ViewVC Help

View File | Revision Log | Show Annotations | Download File | Root Listing

root/jsr166/jsr166/src/jsr166y/ForkJoinPool.java

Comparing jsr166/src/jsr166y/ForkJoinPool.java (file contents):
Revision 1.56 by dl, Thu May 27 16:46:48 2010 UTC vs.
Revision 1.111 by dl, Thu Jan 26 00:08:13 2012 UTC

#	Line 1 | Line 1
1		/*
2		* Written by Doug Lea with assistance from members of JCP JSR-166
3		* Expert Group and released to the public domain, as explained at
4	<	* http://creativecommons.org/licenses/publicdomain
4	>	* http://creativecommons.org/publicdomain/zero/1.0/
5		*/
6
7		package jsr166y;
8	–
9	–	import java.util.concurrent.*;
10	–
8		import java.util.ArrayList;
9		import java.util.Arrays;
10		import java.util.Collection;
11		import java.util.Collections;
12		import java.util.List;
13	<	import java.util.concurrent.locks.LockSupport;
14	<	import java.util.concurrent.locks.ReentrantLock;
13	>	import java.util.Random;
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.atomic.AtomicInteger;
22	<	import java.util.concurrent.CountDownLatch;
22	>	import java.util.concurrent.atomic.AtomicLong;
23	>	import java.util.concurrent.locks.ReentrantLock;
24	>	import java.util.concurrent.locks.Condition;
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
33		* ExecutorService} mainly by virtue of employing
34		* <em>work-stealing</em>: all threads in the pool attempt to find and
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
41	<	* {@linkplain #setAsyncMode async mode}, a {@code ForkJoinPool} may
42	<	* 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.
35	>	* execute tasks submitted to the pool and/or created by other active
36	>	* tasks (eventually blocking waiting for work if none exist). This
37	>	* enables efficient processing when most tasks spawn other subtasks
38	>	* (as do most {@code ForkJoinTask}s), as well as when many small
39	>	* tasks are submitted to the pool from external clients.  Especially
40	>	* when setting <em>asyncMode</em> to true in constructors, {@code
41	>	* ForkJoinPool}s may also be appropriate for use with event-style
42	>	* tasks that are never joined.
43		*
44		* <p>A {@code ForkJoinPool} is constructed with a given target
45		* parallelism level; by default, equal to the number of available
46	<	* processors. Unless configured otherwise via {@link
47	<	* #setMaintainsParallelism}, the pool attempts to maintain this
48	<	* number of active (or available) threads by dynamically adding,
49	<	* suspending, or resuming internal worker threads, even if some tasks
50	<	* are stalled waiting to join others. However, no such adjustments
51	<	* are performed in the face of blocked IO or other unmanaged
52	<	* 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.
46	>	* processors. The pool attempts to maintain enough active (or
47	>	* available) threads by dynamically adding, suspending, or resuming
48	>	* internal worker threads, even if some tasks are stalled waiting to
49	>	* join others. However, no such adjustments are guaranteed in the
50	>	* face of blocked IO or other unmanaged synchronization. The nested
51	>	* {@link ManagedBlocker} interface enables extension of the kinds of
52	>	* synchronization accommodated.
53		*
54		* <p>In addition to execution and lifecycle control methods, this
55		* class provides status check methods (for example
#	Line 65 | Line 58 | import java.util.concurrent.CountDownLat
58		* {@link #toString} returns indications of pool state in a
59		* convenient form for informal monitoring.
60		*
61	+	* <p> As is the case with other ExecutorServices, there are three
62	+	* main task execution methods summarized in the following
63	+	* table. These are designed to be used primarily by clients not
64	+	* already engaged in fork/join computations in the current pool.  The
65	+	* main forms of these methods accept instances of {@code
66	+	* ForkJoinTask}, but overloaded forms also allow mixed execution of
67	+	* plain {@code Runnable}- or {@code Callable}- based activities as
68	+	* well.  However, tasks that are already executing in a pool should
69	+	* normally instead use the within-computation forms listed in the
70	+	* table unless using async event-style tasks that are not usually
71	+	* joined, in which case there is little difference among choice of
72	+	* methods.
73	+	*
74	+	* <table BORDER CELLPADDING=3 CELLSPACING=1>
75	+	*  <tr>
76	+	*    <td></td>
77	+	*    <td ALIGN=CENTER> <b>Call from non-fork/join clients</b></td>
78	+	*    <td ALIGN=CENTER> <b>Call from within fork/join computations</b></td>
79	+	*  </tr>
80	+	*  <tr>
81	+	*    <td> <b>Arrange async execution</td>
82	+	*    <td> {@link #execute(ForkJoinTask)}</td>
83	+	*    <td> {@link ForkJoinTask#fork}</td>
84	+	*  </tr>
85	+	*  <tr>
86	+	*    <td> <b>Await and obtain result</td>
87	+	*    <td> {@link #invoke(ForkJoinTask)}</td>
88	+	*    <td> {@link ForkJoinTask#invoke}</td>
89	+	*  </tr>
90	+	*  <tr>
91	+	*    <td> <b>Arrange exec and obtain Future</td>
92	+	*    <td> {@link #submit(ForkJoinTask)}</td>
93	+	*    <td> {@link ForkJoinTask#fork} (ForkJoinTasks <em>are</em> Futures)</td>
94	+	*  </tr>
95	+	* </table>
96	+	*
97		* <p><b>Sample Usage.</b> Normally a single {@code ForkJoinPool} is
98		* used for all parallel task execution in a program or subsystem.
99		* Otherwise, use would not usually outweigh the construction and
#	Line 75 | Line 104 | import java.util.concurrent.CountDownLat
104		* daemon} mode, there is typically no need to explicitly {@link
105		* #shutdown} such a pool upon program exit.
106		*
107	<	* <pre>
107	>	*  <pre> {@code
108		* static final ForkJoinPool mainPool = new ForkJoinPool();
109		* ...
110		* public void sort(long[] array) {
111		*   mainPool.invoke(new SortTask(array, 0, array.length));
112	<	* }
84	<	* </pre>
112	>	* }}</pre>
113		*
114		* <p><b>Implementation notes</b>: This implementation restricts the
115		* maximum number of running threads to 32767. Attempts to create
#	Line 89 | Line 117 | import java.util.concurrent.CountDownLat
117		* {@code IllegalArgumentException}.
118		*
119		* <p>This implementation rejects submitted tasks (that is, by throwing
120	<	* {@link RejectedExecutionException}) only when the pool is shut down.
120	>	* {@link RejectedExecutionException}) only when the pool is shut down
121	>	* or internal resources have been exhausted.
122		*
123		* @since 1.7
124		* @author Doug Lea
#	Line 99 | Line 128 | public class ForkJoinPool extends Abstra
128		/*
129		* Implementation Overview
130		*
131	<	* This class provides the central bookkeeping and control for a
132	<	* set of worker threads: Submissions from non-FJ threads enter
133	<	* into a submission queue. Workers take these tasks and typically
134	<	* split them into subtasks that may be stolen by other workers.
135	<	* The main work-stealing mechanics implemented in class
136	<	* ForkJoinWorkerThread give first priority to processing tasks
137	<	* from their own queues (LIFO or FIFO, depending on mode), then
138	<	* to randomized FIFO steals of tasks in other worker queues, and
139	<	* lastly to new submissions. These mechanics do not consider
140	<	* affinities, loads, cache localities, etc, so rarely provide the
141	<	* best possible performance on a given machine, but portably
142	<	* provide good throughput by averaging over these factors.
143	<	* (Further, even if we did try to use such information, we do not
144	<	* usually have a basis for exploiting it. For example, some sets
145	<	* of tasks profit from cache affinities, but others are harmed by
146	<	* cache pollution effects.)
131	>	* This class and its nested classes provide the main
132	>	* functionality and control for a set of worker threads:
133	>	* Submissions from non-FJ threads enter into submission
134	>	* queues. Workers take these tasks and typically split them into
135	>	* subtasks that may be stolen by other workers.  Preference rules
136	>	* give first priority to processing tasks from their own queues
137	>	* (LIFO or FIFO, depending on mode), then to randomized FIFO
138	>	* steals of tasks in other queues.
139	>	*
140	>	* WorkQueues.
141	>	* ==========
142	>	*
143	>	* Most operations occur within work-stealing queues (in nested
144	>	* class WorkQueue).  These are special forms of Deques that
145	>	* support only three of the four possible end-operations -- push,
146	>	* pop, and poll (aka steal), under the further constraints that
147	>	* push and pop are called only from the owning thread (or, as
148	>	* extended here, under a lock), while poll may be called from
149	>	* other threads.  (If you are unfamiliar with them, you probably
150	>	* want to read Herlihy and Shavit's book "The Art of
151	>	* Multiprocessor programming", chapter 16 describing these in
152	>	* more detail before proceeding.)  The main work-stealing queue
153	>	* design is roughly similar to those in the papers "Dynamic
154	>	* Circular Work-Stealing Deque" by Chase and Lev, SPAA 2005
155	>	* (http://research.sun.com/scalable/pubs/index.html) and
156	>	* "Idempotent work stealing" by Michael, Saraswat, and Vechev,
157	>	* PPoPP 2009 (http://portal.acm.org/citation.cfm?id=1504186).
158	>	* The main differences ultimately stem from gc requirements that
159	>	* we null out taken slots as soon as we can, to maintain as small
160	>	* a footprint as possible even in programs generating huge
161	>	* numbers of tasks. To accomplish this, we shift the CAS
162	>	* arbitrating pop vs poll (steal) from being on the indices
163	>	* ("base" and "top") to the slots themselves.  So, both a
164	>	* successful pop and poll mainly entail a CAS of a slot from
165	>	* non-null to null.  Because we rely on CASes of references, we
166	>	* do not need tag bits on base or top.  They are simple ints as
167	>	* used in any circular array-based queue (see for example
168	>	* ArrayDeque).  Updates to the indices must still be ordered in a
169	>	* way that guarantees that top == base means the queue is empty,
170	>	* but otherwise may err on the side of possibly making the queue
171	>	* appear nonempty when a push, pop, or poll have not fully
172	>	* committed. Note that this means that the poll operation,
173	>	* considered individually, is not wait-free. One thief cannot
174	>	* successfully continue until another in-progress one (or, if
175	>	* previously empty, a push) completes.  However, in the
176	>	* aggregate, we ensure at least probabilistic non-blockingness.
177	>	* If an attempted steal fails, a thief always chooses a different
178	>	* random victim target to try next. So, in order for one thief to
179	>	* progress, it suffices for any in-progress poll or new push on
180	>	* any empty queue to complete.
181	>	*
182	>	* This approach also enables support a user mode in which local
183	>	* task processing is in FIFO, not LIFO order, simply by using
184	>	* poll rather than pop.  This can be useful in message-passing
185	>	* frameworks in which tasks are never joined.  However neither
186	>	* mode considers affinities, loads, cache localities, etc, so
187	>	* rarely provide the best possible performance on a given
188	>	* machine, but portably provide good throughput by averaging over
189	>	* these factors.  (Further, even if we did try to use such
190	>	* information, we do not usually have a basis for exploiting
191	>	* it. For example, some sets of tasks profit from cache
192	>	* affinities, but others are harmed by cache pollution effects.)
193	>	*
194	>	* WorkQueues are also used in a similar way for tasks submitted
195	>	* to the pool. We cannot mix these tasks in the same queues used
196	>	* for work-stealing (this would contaminate lifo/fifo
197	>	* processing). Instead, we loosely associate (via hashing)
198	>	* submission queues with submitting threads, and randomly scan
199	>	* these queues as well when looking for work. In essence,
200	>	* submitters act like workers except that they never take tasks,
201	>	* and they are multiplexed on to a finite number of shared work
202	>	* queues. However, classes are set up so that future extensions
203	>	* could allow submitters to optionally help perform tasks as
204	>	* well. Pool submissions from internal workers are also allowed,
205	>	* but use randomized rather than thread-hashed queue indices to
206	>	* avoid imbalance.  Insertion of tasks in shared mode requires a
207	>	* lock (mainly to protect in the case of resizing) but we use
208	>	* only a simple spinlock (using bits in field runState), because
209	>	* submitters encountering a busy queue try others so never block.
210	>	*
211	>	* Management.
212	>	* ==========
213		*
214		* The main throughput advantages of work-stealing stem from
215	<	* decentralized control -- workers mostly steal tasks from each
216	<	* other. We do not want to negate this by creating bottlenecks
217	<	* implementing the management responsibilities of this class. So
218	<	* we use a collection of techniques that avoid, reduce, or cope
219	<	* well with contention. These entail several instances of
220	<	* bit-packing into CASable fields to maintain only the minimally
221	<	* required atomicity. To enable such packing, we restrict maximum
222	<	* parallelism to (1<<15)-1 (enabling twice this to fit into a 16
223	<	* bit field), which is far in excess of normal operating range.
224	<	* Even though updates to some of these bookkeeping fields do
225	<	* sometimes contend with each other, they don't normally
226	<	* cache-contend with updates to others enough to warrant memory
227	<	* padding or isolation. So they are all held as fields of
228	<	* ForkJoinPool objects.  The main capabilities are as follows:
229	<	*
230	<	* 1. Creating and removing workers. Workers are recorded in the
231	<	* "workers" array. This is an array as opposed to some other data
232	<	* structure to support index-based random steals by workers.
233	<	* Updates to the array recording new workers and unrecording
234	<	* terminated ones are protected from each other by a lock
235	<	* (workerLock) but the array is otherwise concurrently readable,
236	<	* and accessed directly by workers. To simplify index-based
215	>	* decentralized control -- workers mostly take tasks from
216	>	* themselves or each other. We cannot negate this in the
217	>	* implementation of other management responsibilities. The main
218	>	* tactic for avoiding bottlenecks is packing nearly all
219	>	* essentially atomic control state into two volatile variables
220	>	* that are by far most often read (not written) as status and
221	>	* consistency checks
222	>	*
223	>	* Field "ctl" contains 64 bits holding all the information needed
224	>	* to atomically decide to add, inactivate, enqueue (on an event
225	>	* queue), dequeue, and/or re-activate workers.  To enable this
226	>	* packing, we restrict maximum parallelism to (1<<15)-1 (which is
227	>	* far in excess of normal operating range) to allow ids, counts,
228	>	* and their negations (used for thresholding) to fit into 16bit
229	>	* fields.
230	>	*
231	>	* Field "runState" contains 32 bits needed to register and
232	>	* deregister WorkQueues, as well as to enable shutdown. It is
233	>	* only modified under a lock (normally briefly held, but
234	>	* occasionally protecting allocations and resizings) but even
235	>	* when locked remains available to check consistency.
236	>	*
237	>	* Recording WorkQueues.  WorkQueues are recorded in the
238	>	* "workQueues" array that is created upon pool construction and
239	>	* expanded if necessary.  Updates to the array while recording
240	>	* new workers and unrecording terminated ones are protected from
241	>	* each other by a lock but the array is otherwise concurrently
242	>	* readable, and accessed directly.  To simplify index-based
243		* operations, the array size is always a power of two, and all
244	<	* readers must tolerate null slots. Currently, all worker thread
245	<	* creation is on-demand, triggered by task submissions,
246	<	* replacement of terminated workers, and/or compensation for
247	<	* blocked workers. However, all other support code is set up to
248	<	* work with other policies.
249	<	*
250	<	* 2. Bookkeeping for dynamically adding and removing workers. We
251	<	* maintain a given level of parallelism (or, if
252	<	* maintainsParallelism is false, at least avoid starvation). When
253	<	* some workers are known to be blocked (on joins or via
254	<	* ManagedBlocker), we may create or resume others to take their
255	<	* place until they unblock (see below). Implementing this
256	<	* requires counts of the number of "running" threads (i.e., those
257	<	* that are neither blocked nor artifically suspended) as well as
258	<	* the total number.  These two values are packed into one field,
259	<	* "workerCounts" because we need accurate snapshots when deciding
260	<	* to create, resume or suspend.  To support these decisions,
261	<	* updates to spare counts must be prospective (not
262	<	* retrospective).  For example, the running count is decremented
263	<	* before blocking by a thread about to block as a spare, but
264	<	* incremented by the thread about to unblock it. Updates upon
265	<	* resumption ofr threads blocking in awaitJoin or awaitBlocker
266	<	* cannot usually be prospective, so the running count is in
267	<	* general an upper bound of the number of productively running
268	<	* threads Updates to the workerCounts field sometimes transiently
269	<	* encounter a fair amount of contention when join dependencies
270	<	* are such that many threads block or unblock at about the same
271	<	* time. We alleviate this by sometimes bundling updates (for
272	<	* example blocking one thread on join and resuming a spare cancel
273	<	* each other out), and in most other cases performing an
274	<	* alternative action like releasing waiters or locating spares.
275	<	*
276	<	* 3. Maintaining global run state. The run state of the pool
277	<	* consists of a runLevel (SHUTDOWN, TERMINATING, etc) similar to
278	<	* those in other Executor implementations, as well as a count of
279	<	* "active" workers -- those that are, or soon will be, or
280	<	* recently were executing tasks. The runLevel and active count
281	<	* are packed together in order to correctly trigger shutdown and
282	<	* termination. Without care, active counts can be subject to very
283	<	* high contention.  We substantially reduce this contention by
284	<	* relaxing update rules.  A worker must claim active status
285	<	* prospectively, by activating if it sees that a submitted or
286	<	* stealable task exists (it may find after activating that the
287	<	* task no longer exists). It stays active while processing this
288	<	* task (if it exists) and any other local subtasks it produces,
289	<	* until it cannot find any other tasks. It then tries
290	<	* inactivating (see method preStep), but upon update contention
291	<	* instead scans for more tasks, later retrying inactivation if it
292	<	* doesn't find any.
293	<	*
294	<	* 4. Managing idle workers waiting for tasks. We cannot let
295	<	* workers spin indefinitely scanning for tasks when none are
296	<	* available. On the other hand, we must quickly prod them into
297	<	* action when new tasks are submitted or generated.  We
298	<	* park/unpark these idle workers using an event-count scheme.
299	<	* Field eventCount is incremented upon events that may enable
300	<	* workers that previously could not find a task to now find one:
301	<	* Submission of a new task to the pool, or another worker pushing
302	<	* a task onto a previously empty queue.  (We also use this
303	<	* mechanism for termination and reconfiguration actions that
304	<	* require wakeups of idle workers).  Each worker maintains its
305	<	* last known event count, and blocks when a scan for work did not
306	<	* find a task AND its lastEventCount matches the current
307	<	* eventCount. Waiting idle workers are recorded in a variant of
308	<	* Treiber stack headed by field eventWaiters which, when nonzero,
309	<	* encodes the thread index and count awaited for by the worker
310	<	* thread most recently calling eventSync. This thread in turn has
311	<	* a record (field nextEventWaiter) for the next waiting worker.
312	<	* In addition to allowing simpler decisions about need for
313	<	* wakeup, the event count bits in eventWaiters serve the role of
314	<	* tags to avoid ABA errors in Treiber stacks.  To reduce delays
315	<	* in task diffusion, workers not otherwise occupied may invoke
316	<	* method releaseWaiters, that removes and signals (unparks)
317	<	* workers not waiting on current count. To minimize task
318	<	* production stalls associate with signalling, any worker pushing
319	<	* a task on an empty queue invokes the weaker method signalWork,
320	<	* that only releases idle workers until it detects interference
321	<	* by other threads trying to release, and lets them take
322	<	* over. The net effect is a tree-like diffusion of signals, where
323	<	* released threads (and possibly others) help with unparks.  To
324	<	* further reduce contention effects a bit, failed CASes to
325	<	* increment field eventCount are tolerated without retries.
326	<	* Conceptually they are merged into the same event, which is OK
327	<	* when their only purpose is to enable workers to scan for work.
328	<	*
329	<	* 5. Managing suspension of extra workers. When a worker is about
330	<	* to block waiting for a join (or via ManagedBlockers), we may
331	<	* create a new thread to maintain parallelism level, or at least
332	<	* avoid starvation (see below). Usually, extra threads are needed
333	<	* for only very short periods, yet join dependencies are such
334	<	* that we sometimes need them in bursts. Rather than create new
335	<	* threads each time this happens, we suspend no-longer-needed
336	<	* extra ones as "spares". For most purposes, we don't distinguish
337	<	* "extra" spare threads from normal "core" threads: On each call
338	<	* to preStep (the only point at which we can do this) a worker
339	<	* checks to see if there are now too many running workers, and if
340	<	* so, suspends itself.  Methods awaitJoin and awaitBlocker look
341	<	* for suspended threads to resume before considering creating a
342	<	* new replacement. We don't need a special data structure to
343	<	* maintain spares; simply scanning the workers array looking for
344	<	* worker.isSuspended() is fine because the calling thread is
345	<	* otherwise not doing anything useful anyway; we are at least as
346	<	* happy if after locating a spare, the caller doesn't actually
347	<	* block because the join is ready before we try to adjust and
348	<	* compensate.  Note that this is intrinsically racy.  One thread
349	<	* may become a spare at about the same time as another is
350	<	* needlessly being created. We counteract this and related slop
351	<	* in part by requiring resumed spares to immediately recheck (in
352	<	* preStep) to see whether they they should re-suspend. The only
353	<	* effective difference between "extra" and "core" threads is that
354	<	* we allow the "extra" ones to time out and die if they are not
355	<	* resumed within a keep-alive interval of a few seconds. This is
356	<	* implemented mainly within ForkJoinWorkerThread, but requires
357	<	* some coordination (isTrimmed() -- meaning killed while
358	<	* suspended) to correctly maintain pool counts.
359	<	*
360	<	* 6. Deciding when to create new workers. The main dynamic
361	<	* control in this class is deciding when to create extra threads,
362	<	* in methods awaitJoin and awaitBlocker. We always
363	<	* need to create one when the number of running threads becomes
364	<	* zero. But because blocked joins are typically dependent, we
365	<	* don't necessarily need or want one-to-one replacement. Using a
366	<	* one-to-one compensation rule often leads to enough useless
367	<	* overhead creating, suspending, resuming, and/or killing threads
368	<	* to signficantly degrade throughput.  We use a rule reflecting
369	<	* the idea that, the more spare threads you already have, the
370	<	* more evidence you need to create another one. The "evidence"
371	<	* here takes two forms: (1) Using a creation threshold expressed
372	<	* in terms of the current deficit -- target minus running
373	<	* threads. To reduce flickering and drift around target values,
374	<	* the relation is quadratic: adding a spare if (dc*dc)>=(sc*pc)
375	<	* (where dc is deficit, sc is number of spare threads and pc is
376	<	* target parallelism.)  (2) Using a form of adaptive
377	<	* spionning. requiring a number of threshold checks proportional
378	<	* to the number of spare threads.  This effectively reduces churn
379	<	* at the price of systematically undershooting target parallelism
380	<	* when many threads are blocked.  However, biasing toward
381	<	* undeshooting partially compensates for the above mechanics to
382	<	* suspend extra threads, that normally lead to overshoot because
383	<	* we can only suspend workers in-between top-level actions. It
384	<	* also better copes with the fact that some of the methods in
385	<	* this class tend to never become compiled (but are interpreted),
386	<	* so some components of the entire set of controls might execute
387	<	* many times faster than others. And similarly for cases where
388	<	* the apparent lack of work is just due to GC stalls and other
389	<	* transient system activity.
390	<	*
391	<	* 7. Maintaining other configuration parameters and monitoring
392	<	* statistics. Updates to fields controlling parallelism level,
393	<	* max size, etc can only meaningfully take effect for individual
394	<	* threads upon their next top-level actions; i.e., between
395	<	* stealing/running tasks/submission, which are separated by calls
396	<	* to preStep.  Memory ordering for these (assumed infrequent)
397	<	* reconfiguration calls is ensured by using reads and writes to
398	<	* volatile field workerCounts (that must be read in preStep anyway)
399	<	* as "fences" -- user-level reads are preceded by reads of
400	<	* workCounts, and writes are followed by no-op CAS to
401	<	* workerCounts. The values reported by other management and
402	<	* monitoring methods are either computed on demand, or are kept
403	<	* in fields that are only updated when threads are otherwise
404	<	* idle.
244	>	* readers must tolerate null slots. Shared (submission) queues
245	>	* are at even indices, worker queues at odd indices. Grouping
246	>	* them together in this way simplifies and speeds up task
247	>	* scanning. To avoid flailing during start-up, the array is
248	>	* presized to hold twice #parallelism workers (which is unlikely
249	>	* to need further resizing during execution). But to avoid
250	>	* dealing with so many null slots, variable runState includes a
251	>	* mask for the nearest power of two that contains all current
252	>	* workers.  All worker thread creation is on-demand, triggered by
253	>	* task submissions, replacement of terminated workers, and/or
254	>	* compensation for blocked workers. However, all other support
255	>	* code is set up to work with other policies.  To ensure that we
256	>	* do not hold on to worker references that would prevent GC, ALL
257	>	* accesses to workQueues are via indices into the workQueues
258	>	* array (which is one source of some of the messy code
259	>	* constructions here). In essence, the workQueues array serves as
260	>	* a weak reference mechanism. Thus for example the wait queue
261	>	* field of ctl stores indices, not references.  Access to the
262	>	* workQueues in associated methods (for example signalWork) must
263	>	* both index-check and null-check the IDs. All such accesses
264	>	* ignore bad IDs by returning out early from what they are doing,
265	>	* since this can only be associated with termination, in which
266	>	* case it is OK to give up.
267	>	*
268	>	* All uses of the workQueues array check that it is non-null
269	>	* (even if previously non-null). This allows nulling during
270	>	* termination, which is currently not necessary, but remains an
271	>	* option for resource-revocation-based shutdown schemes. It also
272	>	* helps reduce JIT issuance of uncommon-trap code, which tends to
273	>	* unnecessarily complicate control flow in some methods.
274	>	*
275	>	* Event Queuing. Unlike HPC work-stealing frameworks, we cannot
276	>	* let workers spin indefinitely scanning for tasks when none can
277	>	* be found immediately, and we cannot start/resume workers unless
278	>	* there appear to be tasks available.  On the other hand, we must
279	>	* quickly prod them into action when new tasks are submitted or
280	>	* generated. In many usages, ramp-up time to activate workers is
281	>	* the main limiting factor in overall performance (this is
282	>	* compounded at program start-up by JIT compilation and
283	>	* allocation). So we try to streamline this as much as possible.
284	>	* We park/unpark workers after placing in an event wait queue
285	>	* when they cannot find work. This "queue" is actually a simple
286	>	* Treiber stack, headed by the "id" field of ctl, plus a 15bit
287	>	* counter value (that reflects the number of times a worker has
288	>	* been inactivated) to avoid ABA effects (we need only as many
289	>	* version numbers as worker threads). Successors are held in
290	>	* field WorkQueue.nextWait.  Queuing deals with several intrinsic
291	>	* races, mainly that a task-producing thread can miss seeing (and
292	>	* signalling) another thread that gave up looking for work but
293	>	* has not yet entered the wait queue. We solve this by requiring
294	>	* a full sweep of all workers (via repeated calls to method
295	>	* scan()) both before and after a newly waiting worker is added
296	>	* to the wait queue. During a rescan, the worker might release
297	>	* some other queued worker rather than itself, which has the same
298	>	* net effect. Because enqueued workers may actually be rescanning
299	>	* rather than waiting, we set and clear the "parker" field of
300	>	* Workqueues to reduce unnecessary calls to unpark.  (this
301	>	* requires a secondary recheck to avoid missed signals.)  Note
302	>	* the unusual conventions about Thread.interrupts surrounding
303	>	* parking and other blocking: Because interrupts are used solely
304	>	* to alert threads to check termination, which is checked anyway
305	>	* upon blocking, we clear status (using Thread.interrupted)
306	>	* before any call to park, so that park does not immediately
307	>	* return due to status being set via some other unrelated call to
308	>	* interrupt in user code.
309	>	*
310	>	* Signalling.  We create or wake up workers only when there
311	>	* appears to be at least one task they might be able to find and
312	>	* execute.  When a submission is added or another worker adds a
313	>	* task to a queue that previously had fewer than two tasks, they
314	>	* signal waiting workers (or trigger creation of new ones if
315	>	* fewer than the given parallelism level -- see signalWork).
316	>	* These primary signals are buttressed by signals during rescans;
317	>	* together these cover the signals needed in cases when more
318	>	* tasks are pushed but untaken, and improve performance compared
319	>	* to having one thread wake up all workers.
320	>	*
321	>	* Trimming workers. To release resources after periods of lack of
322	>	* use, a worker starting to wait when the pool is quiescent will
323	>	* time out and terminate if the pool has remained quiescent for
324	>	* SHRINK_RATE nanosecs. This will slowly propagate, eventually
325	>	* terminating all workers after long periods of non-use.
326	>	*
327	>	* Shutdown and Termination. A call to shutdownNow atomically sets
328	>	* a runState bit and then (non-atomically) sets each workers
329	>	* runState status, cancels all unprocessed tasks, and wakes up
330	>	* all waiting workers.  Detecting whether termination should
331	>	* commence after a non-abrupt shutdown() call requires more work
332	>	* and bookkeeping. We need consensus about quiescence (i.e., that
333	>	* there is no more work). The active count provides a primary
334	>	* indication but non-abrupt shutdown still requires a rechecking
335	>	* scan for any workers that are inactive but not queued.
336	>	*
337	>	* Joining Tasks.
338	>	* ==============
339	>	*
340	>	* Any of several actions may be taken when one worker is waiting
341	>	* to join a task stolen (or always held by) another.  Because we
342	>	* are multiplexing many tasks on to a pool of workers, we can't
343	>	* just let them block (as in Thread.join).  We also cannot just
344	>	* reassign the joiner's run-time stack with another and replace
345	>	* it later, which would be a form of "continuation", that even if
346	>	* possible is not necessarily a good idea since we sometimes need
347	>	* both an unblocked task and its continuation to
348	>	* progress. Instead we combine two tactics:
349	>	*
350	>	*   Helping: Arranging for the joiner to execute some task that it
351	>	*      would be running if the steal had not occurred.
352	>	*
353	>	*   Compensating: Unless there are already enough live threads,
354	>	*      method tryCompensate() may create or re-activate a spare
355	>	*      thread to compensate for blocked joiners until they unblock.
356	>	*
357	>	* A third form (implemented in tryRemoveAndExec and
358	>	* tryPollForAndExec) amounts to helping a hypothetical
359	>	* compensator: If we can readily tell that a possible action of a
360	>	* compensator is to steal and execute the task being joined, the
361	>	* joining thread can do so directly, without the need for a
362	>	* compensation thread (although at the expense of larger run-time
363	>	* stacks, but the tradeoff is typically worthwhile).
364	>	*
365	>	* The ManagedBlocker extension API can't use helping so relies
366	>	* only on compensation in method awaitBlocker.
367	>	*
368	>	* The algorithm in tryHelpStealer entails a form of "linear"
369	>	* helping: Each worker records (in field currentSteal) the most
370	>	* recent task it stole from some other worker. Plus, it records
371	>	* (in field currentJoin) the task it is currently actively
372	>	* joining. Method tryHelpStealer uses these markers to try to
373	>	* find a worker to help (i.e., steal back a task from and execute
374	>	* it) that could hasten completion of the actively joined task.
375	>	* In essence, the joiner executes a task that would be on its own
376	>	* local deque had the to-be-joined task not been stolen. This may
377	>	* be seen as a conservative variant of the approach in Wagner &
378	>	* Calder "Leapfrogging: a portable technique for implementing
379	>	* efficient futures" SIGPLAN Notices, 1993
380	>	* (http://portal.acm.org/citation.cfm?id=155354). It differs in
381	>	* that: (1) We only maintain dependency links across workers upon
382	>	* steals, rather than use per-task bookkeeping.  This sometimes
383	>	* requires a linear scan of workers array to locate stealers, but
384	>	* often doesn't because stealers leave hints (that may become
385	>	* stale/wrong) of where to locate them.  A stealHint is only a
386	>	* hint because a worker might have had multiple steals and the
387	>	* hint records only one of them (usually the most current).
388	>	* Hinting isolates cost to when it is needed, rather than adding
389	>	* to per-task overhead.  (2) It is "shallow", ignoring nesting
390	>	* and potentially cyclic mutual steals.  (3) It is intentionally
391	>	* racy: field currentJoin is updated only while actively joining,
392	>	* which means that we miss links in the chain during long-lived
393	>	* tasks, GC stalls etc (which is OK since blocking in such cases
394	>	* is usually a good idea).  (4) We bound the number of attempts
395	>	* to find work (see MAX_HELP_DEPTH) and fall back to suspending
396	>	* the worker and if necessary replacing it with another.
397	>	*
398	>	* It is impossible to keep exactly the target parallelism number
399	>	* of threads running at any given time.  Determining the
400	>	* existence of conservatively safe helping targets, the
401	>	* availability of already-created spares, and the apparent need
402	>	* to create new spares are all racy, so we rely on multiple
403	>	* retries of each.  Currently, in keeping with on-demand
404	>	* signalling policy, we compensate only if blocking would leave
405	>	* less than one active (non-waiting, non-blocked) worker.
406	>	* Additionally, to avoid some false alarms due to GC, lagging
407	>	* counters, system activity, etc, compensated blocking for joins
408	>	* is only attempted after rechecks stabilize in
409	>	* ForkJoinTask.awaitJoin. (Retries are interspersed with
410	>	* Thread.yield, for good citizenship.)
411		*
412	<	* Beware that there is a lot of representation-level coupling
412	>	* Style notes: There is a lot of representation-level coupling
413		* among classes ForkJoinPool, ForkJoinWorkerThread, and
414	<	* ForkJoinTask.  For example, direct access to "workers" array by
415	<	* workers, and direct access to ForkJoinTask.status by both
416	<	* ForkJoinPool and ForkJoinWorkerThread.  There is little point
417	<	* trying to reduce this, since any associated future changes in
418	<	* representations will need to be accompanied by algorithmic
419	<	* changes anyway.
420	<	*
421	<	* Style notes: There are lots of inline assignments (of form
422	<	* "while ((local = field) != 0)") which are usually the simplest
423	<	* way to ensure read orderings. Also several occurrences of the
424	<	* unusual "do {} while(!cas...)" which is the simplest way to
425	<	* force an update of a CAS'ed variable. There are also a few
426	<	* other coding oddities that help some methods perform reasonably
427	<	* even when interpreted (not compiled).
428	<	*
429	<	* The order of declarations in this file is: (1) statics (2)
430	<	* fields (along with constants used when unpacking some of them)
431	<	* (3) internal control methods (4) callbacks and other support
432	<	* for ForkJoinTask and ForkJoinWorkerThread classes, (5) exported
433	<	* methods (plus a few little helpers).
414	>	* ForkJoinTask.  The fields of WorkQueue maintain data structures
415	>	* managed by ForkJoinPool, so are directly accessed.  There is
416	>	* little point trying to reduce this, since any associated future
417	>	* changes in representations will need to be accompanied by
418	>	* algorithmic changes anyway. All together, these low-level
419	>	* implementation choices produce as much as a factor of 4
420	>	* performance improvement compared to naive implementations, and
421	>	* enable the processing of billions of tasks per second, at the
422	>	* expense of some ugliness.
423	>	*
424	>	* Methods signalWork() and scan() are the main bottlenecks so are
425	>	* especially heavily micro-optimized/mangled.  There are lots of
426	>	* inline assignments (of form "while ((local = field) != 0)")
427	>	* which are usually the simplest way to ensure the required read
428	>	* orderings (which are sometimes critical). This leads to a
429	>	* "C"-like style of listing declarations of these locals at the
430	>	* heads of methods or blocks.  There are several occurrences of
431	>	* the unusual "do {} while (!cas...)"  which is the simplest way
432	>	* to force an update of a CAS'ed variable. There are also other
433	>	* coding oddities that help some methods perform reasonably even
434	>	* when interpreted (not compiled).
435	>	*
436	>	* The order of declarations in this file is: (1) declarations of
437	>	* statics (2) fields (along with constants used when unpacking
438	>	* some of them), listed in an order that tends to reduce
439	>	* contention among them a bit under most JVMs; (3) nested
440	>	* classes; (4) internal control methods; (5) callbacks and other
441	>	* support for ForkJoinTask methods; (6) exported methods (plus a
442	>	* few little helpers); (7) static block initializing all statics
443	>	* in a minimally dependent order.
444		*/
445
446		/**
#	Line 346 | Line 463 | public class ForkJoinPool extends Abstra
463		* Default ForkJoinWorkerThreadFactory implementation; creates a
464		* new ForkJoinWorkerThread.
465		*/
466	<	static class  DefaultForkJoinWorkerThreadFactory
466	>	static class DefaultForkJoinWorkerThreadFactory
467		implements ForkJoinWorkerThreadFactory {
468		public ForkJoinWorkerThread newThread(ForkJoinPool pool) {
469		return new ForkJoinWorkerThread(pool);
#	Line 358 | Line 475 | public class ForkJoinPool extends Abstra
475		* overridden in ForkJoinPool constructors.
476		*/
477		public static final ForkJoinWorkerThreadFactory
478	<	defaultForkJoinWorkerThreadFactory =
362	<	new DefaultForkJoinWorkerThreadFactory();
478	>	defaultForkJoinWorkerThreadFactory;
479
480		/**
481		* Permission required for callers of methods that may start or
482		* kill threads.
483		*/
484	<	private static final RuntimePermission modifyThreadPermission =
369	<	new RuntimePermission("modifyThread");
484	>	private static final RuntimePermission modifyThreadPermission;
485
486		/**
487		* If there is a security manager, makes sure caller has
#	Line 381 | Line 496 | public class ForkJoinPool extends Abstra
496		/**
497		* Generator for assigning sequence numbers as pool names.
498		*/
499	<	private static final AtomicInteger poolNumberGenerator =
385	<	new AtomicInteger();
499	>	private static final AtomicInteger poolNumberGenerator;
500
501		/**
502	<	* Absolute bound for parallelism level. Twice this number must
503	<	* fit into a 16bit field to enable word-packing for some counts.
502	>	* Bits and masks for control variables
503	>	*
504	>	* Field ctl is a long packed with:
505	>	* AC: Number of active running workers minus target parallelism (16 bits)
506	>	* TC: Number of total workers minus target parallelism (16 bits)
507	>	* ST: true if pool is terminating (1 bit)
508	>	* EC: the wait count of top waiting thread (15 bits)
509	>	* ID: ~(poolIndex >>> 1) of top of Treiber stack of waiters (16 bits)
510	>	*
511	>	* When convenient, we can extract the upper 32 bits of counts and
512	>	* the lower 32 bits of queue state, u = (int)(ctl >>> 32) and e =
513	>	* (int)ctl.  The ec field is never accessed alone, but always
514	>	* together with id and st. The offsets of counts by the target
515	>	* parallelism and the positionings of fields makes it possible to
516	>	* perform the most common checks via sign tests of fields: When
517	>	* ac is negative, there are not enough active workers, when tc is
518	>	* negative, there are not enough total workers, when id is
519	>	* negative, there is at least one waiting worker, and when e is
520	>	* negative, the pool is terminating.  To deal with these possibly
521	>	* negative fields, we use casts in and out of "short" and/or
522	>	* signed shifts to maintain signedness.
523	>	*
524	>	* When a thread is queued (inactivated), its eventCount field is
525	>	* negative, which is the only way to tell if a worker is
526	>	* prevented from executing tasks, even though it must continue to
527	>	* scan for them to avoid queuing races.
528	>	*
529	>	* Field runState is an int packed with:
530	>	* SHUTDOWN: true if shutdown is enabled (1 bit)
531	>	* SEQ:  a sequence number updated upon (de)registering workers (15 bits)
532	>	* MASK: mask (power of 2 - 1) covering all registered poolIndexes (16 bits)
533	>	*
534	>	* The combination of mask and sequence number enables simple
535	>	* consistency checks: Staleness of read-only operations on the
536	>	* workers and queues arrays can be checked by comparing runState
537	>	* before vs after the reads. The low 16 bits (i.e, anding with
538	>	* SMASK) hold (the smallest power of two covering all worker
539	>	* indices, minus one.  The mask for queues (vs workers) is twice
540	>	* this value plus 1.
541	>	*/
542	>
543	>	// bit positions/shifts for fields
544	>	private static final int  AC_SHIFT   = 48;
545	>	private static final int  TC_SHIFT   = 32;
546	>	private static final int  ST_SHIFT   = 31;
547	>	private static final int  EC_SHIFT   = 16;
548	>
549	>	// bounds
550	>	private static final int  MAX_ID     = 0x7fff;  // max poolIndex
551	>	private static final int  SMASK      = 0xffff;  // mask short bits
552	>	private static final int  SHORT_SIGN = 1 << 15;
553	>	private static final int  INT_SIGN   = 1 << 31;
554	>
555	>	// masks
556	>	private static final long STOP_BIT   = 0x0001L << ST_SHIFT;
557	>	private static final long AC_MASK    = ((long)SMASK) << AC_SHIFT;
558	>	private static final long TC_MASK    = ((long)SMASK) << TC_SHIFT;
559	>
560	>	// units for incrementing and decrementing
561	>	private static final long TC_UNIT    = 1L << TC_SHIFT;
562	>	private static final long AC_UNIT    = 1L << AC_SHIFT;
563	>
564	>	// masks and units for dealing with u = (int)(ctl >>> 32)
565	>	private static final int  UAC_SHIFT  = AC_SHIFT - 32;
566	>	private static final int  UTC_SHIFT  = TC_SHIFT - 32;
567	>	private static final int  UAC_MASK   = SMASK << UAC_SHIFT;
568	>	private static final int  UTC_MASK   = SMASK << UTC_SHIFT;
569	>	private static final int  UAC_UNIT   = 1 << UAC_SHIFT;
570	>	private static final int  UTC_UNIT   = 1 << UTC_SHIFT;
571	>
572	>	// masks and units for dealing with e = (int)ctl
573	>	private static final int E_MASK      = 0x7fffffff; // no STOP_BIT
574	>	private static final int E_SEQ       = 1 << EC_SHIFT;
575	>
576	>	// runState bits
577	>	private static final int SHUTDOWN    = 1 << 31;
578	>	private static final int RS_SEQ      = 1 << 16;
579	>	private static final int RS_SEQ_MASK = 0x7fff0000;
580	>
581	>	// access mode for WorkQueue
582	>	static final int LIFO_QUEUE          =  0;
583	>	static final int FIFO_QUEUE          =  1;
584	>	static final int SHARED_QUEUE        = -1;
585	>
586	>	/**
587	>	* The wakeup interval (in nanoseconds) for a worker waiting for a
588	>	* task when the pool is quiescent to instead try to shrink the
589	>	* number of workers.  The exact value does not matter too
590	>	* much. It must be short enough to release resources during
591	>	* sustained periods of idleness, but not so short that threads
592	>	* are continually re-created.
593	>	*/
594	>	private static final long SHRINK_RATE =
595	>	4L * 1000L * 1000L * 1000L; // 4 seconds
596	>
597	>	/**
598	>	* The timeout value for attempted shrinkage, includes
599	>	* some slop to cope with system timer imprecision.
600	>	*/
601	>	private static final long SHRINK_TIMEOUT = SHRINK_RATE - (SHRINK_RATE / 10);
602	>
603	>	/**
604	>	* The maximum stolen->joining link depth allowed in tryHelpStealer.
605	>	* Depths for legitimate chains are unbounded, but we use a fixed
606	>	* constant to avoid (otherwise unchecked) cycles and to bound
607	>	* staleness of traversal parameters at the expense of sometimes
608	>	* blocking when we could be helping.
609		*/
610	<	private static final int MAX_THREADS = 0x7fff;
610	>	private static final int MAX_HELP_DEPTH = 16;
611
612	<	/**
613	<	* Array holding all worker threads in the pool.  Array size must
614	<	* be a power of two.  Updates and replacements are protected by
615	<	* workerLock, but the array is always kept in a consistent enough
616	<	* state to be randomly accessed without locking by workers
617	<	* performing work-stealing, as well as other traversal-based
618	<	* methods in this class. All readers must tolerate that some
619	<	* array slots may be null.
612	>	/*
613	>	* Field layout order in this class tends to matter more than one
614	>	* would like. Runtime layout order is only loosely related to
615	>	* declaration order and may differ across JVMs, but the following
616	>	* empirically works OK on current JVMs.
617	>	*/
618	>
619	>	volatile long ctl;                       // main pool control
620	>	final int parallelism;                   // parallelism level
621	>	final int localMode;                     // per-worker scheduling mode
622	>	int nextPoolIndex;                       // hint used in registerWorker
623	>	volatile int runState;                   // shutdown status, seq, and mask
624	>	WorkQueue[] workQueues;                  // main registry
625	>	final ReentrantLock lock;                // for registration
626	>	final Condition termination;             // for awaitTermination
627	>	final ForkJoinWorkerThreadFactory factory; // factory for new workers
628	>	final Thread.UncaughtExceptionHandler ueh; // per-worker UEH
629	>	final AtomicLong stealCount;             // collect counts when terminated
630	>	final AtomicInteger nextWorkerNumber;    // to create worker name string
631	>	final String workerNamePrefix;           // Prefix for assigning worker names
632	>
633	>	/**
634	>	* Queues supporting work-stealing as well as external task
635	>	* submission. See above for main rationale and algorithms.
636	>	* Implementation relies heavily on "Unsafe" intrinsics
637	>	* and selective use of "volatile":
638	>	*
639	>	* Field "base" is the index (mod array.length) of the least valid
640	>	* queue slot, which is always the next position to steal (poll)
641	>	* from if nonempty. Reads and writes require volatile orderings
642	>	* but not CAS, because updates are only performed after slot
643	>	* CASes.
644	>	*
645	>	* Field "top" is the index (mod array.length) of the next queue
646	>	* slot to push to or pop from. It is written only by owner thread
647	>	* for push, or under lock for trySharedPush, and accessed by
648	>	* other threads only after reading (volatile) base.  Both top and
649	>	* base are allowed to wrap around on overflow, but (top - base)
650	>	* (or more comonly -(base - top) to force volatile read of base
651	>	* before top) still estimates size.
652	>	*
653	>	* The array slots are read and written using the emulation of
654	>	* volatiles/atomics provided by Unsafe. Insertions must in
655	>	* general use putOrderedObject as a form of releasing store to
656	>	* ensure that all writes to the task object are ordered before
657	>	* its publication in the queue. (Although we can avoid one case
658	>	* of this when locked in trySharedPush.) All removals entail a
659	>	* CAS to null.  The array is always a power of two. To ensure
660	>	* safety of Unsafe array operations, all accesses perform
661	>	* explicit null checks and implicit bounds checks via
662	>	* power-of-two masking.
663	>	*
664	>	* In addition to basic queuing support, this class contains
665	>	* fields described elsewhere to control execution. It turns out
666	>	* to work better memory-layout-wise to include them in this
667	>	* class rather than a separate class.
668	>	*
669	>	* Performance on most platforms is very sensitive to placement of
670	>	* instances of both WorkQueues and their arrays -- we absolutely
671	>	* do not want multiple WorkQueue instances or multiple queue
672	>	* arrays sharing cache lines. (It would be best for queue objects
673	>	* and their arrays to share, but there is nothing available to
674	>	* help arrange that).  Unfortunately, because they are recorded
675	>	* in a common array, WorkQueue instances are often moved to be
676	>	* adjacent by garbage collectors. To reduce impact, we use field
677	>	* padding that works OK on common platforms; this effectively
678	>	* trades off slightly slower average field access for the sake of
679	>	* avoiding really bad worst-case access. (Until better JVM
680	>	* support is in place, this padding is dependent on transient
681	>	* properties of JVM field layout rules.)  We also take care in
682	>	* allocating and sizing and resizing the array. Non-shared queue
683	>	* arrays are initialized (via method growArray) by workers before
684	>	* use. Others are allocated on first use.
685		*/
686	<	volatile ForkJoinWorkerThread[] workers;
686	>	static final class WorkQueue {
687	>	/**
688	>	* Capacity of work-stealing queue array upon initialization.
689	>	* Must be a power of two; at least 4, but set larger to
690	>	* reduce cacheline sharing among queues.
691	>	*/
692	>	static final int INITIAL_QUEUE_CAPACITY = 1 << 8;
693
694	<	/**
695	<	* Queue for external submissions.
696	<	*/
697	<	private final LinkedTransferQueue<ForkJoinTask<?>> submissionQueue;
694	>	/**
695	>	* Maximum size for queue arrays. Must be a power of two less
696	>	* than or equal to 1 << (31 - width of array entry) to ensure
697	>	* lack of wraparound of index calculations, but defined to a
698	>	* value a bit less than this to help users trap runaway
699	>	* programs before saturating systems.
700	>	*/
701	>	static final int MAXIMUM_QUEUE_CAPACITY = 1 << 26; // 64M
702
703	<	/**
704	<	* Lock protecting updates to workers array.
705	<	*/
706	<	private final ReentrantLock workerLock;
703	>	volatile long totalSteals; // cumulative number of steals
704	>	int seed;                  // for random scanning; initialize nonzero
705	>	volatile int eventCount;   // encoded inactivation count; < 0 if inactive
706	>	int nextWait;              // encoded record of next event waiter
707	>	int rescans;               // remaining scans until block
708	>	int nsteals;               // top-level task executions since last idle
709	>	final int mode;            // lifo, fifo, or shared
710	>	int poolIndex;             // index of this queue in pool (or 0)
711	>	int stealHint;             // index of most recent known stealer
712	>	volatile int runState;     // 1: locked, -1: terminate; else 0
713	>	volatile int base;         // index of next slot for poll
714	>	int top;                   // index of next slot for push
715	>	ForkJoinTask<?>[] array;   // the elements (initially unallocated)
716	>	final ForkJoinWorkerThread owner; // owning thread or null if shared
717	>	volatile Thread parker;    // == owner during call to park; else null
718	>	ForkJoinTask<?> currentJoin;  // task being joined in awaitJoin
719	>	ForkJoinTask<?> currentSteal; // current non-local task being executed
720	>	// Heuristic padding to ameliorate unfortunate memory placements
721	>	Object p00, p01, p02, p03, p04, p05, p06, p07, p08, p09, p0a;
722	>
723	>	WorkQueue(ForkJoinWorkerThread owner, int mode) {
724	>	this.owner = owner;
725	>	this.mode = mode;
726	>	// Place indices in the center of array (that is not yet allocated)
727	>	base = top = INITIAL_QUEUE_CAPACITY >>> 1;
728	>	}
729
730	<	/**
731	<	* Latch released upon termination.
732	<	*/
733	<	private final CountDownLatch terminationLatch;
730	>	/**
731	>	* Returns number of tasks in the queue
732	>	*/
733	>	final int queueSize() {
734	>	int n = base - top; // non-owner callers must read base first
735	>	return (n >= 0) ? 0 : -n;
736	>	}
737
738	<	/**
739	<	* Creation factory for worker threads.
740	<	*/
741	<	private final ForkJoinWorkerThreadFactory factory;
738	>	/**
739	>	* Pushes a task. Call only by owner in unshared queues.
740	>	*
741	>	* @param task the task. Caller must ensure non-null.
742	>	* @param p, if non-null, pool to signal if necessary
743	>	* @throw RejectedExecutionException if array cannot
744	>	* be resized
745	>	*/
746	>	final void push(ForkJoinTask<?> task, ForkJoinPool p) {
747	>	boolean signal = false;
748	>	ForkJoinTask<?>[] a;
749	>	int s = top, m, n;
750	>	if ((a = array) != null) {    // ignore if queue removed
751	>	U.putOrderedObject
752	>	(a, (((m = a.length - 1) & s) << ASHIFT) + ABASE, task);
753	>	if ((n = (top = s + 1) - base) <= 2) {
754	>	if (p != null)
755	>	p.signalWork();
756	>	}
757	>	else if (n >= m)
758	>	growArray(true);
759	>	}
760	>	}
761
762	<	/**
763	<	* Sum of per-thread steal counts, updated only when threads are
764	<	* idle or terminating.
765	<	*/
766	<	private volatile long stealCount;
762	>	/**
763	>	* Pushes a task if lock is free and array is either big
764	>	* enough or can be resized to be big enough.
765	>	*
766	>	* @param task the task. Caller must ensure non-null.
767	>	* @return true if submitted
768	>	*/
769	>	final boolean trySharedPush(ForkJoinTask<?> task) {
770	>	boolean submitted = false;
771	>	if (runState == 0 && U.compareAndSwapInt(this, RUNSTATE, 0, 1)) {
772	>	ForkJoinTask<?>[] a = array;
773	>	int s = top, n = s - base;
774	>	try {
775	>	if ((a != null && n < a.length - 1) ||
776	>	(a = growArray(false)) != null) { // must presize
777	>	int j = (((a.length - 1) & s) << ASHIFT) + ABASE;
778	>	U.putObject(a, (long)j, task);    // don't need "ordered"
779	>	top = s + 1;
780	>	submitted = true;
781	>	}
782	>	} finally {
783	>	runState = 0;                         // unlock
784	>	}
785	>	}
786	>	return submitted;
787	>	}
788
789	<	/**
790	<	* Encoded record of top of treiber stack of threads waiting for
791	<	* events. The top 32 bits contain the count being waited for. The
792	<	* bottom word contains one plus the pool index of waiting worker
793	<	* thread.
794	<	*/
795	<	private volatile long eventWaiters;
789	>	/**
790	>	* Takes next task, if one exists, in FIFO order.
791	>	*/
792	>	final ForkJoinTask<?> poll() {
793	>	ForkJoinTask<?>[] a; int b, i;
794	>	while ((b = base) - top < 0 && (a = array) != null &&
795	>	(i = (a.length - 1) & b) >= 0) {
796	>	int j = (i << ASHIFT) + ABASE;
797	>	ForkJoinTask<?> t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
798	>	if (t != null && base == b &&
799	>	U.compareAndSwapObject(a, j, t, null)) {
800	>	base = b + 1;
801	>	return t;
802	>	}
803	>	}
804	>	return null;
805	>	}
806
807	<	private static final int  EVENT_COUNT_SHIFT = 32;
808	<	private static final long WAITER_INDEX_MASK = (1L << EVENT_COUNT_SHIFT)-1L;
807	>	/**
808	>	* Takes next task, if one exists, in LIFO order.
809	>	* Call only by owner in unshared queues.
810	>	*/
811	>	final ForkJoinTask<?> pop() {
812	>	ForkJoinTask<?> t; int m;
813	>	ForkJoinTask<?>[] a = array;
814	>	if (a != null && (m = a.length - 1) >= 0) {
815	>	for (int s; (s = top - 1) - base >= 0;) {
816	>	int j = ((m & s) << ASHIFT) + ABASE;
817	>	if ((t = (ForkJoinTask<?>)U.getObjectVolatile(a, j)) == null)
818	>	break;
819	>	if (U.compareAndSwapObject(a, j, t, null)) {
820	>	top = s;
821	>	return t;
822	>	}
823	>	}
824	>	}
825	>	return null;
826	>	}
827
828	<	/**
829	<	* A counter for events that may wake up worker threads:
830	<	*   - Submission of a new task to the pool
831	<	*   - A worker pushing a task on an empty queue
832	<	*   - termination and reconfiguration
833	<	*/
447	<	private volatile int eventCount;
828	>	/**
829	>	* Takes next task, if one exists, in order specified by mode.
830	>	*/
831	>	final ForkJoinTask<?> nextLocalTask() {
832	>	return mode == 0 ? pop() : poll();
833	>	}
834
835	<	/**
836	<	* Lifecycle control. The low word contains the number of workers
837	<	* that are (probably) executing tasks. This value is atomically
838	<	* incremented before a worker gets a task to run, and decremented
839	<	* when worker has no tasks and cannot find any.  Bits 16-18
840	<	* contain runLevel value. When all are zero, the pool is
841	<	* running. Level transitions are monotonic (running -> shutdown
842	<	* -> terminating -> terminated) so each transition adds a bit.
843	<	* These are bundled together to ensure consistent read for
844	<	* termination checks (i.e., that runLevel is at least SHUTDOWN
845	<	* and active threads is zero).
460	<	*/
461	<	private volatile int runState;
835	>	/**
836	>	* Returns next task, if one exists, in order specified by mode.
837	>	*/
838	>	final ForkJoinTask<?> peek() {
839	>	ForkJoinTask<?>[] a = array; int m;
840	>	if (a == null || (m = a.length - 1) < 0)
841	>	return null;
842	>	int i = mode == 0 ? top - 1 : base;
843	>	int j = ((i & m) << ASHIFT) + ABASE;
844	>	return (ForkJoinTask<?>)U.getObjectVolatile(a, j);
845	>	}
846
847	<	// Note: The order among run level values matters.
848	<	private static final int RUNLEVEL_SHIFT     = 16;
849	<	private static final int SHUTDOWN           = 1 << RUNLEVEL_SHIFT;
850	<	private static final int TERMINATING        = 1 << (RUNLEVEL_SHIFT + 1);
851	<	private static final int TERMINATED         = 1 << (RUNLEVEL_SHIFT + 2);
852	<	private static final int ACTIVE_COUNT_MASK  = (1 << RUNLEVEL_SHIFT) - 1;
853	<	private static final int ONE_ACTIVE         = 1; // active update delta
847	>	/**
848	>	* Returns task at index b if b is current base of queue.
849	>	*/
850	>	final ForkJoinTask<?> pollAt(int b) {
851	>	ForkJoinTask<?>[] a; int i;
852	>	ForkJoinTask<?> task = null;
853	>	if ((a = array) != null && (i = ((a.length - 1) & b)) >= 0) {
854	>	int j = (i << ASHIFT) + ABASE;
855	>	ForkJoinTask<?> t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
856	>	if (t != null && base == b &&
857	>	U.compareAndSwapObject(a, j, t, null)) {
858	>	base = b + 1;
859	>	task = t;
860	>	}
861	>	}
862	>	return task;
863	>	}
864
865	<	/**
866	<	* Holds number of total (i.e., created and not yet terminated)
867	<	* and running (i.e., not blocked on joins or other managed sync)
868	<	* threads, packed together to ensure consistent snapshot when
869	<	* making decisions about creating and suspending spare
870	<	* threads. Updated only by CAS. Note that adding a new worker
871	<	* requires incrementing both counts, since workers start off in
872	<	* running state.  This field is also used for memory-fencing
873	<	* configuration parameters.
874	<	*/
875	<	private volatile int workerCounts;
865	>	/**
866	>	* Pops the given task only if it is at the current top.
867	>	*/
868	>	final boolean tryUnpush(ForkJoinTask<?> t) {
869	>	ForkJoinTask<?>[] a; int s;
870	>	if ((a = array) != null && (s = top) != base &&
871	>	U.compareAndSwapObject
872	>	(a, (((a.length - 1) & --s) << ASHIFT) + ABASE, t, null)) {
873	>	top = s;
874	>	return true;
875	>	}
876	>	return false;
877	>	}
878
879	<	private static final int TOTAL_COUNT_SHIFT  = 16;
880	<	private static final int RUNNING_COUNT_MASK = (1 << TOTAL_COUNT_SHIFT) - 1;
881	<	private static final int ONE_RUNNING        = 1;
882	<	private static final int ONE_TOTAL          = 1 << TOTAL_COUNT_SHIFT;
879	>	/**
880	>	* Polls the given task only if it is at the current base.
881	>	*/
882	>	final boolean pollFor(ForkJoinTask<?> task) {
883	>	ForkJoinTask<?>[] a; int b, i;
884	>	if ((b = base) - top < 0 && (a = array) != null &&
885	>	(i = (a.length - 1) & b) >= 0) {
886	>	int j = (i << ASHIFT) + ABASE;
887	>	if (U.getObjectVolatile(a, j) == task && base == b &&
888	>	U.compareAndSwapObject(a, j, task, null)) {
889	>	base = b + 1;
890	>	return true;
891	>	}
892	>	}
893	>	return false;
894	>	}
895
896	<	/*
897	<	* Fields parallelism. maxPoolSize, and maintainsParallelism are
898	<	* non-volatile, but external reads/writes use workerCount fences
899	<	* to ensure visability.
900	<	*/
896	>	/**
897	>	* If present, removes from queue and executes the given task, or
898	>	* any other cancelled task. Returns (true) immediately on any CAS
899	>	* or consistency check failure so caller can retry.
900	>	*
901	>	* @return false if no progress can be made
902	>	*/
903	>	final boolean tryRemoveAndExec(ForkJoinTask<?> task) {
904	>	boolean removed = false, empty = true, progress = true;
905	>	ForkJoinTask<?>[] a; int m, s, b, n;
906	>	if ((a = array) != null && (m = a.length - 1) >= 0 &&
907	>	(n = (s = top) - (b = base)) > 0) {
908	>	for (ForkJoinTask<?> t;;) {           // traverse from s to b
909	>	int j = ((--s & m) << ASHIFT) + ABASE;
910	>	t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
911	>	if (t == null)                    // inconsistent length
912	>	break;
913	>	else if (t == task) {
914	>	if (s + 1 == top) {           // pop
915	>	if (!U.compareAndSwapObject(a, j, task, null))
916	>	break;
917	>	top = s;
918	>	removed = true;
919	>	}
920	>	else if (base == b)           // replace with proxy
921	>	removed = U.compareAndSwapObject(a, j, task,
922	>	new EmptyTask());
923	>	break;
924	>	}
925	>	else if (t.status >= 0)
926	>	empty = false;
927	>	else if (s + 1 == top) {          // pop and throw away
928	>	if (U.compareAndSwapObject(a, j, t, null))
929	>	top = s;
930	>	break;
931	>	}
932	>	if (--n == 0) {
933	>	if (!empty && base == b)
934	>	progress = false;
935	>	break;
936	>	}
937	>	}
938	>	}
939	>	if (removed)
940	>	task.doExec();
941	>	return progress;
942	>	}
943
944	<	/**
945	<	* The target parallelism level.
946	<	*/
947	<	private int parallelism;
944	>	/**
945	>	* Initializes or doubles the capacity of array. Call either
946	>	* by owner or with lock held -- it is OK for base, but not
947	>	* top, to move while resizings are in progress.
948	>	*
949	>	* @param rejectOnFailure if true, throw exception if capacity
950	>	* exceeded (relayed ultimately to user); else return null.
951	>	*/
952	>	final ForkJoinTask<?>[] growArray(boolean rejectOnFailure) {
953	>	ForkJoinTask<?>[] oldA = array;
954	>	int size = oldA != null ? oldA.length << 1 : INITIAL_QUEUE_CAPACITY;
955	>	if (size <= MAXIMUM_QUEUE_CAPACITY) {
956	>	int oldMask, t, b;
957	>	ForkJoinTask<?>[] a = array = new ForkJoinTask<?>[size];
958	>	if (oldA != null && (oldMask = oldA.length - 1) >= 0 &&
959	>	(t = top) - (b = base) > 0) {
960	>	int mask = size - 1;
961	>	do {
962	>	ForkJoinTask<?> x;
963	>	int oldj = ((b & oldMask) << ASHIFT) + ABASE;
964	>	int j    = ((b &    mask) << ASHIFT) + ABASE;
965	>	x = (ForkJoinTask<?>)U.getObjectVolatile(oldA, oldj);
966	>	if (x != null &&
967	>	U.compareAndSwapObject(oldA, oldj, x, null))
968	>	U.putObjectVolatile(a, j, x);
969	>	} while (++b != t);
970	>	}
971	>	return a;
972	>	}
973	>	else if (!rejectOnFailure)
974	>	return null;
975	>	else
976	>	throw new RejectedExecutionException("Queue capacity exceeded");
977	>	}
978
979	<	/**
980	<	* The maximum allowed pool size.
981	<	*/
982	<	private int maxPoolSize;
979	>	/**
980	>	* Removes and cancels all known tasks, ignoring any exceptions
981	>	*/
982	>	final void cancelAll() {
983	>	ForkJoinTask.cancelIgnoringExceptions(currentJoin);
984	>	ForkJoinTask.cancelIgnoringExceptions(currentSteal);
985	>	for (ForkJoinTask<?> t; (t = poll()) != null; )
986	>	ForkJoinTask.cancelIgnoringExceptions(t);
987	>	}
988
989	<	/**
505	<	* True if use local fifo, not default lifo, for local polling
506	<	* Replicated by ForkJoinWorkerThreads
507	<	*/
508	<	private volatile boolean locallyFifo;
989	>	// Execution methods
990
991	<	/**
992	<	* Controls whether to add spares to maintain parallelism
993	<	*/
994	<	private boolean maintainsParallelism;
991	>	/**
992	>	* Removes and runs tasks until empty, using local mode
993	>	* ordering.
994	>	*/
995	>	final void runLocalTasks() {
996	>	if (base - top < 0) {
997	>	for (ForkJoinTask<?> t; (t = nextLocalTask()) != null; )
998	>	t.doExec();
999	>	}
1000	>	}
1001
1002	<	/**
1003	<	* The uncaught exception handler used when any worker
1004	<	* abruptly terminates
1005	<	*/
1006	<	private volatile Thread.UncaughtExceptionHandler ueh;
1002	>	/**
1003	>	* Executes a top-level task and any local tasks remaining
1004	>	* after execution.
1005	>	*
1006	>	* @return true unless terminating
1007	>	*/
1008	>	final boolean runTask(ForkJoinTask<?> t) {
1009	>	boolean alive = true;
1010	>	if (t != null) {
1011	>	currentSteal = t;
1012	>	t.doExec();
1013	>	runLocalTasks();
1014	>	++nsteals;
1015	>	currentSteal = null;
1016	>	}
1017	>	else if (runState < 0)            // terminating
1018	>	alive = false;
1019	>	return alive;
1020	>	}
1021
1022	<	/**
1023	<	* Pool number, just for assigning useful names to worker threads
1024	<	*/
1025	<	private final int poolNumber;
1022	>	/**
1023	>	* Executes a non-top-level (stolen) task
1024	>	*/
1025	>	final void runSubtask(ForkJoinTask<?> t) {
1026	>	if (t != null) {
1027	>	ForkJoinTask<?> ps = currentSteal;
1028	>	currentSteal = t;
1029	>	t.doExec();
1030	>	currentSteal = ps;
1031	>	}
1032	>	}
1033
1034	<	// utilities for updating fields
1034	>	/**
1035	>	* Computes next value for random probes.  Scans don't require
1036	>	* a very high quality generator, but also not a crummy one.
1037	>	* Marsaglia xor-shift is cheap and works well enough.  Note:
1038	>	* This is manually inlined in several usages in ForkJoinPool
1039	>	* to avoid writes inside busy scan loops.
1040	>	*/
1041	>	final int nextSeed() {
1042	>	int r = seed;
1043	>	r ^= r << 13;
1044	>	r ^= r >>> 17;
1045	>	r ^= r << 5;
1046	>	return seed = r;
1047	>	}
1048
1049	<	/**
1050	<	* Adds delta to running count.  Used mainly by ForkJoinTask.
1051	<	*/
1052	<	final void updateRunningCount(int delta) {
1053	<	int wc;
1054	<	do {} while (!UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1055	<	wc = workerCounts,
1056	<	wc + delta));
1049	>	// Unsafe mechanics
1050	>	private static final sun.misc.Unsafe U;
1051	>	private static final long RUNSTATE;
1052	>	private static final int ABASE;
1053	>	private static final int ASHIFT;
1054	>	static {
1055	>	int s;
1056	>	try {
1057	>	U = getUnsafe();
1058	>	Class<?> k = WorkQueue.class;
1059	>	Class<?> ak = ForkJoinTask[].class;
1060	>	RUNSTATE = U.objectFieldOffset
1061	>	(k.getDeclaredField("runState"));
1062	>	ABASE = U.arrayBaseOffset(ak);
1063	>	s = U.arrayIndexScale(ak);
1064	>	} catch (Exception e) {
1065	>	throw new Error(e);
1066	>	}
1067	>	if ((s & (s-1)) != 0)
1068	>	throw new Error("data type scale not a power of two");
1069	>	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
1070	>	}
1071		}
1072
1073		/**
1074	<	* Decrements running count unless already zero
1074	>	* Class for artificial tasks that are used to replace the target
1075	>	* of local joins if they are removed from an interior queue slot
1076	>	* in WorkQueue.tryRemoveAndExec. We don't need the proxy to
1077	>	* actually do anything beyond having a unique identity.
1078		*/
1079	<	final boolean tryDecrementRunningCount() {
1080	<	int wc = workerCounts;
1081	<	if ((wc & RUNNING_COUNT_MASK) == 0)
1082	<	return false;
1083	<	return UNSAFE.compareAndSwapInt(this, workerCountsOffset,
546	<	wc, wc - ONE_RUNNING);
1079	>	static final class EmptyTask extends ForkJoinTask<Void> {
1080	>	EmptyTask() { status = ForkJoinTask.NORMAL; } // force done
1081	>	public Void getRawResult() { return null; }
1082	>	public void setRawResult(Void x) {}
1083	>	public boolean exec() { return true; }
1084		}
1085
1086		/**
1087	<	* Write fence for user modifications of pool parameters
1088	<	* (parallelism. etc).  Note that it doesn't matter if CAS fails.
1087	>	* Computes a hash code for the given thread. This method is
1088	>	* expected to provide higher-quality hash codes than those using
1089	>	* method hashCode().
1090		*/
1091	<	private void workerCountWriteFence() {
1092	<	int wc;
1093	<	UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1094	<	wc = workerCounts, wc);
1091	>	static final int hashThread(Thread t) {
1092	>	long id = (t == null)? 0L : t.getId(); // Use MurmurHash of thread id
1093	>	int h = (int)id ^ (int)(id >>> 32);
1094	>	h ^= h >>> 16;
1095	>	h *= 0x85ebca6b;
1096	>	h ^= h >>> 13;
1097	>	h *= 0xc2b2ae35;
1098	>	return h ^ (h >>> 16);
1099		}
1100
1101		/**
1102	<	* Read fence for external reads of pool parameters
561	<	* (parallelism. maxPoolSize, etc).
1102	>	* Top-level runloop for workers
1103		*/
1104	<	private void workerCountReadFence() {
1105	<	int ignore = workerCounts;
1106	<	}
1104	>	final void runWorker(ForkJoinWorkerThread wt) {
1105	>	WorkQueue w = wt.workQueue;
1106	>	w.growArray(false);     // Initialize queue array and seed in this thread
1107	>	w.seed = hashThread(Thread.currentThread()) | (1 << 31); // force < 0
1108
1109	<	/**
568	<	* Tries incrementing active count; fails on contention.
569	<	* Called by workers before executing tasks.
570	<	*
571	<	* @return true on success
572	<	*/
573	<	final boolean tryIncrementActiveCount() {
574	<	int c;
575	<	return UNSAFE.compareAndSwapInt(this, runStateOffset,
576	<	c = runState, c + ONE_ACTIVE);
1109	>	do {} while (w.runTask(scan(w)));
1110		}
1111
1112	<	/**
580	<	* Tries decrementing active count; fails on contention.
581	<	* Called when workers cannot find tasks to run.
582	<	*/
583	<	final boolean tryDecrementActiveCount() {
584	<	int c;
585	<	return UNSAFE.compareAndSwapInt(this, runStateOffset,
586	<	c = runState, c - ONE_ACTIVE);
587	<	}
1112	>	// Creating, registering and deregistering workers
1113
1114		/**
1115	<	* Advances to at least the given level. Returns true if not
591	<	* already in at least the given level.
1115	>	* Tries to create and start a worker
1116		*/
1117	<	private boolean advanceRunLevel(int level) {
1118	<	for (;;) {
1119	<	int s = runState;
1120	<	if ((s & level) != 0)
1121	<	return false;
1122	<	if (UNSAFE.compareAndSwapInt(this, runStateOffset, s, s | level))
1123	<	return true;
1117	>	private void addWorker() {
1118	>	Throwable ex = null;
1119	>	ForkJoinWorkerThread w = null;
1120	>	try {
1121	>	if ((w = factory.newThread(this)) != null) {
1122	>	w.start();
1123	>	return;
1124	>	}
1125	>	} catch (Throwable e) {
1126	>	ex = e;
1127		}
1128	+	deregisterWorker(w, ex);
1129		}
1130
603	–	// workers array maintenance
604	–
1131		/**
1132	<	* Records and returns a workers array index for new worker.
1132	>	* Callback from ForkJoinWorkerThread constructor to assign a
1133	>	* public name. This must be separate from registerWorker because
1134	>	* it is called during the "super" constructor call in
1135	>	* ForkJoinWorkerThread.
1136		*/
1137	<	private int recordWorker(ForkJoinWorkerThread w) {
1138	<	// Try using slot totalCount-1. If not available, scan and/or resize
1139	<	int k = (workerCounts >>> TOTAL_COUNT_SHIFT) - 1;
611	<	final ReentrantLock lock = this.workerLock;
612	<	lock.lock();
613	<	try {
614	<	ForkJoinWorkerThread[] ws = workers;
615	<	int nws = ws.length;
616	<	if (k < 0 || k >= nws || ws[k] != null) {
617	<	for (k = 0; k < nws && ws[k] != null; ++k)
618	<	;
619	<	if (k == nws)
620	<	ws = Arrays.copyOf(ws, nws << 1);
621	<	}
622	<	ws[k] = w;
623	<	workers = ws; // volatile array write ensures slot visibility
624	<	} finally {
625	<	lock.unlock();
626	<	}
627	<	return k;
1137	>	final String nextWorkerName() {
1138	>	return workerNamePrefix.concat
1139	>	(Integer.toString(nextWorkerNumber.addAndGet(1)));
1140		}
1141
1142		/**
1143	<	* Nulls out record of worker in workers array
1143	>	* Callback from ForkJoinWorkerThread constructor to establish and
1144	>	* record its WorkQueue
1145	>	*
1146	>	* @param wt the worker thread
1147		*/
1148	<	private void forgetWorker(ForkJoinWorkerThread w) {
1149	<	int idx = w.poolIndex;
1150	<	// Locking helps method recordWorker avoid unecessary expansion
636	<	final ReentrantLock lock = this.workerLock;
1148	>	final void registerWorker(ForkJoinWorkerThread wt) {
1149	>	WorkQueue w = wt.workQueue;
1150	>	ReentrantLock lock = this.lock;
1151		lock.lock();
1152		try {
1153	<	ForkJoinWorkerThread[] ws = workers;
1154	<	if (idx >= 0 && idx < ws.length && ws[idx] == w) // verify
1155	<	ws[idx] = null;
1153	>	int k = nextPoolIndex;
1154	>	WorkQueue[] ws = workQueues;
1155	>	if (ws != null) {                       // ignore on shutdown
1156	>	int n = ws.length;
1157	>	if (k < 0 || (k & 1) == 0 || k >= n || ws[k] != null) {
1158	>	for (k = 1; k < n && ws[k] != null; k += 2)
1159	>	;                           // workers are at odd indices
1160	>	if (k >= n)                     // resize
1161	>	workQueues = ws = Arrays.copyOf(ws, n << 1);
1162	>	}
1163	>	w.poolIndex = k;
1164	>	w.eventCount = ~(k >>> 1) & SMASK;  // Set up wait count
1165	>	ws[k] = w;                          // record worker
1166	>	nextPoolIndex = k + 2;
1167	>	int rs = runState;
1168	>	int m = rs & SMASK;                 // recalculate runState mask
1169	>	if (k > m)
1170	>	m = (m << 1) + 1;
1171	>	runState = (rs & SHUTDOWN) | ((rs + RS_SEQ) & RS_SEQ_MASK) | m;
1172	>	}
1173		} finally {
1174		lock.unlock();
1175		}
1176		}
1177
647	–	// adding and removing workers
648	–
1178		/**
1179	<	* Tries to create and add new worker. Assumes that worker counts
1180	<	* are already updated to accommodate the worker, so adjusts on
1181	<	* failure.
1179	>	* Final callback from terminating worker, as well as failure to
1180	>	* construct or start a worker in addWorker.  Removes record of
1181	>	* worker from array, and adjusts counts. If pool is shutting
1182	>	* down, tries to complete termination.
1183		*
1184	<	* @return new worker or null if creation failed
1184	>	* @param wt the worker thread or null if addWorker failed
1185	>	* @param ex the exception causing failure, or null if none
1186		*/
1187	<	private ForkJoinWorkerThread addWorker() {
1188	<	ForkJoinWorkerThread w = null;
1189	<	try {
1190	<	w = factory.newThread(this);
1191	<	} finally { // Adjust on either null or exceptional factory return
1192	<	if (w == null) {
1193	<	onWorkerCreationFailure();
1194	<	return null;
1187	>	final void deregisterWorker(ForkJoinWorkerThread wt, Throwable ex) {
1188	>	WorkQueue w = null;
1189	>	if (wt != null && (w = wt.workQueue) != null) {
1190	>	w.runState = -1;                // ensure runState is set
1191	>	stealCount.getAndAdd(w.totalSteals + w.nsteals);
1192	>	int idx = w.poolIndex;
1193	>	ReentrantLock lock = this.lock;
1194	>	lock.lock();
1195	>	try {                           // remove record from array
1196	>	WorkQueue[] ws = workQueues;
1197	>	if (ws != null && idx >= 0 && idx < ws.length && ws[idx] == w)
1198	>	ws[nextPoolIndex = idx] = null;
1199	>	} finally {
1200	>	lock.unlock();
1201		}
1202		}
666	–	w.start(recordWorker(w), locallyFifo, ueh);
667	–	return w;
668	–	}
1203
1204	<	/**
1205	<	* Adjusts counts upon failure to create worker
1206	<	*/
1207	<	private void onWorkerCreationFailure() {
1208	<	for (;;) {
675	<	int wc = workerCounts;
676	<	if ((wc >>> TOTAL_COUNT_SHIFT) > 0 &&
677	<	UNSAFE.compareAndSwapInt(this, workerCountsOffset,
678	<	wc, wc - (ONE_RUNNING|ONE_TOTAL)))
679	<	break;
680	<	}
681	<	tryTerminate(false); // in case of failure during shutdown
682	<	}
1204	>	long c;                             // adjust ctl counts
1205	>	do {} while (!U.compareAndSwapLong
1206	>	(this, CTL, c = ctl, (((c - AC_UNIT) & AC_MASK) |
1207	>	((c - TC_UNIT) & TC_MASK) |
1208	>	(c & ~(AC_MASK|TC_MASK)))));
1209
1210	<	/**
1211	<	* Create enough total workers to establish target parallelism,
1212	<	* giving up if terminating or addWorker fails
1213	<	*/
688	<	private void ensureEnoughTotalWorkers() {
689	<	int wc;
690	<	while (((wc = workerCounts) >>> TOTAL_COUNT_SHIFT) < parallelism &&
691	<	runState < TERMINATING) {
692	<	if ((UNSAFE.compareAndSwapInt(this, workerCountsOffset,
693	<	wc, wc + (ONE_RUNNING|ONE_TOTAL)) &&
694	<	addWorker() == null))
695	<	break;
1210	>	if (!tryTerminate(false) && w != null) {
1211	>	w.cancelAll();                  // cancel remaining tasks
1212	>	if (w.array != null)            // suppress signal if never ran
1213	>	signalWork();               // wake up or create replacement
1214		}
697	–	}
1215
1216	<	/**
1217	<	* Final callback from terminating worker.  Removes record of
701	<	* worker from array, and adjusts counts. If pool is shutting
702	<	* down, tries to complete terminatation, else possibly replaces
703	<	* the worker.
704	<	*
705	<	* @param w the worker
706	<	*/
707	<	final void workerTerminated(ForkJoinWorkerThread w) {
708	<	if (w.active) { // force inactive
709	<	w.active = false;
710	<	do {} while (!tryDecrementActiveCount());
711	<	}
712	<	forgetWorker(w);
713	<
714	<	// Decrement total count, and if was running, running count
715	<	// Spin (waiting for other updates) if either would be negative
716	<	int nr = w.isTrimmed() ? 0 : ONE_RUNNING;
717	<	int unit = ONE_TOTAL + nr;
718	<	for (;;) {
719	<	int wc = workerCounts;
720	<	int rc = wc & RUNNING_COUNT_MASK;
721	<	if (rc - nr < 0 || (wc >>> TOTAL_COUNT_SHIFT) == 0)
722	<	Thread.yield(); // back off if waiting for other updates
723	<	else if (UNSAFE.compareAndSwapInt(this, workerCountsOffset,
724	<	wc, wc - unit))
725	<	break;
726	<	}
727	<
728	<	accumulateStealCount(w); // collect final count
729	<	if (!tryTerminate(false))
730	<	ensureEnoughTotalWorkers();
1216	>	if (ex != null)                     // rethrow
1217	>	U.throwException(ex);
1218		}
1219
733	–	// Waiting for and signalling events
1220
1221	<	/**
736	<	* Ensures eventCount on exit is different (mod 2^32) than on
737	<	* entry.  CAS failures are OK -- any change in count suffices.
738	<	*/
739	<	private void advanceEventCount() {
740	<	int c;
741	<	UNSAFE.compareAndSwapInt(this, eventCountOffset, c = eventCount, c+1);
742	<	}
1221	>	// Maintaining ctl counts
1222
1223		/**
1224	<	* Releases workers blocked on a count not equal to current count.
1224	>	* Increments active count; mainly called upon return from blocking
1225		*/
1226	<	final void releaseWaiters() {
1227	<	long top;
1228	<	int id;
750	<	while ((id = (int)((top = eventWaiters) & WAITER_INDEX_MASK)) > 0 &&
751	<	(int)(top >>> EVENT_COUNT_SHIFT) != eventCount) {
752	<	ForkJoinWorkerThread[] ws = workers;
753	<	ForkJoinWorkerThread w;
754	<	if (ws.length >= id && (w = ws[id - 1]) != null &&
755	<	UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
756	<	top, w.nextWaiter))
757	<	LockSupport.unpark(w);
758	<	}
1226	>	final void incrementActiveCount() {
1227	>	long c;
1228	>	do {} while (!U.compareAndSwapLong(this, CTL, c = ctl, c + AC_UNIT));
1229		}
1230
1231		/**
1232	<	* Advances eventCount and releases waiters until interference by
763	<	* other releasing threads is detected.
1232	>	* Activates or creates a worker
1233		*/
1234		final void signalWork() {
1235	<	int ec;
1236	<	UNSAFE.compareAndSwapInt(this, eventCountOffset, ec=eventCount, ec+1);
1237	<	outer:for (;;) {
1238	<	long top = eventWaiters;
1239	<	ec = eventCount;
1240	<	for (;;) {
1241	<	ForkJoinWorkerThread[] ws; ForkJoinWorkerThread w;
1242	<	int id = (int)(top & WAITER_INDEX_MASK);
1243	<	if (id <= 0 || (int)(top >>> EVENT_COUNT_SHIFT) == ec)
1244	<	return;
1245	<	if ((ws = workers).length < id || (w = ws[id - 1]) == null ||
1246	<	!UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
1247	<	top, top = w.nextWaiter))
1248	<	continue outer;      // possibly stale; reread
1249	<	LockSupport.unpark(w);
1250	<	if (top != eventWaiters) // let someone else take over
1251	<	return;
1235	>	/*
1236	>	* The while condition is true if: (there is are too few total
1237	>	* workers OR there is at least one waiter) AND (there are too
1238	>	* few active workers OR the pool is terminating).  The value
1239	>	* of e distinguishes the remaining cases: zero (no waiters)
1240	>	* for create, negative if terminating (in which case do
1241	>	* nothing), else release a waiter. The secondary checks for
1242	>	* release (non-null array etc) can fail if the pool begins
1243	>	* terminating after the test, and don't impose any added cost
1244	>	* because JVMs must perform null and bounds checks anyway.
1245	>	*/
1246	>	long c; int e, u;
1247	>	while ((((e = (int)(c = ctl)) | (u = (int)(c >>> 32))) &
1248	>	(INT_SIGN|SHORT_SIGN)) == (INT_SIGN|SHORT_SIGN)) {
1249	>	WorkQueue[] ws = workQueues; int i; WorkQueue w; Thread p;
1250	>	if (e == 0) {                    // add a new worker
1251	>	if (U.compareAndSwapLong
1252	>	(this, CTL, c, (long)(((u + UTC_UNIT) & UTC_MASK) |
1253	>	((u + UAC_UNIT) & UAC_MASK)) << 32)) {
1254	>	addWorker();
1255	>	break;
1256	>	}
1257		}
1258	+	else if (e > 0 && ws != null &&
1259	+	(i = ((~e << 1) | 1) & SMASK) < ws.length &&
1260	+	(w = ws[i]) != null &&
1261	+	w.eventCount == (e | INT_SIGN)) {
1262	+	if (U.compareAndSwapLong
1263	+	(this, CTL, c, (((long)(w.nextWait & E_MASK)) |
1264	+	((long)(u + UAC_UNIT) << 32)))) {
1265	+	w.eventCount = (e + E_SEQ) & E_MASK;
1266	+	if ((p = w.parker) != null)
1267	+	U.unpark(p);         // release a waiting worker
1268	+	break;
1269	+	}
1270	+	}
1271	+	else
1272	+	break;
1273		}
1274		}
1275
1276		/**
1277	<	* If worker is inactive, blocks until terminating or event count
1278	<	* advances from last value held by worker; in any case helps
1279	<	* release others.
1280	<	*
1281	<	* @param w the calling worker thread
1282	<	*/
1283	<	private void eventSync(ForkJoinWorkerThread w) {
1284	<	if (!w.active) {
1285	<	int prev = w.lastEventCount;
1286	<	long nextTop = (((long)prev << EVENT_COUNT_SHIFT) |
1287	<	((long)(w.poolIndex + 1)));
1288	<	long top;
1289	<	while ((runState < SHUTDOWN || !tryTerminate(false)) &&
1290	<	(((int)(top = eventWaiters) & WAITER_INDEX_MASK) == 0 ||
1291	<	(int)(top >>> EVENT_COUNT_SHIFT) == prev) &&
1292	<	eventCount == prev) {
1293	<	if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
1294	<	w.nextWaiter = top, nextTop)) {
1295	<	accumulateStealCount(w); // transfer steals while idle
1296	<	Thread.interrupted();    // clear/ignore interrupt
1297	<	while (eventCount == prev)
1298	<	w.doPark();
1299	<	break;
1277	>	* Tries to decrement active count (sometimes implicitly) and
1278	>	* possibly release or create a compensating worker in preparation
1279	>	* for blocking. Fails on contention or termination.
1280	>	*
1281	>	* @return true if the caller can block, else should recheck and retry
1282	>	*/
1283	>	final boolean tryCompensate() {
1284	>	WorkQueue[] ws; WorkQueue w; Thread p;
1285	>	int pc = parallelism, e, u, ac, tc, i;
1286	>	long c = ctl;
1287	>
1288	>	if ((e = (int)c) >= 0) {
1289	>	if ((ac = ((u = (int)(c >>> 32)) >> UAC_SHIFT)) <= 0 &&
1290	>	e != 0 && (ws = workQueues) != null &&
1291	>	(i = ((~e << 1) | 1) & SMASK) < ws.length &&
1292	>	(w = ws[i]) != null) {
1293	>	if (w.eventCount == (e | INT_SIGN) &&
1294	>	U.compareAndSwapLong
1295	>	(this, CTL, c, ((long)(w.nextWait & E_MASK) |
1296	>	(c & (AC_MASK|TC_MASK))))) {
1297	>	w.eventCount = (e + E_SEQ) & E_MASK;
1298	>	if ((p = w.parker) != null)
1299	>	U.unpark(p);
1300	>	return true;             // release an idle worker
1301	>	}
1302	>	}
1303	>	else if ((tc = (short)(u >>> UTC_SHIFT)) >= 0 && ac + pc > 1) {
1304	>	long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK);
1305	>	if (U.compareAndSwapLong(this, CTL, c, nc))
1306	>	return true;             // no compensation needed
1307	>	}
1308	>	else if (tc + pc < MAX_ID) {
1309	>	long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK);
1310	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1311	>	addWorker();
1312	>	return true;             // create replacement
1313		}
1314		}
813	–	w.lastEventCount = eventCount;
1315		}
1316	<	releaseWaiters();
1316	>	return false;
1317		}
1318
1319	+	// Submissions
1320	+
1321		/**
1322	<	* Callback from workers invoked upon each top-level action (i.e.,
1323	<	* stealing a task or taking a submission and running
1324	<	* it). Performs one or both of the following:
1325	<	*
1326	<	* * If the worker cannot find work, updates its active status to
1327	<	* inactive and updates activeCount unless there is contention, in
825	<	* which case it may try again (either in this or a subsequent
826	<	* call).  Additionally, awaits the next task event and/or helps
827	<	* wake up other releasable waiters.
828	<	*
829	<	* * If there are too many running threads, suspends this worker
830	<	* (first forcing inactivation if necessary).  If it is not
831	<	* resumed before a keepAlive elapses, the worker may be "trimmed"
832	<	* -- killed while suspended within suspendAsSpare. Otherwise,
833	<	* upon resume it rechecks to make sure that it is still needed.
834	<	*
835	<	* @param w the worker
836	<	* @param worked false if the worker scanned for work but didn't
837	<	* find any (in which case it may block waiting for work).
1322	>	* Unless shutting down, adds the given task to some submission
1323	>	* queue; using a randomly chosen queue index if the caller is a
1324	>	* ForkJoinWorkerThread, else one based on caller thread's hash
1325	>	* code. If no queue exists at the index, one is created.  If the
1326	>	* queue is busy, another is chosen by sweeping through the queues
1327	>	* array.
1328		*/
1329	<	final void preStep(ForkJoinWorkerThread w, boolean worked) {
1330	<	boolean active = w.active;
1331	<	boolean inactivate = !worked & active;
1329	>	private void doSubmit(ForkJoinTask<?> task) {
1330	>	if (task == null)
1331	>	throw new NullPointerException();
1332	>	Thread t = Thread.currentThread();
1333	>	int r = ((t instanceof ForkJoinWorkerThread) ?
1334	>	((ForkJoinWorkerThread)t).workQueue.nextSeed() : hashThread(t));
1335		for (;;) {
1336	<	if (inactivate) {
1337	<	int c = runState;
1338	<	if (UNSAFE.compareAndSwapInt(this, runStateOffset,
1339	<	c, c - ONE_ACTIVE))
1340	<	inactivate = active = w.active = false;
1341	<	}
1342	<	int wc = workerCounts;
1343	<	if ((wc & RUNNING_COUNT_MASK) <= parallelism) {
1344	<	if (!worked)
1345	<	eventSync(w);
1346	<	return;
1336	>	int rs = runState, m = rs & SMASK;
1337	>	int j = r &= (m & ~1);                      // even numbered queues
1338	>	WorkQueue[] ws = workQueues;
1339	>	if (rs < 0 || ws == null)
1340	>	throw new RejectedExecutionException(); // shutting down
1341	>	if (ws.length > m) {                        // consistency check
1342	>	for (WorkQueue q;;) {                   // circular sweep
1343	>	if (((q = ws[j]) != null ||
1344	>	(q = tryAddSharedQueue(j)) != null) &&
1345	>	q.trySharedPush(task)) {
1346	>	signalWork();
1347	>	return;
1348	>	}
1349	>	if ((j = (j + 2) & m) == r) {
1350	>	Thread.yield();                 // all queues busy
1351	>	break;
1352	>	}
1353	>	}
1354		}
855	–	if (!(inactivate |= active) &&  // must inactivate to suspend
856	–	UNSAFE.compareAndSwapInt(this, workerCountsOffset,
857	–	wc, wc - ONE_RUNNING) &&
858	–	!w.suspendAsSpare())        // false if trimmed
859	–	return;
1355		}
1356		}
1357
1358		/**
1359	<	* Adjusts counts and creates or resumes compensating threads for
865	<	* a worker blocking on task joinMe.  First tries resuming an
866	<	* existing spare (which usually also avoids any count
867	<	* adjustment), but must then decrement running count to determine
868	<	* whether a new thread is needed. See above for fuller
869	<	* explanation. This code is sprawled out non-modularly mainly
870	<	* because adaptive spinning works best if the entire method is
871	<	* either interpreted or compiled vs having only some pieces of it
872	<	* compiled.
1359	>	* Tries to add and register a new queue at the given index.
1360		*
1361	<	* @param joinMe the task to join
1362	<	* @return task status on exit (to simplify usage by callers)
1363	<	*/
1364	<	final int awaitJoin(ForkJoinTask<?> joinMe) {
1365	<	int pc = parallelism;
1366	<	boolean adj = false;        // true when running count adjusted
1367	<	int scans = 0;
1368	<
1369	<	while (joinMe.status >= 0) {
1370	<	ForkJoinWorkerThread spare = null;
1371	<	if ((workerCounts & RUNNING_COUNT_MASK) < pc) {
1372	<	ForkJoinWorkerThread[] ws = workers;
1373	<	int nws = ws.length;
1374	<	for (int i = 0; i < nws; ++i) {
1375	<	ForkJoinWorkerThread w = ws[i];
1376	<	if (w != null && w.isSuspended()) {
1377	<	spare = w;
1378	<	break;
1361	>	* @param idx the workQueues array index to register the queue
1362	>	* @return the queue, or null if could not add because could
1363	>	* not acquire lock or idx is unusable
1364	>	*/
1365	>	private WorkQueue tryAddSharedQueue(int idx) {
1366	>	WorkQueue q = null;
1367	>	ReentrantLock lock = this.lock;
1368	>	if (idx >= 0 && (idx & 1) == 0 && !lock.isLocked()) {
1369	>	// create queue outside of lock but only if apparently free
1370	>	WorkQueue nq = new WorkQueue(null, SHARED_QUEUE);
1371	>	if (lock.tryLock()) {
1372	>	try {
1373	>	WorkQueue[] ws = workQueues;
1374	>	if (ws != null && idx < ws.length) {
1375	>	if ((q = ws[idx]) == null) {
1376	>	int rs;         // update runState seq
1377	>	ws[idx] = q = nq;
1378	>	runState = (((rs = runState) & SHUTDOWN) |
1379	>	((rs + RS_SEQ) & ~SHUTDOWN));
1380	>	}
1381		}
1382	+	} finally {
1383	+	lock.unlock();
1384		}
894	–	if (joinMe.status < 0)
895	–	break;
1385		}
1386	<	int wc = workerCounts;
1387	<	int rc = wc & RUNNING_COUNT_MASK;
1388	<	int dc = pc - rc;
1389	<	if (dc > 0 && spare != null && spare.tryUnsuspend()) {
1390	<	if (adj) {
1391	<	int c;
1392	<	do {} while (!UNSAFE.compareAndSwapInt
1393	<	(this, workerCountsOffset,
1394	<	c = workerCounts, c + ONE_RUNNING));
1386	>	}
1387	>	return q;
1388	>	}
1389	>
1390	>	// Scanning for tasks
1391	>
1392	>	/**
1393	>	* Scans for and, if found, returns one task, else possibly
1394	>	* inactivates the worker. This method operates on single reads of
1395	>	* volatile state and is designed to be re-invoked continuously in
1396	>	* part because it returns upon detecting inconsistencies,
1397	>	* contention, or state changes that indicate possible success on
1398	>	* re-invocation.
1399	>	*
1400	>	* The scan searches for tasks across queues, randomly selecting
1401	>	* the first #queues probes, favoring steals 2:1 over submissions
1402	>	* (by exploiting even/odd indexing), and then performing a
1403	>	* circular sweep of all queues.  The scan terminates upon either
1404	>	* finding a non-empty queue, or completing a full sweep. If the
1405	>	* worker is not inactivated, it takes and returns a task from
1406	>	* this queue.  On failure to find a task, we take one of the
1407	>	* following actions, after which the caller will retry calling
1408	>	* this method unless terminated.
1409	>	*
1410	>	* * If not a complete sweep, try to release a waiting worker.  If
1411	>	* the scan terminated because the worker is inactivated, then the
1412	>	* released worker will often be the calling worker, and it can
1413	>	* succeed obtaining a task on the next call. Or maybe it is
1414	>	* another worker, but with same net effect. Releasing in other
1415	>	* cases as well ensures that we have enough workers running.
1416	>	*
1417	>	* * If the caller has run a task since the the last empty scan,
1418	>	* return (to allow rescan) if other workers are not also yet
1419	>	* enqueued.  Field WorkQueue.rescans counts down on each scan to
1420	>	* ensure eventual inactivation, and occasional calls to
1421	>	* Thread.yield to help avoid interference with more useful
1422	>	* activities on the system.
1423	>	*
1424	>	* * If pool is terminating, terminate the worker
1425	>	*
1426	>	* * If not already enqueued, try to inactivate and enqueue the
1427	>	* worker on wait queue.
1428	>	*
1429	>	* * If already enqueued and none of the above apply, either park
1430	>	* awaiting signal, or if this is the most recent waiter and pool
1431	>	* is quiescent, relay to idleAwaitWork to check for termination
1432	>	* and possibly shrink pool.
1433	>	*
1434	>	* @param w the worker (via its WorkQueue)
1435	>	* @return a task or null of none found
1436	>	*/
1437	>	private final ForkJoinTask<?> scan(WorkQueue w) {
1438	>	boolean swept = false;                 // true after full empty scan
1439	>	WorkQueue[] ws;                        // volatile read order matters
1440	>	int r = w.seed, ec = w.eventCount;     // ec is negative if inactive
1441	>	int rs = runState, m = rs & SMASK;
1442	>	if ((ws = workQueues) != null && ws.length > m) {
1443	>	ForkJoinTask<?> task = null;
1444	>	for (int k = 0, j = -2 - m; ; ++j) {
1445	>	WorkQueue q; int b;
1446	>	if (j < 0) {                    // random probes while j negative
1447	>	r ^= r << 13; r ^= r >>> 17; k = (r ^= r << 5) | (j & 1);
1448	>	}                               // worker (not submit) for odd j
1449	>	else                            // cyclic scan when j >= 0
1450	>	k += (m >>> 1) | 1;         // step by half to reduce bias
1451	>
1452	>	if ((q = ws[k & m]) != null && (b = q.base) - q.top < 0) {
1453	>	if (ec >= 0)
1454	>	task = q.pollAt(b);     // steal
1455	>	break;
1456		}
1457	<	adj = true;
1458	<	LockSupport.unpark(spare);
1459	<	}
910	<	else if (adj) {
911	<	if (dc <= 0)
1457	>	else if (j > m) {
1458	>	if (rs == runState)        // staleness check
1459	>	swept = true;
1460		break;
913	–	int tc = wc >>> TOTAL_COUNT_SHIFT;
914	–	if (scans > tc) {
915	–	int ts = (tc - pc) * pc;
916	–	if (rc != 0 &&  (dc * dc < ts || !maintainsParallelism))
917	–	break;
918	–	if (scans > ts && tc < maxPoolSize &&
919	–	UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
920	–	wc+(ONE_RUNNING|ONE_TOTAL))){
921	–	addWorker();
922	–	break;
923	–	}
1461		}
1462		}
1463	<	else if (rc != 0)
1464	<	adj = UNSAFE.compareAndSwapInt (this, workerCountsOffset,
1465	<	wc, wc - ONE_RUNNING);
1466	<	if ((scans++ & 1) == 0)
1467	<	releaseWaiters();   // help others progress
1468	<	else
1469	<	Thread.yield();     // avoid starving productive threads
1463	>	w.seed = r;                        // save seed for next scan
1464	>	if (task != null)
1465	>	return task;
1466	>	}
1467	>
1468	>	// Decode ctl on empty scan
1469	>	long c = ctl; int e = (int)c, a = (int)(c >> AC_SHIFT), nr, ns;
1470	>	if (!swept) {                          // try to release a waiter
1471	>	WorkQueue v; Thread p;
1472	>	if (e > 0 && a < 0 && ws != null &&
1473	>	(v = ws[((~e << 1) | 1) & m]) != null &&
1474	>	v.eventCount == (e | INT_SIGN) && U.compareAndSwapLong
1475	>	(this, CTL, c, ((long)(v.nextWait & E_MASK) |
1476	>	((c + AC_UNIT) & (AC_MASK|TC_MASK))))) {
1477	>	v.eventCount = (e + E_SEQ) & E_MASK;
1478	>	if ((p = v.parker) != null)
1479	>	U.unpark(p);
1480	>	}
1481		}
1482	<
1483	<	if (adj) {
1484	<	joinMe.internalAwaitDone();
1485	<	int c;
1486	<	do {} while (!UNSAFE.compareAndSwapInt
1487	<	(this, workerCountsOffset,
1488	<	c = workerCounts, c + ONE_RUNNING));
1482	>	else if ((nr = w.rescans) > 0) {       // continue rescanning
1483	>	int ac = a + parallelism;
1484	>	if ((w.rescans = (ac < nr) ? ac : nr - 1) > 0 && w.seed < 0 &&
1485	>	w.eventCount == ec)
1486	>	Thread.yield();                // 1 bit randomness for yield call
1487	>	}
1488	>	else if (e < 0)                        // pool is terminating
1489	>	w.runState = -1;
1490	>	else if (ec >= 0) {                    // try to enqueue
1491	>	long nc = (long)ec | ((c - AC_UNIT) & (AC_MASK|TC_MASK));
1492	>	w.nextWait = e;
1493	>	w.eventCount = ec | INT_SIGN;      // mark as inactive
1494	>	if (!U.compareAndSwapLong(this, CTL, c, nc))
1495	>	w.eventCount = ec;             // back out on CAS failure
1496	>	else if ((ns = w.nsteals) != 0) {  // set rescans if ran task
1497	>	if (a <= 0)                    // ... unless too many active
1498	>	w.rescans = a + parallelism;
1499	>	w.nsteals = 0;
1500	>	w.totalSteals += ns;
1501	>	}
1502	>	}
1503	>	else{                                  // already queued
1504	>	if (parallelism == -a)
1505	>	idleAwaitWork(w);              // quiescent
1506	>	if (w.eventCount == ec) {
1507	>	Thread.interrupted();          // clear status
1508	>	ForkJoinWorkerThread wt = w.owner;
1509	>	U.putObject(wt, PARKBLOCKER, this);
1510	>	w.parker = wt;                 // emulate LockSupport.park
1511	>	if (w.eventCount == ec)        // recheck
1512	>	U.park(false, 0L);         // block
1513	>	w.parker = null;
1514	>	U.putObject(wt, PARKBLOCKER, null);
1515	>	}
1516		}
1517	<	return joinMe.status;
1517	>	return null;
1518		}
1519
1520		/**
1521	<	* Same idea as awaitJoin
1522	<	*/
1523	<	final void awaitBlocker(ManagedBlocker blocker, boolean maintainPar)
1524	<	throws InterruptedException {
1525	<	maintainPar &= maintainsParallelism;
1526	<	int pc = parallelism;
1527	<	boolean adj = false;        // true when running count adjusted
1528	<	int scans = 0;
1529	<	boolean done;
1530	<
1531	<	for (;;) {
1532	<	if (done = blocker.isReleasable())
1533	<	break;
1534	<	ForkJoinWorkerThread spare = null;
1535	<	if ((workerCounts & RUNNING_COUNT_MASK) < pc) {
1536	<	ForkJoinWorkerThread[] ws = workers;
1537	<	int nws = ws.length;
1538	<	for (int i = 0; i < nws; ++i) {
1539	<	ForkJoinWorkerThread w = ws[i];
1540	<	if (w != null && w.isSuspended()) {
1541	<	spare = w;
1542	<	break;
1543	<	}
1544	<	}
1545	<	if (done = blocker.isReleasable())
1521	>	* If inactivating worker w has caused pool to become quiescent,
1522	>	* check for pool termination, and, so long as this is not the
1523	>	* only worker, wait for event for up to SHRINK_RATE nanosecs On
1524	>	* timeout, if ctl has not changed, terminate the worker, which
1525	>	* will in turn wake up another worker to possibly repeat this
1526	>	* process.
1527	>	*
1528	>	* @param w the calling worker
1529	>	*/
1530	>	private void idleAwaitWork(WorkQueue w) {
1531	>	long c; int nw, ec;
1532	>	if (!tryTerminate(false) &&
1533	>	(int)((c = ctl) >> AC_SHIFT) + parallelism == 0 &&
1534	>	(ec = w.eventCount) == ((int)c | INT_SIGN) &&
1535	>	(nw = w.nextWait) != 0) {
1536	>	long nc = ((long)(nw & E_MASK) | // ctl to restore on timeout
1537	>	((c + AC_UNIT) & AC_MASK) | (c & TC_MASK));
1538	>	ForkJoinTask.helpExpungeStaleExceptions(); // help clean
1539	>	ForkJoinWorkerThread wt = w.owner;
1540	>	while (ctl == c) {
1541	>	long startTime = System.nanoTime();
1542	>	Thread.interrupted();  // timed variant of version in scan()
1543	>	U.putObject(wt, PARKBLOCKER, this);
1544	>	w.parker = wt;
1545	>	if (ctl == c)
1546	>	U.park(false, SHRINK_RATE);
1547	>	w.parker = null;
1548	>	U.putObject(wt, PARKBLOCKER, null);
1549	>	if (ctl != c)
1550		break;
1551	<	}
1552	<	int wc = workerCounts;
1553	<	int rc = wc & RUNNING_COUNT_MASK;
1554	<	int dc = pc - rc;
976	<	if (dc > 0 && spare != null && spare.tryUnsuspend()) {
977	<	if (adj) {
978	<	int c;
979	<	do {} while (!UNSAFE.compareAndSwapInt
980	<	(this, workerCountsOffset,
981	<	c = workerCounts, c + ONE_RUNNING));
982	<	}
983	<	adj = true;
984	<	LockSupport.unpark(spare);
985	<	}
986	<	else if (adj) {
987	<	if (dc <= 0)
1551	>	if (System.nanoTime() - startTime >= SHRINK_TIMEOUT &&
1552	>	U.compareAndSwapLong(this, CTL, c, nc)) {
1553	>	w.runState = -1;          // shrink
1554	>	w.eventCount = (ec + E_SEQ) | E_MASK;
1555		break;
989	–	int tc = wc >>> TOTAL_COUNT_SHIFT;
990	–	if (scans > tc) {
991	–	int ts = (tc - pc) * pc;
992	–	if (rc != 0 &&  (dc * dc < ts || !maintainPar))
993	–	break;
994	–	if (scans > ts && tc < maxPoolSize &&
995	–	UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
996	–	wc+(ONE_RUNNING|ONE_TOTAL))){
997	–	addWorker();
998	–	break;
999	–	}
1556		}
1557		}
1002	–	else if (rc != 0)
1003	–	adj = UNSAFE.compareAndSwapInt (this, workerCountsOffset,
1004	–	wc, wc - ONE_RUNNING);
1005	–	if ((++scans & 1) == 0)
1006	–	releaseWaiters();   // help others progress
1007	–	else
1008	–	Thread.yield();     // avoid starving productive threads
1009	–	}
1010	–
1011	–	try {
1012	–	if (!done)
1013	–	do {} while (!blocker.isReleasable() && !blocker.block());
1014	–	} finally {
1015	–	if (adj) {
1016	–	int c;
1017	–	do {} while (!UNSAFE.compareAndSwapInt
1018	–	(this, workerCountsOffset,
1019	–	c = workerCounts, c + ONE_RUNNING));
1020	–	}
1558		}
1559		}
1560
1561		/**
1562	<	* Unless there are not enough other running threads, adjusts
1563	<	* counts and blocks a worker performing helpJoin that cannot find
1564	<	* any work.
1565	<	*
1566	<	* @return true if joinMe now done
1567	<	*/
1568	<	final boolean tryAwaitBusyJoin(ForkJoinTask<?> joinMe) {
1569	<	int pc = parallelism;
1570	<	outer:for (;;) {
1571	<	releaseWaiters();
1572	<	if ((workerCounts & RUNNING_COUNT_MASK) < pc) {
1573	<	ForkJoinWorkerThread[] ws = workers;
1574	<	int nws = ws.length;
1575	<	for (int i = 0; i < nws; ++i) {
1576	<	ForkJoinWorkerThread w = ws[i];
1577	<	if (w != null && w.isSuspended()) {
1578	<	if (joinMe.status < 0)
1579	<	return true;
1580	<	if ((workerCounts & RUNNING_COUNT_MASK) > pc)
1562	>	* Tries to locate and execute tasks for a stealer of the given
1563	>	* task, or in turn one of its stealers, Traces currentSteal ->
1564	>	* currentJoin links looking for a thread working on a descendant
1565	>	* of the given task and with a non-empty queue to steal back and
1566	>	* execute tasks from. The first call to this method upon a
1567	>	* waiting join will often entail scanning/search, (which is OK
1568	>	* because the joiner has nothing better to do), but this method
1569	>	* leaves hints in workers to speed up subsequent calls. The
1570	>	* implementation is very branchy to cope with potential
1571	>	* inconsistencies or loops encountering chains that are stale,
1572	>	* unknown, or of length greater than MAX_HELP_DEPTH links.  All
1573	>	* of these cases are dealt with by just retrying by caller.
1574	>	*
1575	>	* @param joiner the joining worker
1576	>	* @param task the task to join
1577	>	* @return true if found or ran a task (and so is immediately retryable)
1578	>	*/
1579	>	final boolean tryHelpStealer(WorkQueue joiner, ForkJoinTask<?> task) {
1580	>	ForkJoinTask<?> subtask;    // current target
1581	>	boolean progress = false;
1582	>	int depth = 0;              // current chain depth
1583	>	int m = runState & SMASK;
1584	>	WorkQueue[] ws = workQueues;
1585	>
1586	>	if (ws != null && ws.length > m && (subtask = task).status >= 0) {
1587	>	outer:for (WorkQueue j = joiner;;) {
1588	>	// Try to find the stealer of subtask, by first using hint
1589	>	WorkQueue stealer = null;
1590	>	WorkQueue v = ws[j.stealHint & m];
1591	>	if (v != null && v.currentSteal == subtask)
1592	>	stealer = v;
1593	>	else {
1594	>	for (int i = 1; i <= m; i += 2) {
1595	>	if ((v = ws[i]) != null && v.currentSteal == subtask) {
1596	>	stealer = v;
1597	>	j.stealHint = i; // save hint
1598		break;
1045	–	if (w.tryUnsuspend()) {
1046	–	LockSupport.unpark(w);
1047	–	break outer;
1599		}
1600	<	continue outer;
1600	>	}
1601	>	if (stealer == null)
1602	>	break;
1603	>	}
1604	>
1605	>	for (WorkQueue q = stealer;;) { // Try to help stealer
1606	>	ForkJoinTask<?> t; int b;
1607	>	if (task.status < 0)
1608	>	break outer;
1609	>	if ((b = q.base) - q.top < 0) {
1610	>	progress = true;
1611	>	if (subtask.status < 0)
1612	>	break outer;               // stale
1613	>	if ((t = q.pollAt(b)) != null) {
1614	>	stealer.stealHint = joiner.poolIndex;
1615	>	joiner.runSubtask(t);
1616	>	}
1617	>	}
1618	>	else { // empty - try to descend to find stealer's stealer
1619	>	ForkJoinTask<?> next = stealer.currentJoin;
1620	>	if (++depth == MAX_HELP_DEPTH || subtask.status < 0 ||
1621	>	next == null || next == subtask)
1622	>	break outer;  // max depth, stale, dead-end, cyclic
1623	>	subtask = next;
1624	>	j = stealer;
1625	>	break;
1626		}
1627		}
1628		}
1053	–	if (joinMe.status < 0)
1054	–	return true;
1055	–	int wc = workerCounts;
1056	–	if ((wc & RUNNING_COUNT_MASK) <= 2 ||
1057	–	(wc >>> TOTAL_COUNT_SHIFT) < pc)
1058	–	return false;  // keep this thread alive
1059	–	if (UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1060	–	wc, wc - ONE_RUNNING))
1061	–	break;
1629		}
1630	<
1064	<	joinMe.internalAwaitDone();
1065	<	int c;
1066	<	do {} while (!UNSAFE.compareAndSwapInt
1067	<	(this, workerCountsOffset,
1068	<	c = workerCounts, c + ONE_RUNNING));
1069	<	return true;
1630	>	return progress;
1631		}
1632
1633		/**
1634	<	* Possibly initiates and/or completes termination.
1634	>	* If task is at base of some steal queue, steals and executes it.
1635		*
1636	<	* @param now if true, unconditionally terminate, else only
1637	<	* if shutdown and empty queue and no active workers
1077	<	* @return true if now terminating or terminated
1636	>	* @param joiner the joining worker
1637	>	* @param task the task
1638		*/
1639	<	private boolean tryTerminate(boolean now) {
1640	<	if (now)
1641	<	advanceRunLevel(SHUTDOWN); // ensure at least SHUTDOWN
1642	<	else if (runState < SHUTDOWN ||
1643	<	!submissionQueue.isEmpty() ||
1644	<	(runState & ACTIVE_COUNT_MASK) != 0)
1645	<	return false;
1646	<
1647	<	if (advanceRunLevel(TERMINATING))
1648	<	startTerminating();
1649	<
1090	<	// Finish now if all threads terminated; else in some subsequent call
1091	<	if ((workerCounts >>> TOTAL_COUNT_SHIFT) == 0) {
1092	<	advanceRunLevel(TERMINATED);
1093	<	terminationLatch.countDown();
1639	>	final void tryPollForAndExec(WorkQueue joiner, ForkJoinTask<?> task) {
1640	>	WorkQueue[] ws;
1641	>	int m = runState & SMASK;
1642	>	if ((ws = workQueues) != null && ws.length > m) {
1643	>	for (int j = 1; j <= m && task.status >= 0; j += 2) {
1644	>	WorkQueue q = ws[j];
1645	>	if (q != null && q.pollFor(task)) {
1646	>	joiner.runSubtask(task);
1647	>	break;
1648	>	}
1649	>	}
1650		}
1095	–	return true;
1651		}
1652
1653		/**
1654	<	* Actions on transition to TERMINATING
1655	<	*/
1656	<	private void startTerminating() {
1657	<	for (int i = 0; i < 2; ++i) { // twice to mop up newly created workers
1658	<	cancelSubmissions();
1659	<	shutdownWorkers();
1660	<	cancelWorkerTasks();
1661	<	advanceEventCount();
1662	<	releaseWaiters();
1663	<	interruptWorkers();
1654	>	* Returns a non-empty steal queue, if is found during a random,
1655	>	* then cyclic scan, else null.  This method must be retried by
1656	>	* caller if, by the time it tries to use the queue, it is empty.
1657	>	*/
1658	>	private WorkQueue findNonEmptyStealQueue(WorkQueue w) {
1659	>	int r = w.seed;    // Same idea as scan(), but ignoring submissions
1660	>	for (WorkQueue[] ws;;) {
1661	>	int m = runState & SMASK;
1662	>	if ((ws = workQueues) == null)
1663	>	return null;
1664	>	if (ws.length > m) {
1665	>	WorkQueue q;
1666	>	for (int n = m << 2, k = r, j = -n;;) {
1667	>	r ^= r << 13; r ^= r >>> 17; r ^= r << 5;
1668	>	if ((q = ws[(k | 1) & m]) != null && q.base - q.top < 0) {
1669	>	w.seed = r;
1670	>	return q;
1671	>	}
1672	>	else if (j > n)
1673	>	return null;
1674	>	else
1675	>	k = (j++ < 0) ? r : k + ((m >>> 1) | 1);
1676	>
1677	>	}
1678	>	}
1679		}
1680		}
1681
1682		/**
1683	<	* Clear out and cancel submissions, ignoring exceptions
1684	<	*/
1685	<	private void cancelSubmissions() {
1686	<	ForkJoinTask<?> task;
1687	<	while ((task = submissionQueue.poll()) != null) {
1688	<	try {
1689	<	task.cancel(false);
1690	<	} catch (Throwable ignore) {
1683	>	* Runs tasks until {@code isQuiescent()}. We piggyback on
1684	>	* active count ctl maintenance, but rather than blocking
1685	>	* when tasks cannot be found, we rescan until all others cannot
1686	>	* find tasks either.
1687	>	*/
1688	>	final void helpQuiescePool(WorkQueue w) {
1689	>	for (boolean active = true;;) {
1690	>	w.runLocalTasks();      // exhaust local queue
1691	>	WorkQueue q = findNonEmptyStealQueue(w);
1692	>	if (q != null) {
1693	>	ForkJoinTask<?> t;
1694	>	if (!active) {      // re-establish active count
1695	>	long c;
1696	>	active = true;
1697	>	do {} while (!U.compareAndSwapLong
1698	>	(this, CTL, c = ctl, c + AC_UNIT));
1699	>	}
1700	>	if ((t = q.poll()) != null)
1701	>	w.runSubtask(t);
1702	>	}
1703	>	else {
1704	>	long c;
1705	>	if (active) {       // decrement active count without queuing
1706	>	active = false;
1707	>	do {} while (!U.compareAndSwapLong
1708	>	(this, CTL, c = ctl, c -= AC_UNIT));
1709	>	}
1710	>	else
1711	>	c = ctl;        // re-increment on exit
1712	>	if ((int)(c >> AC_SHIFT) + parallelism == 0) {
1713	>	do {} while (!U.compareAndSwapLong
1714	>	(this, CTL, c = ctl, c + AC_UNIT));
1715	>	break;
1716	>	}
1717		}
1718		}
1719		}
1720
1721		/**
1722	<	* Sets all worker run states to at least shutdown,
1723	<	* also resuming suspended workers
1722	>	* Gets and removes a local or stolen task for the given worker
1723	>	*
1724	>	* @return a task, if available
1725		*/
1726	<	private void shutdownWorkers() {
1727	<	ForkJoinWorkerThread[] ws = workers;
1728	<	int nws = ws.length;
1729	<	for (int i = 0; i < nws; ++i) {
1730	<	ForkJoinWorkerThread w = ws[i];
1731	<	if (w != null)
1732	<	w.shutdown();
1726	>	final ForkJoinTask<?> nextTaskFor(WorkQueue w) {
1727	>	for (ForkJoinTask<?> t;;) {
1728	>	WorkQueue q;
1729	>	if ((t = w.nextLocalTask()) != null)
1730	>	return t;
1731	>	if ((q = findNonEmptyStealQueue(w)) == null)
1732	>	return null;
1733	>	if ((t = q.poll()) != null)
1734	>	return t;
1735		}
1736		}
1737
1738		/**
1739	<	* Clears out and cancels all locally queued tasks
1739	>	* Returns the approximate (non-atomic) number of idle threads per
1740	>	* active thread to offset steal queue size for method
1741	>	* ForkJoinTask.getSurplusQueuedTaskCount().
1742		*/
1743	<	private void cancelWorkerTasks() {
1744	<	ForkJoinWorkerThread[] ws = workers;
1745	<	int nws = ws.length;
1746	<	for (int i = 0; i < nws; ++i) {
1747	<	ForkJoinWorkerThread w = ws[i];
1748	<	if (w != null)
1749	<	w.cancelTasks();
1743	>	final int idlePerActive() {
1744	>	// Approximate at powers of two for small values, saturate past 4
1745	>	int p = parallelism;
1746	>	int a = p + (int)(ctl >> AC_SHIFT);
1747	>	return (a > (p >>>= 1) ? 0 :
1748	>	a > (p >>>= 1) ? 1 :
1749	>	a > (p >>>= 1) ? 2 :
1750	>	a > (p >>>= 1) ? 4 :
1751	>	8);
1752	>	}
1753	>
1754	>	// Termination
1755	>
1756	>	/**
1757	>	* Sets SHUTDOWN bit of runState under lock
1758	>	*/
1759	>	private void enableShutdown() {
1760	>	ReentrantLock lock = this.lock;
1761	>	if (runState >= 0) {
1762	>	lock.lock();                       // don't need try/finally
1763	>	runState |= SHUTDOWN;
1764	>	lock.unlock();
1765		}
1766		}
1767
1768		/**
1769	<	* Unsticks all workers blocked on joins etc
1769	>	* Possibly initiates and/or completes termination.  Upon
1770	>	* termination, cancels all queued tasks and then
1771	>	*
1772	>	* @param now if true, unconditionally terminate, else only
1773	>	* if no work and no active workers
1774	>	* @return true if now terminating or terminated
1775		*/
1776	<	private void interruptWorkers() {
1777	<	ForkJoinWorkerThread[] ws = workers;
1778	<	int nws = ws.length;
1779	<	for (int i = 0; i < nws; ++i) {
1780	<	ForkJoinWorkerThread w = ws[i];
1781	<	if (w != null && !w.isTerminated()) {
1782	<	try {
1783	<	w.interrupt();
1784	<	} catch (SecurityException ignore) {
1776	>	private boolean tryTerminate(boolean now) {
1777	>	for (long c;;) {
1778	>	if (((c = ctl) & STOP_BIT) != 0) {      // already terminating
1779	>	if ((short)(c >>> TC_SHIFT) == -parallelism) {
1780	>	ReentrantLock lock = this.lock; // signal when no workers
1781	>	lock.lock();                    // don't need try/finally
1782	>	termination.signalAll();        // signal when 0 workers
1783	>	lock.unlock();
1784	>	}
1785	>	return true;
1786	>	}
1787	>	if (!now) {
1788	>	if ((int)(c >> AC_SHIFT) != -parallelism || runState >= 0 ||
1789	>	hasQueuedSubmissions())
1790	>	return false;
1791	>	// Check for unqueued inactive workers. One pass suffices.
1792	>	WorkQueue[] ws = workQueues; WorkQueue w;
1793	>	if (ws != null) {
1794	>	int n = ws.length;
1795	>	for (int i = 1; i < n; i += 2) {
1796	>	if ((w = ws[i]) != null && w.eventCount >= 0)
1797	>	return false;
1798	>	}
1799		}
1800		}
1801	+	if (U.compareAndSwapLong(this, CTL, c, c | STOP_BIT))
1802	+	startTerminating();
1803		}
1804		}
1805
1169	–	// misc support for ForkJoinWorkerThread
1170	–
1171	–	/**
1172	–	* Returns pool number
1173	–	*/
1174	–	final int getPoolNumber() {
1175	–	return poolNumber;
1176	–	}
1177	–
1806		/**
1807	<	* Accumulates steal count from a worker, clearing
1808	<	* the worker's value
1807	>	* Initiates termination: Runs three passes through workQueues:
1808	>	* (0) Setting termination status, followed by wakeups of queued
1809	>	* workers; (1) cancelling all tasks; (2) interrupting lagging
1810	>	* threads (likely in external tasks, but possibly also blocked in
1811	>	* joins).  Each pass repeats previous steps because of potential
1812	>	* lagging thread creation.
1813		*/
1814	<	final void accumulateStealCount(ForkJoinWorkerThread w) {
1815	<	int sc = w.stealCount;
1816	<	if (sc != 0) {
1817	<	long c;
1818	<	w.stealCount = 0;
1819	<	do {} while (!UNSAFE.compareAndSwapLong(this, stealCountOffset,
1820	<	c = stealCount, c + sc));
1814	>	private void startTerminating() {
1815	>	for (int pass = 0; pass < 3; ++pass) {
1816	>	WorkQueue[] ws = workQueues;
1817	>	if (ws != null) {
1818	>	WorkQueue w; Thread wt;
1819	>	int n = ws.length;
1820	>	for (int j = 0; j < n; ++j) {
1821	>	if ((w = ws[j]) != null) {
1822	>	w.runState = -1;
1823	>	if (pass > 0) {
1824	>	w.cancelAll();
1825	>	if (pass > 1 && (wt = w.owner) != null &&
1826	>	!wt.isInterrupted()) {
1827	>	try {
1828	>	wt.interrupt();
1829	>	} catch (SecurityException ignore) {
1830	>	}
1831	>	}
1832	>	}
1833	>	}
1834	>	}
1835	>	// Wake up workers parked on event queue
1836	>	int i, e; long c; Thread p;
1837	>	while ((i = ((~(e = (int)(c = ctl)) << 1) | 1) & SMASK) < n &&
1838	>	(w = ws[i]) != null &&
1839	>	w.eventCount == (e | INT_SIGN)) {
1840	>	long nc = ((long)(w.nextWait & E_MASK) |
1841	>	((c + AC_UNIT) & AC_MASK) |
1842	>	(c & (TC_MASK|STOP_BIT)));
1843	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1844	>	w.eventCount = (e + E_SEQ) & E_MASK;
1845	>	if ((p = w.parker) != null)
1846	>	U.unpark(p);
1847	>	}
1848	>	}
1849	>	}
1850		}
1851		}
1852
1853	<	/**
1193	<	* Returns the approximate (non-atomic) number of idle threads per
1194	<	* active thread.
1195	<	*/
1196	<	final int idlePerActive() {
1197	<	int ac = runState;    // no mask -- artifically boosts during shutdown
1198	<	int pc = parallelism; // use targeted parallelism, not rc
1199	<	// Use exact results for small values, saturate past 4
1200	<	return pc <= ac? 0 : pc >>> 1 <= ac? 1 : pc >>> 2 <= ac? 3 : pc >>> 3;
1201	<	}
1202	<
1203	<	// Public and protected methods
1853	>	// Exported methods
1854
1855		// Constructors
1856
1857		/**
1858		* Creates a {@code ForkJoinPool} with parallelism equal to {@link
1859	<	* java.lang.Runtime#availableProcessors}, and using the {@linkplain
1860	<	* #defaultForkJoinWorkerThreadFactory default thread factory}.
1859	>	* java.lang.Runtime#availableProcessors}, using the {@linkplain
1860	>	* #defaultForkJoinWorkerThreadFactory default thread factory},
1861	>	* no UncaughtExceptionHandler, and non-async LIFO processing mode.
1862		*
1863		* @throws SecurityException if a security manager exists and
1864		*         the caller is not permitted to modify threads
#	Line 1216 | Line 1867 | public class ForkJoinPool extends Abstra
1867		*/
1868		public ForkJoinPool() {
1869		this(Runtime.getRuntime().availableProcessors(),
1870	<	defaultForkJoinWorkerThreadFactory);
1870	>	defaultForkJoinWorkerThreadFactory, null, false);
1871		}
1872
1873		/**
1874		* Creates a {@code ForkJoinPool} with the indicated parallelism
1875	<	* level and using the {@linkplain
1876	<	* #defaultForkJoinWorkerThreadFactory default thread factory}.
1875	>	* level, the {@linkplain
1876	>	* #defaultForkJoinWorkerThreadFactory default thread factory},
1877	>	* no UncaughtExceptionHandler, and non-async LIFO processing mode.
1878		*
1879		* @param parallelism the parallelism level
1880		* @throws IllegalArgumentException if parallelism less than or
#	Line 1233 | Line 1885 | public class ForkJoinPool extends Abstra
1885		*         java.lang.RuntimePermission}{@code ("modifyThread")}
1886		*/
1887		public ForkJoinPool(int parallelism) {
1888	<	this(parallelism, defaultForkJoinWorkerThreadFactory);
1888	>	this(parallelism, defaultForkJoinWorkerThreadFactory, null, false);
1889		}
1890
1891		/**
1892	<	* Creates a {@code ForkJoinPool} with parallelism equal to {@link
1241	<	* java.lang.Runtime#availableProcessors}, and using the given
1242	<	* thread factory.
1892	>	* Creates a {@code ForkJoinPool} with the given parameters.
1893		*
1894	<	* @param factory the factory for creating new threads
1895	<	* @throws NullPointerException if the factory is null
1896	<	* @throws SecurityException if a security manager exists and
1897	<	*         the caller is not permitted to modify threads
1898	<	*         because it does not hold {@link
1899	<	*         java.lang.RuntimePermission}{@code ("modifyThread")}
1900	<	*/
1901	<	public ForkJoinPool(ForkJoinWorkerThreadFactory factory) {
1902	<	this(Runtime.getRuntime().availableProcessors(), factory);
1903	<	}
1904	<
1905	<	/**
1906	<	* Creates a {@code ForkJoinPool} with the given parallelism and
1257	<	* thread factory.
1258	<	*
1259	<	* @param parallelism the parallelism level
1260	<	* @param factory the factory for creating new threads
1894	>	* @param parallelism the parallelism level. For default value,
1895	>	* use {@link java.lang.Runtime#availableProcessors}.
1896	>	* @param factory the factory for creating new threads. For default value,
1897	>	* use {@link #defaultForkJoinWorkerThreadFactory}.
1898	>	* @param handler the handler for internal worker threads that
1899	>	* terminate due to unrecoverable errors encountered while executing
1900	>	* tasks. For default value, use {@code null}.
1901	>	* @param asyncMode if true,
1902	>	* establishes local first-in-first-out scheduling mode for forked
1903	>	* tasks that are never joined. This mode may be more appropriate
1904	>	* than default locally stack-based mode in applications in which
1905	>	* worker threads only process event-style asynchronous tasks.
1906	>	* For default value, use {@code false}.
1907		* @throws IllegalArgumentException if parallelism less than or
1908		*         equal to zero, or greater than implementation limit
1909		* @throws NullPointerException if the factory is null
#	Line 1266 | Line 1912 | public class ForkJoinPool extends Abstra
1912		*         because it does not hold {@link
1913		*         java.lang.RuntimePermission}{@code ("modifyThread")}
1914		*/
1915	<	public ForkJoinPool(int parallelism, ForkJoinWorkerThreadFactory factory) {
1915	>	public ForkJoinPool(int parallelism,
1916	>	ForkJoinWorkerThreadFactory factory,
1917	>	Thread.UncaughtExceptionHandler handler,
1918	>	boolean asyncMode) {
1919		checkPermission();
1920		if (factory == null)
1921		throw new NullPointerException();
1922	<	if (parallelism <= 0 || parallelism > MAX_THREADS)
1922	>	if (parallelism <= 0 || parallelism > MAX_ID)
1923		throw new IllegalArgumentException();
1275	–	this.poolNumber = poolNumberGenerator.incrementAndGet();
1276	–	int arraySize = initialArraySizeFor(parallelism);
1924		this.parallelism = parallelism;
1925		this.factory = factory;
1926	<	this.maxPoolSize = MAX_THREADS;
1927	<	this.maintainsParallelism = true;
1928	<	this.workers = new ForkJoinWorkerThread[arraySize];
1929	<	this.submissionQueue = new LinkedTransferQueue<ForkJoinTask<?>>();
1930	<	this.workerLock = new ReentrantLock();
1931	<	this.terminationLatch = new CountDownLatch(1);
1932	<	}
1933	<
1934	<	/**
1935	<	* Returns initial power of two size for workers array.
1936	<	* @param pc the initial parallelism level
1937	<	*/
1938	<	private static int initialArraySizeFor(int pc) {
1939	<	// See Hackers Delight, sec 3.2. We know MAX_THREADS < (1 >>> 16)
1940	<	int size = pc < MAX_THREADS ? pc + 1 : MAX_THREADS;
1941	<	size |= size >>> 1;
1942	<	size |= size >>> 2;
1943	<	size |= size >>> 4;
1944	<	size |= size >>> 8;
1945	<	return size + 1;
1926	>	this.ueh = handler;
1927	>	this.localMode = asyncMode ? FIFO_QUEUE : LIFO_QUEUE;
1928	>	this.nextPoolIndex = 1;
1929	>	long np = (long)(-parallelism); // offset ctl counts
1930	>	this.ctl = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK);
1931	>	// initialize workQueues array with room for 2*parallelism if possible
1932	>	int n = parallelism << 1;
1933	>	if (n >= MAX_ID)
1934	>	n = MAX_ID;
1935	>	else { // See Hackers Delight, sec 3.2, where n < (1 << 16)
1936	>	n |= n >>> 1; n |= n >>> 2; n |= n >>> 4; n |= n >>> 8;
1937	>	}
1938	>	this.workQueues = new WorkQueue[(n + 1) << 1];
1939	>	ReentrantLock lck = this.lock = new ReentrantLock();
1940	>	this.termination = lck.newCondition();
1941	>	this.stealCount = new AtomicLong();
1942	>	this.nextWorkerNumber = new AtomicInteger();
1943	>	StringBuilder sb = new StringBuilder("ForkJoinPool-");
1944	>	sb.append(poolNumberGenerator.incrementAndGet());
1945	>	sb.append("-worker-");
1946	>	this.workerNamePrefix = sb.toString();
1947	>	// Create initial submission queue
1948	>	WorkQueue sq = tryAddSharedQueue(0);
1949	>	if (sq != null)
1950	>	sq.growArray(false);
1951		}
1952
1953		// Execution methods
1954
1955		/**
1304	–	* Common code for execute, invoke and submit
1305	–	*/
1306	–	private <T> void doSubmit(ForkJoinTask<T> task) {
1307	–	if (task == null)
1308	–	throw new NullPointerException();
1309	–	if (runState >= SHUTDOWN)
1310	–	throw new RejectedExecutionException();
1311	–	submissionQueue.offer(task);
1312	–	advanceEventCount();
1313	–	releaseWaiters();
1314	–	ensureEnoughTotalWorkers();
1315	–	}
1316	–
1317	–	/**
1956		* Performs the given task, returning its result upon completion.
1957	+	* If the computation encounters an unchecked Exception or Error,
1958	+	* it is rethrown as the outcome of this invocation.  Rethrown
1959	+	* exceptions behave in the same way as regular exceptions, but,
1960	+	* when possible, contain stack traces (as displayed for example
1961	+	* using {@code ex.printStackTrace()}) of both the current thread
1962	+	* as well as the thread actually encountering the exception;
1963	+	* minimally only the latter.
1964		*
1965		* @param task the task
1966		* @return the task's result
#	Line 1348 | Line 1993 | public class ForkJoinPool extends Abstra
1993		*         scheduled for execution
1994		*/
1995		public void execute(Runnable task) {
1996	+	if (task == null)
1997	+	throw new NullPointerException();
1998		ForkJoinTask<?> job;
1999		if (task instanceof ForkJoinTask<?>) // avoid re-wrap
2000		job = (ForkJoinTask<?>) task;
#	Line 1357 | Line 2004 | public class ForkJoinPool extends Abstra
2004		}
2005
2006		/**
2007	+	* Submits a ForkJoinTask for execution.
2008	+	*
2009	+	* @param task the task to submit
2010	+	* @return the task
2011	+	* @throws NullPointerException if the task is null
2012	+	* @throws RejectedExecutionException if the task cannot be
2013	+	*         scheduled for execution
2014	+	*/
2015	+	public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) {
2016	+	doSubmit(task);
2017	+	return task;
2018	+	}
2019	+
2020	+	/**
2021		* @throws NullPointerException if the task is null
2022		* @throws RejectedExecutionException if the task cannot be
2023		*         scheduled for execution
2024		*/
2025		public <T> ForkJoinTask<T> submit(Callable<T> task) {
2026	+	if (task == null)
2027	+	throw new NullPointerException();
2028		ForkJoinTask<T> job = ForkJoinTask.adapt(task);
2029		doSubmit(job);
2030		return job;
#	Line 1373 | Line 2036 | public class ForkJoinPool extends Abstra
2036		*         scheduled for execution
2037		*/
2038		public <T> ForkJoinTask<T> submit(Runnable task, T result) {
2039	+	if (task == null)
2040	+	throw new NullPointerException();
2041		ForkJoinTask<T> job = ForkJoinTask.adapt(task, result);
2042		doSubmit(job);
2043		return job;
#	Line 1384 | Line 2049 | public class ForkJoinPool extends Abstra
2049		*         scheduled for execution
2050		*/
2051		public ForkJoinTask<?> submit(Runnable task) {
2052	+	if (task == null)
2053	+	throw new NullPointerException();
2054		ForkJoinTask<?> job;
2055		if (task instanceof ForkJoinTask<?>) // avoid re-wrap
2056		job = (ForkJoinTask<?>) task;
#	Line 1394 | Line 2061 | public class ForkJoinPool extends Abstra
2061		}
2062
2063		/**
1397	–	* Submits a ForkJoinTask for execution.
1398	–	*
1399	–	* @param task the task to submit
1400	–	* @return the task
1401	–	* @throws NullPointerException if the task is null
1402	–	* @throws RejectedExecutionException if the task cannot be
1403	–	*         scheduled for execution
1404	–	*/
1405	–	public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) {
1406	–	doSubmit(task);
1407	–	return task;
1408	–	}
1409	–
1410	–	/**
2064		* @throws NullPointerException       {@inheritDoc}
2065		* @throws RejectedExecutionException {@inheritDoc}
2066		*/
#	Line 1449 | Line 2102 | public class ForkJoinPool extends Abstra
2102		* @return the handler, or {@code null} if none
2103		*/
2104		public Thread.UncaughtExceptionHandler getUncaughtExceptionHandler() {
1452	–	workerCountReadFence();
2105		return ueh;
2106		}
2107
2108		/**
1457	–	* Sets the handler for internal worker threads that terminate due
1458	–	* to unrecoverable errors encountered while executing tasks.
1459	–	* Unless set, the current default or ThreadGroup handler is used
1460	–	* as handler.
1461	–	*
1462	–	* @param h the new handler
1463	–	* @return the old handler, or {@code null} if none
1464	–	* @throws SecurityException if a security manager exists and
1465	–	*         the caller is not permitted to modify threads
1466	–	*         because it does not hold {@link
1467	–	*         java.lang.RuntimePermission}{@code ("modifyThread")}
1468	–	*/
1469	–	public Thread.UncaughtExceptionHandler
1470	–	setUncaughtExceptionHandler(Thread.UncaughtExceptionHandler h) {
1471	–	checkPermission();
1472	–	Thread.UncaughtExceptionHandler old = ueh;
1473	–	if (h != old) {
1474	–	ueh = h;
1475	–	ForkJoinWorkerThread[] ws = workers;
1476	–	int nws = ws.length;
1477	–	for (int i = 0; i < nws; ++i) {
1478	–	ForkJoinWorkerThread w = ws[i];
1479	–	if (w != null)
1480	–	w.setUncaughtExceptionHandler(h);
1481	–	}
1482	–	}
1483	–	return old;
1484	–	}
1485	–
1486	–	/**
1487	–	* Sets the target parallelism level of this pool.
1488	–	*
1489	–	* @param parallelism the target parallelism
1490	–	* @throws IllegalArgumentException if parallelism less than or
1491	–	* equal to zero or greater than maximum size bounds
1492	–	* @throws SecurityException if a security manager exists and
1493	–	*         the caller is not permitted to modify threads
1494	–	*         because it does not hold {@link
1495	–	*         java.lang.RuntimePermission}{@code ("modifyThread")}
1496	–	*/
1497	–	public void setParallelism(int parallelism) {
1498	–	checkPermission();
1499	–	if (parallelism <= 0 || parallelism > maxPoolSize)
1500	–	throw new IllegalArgumentException();
1501	–	workerCountReadFence();
1502	–	int pc = this.parallelism;
1503	–	if (pc != parallelism) {
1504	–	this.parallelism = parallelism;
1505	–	workerCountWriteFence();
1506	–	// Release spares. If too many, some will die after re-suspend
1507	–	ForkJoinWorkerThread[] ws = workers;
1508	–	int nws = ws.length;
1509	–	for (int i = 0; i < nws; ++i) {
1510	–	ForkJoinWorkerThread w = ws[i];
1511	–	if (w != null && w.tryUnsuspend()) {
1512	–	int c;
1513	–	do {} while (!UNSAFE.compareAndSwapInt
1514	–	(this, workerCountsOffset,
1515	–	c = workerCounts, c + ONE_RUNNING));
1516	–	LockSupport.unpark(w);
1517	–	}
1518	–	}
1519	–	ensureEnoughTotalWorkers();
1520	–	advanceEventCount();
1521	–	releaseWaiters(); // force config recheck by existing workers
1522	–	}
1523	–	}
1524	–
1525	–	/**
2109		* Returns the targeted parallelism level of this pool.
2110		*
2111		* @return the targeted parallelism level of this pool
2112		*/
2113		public int getParallelism() {
1531	–	//        workerCountReadFence(); // inlined below
1532	–	int ignore = workerCounts;
2114		return parallelism;
2115		}
2116
2117		/**
2118		* Returns the number of worker threads that have started but not
2119	<	* yet terminated.  This result returned by this method may differ
2119	>	* yet terminated.  The result returned by this method may differ
2120		* from {@link #getParallelism} when threads are created to
2121		* maintain parallelism when others are cooperatively blocked.
2122		*
2123		* @return the number of worker threads
2124		*/
2125		public int getPoolSize() {
2126	<	return workerCounts >>> TOTAL_COUNT_SHIFT;
1546	<	}
1547	<
1548	<	/**
1549	<	* Returns the maximum number of threads allowed to exist in the
1550	<	* pool. Unless set using {@link #setMaximumPoolSize}, the
1551	<	* maximum is an implementation-defined value designed only to
1552	<	* prevent runaway growth.
1553	<	*
1554	<	* @return the maximum
1555	<	*/
1556	<	public int getMaximumPoolSize() {
1557	<	workerCountReadFence();
1558	<	return maxPoolSize;
1559	<	}
1560	<
1561	<	/**
1562	<	* Sets the maximum number of threads allowed to exist in the
1563	<	* pool. The given value should normally be greater than or equal
1564	<	* to the {@link #getParallelism parallelism} level. Setting this
1565	<	* value has no effect on current pool size. It controls
1566	<	* construction of new threads. The use of this method may cause
1567	<	* tasks that intrinsically require extra threads for dependent
1568	<	* computations to indefinitely stall. If you are instead trying
1569	<	* to minimize internal thread creation, consider setting {@link
1570	<	* #setMaintainsParallelism} as false.
1571	<	*
1572	<	* @throws IllegalArgumentException if negative or greater than
1573	<	* internal implementation limit
1574	<	*/
1575	<	public void setMaximumPoolSize(int newMax) {
1576	<	if (newMax < 0 || newMax > MAX_THREADS)
1577	<	throw new IllegalArgumentException();
1578	<	maxPoolSize = newMax;
1579	<	workerCountWriteFence();
1580	<	}
1581	<
1582	<	/**
1583	<	* Returns {@code true} if this pool dynamically maintains its
1584	<	* target parallelism level. If false, new threads are added only
1585	<	* to avoid possible starvation.  This setting is by default true.
1586	<	*
1587	<	* @return {@code true} if maintains parallelism
1588	<	*/
1589	<	public boolean getMaintainsParallelism() {
1590	<	workerCountReadFence();
1591	<	return maintainsParallelism;
1592	<	}
1593	<
1594	<	/**
1595	<	* Sets whether this pool dynamically maintains its target
1596	<	* parallelism level. If false, new threads are added only to
1597	<	* avoid possible starvation.
1598	<	*
1599	<	* @param enable {@code true} to maintain parallelism
1600	<	*/
1601	<	public void setMaintainsParallelism(boolean enable) {
1602	<	maintainsParallelism = enable;
1603	<	workerCountWriteFence();
1604	<	}
1605	<
1606	<	/**
1607	<	* Establishes local first-in-first-out scheduling mode for forked
1608	<	* tasks that are never joined. This mode may be more appropriate
1609	<	* than default locally stack-based mode in applications in which
1610	<	* worker threads only process asynchronous tasks.  This method is
1611	<	* designed to be invoked only when the pool is quiescent, and
1612	<	* typically only before any tasks are submitted. The effects of
1613	<	* invocations at other times may be unpredictable.
1614	<	*
1615	<	* @param async if {@code true}, use locally FIFO scheduling
1616	<	* @return the previous mode
1617	<	* @see #getAsyncMode
1618	<	*/
1619	<	public boolean setAsyncMode(boolean async) {
1620	<	workerCountReadFence();
1621	<	boolean oldMode = locallyFifo;
1622	<	if (oldMode != async) {
1623	<	locallyFifo = async;
1624	<	workerCountWriteFence();
1625	<	ForkJoinWorkerThread[] ws = workers;
1626	<	int nws = ws.length;
1627	<	for (int i = 0; i < nws; ++i) {
1628	<	ForkJoinWorkerThread w = ws[i];
1629	<	if (w != null)
1630	<	w.setAsyncMode(async);
1631	<	}
1632	<	}
1633	<	return oldMode;
2126	>	return parallelism + (short)(ctl >>> TC_SHIFT);
2127		}
2128
2129		/**
#	Line 1638 | Line 2131 | public class ForkJoinPool extends Abstra
2131		* scheduling mode for forked tasks that are never joined.
2132		*
2133		* @return {@code true} if this pool uses async mode
1641	–	* @see #setAsyncMode
2134		*/
2135		public boolean getAsyncMode() {
2136	<	workerCountReadFence();
1645	<	return locallyFifo;
2136	>	return localMode != 0;
2137		}
2138
2139		/**
#	Line 1654 | Line 2145 | public class ForkJoinPool extends Abstra
2145		* @return the number of worker threads
2146		*/
2147		public int getRunningThreadCount() {
2148	<	return workerCounts & RUNNING_COUNT_MASK;
2148	>	int rc = 0;
2149	>	WorkQueue[] ws; WorkQueue w;
2150	>	if ((ws = workQueues) != null) {
2151	>	int n = ws.length;
2152	>	for (int i = 1; i < n; i += 2) {
2153	>	Thread.State s; ForkJoinWorkerThread wt;
2154	>	if ((w = ws[i]) != null && (wt = w.owner) != null &&
2155	>	w.eventCount >= 0 &&
2156	>	(s = wt.getState()) != Thread.State.BLOCKED &&
2157	>	s != Thread.State.WAITING &&
2158	>	s != Thread.State.TIMED_WAITING)
2159	>	++rc;
2160	>	}
2161	>	}
2162	>	return rc;
2163		}
2164
2165		/**
#	Line 1665 | Line 2170 | public class ForkJoinPool extends Abstra
2170		* @return the number of active threads
2171		*/
2172		public int getActiveThreadCount() {
2173	<	return runState & ACTIVE_COUNT_MASK;
2173	>	int r = parallelism + (int)(ctl >> AC_SHIFT);
2174	>	return (r <= 0) ? 0 : r; // suppress momentarily negative values
2175		}
2176
2177		/**
#	Line 1680 | Line 2186 | public class ForkJoinPool extends Abstra
2186		* @return {@code true} if all threads are currently idle
2187		*/
2188		public boolean isQuiescent() {
2189	<	return (runState & ACTIVE_COUNT_MASK) == 0;
2189	>	return (int)(ctl >> AC_SHIFT) + parallelism == 0;
2190		}
2191
2192		/**
#	Line 1695 | Line 2201 | public class ForkJoinPool extends Abstra
2201		* @return the number of steals
2202		*/
2203		public long getStealCount() {
2204	<	return stealCount;
2204	>	long count = stealCount.get();
2205	>	WorkQueue[] ws; WorkQueue w;
2206	>	if ((ws = workQueues) != null) {
2207	>	int n = ws.length;
2208	>	for (int i = 1; i < n; i += 2) {
2209	>	if ((w = ws[i]) != null)
2210	>	count += w.totalSteals;
2211	>	}
2212	>	}
2213	>	return count;
2214		}
2215
2216		/**
#	Line 1710 | Line 2225 | public class ForkJoinPool extends Abstra
2225		*/
2226		public long getQueuedTaskCount() {
2227		long count = 0;
2228	<	ForkJoinWorkerThread[] ws = workers;
2229	<	int nws = ws.length;
2230	<	for (int i = 0; i < nws; ++i) {
2231	<	ForkJoinWorkerThread w = ws[i];
2232	<	if (w != null)
2233	<	count += w.getQueueSize();
2228	>	WorkQueue[] ws; WorkQueue w;
2229	>	if ((ws = workQueues) != null) {
2230	>	int n = ws.length;
2231	>	for (int i = 1; i < n; i += 2) {
2232	>	if ((w = ws[i]) != null)
2233	>	count += w.queueSize();
2234	>	}
2235		}
2236		return count;
2237		}
2238
2239		/**
2240		* Returns an estimate of the number of tasks submitted to this
2241	<	* pool that have not yet begun executing.  This method takes time
2242	<	* proportional to the number of submissions.
2241	>	* pool that have not yet begun executing.  This method may take
2242	>	* time proportional to the number of submissions.
2243		*
2244		* @return the number of queued submissions
2245		*/
2246		public int getQueuedSubmissionCount() {
2247	<	return submissionQueue.size();
2247	>	int count = 0;
2248	>	WorkQueue[] ws; WorkQueue w;
2249	>	if ((ws = workQueues) != null) {
2250	>	int n = ws.length;
2251	>	for (int i = 0; i < n; i += 2) {
2252	>	if ((w = ws[i]) != null)
2253	>	count += w.queueSize();
2254	>	}
2255	>	}
2256	>	return count;
2257		}
2258
2259		/**
#	Line 1738 | Line 2263 | public class ForkJoinPool extends Abstra
2263		* @return {@code true} if there are any queued submissions
2264		*/
2265		public boolean hasQueuedSubmissions() {
2266	<	return !submissionQueue.isEmpty();
2266	>	WorkQueue[] ws; WorkQueue w;
2267	>	if ((ws = workQueues) != null) {
2268	>	int n = ws.length;
2269	>	for (int i = 0; i < n; i += 2) {
2270	>	if ((w = ws[i]) != null && w.queueSize() != 0)
2271	>	return true;
2272	>	}
2273	>	}
2274	>	return false;
2275		}
2276
2277		/**
#	Line 1749 | Line 2282 | public class ForkJoinPool extends Abstra
2282		* @return the next submission, or {@code null} if none
2283		*/
2284		protected ForkJoinTask<?> pollSubmission() {
2285	<	return submissionQueue.poll();
2285	>	WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2286	>	if ((ws = workQueues) != null) {
2287	>	int n = ws.length;
2288	>	for (int i = 0; i < n; i += 2) {
2289	>	if ((w = ws[i]) != null && (t = w.poll()) != null)
2290	>	return t;
2291	>	}
2292	>	}
2293	>	return null;
2294		}
2295
2296		/**
#	Line 1770 | Line 2311 | public class ForkJoinPool extends Abstra
2311		* @return the number of elements transferred
2312		*/
2313		protected int drainTasksTo(Collection<? super ForkJoinTask<?>> c) {
2314	<	int n = submissionQueue.drainTo(c);
2315	<	ForkJoinWorkerThread[] ws = workers;
2316	<	int nws = ws.length;
2317	<	for (int i = 0; i < nws; ++i) {
2318	<	ForkJoinWorkerThread w = ws[i];
2319	<	if (w != null)
2320	<	n += w.drainTasksTo(c);
2314	>	int count = 0;
2315	>	WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2316	>	if ((ws = workQueues) != null) {
2317	>	int n = ws.length;
2318	>	for (int i = 0; i < n; ++i) {
2319	>	if ((w = ws[i]) != null) {
2320	>	while ((t = w.poll()) != null) {
2321	>	c.add(t);
2322	>	++count;
2323	>	}
2324	>	}
2325	>	}
2326		}
2327	<	return n;
2327	>	return count;
2328		}
2329
2330		/**
#	Line 1792 | Line 2338 | public class ForkJoinPool extends Abstra
2338		long st = getStealCount();
2339		long qt = getQueuedTaskCount();
2340		long qs = getQueuedSubmissionCount();
2341	<	int wc = workerCounts;
1796	<	int tc = wc >>> TOTAL_COUNT_SHIFT;
1797	<	int rc = wc & RUNNING_COUNT_MASK;
2341	>	int rc = getRunningThreadCount();
2342		int pc = parallelism;
2343	<	int rs = runState;
2344	<	int ac = rs & ACTIVE_COUNT_MASK;
2343	>	long c = ctl;
2344	>	int tc = pc + (short)(c >>> TC_SHIFT);
2345	>	int ac = pc + (int)(c >> AC_SHIFT);
2346	>	if (ac < 0) // ignore transient negative
2347	>	ac = 0;
2348	>	String level;
2349	>	if ((c & STOP_BIT) != 0)
2350	>	level = (tc == 0) ? "Terminated" : "Terminating";
2351	>	else
2352	>	level = runState < 0 ? "Shutting down" : "Running";
2353		return super.toString() +
2354	<	"[" + runLevelToString(rs) +
2354	>	"[" + level +
2355		", parallelism = " + pc +
2356		", size = " + tc +
2357		", active = " + ac +
#	Line 1810 | Line 2362 | public class ForkJoinPool extends Abstra
2362		"]";
2363		}
2364
1813	–	private static String runLevelToString(int s) {
1814	–	return ((s & TERMINATED) != 0 ? "Terminated" :
1815	–	((s & TERMINATING) != 0 ? "Terminating" :
1816	–	((s & SHUTDOWN) != 0 ? "Shutting down" :
1817	–	"Running")));
1818	–	}
1819	–
2365		/**
2366		* Initiates an orderly shutdown in which previously submitted
2367		* tasks are executed, but no new tasks will be accepted.
#	Line 1831 | Line 2376 | public class ForkJoinPool extends Abstra
2376		*/
2377		public void shutdown() {
2378		checkPermission();
2379	<	advanceRunLevel(SHUTDOWN);
2379	>	enableShutdown();
2380		tryTerminate(false);
2381		}
2382
#	Line 1853 | Line 2398 | public class ForkJoinPool extends Abstra
2398		*/
2399		public List<Runnable> shutdownNow() {
2400		checkPermission();
2401	+	enableShutdown();
2402		tryTerminate(true);
2403		return Collections.emptyList();
2404		}
#	Line 1863 | Line 2409 | public class ForkJoinPool extends Abstra
2409		* @return {@code true} if all tasks have completed following shut down
2410		*/
2411		public boolean isTerminated() {
2412	<	return runState >= TERMINATED;
2412	>	long c = ctl;
2413	>	return ((c & STOP_BIT) != 0L &&
2414	>	(short)(c >>> TC_SHIFT) == -parallelism);
2415		}
2416
2417		/**
#	Line 1871 | Line 2419 | public class ForkJoinPool extends Abstra
2419		* commenced but not yet completed.  This method may be useful for
2420		* debugging. A return of {@code true} reported a sufficient
2421		* period after shutdown may indicate that submitted tasks have
2422	<	* ignored or suppressed interruption, causing this executor not
2423	<	* to properly terminate.
2422	>	* ignored or suppressed interruption, or are waiting for IO,
2423	>	* causing this executor not to properly terminate. (See the
2424	>	* advisory notes for class {@link ForkJoinTask} stating that
2425	>	* tasks should not normally entail blocking operations.  But if
2426	>	* they do, they must abort them on interrupt.)
2427		*
2428		* @return {@code true} if terminating but not yet terminated
2429		*/
2430		public boolean isTerminating() {
2431	<	return (runState & (TERMINATING|TERMINATED)) == TERMINATING;
2431	>	long c = ctl;
2432	>	return ((c & STOP_BIT) != 0L &&
2433	>	(short)(c >>> TC_SHIFT) != -parallelism);
2434		}
2435
2436		/**
#	Line 1886 | Line 2439 | public class ForkJoinPool extends Abstra
2439		* @return {@code true} if this pool has been shut down
2440		*/
2441		public boolean isShutdown() {
2442	<	return runState >= SHUTDOWN;
2442	>	return runState < 0;
2443		}
2444
2445		/**
#	Line 1902 | Line 2455 | public class ForkJoinPool extends Abstra
2455		*/
2456		public boolean awaitTermination(long timeout, TimeUnit unit)
2457		throws InterruptedException {
2458	<	return terminationLatch.await(timeout, unit);
2458	>	long nanos = unit.toNanos(timeout);
2459	>	final ReentrantLock lock = this.lock;
2460	>	lock.lock();
2461	>	try {
2462	>	for (;;) {
2463	>	if (isTerminated())
2464	>	return true;
2465	>	if (nanos <= 0)
2466	>	return false;
2467	>	nanos = termination.awaitNanos(nanos);
2468	>	}
2469	>	} finally {
2470	>	lock.unlock();
2471	>	}
2472		}
2473
2474		/**
2475		* Interface for extending managed parallelism for tasks running
2476		* in {@link ForkJoinPool}s.
2477		*
2478	<	* <p>A {@code ManagedBlocker} provides two methods.
2479	<	* Method {@code isReleasable} must return {@code true} if
2480	<	* blocking is not necessary. Method {@code block} blocks the
2481	<	* current thread if necessary (perhaps internally invoking
2482	<	* {@code isReleasable} before actually blocking).
2478	>	* <p>A {@code ManagedBlocker} provides two methods.  Method
2479	>	* {@code isReleasable} must return {@code true} if blocking is
2480	>	* not necessary. Method {@code block} blocks the current thread
2481	>	* if necessary (perhaps internally invoking {@code isReleasable}
2482	>	* before actually blocking). These actions are performed by any
2483	>	* thread invoking {@link ForkJoinPool#managedBlock}.  The
2484	>	* unusual methods in this API accommodate synchronizers that may,
2485	>	* but don't usually, block for long periods. Similarly, they
2486	>	* allow more efficient internal handling of cases in which
2487	>	* additional workers may be, but usually are not, needed to
2488	>	* ensure sufficient parallelism.  Toward this end,
2489	>	* implementations of method {@code isReleasable} must be amenable
2490	>	* to repeated invocation.
2491		*
2492		* <p>For example, here is a ManagedBlocker based on a
2493		* ReentrantLock:
#	Line 1931 | Line 2505 | public class ForkJoinPool extends Abstra
2505		*     return hasLock || (hasLock = lock.tryLock());
2506		*   }
2507		* }}</pre>
2508	+	*
2509	+	* <p>Here is a class that possibly blocks waiting for an
2510	+	* item on a given queue:
2511	+	*  <pre> {@code
2512	+	* class QueueTaker<E> implements ManagedBlocker {
2513	+	*   final BlockingQueue<E> queue;
2514	+	*   volatile E item = null;
2515	+	*   QueueTaker(BlockingQueue<E> q) { this.queue = q; }
2516	+	*   public boolean block() throws InterruptedException {
2517	+	*     if (item == null)
2518	+	*       item = queue.take();
2519	+	*     return true;
2520	+	*   }
2521	+	*   public boolean isReleasable() {
2522	+	*     return item != null || (item = queue.poll()) != null;
2523	+	*   }
2524	+	*   public E getItem() { // call after pool.managedBlock completes
2525	+	*     return item;
2526	+	*   }
2527	+	* }}</pre>
2528		*/
2529		public static interface ManagedBlocker {
2530		/**
#	Line 1954 | Line 2548 | public class ForkJoinPool extends Abstra
2548		* Blocks in accord with the given blocker.  If the current thread
2549		* is a {@link ForkJoinWorkerThread}, this method possibly
2550		* arranges for a spare thread to be activated if necessary to
2551	<	* ensure parallelism while the current thread is blocked.
1958	<	*
1959	<	* <p>If {@code maintainParallelism} is {@code true} and the pool
1960	<	* supports it ({@link #getMaintainsParallelism}), this method
1961	<	* attempts to maintain the pool's nominal parallelism. Otherwise
1962	<	* it activates a thread only if necessary to avoid complete
1963	<	* starvation. This option may be preferable when blockages use
1964	<	* timeouts, or are almost always brief.
2551	>	* ensure sufficient parallelism while the current thread is blocked.
2552		*
2553		* <p>If the caller is not a {@link ForkJoinTask}, this method is
2554		* behaviorally equivalent to
2555	<	*  <pre> {@code
2555	>	a     *  <pre> {@code
2556		* while (!blocker.isReleasable())
2557		*   if (blocker.block())
2558		*     return;
#	Line 1975 | Line 2562 | public class ForkJoinPool extends Abstra
2562		* first be expanded to ensure parallelism, and later adjusted.
2563		*
2564		* @param blocker the blocker
1978	–	* @param maintainParallelism if {@code true} and supported by
1979	–	* this pool, attempt to maintain the pool's nominal parallelism;
1980	–	* otherwise activate a thread only if necessary to avoid
1981	–	* complete starvation.
2565		* @throws InterruptedException if blocker.block did so
2566		*/
2567	<	public static void managedBlock(ManagedBlocker blocker,
1985	<	boolean maintainParallelism)
2567	>	public static void managedBlock(ManagedBlocker blocker)
2568		throws InterruptedException {
2569		Thread t = Thread.currentThread();
2570	<	if (t instanceof ForkJoinWorkerThread)
2571	<	((ForkJoinWorkerThread) t).pool.
2572	<	awaitBlocker(blocker, maintainParallelism);
2573	<	else
2574	<	awaitBlocker(blocker);
2575	<	}
2576	<
2577	<	/**
2578	<	* Performs Non-FJ blocking
2579	<	*/
2580	<	private static void awaitBlocker(ManagedBlocker blocker)
2581	<	throws InterruptedException {
2582	<	do {} while (!blocker.isReleasable() && !blocker.block());
2570	>	ForkJoinPool p = ((t instanceof ForkJoinWorkerThread) ?
2571	>	((ForkJoinWorkerThread)t).pool : null);
2572	>	while (!blocker.isReleasable()) {
2573	>	if (p == null || p.tryCompensate()) {
2574	>	try {
2575	>	do {} while (!blocker.isReleasable() && !blocker.block());
2576	>	} finally {
2577	>	if (p != null)
2578	>	p.incrementActiveCount();
2579	>	}
2580	>	break;
2581	>	}
2582	>	}
2583		}
2584
2585		// AbstractExecutorService overrides.  These rely on undocumented
#	Line 2013 | Line 2595 | public class ForkJoinPool extends Abstra
2595		}
2596
2597		// Unsafe mechanics
2598	<
2599	<	private static final sun.misc.Unsafe UNSAFE = getUnsafe();
2600	<	private static final long workerCountsOffset =
2601	<	objectFieldOffset("workerCounts", ForkJoinPool.class);
2602	<	private static final long runStateOffset =
2603	<	objectFieldOffset("runState", ForkJoinPool.class);
2604	<	private static final long eventCountOffset =
2605	<	objectFieldOffset("eventCount", ForkJoinPool.class);
2606	<	private static final long eventWaitersOffset =
2607	<	objectFieldOffset("eventWaiters",ForkJoinPool.class);
2608	<	private static final long stealCountOffset =
2027	<	objectFieldOffset("stealCount",ForkJoinPool.class);
2028	<
2029	<
2030	<	private static long objectFieldOffset(String field, Class<?> klazz) {
2598	>	private static final sun.misc.Unsafe U;
2599	>	private static final long CTL;
2600	>	private static final long RUNSTATE;
2601	>	private static final long PARKBLOCKER;
2602	>
2603	>	static {
2604	>	poolNumberGenerator = new AtomicInteger();
2605	>	modifyThreadPermission = new RuntimePermission("modifyThread");
2606	>	defaultForkJoinWorkerThreadFactory =
2607	>	new DefaultForkJoinWorkerThreadFactory();
2608	>	int s;
2609		try {
2610	<	return UNSAFE.objectFieldOffset(klazz.getDeclaredField(field));
2611	<	} catch (NoSuchFieldException e) {
2612	<	// Convert Exception to corresponding Error
2613	<	NoSuchFieldError error = new NoSuchFieldError(field);
2614	<	error.initCause(e);
2615	<	throw error;
2610	>	U = getUnsafe();
2611	>	Class<?> k = ForkJoinPool.class;
2612	>	Class<?> tk = Thread.class;
2613	>	CTL = U.objectFieldOffset
2614	>	(k.getDeclaredField("ctl"));
2615	>	RUNSTATE = U.objectFieldOffset
2616	>	(k.getDeclaredField("runState"));
2617	>	PARKBLOCKER = U.objectFieldOffset
2618	>	(tk.getDeclaredField("parkBlocker"));
2619	>	} catch (Exception e) {
2620	>	throw new Error(e);
2621		}
2622		}
2623

Diff Legend

–	Removed lines
+	Added lines
<	Changed lines
>	Changed lines