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 1.55 by dl, Sun Apr 18 13:59:57 2010 UTC vs.
Revision 1.115 by jsr166, Thu Jan 26 19:10:27 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	–
9		import java.util.ArrayList;
10		import java.util.Arrays;
11		import java.util.Collection;
12		import java.util.Collections;
13		import java.util.List;
14	<	import java.util.concurrent.locks.LockSupport;
15	<	import java.util.concurrent.locks.ReentrantLock;
14	>	import java.util.Random;
15	>	import java.util.concurrent.AbstractExecutorService;
16	>	import java.util.concurrent.Callable;
17	>	import java.util.concurrent.ExecutorService;
18	>	import java.util.concurrent.Future;
19	>	import java.util.concurrent.RejectedExecutionException;
20	>	import java.util.concurrent.RunnableFuture;
21	>	import java.util.concurrent.TimeUnit;
22		import java.util.concurrent.atomic.AtomicInteger;
23	<	import java.util.concurrent.CountDownLatch;
23	>	import java.util.concurrent.atomic.AtomicLong;
24	>	import java.util.concurrent.locks.ReentrantLock;
25	>	import java.util.concurrent.locks.Condition;
26
27		/**
28		* An {@link ExecutorService} for running {@link ForkJoinTask}s.
29		* A {@code ForkJoinPool} provides the entry point for submissions
30	<	* from non-{@code ForkJoinTask}s, as well as management and
30	>	* from non-{@code ForkJoinTask} clients, as well as management and
31		* monitoring operations.
32		*
33		* <p>A {@code ForkJoinPool} differs from other kinds of {@link
34		* ExecutorService} mainly by virtue of employing
35		* <em>work-stealing</em>: all threads in the pool attempt to find and
36	<	* execute subtasks created by other active tasks (eventually blocking
37	<	* waiting for work if none exist). This enables efficient processing
38	<	* when most tasks spawn other subtasks (as do most {@code
39	<	* ForkJoinTask}s). A {@code ForkJoinPool} may also be used for mixed
40	<	* execution of some plain {@code Runnable}- or {@code Callable}-
41	<	* based activities along with {@code ForkJoinTask}s. When setting
42	<	* {@linkplain #setAsyncMode async mode}, a {@code ForkJoinPool} may
43	<	* 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.
36	>	* execute tasks submitted to the pool and/or created by other active
37	>	* tasks (eventually blocking waiting for work if none exist). This
38	>	* enables efficient processing when most tasks spawn other subtasks
39	>	* (as do most {@code ForkJoinTask}s), as well as when many small
40	>	* tasks are submitted to the pool from external clients.  Especially
41	>	* when setting <em>asyncMode</em> to true in constructors, {@code
42	>	* ForkJoinPool}s may also be appropriate for use with event-style
43	>	* tasks that are never joined.
44		*
45		* <p>A {@code ForkJoinPool} is constructed with a given target
46		* parallelism level; by default, equal to the number of available
47	<	* processors. Unless configured otherwise via {@link
48	<	* #setMaintainsParallelism}, the pool attempts to maintain this
49	<	* number of active (or available) threads by dynamically adding,
50	<	* suspending, or resuming internal worker threads, even if some tasks
51	<	* are stalled waiting to join others. However, no such adjustments
52	<	* are performed in the face of blocked IO or other unmanaged
53	<	* 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.
47	>	* processors. The pool attempts to maintain enough active (or
48	>	* available) threads by dynamically adding, suspending, or resuming
49	>	* internal worker threads, even if some tasks are stalled waiting to
50	>	* join others. However, no such adjustments are guaranteed in the
51	>	* face of blocked IO or other unmanaged synchronization. The nested
52	>	* {@link ManagedBlocker} interface enables extension of the kinds of
53	>	* synchronization accommodated.
54		*
55		* <p>In addition to execution and lifecycle control methods, this
56		* class provides status check methods (for example
#	Line 65 | Line 59 | import java.util.concurrent.CountDownLat
59		* {@link #toString} returns indications of pool state in a
60		* convenient form for informal monitoring.
61		*
62	+	* <p> As is the case with other ExecutorServices, there are three
63	+	* main task execution methods summarized in the following
64	+	* table. These are designed to be used primarily by clients not
65	+	* already engaged in fork/join computations in the current pool.  The
66	+	* main forms of these methods accept instances of {@code
67	+	* ForkJoinTask}, but overloaded forms also allow mixed execution of
68	+	* plain {@code Runnable}- or {@code Callable}- based activities as
69	+	* well.  However, tasks that are already executing in a pool should
70	+	* normally instead use the within-computation forms listed in the
71	+	* table unless using async event-style tasks that are not usually
72	+	* joined, in which case there is little difference among choice of
73	+	* methods.
74	+	*
75	+	* <table BORDER CELLPADDING=3 CELLSPACING=1>
76	+	*  <tr>
77	+	*    <td></td>
78	+	*    <td ALIGN=CENTER> <b>Call from non-fork/join clients</b></td>
79	+	*    <td ALIGN=CENTER> <b>Call from within fork/join computations</b></td>
80	+	*  </tr>
81	+	*  <tr>
82	+	*    <td> <b>Arrange async execution</td>
83	+	*    <td> {@link #execute(ForkJoinTask)}</td>
84	+	*    <td> {@link ForkJoinTask#fork}</td>
85	+	*  </tr>
86	+	*  <tr>
87	+	*    <td> <b>Await and obtain result</td>
88	+	*    <td> {@link #invoke(ForkJoinTask)}</td>
89	+	*    <td> {@link ForkJoinTask#invoke}</td>
90	+	*  </tr>
91	+	*  <tr>
92	+	*    <td> <b>Arrange exec and obtain Future</td>
93	+	*    <td> {@link #submit(ForkJoinTask)}</td>
94	+	*    <td> {@link ForkJoinTask#fork} (ForkJoinTasks <em>are</em> Futures)</td>
95	+	*  </tr>
96	+	* </table>
97	+	*
98		* <p><b>Sample Usage.</b> Normally a single {@code ForkJoinPool} is
99		* used for all parallel task execution in a program or subsystem.
100		* Otherwise, use would not usually outweigh the construction and
#	Line 75 | Line 105 | import java.util.concurrent.CountDownLat
105		* daemon} mode, there is typically no need to explicitly {@link
106		* #shutdown} such a pool upon program exit.
107		*
108	<	* <pre>
108	>	*  <pre> {@code
109		* static final ForkJoinPool mainPool = new ForkJoinPool();
110		* ...
111		* public void sort(long[] array) {
112		*   mainPool.invoke(new SortTask(array, 0, array.length));
113	<	* }
84	<	* </pre>
113	>	* }}</pre>
114		*
115		* <p><b>Implementation notes</b>: This implementation restricts the
116		* maximum number of running threads to 32767. Attempts to create
#	Line 89 | Line 118 | import java.util.concurrent.CountDownLat
118		* {@code IllegalArgumentException}.
119		*
120		* <p>This implementation rejects submitted tasks (that is, by throwing
121	<	* {@link RejectedExecutionException}) only when the pool is shut down.
121	>	* {@link RejectedExecutionException}) only when the pool is shut down
122	>	* or internal resources have been exhausted.
123		*
124		* @since 1.7
125		* @author Doug Lea
#	Line 99 | Line 129 | public class ForkJoinPool extends Abstra
129		/*
130		* Implementation Overview
131		*
132	<	* This class provides the central bookkeeping and control for a
133	<	* set of worker threads: Submissions from non-FJ threads enter
134	<	* into a submission queue. Workers take these tasks and typically
135	<	* split them into subtasks that may be stolen by other workers.
136	<	* The main work-stealing mechanics implemented in class
137	<	* ForkJoinWorkerThread give first priority to processing tasks
138	<	* from their own queues (LIFO or FIFO, depending on mode), then
139	<	* to randomized FIFO steals of tasks in other worker queues, and
140	<	* lastly to new submissions. These mechanics do not consider
141	<	* affinities, loads, cache localities, etc, so rarely provide the
142	<	* best possible performance on a given machine, but portably
143	<	* provide good throughput by averaging over these factors.
144	<	* (Further, even if we did try to use such information, we do not
145	<	* usually have a basis for exploiting it. For example, some sets
146	<	* of tasks profit from cache affinities, but others are harmed by
147	<	* cache pollution effects.)
132	>	* This class and its nested classes provide the main
133	>	* functionality and control for a set of worker threads:
134	>	* Submissions from non-FJ threads enter into submission
135	>	* queues. Workers take these tasks and typically split them into
136	>	* subtasks that may be stolen by other workers.  Preference rules
137	>	* give first priority to processing tasks from their own queues
138	>	* (LIFO or FIFO, depending on mode), then to randomized FIFO
139	>	* steals of tasks in other queues.
140	>	*
141	>	* WorkQueues.
142	>	* ==========
143	>	*
144	>	* Most operations occur within work-stealing queues (in nested
145	>	* class WorkQueue).  These are special forms of Deques that
146	>	* support only three of the four possible end-operations -- push,
147	>	* pop, and poll (aka steal), under the further constraints that
148	>	* push and pop are called only from the owning thread (or, as
149	>	* extended here, under a lock), while poll may be called from
150	>	* other threads.  (If you are unfamiliar with them, you probably
151	>	* want to read Herlihy and Shavit's book "The Art of
152	>	* Multiprocessor programming", chapter 16 describing these in
153	>	* more detail before proceeding.)  The main work-stealing queue
154	>	* design is roughly similar to those in the papers "Dynamic
155	>	* Circular Work-Stealing Deque" by Chase and Lev, SPAA 2005
156	>	* (http://research.sun.com/scalable/pubs/index.html) and
157	>	* "Idempotent work stealing" by Michael, Saraswat, and Vechev,
158	>	* PPoPP 2009 (http://portal.acm.org/citation.cfm?id=1504186).
159	>	* The main differences ultimately stem from gc requirements that
160	>	* we null out taken slots as soon as we can, to maintain as small
161	>	* a footprint as possible even in programs generating huge
162	>	* numbers of tasks. To accomplish this, we shift the CAS
163	>	* arbitrating pop vs poll (steal) from being on the indices
164	>	* ("base" and "top") to the slots themselves.  So, both a
165	>	* successful pop and poll mainly entail a CAS of a slot from
166	>	* non-null to null.  Because we rely on CASes of references, we
167	>	* do not need tag bits on base or top.  They are simple ints as
168	>	* used in any circular array-based queue (see for example
169	>	* ArrayDeque).  Updates to the indices must still be ordered in a
170	>	* way that guarantees that top == base means the queue is empty,
171	>	* but otherwise may err on the side of possibly making the queue
172	>	* appear nonempty when a push, pop, or poll have not fully
173	>	* committed. Note that this means that the poll operation,
174	>	* considered individually, is not wait-free. One thief cannot
175	>	* successfully continue until another in-progress one (or, if
176	>	* previously empty, a push) completes.  However, in the
177	>	* aggregate, we ensure at least probabilistic non-blockingness.
178	>	* If an attempted steal fails, a thief always chooses a different
179	>	* random victim target to try next. So, in order for one thief to
180	>	* progress, it suffices for any in-progress poll or new push on
181	>	* any empty queue to complete.
182	>	*
183	>	* This approach also enables support of a user mode in which local
184	>	* task processing is in FIFO, not LIFO order, simply by using
185	>	* poll rather than pop.  This can be useful in message-passing
186	>	* frameworks in which tasks are never joined.  However neither
187	>	* mode considers affinities, loads, cache localities, etc, so
188	>	* rarely provide the best possible performance on a given
189	>	* machine, but portably provide good throughput by averaging over
190	>	* these factors.  (Further, even if we did try to use such
191	>	* information, we do not usually have a basis for exploiting
192	>	* it. For example, some sets of tasks profit from cache
193	>	* affinities, but others are harmed by cache pollution effects.)
194	>	*
195	>	* WorkQueues are also used in a similar way for tasks submitted
196	>	* to the pool. We cannot mix these tasks in the same queues used
197	>	* for work-stealing (this would contaminate lifo/fifo
198	>	* processing). Instead, we loosely associate (via hashing)
199	>	* submission queues with submitting threads, and randomly scan
200	>	* these queues as well when looking for work. In essence,
201	>	* submitters act like workers except that they never take tasks,
202	>	* and they are multiplexed on to a finite number of shared work
203	>	* queues. However, classes are set up so that future extensions
204	>	* could allow submitters to optionally help perform tasks as
205	>	* well. Pool submissions from internal workers are also allowed,
206	>	* but use randomized rather than thread-hashed queue indices to
207	>	* avoid imbalance.  Insertion of tasks in shared mode requires a
208	>	* lock (mainly to protect in the case of resizing) but we use
209	>	* only a simple spinlock (using bits in field runState), because
210	>	* submitters encountering a busy queue try or create others so
211	>	* never block.
212	>	*
213	>	* Management.
214	>	* ==========
215		*
216		* The main throughput advantages of work-stealing stem from
217	<	* decentralized control -- workers mostly steal tasks from each
218	<	* other. We do not want to negate this by creating bottlenecks
219	<	* implementing the management responsibilities of this class. So
220	<	* we use a collection of techniques that avoid, reduce, or cope
221	<	* well with contention. These entail several instances of
222	<	* bit-packing into CASable fields to maintain only the minimally
223	<	* required atomicity. To enable such packing, we restrict maximum
224	<	* parallelism to (1<<15)-1 (enabling twice this to fit into a 16
225	<	* bit field), which is far in excess of normal operating range.
226	<	* Even though updates to some of these bookkeeping fields do
227	<	* sometimes contend with each other, they don't normally
228	<	* cache-contend with updates to others enough to warrant memory
229	<	* padding or isolation. So they are all held as fields of
230	<	* ForkJoinPool objects.  The main capabilities are as follows:
231	<	*
232	<	* 1. Creating and removing workers. Workers are recorded in the
233	<	* "workers" array. This is an array as opposed to some other data
234	<	* structure to support index-based random steals by workers.
235	<	* Updates to the array recording new workers and unrecording
236	<	* terminated ones are protected from each other by a lock
237	<	* (workerLock) but the array is otherwise concurrently readable,
238	<	* and accessed directly by workers. To simplify index-based
217	>	* decentralized control -- workers mostly take tasks from
218	>	* themselves or each other. We cannot negate this in the
219	>	* implementation of other management responsibilities. The main
220	>	* tactic for avoiding bottlenecks is packing nearly all
221	>	* essentially atomic control state into two volatile variables
222	>	* that are by far most often read (not written) as status and
223	>	* consistency checks
224	>	*
225	>	* Field "ctl" contains 64 bits holding all the information needed
226	>	* to atomically decide to add, inactivate, enqueue (on an event
227	>	* queue), dequeue, and/or re-activate workers.  To enable this
228	>	* packing, we restrict maximum parallelism to (1<<15)-1 (which is
229	>	* far in excess of normal operating range) to allow ids, counts,
230	>	* and their negations (used for thresholding) to fit into 16bit
231	>	* fields.
232	>	*
233	>	* Field "runState" contains 32 bits needed to register and
234	>	* deregister WorkQueues, as well as to enable shutdown. It is
235	>	* only modified under a lock (normally briefly held, but
236	>	* occasionally protecting allocations and resizings) but even
237	>	* when locked remains available to check consistency.
238	>	*
239	>	* Recording WorkQueues.  WorkQueues are recorded in the
240	>	* "workQueues" array that is created upon pool construction and
241	>	* expanded if necessary.  Updates to the array while recording
242	>	* new workers and unrecording terminated ones are protected from
243	>	* each other by a lock but the array is otherwise concurrently
244	>	* readable, and accessed directly.  To simplify index-based
245		* operations, the array size is always a power of two, and all
246	<	* readers must tolerate null slots. Currently, all but the first
247	<	* worker thread creation is on-demand, triggered by task
248	<	* submissions, replacement of terminated workers, and/or
246	>	* readers must tolerate null slots. Shared (submission) queues
247	>	* are at even indices, worker queues at odd indices. Grouping
248	>	* them together in this way simplifies and speeds up task
249	>	* scanning. To avoid flailing during start-up, the array is
250	>	* presized to hold twice #parallelism workers (which is unlikely
251	>	* to need further resizing during execution). But to avoid
252	>	* dealing with so many null slots, variable runState includes a
253	>	* mask for the nearest power of two that contains all current
254	>	* workers.  All worker thread creation is on-demand, triggered by
255	>	* task submissions, replacement of terminated workers, and/or
256		* compensation for blocked workers. However, all other support
257	<	* code is set up to work with other policies.
257	>	* code is set up to work with other policies.  To ensure that we
258	>	* do not hold on to worker references that would prevent GC, ALL
259	>	* accesses to workQueues are via indices into the workQueues
260	>	* array (which is one source of some of the messy code
261	>	* constructions here). In essence, the workQueues array serves as
262	>	* a weak reference mechanism. Thus for example the wait queue
263	>	* field of ctl stores indices, not references.  Access to the
264	>	* workQueues in associated methods (for example signalWork) must
265	>	* both index-check and null-check the IDs. All such accesses
266	>	* ignore bad IDs by returning out early from what they are doing,
267	>	* since this can only be associated with termination, in which
268	>	* case it is OK to give up.
269	>	*
270	>	* All uses of the workQueues array check that it is non-null
271	>	* (even if previously non-null). This allows nulling during
272	>	* termination, which is currently not necessary, but remains an
273	>	* option for resource-revocation-based shutdown schemes. It also
274	>	* helps reduce JIT issuance of uncommon-trap code, which tends to
275	>	* unnecessarily complicate control flow in some methods.
276	>	*
277	>	* Event Queuing. Unlike HPC work-stealing frameworks, we cannot
278	>	* let workers spin indefinitely scanning for tasks when none can
279	>	* be found immediately, and we cannot start/resume workers unless
280	>	* there appear to be tasks available.  On the other hand, we must
281	>	* quickly prod them into action when new tasks are submitted or
282	>	* generated. In many usages, ramp-up time to activate workers is
283	>	* the main limiting factor in overall performance (this is
284	>	* compounded at program start-up by JIT compilation and
285	>	* allocation). So we try to streamline this as much as possible.
286	>	* We park/unpark workers after placing in an event wait queue
287	>	* when they cannot find work. This "queue" is actually a simple
288	>	* Treiber stack, headed by the "id" field of ctl, plus a 15bit
289	>	* counter value (that reflects the number of times a worker has
290	>	* been inactivated) to avoid ABA effects (we need only as many
291	>	* version numbers as worker threads). Successors are held in
292	>	* field WorkQueue.nextWait.  Queuing deals with several intrinsic
293	>	* races, mainly that a task-producing thread can miss seeing (and
294	>	* signalling) another thread that gave up looking for work but
295	>	* has not yet entered the wait queue. We solve this by requiring
296	>	* a full sweep of all workers (via repeated calls to method
297	>	* scan()) both before and after a newly waiting worker is added
298	>	* to the wait queue. During a rescan, the worker might release
299	>	* some other queued worker rather than itself, which has the same
300	>	* net effect. Because enqueued workers may actually be rescanning
301	>	* rather than waiting, we set and clear the "parker" field of
302	>	* Workqueues to reduce unnecessary calls to unpark.  (This
303	>	* requires a secondary recheck to avoid missed signals.)  Note
304	>	* the unusual conventions about Thread.interrupts surrounding
305	>	* parking and other blocking: Because interrupts are used solely
306	>	* to alert threads to check termination, which is checked anyway
307	>	* upon blocking, we clear status (using Thread.interrupted)
308	>	* before any call to park, so that park does not immediately
309	>	* return due to status being set via some other unrelated call to
310	>	* interrupt in user code.
311	>	*
312	>	* Signalling.  We create or wake up workers only when there
313	>	* appears to be at least one task they might be able to find and
314	>	* execute.  When a submission is added or another worker adds a
315	>	* task to a queue that previously had fewer than two tasks, they
316	>	* signal waiting workers (or trigger creation of new ones if
317	>	* fewer than the given parallelism level -- see signalWork).
318	>	* These primary signals are buttressed by signals during rescans;
319	>	* together these cover the signals needed in cases when more
320	>	* tasks are pushed but untaken, and improve performance compared
321	>	* to having one thread wake up all workers.
322	>	*
323	>	* Trimming workers. To release resources after periods of lack of
324	>	* use, a worker starting to wait when the pool is quiescent will
325	>	* time out and terminate if the pool has remained quiescent for
326	>	* SHRINK_RATE nanosecs. This will slowly propagate, eventually
327	>	* terminating all workers after long periods of non-use.
328	>	*
329	>	* Shutdown and Termination. A call to shutdownNow atomically sets
330	>	* a runState bit and then (non-atomically) sets each workers
331	>	* runState status, cancels all unprocessed tasks, and wakes up
332	>	* all waiting workers.  Detecting whether termination should
333	>	* commence after a non-abrupt shutdown() call requires more work
334	>	* and bookkeeping. We need consensus about quiescence (i.e., that
335	>	* there is no more work). The active count provides a primary
336	>	* indication but non-abrupt shutdown still requires a rechecking
337	>	* scan for any workers that are inactive but not queued.
338	>	*
339	>	* Joining Tasks.
340	>	* ==============
341	>	*
342	>	* Any of several actions may be taken when one worker is waiting
343	>	* to join a task stolen (or always held by) another.  Because we
344	>	* are multiplexing many tasks on to a pool of workers, we can't
345	>	* just let them block (as in Thread.join).  We also cannot just
346	>	* reassign the joiner's run-time stack with another and replace
347	>	* it later, which would be a form of "continuation", that even if
348	>	* possible is not necessarily a good idea since we sometimes need
349	>	* both an unblocked task and its continuation to
350	>	* progress. Instead we combine two tactics:
351	>	*
352	>	*   Helping: Arranging for the joiner to execute some task that it
353	>	*      would be running if the steal had not occurred.
354	>	*
355	>	*   Compensating: Unless there are already enough live threads,
356	>	*      method tryCompensate() may create or re-activate a spare
357	>	*      thread to compensate for blocked joiners until they unblock.
358	>	*
359	>	* A third form (implemented in tryRemoveAndExec and
360	>	* tryPollForAndExec) amounts to helping a hypothetical
361	>	* compensator: If we can readily tell that a possible action of a
362	>	* compensator is to steal and execute the task being joined, the
363	>	* joining thread can do so directly, without the need for a
364	>	* compensation thread (although at the expense of larger run-time
365	>	* stacks, but the tradeoff is typically worthwhile).
366	>	*
367	>	* The ManagedBlocker extension API can't use helping so relies
368	>	* only on compensation in method awaitBlocker.
369	>	*
370	>	* The algorithm in tryHelpStealer entails a form of "linear"
371	>	* helping: Each worker records (in field currentSteal) the most
372	>	* recent task it stole from some other worker. Plus, it records
373	>	* (in field currentJoin) the task it is currently actively
374	>	* joining. Method tryHelpStealer uses these markers to try to
375	>	* find a worker to help (i.e., steal back a task from and execute
376	>	* it) that could hasten completion of the actively joined task.
377	>	* In essence, the joiner executes a task that would be on its own
378	>	* local deque had the to-be-joined task not been stolen. This may
379	>	* be seen as a conservative variant of the approach in Wagner &
380	>	* Calder "Leapfrogging: a portable technique for implementing
381	>	* efficient futures" SIGPLAN Notices, 1993
382	>	* (http://portal.acm.org/citation.cfm?id=155354). It differs in
383	>	* that: (1) We only maintain dependency links across workers upon
384	>	* steals, rather than use per-task bookkeeping.  This sometimes
385	>	* requires a linear scan of workers array to locate stealers, but
386	>	* often doesn't because stealers leave hints (that may become
387	>	* stale/wrong) of where to locate them.  A stealHint is only a
388	>	* hint because a worker might have had multiple steals and the
389	>	* hint records only one of them (usually the most current).
390	>	* Hinting isolates cost to when it is needed, rather than adding
391	>	* to per-task overhead.  (2) It is "shallow", ignoring nesting
392	>	* and potentially cyclic mutual steals.  (3) It is intentionally
393	>	* racy: field currentJoin is updated only while actively joining,
394	>	* which means that we miss links in the chain during long-lived
395	>	* tasks, GC stalls etc (which is OK since blocking in such cases
396	>	* is usually a good idea).  (4) We bound the number of attempts
397	>	* to find work (see MAX_HELP_DEPTH) and fall back to suspending
398	>	* the worker and if necessary replacing it with another.
399	>	*
400	>	* It is impossible to keep exactly the target parallelism number
401	>	* of threads running at any given time.  Determining the
402	>	* existence of conservatively safe helping targets, the
403	>	* availability of already-created spares, and the apparent need
404	>	* to create new spares are all racy, so we rely on multiple
405	>	* retries of each.  Currently, in keeping with on-demand
406	>	* signalling policy, we compensate only if blocking would leave
407	>	* less than one active (non-waiting, non-blocked) worker.
408	>	* Additionally, to avoid some false alarms due to GC, lagging
409	>	* counters, system activity, etc, compensated blocking for joins
410	>	* is only attempted after rechecks stabilize in
411	>	* ForkJoinTask.awaitJoin. (Retries are interspersed with
412	>	* Thread.yield, for good citizenship.)
413		*
414	<	* 2. Bookkeeping for dynamically adding and removing workers. We
150	<	* maintain a given level of parallelism (or, if
151	<	* maintainsParallelism is false, at least avoid starvation). When
152	<	* some workers are known to be blocked (on joins or via
153	<	* ManagedBlocker), we may create or resume others to take their
154	<	* place until they unblock (see below). Implementing this
155	<	* requires counts of the number of "running" threads (i.e., those
156	<	* that are neither blocked nor artifically suspended) as well as
157	<	* the total number.  These two values are packed into one field,
158	<	* "workerCounts" because we need accurate snapshots when deciding
159	<	* to create, resume or suspend.  To support these decisions,
160	<	* updates must be prospective (not retrospective).  For example,
161	<	* the running count is decremented before blocking by a thread
162	<	* about to block, but incremented by the thread about to unblock
163	<	* it. (In a few cases, these prospective updates may need to be
164	<	* rolled back, for example when deciding to create a new worker
165	<	* but the thread factory fails or returns null. In these cases,
166	<	* we are no worse off wrt other decisions than we would be
167	<	* otherwise.)  Updates to the workerCounts field sometimes
168	<	* transiently encounter a fair amount of contention when join
169	<	* dependencies are such that many threads block or unblock at
170	<	* about the same time. We alleviate this by sometimes bundling
171	<	* updates (for example blocking one thread on join and resuming a
172	<	* spare cancel each other out), and in most other cases
173	<	* performing an alternative action (like releasing waiters and
174	<	* finding spares; see below) as a more productive form of
175	<	* backoff.
176	<	*
177	<	* 3. Maintaining global run state. The run state of the pool
178	<	* consists of a runLevel (SHUTDOWN, TERMINATING, etc) similar to
179	<	* those in other Executor implementations, as well as a count of
180	<	* "active" workers -- those that are, or soon will be, or
181	<	* recently were executing tasks. The runLevel and active count
182	<	* are packed together in order to correctly trigger shutdown and
183	<	* termination. Without care, active counts can be subject to very
184	<	* high contention.  We substantially reduce this contention by
185	<	* relaxing update rules.  A worker must claim active status
186	<	* prospectively, by activating if it sees that a submitted or
187	<	* stealable task exists (it may find after activating that the
188	<	* task no longer exists). It stays active while processing this
189	<	* task (if it exists) and any other local subtasks it produces,
190	<	* until it cannot find any other tasks. It then tries
191	<	* inactivating (see method preStep), but upon update contention
192	<	* instead scans for more tasks, later retrying inactivation if it
193	<	* doesn't find any.
194	<	*
195	<	* 4. Managing idle workers waiting for tasks. We cannot let
196	<	* workers spin indefinitely scanning for tasks when none are
197	<	* available. On the other hand, we must quickly prod them into
198	<	* action when new tasks are submitted or generated.  We
199	<	* park/unpark these idle workers using an event-count scheme.
200	<	* Field eventCount is incremented upon events that may enable
201	<	* workers that previously could not find a task to now find one:
202	<	* Submission of a new task to the pool, or another worker pushing
203	<	* a task onto a previously empty queue.  (We also use this
204	<	* mechanism for termination and reconfiguration actions that
205	<	* require wakeups of idle workers).  Each worker maintains its
206	<	* last known event count, and blocks when a scan for work did not
207	<	* find a task AND its lastEventCount matches the current
208	<	* eventCount. Waiting idle workers are recorded in a variant of
209	<	* Treiber stack headed by field eventWaiters which, when nonzero,
210	<	* encodes the thread index and count awaited for by the worker
211	<	* thread most recently calling eventSync. This thread in turn has
212	<	* a record (field nextEventWaiter) for the next waiting worker.
213	<	* In addition to allowing simpler decisions about need for
214	<	* wakeup, the event count bits in eventWaiters serve the role of
215	<	* tags to avoid ABA errors in Treiber stacks.  To reduce delays
216	<	* in task diffusion, workers not otherwise occupied may invoke
217	<	* method releaseWaiters, that removes and signals (unparks)
218	<	* workers not waiting on current count. To minimize task
219	<	* production stalls associate with signalling, any worker pushing
220	<	* a task on an empty queue invokes the weaker method signalWork,
221	<	* that only releases idle workers until it detects interference
222	<	* by other threads trying to release, and lets them take
223	<	* over. The net effect is a tree-like diffusion of signals, where
224	<	* released threads and possibly others) help with unparks.  To
225	<	* further reduce contention effects a bit, failed CASes to
226	<	* increment field eventCount are tolerated without retries.
227	<	* Conceptually they are merged into the same event, which is OK
228	<	* when their only purpose is to enable workers to scan for work.
229	<	*
230	<	* 5. Managing suspension of extra workers. When a worker is about
231	<	* to block waiting for a join (or via ManagedBlockers), we may
232	<	* create a new thread to maintain parallelism level, or at least
233	<	* avoid starvation (see below). Usually, extra threads are needed
234	<	* for only very short periods, yet join dependencies are such
235	<	* that we sometimes need them in bursts. Rather than create new
236	<	* threads each time this happens, we suspend no-longer-needed
237	<	* extra ones as "spares". For most purposes, we don't distinguish
238	<	* "extra" spare threads from normal "core" threads: On each call
239	<	* to preStep (the only point at which we can do this) a worker
240	<	* checks to see if there are now too many running workers, and if
241	<	* so, suspends itself.  Methods preJoin and doBlock look for
242	<	* suspended threads to resume before considering creating a new
243	<	* replacement. We don't need a special data structure to maintain
244	<	* spares; simply scanning the workers array looking for
245	<	* worker.isSuspended() is fine because the calling thread is
246	<	* otherwise not doing anything useful anyway; we are at least as
247	<	* happy if after locating a spare, the caller doesn't actually
248	<	* block because the join is ready before we try to adjust and
249	<	* compensate.  Note that this is intrinsically racy.  One thread
250	<	* may become a spare at about the same time as another is
251	<	* needlessly being created. We counteract this and related slop
252	<	* in part by requiring resumed spares to immediately recheck (in
253	<	* preStep) to see whether they they should re-suspend. The only
254	<	* effective difference between "extra" and "core" threads is that
255	<	* we allow the "extra" ones to time out and die if they are not
256	<	* resumed within a keep-alive interval of a few seconds. This is
257	<	* implemented mainly within ForkJoinWorkerThread, but requires
258	<	* some coordination (isTrimmed() -- meaning killed while
259	<	* suspended) to correctly maintain pool counts.
260	<	*
261	<	* 6. Deciding when to create new workers. The main dynamic
262	<	* control in this class is deciding when to create extra threads,
263	<	* in methods preJoin and doBlock. We always need to create one
264	<	* when the number of running threads becomes zero. But because
265	<	* blocked joins are typically dependent, we don't necessarily
266	<	* need or want one-to-one replacement. Using a one-to-one
267	<	* compensation rule often leads to enough useless overhead
268	<	* creating, suspending, resuming, and/or killing threads to
269	<	* signficantly degrade throughput.  We use a rule reflecting the
270	<	* idea that, the more spare threads you already have, the more
271	<	* evidence you need to create another one; where "evidence" is
272	<	* expressed as the current deficit -- target minus running
273	<	* threads. To reduce flickering and drift around target values,
274	<	* the relation is quadratic: adding a spare if (dc*dc)>=(sc*pc)
275	<	* (where dc is deficit, sc is number of spare threads and pc is
276	<	* target parallelism.)  This effectively reduces churn at the
277	<	* price of systematically undershooting target parallelism when
278	<	* many threads are blocked.  However, biasing toward undeshooting
279	<	* partially compensates for the above mechanics to suspend extra
280	<	* threads, that normally lead to overshoot because we can only
281	<	* suspend workers in-between top-level actions. It also better
282	<	* copes with the fact that some of the methods in this class tend
283	<	* to never become compiled (but are interpreted), so some
284	<	* components of the entire set of controls might execute many
285	<	* times faster than others. And similarly for cases where the
286	<	* apparent lack of work is just due to GC stalls and other
287	<	* transient system activity.
288	<	*
289	<	* 7. Maintaining other configuration parameters and monitoring
290	<	* statistics. Updates to fields controlling parallelism level,
291	<	* max size, etc can only meaningfully take effect for individual
292	<	* threads upon their next top-level actions; i.e., between
293	<	* stealing/running tasks/submission, which are separated by calls
294	<	* to preStep.  Memory ordering for these (assumed infrequent)
295	<	* reconfiguration calls is ensured by using reads and writes to
296	<	* volatile field workerCounts (that must be read in preStep anyway)
297	<	* as "fences" -- user-level reads are preceded by reads of
298	<	* workCounts, and writes are followed by no-op CAS to
299	<	* workerCounts. The values reported by other management and
300	<	* monitoring methods are either computed on demand, or are kept
301	<	* in fields that are only updated when threads are otherwise
302	<	* idle.
303	<	*
304	<	* Beware that there is a lot of representation-level coupling
414	>	* Style notes: There is a lot of representation-level coupling
415		* among classes ForkJoinPool, ForkJoinWorkerThread, and
416	<	* ForkJoinTask.  For example, direct access to "workers" array by
417	<	* workers, and direct access to ForkJoinTask.status by both
418	<	* ForkJoinPool and ForkJoinWorkerThread.  There is little point
419	<	* trying to reduce this, since any associated future changes in
420	<	* representations will need to be accompanied by algorithmic
421	<	* changes anyway.
422	<	*
423	<	* Style notes: There are lots of inline assignments (of form
424	<	* "while ((local = field) != 0)") which are usually the simplest
425	<	* way to ensure read orderings. Also several occurrences of the
426	<	* unusual "do {} while(!cas...)" which is the simplest way to
427	<	* force an update of a CAS'ed variable. There are also a few
428	<	* other coding oddities that help some methods perform reasonably
429	<	* even when interpreted (not compiled).
430	<	*
431	<	* The order of declarations in this file is: (1) statics (2)
432	<	* fields (along with constants used when unpacking some of them)
433	<	* (3) internal control methods (4) callbacks and other support
434	<	* for ForkJoinTask and ForkJoinWorkerThread classes, (5) exported
435	<	* methods (plus a few little helpers).
416	>	* ForkJoinTask.  The fields of WorkQueue maintain data structures
417	>	* managed by ForkJoinPool, so are directly accessed.  There is
418	>	* little point trying to reduce this, since any associated future
419	>	* changes in representations will need to be accompanied by
420	>	* algorithmic changes anyway. All together, these low-level
421	>	* implementation choices produce as much as a factor of 4
422	>	* performance improvement compared to naive implementations, and
423	>	* enable the processing of billions of tasks per second, at the
424	>	* expense of some ugliness.
425	>	*
426	>	* Methods signalWork() and scan() are the main bottlenecks so are
427	>	* especially heavily micro-optimized/mangled.  There are lots of
428	>	* inline assignments (of form "while ((local = field) != 0)")
429	>	* which are usually the simplest way to ensure the required read
430	>	* orderings (which are sometimes critical). This leads to a
431	>	* "C"-like style of listing declarations of these locals at the
432	>	* heads of methods or blocks.  There are several occurrences of
433	>	* the unusual "do {} while (!cas...)"  which is the simplest way
434	>	* to force an update of a CAS'ed variable. There are also other
435	>	* coding oddities that help some methods perform reasonably even
436	>	* when interpreted (not compiled).
437	>	*
438	>	* The order of declarations in this file is: (1) declarations of
439	>	* statics (2) fields (along with constants used when unpacking
440	>	* some of them), listed in an order that tends to reduce
441	>	* contention among them a bit under most JVMs; (3) nested
442	>	* classes; (4) internal control methods; (5) callbacks and other
443	>	* support for ForkJoinTask methods; (6) exported methods (plus a
444	>	* few little helpers); (7) static block initializing all statics
445	>	* in a minimally dependent order.
446		*/
447
448		/**
#	Line 345 | Line 465 | public class ForkJoinPool extends Abstra
465		* Default ForkJoinWorkerThreadFactory implementation; creates a
466		* new ForkJoinWorkerThread.
467		*/
468	<	static class  DefaultForkJoinWorkerThreadFactory
468	>	static class DefaultForkJoinWorkerThreadFactory
469		implements ForkJoinWorkerThreadFactory {
470		public ForkJoinWorkerThread newThread(ForkJoinPool pool) {
471		return new ForkJoinWorkerThread(pool);
#	Line 357 | Line 477 | public class ForkJoinPool extends Abstra
477		* overridden in ForkJoinPool constructors.
478		*/
479		public static final ForkJoinWorkerThreadFactory
480	<	defaultForkJoinWorkerThreadFactory =
361	<	new DefaultForkJoinWorkerThreadFactory();
480	>	defaultForkJoinWorkerThreadFactory;
481
482		/**
483		* Permission required for callers of methods that may start or
484		* kill threads.
485		*/
486	<	private static final RuntimePermission modifyThreadPermission =
368	<	new RuntimePermission("modifyThread");
486	>	private static final RuntimePermission modifyThreadPermission;
487
488		/**
489		* If there is a security manager, makes sure caller has
#	Line 380 | Line 498 | public class ForkJoinPool extends Abstra
498		/**
499		* Generator for assigning sequence numbers as pool names.
500		*/
501	<	private static final AtomicInteger poolNumberGenerator =
384	<	new AtomicInteger();
501	>	private static final AtomicInteger poolNumberGenerator;
502
503		/**
504	<	* Absolute bound for parallelism level. Twice this number must
505	<	* fit into a 16bit field to enable word-packing for some counts.
506	<	*/
507	<	private static final int MAX_THREADS = 0x7fff;
508	<
509	<	/**
510	<	* Array holding all worker threads in the pool.  Array size must
511	<	* be a power of two.  Updates and replacements are protected by
512	<	* workerLock, but the array is always kept in a consistent enough
513	<	* state to be randomly accessed without locking by workers
514	<	* performing work-stealing, as well as other traversal-based
515	<	* methods in this class. All readers must tolerate that some
516	<	* array slots may be null.
504	>	* Bits and masks for control variables
505	>	*
506	>	* Field ctl is a long packed with:
507	>	* AC: Number of active running workers minus target parallelism (16 bits)
508	>	* TC: Number of total workers minus target parallelism (16 bits)
509	>	* ST: true if pool is terminating (1 bit)
510	>	* EC: the wait count of top waiting thread (15 bits)
511	>	* ID: ~(poolIndex >>> 1) of top of Treiber stack of waiters (16 bits)
512	>	*
513	>	* When convenient, we can extract the upper 32 bits of counts and
514	>	* the lower 32 bits of queue state, u = (int)(ctl >>> 32) and e =
515	>	* (int)ctl.  The ec field is never accessed alone, but always
516	>	* together with id and st. The offsets of counts by the target
517	>	* parallelism and the positionings of fields makes it possible to
518	>	* perform the most common checks via sign tests of fields: When
519	>	* ac is negative, there are not enough active workers, when tc is
520	>	* negative, there are not enough total workers, when id is
521	>	* negative, there is at least one waiting worker, and when e is
522	>	* negative, the pool is terminating.  To deal with these possibly
523	>	* negative fields, we use casts in and out of "short" and/or
524	>	* signed shifts to maintain signedness.
525	>	*
526	>	* When a thread is queued (inactivated), its eventCount field is
527	>	* negative, which is the only way to tell if a worker is
528	>	* prevented from executing tasks, even though it must continue to
529	>	* scan for them to avoid queuing races.
530	>	*
531	>	* Field runState is an int packed with:
532	>	* SHUTDOWN: true if shutdown is enabled (1 bit)
533	>	* SEQ:  a sequence number updated upon (de)registering workers (15 bits)
534	>	* MASK: mask (power of 2 - 1) covering all registered poolIndexes (16 bits)
535	>	*
536	>	* The combination of mask and sequence number enables simple
537	>	* consistency checks: Staleness of read-only operations on the
538	>	* workers and queues arrays can be checked by comparing runState
539	>	* before vs after the reads. The low 16 bits (i.e, anding with
540	>	* SMASK) hold (the smallest power of two covering all worker
541	>	* indices, minus one.  The mask for queues (vs workers) is twice
542	>	* this value plus 1.
543	>	*/
544	>
545	>	// bit positions/shifts for fields
546	>	private static final int  AC_SHIFT   = 48;
547	>	private static final int  TC_SHIFT   = 32;
548	>	private static final int  ST_SHIFT   = 31;
549	>	private static final int  EC_SHIFT   = 16;
550	>
551	>	// bounds
552	>	private static final int  MAX_ID     = 0x7fff;  // max poolIndex
553	>	private static final int  SMASK      = 0xffff;  // mask short bits
554	>	private static final int  SHORT_SIGN = 1 << 15;
555	>	private static final int  INT_SIGN   = 1 << 31;
556	>
557	>	// masks
558	>	private static final long STOP_BIT   = 0x0001L << ST_SHIFT;
559	>	private static final long AC_MASK    = ((long)SMASK) << AC_SHIFT;
560	>	private static final long TC_MASK    = ((long)SMASK) << TC_SHIFT;
561	>
562	>	// units for incrementing and decrementing
563	>	private static final long TC_UNIT    = 1L << TC_SHIFT;
564	>	private static final long AC_UNIT    = 1L << AC_SHIFT;
565	>
566	>	// masks and units for dealing with u = (int)(ctl >>> 32)
567	>	private static final int  UAC_SHIFT  = AC_SHIFT - 32;
568	>	private static final int  UTC_SHIFT  = TC_SHIFT - 32;
569	>	private static final int  UAC_MASK   = SMASK << UAC_SHIFT;
570	>	private static final int  UTC_MASK   = SMASK << UTC_SHIFT;
571	>	private static final int  UAC_UNIT   = 1 << UAC_SHIFT;
572	>	private static final int  UTC_UNIT   = 1 << UTC_SHIFT;
573	>
574	>	// masks and units for dealing with e = (int)ctl
575	>	private static final int E_MASK      = 0x7fffffff; // no STOP_BIT
576	>	private static final int E_SEQ       = 1 << EC_SHIFT;
577	>
578	>	// runState bits
579	>	private static final int SHUTDOWN    = 1 << 31;
580	>	private static final int RS_SEQ      = 1 << 16;
581	>	private static final int RS_SEQ_MASK = 0x7fff0000;
582	>
583	>	// access mode for WorkQueue
584	>	static final int LIFO_QUEUE          =  0;
585	>	static final int FIFO_QUEUE          =  1;
586	>	static final int SHARED_QUEUE        = -1;
587	>
588	>	/**
589	>	* The wakeup interval (in nanoseconds) for a worker waiting for a
590	>	* task when the pool is quiescent to instead try to shrink the
591	>	* number of workers.  The exact value does not matter too
592	>	* much. It must be short enough to release resources during
593	>	* sustained periods of idleness, but not so short that threads
594	>	* are continually re-created.
595	>	*/
596	>	private static final long SHRINK_RATE =
597	>	4L * 1000L * 1000L * 1000L; // 4 seconds
598	>
599	>	/**
600	>	* The timeout value for attempted shrinkage, includes
601	>	* some slop to cope with system timer imprecision.
602	>	*/
603	>	private static final long SHRINK_TIMEOUT = SHRINK_RATE - (SHRINK_RATE / 10);
604	>
605	>	/**
606	>	* The maximum stolen->joining link depth allowed in tryHelpStealer.
607	>	* Depths for legitimate chains are unbounded, but we use a fixed
608	>	* constant to avoid (otherwise unchecked) cycles and to bound
609	>	* staleness of traversal parameters at the expense of sometimes
610	>	* blocking when we could be helping.
611		*/
612	<	volatile ForkJoinWorkerThread[] workers;
612	>	private static final int MAX_HELP_DEPTH = 16;
613
614	<	/**
615	<	* Queue for external submissions.
614	>	/*
615	>	* Field layout order in this class tends to matter more than one
616	>	* would like. Runtime layout order is only loosely related to
617	>	* declaration order and may differ across JVMs, but the following
618	>	* empirically works OK on current JVMs.
619	>	*/
620	>
621	>	volatile long ctl;                       // main pool control
622	>	final int parallelism;                   // parallelism level
623	>	final int localMode;                     // per-worker scheduling mode
624	>	int nextPoolIndex;                       // hint used in registerWorker
625	>	volatile int runState;                   // shutdown status, seq, and mask
626	>	WorkQueue[] workQueues;                  // main registry
627	>	final ReentrantLock lock;                // for registration
628	>	final Condition termination;             // for awaitTermination
629	>	final ForkJoinWorkerThreadFactory factory; // factory for new workers
630	>	final Thread.UncaughtExceptionHandler ueh; // per-worker UEH
631	>	final AtomicLong stealCount;             // collect counts when terminated
632	>	final AtomicInteger nextWorkerNumber;    // to create worker name string
633	>	final String workerNamePrefix;           // Prefix for assigning worker names
634	>
635	>	/**
636	>	* Queues supporting work-stealing as well as external task
637	>	* submission. See above for main rationale and algorithms.
638	>	* Implementation relies heavily on "Unsafe" intrinsics
639	>	* and selective use of "volatile":
640	>	*
641	>	* Field "base" is the index (mod array.length) of the least valid
642	>	* queue slot, which is always the next position to steal (poll)
643	>	* from if nonempty. Reads and writes require volatile orderings
644	>	* but not CAS, because updates are only performed after slot
645	>	* CASes.
646	>	*
647	>	* Field "top" is the index (mod array.length) of the next queue
648	>	* slot to push to or pop from. It is written only by owner thread
649	>	* for push, or under lock for trySharedPush, and accessed by
650	>	* other threads only after reading (volatile) base.  Both top and
651	>	* base are allowed to wrap around on overflow, but (top - base)
652	>	* (or more commonly -(base - top) to force volatile read of base
653	>	* before top) still estimates size.
654	>	*
655	>	* The array slots are read and written using the emulation of
656	>	* volatiles/atomics provided by Unsafe. Insertions must in
657	>	* general use putOrderedObject as a form of releasing store to
658	>	* ensure that all writes to the task object are ordered before
659	>	* its publication in the queue. (Although we can avoid one case
660	>	* of this when locked in trySharedPush.) All removals entail a
661	>	* CAS to null.  The array is always a power of two. To ensure
662	>	* safety of Unsafe array operations, all accesses perform
663	>	* explicit null checks and implicit bounds checks via
664	>	* power-of-two masking.
665	>	*
666	>	* In addition to basic queuing support, this class contains
667	>	* fields described elsewhere to control execution. It turns out
668	>	* to work better memory-layout-wise to include them in this
669	>	* class rather than a separate class.
670	>	*
671	>	* Performance on most platforms is very sensitive to placement of
672	>	* instances of both WorkQueues and their arrays -- we absolutely
673	>	* do not want multiple WorkQueue instances or multiple queue
674	>	* arrays sharing cache lines. (It would be best for queue objects
675	>	* and their arrays to share, but there is nothing available to
676	>	* help arrange that).  Unfortunately, because they are recorded
677	>	* in a common array, WorkQueue instances are often moved to be
678	>	* adjacent by garbage collectors. To reduce impact, we use field
679	>	* padding that works OK on common platforms; this effectively
680	>	* trades off slightly slower average field access for the sake of
681	>	* avoiding really bad worst-case access. (Until better JVM
682	>	* support is in place, this padding is dependent on transient
683	>	* properties of JVM field layout rules.)  We also take care in
684	>	* allocating and sizing and resizing the array. Non-shared queue
685	>	* arrays are initialized (via method growArray) by workers before
686	>	* use. Others are allocated on first use.
687		*/
688	<	private final LinkedTransferQueue<ForkJoinTask<?>> submissionQueue;
688	>	static final class WorkQueue {
689	>	/**
690	>	* Capacity of work-stealing queue array upon initialization.
691	>	* Must be a power of two; at least 4, but set larger to
692	>	* reduce cacheline sharing among queues.
693	>	*/
694	>	static final int INITIAL_QUEUE_CAPACITY = 1 << 8;
695
696	<	/**
697	<	* Lock protecting updates to workers array.
698	<	*/
699	<	private final ReentrantLock workerLock;
696	>	/**
697	>	* Maximum size for queue arrays. Must be a power of two less
698	>	* than or equal to 1 << (31 - width of array entry) to ensure
699	>	* lack of wraparound of index calculations, but defined to a
700	>	* value a bit less than this to help users trap runaway
701	>	* programs before saturating systems.
702	>	*/
703	>	static final int MAXIMUM_QUEUE_CAPACITY = 1 << 26; // 64M
704
705	<	/**
706	<	* Latch released upon termination.
707	<	*/
708	<	private final CountDownLatch terminationLatch;
705	>	volatile long totalSteals; // cumulative number of steals
706	>	int seed;                  // for random scanning; initialize nonzero
707	>	volatile int eventCount;   // encoded inactivation count; < 0 if inactive
708	>	int nextWait;              // encoded record of next event waiter
709	>	int rescans;               // remaining scans until block
710	>	int nsteals;               // top-level task executions since last idle
711	>	final int mode;            // lifo, fifo, or shared
712	>	int poolIndex;             // index of this queue in pool (or 0)
713	>	int stealHint;             // index of most recent known stealer
714	>	volatile int runState;     // 1: locked, -1: terminate; else 0
715	>	volatile int base;         // index of next slot for poll
716	>	int top;                   // index of next slot for push
717	>	ForkJoinTask<?>[] array;   // the elements (initially unallocated)
718	>	final ForkJoinWorkerThread owner; // owning thread or null if shared
719	>	volatile Thread parker;    // == owner during call to park; else null
720	>	ForkJoinTask<?> currentJoin;  // task being joined in awaitJoin
721	>	ForkJoinTask<?> currentSteal; // current non-local task being executed
722	>	// Heuristic padding to ameliorate unfortunate memory placements
723	>	Object p00, p01, p02, p03, p04, p05, p06, p07, p08, p09, p0a;
724	>
725	>	WorkQueue(ForkJoinWorkerThread owner, int mode) {
726	>	this.owner = owner;
727	>	this.mode = mode;
728	>	// Place indices in the center of array (that is not yet allocated)
729	>	base = top = INITIAL_QUEUE_CAPACITY >>> 1;
730	>	}
731
732	<	/**
733	<	* Creation factory for worker threads.
734	<	*/
735	<	private final ForkJoinWorkerThreadFactory factory;
732	>	/**
733	>	* Returns number of tasks in the queue
734	>	*/
735	>	final int queueSize() {
736	>	int n = base - top; // non-owner callers must read base first
737	>	return (n >= 0) ? 0 : -n;
738	>	}
739
740	<	/**
741	<	* Sum of per-thread steal counts, updated only when threads are
742	<	* idle or terminating.
743	<	*/
744	<	private volatile long stealCount;
740	>	/**
741	>	* Pushes a task. Call only by owner in unshared queues.
742	>	*
743	>	* @param task the task. Caller must ensure non-null.
744	>	* @param p, if non-null, pool to signal if necessary
745	>	* @throw RejectedExecutionException if array cannot
746	>	* be resized
747	>	*/
748	>	final void push(ForkJoinTask<?> task, ForkJoinPool p) {
749	>	ForkJoinTask<?>[] a;
750	>	int s = top, m, n;
751	>	if ((a = array) != null) {    // ignore if queue removed
752	>	U.putOrderedObject
753	>	(a, (((m = a.length - 1) & s) << ASHIFT) + ABASE, task);
754	>	if ((n = (top = s + 1) - base) <= 2) {
755	>	if (p != null)
756	>	p.signalWork();
757	>	}
758	>	else if (n >= m)
759	>	growArray(true);
760	>	}
761	>	}
762
763	<	/**
764	<	* Encoded record of top of treiber stack of threads waiting for
765	<	* events. The top 32 bits contain the count being waited for. The
766	<	* bottom word contains one plus the pool index of waiting worker
767	<	* thread.
768	<	*/
769	<	private volatile long eventWaiters;
763	>	/**
764	>	* Pushes a task if lock is free and array is either big
765	>	* enough or can be resized to be big enough.
766	>	*
767	>	* @param task the task. Caller must ensure non-null.
768	>	* @return true if submitted
769	>	*/
770	>	final boolean trySharedPush(ForkJoinTask<?> task) {
771	>	boolean submitted = false;
772	>	if (runState == 0 && U.compareAndSwapInt(this, RUNSTATE, 0, 1)) {
773	>	ForkJoinTask<?>[] a = array;
774	>	int s = top, n = s - base;
775	>	try {
776	>	if ((a != null && n < a.length - 1) ||
777	>	(a = growArray(false)) != null) { // must presize
778	>	int j = (((a.length - 1) & s) << ASHIFT) + ABASE;
779	>	U.putObject(a, (long)j, task);    // don't need "ordered"
780	>	top = s + 1;
781	>	submitted = true;
782	>	}
783	>	} finally {
784	>	runState = 0;                         // unlock
785	>	}
786	>	}
787	>	return submitted;
788	>	}
789
790	<	private static final int  EVENT_COUNT_SHIFT = 32;
791	<	private static final long WAITER_INDEX_MASK = (1L << EVENT_COUNT_SHIFT)-1L;
790	>	/**
791	>	* Takes next task, if one exists, in FIFO order.
792	>	*/
793	>	final ForkJoinTask<?> poll() {
794	>	ForkJoinTask<?>[] a; int b, i;
795	>	while ((b = base) - top < 0 && (a = array) != null &&
796	>	(i = (a.length - 1) & b) >= 0) {
797	>	int j = (i << ASHIFT) + ABASE;
798	>	ForkJoinTask<?> t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
799	>	if (t != null && base == b &&
800	>	U.compareAndSwapObject(a, j, t, null)) {
801	>	base = b + 1;
802	>	return t;
803	>	}
804	>	}
805	>	return null;
806	>	}
807
808	<	/**
809	<	* A counter for events that may wake up worker threads:
810	<	*   - Submission of a new task to the pool
811	<	*   - A worker pushing a task on an empty queue
812	<	*   - termination and reconfiguration
813	<	*/
814	<	private volatile int eventCount;
808	>	/**
809	>	* Takes next task, if one exists, in LIFO order.
810	>	* Call only by owner in unshared queues.
811	>	*/
812	>	final ForkJoinTask<?> pop() {
813	>	ForkJoinTask<?> t; int m;
814	>	ForkJoinTask<?>[] a = array;
815	>	if (a != null && (m = a.length - 1) >= 0) {
816	>	for (int s; (s = top - 1) - base >= 0;) {
817	>	int j = ((m & s) << ASHIFT) + ABASE;
818	>	if ((t = (ForkJoinTask<?>)U.getObjectVolatile(a, j)) == null)
819	>	break;
820	>	if (U.compareAndSwapObject(a, j, t, null)) {
821	>	top = s;
822	>	return t;
823	>	}
824	>	}
825	>	}
826	>	return null;
827	>	}
828
829	<	/**
830	<	* Lifecycle control. The low word contains the number of workers
831	<	* that are (probably) executing tasks. This value is atomically
832	<	* incremented before a worker gets a task to run, and decremented
833	<	* when worker has no tasks and cannot find any.  Bits 16-18
834	<	* contain runLevel value. When all are zero, the pool is
454	<	* running. Level transitions are monotonic (running -> shutdown
455	<	* -> terminating -> terminated) so each transition adds a bit.
456	<	* These are bundled together to ensure consistent read for
457	<	* termination checks (i.e., that runLevel is at least SHUTDOWN
458	<	* and active threads is zero).
459	<	*/
460	<	private volatile int runState;
829	>	/**
830	>	* Takes next task, if one exists, in order specified by mode.
831	>	*/
832	>	final ForkJoinTask<?> nextLocalTask() {
833	>	return mode == 0 ? pop() : poll();
834	>	}
835
836	<	// Note: The order among run level values matters.
837	<	private static final int RUNLEVEL_SHIFT     = 16;
838	<	private static final int SHUTDOWN           = 1 << RUNLEVEL_SHIFT;
839	<	private static final int TERMINATING        = 1 << (RUNLEVEL_SHIFT + 1);
840	<	private static final int TERMINATED         = 1 << (RUNLEVEL_SHIFT + 2);
841	<	private static final int ACTIVE_COUNT_MASK  = (1 << RUNLEVEL_SHIFT) - 1;
842	<	private static final int ONE_ACTIVE         = 1; // active update delta
836	>	/**
837	>	* Returns next task, if one exists, in order specified by mode.
838	>	*/
839	>	final ForkJoinTask<?> peek() {
840	>	ForkJoinTask<?>[] a = array; int m;
841	>	if (a == null || (m = a.length - 1) < 0)
842	>	return null;
843	>	int i = mode == 0 ? top - 1 : base;
844	>	int j = ((i & m) << ASHIFT) + ABASE;
845	>	return (ForkJoinTask<?>)U.getObjectVolatile(a, j);
846	>	}
847
848	<	/**
849	<	* Holds number of total (i.e., created and not yet terminated)
850	<	* and running (i.e., not blocked on joins or other managed sync)
851	<	* threads, packed together to ensure consistent snapshot when
852	<	* making decisions about creating and suspending spare
853	<	* threads. Updated only by CAS. Note that adding a new worker
854	<	* requires incrementing both counts, since workers start off in
855	<	* running state.  This field is also used for memory-fencing
856	<	* configuration parameters.
857	<	*/
858	<	private volatile int workerCounts;
848	>	/**
849	>	* Returns task at index b if b is current base of queue.
850	>	*/
851	>	final ForkJoinTask<?> pollAt(int b) {
852	>	ForkJoinTask<?>[] a; int i;
853	>	ForkJoinTask<?> task = null;
854	>	if ((a = array) != null && (i = ((a.length - 1) & b)) >= 0) {
855	>	int j = (i << ASHIFT) + ABASE;
856	>	ForkJoinTask<?> t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
857	>	if (t != null && base == b &&
858	>	U.compareAndSwapObject(a, j, t, null)) {
859	>	base = b + 1;
860	>	task = t;
861	>	}
862	>	}
863	>	return task;
864	>	}
865
866	<	private static final int TOTAL_COUNT_SHIFT  = 16;
867	<	private static final int RUNNING_COUNT_MASK = (1 << TOTAL_COUNT_SHIFT) - 1;
868	<	private static final int ONE_RUNNING        = 1;
869	<	private static final int ONE_TOTAL          = 1 << TOTAL_COUNT_SHIFT;
866	>	/**
867	>	* Pops the given task only if it is at the current top.
868	>	*/
869	>	final boolean tryUnpush(ForkJoinTask<?> t) {
870	>	ForkJoinTask<?>[] a; int s;
871	>	if ((a = array) != null && (s = top) != base &&
872	>	U.compareAndSwapObject
873	>	(a, (((a.length - 1) & --s) << ASHIFT) + ABASE, t, null)) {
874	>	top = s;
875	>	return true;
876	>	}
877	>	return false;
878	>	}
879
880	<	/*
881	<	* Fields parallelism. maxPoolSize, locallyFifo,
882	<	* maintainsParallelism, and ueh are non-volatile, but external
883	<	* reads/writes use workerCount fences to ensure visability.
884	<	*/
880	>	/**
881	>	* Polls the given task only if it is at the current base.
882	>	*/
883	>	final boolean pollFor(ForkJoinTask<?> task) {
884	>	ForkJoinTask<?>[] a; int b, i;
885	>	if ((b = base) - top < 0 && (a = array) != null &&
886	>	(i = (a.length - 1) & b) >= 0) {
887	>	int j = (i << ASHIFT) + ABASE;
888	>	if (U.getObjectVolatile(a, j) == task && base == b &&
889	>	U.compareAndSwapObject(a, j, task, null)) {
890	>	base = b + 1;
891	>	return true;
892	>	}
893	>	}
894	>	return false;
895	>	}
896
897	<	/**
898	<	* The target parallelism level.
899	<	*/
900	<	private int parallelism;
897	>	/**
898	>	* If present, removes from queue and executes the given task, or
899	>	* any other cancelled task. Returns (true) immediately on any CAS
900	>	* or consistency check failure so caller can retry.
901	>	*
902	>	* @return false if no progress can be made
903	>	*/
904	>	final boolean tryRemoveAndExec(ForkJoinTask<?> task) {
905	>	boolean removed = false, empty = true, progress = true;
906	>	ForkJoinTask<?>[] a; int m, s, b, n;
907	>	if ((a = array) != null && (m = a.length - 1) >= 0 &&
908	>	(n = (s = top) - (b = base)) > 0) {
909	>	for (ForkJoinTask<?> t;;) {           // traverse from s to b
910	>	int j = ((--s & m) << ASHIFT) + ABASE;
911	>	t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
912	>	if (t == null)                    // inconsistent length
913	>	break;
914	>	else if (t == task) {
915	>	if (s + 1 == top) {           // pop
916	>	if (!U.compareAndSwapObject(a, j, task, null))
917	>	break;
918	>	top = s;
919	>	removed = true;
920	>	}
921	>	else if (base == b)           // replace with proxy
922	>	removed = U.compareAndSwapObject(a, j, task,
923	>	new EmptyTask());
924	>	break;
925	>	}
926	>	else if (t.status >= 0)
927	>	empty = false;
928	>	else if (s + 1 == top) {          // pop and throw away
929	>	if (U.compareAndSwapObject(a, j, t, null))
930	>	top = s;
931	>	break;
932	>	}
933	>	if (--n == 0) {
934	>	if (!empty && base == b)
935	>	progress = false;
936	>	break;
937	>	}
938	>	}
939	>	}
940	>	if (removed)
941	>	task.doExec();
942	>	return progress;
943	>	}
944
945	<	/**
946	<	* The maximum allowed pool size.
947	<	*/
948	<	private int maxPoolSize;
945	>	/**
946	>	* Initializes or doubles the capacity of array. Call either
947	>	* by owner or with lock held -- it is OK for base, but not
948	>	* top, to move while resizings are in progress.
949	>	*
950	>	* @param rejectOnFailure if true, throw exception if capacity
951	>	* exceeded (relayed ultimately to user); else return null.
952	>	*/
953	>	final ForkJoinTask<?>[] growArray(boolean rejectOnFailure) {
954	>	ForkJoinTask<?>[] oldA = array;
955	>	int size = oldA != null ? oldA.length << 1 : INITIAL_QUEUE_CAPACITY;
956	>	if (size <= MAXIMUM_QUEUE_CAPACITY) {
957	>	int oldMask, t, b;
958	>	ForkJoinTask<?>[] a = array = new ForkJoinTask<?>[size];
959	>	if (oldA != null && (oldMask = oldA.length - 1) >= 0 &&
960	>	(t = top) - (b = base) > 0) {
961	>	int mask = size - 1;
962	>	do {
963	>	ForkJoinTask<?> x;
964	>	int oldj = ((b & oldMask) << ASHIFT) + ABASE;
965	>	int j    = ((b &    mask) << ASHIFT) + ABASE;
966	>	x = (ForkJoinTask<?>)U.getObjectVolatile(oldA, oldj);
967	>	if (x != null &&
968	>	U.compareAndSwapObject(oldA, oldj, x, null))
969	>	U.putObjectVolatile(a, j, x);
970	>	} while (++b != t);
971	>	}
972	>	return a;
973	>	}
974	>	else if (!rejectOnFailure)
975	>	return null;
976	>	else
977	>	throw new RejectedExecutionException("Queue capacity exceeded");
978	>	}
979
980	<	/**
981	<	* True if use local fifo, not default lifo, for local polling
982	<	* Replicated by ForkJoinWorkerThreads
983	<	*/
984	<	private boolean locallyFifo;
980	>	/**
981	>	* Removes and cancels all known tasks, ignoring any exceptions
982	>	*/
983	>	final void cancelAll() {
984	>	ForkJoinTask.cancelIgnoringExceptions(currentJoin);
985	>	ForkJoinTask.cancelIgnoringExceptions(currentSteal);
986	>	for (ForkJoinTask<?> t; (t = poll()) != null; )
987	>	ForkJoinTask.cancelIgnoringExceptions(t);
988	>	}
989
990	<	/**
510	<	* Controls whether to add spares to maintain parallelism
511	<	*/
512	<	private boolean maintainsParallelism;
990	>	// Execution methods
991
992	<	/**
993	<	* The uncaught exception handler used when any worker
994	<	* abruptly terminates
995	<	*/
996	<	private Thread.UncaughtExceptionHandler ueh;
992	>	/**
993	>	* Removes and runs tasks until empty, using local mode
994	>	* ordering.
995	>	*/
996	>	final void runLocalTasks() {
997	>	if (base - top < 0) {
998	>	for (ForkJoinTask<?> t; (t = nextLocalTask()) != null; )
999	>	t.doExec();
1000	>	}
1001	>	}
1002
1003	<	/**
1004	<	* Pool number, just for assigning useful names to worker threads
1005	<	*/
1006	<	private final int poolNumber;
1003	>	/**
1004	>	* Executes a top-level task and any local tasks remaining
1005	>	* after execution.
1006	>	*
1007	>	* @return true unless terminating
1008	>	*/
1009	>	final boolean runTask(ForkJoinTask<?> t) {
1010	>	boolean alive = true;
1011	>	if (t != null) {
1012	>	currentSteal = t;
1013	>	t.doExec();
1014	>	runLocalTasks();
1015	>	++nsteals;
1016	>	currentSteal = null;
1017	>	}
1018	>	else if (runState < 0)            // terminating
1019	>	alive = false;
1020	>	return alive;
1021	>	}
1022
1023	<	// utilities for updating fields
1023	>	/**
1024	>	* Executes a non-top-level (stolen) task
1025	>	*/
1026	>	final void runSubtask(ForkJoinTask<?> t) {
1027	>	if (t != null) {
1028	>	ForkJoinTask<?> ps = currentSteal;
1029	>	currentSteal = t;
1030	>	t.doExec();
1031	>	currentSteal = ps;
1032	>	}
1033	>	}
1034
1035	<	/**
1036	<	* Adds delta to running count.  Used mainly by ForkJoinTask.
1037	<	*
1038	<	* @param delta the number to add
1039	<	*/
1040	<	final void updateRunningCount(int delta) {
1041	<	int wc;
1042	<	do {} while (!UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1043	<	wc = workerCounts,
1044	<	wc + delta));
1045	<	}
1035	>	/**
1036	>	* Computes next value for random probes.  Scans don't require
1037	>	* a very high quality generator, but also not a crummy one.
1038	>	* Marsaglia xor-shift is cheap and works well enough.  Note:
1039	>	* This is manually inlined in several usages in ForkJoinPool
1040	>	* to avoid writes inside busy scan loops.
1041	>	*/
1042	>	final int nextSeed() {
1043	>	int r = seed;
1044	>	r ^= r << 13;
1045	>	r ^= r >>> 17;
1046	>	r ^= r << 5;
1047	>	return seed = r;
1048	>	}
1049
1050	<	/**
1051	<	* Write fence for user modifications of pool parameters
1052	<	* (parallelism. etc).  Note that it doesn't matter if CAS fails.
1053	<	*/
1054	<	private void workerCountWriteFence() {
1055	<	int wc;
1056	<	UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1057	<	wc = workerCounts, wc);
1050	>	// Unsafe mechanics
1051	>	private static final sun.misc.Unsafe U;
1052	>	private static final long RUNSTATE;
1053	>	private static final int ABASE;
1054	>	private static final int ASHIFT;
1055	>	static {
1056	>	int s;
1057	>	try {
1058	>	U = getUnsafe();
1059	>	Class<?> k = WorkQueue.class;
1060	>	Class<?> ak = ForkJoinTask[].class;
1061	>	RUNSTATE = U.objectFieldOffset
1062	>	(k.getDeclaredField("runState"));
1063	>	ABASE = U.arrayBaseOffset(ak);
1064	>	s = U.arrayIndexScale(ak);
1065	>	} catch (Exception e) {
1066	>	throw new Error(e);
1067	>	}
1068	>	if ((s & (s-1)) != 0)
1069	>	throw new Error("data type scale not a power of two");
1070	>	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
1071	>	}
1072		}
1073
1074		/**
1075	<	* Read fence for external reads of pool parameters
1076	<	* (parallelism. maxPoolSize, etc).
1075	>	* Class for artificial tasks that are used to replace the target
1076	>	* of local joins if they are removed from an interior queue slot
1077	>	* in WorkQueue.tryRemoveAndExec. We don't need the proxy to
1078	>	* actually do anything beyond having a unique identity.
1079		*/
1080	<	private void workerCountReadFence() {
1081	<	int ignore = workerCounts;
1080	>	static final class EmptyTask extends ForkJoinTask<Void> {
1081	>	EmptyTask() { status = ForkJoinTask.NORMAL; } // force done
1082	>	public Void getRawResult() { return null; }
1083	>	public void setRawResult(Void x) {}
1084	>	public boolean exec() { return true; }
1085		}
1086
1087		/**
1088	<	* Tries incrementing active count; fails on contention.
1089	<	* Called by workers before executing tasks.
1090	<	*
561	<	* @return true on success
1088	>	* Computes a hash code for the given thread. This method is
1089	>	* expected to provide higher-quality hash codes than those using
1090	>	* method hashCode().
1091		*/
1092	<	final boolean tryIncrementActiveCount() {
1093	<	int c;
1094	<	return UNSAFE.compareAndSwapInt(this, runStateOffset,
1095	<	c = runState, c + ONE_ACTIVE);
1092	>	static final int hashThread(Thread t) {
1093	>	long id = (t == null) ? 0L : t.getId(); // Use MurmurHash of thread id
1094	>	int h = (int)id ^ (int)(id >>> 32);
1095	>	h ^= h >>> 16;
1096	>	h *= 0x85ebca6b;
1097	>	h ^= h >>> 13;
1098	>	h *= 0xc2b2ae35;
1099	>	return h ^ (h >>> 16);
1100		}
1101
1102		/**
1103	<	* Tries decrementing active count; fails on contention.
571	<	* Called when workers cannot find tasks to run.
1103	>	* Top-level runloop for workers
1104		*/
1105	<	final boolean tryDecrementActiveCount() {
1106	<	int c;
1107	<	return UNSAFE.compareAndSwapInt(this, runStateOffset,
1108	<	c = runState, c - ONE_ACTIVE);
577	<	}
1105	>	final void runWorker(ForkJoinWorkerThread wt) {
1106	>	WorkQueue w = wt.workQueue;
1107	>	w.growArray(false);     // Initialize queue array and seed in this thread
1108	>	w.seed = hashThread(Thread.currentThread()) | (1 << 31); // force < 0
1109
1110	<	/**
580	<	* Advances to at least the given level. Returns true if not
581	<	* already in at least the given level.
582	<	*/
583	<	private boolean advanceRunLevel(int level) {
584	<	for (;;) {
585	<	int s = runState;
586	<	if ((s & level) != 0)
587	<	return false;
588	<	if (UNSAFE.compareAndSwapInt(this, runStateOffset, s, s | level))
589	<	return true;
590	<	}
1110	>	do {} while (w.runTask(scan(w)));
1111		}
1112
1113	<	// workers array maintenance
1113	>	// Creating, registering and deregistering workers
1114
1115		/**
1116	<	* Records and returns a workers array index for new worker.
1116	>	* Tries to create and start a worker
1117		*/
1118	<	private int recordWorker(ForkJoinWorkerThread w) {
1119	<	// Try using slot totalCount-1. If not available, scan and/or resize
1120	<	int k = (workerCounts >>> TOTAL_COUNT_SHIFT) - 1;
601	<	final ReentrantLock lock = this.workerLock;
602	<	lock.lock();
1118	>	private void addWorker() {
1119	>	Throwable ex = null;
1120	>	ForkJoinWorkerThread w = null;
1121		try {
1122	<	ForkJoinWorkerThread[] ws = workers;
1123	<	int len = ws.length;
1124	<	if (k < 0 || k >= len || ws[k] != null) {
607	<	for (k = 0; k < len && ws[k] != null; ++k)
608	<	;
609	<	if (k == len)
610	<	ws = Arrays.copyOf(ws, len << 1);
1122	>	if ((w = factory.newThread(this)) != null) {
1123	>	w.start();
1124	>	return;
1125		}
1126	<	ws[k] = w;
1127	<	workers = ws; // volatile array write ensures slot visibility
614	<	} finally {
615	<	lock.unlock();
1126	>	} catch (Throwable e) {
1127	>	ex = e;
1128		}
1129	<	return k;
1129	>	deregisterWorker(w, ex);
1130		}
1131
1132		/**
1133	<	* Nulls out record of worker in workers array
1133	>	* Callback from ForkJoinWorkerThread constructor to assign a
1134	>	* public name. This must be separate from registerWorker because
1135	>	* it is called during the "super" constructor call in
1136	>	* ForkJoinWorkerThread.
1137		*/
1138	<	private void forgetWorker(ForkJoinWorkerThread w) {
1139	<	int idx = w.poolIndex;
1140	<	// Locking helps method recordWorker avoid unecessary expansion
626	<	final ReentrantLock lock = this.workerLock;
627	<	lock.lock();
628	<	try {
629	<	ForkJoinWorkerThread[] ws = workers;
630	<	if (idx >= 0 && idx < ws.length && ws[idx] == w) // verify
631	<	ws[idx] = null;
632	<	} finally {
633	<	lock.unlock();
634	<	}
1138	>	final String nextWorkerName() {
1139	>	return workerNamePrefix.concat
1140	>	(Integer.toString(nextWorkerNumber.addAndGet(1)));
1141		}
1142
637	–	// adding and removing workers
638	–
1143		/**
1144	<	* Tries to create and add new worker. Assumes that worker counts
1145	<	* are already updated to accommodate the worker, so adjusts on
642	<	* failure.
1144	>	* Callback from ForkJoinWorkerThread constructor to establish and
1145	>	* record its WorkQueue
1146		*
1147	<	* @return new worker or null if creation failed
1147	>	* @param wt the worker thread
1148		*/
1149	<	private ForkJoinWorkerThread addWorker() {
1150	<	ForkJoinWorkerThread w = null;
1149	>	final void registerWorker(ForkJoinWorkerThread wt) {
1150	>	WorkQueue w = wt.workQueue;
1151	>	ReentrantLock lock = this.lock;
1152	>	lock.lock();
1153		try {
1154	<	w = factory.newThread(this);
1155	<	} finally { // Adjust on either null or exceptional factory return
1156	<	if (w == null) {
1157	<	onWorkerCreationFailure();
1158	<	return null;
1154	>	int k = nextPoolIndex;
1155	>	WorkQueue[] ws = workQueues;
1156	>	if (ws != null) {                       // ignore on shutdown
1157	>	int n = ws.length;
1158	>	if (k < 0 || (k & 1) == 0 || k >= n || ws[k] != null) {
1159	>	for (k = 1; k < n && ws[k] != null; k += 2)
1160	>	;                           // workers are at odd indices
1161	>	if (k >= n)                     // resize
1162	>	workQueues = ws = Arrays.copyOf(ws, n << 1);
1163	>	}
1164	>	w.poolIndex = k;
1165	>	w.eventCount = ~(k >>> 1) & SMASK;  // Set up wait count
1166	>	ws[k] = w;                          // record worker
1167	>	nextPoolIndex = k + 2;
1168	>	int rs = runState;
1169	>	int m = rs & SMASK;                 // recalculate runState mask
1170	>	if (k > m)
1171	>	m = (m << 1) + 1;
1172	>	runState = (rs & SHUTDOWN) | ((rs + RS_SEQ) & RS_SEQ_MASK) | m;
1173		}
1174	<	}
1175	<	w.start(recordWorker(w), locallyFifo, ueh);
657	<	return w;
658	<	}
659	<
660	<	/**
661	<	* Adjusts counts upon failure to create worker
662	<	*/
663	<	private void onWorkerCreationFailure() {
664	<	int c;
665	<	do {} while (!UNSAFE.compareAndSwapInt(this, workerCountsOffset,
666	<	c = workerCounts,
667	<	c - (ONE_RUNNING|ONE_TOTAL)));
668	<	tryTerminate(false); // in case of failure during shutdown
669	<	}
670	<
671	<	/**
672	<	* Create enough total workers to establish target parallelism,
673	<	* giving up if terminating or addWorker fails
674	<	*/
675	<	private void ensureEnoughTotalWorkers() {
676	<	int wc;
677	<	while (runState < TERMINATING &&
678	<	((wc = workerCounts) >>> TOTAL_COUNT_SHIFT) < parallelism) {
679	<	if ((UNSAFE.compareAndSwapInt(this, workerCountsOffset,
680	<	wc, wc + (ONE_RUNNING|ONE_TOTAL)) &&
681	<	addWorker() == null))
682	<	break;
1174	>	} finally {
1175	>	lock.unlock();
1176		}
1177		}
1178
1179		/**
1180	<	* Final callback from terminating worker.  Removes record of
1180	>	* Final callback from terminating worker, as well as failure to
1181	>	* construct or start a worker in addWorker.  Removes record of
1182		* worker from array, and adjusts counts. If pool is shutting
1183	<	* down, tries to complete terminatation, else possibly replaces
690	<	* the worker.
1183	>	* down, tries to complete termination.
1184		*
1185	<	* @param w the worker
1185	>	* @param wt the worker thread or null if addWorker failed
1186	>	* @param ex the exception causing failure, or null if none
1187		*/
1188	<	final void workerTerminated(ForkJoinWorkerThread w) {
1189	<	if (w.active) { // force inactive
1190	<	w.active = false;
1191	<	do {} while (!tryDecrementActiveCount());
1188	>	final void deregisterWorker(ForkJoinWorkerThread wt, Throwable ex) {
1189	>	WorkQueue w = null;
1190	>	if (wt != null && (w = wt.workQueue) != null) {
1191	>	w.runState = -1;                // ensure runState is set
1192	>	stealCount.getAndAdd(w.totalSteals + w.nsteals);
1193	>	int idx = w.poolIndex;
1194	>	ReentrantLock lock = this.lock;
1195	>	lock.lock();
1196	>	try {                           // remove record from array
1197	>	WorkQueue[] ws = workQueues;
1198	>	if (ws != null && idx >= 0 && idx < ws.length && ws[idx] == w)
1199	>	ws[nextPoolIndex = idx] = null;
1200	>	} finally {
1201	>	lock.unlock();
1202	>	}
1203		}
699	–	forgetWorker(w);
1204
1205	<	// decrement total count, and if was running, running count
1206	<	int unit = w.isTrimmed()? ONE_TOTAL : (ONE_RUNNING|ONE_TOTAL);
1207	<	int wc;
1208	<	do {} while (!UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1209	<	wc = workerCounts, wc - unit));
1205	>	long c;                             // adjust ctl counts
1206	>	do {} while (!U.compareAndSwapLong
1207	>	(this, CTL, c = ctl, (((c - AC_UNIT) & AC_MASK) |
1208	>	((c - TC_UNIT) & TC_MASK) |
1209	>	(c & ~(AC_MASK|TC_MASK)))));
1210
1211	<	accumulateStealCount(w); // collect final count
1212	<	if (!tryTerminate(false))
1213	<	ensureEnoughTotalWorkers();
1211	>	if (!tryTerminate(false) && w != null) {
1212	>	w.cancelAll();                  // cancel remaining tasks
1213	>	if (w.array != null)            // suppress signal if never ran
1214	>	signalWork();               // wake up or create replacement
1215	>	}
1216	>
1217	>	if (ex != null)                     // rethrow
1218	>	U.throwException(ex);
1219		}
1220
712	–	// Waiting for and signalling events
1221
1222	<	/**
715	<	* Ensures eventCount on exit is different (mod 2^32) than on
716	<	* entry.  CAS failures are OK -- any change in count suffices.
717	<	*/
718	<	private void advanceEventCount() {
719	<	int c;
720	<	UNSAFE.compareAndSwapInt(this, eventCountOffset, c = eventCount, c+1);
721	<	}
1222	>	// Maintaining ctl counts
1223
1224		/**
1225	<	* Releases workers blocked on a count not equal to current count.
1225	>	* Increments active count; mainly called upon return from blocking
1226		*/
1227	<	final void releaseWaiters() {
1228	<	long top;
1229	<	int id;
729	<	while ((id = (int)((top = eventWaiters) & WAITER_INDEX_MASK)) > 0 &&
730	<	(int)(top >>> EVENT_COUNT_SHIFT) != eventCount) {
731	<	ForkJoinWorkerThread[] ws = workers;
732	<	ForkJoinWorkerThread w;
733	<	if (ws.length >= id && (w = ws[id - 1]) != null &&
734	<	UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
735	<	top, w.nextWaiter))
736	<	LockSupport.unpark(w);
737	<	}
1227	>	final void incrementActiveCount() {
1228	>	long c;
1229	>	do {} while (!U.compareAndSwapLong(this, CTL, c = ctl, c + AC_UNIT));
1230		}
1231
1232		/**
1233	<	* Advances eventCount and releases waiters until interference by
742	<	* other releasing threads is detected.
1233	>	* Activates or creates a worker
1234		*/
1235		final void signalWork() {
1236	<	int ec;
1237	<	UNSAFE.compareAndSwapInt(this, eventCountOffset, ec=eventCount, ec+1);
1238	<	outer:for (;;) {
1239	<	long top = eventWaiters;
1240	<	ec = eventCount;
1241	<	for (;;) {
1242	<	ForkJoinWorkerThread[] ws; ForkJoinWorkerThread w;
1243	<	int id = (int)(top & WAITER_INDEX_MASK);
1244	<	if (id <= 0 || (int)(top >>> EVENT_COUNT_SHIFT) == ec)
1245	<	return;
1246	<	if ((ws = workers).length < id || (w = ws[id - 1]) == null ||
1247	<	!UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
1248	<	top, top = w.nextWaiter))
1249	<	continue outer;      // possibly stale; reread
1250	<	LockSupport.unpark(w);
1251	<	if (top != eventWaiters) // let someone else take over
1252	<	return;
1236	>	/*
1237	>	* The while condition is true if: (there is are too few total
1238	>	* workers OR there is at least one waiter) AND (there are too
1239	>	* few active workers OR the pool is terminating).  The value
1240	>	* of e distinguishes the remaining cases: zero (no waiters)
1241	>	* for create, negative if terminating (in which case do
1242	>	* nothing), else release a waiter. The secondary checks for
1243	>	* release (non-null array etc) can fail if the pool begins
1244	>	* terminating after the test, and don't impose any added cost
1245	>	* because JVMs must perform null and bounds checks anyway.
1246	>	*/
1247	>	long c; int e, u;
1248	>	while ((((e = (int)(c = ctl)) | (u = (int)(c >>> 32))) &
1249	>	(INT_SIGN|SHORT_SIGN)) == (INT_SIGN|SHORT_SIGN)) {
1250	>	WorkQueue[] ws = workQueues; int i; WorkQueue w; Thread p;
1251	>	if (e == 0) {                    // add a new worker
1252	>	if (U.compareAndSwapLong
1253	>	(this, CTL, c, (long)(((u + UTC_UNIT) & UTC_MASK) |
1254	>	((u + UAC_UNIT) & UAC_MASK)) << 32)) {
1255	>	addWorker();
1256	>	break;
1257	>	}
1258		}
1259	<	}
1260	<	}
1261	<
1262	<	/**
1263	<	* If worker is inactive, blocks until terminating or event count
1264	<	* advances from last value held by worker; in any case helps
1265	<	* release others.
1266	<	*
1267	<	* @param w the calling worker thread
1268	<	*/
773	<	private void eventSync(ForkJoinWorkerThread w) {
774	<	if (!w.active) {
775	<	int prev = w.lastEventCount;
776	<	long nextTop = (((long)prev << EVENT_COUNT_SHIFT) |
777	<	((long)(w.poolIndex + 1)));
778	<	long top;
779	<	while ((runState < SHUTDOWN || !tryTerminate(false)) &&
780	<	(((int)(top = eventWaiters) & WAITER_INDEX_MASK) == 0 ||
781	<	(int)(top >>> EVENT_COUNT_SHIFT) == prev) &&
782	<	eventCount == prev) {
783	<	if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
784	<	w.nextWaiter = top, nextTop)) {
785	<	accumulateStealCount(w); // transfer steals while idle
786	<	Thread.interrupted();    // clear/ignore interrupt
787	<	while (eventCount == prev)
788	<	w.doPark();
1259	>	else if (e > 0 && ws != null &&
1260	>	(i = ((~e << 1) | 1) & SMASK) < ws.length &&
1261	>	(w = ws[i]) != null &&
1262	>	w.eventCount == (e | INT_SIGN)) {
1263	>	if (U.compareAndSwapLong
1264	>	(this, CTL, c, (((long)(w.nextWait & E_MASK)) |
1265	>	((long)(u + UAC_UNIT) << 32)))) {
1266	>	w.eventCount = (e + E_SEQ) & E_MASK;
1267	>	if ((p = w.parker) != null)
1268	>	U.unpark(p);         // release a waiting worker
1269		break;
1270		}
1271		}
1272	<	w.lastEventCount = eventCount;
1272	>	else
1273	>	break;
1274		}
794	–	releaseWaiters();
1275		}
1276
1277		/**
1278	<	* Callback from workers invoked upon each top-level action (i.e.,
1279	<	* stealing a task or taking a submission and running
1280	<	* it). Performs one or both of the following:
1281	<	*
1282	<	* * If the worker cannot find work, updates its active status to
1283	<	* inactive and updates activeCount unless there is contention, in
1284	<	* which case it may try again (either in this or a subsequent
1285	<	* call).  Additionally, awaits the next task event and/or helps
1286	<	* wake up other releasable waiters.
1287	<	*
1288	<	* * If there are too many running threads, suspends this worker
1289	<	* (first forcing inactivation if necessary).  If it is not
1290	<	* resumed before a keepAlive elapses, the worker may be "trimmed"
1291	<	* -- killed while suspended within suspendAsSpare. Otherwise,
1292	<	* upon resume it rechecks to make sure that it is still needed.
1293	<	*
1294	<	* @param w the worker
1295	<	* @param worked false if the worker scanned for work but didn't
1296	<	* find any (in which case it may block waiting for work).
1297	<	*/
1298	<	final void preStep(ForkJoinWorkerThread w, boolean worked) {
1299	<	boolean active = w.active;
1300	<	boolean inactivate = !worked & active;
1301	<	for (;;) {
1302	<	if (inactivate) {
1303	<	int c = runState;
1304	<	if (UNSAFE.compareAndSwapInt(this, runStateOffset,
1305	<	c, c - ONE_ACTIVE))
1306	<	inactivate = active = w.active = false;
1307	<	}
1308	<	int wc = workerCounts;
1309	<	if ((wc & RUNNING_COUNT_MASK) <= parallelism) {
1310	<	if (!worked)
1311	<	eventSync(w);
1312	<	return;
1278	>	* Tries to decrement active count (sometimes implicitly) and
1279	>	* possibly release or create a compensating worker in preparation
1280	>	* for blocking. Fails on contention or termination.
1281	>	*
1282	>	* @return true if the caller can block, else should recheck and retry
1283	>	*/
1284	>	final boolean tryCompensate() {
1285	>	WorkQueue[] ws; WorkQueue w; Thread p;
1286	>	int pc = parallelism, e, u, ac, tc, i;
1287	>	long c = ctl;
1288	>
1289	>	if ((e = (int)c) >= 0) {
1290	>	if ((ac = ((u = (int)(c >>> 32)) >> UAC_SHIFT)) <= 0 &&
1291	>	e != 0 && (ws = workQueues) != null &&
1292	>	(i = ((~e << 1) | 1) & SMASK) < ws.length &&
1293	>	(w = ws[i]) != null) {
1294	>	if (w.eventCount == (e | INT_SIGN) &&
1295	>	U.compareAndSwapLong
1296	>	(this, CTL, c, ((long)(w.nextWait & E_MASK) |
1297	>	(c & (AC_MASK|TC_MASK))))) {
1298	>	w.eventCount = (e + E_SEQ) & E_MASK;
1299	>	if ((p = w.parker) != null)
1300	>	U.unpark(p);
1301	>	return true;             // release an idle worker
1302	>	}
1303	>	}
1304	>	else if ((tc = (short)(u >>> UTC_SHIFT)) >= 0 && ac + pc > 1) {
1305	>	long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK);
1306	>	if (U.compareAndSwapLong(this, CTL, c, nc))
1307	>	return true;             // no compensation needed
1308	>	}
1309	>	else if (tc + pc < MAX_ID) {
1310	>	long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK);
1311	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1312	>	addWorker();
1313	>	return true;             // create replacement
1314	>	}
1315		}
834	–	if (!(inactivate |= active) &&  // must inactivate to suspend
835	–	UNSAFE.compareAndSwapInt(this, workerCountsOffset,
836	–	wc, wc - ONE_RUNNING) &&
837	–	!w.suspendAsSpare())        // false if trimmed
838	–	return;
1316		}
1317	+	return false;
1318		}
1319
1320	+	// Submissions
1321	+
1322		/**
1323	<	* Adjusts counts and creates or resumes compensating threads for
1324	<	* a worker about to block on task joinMe, returning early if
1325	<	* joinMe becomes ready. First tries resuming an existing spare
1326	<	* (which usually also avoids any count adjustment), but must then
1327	<	* decrement running count to determine whether a new thread is
1328	<	* needed. See above for fuller explanation.
1323	>	* Unless shutting down, adds the given task to some submission
1324	>	* queue; using a randomly chosen queue index if the caller is a
1325	>	* ForkJoinWorkerThread, else one based on caller thread's hash
1326	>	* code. If no queue exists at the index, one is created.  If the
1327	>	* queue is busy, another is chosen by sweeping through the queues
1328	>	* array.
1329		*/
1330	<	final void preJoin(ForkJoinTask<?> joinMe) {
1331	<	boolean dec = false;       // true when running count decremented
1330	>	private void doSubmit(ForkJoinTask<?> task) {
1331	>	if (task == null)
1332	>	throw new NullPointerException();
1333	>	Thread t = Thread.currentThread();
1334	>	int r = ((t instanceof ForkJoinWorkerThread) ?
1335	>	((ForkJoinWorkerThread)t).workQueue.nextSeed() : hashThread(t));
1336		for (;;) {
1337	<	releaseWaiters();      // help other threads progress
1338	<
1339	<	if (joinMe.status < 0) // surround spare search with done checks
1340	<	return;
1341	<	ForkJoinWorkerThread spare = null;
1342	<	for (ForkJoinWorkerThread w : workers) {
1343	<	if (w != null && w.isSuspended()) {
1344	<	spare = w;
1345	<	break;
1337	>	int rs = runState, m = rs & SMASK;
1338	>	int j = r &= (m & ~1);                      // even numbered queues
1339	>	WorkQueue[] ws = workQueues;
1340	>	if (rs < 0 || ws == null)
1341	>	throw new RejectedExecutionException(); // shutting down
1342	>	if (ws.length > m) {                        // consistency check
1343	>	for (WorkQueue q;;) {                   // circular sweep
1344	>	if (((q = ws[j]) != null ||
1345	>	(q = tryAddSharedQueue(j)) != null) &&
1346	>	q.trySharedPush(task)) {
1347	>	signalWork();
1348	>	return;
1349	>	}
1350	>	if ((j = (j + 2) & m) == r) {
1351	>	Thread.yield();                 // all queues busy
1352	>	break;
1353	>	}
1354		}
1355		}
1356	<	if (joinMe.status < 0)
1357	<	return;
866	<
867	<	if (spare != null && spare.tryUnsuspend()) {
868	<	if (dec || joinMe.requestSignal() < 0) {
869	<	int c;
870	<	do {} while (!UNSAFE.compareAndSwapInt(this,
871	<	workerCountsOffset,
872	<	c = workerCounts,
873	<	c + ONE_RUNNING));
874	<	} // else no net count change
875	<	LockSupport.unpark(spare);
876	<	return;
877	<	}
878	<
879	<	int wc = workerCounts; // decrement running count
880	<	if (!dec && (wc & RUNNING_COUNT_MASK) != 0 &&
881	<	(dec = UNSAFE.compareAndSwapInt(this, workerCountsOffset,
882	<	wc, wc -= ONE_RUNNING)) &&
883	<	joinMe.requestSignal() < 0) { // cannot block
884	<	int c;                        // back out
885	<	do {} while (!UNSAFE.compareAndSwapInt(this,
886	<	workerCountsOffset,
887	<	c = workerCounts,
888	<	c + ONE_RUNNING));
889	<	return;
890	<	}
1356	>	}
1357	>	}
1358
1359	<	if (dec) {
1360	<	int tc = wc >>> TOTAL_COUNT_SHIFT;
1361	<	int pc = parallelism;
1362	<	int dc = pc - (wc & RUNNING_COUNT_MASK); // deficit count
1363	<	if ((dc < pc && (dc <= 0 || (dc * dc < (tc - pc) * pc) ||
1364	<	!maintainsParallelism)) ||
1365	<	tc >= maxPoolSize) // cannot add
1366	<	return;
1367	<	if (spare == null &&
1368	<	UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
1369	<	wc + (ONE_RUNNING|ONE_TOTAL))) {
1370	<	addWorker();
1371	<	return;
1359	>	/**
1360	>	* Tries to add and register a new queue at the given index.
1361	>	*
1362	>	* @param idx the workQueues array index to register the queue
1363	>	* @return the queue, or null if could not add because could
1364	>	* not acquire lock or idx is unusable
1365	>	*/
1366	>	private WorkQueue tryAddSharedQueue(int idx) {
1367	>	WorkQueue q = null;
1368	>	ReentrantLock lock = this.lock;
1369	>	if (idx >= 0 && (idx & 1) == 0 && !lock.isLocked()) {
1370	>	// create queue outside of lock but only if apparently free
1371	>	WorkQueue nq = new WorkQueue(null, SHARED_QUEUE);
1372	>	if (lock.tryLock()) {
1373	>	try {
1374	>	WorkQueue[] ws = workQueues;
1375	>	if (ws != null && idx < ws.length) {
1376	>	if ((q = ws[idx]) == null) {
1377	>	int rs;         // update runState seq
1378	>	ws[idx] = q = nq;
1379	>	runState = (((rs = runState) & SHUTDOWN) |
1380	>	((rs + RS_SEQ) & ~SHUTDOWN));
1381	>	}
1382	>	}
1383	>	} finally {
1384	>	lock.unlock();
1385		}
1386		}
1387		}
1388	+	return q;
1389		}
1390
1391	+	// Scanning for tasks
1392	+
1393		/**
1394	<	* Same idea as preJoin but with too many differing details to
1395	<	* integrate: There are no task-based signal counts, and only one
1396	<	* way to do the actual blocking. So for simplicity it is directly
1397	<	* incorporated into this method.
1398	<	*/
1399	<	final void doBlock(ManagedBlocker blocker, boolean maintainPar)
1400	<	throws InterruptedException {
1401	<	maintainPar &= maintainsParallelism; // override
1402	<	boolean dec = false;
1403	<	boolean done = false;
1404	<	for (;;) {
1405	<	releaseWaiters();
1406	<	if (done = blocker.isReleasable())
1407	<	break;
1408	<	ForkJoinWorkerThread spare = null;
1409	<	for (ForkJoinWorkerThread w : workers) {
1410	<	if (w != null && w.isSuspended()) {
1411	<	spare = w;
1394	>	* Scans for and, if found, returns one task, else possibly
1395	>	* inactivates the worker. This method operates on single reads of
1396	>	* volatile state and is designed to be re-invoked continuously in
1397	>	* part because it returns upon detecting inconsistencies,
1398	>	* contention, or state changes that indicate possible success on
1399	>	* re-invocation.
1400	>	*
1401	>	* The scan searches for tasks across queues, randomly selecting
1402	>	* the first #queues probes, favoring steals 2:1 over submissions
1403	>	* (by exploiting even/odd indexing), and then performing a
1404	>	* circular sweep of all queues.  The scan terminates upon either
1405	>	* finding a non-empty queue, or completing a full sweep. If the
1406	>	* worker is not inactivated, it takes and returns a task from
1407	>	* this queue.  On failure to find a task, we take one of the
1408	>	* following actions, after which the caller will retry calling
1409	>	* this method unless terminated.
1410	>	*
1411	>	* * If not a complete sweep, try to release a waiting worker.  If
1412	>	* the scan terminated because the worker is inactivated, then the
1413	>	* released worker will often be the calling worker, and it can
1414	>	* succeed obtaining a task on the next call. Or maybe it is
1415	>	* another worker, but with same net effect. Releasing in other
1416	>	* cases as well ensures that we have enough workers running.
1417	>	*
1418	>	* * If the caller has run a task since the the last empty scan,
1419	>	* return (to allow rescan) if other workers are not also yet
1420	>	* enqueued.  Field WorkQueue.rescans counts down on each scan to
1421	>	* ensure eventual inactivation, and occasional calls to
1422	>	* Thread.yield to help avoid interference with more useful
1423	>	* activities on the system.
1424	>	*
1425	>	* * If pool is terminating, terminate the worker
1426	>	*
1427	>	* * If not already enqueued, try to inactivate and enqueue the
1428	>	* worker on wait queue.
1429	>	*
1430	>	* * If already enqueued and none of the above apply, either park
1431	>	* awaiting signal, or if this is the most recent waiter and pool
1432	>	* is quiescent, relay to idleAwaitWork to check for termination
1433	>	* and possibly shrink pool.
1434	>	*
1435	>	* @param w the worker (via its WorkQueue)
1436	>	* @return a task or null of none found
1437	>	*/
1438	>	private final ForkJoinTask<?> scan(WorkQueue w) {
1439	>	boolean swept = false;                 // true after full empty scan
1440	>	WorkQueue[] ws;                        // volatile read order matters
1441	>	int r = w.seed, ec = w.eventCount;     // ec is negative if inactive
1442	>	int rs = runState, m = rs & SMASK;
1443	>	if ((ws = workQueues) != null && ws.length > m) {
1444	>	ForkJoinTask<?> task = null;
1445	>	for (int k = 0, j = -2 - m; ; ++j) {
1446	>	WorkQueue q; int b;
1447	>	if (j < 0) {                    // random probes while j negative
1448	>	r ^= r << 13; r ^= r >>> 17; k = (r ^= r << 5) | (j & 1);
1449	>	}                               // worker (not submit) for odd j
1450	>	else                            // cyclic scan when j >= 0
1451	>	k += (m >>> 1) | 1;         // step by half to reduce bias
1452	>
1453	>	if ((q = ws[k & m]) != null && (b = q.base) - q.top < 0) {
1454	>	if (ec >= 0)
1455	>	task = q.pollAt(b);     // steal
1456		break;
1457		}
1458	<	}
1459	<	if (done = blocker.isReleasable())
1460	<	break;
1461	<	if (spare != null && spare.tryUnsuspend()) {
935	<	if (dec) {
936	<	int c;
937	<	do {} while (!UNSAFE.compareAndSwapInt(this,
938	<	workerCountsOffset,
939	<	c = workerCounts,
940	<	c + ONE_RUNNING));
1458	>	else if (j > m) {
1459	>	if (rs == runState)        // staleness check
1460	>	swept = true;
1461	>	break;
1462		}
942	–	LockSupport.unpark(spare);
943	–	break;
1463		}
1464	<	int wc = workerCounts;
1465	<	if (!dec && (wc & RUNNING_COUNT_MASK) != 0)
1466	<	dec = UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1467	<	wc, wc -= ONE_RUNNING);
1468	<	if (dec) {
1469	<	int tc = wc >>> TOTAL_COUNT_SHIFT;
1470	<	int pc = parallelism;
1471	<	int dc = pc - (wc & RUNNING_COUNT_MASK);
1472	<	if ((dc < pc && (dc <= 0 || (dc * dc < (tc - pc) * pc) ||
1473	<	!maintainPar)) ||
1474	<	tc >= maxPoolSize)
1464	>	w.seed = r;                        // save seed for next scan
1465	>	if (task != null)
1466	>	return task;
1467	>	}
1468	>
1469	>	// Decode ctl on empty scan
1470	>	long c = ctl; int e = (int)c, a = (int)(c >> AC_SHIFT), nr, ns;
1471	>	if (!swept) {                          // try to release a waiter
1472	>	WorkQueue v; Thread p;
1473	>	if (e > 0 && a < 0 && ws != null &&
1474	>	(v = ws[((~e << 1) | 1) & m]) != null &&
1475	>	v.eventCount == (e | INT_SIGN) && U.compareAndSwapLong
1476	>	(this, CTL, c, ((long)(v.nextWait & E_MASK) |
1477	>	((c + AC_UNIT) & (AC_MASK|TC_MASK))))) {
1478	>	v.eventCount = (e + E_SEQ) & E_MASK;
1479	>	if ((p = v.parker) != null)
1480	>	U.unpark(p);
1481	>	}
1482	>	}
1483	>	else if ((nr = w.rescans) > 0) {       // continue rescanning
1484	>	int ac = a + parallelism;
1485	>	if ((w.rescans = (ac < nr) ? ac : nr - 1) > 0 && w.seed < 0 &&
1486	>	w.eventCount == ec)
1487	>	Thread.yield();                // 1 bit randomness for yield call
1488	>	}
1489	>	else if (e < 0)                        // pool is terminating
1490	>	w.runState = -1;
1491	>	else if (ec >= 0) {                    // try to enqueue
1492	>	long nc = (long)ec | ((c - AC_UNIT) & (AC_MASK|TC_MASK));
1493	>	w.nextWait = e;
1494	>	w.eventCount = ec | INT_SIGN;      // mark as inactive
1495	>	if (!U.compareAndSwapLong(this, CTL, c, nc))
1496	>	w.eventCount = ec;             // back out on CAS failure
1497	>	else if ((ns = w.nsteals) != 0) {  // set rescans if ran task
1498	>	if (a <= 0)                    // ... unless too many active
1499	>	w.rescans = a + parallelism;
1500	>	w.nsteals = 0;
1501	>	w.totalSteals += ns;
1502	>	}
1503	>	}
1504	>	else{                                  // already queued
1505	>	if (parallelism == -a)
1506	>	idleAwaitWork(w);              // quiescent
1507	>	if (w.eventCount == ec) {
1508	>	Thread.interrupted();          // clear status
1509	>	ForkJoinWorkerThread wt = w.owner;
1510	>	U.putObject(wt, PARKBLOCKER, this);
1511	>	w.parker = wt;                 // emulate LockSupport.park
1512	>	if (w.eventCount == ec)        // recheck
1513	>	U.park(false, 0L);         // block
1514	>	w.parker = null;
1515	>	U.putObject(wt, PARKBLOCKER, null);
1516	>	}
1517	>	}
1518	>	return null;
1519	>	}
1520	>
1521	>	/**
1522	>	* If inactivating worker w has caused pool to become quiescent,
1523	>	* check for pool termination, and, so long as this is not the
1524	>	* only worker, wait for event for up to SHRINK_RATE nanosecs On
1525	>	* timeout, if ctl has not changed, terminate the worker, which
1526	>	* will in turn wake up another worker to possibly repeat this
1527	>	* process.
1528	>	*
1529	>	* @param w the calling worker
1530	>	*/
1531	>	private void idleAwaitWork(WorkQueue w) {
1532	>	long c; int nw, ec;
1533	>	if (!tryTerminate(false) &&
1534	>	(int)((c = ctl) >> AC_SHIFT) + parallelism == 0 &&
1535	>	(ec = w.eventCount) == ((int)c | INT_SIGN) &&
1536	>	(nw = w.nextWait) != 0) {
1537	>	long nc = ((long)(nw & E_MASK) | // ctl to restore on timeout
1538	>	((c + AC_UNIT) & AC_MASK) | (c & TC_MASK));
1539	>	ForkJoinTask.helpExpungeStaleExceptions(); // help clean
1540	>	ForkJoinWorkerThread wt = w.owner;
1541	>	while (ctl == c) {
1542	>	long startTime = System.nanoTime();
1543	>	Thread.interrupted();  // timed variant of version in scan()
1544	>	U.putObject(wt, PARKBLOCKER, this);
1545	>	w.parker = wt;
1546	>	if (ctl == c)
1547	>	U.park(false, SHRINK_RATE);
1548	>	w.parker = null;
1549	>	U.putObject(wt, PARKBLOCKER, null);
1550	>	if (ctl != c)
1551		break;
1552	<	if (spare == null &&
1553	<	UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
1554	<	wc + (ONE_RUNNING|ONE_TOTAL))){
1555	<	addWorker();
1552	>	if (System.nanoTime() - startTime >= SHRINK_TIMEOUT &&
1553	>	U.compareAndSwapLong(this, CTL, c, nc)) {
1554	>	w.runState = -1;          // shrink
1555	>	w.eventCount = (ec + E_SEQ) | E_MASK;
1556		break;
1557		}
1558		}
1559		}
965	–
966	–	try {
967	–	if (!done)
968	–	do {} while (!blocker.isReleasable() && !blocker.block());
969	–	} finally {
970	–	if (dec) {
971	–	int c;
972	–	do {} while (!UNSAFE.compareAndSwapInt(this,
973	–	workerCountsOffset,
974	–	c = workerCounts,
975	–	c + ONE_RUNNING));
976	–	}
977	–	}
1560		}
1561
1562		/**
1563	<	* Unless there are not enough other running threads, adjusts
1564	<	* counts for a a worker in performing helpJoin that cannot find
1565	<	* any work, so that this worker can now block.
1566	<	*
1567	<	* @return true if worker may block
1568	<	*/
1569	<	final boolean preBlockHelpingJoin(ForkJoinTask<?> joinMe) {
1570	<	while (joinMe.status >= 0) {
1571	<	releaseWaiters(); // help other threads progress
1572	<
1573	<	// if a spare exists, resume it to maintain parallelism level
1574	<	if ((workerCounts & RUNNING_COUNT_MASK) <= parallelism) {
1575	<	ForkJoinWorkerThread spare = null;
1576	<	for (ForkJoinWorkerThread w : workers) {
1577	<	if (w != null && w.isSuspended()) {
1578	<	spare = w;
1579	<	break;
1563	>	* Tries to locate and execute tasks for a stealer of the given
1564	>	* task, or in turn one of its stealers, Traces currentSteal ->
1565	>	* currentJoin links looking for a thread working on a descendant
1566	>	* of the given task and with a non-empty queue to steal back and
1567	>	* execute tasks from. The first call to this method upon a
1568	>	* waiting join will often entail scanning/search, (which is OK
1569	>	* because the joiner has nothing better to do), but this method
1570	>	* leaves hints in workers to speed up subsequent calls. The
1571	>	* implementation is very branchy to cope with potential
1572	>	* inconsistencies or loops encountering chains that are stale,
1573	>	* unknown, or of length greater than MAX_HELP_DEPTH links.  All
1574	>	* of these cases are dealt with by just retrying by caller.
1575	>	*
1576	>	* @param joiner the joining worker
1577	>	* @param task the task to join
1578	>	* @return true if found or ran a task (and so is immediately retryable)
1579	>	*/
1580	>	final boolean tryHelpStealer(WorkQueue joiner, ForkJoinTask<?> task) {
1581	>	ForkJoinTask<?> subtask;    // current target
1582	>	boolean progress = false;
1583	>	int depth = 0;              // current chain depth
1584	>	int m = runState & SMASK;
1585	>	WorkQueue[] ws = workQueues;
1586	>
1587	>	if (ws != null && ws.length > m && (subtask = task).status >= 0) {
1588	>	outer:for (WorkQueue j = joiner;;) {
1589	>	// Try to find the stealer of subtask, by first using hint
1590	>	WorkQueue stealer = null;
1591	>	WorkQueue v = ws[j.stealHint & m];
1592	>	if (v != null && v.currentSteal == subtask)
1593	>	stealer = v;
1594	>	else {
1595	>	for (int i = 1; i <= m; i += 2) {
1596	>	if ((v = ws[i]) != null && v.currentSteal == subtask) {
1597	>	stealer = v;
1598	>	j.stealHint = i; // save hint
1599	>	break;
1600	>	}
1601		}
1602	+	if (stealer == null)
1603	+	break;
1604		}
1605	<	if (joinMe.status < 0)
1606	<	break;
1607	<	if (spare != null) {
1608	<	if (spare.tryUnsuspend()) {
1609	<	boolean canBlock = true;
1610	<	if (joinMe.requestSignal() < 0) {
1611	<	canBlock = false; // already done
1612	<	int c;
1613	<	do {} while (!UNSAFE.compareAndSwapInt
1614	<	(this, workerCountsOffset,
1615	<	c = workerCounts, c + ONE_RUNNING));
1605	>
1606	>	for (WorkQueue q = stealer;;) { // Try to help stealer
1607	>	ForkJoinTask<?> t; int b;
1608	>	if (task.status < 0)
1609	>	break outer;
1610	>	if ((b = q.base) - q.top < 0) {
1611	>	progress = true;
1612	>	if (subtask.status < 0)
1613	>	break outer;               // stale
1614	>	if ((t = q.pollAt(b)) != null) {
1615	>	stealer.stealHint = joiner.poolIndex;
1616	>	joiner.runSubtask(t);
1617		}
1012	–	LockSupport.unpark(spare);
1013	–	return canBlock;
1618		}
1619	<	continue; // recheck -- another spare may exist
1619	>	else { // empty - try to descend to find stealer's stealer
1620	>	ForkJoinTask<?> next = stealer.currentJoin;
1621	>	if (++depth == MAX_HELP_DEPTH || subtask.status < 0 ||
1622	>	next == null || next == subtask)
1623	>	break outer;  // max depth, stale, dead-end, cyclic
1624	>	subtask = next;
1625	>	j = stealer;
1626	>	break;
1627	>	}
1628		}
1629		}
1018	–
1019	–	int wc = workerCounts; // reread to shorten CAS window
1020	–	int rc = wc & RUNNING_COUNT_MASK;
1021	–	if (rc <= 2) // keep this and at most one other thread alive
1022	–	break;
1023	–
1024	–	if (UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1025	–	wc, wc - ONE_RUNNING)) {
1026	–	if (joinMe.requestSignal() >= 0)
1027	–	return true;
1028	–	int c;                        // back out
1029	–	do {} while (!UNSAFE.compareAndSwapInt
1030	–	(this, workerCountsOffset,
1031	–	c = workerCounts, c + ONE_RUNNING));
1032	–	break;
1033	–	}
1630		}
1631	<	return false;
1631	>	return progress;
1632		}
1633
1634		/**
1635	<	* Possibly initiates and/or completes termination.
1635	>	* If task is at base of some steal queue, steals and executes it.
1636		*
1637	<	* @param now if true, unconditionally terminate, else only
1638	<	* if shutdown and empty queue and no active workers
1043	<	* @return true if now terminating or terminated
1637	>	* @param joiner the joining worker
1638	>	* @param task the task
1639		*/
1640	<	private boolean tryTerminate(boolean now) {
1641	<	if (now)
1642	<	advanceRunLevel(SHUTDOWN); // ensure at least SHUTDOWN
1643	<	else if (runState < SHUTDOWN ||
1644	<	!submissionQueue.isEmpty() ||
1645	<	(runState & ACTIVE_COUNT_MASK) != 0)
1646	<	return false;
1640	>	final void tryPollForAndExec(WorkQueue joiner, ForkJoinTask<?> task) {
1641	>	WorkQueue[] ws;
1642	>	int m = runState & SMASK;
1643	>	if ((ws = workQueues) != null && ws.length > m) {
1644	>	for (int j = 1; j <= m && task.status >= 0; j += 2) {
1645	>	WorkQueue q = ws[j];
1646	>	if (q != null && q.pollFor(task)) {
1647	>	joiner.runSubtask(task);
1648	>	break;
1649	>	}
1650	>	}
1651	>	}
1652	>	}
1653
1654	<	if (advanceRunLevel(TERMINATING))
1655	<	startTerminating();
1654	>	/**
1655	>	* Returns a non-empty steal queue, if one is found during a random,
1656	>	* then cyclic scan, else null.  This method must be retried by
1657	>	* caller if, by the time it tries to use the queue, it is empty.
1658	>	*/
1659	>	private WorkQueue findNonEmptyStealQueue(WorkQueue w) {
1660	>	int r = w.seed;    // Same idea as scan(), but ignoring submissions
1661	>	for (WorkQueue[] ws;;) {
1662	>	int m = runState & SMASK;
1663	>	if ((ws = workQueues) == null)
1664	>	return null;
1665	>	if (ws.length > m) {
1666	>	WorkQueue q;
1667	>	for (int n = m << 2, k = r, j = -n;;) {
1668	>	r ^= r << 13; r ^= r >>> 17; r ^= r << 5;
1669	>	if ((q = ws[(k | 1) & m]) != null && q.base - q.top < 0) {
1670	>	w.seed = r;
1671	>	return q;
1672	>	}
1673	>	else if (j > n)
1674	>	return null;
1675	>	else
1676	>	k = (j++ < 0) ? r : k + ((m >>> 1) | 1);
1677
1678	<	// Finish now if all threads terminated; else in some subsequent call
1679	<	if ((workerCounts >>> TOTAL_COUNT_SHIFT) == 0) {
1058	<	advanceRunLevel(TERMINATED);
1059	<	terminationLatch.countDown();
1678	>	}
1679	>	}
1680		}
1061	–	return true;
1681		}
1682
1683		/**
1684	<	* Actions on transition to TERMINATING
1685	<	*/
1686	<	private void startTerminating() {
1687	<	// Clear out and cancel submissions, ignoring exceptions
1688	<	ForkJoinTask<?> task;
1689	<	while ((task = submissionQueue.poll()) != null) {
1690	<	try {
1691	<	task.cancel(false);
1692	<	} catch (Throwable ignore) {
1684	>	* Runs tasks until {@code isQuiescent()}. We piggyback on
1685	>	* active count ctl maintenance, but rather than blocking
1686	>	* when tasks cannot be found, we rescan until all others cannot
1687	>	* find tasks either.
1688	>	*/
1689	>	final void helpQuiescePool(WorkQueue w) {
1690	>	for (boolean active = true;;) {
1691	>	w.runLocalTasks();      // exhaust local queue
1692	>	WorkQueue q = findNonEmptyStealQueue(w);
1693	>	if (q != null) {
1694	>	ForkJoinTask<?> t;
1695	>	if (!active) {      // re-establish active count
1696	>	long c;
1697	>	active = true;
1698	>	do {} while (!U.compareAndSwapLong
1699	>	(this, CTL, c = ctl, c + AC_UNIT));
1700	>	}
1701	>	if ((t = q.poll()) != null)
1702	>	w.runSubtask(t);
1703		}
1704	<	}
1705	<	// Propagate run level
1706	<	for (ForkJoinWorkerThread w : workers) {
1707	<	if (w != null)
1708	<	w.shutdown();    // also resumes suspended workers
1709	<	}
1710	<	// Ensure no straggling local tasks
1711	<	for (ForkJoinWorkerThread w : workers) {
1712	<	if (w != null)
1713	<	w.cancelTasks();
1714	<	}
1715	<	// Wake up idle workers
1716	<	advanceEventCount();
1088	<	releaseWaiters();
1089	<	// Unstick pending joins
1090	<	for (ForkJoinWorkerThread w : workers) {
1091	<	if (w != null && !w.isTerminated()) {
1092	<	try {
1093	<	w.interrupt();
1094	<	} catch (SecurityException ignore) {
1704	>	else {
1705	>	long c;
1706	>	if (active) {       // decrement active count without queuing
1707	>	active = false;
1708	>	do {} while (!U.compareAndSwapLong
1709	>	(this, CTL, c = ctl, c -= AC_UNIT));
1710	>	}
1711	>	else
1712	>	c = ctl;        // re-increment on exit
1713	>	if ((int)(c >> AC_SHIFT) + parallelism == 0) {
1714	>	do {} while (!U.compareAndSwapLong
1715	>	(this, CTL, c = ctl, c + AC_UNIT));
1716	>	break;
1717		}
1718		}
1719		}
1720		}
1721
1722	<	// misc support for ForkJoinWorkerThread
1722	>	/**
1723	>	* Gets and removes a local or stolen task for the given worker
1724	>	*
1725	>	* @return a task, if available
1726	>	*/
1727	>	final ForkJoinTask<?> nextTaskFor(WorkQueue w) {
1728	>	for (ForkJoinTask<?> t;;) {
1729	>	WorkQueue q;
1730	>	if ((t = w.nextLocalTask()) != null)
1731	>	return t;
1732	>	if ((q = findNonEmptyStealQueue(w)) == null)
1733	>	return null;
1734	>	if ((t = q.poll()) != null)
1735	>	return t;
1736	>	}
1737	>	}
1738
1739		/**
1740	<	* Returns pool number
1740	>	* Returns the approximate (non-atomic) number of idle threads per
1741	>	* active thread to offset steal queue size for method
1742	>	* ForkJoinTask.getSurplusQueuedTaskCount().
1743		*/
1744	<	final int getPoolNumber() {
1745	<	return poolNumber;
1744	>	final int idlePerActive() {
1745	>	// Approximate at powers of two for small values, saturate past 4
1746	>	int p = parallelism;
1747	>	int a = p + (int)(ctl >> AC_SHIFT);
1748	>	return (a > (p >>>= 1) ? 0 :
1749	>	a > (p >>>= 1) ? 1 :
1750	>	a > (p >>>= 1) ? 2 :
1751	>	a > (p >>>= 1) ? 4 :
1752	>	8);
1753	>	}
1754	>
1755	>	// Termination
1756	>
1757	>	/**
1758	>	* Sets SHUTDOWN bit of runState under lock
1759	>	*/
1760	>	private void enableShutdown() {
1761	>	ReentrantLock lock = this.lock;
1762	>	if (runState >= 0) {
1763	>	lock.lock();                       // don't need try/finally
1764	>	runState |= SHUTDOWN;
1765	>	lock.unlock();
1766	>	}
1767		}
1768
1769		/**
1770	<	* Accumulates steal count from a worker, clearing
1771	<	* the worker's value
1770	>	* Possibly initiates and/or completes termination.  Upon
1771	>	* termination, cancels all queued tasks and then
1772	>	*
1773	>	* @param now if true, unconditionally terminate, else only
1774	>	* if no work and no active workers
1775	>	* @return true if now terminating or terminated
1776		*/
1777	<	final void accumulateStealCount(ForkJoinWorkerThread w) {
1778	<	int sc = w.stealCount;
1779	<	if (sc != 0) {
1780	<	long c;
1781	<	w.stealCount = 0;
1782	<	do {} while (!UNSAFE.compareAndSwapLong(this, stealCountOffset,
1783	<	c = stealCount, c + sc));
1777	>	private boolean tryTerminate(boolean now) {
1778	>	for (long c;;) {
1779	>	if (((c = ctl) & STOP_BIT) != 0) {      // already terminating
1780	>	if ((short)(c >>> TC_SHIFT) == -parallelism) {
1781	>	ReentrantLock lock = this.lock; // signal when no workers
1782	>	lock.lock();                    // don't need try/finally
1783	>	termination.signalAll();        // signal when 0 workers
1784	>	lock.unlock();
1785	>	}
1786	>	return true;
1787	>	}
1788	>	if (!now) {
1789	>	if ((int)(c >> AC_SHIFT) != -parallelism || runState >= 0 ||
1790	>	hasQueuedSubmissions())
1791	>	return false;
1792	>	// Check for unqueued inactive workers. One pass suffices.
1793	>	WorkQueue[] ws = workQueues; WorkQueue w;
1794	>	if (ws != null) {
1795	>	int n = ws.length;
1796	>	for (int i = 1; i < n; i += 2) {
1797	>	if ((w = ws[i]) != null && w.eventCount >= 0)
1798	>	return false;
1799	>	}
1800	>	}
1801	>	}
1802	>	if (U.compareAndSwapLong(this, CTL, c, c | STOP_BIT))
1803	>	startTerminating();
1804		}
1805		}
1806
1807		/**
1808	<	* Returns the approximate (non-atomic) number of idle threads per
1809	<	* active thread.
1808	>	* Initiates termination: Runs three passes through workQueues:
1809	>	* (0) Setting termination status, followed by wakeups of queued
1810	>	* workers; (1) cancelling all tasks; (2) interrupting lagging
1811	>	* threads (likely in external tasks, but possibly also blocked in
1812	>	* joins).  Each pass repeats previous steps because of potential
1813	>	* lagging thread creation.
1814		*/
1815	<	final int idlePerActive() {
1816	<	int ac = runState;    // no mask -- artifically boosts during shutdown
1817	<	int pc = parallelism; // use targeted parallelism, not rc
1818	<	// Use exact results for small values, saturate past 4
1819	<	return pc <= ac? 0 : pc >>> 1 <= ac? 1 : pc >>> 2 <= ac? 3 : pc >>> 3;
1815	>	private void startTerminating() {
1816	>	for (int pass = 0; pass < 3; ++pass) {
1817	>	WorkQueue[] ws = workQueues;
1818	>	if (ws != null) {
1819	>	WorkQueue w; Thread wt;
1820	>	int n = ws.length;
1821	>	for (int j = 0; j < n; ++j) {
1822	>	if ((w = ws[j]) != null) {
1823	>	w.runState = -1;
1824	>	if (pass > 0) {
1825	>	w.cancelAll();
1826	>	if (pass > 1 && (wt = w.owner) != null &&
1827	>	!wt.isInterrupted()) {
1828	>	try {
1829	>	wt.interrupt();
1830	>	} catch (SecurityException ignore) {
1831	>	}
1832	>	}
1833	>	}
1834	>	}
1835	>	}
1836	>	// Wake up workers parked on event queue
1837	>	int i, e; long c; Thread p;
1838	>	while ((i = ((~(e = (int)(c = ctl)) << 1) | 1) & SMASK) < n &&
1839	>	(w = ws[i]) != null &&
1840	>	w.eventCount == (e | INT_SIGN)) {
1841	>	long nc = ((long)(w.nextWait & E_MASK) |
1842	>	((c + AC_UNIT) & AC_MASK) |
1843	>	(c & (TC_MASK|STOP_BIT)));
1844	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1845	>	w.eventCount = (e + E_SEQ) & E_MASK;
1846	>	if ((p = w.parker) != null)
1847	>	U.unpark(p);
1848	>	}
1849	>	}
1850	>	}
1851	>	}
1852		}
1853
1854	<	// Public and protected methods
1854	>	// Exported methods
1855
1856		// Constructors
1857
1858		/**
1859		* Creates a {@code ForkJoinPool} with parallelism equal to {@link
1860	<	* java.lang.Runtime#availableProcessors}, and using the {@linkplain
1861	<	* #defaultForkJoinWorkerThreadFactory default thread factory}.
1860	>	* java.lang.Runtime#availableProcessors}, using the {@linkplain
1861	>	* #defaultForkJoinWorkerThreadFactory default thread factory},
1862	>	* no UncaughtExceptionHandler, and non-async LIFO processing mode.
1863		*
1864		* @throws SecurityException if a security manager exists and
1865		*         the caller is not permitted to modify threads
#	Line 1147 | Line 1868 | public class ForkJoinPool extends Abstra
1868		*/
1869		public ForkJoinPool() {
1870		this(Runtime.getRuntime().availableProcessors(),
1871	<	defaultForkJoinWorkerThreadFactory);
1871	>	defaultForkJoinWorkerThreadFactory, null, false);
1872		}
1873
1874		/**
1875		* Creates a {@code ForkJoinPool} with the indicated parallelism
1876	<	* level and using the {@linkplain
1877	<	* #defaultForkJoinWorkerThreadFactory default thread factory}.
1876	>	* level, the {@linkplain
1877	>	* #defaultForkJoinWorkerThreadFactory default thread factory},
1878	>	* no UncaughtExceptionHandler, and non-async LIFO processing mode.
1879		*
1880		* @param parallelism the parallelism level
1881		* @throws IllegalArgumentException if parallelism less than or
#	Line 1164 | Line 1886 | public class ForkJoinPool extends Abstra
1886		*         java.lang.RuntimePermission}{@code ("modifyThread")}
1887		*/
1888		public ForkJoinPool(int parallelism) {
1889	<	this(parallelism, defaultForkJoinWorkerThreadFactory);
1889	>	this(parallelism, defaultForkJoinWorkerThreadFactory, null, false);
1890		}
1891
1892		/**
1893	<	* Creates a {@code ForkJoinPool} with parallelism equal to {@link
1172	<	* java.lang.Runtime#availableProcessors}, and using the given
1173	<	* thread factory.
1893	>	* Creates a {@code ForkJoinPool} with the given parameters.
1894		*
1895	<	* @param factory the factory for creating new threads
1896	<	* @throws NullPointerException if the factory is null
1897	<	* @throws SecurityException if a security manager exists and
1898	<	*         the caller is not permitted to modify threads
1899	<	*         because it does not hold {@link
1900	<	*         java.lang.RuntimePermission}{@code ("modifyThread")}
1901	<	*/
1902	<	public ForkJoinPool(ForkJoinWorkerThreadFactory factory) {
1903	<	this(Runtime.getRuntime().availableProcessors(), factory);
1904	<	}
1905	<
1906	<	/**
1907	<	* Creates a {@code ForkJoinPool} with the given parallelism and
1188	<	* thread factory.
1189	<	*
1190	<	* @param parallelism the parallelism level
1191	<	* @param factory the factory for creating new threads
1895	>	* @param parallelism the parallelism level. For default value,
1896	>	* use {@link java.lang.Runtime#availableProcessors}.
1897	>	* @param factory the factory for creating new threads. For default value,
1898	>	* use {@link #defaultForkJoinWorkerThreadFactory}.
1899	>	* @param handler the handler for internal worker threads that
1900	>	* terminate due to unrecoverable errors encountered while executing
1901	>	* tasks. For default value, use {@code null}.
1902	>	* @param asyncMode if true,
1903	>	* establishes local first-in-first-out scheduling mode for forked
1904	>	* tasks that are never joined. This mode may be more appropriate
1905	>	* than default locally stack-based mode in applications in which
1906	>	* worker threads only process event-style asynchronous tasks.
1907	>	* For default value, use {@code false}.
1908		* @throws IllegalArgumentException if parallelism less than or
1909		*         equal to zero, or greater than implementation limit
1910		* @throws NullPointerException if the factory is null
#	Line 1197 | Line 1913 | public class ForkJoinPool extends Abstra
1913		*         because it does not hold {@link
1914		*         java.lang.RuntimePermission}{@code ("modifyThread")}
1915		*/
1916	<	public ForkJoinPool(int parallelism, ForkJoinWorkerThreadFactory factory) {
1916	>	public ForkJoinPool(int parallelism,
1917	>	ForkJoinWorkerThreadFactory factory,
1918	>	Thread.UncaughtExceptionHandler handler,
1919	>	boolean asyncMode) {
1920		checkPermission();
1921		if (factory == null)
1922		throw new NullPointerException();
1923	<	if (parallelism <= 0 || parallelism > MAX_THREADS)
1923	>	if (parallelism <= 0 || parallelism > MAX_ID)
1924		throw new IllegalArgumentException();
1206	–	this.poolNumber = poolNumberGenerator.incrementAndGet();
1207	–	int arraySize = initialArraySizeFor(parallelism);
1925		this.parallelism = parallelism;
1926		this.factory = factory;
1927	<	this.maxPoolSize = MAX_THREADS;
1928	<	this.maintainsParallelism = true;
1929	<	this.workers = new ForkJoinWorkerThread[arraySize];
1930	<	this.submissionQueue = new LinkedTransferQueue<ForkJoinTask<?>>();
1931	<	this.workerLock = new ReentrantLock();
1932	<	this.terminationLatch = new CountDownLatch(1);
1933	<	// Start first worker; remaining workers added upon first submission
1934	<	workerCounts = ONE_RUNNING | ONE_TOTAL;
1935	<	addWorker();
1936	<	}
1937	<
1938	<	/**
1939	<	* Returns initial power of two size for workers array.
1940	<	* @param pc the initial parallelism level
1941	<	*/
1942	<	private static int initialArraySizeFor(int pc) {
1943	<	// See Hackers Delight, sec 3.2. We know MAX_THREADS < (1 >>> 16)
1944	<	int size = pc < MAX_THREADS ? pc + 1 : MAX_THREADS;
1945	<	size |= size >>> 1;
1946	<	size |= size >>> 2;
1947	<	size |= size >>> 4;
1948	<	size |= size >>> 8;
1949	<	return size + 1;
1927	>	this.ueh = handler;
1928	>	this.localMode = asyncMode ? FIFO_QUEUE : LIFO_QUEUE;
1929	>	this.nextPoolIndex = 1;
1930	>	long np = (long)(-parallelism); // offset ctl counts
1931	>	this.ctl = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK);
1932	>	// initialize workQueues array with room for 2*parallelism if possible
1933	>	int n = parallelism << 1;
1934	>	if (n >= MAX_ID)
1935	>	n = MAX_ID;
1936	>	else { // See Hackers Delight, sec 3.2, where n < (1 << 16)
1937	>	n |= n >>> 1; n |= n >>> 2; n |= n >>> 4; n |= n >>> 8;
1938	>	}
1939	>	this.workQueues = new WorkQueue[(n + 1) << 1];
1940	>	ReentrantLock lck = this.lock = new ReentrantLock();
1941	>	this.termination = lck.newCondition();
1942	>	this.stealCount = new AtomicLong();
1943	>	this.nextWorkerNumber = new AtomicInteger();
1944	>	StringBuilder sb = new StringBuilder("ForkJoinPool-");
1945	>	sb.append(poolNumberGenerator.incrementAndGet());
1946	>	sb.append("-worker-");
1947	>	this.workerNamePrefix = sb.toString();
1948	>	// Create initial submission queue
1949	>	WorkQueue sq = tryAddSharedQueue(0);
1950	>	if (sq != null)
1951	>	sq.growArray(false);
1952		}
1953
1954		// Execution methods
1955
1956		/**
1238	–	* Common code for execute, invoke and submit
1239	–	*/
1240	–	private <T> void doSubmit(ForkJoinTask<T> task) {
1241	–	if (task == null)
1242	–	throw new NullPointerException();
1243	–	if (runState >= SHUTDOWN)
1244	–	throw new RejectedExecutionException();
1245	–	submissionQueue.offer(task);
1246	–	advanceEventCount();
1247	–	releaseWaiters();
1248	–	if ((workerCounts >>> TOTAL_COUNT_SHIFT) < parallelism)
1249	–	ensureEnoughTotalWorkers();
1250	–	}
1251	–
1252	–	/**
1957		* Performs the given task, returning its result upon completion.
1958	+	* If the computation encounters an unchecked Exception or Error,
1959	+	* it is rethrown as the outcome of this invocation.  Rethrown
1960	+	* exceptions behave in the same way as regular exceptions, but,
1961	+	* when possible, contain stack traces (as displayed for example
1962	+	* using {@code ex.printStackTrace()}) of both the current thread
1963	+	* as well as the thread actually encountering the exception;
1964	+	* minimally only the latter.
1965		*
1966		* @param task the task
1967		* @return the task's result
#	Line 1283 | Line 1994 | public class ForkJoinPool extends Abstra
1994		*         scheduled for execution
1995		*/
1996		public void execute(Runnable task) {
1997	+	if (task == null)
1998	+	throw new NullPointerException();
1999		ForkJoinTask<?> job;
2000		if (task instanceof ForkJoinTask<?>) // avoid re-wrap
2001		job = (ForkJoinTask<?>) task;
#	Line 1292 | Line 2005 | public class ForkJoinPool extends Abstra
2005		}
2006
2007		/**
2008	+	* Submits a ForkJoinTask for execution.
2009	+	*
2010	+	* @param task the task to submit
2011	+	* @return the task
2012	+	* @throws NullPointerException if the task is null
2013	+	* @throws RejectedExecutionException if the task cannot be
2014	+	*         scheduled for execution
2015	+	*/
2016	+	public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) {
2017	+	doSubmit(task);
2018	+	return task;
2019	+	}
2020	+
2021	+	/**
2022		* @throws NullPointerException if the task is null
2023		* @throws RejectedExecutionException if the task cannot be
2024		*         scheduled for execution
2025		*/
2026		public <T> ForkJoinTask<T> submit(Callable<T> task) {
2027	+	if (task == null)
2028	+	throw new NullPointerException();
2029		ForkJoinTask<T> job = ForkJoinTask.adapt(task);
2030		doSubmit(job);
2031		return job;
#	Line 1308 | Line 2037 | public class ForkJoinPool extends Abstra
2037		*         scheduled for execution
2038		*/
2039		public <T> ForkJoinTask<T> submit(Runnable task, T result) {
2040	+	if (task == null)
2041	+	throw new NullPointerException();
2042		ForkJoinTask<T> job = ForkJoinTask.adapt(task, result);
2043		doSubmit(job);
2044		return job;
#	Line 1319 | Line 2050 | public class ForkJoinPool extends Abstra
2050		*         scheduled for execution
2051		*/
2052		public ForkJoinTask<?> submit(Runnable task) {
2053	+	if (task == null)
2054	+	throw new NullPointerException();
2055		ForkJoinTask<?> job;
2056		if (task instanceof ForkJoinTask<?>) // avoid re-wrap
2057		job = (ForkJoinTask<?>) task;
#	Line 1329 | Line 2062 | public class ForkJoinPool extends Abstra
2062		}
2063
2064		/**
1332	–	* Submits a ForkJoinTask for execution.
1333	–	*
1334	–	* @param task the task to submit
1335	–	* @return the task
1336	–	* @throws NullPointerException if the task is null
1337	–	* @throws RejectedExecutionException if the task cannot be
1338	–	*         scheduled for execution
1339	–	*/
1340	–	public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) {
1341	–	doSubmit(task);
1342	–	return task;
1343	–	}
1344	–
1345	–	/**
2065		* @throws NullPointerException       {@inheritDoc}
2066		* @throws RejectedExecutionException {@inheritDoc}
2067		*/
#	Line 1384 | Line 2103 | public class ForkJoinPool extends Abstra
2103		* @return the handler, or {@code null} if none
2104		*/
2105		public Thread.UncaughtExceptionHandler getUncaughtExceptionHandler() {
1387	–	workerCountReadFence();
2106		return ueh;
2107		}
2108
2109		/**
1392	–	* Sets the handler for internal worker threads that terminate due
1393	–	* to unrecoverable errors encountered while executing tasks.
1394	–	* Unless set, the current default or ThreadGroup handler is used
1395	–	* as handler.
1396	–	*
1397	–	* @param h the new handler
1398	–	* @return the old handler, or {@code null} if none
1399	–	* @throws SecurityException if a security manager exists and
1400	–	*         the caller is not permitted to modify threads
1401	–	*         because it does not hold {@link
1402	–	*         java.lang.RuntimePermission}{@code ("modifyThread")}
1403	–	*/
1404	–	public Thread.UncaughtExceptionHandler
1405	–	setUncaughtExceptionHandler(Thread.UncaughtExceptionHandler h) {
1406	–	checkPermission();
1407	–	workerCountReadFence();
1408	–	Thread.UncaughtExceptionHandler old = ueh;
1409	–	if (h != old) {
1410	–	ueh = h;
1411	–	workerCountWriteFence();
1412	–	for (ForkJoinWorkerThread w : workers) {
1413	–	if (w != null)
1414	–	w.setUncaughtExceptionHandler(h);
1415	–	}
1416	–	}
1417	–	return old;
1418	–	}
1419	–
1420	–	/**
1421	–	* Sets the target parallelism level of this pool.
1422	–	*
1423	–	* @param parallelism the target parallelism
1424	–	* @throws IllegalArgumentException if parallelism less than or
1425	–	* equal to zero or greater than maximum size bounds
1426	–	* @throws SecurityException if a security manager exists and
1427	–	*         the caller is not permitted to modify threads
1428	–	*         because it does not hold {@link
1429	–	*         java.lang.RuntimePermission}{@code ("modifyThread")}
1430	–	*/
1431	–	public void setParallelism(int parallelism) {
1432	–	checkPermission();
1433	–	if (parallelism <= 0 || parallelism > maxPoolSize)
1434	–	throw new IllegalArgumentException();
1435	–	workerCountReadFence();
1436	–	int pc = this.parallelism;
1437	–	if (pc != parallelism) {
1438	–	this.parallelism = parallelism;
1439	–	workerCountWriteFence();
1440	–	// Release spares. If too many, some will die after re-suspend
1441	–	for (ForkJoinWorkerThread w : workers) {
1442	–	if (w != null && w.tryUnsuspend()) {
1443	–	updateRunningCount(1);
1444	–	LockSupport.unpark(w);
1445	–	}
1446	–	}
1447	–	ensureEnoughTotalWorkers();
1448	–	advanceEventCount();
1449	–	releaseWaiters(); // force config recheck by existing workers
1450	–	}
1451	–	}
1452	–
1453	–	/**
2110		* Returns the targeted parallelism level of this pool.
2111		*
2112		* @return the targeted parallelism level of this pool
2113		*/
2114		public int getParallelism() {
1459	–	//        workerCountReadFence(); // inlined below
1460	–	int ignore = workerCounts;
2115		return parallelism;
2116		}
2117
2118		/**
2119		* Returns the number of worker threads that have started but not
2120	<	* yet terminated.  This result returned by this method may differ
2120	>	* yet terminated.  The result returned by this method may differ
2121		* from {@link #getParallelism} when threads are created to
2122		* maintain parallelism when others are cooperatively blocked.
2123		*
2124		* @return the number of worker threads
2125		*/
2126		public int getPoolSize() {
2127	<	return workerCounts >>> TOTAL_COUNT_SHIFT;
1474	<	}
1475	<
1476	<	/**
1477	<	* Returns the maximum number of threads allowed to exist in the
1478	<	* pool. Unless set using {@link #setMaximumPoolSize}, the
1479	<	* maximum is an implementation-defined value designed only to
1480	<	* prevent runaway growth.
1481	<	*
1482	<	* @return the maximum
1483	<	*/
1484	<	public int getMaximumPoolSize() {
1485	<	workerCountReadFence();
1486	<	return maxPoolSize;
1487	<	}
1488	<
1489	<	/**
1490	<	* Sets the maximum number of threads allowed to exist in the
1491	<	* pool. The given value should normally be greater than or equal
1492	<	* to the {@link #getParallelism parallelism} level. Setting this
1493	<	* value has no effect on current pool size. It controls
1494	<	* construction of new threads. The use of this method may cause
1495	<	* tasks that intrinsically require extra threads for dependent
1496	<	* computations to indefinitely stall. If you are instead trying
1497	<	* to minimize internal thread creation, consider setting {@link
1498	<	* #setMaintainsParallelism} as false.
1499	<	*
1500	<	* @throws IllegalArgumentException if negative or greater than
1501	<	* internal implementation limit
1502	<	*/
1503	<	public void setMaximumPoolSize(int newMax) {
1504	<	if (newMax < 0 || newMax > MAX_THREADS)
1505	<	throw new IllegalArgumentException();
1506	<	maxPoolSize = newMax;
1507	<	workerCountWriteFence();
1508	<	}
1509	<
1510	<	/**
1511	<	* Returns {@code true} if this pool dynamically maintains its
1512	<	* target parallelism level. If false, new threads are added only
1513	<	* to avoid possible starvation.  This setting is by default true.
1514	<	*
1515	<	* @return {@code true} if maintains parallelism
1516	<	*/
1517	<	public boolean getMaintainsParallelism() {
1518	<	workerCountReadFence();
1519	<	return maintainsParallelism;
1520	<	}
1521	<
1522	<	/**
1523	<	* Sets whether this pool dynamically maintains its target
1524	<	* parallelism level. If false, new threads are added only to
1525	<	* avoid possible starvation.
1526	<	*
1527	<	* @param enable {@code true} to maintain parallelism
1528	<	*/
1529	<	public void setMaintainsParallelism(boolean enable) {
1530	<	maintainsParallelism = enable;
1531	<	workerCountWriteFence();
1532	<	}
1533	<
1534	<	/**
1535	<	* Establishes local first-in-first-out scheduling mode for forked
1536	<	* tasks that are never joined. This mode may be more appropriate
1537	<	* than default locally stack-based mode in applications in which
1538	<	* worker threads only process asynchronous tasks.  This method is
1539	<	* designed to be invoked only when the pool is quiescent, and
1540	<	* typically only before any tasks are submitted. The effects of
1541	<	* invocations at other times may be unpredictable.
1542	<	*
1543	<	* @param async if {@code true}, use locally FIFO scheduling
1544	<	* @return the previous mode
1545	<	* @see #getAsyncMode
1546	<	*/
1547	<	public boolean setAsyncMode(boolean async) {
1548	<	workerCountReadFence();
1549	<	boolean oldMode = locallyFifo;
1550	<	if (oldMode != async) {
1551	<	locallyFifo = async;
1552	<	workerCountWriteFence();
1553	<	for (ForkJoinWorkerThread w : workers) {
1554	<	if (w != null)
1555	<	w.setAsyncMode(async);
1556	<	}
1557	<	}
1558	<	return oldMode;
2127	>	return parallelism + (short)(ctl >>> TC_SHIFT);
2128		}
2129
2130		/**
#	Line 1563 | Line 2132 | public class ForkJoinPool extends Abstra
2132		* scheduling mode for forked tasks that are never joined.
2133		*
2134		* @return {@code true} if this pool uses async mode
1566	–	* @see #setAsyncMode
2135		*/
2136		public boolean getAsyncMode() {
2137	<	workerCountReadFence();
1570	<	return locallyFifo;
2137	>	return localMode != 0;
2138		}
2139
2140		/**
#	Line 1579 | Line 2146 | public class ForkJoinPool extends Abstra
2146		* @return the number of worker threads
2147		*/
2148		public int getRunningThreadCount() {
2149	<	return workerCounts & RUNNING_COUNT_MASK;
2149	>	int rc = 0;
2150	>	WorkQueue[] ws; WorkQueue w;
2151	>	if ((ws = workQueues) != null) {
2152	>	int n = ws.length;
2153	>	for (int i = 1; i < n; i += 2) {
2154	>	Thread.State s; ForkJoinWorkerThread wt;
2155	>	if ((w = ws[i]) != null && (wt = w.owner) != null &&
2156	>	w.eventCount >= 0 &&
2157	>	(s = wt.getState()) != Thread.State.BLOCKED &&
2158	>	s != Thread.State.WAITING &&
2159	>	s != Thread.State.TIMED_WAITING)
2160	>	++rc;
2161	>	}
2162	>	}
2163	>	return rc;
2164		}
2165
2166		/**
#	Line 1590 | Line 2171 | public class ForkJoinPool extends Abstra
2171		* @return the number of active threads
2172		*/
2173		public int getActiveThreadCount() {
2174	<	return runState & ACTIVE_COUNT_MASK;
2174	>	int r = parallelism + (int)(ctl >> AC_SHIFT);
2175	>	return (r <= 0) ? 0 : r; // suppress momentarily negative values
2176		}
2177
2178		/**
#	Line 1605 | Line 2187 | public class ForkJoinPool extends Abstra
2187		* @return {@code true} if all threads are currently idle
2188		*/
2189		public boolean isQuiescent() {
2190	<	return (runState & ACTIVE_COUNT_MASK) == 0;
2190	>	return (int)(ctl >> AC_SHIFT) + parallelism == 0;
2191		}
2192
2193		/**
#	Line 1620 | Line 2202 | public class ForkJoinPool extends Abstra
2202		* @return the number of steals
2203		*/
2204		public long getStealCount() {
2205	<	return stealCount;
2205	>	long count = stealCount.get();
2206	>	WorkQueue[] ws; WorkQueue w;
2207	>	if ((ws = workQueues) != null) {
2208	>	int n = ws.length;
2209	>	for (int i = 1; i < n; i += 2) {
2210	>	if ((w = ws[i]) != null)
2211	>	count += w.totalSteals;
2212	>	}
2213	>	}
2214	>	return count;
2215		}
2216
2217		/**
#	Line 1635 | Line 2226 | public class ForkJoinPool extends Abstra
2226		*/
2227		public long getQueuedTaskCount() {
2228		long count = 0;
2229	<	for (ForkJoinWorkerThread w : workers) {
2230	<	if (w != null)
2231	<	count += w.getQueueSize();
2229	>	WorkQueue[] ws; WorkQueue w;
2230	>	if ((ws = workQueues) != null) {
2231	>	int n = ws.length;
2232	>	for (int i = 1; i < n; i += 2) {
2233	>	if ((w = ws[i]) != null)
2234	>	count += w.queueSize();
2235	>	}
2236		}
2237		return count;
2238		}
2239
2240		/**
2241		* Returns an estimate of the number of tasks submitted to this
2242	<	* pool that have not yet begun executing.  This method takes time
2243	<	* proportional to the number of submissions.
2242	>	* pool that have not yet begun executing.  This method may take
2243	>	* time proportional to the number of submissions.
2244		*
2245		* @return the number of queued submissions
2246		*/
2247		public int getQueuedSubmissionCount() {
2248	<	return submissionQueue.size();
2248	>	int count = 0;
2249	>	WorkQueue[] ws; WorkQueue w;
2250	>	if ((ws = workQueues) != null) {
2251	>	int n = ws.length;
2252	>	for (int i = 0; i < n; i += 2) {
2253	>	if ((w = ws[i]) != null)
2254	>	count += w.queueSize();
2255	>	}
2256	>	}
2257	>	return count;
2258		}
2259
2260		/**
#	Line 1660 | Line 2264 | public class ForkJoinPool extends Abstra
2264		* @return {@code true} if there are any queued submissions
2265		*/
2266		public boolean hasQueuedSubmissions() {
2267	<	return !submissionQueue.isEmpty();
2267	>	WorkQueue[] ws; WorkQueue w;
2268	>	if ((ws = workQueues) != null) {
2269	>	int n = ws.length;
2270	>	for (int i = 0; i < n; i += 2) {
2271	>	if ((w = ws[i]) != null && w.queueSize() != 0)
2272	>	return true;
2273	>	}
2274	>	}
2275	>	return false;
2276		}
2277
2278		/**
#	Line 1671 | Line 2283 | public class ForkJoinPool extends Abstra
2283		* @return the next submission, or {@code null} if none
2284		*/
2285		protected ForkJoinTask<?> pollSubmission() {
2286	<	return submissionQueue.poll();
2286	>	WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2287	>	if ((ws = workQueues) != null) {
2288	>	int n = ws.length;
2289	>	for (int i = 0; i < n; i += 2) {
2290	>	if ((w = ws[i]) != null && (t = w.poll()) != null)
2291	>	return t;
2292	>	}
2293	>	}
2294	>	return null;
2295		}
2296
2297		/**
#	Line 1692 | Line 2312 | public class ForkJoinPool extends Abstra
2312		* @return the number of elements transferred
2313		*/
2314		protected int drainTasksTo(Collection<? super ForkJoinTask<?>> c) {
2315	<	int n = submissionQueue.drainTo(c);
2316	<	for (ForkJoinWorkerThread w : workers) {
2317	<	if (w != null)
2318	<	n += w.drainTasksTo(c);
2315	>	int count = 0;
2316	>	WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2317	>	if ((ws = workQueues) != null) {
2318	>	int n = ws.length;
2319	>	for (int i = 0; i < n; ++i) {
2320	>	if ((w = ws[i]) != null) {
2321	>	while ((t = w.poll()) != null) {
2322	>	c.add(t);
2323	>	++count;
2324	>	}
2325	>	}
2326	>	}
2327		}
2328	<	return n;
2328	>	return count;
2329		}
2330
2331		/**
#	Line 1711 | Line 2339 | public class ForkJoinPool extends Abstra
2339		long st = getStealCount();
2340		long qt = getQueuedTaskCount();
2341		long qs = getQueuedSubmissionCount();
2342	<	int wc = workerCounts;
1715	<	int tc = wc >>> TOTAL_COUNT_SHIFT;
1716	<	int rc = wc & RUNNING_COUNT_MASK;
2342	>	int rc = getRunningThreadCount();
2343		int pc = parallelism;
2344	<	int rs = runState;
2345	<	int ac = rs & ACTIVE_COUNT_MASK;
2344	>	long c = ctl;
2345	>	int tc = pc + (short)(c >>> TC_SHIFT);
2346	>	int ac = pc + (int)(c >> AC_SHIFT);
2347	>	if (ac < 0) // ignore transient negative
2348	>	ac = 0;
2349	>	String level;
2350	>	if ((c & STOP_BIT) != 0)
2351	>	level = (tc == 0) ? "Terminated" : "Terminating";
2352	>	else
2353	>	level = runState < 0 ? "Shutting down" : "Running";
2354		return super.toString() +
2355	<	"[" + runLevelToString(rs) +
2355	>	"[" + level +
2356		", parallelism = " + pc +
2357		", size = " + tc +
2358		", active = " + ac +
#	Line 1729 | Line 2363 | public class ForkJoinPool extends Abstra
2363		"]";
2364		}
2365
1732	–	private static String runLevelToString(int s) {
1733	–	return ((s & TERMINATED) != 0 ? "Terminated" :
1734	–	((s & TERMINATING) != 0 ? "Terminating" :
1735	–	((s & SHUTDOWN) != 0 ? "Shutting down" :
1736	–	"Running")));
1737	–	}
1738	–
2366		/**
2367		* Initiates an orderly shutdown in which previously submitted
2368		* tasks are executed, but no new tasks will be accepted.
#	Line 1750 | Line 2377 | public class ForkJoinPool extends Abstra
2377		*/
2378		public void shutdown() {
2379		checkPermission();
2380	<	advanceRunLevel(SHUTDOWN);
2380	>	enableShutdown();
2381		tryTerminate(false);
2382		}
2383
#	Line 1772 | Line 2399 | public class ForkJoinPool extends Abstra
2399		*/
2400		public List<Runnable> shutdownNow() {
2401		checkPermission();
2402	+	enableShutdown();
2403		tryTerminate(true);
2404		return Collections.emptyList();
2405		}
#	Line 1782 | Line 2410 | public class ForkJoinPool extends Abstra
2410		* @return {@code true} if all tasks have completed following shut down
2411		*/
2412		public boolean isTerminated() {
2413	<	return runState >= TERMINATED;
2413	>	long c = ctl;
2414	>	return ((c & STOP_BIT) != 0L &&
2415	>	(short)(c >>> TC_SHIFT) == -parallelism);
2416		}
2417
2418		/**
#	Line 1790 | Line 2420 | public class ForkJoinPool extends Abstra
2420		* commenced but not yet completed.  This method may be useful for
2421		* debugging. A return of {@code true} reported a sufficient
2422		* period after shutdown may indicate that submitted tasks have
2423	<	* ignored or suppressed interruption, causing this executor not
2424	<	* to properly terminate.
2423	>	* ignored or suppressed interruption, or are waiting for IO,
2424	>	* causing this executor not to properly terminate. (See the
2425	>	* advisory notes for class {@link ForkJoinTask} stating that
2426	>	* tasks should not normally entail blocking operations.  But if
2427	>	* they do, they must abort them on interrupt.)
2428		*
2429		* @return {@code true} if terminating but not yet terminated
2430		*/
2431		public boolean isTerminating() {
2432	<	return (runState & (TERMINATING|TERMINATED)) == TERMINATING;
2432	>	long c = ctl;
2433	>	return ((c & STOP_BIT) != 0L &&
2434	>	(short)(c >>> TC_SHIFT) != -parallelism);
2435		}
2436
2437		/**
#	Line 1805 | Line 2440 | public class ForkJoinPool extends Abstra
2440		* @return {@code true} if this pool has been shut down
2441		*/
2442		public boolean isShutdown() {
2443	<	return runState >= SHUTDOWN;
2443	>	return runState < 0;
2444		}
2445
2446		/**
#	Line 1821 | Line 2456 | public class ForkJoinPool extends Abstra
2456		*/
2457		public boolean awaitTermination(long timeout, TimeUnit unit)
2458		throws InterruptedException {
2459	<	return terminationLatch.await(timeout, unit);
2459	>	long nanos = unit.toNanos(timeout);
2460	>	final ReentrantLock lock = this.lock;
2461	>	lock.lock();
2462	>	try {
2463	>	for (;;) {
2464	>	if (isTerminated())
2465	>	return true;
2466	>	if (nanos <= 0)
2467	>	return false;
2468	>	nanos = termination.awaitNanos(nanos);
2469	>	}
2470	>	} finally {
2471	>	lock.unlock();
2472	>	}
2473		}
2474
2475		/**
2476		* Interface for extending managed parallelism for tasks running
2477		* in {@link ForkJoinPool}s.
2478		*
2479	<	* <p>A {@code ManagedBlocker} provides two methods.
2480	<	* Method {@code isReleasable} must return {@code true} if
2481	<	* blocking is not necessary. Method {@code block} blocks the
2482	<	* current thread if necessary (perhaps internally invoking
2483	<	* {@code isReleasable} before actually blocking).
2479	>	* <p>A {@code ManagedBlocker} provides two methods.  Method
2480	>	* {@code isReleasable} must return {@code true} if blocking is
2481	>	* not necessary. Method {@code block} blocks the current thread
2482	>	* if necessary (perhaps internally invoking {@code isReleasable}
2483	>	* before actually blocking). These actions are performed by any
2484	>	* thread invoking {@link ForkJoinPool#managedBlock}.  The
2485	>	* unusual methods in this API accommodate synchronizers that may,
2486	>	* but don't usually, block for long periods. Similarly, they
2487	>	* allow more efficient internal handling of cases in which
2488	>	* additional workers may be, but usually are not, needed to
2489	>	* ensure sufficient parallelism.  Toward this end,
2490	>	* implementations of method {@code isReleasable} must be amenable
2491	>	* to repeated invocation.
2492		*
2493		* <p>For example, here is a ManagedBlocker based on a
2494		* ReentrantLock:
#	Line 1850 | Line 2506 | public class ForkJoinPool extends Abstra
2506		*     return hasLock || (hasLock = lock.tryLock());
2507		*   }
2508		* }}</pre>
2509	+	*
2510	+	* <p>Here is a class that possibly blocks waiting for an
2511	+	* item on a given queue:
2512	+	*  <pre> {@code
2513	+	* class QueueTaker<E> implements ManagedBlocker {
2514	+	*   final BlockingQueue<E> queue;
2515	+	*   volatile E item = null;
2516	+	*   QueueTaker(BlockingQueue<E> q) { this.queue = q; }
2517	+	*   public boolean block() throws InterruptedException {
2518	+	*     if (item == null)
2519	+	*       item = queue.take();
2520	+	*     return true;
2521	+	*   }
2522	+	*   public boolean isReleasable() {
2523	+	*     return item != null || (item = queue.poll()) != null;
2524	+	*   }
2525	+	*   public E getItem() { // call after pool.managedBlock completes
2526	+	*     return item;
2527	+	*   }
2528	+	* }}</pre>
2529		*/
2530		public static interface ManagedBlocker {
2531		/**
#	Line 1873 | Line 2549 | public class ForkJoinPool extends Abstra
2549		* Blocks in accord with the given blocker.  If the current thread
2550		* is a {@link ForkJoinWorkerThread}, this method possibly
2551		* arranges for a spare thread to be activated if necessary to
2552	<	* ensure parallelism while the current thread is blocked.
1877	<	*
1878	<	* <p>If {@code maintainParallelism} is {@code true} and the pool
1879	<	* supports it ({@link #getMaintainsParallelism}), this method
1880	<	* attempts to maintain the pool's nominal parallelism. Otherwise
1881	<	* it activates a thread only if necessary to avoid complete
1882	<	* starvation. This option may be preferable when blockages use
1883	<	* timeouts, or are almost always brief.
2552	>	* ensure sufficient parallelism while the current thread is blocked.
2553		*
2554		* <p>If the caller is not a {@link ForkJoinTask}, this method is
2555		* behaviorally equivalent to
#	Line 1894 | Line 2563 | public class ForkJoinPool extends Abstra
2563		* first be expanded to ensure parallelism, and later adjusted.
2564		*
2565		* @param blocker the blocker
1897	–	* @param maintainParallelism if {@code true} and supported by
1898	–	* this pool, attempt to maintain the pool's nominal parallelism;
1899	–	* otherwise activate a thread only if necessary to avoid
1900	–	* complete starvation.
2566		* @throws InterruptedException if blocker.block did so
2567		*/
2568	<	public static void managedBlock(ManagedBlocker blocker,
1904	<	boolean maintainParallelism)
2568	>	public static void managedBlock(ManagedBlocker blocker)
2569		throws InterruptedException {
2570		Thread t = Thread.currentThread();
2571	<	if (t instanceof ForkJoinWorkerThread)
2572	<	((ForkJoinWorkerThread) t).pool.
2573	<	doBlock(blocker, maintainParallelism);
2574	<	else
2575	<	awaitBlocker(blocker);
2576	<	}
2577	<
2578	<	/**
2579	<	* Performs Non-FJ blocking
2580	<	*/
2581	<	private static void awaitBlocker(ManagedBlocker blocker)
2582	<	throws InterruptedException {
2583	<	do {} while (!blocker.isReleasable() && !blocker.block());
2571	>	ForkJoinPool p = ((t instanceof ForkJoinWorkerThread) ?
2572	>	((ForkJoinWorkerThread)t).pool : null);
2573	>	while (!blocker.isReleasable()) {
2574	>	if (p == null || p.tryCompensate()) {
2575	>	try {
2576	>	do {} while (!blocker.isReleasable() && !blocker.block());
2577	>	} finally {
2578	>	if (p != null)
2579	>	p.incrementActiveCount();
2580	>	}
2581	>	break;
2582	>	}
2583	>	}
2584		}
2585
2586		// AbstractExecutorService overrides.  These rely on undocumented
#	Line 1932 | Line 2596 | public class ForkJoinPool extends Abstra
2596		}
2597
2598		// Unsafe mechanics
2599	<
2600	<	private static final sun.misc.Unsafe UNSAFE = getUnsafe();
2601	<	private static final long workerCountsOffset =
2602	<	objectFieldOffset("workerCounts", ForkJoinPool.class);
2603	<	private static final long runStateOffset =
2604	<	objectFieldOffset("runState", ForkJoinPool.class);
2605	<	private static final long eventCountOffset =
2606	<	objectFieldOffset("eventCount", ForkJoinPool.class);
2607	<	private static final long eventWaitersOffset =
2608	<	objectFieldOffset("eventWaiters",ForkJoinPool.class);
2609	<	private static final long stealCountOffset =
1946	<	objectFieldOffset("stealCount",ForkJoinPool.class);
1947	<
1948	<
1949	<	private static long objectFieldOffset(String field, Class<?> klazz) {
2599	>	private static final sun.misc.Unsafe U;
2600	>	private static final long CTL;
2601	>	private static final long RUNSTATE;
2602	>	private static final long PARKBLOCKER;
2603	>
2604	>	static {
2605	>	poolNumberGenerator = new AtomicInteger();
2606	>	modifyThreadPermission = new RuntimePermission("modifyThread");
2607	>	defaultForkJoinWorkerThreadFactory =
2608	>	new DefaultForkJoinWorkerThreadFactory();
2609	>	int s;
2610		try {
2611	<	return UNSAFE.objectFieldOffset(klazz.getDeclaredField(field));
2612	<	} catch (NoSuchFieldException e) {
2613	<	// Convert Exception to corresponding Error
2614	<	NoSuchFieldError error = new NoSuchFieldError(field);
2615	<	error.initCause(e);
2616	<	throw error;
2611	>	U = getUnsafe();
2612	>	Class<?> k = ForkJoinPool.class;
2613	>	Class<?> tk = Thread.class;
2614	>	CTL = U.objectFieldOffset
2615	>	(k.getDeclaredField("ctl"));
2616	>	RUNSTATE = U.objectFieldOffset
2617	>	(k.getDeclaredField("runState"));
2618	>	PARKBLOCKER = U.objectFieldOffset
2619	>	(tk.getDeclaredField("parkBlocker"));
2620	>	} catch (Exception e) {
2621	>	throw new Error(e);
2622		}
2623		}
2624
#	Line 1984 | Line 2649 | public class ForkJoinPool extends Abstra
2649		}
2650		}
2651		}
2652	+
2653		}

Diff Legend

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