ViewVC Help

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

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

Comparing jsr166/src/jsr166y/ForkJoinPool.java (file contents):
Revision 1.56 by dl, Thu May 27 16:46:48 2010 UTC vs.
Revision 1.174 by jsr166, Sun Jan 13 21:56:12 2013 UTC

#	Line 1 | Line 1
1		/*
2		* Written by Doug Lea with assistance from members of JCP JSR-166
3		* Expert Group and released to the public domain, as explained at
4	<	* http://creativecommons.org/licenses/publicdomain
4	>	* http://creativecommons.org/publicdomain/zero/1.0/
5		*/
6
7		package jsr166y;
8
9	–	import java.util.concurrent.*;
10	–
9		import java.util.ArrayList;
10		import java.util.Arrays;
11		import java.util.Collection;
12		import java.util.Collections;
13		import java.util.List;
14	<	import java.util.concurrent.locks.LockSupport;
15	<	import java.util.concurrent.locks.ReentrantLock;
16	<	import java.util.concurrent.atomic.AtomicInteger;
17	<	import java.util.concurrent.CountDownLatch;
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
22		/**
23		* An {@link ExecutorService} for running {@link ForkJoinTask}s.
24		* A {@code ForkJoinPool} provides the entry point for submissions
25	<	* from non-{@code ForkJoinTask}s, as well as management and
25	>	* from non-{@code ForkJoinTask} clients, as well as management and
26		* monitoring operations.
27		*
28		* <p>A {@code ForkJoinPool} differs from other kinds of {@link
29		* ExecutorService} mainly by virtue of employing
30		* <em>work-stealing</em>: all threads in the pool attempt to find and
31	<	* execute subtasks created by other active tasks (eventually blocking
32	<	* waiting for work if none exist). This enables efficient processing
33	<	* when most tasks spawn other subtasks (as do most {@code
34	<	* ForkJoinTask}s). A {@code ForkJoinPool} may also be used for mixed
35	<	* execution of some plain {@code Runnable}- or {@code Callable}-
36	<	* based activities along with {@code ForkJoinTask}s. When setting
37	<	* {@linkplain #setAsyncMode async mode}, a {@code ForkJoinPool} may
38	<	* also be appropriate for use with fine-grained tasks of any form
39	<	* that are never joined. Otherwise, other {@code ExecutorService}
40	<	* implementations are typically more appropriate choices.
31	>	* execute tasks submitted to the pool and/or created by other active
32	>	* tasks (eventually blocking waiting for work if none exist). This
33	>	* enables efficient processing when most tasks spawn other subtasks
34	>	* (as do most {@code ForkJoinTask}s), as well as when many small
35	>	* tasks are submitted to the pool from external clients.  Especially
36	>	* when setting <em>asyncMode</em> to true in constructors, {@code
37	>	* ForkJoinPool}s may also be appropriate for use with event-style
38	>	* tasks that are never joined.
39	>	*
40	>	* <p>A static {@link #commonPool()} is available and appropriate for
41	>	* most applications. The common pool is used by any ForkJoinTask that
42	>	* is not explicitly submitted to a specified pool. Using the common
43	>	* pool normally reduces resource usage (its threads are slowly
44	>	* reclaimed during periods of non-use, and reinstated upon subsequent
45	>	* use).
46		*
47	<	* <p>A {@code ForkJoinPool} is constructed with a given target
48	<	* parallelism level; by default, equal to the number of available
49	<	* processors. Unless configured otherwise via {@link
50	<	* #setMaintainsParallelism}, the pool attempts to maintain this
51	<	* number of active (or available) threads by dynamically adding,
52	<	* suspending, or resuming internal worker threads, even if some tasks
53	<	* are stalled waiting to join others. However, no such adjustments
54	<	* are performed in the face of blocked IO or other unmanaged
55	<	* synchronization. The nested {@link ManagedBlocker} interface
56	<	* enables extension of the kinds of synchronization accommodated.
51	<	* The target parallelism level may also be changed dynamically
52	<	* ({@link #setParallelism}). The total number of threads may be
53	<	* limited using method {@link #setMaximumPoolSize}, in which case it
54	<	* may become possible for the activities of a pool to stall due to
55	<	* the lack of available threads to process new tasks. When the pool
56	<	* is executing tasks, these and other configuration setting methods
57	<	* may only gradually affect actual pool sizes. It is normally best
58	<	* practice to invoke these methods only when the pool is known to be
59	<	* quiescent.
47	>	* <p>For applications that require separate or custom pools, a {@code
48	>	* ForkJoinPool} may be constructed with a given target parallelism
49	>	* level; by default, equal to the number of available processors. The
50	>	* pool attempts to maintain enough active (or available) threads by
51	>	* dynamically adding, suspending, or resuming internal worker
52	>	* threads, even if some tasks are stalled waiting to join
53	>	* others. However, no such adjustments are guaranteed in the face of
54	>	* blocked I/O or other unmanaged synchronization. The nested {@link
55	>	* ManagedBlocker} interface enables extension of the kinds of
56	>	* synchronization accommodated.
57		*
58		* <p>In addition to execution and lifecycle control methods, this
59		* class provides status check methods (for example
#	Line 65 | Line 62 | import java.util.concurrent.CountDownLat
62		* {@link #toString} returns indications of pool state in a
63		* convenient form for informal monitoring.
64		*
65	<	* <p><b>Sample Usage.</b> Normally a single {@code ForkJoinPool} is
66	<	* used for all parallel task execution in a program or subsystem.
67	<	* Otherwise, use would not usually outweigh the construction and
68	<	* bookkeeping overhead of creating a large set of threads. For
69	<	* example, a common pool could be used for the {@code SortTasks}
70	<	* illustrated in {@link RecursiveAction}. Because {@code
71	<	* ForkJoinPool} uses threads in {@linkplain java.lang.Thread#isDaemon
72	<	* daemon} mode, there is typically no need to explicitly {@link
73	<	* #shutdown} such a pool upon program exit.
65	>	* <p>As is the case with other ExecutorServices, there are three
66	>	* main task execution methods summarized in the following table.
67	>	* These are designed to be used primarily by clients not already
68	>	* engaged in fork/join computations in the current pool.  The main
69	>	* forms of these methods accept instances of {@code ForkJoinTask},
70	>	* but overloaded forms also allow mixed execution of plain {@code
71	>	* Runnable}- or {@code Callable}- based activities as well.  However,
72	>	* tasks that are already executing in a pool should normally instead
73	>	* use the within-computation forms listed in the table unless using
74	>	* async event-style tasks that are not usually joined, in which case
75	>	* there is little difference among choice of methods.
76	>	*
77	>	* <table BORDER CELLPADDING=3 CELLSPACING=1>
78	>	*  <tr>
79	>	*    <td></td>
80	>	*    <td ALIGN=CENTER> <b>Call from non-fork/join clients</b></td>
81	>	*    <td ALIGN=CENTER> <b>Call from within fork/join computations</b></td>
82	>	*  </tr>
83	>	*  <tr>
84	>	*    <td> <b>Arrange async execution</td>
85	>	*    <td> {@link #execute(ForkJoinTask)}</td>
86	>	*    <td> {@link ForkJoinTask#fork}</td>
87	>	*  </tr>
88	>	*  <tr>
89	>	*    <td> <b>Await and obtain result</td>
90	>	*    <td> {@link #invoke(ForkJoinTask)}</td>
91	>	*    <td> {@link ForkJoinTask#invoke}</td>
92	>	*  </tr>
93	>	*  <tr>
94	>	*    <td> <b>Arrange exec and obtain Future</td>
95	>	*    <td> {@link #submit(ForkJoinTask)}</td>
96	>	*    <td> {@link ForkJoinTask#fork} (ForkJoinTasks <em>are</em> Futures)</td>
97	>	*  </tr>
98	>	* </table>
99		*
100	<	* <pre>
101	<	* static final ForkJoinPool mainPool = new ForkJoinPool();
102	<	* ...
103	<	* public void sort(long[] array) {
104	<	*   mainPool.invoke(new SortTask(array, 0, array.length));
105	<	* }
106	<	* </pre>
100	>	* <p>The common pool is by default constructed with default
101	>	* parameters, but these may be controlled by setting three {@link
102	>	* System#getProperty system properties} with prefix {@code
103	>	* java.util.concurrent.ForkJoinPool.common}: {@code parallelism} --
104	>	* an integer greater than zero, {@code threadFactory} -- the class
105	>	* name of a {@link ForkJoinWorkerThreadFactory}, and {@code
106	>	* exceptionHandler} -- the class name of a {@link
107	>	* java.lang.Thread.UncaughtExceptionHandler
108	>	* Thread.UncaughtExceptionHandler}. Upon any error in establishing
109	>	* these settings, default parameters are used.
110		*
111		* <p><b>Implementation notes</b>: This implementation restricts the
112		* maximum number of running threads to 32767. Attempts to create
#	Line 89 | Line 114 | import java.util.concurrent.CountDownLat
114		* {@code IllegalArgumentException}.
115		*
116		* <p>This implementation rejects submitted tasks (that is, by throwing
117	<	* {@link RejectedExecutionException}) only when the pool is shut down.
117	>	* {@link RejectedExecutionException}) only when the pool is shut down
118	>	* or internal resources have been exhausted.
119		*
120		* @since 1.7
121		* @author Doug Lea
#	Line 99 | Line 125 | public class ForkJoinPool extends Abstra
125		/*
126		* Implementation Overview
127		*
128	<	* This class provides the central bookkeeping and control for a
129	<	* set of worker threads: Submissions from non-FJ threads enter
130	<	* into a submission queue. Workers take these tasks and typically
131	<	* split them into subtasks that may be stolen by other workers.
132	<	* The main work-stealing mechanics implemented in class
133	<	* ForkJoinWorkerThread give first priority to processing tasks
134	<	* from their own queues (LIFO or FIFO, depending on mode), then
135	<	* to randomized FIFO steals of tasks in other worker queues, and
136	<	* lastly to new submissions. These mechanics do not consider
137	<	* affinities, loads, cache localities, etc, so rarely provide the
138	<	* best possible performance on a given machine, but portably
139	<	* provide good throughput by averaging over these factors.
140	<	* (Further, even if we did try to use such information, we do not
141	<	* usually have a basis for exploiting it. For example, some sets
142	<	* of tasks profit from cache affinities, but others are harmed by
143	<	* cache pollution effects.)
128	>	* This class and its nested classes provide the main
129	>	* functionality and control for a set of worker threads:
130	>	* Submissions from non-FJ threads enter into submission queues.
131	>	* Workers take these tasks and typically split them into subtasks
132	>	* that may be stolen by other workers.  Preference rules give
133	>	* first priority to processing tasks from their own queues (LIFO
134	>	* or FIFO, depending on mode), then to randomized FIFO steals of
135	>	* tasks in other queues.
136	>	*
137	>	* WorkQueues
138	>	* ==========
139	>	*
140	>	* Most operations occur within work-stealing queues (in nested
141	>	* class WorkQueue).  These are special forms of Deques that
142	>	* support only three of the four possible end-operations -- push,
143	>	* pop, and poll (aka steal), under the further constraints that
144	>	* push and pop are called only from the owning thread (or, as
145	>	* extended here, under a lock), while poll may be called from
146	>	* other threads.  (If you are unfamiliar with them, you probably
147	>	* want to read Herlihy and Shavit's book "The Art of
148	>	* Multiprocessor programming", chapter 16 describing these in
149	>	* more detail before proceeding.)  The main work-stealing queue
150	>	* design is roughly similar to those in the papers "Dynamic
151	>	* Circular Work-Stealing Deque" by Chase and Lev, SPAA 2005
152	>	* (http://research.sun.com/scalable/pubs/index.html) and
153	>	* "Idempotent work stealing" by Michael, Saraswat, and Vechev,
154	>	* PPoPP 2009 (http://portal.acm.org/citation.cfm?id=1504186).
155	>	* The main differences ultimately stem from GC requirements that
156	>	* we null out taken slots as soon as we can, to maintain as small
157	>	* a footprint as possible even in programs generating huge
158	>	* numbers of tasks. To accomplish this, we shift the CAS
159	>	* arbitrating pop vs poll (steal) from being on the indices
160	>	* ("base" and "top") to the slots themselves.  So, both a
161	>	* successful pop and poll mainly entail a CAS of a slot from
162	>	* non-null to null.  Because we rely on CASes of references, we
163	>	* do not need tag bits on base or top.  They are simple ints as
164	>	* used in any circular array-based queue (see for example
165	>	* ArrayDeque).  Updates to the indices must still be ordered in a
166	>	* way that guarantees that top == base means the queue is empty,
167	>	* but otherwise may err on the side of possibly making the queue
168	>	* appear nonempty when a push, pop, or poll have not fully
169	>	* committed. Note that this means that the poll operation,
170	>	* considered individually, is not wait-free. One thief cannot
171	>	* successfully continue until another in-progress one (or, if
172	>	* previously empty, a push) completes.  However, in the
173	>	* aggregate, we ensure at least probabilistic non-blockingness.
174	>	* If an attempted steal fails, a thief always chooses a different
175	>	* random victim target to try next. So, in order for one thief to
176	>	* progress, it suffices for any in-progress poll or new push on
177	>	* any empty queue to complete. (This is why we normally use
178	>	* method pollAt and its variants that try once at the apparent
179	>	* base index, else consider alternative actions, rather than
180	>	* method poll.)
181	>	*
182	>	* This approach also enables support of 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 it.
191	>	* For example, some sets of tasks profit from cache affinities,
192	>	* 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 randomly associate submission queues
198	>	* with submitting threads, using a form of hashing.  The
199	>	* ThreadLocal Submitter class contains a value initially used as
200	>	* a hash code for choosing existing queues, but may be randomly
201	>	* repositioned upon contention with other submitters.  In
202	>	* essence, submitters act like workers except that they are
203	>	* restricted to executing local tasks that they submitted (or in
204	>	* the case of CountedCompleters, others with the same root task).
205	>	* However, because most shared/external queue operations are more
206	>	* expensive than internal, and because, at steady state, external
207	>	* submitters will compete for CPU with workers, ForkJoinTask.join
208	>	* and related methods disable them from repeatedly helping to
209	>	* process tasks if all workers are active.  Insertion of tasks in
210	>	* shared mode requires a lock (mainly to protect in the case of
211	>	* resizing) but we use only a simple spinlock (using bits in
212	>	* field qlock), because submitters encountering a busy queue move
213	>	* on to try or create other queues -- they block only when
214	>	* creating and registering new queues.
215	>	*
216	>	* Management
217	>	* ==========
218		*
219		* The main throughput advantages of work-stealing stem from
220	<	* decentralized control -- workers mostly steal tasks from each
221	<	* other. We do not want to negate this by creating bottlenecks
222	<	* implementing the management responsibilities of this class. So
223	<	* we use a collection of techniques that avoid, reduce, or cope
224	<	* well with contention. These entail several instances of
225	<	* bit-packing into CASable fields to maintain only the minimally
226	<	* required atomicity. To enable such packing, we restrict maximum
227	<	* parallelism to (1<<15)-1 (enabling twice this to fit into a 16
228	<	* bit field), which is far in excess of normal operating range.
229	<	* Even though updates to some of these bookkeeping fields do
230	<	* sometimes contend with each other, they don't normally
231	<	* cache-contend with updates to others enough to warrant memory
232	<	* padding or isolation. So they are all held as fields of
233	<	* ForkJoinPool objects.  The main capabilities are as follows:
234	<	*
235	<	* 1. Creating and removing workers. Workers are recorded in the
236	<	* "workers" array. This is an array as opposed to some other data
237	<	* structure to support index-based random steals by workers.
238	<	* Updates to the array recording new workers and unrecording
239	<	* terminated ones are protected from each other by a lock
240	<	* (workerLock) but the array is otherwise concurrently readable,
241	<	* and accessed directly by workers. To simplify index-based
242	<	* operations, the array size is always a power of two, and all
243	<	* readers must tolerate null slots. Currently, all worker thread
244	<	* creation is on-demand, triggered by task submissions,
245	<	* replacement of terminated workers, and/or compensation for
246	<	* blocked workers. However, all other support code is set up to
247	<	* work with other policies.
248	<	*
249	<	* 2. Bookkeeping for dynamically adding and removing workers. We
250	<	* maintain a given level of parallelism (or, if
251	<	* maintainsParallelism is false, at least avoid starvation). When
252	<	* some workers are known to be blocked (on joins or via
253	<	* ManagedBlocker), we may create or resume others to take their
254	<	* place until they unblock (see below). Implementing this
255	<	* requires counts of the number of "running" threads (i.e., those
256	<	* that are neither blocked nor artifically suspended) as well as
257	<	* the total number.  These two values are packed into one field,
258	<	* "workerCounts" because we need accurate snapshots when deciding
259	<	* to create, resume or suspend.  To support these decisions,
260	<	* updates to spare counts must be prospective (not
261	<	* retrospective).  For example, the running count is decremented
262	<	* before blocking by a thread about to block as a spare, but
263	<	* incremented by the thread about to unblock it. Updates upon
264	<	* resumption ofr threads blocking in awaitJoin or awaitBlocker
265	<	* cannot usually be prospective, so the running count is in
266	<	* general an upper bound of the number of productively running
267	<	* threads Updates to the workerCounts field sometimes transiently
268	<	* encounter a fair amount of contention when join dependencies
269	<	* are such that many threads block or unblock at about the same
270	<	* time. We alleviate this by sometimes bundling updates (for
271	<	* example blocking one thread on join and resuming a spare cancel
272	<	* each other out), and in most other cases performing an
273	<	* alternative action like releasing waiters or locating spares.
274	<	*
275	<	* 3. Maintaining global run state. The run state of the pool
276	<	* consists of a runLevel (SHUTDOWN, TERMINATING, etc) similar to
277	<	* those in other Executor implementations, as well as a count of
278	<	* "active" workers -- those that are, or soon will be, or
279	<	* recently were executing tasks. The runLevel and active count
280	<	* are packed together in order to correctly trigger shutdown and
281	<	* termination. Without care, active counts can be subject to very
282	<	* high contention.  We substantially reduce this contention by
283	<	* relaxing update rules.  A worker must claim active status
284	<	* prospectively, by activating if it sees that a submitted or
285	<	* stealable task exists (it may find after activating that the
286	<	* task no longer exists). It stays active while processing this
287	<	* task (if it exists) and any other local subtasks it produces,
288	<	* until it cannot find any other tasks. It then tries
289	<	* inactivating (see method preStep), but upon update contention
290	<	* instead scans for more tasks, later retrying inactivation if it
291	<	* doesn't find any.
292	<	*
293	<	* 4. Managing idle workers waiting for tasks. We cannot let
294	<	* workers spin indefinitely scanning for tasks when none are
295	<	* available. On the other hand, we must quickly prod them into
296	<	* action when new tasks are submitted or generated.  We
297	<	* park/unpark these idle workers using an event-count scheme.
298	<	* Field eventCount is incremented upon events that may enable
299	<	* workers that previously could not find a task to now find one:
300	<	* Submission of a new task to the pool, or another worker pushing
301	<	* a task onto a previously empty queue.  (We also use this
302	<	* mechanism for termination and reconfiguration actions that
303	<	* require wakeups of idle workers).  Each worker maintains its
304	<	* last known event count, and blocks when a scan for work did not
305	<	* find a task AND its lastEventCount matches the current
306	<	* eventCount. Waiting idle workers are recorded in a variant of
307	<	* Treiber stack headed by field eventWaiters which, when nonzero,
308	<	* encodes the thread index and count awaited for by the worker
309	<	* thread most recently calling eventSync. This thread in turn has
310	<	* a record (field nextEventWaiter) for the next waiting worker.
311	<	* In addition to allowing simpler decisions about need for
312	<	* wakeup, the event count bits in eventWaiters serve the role of
313	<	* tags to avoid ABA errors in Treiber stacks.  To reduce delays
314	<	* in task diffusion, workers not otherwise occupied may invoke
315	<	* method releaseWaiters, that removes and signals (unparks)
316	<	* workers not waiting on current count. To minimize task
317	<	* production stalls associate with signalling, any worker pushing
318	<	* a task on an empty queue invokes the weaker method signalWork,
319	<	* that only releases idle workers until it detects interference
320	<	* by other threads trying to release, and lets them take
321	<	* over. The net effect is a tree-like diffusion of signals, where
322	<	* released threads (and possibly others) help with unparks.  To
323	<	* further reduce contention effects a bit, failed CASes to
324	<	* increment field eventCount are tolerated without retries.
325	<	* Conceptually they are merged into the same event, which is OK
326	<	* when their only purpose is to enable workers to scan for work.
327	<	*
328	<	* 5. Managing suspension of extra workers. When a worker is about
329	<	* to block waiting for a join (or via ManagedBlockers), we may
330	<	* create a new thread to maintain parallelism level, or at least
331	<	* avoid starvation (see below). Usually, extra threads are needed
332	<	* for only very short periods, yet join dependencies are such
333	<	* that we sometimes need them in bursts. Rather than create new
334	<	* threads each time this happens, we suspend no-longer-needed
335	<	* extra ones as "spares". For most purposes, we don't distinguish
336	<	* "extra" spare threads from normal "core" threads: On each call
337	<	* to preStep (the only point at which we can do this) a worker
338	<	* checks to see if there are now too many running workers, and if
339	<	* so, suspends itself.  Methods awaitJoin and awaitBlocker look
340	<	* for suspended threads to resume before considering creating a
341	<	* new replacement. We don't need a special data structure to
342	<	* maintain spares; simply scanning the workers array looking for
343	<	* worker.isSuspended() is fine because the calling thread is
344	<	* otherwise not doing anything useful anyway; we are at least as
345	<	* happy if after locating a spare, the caller doesn't actually
346	<	* block because the join is ready before we try to adjust and
347	<	* compensate.  Note that this is intrinsically racy.  One thread
348	<	* may become a spare at about the same time as another is
349	<	* needlessly being created. We counteract this and related slop
350	<	* in part by requiring resumed spares to immediately recheck (in
351	<	* preStep) to see whether they they should re-suspend. The only
352	<	* effective difference between "extra" and "core" threads is that
353	<	* we allow the "extra" ones to time out and die if they are not
354	<	* resumed within a keep-alive interval of a few seconds. This is
355	<	* implemented mainly within ForkJoinWorkerThread, but requires
356	<	* some coordination (isTrimmed() -- meaning killed while
357	<	* suspended) to correctly maintain pool counts.
358	<	*
359	<	* 6. Deciding when to create new workers. The main dynamic
360	<	* control in this class is deciding when to create extra threads,
361	<	* in methods awaitJoin and awaitBlocker. We always
362	<	* need to create one when the number of running threads becomes
363	<	* zero. But because blocked joins are typically dependent, we
364	<	* don't necessarily need or want one-to-one replacement. Using a
365	<	* one-to-one compensation rule often leads to enough useless
366	<	* overhead creating, suspending, resuming, and/or killing threads
367	<	* to signficantly degrade throughput.  We use a rule reflecting
368	<	* the idea that, the more spare threads you already have, the
369	<	* more evidence you need to create another one. The "evidence"
370	<	* here takes two forms: (1) Using a creation threshold expressed
371	<	* in terms of the current deficit -- target minus running
372	<	* threads. To reduce flickering and drift around target values,
373	<	* the relation is quadratic: adding a spare if (dc*dc)>=(sc*pc)
374	<	* (where dc is deficit, sc is number of spare threads and pc is
375	<	* target parallelism.)  (2) Using a form of adaptive
376	<	* spionning. requiring a number of threshold checks proportional
377	<	* to the number of spare threads.  This effectively reduces churn
378	<	* at the price of systematically undershooting target parallelism
379	<	* when many threads are blocked.  However, biasing toward
380	<	* undeshooting partially compensates for the above mechanics to
381	<	* suspend extra threads, that normally lead to overshoot because
382	<	* we can only suspend workers in-between top-level actions. It
383	<	* also better copes with the fact that some of the methods in
384	<	* this class tend to never become compiled (but are interpreted),
385	<	* so some components of the entire set of controls might execute
386	<	* many times faster than others. And similarly for cases where
387	<	* the apparent lack of work is just due to GC stalls and other
388	<	* transient system activity.
389	<	*
390	<	* 7. Maintaining other configuration parameters and monitoring
391	<	* statistics. Updates to fields controlling parallelism level,
392	<	* max size, etc can only meaningfully take effect for individual
393	<	* threads upon their next top-level actions; i.e., between
394	<	* stealing/running tasks/submission, which are separated by calls
395	<	* to preStep.  Memory ordering for these (assumed infrequent)
396	<	* reconfiguration calls is ensured by using reads and writes to
397	<	* volatile field workerCounts (that must be read in preStep anyway)
398	<	* as "fences" -- user-level reads are preceded by reads of
399	<	* workCounts, and writes are followed by no-op CAS to
400	<	* workerCounts. The values reported by other management and
401	<	* monitoring methods are either computed on demand, or are kept
402	<	* in fields that are only updated when threads are otherwise
403	<	* idle.
404	<	*
405	<	* Beware that there is a lot of representation-level coupling
406	<	* among classes ForkJoinPool, ForkJoinWorkerThread, and
407	<	* ForkJoinTask.  For example, direct access to "workers" array by
408	<	* workers, and direct access to ForkJoinTask.status by both
409	<	* ForkJoinPool and ForkJoinWorkerThread.  There is little point
220	>	* decentralized control -- workers mostly take tasks from
221	>	* themselves or each other. We cannot negate this in the
222	>	* implementation of other management responsibilities. The main
223	>	* tactic for avoiding bottlenecks is packing nearly all
224	>	* essentially atomic control state into two volatile variables
225	>	* that are by far most often read (not written) as status and
226	>	* consistency checks.
227	>	*
228	>	* Field "ctl" contains 64 bits holding all the information needed
229	>	* to atomically decide to add, inactivate, enqueue (on an event
230	>	* queue), dequeue, and/or re-activate workers.  To enable this
231	>	* packing, we restrict maximum parallelism to (1<<15)-1 (which is
232	>	* far in excess of normal operating range) to allow ids, counts,
233	>	* and their negations (used for thresholding) to fit into 16bit
234	>	* fields.
235	>	*
236	>	* Field "plock" is a form of sequence lock with a saturating
237	>	* shutdown bit (similarly for per-queue "qlocks"), mainly
238	>	* protecting updates to the workQueues array, as well as to
239	>	* enable shutdown.  When used as a lock, it is normally only very
240	>	* briefly held, so is nearly always available after at most a
241	>	* brief spin, but we use a monitor-based backup strategy to
242	>	* block when needed.
243	>	*
244	>	* Recording WorkQueues.  WorkQueues are recorded in the
245	>	* "workQueues" array that is created upon first use and expanded
246	>	* if necessary.  Updates to the array while recording new workers
247	>	* and unrecording terminated ones are protected from each other
248	>	* by a lock but the array is otherwise concurrently readable, and
249	>	* accessed directly.  To simplify index-based operations, the
250	>	* array size is always a power of two, and all readers must
251	>	* tolerate null slots. Worker queues are at odd indices. Shared
252	>	* (submission) queues are at even indices, up to a maximum of 64
253	>	* slots, to limit growth even if array needs to expand to add
254	>	* more workers. Grouping them together in this way simplifies and
255	>	* speeds up task scanning.
256	>	*
257	>	* All worker thread creation is on-demand, triggered by task
258	>	* submissions, replacement of terminated workers, and/or
259	>	* compensation for blocked workers. However, all other support
260	>	* code is set up to work with other policies.  To ensure that we
261	>	* do not hold on to worker references that would prevent GC, ALL
262	>	* accesses to workQueues are via indices into the workQueues
263	>	* array (which is one source of some of the messy code
264	>	* constructions here). In essence, the workQueues array serves as
265	>	* a weak reference mechanism. Thus for example the wait queue
266	>	* field of ctl stores indices, not references.  Access to the
267	>	* workQueues in associated methods (for example signalWork) must
268	>	* both index-check and null-check the IDs. All such accesses
269	>	* ignore bad IDs by returning out early from what they are doing,
270	>	* since this can only be associated with termination, in which
271	>	* case it is OK to give up.  All uses of the workQueues array
272	>	* also check that it is non-null (even if previously
273	>	* non-null). This allows nulling during termination, which is
274	>	* currently not necessary, but remains an option for
275	>	* resource-revocation-based shutdown schemes. It also helps
276	>	* reduce JIT issuance of uncommon-trap code, which tends to
277	>	* unnecessarily complicate control flow in some methods.
278	>	*
279	>	* Event Queuing. Unlike HPC work-stealing frameworks, we cannot
280	>	* let workers spin indefinitely scanning for tasks when none can
281	>	* be found immediately, and we cannot start/resume workers unless
282	>	* there appear to be tasks available.  On the other hand, we must
283	>	* quickly prod them into action when new tasks are submitted or
284	>	* generated. In many usages, ramp-up time to activate workers is
285	>	* the main limiting factor in overall performance (this is
286	>	* compounded at program start-up by JIT compilation and
287	>	* allocation). So we try to streamline this as much as possible.
288	>	* We park/unpark workers after placing in an event wait queue
289	>	* when they cannot find work. This "queue" is actually a simple
290	>	* Treiber stack, headed by the "id" field of ctl, plus a 15bit
291	>	* counter value (that reflects the number of times a worker has
292	>	* been inactivated) to avoid ABA effects (we need only as many
293	>	* version numbers as worker threads). Successors are held in
294	>	* field WorkQueue.nextWait.  Queuing deals with several intrinsic
295	>	* races, mainly that a task-producing thread can miss seeing (and
296	>	* signalling) another thread that gave up looking for work but
297	>	* has not yet entered the wait queue. We solve this by requiring
298	>	* a full sweep of all workers (via repeated calls to method
299	>	* scan()) both before and after a newly waiting worker is added
300	>	* to the wait queue. During a rescan, the worker might release
301	>	* some other queued worker rather than itself, which has the same
302	>	* net effect. Because enqueued workers may actually be rescanning
303	>	* rather than waiting, we set and clear the "parker" field of
304	>	* WorkQueues to reduce unnecessary calls to unpark.  (This
305	>	* requires a secondary recheck to avoid missed signals.)  Note
306	>	* the unusual conventions about Thread.interrupts surrounding
307	>	* parking and other blocking: Because interrupts are used solely
308	>	* to alert threads to check termination, which is checked anyway
309	>	* upon blocking, we clear status (using Thread.interrupted)
310	>	* before any call to park, so that park does not immediately
311	>	* return due to status being set via some other unrelated call to
312	>	* interrupt in user code.
313	>	*
314	>	* Signalling.  We create or wake up workers only when there
315	>	* appears to be at least one task they might be able to find and
316	>	* execute. However, many other threads may notice the same task
317	>	* and each signal to wake up a thread that might take it. So in
318	>	* general, pools will be over-signalled.  When a submission is
319	>	* added or another worker adds a task to a queue that has fewer
320	>	* than two tasks, they signal waiting workers (or trigger
321	>	* creation of new ones if fewer than the given parallelism level
322	>	* -- signalWork), and may leave a hint to the unparked worker to
323	>	* help signal others upon wakeup).  These primary signals are
324	>	* buttressed by others (see method helpSignal) whenever other
325	>	* threads scan for work or do not have a task to process.  On
326	>	* most platforms, signalling (unpark) overhead time is noticeably
327	>	* long, and the time between signalling a thread and it actually
328	>	* making progress can be very noticeably long, so it is worth
329	>	* offloading these delays from critical paths as much as
330	>	* possible.
331	>	*
332	>	* Trimming workers. To release resources after periods of lack of
333	>	* use, a worker starting to wait when the pool is quiescent will
334	>	* time out and terminate if the pool has remained quiescent for a
335	>	* given period -- a short period if there are more threads than
336	>	* parallelism, longer as the number of threads decreases. This
337	>	* will slowly propagate, eventually terminating all workers after
338	>	* periods of non-use.
339	>	*
340	>	* Shutdown and Termination. A call to shutdownNow atomically sets
341	>	* a plock bit and then (non-atomically) sets each worker's
342	>	* qlock status, cancels all unprocessed tasks, and wakes up
343	>	* all waiting workers.  Detecting whether termination should
344	>	* commence after a non-abrupt shutdown() call requires more work
345	>	* and bookkeeping. We need consensus about quiescence (i.e., that
346	>	* there is no more work). The active count provides a primary
347	>	* indication but non-abrupt shutdown still requires a rechecking
348	>	* scan for any workers that are inactive but not queued.
349	>	*
350	>	* Joining Tasks
351	>	* =============
352	>	*
353	>	* Any of several actions may be taken when one worker is waiting
354	>	* to join a task stolen (or always held) by another.  Because we
355	>	* are multiplexing many tasks on to a pool of workers, we can't
356	>	* just let them block (as in Thread.join).  We also cannot just
357	>	* reassign the joiner's run-time stack with another and replace
358	>	* it later, which would be a form of "continuation", that even if
359	>	* possible is not necessarily a good idea since we sometimes need
360	>	* both an unblocked task and its continuation to progress.
361	>	* Instead we combine two tactics:
362	>	*
363	>	*   Helping: Arranging for the joiner to execute some task that it
364	>	*      would be running if the steal had not occurred.
365	>	*
366	>	*   Compensating: Unless there are already enough live threads,
367	>	*      method tryCompensate() may create or re-activate a spare
368	>	*      thread to compensate for blocked joiners until they unblock.
369	>	*
370	>	* A third form (implemented in tryRemoveAndExec) amounts to
371	>	* helping a hypothetical compensator: If we can readily tell that
372	>	* a possible action of a compensator is to steal and execute the
373	>	* task being joined, the joining thread can do so directly,
374	>	* without the need for a compensation thread (although at the
375	>	* expense of larger run-time stacks, but the tradeoff is
376	>	* typically worthwhile).
377	>	*
378	>	* The ManagedBlocker extension API can't use helping so relies
379	>	* only on compensation in method awaitBlocker.
380	>	*
381	>	* The algorithm in tryHelpStealer entails a form of "linear"
382	>	* helping: Each worker records (in field currentSteal) the most
383	>	* recent task it stole from some other worker. Plus, it records
384	>	* (in field currentJoin) the task it is currently actively
385	>	* joining. Method tryHelpStealer uses these markers to try to
386	>	* find a worker to help (i.e., steal back a task from and execute
387	>	* it) that could hasten completion of the actively joined task.
388	>	* In essence, the joiner executes a task that would be on its own
389	>	* local deque had the to-be-joined task not been stolen. This may
390	>	* be seen as a conservative variant of the approach in Wagner &
391	>	* Calder "Leapfrogging: a portable technique for implementing
392	>	* efficient futures" SIGPLAN Notices, 1993
393	>	* (http://portal.acm.org/citation.cfm?id=155354). It differs in
394	>	* that: (1) We only maintain dependency links across workers upon
395	>	* steals, rather than use per-task bookkeeping.  This sometimes
396	>	* requires a linear scan of workQueues array to locate stealers,
397	>	* but often doesn't because stealers leave hints (that may become
398	>	* stale/wrong) of where to locate them.  It is only a hint
399	>	* because a worker might have had multiple steals and the hint
400	>	* records only one of them (usually the most current).  Hinting
401	>	* isolates cost to when it is needed, rather than adding to
402	>	* per-task overhead.  (2) It is "shallow", ignoring nesting and
403	>	* potentially cyclic mutual steals.  (3) It is intentionally
404	>	* racy: field currentJoin is updated only while actively joining,
405	>	* which means that we miss links in the chain during long-lived
406	>	* tasks, GC stalls etc (which is OK since blocking in such cases
407	>	* is usually a good idea).  (4) We bound the number of attempts
408	>	* to find work (see MAX_HELP) and fall back to suspending the
409	>	* worker and if necessary replacing it with another.
410	>	*
411	>	* Helping actions for CountedCompleters are much simpler: Method
412	>	* helpComplete can take and execute any task with the same root
413	>	* as the task being waited on. However, this still entails some
414	>	* traversal of completer chains, so is less efficient than using
415	>	* CountedCompleters without explicit joins.
416	>	*
417	>	* It is impossible to keep exactly the target parallelism number
418	>	* of threads running at any given time.  Determining the
419	>	* existence of conservatively safe helping targets, the
420	>	* availability of already-created spares, and the apparent need
421	>	* to create new spares are all racy, so we rely on multiple
422	>	* retries of each.  Compensation in the apparent absence of
423	>	* helping opportunities is challenging to control on JVMs, where
424	>	* GC and other activities can stall progress of tasks that in
425	>	* turn stall out many other dependent tasks, without us being
426	>	* able to determine whether they will ever require compensation.
427	>	* Even though work-stealing otherwise encounters little
428	>	* degradation in the presence of more threads than cores,
429	>	* aggressively adding new threads in such cases entails risk of
430	>	* unwanted positive feedback control loops in which more threads
431	>	* cause more dependent stalls (as well as delayed progress of
432	>	* unblocked threads to the point that we know they are available)
433	>	* leading to more situations requiring more threads, and so
434	>	* on. This aspect of control can be seen as an (analytically
435	>	* intractable) game with an opponent that may choose the worst
436	>	* (for us) active thread to stall at any time.  We take several
437	>	* precautions to bound losses (and thus bound gains), mainly in
438	>	* methods tryCompensate and awaitJoin.
439	>	*
440	>	* Common Pool
441	>	* ===========
442	>	*
443	>	* The static common Pool always exists after static
444	>	* initialization.  Since it (or any other created pool) need
445	>	* never be used, we minimize initial construction overhead and
446	>	* footprint to the setup of about a dozen fields, with no nested
447	>	* allocation. Most bootstrapping occurs within method
448	>	* fullExternalPush during the first submission to the pool.
449	>	*
450	>	* When external threads submit to the common pool, they can
451	>	* perform some subtask processing (see externalHelpJoin and
452	>	* related methods).  We do not need to record whether these
453	>	* submissions are to the common pool -- if not, externalHelpJoin
454	>	* returns quickly (at the most helping to signal some common pool
455	>	* workers). These submitters would otherwise be blocked waiting
456	>	* for completion, so the extra effort (with liberally sprinkled
457	>	* task status checks) in inapplicable cases amounts to an odd
458	>	* form of limited spin-wait before blocking in ForkJoinTask.join.
459	>	*
460	>	* Style notes
461	>	* ===========
462	>	*
463	>	* There is a lot of representation-level coupling among classes
464	>	* ForkJoinPool, ForkJoinWorkerThread, and ForkJoinTask.  The
465	>	* fields of WorkQueue maintain data structures managed by
466	>	* ForkJoinPool, so are directly accessed.  There is little point
467		* trying to reduce this, since any associated future changes in
468		* representations will need to be accompanied by algorithmic
469	<	* changes anyway.
470	<	*
471	<	* Style notes: There are lots of inline assignments (of form
472	<	* "while ((local = field) != 0)") which are usually the simplest
473	<	* way to ensure read orderings. Also several occurrences of the
474	<	* unusual "do {} while(!cas...)" which is the simplest way to
475	<	* force an update of a CAS'ed variable. There are also a few
476	<	* other coding oddities that help some methods perform reasonably
477	<	* even when interpreted (not compiled).
478	<	*
479	<	* The order of declarations in this file is: (1) statics (2)
480	<	* fields (along with constants used when unpacking some of them)
481	<	* (3) internal control methods (4) callbacks and other support
482	<	* for ForkJoinTask and ForkJoinWorkerThread classes, (5) exported
483	<	* methods (plus a few little helpers).
469	>	* changes anyway. Several methods intrinsically sprawl because
470	>	* they must accumulate sets of consistent reads of volatiles held
471	>	* in local variables.  Methods signalWork() and scan() are the
472	>	* main bottlenecks, so are especially heavily
473	>	* micro-optimized/mangled.  There are lots of inline assignments
474	>	* (of form "while ((local = field) != 0)") which are usually the
475	>	* simplest way to ensure the required read orderings (which are
476	>	* sometimes critical). This leads to a "C"-like style of listing
477	>	* declarations of these locals at the heads of methods or blocks.
478	>	* There are several occurrences of the unusual "do {} while
479	>	* (!cas...)"  which is the simplest way to force an update of a
480	>	* CAS'ed variable. There are also other coding oddities (including
481	>	* several unnecessary-looking hoisted null checks) that help
482	>	* some methods perform reasonably even when interpreted (not
483	>	* compiled).
484	>	*
485	>	* The order of declarations in this file is:
486	>	* (1) Static utility functions
487	>	* (2) Nested (static) classes
488	>	* (3) Static fields
489	>	* (4) Fields, along with constants used when unpacking some of them
490	>	* (5) Internal control methods
491	>	* (6) Callbacks and other support for ForkJoinTask methods
492	>	* (7) Exported methods
493	>	* (8) Static block initializing statics in minimally dependent order
494	>	*/
495	>
496	>	// Static utilities
497	>
498	>	/**
499	>	* If there is a security manager, makes sure caller has
500	>	* permission to modify threads.
501		*/
502	+	private static void checkPermission() {
503	+	SecurityManager security = System.getSecurityManager();
504	+	if (security != null)
505	+	security.checkPermission(modifyThreadPermission);
506	+	}
507	+
508	+	// Nested classes
509
510		/**
511		* Factory for creating new {@link ForkJoinWorkerThread}s.
#	Line 346 | Line 527 | public class ForkJoinPool extends Abstra
527		* Default ForkJoinWorkerThreadFactory implementation; creates a
528		* new ForkJoinWorkerThread.
529		*/
530	<	static class  DefaultForkJoinWorkerThreadFactory
530	>	static final class DefaultForkJoinWorkerThreadFactory
531		implements ForkJoinWorkerThreadFactory {
532	<	public ForkJoinWorkerThread newThread(ForkJoinPool pool) {
532	>	public final ForkJoinWorkerThread newThread(ForkJoinPool pool) {
533		return new ForkJoinWorkerThread(pool);
534		}
535		}
536
537		/**
538	<	* Creates a new ForkJoinWorkerThread. This factory is used unless
539	<	* overridden in ForkJoinPool constructors.
538	>	* Per-thread records for threads that submit to pools. Currently
539	>	* holds only pseudo-random seed / index that is used to choose
540	>	* submission queues in method externalPush. In the future, this may
541	>	* also incorporate a means to implement different task rejection
542	>	* and resubmission policies.
543	>	*
544	>	* Seeds for submitters and workers/workQueues work in basically
545	>	* the same way but are initialized and updated using slightly
546	>	* different mechanics. Both are initialized using the same
547	>	* approach as in class ThreadLocal, where successive values are
548	>	* unlikely to collide with previous values. Seeds are then
549	>	* randomly modified upon collisions using xorshifts, which
550	>	* requires a non-zero seed.
551	>	*/
552	>	static final class Submitter {
553	>	int seed;
554	>	Submitter(int s) { seed = s; }
555	>	}
556	>
557	>	/**
558	>	* Class for artificial tasks that are used to replace the target
559	>	* of local joins if they are removed from an interior queue slot
560	>	* in WorkQueue.tryRemoveAndExec. We don't need the proxy to
561	>	* actually do anything beyond having a unique identity.
562	>	*/
563	>	static final class EmptyTask extends ForkJoinTask<Void> {
564	>	private static final long serialVersionUID = -7721805057305804111L;
565	>	EmptyTask() { status = ForkJoinTask.NORMAL; } // force done
566	>	public final Void getRawResult() { return null; }
567	>	public final void setRawResult(Void x) {}
568	>	public final boolean exec() { return true; }
569	>	}
570	>
571	>	/**
572	>	* Queues supporting work-stealing as well as external task
573	>	* submission. See above for main rationale and algorithms.
574	>	* Implementation relies heavily on "Unsafe" intrinsics
575	>	* and selective use of "volatile":
576	>	*
577	>	* Field "base" is the index (mod array.length) of the least valid
578	>	* queue slot, which is always the next position to steal (poll)
579	>	* from if nonempty. Reads and writes require volatile orderings
580	>	* but not CAS, because updates are only performed after slot
581	>	* CASes.
582	>	*
583	>	* Field "top" is the index (mod array.length) of the next queue
584	>	* slot to push to or pop from. It is written only by owner thread
585	>	* for push, or under lock for external/shared push, and accessed
586	>	* by other threads only after reading (volatile) base.  Both top
587	>	* and base are allowed to wrap around on overflow, but (top -
588	>	* base) (or more commonly -(base - top) to force volatile read of
589	>	* base before top) still estimates size. The lock ("qlock") is
590	>	* forced to -1 on termination, causing all further lock attempts
591	>	* to fail. (Note: we don't need CAS for termination state because
592	>	* upon pool shutdown, all shared-queues will stop being used
593	>	* anyway.)  Nearly all lock bodies are set up so that exceptions
594	>	* within lock bodies are "impossible" (modulo JVM errors that
595	>	* would cause failure anyway.)
596	>	*
597	>	* The array slots are read and written using the emulation of
598	>	* volatiles/atomics provided by Unsafe. Insertions must in
599	>	* general use putOrderedObject as a form of releasing store to
600	>	* ensure that all writes to the task object are ordered before
601	>	* its publication in the queue.  All removals entail a CAS to
602	>	* null.  The array is always a power of two. To ensure safety of
603	>	* Unsafe array operations, all accesses perform explicit null
604	>	* checks and implicit bounds checks via power-of-two masking.
605	>	*
606	>	* In addition to basic queuing support, this class contains
607	>	* fields described elsewhere to control execution. It turns out
608	>	* to work better memory-layout-wise to include them in this class
609	>	* rather than a separate class.
610	>	*
611	>	* Performance on most platforms is very sensitive to placement of
612	>	* instances of both WorkQueues and their arrays -- we absolutely
613	>	* do not want multiple WorkQueue instances or multiple queue
614	>	* arrays sharing cache lines. (It would be best for queue objects
615	>	* and their arrays to share, but there is nothing available to
616	>	* help arrange that).  Unfortunately, because they are recorded
617	>	* in a common array, WorkQueue instances are often moved to be
618	>	* adjacent by garbage collectors. To reduce impact, we use field
619	>	* padding that works OK on common platforms; this effectively
620	>	* trades off slightly slower average field access for the sake of
621	>	* avoiding really bad worst-case access. (Until better JVM
622	>	* support is in place, this padding is dependent on transient
623	>	* properties of JVM field layout rules.) We also take care in
624	>	* allocating, sizing and resizing the array. Non-shared queue
625	>	* arrays are initialized by workers before use. Others are
626	>	* allocated on first use.
627		*/
628	<	public static final ForkJoinWorkerThreadFactory
629	<	defaultForkJoinWorkerThreadFactory =
630	<	new DefaultForkJoinWorkerThreadFactory();
628	>	static final class WorkQueue {
629	>	/**
630	>	* Capacity of work-stealing queue array upon initialization.
631	>	* Must be a power of two; at least 4, but should be larger to
632	>	* reduce or eliminate cacheline sharing among queues.
633	>	* Currently, it is much larger, as a partial workaround for
634	>	* the fact that JVMs often place arrays in locations that
635	>	* share GC bookkeeping (especially cardmarks) such that
636	>	* per-write accesses encounter serious memory contention.
637	>	*/
638	>	static final int INITIAL_QUEUE_CAPACITY = 1 << 13;
639
640	<	/**
641	<	* Permission required for callers of methods that may start or
642	<	* kill threads.
643	<	*/
644	<	private static final RuntimePermission modifyThreadPermission =
645	<	new RuntimePermission("modifyThread");
640	>	/**
641	>	* Maximum size for queue arrays. Must be a power of two less
642	>	* than or equal to 1 << (31 - width of array entry) to ensure
643	>	* lack of wraparound of index calculations, but defined to a
644	>	* value a bit less than this to help users trap runaway
645	>	* programs before saturating systems.
646	>	*/
647	>	static final int MAXIMUM_QUEUE_CAPACITY = 1 << 26; // 64M
648
649	<	/**
650	<	* If there is a security manager, makes sure caller has
373	<	* permission to modify threads.
374	<	*/
375	<	private static void checkPermission() {
376	<	SecurityManager security = System.getSecurityManager();
377	<	if (security != null)
378	<	security.checkPermission(modifyThreadPermission);
379	<	}
649	>	// Heuristic padding to ameliorate unfortunate memory placements
650	>	volatile long pad00, pad01, pad02, pad03, pad04, pad05, pad06;
651
652	<	/**
653	<	* Generator for assigning sequence numbers as pool names.
654	<	*/
655	<	private static final AtomicInteger poolNumberGenerator =
656	<	new AtomicInteger();
652	>	int seed;                  // for random scanning; initialize nonzero
653	>	volatile int eventCount;   // encoded inactivation count; < 0 if inactive
654	>	int nextWait;              // encoded record of next event waiter
655	>	int hint;                  // steal or signal hint (index)
656	>	int poolIndex;             // index of this queue in pool (or 0)
657	>	final int mode;            // 0: lifo, > 0: fifo, < 0: shared
658	>	int nsteals;               // number of steals
659	>	volatile int qlock;        // 1: locked, -1: terminate; else 0
660	>	volatile int base;         // index of next slot for poll
661	>	int top;                   // index of next slot for push
662	>	ForkJoinTask<?>[] array;   // the elements (initially unallocated)
663	>	final ForkJoinPool pool;   // the containing pool (may be null)
664	>	final ForkJoinWorkerThread owner; // owning thread or null if shared
665	>	volatile Thread parker;    // == owner during call to park; else null
666	>	volatile ForkJoinTask<?> currentJoin;  // task being joined in awaitJoin
667	>	ForkJoinTask<?> currentSteal; // current non-local task being executed
668	>
669	>	volatile Object pad10, pad11, pad12, pad13, pad14, pad15, pad16, pad17;
670	>	volatile Object pad18, pad19, pad1a, pad1b, pad1c, pad1d;
671	>
672	>	WorkQueue(ForkJoinPool pool, ForkJoinWorkerThread owner, int mode,
673	>	int seed) {
674	>	this.pool = pool;
675	>	this.owner = owner;
676	>	this.mode = mode;
677	>	this.seed = seed;
678	>	// Place indices in the center of array (that is not yet allocated)
679	>	base = top = INITIAL_QUEUE_CAPACITY >>> 1;
680	>	}
681
682	<	/**
683	<	* Absolute bound for parallelism level. Twice this number must
684	<	* fit into a 16bit field to enable word-packing for some counts.
685	<	*/
686	<	private static final int MAX_THREADS = 0x7fff;
682	>	/**
683	>	* Returns the approximate number of tasks in the queue.
684	>	*/
685	>	final int queueSize() {
686	>	int n = base - top;       // non-owner callers must read base first
687	>	return (n >= 0) ? 0 : -n; // ignore transient negative
688	>	}
689
690	<	/**
691	<	* Array holding all worker threads in the pool.  Array size must
692	<	* be a power of two.  Updates and replacements are protected by
693	<	* workerLock, but the array is always kept in a consistent enough
694	<	* state to be randomly accessed without locking by workers
695	<	* performing work-stealing, as well as other traversal-based
696	<	* methods in this class. All readers must tolerate that some
697	<	* array slots may be null.
698	<	*/
699	<	volatile ForkJoinWorkerThread[] workers;
690	>	/**
691	>	* Provides a more accurate estimate of whether this queue has
692	>	* any tasks than does queueSize, by checking whether a
693	>	* near-empty queue has at least one unclaimed task.
694	>	*/
695	>	final boolean isEmpty() {
696	>	ForkJoinTask<?>[] a; int m, s;
697	>	int n = base - (s = top);
698	>	return (n >= 0 ||
699	>	(n == -1 &&
700	>	((a = array) == null ||
701	>	(m = a.length - 1) < 0 ||
702	>	U.getObject
703	>	(a, (long)((m & (s - 1)) << ASHIFT) + ABASE) == null)));
704	>	}
705
706	<	/**
707	<	* Queue for external submissions.
708	<	*/
709	<	private final LinkedTransferQueue<ForkJoinTask<?>> submissionQueue;
706	>	/**
707	>	* Pushes a task. Call only by owner in unshared queues.  (The
708	>	* shared-queue version is embedded in method externalPush.)
709	>	*
710	>	* @param task the task. Caller must ensure non-null.
711	>	* @throw RejectedExecutionException if array cannot be resized
712	>	*/
713	>	final void push(ForkJoinTask<?> task) {
714	>	ForkJoinTask<?>[] a; ForkJoinPool p;
715	>	int s = top, m, n;
716	>	if ((a = array) != null) {    // ignore if queue removed
717	>	int j = (((m = a.length - 1) & s) << ASHIFT) + ABASE;
718	>	U.putOrderedObject(a, j, task);
719	>	if ((n = (top = s + 1) - base) <= 2) {
720	>	if ((p = pool) != null)
721	>	p.signalWork(this);
722	>	}
723	>	else if (n >= m)
724	>	growArray();
725	>	}
726	>	}
727
728	<	/**
729	<	* Lock protecting updates to workers array.
730	<	*/
731	<	private final ReentrantLock workerLock;
728	>	/**
729	>	* Initializes or doubles the capacity of array. Call either
730	>	* by owner or with lock held -- it is OK for base, but not
731	>	* top, to move while resizings are in progress.
732	>	*/
733	>	final ForkJoinTask<?>[] growArray() {
734	>	ForkJoinTask<?>[] oldA = array;
735	>	int size = oldA != null ? oldA.length << 1 : INITIAL_QUEUE_CAPACITY;
736	>	if (size > MAXIMUM_QUEUE_CAPACITY)
737	>	throw new RejectedExecutionException("Queue capacity exceeded");
738	>	int oldMask, t, b;
739	>	ForkJoinTask<?>[] a = array = new ForkJoinTask<?>[size];
740	>	if (oldA != null && (oldMask = oldA.length - 1) >= 0 &&
741	>	(t = top) - (b = base) > 0) {
742	>	int mask = size - 1;
743	>	do {
744	>	ForkJoinTask<?> x;
745	>	int oldj = ((b & oldMask) << ASHIFT) + ABASE;
746	>	int j    = ((b &    mask) << ASHIFT) + ABASE;
747	>	x = (ForkJoinTask<?>)U.getObjectVolatile(oldA, oldj);
748	>	if (x != null &&
749	>	U.compareAndSwapObject(oldA, oldj, x, null))
750	>	U.putObjectVolatile(a, j, x);
751	>	} while (++b != t);
752	>	}
753	>	return a;
754	>	}
755
756	<	/**
757	<	* Latch released upon termination.
758	<	*/
759	<	private final CountDownLatch terminationLatch;
756	>	/**
757	>	* Takes next task, if one exists, in LIFO order.  Call only
758	>	* by owner in unshared queues.
759	>	*/
760	>	final ForkJoinTask<?> pop() {
761	>	ForkJoinTask<?>[] a; ForkJoinTask<?> t; int m;
762	>	if ((a = array) != null && (m = a.length - 1) >= 0) {
763	>	for (int s; (s = top - 1) - base >= 0;) {
764	>	long j = ((m & s) << ASHIFT) + ABASE;
765	>	if ((t = (ForkJoinTask<?>)U.getObject(a, j)) == null)
766	>	break;
767	>	if (U.compareAndSwapObject(a, j, t, null)) {
768	>	top = s;
769	>	return t;
770	>	}
771	>	}
772	>	}
773	>	return null;
774	>	}
775
776	<	/**
777	<	* Creation factory for worker threads.
778	<	*/
779	<	private final ForkJoinWorkerThreadFactory factory;
776	>	/**
777	>	* Takes a task in FIFO order if b is base of queue and a task
778	>	* can be claimed without contention. Specialized versions
779	>	* appear in ForkJoinPool methods scan and tryHelpStealer.
780	>	*/
781	>	final ForkJoinTask<?> pollAt(int b) {
782	>	ForkJoinTask<?> t; ForkJoinTask<?>[] a;
783	>	if ((a = array) != null) {
784	>	int j = (((a.length - 1) & b) << ASHIFT) + ABASE;
785	>	if ((t = (ForkJoinTask<?>)U.getObjectVolatile(a, j)) != null &&
786	>	base == b &&
787	>	U.compareAndSwapObject(a, j, t, null)) {
788	>	base = b + 1;
789	>	return t;
790	>	}
791	>	}
792	>	return null;
793	>	}
794
795	<	/**
796	<	* Sum of per-thread steal counts, updated only when threads are
797	<	* idle or terminating.
798	<	*/
799	<	private volatile long stealCount;
795	>	/**
796	>	* Takes next task, if one exists, in FIFO order.
797	>	*/
798	>	final ForkJoinTask<?> poll() {
799	>	ForkJoinTask<?>[] a; int b; ForkJoinTask<?> t;
800	>	while ((b = base) - top < 0 && (a = array) != null) {
801	>	int j = (((a.length - 1) & b) << ASHIFT) + ABASE;
802	>	t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
803	>	if (t != null) {
804	>	if (base == b &&
805	>	U.compareAndSwapObject(a, j, t, null)) {
806	>	base = b + 1;
807	>	return t;
808	>	}
809	>	}
810	>	else if (base == b) {
811	>	if (b + 1 == top)
812	>	break;
813	>	Thread.yield(); // wait for lagging update (very rare)
814	>	}
815	>	}
816	>	return null;
817	>	}
818
819	<	/**
820	<	* Encoded record of top of treiber stack of threads waiting for
821	<	* events. The top 32 bits contain the count being waited for. The
822	<	* bottom word contains one plus the pool index of waiting worker
823	<	* thread.
824	<	*/
436	<	private volatile long eventWaiters;
819	>	/**
820	>	* Takes next task, if one exists, in order specified by mode.
821	>	*/
822	>	final ForkJoinTask<?> nextLocalTask() {
823	>	return mode == 0 ? pop() : poll();
824	>	}
825
826	<	private static final int  EVENT_COUNT_SHIFT = 32;
827	<	private static final long WAITER_INDEX_MASK = (1L << EVENT_COUNT_SHIFT)-1L;
826	>	/**
827	>	* Returns next task, if one exists, in order specified by mode.
828	>	*/
829	>	final ForkJoinTask<?> peek() {
830	>	ForkJoinTask<?>[] a = array; int m;
831	>	if (a == null || (m = a.length - 1) < 0)
832	>	return null;
833	>	int i = mode == 0 ? top - 1 : base;
834	>	int j = ((i & m) << ASHIFT) + ABASE;
835	>	return (ForkJoinTask<?>)U.getObjectVolatile(a, j);
836	>	}
837
838	<	/**
839	<	* A counter for events that may wake up worker threads:
840	<	*   - Submission of a new task to the pool
841	<	*   - A worker pushing a task on an empty queue
842	<	*   - termination and reconfiguration
843	<	*/
844	<	private volatile int eventCount;
838	>	/**
839	>	* Pops the given task only if it is at the current top.
840	>	* (A shared version is available only via FJP.tryExternalUnpush)
841	>	*/
842	>	final boolean tryUnpush(ForkJoinTask<?> t) {
843	>	ForkJoinTask<?>[] a; int s;
844	>	if ((a = array) != null && (s = top) != base &&
845	>	U.compareAndSwapObject
846	>	(a, (((a.length - 1) & --s) << ASHIFT) + ABASE, t, null)) {
847	>	top = s;
848	>	return true;
849	>	}
850	>	return false;
851	>	}
852
853	<	/**
854	<	* Lifecycle control. The low word contains the number of workers
855	<	* that are (probably) executing tasks. This value is atomically
856	<	* incremented before a worker gets a task to run, and decremented
857	<	* when worker has no tasks and cannot find any.  Bits 16-18
858	<	* contain runLevel value. When all are zero, the pool is
859	<	* running. Level transitions are monotonic (running -> shutdown
860	<	* -> terminating -> terminated) so each transition adds a bit.
861	<	* These are bundled together to ensure consistent read for
458	<	* termination checks (i.e., that runLevel is at least SHUTDOWN
459	<	* and active threads is zero).
460	<	*/
461	<	private volatile int runState;
853	>	/**
854	>	* Removes and cancels all known tasks, ignoring any exceptions.
855	>	*/
856	>	final void cancelAll() {
857	>	ForkJoinTask.cancelIgnoringExceptions(currentJoin);
858	>	ForkJoinTask.cancelIgnoringExceptions(currentSteal);
859	>	for (ForkJoinTask<?> t; (t = poll()) != null; )
860	>	ForkJoinTask.cancelIgnoringExceptions(t);
861	>	}
862
863	<	// Note: The order among run level values matters.
864	<	private static final int RUNLEVEL_SHIFT     = 16;
865	<	private static final int SHUTDOWN           = 1 << RUNLEVEL_SHIFT;
866	<	private static final int TERMINATING        = 1 << (RUNLEVEL_SHIFT + 1);
867	<	private static final int TERMINATED         = 1 << (RUNLEVEL_SHIFT + 2);
868	<	private static final int ACTIVE_COUNT_MASK  = (1 << RUNLEVEL_SHIFT) - 1;
869	<	private static final int ONE_ACTIVE         = 1; // active update delta
863	>	/**
864	>	* Computes next value for random probes.  Scans don't require
865	>	* a very high quality generator, but also not a crummy one.
866	>	* Marsaglia xor-shift is cheap and works well enough.  Note:
867	>	* This is manually inlined in its usages in ForkJoinPool to
868	>	* avoid writes inside busy scan loops.
869	>	*/
870	>	final int nextSeed() {
871	>	int r = seed;
872	>	r ^= r << 13;
873	>	r ^= r >>> 17;
874	>	return seed = r ^= r << 5;
875	>	}
876
877	<	/**
472	<	* Holds number of total (i.e., created and not yet terminated)
473	<	* and running (i.e., not blocked on joins or other managed sync)
474	<	* threads, packed together to ensure consistent snapshot when
475	<	* making decisions about creating and suspending spare
476	<	* threads. Updated only by CAS. Note that adding a new worker
477	<	* requires incrementing both counts, since workers start off in
478	<	* running state.  This field is also used for memory-fencing
479	<	* configuration parameters.
480	<	*/
481	<	private volatile int workerCounts;
877	>	// Specialized execution methods
878
879	<	private static final int TOTAL_COUNT_SHIFT  = 16;
880	<	private static final int RUNNING_COUNT_MASK = (1 << TOTAL_COUNT_SHIFT) - 1;
881	<	private static final int ONE_RUNNING        = 1;
882	<	private static final int ONE_TOTAL          = 1 << TOTAL_COUNT_SHIFT;
879	>	/**
880	>	* Pops and runs tasks until empty.
881	>	*/
882	>	private void popAndExecAll() {
883	>	// A bit faster than repeated pop calls
884	>	ForkJoinTask<?>[] a; int m, s; long j; ForkJoinTask<?> t;
885	>	while ((a = array) != null && (m = a.length - 1) >= 0 &&
886	>	(s = top - 1) - base >= 0 &&
887	>	(t = ((ForkJoinTask<?>)
888	>	U.getObject(a, j = ((m & s) << ASHIFT) + ABASE)))
889	>	!= null) {
890	>	if (U.compareAndSwapObject(a, j, t, null)) {
891	>	top = s;
892	>	t.doExec();
893	>	}
894	>	}
895	>	}
896
897	<	/*
898	<	* Fields parallelism. maxPoolSize, and maintainsParallelism are
899	<	* non-volatile, but external reads/writes use workerCount fences
900	<	* to ensure visability.
901	<	*/
897	>	/**
898	>	* Polls and runs tasks until empty.
899	>	*/
900	>	private void pollAndExecAll() {
901	>	for (ForkJoinTask<?> t; (t = poll()) != null;)
902	>	t.doExec();
903	>	}
904	>
905	>	/**
906	>	* If present, removes from queue and executes the given task,
907	>	* or any other cancelled task. Returns (true) on any CAS
908	>	* or consistency check failure so caller can retry.
909	>	*
910	>	* @return false if no progress can be made, else true;
911	>	*/
912	>	final boolean tryRemoveAndExec(ForkJoinTask<?> task) {
913	>	boolean stat = true, removed = false, empty = true;
914	>	ForkJoinTask<?>[] a; int m, s, b, n;
915	>	if ((a = array) != null && (m = a.length - 1) >= 0 &&
916	>	(n = (s = top) - (b = base)) > 0) {
917	>	for (ForkJoinTask<?> t;;) {           // traverse from s to b
918	>	int j = ((--s & m) << ASHIFT) + ABASE;
919	>	t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
920	>	if (t == null)                    // inconsistent length
921	>	break;
922	>	else if (t == task) {
923	>	if (s + 1 == top) {           // pop
924	>	if (!U.compareAndSwapObject(a, j, task, null))
925	>	break;
926	>	top = s;
927	>	removed = true;
928	>	}
929	>	else if (base == b)           // replace with proxy
930	>	removed = U.compareAndSwapObject(a, j, task,
931	>	new EmptyTask());
932	>	break;
933	>	}
934	>	else if (t.status >= 0)
935	>	empty = false;
936	>	else if (s + 1 == top) {          // pop and throw away
937	>	if (U.compareAndSwapObject(a, j, t, null))
938	>	top = s;
939	>	break;
940	>	}
941	>	if (--n == 0) {
942	>	if (!empty && base == b)
943	>	stat = false;
944	>	break;
945	>	}
946	>	}
947	>	}
948	>	if (removed)
949	>	task.doExec();
950	>	return stat;
951	>	}
952	>
953	>	/**
954	>	* Polls for and executes the given task or any other task in
955	>	* its CountedCompleter computation
956	>	*/
957	>	final boolean pollAndExecCC(ForkJoinTask<?> root) {
958	>	ForkJoinTask<?>[] a; int b; Object o;
959	>	outer: while ((b = base) - top < 0 && (a = array) != null) {
960	>	long j = (((a.length - 1) & b) << ASHIFT) + ABASE;
961	>	if ((o = U.getObject(a, j)) == null ||
962	>	!(o instanceof CountedCompleter))
963	>	break;
964	>	for (CountedCompleter<?> t = (CountedCompleter<?>)o, r = t;;) {
965	>	if (r == root) {
966	>	if (base == b &&
967	>	U.compareAndSwapObject(a, j, t, null)) {
968	>	base = b + 1;
969	>	t.doExec();
970	>	return true;
971	>	}
972	>	else
973	>	break; // restart
974	>	}
975	>	if ((r = r.completer) == null)
976	>	break outer; // not part of root computation
977	>	}
978	>	}
979	>	return false;
980	>	}
981	>
982	>	/**
983	>	* Executes a top-level task and any local tasks remaining
984	>	* after execution.
985	>	*/
986	>	final void runTask(ForkJoinTask<?> t) {
987	>	if (t != null) {
988	>	(currentSteal = t).doExec();
989	>	currentSteal = null;
990	>	++nsteals;
991	>	if (base - top < 0) {       // process remaining local tasks
992	>	if (mode == 0)
993	>	popAndExecAll();
994	>	else
995	>	pollAndExecAll();
996	>	}
997	>	}
998	>	}
999	>
1000	>	/**
1001	>	* Executes a non-top-level (stolen) task.
1002	>	*/
1003	>	final void runSubtask(ForkJoinTask<?> t) {
1004	>	if (t != null) {
1005	>	ForkJoinTask<?> ps = currentSteal;
1006	>	(currentSteal = t).doExec();
1007	>	currentSteal = ps;
1008	>	}
1009	>	}
1010	>
1011	>	/**
1012	>	* Returns true if owned and not known to be blocked.
1013	>	*/
1014	>	final boolean isApparentlyUnblocked() {
1015	>	Thread wt; Thread.State s;
1016	>	return (eventCount >= 0 &&
1017	>	(wt = owner) != null &&
1018	>	(s = wt.getState()) != Thread.State.BLOCKED &&
1019	>	s != Thread.State.WAITING &&
1020	>	s != Thread.State.TIMED_WAITING);
1021	>	}
1022	>
1023	>	// Unsafe mechanics
1024	>	private static final sun.misc.Unsafe U;
1025	>	private static final long QLOCK;
1026	>	private static final int ABASE;
1027	>	private static final int ASHIFT;
1028	>	static {
1029	>	int s;
1030	>	try {
1031	>	U = getUnsafe();
1032	>	Class<?> k = WorkQueue.class;
1033	>	Class<?> ak = ForkJoinTask[].class;
1034	>	QLOCK = U.objectFieldOffset
1035	>	(k.getDeclaredField("qlock"));
1036	>	ABASE = U.arrayBaseOffset(ak);
1037	>	s = U.arrayIndexScale(ak);
1038	>	} catch (Exception e) {
1039	>	throw new Error(e);
1040	>	}
1041	>	if ((s & (s-1)) != 0)
1042	>	throw new Error("data type scale not a power of two");
1043	>	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
1044	>	}
1045	>	}
1046	>
1047	>	// static fields (initialized in static initializer below)
1048
1049		/**
1050	<	* The target parallelism level.
1050	>	* Creates a new ForkJoinWorkerThread. This factory is used unless
1051	>	* overridden in ForkJoinPool constructors.
1052		*/
1053	<	private int parallelism;
1053	>	public static final ForkJoinWorkerThreadFactory
1054	>	defaultForkJoinWorkerThreadFactory;
1055
1056		/**
1057	<	* The maximum allowed pool size.
1057	>	* Per-thread submission bookkeeping. Shared across all pools
1058	>	* to reduce ThreadLocal pollution and because random motion
1059	>	* to avoid contention in one pool is likely to hold for others.
1060	>	* Lazily initialized on first submission (but null-checked
1061	>	* in other contexts to avoid unnecessary initialization).
1062		*/
1063	<	private int maxPoolSize;
1063	>	static final ThreadLocal<Submitter> submitters;
1064
1065		/**
1066	<	* True if use local fifo, not default lifo, for local polling
1067	<	* Replicated by ForkJoinWorkerThreads
1066	>	* Permission required for callers of methods that may start or
1067	>	* kill threads.
1068		*/
1069	<	private volatile boolean locallyFifo;
1069	>	private static final RuntimePermission modifyThreadPermission;
1070
1071		/**
1072	<	* Controls whether to add spares to maintain parallelism
1072	>	* Common (static) pool. Non-null for public use unless a static
1073	>	* construction exception, but internal usages null-check on use
1074	>	* to paranoically avoid potential initialization circularities
1075	>	* as well as to simplify generated code.
1076		*/
1077	<	private boolean maintainsParallelism;
1077	>	static final ForkJoinPool common;
1078
1079		/**
1080	<	* The uncaught exception handler used when any worker
517	<	* abruptly terminates
1080	>	* Common pool parallelism. Must equal common.parallelism.
1081		*/
1082	<	private volatile Thread.UncaughtExceptionHandler ueh;
1082	>	static final int commonParallelism;
1083
1084		/**
1085	<	* Pool number, just for assigning useful names to worker threads
1085	>	* Sequence number for creating workerNamePrefix.
1086		*/
1087	<	private final int poolNumber;
525	<
526	<	// utilities for updating fields
1087	>	private static int poolNumberSequence;
1088
1089		/**
1090	<	* Adds delta to running count.  Used mainly by ForkJoinTask.
1090	>	* Returns the next sequence number. We don't expect this to
1091	>	* ever contend, so use simple builtin sync.
1092		*/
1093	<	final void updateRunningCount(int delta) {
1094	<	int wc;
533	<	do {} while (!UNSAFE.compareAndSwapInt(this, workerCountsOffset,
534	<	wc = workerCounts,
535	<	wc + delta));
1093	>	private static final synchronized int nextPoolId() {
1094	>	return ++poolNumberSequence;
1095		}
1096
1097	+	// static constants
1098	+
1099		/**
1100	<	* Decrements running count unless already zero
1100	>	* Initial timeout value (in nanoseconds) for the thread
1101	>	* triggering quiescence to park waiting for new work. On timeout,
1102	>	* the thread will instead try to shrink the number of
1103	>	* workers. The value should be large enough to avoid overly
1104	>	* aggressive shrinkage during most transient stalls (long GCs
1105	>	* etc).
1106		*/
1107	<	final boolean tryDecrementRunningCount() {
542	<	int wc = workerCounts;
543	<	if ((wc & RUNNING_COUNT_MASK) == 0)
544	<	return false;
545	<	return UNSAFE.compareAndSwapInt(this, workerCountsOffset,
546	<	wc, wc - ONE_RUNNING);
547	<	}
1107	>	private static final long IDLE_TIMEOUT      = 2000L * 1000L * 1000L; // 2sec
1108
1109		/**
1110	<	* Write fence for user modifications of pool parameters
551	<	* (parallelism. etc).  Note that it doesn't matter if CAS fails.
1110	>	* Timeout value when there are more threads than parallelism level
1111		*/
1112	<	private void workerCountWriteFence() {
554	<	int wc;
555	<	UNSAFE.compareAndSwapInt(this, workerCountsOffset,
556	<	wc = workerCounts, wc);
557	<	}
1112	>	private static final long FAST_IDLE_TIMEOUT =  200L * 1000L * 1000L;
1113
1114		/**
1115	<	* Read fence for external reads of pool parameters
561	<	* (parallelism. maxPoolSize, etc).
1115	>	* Tolerance for idle timeouts, to cope with timer undershoots
1116		*/
1117	<	private void workerCountReadFence() {
564	<	int ignore = workerCounts;
565	<	}
1117	>	private static final long TIMEOUT_SLOP = 2000000L;
1118
1119		/**
1120	<	* Tries incrementing active count; fails on contention.
1121	<	* Called by workers before executing tasks.
1122	<	*
1123	<	* @return true on success
1120	>	* The maximum stolen->joining link depth allowed in method
1121	>	* tryHelpStealer.  Must be a power of two.  Depths for legitimate
1122	>	* chains are unbounded, but we use a fixed constant to avoid
1123	>	* (otherwise unchecked) cycles and to bound staleness of
1124	>	* traversal parameters at the expense of sometimes blocking when
1125	>	* we could be helping.
1126		*/
1127	<	final boolean tryIncrementActiveCount() {
574	<	int c;
575	<	return UNSAFE.compareAndSwapInt(this, runStateOffset,
576	<	c = runState, c + ONE_ACTIVE);
577	<	}
1127	>	private static final int MAX_HELP = 64;
1128
1129		/**
1130	<	* Tries decrementing active count; fails on contention.
1131	<	* Called when workers cannot find tasks to run.
1130	>	* Increment for seed generators. See class ThreadLocal for
1131	>	* explanation.
1132		*/
1133	<	final boolean tryDecrementActiveCount() {
584	<	int c;
585	<	return UNSAFE.compareAndSwapInt(this, runStateOffset,
586	<	c = runState, c - ONE_ACTIVE);
587	<	}
1133	>	private static final int SEED_INCREMENT = 0x61c88647;
1134
1135		/**
1136	<	* Advances to at least the given level. Returns true if not
1137	<	* already in at least the given level.
1136	>	* Bits and masks for control variables
1137	>	*
1138	>	* Field ctl is a long packed with:
1139	>	* AC: Number of active running workers minus target parallelism (16 bits)
1140	>	* TC: Number of total workers minus target parallelism (16 bits)
1141	>	* ST: true if pool is terminating (1 bit)
1142	>	* EC: the wait count of top waiting thread (15 bits)
1143	>	* ID: poolIndex of top of Treiber stack of waiters (16 bits)
1144	>	*
1145	>	* When convenient, we can extract the upper 32 bits of counts and
1146	>	* the lower 32 bits of queue state, u = (int)(ctl >>> 32) and e =
1147	>	* (int)ctl.  The ec field is never accessed alone, but always
1148	>	* together with id and st. The offsets of counts by the target
1149	>	* parallelism and the positionings of fields makes it possible to
1150	>	* perform the most common checks via sign tests of fields: When
1151	>	* ac is negative, there are not enough active workers, when tc is
1152	>	* negative, there are not enough total workers, and when e is
1153	>	* negative, the pool is terminating.  To deal with these possibly
1154	>	* negative fields, we use casts in and out of "short" and/or
1155	>	* signed shifts to maintain signedness.
1156	>	*
1157	>	* When a thread is queued (inactivated), its eventCount field is
1158	>	* set negative, which is the only way to tell if a worker is
1159	>	* prevented from executing tasks, even though it must continue to
1160	>	* scan for them to avoid queuing races. Note however that
1161	>	* eventCount updates lag releases so usage requires care.
1162	>	*
1163	>	* Field plock is an int packed with:
1164	>	* SHUTDOWN: true if shutdown is enabled (1 bit)
1165	>	* SEQ:  a sequence lock, with PL_LOCK bit set if locked (30 bits)
1166	>	* SIGNAL: set when threads may be waiting on the lock (1 bit)
1167	>	*
1168	>	* The sequence number enables simple consistency checks:
1169	>	* Staleness of read-only operations on the workQueues array can
1170	>	* be checked by comparing plock before vs after the reads.
1171		*/
593	–	private boolean advanceRunLevel(int level) {
594	–	for (;;) {
595	–	int s = runState;
596	–	if ((s & level) != 0)
597	–	return false;
598	–	if (UNSAFE.compareAndSwapInt(this, runStateOffset, s, s | level))
599	–	return true;
600	–	}
601	–	}
1172
1173	<	// workers array maintenance
1173	>	// bit positions/shifts for fields
1174	>	private static final int  AC_SHIFT   = 48;
1175	>	private static final int  TC_SHIFT   = 32;
1176	>	private static final int  ST_SHIFT   = 31;
1177	>	private static final int  EC_SHIFT   = 16;
1178
1179	<	/**
1180	<	* Records and returns a workers array index for new worker.
1179	>	// bounds
1180	>	private static final int  SMASK      = 0xffff;  // short bits
1181	>	private static final int  MAX_CAP    = 0x7fff;  // max #workers - 1
1182	>	private static final int  EVENMASK   = 0xfffe;  // even short bits
1183	>	private static final int  SQMASK     = 0x007e;  // max 64 (even) slots
1184	>	private static final int  SHORT_SIGN = 1 << 15;
1185	>	private static final int  INT_SIGN   = 1 << 31;
1186	>
1187	>	// masks
1188	>	private static final long STOP_BIT   = 0x0001L << ST_SHIFT;
1189	>	private static final long AC_MASK    = ((long)SMASK) << AC_SHIFT;
1190	>	private static final long TC_MASK    = ((long)SMASK) << TC_SHIFT;
1191	>
1192	>	// units for incrementing and decrementing
1193	>	private static final long TC_UNIT    = 1L << TC_SHIFT;
1194	>	private static final long AC_UNIT    = 1L << AC_SHIFT;
1195	>
1196	>	// masks and units for dealing with u = (int)(ctl >>> 32)
1197	>	private static final int  UAC_SHIFT  = AC_SHIFT - 32;
1198	>	private static final int  UTC_SHIFT  = TC_SHIFT - 32;
1199	>	private static final int  UAC_MASK   = SMASK << UAC_SHIFT;
1200	>	private static final int  UTC_MASK   = SMASK << UTC_SHIFT;
1201	>	private static final int  UAC_UNIT   = 1 << UAC_SHIFT;
1202	>	private static final int  UTC_UNIT   = 1 << UTC_SHIFT;
1203	>
1204	>	// masks and units for dealing with e = (int)ctl
1205	>	private static final int E_MASK      = 0x7fffffff; // no STOP_BIT
1206	>	private static final int E_SEQ       = 1 << EC_SHIFT;
1207	>
1208	>	// plock bits
1209	>	private static final int SHUTDOWN    = 1 << 31;
1210	>	private static final int PL_LOCK     = 2;
1211	>	private static final int PL_SIGNAL   = 1;
1212	>	private static final int PL_SPINS    = 1 << 8;
1213	>
1214	>	// access mode for WorkQueue
1215	>	static final int LIFO_QUEUE          =  0;
1216	>	static final int FIFO_QUEUE          =  1;
1217	>	static final int SHARED_QUEUE        = -1;
1218	>
1219	>	// bounds for #steps in scan loop -- must be power 2 minus 1
1220	>	private static final int MIN_SCAN    = 0x1ff;   // cover estimation slop
1221	>	private static final int MAX_SCAN    = 0x1ffff; // 4 * max workers
1222	>
1223	>	// Instance fields
1224	>
1225	>	/*
1226	>	* Field layout of this class tends to matter more than one would
1227	>	* like. Runtime layout order is only loosely related to
1228	>	* declaration order and may differ across JVMs, but the following
1229	>	* empirically works OK on current JVMs.
1230	>	*/
1231	>
1232	>	// Heuristic padding to ameliorate unfortunate memory placements
1233	>	volatile long pad00, pad01, pad02, pad03, pad04, pad05, pad06;
1234	>
1235	>	volatile long stealCount;                  // collects worker counts
1236	>	volatile long ctl;                         // main pool control
1237	>	volatile int plock;                        // shutdown status and seqLock
1238	>	volatile int indexSeed;                    // worker/submitter index seed
1239	>	final int config;                          // mode and parallelism level
1240	>	WorkQueue[] workQueues;                    // main registry
1241	>	final ForkJoinWorkerThreadFactory factory;
1242	>	final Thread.UncaughtExceptionHandler ueh; // per-worker UEH
1243	>	final String workerNamePrefix;             // to create worker name string
1244	>
1245	>	volatile Object pad10, pad11, pad12, pad13, pad14, pad15, pad16, pad17;
1246	>	volatile Object pad18, pad19, pad1a, pad1b;
1247	>
1248	>	/*
1249	>	* Acquires the plock lock to protect worker array and related
1250	>	* updates. This method is called only if an initial CAS on plock
1251	>	* fails. This acts as a spinlock for normal cases, but falls back
1252	>	* to builtin monitor to block when (rarely) needed. This would be
1253	>	* a terrible idea for a highly contended lock, but works fine as
1254	>	* a more conservative alternative to a pure spinlock.
1255		*/
1256	<	private int recordWorker(ForkJoinWorkerThread w) {
1257	<	// Try using slot totalCount-1. If not available, scan and/or resize
1258	<	int k = (workerCounts >>> TOTAL_COUNT_SHIFT) - 1;
1259	<	final ReentrantLock lock = this.workerLock;
1260	<	lock.lock();
1261	<	try {
1262	<	ForkJoinWorkerThread[] ws = workers;
1263	<	int nws = ws.length;
1264	<	if (k < 0 || k >= nws || ws[k] != null) {
1265	<	for (k = 0; k < nws && ws[k] != null; ++k)
1266	<	;
1267	<	if (k == nws)
1268	<	ws = Arrays.copyOf(ws, nws << 1);
1256	>	private int acquirePlock() {
1257	>	int spins = PL_SPINS, r = 0, ps, nps;
1258	>	for (;;) {
1259	>	if (((ps = plock) & PL_LOCK) == 0 &&
1260	>	U.compareAndSwapInt(this, PLOCK, ps, nps = ps + PL_LOCK))
1261	>	return nps;
1262	>	else if (r == 0) { // randomize spins if possible
1263	>	Thread t = Thread.currentThread(); WorkQueue w; Submitter z;
1264	>	if ((t instanceof ForkJoinWorkerThread) &&
1265	>	(w = ((ForkJoinWorkerThread)t).workQueue) != null)
1266	>	r = w.seed;
1267	>	else if ((z = submitters.get()) != null)
1268	>	r = z.seed;
1269	>	else
1270	>	r = 1;
1271	>	}
1272	>	else if (spins >= 0) {
1273	>	r ^= r << 1; r ^= r >>> 3; r ^= r << 10; // xorshift
1274	>	if (r >= 0)
1275	>	--spins;
1276	>	}
1277	>	else if (U.compareAndSwapInt(this, PLOCK, ps, ps | PL_SIGNAL)) {
1278	>	synchronized (this) {
1279	>	if ((plock & PL_SIGNAL) != 0) {
1280	>	try {
1281	>	wait();
1282	>	} catch (InterruptedException ie) {
1283	>	try {
1284	>	Thread.currentThread().interrupt();
1285	>	} catch (SecurityException ignore) {
1286	>	}
1287	>	}
1288	>	}
1289	>	else
1290	>	notifyAll();
1291	>	}
1292		}
622	–	ws[k] = w;
623	–	workers = ws; // volatile array write ensures slot visibility
624	–	} finally {
625	–	lock.unlock();
1293		}
627	–	return k;
1294		}
1295
1296		/**
1297	<	* Nulls out record of worker in workers array
1297	>	* Unlocks and signals any thread waiting for plock. Called only
1298	>	* when CAS of seq value for unlock fails.
1299		*/
1300	<	private void forgetWorker(ForkJoinWorkerThread w) {
1301	<	int idx = w.poolIndex;
1302	<	// Locking helps method recordWorker avoid unecessary expansion
1303	<	final ReentrantLock lock = this.workerLock;
1304	<	lock.lock();
1305	<	try {
1306	<	ForkJoinWorkerThread[] ws = workers;
1307	<	if (idx >= 0 && idx < ws.length && ws[idx] == w) // verify
1308	<	ws[idx] = null;
1309	<	} finally {
1310	<	lock.unlock();
1300	>	private void releasePlock(int ps) {
1301	>	plock = ps;
1302	>	synchronized (this) { notifyAll(); }
1303	>	}
1304	>
1305	>	/**
1306	>	* Tries to create and start one worker if fewer than target
1307	>	* parallelism level exist. Adjusts counts etc on failure.
1308	>	*/
1309	>	private void tryAddWorker() {
1310	>	long c; int u;
1311	>	while ((u = (int)((c = ctl) >>> 32)) < 0 &&
1312	>	(u & SHORT_SIGN) != 0 && (int)c == 0) {
1313	>	long nc = (long)(((u + UTC_UNIT) & UTC_MASK) |
1314	>	((u + UAC_UNIT) & UAC_MASK)) << 32;
1315	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1316	>	ForkJoinWorkerThreadFactory fac;
1317	>	Throwable ex = null;
1318	>	ForkJoinWorkerThread wt = null;
1319	>	try {
1320	>	if ((fac = factory) != null &&
1321	>	(wt = fac.newThread(this)) != null) {
1322	>	wt.start();
1323	>	break;
1324	>	}
1325	>	} catch (Throwable e) {
1326	>	ex = e;
1327	>	}
1328	>	deregisterWorker(wt, ex);
1329	>	break;
1330	>	}
1331		}
1332		}
1333
1334	<	// adding and removing workers
1334	>	//  Registering and deregistering workers
1335
1336		/**
1337	<	* Tries to create and add new worker. Assumes that worker counts
1338	<	* are already updated to accommodate the worker, so adjusts on
1339	<	* failure.
1340	<	*
1341	<	* @return new worker or null if creation failed
1342	<	*/
1343	<	private ForkJoinWorkerThread addWorker() {
1344	<	ForkJoinWorkerThread w = null;
1337	>	* Callback from ForkJoinWorkerThread to establish and record its
1338	>	* WorkQueue. To avoid scanning bias due to packing entries in
1339	>	* front of the workQueues array, we treat the array as a simple
1340	>	* power-of-two hash table using per-thread seed as hash,
1341	>	* expanding as needed.
1342	>	*
1343	>	* @param wt the worker thread
1344	>	* @return the worker's queue
1345	>	*/
1346	>	final WorkQueue registerWorker(ForkJoinWorkerThread wt) {
1347	>	Thread.UncaughtExceptionHandler handler; WorkQueue[] ws; int s, ps;
1348	>	wt.setDaemon(true);
1349	>	if ((handler = ueh) != null)
1350	>	wt.setUncaughtExceptionHandler(handler);
1351	>	do {} while (!U.compareAndSwapInt(this, INDEXSEED, s = indexSeed,
1352	>	s += SEED_INCREMENT) ||
1353	>	s == 0); // skip 0
1354	>	WorkQueue w = new WorkQueue(this, wt, config >>> 16, s);
1355	>	if (((ps = plock) & PL_LOCK) != 0 ||
1356	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1357	>	ps = acquirePlock();
1358	>	int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1359		try {
1360	<	w = factory.newThread(this);
1361	<	} finally { // Adjust on either null or exceptional factory return
1362	<	if (w == null) {
1363	<	onWorkerCreationFailure();
1364	<	return null;
1360	>	if ((ws = workQueues) != null) {    // skip if shutting down
1361	>	int n = ws.length, m = n - 1;
1362	>	int r = (s << 1) | 1;           // use odd-numbered indices
1363	>	if (ws[r &= m] != null) {       // collision
1364	>	int probes = 0;             // step by approx half size
1365	>	int step = (n <= 4) ? 2 : ((n >>> 1) & EVENMASK) + 2;
1366	>	while (ws[r = (r + step) & m] != null) {
1367	>	if (++probes >= n) {
1368	>	workQueues = ws = Arrays.copyOf(ws, n <<= 1);
1369	>	m = n - 1;
1370	>	probes = 0;
1371	>	}
1372	>	}
1373	>	}
1374	>	w.eventCount = w.poolIndex = r; // volatile write orders
1375	>	ws[r] = w;
1376		}
1377	+	} finally {
1378	+	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1379	+	releasePlock(nps);
1380		}
1381	<	w.start(recordWorker(w), locallyFifo, ueh);
1381	>	wt.setName(workerNamePrefix.concat(Integer.toString(w.poolIndex)));
1382		return w;
1383		}
1384
1385		/**
1386	<	* Adjusts counts upon failure to create worker
1387	<	*/
1388	<	private void onWorkerCreationFailure() {
1389	<	for (;;) {
1390	<	int wc = workerCounts;
1391	<	if ((wc >>> TOTAL_COUNT_SHIFT) > 0 &&
1392	<	UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1393	<	wc, wc - (ONE_RUNNING|ONE_TOTAL)))
1394	<	break;
1386	>	* Final callback from terminating worker, as well as upon failure
1387	>	* to construct or start a worker.  Removes record of worker from
1388	>	* array, and adjusts counts. If pool is shutting down, tries to
1389	>	* complete termination.
1390	>	*
1391	>	* @param wt the worker thread or null if construction failed
1392	>	* @param ex the exception causing failure, or null if none
1393	>	*/
1394	>	final void deregisterWorker(ForkJoinWorkerThread wt, Throwable ex) {
1395	>	WorkQueue w = null;
1396	>	if (wt != null && (w = wt.workQueue) != null) {
1397	>	int ps;
1398	>	w.qlock = -1;                // ensure set
1399	>	long ns = w.nsteals, sc;     // collect steal count
1400	>	do {} while (!U.compareAndSwapLong(this, STEALCOUNT,
1401	>	sc = stealCount, sc + ns));
1402	>	if (((ps = plock) & PL_LOCK) != 0 ||
1403	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1404	>	ps = acquirePlock();
1405	>	int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1406	>	try {
1407	>	int idx = w.poolIndex;
1408	>	WorkQueue[] ws = workQueues;
1409	>	if (ws != null && idx >= 0 && idx < ws.length && ws[idx] == w)
1410	>	ws[idx] = null;
1411	>	} finally {
1412	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1413	>	releasePlock(nps);
1414	>	}
1415		}
681	–	tryTerminate(false); // in case of failure during shutdown
682	–	}
1416
1417	<	/**
1418	<	* Create enough total workers to establish target parallelism,
1419	<	* giving up if terminating or addWorker fails
1420	<	*/
1421	<	private void ensureEnoughTotalWorkers() {
1422	<	int wc;
1423	<	while (((wc = workerCounts) >>> TOTAL_COUNT_SHIFT) < parallelism &&
1424	<	runState < TERMINATING) {
1425	<	if ((UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1426	<	wc, wc + (ONE_RUNNING|ONE_TOTAL)) &&
1427	<	addWorker() == null))
1428	<	break;
1417	>	long c;                          // adjust ctl counts
1418	>	do {} while (!U.compareAndSwapLong
1419	>	(this, CTL, c = ctl, (((c - AC_UNIT) & AC_MASK) |
1420	>	((c - TC_UNIT) & TC_MASK) |
1421	>	(c & ~(AC_MASK|TC_MASK)))));
1422	>
1423	>	if (!tryTerminate(false, false) && w != null && w.array != null) {
1424	>	w.cancelAll();               // cancel remaining tasks
1425	>	WorkQueue[] ws; WorkQueue v; Thread p; int u, i, e;
1426	>	while ((u = (int)((c = ctl) >>> 32)) < 0 && (e = (int)c) >= 0) {
1427	>	if (e > 0) {             // activate or create replacement
1428	>	if ((ws = workQueues) == null ||
1429	>	(i = e & SMASK) >= ws.length ||
1430	>	(v = ws[i]) == null)
1431	>	break;
1432	>	long nc = (((long)(v.nextWait & E_MASK)) |
1433	>	((long)(u + UAC_UNIT) << 32));
1434	>	if (v.eventCount != (e | INT_SIGN))
1435	>	break;
1436	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1437	>	v.eventCount = (e + E_SEQ) & E_MASK;
1438	>	if ((p = v.parker) != null)
1439	>	U.unpark(p);
1440	>	break;
1441	>	}
1442	>	}
1443	>	else {
1444	>	if ((short)u < 0)
1445	>	tryAddWorker();
1446	>	break;
1447	>	}
1448	>	}
1449	>	}
1450	>	if (ex == null)                     // help clean refs on way out
1451	>	ForkJoinTask.helpExpungeStaleExceptions();
1452	>	else                                // rethrow
1453	>	ForkJoinTask.rethrow(ex);
1454	>	}
1455	>
1456	>	// Submissions
1457	>
1458	>	/**
1459	>	* Unless shutting down, adds the given task to a submission queue
1460	>	* at submitter's current queue index (modulo submission
1461	>	* range). Only the most common path is directly handled in this
1462	>	* method. All others are relayed to fullExternalPush.
1463	>	*
1464	>	* @param task the task. Caller must ensure non-null.
1465	>	*/
1466	>	final void externalPush(ForkJoinTask<?> task) {
1467	>	WorkQueue[] ws; WorkQueue q; Submitter z; int m; ForkJoinTask<?>[] a;
1468	>	if ((z = submitters.get()) != null && plock > 0 &&
1469	>	(ws = workQueues) != null && (m = (ws.length - 1)) >= 0 &&
1470	>	(q = ws[m & z.seed & SQMASK]) != null &&
1471	>	U.compareAndSwapInt(q, QLOCK, 0, 1)) { // lock
1472	>	int b = q.base, s = q.top, n, an;
1473	>	if ((a = q.array) != null && (an = a.length) > (n = s + 1 - b)) {
1474	>	int j = (((an - 1) & s) << ASHIFT) + ABASE;
1475	>	U.putOrderedObject(a, j, task);
1476	>	q.top = s + 1;                     // push on to deque
1477	>	q.qlock = 0;
1478	>	if (n <= 2)
1479	>	signalWork(q);
1480	>	return;
1481	>	}
1482	>	q.qlock = 0;
1483		}
1484	+	fullExternalPush(task);
1485		}
1486
1487		/**
1488	<	* Final callback from terminating worker.  Removes record of
1489	<	* worker from array, and adjusts counts. If pool is shutting
1490	<	* down, tries to complete terminatation, else possibly replaces
1491	<	* the worker.
1492	<	*
1493	<	* @param w the worker
1494	<	*/
1495	<	final void workerTerminated(ForkJoinWorkerThread w) {
1496	<	if (w.active) { // force inactive
1497	<	w.active = false;
1498	<	do {} while (!tryDecrementActiveCount());
1499	<	}
1500	<	forgetWorker(w);
1501	<
1502	<	// Decrement total count, and if was running, running count
1503	<	// Spin (waiting for other updates) if either would be negative
1504	<	int nr = w.isTrimmed() ? 0 : ONE_RUNNING;
1505	<	int unit = ONE_TOTAL + nr;
1506	<	for (;;) {
1507	<	int wc = workerCounts;
1508	<	int rc = wc & RUNNING_COUNT_MASK;
1509	<	if (rc - nr < 0 || (wc >>> TOTAL_COUNT_SHIFT) == 0)
1510	<	Thread.yield(); // back off if waiting for other updates
1511	<	else if (UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1512	<	wc, wc - unit))
1513	<	break;
1488	>	* Full version of externalPush. This method is called, among
1489	>	* other times, upon the first submission of the first task to the
1490	>	* pool, so must perform secondary initialization.  It also
1491	>	* detects first submission by an external thread by looking up
1492	>	* its ThreadLocal, and creates a new shared queue if the one at
1493	>	* index if empty or contended. The plock lock body must be
1494	>	* exception-free (so no try/finally) so we optimistically
1495	>	* allocate new queues outside the lock and throw them away if
1496	>	* (very rarely) not needed.
1497	>	*
1498	>	* Secondary initialization occurs when plock is zero, to create
1499	>	* workQueue array and set plock to a valid value.  This lock body
1500	>	* must also be exception-free. Because the plock seq value can
1501	>	* eventually wrap around zero, this method harmlessly fails to
1502	>	* reinitialize if workQueues exists, while still advancing plock.
1503	>	*/
1504	>	private void fullExternalPush(ForkJoinTask<?> task) {
1505	>	int r = 0; // random index seed
1506	>	for (Submitter z = submitters.get();;) {
1507	>	WorkQueue[] ws; WorkQueue q; int ps, m, k;
1508	>	if (z == null) {
1509	>	if (U.compareAndSwapInt(this, INDEXSEED, r = indexSeed,
1510	>	r += SEED_INCREMENT) && r != 0)
1511	>	submitters.set(z = new Submitter(r));
1512	>	}
1513	>	else if (r == 0) {                  // move to a different index
1514	>	r = z.seed;
1515	>	r ^= r << 13;                   // same xorshift as WorkQueues
1516	>	r ^= r >>> 17;
1517	>	z.seed = r ^ (r << 5);
1518	>	}
1519	>	else if ((ps = plock) < 0)
1520	>	throw new RejectedExecutionException();
1521	>	else if (ps == 0 || (ws = workQueues) == null ||
1522	>	(m = ws.length - 1) < 0) { // initialize workQueues
1523	>	int p = config & SMASK;         // find power of two table size
1524	>	int n = (p > 1) ? p - 1 : 1;    // ensure at least 2 slots
1525	>	n |= n >>> 1; n |= n >>> 2;  n |= n >>> 4;
1526	>	n |= n >>> 8; n |= n >>> 16; n = (n + 1) << 1;
1527	>	WorkQueue[] nws = ((ws = workQueues) == null || ws.length == 0 ?
1528	>	new WorkQueue[n] : null);
1529	>	if (((ps = plock) & PL_LOCK) != 0 ||
1530	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1531	>	ps = acquirePlock();
1532	>	if (((ws = workQueues) == null || ws.length == 0) && nws != null)
1533	>	workQueues = nws;
1534	>	int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1535	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1536	>	releasePlock(nps);
1537	>	}
1538	>	else if ((q = ws[k = r & m & SQMASK]) != null) {
1539	>	if (q.qlock == 0 && U.compareAndSwapInt(q, QLOCK, 0, 1)) {
1540	>	ForkJoinTask<?>[] a = q.array;
1541	>	int s = q.top;
1542	>	boolean submitted = false;
1543	>	try {                      // locked version of push
1544	>	if ((a != null && a.length > s + 1 - q.base) ||
1545	>	(a = q.growArray()) != null) {   // must presize
1546	>	int j = (((a.length - 1) & s) << ASHIFT) + ABASE;
1547	>	U.putOrderedObject(a, j, task);
1548	>	q.top = s + 1;
1549	>	submitted = true;
1550	>	}
1551	>	} finally {
1552	>	q.qlock = 0;  // unlock
1553	>	}
1554	>	if (submitted) {
1555	>	signalWork(q);
1556	>	return;
1557	>	}
1558	>	}
1559	>	r = 0; // move on failure
1560	>	}
1561	>	else if (((ps = plock) & PL_LOCK) == 0) { // create new queue
1562	>	q = new WorkQueue(this, null, SHARED_QUEUE, r);
1563	>	if (((ps = plock) & PL_LOCK) != 0 ||
1564	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1565	>	ps = acquirePlock();
1566	>	if ((ws = workQueues) != null && k < ws.length && ws[k] == null)
1567	>	ws[k] = q;
1568	>	int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1569	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1570	>	releasePlock(nps);
1571	>	}
1572	>	else
1573	>	r = 0; // try elsewhere while lock held
1574		}
727	–
728	–	accumulateStealCount(w); // collect final count
729	–	if (!tryTerminate(false))
730	–	ensureEnoughTotalWorkers();
1575		}
1576
1577	<	// Waiting for and signalling events
1577	>	// Maintaining ctl counts
1578
1579		/**
1580	<	* Ensures eventCount on exit is different (mod 2^32) than on
737	<	* entry.  CAS failures are OK -- any change in count suffices.
1580	>	* Increments active count; mainly called upon return from blocking.
1581		*/
1582	<	private void advanceEventCount() {
1583	<	int c;
1584	<	UNSAFE.compareAndSwapInt(this, eventCountOffset, c = eventCount, c+1);
1582	>	final void incrementActiveCount() {
1583	>	long c;
1584	>	do {} while (!U.compareAndSwapLong(this, CTL, c = ctl, c + AC_UNIT));
1585		}
1586
1587		/**
1588	<	* Releases workers blocked on a count not equal to current count.
1588	>	* Tries to create or activate a worker if too few are active.
1589	>	*
1590	>	* @param q the (non-null) queue holding tasks to be signalled
1591		*/
1592	<	final void releaseWaiters() {
1593	<	long top;
1594	<	int id;
1595	<	while ((id = (int)((top = eventWaiters) & WAITER_INDEX_MASK)) > 0 &&
1596	<	(int)(top >>> EVENT_COUNT_SHIFT) != eventCount) {
1597	<	ForkJoinWorkerThread[] ws = workers;
1598	<	ForkJoinWorkerThread w;
1599	<	if (ws.length >= id && (w = ws[id - 1]) != null &&
1600	<	UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
1601	<	top, w.nextWaiter))
1602	<	LockSupport.unpark(w);
1592	>	final void signalWork(WorkQueue q) {
1593	>	int hint = q.poolIndex;
1594	>	long c; int e, u, i, n; WorkQueue[] ws; WorkQueue w; Thread p;
1595	>	while ((u = (int)((c = ctl) >>> 32)) < 0) {
1596	>	if ((e = (int)c) > 0) {
1597	>	if ((ws = workQueues) != null && ws.length > (i = e & SMASK) &&
1598	>	(w = ws[i]) != null && w.eventCount == (e | INT_SIGN)) {
1599	>	long nc = (((long)(w.nextWait & E_MASK)) |
1600	>	((long)(u + UAC_UNIT) << 32));
1601	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1602	>	w.hint = hint;
1603	>	w.eventCount = (e + E_SEQ) & E_MASK;
1604	>	if ((p = w.parker) != null)
1605	>	U.unpark(p);
1606	>	break;
1607	>	}
1608	>	if (q.top - q.base <= 0)
1609	>	break;
1610	>	}
1611	>	else
1612	>	break;
1613	>	}
1614	>	else {
1615	>	if ((short)u < 0)
1616	>	tryAddWorker();
1617	>	break;
1618	>	}
1619		}
1620		}
1621
1622	+	// Scanning for tasks
1623	+
1624		/**
1625	<	* Advances eventCount and releases waiters until interference by
763	<	* other releasing threads is detected.
1625	>	* Top-level runloop for workers, called by ForkJoinWorkerThread.run.
1626		*/
1627	<	final void signalWork() {
1628	<	int ec;
1629	<	UNSAFE.compareAndSwapInt(this, eventCountOffset, ec=eventCount, ec+1);
1630	<	outer:for (;;) {
1631	<	long top = eventWaiters;
1632	<	ec = eventCount;
1633	<	for (;;) {
1634	<	ForkJoinWorkerThread[] ws; ForkJoinWorkerThread w;
1635	<	int id = (int)(top & WAITER_INDEX_MASK);
1636	<	if (id <= 0 || (int)(top >>> EVENT_COUNT_SHIFT) == ec)
1637	<	return;
1638	<	if ((ws = workers).length < id || (w = ws[id - 1]) == null ||
1639	<	!UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
1640	<	top, top = w.nextWaiter))
1641	<	continue outer;      // possibly stale; reread
1642	<	LockSupport.unpark(w);
1643	<	if (top != eventWaiters) // let someone else take over
1644	<	return;
1627	>	final void runWorker(WorkQueue w) {
1628	>	w.growArray(); // allocate queue
1629	>	do { w.runTask(scan(w)); } while (w.qlock >= 0);
1630	>	}
1631	>
1632	>	/**
1633	>	* Scans for and, if found, returns one task, else possibly
1634	>	* inactivates the worker. This method operates on single reads of
1635	>	* volatile state and is designed to be re-invoked continuously,
1636	>	* in part because it returns upon detecting inconsistencies,
1637	>	* contention, or state changes that indicate possible success on
1638	>	* re-invocation.
1639	>	*
1640	>	* The scan searches for tasks across queues (starting at a random
1641	>	* index, and relying on registerWorker to irregularly scatter
1642	>	* them within array to avoid bias), checking each at least twice.
1643	>	* The scan terminates upon either finding a non-empty queue, or
1644	>	* completing the sweep. If the worker is not inactivated, it
1645	>	* takes and returns a task from this queue. Otherwise, if not
1646	>	* activated, it signals workers (that may include itself) and
1647	>	* returns so caller can retry. Also returns for true if the
1648	>	* worker array may have changed during an empty scan.  On failure
1649	>	* to find a task, we take one of the following actions, after
1650	>	* which the caller will retry calling this method unless
1651	>	* terminated.
1652	>	*
1653	>	* * If pool is terminating, terminate the worker.
1654	>	*
1655	>	* * If not already enqueued, try to inactivate and enqueue the
1656	>	* worker on wait queue. Or, if inactivating has caused the pool
1657	>	* to be quiescent, relay to idleAwaitWork to possibly shrink
1658	>	* pool.
1659	>	*
1660	>	* * If already enqueued and none of the above apply, possibly
1661	>	* park awaiting signal, else lingering to help scan and signal.
1662	>	*
1663	>	* * If a non-empty queue discovered or left as a hint,
1664	>	* help wake up other workers before return.
1665	>	*
1666	>	* @param w the worker (via its WorkQueue)
1667	>	* @return a task or null if none found
1668	>	*/
1669	>	private final ForkJoinTask<?> scan(WorkQueue w) {
1670	>	WorkQueue[] ws; int m;
1671	>	int ps = plock;                          // read plock before ws
1672	>	if (w != null && (ws = workQueues) != null && (m = ws.length - 1) >= 0) {
1673	>	int ec = w.eventCount;               // ec is negative if inactive
1674	>	int r = w.seed; r ^= r << 13; r ^= r >>> 17; w.seed = r ^= r << 5;
1675	>	w.hint = -1;                         // update seed and clear hint
1676	>	int j = ((m + m + 1) | MIN_SCAN) & MAX_SCAN;
1677	>	do {
1678	>	WorkQueue q; ForkJoinTask<?>[] a; int b;
1679	>	if ((q = ws[(r + j) & m]) != null && (b = q.base) - q.top < 0 &&
1680	>	(a = q.array) != null) {     // probably nonempty
1681	>	int i = (((a.length - 1) & b) << ASHIFT) + ABASE;
1682	>	ForkJoinTask<?> t = (ForkJoinTask<?>)
1683	>	U.getObjectVolatile(a, i);
1684	>	if (q.base == b && ec >= 0 && t != null &&
1685	>	U.compareAndSwapObject(a, i, t, null)) {
1686	>	if ((q.base = b + 1) - q.top < 0)
1687	>	signalWork(q);
1688	>	return t;                // taken
1689	>	}
1690	>	else if ((ec < 0 || j < m) && (int)(ctl >> AC_SHIFT) <= 0) {
1691	>	w.hint = (r + j) & m;    // help signal below
1692	>	break;                   // cannot take
1693	>	}
1694	>	}
1695	>	} while (--j >= 0);
1696	>
1697	>	int h, e, ns; long c, sc; WorkQueue q;
1698	>	if ((ns = w.nsteals) != 0) {
1699	>	if (U.compareAndSwapLong(this, STEALCOUNT,
1700	>	sc = stealCount, sc + ns))
1701	>	w.nsteals = 0;               // collect steals and rescan
1702	>	}
1703	>	else if (plock != ps)                // consistency check
1704	>	;                                // skip
1705	>	else if ((e = (int)(c = ctl)) < 0)
1706	>	w.qlock = -1;                    // pool is terminating
1707	>	else {
1708	>	if ((h = w.hint) < 0) {
1709	>	if (ec >= 0) {               // try to enqueue/inactivate
1710	>	long nc = (((long)ec |
1711	>	((c - AC_UNIT) & (AC_MASK|TC_MASK))));
1712	>	w.nextWait = e;          // link and mark inactive
1713	>	w.eventCount = ec | INT_SIGN;
1714	>	if (ctl != c || !U.compareAndSwapLong(this, CTL, c, nc))
1715	>	w.eventCount = ec;   // unmark on CAS failure
1716	>	else if ((int)(c >> AC_SHIFT) == 1 - (config & SMASK))
1717	>	idleAwaitWork(w, nc, c);
1718	>	}
1719	>	else if (w.eventCount < 0 && ctl == c) {
1720	>	Thread wt = Thread.currentThread();
1721	>	Thread.interrupted();    // clear status
1722	>	U.putObject(wt, PARKBLOCKER, this);
1723	>	w.parker = wt;           // emulate LockSupport.park
1724	>	if (w.eventCount < 0)    // recheck
1725	>	U.park(false, 0L);   // block
1726	>	w.parker = null;
1727	>	U.putObject(wt, PARKBLOCKER, null);
1728	>	}
1729	>	}
1730	>	if ((h >= 0 || (h = w.hint) >= 0) &&
1731	>	(ws = workQueues) != null && h < ws.length &&
1732	>	(q = ws[h]) != null) {      // signal others before retry
1733	>	WorkQueue v; Thread p; int u, i, s;
1734	>	for (int n = (config & SMASK) - 1;;) {
1735	>	int idleCount = (w.eventCount < 0) ? 0 : -1;
1736	>	if (((s = idleCount - q.base + q.top) <= n &&
1737	>	(n = s) <= 0) ||
1738	>	(u = (int)((c = ctl) >>> 32)) >= 0 ||
1739	>	(e = (int)c) <= 0 || m < (i = e & SMASK) ||
1740	>	(v = ws[i]) == null)
1741	>	break;
1742	>	long nc = (((long)(v.nextWait & E_MASK)) |
1743	>	((long)(u + UAC_UNIT) << 32));
1744	>	if (v.eventCount != (e | INT_SIGN) ||
1745	>	!U.compareAndSwapLong(this, CTL, c, nc))
1746	>	break;
1747	>	v.hint = h;
1748	>	v.eventCount = (e + E_SEQ) & E_MASK;
1749	>	if ((p = v.parker) != null)
1750	>	U.unpark(p);
1751	>	if (--n <= 0)
1752	>	break;
1753	>	}
1754	>	}
1755		}
1756		}
1757	+	return null;
1758		}
1759
1760		/**
1761	<	* If worker is inactive, blocks until terminating or event count
1762	<	* advances from last value held by worker; in any case helps
1763	<	* release others.
1764	<	*
1765	<	* @param w the calling worker thread
1766	<	*/
1767	<	private void eventSync(ForkJoinWorkerThread w) {
1768	<	if (!w.active) {
1769	<	int prev = w.lastEventCount;
1770	<	long nextTop = (((long)prev << EVENT_COUNT_SHIFT) |
1771	<	((long)(w.poolIndex + 1)));
1772	<	long top;
1773	<	while ((runState < SHUTDOWN || !tryTerminate(false)) &&
1774	<	(((int)(top = eventWaiters) & WAITER_INDEX_MASK) == 0 ||
1775	<	(int)(top >>> EVENT_COUNT_SHIFT) == prev) &&
1776	<	eventCount == prev) {
1777	<	if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
1778	<	w.nextWaiter = top, nextTop)) {
1779	<	accumulateStealCount(w); // transfer steals while idle
1780	<	Thread.interrupted();    // clear/ignore interrupt
1781	<	while (eventCount == prev)
1782	<	w.doPark();
1761	>	* If inactivating worker w has caused the pool to become
1762	>	* quiescent, checks for pool termination, and, so long as this is
1763	>	* not the only worker, waits for event for up to a given
1764	>	* duration.  On timeout, if ctl has not changed, terminates the
1765	>	* worker, which will in turn wake up another worker to possibly
1766	>	* repeat this process.
1767	>	*
1768	>	* @param w the calling worker
1769	>	* @param currentCtl the ctl value triggering possible quiescence
1770	>	* @param prevCtl the ctl value to restore if thread is terminated
1771	>	*/
1772	>	private void idleAwaitWork(WorkQueue w, long currentCtl, long prevCtl) {
1773	>	if (w != null && w.eventCount < 0 &&
1774	>	!tryTerminate(false, false) && (int)prevCtl != 0 &&
1775	>	ctl == currentCtl) {
1776	>	int dc = -(short)(currentCtl >>> TC_SHIFT);
1777	>	long parkTime = dc < 0 ? FAST_IDLE_TIMEOUT: (dc + 1) * IDLE_TIMEOUT;
1778	>	long deadline = System.nanoTime() + parkTime - TIMEOUT_SLOP;
1779	>	Thread wt = Thread.currentThread();
1780	>	while (ctl == currentCtl) {
1781	>	Thread.interrupted();  // timed variant of version in scan()
1782	>	U.putObject(wt, PARKBLOCKER, this);
1783	>	w.parker = wt;
1784	>	if (ctl == currentCtl)
1785	>	U.park(false, parkTime);
1786	>	w.parker = null;
1787	>	U.putObject(wt, PARKBLOCKER, null);
1788	>	if (ctl != currentCtl)
1789	>	break;
1790	>	if (deadline - System.nanoTime() <= 0L &&
1791	>	U.compareAndSwapLong(this, CTL, currentCtl, prevCtl)) {
1792	>	w.eventCount = (w.eventCount + E_SEQ) | E_MASK;
1793	>	w.hint = -1;
1794	>	w.qlock = -1;   // shrink
1795		break;
1796		}
1797		}
813	–	w.lastEventCount = eventCount;
1798		}
815	–	releaseWaiters();
1799		}
1800
1801		/**
1802	<	* Callback from workers invoked upon each top-level action (i.e.,
1803	<	* stealing a task or taking a submission and running
1804	<	* it). Performs one or both of the following:
1805	<	*
1806	<	* * If the worker cannot find work, updates its active status to
1807	<	* inactive and updates activeCount unless there is contention, in
1808	<	* which case it may try again (either in this or a subsequent
1809	<	* call).  Additionally, awaits the next task event and/or helps
1810	<	* wake up other releasable waiters.
1811	<	*
1812	<	* * If there are too many running threads, suspends this worker
1813	<	* (first forcing inactivation if necessary).  If it is not
1814	<	* resumed before a keepAlive elapses, the worker may be "trimmed"
1815	<	* -- killed while suspended within suspendAsSpare. Otherwise,
1816	<	* upon resume it rechecks to make sure that it is still needed.
1817	<	*
1818	<	* @param w the worker
1819	<	* @param worked false if the worker scanned for work but didn't
1820	<	* find any (in which case it may block waiting for work).
1821	<	*/
1822	<	final void preStep(ForkJoinWorkerThread w, boolean worked) {
1823	<	boolean active = w.active;
1824	<	boolean inactivate = !worked & active;
1825	<	for (;;) {
1826	<	if (inactivate) {
1827	<	int c = runState;
1828	<	if (UNSAFE.compareAndSwapInt(this, runStateOffset,
1829	<	c, c - ONE_ACTIVE))
1830	<	inactivate = active = w.active = false;
1831	<	}
1832	<	int wc = workerCounts;
1833	<	if ((wc & RUNNING_COUNT_MASK) <= parallelism) {
1834	<	if (!worked)
1835	<	eventSync(w);
1836	<	return;
1802	>	* Scans through queues looking for work while joining a task; if
1803	>	* any present, signals. May return early if more signalling is
1804	>	* detectably unneeded.
1805	>	*
1806	>	* @param task return early if done
1807	>	* @param origin an index to start scan
1808	>	*/
1809	>	private void helpSignal(ForkJoinTask<?> task, int origin) {
1810	>	WorkQueue[] ws; WorkQueue w; Thread p; long c; int m, u, e, i, s;
1811	>	if (task != null && task.status >= 0 &&
1812	>	(u = (int)(ctl >>> 32)) < 0 && (u >> UAC_SHIFT) < 0 &&
1813	>	(ws = workQueues) != null && (m = ws.length - 1) >= 0) {
1814	>	outer: for (int k = origin, j = m; j >= 0; --j) {
1815	>	WorkQueue q = ws[k++ & m];
1816	>	for (int n = m;;) { // limit to at most m signals
1817	>	if (task.status < 0)
1818	>	break outer;
1819	>	if (q == null ||
1820	>	((s = -q.base + q.top) <= n && (n = s) <= 0))
1821	>	break;
1822	>	if ((u = (int)((c = ctl) >>> 32)) >= 0 ||
1823	>	(e = (int)c) <= 0 || m < (i = e & SMASK) ||
1824	>	(w = ws[i]) == null)
1825	>	break outer;
1826	>	long nc = (((long)(w.nextWait & E_MASK)) |
1827	>	((long)(u + UAC_UNIT) << 32));
1828	>	if (w.eventCount != (e | INT_SIGN))
1829	>	break outer;
1830	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1831	>	w.eventCount = (e + E_SEQ) & E_MASK;
1832	>	if ((p = w.parker) != null)
1833	>	U.unpark(p);
1834	>	if (--n <= 0)
1835	>	break;
1836	>	}
1837	>	}
1838		}
855	–	if (!(inactivate |= active) &&  // must inactivate to suspend
856	–	UNSAFE.compareAndSwapInt(this, workerCountsOffset,
857	–	wc, wc - ONE_RUNNING) &&
858	–	!w.suspendAsSpare())        // false if trimmed
859	–	return;
1839		}
1840		}
1841
1842		/**
1843	<	* Adjusts counts and creates or resumes compensating threads for
1844	<	* a worker blocking on task joinMe.  First tries resuming an
1845	<	* existing spare (which usually also avoids any count
1846	<	* adjustment), but must then decrement running count to determine
1847	<	* whether a new thread is needed. See above for fuller
1848	<	* explanation. This code is sprawled out non-modularly mainly
1849	<	* because adaptive spinning works best if the entire method is
1850	<	* either interpreted or compiled vs having only some pieces of it
1851	<	* compiled.
1852	<	*
1853	<	* @param joinMe the task to join
1854	<	* @return task status on exit (to simplify usage by callers)
1855	<	*/
1856	<	final int awaitJoin(ForkJoinTask<?> joinMe) {
1857	<	int pc = parallelism;
1858	<	boolean adj = false;        // true when running count adjusted
1859	<	int scans = 0;
1860	<
1861	<	while (joinMe.status >= 0) {
1862	<	ForkJoinWorkerThread spare = null;
1863	<	if ((workerCounts & RUNNING_COUNT_MASK) < pc) {
1864	<	ForkJoinWorkerThread[] ws = workers;
1865	<	int nws = ws.length;
1866	<	for (int i = 0; i < nws; ++i) {
1867	<	ForkJoinWorkerThread w = ws[i];
1868	<	if (w != null && w.isSuspended()) {
1869	<	spare = w;
891	<	break;
1843	>	* Tries to locate and execute tasks for a stealer of the given
1844	>	* task, or in turn one of its stealers, Traces currentSteal ->
1845	>	* currentJoin links looking for a thread working on a descendant
1846	>	* of the given task and with a non-empty queue to steal back and
1847	>	* execute tasks from. The first call to this method upon a
1848	>	* waiting join will often entail scanning/search, (which is OK
1849	>	* because the joiner has nothing better to do), but this method
1850	>	* leaves hints in workers to speed up subsequent calls. The
1851	>	* implementation is very branchy to cope with potential
1852	>	* inconsistencies or loops encountering chains that are stale,
1853	>	* unknown, or so long that they are likely cyclic.
1854	>	*
1855	>	* @param joiner the joining worker
1856	>	* @param task the task to join
1857	>	* @return 0 if no progress can be made, negative if task
1858	>	* known complete, else positive
1859	>	*/
1860	>	private int tryHelpStealer(WorkQueue joiner, ForkJoinTask<?> task) {
1861	>	int stat = 0, steps = 0;                    // bound to avoid cycles
1862	>	if (joiner != null && task != null) {       // hoist null checks
1863	>	restart: for (;;) {
1864	>	ForkJoinTask<?> subtask = task;     // current target
1865	>	for (WorkQueue j = joiner, v;;) {   // v is stealer of subtask
1866	>	WorkQueue[] ws; int m, s, h;
1867	>	if ((s = task.status) < 0) {
1868	>	stat = s;
1869	>	break restart;
1870		}
1871	<	}
1872	<	if (joinMe.status < 0)
1873	<	break;
1874	<	}
1875	<	int wc = workerCounts;
1876	<	int rc = wc & RUNNING_COUNT_MASK;
1877	<	int dc = pc - rc;
1878	<	if (dc > 0 && spare != null && spare.tryUnsuspend()) {
1879	<	if (adj) {
1880	<	int c;
1881	<	do {} while (!UNSAFE.compareAndSwapInt
1882	<	(this, workerCountsOffset,
1883	<	c = workerCounts, c + ONE_RUNNING));
1884	<	}
1885	<	adj = true;
1886	<	LockSupport.unpark(spare);
1887	<	}
1888	<	else if (adj) {
1889	<	if (dc <= 0)
1890	<	break;
1891	<	int tc = wc >>> TOTAL_COUNT_SHIFT;
1892	<	if (scans > tc) {
1893	<	int ts = (tc - pc) * pc;
1894	<	if (rc != 0 &&  (dc * dc < ts || !maintainsParallelism))
1895	<	break;
1896	<	if (scans > ts && tc < maxPoolSize &&
1897	<	UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
1898	<	wc+(ONE_RUNNING|ONE_TOTAL))){
1899	<	addWorker();
1900	<	break;
1871	>	if ((ws = workQueues) == null || (m = ws.length - 1) <= 0)
1872	>	break restart;              // shutting down
1873	>	if ((v = ws[h = (j.hint | 1) & m]) == null ||
1874	>	v.currentSteal != subtask) {
1875	>	for (int origin = h;;) {    // find stealer
1876	>	if (((h = (h + 2) & m) & 15) == 1 &&
1877	>	(subtask.status < 0 || j.currentJoin != subtask))
1878	>	continue restart;   // occasional staleness check
1879	>	if ((v = ws[h]) != null &&
1880	>	v.currentSteal == subtask) {
1881	>	j.hint = h;        // save hint
1882	>	break;
1883	>	}
1884	>	if (h == origin)
1885	>	break restart;      // cannot find stealer
1886	>	}
1887	>	}
1888	>	for (;;) { // help stealer or descend to its stealer
1889	>	ForkJoinTask[] a;  int b;
1890	>	if (subtask.status < 0)     // surround probes with
1891	>	continue restart;       //   consistency checks
1892	>	if ((b = v.base) - v.top < 0 && (a = v.array) != null) {
1893	>	int i = (((a.length - 1) & b) << ASHIFT) + ABASE;
1894	>	ForkJoinTask<?> t =
1895	>	(ForkJoinTask<?>)U.getObjectVolatile(a, i);
1896	>	if (subtask.status < 0 || j.currentJoin != subtask ||
1897	>	v.currentSteal != subtask)
1898	>	continue restart;   // stale
1899	>	stat = 1;               // apparent progress
1900	>	if (t != null && v.base == b &&
1901	>	U.compareAndSwapObject(a, i, t, null)) {
1902	>	v.base = b + 1;     // help stealer
1903	>	joiner.runSubtask(t);
1904	>	}
1905	>	else if (v.base == b && ++steps == MAX_HELP)
1906	>	break restart;      // v apparently stalled
1907	>	}
1908	>	else {                      // empty -- try to descend
1909	>	ForkJoinTask<?> next = v.currentJoin;
1910	>	if (subtask.status < 0 || j.currentJoin != subtask ||
1911	>	v.currentSteal != subtask)
1912	>	continue restart;   // stale
1913	>	else if (next == null || ++steps == MAX_HELP)
1914	>	break restart;      // dead-end or maybe cyclic
1915	>	else {
1916	>	subtask = next;
1917	>	j = v;
1918	>	break;
1919	>	}
1920	>	}
1921		}
1922		}
1923		}
926	–	else if (rc != 0)
927	–	adj = UNSAFE.compareAndSwapInt (this, workerCountsOffset,
928	–	wc, wc - ONE_RUNNING);
929	–	if ((scans++ & 1) == 0)
930	–	releaseWaiters();   // help others progress
931	–	else
932	–	Thread.yield();     // avoid starving productive threads
933	–	}
934	–
935	–	if (adj) {
936	–	joinMe.internalAwaitDone();
937	–	int c;
938	–	do {} while (!UNSAFE.compareAndSwapInt
939	–	(this, workerCountsOffset,
940	–	c = workerCounts, c + ONE_RUNNING));
1924		}
1925	<	return joinMe.status;
1925	>	return stat;
1926		}
1927
1928		/**
1929	<	* Same idea as awaitJoin
1930	<	*/
1931	<	final void awaitBlocker(ManagedBlocker blocker, boolean maintainPar)
1932	<	throws InterruptedException {
1933	<	maintainPar &= maintainsParallelism;
1934	<	int pc = parallelism;
1935	<	boolean adj = false;        // true when running count adjusted
1936	<	int scans = 0;
1937	<	boolean done;
1938	<
1939	<	for (;;) {
1940	<	if (done = blocker.isReleasable())
1941	<	break;
1942	<	ForkJoinWorkerThread spare = null;
1943	<	if ((workerCounts & RUNNING_COUNT_MASK) < pc) {
1944	<	ForkJoinWorkerThread[] ws = workers;
1945	<	int nws = ws.length;
1946	<	for (int i = 0; i < nws; ++i) {
964	<	ForkJoinWorkerThread w = ws[i];
965	<	if (w != null && w.isSuspended()) {
966	<	spare = w;
1929	>	* Analog of tryHelpStealer for CountedCompleters. Tries to steal
1930	>	* and run tasks within the target's computation.
1931	>	*
1932	>	* @param task the task to join
1933	>	* @param mode if shared, exit upon completing any task
1934	>	* if all workers are active
1935	>	*/
1936	>	private int helpComplete(ForkJoinTask<?> task, int mode) {
1937	>	WorkQueue[] ws; WorkQueue q; int m, n, s, u;
1938	>	if (task != null && (ws = workQueues) != null &&
1939	>	(m = ws.length - 1) >= 0) {
1940	>	for (int j = 1, origin = j;;) {
1941	>	if ((s = task.status) < 0)
1942	>	return s;
1943	>	if ((q = ws[j & m]) != null && q.pollAndExecCC(task)) {
1944	>	origin = j;
1945	>	if (mode == SHARED_QUEUE &&
1946	>	((u = (int)(ctl >>> 32)) >= 0 || (u >> UAC_SHIFT) >= 0))
1947		break;
968	–	}
1948		}
1949	<	if (done = blocker.isReleasable())
1949	>	else if ((j = (j + 2) & m) == origin)
1950		break;
1951		}
1952	<	int wc = workerCounts;
1953	<	int rc = wc & RUNNING_COUNT_MASK;
1954	<	int dc = pc - rc;
1955	<	if (dc > 0 && spare != null && spare.tryUnsuspend()) {
1956	<	if (adj) {
1957	<	int c;
1958	<	do {} while (!UNSAFE.compareAndSwapInt
1959	<	(this, workerCountsOffset,
1960	<	c = workerCounts, c + ONE_RUNNING));
1952	>	}
1953	>	return 0;
1954	>	}
1955	>
1956	>	/**
1957	>	* Tries to decrement active count (sometimes implicitly) and
1958	>	* possibly release or create a compensating worker in preparation
1959	>	* for blocking. Fails on contention or termination. Otherwise,
1960	>	* adds a new thread if no idle workers are available and pool
1961	>	* may become starved.
1962	>	*/
1963	>	final boolean tryCompensate() {
1964	>	int pc = config & SMASK, e, i, tc; long c;
1965	>	WorkQueue[] ws; WorkQueue w; Thread p;
1966	>	if ((ws = workQueues) != null && (e = (int)(c = ctl)) >= 0) {
1967	>	if (e != 0 && (i = e & SMASK) < ws.length &&
1968	>	(w = ws[i]) != null && w.eventCount == (e | INT_SIGN)) {
1969	>	long nc = ((long)(w.nextWait & E_MASK) |
1970	>	(c & (AC_MASK|TC_MASK)));
1971	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1972	>	w.eventCount = (e + E_SEQ) & E_MASK;
1973	>	if ((p = w.parker) != null)
1974	>	U.unpark(p);
1975	>	return true;   // replace with idle worker
1976		}
983	–	adj = true;
984	–	LockSupport.unpark(spare);
1977		}
1978	<	else if (adj) {
1979	<	if (dc <= 0)
1980	<	break;
1981	<	int tc = wc >>> TOTAL_COUNT_SHIFT;
1982	<	if (scans > tc) {
1983	<	int ts = (tc - pc) * pc;
1984	<	if (rc != 0 &&  (dc * dc < ts || !maintainPar))
1985	<	break;
1986	<	if (scans > ts && tc < maxPoolSize &&
1987	<	UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
1988	<	wc+(ONE_RUNNING|ONE_TOTAL))){
1989	<	addWorker();
1990	<	break;
1978	>	else if ((tc = (short)(c >>> TC_SHIFT)) >= 0 &&
1979	>	(int)(c >> AC_SHIFT) + pc > 1) {
1980	>	long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK);
1981	>	if (U.compareAndSwapLong(this, CTL, c, nc))
1982	>	return true;   // no compensation
1983	>	}
1984	>	else if (tc + pc < MAX_CAP) {
1985	>	long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK);
1986	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1987	>	ForkJoinWorkerThreadFactory fac;
1988	>	Throwable ex = null;
1989	>	ForkJoinWorkerThread wt = null;
1990	>	try {
1991	>	if ((fac = factory) != null &&
1992	>	(wt = fac.newThread(this)) != null) {
1993	>	wt.start();
1994	>	return true;
1995	>	}
1996	>	} catch (Throwable rex) {
1997	>	ex = rex;
1998		}
1999	+	deregisterWorker(wt, ex); // clean up and return false
2000		}
2001		}
1002	–	else if (rc != 0)
1003	–	adj = UNSAFE.compareAndSwapInt (this, workerCountsOffset,
1004	–	wc, wc - ONE_RUNNING);
1005	–	if ((++scans & 1) == 0)
1006	–	releaseWaiters();   // help others progress
1007	–	else
1008	–	Thread.yield();     // avoid starving productive threads
1009	–	}
1010	–
1011	–	try {
1012	–	if (!done)
1013	–	do {} while (!blocker.isReleasable() && !blocker.block());
1014	–	} finally {
1015	–	if (adj) {
1016	–	int c;
1017	–	do {} while (!UNSAFE.compareAndSwapInt
1018	–	(this, workerCountsOffset,
1019	–	c = workerCounts, c + ONE_RUNNING));
1020	–	}
2002		}
2003	+	return false;
2004		}
2005
2006		/**
2007	<	* Unless there are not enough other running threads, adjusts
1026	<	* counts and blocks a worker performing helpJoin that cannot find
1027	<	* any work.
2007	>	* Helps and/or blocks until the given task is done.
2008		*
2009	<	* @return true if joinMe now done
2009	>	* @param joiner the joining worker
2010	>	* @param task the task
2011	>	* @return task status on exit
2012		*/
2013	<	final boolean tryAwaitBusyJoin(ForkJoinTask<?> joinMe) {
2014	<	int pc = parallelism;
2015	<	outer:for (;;) {
2016	<	releaseWaiters();
2017	<	if ((workerCounts & RUNNING_COUNT_MASK) < pc) {
2018	<	ForkJoinWorkerThread[] ws = workers;
2019	<	int nws = ws.length;
2020	<	for (int i = 0; i < nws; ++i) {
2021	<	ForkJoinWorkerThread w = ws[i];
2022	<	if (w != null && w.isSuspended()) {
2023	<	if (joinMe.status < 0)
2024	<	return true;
2025	<	if ((workerCounts & RUNNING_COUNT_MASK) > pc)
2026	<	break;
2027	<	if (w.tryUnsuspend()) {
2028	<	LockSupport.unpark(w);
2029	<	break outer;
2013	>	final int awaitJoin(WorkQueue joiner, ForkJoinTask<?> task) {
2014	>	int s = 0;
2015	>	if (joiner != null && task != null && (s = task.status) >= 0) {
2016	>	ForkJoinTask<?> prevJoin = joiner.currentJoin;
2017	>	joiner.currentJoin = task;
2018	>	do {} while ((s = task.status) >= 0 && !joiner.isEmpty() &&
2019	>	joiner.tryRemoveAndExec(task)); // process local tasks
2020	>	if (s >= 0 && (s = task.status) >= 0) {
2021	>	helpSignal(task, joiner.poolIndex);
2022	>	if ((s = task.status) >= 0 &&
2023	>	(task instanceof CountedCompleter))
2024	>	s = helpComplete(task, LIFO_QUEUE);
2025	>	}
2026	>	while (s >= 0 && (s = task.status) >= 0) {
2027	>	if ((!joiner.isEmpty() ||           // try helping
2028	>	(s = tryHelpStealer(joiner, task)) == 0) &&
2029	>	(s = task.status) >= 0) {
2030	>	helpSignal(task, joiner.poolIndex);
2031	>	if ((s = task.status) >= 0 && tryCompensate()) {
2032	>	if (task.trySetSignal() && (s = task.status) >= 0) {
2033	>	synchronized (task) {
2034	>	if (task.status >= 0) {
2035	>	try {                // see ForkJoinTask
2036	>	task.wait();     //  for explanation
2037	>	} catch (InterruptedException ie) {
2038	>	}
2039	>	}
2040	>	else
2041	>	task.notifyAll();
2042	>	}
2043		}
2044	<	continue outer;
2044	>	long c;                          // re-activate
2045	>	do {} while (!U.compareAndSwapLong
2046	>	(this, CTL, c = ctl, c + AC_UNIT));
2047		}
2048		}
2049		}
2050	<	if (joinMe.status < 0)
1054	<	return true;
1055	<	int wc = workerCounts;
1056	<	if ((wc & RUNNING_COUNT_MASK) <= 2 ||
1057	<	(wc >>> TOTAL_COUNT_SHIFT) < pc)
1058	<	return false;  // keep this thread alive
1059	<	if (UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1060	<	wc, wc - ONE_RUNNING))
1061	<	break;
2050	>	joiner.currentJoin = prevJoin;
2051		}
2052	<
1064	<	joinMe.internalAwaitDone();
1065	<	int c;
1066	<	do {} while (!UNSAFE.compareAndSwapInt
1067	<	(this, workerCountsOffset,
1068	<	c = workerCounts, c + ONE_RUNNING));
1069	<	return true;
2052	>	return s;
2053		}
2054
2055		/**
2056	<	* Possibly initiates and/or completes termination.
2056	>	* Stripped-down variant of awaitJoin used by timed joins. Tries
2057	>	* to help join only while there is continuous progress. (Caller
2058	>	* will then enter a timed wait.)
2059		*
2060	<	* @param now if true, unconditionally terminate, else only
2061	<	* if shutdown and empty queue and no active workers
1077	<	* @return true if now terminating or terminated
1078	<	*/
1079	<	private boolean tryTerminate(boolean now) {
1080	<	if (now)
1081	<	advanceRunLevel(SHUTDOWN); // ensure at least SHUTDOWN
1082	<	else if (runState < SHUTDOWN ||
1083	<	!submissionQueue.isEmpty() ||
1084	<	(runState & ACTIVE_COUNT_MASK) != 0)
1085	<	return false;
1086	<
1087	<	if (advanceRunLevel(TERMINATING))
1088	<	startTerminating();
1089	<
1090	<	// Finish now if all threads terminated; else in some subsequent call
1091	<	if ((workerCounts >>> TOTAL_COUNT_SHIFT) == 0) {
1092	<	advanceRunLevel(TERMINATED);
1093	<	terminationLatch.countDown();
1094	<	}
1095	<	return true;
1096	<	}
1097	<
1098	<	/**
1099	<	* Actions on transition to TERMINATING
2060	>	* @param joiner the joining worker
2061	>	* @param task the task
2062		*/
2063	<	private void startTerminating() {
2064	<	for (int i = 0; i < 2; ++i) { // twice to mop up newly created workers
2065	<	cancelSubmissions();
2066	<	shutdownWorkers();
2067	<	cancelWorkerTasks();
2068	<	advanceEventCount();
2069	<	releaseWaiters();
2070	<	interruptWorkers();
2063	>	final void helpJoinOnce(WorkQueue joiner, ForkJoinTask<?> task) {
2064	>	int s;
2065	>	if (joiner != null && task != null && (s = task.status) >= 0) {
2066	>	ForkJoinTask<?> prevJoin = joiner.currentJoin;
2067	>	joiner.currentJoin = task;
2068	>	do {} while ((s = task.status) >= 0 && !joiner.isEmpty() &&
2069	>	joiner.tryRemoveAndExec(task));
2070	>	if (s >= 0 && (s = task.status) >= 0) {
2071	>	helpSignal(task, joiner.poolIndex);
2072	>	if ((s = task.status) >= 0 &&
2073	>	(task instanceof CountedCompleter))
2074	>	s = helpComplete(task, LIFO_QUEUE);
2075	>	}
2076	>	if (s >= 0 && joiner.isEmpty()) {
2077	>	do {} while (task.status >= 0 &&
2078	>	tryHelpStealer(joiner, task) > 0);
2079	>	}
2080	>	joiner.currentJoin = prevJoin;
2081		}
2082		}
2083
2084		/**
2085	<	* Clear out and cancel submissions, ignoring exceptions
2085	>	* Returns a (probably) non-empty steal queue, if one is found
2086	>	* during a scan, else null.  This method must be retried by
2087	>	* caller if, by the time it tries to use the queue, it is empty.
2088	>	* @param r a (random) seed for scanning
2089		*/
2090	<	private void cancelSubmissions() {
2091	<	ForkJoinTask<?> task;
2092	<	while ((task = submissionQueue.poll()) != null) {
2093	<	try {
2094	<	task.cancel(false);
2095	<	} catch (Throwable ignore) {
2090	>	private WorkQueue findNonEmptyStealQueue(int r) {
2091	>	for (;;) {
2092	>	int ps = plock, m; WorkQueue[] ws; WorkQueue q;
2093	>	if ((ws = workQueues) != null && (m = ws.length - 1) >= 0) {
2094	>	for (int j = (m + 1) << 2; j >= 0; --j) {
2095	>	if ((q = ws[(((r + j) << 1) | 1) & m]) != null &&
2096	>	q.base - q.top < 0)
2097	>	return q;
2098	>	}
2099		}
2100	+	if (plock == ps)
2101	+	return null;
2102		}
2103		}
2104
2105		/**
2106	<	* Sets all worker run states to at least shutdown,
2107	<	* also resuming suspended workers
2108	<	*/
2109	<	private void shutdownWorkers() {
2110	<	ForkJoinWorkerThread[] ws = workers;
2111	<	int nws = ws.length;
2112	<	for (int i = 0; i < nws; ++i) {
2113	<	ForkJoinWorkerThread w = ws[i];
2114	<	if (w != null)
2115	<	w.shutdown();
2106	>	* Runs tasks until {@code isQuiescent()}. We piggyback on
2107	>	* active count ctl maintenance, but rather than blocking
2108	>	* when tasks cannot be found, we rescan until all others cannot
2109	>	* find tasks either.
2110	>	*/
2111	>	final void helpQuiescePool(WorkQueue w) {
2112	>	for (boolean active = true;;) {
2113	>	long c; WorkQueue q; ForkJoinTask<?> t; int b;
2114	>	while ((t = w.nextLocalTask()) != null) {
2115	>	if (w.base - w.top < 0)
2116	>	signalWork(w);
2117	>	t.doExec();
2118	>	}
2119	>	if ((q = findNonEmptyStealQueue(w.nextSeed())) != null) {
2120	>	if (!active) {      // re-establish active count
2121	>	active = true;
2122	>	do {} while (!U.compareAndSwapLong
2123	>	(this, CTL, c = ctl, c + AC_UNIT));
2124	>	}
2125	>	if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null) {
2126	>	if (q.base - q.top < 0)
2127	>	signalWork(q);
2128	>	w.runSubtask(t);
2129	>	}
2130	>	}
2131	>	else if (active) {       // decrement active count without queuing
2132	>	long nc = (c = ctl) - AC_UNIT;
2133	>	if ((int)(nc >> AC_SHIFT) + (config & SMASK) == 0)
2134	>	return;          // bypass decrement-then-increment
2135	>	if (U.compareAndSwapLong(this, CTL, c, nc))
2136	>	active = false;
2137	>	}
2138	>	else if ((int)((c = ctl) >> AC_SHIFT) + (config & SMASK) == 0 &&
2139	>	U.compareAndSwapLong(this, CTL, c, c + AC_UNIT))
2140	>	return;
2141		}
2142		}
2143
2144		/**
2145	<	* Clears out and cancels all locally queued tasks
2145	>	* Gets and removes a local or stolen task for the given worker.
2146	>	*
2147	>	* @return a task, if available
2148		*/
2149	<	private void cancelWorkerTasks() {
2150	<	ForkJoinWorkerThread[] ws = workers;
2151	<	int nws = ws.length;
2152	<	for (int i = 0; i < nws; ++i) {
2153	<	ForkJoinWorkerThread w = ws[i];
2154	<	if (w != null)
2155	<	w.cancelTasks();
2149	>	final ForkJoinTask<?> nextTaskFor(WorkQueue w) {
2150	>	for (ForkJoinTask<?> t;;) {
2151	>	WorkQueue q; int b;
2152	>	if ((t = w.nextLocalTask()) != null)
2153	>	return t;
2154	>	if ((q = findNonEmptyStealQueue(w.nextSeed())) == null)
2155	>	return null;
2156	>	if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null) {
2157	>	if (q.base - q.top < 0)
2158	>	signalWork(q);
2159	>	return t;
2160	>	}
2161		}
2162		}
2163
2164		/**
2165	<	* Unsticks all workers blocked on joins etc
2165	>	* Returns a cheap heuristic guide for task partitioning when
2166	>	* programmers, frameworks, tools, or languages have little or no
2167	>	* idea about task granularity.  In essence by offering this
2168	>	* method, we ask users only about tradeoffs in overhead vs
2169	>	* expected throughput and its variance, rather than how finely to
2170	>	* partition tasks.
2171	>	*
2172	>	* In a steady state strict (tree-structured) computation, each
2173	>	* thread makes available for stealing enough tasks for other
2174	>	* threads to remain active. Inductively, if all threads play by
2175	>	* the same rules, each thread should make available only a
2176	>	* constant number of tasks.
2177	>	*
2178	>	* The minimum useful constant is just 1. But using a value of 1
2179	>	* would require immediate replenishment upon each steal to
2180	>	* maintain enough tasks, which is infeasible.  Further,
2181	>	* partitionings/granularities of offered tasks should minimize
2182	>	* steal rates, which in general means that threads nearer the top
2183	>	* of computation tree should generate more than those nearer the
2184	>	* bottom. In perfect steady state, each thread is at
2185	>	* approximately the same level of computation tree. However,
2186	>	* producing extra tasks amortizes the uncertainty of progress and
2187	>	* diffusion assumptions.
2188	>	*
2189	>	* So, users will want to use values larger, but not much larger
2190	>	* than 1 to both smooth over transient shortages and hedge
2191	>	* against uneven progress; as traded off against the cost of
2192	>	* extra task overhead. We leave the user to pick a threshold
2193	>	* value to compare with the results of this call to guide
2194	>	* decisions, but recommend values such as 3.
2195	>	*
2196	>	* When all threads are active, it is on average OK to estimate
2197	>	* surplus strictly locally. In steady-state, if one thread is
2198	>	* maintaining say 2 surplus tasks, then so are others. So we can
2199	>	* just use estimated queue length.  However, this strategy alone
2200	>	* leads to serious mis-estimates in some non-steady-state
2201	>	* conditions (ramp-up, ramp-down, other stalls). We can detect
2202	>	* many of these by further considering the number of "idle"
2203	>	* threads, that are known to have zero queued tasks, so
2204	>	* compensate by a factor of (#idle/#active) threads.
2205	>	*
2206	>	* Note: The approximation of #busy workers as #active workers is
2207	>	* not very good under current signalling scheme, and should be
2208	>	* improved.
2209	>	*/
2210	>	static int getSurplusQueuedTaskCount() {
2211	>	Thread t; ForkJoinWorkerThread wt; ForkJoinPool pool; WorkQueue q;
2212	>	if (((t = Thread.currentThread()) instanceof ForkJoinWorkerThread)) {
2213	>	int p = (pool = (wt = (ForkJoinWorkerThread)t).pool).config & SMASK;
2214	>	int n = (q = wt.workQueue).top - q.base;
2215	>	int a = (int)(pool.ctl >> AC_SHIFT) + p;
2216	>	return n - (a > (p >>>= 1) ? 0 :
2217	>	a > (p >>>= 1) ? 1 :
2218	>	a > (p >>>= 1) ? 2 :
2219	>	a > (p >>>= 1) ? 4 :
2220	>	8);
2221	>	}
2222	>	return 0;
2223	>	}
2224	>
2225	>	//  Termination
2226	>
2227	>	/**
2228	>	* Possibly initiates and/or completes termination.  The caller
2229	>	* triggering termination runs three passes through workQueues:
2230	>	* (0) Setting termination status, followed by wakeups of queued
2231	>	* workers; (1) cancelling all tasks; (2) interrupting lagging
2232	>	* threads (likely in external tasks, but possibly also blocked in
2233	>	* joins).  Each pass repeats previous steps because of potential
2234	>	* lagging thread creation.
2235	>	*
2236	>	* @param now if true, unconditionally terminate, else only
2237	>	* if no work and no active workers
2238	>	* @param enable if true, enable shutdown when next possible
2239	>	* @return true if now terminating or terminated
2240		*/
2241	<	private void interruptWorkers() {
2242	<	ForkJoinWorkerThread[] ws = workers;
2243	<	int nws = ws.length;
2244	<	for (int i = 0; i < nws; ++i) {
2245	<	ForkJoinWorkerThread w = ws[i];
2246	<	if (w != null && !w.isTerminated()) {
2247	<	try {
2248	<	w.interrupt();
2249	<	} catch (SecurityException ignore) {
2241	>	private boolean tryTerminate(boolean now, boolean enable) {
2242	>	int ps;
2243	>	if (this == common)                    // cannot shut down
2244	>	return false;
2245	>	if ((ps = plock) >= 0) {                   // enable by setting plock
2246	>	if (!enable)
2247	>	return false;
2248	>	if ((ps & PL_LOCK) != 0 ||
2249	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
2250	>	ps = acquirePlock();
2251	>	int nps = ((ps + PL_LOCK) & ~SHUTDOWN) | SHUTDOWN;
2252	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
2253	>	releasePlock(nps);
2254	>	}
2255	>	for (long c;;) {
2256	>	if (((c = ctl) & STOP_BIT) != 0) {     // already terminating
2257	>	if ((short)(c >>> TC_SHIFT) == -(config & SMASK)) {
2258	>	synchronized (this) {
2259	>	notifyAll();               // signal when 0 workers
2260	>	}
2261	>	}
2262	>	return true;
2263	>	}
2264	>	if (!now) {                            // check if idle & no tasks
2265	>	WorkQueue[] ws; WorkQueue w;
2266	>	if ((int)(c >> AC_SHIFT) != -(config & SMASK))
2267	>	return false;
2268	>	if ((ws = workQueues) != null) {
2269	>	for (int i = 0; i < ws.length; ++i) {
2270	>	if ((w = ws[i]) != null) {
2271	>	if (!w.isEmpty()) {    // signal unprocessed tasks
2272	>	signalWork(w);
2273	>	return false;
2274	>	}
2275	>	if ((i & 1) != 0 && w.eventCount >= 0)
2276	>	return false;      // unqueued inactive worker
2277	>	}
2278	>	}
2279	>	}
2280	>	}
2281	>	if (U.compareAndSwapLong(this, CTL, c, c | STOP_BIT)) {
2282	>	for (int pass = 0; pass < 3; ++pass) {
2283	>	WorkQueue[] ws; WorkQueue w; Thread wt;
2284	>	if ((ws = workQueues) != null) {
2285	>	int n = ws.length;
2286	>	for (int i = 0; i < n; ++i) {
2287	>	if ((w = ws[i]) != null) {
2288	>	w.qlock = -1;
2289	>	if (pass > 0) {
2290	>	w.cancelAll();
2291	>	if (pass > 1 && (wt = w.owner) != null) {
2292	>	if (!wt.isInterrupted()) {
2293	>	try {
2294	>	wt.interrupt();
2295	>	} catch (Throwable ignore) {
2296	>	}
2297	>	}
2298	>	U.unpark(wt);
2299	>	}
2300	>	}
2301	>	}
2302	>	}
2303	>	// Wake up workers parked on event queue
2304	>	int i, e; long cc; Thread p;
2305	>	while ((e = (int)(cc = ctl) & E_MASK) != 0 &&
2306	>	(i = e & SMASK) < n && i >= 0 &&
2307	>	(w = ws[i]) != null) {
2308	>	long nc = ((long)(w.nextWait & E_MASK) |
2309	>	((cc + AC_UNIT) & AC_MASK) |
2310	>	(cc & (TC_MASK|STOP_BIT)));
2311	>	if (w.eventCount == (e | INT_SIGN) &&
2312	>	U.compareAndSwapLong(this, CTL, cc, nc)) {
2313	>	w.eventCount = (e + E_SEQ) & E_MASK;
2314	>	w.qlock = -1;
2315	>	if ((p = w.parker) != null)
2316	>	U.unpark(p);
2317	>	}
2318	>	}
2319	>	}
2320		}
2321		}
2322		}
2323		}
2324
2325	<	// misc support for ForkJoinWorkerThread
2325	>	// external operations on common pool
2326
2327		/**
2328	<	* Returns pool number
2328	>	* Returns common pool queue for a thread that has submitted at
2329	>	* least one task.
2330		*/
2331	<	final int getPoolNumber() {
2332	<	return poolNumber;
2331	>	static WorkQueue commonSubmitterQueue() {
2332	>	ForkJoinPool p; WorkQueue[] ws; int m; Submitter z;
2333	>	return ((z = submitters.get()) != null &&
2334	>	(p = common) != null &&
2335	>	(ws = p.workQueues) != null &&
2336	>	(m = ws.length - 1) >= 0) ?
2337	>	ws[m & z.seed & SQMASK] : null;
2338	>	}
2339	>
2340	>	/**
2341	>	* Tries to pop the given task from submitter's queue in common pool.
2342	>	*/
2343	>	static boolean tryExternalUnpush(ForkJoinTask<?> t) {
2344	>	ForkJoinPool p; WorkQueue[] ws; WorkQueue q; Submitter z;
2345	>	ForkJoinTask<?>[] a;  int m, s;
2346	>	if (t != null &&
2347	>	(z = submitters.get()) != null &&
2348	>	(p = common) != null &&
2349	>	(ws = p.workQueues) != null &&
2350	>	(m = ws.length - 1) >= 0 &&
2351	>	(q = ws[m & z.seed & SQMASK]) != null &&
2352	>	(s = q.top) != q.base &&
2353	>	(a = q.array) != null) {
2354	>	long j = (((a.length - 1) & (s - 1)) << ASHIFT) + ABASE;
2355	>	if (U.getObject(a, j) == t &&
2356	>	U.compareAndSwapInt(q, QLOCK, 0, 1)) {
2357	>	if (q.array == a && q.top == s && // recheck
2358	>	U.compareAndSwapObject(a, j, t, null)) {
2359	>	q.top = s - 1;
2360	>	q.qlock = 0;
2361	>	return true;
2362	>	}
2363	>	q.qlock = 0;
2364	>	}
2365	>	}
2366	>	return false;
2367		}
2368
2369		/**
2370	<	* Accumulates steal count from a worker, clearing
2371	<	* the worker's value
2370	>	* Tries to pop and run local tasks within the same computation
2371	>	* as the given root. On failure, tries to help complete from
2372	>	* other queues via helpComplete.
2373		*/
2374	<	final void accumulateStealCount(ForkJoinWorkerThread w) {
2375	<	int sc = w.stealCount;
2376	<	if (sc != 0) {
2377	<	long c;
2378	<	w.stealCount = 0;
2379	<	do {} while (!UNSAFE.compareAndSwapLong(this, stealCountOffset,
2380	<	c = stealCount, c + sc));
2374	>	private void externalHelpComplete(WorkQueue q, ForkJoinTask<?> root) {
2375	>	ForkJoinTask<?>[] a; int m;
2376	>	if (q != null && (a = q.array) != null && (m = (a.length - 1)) >= 0 &&
2377	>	root != null && root.status >= 0) {
2378	>	for (;;) {
2379	>	int s, u; Object o; CountedCompleter<?> task = null;
2380	>	if ((s = q.top) - q.base > 0) {
2381	>	long j = ((m & (s - 1)) << ASHIFT) + ABASE;
2382	>	if ((o = U.getObject(a, j)) != null &&
2383	>	(o instanceof CountedCompleter)) {
2384	>	CountedCompleter<?> t = (CountedCompleter<?>)o, r = t;
2385	>	do {
2386	>	if (r == root) {
2387	>	if (U.compareAndSwapInt(q, QLOCK, 0, 1)) {
2388	>	if (q.array == a && q.top == s &&
2389	>	U.compareAndSwapObject(a, j, t, null)) {
2390	>	q.top = s - 1;
2391	>	task = t;
2392	>	}
2393	>	q.qlock = 0;
2394	>	}
2395	>	break;
2396	>	}
2397	>	} while ((r = r.completer) != null);
2398	>	}
2399	>	}
2400	>	if (task != null)
2401	>	task.doExec();
2402	>	if (root.status < 0 ||
2403	>	(u = (int)(ctl >>> 32)) >= 0 || (u >> UAC_SHIFT) >= 0)
2404	>	break;
2405	>	if (task == null) {
2406	>	helpSignal(root, q.poolIndex);
2407	>	if (root.status >= 0)
2408	>	helpComplete(root, SHARED_QUEUE);
2409	>	break;
2410	>	}
2411	>	}
2412		}
2413		}
2414
2415		/**
2416	<	* Returns the approximate (non-atomic) number of idle threads per
2417	<	* active thread.
2416	>	* Tries to help execute or signal availability of the given task
2417	>	* from submitter's queue in common pool.
2418		*/
2419	<	final int idlePerActive() {
2420	<	int ac = runState;    // no mask -- artifically boosts during shutdown
2421	<	int pc = parallelism; // use targeted parallelism, not rc
2422	<	// Use exact results for small values, saturate past 4
2423	<	return pc <= ac? 0 : pc >>> 1 <= ac? 1 : pc >>> 2 <= ac? 3 : pc >>> 3;
2419	>	static void externalHelpJoin(ForkJoinTask<?> t) {
2420	>	// Some hard-to-avoid overlap with tryExternalUnpush
2421	>	ForkJoinPool p; WorkQueue[] ws; WorkQueue q, w; Submitter z;
2422	>	ForkJoinTask<?>[] a;  int m, s, n;
2423	>	if (t != null &&
2424	>	(z = submitters.get()) != null &&
2425	>	(p = common) != null &&
2426	>	(ws = p.workQueues) != null &&
2427	>	(m = ws.length - 1) >= 0 &&
2428	>	(q = ws[m & z.seed & SQMASK]) != null &&
2429	>	(a = q.array) != null) {
2430	>	int am = a.length - 1;
2431	>	if ((s = q.top) != q.base) {
2432	>	long j = ((am & (s - 1)) << ASHIFT) + ABASE;
2433	>	if (U.getObject(a, j) == t &&
2434	>	U.compareAndSwapInt(q, QLOCK, 0, 1)) {
2435	>	if (q.array == a && q.top == s &&
2436	>	U.compareAndSwapObject(a, j, t, null)) {
2437	>	q.top = s - 1;
2438	>	q.qlock = 0;
2439	>	t.doExec();
2440	>	}
2441	>	else
2442	>	q.qlock = 0;
2443	>	}
2444	>	}
2445	>	if (t.status >= 0) {
2446	>	if (t instanceof CountedCompleter)
2447	>	p.externalHelpComplete(q, t);
2448	>	else
2449	>	p.helpSignal(t, q.poolIndex);
2450	>	}
2451	>	}
2452		}
2453
2454	<	// Public and protected methods
2454	>	// Exported methods
2455
2456		// Constructors
2457
2458		/**
2459		* Creates a {@code ForkJoinPool} with parallelism equal to {@link
2460	<	* java.lang.Runtime#availableProcessors}, and using the {@linkplain
2461	<	* #defaultForkJoinWorkerThreadFactory default thread factory}.
2460	>	* java.lang.Runtime#availableProcessors}, using the {@linkplain
2461	>	* #defaultForkJoinWorkerThreadFactory default thread factory},
2462	>	* no UncaughtExceptionHandler, and non-async LIFO processing mode.
2463		*
2464		* @throws SecurityException if a security manager exists and
2465		*         the caller is not permitted to modify threads
#	Line 1216 | Line 2468 | public class ForkJoinPool extends Abstra
2468		*/
2469		public ForkJoinPool() {
2470		this(Runtime.getRuntime().availableProcessors(),
2471	<	defaultForkJoinWorkerThreadFactory);
2471	>	defaultForkJoinWorkerThreadFactory, null, false);
2472		}
2473
2474		/**
2475		* Creates a {@code ForkJoinPool} with the indicated parallelism
2476	<	* level and using the {@linkplain
2477	<	* #defaultForkJoinWorkerThreadFactory default thread factory}.
2476	>	* level, the {@linkplain
2477	>	* #defaultForkJoinWorkerThreadFactory default thread factory},
2478	>	* no UncaughtExceptionHandler, and non-async LIFO processing mode.
2479		*
2480		* @param parallelism the parallelism level
2481		* @throws IllegalArgumentException if parallelism less than or
#	Line 1233 | Line 2486 | public class ForkJoinPool extends Abstra
2486		*         java.lang.RuntimePermission}{@code ("modifyThread")}
2487		*/
2488		public ForkJoinPool(int parallelism) {
2489	<	this(parallelism, defaultForkJoinWorkerThreadFactory);
1237	<	}
1238	<
1239	<	/**
1240	<	* Creates a {@code ForkJoinPool} with parallelism equal to {@link
1241	<	* java.lang.Runtime#availableProcessors}, and using the given
1242	<	* thread factory.
1243	<	*
1244	<	* @param factory the factory for creating new threads
1245	<	* @throws NullPointerException if the factory is null
1246	<	* @throws SecurityException if a security manager exists and
1247	<	*         the caller is not permitted to modify threads
1248	<	*         because it does not hold {@link
1249	<	*         java.lang.RuntimePermission}{@code ("modifyThread")}
1250	<	*/
1251	<	public ForkJoinPool(ForkJoinWorkerThreadFactory factory) {
1252	<	this(Runtime.getRuntime().availableProcessors(), factory);
2489	>	this(parallelism, defaultForkJoinWorkerThreadFactory, null, false);
2490		}
2491
2492		/**
2493	<	* Creates a {@code ForkJoinPool} with the given parallelism and
1257	<	* thread factory.
2493	>	* Creates a {@code ForkJoinPool} with the given parameters.
2494		*
2495	<	* @param parallelism the parallelism level
2496	<	* @param factory the factory for creating new threads
2495	>	* @param parallelism the parallelism level. For default value,
2496	>	* use {@link java.lang.Runtime#availableProcessors}.
2497	>	* @param factory the factory for creating new threads. For default value,
2498	>	* use {@link #defaultForkJoinWorkerThreadFactory}.
2499	>	* @param handler the handler for internal worker threads that
2500	>	* terminate due to unrecoverable errors encountered while executing
2501	>	* tasks. For default value, use {@code null}.
2502	>	* @param asyncMode if true,
2503	>	* establishes local first-in-first-out scheduling mode for forked
2504	>	* tasks that are never joined. This mode may be more appropriate
2505	>	* than default locally stack-based mode in applications in which
2506	>	* worker threads only process event-style asynchronous tasks.
2507	>	* For default value, use {@code false}.
2508		* @throws IllegalArgumentException if parallelism less than or
2509		*         equal to zero, or greater than implementation limit
2510		* @throws NullPointerException if the factory is null
#	Line 1266 | Line 2513 | public class ForkJoinPool extends Abstra
2513		*         because it does not hold {@link
2514		*         java.lang.RuntimePermission}{@code ("modifyThread")}
2515		*/
2516	<	public ForkJoinPool(int parallelism, ForkJoinWorkerThreadFactory factory) {
2516	>	public ForkJoinPool(int parallelism,
2517	>	ForkJoinWorkerThreadFactory factory,
2518	>	Thread.UncaughtExceptionHandler handler,
2519	>	boolean asyncMode) {
2520		checkPermission();
2521		if (factory == null)
2522		throw new NullPointerException();
2523	<	if (parallelism <= 0 || parallelism > MAX_THREADS)
2523	>	if (parallelism <= 0 || parallelism > MAX_CAP)
2524		throw new IllegalArgumentException();
1275	–	this.poolNumber = poolNumberGenerator.incrementAndGet();
1276	–	int arraySize = initialArraySizeFor(parallelism);
1277	–	this.parallelism = parallelism;
2525		this.factory = factory;
2526	<	this.maxPoolSize = MAX_THREADS;
2527	<	this.maintainsParallelism = true;
2528	<	this.workers = new ForkJoinWorkerThread[arraySize];
2529	<	this.submissionQueue = new LinkedTransferQueue<ForkJoinTask<?>>();
2530	<	this.workerLock = new ReentrantLock();
2531	<	this.terminationLatch = new CountDownLatch(1);
2526	>	this.ueh = handler;
2527	>	this.config = parallelism | (asyncMode ? (FIFO_QUEUE << 16) : 0);
2528	>	long np = (long)(-parallelism); // offset ctl counts
2529	>	this.ctl = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK);
2530	>	int pn = nextPoolId();
2531	>	StringBuilder sb = new StringBuilder("ForkJoinPool-");
2532	>	sb.append(Integer.toString(pn));
2533	>	sb.append("-worker-");
2534	>	this.workerNamePrefix = sb.toString();
2535		}
2536
2537		/**
2538	<	* Returns initial power of two size for workers array.
2539	<	* @param pc the initial parallelism level
2540	<	*/
2541	<	private static int initialArraySizeFor(int pc) {
2542	<	// See Hackers Delight, sec 3.2. We know MAX_THREADS < (1 >>> 16)
2543	<	int size = pc < MAX_THREADS ? pc + 1 : MAX_THREADS;
2544	<	size |= size >>> 1;
2545	<	size |= size >>> 2;
2546	<	size |= size >>> 4;
2547	<	size |= size >>> 8;
2548	<	return size + 1;
2538	>	* Constructor for common pool, suitable only for static initialization.
2539	>	* Basically the same as above, but uses smallest possible initial footprint.
2540	>	*/
2541	>	ForkJoinPool(int parallelism, long ctl,
2542	>	ForkJoinWorkerThreadFactory factory,
2543	>	Thread.UncaughtExceptionHandler handler) {
2544	>	this.config = parallelism;
2545	>	this.ctl = ctl;
2546	>	this.factory = factory;
2547	>	this.ueh = handler;
2548	>	this.workerNamePrefix = "ForkJoinPool.commonPool-worker-";
2549		}
2550
1301	–	// Execution methods
1302	–
2551		/**
2552	<	* Common code for execute, invoke and submit
2553	<	*/
2554	<	private <T> void doSubmit(ForkJoinTask<T> task) {
2555	<	if (task == null)
2556	<	throw new NullPointerException();
2557	<	if (runState >= SHUTDOWN)
2558	<	throw new RejectedExecutionException();
2559	<	submissionQueue.offer(task);
2560	<	advanceEventCount();
2561	<	releaseWaiters();
2562	<	ensureEnoughTotalWorkers();
2552	>	* Returns the common pool instance. This pool is statically
2553	>	* constructed; its run state is unaffected by attempts to {@link
2554	>	* #shutdown} or {@link #shutdownNow}. However this pool and any
2555	>	* ongoing processing are automatically terminated upon program
2556	>	* {@link System#exit}.  Any program that relies on asynchronous
2557	>	* task processing to complete before program termination should
2558	>	* invoke {@code commonPool().}{@link #awaitQuiescence}, before
2559	>	* exit.
2560	>	*
2561	>	* @return the common pool instance
2562	>	* @since 1.8
2563	>	*/
2564	>	public static ForkJoinPool commonPool() {
2565	>	// assert common != null : "static init error";
2566	>	return common;
2567		}
2568
2569	+	// Execution methods
2570	+
2571		/**
2572		* Performs the given task, returning its result upon completion.
2573	+	* If the computation encounters an unchecked Exception or Error,
2574	+	* it is rethrown as the outcome of this invocation.  Rethrown
2575	+	* exceptions behave in the same way as regular exceptions, but,
2576	+	* when possible, contain stack traces (as displayed for example
2577	+	* using {@code ex.printStackTrace()}) of both the current thread
2578	+	* as well as the thread actually encountering the exception;
2579	+	* minimally only the latter.
2580		*
2581		* @param task the task
2582		* @return the task's result
#	Line 1324 | Line 2585 | public class ForkJoinPool extends Abstra
2585		*         scheduled for execution
2586		*/
2587		public <T> T invoke(ForkJoinTask<T> task) {
2588	<	doSubmit(task);
2588	>	if (task == null)
2589	>	throw new NullPointerException();
2590	>	externalPush(task);
2591		return task.join();
2592		}
2593
#	Line 1337 | Line 2600 | public class ForkJoinPool extends Abstra
2600		*         scheduled for execution
2601		*/
2602		public void execute(ForkJoinTask<?> task) {
2603	<	doSubmit(task);
2603	>	if (task == null)
2604	>	throw new NullPointerException();
2605	>	externalPush(task);
2606		}
2607
2608		// AbstractExecutorService methods
#	Line 1348 | Line 2613 | public class ForkJoinPool extends Abstra
2613		*         scheduled for execution
2614		*/
2615		public void execute(Runnable task) {
2616	+	if (task == null)
2617	+	throw new NullPointerException();
2618		ForkJoinTask<?> job;
2619		if (task instanceof ForkJoinTask<?>) // avoid re-wrap
2620		job = (ForkJoinTask<?>) task;
2621		else
2622	<	job = ForkJoinTask.adapt(task, null);
2623	<	doSubmit(job);
2622	>	job = new ForkJoinTask.AdaptedRunnableAction(task);
2623	>	externalPush(job);
2624	>	}
2625	>
2626	>	/**
2627	>	* Submits a ForkJoinTask for execution.
2628	>	*
2629	>	* @param task the task to submit
2630	>	* @return the task
2631	>	* @throws NullPointerException if the task is null
2632	>	* @throws RejectedExecutionException if the task cannot be
2633	>	*         scheduled for execution
2634	>	*/
2635	>	public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) {
2636	>	if (task == null)
2637	>	throw new NullPointerException();
2638	>	externalPush(task);
2639	>	return task;
2640		}
2641
2642		/**
#	Line 1362 | Line 2645 | public class ForkJoinPool extends Abstra
2645		*         scheduled for execution
2646		*/
2647		public <T> ForkJoinTask<T> submit(Callable<T> task) {
2648	<	ForkJoinTask<T> job = ForkJoinTask.adapt(task);
2649	<	doSubmit(job);
2648	>	ForkJoinTask<T> job = new ForkJoinTask.AdaptedCallable<T>(task);
2649	>	externalPush(job);
2650		return job;
2651		}
2652
#	Line 1373 | Line 2656 | public class ForkJoinPool extends Abstra
2656		*         scheduled for execution
2657		*/
2658		public <T> ForkJoinTask<T> submit(Runnable task, T result) {
2659	<	ForkJoinTask<T> job = ForkJoinTask.adapt(task, result);
2660	<	doSubmit(job);
2659	>	ForkJoinTask<T> job = new ForkJoinTask.AdaptedRunnable<T>(task, result);
2660	>	externalPush(job);
2661		return job;
2662		}
2663
#	Line 1384 | Line 2667 | public class ForkJoinPool extends Abstra
2667		*         scheduled for execution
2668		*/
2669		public ForkJoinTask<?> submit(Runnable task) {
2670	+	if (task == null)
2671	+	throw new NullPointerException();
2672		ForkJoinTask<?> job;
2673		if (task instanceof ForkJoinTask<?>) // avoid re-wrap
2674		job = (ForkJoinTask<?>) task;
2675		else
2676	<	job = ForkJoinTask.adapt(task, null);
2677	<	doSubmit(job);
2676	>	job = new ForkJoinTask.AdaptedRunnableAction(task);
2677	>	externalPush(job);
2678		return job;
2679		}
2680
2681		/**
1397	–	* Submits a ForkJoinTask for execution.
1398	–	*
1399	–	* @param task the task to submit
1400	–	* @return the task
1401	–	* @throws NullPointerException if the task is null
1402	–	* @throws RejectedExecutionException if the task cannot be
1403	–	*         scheduled for execution
1404	–	*/
1405	–	public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) {
1406	–	doSubmit(task);
1407	–	return task;
1408	–	}
1409	–
1410	–	/**
2682		* @throws NullPointerException       {@inheritDoc}
2683		* @throws RejectedExecutionException {@inheritDoc}
2684		*/
2685		public <T> List<Future<T>> invokeAll(Collection<? extends Callable<T>> tasks) {
2686	<	ArrayList<ForkJoinTask<T>> forkJoinTasks =
2687	<	new ArrayList<ForkJoinTask<T>>(tasks.size());
2688	<	for (Callable<T> task : tasks)
2689	<	forkJoinTasks.add(ForkJoinTask.adapt(task));
2690	<	invoke(new InvokeAll<T>(forkJoinTasks));
2691	<
2686	>	// In previous versions of this class, this method constructed
2687	>	// a task to run ForkJoinTask.invokeAll, but now external
2688	>	// invocation of multiple tasks is at least as efficient.
2689	>	List<ForkJoinTask<T>> fs = new ArrayList<ForkJoinTask<T>>(tasks.size());
2690	>	// Workaround needed because method wasn't declared with
2691	>	// wildcards in return type but should have been.
2692		@SuppressWarnings({"unchecked", "rawtypes"})
2693	<	List<Future<T>> futures = (List<Future<T>>) (List) forkJoinTasks;
1423	<	return futures;
1424	<	}
2693	>	List<Future<T>> futures = (List<Future<T>>) (List) fs;
2694
2695	<	static final class InvokeAll<T> extends RecursiveAction {
2696	<	final ArrayList<ForkJoinTask<T>> tasks;
2697	<	InvokeAll(ArrayList<ForkJoinTask<T>> tasks) { this.tasks = tasks; }
2698	<	public void compute() {
2699	<	try { invokeAll(tasks); }
2700	<	catch (Exception ignore) {}
2695	>	boolean done = false;
2696	>	try {
2697	>	for (Callable<T> t : tasks) {
2698	>	ForkJoinTask<T> f = new ForkJoinTask.AdaptedCallable<T>(t);
2699	>	externalPush(f);
2700	>	fs.add(f);
2701	>	}
2702	>	for (ForkJoinTask<T> f : fs)
2703	>	f.quietlyJoin();
2704	>	done = true;
2705	>	return futures;
2706	>	} finally {
2707	>	if (!done)
2708	>	for (ForkJoinTask<T> f : fs)
2709	>	f.cancel(false);
2710		}
1433	–	private static final long serialVersionUID = -7914297376763021607L;
2711		}
2712
2713		/**
#	Line 1449 | Line 2726 | public class ForkJoinPool extends Abstra
2726		* @return the handler, or {@code null} if none
2727		*/
2728		public Thread.UncaughtExceptionHandler getUncaughtExceptionHandler() {
1452	–	workerCountReadFence();
2729		return ueh;
2730		}
2731
2732		/**
2733	<	* Sets the handler for internal worker threads that terminate due
1458	<	* to unrecoverable errors encountered while executing tasks.
1459	<	* Unless set, the current default or ThreadGroup handler is used
1460	<	* as handler.
1461	<	*
1462	<	* @param h the new handler
1463	<	* @return the old handler, or {@code null} if none
1464	<	* @throws SecurityException if a security manager exists and
1465	<	*         the caller is not permitted to modify threads
1466	<	*         because it does not hold {@link
1467	<	*         java.lang.RuntimePermission}{@code ("modifyThread")}
1468	<	*/
1469	<	public Thread.UncaughtExceptionHandler
1470	<	setUncaughtExceptionHandler(Thread.UncaughtExceptionHandler h) {
1471	<	checkPermission();
1472	<	Thread.UncaughtExceptionHandler old = ueh;
1473	<	if (h != old) {
1474	<	ueh = h;
1475	<	ForkJoinWorkerThread[] ws = workers;
1476	<	int nws = ws.length;
1477	<	for (int i = 0; i < nws; ++i) {
1478	<	ForkJoinWorkerThread w = ws[i];
1479	<	if (w != null)
1480	<	w.setUncaughtExceptionHandler(h);
1481	<	}
1482	<	}
1483	<	return old;
1484	<	}
1485	<
1486	<	/**
1487	<	* Sets the target parallelism level of this pool.
2733	>	* Returns the targeted parallelism level of this pool.
2734		*
2735	<	* @param parallelism the target parallelism
1490	<	* @throws IllegalArgumentException if parallelism less than or
1491	<	* equal to zero or greater than maximum size bounds
1492	<	* @throws SecurityException if a security manager exists and
1493	<	*         the caller is not permitted to modify threads
1494	<	*         because it does not hold {@link
1495	<	*         java.lang.RuntimePermission}{@code ("modifyThread")}
2735	>	* @return the targeted parallelism level of this pool
2736		*/
2737	<	public void setParallelism(int parallelism) {
2738	<	checkPermission();
1499	<	if (parallelism <= 0 || parallelism > maxPoolSize)
1500	<	throw new IllegalArgumentException();
1501	<	workerCountReadFence();
1502	<	int pc = this.parallelism;
1503	<	if (pc != parallelism) {
1504	<	this.parallelism = parallelism;
1505	<	workerCountWriteFence();
1506	<	// Release spares. If too many, some will die after re-suspend
1507	<	ForkJoinWorkerThread[] ws = workers;
1508	<	int nws = ws.length;
1509	<	for (int i = 0; i < nws; ++i) {
1510	<	ForkJoinWorkerThread w = ws[i];
1511	<	if (w != null && w.tryUnsuspend()) {
1512	<	int c;
1513	<	do {} while (!UNSAFE.compareAndSwapInt
1514	<	(this, workerCountsOffset,
1515	<	c = workerCounts, c + ONE_RUNNING));
1516	<	LockSupport.unpark(w);
1517	<	}
1518	<	}
1519	<	ensureEnoughTotalWorkers();
1520	<	advanceEventCount();
1521	<	releaseWaiters(); // force config recheck by existing workers
1522	<	}
2737	>	public int getParallelism() {
2738	>	return config & SMASK;
2739		}
2740
2741		/**
2742	<	* Returns the targeted parallelism level of this pool.
2742	>	* Returns the targeted parallelism level of the common pool.
2743		*
2744	<	* @return the targeted parallelism level of this pool
2744	>	* @return the targeted parallelism level of the common pool
2745	>	* @since 1.8
2746		*/
2747	<	public int getParallelism() {
2748	<	//        workerCountReadFence(); // inlined below
1532	<	int ignore = workerCounts;
1533	<	return parallelism;
2747	>	public static int getCommonPoolParallelism() {
2748	>	return commonParallelism;
2749		}
2750
2751		/**
2752		* Returns the number of worker threads that have started but not
2753	<	* yet terminated.  This result returned by this method may differ
2753	>	* yet terminated.  The result returned by this method may differ
2754		* from {@link #getParallelism} when threads are created to
2755		* maintain parallelism when others are cooperatively blocked.
2756		*
2757		* @return the number of worker threads
2758		*/
2759		public int getPoolSize() {
2760	<	return workerCounts >>> TOTAL_COUNT_SHIFT;
1546	<	}
1547	<
1548	<	/**
1549	<	* Returns the maximum number of threads allowed to exist in the
1550	<	* pool. Unless set using {@link #setMaximumPoolSize}, the
1551	<	* maximum is an implementation-defined value designed only to
1552	<	* prevent runaway growth.
1553	<	*
1554	<	* @return the maximum
1555	<	*/
1556	<	public int getMaximumPoolSize() {
1557	<	workerCountReadFence();
1558	<	return maxPoolSize;
1559	<	}
1560	<
1561	<	/**
1562	<	* Sets the maximum number of threads allowed to exist in the
1563	<	* pool. The given value should normally be greater than or equal
1564	<	* to the {@link #getParallelism parallelism} level. Setting this
1565	<	* value has no effect on current pool size. It controls
1566	<	* construction of new threads. The use of this method may cause
1567	<	* tasks that intrinsically require extra threads for dependent
1568	<	* computations to indefinitely stall. If you are instead trying
1569	<	* to minimize internal thread creation, consider setting {@link
1570	<	* #setMaintainsParallelism} as false.
1571	<	*
1572	<	* @throws IllegalArgumentException if negative or greater than
1573	<	* internal implementation limit
1574	<	*/
1575	<	public void setMaximumPoolSize(int newMax) {
1576	<	if (newMax < 0 || newMax > MAX_THREADS)
1577	<	throw new IllegalArgumentException();
1578	<	maxPoolSize = newMax;
1579	<	workerCountWriteFence();
1580	<	}
1581	<
1582	<	/**
1583	<	* Returns {@code true} if this pool dynamically maintains its
1584	<	* target parallelism level. If false, new threads are added only
1585	<	* to avoid possible starvation.  This setting is by default true.
1586	<	*
1587	<	* @return {@code true} if maintains parallelism
1588	<	*/
1589	<	public boolean getMaintainsParallelism() {
1590	<	workerCountReadFence();
1591	<	return maintainsParallelism;
1592	<	}
1593	<
1594	<	/**
1595	<	* Sets whether this pool dynamically maintains its target
1596	<	* parallelism level. If false, new threads are added only to
1597	<	* avoid possible starvation.
1598	<	*
1599	<	* @param enable {@code true} to maintain parallelism
1600	<	*/
1601	<	public void setMaintainsParallelism(boolean enable) {
1602	<	maintainsParallelism = enable;
1603	<	workerCountWriteFence();
1604	<	}
1605	<
1606	<	/**
1607	<	* Establishes local first-in-first-out scheduling mode for forked
1608	<	* tasks that are never joined. This mode may be more appropriate
1609	<	* than default locally stack-based mode in applications in which
1610	<	* worker threads only process asynchronous tasks.  This method is
1611	<	* designed to be invoked only when the pool is quiescent, and
1612	<	* typically only before any tasks are submitted. The effects of
1613	<	* invocations at other times may be unpredictable.
1614	<	*
1615	<	* @param async if {@code true}, use locally FIFO scheduling
1616	<	* @return the previous mode
1617	<	* @see #getAsyncMode
1618	<	*/
1619	<	public boolean setAsyncMode(boolean async) {
1620	<	workerCountReadFence();
1621	<	boolean oldMode = locallyFifo;
1622	<	if (oldMode != async) {
1623	<	locallyFifo = async;
1624	<	workerCountWriteFence();
1625	<	ForkJoinWorkerThread[] ws = workers;
1626	<	int nws = ws.length;
1627	<	for (int i = 0; i < nws; ++i) {
1628	<	ForkJoinWorkerThread w = ws[i];
1629	<	if (w != null)
1630	<	w.setAsyncMode(async);
1631	<	}
1632	<	}
1633	<	return oldMode;
2760	>	return (config & SMASK) + (short)(ctl >>> TC_SHIFT);
2761		}
2762
2763		/**
#	Line 1638 | Line 2765 | public class ForkJoinPool extends Abstra
2765		* scheduling mode for forked tasks that are never joined.
2766		*
2767		* @return {@code true} if this pool uses async mode
1641	–	* @see #setAsyncMode
2768		*/
2769		public boolean getAsyncMode() {
2770	<	workerCountReadFence();
1645	<	return locallyFifo;
2770	>	return (config >>> 16) == FIFO_QUEUE;
2771		}
2772
2773		/**
#	Line 1654 | Line 2779 | public class ForkJoinPool extends Abstra
2779		* @return the number of worker threads
2780		*/
2781		public int getRunningThreadCount() {
2782	<	return workerCounts & RUNNING_COUNT_MASK;
2782	>	int rc = 0;
2783	>	WorkQueue[] ws; WorkQueue w;
2784	>	if ((ws = workQueues) != null) {
2785	>	for (int i = 1; i < ws.length; i += 2) {
2786	>	if ((w = ws[i]) != null && w.isApparentlyUnblocked())
2787	>	++rc;
2788	>	}
2789	>	}
2790	>	return rc;
2791		}
2792
2793		/**
#	Line 1665 | Line 2798 | public class ForkJoinPool extends Abstra
2798		* @return the number of active threads
2799		*/
2800		public int getActiveThreadCount() {
2801	<	return runState & ACTIVE_COUNT_MASK;
2801	>	int r = (config & SMASK) + (int)(ctl >> AC_SHIFT);
2802	>	return (r <= 0) ? 0 : r; // suppress momentarily negative values
2803		}
2804
2805		/**
#	Line 1680 | Line 2814 | public class ForkJoinPool extends Abstra
2814		* @return {@code true} if all threads are currently idle
2815		*/
2816		public boolean isQuiescent() {
2817	<	return (runState & ACTIVE_COUNT_MASK) == 0;
2817	>	return (int)(ctl >> AC_SHIFT) + (config & SMASK) == 0;
2818		}
2819
2820		/**
#	Line 1695 | Line 2829 | public class ForkJoinPool extends Abstra
2829		* @return the number of steals
2830		*/
2831		public long getStealCount() {
2832	<	return stealCount;
2832	>	long count = stealCount;
2833	>	WorkQueue[] ws; WorkQueue w;
2834	>	if ((ws = workQueues) != null) {
2835	>	for (int i = 1; i < ws.length; i += 2) {
2836	>	if ((w = ws[i]) != null)
2837	>	count += w.nsteals;
2838	>	}
2839	>	}
2840	>	return count;
2841		}
2842
2843		/**
#	Line 1710 | Line 2852 | public class ForkJoinPool extends Abstra
2852		*/
2853		public long getQueuedTaskCount() {
2854		long count = 0;
2855	<	ForkJoinWorkerThread[] ws = workers;
2856	<	int nws = ws.length;
2857	<	for (int i = 0; i < nws; ++i) {
2858	<	ForkJoinWorkerThread w = ws[i];
2859	<	if (w != null)
2860	<	count += w.getQueueSize();
2855	>	WorkQueue[] ws; WorkQueue w;
2856	>	if ((ws = workQueues) != null) {
2857	>	for (int i = 1; i < ws.length; i += 2) {
2858	>	if ((w = ws[i]) != null)
2859	>	count += w.queueSize();
2860	>	}
2861		}
2862		return count;
2863		}
2864
2865		/**
2866		* Returns an estimate of the number of tasks submitted to this
2867	<	* pool that have not yet begun executing.  This method takes time
2868	<	* proportional to the number of submissions.
2867	>	* pool that have not yet begun executing.  This method may take
2868	>	* time proportional to the number of submissions.
2869		*
2870		* @return the number of queued submissions
2871		*/
2872		public int getQueuedSubmissionCount() {
2873	<	return submissionQueue.size();
2873	>	int count = 0;
2874	>	WorkQueue[] ws; WorkQueue w;
2875	>	if ((ws = workQueues) != null) {
2876	>	for (int i = 0; i < ws.length; i += 2) {
2877	>	if ((w = ws[i]) != null)
2878	>	count += w.queueSize();
2879	>	}
2880	>	}
2881	>	return count;
2882		}
2883
2884		/**
#	Line 1738 | Line 2888 | public class ForkJoinPool extends Abstra
2888		* @return {@code true} if there are any queued submissions
2889		*/
2890		public boolean hasQueuedSubmissions() {
2891	<	return !submissionQueue.isEmpty();
2891	>	WorkQueue[] ws; WorkQueue w;
2892	>	if ((ws = workQueues) != null) {
2893	>	for (int i = 0; i < ws.length; i += 2) {
2894	>	if ((w = ws[i]) != null && !w.isEmpty())
2895	>	return true;
2896	>	}
2897	>	}
2898	>	return false;
2899		}
2900
2901		/**
#	Line 1749 | Line 2906 | public class ForkJoinPool extends Abstra
2906		* @return the next submission, or {@code null} if none
2907		*/
2908		protected ForkJoinTask<?> pollSubmission() {
2909	<	return submissionQueue.poll();
2909	>	WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2910	>	if ((ws = workQueues) != null) {
2911	>	for (int i = 0; i < ws.length; i += 2) {
2912	>	if ((w = ws[i]) != null && (t = w.poll()) != null)
2913	>	return t;
2914	>	}
2915	>	}
2916	>	return null;
2917		}
2918
2919		/**
#	Line 1770 | Line 2934 | public class ForkJoinPool extends Abstra
2934		* @return the number of elements transferred
2935		*/
2936		protected int drainTasksTo(Collection<? super ForkJoinTask<?>> c) {
2937	<	int n = submissionQueue.drainTo(c);
2938	<	ForkJoinWorkerThread[] ws = workers;
2939	<	int nws = ws.length;
2940	<	for (int i = 0; i < nws; ++i) {
2941	<	ForkJoinWorkerThread w = ws[i];
2942	<	if (w != null)
2943	<	n += w.drainTasksTo(c);
2937	>	int count = 0;
2938	>	WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2939	>	if ((ws = workQueues) != null) {
2940	>	for (int i = 0; i < ws.length; ++i) {
2941	>	if ((w = ws[i]) != null) {
2942	>	while ((t = w.poll()) != null) {
2943	>	c.add(t);
2944	>	++count;
2945	>	}
2946	>	}
2947	>	}
2948		}
2949	<	return n;
2949	>	return count;
2950		}
2951
2952		/**
#	Line 1789 | Line 2957 | public class ForkJoinPool extends Abstra
2957		* @return a string identifying this pool, as well as its state
2958		*/
2959		public String toString() {
2960	<	long st = getStealCount();
2961	<	long qt = getQueuedTaskCount();
2962	<	long qs = getQueuedSubmissionCount();
2963	<	int wc = workerCounts;
2964	<	int tc = wc >>> TOTAL_COUNT_SHIFT;
2965	<	int rc = wc & RUNNING_COUNT_MASK;
2966	<	int pc = parallelism;
2967	<	int rs = runState;
2968	<	int ac = rs & ACTIVE_COUNT_MASK;
2960	>	// Use a single pass through workQueues to collect counts
2961	>	long qt = 0L, qs = 0L; int rc = 0;
2962	>	long st = stealCount;
2963	>	long c = ctl;
2964	>	WorkQueue[] ws; WorkQueue w;
2965	>	if ((ws = workQueues) != null) {
2966	>	for (int i = 0; i < ws.length; ++i) {
2967	>	if ((w = ws[i]) != null) {
2968	>	int size = w.queueSize();
2969	>	if ((i & 1) == 0)
2970	>	qs += size;
2971	>	else {
2972	>	qt += size;
2973	>	st += w.nsteals;
2974	>	if (w.isApparentlyUnblocked())
2975	>	++rc;
2976	>	}
2977	>	}
2978	>	}
2979	>	}
2980	>	int pc = (config & SMASK);
2981	>	int tc = pc + (short)(c >>> TC_SHIFT);
2982	>	int ac = pc + (int)(c >> AC_SHIFT);
2983	>	if (ac < 0) // ignore transient negative
2984	>	ac = 0;
2985	>	String level;
2986	>	if ((c & STOP_BIT) != 0)
2987	>	level = (tc == 0) ? "Terminated" : "Terminating";
2988	>	else
2989	>	level = plock < 0 ? "Shutting down" : "Running";
2990		return super.toString() +
2991	<	"[" + runLevelToString(rs) +
2991	>	"[" + level +
2992		", parallelism = " + pc +
2993		", size = " + tc +
2994		", active = " + ac +
#	Line 1810 | Line 2999 | public class ForkJoinPool extends Abstra
2999		"]";
3000		}
3001
1813	–	private static String runLevelToString(int s) {
1814	–	return ((s & TERMINATED) != 0 ? "Terminated" :
1815	–	((s & TERMINATING) != 0 ? "Terminating" :
1816	–	((s & SHUTDOWN) != 0 ? "Shutting down" :
1817	–	"Running")));
1818	–	}
1819	–
3002		/**
3003	<	* Initiates an orderly shutdown in which previously submitted
3004	<	* tasks are executed, but no new tasks will be accepted.
3005	<	* Invocation has no additional effect if already shut down.
3006	<	* Tasks that are in the process of being submitted concurrently
3007	<	* during the course of this method may or may not be rejected.
3003	>	* Possibly initiates an orderly shutdown in which previously
3004	>	* submitted tasks are executed, but no new tasks will be
3005	>	* accepted. Invocation has no effect on execution state if this
3006	>	* is the {@link #commonPool()}, and no additional effect if
3007	>	* already shut down.  Tasks that are in the process of being
3008	>	* submitted concurrently during the course of this method may or
3009	>	* may not be rejected.
3010		*
3011		* @throws SecurityException if a security manager exists and
3012		*         the caller is not permitted to modify threads
#	Line 1831 | Line 3015 | public class ForkJoinPool extends Abstra
3015		*/
3016		public void shutdown() {
3017		checkPermission();
3018	<	advanceRunLevel(SHUTDOWN);
1835	<	tryTerminate(false);
3018	>	tryTerminate(false, true);
3019		}
3020
3021		/**
3022	<	* Attempts to cancel and/or stop all tasks, and reject all
3023	<	* subsequently submitted tasks.  Tasks that are in the process of
3024	<	* being submitted or executed concurrently during the course of
3025	<	* this method may or may not be rejected. This method cancels
3026	<	* both existing and unexecuted tasks, in order to permit
3027	<	* termination in the presence of task dependencies. So the method
3028	<	* always returns an empty list (unlike the case for some other
3029	<	* Executors).
3022	>	* Possibly attempts to cancel and/or stop all tasks, and reject
3023	>	* all subsequently submitted tasks.  Invocation has no effect on
3024	>	* execution state if this is the {@link #commonPool()}, and no
3025	>	* additional effect if already shut down. Otherwise, tasks that
3026	>	* are in the process of being submitted or executed concurrently
3027	>	* during the course of this method may or may not be
3028	>	* rejected. This method cancels both existing and unexecuted
3029	>	* tasks, in order to permit termination in the presence of task
3030	>	* dependencies. So the method always returns an empty list
3031	>	* (unlike the case for some other Executors).
3032		*
3033		* @return an empty list
3034		* @throws SecurityException if a security manager exists and
#	Line 1853 | Line 3038 | public class ForkJoinPool extends Abstra
3038		*/
3039		public List<Runnable> shutdownNow() {
3040		checkPermission();
3041	<	tryTerminate(true);
3041	>	tryTerminate(true, true);
3042		return Collections.emptyList();
3043		}
3044
#	Line 1863 | Line 3048 | public class ForkJoinPool extends Abstra
3048		* @return {@code true} if all tasks have completed following shut down
3049		*/
3050		public boolean isTerminated() {
3051	<	return runState >= TERMINATED;
3051	>	long c = ctl;
3052	>	return ((c & STOP_BIT) != 0L &&
3053	>	(short)(c >>> TC_SHIFT) == -(config & SMASK));
3054		}
3055
3056		/**
#	Line 1871 | Line 3058 | public class ForkJoinPool extends Abstra
3058		* commenced but not yet completed.  This method may be useful for
3059		* debugging. A return of {@code true} reported a sufficient
3060		* period after shutdown may indicate that submitted tasks have
3061	<	* ignored or suppressed interruption, causing this executor not
3062	<	* to properly terminate.
3061	>	* ignored or suppressed interruption, or are waiting for I/O,
3062	>	* causing this executor not to properly terminate. (See the
3063	>	* advisory notes for class {@link ForkJoinTask} stating that
3064	>	* tasks should not normally entail blocking operations.  But if
3065	>	* they do, they must abort them on interrupt.)
3066		*
3067		* @return {@code true} if terminating but not yet terminated
3068		*/
3069		public boolean isTerminating() {
3070	<	return (runState & (TERMINATING|TERMINATED)) == TERMINATING;
3070	>	long c = ctl;
3071	>	return ((c & STOP_BIT) != 0L &&
3072	>	(short)(c >>> TC_SHIFT) != -(config & SMASK));
3073		}
3074
3075		/**
#	Line 1886 | Line 3078 | public class ForkJoinPool extends Abstra
3078		* @return {@code true} if this pool has been shut down
3079		*/
3080		public boolean isShutdown() {
3081	<	return runState >= SHUTDOWN;
3081	>	return plock < 0;
3082		}
3083
3084		/**
3085	<	* Blocks until all tasks have completed execution after a shutdown
3086	<	* request, or the timeout occurs, or the current thread is
3087	<	* interrupted, whichever happens first.
3085	>	* Blocks until all tasks have completed execution after a
3086	>	* shutdown request, or the timeout occurs, or the current thread
3087	>	* is interrupted, whichever happens first. Because the {@link
3088	>	* #commonPool()} never terminates until program shutdown, when
3089	>	* applied to the common pool, this method is equivalent to {@link
3090	>	* #awaitQuiescence} but always returns {@code false}.
3091		*
3092		* @param timeout the maximum time to wait
3093		* @param unit the time unit of the timeout argument
#	Line 1902 | Line 3097 | public class ForkJoinPool extends Abstra
3097		*/
3098		public boolean awaitTermination(long timeout, TimeUnit unit)
3099		throws InterruptedException {
3100	<	return terminationLatch.await(timeout, unit);
3100	>	if (Thread.interrupted())
3101	>	throw new InterruptedException();
3102	>	if (this == common) {
3103	>	awaitQuiescence(timeout, unit);
3104	>	return false;
3105	>	}
3106	>	long nanos = unit.toNanos(timeout);
3107	>	if (isTerminated())
3108	>	return true;
3109	>	long startTime = System.nanoTime();
3110	>	boolean terminated = false;
3111	>	synchronized (this) {
3112	>	for (long waitTime = nanos, millis = 0L;;) {
3113	>	if (terminated = isTerminated() ||
3114	>	waitTime <= 0L ||
3115	>	(millis = unit.toMillis(waitTime)) <= 0L)
3116	>	break;
3117	>	wait(millis);
3118	>	waitTime = nanos - (System.nanoTime() - startTime);
3119	>	}
3120	>	}
3121	>	return terminated;
3122	>	}
3123	>
3124	>	/**
3125	>	* If called by a ForkJoinTask operating in this pool, equivalent
3126	>	* in effect to {@link ForkJoinTask#helpQuiesce}. Otherwise,
3127	>	* waits and/or attempts to assist performing tasks until this
3128	>	* pool {@link #isQuiescent} or the indicated timeout elapses.
3129	>	*
3130	>	* @param timeout the maximum time to wait
3131	>	* @param unit the time unit of the timeout argument
3132	>	* @return {@code true} if quiescent; {@code false} if the
3133	>	* timeout elapsed.
3134	>	*/
3135	>	public boolean awaitQuiescence(long timeout, TimeUnit unit) {
3136	>	long nanos = unit.toNanos(timeout);
3137	>	ForkJoinWorkerThread wt;
3138	>	Thread thread = Thread.currentThread();
3139	>	if ((thread instanceof ForkJoinWorkerThread) &&
3140	>	(wt = (ForkJoinWorkerThread)thread).pool == this) {
3141	>	helpQuiescePool(wt.workQueue);
3142	>	return true;
3143	>	}
3144	>	long startTime = System.nanoTime();
3145	>	WorkQueue[] ws;
3146	>	int r = 0, m;
3147	>	boolean found = true;
3148	>	while (!isQuiescent() && (ws = workQueues) != null &&
3149	>	(m = ws.length - 1) >= 0) {
3150	>	if (!found) {
3151	>	if ((System.nanoTime() - startTime) > nanos)
3152	>	return false;
3153	>	Thread.yield(); // cannot block
3154	>	}
3155	>	found = false;
3156	>	for (int j = (m + 1) << 2; j >= 0; --j) {
3157	>	ForkJoinTask<?> t; WorkQueue q; int b;
3158	>	if ((q = ws[r++ & m]) != null && (b = q.base) - q.top < 0) {
3159	>	found = true;
3160	>	if ((t = q.pollAt(b)) != null) {
3161	>	if (q.base - q.top < 0)
3162	>	signalWork(q);
3163	>	t.doExec();
3164	>	}
3165	>	break;
3166	>	}
3167	>	}
3168	>	}
3169	>	return true;
3170	>	}
3171	>
3172	>	/**
3173	>	* Waits and/or attempts to assist performing tasks indefinitely
3174	>	* until the {@link #commonPool()} {@link #isQuiescent}
3175	>	*/
3176	>	static void quiesceCommonPool() {
3177	>	common.awaitQuiescence(Long.MAX_VALUE, TimeUnit.NANOSECONDS);
3178		}
3179
3180		/**
3181		* Interface for extending managed parallelism for tasks running
3182		* in {@link ForkJoinPool}s.
3183		*
3184	<	* <p>A {@code ManagedBlocker} provides two methods.
3185	<	* Method {@code isReleasable} must return {@code true} if
3186	<	* blocking is not necessary. Method {@code block} blocks the
3187	<	* current thread if necessary (perhaps internally invoking
3188	<	* {@code isReleasable} before actually blocking).
3184	>	* <p>A {@code ManagedBlocker} provides two methods.  Method
3185	>	* {@code isReleasable} must return {@code true} if blocking is
3186	>	* not necessary. Method {@code block} blocks the current thread
3187	>	* if necessary (perhaps internally invoking {@code isReleasable}
3188	>	* before actually blocking). These actions are performed by any
3189	>	* thread invoking {@link ForkJoinPool#managedBlock}.  The
3190	>	* unusual methods in this API accommodate synchronizers that may,
3191	>	* but don't usually, block for long periods. Similarly, they
3192	>	* allow more efficient internal handling of cases in which
3193	>	* additional workers may be, but usually are not, needed to
3194	>	* ensure sufficient parallelism.  Toward this end,
3195	>	* implementations of method {@code isReleasable} must be amenable
3196	>	* to repeated invocation.
3197		*
3198		* <p>For example, here is a ManagedBlocker based on a
3199		* ReentrantLock:
#	Line 1931 | Line 3211 | public class ForkJoinPool extends Abstra
3211		*     return hasLock || (hasLock = lock.tryLock());
3212		*   }
3213		* }}</pre>
3214	+	*
3215	+	* <p>Here is a class that possibly blocks waiting for an
3216	+	* item on a given queue:
3217	+	*  <pre> {@code
3218	+	* class QueueTaker<E> implements ManagedBlocker {
3219	+	*   final BlockingQueue<E> queue;
3220	+	*   volatile E item = null;
3221	+	*   QueueTaker(BlockingQueue<E> q) { this.queue = q; }
3222	+	*   public boolean block() throws InterruptedException {
3223	+	*     if (item == null)
3224	+	*       item = queue.take();
3225	+	*     return true;
3226	+	*   }
3227	+	*   public boolean isReleasable() {
3228	+	*     return item != null || (item = queue.poll()) != null;
3229	+	*   }
3230	+	*   public E getItem() { // call after pool.managedBlock completes
3231	+	*     return item;
3232	+	*   }
3233	+	* }}</pre>
3234		*/
3235		public static interface ManagedBlocker {
3236		/**
#	Line 1954 | Line 3254 | public class ForkJoinPool extends Abstra
3254		* Blocks in accord with the given blocker.  If the current thread
3255		* is a {@link ForkJoinWorkerThread}, this method possibly
3256		* arranges for a spare thread to be activated if necessary to
3257	<	* ensure parallelism while the current thread is blocked.
1958	<	*
1959	<	* <p>If {@code maintainParallelism} is {@code true} and the pool
1960	<	* supports it ({@link #getMaintainsParallelism}), this method
1961	<	* attempts to maintain the pool's nominal parallelism. Otherwise
1962	<	* it activates a thread only if necessary to avoid complete
1963	<	* starvation. This option may be preferable when blockages use
1964	<	* timeouts, or are almost always brief.
3257	>	* ensure sufficient parallelism while the current thread is blocked.
3258		*
3259		* <p>If the caller is not a {@link ForkJoinTask}, this method is
3260		* behaviorally equivalent to
#	Line 1975 | Line 3268 | public class ForkJoinPool extends Abstra
3268		* first be expanded to ensure parallelism, and later adjusted.
3269		*
3270		* @param blocker the blocker
1978	–	* @param maintainParallelism if {@code true} and supported by
1979	–	* this pool, attempt to maintain the pool's nominal parallelism;
1980	–	* otherwise activate a thread only if necessary to avoid
1981	–	* complete starvation.
3271		* @throws InterruptedException if blocker.block did so
3272		*/
3273	<	public static void managedBlock(ManagedBlocker blocker,
1985	<	boolean maintainParallelism)
3273	>	public static void managedBlock(ManagedBlocker blocker)
3274		throws InterruptedException {
3275		Thread t = Thread.currentThread();
3276	<	if (t instanceof ForkJoinWorkerThread)
3277	<	((ForkJoinWorkerThread) t).pool.
3278	<	awaitBlocker(blocker, maintainParallelism);
3279	<	else
3280	<	awaitBlocker(blocker);
3281	<	}
3282	<
3283	<	/**
3284	<	* Performs Non-FJ blocking
3285	<	*/
3286	<	private static void awaitBlocker(ManagedBlocker blocker)
3287	<	throws InterruptedException {
3288	<	do {} while (!blocker.isReleasable() && !blocker.block());
3276	>	if (t instanceof ForkJoinWorkerThread) {
3277	>	ForkJoinPool p = ((ForkJoinWorkerThread)t).pool;
3278	>	while (!blocker.isReleasable()) { // variant of helpSignal
3279	>	WorkQueue[] ws; WorkQueue q; int m, u;
3280	>	if ((ws = p.workQueues) != null && (m = ws.length - 1) >= 0) {
3281	>	for (int i = 0; i <= m; ++i) {
3282	>	if (blocker.isReleasable())
3283	>	return;
3284	>	if ((q = ws[i]) != null && q.base - q.top < 0) {
3285	>	p.signalWork(q);
3286	>	if ((u = (int)(p.ctl >>> 32)) >= 0 ||
3287	>	(u >> UAC_SHIFT) >= 0)
3288	>	break;
3289	>	}
3290	>	}
3291	>	}
3292	>	if (p.tryCompensate()) {
3293	>	try {
3294	>	do {} while (!blocker.isReleasable() &&
3295	>	!blocker.block());
3296	>	} finally {
3297	>	p.incrementActiveCount();
3298	>	}
3299	>	break;
3300	>	}
3301	>	}
3302	>	}
3303	>	else {
3304	>	do {} while (!blocker.isReleasable() &&
3305	>	!blocker.block());
3306	>	}
3307		}
3308
3309		// AbstractExecutorService overrides.  These rely on undocumented
#	Line 2005 | Line 3311 | public class ForkJoinPool extends Abstra
3311		// implement RunnableFuture.
3312
3313		protected <T> RunnableFuture<T> newTaskFor(Runnable runnable, T value) {
3314	<	return (RunnableFuture<T>) ForkJoinTask.adapt(runnable, value);
3314	>	return new ForkJoinTask.AdaptedRunnable<T>(runnable, value);
3315		}
3316
3317		protected <T> RunnableFuture<T> newTaskFor(Callable<T> callable) {
3318	<	return (RunnableFuture<T>) ForkJoinTask.adapt(callable);
3318	>	return new ForkJoinTask.AdaptedCallable<T>(callable);
3319		}
3320
3321		// Unsafe mechanics
3322	+	private static final sun.misc.Unsafe U;
3323	+	private static final long CTL;
3324	+	private static final long PARKBLOCKER;
3325	+	private static final int ABASE;
3326	+	private static final int ASHIFT;
3327	+	private static final long STEALCOUNT;
3328	+	private static final long PLOCK;
3329	+	private static final long INDEXSEED;
3330	+	private static final long QLOCK;
3331
3332	<	private static final sun.misc.Unsafe UNSAFE = getUnsafe();
3333	<	private static final long workerCountsOffset =
3334	<	objectFieldOffset("workerCounts", ForkJoinPool.class);
3335	<	private static final long runStateOffset =
3336	<	objectFieldOffset("runState", ForkJoinPool.class);
3337	<	private static final long eventCountOffset =
3338	<	objectFieldOffset("eventCount", ForkJoinPool.class);
3339	<	private static final long eventWaitersOffset =
3340	<	objectFieldOffset("eventWaiters",ForkJoinPool.class);
3341	<	private static final long stealCountOffset =
3342	<	objectFieldOffset("stealCount",ForkJoinPool.class);
3332	>	static {
3333	>	int s; // initialize field offsets for CAS etc
3334	>	try {
3335	>	U = getUnsafe();
3336	>	Class<?> k = ForkJoinPool.class;
3337	>	CTL = U.objectFieldOffset
3338	>	(k.getDeclaredField("ctl"));
3339	>	STEALCOUNT = U.objectFieldOffset
3340	>	(k.getDeclaredField("stealCount"));
3341	>	PLOCK = U.objectFieldOffset
3342	>	(k.getDeclaredField("plock"));
3343	>	INDEXSEED = U.objectFieldOffset
3344	>	(k.getDeclaredField("indexSeed"));
3345	>	Class<?> tk = Thread.class;
3346	>	PARKBLOCKER = U.objectFieldOffset
3347	>	(tk.getDeclaredField("parkBlocker"));
3348	>	Class<?> wk = WorkQueue.class;
3349	>	QLOCK = U.objectFieldOffset
3350	>	(wk.getDeclaredField("qlock"));
3351	>	Class<?> ak = ForkJoinTask[].class;
3352	>	ABASE = U.arrayBaseOffset(ak);
3353	>	s = U.arrayIndexScale(ak);
3354	>	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
3355	>	} catch (Exception e) {
3356	>	throw new Error(e);
3357	>	}
3358	>	if ((s & (s-1)) != 0)
3359	>	throw new Error("data type scale not a power of two");
3360	>
3361	>	submitters = new ThreadLocal<Submitter>();
3362	>	ForkJoinWorkerThreadFactory fac = defaultForkJoinWorkerThreadFactory =
3363	>	new DefaultForkJoinWorkerThreadFactory();
3364	>	modifyThreadPermission = new RuntimePermission("modifyThread");
3365	>
3366	>	/*
3367	>	* Establish common pool parameters.  For extra caution,
3368	>	* computations to set up common pool state are here; the
3369	>	* constructor just assigns these values to fields.
3370	>	*/
3371
3372	+	int par = 0;
3373	+	Thread.UncaughtExceptionHandler handler = null;
3374	+	try {  // TBD: limit or report ignored exceptions?
3375	+	String pp = System.getProperty
3376	+	("java.util.concurrent.ForkJoinPool.common.parallelism");
3377	+	String hp = System.getProperty
3378	+	("java.util.concurrent.ForkJoinPool.common.exceptionHandler");
3379	+	String fp = System.getProperty
3380	+	("java.util.concurrent.ForkJoinPool.common.threadFactory");
3381	+	if (fp != null)
3382	+	fac = ((ForkJoinWorkerThreadFactory)ClassLoader.
3383	+	getSystemClassLoader().loadClass(fp).newInstance());
3384	+	if (hp != null)
3385	+	handler = ((Thread.UncaughtExceptionHandler)ClassLoader.
3386	+	getSystemClassLoader().loadClass(hp).newInstance());
3387	+	if (pp != null)
3388	+	par = Integer.parseInt(pp);
3389	+	} catch (Exception ignore) {
3390	+	}
3391	+
3392	+	if (par <= 0)
3393	+	par = Runtime.getRuntime().availableProcessors();
3394	+	if (par > MAX_CAP)
3395	+	par = MAX_CAP;
3396	+	commonParallelism = par;
3397	+	long np = (long)(-par); // precompute initial ctl value
3398	+	long ct = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK);
3399
3400	<	private static long objectFieldOffset(String field, Class<?> klazz) {
2031	<	try {
2032	<	return UNSAFE.objectFieldOffset(klazz.getDeclaredField(field));
2033	<	} catch (NoSuchFieldException e) {
2034	<	// Convert Exception to corresponding Error
2035	<	NoSuchFieldError error = new NoSuchFieldError(field);
2036	<	error.initCause(e);
2037	<	throw error;
2038	<	}
3400	>	common = new ForkJoinPool(par, ct, fac, handler);
3401		}
3402
3403		/**
#	Line 2048 | Line 3410 | public class ForkJoinPool extends Abstra
3410		private static sun.misc.Unsafe getUnsafe() {
3411		try {
3412		return sun.misc.Unsafe.getUnsafe();
3413	<	} catch (SecurityException se) {
3414	<	try {
3415	<	return java.security.AccessController.doPrivileged
3416	<	(new java.security
3417	<	.PrivilegedExceptionAction<sun.misc.Unsafe>() {
3418	<	public sun.misc.Unsafe run() throws Exception {
3419	<	java.lang.reflect.Field f = sun.misc
3420	<	.Unsafe.class.getDeclaredField("theUnsafe");
3421	<	f.setAccessible(true);
3422	<	return (sun.misc.Unsafe) f.get(null);
3423	<	}});
3424	<	} catch (java.security.PrivilegedActionException e) {
3425	<	throw new RuntimeException("Could not initialize intrinsics",
3426	<	e.getCause());
3427	<	}
3413	>	} catch (SecurityException tryReflectionInstead) {}
3414	>	try {
3415	>	return java.security.AccessController.doPrivileged
3416	>	(new java.security.PrivilegedExceptionAction<sun.misc.Unsafe>() {
3417	>	public sun.misc.Unsafe run() throws Exception {
3418	>	Class<sun.misc.Unsafe> k = sun.misc.Unsafe.class;
3419	>	for (java.lang.reflect.Field f : k.getDeclaredFields()) {
3420	>	f.setAccessible(true);
3421	>	Object x = f.get(null);
3422	>	if (k.isInstance(x))
3423	>	return k.cast(x);
3424	>	}
3425	>	throw new NoSuchFieldError("the Unsafe");
3426	>	}});
3427	>	} catch (java.security.PrivilegedActionException e) {
3428	>	throw new RuntimeException("Could not initialize intrinsics",
3429	>	e.getCause());
3430		}
3431		}
3432		}

Diff Legend

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