAndSwapInt
851 >                         (this, workerCountsOffset,
852 >                          c = workerCounts, c + ONE_RUNNING));
853 >            if (w.tryUnsuspend())
854                  LockSupport.unpark(w);
855 +            else   // back out if w was shutdown
856 +                decrementWorkerCounts(ONE_RUNNING, 0);
857          }
858      }
859  
860      /**
861 <     * Advances eventCount and releases waiters until interference by
862 <     * other releasing threads is detected.
861 >     * Tries to increase the number of running workers if below target
862 >     * parallelism: If a spare exists tries to resume it via
863 >     * tryResumeSpare.  Otherwise, if not enough total workers or all
864 >     * existing workers are busy, adds a new worker. In all cases also
865 >     * helps wake up releasable workers waiting for work.
866       */
867 <    final void signalWork() {
868 <        int ec;
869 <        UNSAFE.compareAndSwapInt(this, eventCountOffset, ec=eventCount, ec+1);
870 <        outer:for (;;) {
871 <            long top = eventWaiters;
872 <            ec = eventCount;
873 <            for (;;) {
874 <                ForkJoinWorkerThread[] ws; ForkJoinWorkerThread w;
875 <                int id = (int)(top & WAITER_INDEX_MASK);
876 <                if (id <= 0 || (int)(top >>> EVENT_COUNT_SHIFT) == ec)
877 <                    return;
878 <                if ((ws = workers).length < id || (w = ws[id - 1]) == null ||
879 <                    !UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
880 <                                               top, top = w.nextWaiter))
881 <                    continue outer;      // possibly stale; reread
882 <                LockSupport.unpark(w);
883 <                if (top != eventWaiters) // let someone else take over
884 <                    return;
867 >    private void helpMaintainParallelism() {
868 >        int pc = parallelism;
869 >        int wc, rs, tc;
870 >        while (((wc = workerCounts) & RUNNING_COUNT_MASK) < pc &&
871 >               (rs = runState) < TERMINATING) {
872 >            if (spareWaiters != 0)
873 >                tryResumeSpare();
874 >            else if ((tc = wc >>> TOTAL_COUNT_SHIFT) >= MAX_WORKERS ||
875 >                     (tc >= pc && (rs & ACTIVE_COUNT_MASK) != tc))
876 >                break;   // enough total
877 >            else if (runState == rs && workerCounts == wc &&
878 >                     UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
879 >                                              wc + (ONE_RUNNING|ONE_TOTAL))) {
880 >                ForkJoinWorkerThread w = null;
881 >                Throwable fail = null;
882 >                try {
883 >                    w = factory.newThread(this);
884 >                } catch (Throwable ex) {
885 >                    fail = ex;
886 >                }
887 >                if (w == null) { // null or exceptional factory return
888 >                    decrementWorkerCounts(ONE_RUNNING, ONE_TOTAL);
889 >                    tryTerminate(false); // handle failure during shutdown
890 >                    // If originating from an external caller,
891 >                    // propagate exception, else ignore
892 >                    if (fail != null && runState < TERMINATING &&
893 >                        !(Thread.currentThread() instanceof
894 >                          ForkJoinWorkerThread))
895 >                        UNSAFE.throwException(fail);
896 >                    break;
897 >                }
898 >                w.start(recordWorker(w), ueh);
899 >                if ((workerCounts >>> TOTAL_COUNT_SHIFT) >= pc)
900 >                    break; // add at most one unless total below target
901              }
902          }
903 +        if (eventWaiters != 0L)
904 +            releaseEventWaiters();
905      }
906  
907      /**
908 <     * If worker is inactive, blocks until terminating or event count
909 <     * advances from last value held by worker; in any case helps
910 <     * release others.
908 >     * Callback from the oldest waiter in awaitEvent waking up after a
909 >     * period of non-use. If all workers are idle, tries (once) to
910 >     * shutdown an event waiter or a spare, if one exists. Note that
911 >     * we don't need CAS or locks here because the method is called
912 >     * only from one thread occasionally waking (and even misfires are
913 >     * OK). Note that until the shutdown worker fully terminates,
914 >     * workerCounts will overestimate total count, which is tolerable.
915       *
916 <     * @param w the calling worker thread
916 >     * @param ec the event count waited on by caller (to abort
917 >     * attempt if count has since changed).
918       */
919 <    private void eventSync(ForkJoinWorkerThread w) {
920 <        if (!w.active) {
921 <            int prev = w.lastEventCount;
922 <            long nextTop = (((long)prev << EVENT_COUNT_SHIFT) |
923 <                            ((long)(w.poolIndex + 1)));
924 <            long top;
925 <            while ((runState < SHUTDOWN || !tryTerminate(false)) &&
926 <                   (((int)(top = eventWaiters) & WAITER_INDEX_MASK) == 0 ||
927 <                    (int)(top >>> EVENT_COUNT_SHIFT) == prev) &&
928 <                   eventCount == prev) {
929 <                if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
930 <                                              w.nextWaiter = top, nextTop)) {
931 <                    accumulateStealCount(w); // transfer steals while idle
932 <                    Thread.interrupted();    // clear/ignore interrupt
933 <                    while (eventCount == prev)
934 <                        w.doPark();
935 <                    break;
936 <                }
919 >    private void tryShutdownUnusedWorker(int ec) {
920 >        if (runState == 0 && eventCount == ec) { // only trigger if all idle
921 >            ForkJoinWorkerThread[] ws = workers;
922 >            int n = ws.length;
923 >            ForkJoinWorkerThread w = null;
924 >            boolean shutdown = false;
925 >            int sw;
926 >            long h;
927 >            if ((sw = spareWaiters) != 0) { // prefer killing spares
928 >                int id = (sw & SPARE_ID_MASK) - 1;
929 >                if (id >= 0 && id < n && (w = ws[id]) != null &&
930 >                    UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
931 >                                             sw, w.nextSpare))
932 >                    shutdown = true;
933 >            }
934 >            else if ((h = eventWaiters) != 0L) {
935 >                long nh;
936 >                int id = (((int)h) & WAITER_ID_MASK) - 1;
937 >                if (id >= 0 && id < n && (w = ws[id]) != null &&
938 >                    (nh = w.nextWaiter) != 0L && // keep at least one worker
939 >                    UNSAFE.compareAndSwapLong(this, eventWaitersOffset, h, nh))
940 >                    shutdown = true;
941 >            }
942 >            if (w != null && shutdown) {
943 >                w.shutdown();
944 >                LockSupport.unpark(w);
945              }
792            w.lastEventCount = eventCount;
946          }
947 <        releaseWaiters();
947 >        releaseEventWaiters(); // in case of interference
948      }
949  
950      /**
951       * Callback from workers invoked upon each top-level action (i.e.,
952 <     * stealing a task or taking a submission and running
953 <     * it). Performs one or both of the following:
952 >     * stealing a task or taking a submission and running it).
953 >     * Performs one or more of the following:
954       *
955 <     * * If the worker cannot find work, updates its active status to
956 <     * inactive and updates activeCount unless there is contention, in
957 <     * which case it may try again (either in this or a subsequent
958 <     * call).  Additionally, awaits the next task event and/or helps
959 <     * wake up other releasable waiters.
960 <     *
961 <     * * If there are too many running threads, suspends this worker
962 <     * (first forcing inactivation if necessary).  If it is not
963 <     * resumed before a keepAlive elapses, the worker may be "trimmed"
964 <     * -- killed while suspended within suspendAsSpare. Otherwise,
965 <     * upon resume it rechecks to make sure that it is still needed.
955 >     * 1. If the worker is active and either did not run a task
956 >     *    or there are too many workers, try to set its active status
957 >     *    to inactive and update activeCount. On contention, we may
958 >     *    try again in this or a subsequent call.
959 >     *
960 >     * 2. If not enough total workers, help create some.
961 >     *
962 >     * 3. If there are too many running workers, suspend this worker
963 >     *    (first forcing inactive if necessary).  If it is not needed,
964 >     *    it may be shutdown while suspended (via
965 >     *    tryShutdownUnusedWorker).  Otherwise, upon resume it
966 >     *    rechecks running thread count and need for event sync.
967 >     *
968 >     * 4. If worker did not run a task, await the next task event via
969 >     *    eventSync if necessary (first forcing inactivation), upon
970 >     *    which the worker may be shutdown via
971 >     *    tryShutdownUnusedWorker.  Otherwise, help release any
972 >     *    existing event waiters that are now releasable,
973       *
974       * @param w the worker
975 <     * @param worked false if the worker scanned for work but didn't
816 <     * find any (in which case it may block waiting for work).
975 >     * @param ran true if worker ran a task since last call to this method
976       */
977 <    final void preStep(ForkJoinWorkerThread w, boolean worked) {
977 >    final void preStep(ForkJoinWorkerThread w, boolean ran) {
978 >        int wec = w.lastEventCount;
979          boolean active = w.active;
980 <        boolean inactivate = !worked & active;
981 <        for (;;) {
982 <            if (inactivate) {
983 <                int c = runState;
984 <                if (UNSAFE.compareAndSwapInt(this, runStateOffset,
985 <                                             c, c - ONE_ACTIVE))
986 <                    inactivate = active = w.active = false;
827 <            }
828 <            int wc = workerCounts;
829 <            if ((wc & RUNNING_COUNT_MASK) <= parallelism) {
830 <                if (!worked)
831 <                    eventSync(w);
832 <                return;
980 >        boolean inactivate = false;
981 >        int pc = parallelism;
982 >        while (w.runState == 0) {
983 >            int rs = runState;
984 >            if (rs >= TERMINATING) {           // propagate shutdown
985 >                w.shutdown();
986 >                break;
987              }
988 <            if (!(inactivate |= active) &&  // must inactivate to suspend
989 <                UNSAFE.compareAndSwapInt(this, workerCountsOffset,
990 <                                         wc, wc - ONE_RUNNING) &&
991 <                !w.suspendAsSpare())        // false if trimmed
992 <                return;
993 <        }
840 <    }
841 <
842 <    /**
843 <     * Adjusts counts and creates or resumes compensating threads for
844 <     * a worker about to block on task joinMe, returning early if
845 <     * joinMe becomes ready. First tries resuming an existing spare
846 <     * (which usually also avoids any count adjustment), but must then
847 <     * decrement running count to determine whether a new thread is
848 <     * needed. See above for fuller explanation.
849 <     */
850 <    final void preJoin(ForkJoinTask<?> joinMe) {
851 <        boolean dec = false;       // true when running count decremented
852 <        for (;;) {
853 <            releaseWaiters();      // help other threads progress
854 <
855 <            if (joinMe.status < 0) // surround spare search with done checks
856 <                return;
857 <            ForkJoinWorkerThread spare = null;
858 <            for (ForkJoinWorkerThread w : workers) {
859 <                if (w != null && w.isSuspended()) {
860 <                    spare = w;
861 <                    break;
988 >            if ((inactivate || (active && (rs & ACTIVE_COUNT_MASK) >= pc)) &&
989 >                UNSAFE.compareAndSwapInt(this, runStateOffset, rs, --rs)) {
990 >                inactivate = active = w.active = false;
991 >                if (rs == SHUTDOWN) {          // all inactive and shut down
992 >                    tryTerminate(false);
993 >                    continue;
994                  }
995              }
996 <            if (joinMe.status < 0)
997 <                return;
998 <
999 <            if (spare != null && spare.tryUnsuspend()) {
1000 <                if (dec || joinMe.requestSignal() < 0) {
1001 <                    int c;
1002 <                    do {} while (!UNSAFE.compareAndSwapInt(this,
1003 <                                                           workerCountsOffset,
1004 <                                                           c = workerCounts,
1005 <                                                           c + ONE_RUNNING));
1006 <                } // else no net count change
1007 <                LockSupport.unpark(spare);
1008 <                return;
1009 <            }
1010 <
1011 <            int wc = workerCounts; // decrement running count
1012 <            if (!dec && (wc & RUNNING_COUNT_MASK) != 0 &&
1013 <                (dec = UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1014 <                                                wc, wc -= ONE_RUNNING)) &&
1015 <                joinMe.requestSignal() < 0) { // cannot block
884 <                int c;                        // back out
885 <                do {} while (!UNSAFE.compareAndSwapInt(this,
886 <                                                       workerCountsOffset,
887 <                                                       c = workerCounts,
888 <                                                       c + ONE_RUNNING));
889 <                return;
890 <            }
891 <
892 <            if (dec) {
893 <                int tc = wc >>> TOTAL_COUNT_SHIFT;
894 <                int pc = parallelism;
895 <                int dc = pc - (wc & RUNNING_COUNT_MASK); // deficit count
896 <                if ((dc < pc && (dc <= 0 || (dc * dc < (tc - pc) * pc) ||
897 <                                 !maintainsParallelism)) ||
898 <                    tc >= maxPoolSize) // cannot add
899 <                    return;
900 <                if (spare == null &&
901 <                    UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
902 <                                             wc + (ONE_RUNNING|ONE_TOTAL))) {
903 <                    addWorker();
904 <                    return;
996 >            int wc = workerCounts;             // try to suspend as spare
997 >            if ((wc & RUNNING_COUNT_MASK) > pc) {
998 >                if (!(inactivate |= active) && // must inactivate to suspend
999 >                    workerCounts == wc &&
1000 >                    UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1001 >                                             wc, wc - ONE_RUNNING))
1002 >                    w.suspendAsSpare();
1003 >            }
1004 >            else if ((wc >>> TOTAL_COUNT_SHIFT) < pc)
1005 >                helpMaintainParallelism();     // not enough workers
1006 >            else if (ran)
1007 >                break;
1008 >            else {
1009 >                long h = eventWaiters;
1010 >                int ec = eventCount;
1011 >                if (h != 0L && (int)(h >>> EVENT_COUNT_SHIFT) != ec)
1012 >                    releaseEventWaiters();     // release others before waiting
1013 >                else if (ec != wec) {
1014 >                    w.lastEventCount = ec;     // no need to wait
1015 >                    break;
1016                  }
1017 +                else if (!(inactivate |= active))
1018 +                    eventSync(w, wec);         // must inactivate before sync
1019              }
1020          }
1021      }
1022  
1023      /**
1024 <     * Same idea as preJoin but with too many differing details to
1025 <     * integrate: There are no task-based signal counts, and only one
1026 <     * way to do the actual blocking. So for simplicity it is directly
1027 <     * incorporated into this method.
1024 >     * Helps and/or blocks awaiting join of the given task.
1025 >     * See above for explanation.
1026 >     *
1027 >     * @param joinMe the task to join
1028 >     * @param worker the current worker thread
1029 >     * @param timed true if wait should time out
1030 >     * @param nanos timeout value if timed
1031       */
1032 <    final void doBlock(ManagedBlocker blocker, boolean maintainPar)
1033 <        throws InterruptedException {
1034 <        maintainPar &= maintainsParallelism; // override
1035 <        boolean dec = false;
1036 <        boolean done = false;
1037 <        for (;;) {
1038 <            releaseWaiters();
1039 <            if (done = blocker.isReleasable())
1032 >    final void awaitJoin(ForkJoinTask<?> joinMe, ForkJoinWorkerThread worker,
1033 >                         boolean timed, long nanos) {
1034 >        long startTime = timed ? System.nanoTime() : 0L;
1035 >        int retries = 2 + (parallelism >> 2); // #helpJoins before blocking
1036 >        boolean running = true;               // false when count decremented
1037 >        while (joinMe.status >= 0) {
1038 >            if (runState >= TERMINATING) {
1039 >                joinMe.cancelIgnoringExceptions();
1040                  break;
925            ForkJoinWorkerThread spare = null;
926            for (ForkJoinWorkerThread w : workers) {
927                if (w != null && w.isSuspended()) {
928                    spare = w;
929                    break;
930                }
1041              }
1042 <            if (done = blocker.isReleasable())
1043 <                break;
934 <            if (spare != null && spare.tryUnsuspend()) {
935 <                if (dec) {
936 <                    int c;
937 <                    do {} while (!UNSAFE.compareAndSwapInt(this,
938 <                                                           workerCountsOffset,
939 <                                                           c = workerCounts,
940 <                                                           c + ONE_RUNNING));
941 <                }
942 <                LockSupport.unpark(spare);
1042 >            running = worker.helpJoinTask(joinMe, running);
1043 >            if (joinMe.status < 0)
1044                  break;
1045 +            if (retries > 0) {
1046 +                --retries;
1047 +                continue;
1048              }
1049              int wc = workerCounts;
1050 <            if (!dec && (wc & RUNNING_COUNT_MASK) != 0)
1051 <                dec = UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1052 <                                               wc, wc -= ONE_RUNNING);
1053 <            if (dec) {
1054 <                int tc = wc >>> TOTAL_COUNT_SHIFT;
1055 <                int pc = parallelism;
1056 <                int dc = pc - (wc & RUNNING_COUNT_MASK);
1057 <                if ((dc < pc && (dc <= 0 || (dc * dc < (tc - pc) * pc) ||
1058 <                                 !maintainPar)) ||
1059 <                    tc >= maxPoolSize)
1060 <                    break;
1061 <                if (spare == null &&
1062 <                    UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
1063 <                                             wc + (ONE_RUNNING|ONE_TOTAL))){
1064 <                    addWorker();
1065 <                    break;
1050 >            if ((wc & RUNNING_COUNT_MASK) != 0) {
1051 >                if (running) {
1052 >                    if (!UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1053 >                                                  wc, wc - ONE_RUNNING))
1054 >                        continue;
1055 >                    running = false;
1056 >                }
1057 >                long h = eventWaiters;
1058 >                if (h != 0L && (int)(h >>> EVENT_COUNT_SHIFT) != eventCount)
1059 >                    releaseEventWaiters();
1060 >                if ((workerCounts & RUNNING_COUNT_MASK) != 0) {
1061 >                    long ms; int ns;
1062 >                    if (!timed) {
1063 >                        ms = JOIN_TIMEOUT_MILLIS;
1064 >                        ns = 0;
1065 >                    }
1066 >                    else { // at most JOIN_TIMEOUT_MILLIS per wait
1067 >                        long nt = nanos - (System.nanoTime() - startTime);
1068 >                        if (nt <= 0L)
1069 >                            break;
1070 >                        ms = nt / 1000000;
1071 >                        if (ms > JOIN_TIMEOUT_MILLIS) {
1072 >                            ms = JOIN_TIMEOUT_MILLIS;
1073 >                            ns = 0;
1074 >                        }
1075 >                        else
1076 >                            ns = (int) (nt % 1000000);
1077 >                    }
1078 >                    joinMe.internalAwaitDone(ms, ns);
1079                  }
1080 +                if (joinMe.status < 0)
1081 +                    break;
1082              }
1083 +            helpMaintainParallelism();
1084          }
1085 <
1086 <        try {
1087 <            if (!done)
1088 <                do {} while (!blocker.isReleasable() && !blocker.block());
1089 <        } finally {
970 <            if (dec) {
971 <                int c;
972 <                do {} while (!UNSAFE.compareAndSwapInt(this,
973 <                                                       workerCountsOffset,
974 <                                                       c = workerCounts,
975 <                                                       c + ONE_RUNNING));
976 <            }
1085 >        if (!running) {
1086 >            int c;
1087 >            do {} while (!UNSAFE.compareAndSwapInt
1088 >                         (this, workerCountsOffset,
1089 >                          c = workerCounts, c + ONE_RUNNING));
1090          }
1091      }
1092  
1093      /**
1094 <     * Unless there are not enough other running threads, adjusts
982 <     * counts for a a worker in performing helpJoin that cannot find
983 <     * any work, so that this worker can now block.
984 <     *
985 <     * @return true if worker may block
1094 >     * Same idea as awaitJoin, but no helping, retries, or timeouts.
1095       */
1096 <    final boolean preBlockHelpingJoin(ForkJoinTask<?> joinMe) {
1097 <        while (joinMe.status >= 0) {
1098 <            releaseWaiters(); // help other threads progress
1099 <
1100 <            // if a spare exists, resume it to maintain parallelism level
1101 <            if ((workerCounts & RUNNING_COUNT_MASK) <= parallelism) {
1102 <                ForkJoinWorkerThread spare = null;
1103 <                for (ForkJoinWorkerThread w : workers) {
1104 <                    if (w != null && w.isSuspended()) {
1105 <                        spare = w;
1106 <                        break;
1107 <                    }
1108 <                }
1109 <                if (joinMe.status < 0)
1110 <                    break;
1111 <                if (spare != null) {
1112 <                    if (spare.tryUnsuspend()) {
1113 <                        boolean canBlock = true;
1114 <                        if (joinMe.requestSignal() < 0) {
1006 <                            canBlock = false; // already done
1007 <                            int c;
1008 <                            do {} while (!UNSAFE.compareAndSwapInt
1009 <                                         (this, workerCountsOffset,
1010 <                                          c = workerCounts, c + ONE_RUNNING));
1011 <                        }
1012 <                        LockSupport.unpark(spare);
1013 <                        return canBlock;
1096 >    final void awaitBlocker(ManagedBlocker blocker)
1097 >        throws InterruptedException {
1098 >        while (!blocker.isReleasable()) {
1099 >            int wc = workerCounts;
1100 >            if ((wc & RUNNING_COUNT_MASK) == 0)
1101 >                helpMaintainParallelism();
1102 >            else if (UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1103 >                                              wc, wc - ONE_RUNNING)) {
1104 >                try {
1105 >                    while (!blocker.isReleasable()) {
1106 >                        long h = eventWaiters;
1107 >                        if (h != 0L &&
1108 >                            (int)(h >>> EVENT_COUNT_SHIFT) != eventCount)
1109 >                            releaseEventWaiters();
1110 >                        else if ((workerCounts & RUNNING_COUNT_MASK) == 0 &&
1111 >                                 runState < TERMINATING)
1112 >                            helpMaintainParallelism();
1113 >                        else if (blocker.block())
1114 >                            break;
1115                      }
1116 <                    continue; // recheck -- another spare may exist
1116 >                } finally {
1117 >                    int c;
1118 >                    do {} while (!UNSAFE.compareAndSwapInt
1119 >                                 (this, workerCountsOffset,
1120 >                                  c = workerCounts, c + ONE_RUNNING));
1121                  }
1017            }
1018
1019            int wc = workerCounts; // reread to shorten CAS window
1020            int rc = wc & RUNNING_COUNT_MASK;
1021            if (rc <= 2) // keep this and at most one other thread alive
1022                break;
1023
1024            if (UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1025                                         wc, wc - ONE_RUNNING)) {
1026                if (joinMe.requestSignal() >= 0)
1027                    return true;
1028                int c;                        // back out
1029                do {} while (!UNSAFE.compareAndSwapInt
1030                             (this, workerCountsOffset,
1031                              c = workerCounts, c + ONE_RUNNING));
1122                  break;
1123              }
1124          }
1035        return false;
1125      }
1126  
1127      /**
# Line 1056 | Line 1145 | public class ForkJoinPool extends Abstra
1145          // Finish now if all threads terminated; else in some subsequent call
1146          if ((workerCounts >>> TOTAL_COUNT_SHIFT) == 0) {
1147              advanceRunLevel(TERMINATED);
1148 <            terminationLatch.countDown();
1148 >            termination.forceTermination();
1149          }
1150          return true;
1151      }
1152  
1153      /**
1154       * Actions on transition to TERMINATING
1155 +     *
1156 +     * Runs up to four passes through workers: (0) shutting down each
1157 +     * (without waking up if parked) to quickly spread notifications
1158 +     * without unnecessary bouncing around event queues etc (1) wake
1159 +     * up and help cancel tasks (2) interrupt (3) mop up races with
1160 +     * interrupted workers
1161       */
1162      private void startTerminating() {
1163 <        // Clear out and cancel submissions, ignoring exceptions
1163 >        cancelSubmissions();
1164 >        for (int passes = 0; passes < 4 && workerCounts != 0; ++passes) {
1165 >            int c; // advance event count
1166 >            UNSAFE.compareAndSwapInt(this, eventCountOffset,
1167 >                                     c = eventCount, c+1);
1168 >            eventWaiters = 0L; // clobber lists
1169 >            spareWaiters = 0;
1170 >            for (ForkJoinWorkerThread w : workers) {
1171 >                if (w != null) {
1172 >                    w.shutdown();
1173 >                    if (passes > 0 && !w.isTerminated()) {
1174 >                        w.cancelTasks();
1175 >                        LockSupport.unpark(w);
1176 >                        if (passes > 1 && !w.isInterrupted()) {
1177 >                            try {
1178 >                                w.interrupt();
1179 >                            } catch (SecurityException ignore) {
1180 >                            }
1181 >                        }
1182 >                    }
1183 >                }
1184 >            }
1185 >        }
1186 >    }
1187 >
1188 >    /**
1189 >     * Clears out and cancels submissions, ignoring exceptions.
1190 >     */
1191 >    private void cancelSubmissions() {
1192          ForkJoinTask<?> task;
1193          while ((task = submissionQueue.poll()) != null) {
1194              try {
# Line 1073 | Line 1196 | public class ForkJoinPool extends Abstra
1196              } catch (Throwable ignore) {
1197              }
1198          }
1076        // Propagate run level
1077        for (ForkJoinWorkerThread w : workers) {
1078            if (w != null)
1079                w.shutdown();    // also resumes suspended workers
1080        }
1081        // Ensure no straggling local tasks
1082        for (ForkJoinWorkerThread w : workers) {
1083            if (w != null)
1084                w.cancelTasks();
1085        }
1086        // Wake up idle workers
1087        advanceEventCount();
1088        releaseWaiters();
1089        // Unstick pending joins
1090        for (ForkJoinWorkerThread w : workers) {
1091            if (w != null && !w.isTerminated()) {
1092                try {
1093                    w.interrupt();
1094                } catch (SecurityException ignore) {
1095                }
1096            }
1097        }
1199      }
1200  
1201      // misc support for ForkJoinWorkerThread
1202  
1203      /**
1204 <     * Returns pool number
1204 >     * Returns pool number.
1205       */
1206      final int getPoolNumber() {
1207          return poolNumber;
1208      }
1209  
1210      /**
1211 <     * Accumulates steal count from a worker, clearing
1212 <     * the worker's value
1211 >     * Tries to accumulate steal count from a worker, clearing
1212 >     * the worker's value if successful.
1213 >     *
1214 >     * @return true if worker steal count now zero
1215       */
1216 <    final void accumulateStealCount(ForkJoinWorkerThread w) {
1216 >    final boolean tryAccumulateStealCount(ForkJoinWorkerThread w) {
1217          int sc = w.stealCount;
1218 <        if (sc != 0) {
1219 <            long c;
1220 <            w.stealCount = 0;
1221 <            do {} while (!UNSAFE.compareAndSwapLong(this, stealCountOffset,
1222 <                                                    c = stealCount, c + sc));
1218 >        long c = stealCount;
1219 >        // CAS even if zero, for fence effects
1220 >        if (UNSAFE.compareAndSwapLong(this, stealCountOffset, c, c + sc)) {
1221 >            if (sc != 0)
1222 >                w.stealCount = 0;
1223 >            return true;
1224          }
1225 +        return sc == 0;
1226      }
1227  
1228      /**
# Line 1125 | Line 1230 | public class ForkJoinPool extends Abstra
1230       * active thread.
1231       */
1232      final int idlePerActive() {
1233 <        int ac = runState;    // no mask -- artifically boosts during shutdown
1234 <        int pc = parallelism; // use targeted parallelism, not rc
1233 >        int pc = parallelism; // use parallelism, not rc
1234 >        int ac = runState;    // no mask -- artificially boosts during shutdown
1235          // Use exact results for small values, saturate past 4
1236 <        return pc <= ac? 0 : pc >>> 1 <= ac? 1 : pc >>> 2 <= ac? 3 : pc >>> 3;
1236 >        return ((pc <= ac) ? 0 :
1237 >                (pc >>> 1 <= ac) ? 1 :
1238 >                (pc >>> 2 <= ac) ? 3 :
1239 >                pc >>> 3);
1240      }
1241  
1242      // Public and protected methods
# Line 1137 | Line 1245 | public class ForkJoinPool extends Abstra
1245  
1246      /**
1247       * Creates a {@code ForkJoinPool} with parallelism equal to {@link
1248 <     * java.lang.Runtime#availableProcessors}, and using the {@linkplain
1249 <     * #defaultForkJoinWorkerThreadFactory default thread factory}.
1248 >     * java.lang.Runtime#availableProcessors}, using the {@linkplain
1249 >     * #defaultForkJoinWorkerThreadFactory default thread factory},
1250 >     * no UncaughtExceptionHandler, and non-async LIFO processing mode.
1251       *
1252       * @throws SecurityException if a security manager exists and
1253       *         the caller is not permitted to modify threads
# Line 1147 | Line 1256 | public class ForkJoinPool extends Abstra
1256       */
1257      public ForkJoinPool() {
1258          this(Runtime.getRuntime().availableProcessors(),
1259 <             defaultForkJoinWorkerThreadFactory);
1259 >             defaultForkJoinWorkerThreadFactory, null, false);
1260      }
1261  
1262      /**
1263       * Creates a {@code ForkJoinPool} with the indicated parallelism
1264 <     * level and using the {@linkplain
1265 <     * #defaultForkJoinWorkerThreadFactory default thread factory}.
1264 >     * level, the {@linkplain
1265 >     * #defaultForkJoinWorkerThreadFactory default thread factory},
1266 >     * no UncaughtExceptionHandler, and non-async LIFO processing mode.
1267       *
1268       * @param parallelism the parallelism level
1269       * @throws IllegalArgumentException if parallelism less than or
# Line 1164 | Line 1274 | public class ForkJoinPool extends Abstra
1274       *         java.lang.RuntimePermission}{@code ("modifyThread")}
1275       */
1276      public ForkJoinPool(int parallelism) {
1277 <        this(parallelism, defaultForkJoinWorkerThreadFactory);
1168 <    }
1169 <
1170 <    /**
1171 <     * Creates a {@code ForkJoinPool} with parallelism equal to {@link
1172 <     * java.lang.Runtime#availableProcessors}, and using the given
1173 <     * thread factory.
1174 <     *
1175 <     * @param factory the factory for creating new threads
1176 <     * @throws NullPointerException if the factory is null
1177 <     * @throws SecurityException if a security manager exists and
1178 <     *         the caller is not permitted to modify threads
1179 <     *         because it does not hold {@link
1180 <     *         java.lang.RuntimePermission}{@code ("modifyThread")}
1181 <     */
1182 <    public ForkJoinPool(ForkJoinWorkerThreadFactory factory) {
1183 <        this(Runtime.getRuntime().availableProcessors(), factory);
1277 >        this(parallelism, defaultForkJoinWorkerThreadFactory, null, false);
1278      }
1279  
1280      /**
1281 <     * Creates a {@code ForkJoinPool} with the given parallelism and
1188 <     * thread factory.
1281 >     * Creates a {@code ForkJoinPool} with the given parameters.
1282       *
1283 <     * @param parallelism the parallelism level
1284 <     * @param factory the factory for creating new threads
1283 >     * @param parallelism the parallelism level. For default value,
1284 >     * use {@link java.lang.Runtime#availableProcessors}.
1285 >     * @param factory the factory for creating new threads. For default value,
1286 >     * use {@link #defaultForkJoinWorkerThreadFactory}.
1287 >     * @param handler the handler for internal worker threads that
1288 >     * terminate due to unrecoverable errors encountered while executing
1289 >     * tasks. For default value, use {@code null}.
1290 >     * @param asyncMode if true,
1291 >     * establishes local first-in-first-out scheduling mode for forked
1292 >     * tasks that are never joined. This mode may be more appropriate
1293 >     * than default locally stack-based mode in applications in which
1294 >     * worker threads only process event-style asynchronous tasks.
1295 >     * For default value, use {@code false}.
1296       * @throws IllegalArgumentException if parallelism less than or
1297       *         equal to zero, or greater than implementation limit
1298       * @throws NullPointerException if the factory is null
# Line 1197 | Line 1301 | public class ForkJoinPool extends Abstra
1301       *         because it does not hold {@link
1302       *         java.lang.RuntimePermission}{@code ("modifyThread")}
1303       */
1304 <    public ForkJoinPool(int parallelism, ForkJoinWorkerThreadFactory factory) {
1304 >    public ForkJoinPool(int parallelism,
1305 >                        ForkJoinWorkerThreadFactory factory,
1306 >                        Thread.UncaughtExceptionHandler handler,
1307 >                        boolean asyncMode) {
1308          checkPermission();
1309          if (factory == null)
1310              throw new NullPointerException();
1311 <        if (parallelism <= 0 || parallelism > MAX_THREADS)
1311 >        if (parallelism <= 0 || parallelism > MAX_WORKERS)
1312              throw new IllegalArgumentException();
1206        this.poolNumber = poolNumberGenerator.incrementAndGet();
1207        int arraySize = initialArraySizeFor(parallelism);
1313          this.parallelism = parallelism;
1314          this.factory = factory;
1315 <        this.maxPoolSize = MAX_THREADS;
1316 <        this.maintainsParallelism = true;
1315 >        this.ueh = handler;
1316 >        this.locallyFifo = asyncMode;
1317 >        int arraySize = initialArraySizeFor(parallelism);
1318          this.workers = new ForkJoinWorkerThread[arraySize];
1319          this.submissionQueue = new LinkedTransferQueue<ForkJoinTask<?>>();
1320          this.workerLock = new ReentrantLock();
1321 <        this.terminationLatch = new CountDownLatch(1);
1322 <        // Start first worker; remaining workers added upon first submission
1217 <        workerCounts = ONE_RUNNING | ONE_TOTAL;
1218 <        addWorker();
1321 >        this.termination = new Phaser(1);
1322 >        this.poolNumber = poolNumberGenerator.incrementAndGet();
1323      }
1324  
1325      /**
# Line 1223 | Line 1327 | public class ForkJoinPool extends Abstra
1327       * @param pc the initial parallelism level
1328       */
1329      private static int initialArraySizeFor(int pc) {
1330 <        // See Hackers Delight, sec 3.2. We know MAX_THREADS < (1 >>> 16)
1331 <        int size = pc < MAX_THREADS ? pc + 1 : MAX_THREADS;
1330 >        // If possible, initially allocate enough space for one spare
1331 >        int size = pc < MAX_WORKERS ? pc + 1 : MAX_WORKERS;
1332 >        // See Hackers Delight, sec 3.2. We know MAX_WORKERS < (1 >>> 16)
1333          size |= size >>> 1;
1334          size |= size >>> 2;
1335          size |= size >>> 4;
# Line 1235 | Line 1340 | public class ForkJoinPool extends Abstra
1340      // Execution methods
1341  
1342      /**
1343 <     * Common code for execute, invoke and submit
1343 >     * Submits task and creates, starts, or resumes some workers if necessary
1344       */
1345      private <T> void doSubmit(ForkJoinTask<T> task) {
1241        if (task == null)
1242            throw new NullPointerException();
1243        if (runState >= SHUTDOWN)
1244            throw new RejectedExecutionException();
1346          submissionQueue.offer(task);
1347 <        advanceEventCount();
1348 <        releaseWaiters();
1349 <        if ((workerCounts >>> TOTAL_COUNT_SHIFT) < parallelism)
1249 <            ensureEnoughTotalWorkers();
1347 >        int c; // try to increment event count -- CAS failure OK
1348 >        UNSAFE.compareAndSwapInt(this, eventCountOffset, c = eventCount, c+1);
1349 >        helpMaintainParallelism();
1350      }
1351  
1352      /**
# Line 1259 | Line 1359 | public class ForkJoinPool extends Abstra
1359       *         scheduled for execution
1360       */
1361      public <T> T invoke(ForkJoinTask<T> task) {
1362 <        doSubmit(task);
1363 <        return task.join();
1362 >        if (task == null)
1363 >            throw new NullPointerException();
1364 >        if (runState >= SHUTDOWN)
1365 >            throw new RejectedExecutionException();
1366 >        Thread t = Thread.currentThread();
1367 >        if ((t instanceof ForkJoinWorkerThread) &&
1368 >            ((ForkJoinWorkerThread)t).pool == this)
1369 >            return task.invoke();  // bypass submit if in same pool
1370 >        else {
1371 >            doSubmit(task);
1372 >            return task.join();
1373 >        }
1374 >    }
1375 >
1376 >    /**
1377 >     * Unless terminating, forks task if within an ongoing FJ
1378 >     * computation in the current pool, else submits as external task.
1379 >     */
1380 >    private <T> void forkOrSubmit(ForkJoinTask<T> task) {
1381 >        if (runState >= SHUTDOWN)
1382 >            throw new RejectedExecutionException();
1383 >        Thread t = Thread.currentThread();
1384 >        if ((t instanceof ForkJoinWorkerThread) &&
1385 >            ((ForkJoinWorkerThread)t).pool == this)
1386 >            task.fork();
1387 >        else
1388 >            doSubmit(task);
1389      }
1390  
1391      /**
# Line 1272 | Line 1397 | public class ForkJoinPool extends Abstra
1397       *         scheduled for execution
1398       */
1399      public void execute(ForkJoinTask<?> task) {
1400 <        doSubmit(task);
1400 >        if (task == null)
1401 >            throw new NullPointerException();
1402 >        forkOrSubmit(task);
1403      }
1404  
1405      // AbstractExecutorService methods
# Line 1283 | Line 1410 | public class ForkJoinPool extends Abstra
1410       *         scheduled for execution
1411       */
1412      public void execute(Runnable task) {
1413 +        if (task == null)
1414 +            throw new NullPointerException();
1415          ForkJoinTask<?> job;
1416          if (task instanceof ForkJoinTask<?>) // avoid re-wrap
1417              job = (ForkJoinTask<?>) task;
1418          else
1419              job = ForkJoinTask.adapt(task, null);
1420 <        doSubmit(job);
1420 >        forkOrSubmit(job);
1421 >    }
1422 >
1423 >    /**
1424 >     * Submits a ForkJoinTask for execution.
1425 >     *
1426 >     * @param task the task to submit
1427 >     * @return the task
1428 >     * @throws NullPointerException if the task is null
1429 >     * @throws RejectedExecutionException if the task cannot be
1430 >     *         scheduled for execution
1431 >     */
1432 >    public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) {
1433 >        if (task == null)
1434 >            throw new NullPointerException();
1435 >        forkOrSubmit(task);
1436 >        return task;
1437      }
1438  
1439      /**
# Line 1297 | Line 1442 | public class ForkJoinPool extends Abstra
1442       *         scheduled for execution
1443       */
1444      public <T> ForkJoinTask<T> submit(Callable<T> task) {
1445 +        if (task == null)
1446 +            throw new NullPointerException();
1447          ForkJoinTask<T> job = ForkJoinTask.adapt(task);
1448 <        doSubmit(job);
1448 >        forkOrSubmit(job);
1449          return job;
1450      }
1451  
# Line 1308 | Line 1455 | public class ForkJoinPool extends Abstra
1455       *         scheduled for execution
1456       */
1457      public <T> ForkJoinTask<T> submit(Runnable task, T result) {
1458 +        if (task == null)
1459 +            throw new NullPointerException();
1460          ForkJoinTask<T> job = ForkJoinTask.adapt(task, result);
1461 <        doSubmit(job);
1461 >        forkOrSubmit(job);
1462          return job;
1463      }
1464  
# Line 1319 | Line 1468 | public class ForkJoinPool extends Abstra
1468       *         scheduled for execution
1469       */
1470      public ForkJoinTask<?> submit(Runnable task) {
1471 +        if (task == null)
1472 +            throw new NullPointerException();
1473          ForkJoinTask<?> job;
1474          if (task instanceof ForkJoinTask<?>) // avoid re-wrap
1475              job = (ForkJoinTask<?>) task;
1476          else
1477              job = ForkJoinTask.adapt(task, null);
1478 <        doSubmit(job);
1478 >        forkOrSubmit(job);
1479          return job;
1480      }
1481  
1482      /**
1332     * Submits a ForkJoinTask for execution.
1333     *
1334     * @param task the task to submit
1335     * @return the task
1336     * @throws NullPointerException if the task is null
1337     * @throws RejectedExecutionException if the task cannot be
1338     *         scheduled for execution
1339     */
1340    public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) {
1341        doSubmit(task);
1342        return task;
1343    }
1344
1345    /**
1483       * @throws NullPointerException       {@inheritDoc}
1484       * @throws RejectedExecutionException {@inheritDoc}
1485       */
# Line 1384 | Line 1521 | public class ForkJoinPool extends Abstra
1521       * @return the handler, or {@code null} if none
1522       */
1523      public Thread.UncaughtExceptionHandler getUncaughtExceptionHandler() {
1387        workerCountReadFence();
1524          return ueh;
1525      }
1526  
1527      /**
1392     * Sets the handler for internal worker threads that terminate due
1393     * to unrecoverable errors encountered while executing tasks.
1394     * Unless set, the current default or ThreadGroup handler is used
1395     * as handler.
1396     *
1397     * @param h the new handler
1398     * @return the old handler, or {@code null} if none
1399     * @throws SecurityException if a security manager exists and
1400     *         the caller is not permitted to modify threads
1401     *         because it does not hold {@link
1402     *         java.lang.RuntimePermission}{@code ("modifyThread")}
1403     */
1404    public Thread.UncaughtExceptionHandler
1405        setUncaughtExceptionHandler(Thread.UncaughtExceptionHandler h) {
1406        checkPermission();
1407        workerCountReadFence();
1408        Thread.UncaughtExceptionHandler old = ueh;
1409        if (h != old) {
1410            ueh = h;
1411            workerCountWriteFence();
1412            for (ForkJoinWorkerThread w : workers) {
1413                if (w != null)
1414                    w.setUncaughtExceptionHandler(h);
1415            }
1416        }
1417        return old;
1418    }
1419
1420    /**
1421     * Sets the target parallelism level of this pool.
1422     *
1423     * @param parallelism the target parallelism
1424     * @throws IllegalArgumentException if parallelism less than or
1425     * equal to zero or greater than maximum size bounds
1426     * @throws SecurityException if a security manager exists and
1427     *         the caller is not permitted to modify threads
1428     *         because it does not hold {@link
1429     *         java.lang.RuntimePermission}{@code ("modifyThread")}
1430     */
1431    public void setParallelism(int parallelism) {
1432        checkPermission();
1433        if (parallelism <= 0 || parallelism > maxPoolSize)
1434            throw new IllegalArgumentException();
1435        workerCountReadFence();
1436        int pc = this.parallelism;
1437        if (pc != parallelism) {
1438            this.parallelism = parallelism;
1439            workerCountWriteFence();
1440            // Release spares. If too many, some will die after re-suspend
1441            for (ForkJoinWorkerThread w : workers) {
1442                if (w != null && w.tryUnsuspend()) {
1443                    updateRunningCount(1);
1444                    LockSupport.unpark(w);
1445                }
1446            }
1447            ensureEnoughTotalWorkers();
1448            advanceEventCount();
1449            releaseWaiters(); // force config recheck by existing workers
1450        }
1451    }
1452
1453    /**
1528       * Returns the targeted parallelism level of this pool.
1529       *
1530       * @return the targeted parallelism level of this pool
1531       */
1532      public int getParallelism() {
1459        //        workerCountReadFence(); // inlined below
1460        int ignore = workerCounts;
1533          return parallelism;
1534      }
1535  
1536      /**
1537       * Returns the number of worker threads that have started but not
1538 <     * yet terminated.  This result returned by this method may differ
1538 >     * yet terminated.  The result returned by this method may differ
1539       * from {@link #getParallelism} when threads are created to
1540       * maintain parallelism when others are cooperatively blocked.
1541       *
# Line 1474 | Line 1546 | public class ForkJoinPool extends Abstra
1546      }
1547  
1548      /**
1477     * Returns the maximum number of threads allowed to exist in the
1478     * pool. Unless set using {@link #setMaximumPoolSize}, the
1479     * maximum is an implementation-defined value designed only to
1480     * prevent runaway growth.
1481     *
1482     * @return the maximum
1483     */
1484    public int getMaximumPoolSize() {
1485        workerCountReadFence();
1486        return maxPoolSize;
1487    }
1488
1489    /**
1490     * Sets the maximum number of threads allowed to exist in the
1491     * pool. The given value should normally be greater than or equal
1492     * to the {@link #getParallelism parallelism} level. Setting this
1493     * value has no effect on current pool size. It controls
1494     * construction of new threads. The use of this method may cause
1495     * tasks that intrinsically require extra threads for dependent
1496     * computations to indefinitely stall. If you are instead trying
1497     * to minimize internal thread creation, consider setting {@link
1498     * #setMaintainsParallelism} as false.
1499     *
1500     * @throws IllegalArgumentException if negative or greater than
1501     * internal implementation limit
1502     */
1503    public void setMaximumPoolSize(int newMax) {
1504        if (newMax < 0 || newMax > MAX_THREADS)
1505            throw new IllegalArgumentException();
1506        maxPoolSize = newMax;
1507        workerCountWriteFence();
1508    }
1509
1510    /**
1511     * Returns {@code true} if this pool dynamically maintains its
1512     * target parallelism level. If false, new threads are added only
1513     * to avoid possible starvation.  This setting is by default true.
1514     *
1515     * @return {@code true} if maintains parallelism
1516     */
1517    public boolean getMaintainsParallelism() {
1518        workerCountReadFence();
1519        return maintainsParallelism;
1520    }
1521
1522    /**
1523     * Sets whether this pool dynamically maintains its target
1524     * parallelism level. If false, new threads are added only to
1525     * avoid possible starvation.
1526     *
1527     * @param enable {@code true} to maintain parallelism
1528     */
1529    public void setMaintainsParallelism(boolean enable) {
1530        maintainsParallelism = enable;
1531        workerCountWriteFence();
1532    }
1533
1534    /**
1535     * Establishes local first-in-first-out scheduling mode for forked
1536     * tasks that are never joined. This mode may be more appropriate
1537     * than default locally stack-based mode in applications in which
1538     * worker threads only process asynchronous tasks.  This method is
1539     * designed to be invoked only when the pool is quiescent, and
1540     * typically only before any tasks are submitted. The effects of
1541     * invocations at other times may be unpredictable.
1542     *
1543     * @param async if {@code true}, use locally FIFO scheduling
1544     * @return the previous mode
1545     * @see #getAsyncMode
1546     */
1547    public boolean setAsyncMode(boolean async) {
1548        workerCountReadFence();
1549        boolean oldMode = locallyFifo;
1550        if (oldMode != async) {
1551            locallyFifo = async;
1552            workerCountWriteFence();
1553            for (ForkJoinWorkerThread w : workers) {
1554                if (w != null)
1555                    w.setAsyncMode(async);
1556            }
1557        }
1558        return oldMode;
1559    }
1560
1561    /**
1549       * Returns {@code true} if this pool uses local first-in-first-out
1550       * scheduling mode for forked tasks that are never joined.
1551       *
1552       * @return {@code true} if this pool uses async mode
1566     * @see #setAsyncMode
1553       */
1554      public boolean getAsyncMode() {
1569        workerCountReadFence();
1555          return locallyFifo;
1556      }
1557  
# Line 1635 | Line 1620 | public class ForkJoinPool extends Abstra
1620       */
1621      public long getQueuedTaskCount() {
1622          long count = 0;
1623 <        for (ForkJoinWorkerThread w : workers) {
1623 >        for (ForkJoinWorkerThread w : workers)
1624              if (w != null)
1625                  count += w.getQueueSize();
1641        }
1626          return count;
1627      }
1628  
# Line 1692 | Line 1676 | public class ForkJoinPool extends Abstra
1676       * @return the number of elements transferred
1677       */
1678      protected int drainTasksTo(Collection<? super ForkJoinTask<?>> c) {
1679 <        int n = submissionQueue.drainTo(c);
1680 <        for (ForkJoinWorkerThread w : workers) {
1679 >        int count = submissionQueue.drainTo(c);
1680 >        for (ForkJoinWorkerThread w : workers)
1681              if (w != null)
1682 <                n += w.drainTasksTo(c);
1683 <        }
1700 <        return n;
1682 >                count += w.drainTasksTo(c);
1683 >        return count;
1684      }
1685  
1686      /**
# Line 1790 | Line 1773 | public class ForkJoinPool extends Abstra
1773       * commenced but not yet completed.  This method may be useful for
1774       * debugging. A return of {@code true} reported a sufficient
1775       * period after shutdown may indicate that submitted tasks have
1776 <     * ignored or suppressed interruption, causing this executor not
1777 <     * to properly terminate.
1776 >     * ignored or suppressed interruption, or are waiting for IO,
1777 >     * causing this executor not to properly terminate. (See the
1778 >     * advisory notes for class {@link ForkJoinTask} stating that
1779 >     * tasks should not normally entail blocking operations.  But if
1780 >     * they do, they must abort them on interrupt.)
1781       *
1782       * @return {@code true} if terminating but not yet terminated
1783       */
# Line 1800 | Line 1786 | public class ForkJoinPool extends Abstra
1786      }
1787  
1788      /**
1789 +     * Returns true if terminating or terminated. Used by ForkJoinWorkerThread.
1790 +     */
1791 +    final boolean isAtLeastTerminating() {
1792 +        return runState >= TERMINATING;
1793 +    }
1794 +
1795 +    /**
1796       * Returns {@code true} if this pool has been shut down.
1797       *
1798       * @return {@code true} if this pool has been shut down
# Line 1821 | Line 1814 | public class ForkJoinPool extends Abstra
1814       */
1815      public boolean awaitTermination(long timeout, TimeUnit unit)
1816          throws InterruptedException {
1817 <        return terminationLatch.await(timeout, unit);
1817 >        try {
1818 >            termination.awaitAdvanceInterruptibly(0, timeout, unit);
1819 >        } catch (TimeoutException ex) {
1820 >            return false;
1821 >        }
1822 >        return true;
1823      }
1824  
1825      /**
1826       * Interface for extending managed parallelism for tasks running
1827       * in {@link ForkJoinPool}s.
1828       *
1829 <     * <p>A {@code ManagedBlocker} provides two methods.
1830 <     * Method {@code isReleasable} must return {@code true} if
1831 <     * blocking is not necessary. Method {@code block} blocks the
1832 <     * current thread if necessary (perhaps internally invoking
1833 <     * {@code isReleasable} before actually blocking).
1829 >     * <p>A {@code ManagedBlocker} provides two methods.  Method
1830 >     * {@code isReleasable} must return {@code true} if blocking is
1831 >     * not necessary. Method {@code block} blocks the current thread
1832 >     * if necessary (perhaps internally invoking {@code isReleasable}
1833 >     * before actually blocking). The unusual methods in this API
1834 >     * accommodate synchronizers that may, but don't usually, block
1835 >     * for long periods. Similarly, they allow more efficient internal
1836 >     * handling of cases in which additional workers may be, but
1837 >     * usually are not, needed to ensure sufficient parallelism.
1838 >     * Toward this end, implementations of method {@code isReleasable}
1839 >     * must be amenable to repeated invocation.
1840       *
1841       * <p>For example, here is a ManagedBlocker based on a
1842       * ReentrantLock:
# Line 1850 | Line 1854 | public class ForkJoinPool extends Abstra
1854       *     return hasLock || (hasLock = lock.tryLock());
1855       *   }
1856       * }}</pre>
1857 +     *
1858 +     * <p>Here is a class that possibly blocks waiting for an
1859 +     * item on a given queue:
1860 +     *  <pre> {@code
1861 +     * class QueueTaker<E> implements ManagedBlocker {
1862 +     *   final BlockingQueue<E> queue;
1863 +     *   volatile E item = null;
1864 +     *   QueueTaker(BlockingQueue<E> q) { this.queue = q; }
1865 +     *   public boolean block() throws InterruptedException {
1866 +     *     if (item == null)
1867 +     *       item = queue.take();
1868 +     *     return true;
1869 +     *   }
1870 +     *   public boolean isReleasable() {
1871 +     *     return item != null || (item = queue.poll()) != null;
1872 +     *   }
1873 +     *   public E getItem() { // call after pool.managedBlock completes
1874 +     *     return item;
1875 +     *   }
1876 +     * }}</pre>
1877       */
1878      public static interface ManagedBlocker {
1879          /**
# Line 1873 | Line 1897 | public class ForkJoinPool extends Abstra
1897       * Blocks in accord with the given blocker.  If the current thread
1898       * is a {@link ForkJoinWorkerThread}, this method possibly
1899       * arranges for a spare thread to be activated if necessary to
1900 <     * ensure parallelism while the current thread is blocked.
1877 <     *
1878 <     * <p>If {@code maintainParallelism} is {@code true} and the pool
1879 <     * supports it ({@link #getMaintainsParallelism}), this method
1880 <     * attempts to maintain the pool's nominal parallelism. Otherwise
1881 <     * it activates a thread only if necessary to avoid complete
1882 <     * starvation. This option may be preferable when blockages use
1883 <     * timeouts, or are almost always brief.
1900 >     * ensure sufficient parallelism while the current thread is blocked.
1901       *
1902       * <p>If the caller is not a {@link ForkJoinTask}, this method is
1903       * behaviorally equivalent to
# Line 1894 | Line 1911 | public class ForkJoinPool extends Abstra
1911       * first be expanded to ensure parallelism, and later adjusted.
1912       *
1913       * @param blocker the blocker
1897     * @param maintainParallelism if {@code true} and supported by
1898     * this pool, attempt to maintain the pool's nominal parallelism;
1899     * otherwise activate a thread only if necessary to avoid
1900     * complete starvation.
1914       * @throws InterruptedException if blocker.block did so
1915       */
1916 <    public static void managedBlock(ManagedBlocker blocker,
1904 <                                    boolean maintainParallelism)
1916 >    public static void managedBlock(ManagedBlocker blocker)
1917          throws InterruptedException {
1918          Thread t = Thread.currentThread();
1919 <        if (t instanceof ForkJoinWorkerThread)
1920 <            ((ForkJoinWorkerThread) t).pool.
1921 <                doBlock(blocker, maintainParallelism);
1922 <        else
1923 <            awaitBlocker(blocker);
1924 <    }
1925 <
1914 <    /**
1915 <     * Performs Non-FJ blocking
1916 <     */
1917 <    private static void awaitBlocker(ManagedBlocker blocker)
1918 <        throws InterruptedException {
1919 <        do {} while (!blocker.isReleasable() && !blocker.block());
1919 >        if (t instanceof ForkJoinWorkerThread) {
1920 >            ForkJoinWorkerThread w = (ForkJoinWorkerThread) t;
1921 >            w.pool.awaitBlocker(blocker);
1922 >        }
1923 >        else {
1924 >            do {} while (!blocker.isReleasable() && !blocker.block());
1925 >        }
1926      }
1927  
1928      // AbstractExecutorService overrides.  These rely on undocumented
# Line 1941 | Line 1947 | public class ForkJoinPool extends Abstra
1947      private static final long eventCountOffset =
1948          objectFieldOffset("eventCount", ForkJoinPool.class);
1949      private static final long eventWaitersOffset =
1950 <        objectFieldOffset("eventWaiters",ForkJoinPool.class);
1950 >        objectFieldOffset("eventWaiters", ForkJoinPool.class);
1951      private static final long stealCountOffset =
1952 <        objectFieldOffset("stealCount",ForkJoinPool.class);
1953 <
1952 >        objectFieldOffset("stealCount", ForkJoinPool.class);
1953 >    private static final long spareWaitersOffset =
1954 >        objectFieldOffset("spareWaiters", ForkJoinPool.class);
1955  
1956      private static long objectFieldOffset(String field, Class<?> klazz) {
1957          try {

Diff Legend

Removed lines
+ Added lines
< Changed lines
> Changed lines