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 1.57 by dl, Wed Jul 7 19:52:31 2010 UTC vs.
Revision 1.158 by dl, Sat Dec 15 20:21:30 2012 UTC

#	Line 1 | Line 1
1		/*
2		* Written by Doug Lea with assistance from members of JCP JSR-166
3		* Expert Group and released to the public domain, as explained at
4	<	* http://creativecommons.org/licenses/publicdomain
4	>	* http://creativecommons.org/publicdomain/zero/1.0/
5		*/
6
7		package jsr166y;
8
9	–	import java.util.concurrent.*;
10	–
9		import java.util.ArrayList;
10		import java.util.Arrays;
11		import java.util.Collection;
12		import java.util.Collections;
13		import java.util.List;
14	<	import java.util.concurrent.locks.LockSupport;
15	<	import java.util.concurrent.locks.ReentrantLock;
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.
#	Line 27 | Line 28 | import java.util.concurrent.CountDownLat
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). When setting <em>asyncMode</em> to true in
35	<	* constructors, {@code ForkJoinPool}s may also be appropriate for use
36	<	* with event-style tasks that are never joined.
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. The pool attempts to maintain enough active (or
50	<	* available) threads by dynamically adding, suspending, or resuming
51	<	* internal worker threads, even if some tasks are stalled waiting to
52	<	* join others. However, no such adjustments are guaranteed in the
53	<	* face of blocked IO or other unmanaged synchronization. The nested
54	<	* {@link ManagedBlocker} interface enables extension of the kinds of
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
#	Line 51 | 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> As is the case with other ExecutorServices, there are three
66	<	* main task execution methods summarized in the follwoing
67	<	* table. These are designed to be used by clients not already engaged
68	<	* in fork/join computations in the current pool.  The main forms of
69	<	* these methods accept instances of {@code ForkJoinTask}, but
70	<	* overloaded forms also allow mixed execution of plain {@code
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
73	<	* <em>NOT</em> use these pool execution methods, but instead use the
74	<	* within-computation forms listed in the table. To avoid inadvertant
75	<	* cyclic task dependencies and to improve performance, task
65	<	* submissions to the current pool by an ongoing fork/join
66	<	* computations may be implicitly translated to the corresponding
67	<	* ForkJoinTask forms.
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>
#	Line 73 | Line 81 | import java.util.concurrent.CountDownLat
81		*    <td ALIGN=CENTER> <b>Call from within fork/join computations</b></td>
82		*  </tr>
83		*  <tr>
84	<	*    <td> <b>Arange async execution</td>
84	>	*    <td> <b>Arrange async execution</td>
85		*    <td> {@link #execute(ForkJoinTask)}</td>
86		*    <td> {@link ForkJoinTask#fork}</td>
87		*  </tr>
#	Line 88 | Line 96 | import java.util.concurrent.CountDownLat
96		*    <td> {@link ForkJoinTask#fork} (ForkJoinTasks <em>are</em> Futures)</td>
97		*  </tr>
98		* </table>
91	–	*
92	–	* <p><b>Sample Usage.</b> Normally a single {@code ForkJoinPool} is
93	–	* used for all parallel task execution in a program or subsystem.
94	–	* Otherwise, use would not usually outweigh the construction and
95	–	* bookkeeping overhead of creating a large set of threads. For
96	–	* example, a common pool could be used for the {@code SortTasks}
97	–	* illustrated in {@link RecursiveAction}. Because {@code
98	–	* ForkJoinPool} uses threads in {@linkplain java.lang.Thread#isDaemon
99	–	* daemon} mode, there is typically no need to explicitly {@link
100	–	* #shutdown} such a pool upon program exit.
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 113 | 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 123 | 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	<	* aim to approximately maintain the given level of parallelism.
251	<	* When some workers are known to be blocked (on joins or via
252	<	* ManagedBlocker), we may create or resume others to take their
253	<	* place until they unblock (see below). Implementing this
254	<	* requires counts of the number of "running" threads (i.e., those
255	<	* that are neither blocked nor artifically suspended) as well as
256	<	* the total number.  These two values are packed into one field,
257	<	* "workerCounts" because we need accurate snapshots when deciding
258	<	* to create, resume or suspend.  To support these decisions,
259	<	* updates to spare counts must be prospective (not
260	<	* retrospective).  For example, the running count is decremented
261	<	* before blocking by a thread about to block as a spare, but
262	<	* incremented by the thread about to unblock it. Updates upon
263	<	* resumption ofr threads blocking in awaitJoin or awaitBlocker
264	<	* cannot usually be prospective, so the running count is in
265	<	* general an upper bound of the number of productively running
266	<	* threads Updates to the workerCounts field sometimes transiently
267	<	* encounter a fair amount of contention when join dependencies
268	<	* are such that many threads block or unblock at about the same
269	<	* time. We alleviate this by sometimes performing an alternative
270	<	* action on contention like releasing waiters or locating spares.
271	<	*
272	<	* 3. Maintaining global run state. The run state of the pool
273	<	* consists of a runLevel (SHUTDOWN, TERMINATING, etc) similar to
274	<	* those in other Executor implementations, as well as a count of
275	<	* "active" workers -- those that are, or soon will be, or
276	<	* recently were executing tasks. The runLevel and active count
277	<	* are packed together in order to correctly trigger shutdown and
278	<	* termination. Without care, active counts can be subject to very
279	<	* high contention.  We substantially reduce this contention by
280	<	* relaxing update rules.  A worker must claim active status
281	<	* prospectively, by activating if it sees that a submitted or
282	<	* stealable task exists (it may find after activating that the
283	<	* task no longer exists). It stays active while processing this
284	<	* task (if it exists) and any other local subtasks it produces,
285	<	* until it cannot find any other tasks. It then tries
286	<	* inactivating (see method preStep), but upon update contention
287	<	* instead scans for more tasks, later retrying inactivation if it
288	<	* doesn't find any.
289	<	*
290	<	* 4. Managing idle workers waiting for tasks. We cannot let
291	<	* workers spin indefinitely scanning for tasks when none are
292	<	* available. On the other hand, we must quickly prod them into
293	<	* action when new tasks are submitted or generated.  We
294	<	* park/unpark these idle workers using an event-count scheme.
295	<	* Field eventCount is incremented upon events that may enable
296	<	* workers that previously could not find a task to now find one:
297	<	* Submission of a new task to the pool, or another worker pushing
298	<	* a task onto a previously empty queue.  (We also use this
299	<	* mechanism for termination and reconfiguration actions that
300	<	* require wakeups of idle workers).  Each worker maintains its
301	<	* last known event count, and blocks when a scan for work did not
302	<	* find a task AND its lastEventCount matches the current
303	<	* eventCount. Waiting idle workers are recorded in a variant of
304	<	* Treiber stack headed by field eventWaiters which, when nonzero,
305	<	* encodes the thread index and count awaited for by the worker
306	<	* thread most recently calling eventSync. This thread in turn has
307	<	* a record (field nextEventWaiter) for the next waiting worker.
308	<	* In addition to allowing simpler decisions about need for
309	<	* wakeup, the event count bits in eventWaiters serve the role of
310	<	* tags to avoid ABA errors in Treiber stacks.  To reduce delays
311	<	* in task diffusion, workers not otherwise occupied may invoke
312	<	* method releaseWaiters, that removes and signals (unparks)
313	<	* workers not waiting on current count. To minimize task
314	<	* production stalls associate with signalling, any worker pushing
315	<	* a task on an empty queue invokes the weaker method signalWork,
316	<	* that only releases idle workers until it detects interference
317	<	* by other threads trying to release, and lets them take
318	<	* over. The net effect is a tree-like diffusion of signals, where
319	<	* released threads (and possibly others) help with unparks.  To
320	<	* further reduce contention effects a bit, failed CASes to
321	<	* increment field eventCount are tolerated without retries.
322	<	* Conceptually they are merged into the same event, which is OK
323	<	* when their only purpose is to enable workers to scan for work.
324	<	*
325	<	* 5. Managing suspension of extra workers. When a worker is about
326	<	* to block waiting for a join (or via ManagedBlockers), we may
327	<	* create a new thread to maintain parallelism level, or at least
328	<	* avoid starvation (see below). Usually, extra threads are needed
329	<	* for only very short periods, yet join dependencies are such
330	<	* that we sometimes need them in bursts. Rather than create new
331	<	* threads each time this happens, we suspend no-longer-needed
332	<	* extra ones as "spares". For most purposes, we don't distinguish
333	<	* "extra" spare threads from normal "core" threads: On each call
334	<	* to preStep (the only point at which we can do this) a worker
335	<	* checks to see if there are now too many running workers, and if
336	<	* so, suspends itself.  Methods awaitJoin and awaitBlocker look
337	<	* for suspended threads to resume before considering creating a
338	<	* new replacement. We don't need a special data structure to
339	<	* maintain spares; simply scanning the workers array looking for
340	<	* worker.isSuspended() is fine because the calling thread is
341	<	* otherwise not doing anything useful anyway; we are at least as
342	<	* happy if after locating a spare, the caller doesn't actually
343	<	* block because the join is ready before we try to adjust and
344	<	* compensate.  Note that this is intrinsically racy.  One thread
345	<	* may become a spare at about the same time as another is
346	<	* needlessly being created. We counteract this and related slop
347	<	* in part by requiring resumed spares to immediately recheck (in
348	<	* preStep) to see whether they they should re-suspend. The only
349	<	* effective difference between "extra" and "core" threads is that
350	<	* we allow the "extra" ones to time out and die if they are not
351	<	* resumed within a keep-alive interval of a few seconds. This is
352	<	* implemented mainly within ForkJoinWorkerThread, but requires
353	<	* some coordination (isTrimmed() -- meaning killed while
354	<	* suspended) to correctly maintain pool counts.
355	<	*
356	<	* 6. Deciding when to create new workers. The main dynamic
357	<	* control in this class is deciding when to create extra threads,
358	<	* in methods awaitJoin and awaitBlocker. We always need to create
359	<	* one when the number of running threads becomes zero. But
360	<	* because blocked joins are typically dependent, we don't
361	<	* necessarily need or want one-to-one replacement. Instead, we
362	<	* use a combination of heuristics that adds threads only when the
363	<	* pool appears to be approaching starvation.  These effectively
364	<	* reduce churn at the price of systematically undershooting
365	<	* target parallelism when many threads are blocked.  However,
366	<	* biasing toward undeshooting partially compensates for the above
367	<	* mechanics to suspend extra threads, that normally lead to
368	<	* overshoot because we can only suspend workers in-between
369	<	* top-level actions. It also better copes with the fact that some
370	<	* of the methods in this class tend to never become compiled (but
371	<	* are interpreted), so some components of the entire set of
372	<	* controls might execute many times faster than others. And
373	<	* similarly for cases where the apparent lack of work is just due
374	<	* to GC stalls and other transient system activity.
375	<	*
376	<	* Beware that there is a lot of representation-level coupling
377	<	* among classes ForkJoinPool, ForkJoinWorkerThread, and
378	<	* ForkJoinTask.  For example, direct access to "workers" array by
379	<	* workers, and direct access to ForkJoinTask.status by both
380	<	* 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 commonPool 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 341 | 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	+	* 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	+	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	+	* 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	+	// Heuristic padding to ameliorate unfortunate memory placements
650	+	volatile long pad00, pad01, pad02, pad03, pad04, pad05, pad06;
651	+
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	+	* 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	+	* 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	+	* 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	+	* 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	+	* 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	+	* 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	+	* 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	+	* 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	+	/**
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	+	* 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	+	* 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	+	/**
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	+	// Specialized execution methods
878	+
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	+	* 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	+	if (base - top < 0) {       // process remaining local tasks
991	+	if (mode == 0)
992	+	popAndExecAll();
993	+	else
994	+	pollAndExecAll();
995	+	}
996	+	++nsteals;
997	+	hint = -1;
998	+	}
999	+	}
1000	+
1001	+	/**
1002	+	* Executes a non-top-level (stolen) task.
1003	+	*/
1004	+	final void runSubtask(ForkJoinTask<?> t) {
1005	+	if (t != null) {
1006	+	ForkJoinTask<?> ps = currentSteal;
1007	+	(currentSteal = t).doExec();
1008	+	currentSteal = ps;
1009	+	}
1010	+	}
1011	+
1012	+	/**
1013	+	* Returns true if owned and not known to be blocked.
1014	+	*/
1015	+	final boolean isApparentlyUnblocked() {
1016	+	Thread wt; Thread.State s;
1017	+	return (eventCount >= 0 &&
1018	+	(wt = owner) != null &&
1019	+	(s = wt.getState()) != Thread.State.BLOCKED &&
1020	+	s != Thread.State.WAITING &&
1021	+	s != Thread.State.TIMED_WAITING);
1022	+	}
1023	+
1024	+	/**
1025	+	* If this owned and is not already interrupted, try to
1026	+	* interrupt and/or unpark, ignoring exceptions.
1027	+	*/
1028	+	final void interruptOwner() {
1029	+	Thread wt, p;
1030	+	if ((wt = owner) != null && !wt.isInterrupted()) {
1031	+	try {
1032	+	wt.interrupt();
1033	+	} catch (SecurityException ignore) {
1034	+	}
1035	+	}
1036	+	if ((p = parker) != null)
1037	+	U.unpark(p);
1038	+	}
1039	+
1040	+	// Unsafe mechanics
1041	+	private static final sun.misc.Unsafe U;
1042	+	private static final long QLOCK;
1043	+	private static final int ABASE;
1044	+	private static final int ASHIFT;
1045	+	static {
1046	+	int s;
1047	+	try {
1048	+	U = getUnsafe();
1049	+	Class<?> k = WorkQueue.class;
1050	+	Class<?> ak = ForkJoinTask[].class;
1051	+	QLOCK = U.objectFieldOffset
1052	+	(k.getDeclaredField("qlock"));
1053	+	ABASE = U.arrayBaseOffset(ak);
1054	+	s = U.arrayIndexScale(ak);
1055	+	} catch (Exception e) {
1056	+	throw new Error(e);
1057	+	}
1058	+	if ((s & (s-1)) != 0)
1059	+	throw new Error("data type scale not a power of two");
1060	+	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
1061	+	}
1062	+	}
1063	+
1064	+	// static fields (initialized in static initializer below)
1065	+
1066	+	/**
1067		* Creates a new ForkJoinWorkerThread. This factory is used unless
1068		* overridden in ForkJoinPool constructors.
1069		*/
1070		public static final ForkJoinWorkerThreadFactory
1071	<	defaultForkJoinWorkerThreadFactory =
1072	<	new DefaultForkJoinWorkerThreadFactory();
1071	>	defaultForkJoinWorkerThreadFactory;
1072	>
1073	>	/**
1074	>	* Per-thread submission bookkeeping. Shared across all pools
1075	>	* to reduce ThreadLocal pollution and because random motion
1076	>	* to avoid contention in one pool is likely to hold for others.
1077	>	* Lazily initialized on first submission (but null-checked
1078	>	* in other contexts to avoid unnecessary initialization).
1079	>	*/
1080	>	static final ThreadLocal<Submitter> submitters;
1081
1082		/**
1083		* Permission required for callers of methods that may start or
1084		* kill threads.
1085		*/
1086	<	private static final RuntimePermission modifyThreadPermission =
364	<	new RuntimePermission("modifyThread");
1086	>	private static final RuntimePermission modifyThreadPermission;
1087
1088		/**
1089	<	* If there is a security manager, makes sure caller has
1090	<	* permission to modify threads.
1089	>	* Common (static) pool. Non-null for public use unless a static
1090	>	* construction exception, but internal usages null-check on use
1091	>	* to paranoically avoid potential initialization circularities
1092	>	* as well as to simplify generated code.
1093		*/
1094	<	private static void checkPermission() {
371	<	SecurityManager security = System.getSecurityManager();
372	<	if (security != null)
373	<	security.checkPermission(modifyThreadPermission);
374	<	}
1094	>	static final ForkJoinPool commonPool;
1095
1096		/**
1097	<	* Generator for assigning sequence numbers as pool names.
1097	>	* Common pool parallelism. Must equal commonPool.parallelism.
1098		*/
1099	<	private static final AtomicInteger poolNumberGenerator =
380	<	new AtomicInteger();
1099	>	static final int commonPoolParallelism;
1100
1101		/**
1102	<	* Absolute bound for parallelism level. Twice this number must
384	<	* fit into a 16bit field to enable word-packing for some counts.
1102	>	* Sequence number for creating workerNamePrefix.
1103		*/
1104	<	private static final int MAX_THREADS = 0x7fff;
1104	>	private static int poolNumberSequence;
1105
1106		/**
1107	<	* Array holding all worker threads in the pool.  Array size must
1108	<	* be a power of two.  Updates and replacements are protected by
391	<	* workerLock, but the array is always kept in a consistent enough
392	<	* state to be randomly accessed without locking by workers
393	<	* performing work-stealing, as well as other traversal-based
394	<	* methods in this class. All readers must tolerate that some
395	<	* array slots may be null.
1107	>	* Return the next sequence number. We don't expect this to
1108	>	* ever contend so use simple builtin sync.
1109		*/
1110	<	volatile ForkJoinWorkerThread[] workers;
1110	>	private static final synchronized int nextPoolId() {
1111	>	return ++poolNumberSequence;
1112	>	}
1113	>
1114	>	// static constants
1115
1116		/**
1117	<	* Queue for external submissions.
1117	>	* Initial timeout value (in nanoseconds) for the thread
1118	>	* triggering quiescence to park waiting for new work. On timeout,
1119	>	* the thread will instead try to shrink the number of
1120	>	* workers. The value should be large enough to avoid overly
1121	>	* aggressive shrinkage during most transient stalls (long GCs
1122	>	* etc).
1123		*/
1124	<	private final LinkedTransferQueue<ForkJoinTask<?>> submissionQueue;
1124	>	private static final long IDLE_TIMEOUT      = 2000L * 1000L * 1000L; // 2sec
1125
1126		/**
1127	<	* Lock protecting updates to workers array.
1127	>	* Timeout value when there are more threads than parallelism level
1128		*/
1129	<	private final ReentrantLock workerLock;
1129	>	private static final long FAST_IDLE_TIMEOUT =  200L * 1000L * 1000L;
1130
1131		/**
1132	<	* Latch released upon termination.
1132	>	* Tolerance for idle timeouts, to cope with timer undershoots
1133		*/
1134	<	private final Phaser termination;
1134	>	private static final long TIMEOUT_SLOP = 2000000L; // 20ms
1135
1136		/**
1137	<	* Creation factory for worker threads.
1137	>	* The maximum stolen->joining link depth allowed in method
1138	>	* tryHelpStealer.  Must be a power of two.  Depths for legitimate
1139	>	* chains are unbounded, but we use a fixed constant to avoid
1140	>	* (otherwise unchecked) cycles and to bound staleness of
1141	>	* traversal parameters at the expense of sometimes blocking when
1142	>	* we could be helping.
1143		*/
1144	<	private final ForkJoinWorkerThreadFactory factory;
1144	>	private static final int MAX_HELP = 64;
1145
1146		/**
1147	<	* Sum of per-thread steal counts, updated only when threads are
1148	<	* idle or terminating.
1147	>	* Increment for seed generators. See class ThreadLocal for
1148	>	* explanation.
1149		*/
1150	<	private volatile long stealCount;
1150	>	private static final int SEED_INCREMENT = 0x61c88647;
1151
1152		/**
1153	<	* Encoded record of top of treiber stack of threads waiting for
1154	<	* events. The top 32 bits contain the count being waited for. The
1155	<	* bottom word contains one plus the pool index of waiting worker
1156	<	* thread.
1153	>	* Bits and masks for control variables
1154	>	*
1155	>	* Field ctl is a long packed with:
1156	>	* AC: Number of active running workers minus target parallelism (16 bits)
1157	>	* TC: Number of total workers minus target parallelism (16 bits)
1158	>	* ST: true if pool is terminating (1 bit)
1159	>	* EC: the wait count of top waiting thread (15 bits)
1160	>	* ID: poolIndex of top of Treiber stack of waiters (16 bits)
1161	>	*
1162	>	* When convenient, we can extract the upper 32 bits of counts and
1163	>	* the lower 32 bits of queue state, u = (int)(ctl >>> 32) and e =
1164	>	* (int)ctl.  The ec field is never accessed alone, but always
1165	>	* together with id and st. The offsets of counts by the target
1166	>	* parallelism and the positionings of fields makes it possible to
1167	>	* perform the most common checks via sign tests of fields: When
1168	>	* ac is negative, there are not enough active workers, when tc is
1169	>	* negative, there are not enough total workers, and when e is
1170	>	* negative, the pool is terminating.  To deal with these possibly
1171	>	* negative fields, we use casts in and out of "short" and/or
1172	>	* signed shifts to maintain signedness.
1173	>	*
1174	>	* When a thread is queued (inactivated), its eventCount field is
1175	>	* set negative, which is the only way to tell if a worker is
1176	>	* prevented from executing tasks, even though it must continue to
1177	>	* scan for them to avoid queuing races. Note however that
1178	>	* eventCount updates lag releases so usage requires care.
1179	>	*
1180	>	* Field plock is an int packed with:
1181	>	* SHUTDOWN: true if shutdown is enabled (1 bit)
1182	>	* SEQ:  a sequence lock, with PL_LOCK bit set if locked (30 bits)
1183	>	* SIGNAL: set when threads may be waiting on the lock (1 bit)
1184	>	*
1185	>	* The sequence number enables simple consistency checks:
1186	>	* Staleness of read-only operations on the workQueues array can
1187	>	* be checked by comparing plock before vs after the reads.
1188		*/
431	–	private volatile long eventWaiters;
1189
1190	<	private static final int  EVENT_COUNT_SHIFT = 32;
1191	<	private static final long WAITER_INDEX_MASK = (1L << EVENT_COUNT_SHIFT)-1L;
1190	>	// bit positions/shifts for fields
1191	>	private static final int  AC_SHIFT   = 48;
1192	>	private static final int  TC_SHIFT   = 32;
1193	>	private static final int  ST_SHIFT   = 31;
1194	>	private static final int  EC_SHIFT   = 16;
1195
1196	<	/**
1197	<	* A counter for events that may wake up worker threads:
1198	<	*   - Submission of a new task to the pool
1199	<	*   - A worker pushing a task on an empty queue
1200	<	*   - termination and reconfiguration
1201	<	*/
1202	<	private volatile int eventCount;
1196	>	// bounds
1197	>	private static final int  SMASK      = 0xffff;  // short bits
1198	>	private static final int  MAX_CAP    = 0x7fff;  // max #workers - 1
1199	>	private static final int  EVENMASK   = 0xfffe;  // even short bits
1200	>	private static final int  SQMASK     = 0x007e;  // max 64 (even) slots
1201	>	private static final int  SHORT_SIGN = 1 << 15;
1202	>	private static final int  INT_SIGN   = 1 << 31;
1203
1204	<	/**
1205	<	* Lifecycle control. The low word contains the number of workers
1206	<	* that are (probably) executing tasks. This value is atomically
1207	<	* incremented before a worker gets a task to run, and decremented
448	<	* when worker has no tasks and cannot find any.  Bits 16-18
449	<	* contain runLevel value. When all are zero, the pool is
450	<	* running. Level transitions are monotonic (running -> shutdown
451	<	* -> terminating -> terminated) so each transition adds a bit.
452	<	* These are bundled together to ensure consistent read for
453	<	* termination checks (i.e., that runLevel is at least SHUTDOWN
454	<	* and active threads is zero).
455	<	*/
456	<	private volatile int runState;
1204	>	// masks
1205	>	private static final long STOP_BIT   = 0x0001L << ST_SHIFT;
1206	>	private static final long AC_MASK    = ((long)SMASK) << AC_SHIFT;
1207	>	private static final long TC_MASK    = ((long)SMASK) << TC_SHIFT;
1208
1209	<	// Note: The order among run level values matters.
1210	<	private static final int RUNLEVEL_SHIFT     = 16;
1211	<	private static final int SHUTDOWN           = 1 << RUNLEVEL_SHIFT;
461	<	private static final int TERMINATING        = 1 << (RUNLEVEL_SHIFT + 1);
462	<	private static final int TERMINATED         = 1 << (RUNLEVEL_SHIFT + 2);
463	<	private static final int ACTIVE_COUNT_MASK  = (1 << RUNLEVEL_SHIFT) - 1;
464	<	private static final int ONE_ACTIVE         = 1; // active update delta
1209	>	// units for incrementing and decrementing
1210	>	private static final long TC_UNIT    = 1L << TC_SHIFT;
1211	>	private static final long AC_UNIT    = 1L << AC_SHIFT;
1212
1213	<	/**
1214	<	* Holds number of total (i.e., created and not yet terminated)
1215	<	* and running (i.e., not blocked on joins or other managed sync)
1216	<	* threads, packed together to ensure consistent snapshot when
1217	<	* making decisions about creating and suspending spare
1218	<	* threads. Updated only by CAS. Note that adding a new worker
1219	<	* requires incrementing both counts, since workers start off in
473	<	* running state.  This field is also used for memory-fencing
474	<	* configuration parameters.
475	<	*/
476	<	private volatile int workerCounts;
1213	>	// masks and units for dealing with u = (int)(ctl >>> 32)
1214	>	private static final int  UAC_SHIFT  = AC_SHIFT - 32;
1215	>	private static final int  UTC_SHIFT  = TC_SHIFT - 32;
1216	>	private static final int  UAC_MASK   = SMASK << UAC_SHIFT;
1217	>	private static final int  UTC_MASK   = SMASK << UTC_SHIFT;
1218	>	private static final int  UAC_UNIT   = 1 << UAC_SHIFT;
1219	>	private static final int  UTC_UNIT   = 1 << UTC_SHIFT;
1220
1221	<	private static final int TOTAL_COUNT_SHIFT  = 16;
1222	<	private static final int RUNNING_COUNT_MASK = (1 << TOTAL_COUNT_SHIFT) - 1;
1223	<	private static final int ONE_RUNNING        = 1;
481	<	private static final int ONE_TOTAL          = 1 << TOTAL_COUNT_SHIFT;
1221	>	// masks and units for dealing with e = (int)ctl
1222	>	private static final int E_MASK      = 0x7fffffff; // no STOP_BIT
1223	>	private static final int E_SEQ       = 1 << EC_SHIFT;
1224
1225	<	/**
1226	<	* The target parallelism level.
1227	<	* Accessed directly by ForkJoinWorkerThreads.
1228	<	*/
1229	<	final int parallelism;
1225	>	// plock bits
1226	>	private static final int SHUTDOWN    = 1 << 31;
1227	>	private static final int PL_LOCK     = 2;
1228	>	private static final int PL_SIGNAL   = 1;
1229	>	private static final int PL_SPINS    = 1 << 8;
1230
1231	<	/**
1232	<	* True if use local fifo, not default lifo, for local polling
1233	<	* Read by, and replicated by ForkJoinWorkerThreads
1234	<	*/
493	<	final boolean locallyFifo;
1231	>	// access mode for WorkQueue
1232	>	static final int LIFO_QUEUE          =  0;
1233	>	static final int FIFO_QUEUE          =  1;
1234	>	static final int SHARED_QUEUE        = -1;
1235
1236	<	/**
1237	<	* The uncaught exception handler used when any worker abruptly
1238	<	* terminates.
498	<	*/
499	<	private final Thread.UncaughtExceptionHandler ueh;
1236	>	// bounds for #steps in scan loop -- must be power 2 minus 1
1237	>	private static final int MIN_SCAN    = 0x1ff;   // cover estimation slop
1238	>	private static final int MAX_SCAN    = 0x1ffff; // 4 * max workers
1239
1240	<	/**
502	<	* Pool number, just for assigning useful names to worker threads
503	<	*/
504	<	private final int poolNumber;
1240	>	// Instance fields
1241
1242	<	// utilities for updating fields
1242	>	/*
1243	>	* Field layout of this class tends to matter more than one would
1244	>	* like. Runtime layout order is only loosely related to
1245	>	* declaration order and may differ across JVMs, but the following
1246	>	* empirically works OK on current JVMs.
1247	>	*/
1248	>
1249	>	// Heuristic padding to ameliorate unfortunate memory placements
1250	>	volatile long pad00, pad01, pad02, pad03, pad04, pad05, pad06;
1251	>
1252	>	volatile long stealCount;                  // collects worker counts
1253	>	volatile long ctl;                         // main pool control
1254	>	volatile int plock;                        // shutdown status and seqLock
1255	>	volatile int indexSeed;                    // worker/submitter index seed
1256	>	final int config;                          // mode and parallelism level
1257	>	WorkQueue[] workQueues;                    // main registry
1258	>	final ForkJoinWorkerThreadFactory factory;
1259	>	final Thread.UncaughtExceptionHandler ueh; // per-worker UEH
1260	>	final String workerNamePrefix;             // to create worker name string
1261
1262	<	/**
1263	<	* Increments running count.  Also used by ForkJoinTask.
1264	<	*/
1265	<	final void incrementRunningCount() {
1266	<	int c;
1267	<	do {} while (!UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1268	<	c = workerCounts,
1269	<	c + ONE_RUNNING));
1270	<	}
1271	<
1272	<	/**
1273	<	* Tries to decrement running count unless already zero
1262	>	volatile Object pad10, pad11, pad12, pad13, pad14, pad15, pad16, pad17;
1263	>	volatile Object pad18, pad19, pad1a, pad1b;
1264	>
1265	>	/*
1266	>	* Acquires the plock lock to protect worker array and related
1267	>	* updates. This method is called only if an initial CAS on plock
1268	>	* fails. This acts as a spinLock for normal cases, but falls back
1269	>	* to builtin monitor to block when (rarely) needed. This would be
1270	>	* a terrible idea for a highly contended lock, but works fine as
1271	>	* a more conservative alternative to a pure spinlock.  See
1272	>	* internal ConcurrentHashMap documentation for further
1273	>	* explanation of nearly the same construction.
1274		*/
1275	<	final boolean tryDecrementRunningCount() {
1276	<	int wc = workerCounts;
1277	<	if ((wc & RUNNING_COUNT_MASK) == 0)
1278	<	return false;
1279	<	return UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1280	<	wc, wc - ONE_RUNNING);
1275	>	private int acquirePlock() {
1276	>	int spins = PL_SPINS, r = 0, ps, nps;
1277	>	for (;;) {
1278	>	if (((ps = plock) & PL_LOCK) == 0 &&
1279	>	U.compareAndSwapInt(this, PLOCK, ps, nps = ps + PL_LOCK))
1280	>	return nps;
1281	>	else if (r == 0) { // randomize spins if possible
1282	>	Thread t = Thread.currentThread(); WorkQueue w; Submitter z;
1283	>	if ((t instanceof ForkJoinWorkerThread) &&
1284	>	(w = ((ForkJoinWorkerThread)t).workQueue) != null)
1285	>	r = w.seed;
1286	>	else if ((z = submitters.get()) != null)
1287	>	r = z.seed;
1288	>	else
1289	>	r = 1;
1290	>	}
1291	>	else if (spins >= 0) {
1292	>	r ^= r << 1; r ^= r >>> 3; r ^= r << 10; // xorshift
1293	>	if (r >= 0)
1294	>	--spins;
1295	>	}
1296	>	else if (U.compareAndSwapInt(this, PLOCK, ps, ps | PL_SIGNAL)) {
1297	>	synchronized (this) {
1298	>	if ((plock & PL_SIGNAL) != 0) {
1299	>	try {
1300	>	wait();
1301	>	} catch (InterruptedException ie) {
1302	>	try {
1303	>	Thread.currentThread().interrupt();
1304	>	} catch (SecurityException ignore) {
1305	>	}
1306	>	}
1307	>	}
1308	>	else
1309	>	notifyAll();
1310	>	}
1311	>	}
1312	>	}
1313		}
1314
1315		/**
1316	<	* Tries incrementing active count; fails on contention.
1317	<	* Called by workers before executing tasks.
532	<	*
533	<	* @return true on success
1316	>	* Unlocks and signals any thread waiting for plock. Called only
1317	>	* when CAS of seq value for unlock fails.
1318		*/
1319	<	final boolean tryIncrementActiveCount() {
1320	<	int c;
1321	<	return UNSAFE.compareAndSwapInt(this, runStateOffset,
538	<	c = runState, c + ONE_ACTIVE);
1319	>	private void releasePlock(int ps) {
1320	>	plock = ps;
1321	>	synchronized (this) { notifyAll(); }
1322		}
1323
1324		/**
1325	<	* Tries decrementing active count; fails on contention.
1326	<	* Called when workers cannot find tasks to run.
1327	<	*/
1328	<	final boolean tryDecrementActiveCount() {
1329	<	int c;
1330	<	return UNSAFE.compareAndSwapInt(this, runStateOffset,
1331	<	c = runState, c - ONE_ACTIVE);
1325	>	* Performs secondary initialization, called when plock is zero.
1326	>	* Creates workQueue array and sets plock to a valid value.  The
1327	>	* lock body must be exception-free (so no try/finally) so we
1328	>	* optimistically allocate new array outside the lock and throw
1329	>	* away if (very rarely) not needed. (A similar tactic is used in
1330	>	* fullExternalPush.)  Because the plock seq value can eventually
1331	>	* wrap around zero, this method harmlessly fails to reinitialize
1332	>	* if workQueues exists, while still advancing plock.
1333	>	*
1334	>	* Additionally tries to create the first worker.
1335	>	*/
1336	>	private void initWorkers() {
1337	>	WorkQueue[] ws, nws; int ps;
1338	>	int p = config & SMASK;        // find power of two table size
1339	>	int n = (p > 1) ? p - 1 : 1;   // ensure at least 2 slots
1340	>	n |= n >>> 1; n |= n >>> 2; n |= n >>> 4; n |= n >>> 8; n |= n >>> 16;
1341	>	n = (n + 1) << 1;
1342	>	if ((ws = workQueues) == null || ws.length == 0)
1343	>	nws = new WorkQueue[n];
1344	>	else
1345	>	nws = null;
1346	>	if (((ps = plock) & PL_LOCK) != 0 ||
1347	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1348	>	ps = acquirePlock();
1349	>	if (((ws = workQueues) == null || ws.length == 0) && nws != null)
1350	>	workQueues = nws;
1351	>	int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1352	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1353	>	releasePlock(nps);
1354	>	tryAddWorker();
1355		}
1356
1357		/**
1358	<	* Advances to at least the given level. Returns true if not
1359	<	* already in at least the given level.
1360	<	*/
1361	<	private boolean advanceRunLevel(int level) {
1362	<	for (;;) {
1363	<	int s = runState;
1364	<	if ((s & level) != 0)
1365	<	return false;
1366	<	if (UNSAFE.compareAndSwapInt(this, runStateOffset, s, s | level))
1367	<	return true;
1358	>	* Tries to create and start one worker if fewer than target
1359	>	* parallelism level exist. Adjusts counts etc on failure.
1360	>	*/
1361	>	private void tryAddWorker() {
1362	>	long c; int u;
1363	>	while ((u = (int)((c = ctl) >>> 32)) < 0 &&
1364	>	(u & SHORT_SIGN) != 0 && (int)c == 0) {
1365	>	long nc = (long)(((u + UTC_UNIT) & UTC_MASK) |
1366	>	((u + UAC_UNIT) & UAC_MASK)) << 32;
1367	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1368	>	ForkJoinWorkerThreadFactory fac;
1369	>	Throwable ex = null;
1370	>	ForkJoinWorkerThread wt = null;
1371	>	try {
1372	>	if ((fac = factory) != null &&
1373	>	(wt = fac.newThread(this)) != null) {
1374	>	wt.start();
1375	>	break;
1376	>	}
1377	>	} catch (Throwable e) {
1378	>	ex = e;
1379	>	}
1380	>	deregisterWorker(wt, ex);
1381	>	break;
1382	>	}
1383		}
1384		}
1385
1386	<	// workers array maintenance
1386	>	//  Registering and deregistering workers
1387
1388		/**
1389	<	* Records and returns a workers array index for new worker.
1390	<	*/
1391	<	private int recordWorker(ForkJoinWorkerThread w) {
1392	<	// Try using slot totalCount-1. If not available, scan and/or resize
1393	<	int k = (workerCounts >>> TOTAL_COUNT_SHIFT) - 1;
1394	<	final ReentrantLock lock = this.workerLock;
1395	<	lock.lock();
1389	>	* Callback from ForkJoinWorkerThread to establish and record its
1390	>	* WorkQueue. To avoid scanning bias due to packing entries in
1391	>	* front of the workQueues array, we treat the array as a simple
1392	>	* power-of-two hash table using per-thread seed as hash,
1393	>	* expanding as needed.
1394	>	*
1395	>	* @param wt the worker thread
1396	>	* @return the worker's queue
1397	>	*/
1398	>	final WorkQueue registerWorker(ForkJoinWorkerThread wt) {
1399	>	Thread.UncaughtExceptionHandler handler; WorkQueue[] ws; int s, ps;
1400	>	wt.setDaemon(true);
1401	>	if ((handler = ueh) != null)
1402	>	wt.setUncaughtExceptionHandler(handler);
1403	>	do {} while (!U.compareAndSwapInt(this, INDEXSEED, s = indexSeed,
1404	>	s += SEED_INCREMENT) ||
1405	>	s == 0); // skip 0
1406	>	WorkQueue w = new WorkQueue(this, wt, config >>> 16, s);
1407	>	if (((ps = plock) & PL_LOCK) != 0 ||
1408	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1409	>	ps = acquirePlock();
1410	>	int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1411		try {
1412	<	ForkJoinWorkerThread[] ws = workers;
1413	<	int nws = ws.length;
1414	<	if (k < 0 || k >= nws || ws[k] != null) {
1415	<	for (k = 0; k < nws && ws[k] != null; ++k)
1416	<	;
1417	<	if (k == nws)
1418	<	ws = Arrays.copyOf(ws, nws << 1);
1412	>	if ((ws = workQueues) != null) {    // skip if shutting down
1413	>	int n = ws.length, m = n - 1;
1414	>	int r = (s << 1) | 1;           // use odd-numbered indices
1415	>	if (ws[r &= m] != null) {       // collision
1416	>	int probes = 0;             // step by approx half size
1417	>	int step = (n <= 4) ? 2 : ((n >>> 1) & EVENMASK) + 2;
1418	>	while (ws[r = (r + step) & m] != null) {
1419	>	if (++probes >= n) {
1420	>	workQueues = ws = Arrays.copyOf(ws, n <<= 1);
1421	>	m = n - 1;
1422	>	probes = 0;
1423	>	}
1424	>	}
1425	>	}
1426	>	w.eventCount = w.poolIndex = r; // volatile write orders
1427	>	ws[r] = w;
1428		}
584	–	ws[k] = w;
585	–	workers = ws; // volatile array write ensures slot visibility
1429		} finally {
1430	<	lock.unlock();
1430	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1431	>	releasePlock(nps);
1432		}
1433	<	return k;
1433	>	wt.setName(workerNamePrefix.concat(Integer.toString(w.poolIndex)));
1434	>	return w;
1435		}
1436
1437		/**
1438	<	* Nulls out record of worker in workers array
1439	<	*/
1440	<	private void forgetWorker(ForkJoinWorkerThread w) {
1441	<	int idx = w.poolIndex;
1442	<	// Locking helps method recordWorker avoid unecessary expansion
1443	<	final ReentrantLock lock = this.workerLock;
1444	<	lock.lock();
1445	<	try {
1446	<	ForkJoinWorkerThread[] ws = workers;
1447	<	if (idx >= 0 && idx < ws.length && ws[idx] == w) // verify
1448	<	ws[idx] = null;
1449	<	} finally {
1450	<	lock.unlock();
1438	>	* Final callback from terminating worker, as well as upon failure
1439	>	* to construct or start a worker.  Removes record of worker from
1440	>	* array, and adjusts counts. If pool is shutting down, tries to
1441	>	* complete termination.
1442	>	*
1443	>	* @param wt the worker thread or null if construction failed
1444	>	* @param ex the exception causing failure, or null if none
1445	>	*/
1446	>	final void deregisterWorker(ForkJoinWorkerThread wt, Throwable ex) {
1447	>	WorkQueue w = null;
1448	>	if (wt != null && (w = wt.workQueue) != null) {
1449	>	int ps;
1450	>	w.qlock = -1;                // ensure set
1451	>	long ns = w.nsteals, sc;     // collect steal count
1452	>	do {} while (!U.compareAndSwapLong(this, STEALCOUNT,
1453	>	sc = stealCount, sc + ns));
1454	>	if (((ps = plock) & PL_LOCK) != 0 ||
1455	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1456	>	ps = acquirePlock();
1457	>	int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1458	>	try {
1459	>	int idx = w.poolIndex;
1460	>	WorkQueue[] ws = workQueues;
1461	>	if (ws != null && idx >= 0 && idx < ws.length && ws[idx] == w)
1462	>	ws[idx] = null;
1463	>	} finally {
1464	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1465	>	releasePlock(nps);
1466	>	}
1467		}
607	–	}
608	–
609	–	// adding and removing workers
1468
1469	<	/**
1470	<	* Tries to create and add new worker. Assumes that worker counts
1471	<	* are already updated to accommodate the worker, so adjusts on
1472	<	* failure.
1473	<	*
1474	<	* @return new worker or null if creation failed
1475	<	*/
1476	<	private ForkJoinWorkerThread addWorker() {
1477	<	ForkJoinWorkerThread w = null;
1478	<	try {
1479	<	w = factory.newThread(this);
1480	<	} finally { // Adjust on either null or exceptional factory return
1481	<	if (w == null) {
1482	<	onWorkerCreationFailure();
1483	<	return null;
1469	>	long c;                             // adjust ctl counts
1470	>	do {} while (!U.compareAndSwapLong
1471	>	(this, CTL, c = ctl, (((c - AC_UNIT) & AC_MASK) |
1472	>	((c - TC_UNIT) & TC_MASK) |
1473	>	(c & ~(AC_MASK|TC_MASK)))));
1474	>
1475	>	if (!tryTerminate(false, false) && w != null && w.array != null) {
1476	>	w.cancelAll();                  // cancel remaining tasks
1477	>	int e, u, i, n; WorkQueue[] ws; WorkQueue v; Thread p;
1478	>	while ((u = (int)((c = ctl) >>> 32)) < 0) {
1479	>	if ((e = (int)c) > 0) {     // activate or create replacement
1480	>	if ((ws = workQueues) != null &&
1481	>	ws.length > (i = e & SMASK) &&
1482	>	(v = ws[i]) != null && v.eventCount == (e | INT_SIGN)) {
1483	>	long nc = (((long)(v.nextWait & E_MASK)) |
1484	>	((long)(u + UAC_UNIT) << 32));
1485	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1486	>	v.eventCount = (e + E_SEQ) & E_MASK;
1487	>	if ((p = v.parker) != null)
1488	>	U.unpark(p);
1489	>	break;
1490	>	}
1491	>	}
1492	>	else
1493	>	break;
1494	>	}
1495	>	else {
1496	>	if ((short)u < 0)
1497	>	tryAddWorker();
1498	>	break;
1499	>	}
1500		}
1501		}
1502	<	w.start(recordWorker(w), ueh);
1503	<	return w;
1502	>	if (ex == null)                     // help clean refs on way out
1503	>	ForkJoinTask.helpExpungeStaleExceptions();
1504	>	else                                // rethrow
1505	>	ForkJoinTask.rethrow(ex);
1506	>	}
1507	>
1508	>	// Submissions
1509	>
1510	>	/**
1511	>	* Unless shutting down, adds the given task to a submission queue
1512	>	* at submitter's current queue index (modulo submission
1513	>	* range). Only the most common path is directly handled in this
1514	>	* method. All others are relayed to fullExternalPush.
1515	>	*
1516	>	* @param task the task. Caller must ensure non-null.
1517	>	*/
1518	>	final void externalPush(ForkJoinTask<?> task) {
1519	>	WorkQueue[] ws; WorkQueue q; Submitter z; int m; ForkJoinTask<?>[] a;
1520	>	if ((z = submitters.get()) != null && plock > 0 &&
1521	>	(ws = workQueues) != null && (m = (ws.length - 1)) >= 0 &&
1522	>	(q = ws[m & z.seed & SQMASK]) != null &&
1523	>	U.compareAndSwapInt(q, QLOCK, 0, 1)) { // lock
1524	>	int b = q.base, s = q.top, n, an;
1525	>	if ((a = q.array) != null && (an = a.length) > (n = s + 1 - b)) {
1526	>	int j = (((an - 1) & s) << ASHIFT) + ABASE;
1527	>	U.putOrderedObject(a, j, task);
1528	>	q.top = s + 1;                     // push on to deque
1529	>	q.qlock = 0;
1530	>	if (n <= 2)
1531	>	signalWork(q);
1532	>	return;
1533	>	}
1534	>	q.qlock = 0;
1535	>	}
1536	>	fullExternalPush(task);
1537		}
1538
1539		/**
1540	<	* Adjusts counts upon failure to create worker
1541	<	*/
1542	<	private void onWorkerCreationFailure() {
1543	<	for (;;) {
1544	<	int wc = workerCounts;
1545	<	if ((wc >>> TOTAL_COUNT_SHIFT) > 0 &&
1546	<	UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1547	<	wc, wc - (ONE_RUNNING|ONE_TOTAL)))
1548	<	break;
1540	>	* Full version of externalPush. This method is called, among
1541	>	* other times, upon the first submission of the first task to the
1542	>	* pool, so must perform secondary initialization (via
1543	>	* initWorkers). It also detects first submission by an external
1544	>	* thread by looking up its ThreadLocal, and creates a new shared
1545	>	* queue if the one at index if empty or contended. The plock lock
1546	>	* body must be exception-free (so no try/finally) so we
1547	>	* optimistically allocate new queues outside the lock and throw
1548	>	* them away if (very rarely) not needed.
1549	>	*/
1550	>	private void fullExternalPush(ForkJoinTask<?> task) {
1551	>	int r = 0; // random index seed
1552	>	for (Submitter z = submitters.get();;) {
1553	>	WorkQueue[] ws; WorkQueue q; int ps, m, k;
1554	>	if (z == null) {
1555	>	if (U.compareAndSwapInt(this, INDEXSEED, r = indexSeed,
1556	>	r += SEED_INCREMENT) && r != 0)
1557	>	submitters.set(z = new Submitter(r));
1558	>	}
1559	>	else if (r == 0) {               // move to a different index
1560	>	r = z.seed;
1561	>	r ^= r << 13;                // same xorshift as WorkQueues
1562	>	r ^= r >>> 17;
1563	>	z.seed = r ^ (r << 5);
1564	>	}
1565	>	else if ((ps = plock) < 0)
1566	>	throw new RejectedExecutionException();
1567	>	else if (ps == 0 || (ws = workQueues) == null ||
1568	>	(m = ws.length - 1) < 0)
1569	>	initWorkers();
1570	>	else if ((q = ws[k = r & m & SQMASK]) != null) {
1571	>	if (q.qlock == 0 && U.compareAndSwapInt(q, QLOCK, 0, 1)) {
1572	>	ForkJoinTask<?>[] a = q.array;
1573	>	int s = q.top;
1574	>	boolean submitted = false;
1575	>	try {                      // locked version of push
1576	>	if ((a != null && a.length > s + 1 - q.base) ||
1577	>	(a = q.growArray()) != null) {   // must presize
1578	>	int j = (((a.length - 1) & s) << ASHIFT) + ABASE;
1579	>	U.putOrderedObject(a, j, task);
1580	>	q.top = s + 1;
1581	>	submitted = true;
1582	>	}
1583	>	} finally {
1584	>	q.qlock = 0;  // unlock
1585	>	}
1586	>	if (submitted) {
1587	>	signalWork(q);
1588	>	return;
1589	>	}
1590	>	}
1591	>	r = 0; // move on failure
1592	>	}
1593	>	else if (((ps = plock) & PL_LOCK) == 0) { // create new queue
1594	>	q = new WorkQueue(this, null, SHARED_QUEUE, r);
1595	>	if (((ps = plock) & PL_LOCK) != 0 ||
1596	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1597	>	ps = acquirePlock();
1598	>	if ((ws = workQueues) != null && k < ws.length && ws[k] == null)
1599	>	ws[k] = q;
1600	>	int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1601	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1602	>	releasePlock(nps);
1603	>	}
1604	>	else
1605	>	r = 0; // try elsewhere while lock held
1606		}
643	–	tryTerminate(false); // in case of failure during shutdown
1607		}
1608
1609	+	// Maintaining ctl counts
1610	+
1611		/**
1612	<	* Create enough total workers to establish target parallelism,
648	<	* giving up if terminating or addWorker fails
1612	>	* Increments active count; mainly called upon return from blocking.
1613		*/
1614	<	private void ensureEnoughTotalWorkers() {
1615	<	int wc;
1616	<	while (((wc = workerCounts) >>> TOTAL_COUNT_SHIFT) < parallelism &&
653	<	runState < TERMINATING) {
654	<	if ((UNSAFE.compareAndSwapInt(this, workerCountsOffset,
655	<	wc, wc + (ONE_RUNNING|ONE_TOTAL)) &&
656	<	addWorker() == null))
657	<	break;
658	<	}
1614	>	final void incrementActiveCount() {
1615	>	long c;
1616	>	do {} while (!U.compareAndSwapLong(this, CTL, c = ctl, c + AC_UNIT));
1617		}
1618
1619		/**
1620	<	* Final callback from terminating worker.  Removes record of
1621	<	* worker from array, and adjusts counts. If pool is shutting
1622	<	* down, tries to complete terminatation, else possibly replaces
1623	<	* the worker.
1624	<	*
1625	<	* @param w the worker
1626	<	*/
1627	<	final void workerTerminated(ForkJoinWorkerThread w) {
1628	<	if (w.active) { // force inactive
1629	<	w.active = false;
1630	<	do {} while (!tryDecrementActiveCount());
1631	<	}
1632	<	forgetWorker(w);
1633	<
1634	<	// Decrement total count, and if was running, running count
1635	<	// Spin (waiting for other updates) if either would be negative
1636	<	int nr = w.isTrimmed() ? 0 : ONE_RUNNING;
1637	<	int unit = ONE_TOTAL + nr;
1638	<	for (;;) {
1639	<	int wc = workerCounts;
1640	<	int rc = wc & RUNNING_COUNT_MASK;
1641	<	if (rc - nr < 0 || (wc >>> TOTAL_COUNT_SHIFT) == 0)
1642	<	Thread.yield(); // back off if waiting for other updates
1643	<	else if (UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1644	<	wc, wc - unit))
1620	>	* Tries to create or activate a worker if too few are active.
1621	>	*
1622	>	* @param q the (non-null) queue holding tasks to be signalled
1623	>	*/
1624	>	final void signalWork(WorkQueue q) {
1625	>	int hint = q.poolIndex;
1626	>	long c; int e, u, i, n; WorkQueue[] ws; WorkQueue w; Thread p;
1627	>	while ((u = (int)((c = ctl) >>> 32)) < 0) {
1628	>	if ((e = (int)c) > 0) {
1629	>	if ((ws = workQueues) != null && ws.length > (i = e & SMASK) &&
1630	>	(w = ws[i]) != null && w.eventCount == (e | INT_SIGN)) {
1631	>	long nc = (((long)(w.nextWait & E_MASK)) |
1632	>	((long)(u + UAC_UNIT) << 32));
1633	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1634	>	w.hint = hint;
1635	>	w.eventCount = (e + E_SEQ) & E_MASK;
1636	>	if ((p = w.parker) != null)
1637	>	U.unpark(p);
1638	>	break;
1639	>	}
1640	>	if (q.top - q.base <= 0)
1641	>	break;
1642	>	}
1643	>	else
1644	>	break;
1645	>	}
1646	>	else {
1647	>	if ((short)u < 0)
1648	>	tryAddWorker();
1649		break;
1650	+	}
1651		}
689	–
690	–	accumulateStealCount(w); // collect final count
691	–	if (!tryTerminate(false))
692	–	ensureEnoughTotalWorkers();
1652		}
1653
1654	<	// Waiting for and signalling events
696	<
697	<	/**
698	<	* Releases workers blocked on a count not equal to current count.
699	<	*/
700	<	private void releaseWaiters() {
701	<	long top;
702	<	int id;
703	<	while ((id = (int)((top = eventWaiters) & WAITER_INDEX_MASK)) > 0 &&
704	<	(int)(top >>> EVENT_COUNT_SHIFT) != eventCount) {
705	<	ForkJoinWorkerThread[] ws = workers;
706	<	ForkJoinWorkerThread w;
707	<	if (ws.length >= id && (w = ws[id - 1]) != null &&
708	<	UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
709	<	top, w.nextWaiter))
710	<	LockSupport.unpark(w);
711	<	}
712	<	}
1654	>	// Scanning for tasks
1655
1656		/**
1657	<	* Ensures eventCount on exit is different (mod 2^32) than on
716	<	* entry and wakes up all waiters
1657	>	* Top-level runloop for workers, called by ForkJoinWorkerThread.run.
1658		*/
1659	<	private void signalEvent() {
1660	<	int c;
1661	<	do {} while (!UNSAFE.compareAndSwapInt(this, eventCountOffset,
1662	<	c = eventCount, c+1));
1663	<	releaseWaiters();
1664	<	}
1659	>	final void runWorker(WorkQueue w) {
1660	>	w.growArray(); // allocate queue
1661	>	do { w.runTask(scan(w)); } while (w.qlock >= 0);
1662	>	}
1663	>
1664	>	/**
1665	>	* Scans for and, if found, returns one task, else possibly
1666	>	* inactivates the worker. This method operates on single reads of
1667	>	* volatile state and is designed to be re-invoked continuously,
1668	>	* in part because it returns upon detecting inconsistencies,
1669	>	* contention, or state changes that indicate possible success on
1670	>	* re-invocation.
1671	>	*
1672	>	* The scan searches for tasks across queues (starting at a random
1673	>	* index, and relying on registerWorker to irregularly scatter
1674	>	* them within array to avoid bias), checking each at least twice.
1675	>	* The scan terminates upon either finding a non-empty queue, or
1676	>	* completing the sweep. If the worker is not inactivated, it
1677	>	* takes and returns a task from this queue. Otherwise, if not
1678	>	* activated, it signals workers (that may include itself) and
1679	>	* returns so caller can retry. Also returns for true if the
1680	>	* worker array may have changed during an empty scan.  On failure
1681	>	* to find a task, we take one of the following actions, after
1682	>	* which the caller will retry calling this method unless
1683	>	* terminated.
1684	>	*
1685	>	* * If pool is terminating, terminate the worker.
1686	>	*
1687	>	* * If not already enqueued, try to inactivate and enqueue the
1688	>	* worker on wait queue. Or, if inactivating has caused the pool
1689	>	* to be quiescent, relay to idleAwaitWork to check for
1690	>	* termination and possibly shrink pool.
1691	>	*
1692	>	* * If already enqueued and none of the above apply, possibly
1693	>	* (with 1/2 probability) park awaiting signal, else lingering to
1694	>	* help scan and signal.
1695	>	*
1696	>	* @param w the worker (via its WorkQueue)
1697	>	* @return a task or null if none found
1698	>	*/
1699	>	private final ForkJoinTask<?> scan(WorkQueue w) {
1700	>	WorkQueue[] ws; int m;
1701	>	int ps = plock;                          // read plock before ws
1702	>	if (w != null && (ws = workQueues) != null && (m = ws.length - 1) >= 0) {
1703	>	int ec = w.eventCount;               // ec is negative if inactive
1704	>	int r = w.seed; r ^= r << 13; r ^= r >>> 17; w.seed = r ^= r << 5;
1705	>	int j = ((m + m + 1) | MIN_SCAN) & MAX_SCAN;
1706	>	do {
1707	>	WorkQueue q; ForkJoinTask<?>[] a; int b;
1708	>	if ((q = ws[(r + j) & m]) != null && (b = q.base) - q.top < 0 &&
1709	>	(a = q.array) != null) {     // probably nonempty
1710	>	int i = (((a.length - 1) & b) << ASHIFT) + ABASE;
1711	>	ForkJoinTask<?> t = (ForkJoinTask<?>)
1712	>	U.getObjectVolatile(a, i);
1713	>	if (q.base == b && ec >= 0 && t != null &&
1714	>	U.compareAndSwapObject(a, i, t, null)) {
1715	>	if ((q.base = b + 1) - q.top < 0)
1716	>	signalWork(q);
1717	>	return t;                // taken
1718	>	}
1719	>	else if ((ec < 0 || j < m) && (int)(ctl >> AC_SHIFT) <= 0) {
1720	>	w.hint = (r + j) & m;    // help signal below
1721	>	break;                   // cannot take
1722	>	}
1723	>	}
1724	>	} while (--j >= 0);
1725
1726	<	/**
1727	<	* Advances eventCount and releases waiters until interference by
1728	<	* other releasing threads is detected.
1729	<	*/
1730	<	final void signalWork() {
1731	<	// EventCount CAS failures are OK -- any change in count suffices.
1732	<	int ec;
1733	<	UNSAFE.compareAndSwapInt(this, eventCountOffset, ec=eventCount, ec+1);
1734	<	outer:for (;;) {
1735	<	long top = eventWaiters;
1736	<	ec = eventCount;
1737	<	for (;;) {
1738	<	ForkJoinWorkerThread[] ws; ForkJoinWorkerThread w;
1739	<	int id = (int)(top & WAITER_INDEX_MASK);
1740	<	if (id <= 0 || (int)(top >>> EVENT_COUNT_SHIFT) == ec)
1741	<	return;
1742	<	if ((ws = workers).length < id || (w = ws[id - 1]) == null ||
1743	<	!UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
1744	<	top, top = w.nextWaiter))
1745	<	continue outer;      // possibly stale; reread
1746	<	LockSupport.unpark(w);
1747	<	if (top != eventWaiters) // let someone else take over
1748	<	return;
1726	>	long c, sc; int e, ns, h;
1727	>	if ((h = w.hint) < 0) {
1728	>	if ((ns = w.nsteals) != 0) {
1729	>	if (U.compareAndSwapLong(this, STEALCOUNT,
1730	>	sc = stealCount, sc + ns))
1731	>	w.nsteals = 0;           // collect steals
1732	>	}
1733	>	else if (plock != ps)            // consistency check
1734	>	;                            // skip
1735	>	else if ((e = (int)(c = ctl)) < 0)
1736	>	w.qlock = -1;                // pool is terminating
1737	>	else if (ec >= 0) {              // try to enqueue/inactivate
1738	>	long nc = ((long)ec | ((c - AC_UNIT) & (AC_MASK|TC_MASK)));
1739	>	w.nextWait = e;              // link and mark inactive
1740	>	w.eventCount = ec | INT_SIGN;
1741	>	if (ctl != c || !U.compareAndSwapLong(this, CTL, c, nc))
1742	>	w.eventCount = ec;       // unmark on CAS failure
1743	>	else if ((int)(c >> AC_SHIFT) == 1 - (config & SMASK))
1744	>	idleAwaitWork(w, nc, c);
1745	>	}
1746	>	else if (w.eventCount < 0) {     // block
1747	>	Thread wt = Thread.currentThread();
1748	>	Thread.interrupted();        // clear status
1749	>	U.putObject(wt, PARKBLOCKER, this);
1750	>	w.parker = wt;               // emulate LockSupport.park
1751	>	if (w.eventCount < 0)        // recheck
1752	>	U.park(false, 0L);
1753	>	w.parker = null;
1754	>	U.putObject(wt, PARKBLOCKER, null);
1755	>	}
1756		}
1757	+	if (h >= 0 || w.hint >= 0)           // signal others before retry
1758	+	helpSignalHint(w);
1759		}
1760	+	return null;
1761		}
1762
1763		/**
1764	<	* If worker is inactive, blocks until terminating or event count
1765	<	* advances from last value held by worker; in any case helps
1766	<	* release others.
1767	<	*
1768	<	* @param w the calling worker thread
1769	<	*/
1770	<	private void eventSync(ForkJoinWorkerThread w) {
1771	<	if (!w.active) {
1772	<	int prev = w.lastEventCount;
1773	<	long nextTop = (((long)prev << EVENT_COUNT_SHIFT) |
1774	<	((long)(w.poolIndex + 1)));
1775	<	long top;
1776	<	while ((runState < SHUTDOWN || !tryTerminate(false)) &&
1777	<	(((int)(top = eventWaiters) & WAITER_INDEX_MASK) == 0 ||
1778	<	(int)(top >>> EVENT_COUNT_SHIFT) == prev) &&
1779	<	eventCount == prev) {
1780	<	if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
1781	<	w.nextWaiter = top, nextTop)) {
1782	<	accumulateStealCount(w); // transfer steals while idle
1783	<	Thread.interrupted();    // clear/ignore interrupt
1784	<	while (eventCount == prev)
1785	<	w.doPark();
1764	>	* If inactivating worker w has caused the pool to become
1765	>	* quiescent, checks for pool termination, and, so long as this is
1766	>	* not the only worker, waits for event for up to a given
1767	>	* duration.  On timeout, if ctl has not changed, terminates the
1768	>	* worker, which will in turn wake up another worker to possibly
1769	>	* repeat this process.
1770	>	*
1771	>	* @param w the calling worker
1772	>	* @param currentCtl the ctl value triggering possible quiescence
1773	>	* @param prevCtl the ctl value to restore if thread is terminated
1774	>	*/
1775	>	private void idleAwaitWork(WorkQueue w, long currentCtl, long prevCtl) {
1776	>	if (w != null && w.eventCount < 0 &&
1777	>	!tryTerminate(false, false) && (int)prevCtl != 0) {
1778	>	int dc = -(short)(currentCtl >>> TC_SHIFT);
1779	>	long parkTime = dc < 0 ? FAST_IDLE_TIMEOUT: (dc + 1) * IDLE_TIMEOUT;
1780	>	long deadline = System.nanoTime() + parkTime - TIMEOUT_SLOP;
1781	>	Thread wt = Thread.currentThread();
1782	>	while (ctl == currentCtl) {
1783	>	Thread.interrupted();  // timed variant of version in scan()
1784	>	U.putObject(wt, PARKBLOCKER, this);
1785	>	w.parker = wt;
1786	>	if (ctl == currentCtl)
1787	>	U.park(false, parkTime);
1788	>	w.parker = null;
1789	>	U.putObject(wt, PARKBLOCKER, null);
1790	>	if (ctl != currentCtl)
1791	>	break;
1792	>	if (deadline - System.nanoTime() <= 0L &&
1793	>	U.compareAndSwapLong(this, CTL, currentCtl, prevCtl)) {
1794	>	w.eventCount = (w.eventCount + E_SEQ) | E_MASK;
1795	>	w.qlock = -1;   // shrink
1796		break;
1797		}
1798		}
778	–	w.lastEventCount = eventCount;
1799		}
780	–	releaseWaiters();
1800		}
1801
1802		/**
1803	<	* Callback from workers invoked upon each top-level action (i.e.,
1804	<	* stealing a task or taking a submission and running
1805	<	* it). Performs one or both of the following:
1806	<	*
1807	<	* * If the worker cannot find work, updates its active status to
1808	<	* inactive and updates activeCount unless there is contention, in
1809	<	* which case it may try again (either in this or a subsequent
1810	<	* call).  Additionally, awaits the next task event and/or helps
1811	<	* wake up other releasable waiters.
1812	<	*
1813	<	* * If there are too many running threads, suspends this worker
1814	<	* (first forcing inactivation if necessary).  If it is not
1815	<	* resumed before a keepAlive elapses, the worker may be "trimmed"
1816	<	* -- killed while suspended within suspendAsSpare. Otherwise,
1817	<	* upon resume it rechecks to make sure that it is still needed.
1818	<	*
1819	<	* @param w the worker
1820	<	* @param worked false if the worker scanned for work but didn't
1821	<	* find any (in which case it may block waiting for work).
1822	<	*/
1823	<	final void preStep(ForkJoinWorkerThread w, boolean worked) {
1824	<	boolean active = w.active;
1825	<	boolean inactivate = !worked & active;
1826	<	for (;;) {
1827	<	if (inactivate) {
1828	<	int rs = runState;
1829	<	if (UNSAFE.compareAndSwapInt(this, runStateOffset,
1830	<	rs, rs - ONE_ACTIVE))
1831	<	inactivate = active = w.active = false;
1832	<	}
1833	<	int wc = workerCounts;
1834	<	if ((wc & RUNNING_COUNT_MASK) <= parallelism) {
1835	<	if (!worked)
1836	<	eventSync(w);
1837	<	return;
1803	>	* Scans through queues looking for work while joining a task; if
1804	>	* any present, signals. May return early if more signalling is
1805	>	* detectably unneeded.
1806	>	*
1807	>	* @param task return early if done
1808	>	* @param origin an index to start scan
1809	>	*/
1810	>	private void helpSignal(ForkJoinTask<?> task, int origin) {
1811	>	WorkQueue[] ws; WorkQueue w; Thread p; long c; int m, u, e, i, s;
1812	>	if (task != null && task.status >= 0 &&
1813	>	(u = (int)(ctl >>> 32)) < 0 && (u >> UAC_SHIFT) < 0 &&
1814	>	(ws = workQueues) != null && (m = ws.length - 1) >= 0) {
1815	>	outer: for (int k = origin, j = m; j >= 0; --j) {
1816	>	WorkQueue q = ws[k++ & m];
1817	>	for (int n = m;;) { // limit to at most m signals
1818	>	if (task.status < 0)
1819	>	break outer;
1820	>	if (q == null ||
1821	>	((s = -q.base + q.top) <= n && (n = s) <= 0))
1822	>	break;
1823	>	if ((u = (int)((c = ctl) >>> 32)) >= 0 ||
1824	>	(e = (int)c) <= 0 || m < (i = e & SMASK) ||
1825	>	(w = ws[i]) == null)
1826	>	break outer;
1827	>	long nc = (((long)(w.nextWait & E_MASK)) |
1828	>	((long)(u + UAC_UNIT) << 32));
1829	>	if (w.eventCount == (e | INT_SIGN) &&
1830	>	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		}
820	–	if (!(inactivate |= active) &&  // must inactivate to suspend
821	–	UNSAFE.compareAndSwapInt(this, workerCountsOffset,
822	–	wc, wc - ONE_RUNNING) &&
823	–	!w.suspendAsSpare())        // false if trimmed
824	–	return;
1839		}
1840		}
1841
1842		/**
1843	<	* Tries to decrement running count, and if so, possibly creates
830	<	* or resumes compensating threads before blocking on task joinMe.
831	<	* This code is sprawled out with manual inlining to evade some
832	<	* JIT oddities.
1843	>	* Signals other workers if tasks are present in hinted queue.
1844		*
1845	<	* @param joinMe the task to join
835	<	* @return task status on exit
1845	>	* @param caller the worker with the hint
1846		*/
1847	<	final int tryAwaitJoin(ForkJoinTask<?> joinMe) {
1848	<	int cw = workerCounts; // read now to spoil CAS if counts change as ...
1849	<	releaseWaiters();      // ... a byproduct of releaseWaiters
1850	<	int stat = joinMe.status;
1851	<	if (stat >= 0 && // inline variant of tryDecrementRunningCount
1852	<	(cw & RUNNING_COUNT_MASK) > 0 &&
1853	<	UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1854	<	cw, cw - ONE_RUNNING)) {
1855	<	int pc = parallelism;
1856	<	int scans = 0;  // to require confirming passes to add threads
1857	<	outer: while ((workerCounts & RUNNING_COUNT_MASK) < pc) {
1858	<	if ((stat = joinMe.status) < 0)
1859	<	break;
1860	<	ForkJoinWorkerThread spare = null;
851	<	ForkJoinWorkerThread[] ws = workers;
852	<	int nws = ws.length;
853	<	for (int i = 0; i < nws; ++i) {
854	<	ForkJoinWorkerThread w = ws[i];
855	<	if (w != null && w.isSuspended()) {
856	<	spare = w;
1847	>	private void helpSignalHint(WorkQueue caller) {
1848	>	WorkQueue[] ws; WorkQueue q, w; Thread p; long c; int h, m, u, e, i, s;
1849	>	if (caller != null && (h = caller.hint) >= 0) {
1850	>	caller.hint = -1;
1851	>	if ((u = (int)(ctl >>> 32)) < 0 && (u >> UAC_SHIFT) < 0 &&
1852	>	(ws = workQueues) != null && (m = ws.length - 1) >= 0 &&
1853	>	(q = ws[h & m]) != null) {
1854	>	for (int n = (m >>> 2) | 1;;) { // limit signals
1855	>	int idleCount = (caller.eventCount < 0) ? 0 : -1;
1856	>	if (((s = idleCount - q.base + q.top) <= n &&
1857	>	(n = s) <= 0) ||
1858	>	(u = (int)((c = ctl) >>> 32)) >= 0 ||
1859	>	(e = (int)c) <= 0 || m < (i = e & SMASK) ||
1860	>	(w = ws[i]) == null)
1861		break;
1862	+	long nc = (((long)(w.nextWait & E_MASK)) |
1863	+	((long)(u + UAC_UNIT) << 32));
1864	+	if (w.eventCount == (e | INT_SIGN) &&
1865	+	U.compareAndSwapLong(this, CTL, c, nc)) {
1866	+	w.hint = h;
1867	+	w.eventCount = (e + E_SEQ) & E_MASK;
1868	+	if ((p = w.parker) != null)
1869	+	U.unpark(p);
1870	+	if (--n <= 0)
1871	+	break;
1872		}
1873		}
1874	<	if ((stat = joinMe.status) < 0) // recheck to narrow race
1875	<	break;
1876	<	int wc = workerCounts;
1877	<	int rc = wc & RUNNING_COUNT_MASK;
1878	<	if (rc >= pc)
1879	<	break;
1880	<	if (spare != null) {
1881	<	if (spare.tryUnsuspend()) {
1882	<	int c; // inline incrementRunningCount
1883	<	do {} while (!UNSAFE.compareAndSwapInt
1884	<	(this, workerCountsOffset,
1885	<	c = workerCounts, c + ONE_RUNNING));
1886	<	LockSupport.unpark(spare);
1887	<	break;
1874	>	}
1875	>	}
1876	>	}
1877	>
1878	>	/**
1879	>	* Tries to locate and execute tasks for a stealer of the given
1880	>	* task, or in turn one of its stealers, Traces currentSteal ->
1881	>	* currentJoin links looking for a thread working on a descendant
1882	>	* of the given task and with a non-empty queue to steal back and
1883	>	* execute tasks from. The first call to this method upon a
1884	>	* waiting join will often entail scanning/search, (which is OK
1885	>	* because the joiner has nothing better to do), but this method
1886	>	* leaves hints in workers to speed up subsequent calls. The
1887	>	* implementation is very branchy to cope with potential
1888	>	* inconsistencies or loops encountering chains that are stale,
1889	>	* unknown, or so long that they are likely cyclic.
1890	>	*
1891	>	* @param joiner the joining worker
1892	>	* @param task the task to join
1893	>	* @return 0 if no progress can be made, negative if task
1894	>	* known complete, else positive
1895	>	*/
1896	>	private int tryHelpStealer(WorkQueue joiner, ForkJoinTask<?> task) {
1897	>	int stat = 0, steps = 0;                    // bound to avoid cycles
1898	>	if (joiner != null && task != null) {       // hoist null checks
1899	>	restart: for (;;) {
1900	>	ForkJoinTask<?> subtask = task;     // current target
1901	>	for (WorkQueue j = joiner, v;;) {   // v is stealer of subtask
1902	>	WorkQueue[] ws; int m, s, h;
1903	>	if ((s = task.status) < 0) {
1904	>	stat = s;
1905	>	break restart;
1906	>	}
1907	>	if ((ws = workQueues) == null || (m = ws.length - 1) <= 0)
1908	>	break restart;              // shutting down
1909	>	if ((v = ws[h = (j.hint | 1) & m]) == null ||
1910	>	v.currentSteal != subtask) {
1911	>	for (int origin = h;;) {    // find stealer
1912	>	if (((h = (h + 2) & m) & 15) == 1 &&
1913	>	(subtask.status < 0 || j.currentJoin != subtask))
1914	>	continue restart;   // occasional staleness check
1915	>	if ((v = ws[h]) != null &&
1916	>	v.currentSteal == subtask) {
1917	>	j.hint = h;        // save hint
1918	>	break;
1919	>	}
1920	>	if (h == origin)
1921	>	break restart;      // cannot find stealer
1922	>	}
1923	>	}
1924	>	for (;;) { // help stealer or descend to its stealer
1925	>	ForkJoinTask[] a;  int b;
1926	>	if (subtask.status < 0)     // surround probes with
1927	>	continue restart;       //   consistency checks
1928	>	if ((b = v.base) - v.top < 0 && (a = v.array) != null) {
1929	>	int i = (((a.length - 1) & b) << ASHIFT) + ABASE;
1930	>	ForkJoinTask<?> t =
1931	>	(ForkJoinTask<?>)U.getObjectVolatile(a, i);
1932	>	if (subtask.status < 0 || j.currentJoin != subtask ||
1933	>	v.currentSteal != subtask)
1934	>	continue restart;   // stale
1935	>	stat = 1;               // apparent progress
1936	>	if (t != null && v.base == b &&
1937	>	U.compareAndSwapObject(a, i, t, null)) {
1938	>	v.base = b + 1;     // help stealer
1939	>	joiner.runSubtask(t);
1940	>	}
1941	>	else if (v.base == b && ++steps == MAX_HELP)
1942	>	break restart;      // v apparently stalled
1943	>	}
1944	>	else {                      // empty -- try to descend
1945	>	ForkJoinTask<?> next = v.currentJoin;
1946	>	if (subtask.status < 0 || j.currentJoin != subtask ||
1947	>	v.currentSteal != subtask)
1948	>	continue restart;   // stale
1949	>	else if (next == null || ++steps == MAX_HELP)
1950	>	break restart;      // dead-end or maybe cyclic
1951	>	else {
1952	>	subtask = next;
1953	>	j = v;
1954	>	break;
1955	>	}
1956	>	}
1957		}
875	–	continue;
876	–	}
877	–	int tc = wc >>> TOTAL_COUNT_SHIFT;
878	–	int sc = tc - pc;
879	–	if (rc > 0) {
880	–	int p = pc;
881	–	int s = sc;
882	–	while (s-- >= 0) { // try keeping 3/4 live
883	–	if (rc > (p -= (p >>> 2) + 1))
884	–	break outer;
885	–	}
886	–	}
887	–	if (scans++ > sc && tc < MAX_THREADS &&
888	–	UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
889	–	wc + (ONE_RUNNING|ONE_TOTAL))) {
890	–	addWorker();
891	–	break;
1958		}
1959		}
894	–	if (stat >= 0)
895	–	stat = joinMe.internalAwaitDone();
896	–	int c; // inline incrementRunningCount
897	–	do {} while (!UNSAFE.compareAndSwapInt
898	–	(this, workerCountsOffset,
899	–	c = workerCounts, c + ONE_RUNNING));
1960		}
1961		return stat;
1962		}
1963
1964		/**
1965	<	* Same idea as (and mostly pasted from) tryAwaitJoin, but
1966	<	* self-contained
1967	<	*/
1968	<	final void awaitBlocker(ManagedBlocker blocker)
1969	<	throws InterruptedException {
1970	<	for (;;) {
1971	<	if (blocker.isReleasable())
1972	<	return;
1973	<	int cw = workerCounts;
1974	<	releaseWaiters();
1975	<	if ((cw & RUNNING_COUNT_MASK) > 0 &&
1976	<	UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1977	<	cw, cw - ONE_RUNNING))
1978	<	break;
1979	<	}
1980	<	boolean done = false;
1981	<	int pc = parallelism;
1982	<	int scans = 0;
1983	<	outer: while ((workerCounts & RUNNING_COUNT_MASK) < pc) {
1984	<	if (done = blocker.isReleasable())
925	<	break;
926	<	ForkJoinWorkerThread spare = null;
927	<	ForkJoinWorkerThread[] ws = workers;
928	<	int nws = ws.length;
929	<	for (int i = 0; i < nws; ++i) {
930	<	ForkJoinWorkerThread w = ws[i];
931	<	if (w != null && w.isSuspended()) {
932	<	spare = w;
933	<	break;
1965	>	* Analog of tryHelpStealer for CountedCompleters. Tries to steal
1966	>	* and run tasks within the target's computation.
1967	>	*
1968	>	* @param task the task to join
1969	>	* @param mode if shared, exit upon completing any task
1970	>	* if all workers are active
1971	>	*
1972	>	*/
1973	>	private int helpComplete(ForkJoinTask<?> task, int mode) {
1974	>	WorkQueue[] ws; WorkQueue q; int m, n, s, u;
1975	>	if (task != null && (ws = workQueues) != null &&
1976	>	(m = ws.length - 1) >= 0) {
1977	>	for (int j = 1, origin = j;;) {
1978	>	if ((s = task.status) < 0)
1979	>	return s;
1980	>	if ((q = ws[j & m]) != null && q.pollAndExecCC(task)) {
1981	>	origin = j;
1982	>	if (mode == SHARED_QUEUE &&
1983	>	((u = (int)(ctl >>> 32)) >= 0 || (u >> UAC_SHIFT) >= 0))
1984	>	break;
1985		}
1986	<	}
936	<	if (done = blocker.isReleasable())
937	<	break;
938	<	int wc = workerCounts;
939	<	int rc = wc & RUNNING_COUNT_MASK;
940	<	if (rc >= pc)
941	<	break;
942	<	if (spare != null) {
943	<	if (spare.tryUnsuspend()) {
944	<	int c;
945	<	do {} while (!UNSAFE.compareAndSwapInt
946	<	(this, workerCountsOffset,
947	<	c = workerCounts, c + ONE_RUNNING));
948	<	LockSupport.unpark(spare);
1986	>	else if ((j = (j + 2) & m) == origin)
1987		break;
950	–	}
951	–	continue;
1988		}
1989	<	int tc = wc >>> TOTAL_COUNT_SHIFT;
1990	<	int sc = tc - pc;
1991	<	if (rc > 0) {
1992	<	int p = pc;
1993	<	int s = sc;
1994	<	while (s-- >= 0) {
1995	<	if (rc > (p -= (p >>> 2) + 1))
1996	<	break outer;
1989	>	}
1990	>	return 0;
1991	>	}
1992	>
1993	>	/**
1994	>	* Tries to decrement active count (sometimes implicitly) and
1995	>	* possibly release or create a compensating worker in preparation
1996	>	* for blocking. Fails on contention or termination. Otherwise,
1997	>	* adds a new thread if no idle workers are available and pool
1998	>	* may become starved.
1999	>	*/
2000	>	final boolean tryCompensate() {
2001	>	int pc = config & SMASK, e, i, tc; long c;
2002	>	WorkQueue[] ws; WorkQueue w; Thread p;
2003	>	if ((ws = workQueues) != null && (e = (int)(c = ctl)) >= 0) {
2004	>	if (e != 0 && (i = e & SMASK) < ws.length &&
2005	>	(w = ws[i]) != null && w.eventCount == (e | INT_SIGN)) {
2006	>	long nc = ((long)(w.nextWait & E_MASK) |
2007	>	(c & (AC_MASK|TC_MASK)));
2008	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
2009	>	w.eventCount = (e + E_SEQ) & E_MASK;
2010	>	if ((p = w.parker) != null)
2011	>	U.unpark(p);
2012	>	return true;   // replace with idle worker
2013		}
2014		}
2015	<	if (scans++ > sc && tc < MAX_THREADS &&
2016	<	UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
2017	<	wc + (ONE_RUNNING|ONE_TOTAL))) {
2018	<	addWorker();
2019	<	break;
2015	>	else if ((tc = (short)(c >>> TC_SHIFT)) >= 0 &&
2016	>	(int)(c >> AC_SHIFT) + pc > 1) {
2017	>	long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK);
2018	>	if (U.compareAndSwapLong(this, CTL, c, nc))
2019	>	return true;   // no compensation
2020	>	}
2021	>	else if (tc + pc < MAX_CAP) {
2022	>	long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK);
2023	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
2024	>	ForkJoinWorkerThreadFactory fac;
2025	>	Throwable ex = null;
2026	>	ForkJoinWorkerThread wt = null;
2027	>	try {
2028	>	if ((fac = factory) != null &&
2029	>	(wt = fac.newThread(this)) != null) {
2030	>	wt.start();
2031	>	return true;
2032	>	}
2033	>	} catch (Throwable rex) {
2034	>	ex = rex;
2035	>	}
2036	>	deregisterWorker(wt, ex); // clean up and return false
2037	>	}
2038		}
2039		}
2040	<	try {
2041	<	if (!done)
972	<	do {} while (!blocker.isReleasable() &&
973	<	!blocker.block());
974	<	} finally {
975	<	int c;
976	<	do {} while (!UNSAFE.compareAndSwapInt
977	<	(this, workerCountsOffset,
978	<	c = workerCounts, c + ONE_RUNNING));
979	<	}
980	<	}
2040	>	return false;
2041	>	}
2042
2043		/**
2044	<	* Possibly initiates and/or completes termination.
2044	>	* Helps and/or blocks until the given task is done.
2045		*
2046	<	* @param now if true, unconditionally terminate, else only
2047	<	* if shutdown and empty queue and no active workers
2048	<	* @return true if now terminating or terminated
2046	>	* @param joiner the joining worker
2047	>	* @param task the task
2048	>	* @return task status on exit
2049		*/
2050	<	private boolean tryTerminate(boolean now) {
2051	<	if (now)
2052	<	advanceRunLevel(SHUTDOWN); // ensure at least SHUTDOWN
2053	<	else if (runState < SHUTDOWN ||
2054	<	!submissionQueue.isEmpty() ||
2055	<	(runState & ACTIVE_COUNT_MASK) != 0)
2056	<	return false;
2057	<
2058	<	if (advanceRunLevel(TERMINATING))
2059	<	startTerminating();
2060	<
2061	<	// Finish now if all threads terminated; else in some subsequent call
2062	<	if ((workerCounts >>> TOTAL_COUNT_SHIFT) == 0) {
2063	<	advanceRunLevel(TERMINATED);
2064	<	termination.arrive();
2050	>	final int awaitJoin(WorkQueue joiner, ForkJoinTask<?> task) {
2051	>	int s = 0;
2052	>	if (joiner != null && task != null && (s = task.status) >= 0) {
2053	>	ForkJoinTask<?> prevJoin = joiner.currentJoin;
2054	>	joiner.currentJoin = task;
2055	>	do {} while ((s = task.status) >= 0 && !joiner.isEmpty() &&
2056	>	joiner.tryRemoveAndExec(task)); // process local tasks
2057	>	if (s >= 0 && (s = task.status) >= 0) {
2058	>	helpSignal(task, joiner.poolIndex);
2059	>	if ((s = task.status) >= 0 &&
2060	>	(task instanceof CountedCompleter))
2061	>	s = helpComplete(task, LIFO_QUEUE);
2062	>	}
2063	>	while (s >= 0 && (s = task.status) >= 0) {
2064	>	if ((!joiner.isEmpty() ||           // try helping
2065	>	(s = tryHelpStealer(joiner, task)) == 0) &&
2066	>	(s = task.status) >= 0) {
2067	>	helpSignal(task, joiner.poolIndex);
2068	>	if ((s = task.status) >= 0 && tryCompensate()) {
2069	>	if (task.trySetSignal() && (s = task.status) >= 0) {
2070	>	synchronized (task) {
2071	>	if (task.status >= 0) {
2072	>	try {                // see ForkJoinTask
2073	>	task.wait();     //  for explanation
2074	>	} catch (InterruptedException ie) {
2075	>	}
2076	>	}
2077	>	else
2078	>	task.notifyAll();
2079	>	}
2080	>	}
2081	>	long c;                          // re-activate
2082	>	do {} while (!U.compareAndSwapLong
2083	>	(this, CTL, c = ctl, c + AC_UNIT));
2084	>	}
2085	>	}
2086	>	}
2087	>	joiner.currentJoin = prevJoin;
2088		}
2089	<	return true;
2089	>	return s;
2090		}
2091
2092		/**
2093	<	* Actions on transition to TERMINATING
2093	>	* Stripped-down variant of awaitJoin used by timed joins. Tries
2094	>	* to help join only while there is continuous progress. (Caller
2095	>	* will then enter a timed wait.)
2096	>	*
2097	>	* @param joiner the joining worker
2098	>	* @param task the task
2099		*/
2100	<	private void startTerminating() {
2101	<	for (int i = 0; i < 2; ++i) { // twice to mop up newly created workers
2102	<	cancelSubmissions();
2103	<	shutdownWorkers();
2104	<	cancelWorkerTasks();
2105	<	signalEvent();
2106	<	interruptWorkers();
2100	>	final void helpJoinOnce(WorkQueue joiner, ForkJoinTask<?> task) {
2101	>	int s;
2102	>	if (joiner != null && task != null && (s = task.status) >= 0) {
2103	>	ForkJoinTask<?> prevJoin = joiner.currentJoin;
2104	>	joiner.currentJoin = task;
2105	>	do {} while ((s = task.status) >= 0 && !joiner.isEmpty() &&
2106	>	joiner.tryRemoveAndExec(task));
2107	>	if (s >= 0 && (s = task.status) >= 0) {
2108	>	helpSignal(task, joiner.poolIndex);
2109	>	if ((s = task.status) >= 0 &&
2110	>	(task instanceof CountedCompleter))
2111	>	s = helpComplete(task, LIFO_QUEUE);
2112	>	}
2113	>	if (s >= 0 && joiner.isEmpty()) {
2114	>	do {} while (task.status >= 0 &&
2115	>	tryHelpStealer(joiner, task) > 0);
2116	>	}
2117	>	joiner.currentJoin = prevJoin;
2118		}
2119		}
2120
2121		/**
2122	<	* Clear out and cancel submissions, ignoring exceptions
2122	>	* Returns a (probably) non-empty steal queue, if one is found
2123	>	* during a random, then cyclic scan, else null.  This method must
2124	>	* be retried by caller if, by the time it tries to use the queue,
2125	>	* it is empty.
2126	>	* @param r a (random) seed for scanning
2127		*/
2128	<	private void cancelSubmissions() {
2129	<	ForkJoinTask<?> task;
2130	<	while ((task = submissionQueue.poll()) != null) {
2131	<	try {
2132	<	task.cancel(false);
2133	<	} catch (Throwable ignore) {
2128	>	private WorkQueue findNonEmptyStealQueue(int r) {
2129	>	for (WorkQueue[] ws;;) {
2130	>	int ps = plock, m, n;
2131	>	if ((ws = workQueues) == null || (m = ws.length - 1) < 1)
2132	>	return null;
2133	>	for (int j = (m + 1) << 2; ;) {
2134	>	WorkQueue q = ws[(((r + j) << 1) | 1) & m];
2135	>	if (q != null && (n = q.base - q.top) < 0) {
2136	>	if (n < -1)
2137	>	signalWork(q);
2138	>	return q;
2139	>	}
2140	>	else if (--j < 0) {
2141	>	if (plock == ps)
2142	>	return null;
2143	>	break;
2144	>	}
2145		}
2146		}
2147		}
2148
2149		/**
2150	<	* Sets all worker run states to at least shutdown,
2151	<	* also resuming suspended workers
2152	<	*/
2153	<	private void shutdownWorkers() {
2154	<	ForkJoinWorkerThread[] ws = workers;
2155	<	int nws = ws.length;
2156	<	for (int i = 0; i < nws; ++i) {
2157	<	ForkJoinWorkerThread w = ws[i];
2158	<	if (w != null)
2159	<	w.shutdown();
2150	>	* Runs tasks until {@code isQuiescent()}. We piggyback on
2151	>	* active count ctl maintenance, but rather than blocking
2152	>	* when tasks cannot be found, we rescan until all others cannot
2153	>	* find tasks either.
2154	>	*/
2155	>	final void helpQuiescePool(WorkQueue w) {
2156	>	for (boolean active = true;;) {
2157	>	ForkJoinTask<?> localTask; // exhaust local queue
2158	>	while ((localTask = w.nextLocalTask()) != null)
2159	>	localTask.doExec();
2160	>	// Similar to loop in scan(), but ignoring submissions
2161	>	WorkQueue q = findNonEmptyStealQueue(w.nextSeed());
2162	>	if (q != null) {
2163	>	ForkJoinTask<?> t; int b;
2164	>	if (!active) {      // re-establish active count
2165	>	long c;
2166	>	active = true;
2167	>	do {} while (!U.compareAndSwapLong
2168	>	(this, CTL, c = ctl, c + AC_UNIT));
2169	>	}
2170	>	if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null)
2171	>	w.runSubtask(t);
2172	>	}
2173	>	else {
2174	>	long c;
2175	>	if (active) {       // decrement active count without queuing
2176	>	active = false;
2177	>	do {} while (!U.compareAndSwapLong
2178	>	(this, CTL, c = ctl, c -= AC_UNIT));
2179	>	}
2180	>	else
2181	>	c = ctl;        // re-increment on exit
2182	>	if ((int)(c >> AC_SHIFT) + (config & SMASK) == 0) {
2183	>	do {} while (!U.compareAndSwapLong
2184	>	(this, CTL, c = ctl, c + AC_UNIT));
2185	>	break;
2186	>	}
2187	>	}
2188		}
2189		}
2190
2191		/**
2192	<	* Clears out and cancels all locally queued tasks
2192	>	* Gets and removes a local or stolen task for the given worker.
2193	>	*
2194	>	* @return a task, if available
2195		*/
2196	<	private void cancelWorkerTasks() {
2197	<	ForkJoinWorkerThread[] ws = workers;
2198	<	int nws = ws.length;
2199	<	for (int i = 0; i < nws; ++i) {
2200	<	ForkJoinWorkerThread w = ws[i];
2201	<	if (w != null)
2202	<	w.cancelTasks();
2196	>	final ForkJoinTask<?> nextTaskFor(WorkQueue w) {
2197	>	for (ForkJoinTask<?> t;;) {
2198	>	WorkQueue q; int b;
2199	>	if ((t = w.nextLocalTask()) != null)
2200	>	return t;
2201	>	if ((q = findNonEmptyStealQueue(w.nextSeed())) == null)
2202	>	return null;
2203	>	if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null)
2204	>	return t;
2205		}
2206		}
2207
2208		/**
2209	<	* Unsticks all workers blocked on joins etc
2209	>	* Returns a cheap heuristic guide for task partitioning when
2210	>	* programmers, frameworks, tools, or languages have little or no
2211	>	* idea about task granularity.  In essence by offering this
2212	>	* method, we ask users only about tradeoffs in overhead vs
2213	>	* expected throughput and its variance, rather than how finely to
2214	>	* partition tasks.
2215	>	*
2216	>	* In a steady state strict (tree-structured) computation, each
2217	>	* thread makes available for stealing enough tasks for other
2218	>	* threads to remain active. Inductively, if all threads play by
2219	>	* the same rules, each thread should make available only a
2220	>	* constant number of tasks.
2221	>	*
2222	>	* The minimum useful constant is just 1. But using a value of 1
2223	>	* would require immediate replenishment upon each steal to
2224	>	* maintain enough tasks, which is infeasible.  Further,
2225	>	* partitionings/granularities of offered tasks should minimize
2226	>	* steal rates, which in general means that threads nearer the top
2227	>	* of computation tree should generate more than those nearer the
2228	>	* bottom. In perfect steady state, each thread is at
2229	>	* approximately the same level of computation tree. However,
2230	>	* producing extra tasks amortizes the uncertainty of progress and
2231	>	* diffusion assumptions.
2232	>	*
2233	>	* So, users will want to use values larger, but not much larger
2234	>	* than 1 to both smooth over transient shortages and hedge
2235	>	* against uneven progress; as traded off against the cost of
2236	>	* extra task overhead. We leave the user to pick a threshold
2237	>	* value to compare with the results of this call to guide
2238	>	* decisions, but recommend values such as 3.
2239	>	*
2240	>	* When all threads are active, it is on average OK to estimate
2241	>	* surplus strictly locally. In steady-state, if one thread is
2242	>	* maintaining say 2 surplus tasks, then so are others. So we can
2243	>	* just use estimated queue length.  However, this strategy alone
2244	>	* leads to serious mis-estimates in some non-steady-state
2245	>	* conditions (ramp-up, ramp-down, other stalls). We can detect
2246	>	* many of these by further considering the number of "idle"
2247	>	* threads, that are known to have zero queued tasks, so
2248	>	* compensate by a factor of (#idle/#active) threads.
2249	>	*
2250	>	* Note: The approximation of #busy workers as #active workers is
2251	>	* not very good under current signalling scheme, and should be
2252	>	* improved.
2253	>	*/
2254	>	static int getSurplusQueuedTaskCount() {
2255	>	Thread t; ForkJoinWorkerThread wt; ForkJoinPool pool; WorkQueue q;
2256	>	if (((t = Thread.currentThread()) instanceof ForkJoinWorkerThread)) {
2257	>	int p = (pool = (wt = (ForkJoinWorkerThread)t).pool).config & SMASK;
2258	>	int n = (q = wt.workQueue).top - q.base;
2259	>	int a = (int)(pool.ctl >> AC_SHIFT) + p;
2260	>	return n - (a > (p >>>= 1) ? 0 :
2261	>	a > (p >>>= 1) ? 1 :
2262	>	a > (p >>>= 1) ? 2 :
2263	>	a > (p >>>= 1) ? 4 :
2264	>	8);
2265	>	}
2266	>	return 0;
2267	>	}
2268	>
2269	>	//  Termination
2270	>
2271	>	/**
2272	>	* Possibly initiates and/or completes termination.  The caller
2273	>	* triggering termination runs three passes through workQueues:
2274	>	* (0) Setting termination status, followed by wakeups of queued
2275	>	* workers; (1) cancelling all tasks; (2) interrupting lagging
2276	>	* threads (likely in external tasks, but possibly also blocked in
2277	>	* joins).  Each pass repeats previous steps because of potential
2278	>	* lagging thread creation.
2279	>	*
2280	>	* @param now if true, unconditionally terminate, else only
2281	>	* if no work and no active workers
2282	>	* @param enable if true, enable shutdown when next possible
2283	>	* @return true if now terminating or terminated
2284		*/
2285	<	private void interruptWorkers() {
2286	<	ForkJoinWorkerThread[] ws = workers;
2287	<	int nws = ws.length;
2288	<	for (int i = 0; i < nws; ++i) {
2289	<	ForkJoinWorkerThread w = ws[i];
2290	<	if (w != null && !w.isTerminated()) {
2291	<	try {
2292	<	w.interrupt();
2293	<	} catch (SecurityException ignore) {
2285	>	private boolean tryTerminate(boolean now, boolean enable) {
2286	>	if (this == commonPool)                     // cannot shut down
2287	>	return false;
2288	>	for (long c;;) {
2289	>	if (((c = ctl) & STOP_BIT) != 0) {      // already terminating
2290	>	if ((short)(c >>> TC_SHIFT) == -(config & SMASK)) {
2291	>	synchronized (this) {
2292	>	notifyAll();                // signal when 0 workers
2293	>	}
2294	>	}
2295	>	return true;
2296	>	}
2297	>	if (plock >= 0) {                       // not yet enabled
2298	>	int ps;
2299	>	if (!enable)
2300	>	return false;
2301	>	if (((ps = plock) & PL_LOCK) != 0 ||
2302	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
2303	>	ps = acquirePlock();
2304	>	int nps = SHUTDOWN;
2305	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
2306	>	releasePlock(nps);
2307	>	}
2308	>	if (!now) {                             // check if idle & no tasks
2309	>	if ((int)(c >> AC_SHIFT) != -(config & SMASK) ||
2310	>	hasQueuedSubmissions())
2311	>	return false;
2312	>	// Check for unqueued inactive workers. One pass suffices.
2313	>	WorkQueue[] ws = workQueues; WorkQueue w;
2314	>	if (ws != null) {
2315	>	for (int i = 1; i < ws.length; i += 2) {
2316	>	if ((w = ws[i]) != null && w.eventCount >= 0)
2317	>	return false;
2318	>	}
2319	>	}
2320	>	}
2321	>	if (U.compareAndSwapLong(this, CTL, c, c | STOP_BIT)) {
2322	>	for (int pass = 0; pass < 3; ++pass) {
2323	>	WorkQueue[] ws = workQueues;
2324	>	if (ws != null) {
2325	>	WorkQueue w;
2326	>	int n = ws.length;
2327	>	for (int i = 0; i < n; ++i) {
2328	>	if ((w = ws[i]) != null) {
2329	>	w.qlock = -1;
2330	>	if (pass > 0) {
2331	>	w.cancelAll();
2332	>	if (pass > 1)
2333	>	w.interruptOwner();
2334	>	}
2335	>	}
2336	>	}
2337	>	// Wake up workers parked on event queue
2338	>	int i, e; long cc; Thread p;
2339	>	while ((e = (int)(cc = ctl) & E_MASK) != 0 &&
2340	>	(i = e & SMASK) < n &&
2341	>	(w = ws[i]) != null) {
2342	>	long nc = ((long)(w.nextWait & E_MASK) |
2343	>	((cc + AC_UNIT) & AC_MASK) |
2344	>	(cc & (TC_MASK|STOP_BIT)));
2345	>	if (w.eventCount == (e | INT_SIGN) &&
2346	>	U.compareAndSwapLong(this, CTL, cc, nc)) {
2347	>	w.eventCount = (e + E_SEQ) & E_MASK;
2348	>	w.qlock = -1;
2349	>	if ((p = w.parker) != null)
2350	>	U.unpark(p);
2351	>	}
2352	>	}
2353	>	}
2354		}
2355		}
2356		}
2357		}
2358
2359	<	// misc support for ForkJoinWorkerThread
2359	>	// external operations on common pool
2360
2361		/**
2362	<	* Returns pool number
2362	>	* Returns common pool queue for a thread that has submitted at
2363	>	* least one task.
2364		*/
2365	<	final int getPoolNumber() {
2366	<	return poolNumber;
2365	>	static WorkQueue commonSubmitterQueue() {
2366	>	ForkJoinPool p; WorkQueue[] ws; int m; Submitter z;
2367	>	return ((z = submitters.get()) != null &&
2368	>	(p = commonPool) != null &&
2369	>	(ws = p.workQueues) != null &&
2370	>	(m = ws.length - 1) >= 0) ?
2371	>	ws[m & z.seed & SQMASK] : null;
2372		}
2373
2374		/**
2375	<	* Accumulates steal count from a worker, clearing
2376	<	* the worker's value
2375	>	* Tries to pop the given task from submitter's queue in common pool.
2376	>	*/
2377	>	static boolean tryExternalUnpush(ForkJoinTask<?> t) {
2378	>	ForkJoinPool p; WorkQueue[] ws; WorkQueue q; Submitter z;
2379	>	ForkJoinTask<?>[] a;  int m, s;
2380	>	if (t != null &&
2381	>	(z = submitters.get()) != null &&
2382	>	(p = commonPool) != null &&
2383	>	(ws = p.workQueues) != null &&
2384	>	(m = ws.length - 1) >= 0 &&
2385	>	(q = ws[m & z.seed & SQMASK]) != null &&
2386	>	(s = q.top) != q.base &&
2387	>	(a = q.array) != null) {
2388	>	long j = (((a.length - 1) & (s - 1)) << ASHIFT) + ABASE;
2389	>	if (U.getObject(a, j) == t &&
2390	>	U.compareAndSwapInt(q, QLOCK, 0, 1)) {
2391	>	if (q.array == a && q.top == s && // recheck
2392	>	U.compareAndSwapObject(a, j, t, null)) {
2393	>	q.top = s - 1;
2394	>	q.qlock = 0;
2395	>	return true;
2396	>	}
2397	>	q.qlock = 0;
2398	>	}
2399	>	}
2400	>	return false;
2401	>	}
2402	>
2403	>	/**
2404	>	* Tries to pop and run local tasks within the same computation
2405	>	* as the given root. On failure, tries to help complete from
2406	>	* other queues via helpComplete.
2407		*/
2408	<	final void accumulateStealCount(ForkJoinWorkerThread w) {
2409	<	int sc = w.stealCount;
2410	<	if (sc != 0) {
2411	<	long c;
2412	<	w.stealCount = 0;
2413	<	do {} while (!UNSAFE.compareAndSwapLong(this, stealCountOffset,
2414	<	c = stealCount, c + sc));
2408	>	private void externalHelpComplete(WorkQueue q, ForkJoinTask<?> root) {
2409	>	ForkJoinTask<?>[] a; int m;
2410	>	if (q != null && (a = q.array) != null && (m = (a.length - 1)) >= 0 &&
2411	>	root != null && root.status >= 0) {
2412	>	for (;;) {
2413	>	int s, u; Object o; CountedCompleter<?> task = null;
2414	>	if ((s = q.top) - q.base > 0) {
2415	>	long j = ((m & (s - 1)) << ASHIFT) + ABASE;
2416	>	if ((o = U.getObject(a, j)) != null &&
2417	>	(o instanceof CountedCompleter)) {
2418	>	CountedCompleter<?> t = (CountedCompleter<?>)o, r = t;
2419	>	do {
2420	>	if (r == root) {
2421	>	if (U.compareAndSwapInt(q, QLOCK, 0, 1)) {
2422	>	if (q.array == a && q.top == s &&
2423	>	U.compareAndSwapObject(a, j, t, null)) {
2424	>	q.top = s - 1;
2425	>	task = t;
2426	>	}
2427	>	q.qlock = 0;
2428	>	}
2429	>	break;
2430	>	}
2431	>	} while ((r = r.completer) != null);
2432	>	}
2433	>	}
2434	>	if (task != null)
2435	>	task.doExec();
2436	>	if (root.status < 0 ||
2437	>	(u = (int)(ctl >>> 32)) >= 0 || (u >> UAC_SHIFT) >= 0)
2438	>	break;
2439	>	if (task == null) {
2440	>	helpSignal(root, q.poolIndex);
2441	>	if (root.status >= 0)
2442	>	helpComplete(root, SHARED_QUEUE);
2443	>	break;
2444	>	}
2445	>	}
2446		}
2447		}
2448
2449		/**
2450	<	* Returns the approximate (non-atomic) number of idle threads per
2451	<	* active thread.
2450	>	* Tries to help execute or signal availability of the given task
2451	>	* from submitter's queue in common pool.
2452		*/
2453	<	final int idlePerActive() {
2454	<	int pc = parallelism; // use targeted parallelism, not rc
2455	<	int ac = runState;    // no mask -- artifically boosts during shutdown
2456	<	// Use exact results for small values, saturate past 4
2457	<	return pc <= ac? 0 : pc >>> 1 <= ac? 1 : pc >>> 2 <= ac? 3 : pc >>> 3;
2453	>	static void externalHelpJoin(ForkJoinTask<?> t) {
2454	>	// Some hard-to-avoid overlap with tryExternalUnpush
2455	>	ForkJoinPool p; WorkQueue[] ws; WorkQueue q, w; Submitter z;
2456	>	ForkJoinTask<?>[] a;  int m, s, n;
2457	>	if (t != null &&
2458	>	(z = submitters.get()) != null &&
2459	>	(p = commonPool) != null &&
2460	>	(ws = p.workQueues) != null &&
2461	>	(m = ws.length - 1) >= 0 &&
2462	>	(q = ws[m & z.seed & SQMASK]) != null &&
2463	>	(a = q.array) != null) {
2464	>	int am = a.length - 1;
2465	>	if ((s = q.top) != q.base) {
2466	>	long j = ((am & (s - 1)) << ASHIFT) + ABASE;
2467	>	if (U.getObject(a, j) == t &&
2468	>	U.compareAndSwapInt(q, QLOCK, 0, 1)) {
2469	>	if (q.array == a && q.top == s &&
2470	>	U.compareAndSwapObject(a, j, t, null)) {
2471	>	q.top = s - 1;
2472	>	q.qlock = 0;
2473	>	t.doExec();
2474	>	}
2475	>	else
2476	>	q.qlock = 0;
2477	>	}
2478	>	}
2479	>	if (t.status >= 0) {
2480	>	if (t instanceof CountedCompleter)
2481	>	p.externalHelpComplete(q, t);
2482	>	else
2483	>	p.helpSignal(t, q.poolIndex);
2484	>	}
2485	>	}
2486		}
2487
2488	<	// Public and protected methods
2488	>	/**
2489	>	* Restricted version of helpQuiescePool for external callers
2490	>	*/
2491	>	static void externalHelpQuiescePool() {
2492	>	ForkJoinPool p; ForkJoinTask<?> t; WorkQueue q; int b;
2493	>	if ((p = commonPool) != null &&
2494	>	(q = p.findNonEmptyStealQueue(1)) != null &&
2495	>	(b = q.base) - q.top < 0 &&
2496	>	(t = q.pollAt(b)) != null)
2497	>	t.doExec();
2498	>	}
2499	>
2500	>	// Exported methods
2501
2502		// Constructors
2503
#	Line 1154 | Line 2542 | public class ForkJoinPool extends Abstra
2542		* use {@link java.lang.Runtime#availableProcessors}.
2543		* @param factory the factory for creating new threads. For default value,
2544		* use {@link #defaultForkJoinWorkerThreadFactory}.
2545	<	* @param handler the handler for internal worker threads that
2546	<	* terminate due to unrecoverable errors encountered while executing
2547	<	* tasks. For default value, use <code>null</code>.
2548	<	* @param asyncMode if true,
2545	>	* @param handler the handler for internal worker threads that
2546	>	* terminate due to unrecoverable errors encountered while executing
2547	>	* tasks. For default value, use {@code null}.
2548	>	* @param asyncMode if true,
2549		* establishes local first-in-first-out scheduling mode for forked
2550		* tasks that are never joined. This mode may be more appropriate
2551		* than default locally stack-based mode in applications in which
2552		* worker threads only process event-style asynchronous tasks.
2553	<	* For default value, use <code>false</code>.
2553	>	* For default value, use {@code false}.
2554		* @throws IllegalArgumentException if parallelism less than or
2555		*         equal to zero, or greater than implementation limit
2556		* @throws NullPointerException if the factory is null
#	Line 1171 | Line 2559 | public class ForkJoinPool extends Abstra
2559		*         because it does not hold {@link
2560		*         java.lang.RuntimePermission}{@code ("modifyThread")}
2561		*/
2562	<	public ForkJoinPool(int parallelism,
2562	>	public ForkJoinPool(int parallelism,
2563		ForkJoinWorkerThreadFactory factory,
2564		Thread.UncaughtExceptionHandler handler,
2565		boolean asyncMode) {
2566		checkPermission();
2567		if (factory == null)
2568		throw new NullPointerException();
2569	<	if (parallelism <= 0 || parallelism > MAX_THREADS)
2569	>	if (parallelism <= 0 || parallelism > MAX_CAP)
2570		throw new IllegalArgumentException();
1183	–	this.parallelism = parallelism;
2571		this.factory = factory;
2572		this.ueh = handler;
2573	<	this.locallyFifo = asyncMode;
2574	<	int arraySize = initialArraySizeFor(parallelism);
2575	<	this.workers = new ForkJoinWorkerThread[arraySize];
2576	<	this.submissionQueue = new LinkedTransferQueue<ForkJoinTask<?>>();
2577	<	this.workerLock = new ReentrantLock();
2578	<	this.termination = new Phaser(1);
2579	<	this.poolNumber = poolNumberGenerator.incrementAndGet();
2573	>	this.config = parallelism | (asyncMode ? (FIFO_QUEUE << 16) : 0);
2574	>	long np = (long)(-parallelism); // offset ctl counts
2575	>	this.ctl = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK);
2576	>	int pn = nextPoolId();
2577	>	StringBuilder sb = new StringBuilder("ForkJoinPool-");
2578	>	sb.append(Integer.toString(pn));
2579	>	sb.append("-worker-");
2580	>	this.workerNamePrefix = sb.toString();
2581		}
2582
2583		/**
2584	<	* Returns initial power of two size for workers array.
2585	<	* @param pc the initial parallelism level
2586	<	*/
2587	<	private static int initialArraySizeFor(int pc) {
2588	<	// See Hackers Delight, sec 3.2. We know MAX_THREADS < (1 >>> 16)
2589	<	int size = pc < MAX_THREADS ? pc + 1 : MAX_THREADS;
2590	<	size |= size >>> 1;
2591	<	size |= size >>> 2;
2592	<	size |= size >>> 4;
2593	<	size |= size >>> 8;
2594	<	return size + 1;
2584	>	* Constructor for common pool, suitable only for static initialization.
2585	>	* Basically the same as above, but uses smallest possible initial footprint.
2586	>	*/
2587	>	ForkJoinPool(int parallelism, long ctl,
2588	>	ForkJoinWorkerThreadFactory factory,
2589	>	Thread.UncaughtExceptionHandler handler) {
2590	>	this.config = parallelism;
2591	>	this.ctl = ctl;
2592	>	this.factory = factory;
2593	>	this.ueh = handler;
2594	>	this.workerNamePrefix = "ForkJoinPool.commonPool-worker-";
2595		}
2596
1209	–	// Execution methods
1210	–
2597		/**
2598	<	* Common code for execute, invoke and submit
2598	>	* Returns the common pool instance.
2599	>	*
2600	>	* @return the common pool instance
2601		*/
2602	<	private <T> void doSubmit(ForkJoinTask<T> task) {
2603	<	if (task == null)
2604	<	throw new NullPointerException();
1217	<	if (runState >= SHUTDOWN)
1218	<	throw new RejectedExecutionException();
1219	<	// Convert submissions to current pool into forks
1220	<	Thread t = Thread.currentThread();
1221	<	ForkJoinWorkerThread w;
1222	<	if ((t instanceof ForkJoinWorkerThread) &&
1223	<	(w = (ForkJoinWorkerThread) t).pool == this)
1224	<	w.pushTask(task);
1225	<	else {
1226	<	submissionQueue.offer(task);
1227	<	signalEvent();
1228	<	ensureEnoughTotalWorkers();
1229	<	}
2602	>	public static ForkJoinPool commonPool() {
2603	>	// assert commonPool != null : "static init error";
2604	>	return commonPool;
2605		}
2606
2607	+	// Execution methods
2608	+
2609		/**
2610		* Performs the given task, returning its result upon completion.
2611	<	* If the caller is already engaged in a fork/join computation in
2612	<	* the current pool, this method is equivalent in effect to
2613	<	* {@link ForkJoinTask#invoke}.
2611	>	* If the computation encounters an unchecked Exception or Error,
2612	>	* it is rethrown as the outcome of this invocation.  Rethrown
2613	>	* exceptions behave in the same way as regular exceptions, but,
2614	>	* when possible, contain stack traces (as displayed for example
2615	>	* using {@code ex.printStackTrace()}) of both the current thread
2616	>	* as well as the thread actually encountering the exception;
2617	>	* minimally only the latter.
2618		*
2619		* @param task the task
2620		* @return the task's result
#	Line 1242 | Line 2623 | public class ForkJoinPool extends Abstra
2623		*         scheduled for execution
2624		*/
2625		public <T> T invoke(ForkJoinTask<T> task) {
2626	<	doSubmit(task);
2626	>	if (task == null)
2627	>	throw new NullPointerException();
2628	>	externalPush(task);
2629		return task.join();
2630		}
2631
2632		/**
2633		* Arranges for (asynchronous) execution of the given task.
1251	–	* If the caller is already engaged in a fork/join computation in
1252	–	* the current pool, this method is equivalent in effect to
1253	–	* {@link ForkJoinTask#fork}.
2634		*
2635		* @param task the task
2636		* @throws NullPointerException if the task is null
#	Line 1258 | Line 2638 | public class ForkJoinPool extends Abstra
2638		*         scheduled for execution
2639		*/
2640		public void execute(ForkJoinTask<?> task) {
2641	<	doSubmit(task);
2641	>	if (task == null)
2642	>	throw new NullPointerException();
2643	>	externalPush(task);
2644		}
2645
2646		// AbstractExecutorService methods
#	Line 1269 | Line 2651 | public class ForkJoinPool extends Abstra
2651		*         scheduled for execution
2652		*/
2653		public void execute(Runnable task) {
2654	+	if (task == null)
2655	+	throw new NullPointerException();
2656		ForkJoinTask<?> job;
2657		if (task instanceof ForkJoinTask<?>) // avoid re-wrap
2658		job = (ForkJoinTask<?>) task;
2659		else
2660	<	job = ForkJoinTask.adapt(task, null);
2661	<	doSubmit(job);
2660	>	job = new ForkJoinTask.AdaptedRunnableAction(task);
2661	>	externalPush(job);
2662		}
2663
2664		/**
2665		* Submits a ForkJoinTask for execution.
1282	–	* If the caller is already engaged in a fork/join computation in
1283	–	* the current pool, this method is equivalent in effect to
1284	–	* {@link ForkJoinTask#fork}.
2666		*
2667		* @param task the task to submit
2668		* @return the task
#	Line 1290 | Line 2671 | public class ForkJoinPool extends Abstra
2671		*         scheduled for execution
2672		*/
2673		public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) {
2674	<	doSubmit(task);
2674	>	if (task == null)
2675	>	throw new NullPointerException();
2676	>	externalPush(task);
2677		return task;
2678		}
2679
#	Line 1300 | Line 2683 | public class ForkJoinPool extends Abstra
2683		*         scheduled for execution
2684		*/
2685		public <T> ForkJoinTask<T> submit(Callable<T> task) {
2686	<	ForkJoinTask<T> job = ForkJoinTask.adapt(task);
2687	<	doSubmit(job);
2686	>	ForkJoinTask<T> job = new ForkJoinTask.AdaptedCallable<T>(task);
2687	>	externalPush(job);
2688		return job;
2689		}
2690
#	Line 1311 | Line 2694 | public class ForkJoinPool extends Abstra
2694		*         scheduled for execution
2695		*/
2696		public <T> ForkJoinTask<T> submit(Runnable task, T result) {
2697	<	ForkJoinTask<T> job = ForkJoinTask.adapt(task, result);
2698	<	doSubmit(job);
2697	>	ForkJoinTask<T> job = new ForkJoinTask.AdaptedRunnable<T>(task, result);
2698	>	externalPush(job);
2699		return job;
2700		}
2701
#	Line 1322 | Line 2705 | public class ForkJoinPool extends Abstra
2705		*         scheduled for execution
2706		*/
2707		public ForkJoinTask<?> submit(Runnable task) {
2708	+	if (task == null)
2709	+	throw new NullPointerException();
2710		ForkJoinTask<?> job;
2711		if (task instanceof ForkJoinTask<?>) // avoid re-wrap
2712		job = (ForkJoinTask<?>) task;
2713		else
2714	<	job = ForkJoinTask.adapt(task, null);
2715	<	doSubmit(job);
2714	>	job = new ForkJoinTask.AdaptedRunnableAction(task);
2715	>	externalPush(job);
2716		return job;
2717		}
2718
#	Line 1336 | Line 2721 | public class ForkJoinPool extends Abstra
2721		* @throws RejectedExecutionException {@inheritDoc}
2722		*/
2723		public <T> List<Future<T>> invokeAll(Collection<? extends Callable<T>> tasks) {
2724	<	ArrayList<ForkJoinTask<T>> forkJoinTasks =
2725	<	new ArrayList<ForkJoinTask<T>>(tasks.size());
2726	<	for (Callable<T> task : tasks)
2727	<	forkJoinTasks.add(ForkJoinTask.adapt(task));
2728	<	invoke(new InvokeAll<T>(forkJoinTasks));
2729	<
2724	>	// In previous versions of this class, this method constructed
2725	>	// a task to run ForkJoinTask.invokeAll, but now external
2726	>	// invocation of multiple tasks is at least as efficient.
2727	>	List<ForkJoinTask<T>> fs = new ArrayList<ForkJoinTask<T>>(tasks.size());
2728	>	// Workaround needed because method wasn't declared with
2729	>	// wildcards in return type but should have been.
2730		@SuppressWarnings({"unchecked", "rawtypes"})
2731	<	List<Future<T>> futures = (List<Future<T>>) (List) forkJoinTasks;
1347	<	return futures;
1348	<	}
2731	>	List<Future<T>> futures = (List<Future<T>>) (List) fs;
2732
2733	<	static final class InvokeAll<T> extends RecursiveAction {
2734	<	final ArrayList<ForkJoinTask<T>> tasks;
2735	<	InvokeAll(ArrayList<ForkJoinTask<T>> tasks) { this.tasks = tasks; }
2736	<	public void compute() {
2737	<	try { invokeAll(tasks); }
2738	<	catch (Exception ignore) {}
2733	>	boolean done = false;
2734	>	try {
2735	>	for (Callable<T> t : tasks) {
2736	>	ForkJoinTask<T> f = new ForkJoinTask.AdaptedCallable<T>(t);
2737	>	externalPush(f);
2738	>	fs.add(f);
2739	>	}
2740	>	for (ForkJoinTask<T> f : fs)
2741	>	f.quietlyJoin();
2742	>	done = true;
2743	>	return futures;
2744	>	} finally {
2745	>	if (!done)
2746	>	for (ForkJoinTask<T> f : fs)
2747	>	f.cancel(false);
2748		}
1357	–	private static final long serialVersionUID = -7914297376763021607L;
2749		}
2750
2751		/**
#	Line 1382 | Line 2773 | public class ForkJoinPool extends Abstra
2773		* @return the targeted parallelism level of this pool
2774		*/
2775		public int getParallelism() {
2776	<	return parallelism;
2776	>	return config & SMASK;
2777	>	}
2778	>
2779	>	/**
2780	>	* Returns the targeted parallelism level of the common pool.
2781	>	*
2782	>	* @return the targeted parallelism level of the common pool
2783	>	*/
2784	>	public static int getCommonPoolParallelism() {
2785	>	return commonPoolParallelism;
2786		}
2787
2788		/**
2789		* Returns the number of worker threads that have started but not
2790	<	* yet terminated.  This result returned by this method may differ
2790	>	* yet terminated.  The result returned by this method may differ
2791		* from {@link #getParallelism} when threads are created to
2792		* maintain parallelism when others are cooperatively blocked.
2793		*
2794		* @return the number of worker threads
2795		*/
2796		public int getPoolSize() {
2797	<	return workerCounts >>> TOTAL_COUNT_SHIFT;
2797	>	return (config & SMASK) + (short)(ctl >>> TC_SHIFT);
2798		}
2799
2800		/**
#	Line 1404 | Line 2804 | public class ForkJoinPool extends Abstra
2804		* @return {@code true} if this pool uses async mode
2805		*/
2806		public boolean getAsyncMode() {
2807	<	return locallyFifo;
2807	>	return (config >>> 16) == FIFO_QUEUE;
2808		}
2809
2810		/**
#	Line 1416 | Line 2816 | public class ForkJoinPool extends Abstra
2816		* @return the number of worker threads
2817		*/
2818		public int getRunningThreadCount() {
2819	<	return workerCounts & RUNNING_COUNT_MASK;
2819	>	int rc = 0;
2820	>	WorkQueue[] ws; WorkQueue w;
2821	>	if ((ws = workQueues) != null) {
2822	>	for (int i = 1; i < ws.length; i += 2) {
2823	>	if ((w = ws[i]) != null && w.isApparentlyUnblocked())
2824	>	++rc;
2825	>	}
2826	>	}
2827	>	return rc;
2828		}
2829
2830		/**
#	Line 1427 | Line 2835 | public class ForkJoinPool extends Abstra
2835		* @return the number of active threads
2836		*/
2837		public int getActiveThreadCount() {
2838	<	return runState & ACTIVE_COUNT_MASK;
2838	>	int r = (config & SMASK) + (int)(ctl >> AC_SHIFT);
2839	>	return (r <= 0) ? 0 : r; // suppress momentarily negative values
2840		}
2841
2842		/**
#	Line 1442 | Line 2851 | public class ForkJoinPool extends Abstra
2851		* @return {@code true} if all threads are currently idle
2852		*/
2853		public boolean isQuiescent() {
2854	<	return (runState & ACTIVE_COUNT_MASK) == 0;
2854	>	return (int)(ctl >> AC_SHIFT) + (config & SMASK) == 0;
2855		}
2856
2857		/**
#	Line 1457 | Line 2866 | public class ForkJoinPool extends Abstra
2866		* @return the number of steals
2867		*/
2868		public long getStealCount() {
2869	<	return stealCount;
2869	>	long count = stealCount;
2870	>	WorkQueue[] ws; WorkQueue w;
2871	>	if ((ws = workQueues) != null) {
2872	>	for (int i = 1; i < ws.length; i += 2) {
2873	>	if ((w = ws[i]) != null)
2874	>	count += w.nsteals;
2875	>	}
2876	>	}
2877	>	return count;
2878		}
2879
2880		/**
#	Line 1472 | Line 2889 | public class ForkJoinPool extends Abstra
2889		*/
2890		public long getQueuedTaskCount() {
2891		long count = 0;
2892	<	ForkJoinWorkerThread[] ws = workers;
2893	<	int nws = ws.length;
2894	<	for (int i = 0; i < nws; ++i) {
2895	<	ForkJoinWorkerThread w = ws[i];
2896	<	if (w != null)
2897	<	count += w.getQueueSize();
2892	>	WorkQueue[] ws; WorkQueue w;
2893	>	if ((ws = workQueues) != null) {
2894	>	for (int i = 1; i < ws.length; i += 2) {
2895	>	if ((w = ws[i]) != null)
2896	>	count += w.queueSize();
2897	>	}
2898		}
2899		return count;
2900		}
2901
2902		/**
2903		* Returns an estimate of the number of tasks submitted to this
2904	<	* pool that have not yet begun executing.  This method takes time
2905	<	* proportional to the number of submissions.
2904	>	* pool that have not yet begun executing.  This method may take
2905	>	* time proportional to the number of submissions.
2906		*
2907		* @return the number of queued submissions
2908		*/
2909		public int getQueuedSubmissionCount() {
2910	<	return submissionQueue.size();
2910	>	int count = 0;
2911	>	WorkQueue[] ws; WorkQueue w;
2912	>	if ((ws = workQueues) != null) {
2913	>	for (int i = 0; i < ws.length; i += 2) {
2914	>	if ((w = ws[i]) != null)
2915	>	count += w.queueSize();
2916	>	}
2917	>	}
2918	>	return count;
2919		}
2920
2921		/**
#	Line 1500 | Line 2925 | public class ForkJoinPool extends Abstra
2925		* @return {@code true} if there are any queued submissions
2926		*/
2927		public boolean hasQueuedSubmissions() {
2928	<	return !submissionQueue.isEmpty();
2928	>	WorkQueue[] ws; WorkQueue w;
2929	>	if ((ws = workQueues) != null) {
2930	>	for (int i = 0; i < ws.length; i += 2) {
2931	>	if ((w = ws[i]) != null && !w.isEmpty())
2932	>	return true;
2933	>	}
2934	>	}
2935	>	return false;
2936		}
2937
2938		/**
#	Line 1511 | Line 2943 | public class ForkJoinPool extends Abstra
2943		* @return the next submission, or {@code null} if none
2944		*/
2945		protected ForkJoinTask<?> pollSubmission() {
2946	<	return submissionQueue.poll();
2946	>	WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2947	>	if ((ws = workQueues) != null) {
2948	>	for (int i = 0; i < ws.length; i += 2) {
2949	>	if ((w = ws[i]) != null && (t = w.poll()) != null)
2950	>	return t;
2951	>	}
2952	>	}
2953	>	return null;
2954		}
2955
2956		/**
#	Line 1532 | Line 2971 | public class ForkJoinPool extends Abstra
2971		* @return the number of elements transferred
2972		*/
2973		protected int drainTasksTo(Collection<? super ForkJoinTask<?>> c) {
1535	–	int n = submissionQueue.drainTo(c);
1536	–	ForkJoinWorkerThread[] ws = workers;
1537	–	int nws = ws.length;
1538	–	for (int i = 0; i < nws; ++i) {
1539	–	ForkJoinWorkerThread w = ws[i];
1540	–	if (w != null)
1541	–	n += w.drainTasksTo(c);
1542	–	}
1543	–	return n;
1544	–	}
1545	–
1546	–	/**
1547	–	* Returns count of total parks by existing workers.
1548	–	* Used during development only since not meaningful to users.
1549	–	*/
1550	–	private int collectParkCount() {
2974		int count = 0;
2975	<	ForkJoinWorkerThread[] ws = workers;
2976	<	int nws = ws.length;
2977	<	for (int i = 0; i < nws; ++i) {
2978	<	ForkJoinWorkerThread w = ws[i];
2979	<	if (w != null)
2980	<	count += w.parkCount;
2975	>	WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2976	>	if ((ws = workQueues) != null) {
2977	>	for (int i = 0; i < ws.length; ++i) {
2978	>	if ((w = ws[i]) != null) {
2979	>	while ((t = w.poll()) != null) {
2980	>	c.add(t);
2981	>	++count;
2982	>	}
2983	>	}
2984	>	}
2985		}
2986		return count;
2987		}
#	Line 1567 | Line 2994 | public class ForkJoinPool extends Abstra
2994		* @return a string identifying this pool, as well as its state
2995		*/
2996		public String toString() {
2997	<	long st = getStealCount();
2998	<	long qt = getQueuedTaskCount();
2999	<	long qs = getQueuedSubmissionCount();
3000	<	int wc = workerCounts;
3001	<	int tc = wc >>> TOTAL_COUNT_SHIFT;
3002	<	int rc = wc & RUNNING_COUNT_MASK;
3003	<	int pc = parallelism;
3004	<	int rs = runState;
3005	<	int ac = rs & ACTIVE_COUNT_MASK;
3006	<	//        int pk = collectParkCount();
2997	>	// Use a single pass through workQueues to collect counts
2998	>	long qt = 0L, qs = 0L; int rc = 0;
2999	>	long st = stealCount;
3000	>	long c = ctl;
3001	>	WorkQueue[] ws; WorkQueue w;
3002	>	if ((ws = workQueues) != null) {
3003	>	for (int i = 0; i < ws.length; ++i) {
3004	>	if ((w = ws[i]) != null) {
3005	>	int size = w.queueSize();
3006	>	if ((i & 1) == 0)
3007	>	qs += size;
3008	>	else {
3009	>	qt += size;
3010	>	st += w.nsteals;
3011	>	if (w.isApparentlyUnblocked())
3012	>	++rc;
3013	>	}
3014	>	}
3015	>	}
3016	>	}
3017	>	int pc = (config & SMASK);
3018	>	int tc = pc + (short)(c >>> TC_SHIFT);
3019	>	int ac = pc + (int)(c >> AC_SHIFT);
3020	>	if (ac < 0) // ignore transient negative
3021	>	ac = 0;
3022	>	String level;
3023	>	if ((c & STOP_BIT) != 0)
3024	>	level = (tc == 0) ? "Terminated" : "Terminating";
3025	>	else
3026	>	level = plock < 0 ? "Shutting down" : "Running";
3027		return super.toString() +
3028	<	"[" + runLevelToString(rs) +
3028	>	"[" + level +
3029		", parallelism = " + pc +
3030		", size = " + tc +
3031		", active = " + ac +
#	Line 1586 | Line 3033 | public class ForkJoinPool extends Abstra
3033		", steals = " + st +
3034		", tasks = " + qt +
3035		", submissions = " + qs +
1589	–	//            ", parks = " + pk +
3036		"]";
3037		}
3038
1593	–	private static String runLevelToString(int s) {
1594	–	return ((s & TERMINATED) != 0 ? "Terminated" :
1595	–	((s & TERMINATING) != 0 ? "Terminating" :
1596	–	((s & SHUTDOWN) != 0 ? "Shutting down" :
1597	–	"Running")));
1598	–	}
1599	–
3039		/**
3040	<	* Initiates an orderly shutdown in which previously submitted
3041	<	* tasks are executed, but no new tasks will be accepted.
3042	<	* Invocation has no additional effect if already shut down.
3043	<	* Tasks that are in the process of being submitted concurrently
3044	<	* during the course of this method may or may not be rejected.
3040	>	* Possibly initiates an orderly shutdown in which previously
3041	>	* submitted tasks are executed, but no new tasks will be
3042	>	* accepted. Invocation has no effect on execution state if this
3043	>	* is the {@link #commonPool}, and no additional effect if
3044	>	* already shut down.  Tasks that are in the process of being
3045	>	* submitted concurrently during the course of this method may or
3046	>	* may not be rejected.
3047		*
3048		* @throws SecurityException if a security manager exists and
3049		*         the caller is not permitted to modify threads
#	Line 1611 | Line 3052 | public class ForkJoinPool extends Abstra
3052		*/
3053		public void shutdown() {
3054		checkPermission();
3055	<	advanceRunLevel(SHUTDOWN);
1615	<	tryTerminate(false);
3055	>	tryTerminate(false, true);
3056		}
3057
3058		/**
3059	<	* Attempts to cancel and/or stop all tasks, and reject all
3060	<	* subsequently submitted tasks.  Tasks that are in the process of
3061	<	* being submitted or executed concurrently during the course of
3062	<	* this method may or may not be rejected. This method cancels
3063	<	* both existing and unexecuted tasks, in order to permit
3064	<	* termination in the presence of task dependencies. So the method
3065	<	* always returns an empty list (unlike the case for some other
3066	<	* Executors).
3059	>	* Possibly attempts to cancel and/or stop all tasks, and reject
3060	>	* all subsequently submitted tasks.  Invocation has no effect on
3061	>	* execution state if this is the {@link #commonPool}, and no
3062	>	* additional effect if already shut down. Otherwise, tasks that
3063	>	* are in the process of being submitted or executed concurrently
3064	>	* during the course of this method may or may not be
3065	>	* rejected. This method cancels both existing and unexecuted
3066	>	* tasks, in order to permit termination in the presence of task
3067	>	* dependencies. So the method always returns an empty list
3068	>	* (unlike the case for some other Executors).
3069		*
3070		* @return an empty list
3071		* @throws SecurityException if a security manager exists and
#	Line 1633 | Line 3075 | public class ForkJoinPool extends Abstra
3075		*/
3076		public List<Runnable> shutdownNow() {
3077		checkPermission();
3078	<	tryTerminate(true);
3078	>	tryTerminate(true, true);
3079		return Collections.emptyList();
3080		}
3081
#	Line 1643 | Line 3085 | public class ForkJoinPool extends Abstra
3085		* @return {@code true} if all tasks have completed following shut down
3086		*/
3087		public boolean isTerminated() {
3088	<	return runState >= TERMINATED;
3088	>	long c = ctl;
3089	>	return ((c & STOP_BIT) != 0L &&
3090	>	(short)(c >>> TC_SHIFT) == -(config & SMASK));
3091		}
3092
3093		/**
#	Line 1651 | Line 3095 | public class ForkJoinPool extends Abstra
3095		* commenced but not yet completed.  This method may be useful for
3096		* debugging. A return of {@code true} reported a sufficient
3097		* period after shutdown may indicate that submitted tasks have
3098	<	* ignored or suppressed interruption, causing this executor not
3099	<	* to properly terminate.
3098	>	* ignored or suppressed interruption, or are waiting for I/O,
3099	>	* causing this executor not to properly terminate. (See the
3100	>	* advisory notes for class {@link ForkJoinTask} stating that
3101	>	* tasks should not normally entail blocking operations.  But if
3102	>	* they do, they must abort them on interrupt.)
3103		*
3104		* @return {@code true} if terminating but not yet terminated
3105		*/
3106		public boolean isTerminating() {
3107	<	return (runState & (TERMINATING|TERMINATED)) == TERMINATING;
3107	>	long c = ctl;
3108	>	return ((c & STOP_BIT) != 0L &&
3109	>	(short)(c >>> TC_SHIFT) != -(config & SMASK));
3110		}
3111
3112		/**
#	Line 1666 | Line 3115 | public class ForkJoinPool extends Abstra
3115		* @return {@code true} if this pool has been shut down
3116		*/
3117		public boolean isShutdown() {
3118	<	return runState >= SHUTDOWN;
3118	>	return plock < 0;
3119		}
3120
3121		/**
3122	<	* Blocks until all tasks have completed execution after a shutdown
3123	<	* request, or the timeout occurs, or the current thread is
3124	<	* interrupted, whichever happens first.
3122	>	* Blocks until all tasks have completed execution after a
3123	>	* shutdown request, or the timeout occurs, or the current thread
3124	>	* is interrupted, whichever happens first. Note that the {@link
3125	>	* #commonPool()} never terminates until program shutdown so
3126	>	* this method will always time out.
3127		*
3128		* @param timeout the maximum time to wait
3129		* @param unit the time unit of the timeout argument
#	Line 1682 | Line 3133 | public class ForkJoinPool extends Abstra
3133		*/
3134		public boolean awaitTermination(long timeout, TimeUnit unit)
3135		throws InterruptedException {
3136	<	try {
3137	<	return termination.awaitAdvanceInterruptibly(0, timeout, unit) > 0;
3138	<	} catch(TimeoutException ex) {
3139	<	return false;
3136	>	long nanos = unit.toNanos(timeout);
3137	>	if (isTerminated())
3138	>	return true;
3139	>	long startTime = System.nanoTime();
3140	>	boolean terminated = false;
3141	>	synchronized (this) {
3142	>	for (long waitTime = nanos, millis = 0L;;) {
3143	>	if (terminated = isTerminated() ||
3144	>	waitTime <= 0L ||
3145	>	(millis = unit.toMillis(waitTime)) <= 0L)
3146	>	break;
3147	>	wait(millis);
3148	>	waitTime = nanos - (System.nanoTime() - startTime);
3149	>	}
3150		}
3151	+	return terminated;
3152		}
3153
3154		/**
3155		* Interface for extending managed parallelism for tasks running
3156		* in {@link ForkJoinPool}s.
3157		*
3158	<	* <p>A {@code ManagedBlocker} provides two methods.
3159	<	* Method {@code isReleasable} must return {@code true} if
3160	<	* blocking is not necessary. Method {@code block} blocks the
3161	<	* current thread if necessary (perhaps internally invoking
3162	<	* {@code isReleasable} before actually blocking).
3158	>	* <p>A {@code ManagedBlocker} provides two methods.  Method
3159	>	* {@code isReleasable} must return {@code true} if blocking is
3160	>	* not necessary. Method {@code block} blocks the current thread
3161	>	* if necessary (perhaps internally invoking {@code isReleasable}
3162	>	* before actually blocking). These actions are performed by any
3163	>	* thread invoking {@link ForkJoinPool#managedBlock}.  The
3164	>	* unusual methods in this API accommodate synchronizers that may,
3165	>	* but don't usually, block for long periods. Similarly, they
3166	>	* allow more efficient internal handling of cases in which
3167	>	* additional workers may be, but usually are not, needed to
3168	>	* ensure sufficient parallelism.  Toward this end,
3169	>	* implementations of method {@code isReleasable} must be amenable
3170	>	* to repeated invocation.
3171		*
3172		* <p>For example, here is a ManagedBlocker based on a
3173		* ReentrantLock:
#	Line 1715 | Line 3185 | public class ForkJoinPool extends Abstra
3185		*     return hasLock || (hasLock = lock.tryLock());
3186		*   }
3187		* }}</pre>
3188	+	*
3189	+	* <p>Here is a class that possibly blocks waiting for an
3190	+	* item on a given queue:
3191	+	*  <pre> {@code
3192	+	* class QueueTaker<E> implements ManagedBlocker {
3193	+	*   final BlockingQueue<E> queue;
3194	+	*   volatile E item = null;
3195	+	*   QueueTaker(BlockingQueue<E> q) { this.queue = q; }
3196	+	*   public boolean block() throws InterruptedException {
3197	+	*     if (item == null)
3198	+	*       item = queue.take();
3199	+	*     return true;
3200	+	*   }
3201	+	*   public boolean isReleasable() {
3202	+	*     return item != null || (item = queue.poll()) != null;
3203	+	*   }
3204	+	*   public E getItem() { // call after pool.managedBlock completes
3205	+	*     return item;
3206	+	*   }
3207	+	* }}</pre>
3208		*/
3209		public static interface ManagedBlocker {
3210		/**
#	Line 1757 | Line 3247 | public class ForkJoinPool extends Abstra
3247		public static void managedBlock(ManagedBlocker blocker)
3248		throws InterruptedException {
3249		Thread t = Thread.currentThread();
3250	<	if (t instanceof ForkJoinWorkerThread)
3251	<	((ForkJoinWorkerThread) t).pool.awaitBlocker(blocker);
3250	>	if (t instanceof ForkJoinWorkerThread) {
3251	>	ForkJoinPool p = ((ForkJoinWorkerThread)t).pool;
3252	>	while (!blocker.isReleasable()) { // variant of helpSignal
3253	>	WorkQueue[] ws; WorkQueue q; int m, u;
3254	>	if ((ws = p.workQueues) != null && (m = ws.length - 1) >= 0) {
3255	>	for (int i = 0; i <= m; ++i) {
3256	>	if (blocker.isReleasable())
3257	>	return;
3258	>	if ((q = ws[i]) != null && q.base - q.top < 0) {
3259	>	p.signalWork(q);
3260	>	if ((u = (int)(p.ctl >>> 32)) >= 0 ||
3261	>	(u >> UAC_SHIFT) >= 0)
3262	>	break;
3263	>	}
3264	>	}
3265	>	}
3266	>	if (p.tryCompensate()) {
3267	>	try {
3268	>	do {} while (!blocker.isReleasable() &&
3269	>	!blocker.block());
3270	>	} finally {
3271	>	p.incrementActiveCount();
3272	>	}
3273	>	break;
3274	>	}
3275	>	}
3276	>	}
3277		else {
3278	<	do {} while (!blocker.isReleasable() && !blocker.block());
3278	>	do {} while (!blocker.isReleasable() &&
3279	>	!blocker.block());
3280		}
3281		}
3282
#	Line 1769 | Line 3285 | public class ForkJoinPool extends Abstra
3285		// implement RunnableFuture.
3286
3287		protected <T> RunnableFuture<T> newTaskFor(Runnable runnable, T value) {
3288	<	return (RunnableFuture<T>) ForkJoinTask.adapt(runnable, value);
3288	>	return new ForkJoinTask.AdaptedRunnable<T>(runnable, value);
3289		}
3290
3291		protected <T> RunnableFuture<T> newTaskFor(Callable<T> callable) {
3292	<	return (RunnableFuture<T>) ForkJoinTask.adapt(callable);
3292	>	return new ForkJoinTask.AdaptedCallable<T>(callable);
3293		}
3294
3295		// Unsafe mechanics
3296	+	private static final sun.misc.Unsafe U;
3297	+	private static final long CTL;
3298	+	private static final long PARKBLOCKER;
3299	+	private static final int ABASE;
3300	+	private static final int ASHIFT;
3301	+	private static final long STEALCOUNT;
3302	+	private static final long PLOCK;
3303	+	private static final long INDEXSEED;
3304	+	private static final long QLOCK;
3305
3306	<	private static final sun.misc.Unsafe UNSAFE = getUnsafe();
3307	<	private static final long workerCountsOffset =
1783	<	objectFieldOffset("workerCounts", ForkJoinPool.class);
1784	<	private static final long runStateOffset =
1785	<	objectFieldOffset("runState", ForkJoinPool.class);
1786	<	private static final long eventCountOffset =
1787	<	objectFieldOffset("eventCount", ForkJoinPool.class);
1788	<	private static final long eventWaitersOffset =
1789	<	objectFieldOffset("eventWaiters",ForkJoinPool.class);
1790	<	private static final long stealCountOffset =
1791	<	objectFieldOffset("stealCount",ForkJoinPool.class);
1792	<
1793	<	private static long objectFieldOffset(String field, Class<?> klazz) {
3306	>	static {
3307	>	int s; // initialize field offsets for CAS etc
3308		try {
3309	<	return UNSAFE.objectFieldOffset(klazz.getDeclaredField(field));
3310	<	} catch (NoSuchFieldException e) {
3311	<	// Convert Exception to corresponding Error
3312	<	NoSuchFieldError error = new NoSuchFieldError(field);
3313	<	error.initCause(e);
3314	<	throw error;
3315	<	}
3309	>	U = getUnsafe();
3310	>	Class<?> k = ForkJoinPool.class;
3311	>	CTL = U.objectFieldOffset
3312	>	(k.getDeclaredField("ctl"));
3313	>	STEALCOUNT = U.objectFieldOffset
3314	>	(k.getDeclaredField("stealCount"));
3315	>	PLOCK = U.objectFieldOffset
3316	>	(k.getDeclaredField("plock"));
3317	>	INDEXSEED = U.objectFieldOffset
3318	>	(k.getDeclaredField("indexSeed"));
3319	>	Class<?> tk = Thread.class;
3320	>	PARKBLOCKER = U.objectFieldOffset
3321	>	(tk.getDeclaredField("parkBlocker"));
3322	>	Class<?> wk = WorkQueue.class;
3323	>	QLOCK = U.objectFieldOffset
3324	>	(wk.getDeclaredField("qlock"));
3325	>	Class<?> ak = ForkJoinTask[].class;
3326	>	ABASE = U.arrayBaseOffset(ak);
3327	>	s = U.arrayIndexScale(ak);
3328	>	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
3329	>	} catch (Exception e) {
3330	>	throw new Error(e);
3331	>	}
3332	>	if ((s & (s-1)) != 0)
3333	>	throw new Error("data type scale not a power of two");
3334	>
3335	>	submitters = new ThreadLocal<Submitter>();
3336	>	ForkJoinWorkerThreadFactory fac = defaultForkJoinWorkerThreadFactory =
3337	>	new DefaultForkJoinWorkerThreadFactory();
3338	>	modifyThreadPermission = new RuntimePermission("modifyThread");
3339	>
3340	>	/*
3341	>	* Establish common pool parameters.  For extra caution,
3342	>	* computations to set up common pool state are here; the
3343	>	* constructor just assigns these values to fields.
3344	>	*/
3345	>
3346	>	int par = 0;
3347	>	Thread.UncaughtExceptionHandler handler = null;
3348	>	try {  // TBD: limit or report ignored exceptions?
3349	>	String pp = System.getProperty
3350	>	("java.util.concurrent.ForkJoinPool.common.parallelism");
3351	>	String hp = System.getProperty
3352	>	("java.util.concurrent.ForkJoinPool.common.exceptionHandler");
3353	>	String fp = System.getProperty
3354	>	("java.util.concurrent.ForkJoinPool.common.threadFactory");
3355	>	if (fp != null)
3356	>	fac = ((ForkJoinWorkerThreadFactory)ClassLoader.
3357	>	getSystemClassLoader().loadClass(fp).newInstance());
3358	>	if (hp != null)
3359	>	handler = ((Thread.UncaughtExceptionHandler)ClassLoader.
3360	>	getSystemClassLoader().loadClass(hp).newInstance());
3361	>	if (pp != null)
3362	>	par = Integer.parseInt(pp);
3363	>	} catch (Exception ignore) {
3364	>	}
3365	>
3366	>	if (par <= 0)
3367	>	par = Runtime.getRuntime().availableProcessors();
3368	>	if (par > MAX_CAP)
3369	>	par = MAX_CAP;
3370	>	commonPoolParallelism = par;
3371	>	long np = (long)(-par); // precompute initial ctl value
3372	>	long ct = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK);
3373	>
3374	>	commonPool = new ForkJoinPool(par, ct, fac, handler);
3375		}
3376
3377		/**
#	Line 1828 | Line 3401 | public class ForkJoinPool extends Abstra
3401		}
3402		}
3403		}
3404	+
3405		}

Diff Legend

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