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 1.23 by dl, Sat Jul 25 15:50:57 2009 UTC vs.
Revision 1.111 by dl, Thu Jan 26 00:08:13 2012 UTC

#	Line 1 | Line 1
1		/*
2		* Written by Doug Lea with assistance from members of JCP JSR-166
3		* Expert Group and released to the public domain, as explained at
4	<	* http://creativecommons.org/licenses/publicdomain
4	>	* http://creativecommons.org/publicdomain/zero/1.0/
5		*/
6
7		package jsr166y;
8	–
9	–	import java.util.concurrent.*;
10	–
8		import java.util.ArrayList;
9		import java.util.Arrays;
10		import java.util.Collection;
11		import java.util.Collections;
12		import java.util.List;
13	<	import java.util.concurrent.locks.Condition;
14	<	import java.util.concurrent.locks.LockSupport;
15	<	import java.util.concurrent.locks.ReentrantLock;
13	>	import java.util.Random;
14	>	import java.util.concurrent.AbstractExecutorService;
15	>	import java.util.concurrent.Callable;
16	>	import java.util.concurrent.ExecutorService;
17	>	import java.util.concurrent.Future;
18	>	import java.util.concurrent.RejectedExecutionException;
19	>	import java.util.concurrent.RunnableFuture;
20	>	import java.util.concurrent.TimeUnit;
21		import java.util.concurrent.atomic.AtomicInteger;
22		import java.util.concurrent.atomic.AtomicLong;
23	+	import java.util.concurrent.locks.ReentrantLock;
24	+	import java.util.concurrent.locks.Condition;
25
26		/**
27	<	* An {@link ExecutorService} for running {@link ForkJoinTask}s.  A
28	<	* ForkJoinPool provides the entry point for submissions from
29	<	* non-ForkJoinTasks, as well as management and monitoring operations.
30	<	* Normally a single ForkJoinPool is used for a large number of
27	<	* submitted tasks. Otherwise, use would not usually outweigh the
28	<	* construction and bookkeeping overhead of creating a large set of
29	<	* threads.
27	>	* An {@link ExecutorService} for running {@link ForkJoinTask}s.
28	>	* A {@code ForkJoinPool} provides the entry point for submissions
29	>	* from non-{@code ForkJoinTask} clients, as well as management and
30	>	* monitoring operations.
31		*
32	<	* <p>ForkJoinPools differ from other kinds of Executors mainly in
33	<	* that they provide <em>work-stealing</em>: all threads in the pool
34	<	* attempt to find and execute subtasks created by other active tasks
35	<	* (eventually blocking if none exist). This makes them efficient when
36	<	* most tasks spawn other subtasks (as do most ForkJoinTasks), as well
37	<	* as the mixed execution of some plain Runnable- or Callable- based
38	<	* activities along with ForkJoinTasks. When setting
39	<	* {@code setAsyncMode}, a ForkJoinPools may also be appropriate for
40	<	* use with fine-grained tasks that are never joined. Otherwise, other
41	<	* ExecutorService implementations are typically more appropriate
42	<	* choices.
32	>	* <p>A {@code ForkJoinPool} differs from other kinds of {@link
33	>	* ExecutorService} mainly by virtue of employing
34	>	* <em>work-stealing</em>: all threads in the pool attempt to find and
35	>	* execute tasks submitted to the pool and/or created by other active
36	>	* tasks (eventually blocking waiting for work if none exist). This
37	>	* enables efficient processing when most tasks spawn other subtasks
38	>	* (as do most {@code ForkJoinTask}s), as well as when many small
39	>	* tasks are submitted to the pool from external clients.  Especially
40	>	* when setting <em>asyncMode</em> to true in constructors, {@code
41	>	* ForkJoinPool}s may also be appropriate for use with event-style
42	>	* tasks that are never joined.
43		*
44	<	* <p>A ForkJoinPool may be constructed with a given parallelism level
45	<	* (target pool size), which it attempts to maintain by dynamically
46	<	* adding, suspending, or resuming threads, even if some tasks are
47	<	* waiting to join others. However, no such adjustments are performed
48	<	* in the face of blocked IO or other unmanaged synchronization. The
49	<	* nested {@code ManagedBlocker} interface enables extension of
50	<	* the kinds of synchronization accommodated.  The target parallelism
51	<	* level may also be changed dynamically ({@code setParallelism})
52	<	* and thread construction can be limited using methods
52	<	* {@code setMaximumPoolSize} and/or
53	<	* {@code setMaintainsParallelism}.
44	>	* <p>A {@code ForkJoinPool} is constructed with a given target
45	>	* parallelism level; by default, equal to the number of available
46	>	* processors. The pool attempts to maintain enough active (or
47	>	* available) threads by dynamically adding, suspending, or resuming
48	>	* internal worker threads, even if some tasks are stalled waiting to
49	>	* join others. However, no such adjustments are guaranteed in the
50	>	* face of blocked IO or other unmanaged synchronization. The nested
51	>	* {@link ManagedBlocker} interface enables extension of the kinds of
52	>	* synchronization accommodated.
53		*
54		* <p>In addition to execution and lifecycle control methods, this
55		* class provides status check methods (for example
56	<	* {@code getStealCount}) that are intended to aid in developing,
56	>	* {@link #getStealCount}) that are intended to aid in developing,
57		* tuning, and monitoring fork/join applications. Also, method
58	<	* {@code toString} returns indications of pool state in a
58	>	* {@link #toString} returns indications of pool state in a
59		* convenient form for informal monitoring.
60		*
61	+	* <p> As is the case with other ExecutorServices, there are three
62	+	* main task execution methods summarized in the following
63	+	* table. These are designed to be used primarily by clients not
64	+	* already engaged in fork/join computations in the current pool.  The
65	+	* main forms of these methods accept instances of {@code
66	+	* ForkJoinTask}, but overloaded forms also allow mixed execution of
67	+	* plain {@code Runnable}- or {@code Callable}- based activities as
68	+	* well.  However, tasks that are already executing in a pool should
69	+	* normally instead use the within-computation forms listed in the
70	+	* table unless using async event-style tasks that are not usually
71	+	* joined, in which case there is little difference among choice of
72	+	* methods.
73	+	*
74	+	* <table BORDER CELLPADDING=3 CELLSPACING=1>
75	+	*  <tr>
76	+	*    <td></td>
77	+	*    <td ALIGN=CENTER> <b>Call from non-fork/join clients</b></td>
78	+	*    <td ALIGN=CENTER> <b>Call from within fork/join computations</b></td>
79	+	*  </tr>
80	+	*  <tr>
81	+	*    <td> <b>Arrange async execution</td>
82	+	*    <td> {@link #execute(ForkJoinTask)}</td>
83	+	*    <td> {@link ForkJoinTask#fork}</td>
84	+	*  </tr>
85	+	*  <tr>
86	+	*    <td> <b>Await and obtain result</td>
87	+	*    <td> {@link #invoke(ForkJoinTask)}</td>
88	+	*    <td> {@link ForkJoinTask#invoke}</td>
89	+	*  </tr>
90	+	*  <tr>
91	+	*    <td> <b>Arrange exec and obtain Future</td>
92	+	*    <td> {@link #submit(ForkJoinTask)}</td>
93	+	*    <td> {@link ForkJoinTask#fork} (ForkJoinTasks <em>are</em> Futures)</td>
94	+	*  </tr>
95	+	* </table>
96	+	*
97	+	* <p><b>Sample Usage.</b> Normally a single {@code ForkJoinPool} is
98	+	* used for all parallel task execution in a program or subsystem.
99	+	* Otherwise, use would not usually outweigh the construction and
100	+	* bookkeeping overhead of creating a large set of threads. For
101	+	* example, a common pool could be used for the {@code SortTasks}
102	+	* illustrated in {@link RecursiveAction}. Because {@code
103	+	* ForkJoinPool} uses threads in {@linkplain java.lang.Thread#isDaemon
104	+	* daemon} mode, there is typically no need to explicitly {@link
105	+	* #shutdown} such a pool upon program exit.
106	+	*
107	+	*  <pre> {@code
108	+	* static final ForkJoinPool mainPool = new ForkJoinPool();
109	+	* ...
110	+	* public void sort(long[] array) {
111	+	*   mainPool.invoke(new SortTask(array, 0, array.length));
112	+	* }}</pre>
113	+	*
114		* <p><b>Implementation notes</b>: This implementation restricts the
115		* maximum number of running threads to 32767. Attempts to create
116	<	* pools with greater than the maximum result in
117	<	* IllegalArgumentExceptions.
116	>	* pools with greater than the maximum number result in
117	>	* {@code IllegalArgumentException}.
118	>	*
119	>	* <p>This implementation rejects submitted tasks (that is, by throwing
120	>	* {@link RejectedExecutionException}) only when the pool is shut down
121	>	* or internal resources have been exhausted.
122		*
123		* @since 1.7
124		* @author Doug Lea
#	Line 70 | Line 126 | import java.util.concurrent.atomic.Atomi
126		public class ForkJoinPool extends AbstractExecutorService {
127
128		/*
129	<	* See the extended comments interspersed below for design,
130	<	* rationale, and walkthroughs.
131	<	*/
132	<
133	<	/** Mask for packing and unpacking shorts */
134	<	private static final int  shortMask = 0xffff;
135	<
136	<	/** Max pool size -- must be a power of two minus 1 */
137	<	private static final int MAX_THREADS =  0x7FFF;
138	<
139	<	/**
140	<	* Factory for creating new ForkJoinWorkerThreads.  A
141	<	* ForkJoinWorkerThreadFactory must be defined and used for
142	<	* ForkJoinWorkerThread subclasses that extend base functionality
143	<	* or initialize threads with different contexts.
129	>	* Implementation Overview
130	>	*
131	>	* This class and its nested classes provide the main
132	>	* functionality and control for a set of worker threads:
133	>	* Submissions from non-FJ threads enter into submission
134	>	* queues. Workers take these tasks and typically split them into
135	>	* subtasks that may be stolen by other workers.  Preference rules
136	>	* give first priority to processing tasks from their own queues
137	>	* (LIFO or FIFO, depending on mode), then to randomized FIFO
138	>	* steals of tasks in other queues.
139	>	*
140	>	* WorkQueues.
141	>	* ==========
142	>	*
143	>	* Most operations occur within work-stealing queues (in nested
144	>	* class WorkQueue).  These are special forms of Deques that
145	>	* support only three of the four possible end-operations -- push,
146	>	* pop, and poll (aka steal), under the further constraints that
147	>	* push and pop are called only from the owning thread (or, as
148	>	* extended here, under a lock), while poll may be called from
149	>	* other threads.  (If you are unfamiliar with them, you probably
150	>	* want to read Herlihy and Shavit's book "The Art of
151	>	* Multiprocessor programming", chapter 16 describing these in
152	>	* more detail before proceeding.)  The main work-stealing queue
153	>	* design is roughly similar to those in the papers "Dynamic
154	>	* Circular Work-Stealing Deque" by Chase and Lev, SPAA 2005
155	>	* (http://research.sun.com/scalable/pubs/index.html) and
156	>	* "Idempotent work stealing" by Michael, Saraswat, and Vechev,
157	>	* PPoPP 2009 (http://portal.acm.org/citation.cfm?id=1504186).
158	>	* The main differences ultimately stem from gc requirements that
159	>	* we null out taken slots as soon as we can, to maintain as small
160	>	* a footprint as possible even in programs generating huge
161	>	* numbers of tasks. To accomplish this, we shift the CAS
162	>	* arbitrating pop vs poll (steal) from being on the indices
163	>	* ("base" and "top") to the slots themselves.  So, both a
164	>	* successful pop and poll mainly entail a CAS of a slot from
165	>	* non-null to null.  Because we rely on CASes of references, we
166	>	* do not need tag bits on base or top.  They are simple ints as
167	>	* used in any circular array-based queue (see for example
168	>	* ArrayDeque).  Updates to the indices must still be ordered in a
169	>	* way that guarantees that top == base means the queue is empty,
170	>	* but otherwise may err on the side of possibly making the queue
171	>	* appear nonempty when a push, pop, or poll have not fully
172	>	* committed. Note that this means that the poll operation,
173	>	* considered individually, is not wait-free. One thief cannot
174	>	* successfully continue until another in-progress one (or, if
175	>	* previously empty, a push) completes.  However, in the
176	>	* aggregate, we ensure at least probabilistic non-blockingness.
177	>	* If an attempted steal fails, a thief always chooses a different
178	>	* random victim target to try next. So, in order for one thief to
179	>	* progress, it suffices for any in-progress poll or new push on
180	>	* any empty queue to complete.
181	>	*
182	>	* This approach also enables support a user mode in which local
183	>	* task processing is in FIFO, not LIFO order, simply by using
184	>	* poll rather than pop.  This can be useful in message-passing
185	>	* frameworks in which tasks are never joined.  However neither
186	>	* mode considers affinities, loads, cache localities, etc, so
187	>	* rarely provide the best possible performance on a given
188	>	* machine, but portably provide good throughput by averaging over
189	>	* these factors.  (Further, even if we did try to use such
190	>	* information, we do not usually have a basis for exploiting
191	>	* it. For example, some sets of tasks profit from cache
192	>	* affinities, but others are harmed by cache pollution effects.)
193	>	*
194	>	* WorkQueues are also used in a similar way for tasks submitted
195	>	* to the pool. We cannot mix these tasks in the same queues used
196	>	* for work-stealing (this would contaminate lifo/fifo
197	>	* processing). Instead, we loosely associate (via hashing)
198	>	* submission queues with submitting threads, and randomly scan
199	>	* these queues as well when looking for work. In essence,
200	>	* submitters act like workers except that they never take tasks,
201	>	* and they are multiplexed on to a finite number of shared work
202	>	* queues. However, classes are set up so that future extensions
203	>	* could allow submitters to optionally help perform tasks as
204	>	* well. Pool submissions from internal workers are also allowed,
205	>	* but use randomized rather than thread-hashed queue indices to
206	>	* avoid imbalance.  Insertion of tasks in shared mode requires a
207	>	* lock (mainly to protect in the case of resizing) but we use
208	>	* only a simple spinlock (using bits in field runState), because
209	>	* submitters encountering a busy queue try others so never block.
210	>	*
211	>	* Management.
212	>	* ==========
213	>	*
214	>	* The main throughput advantages of work-stealing stem from
215	>	* decentralized control -- workers mostly take tasks from
216	>	* themselves or each other. We cannot negate this in the
217	>	* implementation of other management responsibilities. The main
218	>	* tactic for avoiding bottlenecks is packing nearly all
219	>	* essentially atomic control state into two volatile variables
220	>	* that are by far most often read (not written) as status and
221	>	* consistency checks
222	>	*
223	>	* Field "ctl" contains 64 bits holding all the information needed
224	>	* to atomically decide to add, inactivate, enqueue (on an event
225	>	* queue), dequeue, and/or re-activate workers.  To enable this
226	>	* packing, we restrict maximum parallelism to (1<<15)-1 (which is
227	>	* far in excess of normal operating range) to allow ids, counts,
228	>	* and their negations (used for thresholding) to fit into 16bit
229	>	* fields.
230	>	*
231	>	* Field "runState" contains 32 bits needed to register and
232	>	* deregister WorkQueues, as well as to enable shutdown. It is
233	>	* only modified under a lock (normally briefly held, but
234	>	* occasionally protecting allocations and resizings) but even
235	>	* when locked remains available to check consistency.
236	>	*
237	>	* Recording WorkQueues.  WorkQueues are recorded in the
238	>	* "workQueues" array that is created upon pool construction and
239	>	* expanded if necessary.  Updates to the array while recording
240	>	* new workers and unrecording terminated ones are protected from
241	>	* each other by a lock but the array is otherwise concurrently
242	>	* readable, and accessed directly.  To simplify index-based
243	>	* operations, the array size is always a power of two, and all
244	>	* readers must tolerate null slots. Shared (submission) queues
245	>	* are at even indices, worker queues at odd indices. Grouping
246	>	* them together in this way simplifies and speeds up task
247	>	* scanning. To avoid flailing during start-up, the array is
248	>	* presized to hold twice #parallelism workers (which is unlikely
249	>	* to need further resizing during execution). But to avoid
250	>	* dealing with so many null slots, variable runState includes a
251	>	* mask for the nearest power of two that contains all current
252	>	* workers.  All worker thread creation is on-demand, triggered by
253	>	* task submissions, replacement of terminated workers, and/or
254	>	* compensation for blocked workers. However, all other support
255	>	* code is set up to work with other policies.  To ensure that we
256	>	* do not hold on to worker references that would prevent GC, ALL
257	>	* accesses to workQueues are via indices into the workQueues
258	>	* array (which is one source of some of the messy code
259	>	* constructions here). In essence, the workQueues array serves as
260	>	* a weak reference mechanism. Thus for example the wait queue
261	>	* field of ctl stores indices, not references.  Access to the
262	>	* workQueues in associated methods (for example signalWork) must
263	>	* both index-check and null-check the IDs. All such accesses
264	>	* ignore bad IDs by returning out early from what they are doing,
265	>	* since this can only be associated with termination, in which
266	>	* case it is OK to give up.
267	>	*
268	>	* All uses of the workQueues array check that it is non-null
269	>	* (even if previously non-null). This allows nulling during
270	>	* termination, which is currently not necessary, but remains an
271	>	* option for resource-revocation-based shutdown schemes. It also
272	>	* helps reduce JIT issuance of uncommon-trap code, which tends to
273	>	* unnecessarily complicate control flow in some methods.
274	>	*
275	>	* Event Queuing. Unlike HPC work-stealing frameworks, we cannot
276	>	* let workers spin indefinitely scanning for tasks when none can
277	>	* be found immediately, and we cannot start/resume workers unless
278	>	* there appear to be tasks available.  On the other hand, we must
279	>	* quickly prod them into action when new tasks are submitted or
280	>	* generated. In many usages, ramp-up time to activate workers is
281	>	* the main limiting factor in overall performance (this is
282	>	* compounded at program start-up by JIT compilation and
283	>	* allocation). So we try to streamline this as much as possible.
284	>	* We park/unpark workers after placing in an event wait queue
285	>	* when they cannot find work. This "queue" is actually a simple
286	>	* Treiber stack, headed by the "id" field of ctl, plus a 15bit
287	>	* counter value (that reflects the number of times a worker has
288	>	* been inactivated) to avoid ABA effects (we need only as many
289	>	* version numbers as worker threads). Successors are held in
290	>	* field WorkQueue.nextWait.  Queuing deals with several intrinsic
291	>	* races, mainly that a task-producing thread can miss seeing (and
292	>	* signalling) another thread that gave up looking for work but
293	>	* has not yet entered the wait queue. We solve this by requiring
294	>	* a full sweep of all workers (via repeated calls to method
295	>	* scan()) both before and after a newly waiting worker is added
296	>	* to the wait queue. During a rescan, the worker might release
297	>	* some other queued worker rather than itself, which has the same
298	>	* net effect. Because enqueued workers may actually be rescanning
299	>	* rather than waiting, we set and clear the "parker" field of
300	>	* Workqueues to reduce unnecessary calls to unpark.  (this
301	>	* requires a secondary recheck to avoid missed signals.)  Note
302	>	* the unusual conventions about Thread.interrupts surrounding
303	>	* parking and other blocking: Because interrupts are used solely
304	>	* to alert threads to check termination, which is checked anyway
305	>	* upon blocking, we clear status (using Thread.interrupted)
306	>	* before any call to park, so that park does not immediately
307	>	* return due to status being set via some other unrelated call to
308	>	* interrupt in user code.
309	>	*
310	>	* Signalling.  We create or wake up workers only when there
311	>	* appears to be at least one task they might be able to find and
312	>	* execute.  When a submission is added or another worker adds a
313	>	* task to a queue that previously had fewer than two tasks, they
314	>	* signal waiting workers (or trigger creation of new ones if
315	>	* fewer than the given parallelism level -- see signalWork).
316	>	* These primary signals are buttressed by signals during rescans;
317	>	* together these cover the signals needed in cases when more
318	>	* tasks are pushed but untaken, and improve performance compared
319	>	* to having one thread wake up all workers.
320	>	*
321	>	* Trimming workers. To release resources after periods of lack of
322	>	* use, a worker starting to wait when the pool is quiescent will
323	>	* time out and terminate if the pool has remained quiescent for
324	>	* SHRINK_RATE nanosecs. This will slowly propagate, eventually
325	>	* terminating all workers after long periods of non-use.
326	>	*
327	>	* Shutdown and Termination. A call to shutdownNow atomically sets
328	>	* a runState bit and then (non-atomically) sets each workers
329	>	* runState status, cancels all unprocessed tasks, and wakes up
330	>	* all waiting workers.  Detecting whether termination should
331	>	* commence after a non-abrupt shutdown() call requires more work
332	>	* and bookkeeping. We need consensus about quiescence (i.e., that
333	>	* there is no more work). The active count provides a primary
334	>	* indication but non-abrupt shutdown still requires a rechecking
335	>	* scan for any workers that are inactive but not queued.
336	>	*
337	>	* Joining Tasks.
338	>	* ==============
339	>	*
340	>	* Any of several actions may be taken when one worker is waiting
341	>	* to join a task stolen (or always held by) another.  Because we
342	>	* are multiplexing many tasks on to a pool of workers, we can't
343	>	* just let them block (as in Thread.join).  We also cannot just
344	>	* reassign the joiner's run-time stack with another and replace
345	>	* it later, which would be a form of "continuation", that even if
346	>	* possible is not necessarily a good idea since we sometimes need
347	>	* both an unblocked task and its continuation to
348	>	* progress. Instead we combine two tactics:
349	>	*
350	>	*   Helping: Arranging for the joiner to execute some task that it
351	>	*      would be running if the steal had not occurred.
352	>	*
353	>	*   Compensating: Unless there are already enough live threads,
354	>	*      method tryCompensate() may create or re-activate a spare
355	>	*      thread to compensate for blocked joiners until they unblock.
356	>	*
357	>	* A third form (implemented in tryRemoveAndExec and
358	>	* tryPollForAndExec) amounts to helping a hypothetical
359	>	* compensator: If we can readily tell that a possible action of a
360	>	* compensator is to steal and execute the task being joined, the
361	>	* joining thread can do so directly, without the need for a
362	>	* compensation thread (although at the expense of larger run-time
363	>	* stacks, but the tradeoff is typically worthwhile).
364	>	*
365	>	* The ManagedBlocker extension API can't use helping so relies
366	>	* only on compensation in method awaitBlocker.
367	>	*
368	>	* The algorithm in tryHelpStealer entails a form of "linear"
369	>	* helping: Each worker records (in field currentSteal) the most
370	>	* recent task it stole from some other worker. Plus, it records
371	>	* (in field currentJoin) the task it is currently actively
372	>	* joining. Method tryHelpStealer uses these markers to try to
373	>	* find a worker to help (i.e., steal back a task from and execute
374	>	* it) that could hasten completion of the actively joined task.
375	>	* In essence, the joiner executes a task that would be on its own
376	>	* local deque had the to-be-joined task not been stolen. This may
377	>	* be seen as a conservative variant of the approach in Wagner &
378	>	* Calder "Leapfrogging: a portable technique for implementing
379	>	* efficient futures" SIGPLAN Notices, 1993
380	>	* (http://portal.acm.org/citation.cfm?id=155354). It differs in
381	>	* that: (1) We only maintain dependency links across workers upon
382	>	* steals, rather than use per-task bookkeeping.  This sometimes
383	>	* requires a linear scan of workers array to locate stealers, but
384	>	* often doesn't because stealers leave hints (that may become
385	>	* stale/wrong) of where to locate them.  A stealHint is only a
386	>	* hint because a worker might have had multiple steals and the
387	>	* hint records only one of them (usually the most current).
388	>	* Hinting isolates cost to when it is needed, rather than adding
389	>	* to per-task overhead.  (2) It is "shallow", ignoring nesting
390	>	* and potentially cyclic mutual steals.  (3) It is intentionally
391	>	* racy: field currentJoin is updated only while actively joining,
392	>	* which means that we miss links in the chain during long-lived
393	>	* tasks, GC stalls etc (which is OK since blocking in such cases
394	>	* is usually a good idea).  (4) We bound the number of attempts
395	>	* to find work (see MAX_HELP_DEPTH) and fall back to suspending
396	>	* the worker and if necessary replacing it with another.
397	>	*
398	>	* It is impossible to keep exactly the target parallelism number
399	>	* of threads running at any given time.  Determining the
400	>	* existence of conservatively safe helping targets, the
401	>	* availability of already-created spares, and the apparent need
402	>	* to create new spares are all racy, so we rely on multiple
403	>	* retries of each.  Currently, in keeping with on-demand
404	>	* signalling policy, we compensate only if blocking would leave
405	>	* less than one active (non-waiting, non-blocked) worker.
406	>	* Additionally, to avoid some false alarms due to GC, lagging
407	>	* counters, system activity, etc, compensated blocking for joins
408	>	* is only attempted after rechecks stabilize in
409	>	* ForkJoinTask.awaitJoin. (Retries are interspersed with
410	>	* Thread.yield, for good citizenship.)
411	>	*
412	>	* Style notes: There is a lot of representation-level coupling
413	>	* among classes ForkJoinPool, ForkJoinWorkerThread, and
414	>	* ForkJoinTask.  The fields of WorkQueue maintain data structures
415	>	* managed by ForkJoinPool, so are directly accessed.  There is
416	>	* little point trying to reduce this, since any associated future
417	>	* changes in representations will need to be accompanied by
418	>	* algorithmic changes anyway. All together, these low-level
419	>	* implementation choices produce as much as a factor of 4
420	>	* performance improvement compared to naive implementations, and
421	>	* enable the processing of billions of tasks per second, at the
422	>	* expense of some ugliness.
423	>	*
424	>	* Methods signalWork() and scan() are the main bottlenecks so are
425	>	* especially heavily micro-optimized/mangled.  There are lots of
426	>	* inline assignments (of form "while ((local = field) != 0)")
427	>	* which are usually the simplest way to ensure the required read
428	>	* orderings (which are sometimes critical). This leads to a
429	>	* "C"-like style of listing declarations of these locals at the
430	>	* heads of methods or blocks.  There are several occurrences of
431	>	* the unusual "do {} while (!cas...)"  which is the simplest way
432	>	* to force an update of a CAS'ed variable. There are also other
433	>	* coding oddities that help some methods perform reasonably even
434	>	* when interpreted (not compiled).
435	>	*
436	>	* The order of declarations in this file is: (1) declarations of
437	>	* statics (2) fields (along with constants used when unpacking
438	>	* some of them), listed in an order that tends to reduce
439	>	* contention among them a bit under most JVMs; (3) nested
440	>	* classes; (4) internal control methods; (5) callbacks and other
441	>	* support for ForkJoinTask methods; (6) exported methods (plus a
442	>	* few little helpers); (7) static block initializing all statics
443	>	* in a minimally dependent order.
444	>	*/
445	>
446	>	/**
447	>	* Factory for creating new {@link ForkJoinWorkerThread}s.
448	>	* A {@code ForkJoinWorkerThreadFactory} must be defined and used
449	>	* for {@code ForkJoinWorkerThread} subclasses that extend base
450	>	* functionality or initialize threads with different contexts.
451		*/
452		public static interface ForkJoinWorkerThreadFactory {
453		/**
454		* Returns a new worker thread operating in the given pool.
455		*
456		* @param pool the pool this thread works in
457	<	* @throws NullPointerException if pool is null
457	>	* @throws NullPointerException if the pool is null
458		*/
459		public ForkJoinWorkerThread newThread(ForkJoinPool pool);
460		}
#	Line 100 | Line 463 | public class ForkJoinPool extends Abstra
463		* Default ForkJoinWorkerThreadFactory implementation; creates a
464		* new ForkJoinWorkerThread.
465		*/
466	<	static class  DefaultForkJoinWorkerThreadFactory
466	>	static class DefaultForkJoinWorkerThreadFactory
467		implements ForkJoinWorkerThreadFactory {
468		public ForkJoinWorkerThread newThread(ForkJoinPool pool) {
469	<	try {
107	<	return new ForkJoinWorkerThread(pool);
108	<	} catch (OutOfMemoryError oom)  {
109	<	return null;
110	<	}
469	>	return new ForkJoinWorkerThread(pool);
470		}
471		}
472
#	Line 116 | Line 475 | public class ForkJoinPool extends Abstra
475		* overridden in ForkJoinPool constructors.
476		*/
477		public static final ForkJoinWorkerThreadFactory
478	<	defaultForkJoinWorkerThreadFactory =
120	<	new DefaultForkJoinWorkerThreadFactory();
478	>	defaultForkJoinWorkerThreadFactory;
479
480		/**
481		* Permission required for callers of methods that may start or
482		* kill threads.
483		*/
484	<	private static final RuntimePermission modifyThreadPermission =
127	<	new RuntimePermission("modifyThread");
484	>	private static final RuntimePermission modifyThreadPermission;
485
486		/**
487		* If there is a security manager, makes sure caller has
#	Line 139 | Line 496 | public class ForkJoinPool extends Abstra
496		/**
497		* Generator for assigning sequence numbers as pool names.
498		*/
499	<	private static final AtomicInteger poolNumberGenerator =
143	<	new AtomicInteger();
499	>	private static final AtomicInteger poolNumberGenerator;
500
501		/**
502	<	* Array holding all worker threads in the pool. Initialized upon
503	<	* first use. Array size must be a power of two.  Updates and
504	<	* replacements are protected by workerLock, but it is always kept
505	<	* in a consistent enough state to be randomly accessed without
506	<	* locking by workers performing work-stealing.
502	>	* Bits and masks for control variables
503	>	*
504	>	* Field ctl is a long packed with:
505	>	* AC: Number of active running workers minus target parallelism (16 bits)
506	>	* TC: Number of total workers minus target parallelism (16 bits)
507	>	* ST: true if pool is terminating (1 bit)
508	>	* EC: the wait count of top waiting thread (15 bits)
509	>	* ID: ~(poolIndex >>> 1) of top of Treiber stack of waiters (16 bits)
510	>	*
511	>	* When convenient, we can extract the upper 32 bits of counts and
512	>	* the lower 32 bits of queue state, u = (int)(ctl >>> 32) and e =
513	>	* (int)ctl.  The ec field is never accessed alone, but always
514	>	* together with id and st. The offsets of counts by the target
515	>	* parallelism and the positionings of fields makes it possible to
516	>	* perform the most common checks via sign tests of fields: When
517	>	* ac is negative, there are not enough active workers, when tc is
518	>	* negative, there are not enough total workers, when id is
519	>	* negative, there is at least one waiting worker, and when e is
520	>	* negative, the pool is terminating.  To deal with these possibly
521	>	* negative fields, we use casts in and out of "short" and/or
522	>	* signed shifts to maintain signedness.
523	>	*
524	>	* When a thread is queued (inactivated), its eventCount field is
525	>	* negative, which is the only way to tell if a worker is
526	>	* prevented from executing tasks, even though it must continue to
527	>	* scan for them to avoid queuing races.
528	>	*
529	>	* Field runState is an int packed with:
530	>	* SHUTDOWN: true if shutdown is enabled (1 bit)
531	>	* SEQ:  a sequence number updated upon (de)registering workers (15 bits)
532	>	* MASK: mask (power of 2 - 1) covering all registered poolIndexes (16 bits)
533	>	*
534	>	* The combination of mask and sequence number enables simple
535	>	* consistency checks: Staleness of read-only operations on the
536	>	* workers and queues arrays can be checked by comparing runState
537	>	* before vs after the reads. The low 16 bits (i.e, anding with
538	>	* SMASK) hold (the smallest power of two covering all worker
539	>	* indices, minus one.  The mask for queues (vs workers) is twice
540	>	* this value plus 1.
541	>	*/
542	>
543	>	// bit positions/shifts for fields
544	>	private static final int  AC_SHIFT   = 48;
545	>	private static final int  TC_SHIFT   = 32;
546	>	private static final int  ST_SHIFT   = 31;
547	>	private static final int  EC_SHIFT   = 16;
548	>
549	>	// bounds
550	>	private static final int  MAX_ID     = 0x7fff;  // max poolIndex
551	>	private static final int  SMASK      = 0xffff;  // mask short bits
552	>	private static final int  SHORT_SIGN = 1 << 15;
553	>	private static final int  INT_SIGN   = 1 << 31;
554	>
555	>	// masks
556	>	private static final long STOP_BIT   = 0x0001L << ST_SHIFT;
557	>	private static final long AC_MASK    = ((long)SMASK) << AC_SHIFT;
558	>	private static final long TC_MASK    = ((long)SMASK) << TC_SHIFT;
559	>
560	>	// units for incrementing and decrementing
561	>	private static final long TC_UNIT    = 1L << TC_SHIFT;
562	>	private static final long AC_UNIT    = 1L << AC_SHIFT;
563	>
564	>	// masks and units for dealing with u = (int)(ctl >>> 32)
565	>	private static final int  UAC_SHIFT  = AC_SHIFT - 32;
566	>	private static final int  UTC_SHIFT  = TC_SHIFT - 32;
567	>	private static final int  UAC_MASK   = SMASK << UAC_SHIFT;
568	>	private static final int  UTC_MASK   = SMASK << UTC_SHIFT;
569	>	private static final int  UAC_UNIT   = 1 << UAC_SHIFT;
570	>	private static final int  UTC_UNIT   = 1 << UTC_SHIFT;
571	>
572	>	// masks and units for dealing with e = (int)ctl
573	>	private static final int E_MASK      = 0x7fffffff; // no STOP_BIT
574	>	private static final int E_SEQ       = 1 << EC_SHIFT;
575	>
576	>	// runState bits
577	>	private static final int SHUTDOWN    = 1 << 31;
578	>	private static final int RS_SEQ      = 1 << 16;
579	>	private static final int RS_SEQ_MASK = 0x7fff0000;
580	>
581	>	// access mode for WorkQueue
582	>	static final int LIFO_QUEUE          =  0;
583	>	static final int FIFO_QUEUE          =  1;
584	>	static final int SHARED_QUEUE        = -1;
585	>
586	>	/**
587	>	* The wakeup interval (in nanoseconds) for a worker waiting for a
588	>	* task when the pool is quiescent to instead try to shrink the
589	>	* number of workers.  The exact value does not matter too
590	>	* much. It must be short enough to release resources during
591	>	* sustained periods of idleness, but not so short that threads
592	>	* are continually re-created.
593	>	*/
594	>	private static final long SHRINK_RATE =
595	>	4L * 1000L * 1000L * 1000L; // 4 seconds
596	>
597	>	/**
598	>	* The timeout value for attempted shrinkage, includes
599	>	* some slop to cope with system timer imprecision.
600	>	*/
601	>	private static final long SHRINK_TIMEOUT = SHRINK_RATE - (SHRINK_RATE / 10);
602	>
603	>	/**
604	>	* The maximum stolen->joining link depth allowed in tryHelpStealer.
605	>	* Depths for legitimate chains are unbounded, but we use a fixed
606	>	* constant to avoid (otherwise unchecked) cycles and to bound
607	>	* staleness of traversal parameters at the expense of sometimes
608	>	* blocking when we could be helping.
609		*/
610	<	volatile ForkJoinWorkerThread[] workers;
610	>	private static final int MAX_HELP_DEPTH = 16;
611
612	<	/**
613	<	* Lock protecting access to workers.
612	>	/*
613	>	* Field layout order in this class tends to matter more than one
614	>	* would like. Runtime layout order is only loosely related to
615	>	* declaration order and may differ across JVMs, but the following
616	>	* empirically works OK on current JVMs.
617	>	*/
618	>
619	>	volatile long ctl;                       // main pool control
620	>	final int parallelism;                   // parallelism level
621	>	final int localMode;                     // per-worker scheduling mode
622	>	int nextPoolIndex;                       // hint used in registerWorker
623	>	volatile int runState;                   // shutdown status, seq, and mask
624	>	WorkQueue[] workQueues;                  // main registry
625	>	final ReentrantLock lock;                // for registration
626	>	final Condition termination;             // for awaitTermination
627	>	final ForkJoinWorkerThreadFactory factory; // factory for new workers
628	>	final Thread.UncaughtExceptionHandler ueh; // per-worker UEH
629	>	final AtomicLong stealCount;             // collect counts when terminated
630	>	final AtomicInteger nextWorkerNumber;    // to create worker name string
631	>	final String workerNamePrefix;           // Prefix for assigning worker names
632	>
633	>	/**
634	>	* Queues supporting work-stealing as well as external task
635	>	* submission. See above for main rationale and algorithms.
636	>	* Implementation relies heavily on "Unsafe" intrinsics
637	>	* and selective use of "volatile":
638	>	*
639	>	* Field "base" is the index (mod array.length) of the least valid
640	>	* queue slot, which is always the next position to steal (poll)
641	>	* from if nonempty. Reads and writes require volatile orderings
642	>	* but not CAS, because updates are only performed after slot
643	>	* CASes.
644	>	*
645	>	* Field "top" is the index (mod array.length) of the next queue
646	>	* slot to push to or pop from. It is written only by owner thread
647	>	* for push, or under lock for trySharedPush, and accessed by
648	>	* other threads only after reading (volatile) base.  Both top and
649	>	* base are allowed to wrap around on overflow, but (top - base)
650	>	* (or more comonly -(base - top) to force volatile read of base
651	>	* before top) still estimates size.
652	>	*
653	>	* The array slots are read and written using the emulation of
654	>	* volatiles/atomics provided by Unsafe. Insertions must in
655	>	* general use putOrderedObject as a form of releasing store to
656	>	* ensure that all writes to the task object are ordered before
657	>	* its publication in the queue. (Although we can avoid one case
658	>	* of this when locked in trySharedPush.) All removals entail a
659	>	* CAS to null.  The array is always a power of two. To ensure
660	>	* safety of Unsafe array operations, all accesses perform
661	>	* explicit null checks and implicit bounds checks via
662	>	* power-of-two masking.
663	>	*
664	>	* In addition to basic queuing support, this class contains
665	>	* fields described elsewhere to control execution. It turns out
666	>	* to work better memory-layout-wise to include them in this
667	>	* class rather than a separate class.
668	>	*
669	>	* Performance on most platforms is very sensitive to placement of
670	>	* instances of both WorkQueues and their arrays -- we absolutely
671	>	* do not want multiple WorkQueue instances or multiple queue
672	>	* arrays sharing cache lines. (It would be best for queue objects
673	>	* and their arrays to share, but there is nothing available to
674	>	* help arrange that).  Unfortunately, because they are recorded
675	>	* in a common array, WorkQueue instances are often moved to be
676	>	* adjacent by garbage collectors. To reduce impact, we use field
677	>	* padding that works OK on common platforms; this effectively
678	>	* trades off slightly slower average field access for the sake of
679	>	* avoiding really bad worst-case access. (Until better JVM
680	>	* support is in place, this padding is dependent on transient
681	>	* properties of JVM field layout rules.)  We also take care in
682	>	* allocating and sizing and resizing the array. Non-shared queue
683	>	* arrays are initialized (via method growArray) by workers before
684	>	* use. Others are allocated on first use.
685		*/
686	<	private final ReentrantLock workerLock;
686	>	static final class WorkQueue {
687	>	/**
688	>	* Capacity of work-stealing queue array upon initialization.
689	>	* Must be a power of two; at least 4, but set larger to
690	>	* reduce cacheline sharing among queues.
691	>	*/
692	>	static final int INITIAL_QUEUE_CAPACITY = 1 << 8;
693
694	<	/**
695	<	* Condition for awaitTermination.
696	<	*/
697	<	private final Condition termination;
694	>	/**
695	>	* Maximum size for queue arrays. Must be a power of two less
696	>	* than or equal to 1 << (31 - width of array entry) to ensure
697	>	* lack of wraparound of index calculations, but defined to a
698	>	* value a bit less than this to help users trap runaway
699	>	* programs before saturating systems.
700	>	*/
701	>	static final int MAXIMUM_QUEUE_CAPACITY = 1 << 26; // 64M
702	>
703	>	volatile long totalSteals; // cumulative number of steals
704	>	int seed;                  // for random scanning; initialize nonzero
705	>	volatile int eventCount;   // encoded inactivation count; < 0 if inactive
706	>	int nextWait;              // encoded record of next event waiter
707	>	int rescans;               // remaining scans until block
708	>	int nsteals;               // top-level task executions since last idle
709	>	final int mode;            // lifo, fifo, or shared
710	>	int poolIndex;             // index of this queue in pool (or 0)
711	>	int stealHint;             // index of most recent known stealer
712	>	volatile int runState;     // 1: locked, -1: terminate; else 0
713	>	volatile int base;         // index of next slot for poll
714	>	int top;                   // index of next slot for push
715	>	ForkJoinTask<?>[] array;   // the elements (initially unallocated)
716	>	final ForkJoinWorkerThread owner; // owning thread or null if shared
717	>	volatile Thread parker;    // == owner during call to park; else null
718	>	ForkJoinTask<?> currentJoin;  // task being joined in awaitJoin
719	>	ForkJoinTask<?> currentSteal; // current non-local task being executed
720	>	// Heuristic padding to ameliorate unfortunate memory placements
721	>	Object p00, p01, p02, p03, p04, p05, p06, p07, p08, p09, p0a;
722	>
723	>	WorkQueue(ForkJoinWorkerThread owner, int mode) {
724	>	this.owner = owner;
725	>	this.mode = mode;
726	>	// Place indices in the center of array (that is not yet allocated)
727	>	base = top = INITIAL_QUEUE_CAPACITY >>> 1;
728	>	}
729	>
730	>	/**
731	>	* Returns number of tasks in the queue
732	>	*/
733	>	final int queueSize() {
734	>	int n = base - top; // non-owner callers must read base first
735	>	return (n >= 0) ? 0 : -n;
736	>	}
737	>
738	>	/**
739	>	* Pushes a task. Call only by owner in unshared queues.
740	>	*
741	>	* @param task the task. Caller must ensure non-null.
742	>	* @param p, if non-null, pool to signal if necessary
743	>	* @throw RejectedExecutionException if array cannot
744	>	* be resized
745	>	*/
746	>	final void push(ForkJoinTask<?> task, ForkJoinPool p) {
747	>	boolean signal = false;
748	>	ForkJoinTask<?>[] a;
749	>	int s = top, m, n;
750	>	if ((a = array) != null) {    // ignore if queue removed
751	>	U.putOrderedObject
752	>	(a, (((m = a.length - 1) & s) << ASHIFT) + ABASE, task);
753	>	if ((n = (top = s + 1) - base) <= 2) {
754	>	if (p != null)
755	>	p.signalWork();
756	>	}
757	>	else if (n >= m)
758	>	growArray(true);
759	>	}
760	>	}
761	>
762	>	/**
763	>	* Pushes a task if lock is free and array is either big
764	>	* enough or can be resized to be big enough.
765	>	*
766	>	* @param task the task. Caller must ensure non-null.
767	>	* @return true if submitted
768	>	*/
769	>	final boolean trySharedPush(ForkJoinTask<?> task) {
770	>	boolean submitted = false;
771	>	if (runState == 0 && U.compareAndSwapInt(this, RUNSTATE, 0, 1)) {
772	>	ForkJoinTask<?>[] a = array;
773	>	int s = top, n = s - base;
774	>	try {
775	>	if ((a != null && n < a.length - 1) ||
776	>	(a = growArray(false)) != null) { // must presize
777	>	int j = (((a.length - 1) & s) << ASHIFT) + ABASE;
778	>	U.putObject(a, (long)j, task);    // don't need "ordered"
779	>	top = s + 1;
780	>	submitted = true;
781	>	}
782	>	} finally {
783	>	runState = 0;                         // unlock
784	>	}
785	>	}
786	>	return submitted;
787	>	}
788	>
789	>	/**
790	>	* Takes next task, if one exists, in FIFO order.
791	>	*/
792	>	final ForkJoinTask<?> poll() {
793	>	ForkJoinTask<?>[] a; int b, i;
794	>	while ((b = base) - top < 0 && (a = array) != null &&
795	>	(i = (a.length - 1) & b) >= 0) {
796	>	int j = (i << ASHIFT) + ABASE;
797	>	ForkJoinTask<?> t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
798	>	if (t != null && base == b &&
799	>	U.compareAndSwapObject(a, j, t, null)) {
800	>	base = b + 1;
801	>	return t;
802	>	}
803	>	}
804	>	return null;
805	>	}
806	>
807	>	/**
808	>	* Takes next task, if one exists, in LIFO order.
809	>	* Call only by owner in unshared queues.
810	>	*/
811	>	final ForkJoinTask<?> pop() {
812	>	ForkJoinTask<?> t; int m;
813	>	ForkJoinTask<?>[] a = array;
814	>	if (a != null && (m = a.length - 1) >= 0) {
815	>	for (int s; (s = top - 1) - base >= 0;) {
816	>	int j = ((m & s) << ASHIFT) + ABASE;
817	>	if ((t = (ForkJoinTask<?>)U.getObjectVolatile(a, j)) == null)
818	>	break;
819	>	if (U.compareAndSwapObject(a, j, t, null)) {
820	>	top = s;
821	>	return t;
822	>	}
823	>	}
824	>	}
825	>	return null;
826	>	}
827	>
828	>	/**
829	>	* Takes next task, if one exists, in order specified by mode.
830	>	*/
831	>	final ForkJoinTask<?> nextLocalTask() {
832	>	return mode == 0 ? pop() : poll();
833	>	}
834	>
835	>	/**
836	>	* Returns next task, if one exists, in order specified by mode.
837	>	*/
838	>	final ForkJoinTask<?> peek() {
839	>	ForkJoinTask<?>[] a = array; int m;
840	>	if (a == null || (m = a.length - 1) < 0)
841	>	return null;
842	>	int i = mode == 0 ? top - 1 : base;
843	>	int j = ((i & m) << ASHIFT) + ABASE;
844	>	return (ForkJoinTask<?>)U.getObjectVolatile(a, j);
845	>	}
846	>
847	>	/**
848	>	* Returns task at index b if b is current base of queue.
849	>	*/
850	>	final ForkJoinTask<?> pollAt(int b) {
851	>	ForkJoinTask<?>[] a; int i;
852	>	ForkJoinTask<?> task = null;
853	>	if ((a = array) != null && (i = ((a.length - 1) & b)) >= 0) {
854	>	int j = (i << ASHIFT) + ABASE;
855	>	ForkJoinTask<?> t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
856	>	if (t != null && base == b &&
857	>	U.compareAndSwapObject(a, j, t, null)) {
858	>	base = b + 1;
859	>	task = t;
860	>	}
861	>	}
862	>	return task;
863	>	}
864	>
865	>	/**
866	>	* Pops the given task only if it is at the current top.
867	>	*/
868	>	final boolean tryUnpush(ForkJoinTask<?> t) {
869	>	ForkJoinTask<?>[] a; int s;
870	>	if ((a = array) != null && (s = top) != base &&
871	>	U.compareAndSwapObject
872	>	(a, (((a.length - 1) & --s) << ASHIFT) + ABASE, t, null)) {
873	>	top = s;
874	>	return true;
875	>	}
876	>	return false;
877	>	}
878	>
879	>	/**
880	>	* Polls the given task only if it is at the current base.
881	>	*/
882	>	final boolean pollFor(ForkJoinTask<?> task) {
883	>	ForkJoinTask<?>[] a; int b, i;
884	>	if ((b = base) - top < 0 && (a = array) != null &&
885	>	(i = (a.length - 1) & b) >= 0) {
886	>	int j = (i << ASHIFT) + ABASE;
887	>	if (U.getObjectVolatile(a, j) == task && base == b &&
888	>	U.compareAndSwapObject(a, j, task, null)) {
889	>	base = b + 1;
890	>	return true;
891	>	}
892	>	}
893	>	return false;
894	>	}
895	>
896	>	/**
897	>	* If present, removes from queue and executes the given task, or
898	>	* any other cancelled task. Returns (true) immediately on any CAS
899	>	* or consistency check failure so caller can retry.
900	>	*
901	>	* @return false if no progress can be made
902	>	*/
903	>	final boolean tryRemoveAndExec(ForkJoinTask<?> task) {
904	>	boolean removed = false, empty = true, progress = true;
905	>	ForkJoinTask<?>[] a; int m, s, b, n;
906	>	if ((a = array) != null && (m = a.length - 1) >= 0 &&
907	>	(n = (s = top) - (b = base)) > 0) {
908	>	for (ForkJoinTask<?> t;;) {           // traverse from s to b
909	>	int j = ((--s & m) << ASHIFT) + ABASE;
910	>	t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
911	>	if (t == null)                    // inconsistent length
912	>	break;
913	>	else if (t == task) {
914	>	if (s + 1 == top) {           // pop
915	>	if (!U.compareAndSwapObject(a, j, task, null))
916	>	break;
917	>	top = s;
918	>	removed = true;
919	>	}
920	>	else if (base == b)           // replace with proxy
921	>	removed = U.compareAndSwapObject(a, j, task,
922	>	new EmptyTask());
923	>	break;
924	>	}
925	>	else if (t.status >= 0)
926	>	empty = false;
927	>	else if (s + 1 == top) {          // pop and throw away
928	>	if (U.compareAndSwapObject(a, j, t, null))
929	>	top = s;
930	>	break;
931	>	}
932	>	if (--n == 0) {
933	>	if (!empty && base == b)
934	>	progress = false;
935	>	break;
936	>	}
937	>	}
938	>	}
939	>	if (removed)
940	>	task.doExec();
941	>	return progress;
942	>	}
943	>
944	>	/**
945	>	* Initializes or doubles the capacity of array. Call either
946	>	* by owner or with lock held -- it is OK for base, but not
947	>	* top, to move while resizings are in progress.
948	>	*
949	>	* @param rejectOnFailure if true, throw exception if capacity
950	>	* exceeded (relayed ultimately to user); else return null.
951	>	*/
952	>	final ForkJoinTask<?>[] growArray(boolean rejectOnFailure) {
953	>	ForkJoinTask<?>[] oldA = array;
954	>	int size = oldA != null ? oldA.length << 1 : INITIAL_QUEUE_CAPACITY;
955	>	if (size <= MAXIMUM_QUEUE_CAPACITY) {
956	>	int oldMask, t, b;
957	>	ForkJoinTask<?>[] a = array = new ForkJoinTask<?>[size];
958	>	if (oldA != null && (oldMask = oldA.length - 1) >= 0 &&
959	>	(t = top) - (b = base) > 0) {
960	>	int mask = size - 1;
961	>	do {
962	>	ForkJoinTask<?> x;
963	>	int oldj = ((b & oldMask) << ASHIFT) + ABASE;
964	>	int j    = ((b &    mask) << ASHIFT) + ABASE;
965	>	x = (ForkJoinTask<?>)U.getObjectVolatile(oldA, oldj);
966	>	if (x != null &&
967	>	U.compareAndSwapObject(oldA, oldj, x, null))
968	>	U.putObjectVolatile(a, j, x);
969	>	} while (++b != t);
970	>	}
971	>	return a;
972	>	}
973	>	else if (!rejectOnFailure)
974	>	return null;
975	>	else
976	>	throw new RejectedExecutionException("Queue capacity exceeded");
977	>	}
978	>
979	>	/**
980	>	* Removes and cancels all known tasks, ignoring any exceptions
981	>	*/
982	>	final void cancelAll() {
983	>	ForkJoinTask.cancelIgnoringExceptions(currentJoin);
984	>	ForkJoinTask.cancelIgnoringExceptions(currentSteal);
985	>	for (ForkJoinTask<?> t; (t = poll()) != null; )
986	>	ForkJoinTask.cancelIgnoringExceptions(t);
987	>	}
988	>
989	>	// Execution methods
990	>
991	>	/**
992	>	* Removes and runs tasks until empty, using local mode
993	>	* ordering.
994	>	*/
995	>	final void runLocalTasks() {
996	>	if (base - top < 0) {
997	>	for (ForkJoinTask<?> t; (t = nextLocalTask()) != null; )
998	>	t.doExec();
999	>	}
1000	>	}
1001	>
1002	>	/**
1003	>	* Executes a top-level task and any local tasks remaining
1004	>	* after execution.
1005	>	*
1006	>	* @return true unless terminating
1007	>	*/
1008	>	final boolean runTask(ForkJoinTask<?> t) {
1009	>	boolean alive = true;
1010	>	if (t != null) {
1011	>	currentSteal = t;
1012	>	t.doExec();
1013	>	runLocalTasks();
1014	>	++nsteals;
1015	>	currentSteal = null;
1016	>	}
1017	>	else if (runState < 0)            // terminating
1018	>	alive = false;
1019	>	return alive;
1020	>	}
1021	>
1022	>	/**
1023	>	* Executes a non-top-level (stolen) task
1024	>	*/
1025	>	final void runSubtask(ForkJoinTask<?> t) {
1026	>	if (t != null) {
1027	>	ForkJoinTask<?> ps = currentSteal;
1028	>	currentSteal = t;
1029	>	t.doExec();
1030	>	currentSteal = ps;
1031	>	}
1032	>	}
1033	>
1034	>	/**
1035	>	* Computes next value for random probes.  Scans don't require
1036	>	* a very high quality generator, but also not a crummy one.
1037	>	* Marsaglia xor-shift is cheap and works well enough.  Note:
1038	>	* This is manually inlined in several usages in ForkJoinPool
1039	>	* to avoid writes inside busy scan loops.
1040	>	*/
1041	>	final int nextSeed() {
1042	>	int r = seed;
1043	>	r ^= r << 13;
1044	>	r ^= r >>> 17;
1045	>	r ^= r << 5;
1046	>	return seed = r;
1047	>	}
1048	>
1049	>	// Unsafe mechanics
1050	>	private static final sun.misc.Unsafe U;
1051	>	private static final long RUNSTATE;
1052	>	private static final int ABASE;
1053	>	private static final int ASHIFT;
1054	>	static {
1055	>	int s;
1056	>	try {
1057	>	U = getUnsafe();
1058	>	Class<?> k = WorkQueue.class;
1059	>	Class<?> ak = ForkJoinTask[].class;
1060	>	RUNSTATE = U.objectFieldOffset
1061	>	(k.getDeclaredField("runState"));
1062	>	ABASE = U.arrayBaseOffset(ak);
1063	>	s = U.arrayIndexScale(ak);
1064	>	} catch (Exception e) {
1065	>	throw new Error(e);
1066	>	}
1067	>	if ((s & (s-1)) != 0)
1068	>	throw new Error("data type scale not a power of two");
1069	>	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
1070	>	}
1071	>	}
1072
1073		/**
1074	<	* The uncaught exception handler used when any worker
1075	<	* abruptly terminates
1074	>	* Class for artificial tasks that are used to replace the target
1075	>	* of local joins if they are removed from an interior queue slot
1076	>	* in WorkQueue.tryRemoveAndExec. We don't need the proxy to
1077	>	* actually do anything beyond having a unique identity.
1078		*/
1079	<	private Thread.UncaughtExceptionHandler ueh;
1079	>	static final class EmptyTask extends ForkJoinTask<Void> {
1080	>	EmptyTask() { status = ForkJoinTask.NORMAL; } // force done
1081	>	public Void getRawResult() { return null; }
1082	>	public void setRawResult(Void x) {}
1083	>	public boolean exec() { return true; }
1084	>	}
1085
1086		/**
1087	<	* Creation factory for worker threads.
1087	>	* Computes a hash code for the given thread. This method is
1088	>	* expected to provide higher-quality hash codes than those using
1089	>	* method hashCode().
1090		*/
1091	<	private final ForkJoinWorkerThreadFactory factory;
1091	>	static final int hashThread(Thread t) {
1092	>	long id = (t == null)? 0L : t.getId(); // Use MurmurHash of thread id
1093	>	int h = (int)id ^ (int)(id >>> 32);
1094	>	h ^= h >>> 16;
1095	>	h *= 0x85ebca6b;
1096	>	h ^= h >>> 13;
1097	>	h *= 0xc2b2ae35;
1098	>	return h ^ (h >>> 16);
1099	>	}
1100
1101		/**
1102	<	* Head of stack of threads that were created to maintain
177	<	* parallelism when other threads blocked, but have since
178	<	* suspended when the parallelism level rose.
1102	>	* Top-level runloop for workers
1103		*/
1104	<	private volatile WaitQueueNode spareStack;
1104	>	final void runWorker(ForkJoinWorkerThread wt) {
1105	>	WorkQueue w = wt.workQueue;
1106	>	w.growArray(false);     // Initialize queue array and seed in this thread
1107	>	w.seed = hashThread(Thread.currentThread()) | (1 << 31); // force < 0
1108	>
1109	>	do {} while (w.runTask(scan(w)));
1110	>	}
1111	>
1112	>	// Creating, registering and deregistering workers
1113
1114		/**
1115	<	* Sum of per-thread steal counts, updated only when threads are
184	<	* idle or terminating.
1115	>	* Tries to create and start a worker
1116		*/
1117	<	private final AtomicLong stealCount;
1117	>	private void addWorker() {
1118	>	Throwable ex = null;
1119	>	ForkJoinWorkerThread w = null;
1120	>	try {
1121	>	if ((w = factory.newThread(this)) != null) {
1122	>	w.start();
1123	>	return;
1124	>	}
1125	>	} catch (Throwable e) {
1126	>	ex = e;
1127	>	}
1128	>	deregisterWorker(w, ex);
1129	>	}
1130
1131		/**
1132	<	* Queue for external submissions.
1132	>	* Callback from ForkJoinWorkerThread constructor to assign a
1133	>	* public name. This must be separate from registerWorker because
1134	>	* it is called during the "super" constructor call in
1135	>	* ForkJoinWorkerThread.
1136		*/
1137	<	private final LinkedTransferQueue<ForkJoinTask<?>> submissionQueue;
1137	>	final String nextWorkerName() {
1138	>	return workerNamePrefix.concat
1139	>	(Integer.toString(nextWorkerNumber.addAndGet(1)));
1140	>	}
1141
1142		/**
1143	<	* Head of Treiber stack for barrier sync. See below for explanation.
1143	>	* Callback from ForkJoinWorkerThread constructor to establish and
1144	>	* record its WorkQueue
1145	>	*
1146	>	* @param wt the worker thread
1147		*/
1148	<	private volatile WaitQueueNode syncStack;
1148	>	final void registerWorker(ForkJoinWorkerThread wt) {
1149	>	WorkQueue w = wt.workQueue;
1150	>	ReentrantLock lock = this.lock;
1151	>	lock.lock();
1152	>	try {
1153	>	int k = nextPoolIndex;
1154	>	WorkQueue[] ws = workQueues;
1155	>	if (ws != null) {                       // ignore on shutdown
1156	>	int n = ws.length;
1157	>	if (k < 0 || (k & 1) == 0 || k >= n || ws[k] != null) {
1158	>	for (k = 1; k < n && ws[k] != null; k += 2)
1159	>	;                           // workers are at odd indices
1160	>	if (k >= n)                     // resize
1161	>	workQueues = ws = Arrays.copyOf(ws, n << 1);
1162	>	}
1163	>	w.poolIndex = k;
1164	>	w.eventCount = ~(k >>> 1) & SMASK;  // Set up wait count
1165	>	ws[k] = w;                          // record worker
1166	>	nextPoolIndex = k + 2;
1167	>	int rs = runState;
1168	>	int m = rs & SMASK;                 // recalculate runState mask
1169	>	if (k > m)
1170	>	m = (m << 1) + 1;
1171	>	runState = (rs & SHUTDOWN) | ((rs + RS_SEQ) & RS_SEQ_MASK) | m;
1172	>	}
1173	>	} finally {
1174	>	lock.unlock();
1175	>	}
1176	>	}
1177
1178		/**
1179	<	* The count for event barrier
1180	<	*/
1181	<	private volatile long eventCount;
1179	>	* Final callback from terminating worker, as well as failure to
1180	>	* construct or start a worker in addWorker.  Removes record of
1181	>	* worker from array, and adjusts counts. If pool is shutting
1182	>	* down, tries to complete termination.
1183	>	*
1184	>	* @param wt the worker thread or null if addWorker failed
1185	>	* @param ex the exception causing failure, or null if none
1186	>	*/
1187	>	final void deregisterWorker(ForkJoinWorkerThread wt, Throwable ex) {
1188	>	WorkQueue w = null;
1189	>	if (wt != null && (w = wt.workQueue) != null) {
1190	>	w.runState = -1;                // ensure runState is set
1191	>	stealCount.getAndAdd(w.totalSteals + w.nsteals);
1192	>	int idx = w.poolIndex;
1193	>	ReentrantLock lock = this.lock;
1194	>	lock.lock();
1195	>	try {                           // remove record from array
1196	>	WorkQueue[] ws = workQueues;
1197	>	if (ws != null && idx >= 0 && idx < ws.length && ws[idx] == w)
1198	>	ws[nextPoolIndex = idx] = null;
1199	>	} finally {
1200	>	lock.unlock();
1201	>	}
1202	>	}
1203	>
1204	>	long c;                             // adjust ctl counts
1205	>	do {} while (!U.compareAndSwapLong
1206	>	(this, CTL, c = ctl, (((c - AC_UNIT) & AC_MASK) |
1207	>	((c - TC_UNIT) & TC_MASK) |
1208	>	(c & ~(AC_MASK|TC_MASK)))));
1209	>
1210	>	if (!tryTerminate(false) && w != null) {
1211	>	w.cancelAll();                  // cancel remaining tasks
1212	>	if (w.array != null)            // suppress signal if never ran
1213	>	signalWork();               // wake up or create replacement
1214	>	}
1215	>
1216	>	if (ex != null)                     // rethrow
1217	>	U.throwException(ex);
1218	>	}
1219	>
1220	>
1221	>	// Maintaining ctl counts
1222
1223		/**
1224	<	* Pool number, just for assigning useful names to worker threads
1224	>	* Increments active count; mainly called upon return from blocking
1225		*/
1226	<	private final int poolNumber;
1226	>	final void incrementActiveCount() {
1227	>	long c;
1228	>	do {} while (!U.compareAndSwapLong(this, CTL, c = ctl, c + AC_UNIT));
1229	>	}
1230
1231		/**
1232	<	* The maximum allowed pool size
1232	>	* Activates or creates a worker
1233		*/
1234	<	private volatile int maxPoolSize;
1234	>	final void signalWork() {
1235	>	/*
1236	>	* The while condition is true if: (there is are too few total
1237	>	* workers OR there is at least one waiter) AND (there are too
1238	>	* few active workers OR the pool is terminating).  The value
1239	>	* of e distinguishes the remaining cases: zero (no waiters)
1240	>	* for create, negative if terminating (in which case do
1241	>	* nothing), else release a waiter. The secondary checks for
1242	>	* release (non-null array etc) can fail if the pool begins
1243	>	* terminating after the test, and don't impose any added cost
1244	>	* because JVMs must perform null and bounds checks anyway.
1245	>	*/
1246	>	long c; int e, u;
1247	>	while ((((e = (int)(c = ctl)) | (u = (int)(c >>> 32))) &
1248	>	(INT_SIGN|SHORT_SIGN)) == (INT_SIGN|SHORT_SIGN)) {
1249	>	WorkQueue[] ws = workQueues; int i; WorkQueue w; Thread p;
1250	>	if (e == 0) {                    // add a new worker
1251	>	if (U.compareAndSwapLong
1252	>	(this, CTL, c, (long)(((u + UTC_UNIT) & UTC_MASK) |
1253	>	((u + UAC_UNIT) & UAC_MASK)) << 32)) {
1254	>	addWorker();
1255	>	break;
1256	>	}
1257	>	}
1258	>	else if (e > 0 && ws != null &&
1259	>	(i = ((~e << 1) | 1) & SMASK) < ws.length &&
1260	>	(w = ws[i]) != null &&
1261	>	w.eventCount == (e | INT_SIGN)) {
1262	>	if (U.compareAndSwapLong
1263	>	(this, CTL, c, (((long)(w.nextWait & E_MASK)) |
1264	>	((long)(u + UAC_UNIT) << 32)))) {
1265	>	w.eventCount = (e + E_SEQ) & E_MASK;
1266	>	if ((p = w.parker) != null)
1267	>	U.unpark(p);         // release a waiting worker
1268	>	break;
1269	>	}
1270	>	}
1271	>	else
1272	>	break;
1273	>	}
1274	>	}
1275
1276		/**
1277	<	* The desired parallelism level, updated only under workerLock.
1278	<	*/
1279	<	private volatile int parallelism;
1277	>	* Tries to decrement active count (sometimes implicitly) and
1278	>	* possibly release or create a compensating worker in preparation
1279	>	* for blocking. Fails on contention or termination.
1280	>	*
1281	>	* @return true if the caller can block, else should recheck and retry
1282	>	*/
1283	>	final boolean tryCompensate() {
1284	>	WorkQueue[] ws; WorkQueue w; Thread p;
1285	>	int pc = parallelism, e, u, ac, tc, i;
1286	>	long c = ctl;
1287	>
1288	>	if ((e = (int)c) >= 0) {
1289	>	if ((ac = ((u = (int)(c >>> 32)) >> UAC_SHIFT)) <= 0 &&
1290	>	e != 0 && (ws = workQueues) != null &&
1291	>	(i = ((~e << 1) | 1) & SMASK) < ws.length &&
1292	>	(w = ws[i]) != null) {
1293	>	if (w.eventCount == (e | INT_SIGN) &&
1294	>	U.compareAndSwapLong
1295	>	(this, CTL, c, ((long)(w.nextWait & E_MASK) |
1296	>	(c & (AC_MASK|TC_MASK))))) {
1297	>	w.eventCount = (e + E_SEQ) & E_MASK;
1298	>	if ((p = w.parker) != null)
1299	>	U.unpark(p);
1300	>	return true;             // release an idle worker
1301	>	}
1302	>	}
1303	>	else if ((tc = (short)(u >>> UTC_SHIFT)) >= 0 && ac + pc > 1) {
1304	>	long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK);
1305	>	if (U.compareAndSwapLong(this, CTL, c, nc))
1306	>	return true;             // no compensation needed
1307	>	}
1308	>	else if (tc + pc < MAX_ID) {
1309	>	long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK);
1310	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1311	>	addWorker();
1312	>	return true;             // create replacement
1313	>	}
1314	>	}
1315	>	}
1316	>	return false;
1317	>	}
1318	>
1319	>	// Submissions
1320
1321		/**
1322	<	* True if use local fifo, not default lifo, for local polling
1322	>	* Unless shutting down, adds the given task to some submission
1323	>	* queue; using a randomly chosen queue index if the caller is a
1324	>	* ForkJoinWorkerThread, else one based on caller thread's hash
1325	>	* code. If no queue exists at the index, one is created.  If the
1326	>	* queue is busy, another is chosen by sweeping through the queues
1327	>	* array.
1328		*/
1329	<	private volatile boolean locallyFifo;
1329	>	private void doSubmit(ForkJoinTask<?> task) {
1330	>	if (task == null)
1331	>	throw new NullPointerException();
1332	>	Thread t = Thread.currentThread();
1333	>	int r = ((t instanceof ForkJoinWorkerThread) ?
1334	>	((ForkJoinWorkerThread)t).workQueue.nextSeed() : hashThread(t));
1335	>	for (;;) {
1336	>	int rs = runState, m = rs & SMASK;
1337	>	int j = r &= (m & ~1);                      // even numbered queues
1338	>	WorkQueue[] ws = workQueues;
1339	>	if (rs < 0 || ws == null)
1340	>	throw new RejectedExecutionException(); // shutting down
1341	>	if (ws.length > m) {                        // consistency check
1342	>	for (WorkQueue q;;) {                   // circular sweep
1343	>	if (((q = ws[j]) != null ||
1344	>	(q = tryAddSharedQueue(j)) != null) &&
1345	>	q.trySharedPush(task)) {
1346	>	signalWork();
1347	>	return;
1348	>	}
1349	>	if ((j = (j + 2) & m) == r) {
1350	>	Thread.yield();                 // all queues busy
1351	>	break;
1352	>	}
1353	>	}
1354	>	}
1355	>	}
1356	>	}
1357
1358		/**
1359	<	* Holds number of total (i.e., created and not yet terminated)
1360	<	* and running (i.e., not blocked on joins or other managed sync)
1361	<	* threads, packed into one int to ensure consistent snapshot when
1362	<	* making decisions about creating and suspending spare
1363	<	* threads. Updated only by CAS.  Note: CASes in
1364	<	* updateRunningCount and preJoin assume that running active count
1365	<	* is in low word, so need to be modified if this changes.
1366	<	*/
1367	<	private volatile int workerCounts;
1359	>	* Tries to add and register a new queue at the given index.
1360	>	*
1361	>	* @param idx the workQueues array index to register the queue
1362	>	* @return the queue, or null if could not add because could
1363	>	* not acquire lock or idx is unusable
1364	>	*/
1365	>	private WorkQueue tryAddSharedQueue(int idx) {
1366	>	WorkQueue q = null;
1367	>	ReentrantLock lock = this.lock;
1368	>	if (idx >= 0 && (idx & 1) == 0 && !lock.isLocked()) {
1369	>	// create queue outside of lock but only if apparently free
1370	>	WorkQueue nq = new WorkQueue(null, SHARED_QUEUE);
1371	>	if (lock.tryLock()) {
1372	>	try {
1373	>	WorkQueue[] ws = workQueues;
1374	>	if (ws != null && idx < ws.length) {
1375	>	if ((q = ws[idx]) == null) {
1376	>	int rs;         // update runState seq
1377	>	ws[idx] = q = nq;
1378	>	runState = (((rs = runState) & SHUTDOWN) |
1379	>	((rs + RS_SEQ) & ~SHUTDOWN));
1380	>	}
1381	>	}
1382	>	} finally {
1383	>	lock.unlock();
1384	>	}
1385	>	}
1386	>	}
1387	>	return q;
1388	>	}
1389
1390	<	private static int totalCountOf(int s)           { return s >>> 16;  }
235	<	private static int runningCountOf(int s)         { return s & shortMask; }
236	<	private static int workerCountsFor(int t, int r) { return (t << 16) + r; }
1390	>	// Scanning for tasks
1391
1392		/**
1393	<	* Adds delta (which may be negative) to running count.  This must
1394	<	* be called before (with negative arg) and after (with positive)
1395	<	* any managed synchronization (i.e., mainly, joins).
1396	<	*
1397	<	* @param delta the number to add
1398	<	*/
1399	<	final void updateRunningCount(int delta) {
1400	<	int s;
1401	<	do {} while (!casWorkerCounts(s = workerCounts, s + delta));
1393	>	* Scans for and, if found, returns one task, else possibly
1394	>	* inactivates the worker. This method operates on single reads of
1395	>	* volatile state and is designed to be re-invoked continuously in
1396	>	* part because it returns upon detecting inconsistencies,
1397	>	* contention, or state changes that indicate possible success on
1398	>	* re-invocation.
1399	>	*
1400	>	* The scan searches for tasks across queues, randomly selecting
1401	>	* the first #queues probes, favoring steals 2:1 over submissions
1402	>	* (by exploiting even/odd indexing), and then performing a
1403	>	* circular sweep of all queues.  The scan terminates upon either
1404	>	* finding a non-empty queue, or completing a full sweep. If the
1405	>	* worker is not inactivated, it takes and returns a task from
1406	>	* this queue.  On failure to find a task, we take one of the
1407	>	* following actions, after which the caller will retry calling
1408	>	* this method unless terminated.
1409	>	*
1410	>	* * If not a complete sweep, try to release a waiting worker.  If
1411	>	* the scan terminated because the worker is inactivated, then the
1412	>	* released worker will often be the calling worker, and it can
1413	>	* succeed obtaining a task on the next call. Or maybe it is
1414	>	* another worker, but with same net effect. Releasing in other
1415	>	* cases as well ensures that we have enough workers running.
1416	>	*
1417	>	* * If the caller has run a task since the the last empty scan,
1418	>	* return (to allow rescan) if other workers are not also yet
1419	>	* enqueued.  Field WorkQueue.rescans counts down on each scan to
1420	>	* ensure eventual inactivation, and occasional calls to
1421	>	* Thread.yield to help avoid interference with more useful
1422	>	* activities on the system.
1423	>	*
1424	>	* * If pool is terminating, terminate the worker
1425	>	*
1426	>	* * If not already enqueued, try to inactivate and enqueue the
1427	>	* worker on wait queue.
1428	>	*
1429	>	* * If already enqueued and none of the above apply, either park
1430	>	* awaiting signal, or if this is the most recent waiter and pool
1431	>	* is quiescent, relay to idleAwaitWork to check for termination
1432	>	* and possibly shrink pool.
1433	>	*
1434	>	* @param w the worker (via its WorkQueue)
1435	>	* @return a task or null of none found
1436	>	*/
1437	>	private final ForkJoinTask<?> scan(WorkQueue w) {
1438	>	boolean swept = false;                 // true after full empty scan
1439	>	WorkQueue[] ws;                        // volatile read order matters
1440	>	int r = w.seed, ec = w.eventCount;     // ec is negative if inactive
1441	>	int rs = runState, m = rs & SMASK;
1442	>	if ((ws = workQueues) != null && ws.length > m) {
1443	>	ForkJoinTask<?> task = null;
1444	>	for (int k = 0, j = -2 - m; ; ++j) {
1445	>	WorkQueue q; int b;
1446	>	if (j < 0) {                    // random probes while j negative
1447	>	r ^= r << 13; r ^= r >>> 17; k = (r ^= r << 5) | (j & 1);
1448	>	}                               // worker (not submit) for odd j
1449	>	else                            // cyclic scan when j >= 0
1450	>	k += (m >>> 1) | 1;         // step by half to reduce bias
1451	>
1452	>	if ((q = ws[k & m]) != null && (b = q.base) - q.top < 0) {
1453	>	if (ec >= 0)
1454	>	task = q.pollAt(b);     // steal
1455	>	break;
1456	>	}
1457	>	else if (j > m) {
1458	>	if (rs == runState)        // staleness check
1459	>	swept = true;
1460	>	break;
1461	>	}
1462	>	}
1463	>	w.seed = r;                        // save seed for next scan
1464	>	if (task != null)
1465	>	return task;
1466	>	}
1467	>
1468	>	// Decode ctl on empty scan
1469	>	long c = ctl; int e = (int)c, a = (int)(c >> AC_SHIFT), nr, ns;
1470	>	if (!swept) {                          // try to release a waiter
1471	>	WorkQueue v; Thread p;
1472	>	if (e > 0 && a < 0 && ws != null &&
1473	>	(v = ws[((~e << 1) | 1) & m]) != null &&
1474	>	v.eventCount == (e | INT_SIGN) && U.compareAndSwapLong
1475	>	(this, CTL, c, ((long)(v.nextWait & E_MASK) |
1476	>	((c + AC_UNIT) & (AC_MASK|TC_MASK))))) {
1477	>	v.eventCount = (e + E_SEQ) & E_MASK;
1478	>	if ((p = v.parker) != null)
1479	>	U.unpark(p);
1480	>	}
1481	>	}
1482	>	else if ((nr = w.rescans) > 0) {       // continue rescanning
1483	>	int ac = a + parallelism;
1484	>	if ((w.rescans = (ac < nr) ? ac : nr - 1) > 0 && w.seed < 0 &&
1485	>	w.eventCount == ec)
1486	>	Thread.yield();                // 1 bit randomness for yield call
1487	>	}
1488	>	else if (e < 0)                        // pool is terminating
1489	>	w.runState = -1;
1490	>	else if (ec >= 0) {                    // try to enqueue
1491	>	long nc = (long)ec | ((c - AC_UNIT) & (AC_MASK|TC_MASK));
1492	>	w.nextWait = e;
1493	>	w.eventCount = ec | INT_SIGN;      // mark as inactive
1494	>	if (!U.compareAndSwapLong(this, CTL, c, nc))
1495	>	w.eventCount = ec;             // back out on CAS failure
1496	>	else if ((ns = w.nsteals) != 0) {  // set rescans if ran task
1497	>	if (a <= 0)                    // ... unless too many active
1498	>	w.rescans = a + parallelism;
1499	>	w.nsteals = 0;
1500	>	w.totalSteals += ns;
1501	>	}
1502	>	}
1503	>	else{                                  // already queued
1504	>	if (parallelism == -a)
1505	>	idleAwaitWork(w);              // quiescent
1506	>	if (w.eventCount == ec) {
1507	>	Thread.interrupted();          // clear status
1508	>	ForkJoinWorkerThread wt = w.owner;
1509	>	U.putObject(wt, PARKBLOCKER, this);
1510	>	w.parker = wt;                 // emulate LockSupport.park
1511	>	if (w.eventCount == ec)        // recheck
1512	>	U.park(false, 0L);         // block
1513	>	w.parker = null;
1514	>	U.putObject(wt, PARKBLOCKER, null);
1515	>	}
1516	>	}
1517	>	return null;
1518	>	}
1519	>
1520	>	/**
1521	>	* If inactivating worker w has caused pool to become quiescent,
1522	>	* check for pool termination, and, so long as this is not the
1523	>	* only worker, wait for event for up to SHRINK_RATE nanosecs On
1524	>	* timeout, if ctl has not changed, terminate the worker, which
1525	>	* will in turn wake up another worker to possibly repeat this
1526	>	* process.
1527	>	*
1528	>	* @param w the calling worker
1529	>	*/
1530	>	private void idleAwaitWork(WorkQueue w) {
1531	>	long c; int nw, ec;
1532	>	if (!tryTerminate(false) &&
1533	>	(int)((c = ctl) >> AC_SHIFT) + parallelism == 0 &&
1534	>	(ec = w.eventCount) == ((int)c | INT_SIGN) &&
1535	>	(nw = w.nextWait) != 0) {
1536	>	long nc = ((long)(nw & E_MASK) | // ctl to restore on timeout
1537	>	((c + AC_UNIT) & AC_MASK) | (c & TC_MASK));
1538	>	ForkJoinTask.helpExpungeStaleExceptions(); // help clean
1539	>	ForkJoinWorkerThread wt = w.owner;
1540	>	while (ctl == c) {
1541	>	long startTime = System.nanoTime();
1542	>	Thread.interrupted();  // timed variant of version in scan()
1543	>	U.putObject(wt, PARKBLOCKER, this);
1544	>	w.parker = wt;
1545	>	if (ctl == c)
1546	>	U.park(false, SHRINK_RATE);
1547	>	w.parker = null;
1548	>	U.putObject(wt, PARKBLOCKER, null);
1549	>	if (ctl != c)
1550	>	break;
1551	>	if (System.nanoTime() - startTime >= SHRINK_TIMEOUT &&
1552	>	U.compareAndSwapLong(this, CTL, c, nc)) {
1553	>	w.runState = -1;          // shrink
1554	>	w.eventCount = (ec + E_SEQ) | E_MASK;
1555	>	break;
1556	>	}
1557	>	}
1558	>	}
1559	>	}
1560	>
1561	>	/**
1562	>	* Tries to locate and execute tasks for a stealer of the given
1563	>	* task, or in turn one of its stealers, Traces currentSteal ->
1564	>	* currentJoin links looking for a thread working on a descendant
1565	>	* of the given task and with a non-empty queue to steal back and
1566	>	* execute tasks from. The first call to this method upon a
1567	>	* waiting join will often entail scanning/search, (which is OK
1568	>	* because the joiner has nothing better to do), but this method
1569	>	* leaves hints in workers to speed up subsequent calls. The
1570	>	* implementation is very branchy to cope with potential
1571	>	* inconsistencies or loops encountering chains that are stale,
1572	>	* unknown, or of length greater than MAX_HELP_DEPTH links.  All
1573	>	* of these cases are dealt with by just retrying by caller.
1574	>	*
1575	>	* @param joiner the joining worker
1576	>	* @param task the task to join
1577	>	* @return true if found or ran a task (and so is immediately retryable)
1578	>	*/
1579	>	final boolean tryHelpStealer(WorkQueue joiner, ForkJoinTask<?> task) {
1580	>	ForkJoinTask<?> subtask;    // current target
1581	>	boolean progress = false;
1582	>	int depth = 0;              // current chain depth
1583	>	int m = runState & SMASK;
1584	>	WorkQueue[] ws = workQueues;
1585	>
1586	>	if (ws != null && ws.length > m && (subtask = task).status >= 0) {
1587	>	outer:for (WorkQueue j = joiner;;) {
1588	>	// Try to find the stealer of subtask, by first using hint
1589	>	WorkQueue stealer = null;
1590	>	WorkQueue v = ws[j.stealHint & m];
1591	>	if (v != null && v.currentSteal == subtask)
1592	>	stealer = v;
1593	>	else {
1594	>	for (int i = 1; i <= m; i += 2) {
1595	>	if ((v = ws[i]) != null && v.currentSteal == subtask) {
1596	>	stealer = v;
1597	>	j.stealHint = i; // save hint
1598	>	break;
1599	>	}
1600	>	}
1601	>	if (stealer == null)
1602	>	break;
1603	>	}
1604	>
1605	>	for (WorkQueue q = stealer;;) { // Try to help stealer
1606	>	ForkJoinTask<?> t; int b;
1607	>	if (task.status < 0)
1608	>	break outer;
1609	>	if ((b = q.base) - q.top < 0) {
1610	>	progress = true;
1611	>	if (subtask.status < 0)
1612	>	break outer;               // stale
1613	>	if ((t = q.pollAt(b)) != null) {
1614	>	stealer.stealHint = joiner.poolIndex;
1615	>	joiner.runSubtask(t);
1616	>	}
1617	>	}
1618	>	else { // empty - try to descend to find stealer's stealer
1619	>	ForkJoinTask<?> next = stealer.currentJoin;
1620	>	if (++depth == MAX_HELP_DEPTH || subtask.status < 0 ||
1621	>	next == null || next == subtask)
1622	>	break outer;  // max depth, stale, dead-end, cyclic
1623	>	subtask = next;
1624	>	j = stealer;
1625	>	break;
1626	>	}
1627	>	}
1628	>	}
1629	>	}
1630	>	return progress;
1631		}
1632
1633		/**
1634	<	* Adds delta (which may be negative) to both total and running
252	<	* count.  This must be called upon creation and termination of
253	<	* worker threads.
1634	>	* If task is at base of some steal queue, steals and executes it.
1635		*
1636	<	* @param delta the number to add
1636	>	* @param joiner the joining worker
1637	>	* @param task the task
1638		*/
1639	<	private void updateWorkerCount(int delta) {
1640	<	int d = delta + (delta << 16); // add to both lo and hi parts
1641	<	int s;
1642	<	do {} while (!casWorkerCounts(s = workerCounts, s + d));
1639	>	final void tryPollForAndExec(WorkQueue joiner, ForkJoinTask<?> task) {
1640	>	WorkQueue[] ws;
1641	>	int m = runState & SMASK;
1642	>	if ((ws = workQueues) != null && ws.length > m) {
1643	>	for (int j = 1; j <= m && task.status >= 0; j += 2) {
1644	>	WorkQueue q = ws[j];
1645	>	if (q != null && q.pollFor(task)) {
1646	>	joiner.runSubtask(task);
1647	>	break;
1648	>	}
1649	>	}
1650	>	}
1651		}
1652
1653		/**
1654	<	* Lifecycle control. High word contains runState, low word
1655	<	* contains the number of workers that are (probably) executing
1656	<	* tasks. This value is atomically incremented before a worker
1657	<	* gets a task to run, and decremented when worker has no tasks
1658	<	* and cannot find any. These two fields are bundled together to
1659	<	* support correct termination triggering.  Note: activeCount
1660	<	* CAS'es cheat by assuming active count is in low word, so need
1661	<	* to be modified if this changes
1662	<	*/
1663	<	private volatile int runControl;
1654	>	* Returns a non-empty steal queue, if is found during a random,
1655	>	* then cyclic scan, else null.  This method must be retried by
1656	>	* caller if, by the time it tries to use the queue, it is empty.
1657	>	*/
1658	>	private WorkQueue findNonEmptyStealQueue(WorkQueue w) {
1659	>	int r = w.seed;    // Same idea as scan(), but ignoring submissions
1660	>	for (WorkQueue[] ws;;) {
1661	>	int m = runState & SMASK;
1662	>	if ((ws = workQueues) == null)
1663	>	return null;
1664	>	if (ws.length > m) {
1665	>	WorkQueue q;
1666	>	for (int n = m << 2, k = r, j = -n;;) {
1667	>	r ^= r << 13; r ^= r >>> 17; r ^= r << 5;
1668	>	if ((q = ws[(k | 1) & m]) != null && q.base - q.top < 0) {
1669	>	w.seed = r;
1670	>	return q;
1671	>	}
1672	>	else if (j > n)
1673	>	return null;
1674	>	else
1675	>	k = (j++ < 0) ? r : k + ((m >>> 1) | 1);
1676
1677	<	// RunState values. Order among values matters
1678	<	private static final int RUNNING     = 0;
1679	<	private static final int SHUTDOWN    = 1;
1680	<	private static final int TERMINATING = 2;
279	<	private static final int TERMINATED  = 3;
1677	>	}
1678	>	}
1679	>	}
1680	>	}
1681
1682	<	private static int runStateOf(int c)             { return c >>> 16; }
1683	<	private static int activeCountOf(int c)          { return c & shortMask; }
1684	<	private static int runControlFor(int r, int a)   { return (r << 16) + a; }
1682	>	/**
1683	>	* Runs tasks until {@code isQuiescent()}. We piggyback on
1684	>	* active count ctl maintenance, but rather than blocking
1685	>	* when tasks cannot be found, we rescan until all others cannot
1686	>	* find tasks either.
1687	>	*/
1688	>	final void helpQuiescePool(WorkQueue w) {
1689	>	for (boolean active = true;;) {
1690	>	w.runLocalTasks();      // exhaust local queue
1691	>	WorkQueue q = findNonEmptyStealQueue(w);
1692	>	if (q != null) {
1693	>	ForkJoinTask<?> t;
1694	>	if (!active) {      // re-establish active count
1695	>	long c;
1696	>	active = true;
1697	>	do {} while (!U.compareAndSwapLong
1698	>	(this, CTL, c = ctl, c + AC_UNIT));
1699	>	}
1700	>	if ((t = q.poll()) != null)
1701	>	w.runSubtask(t);
1702	>	}
1703	>	else {
1704	>	long c;
1705	>	if (active) {       // decrement active count without queuing
1706	>	active = false;
1707	>	do {} while (!U.compareAndSwapLong
1708	>	(this, CTL, c = ctl, c -= AC_UNIT));
1709	>	}
1710	>	else
1711	>	c = ctl;        // re-increment on exit
1712	>	if ((int)(c >> AC_SHIFT) + parallelism == 0) {
1713	>	do {} while (!U.compareAndSwapLong
1714	>	(this, CTL, c = ctl, c + AC_UNIT));
1715	>	break;
1716	>	}
1717	>	}
1718	>	}
1719	>	}
1720
1721		/**
1722	<	* Tries incrementing active count; fails on contention.
287	<	* Called by workers before/during executing tasks.
1722	>	* Gets and removes a local or stolen task for the given worker
1723		*
1724	<	* @return true on success
1724	>	* @return a task, if available
1725		*/
1726	<	final boolean tryIncrementActiveCount() {
1727	<	int c = runControl;
1728	<	return casRunControl(c, c+1);
1726	>	final ForkJoinTask<?> nextTaskFor(WorkQueue w) {
1727	>	for (ForkJoinTask<?> t;;) {
1728	>	WorkQueue q;
1729	>	if ((t = w.nextLocalTask()) != null)
1730	>	return t;
1731	>	if ((q = findNonEmptyStealQueue(w)) == null)
1732	>	return null;
1733	>	if ((t = q.poll()) != null)
1734	>	return t;
1735	>	}
1736		}
1737
1738		/**
1739	<	* Tries decrementing active count; fails on contention.
1740	<	* Possibly triggers termination on success.
1741	<	* Called by workers when they can't find tasks.
300	<	*
301	<	* @return true on success
1739	>	* Returns the approximate (non-atomic) number of idle threads per
1740	>	* active thread to offset steal queue size for method
1741	>	* ForkJoinTask.getSurplusQueuedTaskCount().
1742		*/
1743	<	final boolean tryDecrementActiveCount() {
1744	<	int c = runControl;
1745	<	int nextc = c - 1;
1746	<	if (!casRunControl(c, nextc))
1747	<	return false;
1748	<	if (canTerminateOnShutdown(nextc))
1749	<	terminateOnShutdown();
1750	<	return true;
1743	>	final int idlePerActive() {
1744	>	// Approximate at powers of two for small values, saturate past 4
1745	>	int p = parallelism;
1746	>	int a = p + (int)(ctl >> AC_SHIFT);
1747	>	return (a > (p >>>= 1) ? 0 :
1748	>	a > (p >>>= 1) ? 1 :
1749	>	a > (p >>>= 1) ? 2 :
1750	>	a > (p >>>= 1) ? 4 :
1751	>	8);
1752		}
1753
1754	+	// Termination
1755	+
1756		/**
1757	<	* Returns true if argument represents zero active count and
315	<	* nonzero runstate, which is the triggering condition for
316	<	* terminating on shutdown.
1757	>	* Sets SHUTDOWN bit of runState under lock
1758		*/
1759	<	private static boolean canTerminateOnShutdown(int c) {
1760	<	// i.e. least bit is nonzero runState bit
1761	<	return ((c & -c) >>> 16) != 0;
1759	>	private void enableShutdown() {
1760	>	ReentrantLock lock = this.lock;
1761	>	if (runState >= 0) {
1762	>	lock.lock();                       // don't need try/finally
1763	>	runState |= SHUTDOWN;
1764	>	lock.unlock();
1765	>	}
1766		}
1767
1768		/**
1769	<	* Transition run state to at least the given state. Return true
1770	<	* if not already at least given state.
1771	<	*/
1772	<	private boolean transitionRunStateTo(int state) {
1773	<	for (;;) {
1774	<	int c = runControl;
1775	<	if (runStateOf(c) >= state)
1776	<	return false;
1777	<	if (casRunControl(c, runControlFor(state, activeCountOf(c))))
1769	>	* Possibly initiates and/or completes termination.  Upon
1770	>	* termination, cancels all queued tasks and then
1771	>	*
1772	>	* @param now if true, unconditionally terminate, else only
1773	>	* if no work and no active workers
1774	>	* @return true if now terminating or terminated
1775	>	*/
1776	>	private boolean tryTerminate(boolean now) {
1777	>	for (long c;;) {
1778	>	if (((c = ctl) & STOP_BIT) != 0) {      // already terminating
1779	>	if ((short)(c >>> TC_SHIFT) == -parallelism) {
1780	>	ReentrantLock lock = this.lock; // signal when no workers
1781	>	lock.lock();                    // don't need try/finally
1782	>	termination.signalAll();        // signal when 0 workers
1783	>	lock.unlock();
1784	>	}
1785		return true;
1786	+	}
1787	+	if (!now) {
1788	+	if ((int)(c >> AC_SHIFT) != -parallelism || runState >= 0 ||
1789	+	hasQueuedSubmissions())
1790	+	return false;
1791	+	// Check for unqueued inactive workers. One pass suffices.
1792	+	WorkQueue[] ws = workQueues; WorkQueue w;
1793	+	if (ws != null) {
1794	+	int n = ws.length;
1795	+	for (int i = 1; i < n; i += 2) {
1796	+	if ((w = ws[i]) != null && w.eventCount >= 0)
1797	+	return false;
1798	+	}
1799	+	}
1800	+	}
1801	+	if (U.compareAndSwapLong(this, CTL, c, c | STOP_BIT))
1802	+	startTerminating();
1803		}
1804		}
1805
1806		/**
1807	<	* Controls whether to add spares to maintain parallelism
1807	>	* Initiates termination: Runs three passes through workQueues:
1808	>	* (0) Setting termination status, followed by wakeups of queued
1809	>	* workers; (1) cancelling all tasks; (2) interrupting lagging
1810	>	* threads (likely in external tasks, but possibly also blocked in
1811	>	* joins).  Each pass repeats previous steps because of potential
1812	>	* lagging thread creation.
1813		*/
1814	<	private volatile boolean maintainsParallelism;
1814	>	private void startTerminating() {
1815	>	for (int pass = 0; pass < 3; ++pass) {
1816	>	WorkQueue[] ws = workQueues;
1817	>	if (ws != null) {
1818	>	WorkQueue w; Thread wt;
1819	>	int n = ws.length;
1820	>	for (int j = 0; j < n; ++j) {
1821	>	if ((w = ws[j]) != null) {
1822	>	w.runState = -1;
1823	>	if (pass > 0) {
1824	>	w.cancelAll();
1825	>	if (pass > 1 && (wt = w.owner) != null &&
1826	>	!wt.isInterrupted()) {
1827	>	try {
1828	>	wt.interrupt();
1829	>	} catch (SecurityException ignore) {
1830	>	}
1831	>	}
1832	>	}
1833	>	}
1834	>	}
1835	>	// Wake up workers parked on event queue
1836	>	int i, e; long c; Thread p;
1837	>	while ((i = ((~(e = (int)(c = ctl)) << 1) | 1) & SMASK) < n &&
1838	>	(w = ws[i]) != null &&
1839	>	w.eventCount == (e | INT_SIGN)) {
1840	>	long nc = ((long)(w.nextWait & E_MASK) |
1841	>	((c + AC_UNIT) & AC_MASK) |
1842	>	(c & (TC_MASK|STOP_BIT)));
1843	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1844	>	w.eventCount = (e + E_SEQ) & E_MASK;
1845	>	if ((p = w.parker) != null)
1846	>	U.unpark(p);
1847	>	}
1848	>	}
1849	>	}
1850	>	}
1851	>	}
1852	>
1853	>	// Exported methods
1854
1855		// Constructors
1856
1857		/**
1858	<	* Creates a ForkJoinPool with a pool size equal to the number of
1859	<	* processors available on the system, using the default
1860	<	* ForkJoinWorkerThreadFactory.
1858	>	* Creates a {@code ForkJoinPool} with parallelism equal to {@link
1859	>	* java.lang.Runtime#availableProcessors}, using the {@linkplain
1860	>	* #defaultForkJoinWorkerThreadFactory default thread factory},
1861	>	* no UncaughtExceptionHandler, and non-async LIFO processing mode.
1862		*
1863		* @throws SecurityException if a security manager exists and
1864		*         the caller is not permitted to modify threads
#	Line 353 | Line 1867 | public class ForkJoinPool extends Abstra
1867		*/
1868		public ForkJoinPool() {
1869		this(Runtime.getRuntime().availableProcessors(),
1870	<	defaultForkJoinWorkerThreadFactory);
1870	>	defaultForkJoinWorkerThreadFactory, null, false);
1871		}
1872
1873		/**
1874	<	* Creates a ForkJoinPool with the indicated parallelism level
1875	<	* threads and using the default ForkJoinWorkerThreadFactory.
1874	>	* Creates a {@code ForkJoinPool} with the indicated parallelism
1875	>	* level, the {@linkplain
1876	>	* #defaultForkJoinWorkerThreadFactory default thread factory},
1877	>	* no UncaughtExceptionHandler, and non-async LIFO processing mode.
1878		*
1879	<	* @param parallelism the number of worker threads
1879	>	* @param parallelism the parallelism level
1880		* @throws IllegalArgumentException if parallelism less than or
1881	<	* equal to zero
1881	>	*         equal to zero, or greater than implementation limit
1882		* @throws SecurityException if a security manager exists and
1883		*         the caller is not permitted to modify threads
1884		*         because it does not hold {@link
1885		*         java.lang.RuntimePermission}{@code ("modifyThread")}
1886		*/
1887		public ForkJoinPool(int parallelism) {
1888	<	this(parallelism, defaultForkJoinWorkerThreadFactory);
373	<	}
374	<
375	<	/**
376	<	* Creates a ForkJoinPool with parallelism equal to the number of
377	<	* processors available on the system and using the given
378	<	* ForkJoinWorkerThreadFactory.
379	<	*
380	<	* @param factory the factory for creating new threads
381	<	* @throws NullPointerException if factory is null
382	<	* @throws SecurityException if a security manager exists and
383	<	*         the caller is not permitted to modify threads
384	<	*         because it does not hold {@link
385	<	*         java.lang.RuntimePermission}{@code ("modifyThread")}
386	<	*/
387	<	public ForkJoinPool(ForkJoinWorkerThreadFactory factory) {
388	<	this(Runtime.getRuntime().availableProcessors(), factory);
1888	>	this(parallelism, defaultForkJoinWorkerThreadFactory, null, false);
1889		}
1890
1891		/**
1892	<	* Creates a ForkJoinPool with the given parallelism and factory.
1892	>	* Creates a {@code ForkJoinPool} with the given parameters.
1893		*
1894	<	* @param parallelism the targeted number of worker threads
1895	<	* @param factory the factory for creating new threads
1894	>	* @param parallelism the parallelism level. For default value,
1895	>	* use {@link java.lang.Runtime#availableProcessors}.
1896	>	* @param factory the factory for creating new threads. For default value,
1897	>	* use {@link #defaultForkJoinWorkerThreadFactory}.
1898	>	* @param handler the handler for internal worker threads that
1899	>	* terminate due to unrecoverable errors encountered while executing
1900	>	* tasks. For default value, use {@code null}.
1901	>	* @param asyncMode if true,
1902	>	* establishes local first-in-first-out scheduling mode for forked
1903	>	* tasks that are never joined. This mode may be more appropriate
1904	>	* than default locally stack-based mode in applications in which
1905	>	* worker threads only process event-style asynchronous tasks.
1906	>	* For default value, use {@code false}.
1907		* @throws IllegalArgumentException if parallelism less than or
1908	<	* equal to zero, or greater than implementation limit
1909	<	* @throws NullPointerException if factory is null
1908	>	*         equal to zero, or greater than implementation limit
1909	>	* @throws NullPointerException if the factory is null
1910		* @throws SecurityException if a security manager exists and
1911		*         the caller is not permitted to modify threads
1912		*         because it does not hold {@link
1913		*         java.lang.RuntimePermission}{@code ("modifyThread")}
1914		*/
1915	<	public ForkJoinPool(int parallelism, ForkJoinWorkerThreadFactory factory) {
1916	<	if (parallelism <= 0 || parallelism > MAX_THREADS)
1917	<	throw new IllegalArgumentException();
1915	>	public ForkJoinPool(int parallelism,
1916	>	ForkJoinWorkerThreadFactory factory,
1917	>	Thread.UncaughtExceptionHandler handler,
1918	>	boolean asyncMode) {
1919	>	checkPermission();
1920		if (factory == null)
1921		throw new NullPointerException();
1922	<	checkPermission();
1923	<	this.factory = factory;
1922	>	if (parallelism <= 0 || parallelism > MAX_ID)
1923	>	throw new IllegalArgumentException();
1924		this.parallelism = parallelism;
1925	<	this.maxPoolSize = MAX_THREADS;
1926	<	this.maintainsParallelism = true;
1927	<	this.poolNumber = poolNumberGenerator.incrementAndGet();
1928	<	this.workerLock = new ReentrantLock();
1929	<	this.termination = workerLock.newCondition();
1925	>	this.factory = factory;
1926	>	this.ueh = handler;
1927	>	this.localMode = asyncMode ? FIFO_QUEUE : LIFO_QUEUE;
1928	>	this.nextPoolIndex = 1;
1929	>	long np = (long)(-parallelism); // offset ctl counts
1930	>	this.ctl = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK);
1931	>	// initialize workQueues array with room for 2*parallelism if possible
1932	>	int n = parallelism << 1;
1933	>	if (n >= MAX_ID)
1934	>	n = MAX_ID;
1935	>	else { // See Hackers Delight, sec 3.2, where n < (1 << 16)
1936	>	n |= n >>> 1; n |= n >>> 2; n |= n >>> 4; n |= n >>> 8;
1937	>	}
1938	>	this.workQueues = new WorkQueue[(n + 1) << 1];
1939	>	ReentrantLock lck = this.lock = new ReentrantLock();
1940	>	this.termination = lck.newCondition();
1941		this.stealCount = new AtomicLong();
1942	<	this.submissionQueue = new LinkedTransferQueue<ForkJoinTask<?>>();
1943	<	// worker array and workers are lazily constructed
1944	<	}
1945	<
1946	<	/**
1947	<	* Creates a new worker thread using factory.
1948	<	*
1949	<	* @param index the index to assign worker
1950	<	* @return new worker, or null of factory failed
427	<	*/
428	<	private ForkJoinWorkerThread createWorker(int index) {
429	<	Thread.UncaughtExceptionHandler h = ueh;
430	<	ForkJoinWorkerThread w = factory.newThread(this);
431	<	if (w != null) {
432	<	w.poolIndex = index;
433	<	w.setDaemon(true);
434	<	w.setAsyncMode(locallyFifo);
435	<	w.setName("ForkJoinPool-" + poolNumber + "-worker-" + index);
436	<	if (h != null)
437	<	w.setUncaughtExceptionHandler(h);
438	<	}
439	<	return w;
440	<	}
441	<
442	<	/**
443	<	* Returns a good size for worker array given pool size.
444	<	* Currently requires size to be a power of two.
445	<	*/
446	<	private static int arraySizeFor(int poolSize) {
447	<	return (poolSize <= 1) ? 1 :
448	<	(1 << (32 - Integer.numberOfLeadingZeros(poolSize-1)));
449	<	}
450	<
451	<	/**
452	<	* Creates or resizes array if necessary to hold newLength.
453	<	* Call only under exclusion.
454	<	*
455	<	* @return the array
456	<	*/
457	<	private ForkJoinWorkerThread[] ensureWorkerArrayCapacity(int newLength) {
458	<	ForkJoinWorkerThread[] ws = workers;
459	<	if (ws == null)
460	<	return workers = new ForkJoinWorkerThread[arraySizeFor(newLength)];
461	<	else if (newLength > ws.length)
462	<	return workers = Arrays.copyOf(ws, arraySizeFor(newLength));
463	<	else
464	<	return ws;
465	<	}
466	<
467	<	/**
468	<	* Tries to shrink workers into smaller array after one or more terminate.
469	<	*/
470	<	private void tryShrinkWorkerArray() {
471	<	ForkJoinWorkerThread[] ws = workers;
472	<	if (ws != null) {
473	<	int len = ws.length;
474	<	int last = len - 1;
475	<	while (last >= 0 && ws[last] == null)
476	<	--last;
477	<	int newLength = arraySizeFor(last+1);
478	<	if (newLength < len)
479	<	workers = Arrays.copyOf(ws, newLength);
480	<	}
481	<	}
482	<
483	<	/**
484	<	* Initializes workers if necessary.
485	<	*/
486	<	final void ensureWorkerInitialization() {
487	<	ForkJoinWorkerThread[] ws = workers;
488	<	if (ws == null) {
489	<	final ReentrantLock lock = this.workerLock;
490	<	lock.lock();
491	<	try {
492	<	ws = workers;
493	<	if (ws == null) {
494	<	int ps = parallelism;
495	<	ws = ensureWorkerArrayCapacity(ps);
496	<	for (int i = 0; i < ps; ++i) {
497	<	ForkJoinWorkerThread w = createWorker(i);
498	<	if (w != null) {
499	<	ws[i] = w;
500	<	w.start();
501	<	updateWorkerCount(1);
502	<	}
503	<	}
504	<	}
505	<	} finally {
506	<	lock.unlock();
507	<	}
508	<	}
509	<	}
510	<
511	<	/**
512	<	* Worker creation and startup for threads added via setParallelism.
513	<	*/
514	<	private void createAndStartAddedWorkers() {
515	<	resumeAllSpares();  // Allow spares to convert to nonspare
516	<	int ps = parallelism;
517	<	ForkJoinWorkerThread[] ws = ensureWorkerArrayCapacity(ps);
518	<	int len = ws.length;
519	<	// Sweep through slots, to keep lowest indices most populated
520	<	int k = 0;
521	<	while (k < len) {
522	<	if (ws[k] != null) {
523	<	++k;
524	<	continue;
525	<	}
526	<	int s = workerCounts;
527	<	int tc = totalCountOf(s);
528	<	int rc = runningCountOf(s);
529	<	if (rc >= ps || tc >= ps)
530	<	break;
531	<	if (casWorkerCounts (s, workerCountsFor(tc+1, rc+1))) {
532	<	ForkJoinWorkerThread w = createWorker(k);
533	<	if (w != null) {
534	<	ws[k++] = w;
535	<	w.start();
536	<	}
537	<	else {
538	<	updateWorkerCount(-1); // back out on failed creation
539	<	break;
540	<	}
541	<	}
542	<	}
1942	>	this.nextWorkerNumber = new AtomicInteger();
1943	>	StringBuilder sb = new StringBuilder("ForkJoinPool-");
1944	>	sb.append(poolNumberGenerator.incrementAndGet());
1945	>	sb.append("-worker-");
1946	>	this.workerNamePrefix = sb.toString();
1947	>	// Create initial submission queue
1948	>	WorkQueue sq = tryAddSharedQueue(0);
1949	>	if (sq != null)
1950	>	sq.growArray(false);
1951		}
1952
1953		// Execution methods
1954
1955		/**
548	–	* Common code for execute, invoke and submit
549	–	*/
550	–	private <T> void doSubmit(ForkJoinTask<T> task) {
551	–	if (task == null)
552	–	throw new NullPointerException();
553	–	if (isShutdown())
554	–	throw new RejectedExecutionException();
555	–	if (workers == null)
556	–	ensureWorkerInitialization();
557	–	submissionQueue.offer(task);
558	–	signalIdleWorkers();
559	–	}
560	–
561	–	/**
1956		* Performs the given task, returning its result upon completion.
1957	+	* If the computation encounters an unchecked Exception or Error,
1958	+	* it is rethrown as the outcome of this invocation.  Rethrown
1959	+	* exceptions behave in the same way as regular exceptions, but,
1960	+	* when possible, contain stack traces (as displayed for example
1961	+	* using {@code ex.printStackTrace()}) of both the current thread
1962	+	* as well as the thread actually encountering the exception;
1963	+	* minimally only the latter.
1964		*
1965		* @param task the task
1966		* @return the task's result
1967	<	* @throws NullPointerException if task is null
1968	<	* @throws RejectedExecutionException if pool is shut down
1967	>	* @throws NullPointerException if the task is null
1968	>	* @throws RejectedExecutionException if the task cannot be
1969	>	*         scheduled for execution
1970		*/
1971		public <T> T invoke(ForkJoinTask<T> task) {
1972		doSubmit(task);
#	Line 575 | Line 1977 | public class ForkJoinPool extends Abstra
1977		* Arranges for (asynchronous) execution of the given task.
1978		*
1979		* @param task the task
1980	<	* @throws NullPointerException if task is null
1981	<	* @throws RejectedExecutionException if pool is shut down
1980	>	* @throws NullPointerException if the task is null
1981	>	* @throws RejectedExecutionException if the task cannot be
1982	>	*         scheduled for execution
1983		*/
1984	<	public <T> void execute(ForkJoinTask<T> task) {
1984	>	public void execute(ForkJoinTask<?> task) {
1985		doSubmit(task);
1986		}
1987
1988		// AbstractExecutorService methods
1989
1990	+	/**
1991	+	* @throws NullPointerException if the task is null
1992	+	* @throws RejectedExecutionException if the task cannot be
1993	+	*         scheduled for execution
1994	+	*/
1995		public void execute(Runnable task) {
1996	+	if (task == null)
1997	+	throw new NullPointerException();
1998		ForkJoinTask<?> job;
1999	<	if (task instanceof AdaptedCallable) // avoid re-wrap
2000	<	job = (AdaptedCallable<?>)task;
591	<	else if (task instanceof AdaptedRunnable)
592	<	job = (AdaptedRunnable<?>)task;
593	<	else
594	<	job = new AdaptedRunnable<Void>(task, null);
595	<	doSubmit(job);
596	<	}
597	<
598	<	public <T> ForkJoinTask<T> submit(Callable<T> task) {
599	<	ForkJoinTask<T> job = new AdaptedCallable<T>(task);
600	<	doSubmit(job);
601	<	return job;
602	<	}
603	<
604	<	public <T> ForkJoinTask<T> submit(Runnable task, T result) {
605	<	ForkJoinTask<T> job = new AdaptedRunnable<T>(task, result);
606	<	doSubmit(job);
607	<	return job;
608	<	}
609	<
610	<	public ForkJoinTask<?> submit(Runnable task) {
611	<	ForkJoinTask<?> job;
612	<	if (task instanceof AdaptedCallable) // avoid re-wrap
613	<	job = (AdaptedCallable<?>)task;
614	<	else if (task instanceof AdaptedRunnable)
615	<	job = (AdaptedRunnable<?>)task;
1999	>	if (task instanceof ForkJoinTask<?>) // avoid re-wrap
2000	>	job = (ForkJoinTask<?>) task;
2001		else
2002	<	job = new AdaptedRunnable<Void>(task, null);
2002	>	job = ForkJoinTask.adapt(task, null);
2003		doSubmit(job);
619	–	return job;
2004		}
2005
2006		/**
#	Line 624 | Line 2008 | public class ForkJoinPool extends Abstra
2008		*
2009		* @param task the task to submit
2010		* @return the task
2011	+	* @throws NullPointerException if the task is null
2012		* @throws RejectedExecutionException if the task cannot be
2013		*         scheduled for execution
629	–	* @throws NullPointerException if the task is null
2014		*/
2015		public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) {
2016		doSubmit(task);
#	Line 634 | Line 2018 | public class ForkJoinPool extends Abstra
2018		}
2019
2020		/**
2021	<	* Adaptor for Runnables. This implements RunnableFuture
2022	<	* to be compliant with AbstractExecutorService constraints.
2021	>	* @throws NullPointerException if the task is null
2022	>	* @throws RejectedExecutionException if the task cannot be
2023	>	*         scheduled for execution
2024		*/
2025	<	static final class AdaptedRunnable<T> extends ForkJoinTask<T>
2026	<	implements RunnableFuture<T> {
2027	<	final Runnable runnable;
2028	<	final T resultOnCompletion;
2029	<	T result;
2030	<	AdaptedRunnable(Runnable runnable, T result) {
646	<	if (runnable == null) throw new NullPointerException();
647	<	this.runnable = runnable;
648	<	this.resultOnCompletion = result;
649	<	}
650	<	public T getRawResult() { return result; }
651	<	public void setRawResult(T v) { result = v; }
652	<	public boolean exec() {
653	<	runnable.run();
654	<	result = resultOnCompletion;
655	<	return true;
656	<	}
657	<	public void run() { invoke(); }
658	<	private static final long serialVersionUID = 5232453952276885070L;
2025	>	public <T> ForkJoinTask<T> submit(Callable<T> task) {
2026	>	if (task == null)
2027	>	throw new NullPointerException();
2028	>	ForkJoinTask<T> job = ForkJoinTask.adapt(task);
2029	>	doSubmit(job);
2030	>	return job;
2031		}
2032
2033		/**
2034	<	* Adaptor for Callables
2035	<	*/
2036	<	static final class AdaptedCallable<T> extends ForkJoinTask<T>
2037	<	implements RunnableFuture<T> {
2038	<	final Callable<T> callable;
2039	<	T result;
2040	<	AdaptedCallable(Callable<T> callable) {
2041	<	if (callable == null) throw new NullPointerException();
2042	<	this.callable = callable;
2043	<	}
672	<	public T getRawResult() { return result; }
673	<	public void setRawResult(T v) { result = v; }
674	<	public boolean exec() {
675	<	try {
676	<	result = callable.call();
677	<	return true;
678	<	} catch (Error err) {
679	<	throw err;
680	<	} catch (RuntimeException rex) {
681	<	throw rex;
682	<	} catch (Exception ex) {
683	<	throw new RuntimeException(ex);
684	<	}
685	<	}
686	<	public void run() { invoke(); }
687	<	private static final long serialVersionUID = 2838392045355241008L;
2034	>	* @throws NullPointerException if the task is null
2035	>	* @throws RejectedExecutionException if the task cannot be
2036	>	*         scheduled for execution
2037	>	*/
2038	>	public <T> ForkJoinTask<T> submit(Runnable task, T result) {
2039	>	if (task == null)
2040	>	throw new NullPointerException();
2041	>	ForkJoinTask<T> job = ForkJoinTask.adapt(task, result);
2042	>	doSubmit(job);
2043	>	return job;
2044		}
2045
2046	+	/**
2047	+	* @throws NullPointerException if the task is null
2048	+	* @throws RejectedExecutionException if the task cannot be
2049	+	*         scheduled for execution
2050	+	*/
2051	+	public ForkJoinTask<?> submit(Runnable task) {
2052	+	if (task == null)
2053	+	throw new NullPointerException();
2054	+	ForkJoinTask<?> job;
2055	+	if (task instanceof ForkJoinTask<?>) // avoid re-wrap
2056	+	job = (ForkJoinTask<?>) task;
2057	+	else
2058	+	job = ForkJoinTask.adapt(task, null);
2059	+	doSubmit(job);
2060	+	return job;
2061	+	}
2062	+
2063	+	/**
2064	+	* @throws NullPointerException       {@inheritDoc}
2065	+	* @throws RejectedExecutionException {@inheritDoc}
2066	+	*/
2067		public <T> List<Future<T>> invokeAll(Collection<? extends Callable<T>> tasks) {
2068		ArrayList<ForkJoinTask<T>> forkJoinTasks =
2069		new ArrayList<ForkJoinTask<T>>(tasks.size());
2070		for (Callable<T> task : tasks)
2071	<	forkJoinTasks.add(new AdaptedCallable<T>(task));
2071	>	forkJoinTasks.add(ForkJoinTask.adapt(task));
2072		invoke(new InvokeAll<T>(forkJoinTasks));
2073
2074		@SuppressWarnings({"unchecked", "rawtypes"})
2075	<	List<Future<T>> futures = (List<Future<T>>) (List) forkJoinTasks;
2075	>	List<Future<T>> futures = (List<Future<T>>) (List) forkJoinTasks;
2076		return futures;
2077		}
2078
#	Line 709 | Line 2086 | public class ForkJoinPool extends Abstra
2086		private static final long serialVersionUID = -7914297376763021607L;
2087		}
2088
712	–	// Configuration and status settings and queries
713	–
2089		/**
2090		* Returns the factory used for constructing new workers.
2091		*
#	Line 724 | Line 2099 | public class ForkJoinPool extends Abstra
2099		* Returns the handler for internal worker threads that terminate
2100		* due to unrecoverable errors encountered while executing tasks.
2101		*
2102	<	* @return the handler, or null if none
2102	>	* @return the handler, or {@code null} if none
2103		*/
2104		public Thread.UncaughtExceptionHandler getUncaughtExceptionHandler() {
2105	<	Thread.UncaughtExceptionHandler h;
731	<	final ReentrantLock lock = this.workerLock;
732	<	lock.lock();
733	<	try {
734	<	h = ueh;
735	<	} finally {
736	<	lock.unlock();
737	<	}
738	<	return h;
2105	>	return ueh;
2106		}
2107
2108		/**
2109	<	* Sets the handler for internal worker threads that terminate due
743	<	* to unrecoverable errors encountered while executing tasks.
744	<	* Unless set, the current default or ThreadGroup handler is used
745	<	* as handler.
2109	>	* Returns the targeted parallelism level of this pool.
2110		*
2111	<	* @param h the new handler
748	<	* @return the old handler, or null if none
749	<	* @throws SecurityException if a security manager exists and
750	<	*         the caller is not permitted to modify threads
751	<	*         because it does not hold {@link
752	<	*         java.lang.RuntimePermission}{@code ("modifyThread")}
753	<	*/
754	<	public Thread.UncaughtExceptionHandler
755	<	setUncaughtExceptionHandler(Thread.UncaughtExceptionHandler h) {
756	<	checkPermission();
757	<	Thread.UncaughtExceptionHandler old = null;
758	<	final ReentrantLock lock = this.workerLock;
759	<	lock.lock();
760	<	try {
761	<	old = ueh;
762	<	ueh = h;
763	<	ForkJoinWorkerThread[] ws = workers;
764	<	if (ws != null) {
765	<	for (int i = 0; i < ws.length; ++i) {
766	<	ForkJoinWorkerThread w = ws[i];
767	<	if (w != null)
768	<	w.setUncaughtExceptionHandler(h);
769	<	}
770	<	}
771	<	} finally {
772	<	lock.unlock();
773	<	}
774	<	return old;
775	<	}
776	<
777	<
778	<	/**
779	<	* Sets the target parallelism level of this pool.
780	<	*
781	<	* @param parallelism the target parallelism
782	<	* @throws IllegalArgumentException if parallelism less than or
783	<	* equal to zero or greater than maximum size bounds
784	<	* @throws SecurityException if a security manager exists and
785	<	*         the caller is not permitted to modify threads
786	<	*         because it does not hold {@link
787	<	*         java.lang.RuntimePermission}{@code ("modifyThread")}
788	<	*/
789	<	public void setParallelism(int parallelism) {
790	<	checkPermission();
791	<	if (parallelism <= 0 || parallelism > maxPoolSize)
792	<	throw new IllegalArgumentException();
793	<	final ReentrantLock lock = this.workerLock;
794	<	lock.lock();
795	<	try {
796	<	if (!isTerminating()) {
797	<	int p = this.parallelism;
798	<	this.parallelism = parallelism;
799	<	if (parallelism > p)
800	<	createAndStartAddedWorkers();
801	<	else
802	<	trimSpares();
803	<	}
804	<	} finally {
805	<	lock.unlock();
806	<	}
807	<	signalIdleWorkers();
808	<	}
809	<
810	<	/**
811	<	* Returns the targeted number of worker threads in this pool.
812	<	*
813	<	* @return the targeted number of worker threads in this pool
2111	>	* @return the targeted parallelism level of this pool
2112		*/
2113		public int getParallelism() {
2114		return parallelism;
#	Line 818 | Line 2116 | public class ForkJoinPool extends Abstra
2116
2117		/**
2118		* Returns the number of worker threads that have started but not
2119	<	* yet terminated.  This result returned by this method may differ
2120	<	* from {@code getParallelism} when threads are created to
2119	>	* yet terminated.  The result returned by this method may differ
2120	>	* from {@link #getParallelism} when threads are created to
2121		* maintain parallelism when others are cooperatively blocked.
2122		*
2123		* @return the number of worker threads
2124		*/
2125		public int getPoolSize() {
2126	<	return totalCountOf(workerCounts);
829	<	}
830	<
831	<	/**
832	<	* Returns the maximum number of threads allowed to exist in the
833	<	* pool, even if there are insufficient unblocked running threads.
834	<	*
835	<	* @return the maximum
836	<	*/
837	<	public int getMaximumPoolSize() {
838	<	return maxPoolSize;
839	<	}
840	<
841	<	/**
842	<	* Sets the maximum number of threads allowed to exist in the
843	<	* pool, even if there are insufficient unblocked running threads.
844	<	* Setting this value has no effect on current pool size. It
845	<	* controls construction of new threads.
846	<	*
847	<	* @throws IllegalArgumentException if negative or greater then
848	<	* internal implementation limit
849	<	*/
850	<	public void setMaximumPoolSize(int newMax) {
851	<	if (newMax < 0 || newMax > MAX_THREADS)
852	<	throw new IllegalArgumentException();
853	<	maxPoolSize = newMax;
854	<	}
855	<
856	<
857	<	/**
858	<	* Returns true if this pool dynamically maintains its target
859	<	* parallelism level. If false, new threads are added only to
860	<	* avoid possible starvation.
861	<	* This setting is by default true.
862	<	*
863	<	* @return true if maintains parallelism
864	<	*/
865	<	public boolean getMaintainsParallelism() {
866	<	return maintainsParallelism;
2126	>	return parallelism + (short)(ctl >>> TC_SHIFT);
2127		}
2128
2129		/**
2130	<	* Sets whether this pool dynamically maintains its target
871	<	* parallelism level. If false, new threads are added only to
872	<	* avoid possible starvation.
873	<	*
874	<	* @param enable true to maintains parallelism
875	<	*/
876	<	public void setMaintainsParallelism(boolean enable) {
877	<	maintainsParallelism = enable;
878	<	}
879	<
880	<	/**
881	<	* Establishes local first-in-first-out scheduling mode for forked
882	<	* tasks that are never joined. This mode may be more appropriate
883	<	* than default locally stack-based mode in applications in which
884	<	* worker threads only process asynchronous tasks.  This method is
885	<	* designed to be invoked only when pool is quiescent, and
886	<	* typically only before any tasks are submitted. The effects of
887	<	* invocations at other times may be unpredictable.
888	<	*
889	<	* @param async if true, use locally FIFO scheduling
890	<	* @return the previous mode
891	<	*/
892	<	public boolean setAsyncMode(boolean async) {
893	<	boolean oldMode = locallyFifo;
894	<	locallyFifo = async;
895	<	ForkJoinWorkerThread[] ws = workers;
896	<	if (ws != null) {
897	<	for (int i = 0; i < ws.length; ++i) {
898	<	ForkJoinWorkerThread t = ws[i];
899	<	if (t != null)
900	<	t.setAsyncMode(async);
901	<	}
902	<	}
903	<	return oldMode;
904	<	}
905	<
906	<	/**
907	<	* Returns true if this pool uses local first-in-first-out
2130	>	* Returns {@code true} if this pool uses local first-in-first-out
2131		* scheduling mode for forked tasks that are never joined.
2132		*
2133	<	* @return true if this pool uses async mode
2133	>	* @return {@code true} if this pool uses async mode
2134		*/
2135		public boolean getAsyncMode() {
2136	<	return locallyFifo;
2136	>	return localMode != 0;
2137		}
2138
2139		/**
2140		* Returns an estimate of the number of worker threads that are
2141		* not blocked waiting to join tasks or for other managed
2142	<	* synchronization.
2142	>	* synchronization. This method may overestimate the
2143	>	* number of running threads.
2144		*
2145		* @return the number of worker threads
2146		*/
2147		public int getRunningThreadCount() {
2148	<	return runningCountOf(workerCounts);
2148	>	int rc = 0;
2149	>	WorkQueue[] ws; WorkQueue w;
2150	>	if ((ws = workQueues) != null) {
2151	>	int n = ws.length;
2152	>	for (int i = 1; i < n; i += 2) {
2153	>	Thread.State s; ForkJoinWorkerThread wt;
2154	>	if ((w = ws[i]) != null && (wt = w.owner) != null &&
2155	>	w.eventCount >= 0 &&
2156	>	(s = wt.getState()) != Thread.State.BLOCKED &&
2157	>	s != Thread.State.WAITING &&
2158	>	s != Thread.State.TIMED_WAITING)
2159	>	++rc;
2160	>	}
2161	>	}
2162	>	return rc;
2163		}
2164
2165		/**
#	Line 932 | Line 2170 | public class ForkJoinPool extends Abstra
2170		* @return the number of active threads
2171		*/
2172		public int getActiveThreadCount() {
2173	<	return activeCountOf(runControl);
2174	<	}
937	<
938	<	/**
939	<	* Returns an estimate of the number of threads that are currently
940	<	* idle waiting for tasks. This method may underestimate the
941	<	* number of idle threads.
942	<	*
943	<	* @return the number of idle threads
944	<	*/
945	<	final int getIdleThreadCount() {
946	<	int c = runningCountOf(workerCounts) - activeCountOf(runControl);
947	<	return (c <= 0) ? 0 : c;
2173	>	int r = parallelism + (int)(ctl >> AC_SHIFT);
2174	>	return (r <= 0) ? 0 : r; // suppress momentarily negative values
2175		}
2176
2177		/**
2178	<	* Returns true if all worker threads are currently idle. An idle
2179	<	* worker is one that cannot obtain a task to execute because none
2180	<	* are available to steal from other threads, and there are no
2181	<	* pending submissions to the pool. This method is conservative;
2182	<	* it might not return true immediately upon idleness of all
2183	<	* threads, but will eventually become true if threads remain
2184	<	* inactive.
2178	>	* Returns {@code true} if all worker threads are currently idle.
2179	>	* An idle worker is one that cannot obtain a task to execute
2180	>	* because none are available to steal from other threads, and
2181	>	* there are no pending submissions to the pool. This method is
2182	>	* conservative; it might not return {@code true} immediately upon
2183	>	* idleness of all threads, but will eventually become true if
2184	>	* threads remain inactive.
2185		*
2186	<	* @return true if all threads are currently idle
2186	>	* @return {@code true} if all threads are currently idle
2187		*/
2188		public boolean isQuiescent() {
2189	<	return activeCountOf(runControl) == 0;
2189	>	return (int)(ctl >> AC_SHIFT) + parallelism == 0;
2190		}
2191
2192		/**
#	Line 974 | Line 2201 | public class ForkJoinPool extends Abstra
2201		* @return the number of steals
2202		*/
2203		public long getStealCount() {
2204	<	return stealCount.get();
2205	<	}
2206	<
2207	<	/**
2208	<	* Accumulates steal count from a worker.
2209	<	* Call only when worker known to be idle.
2210	<	*/
2211	<	private void updateStealCount(ForkJoinWorkerThread w) {
2212	<	int sc = w.getAndClearStealCount();
2213	<	if (sc != 0)
987	<	stealCount.addAndGet(sc);
2204	>	long count = stealCount.get();
2205	>	WorkQueue[] ws; WorkQueue w;
2206	>	if ((ws = workQueues) != null) {
2207	>	int n = ws.length;
2208	>	for (int i = 1; i < n; i += 2) {
2209	>	if ((w = ws[i]) != null)
2210	>	count += w.totalSteals;
2211	>	}
2212	>	}
2213	>	return count;
2214		}
2215
2216		/**
#	Line 999 | Line 2225 | public class ForkJoinPool extends Abstra
2225		*/
2226		public long getQueuedTaskCount() {
2227		long count = 0;
2228	<	ForkJoinWorkerThread[] ws = workers;
2229	<	if (ws != null) {
2230	<	for (int i = 0; i < ws.length; ++i) {
2231	<	ForkJoinWorkerThread t = ws[i];
2232	<	if (t != null)
2233	<	count += t.getQueueSize();
2228	>	WorkQueue[] ws; WorkQueue w;
2229	>	if ((ws = workQueues) != null) {
2230	>	int n = ws.length;
2231	>	for (int i = 1; i < n; i += 2) {
2232	>	if ((w = ws[i]) != null)
2233	>	count += w.queueSize();
2234		}
2235		}
2236		return count;
2237		}
2238
2239		/**
2240	<	* Returns an estimate of the number tasks submitted to this pool
2241	<	* that have not yet begun executing. This method takes time
2242	<	* proportional to the number of submissions.
2240	>	* Returns an estimate of the number of tasks submitted to this
2241	>	* pool that have not yet begun executing.  This method may take
2242	>	* time proportional to the number of submissions.
2243		*
2244		* @return the number of queued submissions
2245		*/
2246		public int getQueuedSubmissionCount() {
2247	<	return submissionQueue.size();
2247	>	int count = 0;
2248	>	WorkQueue[] ws; WorkQueue w;
2249	>	if ((ws = workQueues) != null) {
2250	>	int n = ws.length;
2251	>	for (int i = 0; i < n; i += 2) {
2252	>	if ((w = ws[i]) != null)
2253	>	count += w.queueSize();
2254	>	}
2255	>	}
2256	>	return count;
2257		}
2258
2259		/**
2260	<	* Returns true if there are any tasks submitted to this pool
2261	<	* that have not yet begun executing.
2260	>	* Returns {@code true} if there are any tasks submitted to this
2261	>	* pool that have not yet begun executing.
2262		*
2263		* @return {@code true} if there are any queued submissions
2264		*/
2265		public boolean hasQueuedSubmissions() {
2266	<	return !submissionQueue.isEmpty();
2266	>	WorkQueue[] ws; WorkQueue w;
2267	>	if ((ws = workQueues) != null) {
2268	>	int n = ws.length;
2269	>	for (int i = 0; i < n; i += 2) {
2270	>	if ((w = ws[i]) != null && w.queueSize() != 0)
2271	>	return true;
2272	>	}
2273	>	}
2274	>	return false;
2275		}
2276
2277		/**
#	Line 1036 | Line 2279 | public class ForkJoinPool extends Abstra
2279		* available.  This method may be useful in extensions to this
2280		* class that re-assign work in systems with multiple pools.
2281		*
2282	<	* @return the next submission, or null if none
2282	>	* @return the next submission, or {@code null} if none
2283		*/
2284		protected ForkJoinTask<?> pollSubmission() {
2285	<	return submissionQueue.poll();
2285	>	WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2286	>	if ((ws = workQueues) != null) {
2287	>	int n = ws.length;
2288	>	for (int i = 0; i < n; i += 2) {
2289	>	if ((w = ws[i]) != null && (t = w.poll()) != null)
2290	>	return t;
2291	>	}
2292	>	}
2293	>	return null;
2294		}
2295
2296		/**
2297		* Removes all available unexecuted submitted and forked tasks
2298		* from scheduling queues and adds them to the given collection,
2299		* without altering their execution status. These may include
2300	<	* artificially generated or wrapped tasks. This method is designed
2301	<	* to be invoked only when the pool is known to be
2300	>	* artificially generated or wrapped tasks. This method is
2301	>	* designed to be invoked only when the pool is known to be
2302		* quiescent. Invocations at other times may not remove all
2303		* tasks. A failure encountered while attempting to add elements
2304		* to collection {@code c} may result in elements being in
#	Line 1059 | Line 2310 | public class ForkJoinPool extends Abstra
2310		* @param c the collection to transfer elements into
2311		* @return the number of elements transferred
2312		*/
2313	<	protected int drainTasksTo(Collection<ForkJoinTask<?>> c) {
2314	<	int n = submissionQueue.drainTo(c);
2315	<	ForkJoinWorkerThread[] ws = workers;
2316	<	if (ws != null) {
2317	<	for (int i = 0; i < ws.length; ++i) {
2318	<	ForkJoinWorkerThread w = ws[i];
2319	<	if (w != null)
2320	<	n += w.drainTasksTo(c);
2313	>	protected int drainTasksTo(Collection<? super ForkJoinTask<?>> c) {
2314	>	int count = 0;
2315	>	WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2316	>	if ((ws = workQueues) != null) {
2317	>	int n = ws.length;
2318	>	for (int i = 0; i < n; ++i) {
2319	>	if ((w = ws[i]) != null) {
2320	>	while ((t = w.poll()) != null) {
2321	>	c.add(t);
2322	>	++count;
2323	>	}
2324	>	}
2325		}
2326		}
2327	<	return n;
2327	>	return count;
2328		}
2329
2330		/**
#	Line 1080 | Line 2335 | public class ForkJoinPool extends Abstra
2335		* @return a string identifying this pool, as well as its state
2336		*/
2337		public String toString() {
1083	–	int ps = parallelism;
1084	–	int wc = workerCounts;
1085	–	int rc = runControl;
2338		long st = getStealCount();
2339		long qt = getQueuedTaskCount();
2340		long qs = getQueuedSubmissionCount();
2341	+	int rc = getRunningThreadCount();
2342	+	int pc = parallelism;
2343	+	long c = ctl;
2344	+	int tc = pc + (short)(c >>> TC_SHIFT);
2345	+	int ac = pc + (int)(c >> AC_SHIFT);
2346	+	if (ac < 0) // ignore transient negative
2347	+	ac = 0;
2348	+	String level;
2349	+	if ((c & STOP_BIT) != 0)
2350	+	level = (tc == 0) ? "Terminated" : "Terminating";
2351	+	else
2352	+	level = runState < 0 ? "Shutting down" : "Running";
2353		return super.toString() +
2354	<	"[" + runStateToString(runStateOf(rc)) +
2355	<	", parallelism = " + ps +
2356	<	", size = " + totalCountOf(wc) +
2357	<	", active = " + activeCountOf(rc) +
2358	<	", running = " + runningCountOf(wc) +
2354	>	"[" + level +
2355	>	", parallelism = " + pc +
2356	>	", size = " + tc +
2357	>	", active = " + ac +
2358	>	", running = " + rc +
2359		", steals = " + st +
2360		", tasks = " + qt +
2361		", submissions = " + qs +
2362		"]";
2363		}
2364
1101	–	private static String runStateToString(int rs) {
1102	–	switch(rs) {
1103	–	case RUNNING: return "Running";
1104	–	case SHUTDOWN: return "Shutting down";
1105	–	case TERMINATING: return "Terminating";
1106	–	case TERMINATED: return "Terminated";
1107	–	default: throw new Error("Unknown run state");
1108	–	}
1109	–	}
1110	–
1111	–	// lifecycle control
1112	–
2365		/**
2366		* Initiates an orderly shutdown in which previously submitted
2367		* tasks are executed, but no new tasks will be accepted.
#	Line 1124 | Line 2376 | public class ForkJoinPool extends Abstra
2376		*/
2377		public void shutdown() {
2378		checkPermission();
2379	<	transitionRunStateTo(SHUTDOWN);
2380	<	if (canTerminateOnShutdown(runControl))
1129	<	terminateOnShutdown();
2379	>	enableShutdown();
2380	>	tryTerminate(false);
2381		}
2382
2383		/**
2384	<	* Attempts to stop all actively executing tasks, and cancels all
2385	<	* waiting tasks.  Tasks that are in the process of being
2386	<	* submitted or executed concurrently during the course of this
2387	<	* method may or may not be rejected. Unlike some other executors,
2388	<	* this method cancels rather than collects non-executed tasks
2389	<	* upon termination, so always returns an empty list. However, you
2390	<	* can use method {@code drainTasksTo} before invoking this
2391	<	* method to transfer unexecuted tasks to another collection.
2384	>	* Attempts to cancel and/or stop all tasks, and reject all
2385	>	* subsequently submitted tasks.  Tasks that are in the process of
2386	>	* being submitted or executed concurrently during the course of
2387	>	* this method may or may not be rejected. This method cancels
2388	>	* both existing and unexecuted tasks, in order to permit
2389	>	* termination in the presence of task dependencies. So the method
2390	>	* always returns an empty list (unlike the case for some other
2391	>	* Executors).
2392		*
2393		* @return an empty list
2394		* @throws SecurityException if a security manager exists and
#	Line 1147 | Line 2398 | public class ForkJoinPool extends Abstra
2398		*/
2399		public List<Runnable> shutdownNow() {
2400		checkPermission();
2401	<	terminate();
2401	>	enableShutdown();
2402	>	tryTerminate(true);
2403		return Collections.emptyList();
2404		}
2405
#	Line 1157 | Line 2409 | public class ForkJoinPool extends Abstra
2409		* @return {@code true} if all tasks have completed following shut down
2410		*/
2411		public boolean isTerminated() {
2412	<	return runStateOf(runControl) == TERMINATED;
2412	>	long c = ctl;
2413	>	return ((c & STOP_BIT) != 0L &&
2414	>	(short)(c >>> TC_SHIFT) == -parallelism);
2415		}
2416
2417		/**
2418		* Returns {@code true} if the process of termination has
2419	<	* commenced but possibly not yet completed.
2419	>	* commenced but not yet completed.  This method may be useful for
2420	>	* debugging. A return of {@code true} reported a sufficient
2421	>	* period after shutdown may indicate that submitted tasks have
2422	>	* ignored or suppressed interruption, or are waiting for IO,
2423	>	* causing this executor not to properly terminate. (See the
2424	>	* advisory notes for class {@link ForkJoinTask} stating that
2425	>	* tasks should not normally entail blocking operations.  But if
2426	>	* they do, they must abort them on interrupt.)
2427		*
2428	<	* @return {@code true} if terminating
2428	>	* @return {@code true} if terminating but not yet terminated
2429		*/
2430		public boolean isTerminating() {
2431	<	return runStateOf(runControl) >= TERMINATING;
2431	>	long c = ctl;
2432	>	return ((c & STOP_BIT) != 0L &&
2433	>	(short)(c >>> TC_SHIFT) != -parallelism);
2434		}
2435
2436		/**
#	Line 1176 | Line 2439 | public class ForkJoinPool extends Abstra
2439		* @return {@code true} if this pool has been shut down
2440		*/
2441		public boolean isShutdown() {
2442	<	return runStateOf(runControl) >= SHUTDOWN;
2442	>	return runState < 0;
2443		}
2444
2445		/**
#	Line 1193 | Line 2456 | public class ForkJoinPool extends Abstra
2456		public boolean awaitTermination(long timeout, TimeUnit unit)
2457		throws InterruptedException {
2458		long nanos = unit.toNanos(timeout);
2459	<	final ReentrantLock lock = this.workerLock;
2459	>	final ReentrantLock lock = this.lock;
2460		lock.lock();
2461		try {
2462		for (;;) {
#	Line 1208 | Line 2471 | public class ForkJoinPool extends Abstra
2471		}
2472		}
2473
1211	–	// Shutdown and termination support
1212	–
1213	–	/**
1214	–	* Callback from terminating worker. Nulls out the corresponding
1215	–	* workers slot, and if terminating, tries to terminate; else
1216	–	* tries to shrink workers array.
1217	–	*
1218	–	* @param w the worker
1219	–	*/
1220	–	final void workerTerminated(ForkJoinWorkerThread w) {
1221	–	updateStealCount(w);
1222	–	updateWorkerCount(-1);
1223	–	final ReentrantLock lock = this.workerLock;
1224	–	lock.lock();
1225	–	try {
1226	–	ForkJoinWorkerThread[] ws = workers;
1227	–	if (ws != null) {
1228	–	int idx = w.poolIndex;
1229	–	if (idx >= 0 && idx < ws.length && ws[idx] == w)
1230	–	ws[idx] = null;
1231	–	if (totalCountOf(workerCounts) == 0) {
1232	–	terminate(); // no-op if already terminating
1233	–	transitionRunStateTo(TERMINATED);
1234	–	termination.signalAll();
1235	–	}
1236	–	else if (!isTerminating()) {
1237	–	tryShrinkWorkerArray();
1238	–	tryResumeSpare(true); // allow replacement
1239	–	}
1240	–	}
1241	–	} finally {
1242	–	lock.unlock();
1243	–	}
1244	–	signalIdleWorkers();
1245	–	}
1246	–
1247	–	/**
1248	–	* Initiates termination.
1249	–	*/
1250	–	private void terminate() {
1251	–	if (transitionRunStateTo(TERMINATING)) {
1252	–	stopAllWorkers();
1253	–	resumeAllSpares();
1254	–	signalIdleWorkers();
1255	–	cancelQueuedSubmissions();
1256	–	cancelQueuedWorkerTasks();
1257	–	interruptUnterminatedWorkers();
1258	–	signalIdleWorkers(); // resignal after interrupt
1259	–	}
1260	–	}
1261	–
1262	–	/**
1263	–	* Possibly terminates when on shutdown state.
1264	–	*/
1265	–	private void terminateOnShutdown() {
1266	–	if (!hasQueuedSubmissions() && canTerminateOnShutdown(runControl))
1267	–	terminate();
1268	–	}
1269	–
1270	–	/**
1271	–	* Clears out and cancels submissions.
1272	–	*/
1273	–	private void cancelQueuedSubmissions() {
1274	–	ForkJoinTask<?> task;
1275	–	while ((task = pollSubmission()) != null)
1276	–	task.cancel(false);
1277	–	}
1278	–
1279	–	/**
1280	–	* Cleans out worker queues.
1281	–	*/
1282	–	private void cancelQueuedWorkerTasks() {
1283	–	final ReentrantLock lock = this.workerLock;
1284	–	lock.lock();
1285	–	try {
1286	–	ForkJoinWorkerThread[] ws = workers;
1287	–	if (ws != null) {
1288	–	for (int i = 0; i < ws.length; ++i) {
1289	–	ForkJoinWorkerThread t = ws[i];
1290	–	if (t != null)
1291	–	t.cancelTasks();
1292	–	}
1293	–	}
1294	–	} finally {
1295	–	lock.unlock();
1296	–	}
1297	–	}
1298	–
1299	–	/**
1300	–	* Sets each worker's status to terminating. Requires lock to avoid
1301	–	* conflicts with add/remove.
1302	–	*/
1303	–	private void stopAllWorkers() {
1304	–	final ReentrantLock lock = this.workerLock;
1305	–	lock.lock();
1306	–	try {
1307	–	ForkJoinWorkerThread[] ws = workers;
1308	–	if (ws != null) {
1309	–	for (int i = 0; i < ws.length; ++i) {
1310	–	ForkJoinWorkerThread t = ws[i];
1311	–	if (t != null)
1312	–	t.shutdownNow();
1313	–	}
1314	–	}
1315	–	} finally {
1316	–	lock.unlock();
1317	–	}
1318	–	}
1319	–
1320	–	/**
1321	–	* Interrupts all unterminated workers.  This is not required for
1322	–	* sake of internal control, but may help unstick user code during
1323	–	* shutdown.
1324	–	*/
1325	–	private void interruptUnterminatedWorkers() {
1326	–	final ReentrantLock lock = this.workerLock;
1327	–	lock.lock();
1328	–	try {
1329	–	ForkJoinWorkerThread[] ws = workers;
1330	–	if (ws != null) {
1331	–	for (int i = 0; i < ws.length; ++i) {
1332	–	ForkJoinWorkerThread t = ws[i];
1333	–	if (t != null && !t.isTerminated()) {
1334	–	try {
1335	–	t.interrupt();
1336	–	} catch (SecurityException ignore) {
1337	–	}
1338	–	}
1339	–	}
1340	–	}
1341	–	} finally {
1342	–	lock.unlock();
1343	–	}
1344	–	}
1345	–
1346	–
1347	–	/*
1348	–	* Nodes for event barrier to manage idle threads.  Queue nodes
1349	–	* are basic Treiber stack nodes, also used for spare stack.
1350	–	*
1351	–	* The event barrier has an event count and a wait queue (actually
1352	–	* a Treiber stack).  Workers are enabled to look for work when
1353	–	* the eventCount is incremented. If they fail to find work, they
1354	–	* may wait for next count. Upon release, threads help others wake
1355	–	* up.
1356	–	*
1357	–	* Synchronization events occur only in enough contexts to
1358	–	* maintain overall liveness:
1359	–	*
1360	–	*   - Submission of a new task to the pool
1361	–	*   - Resizes or other changes to the workers array
1362	–	*   - pool termination
1363	–	*   - A worker pushing a task on an empty queue
1364	–	*
1365	–	* The case of pushing a task occurs often enough, and is heavy
1366	–	* enough compared to simple stack pushes, to require special
1367	–	* handling: Method signalWork returns without advancing count if
1368	–	* the queue appears to be empty.  This would ordinarily result in
1369	–	* races causing some queued waiters not to be woken up. To avoid
1370	–	* this, the first worker enqueued in method sync (see
1371	–	* syncIsReleasable) rescans for tasks after being enqueued, and
1372	–	* helps signal if any are found. This works well because the
1373	–	* worker has nothing better to do, and so might as well help
1374	–	* alleviate the overhead and contention on the threads actually
1375	–	* doing work.  Also, since event counts increments on task
1376	–	* availability exist to maintain liveness (rather than to force
1377	–	* refreshes etc), it is OK for callers to exit early if
1378	–	* contending with another signaller.
1379	–	*/
1380	–	static final class WaitQueueNode {
1381	–	WaitQueueNode next; // only written before enqueued
1382	–	volatile ForkJoinWorkerThread thread; // nulled to cancel wait
1383	–	final long count; // unused for spare stack
1384	–
1385	–	WaitQueueNode(long c, ForkJoinWorkerThread w) {
1386	–	count = c;
1387	–	thread = w;
1388	–	}
1389	–
1390	–	/**
1391	–	* Wakes up waiter, returning false if known to already
1392	–	*/
1393	–	boolean signal() {
1394	–	ForkJoinWorkerThread t = thread;
1395	–	if (t == null)
1396	–	return false;
1397	–	thread = null;
1398	–	LockSupport.unpark(t);
1399	–	return true;
1400	–	}
1401	–
1402	–	/**
1403	–	* Awaits release on sync.
1404	–	*/
1405	–	void awaitSyncRelease(ForkJoinPool p) {
1406	–	while (thread != null && !p.syncIsReleasable(this))
1407	–	LockSupport.park(this);
1408	–	}
1409	–
1410	–	/**
1411	–	* Awaits resumption as spare.
1412	–	*/
1413	–	void awaitSpareRelease() {
1414	–	while (thread != null) {
1415	–	if (!Thread.interrupted())
1416	–	LockSupport.park(this);
1417	–	}
1418	–	}
1419	–	}
1420	–
1421	–	/**
1422	–	* Ensures that no thread is waiting for count to advance from the
1423	–	* current value of eventCount read on entry to this method, by
1424	–	* releasing waiting threads if necessary.
1425	–	*
1426	–	* @return the count
1427	–	*/
1428	–	final long ensureSync() {
1429	–	long c = eventCount;
1430	–	WaitQueueNode q;
1431	–	while ((q = syncStack) != null && q.count < c) {
1432	–	if (casBarrierStack(q, null)) {
1433	–	do {
1434	–	q.signal();
1435	–	} while ((q = q.next) != null);
1436	–	break;
1437	–	}
1438	–	}
1439	–	return c;
1440	–	}
1441	–
1442	–	/**
1443	–	* Increments event count and releases waiting threads.
1444	–	*/
1445	–	private void signalIdleWorkers() {
1446	–	long c;
1447	–	do {} while (!casEventCount(c = eventCount, c+1));
1448	–	ensureSync();
1449	–	}
1450	–
1451	–	/**
1452	–	* Signals threads waiting to poll a task. Because method sync
1453	–	* rechecks availability, it is OK to only proceed if queue
1454	–	* appears to be non-empty, and OK to skip under contention to
1455	–	* increment count (since some other thread succeeded).
1456	–	*/
1457	–	final void signalWork() {
1458	–	long c;
1459	–	WaitQueueNode q;
1460	–	if (syncStack != null &&
1461	–	casEventCount(c = eventCount, c+1) &&
1462	–	(((q = syncStack) != null && q.count <= c) &&
1463	–	(!casBarrierStack(q, q.next) || !q.signal())))
1464	–	ensureSync();
1465	–	}
1466	–
1467	–	/**
1468	–	* Waits until event count advances from last value held by
1469	–	* caller, or if excess threads, caller is resumed as spare, or
1470	–	* caller or pool is terminating. Updates caller's event on exit.
1471	–	*
1472	–	* @param w the calling worker thread
1473	–	*/
1474	–	final void sync(ForkJoinWorkerThread w) {
1475	–	updateStealCount(w); // Transfer w's count while it is idle
1476	–
1477	–	while (!w.isShutdown() && !isTerminating() && !suspendIfSpare(w)) {
1478	–	long prev = w.lastEventCount;
1479	–	WaitQueueNode node = null;
1480	–	WaitQueueNode h;
1481	–	while (eventCount == prev &&
1482	–	((h = syncStack) == null || h.count == prev)) {
1483	–	if (node == null)
1484	–	node = new WaitQueueNode(prev, w);
1485	–	if (casBarrierStack(node.next = h, node)) {
1486	–	node.awaitSyncRelease(this);
1487	–	break;
1488	–	}
1489	–	}
1490	–	long ec = ensureSync();
1491	–	if (ec != prev) {
1492	–	w.lastEventCount = ec;
1493	–	break;
1494	–	}
1495	–	}
1496	–	}
1497	–
1498	–	/**
1499	–	* Returns true if worker waiting on sync can proceed:
1500	–	*  - on signal (thread == null)
1501	–	*  - on event count advance (winning race to notify vs signaller)
1502	–	*  - on interrupt
1503	–	*  - if the first queued node, we find work available
1504	–	* If node was not signalled and event count not advanced on exit,
1505	–	* then we also help advance event count.
1506	–	*
1507	–	* @return true if node can be released
1508	–	*/
1509	–	final boolean syncIsReleasable(WaitQueueNode node) {
1510	–	long prev = node.count;
1511	–	if (!Thread.interrupted() && node.thread != null &&
1512	–	(node.next != null ||
1513	–	!ForkJoinWorkerThread.hasQueuedTasks(workers)) &&
1514	–	eventCount == prev)
1515	–	return false;
1516	–	if (node.thread != null) {
1517	–	node.thread = null;
1518	–	long ec = eventCount;
1519	–	if (prev <= ec) // help signal
1520	–	casEventCount(ec, ec+1);
1521	–	}
1522	–	return true;
1523	–	}
1524	–
1525	–	/**
1526	–	* Returns true if a new sync event occurred since last call to
1527	–	* sync or this method, if so, updating caller's count.
1528	–	*/
1529	–	final boolean hasNewSyncEvent(ForkJoinWorkerThread w) {
1530	–	long lc = w.lastEventCount;
1531	–	long ec = ensureSync();
1532	–	if (ec == lc)
1533	–	return false;
1534	–	w.lastEventCount = ec;
1535	–	return true;
1536	–	}
1537	–
1538	–	//  Parallelism maintenance
1539	–
1540	–	/**
1541	–	* Decrements running count; if too low, adds spare.
1542	–	*
1543	–	* Conceptually, all we need to do here is add or resume a
1544	–	* spare thread when one is about to block (and remove or
1545	–	* suspend it later when unblocked -- see suspendIfSpare).
1546	–	* However, implementing this idea requires coping with
1547	–	* several problems: we have imperfect information about the
1548	–	* states of threads. Some count updates can and usually do
1549	–	* lag run state changes, despite arrangements to keep them
1550	–	* accurate (for example, when possible, updating counts
1551	–	* before signalling or resuming), especially when running on
1552	–	* dynamic JVMs that don't optimize the infrequent paths that
1553	–	* update counts. Generating too many threads can make these
1554	–	* problems become worse, because excess threads are more
1555	–	* likely to be context-switched with others, slowing them all
1556	–	* down, especially if there is no work available, so all are
1557	–	* busy scanning or idling.  Also, excess spare threads can
1558	–	* only be suspended or removed when they are idle, not
1559	–	* immediately when they aren't needed. So adding threads will
1560	–	* raise parallelism level for longer than necessary.  Also,
1561	–	* FJ applications often encounter highly transient peaks when
1562	–	* many threads are blocked joining, but for less time than it
1563	–	* takes to create or resume spares.
1564	–	*
1565	–	* @param joinMe if non-null, return early if done
1566	–	* @param maintainParallelism if true, try to stay within
1567	–	* target counts, else create only to avoid starvation
1568	–	* @return true if joinMe known to be done
1569	–	*/
1570	–	final boolean preJoin(ForkJoinTask<?> joinMe,
1571	–	boolean maintainParallelism) {
1572	–	maintainParallelism &= maintainsParallelism; // overrride
1573	–	boolean dec = false;  // true when running count decremented
1574	–	while (spareStack == null || !tryResumeSpare(dec)) {
1575	–	int counts = workerCounts;
1576	–	if (dec || (dec = casWorkerCounts(counts, --counts))) {
1577	–	// CAS cheat
1578	–	if (!needSpare(counts, maintainParallelism))
1579	–	break;
1580	–	if (joinMe.status < 0)
1581	–	return true;
1582	–	if (tryAddSpare(counts))
1583	–	break;
1584	–	}
1585	–	}
1586	–	return false;
1587	–	}
1588	–
1589	–	/**
1590	–	* Same idea as preJoin
1591	–	*/
1592	–	final boolean preBlock(ManagedBlocker blocker,
1593	–	boolean maintainParallelism) {
1594	–	maintainParallelism &= maintainsParallelism;
1595	–	boolean dec = false;
1596	–	while (spareStack == null || !tryResumeSpare(dec)) {
1597	–	int counts = workerCounts;
1598	–	if (dec || (dec = casWorkerCounts(counts, --counts))) {
1599	–	if (!needSpare(counts, maintainParallelism))
1600	–	break;
1601	–	if (blocker.isReleasable())
1602	–	return true;
1603	–	if (tryAddSpare(counts))
1604	–	break;
1605	–	}
1606	–	}
1607	–	return false;
1608	–	}
1609	–
1610	–	/**
1611	–	* Returns true if a spare thread appears to be needed.  If
1612	–	* maintaining parallelism, returns true when the deficit in
1613	–	* running threads is more than the surplus of total threads, and
1614	–	* there is apparently some work to do.  This self-limiting rule
1615	–	* means that the more threads that have already been added, the
1616	–	* less parallelism we will tolerate before adding another.
1617	–	*
1618	–	* @param counts current worker counts
1619	–	* @param maintainParallelism try to maintain parallelism
1620	–	*/
1621	–	private boolean needSpare(int counts, boolean maintainParallelism) {
1622	–	int ps = parallelism;
1623	–	int rc = runningCountOf(counts);
1624	–	int tc = totalCountOf(counts);
1625	–	int runningDeficit = ps - rc;
1626	–	int totalSurplus = tc - ps;
1627	–	return (tc < maxPoolSize &&
1628	–	(rc == 0 || totalSurplus < 0 ||
1629	–	(maintainParallelism &&
1630	–	runningDeficit > totalSurplus &&
1631	–	ForkJoinWorkerThread.hasQueuedTasks(workers))));
1632	–	}
1633	–
1634	–	/**
1635	–	* Adds a spare worker if lock available and no more than the
1636	–	* expected numbers of threads exist.
1637	–	*
1638	–	* @return true if successful
1639	–	*/
1640	–	private boolean tryAddSpare(int expectedCounts) {
1641	–	final ReentrantLock lock = this.workerLock;
1642	–	int expectedRunning = runningCountOf(expectedCounts);
1643	–	int expectedTotal = totalCountOf(expectedCounts);
1644	–	boolean success = false;
1645	–	boolean locked = false;
1646	–	// confirm counts while locking; CAS after obtaining lock
1647	–	try {
1648	–	for (;;) {
1649	–	int s = workerCounts;
1650	–	int tc = totalCountOf(s);
1651	–	int rc = runningCountOf(s);
1652	–	if (rc > expectedRunning || tc > expectedTotal)
1653	–	break;
1654	–	if (!locked && !(locked = lock.tryLock()))
1655	–	break;
1656	–	if (casWorkerCounts(s, workerCountsFor(tc+1, rc+1))) {
1657	–	createAndStartSpare(tc);
1658	–	success = true;
1659	–	break;
1660	–	}
1661	–	}
1662	–	} finally {
1663	–	if (locked)
1664	–	lock.unlock();
1665	–	}
1666	–	return success;
1667	–	}
1668	–
1669	–	/**
1670	–	* Adds the kth spare worker. On entry, pool counts are already
1671	–	* adjusted to reflect addition.
1672	–	*/
1673	–	private void createAndStartSpare(int k) {
1674	–	ForkJoinWorkerThread w = null;
1675	–	ForkJoinWorkerThread[] ws = ensureWorkerArrayCapacity(k + 1);
1676	–	int len = ws.length;
1677	–	// Probably, we can place at slot k. If not, find empty slot
1678	–	if (k < len && ws[k] != null) {
1679	–	for (k = 0; k < len && ws[k] != null; ++k)
1680	–	;
1681	–	}
1682	–	if (k < len && !isTerminating() && (w = createWorker(k)) != null) {
1683	–	ws[k] = w;
1684	–	w.start();
1685	–	}
1686	–	else
1687	–	updateWorkerCount(-1); // adjust on failure
1688	–	signalIdleWorkers();
1689	–	}
1690	–
1691	–	/**
1692	–	* Suspends calling thread w if there are excess threads.  Called
1693	–	* only from sync.  Spares are enqueued in a Treiber stack using
1694	–	* the same WaitQueueNodes as barriers.  They are resumed mainly
1695	–	* in preJoin, but are also woken on pool events that require all
1696	–	* threads to check run state.
1697	–	*
1698	–	* @param w the caller
1699	–	*/
1700	–	private boolean suspendIfSpare(ForkJoinWorkerThread w) {
1701	–	WaitQueueNode node = null;
1702	–	int s;
1703	–	while (parallelism < runningCountOf(s = workerCounts)) {
1704	–	if (node == null)
1705	–	node = new WaitQueueNode(0, w);
1706	–	if (casWorkerCounts(s, s-1)) { // representation-dependent
1707	–	// push onto stack
1708	–	do {} while (!casSpareStack(node.next = spareStack, node));
1709	–	// block until released by resumeSpare
1710	–	node.awaitSpareRelease();
1711	–	return true;
1712	–	}
1713	–	}
1714	–	return false;
1715	–	}
1716	–
1717	–	/**
1718	–	* Tries to pop and resume a spare thread.
1719	–	*
1720	–	* @param updateCount if true, increment running count on success
1721	–	* @return true if successful
1722	–	*/
1723	–	private boolean tryResumeSpare(boolean updateCount) {
1724	–	WaitQueueNode q;
1725	–	while ((q = spareStack) != null) {
1726	–	if (casSpareStack(q, q.next)) {
1727	–	if (updateCount)
1728	–	updateRunningCount(1);
1729	–	q.signal();
1730	–	return true;
1731	–	}
1732	–	}
1733	–	return false;
1734	–	}
1735	–
1736	–	/**
1737	–	* Pops and resumes all spare threads. Same idea as ensureSync.
1738	–	*
1739	–	* @return true if any spares released
1740	–	*/
1741	–	private boolean resumeAllSpares() {
1742	–	WaitQueueNode q;
1743	–	while ( (q = spareStack) != null) {
1744	–	if (casSpareStack(q, null)) {
1745	–	do {
1746	–	updateRunningCount(1);
1747	–	q.signal();
1748	–	} while ((q = q.next) != null);
1749	–	return true;
1750	–	}
1751	–	}
1752	–	return false;
1753	–	}
1754	–
1755	–	/**
1756	–	* Pops and shuts down excessive spare threads. Call only while
1757	–	* holding lock. This is not guaranteed to eliminate all excess
1758	–	* threads, only those suspended as spares, which are the ones
1759	–	* unlikely to be needed in the future.
1760	–	*/
1761	–	private void trimSpares() {
1762	–	int surplus = totalCountOf(workerCounts) - parallelism;
1763	–	WaitQueueNode q;
1764	–	while (surplus > 0 && (q = spareStack) != null) {
1765	–	if (casSpareStack(q, null)) {
1766	–	do {
1767	–	updateRunningCount(1);
1768	–	ForkJoinWorkerThread w = q.thread;
1769	–	if (w != null && surplus > 0 &&
1770	–	runningCountOf(workerCounts) > 0 && w.shutdown())
1771	–	--surplus;
1772	–	q.signal();
1773	–	} while ((q = q.next) != null);
1774	–	}
1775	–	}
1776	–	}
1777	–
2474		/**
2475		* Interface for extending managed parallelism for tasks running
2476	<	* in ForkJoinPools. A ManagedBlocker provides two methods.
2477	<	* Method {@code isReleasable} must return true if blocking is not
2478	<	* necessary. Method {@code block} blocks the current thread if
2479	<	* necessary (perhaps internally invoking {@code isReleasable}
2480	<	* before actually blocking.).
2476	>	* in {@link ForkJoinPool}s.
2477	>	*
2478	>	* <p>A {@code ManagedBlocker} provides two methods.  Method
2479	>	* {@code isReleasable} must return {@code true} if blocking is
2480	>	* not necessary. Method {@code block} blocks the current thread
2481	>	* if necessary (perhaps internally invoking {@code isReleasable}
2482	>	* before actually blocking). These actions are performed by any
2483	>	* thread invoking {@link ForkJoinPool#managedBlock}.  The
2484	>	* unusual methods in this API accommodate synchronizers that may,
2485	>	* but don't usually, block for long periods. Similarly, they
2486	>	* allow more efficient internal handling of cases in which
2487	>	* additional workers may be, but usually are not, needed to
2488	>	* ensure sufficient parallelism.  Toward this end,
2489	>	* implementations of method {@code isReleasable} must be amenable
2490	>	* to repeated invocation.
2491		*
2492		* <p>For example, here is a ManagedBlocker based on a
2493		* ReentrantLock:
#	Line 1799 | Line 2505 | public class ForkJoinPool extends Abstra
2505		*     return hasLock || (hasLock = lock.tryLock());
2506		*   }
2507		* }}</pre>
2508	+	*
2509	+	* <p>Here is a class that possibly blocks waiting for an
2510	+	* item on a given queue:
2511	+	*  <pre> {@code
2512	+	* class QueueTaker<E> implements ManagedBlocker {
2513	+	*   final BlockingQueue<E> queue;
2514	+	*   volatile E item = null;
2515	+	*   QueueTaker(BlockingQueue<E> q) { this.queue = q; }
2516	+	*   public boolean block() throws InterruptedException {
2517	+	*     if (item == null)
2518	+	*       item = queue.take();
2519	+	*     return true;
2520	+	*   }
2521	+	*   public boolean isReleasable() {
2522	+	*     return item != null || (item = queue.poll()) != null;
2523	+	*   }
2524	+	*   public E getItem() { // call after pool.managedBlock completes
2525	+	*     return item;
2526	+	*   }
2527	+	* }}</pre>
2528		*/
2529		public static interface ManagedBlocker {
2530		/**
2531		* Possibly blocks the current thread, for example waiting for
2532		* a lock or condition.
2533		*
2534	<	* @return true if no additional blocking is necessary (i.e.,
2535	<	* if isReleasable would return true)
2534	>	* @return {@code true} if no additional blocking is necessary
2535	>	* (i.e., if isReleasable would return true)
2536		* @throws InterruptedException if interrupted while waiting
2537		* (the method is not required to do so, but is allowed to)
2538		*/
2539		boolean block() throws InterruptedException;
2540
2541		/**
2542	<	* Returns true if blocking is unnecessary.
2542	>	* Returns {@code true} if blocking is unnecessary.
2543		*/
2544		boolean isReleasable();
2545		}
2546
2547		/**
2548		* Blocks in accord with the given blocker.  If the current thread
2549	<	* is a ForkJoinWorkerThread, this method possibly arranges for a
2550	<	* spare thread to be activated if necessary to ensure parallelism
2551	<	* while the current thread is blocked.  If
2552	<	* {@code maintainParallelism} is true and the pool supports
2553	<	* it ({@link #getMaintainsParallelism}), this method attempts to
2554	<	* maintain the pool's nominal parallelism. Otherwise it activates
2555	<	* a thread only if necessary to avoid complete starvation. This
1830	<	* option may be preferable when blockages use timeouts, or are
1831	<	* almost always brief.
1832	<	*
1833	<	* <p> If the caller is not a ForkJoinTask, this method is behaviorally
1834	<	* equivalent to
1835	<	*  <pre> {@code
2549	>	* is a {@link ForkJoinWorkerThread}, this method possibly
2550	>	* arranges for a spare thread to be activated if necessary to
2551	>	* ensure sufficient parallelism while the current thread is blocked.
2552	>	*
2553	>	* <p>If the caller is not a {@link ForkJoinTask}, this method is
2554	>	* behaviorally equivalent to
2555	>	a     *  <pre> {@code
2556		* while (!blocker.isReleasable())
2557		*   if (blocker.block())
2558		*     return;
2559		* }</pre>
2560	<	* If the caller is a ForkJoinTask, then the pool may first
2561	<	* be expanded to ensure parallelism, and later adjusted.
2560	>	*
2561	>	* If the caller is a {@code ForkJoinTask}, then the pool may
2562	>	* first be expanded to ensure parallelism, and later adjusted.
2563		*
2564		* @param blocker the blocker
1844	–	* @param maintainParallelism if true and supported by this pool,
1845	–	* attempt to maintain the pool's nominal parallelism; otherwise
1846	–	* activate a thread only if necessary to avoid complete
1847	–	* starvation.
2565		* @throws InterruptedException if blocker.block did so
2566		*/
2567	<	public static void managedBlock(ManagedBlocker blocker,
1851	<	boolean maintainParallelism)
2567	>	public static void managedBlock(ManagedBlocker blocker)
2568		throws InterruptedException {
2569		Thread t = Thread.currentThread();
2570	<	ForkJoinPool pool = ((t instanceof ForkJoinWorkerThread) ?
2571	<	((ForkJoinWorkerThread) t).pool : null);
2572	<	if (!blocker.isReleasable()) {
2573	<	try {
2574	<	if (pool == null ||
2575	<	!pool.preBlock(blocker, maintainParallelism))
2576	<	awaitBlocker(blocker);
2577	<	} finally {
2578	<	if (pool != null)
2579	<	pool.updateRunningCount(1);
2570	>	ForkJoinPool p = ((t instanceof ForkJoinWorkerThread) ?
2571	>	((ForkJoinWorkerThread)t).pool : null);
2572	>	while (!blocker.isReleasable()) {
2573	>	if (p == null || p.tryCompensate()) {
2574	>	try {
2575	>	do {} while (!blocker.isReleasable() && !blocker.block());
2576	>	} finally {
2577	>	if (p != null)
2578	>	p.incrementActiveCount();
2579	>	}
2580	>	break;
2581		}
2582		}
2583		}
2584
2585	<	private static void awaitBlocker(ManagedBlocker blocker)
2586	<	throws InterruptedException {
2587	<	do {} while (!blocker.isReleasable() && !blocker.block());
1871	<	}
1872	<
1873	<	// AbstractExecutorService overrides
2585	>	// AbstractExecutorService overrides.  These rely on undocumented
2586	>	// fact that ForkJoinTask.adapt returns ForkJoinTasks that also
2587	>	// implement RunnableFuture.
2588
2589		protected <T> RunnableFuture<T> newTaskFor(Runnable runnable, T value) {
2590	<	return new AdaptedRunnable<T>(runnable, value);
2590	>	return (RunnableFuture<T>) ForkJoinTask.adapt(runnable, value);
2591		}
2592
2593		protected <T> RunnableFuture<T> newTaskFor(Callable<T> callable) {
2594	<	return new AdaptedCallable<T>(callable);
2594	>	return (RunnableFuture<T>) ForkJoinTask.adapt(callable);
2595		}
2596
2597	+	// Unsafe mechanics
2598	+	private static final sun.misc.Unsafe U;
2599	+	private static final long CTL;
2600	+	private static final long RUNSTATE;
2601	+	private static final long PARKBLOCKER;
2602	+
2603	+	static {
2604	+	poolNumberGenerator = new AtomicInteger();
2605	+	modifyThreadPermission = new RuntimePermission("modifyThread");
2606	+	defaultForkJoinWorkerThreadFactory =
2607	+	new DefaultForkJoinWorkerThreadFactory();
2608	+	int s;
2609	+	try {
2610	+	U = getUnsafe();
2611	+	Class<?> k = ForkJoinPool.class;
2612	+	Class<?> tk = Thread.class;
2613	+	CTL = U.objectFieldOffset
2614	+	(k.getDeclaredField("ctl"));
2615	+	RUNSTATE = U.objectFieldOffset
2616	+	(k.getDeclaredField("runState"));
2617	+	PARKBLOCKER = U.objectFieldOffset
2618	+	(tk.getDeclaredField("parkBlocker"));
2619	+	} catch (Exception e) {
2620	+	throw new Error(e);
2621	+	}
2622	+	}
2623
2624	<	// Unsafe mechanics for jsr166y 3rd party package.
2624	>	/**
2625	>	* Returns a sun.misc.Unsafe.  Suitable for use in a 3rd party package.
2626	>	* Replace with a simple call to Unsafe.getUnsafe when integrating
2627	>	* into a jdk.
2628	>	*
2629	>	* @return a sun.misc.Unsafe
2630	>	*/
2631		private static sun.misc.Unsafe getUnsafe() {
2632		try {
2633		return sun.misc.Unsafe.getUnsafe();
2634		} catch (SecurityException se) {
2635		try {
2636		return java.security.AccessController.doPrivileged
2637	<	(new java.security.PrivilegedExceptionAction<sun.misc.Unsafe>() {
2637	>	(new java.security
2638	>	.PrivilegedExceptionAction<sun.misc.Unsafe>() {
2639		public sun.misc.Unsafe run() throws Exception {
2640	<	return getUnsafeByReflection();
2640	>	java.lang.reflect.Field f = sun.misc
2641	>	.Unsafe.class.getDeclaredField("theUnsafe");
2642	>	f.setAccessible(true);
2643	>	return (sun.misc.Unsafe) f.get(null);
2644		}});
2645		} catch (java.security.PrivilegedActionException e) {
2646		throw new RuntimeException("Could not initialize intrinsics",
#	Line 1898 | Line 2648 | public class ForkJoinPool extends Abstra
2648		}
2649		}
2650		}
1901	–
1902	–	private static sun.misc.Unsafe getUnsafeByReflection()
1903	–	throws NoSuchFieldException, IllegalAccessException {
1904	–	java.lang.reflect.Field f =
1905	–	sun.misc.Unsafe.class.getDeclaredField("theUnsafe");
1906	–	f.setAccessible(true);
1907	–	return (sun.misc.Unsafe) f.get(null);
1908	–	}
1909	–
1910	–	private static long fieldOffset(String fieldName, Class<?> klazz) {
1911	–	try {
1912	–	return UNSAFE.objectFieldOffset(klazz.getDeclaredField(fieldName));
1913	–	} catch (NoSuchFieldException e) {
1914	–	// Convert Exception to Error
1915	–	NoSuchFieldError error = new NoSuchFieldError(fieldName);
1916	–	error.initCause(e);
1917	–	throw error;
1918	–	}
1919	–	}
1920	–
1921	–	private static final sun.misc.Unsafe UNSAFE = getUnsafe();
1922	–	static final long eventCountOffset =
1923	–	fieldOffset("eventCount", ForkJoinPool.class);
1924	–	static final long workerCountsOffset =
1925	–	fieldOffset("workerCounts", ForkJoinPool.class);
1926	–	static final long runControlOffset =
1927	–	fieldOffset("runControl", ForkJoinPool.class);
1928	–	static final long syncStackOffset =
1929	–	fieldOffset("syncStack",ForkJoinPool.class);
1930	–	static final long spareStackOffset =
1931	–	fieldOffset("spareStack", ForkJoinPool.class);
1932	–
1933	–	private boolean casEventCount(long cmp, long val) {
1934	–	return UNSAFE.compareAndSwapLong(this, eventCountOffset, cmp, val);
1935	–	}
1936	–	private boolean casWorkerCounts(int cmp, int val) {
1937	–	return UNSAFE.compareAndSwapInt(this, workerCountsOffset, cmp, val);
1938	–	}
1939	–	private boolean casRunControl(int cmp, int val) {
1940	–	return UNSAFE.compareAndSwapInt(this, runControlOffset, cmp, val);
1941	–	}
1942	–	private boolean casSpareStack(WaitQueueNode cmp, WaitQueueNode val) {
1943	–	return UNSAFE.compareAndSwapObject(this, spareStackOffset, cmp, val);
1944	–	}
1945	–	private boolean casBarrierStack(WaitQueueNode cmp, WaitQueueNode val) {
1946	–	return UNSAFE.compareAndSwapObject(this, syncStackOffset, cmp, val);
1947	–	}
2651		}

Diff Legend

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