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 1.53 by dl, Mon Apr 5 15:52:26 2010 UTC vs.
Revision 1.135 by dl, Sun Oct 28 22:36:01 2012 UTC

#	Line 1 | Line 1
1		/*
2		* Written by Doug Lea with assistance from members of JCP JSR-166
3		* Expert Group and released to the public domain, as explained at
4	<	* http://creativecommons.org/licenses/publicdomain
4	>	* http://creativecommons.org/publicdomain/zero/1.0/
5		*/
6
7		package jsr166y;
8
9	–	import java.util.concurrent.*;
10	–
9		import java.util.ArrayList;
10		import java.util.Arrays;
11		import java.util.Collection;
12		import java.util.Collections;
13		import java.util.List;
14	<	import java.util.concurrent.locks.LockSupport;
15	<	import java.util.concurrent.locks.ReentrantLock;
14	>	import java.util.Random;
15	>	import java.util.concurrent.AbstractExecutorService;
16	>	import java.util.concurrent.Callable;
17	>	import java.util.concurrent.ExecutorService;
18	>	import java.util.concurrent.Future;
19	>	import java.util.concurrent.RejectedExecutionException;
20	>	import java.util.concurrent.RunnableFuture;
21	>	import java.util.concurrent.TimeUnit;
22		import java.util.concurrent.atomic.AtomicInteger;
23	<	import java.util.concurrent.CountDownLatch;
23	>	import java.util.concurrent.atomic.AtomicLong;
24	>	import java.util.concurrent.locks.AbstractQueuedSynchronizer;
25	>	import java.util.concurrent.locks.Condition;
26
27		/**
28		* An {@link ExecutorService} for running {@link ForkJoinTask}s.
29		* A {@code ForkJoinPool} provides the entry point for submissions
30	<	* from non-{@code ForkJoinTask}s, as well as management and
30	>	* from non-{@code ForkJoinTask} clients, as well as management and
31		* monitoring operations.
32		*
33		* <p>A {@code ForkJoinPool} differs from other kinds of {@link
34		* ExecutorService} mainly by virtue of employing
35		* <em>work-stealing</em>: all threads in the pool attempt to find and
36	<	* execute subtasks created by other active tasks (eventually blocking
37	<	* waiting for work if none exist). This enables efficient processing
38	<	* when most tasks spawn other subtasks (as do most {@code
39	<	* ForkJoinTask}s). A {@code ForkJoinPool} may also be used for mixed
40	<	* execution of some plain {@code Runnable}- or {@code Callable}-
41	<	* based activities along with {@code ForkJoinTask}s. When setting
42	<	* {@linkplain #setAsyncMode async mode}, a {@code ForkJoinPool} may
43	<	* also be appropriate for use with fine-grained tasks of any form
44	<	* that are never joined. Otherwise, other {@code ExecutorService}
45	<	* implementations are typically more appropriate choices.
36	>	* execute tasks submitted to the pool and/or created by other active
37	>	* tasks (eventually blocking waiting for work if none exist). This
38	>	* enables efficient processing when most tasks spawn other subtasks
39	>	* (as do most {@code ForkJoinTask}s), as well as when many small
40	>	* tasks are submitted to the pool from external clients.  Especially
41	>	* when setting <em>asyncMode</em> to true in constructors, {@code
42	>	* ForkJoinPool}s may also be appropriate for use with event-style
43	>	* tasks that are never joined.
44	>	*
45	>	* <p>A static {@link #commonPool} is available and appropriate for
46	>	* most applications. The common pool is constructed upon first
47	>	* access, or upon usage by any ForkJoinTask that is not explictly
48	>	* submitted to a specified pool. Using the common pool normally
49	>	* reduces resource usage (its threads are slowly reclaimed during
50	>	* periods of non-use, and reinstated upon subsequent use).  The
51	>	* common pool is by default constructed with default parameters, but
52	>	* these may be controlled by setting any or all of the three
53	>	* properties {@code
54	>	* java.util.concurrent.ForkJoinPool.common.{parallelism,
55	>	* threadFactory, exceptionHandler}}.
56		*
57	<	* <p>A {@code ForkJoinPool} is constructed with a given target
58	<	* parallelism level; by default, equal to the number of available
59	<	* processors. Unless configured otherwise via {@link
60	<	* #setMaintainsParallelism}, the pool attempts to maintain this
61	<	* number of active (or available) threads by dynamically adding,
62	<	* suspending, or resuming internal worker threads, even if some tasks
63	<	* are stalled waiting to join others. However, no such adjustments
64	<	* are performed in the face of blocked IO or other unmanaged
65	<	* synchronization. The nested {@link ManagedBlocker} interface
66	<	* enables extension of the kinds of synchronization accommodated.
51	<	* The target parallelism level may also be changed dynamically
52	<	* ({@link #setParallelism}). The total number of threads may be
53	<	* limited using method {@link #setMaximumPoolSize}, in which case it
54	<	* may become possible for the activities of a pool to stall due to
55	<	* the lack of available threads to process new tasks. When the pool
56	<	* is executing tasks, these and other configuration setting methods
57	<	* may only gradually affect actual pool sizes. It is normally best
58	<	* practice to invoke these methods only when the pool is known to be
59	<	* quiescent.
57	>	* <p>For applications that require separate or custom pools, a {@code
58	>	* ForkJoinPool} may be constructed with a given target parallelism
59	>	* level; by default, equal to the number of available processors. The
60	>	* pool attempts to maintain enough active (or available) threads by
61	>	* dynamically adding, suspending, or resuming internal worker
62	>	* threads, even if some tasks are stalled waiting to join
63	>	* others. However, no such adjustments are guaranteed in the face of
64	>	* blocked IO or other unmanaged synchronization. The nested {@link
65	>	* ManagedBlocker} interface enables extension of the kinds of
66	>	* synchronization accommodated.
67		*
68		* <p>In addition to execution and lifecycle control methods, this
69		* class provides status check methods (for example
#	Line 65 | Line 72 | import java.util.concurrent.CountDownLat
72		* {@link #toString} returns indications of pool state in a
73		* convenient form for informal monitoring.
74		*
75	<	* <p><b>Sample Usage.</b> Normally a single {@code ForkJoinPool} is
76	<	* used for all parallel task execution in a program or subsystem.
77	<	* Otherwise, use would not usually outweigh the construction and
78	<	* bookkeeping overhead of creating a large set of threads. For
79	<	* example, a common pool could be used for the {@code SortTasks}
80	<	* illustrated in {@link RecursiveAction}. Because {@code
81	<	* ForkJoinPool} uses threads in {@linkplain java.lang.Thread#isDaemon
82	<	* daemon} mode, there is typically no need to explicitly {@link
83	<	* #shutdown} such a pool upon program exit.
75	>	* <p> As is the case with other ExecutorServices, there are three
76	>	* main task execution methods summarized in the following table.
77	>	* These are designed to be used primarily by clients not already
78	>	* engaged in fork/join computations in the current pool.  The main
79	>	* forms of these methods accept instances of {@code ForkJoinTask},
80	>	* but overloaded forms also allow mixed execution of plain {@code
81	>	* Runnable}- or {@code Callable}- based activities as well.  However,
82	>	* tasks that are already executing in a pool should normally instead
83	>	* use the within-computation forms listed in the table unless using
84	>	* async event-style tasks that are not usually joined, in which case
85	>	* there is little difference among choice of methods.
86		*
87	<	* <pre>
88	<	* static final ForkJoinPool mainPool = new ForkJoinPool();
89	<	* ...
90	<	* public void sort(long[] array) {
91	<	*   mainPool.invoke(new SortTask(array, 0, array.length));
92	<	* }
93	<	* </pre>
87	>	* <table BORDER CELLPADDING=3 CELLSPACING=1>
88	>	*  <tr>
89	>	*    <td></td>
90	>	*    <td ALIGN=CENTER> <b>Call from non-fork/join clients</b></td>
91	>	*    <td ALIGN=CENTER> <b>Call from within fork/join computations</b></td>
92	>	*  </tr>
93	>	*  <tr>
94	>	*    <td> <b>Arrange async execution</td>
95	>	*    <td> {@link #execute(ForkJoinTask)}</td>
96	>	*    <td> {@link ForkJoinTask#fork}</td>
97	>	*  </tr>
98	>	*  <tr>
99	>	*    <td> <b>Await and obtain result</td>
100	>	*    <td> {@link #invoke(ForkJoinTask)}</td>
101	>	*    <td> {@link ForkJoinTask#invoke}</td>
102	>	*  </tr>
103	>	*  <tr>
104	>	*    <td> <b>Arrange exec and obtain Future</td>
105	>	*    <td> {@link #submit(ForkJoinTask)}</td>
106	>	*    <td> {@link ForkJoinTask#fork} (ForkJoinTasks <em>are</em> Futures)</td>
107	>	*  </tr>
108	>	* </table>
109		*
110		* <p><b>Implementation notes</b>: This implementation restricts the
111		* maximum number of running threads to 32767. Attempts to create
#	Line 89 | Line 113 | import java.util.concurrent.CountDownLat
113		* {@code IllegalArgumentException}.
114		*
115		* <p>This implementation rejects submitted tasks (that is, by throwing
116	<	* {@link RejectedExecutionException}) only when the pool is shut down.
116	>	* {@link RejectedExecutionException}) only when the pool is shut down
117	>	* or internal resources have been exhausted.
118		*
119		* @since 1.7
120		* @author Doug Lea
#	Line 99 | Line 124 | public class ForkJoinPool extends Abstra
124		/*
125		* Implementation Overview
126		*
127	<	* This class provides the central bookkeeping and control for a
128	<	* set of worker threads: Submissions from non-FJ threads enter
129	<	* into a submission queue. Workers take these tasks and typically
130	<	* split them into subtasks that may be stolen by other workers.
131	<	* The main work-stealing mechanics implemented in class
132	<	* ForkJoinWorkerThread give first priority to processing tasks
133	<	* from their own queues (LIFO or FIFO, depending on mode), then
134	<	* to randomized FIFO steals of tasks in other worker queues, and
135	<	* lastly to new submissions. These mechanics do not consider
136	<	* affinities, loads, cache localities, etc, so rarely provide the
137	<	* best possible performance on a given machine, but portably
138	<	* provide good throughput by averaging over these factors.
139	<	* (Further, even if we did try to use such information, we do not
140	<	* usually have a basis for exploiting it. For example, some sets
141	<	* of tasks profit from cache affinities, but others are harmed by
142	<	* cache pollution effects.)
127	>	* This class and its nested classes provide the main
128	>	* functionality and control for a set of worker threads:
129	>	* Submissions from non-FJ threads enter into submission queues.
130	>	* Workers take these tasks and typically split them into subtasks
131	>	* that may be stolen by other workers.  Preference rules give
132	>	* first priority to processing tasks from their own queues (LIFO
133	>	* or FIFO, depending on mode), then to randomized FIFO steals of
134	>	* tasks in other queues.
135	>	*
136	>	* WorkQueues
137	>	* ==========
138	>	*
139	>	* Most operations occur within work-stealing queues (in nested
140	>	* class WorkQueue).  These are special forms of Deques that
141	>	* support only three of the four possible end-operations -- push,
142	>	* pop, and poll (aka steal), under the further constraints that
143	>	* push and pop are called only from the owning thread (or, as
144	>	* extended here, under a lock), while poll may be called from
145	>	* other threads.  (If you are unfamiliar with them, you probably
146	>	* want to read Herlihy and Shavit's book "The Art of
147	>	* Multiprocessor programming", chapter 16 describing these in
148	>	* more detail before proceeding.)  The main work-stealing queue
149	>	* design is roughly similar to those in the papers "Dynamic
150	>	* Circular Work-Stealing Deque" by Chase and Lev, SPAA 2005
151	>	* (http://research.sun.com/scalable/pubs/index.html) and
152	>	* "Idempotent work stealing" by Michael, Saraswat, and Vechev,
153	>	* PPoPP 2009 (http://portal.acm.org/citation.cfm?id=1504186).
154	>	* The main differences ultimately stem from GC requirements that
155	>	* we null out taken slots as soon as we can, to maintain as small
156	>	* a footprint as possible even in programs generating huge
157	>	* numbers of tasks. To accomplish this, we shift the CAS
158	>	* arbitrating pop vs poll (steal) from being on the indices
159	>	* ("base" and "top") to the slots themselves.  So, both a
160	>	* successful pop and poll mainly entail a CAS of a slot from
161	>	* non-null to null.  Because we rely on CASes of references, we
162	>	* do not need tag bits on base or top.  They are simple ints as
163	>	* used in any circular array-based queue (see for example
164	>	* ArrayDeque).  Updates to the indices must still be ordered in a
165	>	* way that guarantees that top == base means the queue is empty,
166	>	* but otherwise may err on the side of possibly making the queue
167	>	* appear nonempty when a push, pop, or poll have not fully
168	>	* committed. Note that this means that the poll operation,
169	>	* considered individually, is not wait-free. One thief cannot
170	>	* successfully continue until another in-progress one (or, if
171	>	* previously empty, a push) completes.  However, in the
172	>	* aggregate, we ensure at least probabilistic non-blockingness.
173	>	* If an attempted steal fails, a thief always chooses a different
174	>	* random victim target to try next. So, in order for one thief to
175	>	* progress, it suffices for any in-progress poll or new push on
176	>	* any empty queue to complete. (This is why we normally use
177	>	* method pollAt and its variants that try once at the apparent
178	>	* base index, else consider alternative actions, rather than
179	>	* method poll.)
180	>	*
181	>	* This approach also enables support of a user mode in which local
182	>	* task processing is in FIFO, not LIFO order, simply by using
183	>	* poll rather than pop.  This can be useful in message-passing
184	>	* frameworks in which tasks are never joined.  However neither
185	>	* mode considers affinities, loads, cache localities, etc, so
186	>	* rarely provide the best possible performance on a given
187	>	* machine, but portably provide good throughput by averaging over
188	>	* these factors.  (Further, even if we did try to use such
189	>	* information, we do not usually have a basis for exploiting it.
190	>	* For example, some sets of tasks profit from cache affinities,
191	>	* but others are harmed by cache pollution effects.)
192	>	*
193	>	* WorkQueues are also used in a similar way for tasks submitted
194	>	* to the pool. We cannot mix these tasks in the same queues used
195	>	* for work-stealing (this would contaminate lifo/fifo
196	>	* processing). Instead, we loosely associate submission queues
197	>	* with submitting threads, using a form of hashing.  The
198	>	* ThreadLocal Submitter class contains a value initially used as
199	>	* a hash code for choosing existing queues, but may be randomly
200	>	* repositioned upon contention with other submitters.  In
201	>	* essence, submitters act like workers except that they never
202	>	* take tasks, and they are multiplexed on to a finite number of
203	>	* shared work queues. However, classes are set up so that future
204	>	* extensions could allow submitters to optionally help perform
205	>	* tasks as well. Insertion of tasks in shared mode requires a
206	>	* lock (mainly to protect in the case of resizing) but we use
207	>	* only a simple spinlock (using bits in field runState), because
208	>	* submitters encountering a busy queue move on to try or create
209	>	* other queues -- they block only when creating and registering
210	>	* new queues.
211	>	*
212	>	* Management
213	>	* ==========
214		*
215		* The main throughput advantages of work-stealing stem from
216	<	* decentralized control -- workers mostly steal tasks from each
217	<	* other. We do not want to negate this by creating bottlenecks
218	<	* implementing the management responsibilities of this class. So
219	<	* we use a collection of techniques that avoid, reduce, or cope
220	<	* well with contention. These entail several instances of
221	<	* bit-packing into CASable fields to maintain only the minimally
222	<	* required atomicity. To enable such packing, we restrict maximum
223	<	* parallelism to (1<<15)-1 (enabling twice this to fit into a 16
224	<	* bit field), which is far in excess of normal operating range.
225	<	* Even though updates to some of these bookkeeping fields do
226	<	* sometimes contend with each other, they don't normally
227	<	* cache-contend with updates to others enough to warrant memory
228	<	* padding or isolation. So they are all held as fields of
229	<	* ForkJoinPool objects.  The main capabilities are as follows:
230	<	*
231	<	* 1. Creating and removing workers. Workers are recorded in the
232	<	* "workers" array. This is an array as opposed to some other data
233	<	* structure to support index-based random steals by workers.
234	<	* Updates to the array recording new workers and unrecording
235	<	* terminated ones are protected from each other by a lock
236	<	* (workerLock) but the array is otherwise concurrently readable,
237	<	* and accessed directly by workers. To simplify index-based
216	>	* decentralized control -- workers mostly take tasks from
217	>	* themselves or each other. We cannot negate this in the
218	>	* implementation of other management responsibilities. The main
219	>	* tactic for avoiding bottlenecks is packing nearly all
220	>	* essentially atomic control state into two volatile variables
221	>	* that are by far most often read (not written) as status and
222	>	* consistency checks.
223	>	*
224	>	* Field "ctl" contains 64 bits holding all the information needed
225	>	* to atomically decide to add, inactivate, enqueue (on an event
226	>	* queue), dequeue, and/or re-activate workers.  To enable this
227	>	* packing, we restrict maximum parallelism to (1<<15)-1 (which is
228	>	* far in excess of normal operating range) to allow ids, counts,
229	>	* and their negations (used for thresholding) to fit into 16bit
230	>	* fields.
231	>	*
232	>	* Field "runState" contains 32 bits needed to register and
233	>	* deregister WorkQueues, as well as to enable shutdown. It is
234	>	* only modified under a lock (normally briefly held, but
235	>	* occasionally protecting allocations and resizings) but even
236	>	* when locked remains available to check consistency.
237	>	*
238	>	* Recording WorkQueues.  WorkQueues are recorded in the
239	>	* "workQueues" array that is created upon pool construction and
240	>	* expanded if necessary.  Updates to the array while recording
241	>	* new workers and unrecording terminated ones are protected from
242	>	* each other by a lock but the array is otherwise concurrently
243	>	* readable, and accessed directly.  To simplify index-based
244		* operations, the array size is always a power of two, and all
245	<	* readers must tolerate null slots. Currently, all but the first
246	<	* worker thread creation is on-demand, triggered by task
245	>	* readers must tolerate null slots. Shared (submission) queues
246	>	* are at even indices, worker queues at odd indices. Grouping
247	>	* them together in this way simplifies and speeds up task
248	>	* scanning.
249	>	*
250	>	* All worker thread creation is on-demand, triggered by task
251		* submissions, replacement of terminated workers, and/or
252		* compensation for blocked workers. However, all other support
253	<	* code is set up to work with other policies.
254	<	*
255	<	* 2. Bookkeeping for dynamically adding and removing workers. We
256	<	* maintain a given level of parallelism (or, if
257	<	* maintainsParallelism is false, at least avoid starvation). When
258	<	* some workers are known to be blocked (on joins or via
259	<	* ManagedBlocker), we may create or resume others to take their
260	<	* place until they unblock (see below). Implementing this
261	<	* requires counts of the number of "running" threads (i.e., those
262	<	* that are neither blocked nor artifically suspended) as well as
263	<	* the total number.  These two values are packed into one field,
264	<	* "workerCounts" because we need accurate snapshots when deciding
265	<	* to create, resume or suspend.  To support these decisions,
266	<	* updates must be prospective (not retrospective).  For example,
267	<	* the running count is decremented before blocking by a thread
268	<	* about to block, but incremented by the thread about to unblock
269	<	* it. (In a few cases, these prospective updates may need to be
270	<	* rolled back, for example when deciding to create a new worker
271	<	* but the thread factory fails or returns null. In these cases,
272	<	* we are no worse off wrt other decisions than we would be
273	<	* otherwise.)  Updates to the workerCounts field sometimes
274	<	* transiently encounter a fair amount of contention when join
275	<	* dependencies are such that many threads block or unblock at
276	<	* about the same time. We alleviate this by sometimes bundling
277	<	* updates (for example blocking one thread on join and resuming a
278	<	* spare cancel each other out), and in most other cases
279	<	* performing an alternative action (like releasing waiters and
280	<	* finding spares; see below) as a more productive form of
281	<	* backoff.
282	<	*
283	<	* 3. Maintaining global run state. The run state of the pool
284	<	* consists of a runLevel (SHUTDOWN, TERMINATING, etc) similar to
285	<	* those in other Executor implementations, as well as a count of
286	<	* "active" workers -- those that are, or soon will be, or
287	<	* recently were executing tasks. The runLevel and active count
288	<	* are packed together in order to correctly trigger shutdown and
289	<	* termination. Without care, active counts can be subject to very
290	<	* high contention.  We substantially reduce this contention by
291	<	* relaxing update rules.  A worker must claim active status
292	<	* prospectively, by activating if it sees that a submitted or
293	<	* stealable task exists (it may find after activating that the
294	<	* task no longer exists). It stays active while processing this
295	<	* task (if it exists) and any other local subtasks it produces,
296	<	* until it cannot find any other tasks. It then tries
297	<	* inactivating (see method preStep), but upon update contention
298	<	* instead scans for more tasks, later retrying inactivation if it
299	<	* doesn't find any.
300	<	*
301	<	* 4. Managing idle workers waiting for tasks. We cannot let
302	<	* workers spin indefinitely scanning for tasks when none are
303	<	* available. On the other hand, we must quickly prod them into
304	<	* action when new tasks are submitted or generated.  We
305	<	* park/unpark these idle workers using an event-count scheme.
306	<	* Field eventCount is incremented upon events that may enable
307	<	* workers that previously could not find a task to now find one:
308	<	* Submission of a new task to the pool, or another worker pushing
309	<	* a task onto a previously empty queue.  (We also use this
310	<	* mechanism for termination and reconfiguration actions that
311	<	* require wakeups of idle workers).  Each worker maintains its
312	<	* last known event count, and blocks when a scan for work did not
313	<	* find a task AND its lastEventCount matches the current
314	<	* eventCount. Waiting idle workers are recorded in a variant of
315	<	* Treiber stack headed by field eventWaiters which, when nonzero,
316	<	* encodes the thread index and count awaited for by the worker
317	<	* thread most recently calling eventSync. This thread in turn has
318	<	* a record (field nextEventWaiter) for the next waiting worker.
319	<	* In addition to allowing simpler decisions about need for
320	<	* wakeup, the event count bits in eventWaiters serve the role of
321	<	* tags to avoid ABA errors in Treiber stacks.  To reduce delays
322	<	* in task diffusion, workers not otherwise occupied may invoke
323	<	* method releaseWaiters, that removes and signals (unparks)
324	<	* workers not waiting on current count. To minimize task
325	<	* production stalls associate with signalling, any worker pushing
326	<	* a task on an empty queue invokes the weaker method signalWork,
327	<	* that only releases idle workers until it detects interference
328	<	* by other threads trying to release, and lets them take
329	<	* over. The net effect is a tree-like diffusion of signals, where
330	<	* released threads and possibly others) help with unparks.  To
331	<	* further reduce contention effects a bit, failed CASes to
332	<	* increment field eventCount are tolerated without retries.
333	<	* Conceptually they are merged into the same event, which is OK
334	<	* when their only purpose is to enable workers to scan for work.
335	<	*
336	<	* 5. Managing suspension of extra workers. When a worker is about
337	<	* to block waiting for a join (or via ManagedBlockers), we may
338	<	* create a new thread to maintain parallelism level, or at least
339	<	* avoid starvation (see below). Usually, extra threads are needed
340	<	* for only very short periods, yet join dependencies are such
341	<	* that we sometimes need them in bursts. Rather than create new
342	<	* threads each time this happens, we suspend no-longer-needed
343	<	* extra ones as "spares". For most purposes, we don't distinguish
344	<	* "extra" spare threads from normal "core" threads: On each call
345	<	* to preStep (the only point at which we can do this) a worker
346	<	* checks to see if there are now too many running workers, and if
347	<	* so, suspends itself.  Methods preJoin and doBlock look for
348	<	* suspended threads to resume before considering creating a new
349	<	* replacement. We don't need a special data structure to maintain
350	<	* spares; simply scanning the workers array looking for
351	<	* worker.isSuspended() is fine because the calling thread is
352	<	* otherwise not doing anything useful anyway; we are at least as
353	<	* happy if after locating a spare, the caller doesn't actually
354	<	* block because the join is ready before we try to adjust and
355	<	* compensate.  Note that this is intrinsically racy.  One thread
356	<	* may become a spare at about the same time as another is
357	<	* needlessly being created. We counteract this and related slop
358	<	* in part by requiring resumed spares to immediately recheck (in
359	<	* preStep) to see whether they they should re-suspend. The only
360	<	* effective difference between "extra" and "core" threads is that
361	<	* we allow the "extra" ones to time out and die if they are not
362	<	* resumed within a keep-alive interval of a few seconds. This is
363	<	* implemented mainly within ForkJoinWorkerThread, but requires
364	<	* some coordination (isTrimmed() -- meaning killed while
365	<	* suspended) to correctly maintain pool counts.
366	<	*
367	<	* 6. Deciding when to create new workers. The main dynamic
368	<	* control in this class is deciding when to create extra threads,
369	<	* in methods preJoin and doBlock. We always need to create one
370	<	* when the number of running threads becomes zero. But because
371	<	* blocked joins are typically dependent, we don't necessarily
372	<	* need or want one-to-one replacement. Using a one-to-one
373	<	* compensation rule often leads to enough useless overhead
374	<	* creating, suspending, resuming, and/or killing threads to
375	<	* signficantly degrade throughput.  We use a rule reflecting the
376	<	* idea that, the more spare threads you already have, the more
377	<	* evidence you need to create another one; where "evidence" is
378	<	* expressed as the current deficit -- target minus running
379	<	* threads. To reduce flickering and drift around target values,
380	<	* the relation is quadratic: adding a spare if (dc*dc)>=(sc*pc)
381	<	* (where dc is deficit, sc is number of spare threads and pc is
382	<	* target parallelism.)  This effectively reduces churn at the
383	<	* price of systematically undershooting target parallelism when
384	<	* many threads are blocked.  However, biasing toward undeshooting
385	<	* partially compensates for the above mechanics to suspend extra
386	<	* threads, that normally lead to overshoot because we can only
387	<	* suspend workers in-between top-level actions. It also better
388	<	* copes with the fact that some of the methods in this class tend
389	<	* to never become compiled (but are interpreted), so some
390	<	* components of the entire set of controls might execute many
391	<	* times faster than others. And similarly for cases where the
392	<	* apparent lack of work is just due to GC stalls and other
393	<	* transient system activity.
394	<	*
395	<	* 7. Maintaining other configuration parameters and monitoring
396	<	* statistics. Updates to fields controlling parallelism level,
397	<	* max size, etc can only meaningfully take effect for individual
398	<	* threads upon their next top-level actions; i.e., between
399	<	* stealing/running tasks/submission, which are separated by calls
400	<	* to preStep.  Memory ordering for these (assumed infrequent)
401	<	* reconfiguration calls is ensured by using reads and writes to
402	<	* volatile field workerCounts (that must be read in preStep anyway)
403	<	* as "fences" -- user-level reads are preceded by reads of
404	<	* workCounts, and writes are followed by no-op CAS to
405	<	* workerCounts. The values reported by other management and
406	<	* monitoring methods are either computed on demand, or are kept
407	<	* in fields that are only updated when threads are otherwise
408	<	* idle.
253	>	* code is set up to work with other policies.  To ensure that we
254	>	* do not hold on to worker references that would prevent GC, ALL
255	>	* accesses to workQueues are via indices into the workQueues
256	>	* array (which is one source of some of the messy code
257	>	* constructions here). In essence, the workQueues array serves as
258	>	* a weak reference mechanism. Thus for example the wait queue
259	>	* field of ctl stores indices, not references.  Access to the
260	>	* workQueues in associated methods (for example signalWork) must
261	>	* both index-check and null-check the IDs. All such accesses
262	>	* ignore bad IDs by returning out early from what they are doing,
263	>	* since this can only be associated with termination, in which
264	>	* case it is OK to give up.  All uses of the workQueues array
265	>	* also check that it is non-null (even if previously
266	>	* non-null). This allows nulling during termination, which is
267	>	* currently not necessary, but remains an option for
268	>	* resource-revocation-based shutdown schemes. It also helps
269	>	* reduce JIT issuance of uncommon-trap code, which tends to
270	>	* unnecessarily complicate control flow in some methods.
271	>	*
272	>	* Event Queuing. Unlike HPC work-stealing frameworks, we cannot
273	>	* let workers spin indefinitely scanning for tasks when none can
274	>	* be found immediately, and we cannot start/resume workers unless
275	>	* there appear to be tasks available.  On the other hand, we must
276	>	* quickly prod them into action when new tasks are submitted or
277	>	* generated. In many usages, ramp-up time to activate workers is
278	>	* the main limiting factor in overall performance (this is
279	>	* compounded at program start-up by JIT compilation and
280	>	* allocation). So we try to streamline this as much as possible.
281	>	* We park/unpark workers after placing in an event wait queue
282	>	* when they cannot find work. This "queue" is actually a simple
283	>	* Treiber stack, headed by the "id" field of ctl, plus a 15bit
284	>	* counter value (that reflects the number of times a worker has
285	>	* been inactivated) to avoid ABA effects (we need only as many
286	>	* version numbers as worker threads). Successors are held in
287	>	* field WorkQueue.nextWait.  Queuing deals with several intrinsic
288	>	* races, mainly that a task-producing thread can miss seeing (and
289	>	* signalling) another thread that gave up looking for work but
290	>	* has not yet entered the wait queue. We solve this by requiring
291	>	* a full sweep of all workers (via repeated calls to method
292	>	* scan()) both before and after a newly waiting worker is added
293	>	* to the wait queue. During a rescan, the worker might release
294	>	* some other queued worker rather than itself, which has the same
295	>	* net effect. Because enqueued workers may actually be rescanning
296	>	* rather than waiting, we set and clear the "parker" field of
297	>	* WorkQueues to reduce unnecessary calls to unpark.  (This
298	>	* requires a secondary recheck to avoid missed signals.)  Note
299	>	* the unusual conventions about Thread.interrupts surrounding
300	>	* parking and other blocking: Because interrupts are used solely
301	>	* to alert threads to check termination, which is checked anyway
302	>	* upon blocking, we clear status (using Thread.interrupted)
303	>	* before any call to park, so that park does not immediately
304	>	* return due to status being set via some other unrelated call to
305	>	* interrupt in user code.
306	>	*
307	>	* Signalling.  We create or wake up workers only when there
308	>	* appears to be at least one task they might be able to find and
309	>	* execute.  When a submission is added or another worker adds a
310	>	* task to a queue that previously had fewer than two tasks, they
311	>	* signal waiting workers (or trigger creation of new ones if
312	>	* fewer than the given parallelism level -- see signalWork).
313	>	* These primary signals are buttressed by signals during rescans;
314	>	* together these cover the signals needed in cases when more
315	>	* tasks are pushed but untaken, and improve performance compared
316	>	* to having one thread wake up all workers.
317	>	*
318	>	* Trimming workers. To release resources after periods of lack of
319	>	* use, a worker starting to wait when the pool is quiescent will
320	>	* time out and terminate if the pool has remained quiescent for a
321	>	* given period -- a short period if there are more threads than
322	>	* parallelism, longer as the number of threads decreases. This
323	>	* will slowly propagate, eventually terminating all workers after
324	>	* periods of non-use.
325	>	*
326	>	* Shutdown and Termination. A call to shutdownNow atomically sets
327	>	* a runState bit and then (non-atomically) sets each worker's
328	>	* runState status, cancels all unprocessed tasks, and wakes up
329	>	* all waiting workers.  Detecting whether termination should
330	>	* commence after a non-abrupt shutdown() call requires more work
331	>	* and bookkeeping. We need consensus about quiescence (i.e., that
332	>	* there is no more work). The active count provides a primary
333	>	* indication but non-abrupt shutdown still requires a rechecking
334	>	* scan for any workers that are inactive but not queued.
335	>	*
336	>	* Joining Tasks
337	>	* =============
338	>	*
339	>	* Any of several actions may be taken when one worker is waiting
340	>	* to join a task stolen (or always held) by another.  Because we
341	>	* are multiplexing many tasks on to a pool of workers, we can't
342	>	* just let them block (as in Thread.join).  We also cannot just
343	>	* reassign the joiner's run-time stack with another and replace
344	>	* it later, which would be a form of "continuation", that even if
345	>	* possible is not necessarily a good idea since we sometimes need
346	>	* both an unblocked task and its continuation to progress.
347	>	* Instead we combine two tactics:
348	>	*
349	>	*   Helping: Arranging for the joiner to execute some task that it
350	>	*      would be running if the steal had not occurred.
351	>	*
352	>	*   Compensating: Unless there are already enough live threads,
353	>	*      method tryCompensate() may create or re-activate a spare
354	>	*      thread to compensate for blocked joiners until they unblock.
355	>	*
356	>	* A third form (implemented in tryRemoveAndExec and
357	>	* tryPollForAndExec) amounts to helping a hypothetical
358	>	* compensator: If we can readily tell that a possible action of a
359	>	* compensator is to steal and execute the task being joined, the
360	>	* joining thread can do so directly, without the need for a
361	>	* compensation thread (although at the expense of larger run-time
362	>	* stacks, but the tradeoff is typically worthwhile).
363	>	*
364	>	* The ManagedBlocker extension API can't use helping so relies
365	>	* only on compensation in method awaitBlocker.
366	>	*
367	>	* The algorithm in tryHelpStealer entails a form of "linear"
368	>	* helping: Each worker records (in field currentSteal) the most
369	>	* recent task it stole from some other worker. Plus, it records
370	>	* (in field currentJoin) the task it is currently actively
371	>	* joining. Method tryHelpStealer uses these markers to try to
372	>	* find a worker to help (i.e., steal back a task from and execute
373	>	* it) that could hasten completion of the actively joined task.
374	>	* In essence, the joiner executes a task that would be on its own
375	>	* local deque had the to-be-joined task not been stolen. This may
376	>	* be seen as a conservative variant of the approach in Wagner &
377	>	* Calder "Leapfrogging: a portable technique for implementing
378	>	* efficient futures" SIGPLAN Notices, 1993
379	>	* (http://portal.acm.org/citation.cfm?id=155354). It differs in
380	>	* that: (1) We only maintain dependency links across workers upon
381	>	* steals, rather than use per-task bookkeeping.  This sometimes
382	>	* requires a linear scan of workQueues array to locate stealers,
383	>	* but often doesn't because stealers leave hints (that may become
384	>	* stale/wrong) of where to locate them.  A stealHint is only a
385	>	* hint because a worker might have had multiple steals and the
386	>	* hint records only one of them (usually the most current).
387	>	* Hinting isolates cost to when it is needed, rather than adding
388	>	* to per-task overhead.  (2) It is "shallow", ignoring nesting
389	>	* and potentially cyclic mutual steals.  (3) It is intentionally
390	>	* racy: field currentJoin is updated only while actively joining,
391	>	* which means that we miss links in the chain during long-lived
392	>	* tasks, GC stalls etc (which is OK since blocking in such cases
393	>	* is usually a good idea).  (4) We bound the number of attempts
394	>	* to find work (see MAX_HELP) and fall back to suspending the
395	>	* worker and if necessary replacing it with another.
396	>	*
397	>	* It is impossible to keep exactly the target parallelism number
398	>	* of threads running at any given time.  Determining the
399	>	* existence of conservatively safe helping targets, the
400	>	* availability of already-created spares, and the apparent need
401	>	* to create new spares are all racy, so we rely on multiple
402	>	* retries of each.  Compensation in the apparent absence of
403	>	* helping opportunities is challenging to control on JVMs, where
404	>	* GC and other activities can stall progress of tasks that in
405	>	* turn stall out many other dependent tasks, without us being
406	>	* able to determine whether they will ever require compensation.
407	>	* Even though work-stealing otherwise encounters little
408	>	* degradation in the presence of more threads than cores,
409	>	* aggressively adding new threads in such cases entails risk of
410	>	* unwanted positive feedback control loops in which more threads
411	>	* cause more dependent stalls (as well as delayed progress of
412	>	* unblocked threads to the point that we know they are available)
413	>	* leading to more situations requiring more threads, and so
414	>	* on. This aspect of control can be seen as an (analytically
415	>	* intractable) game with an opponent that may choose the worst
416	>	* (for us) active thread to stall at any time.  We take several
417	>	* precautions to bound losses (and thus bound gains), mainly in
418	>	* methods tryCompensate and awaitJoin: (1) We only try
419	>	* compensation after attempting enough helping steps (measured
420	>	* via counting and timing) that we have already consumed the
421	>	* estimated cost of creating and activating a new thread.  (2) We
422	>	* allow up to 50% of threads to be blocked before initially
423	>	* adding any others, and unless completely saturated, check that
424	>	* some work is available for a new worker before adding. Also, we
425	>	* create up to only 50% more threads until entering a mode that
426	>	* only adds a thread if all others are possibly blocked.  All
427	>	* together, this means that we might be half as fast to react,
428	>	* and create half as many threads as possible in the ideal case,
429	>	* but present vastly fewer anomalies in all other cases compared
430	>	* to both more aggressive and more conservative alternatives.
431		*
432	<	* Beware that there is a lot of representation-level coupling
432	>	* Style notes: There is a lot of representation-level coupling
433		* among classes ForkJoinPool, ForkJoinWorkerThread, and
434	<	* ForkJoinTask.  For example, direct access to "workers" array by
435	<	* workers, and direct access to ForkJoinTask.status by both
436	<	* ForkJoinPool and ForkJoinWorkerThread.  There is little point
437	<	* trying to reduce this, since any associated future changes in
438	<	* representations will need to be accompanied by algorithmic
439	<	* changes anyway.
440	<	*
441	<	* Style notes: There are lots of inline assignments (of form
442	<	* "while ((local = field) != 0)") which are usually the simplest
443	<	* way to ensure read orderings. Also several occurrences of the
444	<	* unusual "do {} while(!cas...)" which is the simplest way to
445	<	* force an update of a CAS'ed variable. There are also a few
446	<	* other coding oddities that help some methods perform reasonably
447	<	* even when interpreted (not compiled).
448	<	*
449	<	* The order of declarations in this file is: (1) statics (2)
450	<	* fields (along with constants used when unpacking some of them)
451	<	* (3) internal control methods (4) callbacks and other support
452	<	* for ForkJoinTask and ForkJoinWorkerThread classes, (5) exported
453	<	* methods (plus a few little helpers).
434	>	* ForkJoinTask.  The fields of WorkQueue maintain data structures
435	>	* managed by ForkJoinPool, so are directly accessed.  There is
436	>	* little point trying to reduce this, since any associated future
437	>	* changes in representations will need to be accompanied by
438	>	* algorithmic changes anyway. Several methods intrinsically
439	>	* sprawl because they must accumulate sets of consistent reads of
440	>	* volatiles held in local variables.  Methods signalWork() and
441	>	* scan() are the main bottlenecks, so are especially heavily
442	>	* micro-optimized/mangled.  There are lots of inline assignments
443	>	* (of form "while ((local = field) != 0)") which are usually the
444	>	* simplest way to ensure the required read orderings (which are
445	>	* sometimes critical). This leads to a "C"-like style of listing
446	>	* declarations of these locals at the heads of methods or blocks.
447	>	* There are several occurrences of the unusual "do {} while
448	>	* (!cas...)"  which is the simplest way to force an update of a
449	>	* CAS'ed variable. There are also other coding oddities that help
450	>	* some methods perform reasonably even when interpreted (not
451	>	* compiled).
452	>	*
453	>	* The order of declarations in this file is:
454	>	* (1) Static utility functions
455	>	* (2) Nested (static) classes
456	>	* (3) Static fields
457	>	* (4) Fields, along with constants used when unpacking some of them
458	>	* (5) Internal control methods
459	>	* (6) Callbacks and other support for ForkJoinTask methods
460	>	* (7) Exported methods
461	>	* (8) Static block initializing statics in minimally dependent order
462		*/
463
464	+	// Static utilities
465	+
466	+	/**
467	+	* If there is a security manager, makes sure caller has
468	+	* permission to modify threads.
469	+	*/
470	+	private static void checkPermission() {
471	+	SecurityManager security = System.getSecurityManager();
472	+	if (security != null)
473	+	security.checkPermission(modifyThreadPermission);
474	+	}
475	+
476	+	// Nested classes
477	+
478		/**
479		* Factory for creating new {@link ForkJoinWorkerThread}s.
480		* A {@code ForkJoinWorkerThreadFactory} must be defined and used
#	Line 345 | Line 495 | public class ForkJoinPool extends Abstra
495		* Default ForkJoinWorkerThreadFactory implementation; creates a
496		* new ForkJoinWorkerThread.
497		*/
498	<	static class  DefaultForkJoinWorkerThreadFactory
498	>	static class DefaultForkJoinWorkerThreadFactory
499		implements ForkJoinWorkerThreadFactory {
500		public ForkJoinWorkerThread newThread(ForkJoinPool pool) {
501		return new ForkJoinWorkerThread(pool);
#	Line 353 | Line 503 | public class ForkJoinPool extends Abstra
503		}
504
505		/**
506	<	* Creates a new ForkJoinWorkerThread. This factory is used unless
507	<	* overridden in ForkJoinPool constructors.
506	>	* A simple non-reentrant lock used for exclusion when managing
507	>	* queues and workers. We use a custom lock so that we can readily
508	>	* probe lock state in constructions that check among alternative
509	>	* actions. The lock is normally only very briefly held, and
510	>	* sometimes treated as a spinlock, but other usages block to
511	>	* reduce overall contention in those cases where locked code
512	>	* bodies perform allocation/resizing.
513	>	*/
514	>	static final class Mutex extends AbstractQueuedSynchronizer {
515	>	public final boolean tryAcquire(int ignore) {
516	>	return compareAndSetState(0, 1);
517	>	}
518	>	public final boolean tryRelease(int ignore) {
519	>	setState(0);
520	>	return true;
521	>	}
522	>	public final void lock() { acquire(0); }
523	>	public final void unlock() { release(0); }
524	>	public final boolean isHeldExclusively() { return getState() == 1; }
525	>	public final Condition newCondition() { return new ConditionObject(); }
526	>	}
527	>
528	>	/**
529	>	* Class for artificial tasks that are used to replace the target
530	>	* of local joins if they are removed from an interior queue slot
531	>	* in WorkQueue.tryRemoveAndExec. We don't need the proxy to
532	>	* actually do anything beyond having a unique identity.
533	>	*/
534	>	static final class EmptyTask extends ForkJoinTask<Void> {
535	>	EmptyTask() { status = ForkJoinTask.NORMAL; } // force done
536	>	public final Void getRawResult() { return null; }
537	>	public final void setRawResult(Void x) {}
538	>	public final boolean exec() { return true; }
539	>	}
540	>
541	>	/**
542	>	* Queues supporting work-stealing as well as external task
543	>	* submission. See above for main rationale and algorithms.
544	>	* Implementation relies heavily on "Unsafe" intrinsics
545	>	* and selective use of "volatile":
546	>	*
547	>	* Field "base" is the index (mod array.length) of the least valid
548	>	* queue slot, which is always the next position to steal (poll)
549	>	* from if nonempty. Reads and writes require volatile orderings
550	>	* but not CAS, because updates are only performed after slot
551	>	* CASes.
552	>	*
553	>	* Field "top" is the index (mod array.length) of the next queue
554	>	* slot to push to or pop from. It is written only by owner thread
555	>	* for push, or under lock for trySharedPush, and accessed by
556	>	* other threads only after reading (volatile) base.  Both top and
557	>	* base are allowed to wrap around on overflow, but (top - base)
558	>	* (or more commonly -(base - top) to force volatile read of base
559	>	* before top) still estimates size.
560	>	*
561	>	* The array slots are read and written using the emulation of
562	>	* volatiles/atomics provided by Unsafe. Insertions must in
563	>	* general use putOrderedObject as a form of releasing store to
564	>	* ensure that all writes to the task object are ordered before
565	>	* its publication in the queue. (Although we can avoid one case
566	>	* of this when locked in trySharedPush.) All removals entail a
567	>	* CAS to null.  The array is always a power of two. To ensure
568	>	* safety of Unsafe array operations, all accesses perform
569	>	* explicit null checks and implicit bounds checks via
570	>	* power-of-two masking.
571	>	*
572	>	* In addition to basic queuing support, this class contains
573	>	* fields described elsewhere to control execution. It turns out
574	>	* to work better memory-layout-wise to include them in this
575	>	* class rather than a separate class.
576	>	*
577	>	* Performance on most platforms is very sensitive to placement of
578	>	* instances of both WorkQueues and their arrays -- we absolutely
579	>	* do not want multiple WorkQueue instances or multiple queue
580	>	* arrays sharing cache lines. (It would be best for queue objects
581	>	* and their arrays to share, but there is nothing available to
582	>	* help arrange that).  Unfortunately, because they are recorded
583	>	* in a common array, WorkQueue instances are often moved to be
584	>	* adjacent by garbage collectors. To reduce impact, we use field
585	>	* padding that works OK on common platforms; this effectively
586	>	* trades off slightly slower average field access for the sake of
587	>	* avoiding really bad worst-case access. (Until better JVM
588	>	* support is in place, this padding is dependent on transient
589	>	* properties of JVM field layout rules.)  We also take care in
590	>	* allocating, sizing and resizing the array. Non-shared queue
591	>	* arrays are initialized (via method growArray) by workers before
592	>	* use. Others are allocated on first use.
593		*/
594	<	public static final ForkJoinWorkerThreadFactory
595	<	defaultForkJoinWorkerThreadFactory =
596	<	new DefaultForkJoinWorkerThreadFactory();
594	>	static final class WorkQueue {
595	>	/**
596	>	* Capacity of work-stealing queue array upon initialization.
597	>	* Must be a power of two; at least 4, but should be larger to
598	>	* reduce or eliminate cacheline sharing among queues.
599	>	* Currently, it is much larger, as a partial workaround for
600	>	* the fact that JVMs often place arrays in locations that
601	>	* share GC bookkeeping (especially cardmarks) such that
602	>	* per-write accesses encounter serious memory contention.
603	>	*/
604	>	static final int INITIAL_QUEUE_CAPACITY = 1 << 13;
605
606	<	/**
607	<	* Permission required for callers of methods that may start or
608	<	* kill threads.
609	<	*/
610	<	private static final RuntimePermission modifyThreadPermission =
611	<	new RuntimePermission("modifyThread");
606	>	/**
607	>	* Maximum size for queue arrays. Must be a power of two less
608	>	* than or equal to 1 << (31 - width of array entry) to ensure
609	>	* lack of wraparound of index calculations, but defined to a
610	>	* value a bit less than this to help users trap runaway
611	>	* programs before saturating systems.
612	>	*/
613	>	static final int MAXIMUM_QUEUE_CAPACITY = 1 << 26; // 64M
614
615	<	/**
616	<	* If there is a security manager, makes sure caller has
617	<	* permission to modify threads.
618	<	*/
619	<	private static void checkPermission() {
620	<	SecurityManager security = System.getSecurityManager();
621	<	if (security != null)
622	<	security.checkPermission(modifyThreadPermission);
623	<	}
615	>	volatile long totalSteals; // cumulative number of steals
616	>	int seed;                  // for random scanning; initialize nonzero
617	>	volatile int eventCount;   // encoded inactivation count; < 0 if inactive
618	>	int nextWait;              // encoded record of next event waiter
619	>	int rescans;               // remaining scans until block
620	>	int nsteals;               // top-level task executions since last idle
621	>	final int mode;            // lifo, fifo, or shared
622	>	int poolIndex;             // index of this queue in pool (or 0)
623	>	int stealHint;             // index of most recent known stealer
624	>	volatile int runState;     // 1: locked, -1: terminate; else 0
625	>	volatile int base;         // index of next slot for poll
626	>	int top;                   // index of next slot for push
627	>	ForkJoinTask<?>[] array;   // the elements (initially unallocated)
628	>	final ForkJoinPool pool;   // the containing pool (may be null)
629	>	final ForkJoinWorkerThread owner; // owning thread or null if shared
630	>	volatile Thread parker;    // == owner during call to park; else null
631	>	volatile ForkJoinTask<?> currentJoin;  // task being joined in awaitJoin
632	>	ForkJoinTask<?> currentSteal; // current non-local task being executed
633	>	// Heuristic padding to ameliorate unfortunate memory placements
634	>	Object p00, p01, p02, p03, p04, p05, p06, p07;
635	>	Object p08, p09, p0a, p0b, p0c, p0d, p0e;
636	>
637	>	WorkQueue(ForkJoinPool pool, ForkJoinWorkerThread owner, int mode) {
638	>	this.mode = mode;
639	>	this.pool = pool;
640	>	this.owner = owner;
641	>	// Place indices in the center of array (that is not yet allocated)
642	>	base = top = INITIAL_QUEUE_CAPACITY >>> 1;
643	>	}
644
645	<	/**
646	<	* Generator for assigning sequence numbers as pool names.
647	<	*/
648	<	private static final AtomicInteger poolNumberGenerator =
649	<	new AtomicInteger();
645	>	/**
646	>	* Returns the approximate number of tasks in the queue.
647	>	*/
648	>	final int queueSize() {
649	>	int n = base - top;       // non-owner callers must read base first
650	>	return (n >= 0) ? 0 : -n; // ignore transient negative
651	>	}
652
653	<	/**
654	<	* Absolute bound for parallelism level. Twice this number must
655	<	* fit into a 16bit field to enable word-packing for some counts.
656	<	*/
657	<	private static final int MAX_THREADS = 0x7fff;
653	>	/**
654	>	* Provides a more accurate estimate of whether this queue has
655	>	* any tasks than does queueSize, by checking whether a
656	>	* near-empty queue has at least one unclaimed task.
657	>	*/
658	>	final boolean isEmpty() {
659	>	ForkJoinTask<?>[] a; int m, s;
660	>	int n = base - (s = top);
661	>	return (n >= 0 ||
662	>	(n == -1 &&
663	>	((a = array) == null ||
664	>	(m = a.length - 1) < 0 ||
665	>	U.getObjectVolatile
666	>	(a, ((m & (s - 1)) << ASHIFT) + ABASE) == null)));
667	>	}
668
669	<	/**
670	<	* Array holding all worker threads in the pool.  Array size must
671	<	* be a power of two.  Updates and replacements are protected by
672	<	* workerLock, but the array is always kept in a consistent enough
673	<	* state to be randomly accessed without locking by workers
674	<	* performing work-stealing, as well as other traversal-based
675	<	* methods in this class. All readers must tolerate that some
676	<	* array slots may be null.
677	<	*/
678	<	volatile ForkJoinWorkerThread[] workers;
669	>	/**
670	>	* Pushes a task. Call only by owner in unshared queues.
671	>	*
672	>	* @param task the task. Caller must ensure non-null.
673	>	* @throw RejectedExecutionException if array cannot be resized
674	>	*/
675	>	final void push(ForkJoinTask<?> task) {
676	>	ForkJoinTask<?>[] a; ForkJoinPool p;
677	>	int s = top, m, n;
678	>	if ((a = array) != null) {    // ignore if queue removed
679	>	U.putOrderedObject
680	>	(a, (((m = a.length - 1) & s) << ASHIFT) + ABASE, task);
681	>	if ((n = (top = s + 1) - base) <= 2) {
682	>	if ((p = pool) != null)
683	>	p.signalWork();
684	>	}
685	>	else if (n >= m)
686	>	growArray(true);
687	>	}
688	>	}
689
690	<	/**
691	<	* Queue for external submissions.
692	<	*/
693	<	private final LinkedTransferQueue<ForkJoinTask<?>> submissionQueue;
690	>	/**
691	>	* Pushes a task if lock is free and array is either big
692	>	* enough or can be resized to be big enough.
693	>	*
694	>	* @param task the task. Caller must ensure non-null.
695	>	* @return true if submitted
696	>	*/
697	>	final boolean trySharedPush(ForkJoinTask<?> task) {
698	>	boolean submitted = false;
699	>	if (runState == 0 && U.compareAndSwapInt(this, RUNSTATE, 0, 1)) {
700	>	ForkJoinTask<?>[] a = array;
701	>	int s = top;
702	>	try {
703	>	if ((a != null && a.length > s + 1 - base) ||
704	>	(a = growArray(false)) != null) { // must presize
705	>	int j = (((a.length - 1) & s) << ASHIFT) + ABASE;
706	>	U.putObject(a, (long)j, task);    // don't need "ordered"
707	>	top = s + 1;
708	>	submitted = true;
709	>	}
710	>	} finally {
711	>	runState = 0;                         // unlock
712	>	}
713	>	}
714	>	return submitted;
715	>	}
716
717	<	/**
718	<	* Lock protecting updates to workers array.
719	<	*/
720	<	private final ReentrantLock workerLock;
717	>	/**
718	>	* Takes next task, if one exists, in LIFO order.  Call only
719	>	* by owner in unshared queues. (We do not have a shared
720	>	* version of this method because it is never needed.)
721	>	*/
722	>	final ForkJoinTask<?> pop() {
723	>	ForkJoinTask<?>[] a; ForkJoinTask<?> t; int m;
724	>	if ((a = array) != null && (m = a.length - 1) >= 0) {
725	>	for (int s; (s = top - 1) - base >= 0;) {
726	>	long j = ((m & s) << ASHIFT) + ABASE;
727	>	if ((t = (ForkJoinTask<?>)U.getObject(a, j)) == null)
728	>	break;
729	>	if (U.compareAndSwapObject(a, j, t, null)) {
730	>	top = s;
731	>	return t;
732	>	}
733	>	}
734	>	}
735	>	return null;
736	>	}
737
738	<	/**
739	<	* Latch released upon termination.
740	<	*/
741	<	private final CountDownLatch terminationLatch;
738	>	/**
739	>	* Takes a task in FIFO order if b is base of queue and a task
740	>	* can be claimed without contention. Specialized versions
741	>	* appear in ForkJoinPool methods scan and tryHelpStealer.
742	>	*/
743	>	final ForkJoinTask<?> pollAt(int b) {
744	>	ForkJoinTask<?> t; ForkJoinTask<?>[] a;
745	>	if ((a = array) != null) {
746	>	int j = (((a.length - 1) & b) << ASHIFT) + ABASE;
747	>	if ((t = (ForkJoinTask<?>)U.getObjectVolatile(a, j)) != null &&
748	>	base == b &&
749	>	U.compareAndSwapObject(a, j, t, null)) {
750	>	base = b + 1;
751	>	return t;
752	>	}
753	>	}
754	>	return null;
755	>	}
756
757	<	/**
758	<	* Creation factory for worker threads.
759	<	*/
760	<	private final ForkJoinWorkerThreadFactory factory;
757	>	/**
758	>	* Takes next task, if one exists, in FIFO order.
759	>	*/
760	>	final ForkJoinTask<?> poll() {
761	>	ForkJoinTask<?>[] a; int b; ForkJoinTask<?> t;
762	>	while ((b = base) - top < 0 && (a = array) != null) {
763	>	int j = (((a.length - 1) & b) << ASHIFT) + ABASE;
764	>	t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
765	>	if (t != null) {
766	>	if (base == b &&
767	>	U.compareAndSwapObject(a, j, t, null)) {
768	>	base = b + 1;
769	>	return t;
770	>	}
771	>	}
772	>	else if (base == b) {
773	>	if (b + 1 == top)
774	>	break;
775	>	Thread.yield(); // wait for lagging update
776	>	}
777	>	}
778	>	return null;
779	>	}
780
781	<	/**
782	<	* Sum of per-thread steal counts, updated only when threads are
783	<	* idle or terminating.
784	<	*/
785	<	private volatile long stealCount;
781	>	/**
782	>	* Takes next task, if one exists, in order specified by mode.
783	>	*/
784	>	final ForkJoinTask<?> nextLocalTask() {
785	>	return mode == 0 ? pop() : poll();
786	>	}
787
788	<	/**
789	<	* Encoded record of top of treiber stack of threads waiting for
790	<	* events. The top 32 bits contain the count being waited for. The
791	<	* bottom word contains one plus the pool index of waiting worker
792	<	* thread.
793	<	*/
794	<	private volatile long eventWaiters;
788	>	/**
789	>	* Returns next task, if one exists, in order specified by mode.
790	>	*/
791	>	final ForkJoinTask<?> peek() {
792	>	ForkJoinTask<?>[] a = array; int m;
793	>	if (a == null || (m = a.length - 1) < 0)
794	>	return null;
795	>	int i = mode == 0 ? top - 1 : base;
796	>	int j = ((i & m) << ASHIFT) + ABASE;
797	>	return (ForkJoinTask<?>)U.getObjectVolatile(a, j);
798	>	}
799
800	<	private static final int  EVENT_COUNT_SHIFT = 32;
801	<	private static final long WAITER_INDEX_MASK = (1L << EVENT_COUNT_SHIFT)-1L;
800	>	/**
801	>	* Pops the given task only if it is at the current top.
802	>	*/
803	>	final boolean tryUnpush(ForkJoinTask<?> t) {
804	>	ForkJoinTask<?>[] a; int s;
805	>	if ((a = array) != null && (s = top) != base &&
806	>	U.compareAndSwapObject
807	>	(a, (((a.length - 1) & --s) << ASHIFT) + ABASE, t, null)) {
808	>	top = s;
809	>	return true;
810	>	}
811	>	return false;
812	>	}
813
814	<	/**
815	<	* A counter for events that may wake up worker threads:
816	<	*   - Submission of a new task to the pool
817	<	*   - A worker pushing a task on an empty queue
818	<	*   - termination and reconfiguration
819	<	*/
820	<	private volatile int eventCount;
814	>	/**
815	>	* Version of tryUnpush for shared queues; called by non-FJ
816	>	* submitters. Conservatively fails to unpush if all workers
817	>	* are active unless there are multiple tasks in queue.
818	>	*/
819	>	final boolean trySharedUnpush(ForkJoinTask<?> task, ForkJoinPool p) {
820	>	boolean success = false;
821	>	if (task != null && top != base && runState == 0 &&
822	>	U.compareAndSwapInt(this, RUNSTATE, 0, 1)) {
823	>	try {
824	>	ForkJoinTask<?>[] a; int n, s;
825	>	if ((a = array) != null && (n = (s = top) - base) > 0 &&
826	>	(n > 1 || p == null || (int)(p.ctl >> AC_SHIFT) < 0)) {
827	>	int j = (((a.length - 1) & --s) << ASHIFT) + ABASE;
828	>	if (U.getObjectVolatile(a, j) == task &&
829	>	U.compareAndSwapObject(a, j, task, null)) {
830	>	top = s;
831	>	success = true;
832	>	}
833	>	}
834	>	} finally {
835	>	runState = 0;                         // unlock
836	>	}
837	>	}
838	>	return success;
839	>	}
840
841	<	/**
842	<	* Lifecycle control. The low word contains the number of workers
843	<	* that are (probably) executing tasks. This value is atomically
844	<	* incremented before a worker gets a task to run, and decremented
845	<	* when worker has no tasks and cannot find any.  Bits 16-18
846	<	* contain runLevel value. When all are zero, the pool is
847	<	* running. Level transitions are monotonic (running -> shutdown
848	<	* -> terminating -> terminated) so each transition adds a bit.
849	<	* These are bundled together to ensure consistent read for
850	<	* termination checks (i.e., that runLevel is at least SHUTDOWN
851	<	* and active threads is zero).
852	<	*/
853	<	private volatile int runState;
841	>	/**
842	>	* Polls the given task only if it is at the current base.
843	>	*/
844	>	final boolean pollFor(ForkJoinTask<?> task) {
845	>	ForkJoinTask<?>[] a; int b;
846	>	if ((b = base) - top < 0 && (a = array) != null) {
847	>	int j = (((a.length - 1) & b) << ASHIFT) + ABASE;
848	>	if (U.getObjectVolatile(a, j) == task && base == b &&
849	>	U.compareAndSwapObject(a, j, task, null)) {
850	>	base = b + 1;
851	>	return true;
852	>	}
853	>	}
854	>	return false;
855	>	}
856
857	<	// Note: The order among run level values matters.
858	<	private static final int RUNLEVEL_SHIFT     = 16;
859	<	private static final int SHUTDOWN           = 1 << RUNLEVEL_SHIFT;
860	<	private static final int TERMINATING        = 1 << (RUNLEVEL_SHIFT + 1);
861	<	private static final int TERMINATED         = 1 << (RUNLEVEL_SHIFT + 2);
862	<	private static final int ACTIVE_COUNT_MASK  = (1 << RUNLEVEL_SHIFT) - 1;
863	<	private static final int ONE_ACTIVE         = 1; // active update delta
857	>	/**
858	>	* Initializes or doubles the capacity of array. Call either
859	>	* by owner or with lock held -- it is OK for base, but not
860	>	* top, to move while resizings are in progress.
861	>	*
862	>	* @param rejectOnFailure if true, throw exception if capacity
863	>	* exceeded (relayed ultimately to user); else return null.
864	>	*/
865	>	final ForkJoinTask<?>[] growArray(boolean rejectOnFailure) {
866	>	ForkJoinTask<?>[] oldA = array;
867	>	int size = oldA != null ? oldA.length << 1 : INITIAL_QUEUE_CAPACITY;
868	>	if (size <= MAXIMUM_QUEUE_CAPACITY) {
869	>	int oldMask, t, b;
870	>	ForkJoinTask<?>[] a = array = new ForkJoinTask<?>[size];
871	>	if (oldA != null && (oldMask = oldA.length - 1) >= 0 &&
872	>	(t = top) - (b = base) > 0) {
873	>	int mask = size - 1;
874	>	do {
875	>	ForkJoinTask<?> x;
876	>	int oldj = ((b & oldMask) << ASHIFT) + ABASE;
877	>	int j    = ((b &    mask) << ASHIFT) + ABASE;
878	>	x = (ForkJoinTask<?>)U.getObjectVolatile(oldA, oldj);
879	>	if (x != null &&
880	>	U.compareAndSwapObject(oldA, oldj, x, null))
881	>	U.putObjectVolatile(a, j, x);
882	>	} while (++b != t);
883	>	}
884	>	return a;
885	>	}
886	>	else if (!rejectOnFailure)
887	>	return null;
888	>	else
889	>	throw new RejectedExecutionException("Queue capacity exceeded");
890	>	}
891
892	<	/**
893	<	* Holds number of total (i.e., created and not yet terminated)
894	<	* and running (i.e., not blocked on joins or other managed sync)
895	<	* threads, packed together to ensure consistent snapshot when
896	<	* making decisions about creating and suspending spare
897	<	* threads. Updated only by CAS. Note that adding a new worker
898	<	* requires incrementing both counts, since workers start off in
899	<	* running state.  This field is also used for memory-fencing
900	<	* configuration parameters.
479	<	*/
480	<	private volatile int workerCounts;
892	>	/**
893	>	* Removes and cancels all known tasks, ignoring any exceptions.
894	>	*/
895	>	final void cancelAll() {
896	>	ForkJoinTask.cancelIgnoringExceptions(currentJoin);
897	>	ForkJoinTask.cancelIgnoringExceptions(currentSteal);
898	>	for (ForkJoinTask<?> t; (t = poll()) != null; )
899	>	ForkJoinTask.cancelIgnoringExceptions(t);
900	>	}
901
902	<	private static final int TOTAL_COUNT_SHIFT  = 16;
903	<	private static final int RUNNING_COUNT_MASK = (1 << TOTAL_COUNT_SHIFT) - 1;
904	<	private static final int ONE_RUNNING        = 1;
905	<	private static final int ONE_TOTAL          = 1 << TOTAL_COUNT_SHIFT;
902	>	/**
903	>	* Computes next value for random probes.  Scans don't require
904	>	* a very high quality generator, but also not a crummy one.
905	>	* Marsaglia xor-shift is cheap and works well enough.  Note:
906	>	* This is manually inlined in its usages in ForkJoinPool to
907	>	* avoid writes inside busy scan loops.
908	>	*/
909	>	final int nextSeed() {
910	>	int r = seed;
911	>	r ^= r << 13;
912	>	r ^= r >>> 17;
913	>	return seed = r ^= r << 5;
914	>	}
915
916	<	/*
917	<	* Fields parallelism. maxPoolSize, locallyFifo,
918	<	* maintainsParallelism, and ueh are non-volatile, but external
919	<	* reads/writes use workerCount fences to ensure visability.
920	<	*/
916	>	// Execution methods
917	>
918	>	/**
919	>	* Pops and runs tasks until empty.
920	>	*/
921	>	private void popAndExecAll() {
922	>	// A bit faster than repeated pop calls
923	>	ForkJoinTask<?>[] a; int m, s; long j; ForkJoinTask<?> t;
924	>	while ((a = array) != null && (m = a.length - 1) >= 0 &&
925	>	(s = top - 1) - base >= 0 &&
926	>	(t = ((ForkJoinTask<?>)
927	>	U.getObject(a, j = ((m & s) << ASHIFT) + ABASE)))
928	>	!= null) {
929	>	if (U.compareAndSwapObject(a, j, t, null)) {
930	>	top = s;
931	>	t.doExec();
932	>	}
933	>	}
934	>	}
935	>
936	>	/**
937	>	* Polls and runs tasks until empty.
938	>	*/
939	>	private void pollAndExecAll() {
940	>	for (ForkJoinTask<?> t; (t = poll()) != null;)
941	>	t.doExec();
942	>	}
943	>
944	>	/**
945	>	* If present, removes from queue and executes the given task, or
946	>	* any other cancelled task. Returns (true) immediately on any CAS
947	>	* or consistency check failure so caller can retry.
948	>	*
949	>	* @return 0 if no progress can be made, else positive
950	>	* (this unusual convention simplifies use with tryHelpStealer.)
951	>	*/
952	>	final int tryRemoveAndExec(ForkJoinTask<?> task) {
953	>	int stat = 1;
954	>	boolean removed = false, empty = true;
955	>	ForkJoinTask<?>[] a; int m, s, b, n;
956	>	if ((a = array) != null && (m = a.length - 1) >= 0 &&
957	>	(n = (s = top) - (b = base)) > 0) {
958	>	for (ForkJoinTask<?> t;;) {           // traverse from s to b
959	>	int j = ((--s & m) << ASHIFT) + ABASE;
960	>	t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
961	>	if (t == null)                    // inconsistent length
962	>	break;
963	>	else if (t == task) {
964	>	if (s + 1 == top) {           // pop
965	>	if (!U.compareAndSwapObject(a, j, task, null))
966	>	break;
967	>	top = s;
968	>	removed = true;
969	>	}
970	>	else if (base == b)           // replace with proxy
971	>	removed = U.compareAndSwapObject(a, j, task,
972	>	new EmptyTask());
973	>	break;
974	>	}
975	>	else if (t.status >= 0)
976	>	empty = false;
977	>	else if (s + 1 == top) {          // pop and throw away
978	>	if (U.compareAndSwapObject(a, j, t, null))
979	>	top = s;
980	>	break;
981	>	}
982	>	if (--n == 0) {
983	>	if (!empty && base == b)
984	>	stat = 0;
985	>	break;
986	>	}
987	>	}
988	>	}
989	>	if (removed)
990	>	task.doExec();
991	>	return stat;
992	>	}
993	>
994	>	/**
995	>	* Executes a top-level task and any local tasks remaining
996	>	* after execution.
997	>	*/
998	>	final void runTask(ForkJoinTask<?> t) {
999	>	if (t != null) {
1000	>	currentSteal = t;
1001	>	t.doExec();
1002	>	if (top != base) {       // process remaining local tasks
1003	>	if (mode == 0)
1004	>	popAndExecAll();
1005	>	else
1006	>	pollAndExecAll();
1007	>	}
1008	>	++nsteals;
1009	>	currentSteal = null;
1010	>	}
1011	>	}
1012	>
1013	>	/**
1014	>	* Executes a non-top-level (stolen) task.
1015	>	*/
1016	>	final void runSubtask(ForkJoinTask<?> t) {
1017	>	if (t != null) {
1018	>	ForkJoinTask<?> ps = currentSteal;
1019	>	currentSteal = t;
1020	>	t.doExec();
1021	>	currentSteal = ps;
1022	>	}
1023	>	}
1024	>
1025	>	/**
1026	>	* Returns true if owned and not known to be blocked.
1027	>	*/
1028	>	final boolean isApparentlyUnblocked() {
1029	>	Thread wt; Thread.State s;
1030	>	return (eventCount >= 0 &&
1031	>	(wt = owner) != null &&
1032	>	(s = wt.getState()) != Thread.State.BLOCKED &&
1033	>	s != Thread.State.WAITING &&
1034	>	s != Thread.State.TIMED_WAITING);
1035	>	}
1036	>
1037	>	/**
1038	>	* If this owned and is not already interrupted, try to
1039	>	* interrupt and/or unpark, ignoring exceptions.
1040	>	*/
1041	>	final void interruptOwner() {
1042	>	Thread wt, p;
1043	>	if ((wt = owner) != null && !wt.isInterrupted()) {
1044	>	try {
1045	>	wt.interrupt();
1046	>	} catch (SecurityException ignore) {
1047	>	}
1048	>	}
1049	>	if ((p = parker) != null)
1050	>	U.unpark(p);
1051	>	}
1052	>
1053	>	// Unsafe mechanics
1054	>	private static final sun.misc.Unsafe U;
1055	>	private static final long RUNSTATE;
1056	>	private static final int ABASE;
1057	>	private static final int ASHIFT;
1058	>	static {
1059	>	int s;
1060	>	try {
1061	>	U = getUnsafe();
1062	>	Class<?> k = WorkQueue.class;
1063	>	Class<?> ak = ForkJoinTask[].class;
1064	>	RUNSTATE = U.objectFieldOffset
1065	>	(k.getDeclaredField("runState"));
1066	>	ABASE = U.arrayBaseOffset(ak);
1067	>	s = U.arrayIndexScale(ak);
1068	>	} catch (Exception e) {
1069	>	throw new Error(e);
1070	>	}
1071	>	if ((s & (s-1)) != 0)
1072	>	throw new Error("data type scale not a power of two");
1073	>	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
1074	>	}
1075	>	}
1076
1077		/**
1078	<	* The target parallelism level.
1078	>	* Per-thread records for threads that submit to pools. Currently
1079	>	* holds only pseudo-random seed / index that is used to choose
1080	>	* submission queues in method doSubmit. In the future, this may
1081	>	* also incorporate a means to implement different task rejection
1082	>	* and resubmission policies.
1083	>	*
1084	>	* Seeds for submitters and workers/workQueues work in basically
1085	>	* the same way but are initialized and updated using slightly
1086	>	* different mechanics. Both are initialized using the same
1087	>	* approach as in class ThreadLocal, where successive values are
1088	>	* unlikely to collide with previous values. This is done during
1089	>	* registration for workers, but requires a separate AtomicInteger
1090	>	* for submitters. Seeds are then randomly modified upon
1091	>	* collisions using xorshifts, which requires a non-zero seed.
1092		*/
1093	<	private int parallelism;
1093	>	static final class Submitter {
1094	>	int seed;
1095	>	Submitter() {
1096	>	int s = nextSubmitterSeed.getAndAdd(SEED_INCREMENT);
1097	>	seed = (s == 0) ? 1 : s; // ensure non-zero
1098	>	}
1099	>	}
1100	>
1101	>	/** ThreadLocal class for Submitters */
1102	>	static final class ThreadSubmitter extends ThreadLocal<Submitter> {
1103	>	public Submitter initialValue() { return new Submitter(); }
1104	>	}
1105	>
1106	>	// static fields (initialized in static initializer below)
1107
1108		/**
1109	<	* The maximum allowed pool size.
1109	>	* Creates a new ForkJoinWorkerThread. This factory is used unless
1110	>	* overridden in ForkJoinPool constructors.
1111		*/
1112	<	private int maxPoolSize;
1112	>	public static final ForkJoinWorkerThreadFactory
1113	>	defaultForkJoinWorkerThreadFactory;
1114
1115		/**
1116	<	* True if use local fifo, not default lifo, for local polling
505	<	* Replicated by ForkJoinWorkerThreads
1116	>	* Generator for assigning sequence numbers as pool names.
1117		*/
1118	<	private boolean locallyFifo;
1118	>	private static final AtomicInteger poolNumberGenerator;
1119
1120		/**
1121	<	* Controls whether to add spares to maintain parallelism
1121	>	* Generator for initial hashes/seeds for submitters. Accessed by
1122	>	* Submitter class constructor.
1123		*/
1124	<	private boolean maintainsParallelism;
1124	>	static final AtomicInteger nextSubmitterSeed;
1125
1126		/**
1127	<	* The uncaught exception handler used when any worker
1128	<	* abruptly terminates
1127	>	* Permission required for callers of methods that may start or
1128	>	* kill threads.
1129		*/
1130	<	private Thread.UncaughtExceptionHandler ueh;
1130	>	private static final RuntimePermission modifyThreadPermission;
1131
1132		/**
1133	<	* Pool number, just for assigning useful names to worker threads
1133	>	* Per-thread submission bookkeeping. Shared across all pools
1134	>	* to reduce ThreadLocal pollution and because random motion
1135	>	* to avoid contention in one pool is likely to hold for others.
1136		*/
1137	<	private final int poolNumber;
1137	>	private static final ThreadSubmitter submitters;
1138	>
1139	>	/** Common default pool */
1140	>	static volatile ForkJoinPool commonPool;
1141	>
1142	>	// commonPool construction parameters
1143	>	private static final String propPrefix =
1144	>	"java.util.concurrent.ForkJoinPool.common.";
1145	>	private static final Thread.UncaughtExceptionHandler commonPoolUEH;
1146	>	private static final ForkJoinWorkerThreadFactory commonPoolFactory;
1147	>	static final int commonPoolParallelism;
1148
1149	<	// utilities for updating fields
1149	>	/** Static initialization lock */
1150	>	private static final Mutex initializationLock;
1151	>
1152	>	// static constants
1153
1154		/**
1155	<	* Adds delta to running count.  Used mainly by ForkJoinTask.
1156	<	*
1157	<	* @param delta the number to add
1155	>	* Initial timeout value (in nanoseconds) for the tread triggering
1156	>	* quiescence to park waiting for new work. On timeout, the thread
1157	>	* will instead try to shrink the number of workers.
1158		*/
1159	<	final void updateRunningCount(int delta) {
533	<	int wc;
534	<	do {} while (!UNSAFE.compareAndSwapInt(this, workerCountsOffset,
535	<	wc = workerCounts,
536	<	wc + delta));
537	<	}
1159	>	private static final long IDLE_TIMEOUT      = 1000L * 1000L * 1000L; // 1sec
1160
1161		/**
1162	<	* Write fence for user modifications of pool parameters
541	<	* (parallelism. etc).  Note that it doesn't matter if CAS fails.
1162	>	* Timeout value when there are more threads than parallelism level
1163		*/
1164	<	private void workerCountWriteFence() {
544	<	int wc;
545	<	UNSAFE.compareAndSwapInt(this, workerCountsOffset,
546	<	wc = workerCounts, wc);
547	<	}
1164	>	private static final long FAST_IDLE_TIMEOUT =  100L * 1000L * 1000L;
1165
1166		/**
1167	<	* Read fence for external reads of pool parameters
1168	<	* (parallelism. maxPoolSize, etc).
1167	>	* The maximum stolen->joining link depth allowed in method
1168	>	* tryHelpStealer.  Must be a power of two. This value also
1169	>	* controls the maximum number of times to try to help join a task
1170	>	* without any apparent progress or change in pool state before
1171	>	* giving up and blocking (see awaitJoin).  Depths for legitimate
1172	>	* chains are unbounded, but we use a fixed constant to avoid
1173	>	* (otherwise unchecked) cycles and to bound staleness of
1174	>	* traversal parameters at the expense of sometimes blocking when
1175	>	* we could be helping.
1176		*/
1177	<	private void workerCountReadFence() {
554	<	int ignore = workerCounts;
555	<	}
1177	>	private static final int MAX_HELP = 64;
1178
1179		/**
1180	<	* Tries incrementing active count; fails on contention.
1181	<	* Called by workers before executing tasks.
1182	<	*
1183	<	* @return true on success
1180	>	* Secondary time-based bound (in nanosecs) for helping attempts
1181	>	* before trying compensated blocking in awaitJoin. Used in
1182	>	* conjunction with MAX_HELP to reduce variance due to different
1183	>	* polling rates associated with different helping options. The
1184	>	* value should roughly approximate the time required to create
1185	>	* and/or activate a worker thread.
1186		*/
1187	<	final boolean tryIncrementActiveCount() {
564	<	int c;
565	<	return UNSAFE.compareAndSwapInt(this, runStateOffset,
566	<	c = runState, c + ONE_ACTIVE);
567	<	}
1187	>	private static final long COMPENSATION_DELAY = 1L << 18; // ~0.25 millisec
1188
1189		/**
1190	<	* Tries decrementing active count; fails on contention.
1191	<	* Called when workers cannot find tasks to run.
1190	>	* Increment for seed generators. See class ThreadLocal for
1191	>	* explanation.
1192		*/
1193	<	final boolean tryDecrementActiveCount() {
574	<	int c;
575	<	return UNSAFE.compareAndSwapInt(this, runStateOffset,
576	<	c = runState, c - ONE_ACTIVE);
577	<	}
1193	>	private static final int SEED_INCREMENT = 0x61c88647;
1194
1195		/**
1196	<	* Advances to at least the given level. Returns true if not
1197	<	* already in at least the given level.
1196	>	* Bits and masks for control variables
1197	>	*
1198	>	* Field ctl is a long packed with:
1199	>	* AC: Number of active running workers minus target parallelism (16 bits)
1200	>	* TC: Number of total workers minus target parallelism (16 bits)
1201	>	* ST: true if pool is terminating (1 bit)
1202	>	* EC: the wait count of top waiting thread (15 bits)
1203	>	* ID: poolIndex of top of Treiber stack of waiters (16 bits)
1204	>	*
1205	>	* When convenient, we can extract the upper 32 bits of counts and
1206	>	* the lower 32 bits of queue state, u = (int)(ctl >>> 32) and e =
1207	>	* (int)ctl.  The ec field is never accessed alone, but always
1208	>	* together with id and st. The offsets of counts by the target
1209	>	* parallelism and the positionings of fields makes it possible to
1210	>	* perform the most common checks via sign tests of fields: When
1211	>	* ac is negative, there are not enough active workers, when tc is
1212	>	* negative, there are not enough total workers, and when e is
1213	>	* negative, the pool is terminating.  To deal with these possibly
1214	>	* negative fields, we use casts in and out of "short" and/or
1215	>	* signed shifts to maintain signedness.
1216	>	*
1217	>	* When a thread is queued (inactivated), its eventCount field is
1218	>	* set negative, which is the only way to tell if a worker is
1219	>	* prevented from executing tasks, even though it must continue to
1220	>	* scan for them to avoid queuing races. Note however that
1221	>	* eventCount updates lag releases so usage requires care.
1222	>	*
1223	>	* Field runState is an int packed with:
1224	>	* SHUTDOWN: true if shutdown is enabled (1 bit)
1225	>	* SEQ:  a sequence number updated upon (de)registering workers (30 bits)
1226	>	* INIT: set true after workQueues array construction (1 bit)
1227	>	*
1228	>	* The sequence number enables simple consistency checks:
1229	>	* Staleness of read-only operations on the workQueues array can
1230	>	* be checked by comparing runState before vs after the reads.
1231		*/
583	–	private boolean advanceRunLevel(int level) {
584	–	for (;;) {
585	–	int s = runState;
586	–	if ((s & level) != 0)
587	–	return false;
588	–	if (UNSAFE.compareAndSwapInt(this, runStateOffset, s, s | level))
589	–	return true;
590	–	}
591	–	}
1232
1233	<	// workers array maintenance
1233	>	// bit positions/shifts for fields
1234	>	private static final int  AC_SHIFT   = 48;
1235	>	private static final int  TC_SHIFT   = 32;
1236	>	private static final int  ST_SHIFT   = 31;
1237	>	private static final int  EC_SHIFT   = 16;
1238
1239	<	/**
1240	<	* Records and returns a workers array index for new worker.
1241	<	*/
1242	<	private int recordWorker(ForkJoinWorkerThread w) {
1243	<	// Try using slot totalCount-1. If not available, scan and/or resize
1244	<	int k = (workerCounts >>> TOTAL_COUNT_SHIFT) - 1;
1245	<	final ReentrantLock lock = this.workerLock;
1246	<	lock.lock();
1239	>	// bounds
1240	>	private static final int  SMASK      = 0xffff;  // short bits
1241	>	private static final int  MAX_CAP    = 0x7fff;  // max #workers - 1
1242	>	private static final int  SQMASK     = 0xfffe;  // even short bits
1243	>	private static final int  SHORT_SIGN = 1 << 15;
1244	>	private static final int  INT_SIGN   = 1 << 31;
1245	>
1246	>	// masks
1247	>	private static final long STOP_BIT   = 0x0001L << ST_SHIFT;
1248	>	private static final long AC_MASK    = ((long)SMASK) << AC_SHIFT;
1249	>	private static final long TC_MASK    = ((long)SMASK) << TC_SHIFT;
1250	>
1251	>	// units for incrementing and decrementing
1252	>	private static final long TC_UNIT    = 1L << TC_SHIFT;
1253	>	private static final long AC_UNIT    = 1L << AC_SHIFT;
1254	>
1255	>	// masks and units for dealing with u = (int)(ctl >>> 32)
1256	>	private static final int  UAC_SHIFT  = AC_SHIFT - 32;
1257	>	private static final int  UTC_SHIFT  = TC_SHIFT - 32;
1258	>	private static final int  UAC_MASK   = SMASK << UAC_SHIFT;
1259	>	private static final int  UTC_MASK   = SMASK << UTC_SHIFT;
1260	>	private static final int  UAC_UNIT   = 1 << UAC_SHIFT;
1261	>	private static final int  UTC_UNIT   = 1 << UTC_SHIFT;
1262	>
1263	>	// masks and units for dealing with e = (int)ctl
1264	>	private static final int E_MASK      = 0x7fffffff; // no STOP_BIT
1265	>	private static final int E_SEQ       = 1 << EC_SHIFT;
1266	>
1267	>	// runState bits
1268	>	private static final int SHUTDOWN    = 1 << 31;
1269	>
1270	>	// access mode for WorkQueue
1271	>	static final int LIFO_QUEUE          =  0;
1272	>	static final int FIFO_QUEUE          =  1;
1273	>	static final int SHARED_QUEUE        = -1;
1274	>
1275	>	// Instance fields
1276	>
1277	>	/*
1278	>	* Field layout order in this class tends to matter more than one
1279	>	* would like. Runtime layout order is only loosely related to
1280	>	* declaration order and may differ across JVMs, but the following
1281	>	* empirically works OK on current JVMs.
1282	>	*/
1283	>
1284	>	volatile long ctl;                         // main pool control
1285	>	final int parallelism;                     // parallelism level
1286	>	final int localMode;                       // per-worker scheduling mode
1287	>	final int submitMask;                      // submit queue index bound
1288	>	int nextSeed;                              // for initializing worker seeds
1289	>	volatile int runState;                     // shutdown status and seq
1290	>	WorkQueue[] workQueues;                    // main registry
1291	>	final Mutex lock;                          // for registration
1292	>	final Condition termination;               // for awaitTermination
1293	>	final ForkJoinWorkerThreadFactory factory; // factory for new workers
1294	>	final Thread.UncaughtExceptionHandler ueh; // per-worker UEH
1295	>	final AtomicLong stealCount;               // collect counts when terminated
1296	>	final AtomicInteger nextWorkerNumber;      // to create worker name string
1297	>	String workerNamePrefix;                   // to create worker name string
1298	>
1299	>	//  Creating, registering, and deregistering workers
1300	>
1301	>	/**
1302	>	* Tries to create and start a worker
1303	>	*/
1304	>	private void addWorker() {
1305	>	Throwable ex = null;
1306	>	ForkJoinWorkerThread wt = null;
1307		try {
1308	<	ForkJoinWorkerThread[] ws = workers;
1309	<	int len = ws.length;
1310	<	if (k < 0 || k >= len || ws[k] != null) {
607	<	for (k = 0; k < len && ws[k] != null; ++k)
608	<	;
609	<	if (k == len)
610	<	ws = Arrays.copyOf(ws, len << 1);
1308	>	if ((wt = factory.newThread(this)) != null) {
1309	>	wt.start();
1310	>	return;
1311		}
1312	<	ws[k] = w;
1313	<	workers = ws; // volatile array write ensures slot visibility
614	<	} finally {
615	<	lock.unlock();
1312	>	} catch (Throwable e) {
1313	>	ex = e;
1314		}
1315	<	return k;
1315	>	deregisterWorker(wt, ex); // adjust counts etc on failure
1316	>	}
1317	>
1318	>	/**
1319	>	* Callback from ForkJoinWorkerThread constructor to assign a
1320	>	* public name. This must be separate from registerWorker because
1321	>	* it is called during the "super" constructor call in
1322	>	* ForkJoinWorkerThread.
1323	>	*/
1324	>	final String nextWorkerName() {
1325	>	return workerNamePrefix.concat
1326	>	(Integer.toString(nextWorkerNumber.addAndGet(1)));
1327		}
1328
1329		/**
1330	<	* Nulls out record of worker in workers array
1330	>	* Callback from ForkJoinWorkerThread constructor to establish its
1331	>	* poolIndex and record its WorkQueue. To avoid scanning bias due
1332	>	* to packing entries in front of the workQueues array, we treat
1333	>	* the array as a simple power-of-two hash table using per-thread
1334	>	* seed as hash, expanding as needed.
1335	>	*
1336	>	* @param w the worker's queue
1337		*/
1338	<	private void forgetWorker(ForkJoinWorkerThread w) {
1339	<	int idx = w.poolIndex;
625	<	// Locking helps method recordWorker avoid unecessary expansion
626	<	final ReentrantLock lock = this.workerLock;
1338	>	final void registerWorker(WorkQueue w) {
1339	>	Mutex lock = this.lock;
1340		lock.lock();
1341		try {
1342	<	ForkJoinWorkerThread[] ws = workers;
1343	<	if (idx >= 0 && idx < ws.length && ws[idx] == w) // verify
1344	<	ws[idx] = null;
1342	>	WorkQueue[] ws = workQueues;
1343	>	if (w != null && ws != null) {          // skip on shutdown/failure
1344	>	int rs, n = ws.length, m = n - 1;
1345	>	int s = nextSeed += SEED_INCREMENT; // rarely-colliding sequence
1346	>	w.seed = (s == 0) ? 1 : s;          // ensure non-zero seed
1347	>	int r = (s << 1) | 1;               // use odd-numbered indices
1348	>	if (ws[r &= m] != null) {           // collision
1349	>	int probes = 0;                 // step by approx half size
1350	>	int step = (n <= 4) ? 2 : ((n >>> 1) & SQMASK) + 2;
1351	>	while (ws[r = (r + step) & m] != null) {
1352	>	if (++probes >= n) {
1353	>	workQueues = ws = Arrays.copyOf(ws, n <<= 1);
1354	>	m = n - 1;
1355	>	probes = 0;
1356	>	}
1357	>	}
1358	>	}
1359	>	w.eventCount = w.poolIndex = r;     // establish before recording
1360	>	ws[r] = w;                          // also update seq
1361	>	runState = ((rs = runState) & SHUTDOWN) | ((rs + 2) & ~SHUTDOWN);
1362	>	}
1363		} finally {
1364		lock.unlock();
1365		}
1366		}
1367
637	–	// adding and removing workers
638	–
1368		/**
1369	<	* Tries to create and add new worker. Assumes that worker counts
1370	<	* are already updated to accommodate the worker, so adjusts on
1371	<	* failure.
1369	>	* Final callback from terminating worker, as well as upon failure
1370	>	* to construct or start a worker in addWorker.  Removes record of
1371	>	* worker from array, and adjusts counts. If pool is shutting
1372	>	* down, tries to complete termination.
1373		*
1374	<	* @return new worker or null if creation failed
1374	>	* @param wt the worker thread or null if addWorker failed
1375	>	* @param ex the exception causing failure, or null if none
1376		*/
1377	<	private ForkJoinWorkerThread addWorker() {
1378	<	ForkJoinWorkerThread w = null;
1379	<	try {
1380	<	w = factory.newThread(this);
1381	<	} finally { // Adjust on either null or exceptional factory return
1382	<	if (w == null) {
1383	<	onWorkerCreationFailure();
1384	<	return null;
1377	>	final void deregisterWorker(ForkJoinWorkerThread wt, Throwable ex) {
1378	>	Mutex lock = this.lock;
1379	>	WorkQueue w = null;
1380	>	if (wt != null && (w = wt.workQueue) != null) {
1381	>	w.runState = -1;                // ensure runState is set
1382	>	stealCount.getAndAdd(w.totalSteals + w.nsteals);
1383	>	int idx = w.poolIndex;
1384	>	lock.lock();
1385	>	try {                           // remove record from array
1386	>	WorkQueue[] ws = workQueues;
1387	>	if (ws != null && idx >= 0 && idx < ws.length && ws[idx] == w)
1388	>	ws[idx] = null;
1389	>	} finally {
1390	>	lock.unlock();
1391		}
1392		}
656	–	w.start(recordWorker(w), locallyFifo, ueh);
657	–	return w;
658	–	}
1393
1394	<	/**
1395	<	* Adjusts counts upon failure to create worker
1396	<	*/
1397	<	private void onWorkerCreationFailure() {
1398	<	int c;
1399	<	do {} while (!UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1400	<	c = workerCounts,
1401	<	c - (ONE_RUNNING|ONE_TOTAL)));
1402	<	tryTerminate(false); // in case of failure during shutdown
1394	>	long c;                             // adjust ctl counts
1395	>	do {} while (!U.compareAndSwapLong
1396	>	(this, CTL, c = ctl, (((c - AC_UNIT) & AC_MASK) |
1397	>	((c - TC_UNIT) & TC_MASK) |
1398	>	(c & ~(AC_MASK|TC_MASK)))));
1399	>
1400	>	if (!tryTerminate(false, false) && w != null) {
1401	>	w.cancelAll();                  // cancel remaining tasks
1402	>	if (w.array != null)            // suppress signal if never ran
1403	>	signalWork();               // wake up or create replacement
1404	>	if (ex == null)                 // help clean refs on way out
1405	>	ForkJoinTask.helpExpungeStaleExceptions();
1406	>	}
1407	>
1408	>	if (ex != null)                     // rethrow
1409	>	U.throwException(ex);
1410	>	}
1411	>
1412	>	// Submissions
1413	>
1414	>	/**
1415	>	* Unless shutting down, adds the given task to a submission queue
1416	>	* at submitter's current queue index (modulo submission
1417	>	* range). If no queue exists at the index, one is created.  If
1418	>	* the queue is busy, another index is randomly chosen. The
1419	>	* submitMask bounds the effective number of queues to the
1420	>	* (nearest power of two for) parallelism level.
1421	>	*
1422	>	* @param task the task. Caller must ensure non-null.
1423	>	*/
1424	>	private void doSubmit(ForkJoinTask<?> task) {
1425	>	Submitter s = submitters.get();
1426	>	for (int r = s.seed, m = submitMask;;) {
1427	>	WorkQueue[] ws; WorkQueue q;
1428	>	int k = r & m & SQMASK;          // use only even indices
1429	>	if (runState < 0 || (ws = workQueues) == null || ws.length <= k)
1430	>	throw new RejectedExecutionException(); // shutting down
1431	>	else if ((q = ws[k]) == null) {  // create new queue
1432	>	WorkQueue nq = new WorkQueue(this, null, SHARED_QUEUE);
1433	>	Mutex lock = this.lock;      // construct outside lock
1434	>	lock.lock();
1435	>	try {                        // recheck under lock
1436	>	int rs = runState;       // to update seq
1437	>	if (ws == workQueues && ws[k] == null) {
1438	>	ws[k] = nq;
1439	>	runState = ((rs & SHUTDOWN) | ((rs + 2) & ~SHUTDOWN));
1440	>	}
1441	>	} finally {
1442	>	lock.unlock();
1443	>	}
1444	>	}
1445	>	else if (q.trySharedPush(task)) {
1446	>	signalWork();
1447	>	return;
1448	>	}
1449	>	else if (m > 1) {                // move to a different index
1450	>	r ^= r << 13;                // same xorshift as WorkQueues
1451	>	r ^= r >>> 17;
1452	>	s.seed = r ^= r << 5;
1453	>	}
1454	>	else
1455	>	Thread.yield();              // yield if no alternatives
1456	>	}
1457		}
1458
1459		/**
1460	<	* Create enough total workers to establish target parallelism,
673	<	* giving up if terminating or addWorker fails
1460	>	* Submits the given (non-null) task to the common pool, if possible.
1461		*/
1462	<	private void ensureEnoughTotalWorkers() {
1463	<	int wc;
1464	<	while (runState < TERMINATING &&
1465	<	((wc = workerCounts) >>> TOTAL_COUNT_SHIFT) < parallelism) {
1466	<	if ((UNSAFE.compareAndSwapInt(this, workerCountsOffset,
680	<	wc, wc + (ONE_RUNNING|ONE_TOTAL)) &&
681	<	addWorker() == null))
682	<	break;
683	<	}
1462	>	static void submitToCommonPool(ForkJoinTask<?> task) {
1463	>	ForkJoinPool p;
1464	>	if ((p = commonPool) == null)
1465	>	p = ensureCommonPool();
1466	>	p.doSubmit(task);
1467		}
1468
1469		/**
1470	<	* Final callback from terminating worker.  Removes record of
1471	<	* worker from array, and adjusts counts. If pool is shutting
1472	<	* down, tries to complete terminatation, else possibly replaces
1473	<	* the worker.
1470	>	* Returns true if the given task was submitted to common pool
1471	>	* and has not yet commenced execution, and is available for
1472	>	* removal according to execution policies; if so removing the
1473	>	* submission from the pool.
1474		*
1475	<	* @param w the worker
1475	>	* @param task the task
1476	>	* @return true if successful
1477		*/
1478	<	final void workerTerminated(ForkJoinWorkerThread w) {
1479	<	if (w.active) { // force inactive
1480	<	w.active = false;
1481	<	do {} while (!tryDecrementActiveCount());
1482	<	}
1483	<	forgetWorker(w);
1484	<
1485	<	// decrement total count, and if was running, running count
702	<	int unit = w.isTrimmed()? ONE_TOTAL : (ONE_RUNNING|ONE_TOTAL);
703	<	int wc;
704	<	do {} while (!UNSAFE.compareAndSwapInt(this, workerCountsOffset,
705	<	wc = workerCounts, wc - unit));
706	<
707	<	accumulateStealCount(w); // collect final count
708	<	if (!tryTerminate(false))
709	<	ensureEnoughTotalWorkers();
1478	>	static boolean tryUnsubmitFromCommonPool(ForkJoinTask<?> task) {
1479	>	ForkJoinPool p; WorkQueue[] ws; WorkQueue q;
1480	>	int k = submitters.get().seed & SQMASK;
1481	>	return ((p = commonPool) != null &&
1482	>	(ws = p.workQueues) != null &&
1483	>	ws.length > (k &= p.submitMask) &&
1484	>	(q = ws[k]) != null &&
1485	>	q.trySharedUnpush(task, p));
1486		}
1487
1488	<	// Waiting for and signalling events
1488	>	// Maintaining ctl counts
1489
1490		/**
1491	<	* Ensures eventCount on exit is different (mod 2^32) than on
716	<	* entry.  CAS failures are OK -- any change in count suffices.
1491	>	* Increments active count; mainly called upon return from blocking.
1492		*/
1493	<	private void advanceEventCount() {
1494	<	int c;
1495	<	UNSAFE.compareAndSwapInt(this, eventCountOffset, c = eventCount, c+1);
1493	>	final void incrementActiveCount() {
1494	>	long c;
1495	>	do {} while (!U.compareAndSwapLong(this, CTL, c = ctl, c + AC_UNIT));
1496		}
1497
1498		/**
1499	<	* Releases workers blocked on a count not equal to current count.
1499	>	* Tries to create one or activate one or more workers if too few are active.
1500		*/
1501	<	final void releaseWaiters() {
1502	<	long top;
1503	<	int id;
1504	<	while ((id = (int)((top = eventWaiters) & WAITER_INDEX_MASK)) > 0 &&
1505	<	(int)(top >>> EVENT_COUNT_SHIFT) != eventCount) {
1506	<	ForkJoinWorkerThread[] ws = workers;
1507	<	ForkJoinWorkerThread w;
1508	<	if (ws.length >= id && (w = ws[id - 1]) != null &&
1509	<	UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
1510	<	top, w.nextWaiter))
1511	<	LockSupport.unpark(w);
1501	>	final void signalWork() {
1502	>	long c; int u;
1503	>	while ((u = (int)((c = ctl) >>> 32)) < 0) {     // too few active
1504	>	WorkQueue[] ws = workQueues; int e, i; WorkQueue w; Thread p;
1505	>	if ((e = (int)c) > 0) {                     // at least one waiting
1506	>	if (ws != null && (i = e & SMASK) < ws.length &&
1507	>	(w = ws[i]) != null && w.eventCount == (e | INT_SIGN)) {
1508	>	long nc = (((long)(w.nextWait & E_MASK)) |
1509	>	((long)(u + UAC_UNIT) << 32));
1510	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1511	>	w.eventCount = (e + E_SEQ) & E_MASK;
1512	>	if ((p = w.parker) != null)
1513	>	U.unpark(p);                // activate and release
1514	>	break;
1515	>	}
1516	>	}
1517	>	else
1518	>	break;
1519	>	}
1520	>	else if (e == 0 && (u & SHORT_SIGN) != 0) { // too few total
1521	>	long nc = (long)(((u + UTC_UNIT) & UTC_MASK) |
1522	>	((u + UAC_UNIT) & UAC_MASK)) << 32;
1523	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1524	>	addWorker();
1525	>	break;
1526	>	}
1527	>	}
1528	>	else
1529	>	break;
1530		}
1531		}
1532
1533	+	// Scanning for tasks
1534	+
1535		/**
1536	<	* Advances eventCount and releases waiters until interference by
742	<	* other releasing threads is detected.
1536	>	* Top-level runloop for workers, called by ForkJoinWorkerThread.run.
1537		*/
1538	<	final void signalWork() {
1539	<	int ec;
1540	<	UNSAFE.compareAndSwapInt(this, eventCountOffset, ec=eventCount, ec+1);
1541	<	outer:for (;;) {
1542	<	long top = eventWaiters;
1543	<	ec = eventCount;
1544	<	for (;;) {
1545	<	ForkJoinWorkerThread[] ws; ForkJoinWorkerThread w;
1546	<	int id = (int)(top & WAITER_INDEX_MASK);
1547	<	if (id <= 0 || (int)(top >>> EVENT_COUNT_SHIFT) == ec)
1548	<	return;
1549	<	if ((ws = workers).length < id || (w = ws[id - 1]) == null ||
1550	<	!UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
1551	<	top, top = w.nextWaiter))
1552	<	continue outer;      // possibly stale; reread
1553	<	LockSupport.unpark(w);
1554	<	if (top != eventWaiters) // let someone else take over
1555	<	return;
1538	>	final void runWorker(WorkQueue w) {
1539	>	w.growArray(false);         // initialize queue array in this thread
1540	>	do { w.runTask(scan(w)); } while (w.runState >= 0);
1541	>	}
1542	>
1543	>	/**
1544	>	* Scans for and, if found, returns one task, else possibly
1545	>	* inactivates the worker. This method operates on single reads of
1546	>	* volatile state and is designed to be re-invoked continuously,
1547	>	* in part because it returns upon detecting inconsistencies,
1548	>	* contention, or state changes that indicate possible success on
1549	>	* re-invocation.
1550	>	*
1551	>	* The scan searches for tasks across a random permutation of
1552	>	* queues (starting at a random index and stepping by a random
1553	>	* relative prime, checking each at least once).  The scan
1554	>	* terminates upon either finding a non-empty queue, or completing
1555	>	* the sweep. If the worker is not inactivated, it takes and
1556	>	* returns a task from this queue.  On failure to find a task, we
1557	>	* take one of the following actions, after which the caller will
1558	>	* retry calling this method unless terminated.
1559	>	*
1560	>	* * If pool is terminating, terminate the worker.
1561	>	*
1562	>	* * If not a complete sweep, try to release a waiting worker.  If
1563	>	* the scan terminated because the worker is inactivated, then the
1564	>	* released worker will often be the calling worker, and it can
1565	>	* succeed obtaining a task on the next call. Or maybe it is
1566	>	* another worker, but with same net effect. Releasing in other
1567	>	* cases as well ensures that we have enough workers running.
1568	>	*
1569	>	* * If not already enqueued, try to inactivate and enqueue the
1570	>	* worker on wait queue. Or, if inactivating has caused the pool
1571	>	* to be quiescent, relay to idleAwaitWork to check for
1572	>	* termination and possibly shrink pool.
1573	>	*
1574	>	* * If already inactive, and the caller has run a task since the
1575	>	* last empty scan, return (to allow rescan) unless others are
1576	>	* also inactivated.  Field WorkQueue.rescans counts down on each
1577	>	* scan to ensure eventual inactivation and blocking.
1578	>	*
1579	>	* * If already enqueued and none of the above apply, park
1580	>	* awaiting signal,
1581	>	*
1582	>	* @param w the worker (via its WorkQueue)
1583	>	* @return a task or null if none found
1584	>	*/
1585	>	private final ForkJoinTask<?> scan(WorkQueue w) {
1586	>	WorkQueue[] ws;                       // first update random seed
1587	>	int r = w.seed; r ^= r << 13; r ^= r >>> 17; w.seed = r ^= r << 5;
1588	>	int rs = runState, m;                 // volatile read order matters
1589	>	if ((ws = workQueues) != null && (m = ws.length - 1) > 0) {
1590	>	int ec = w.eventCount;            // ec is negative if inactive
1591	>	int step = (r >>> 16) | 1;        // relative prime
1592	>	for (int j = (m + 1) << 2; ; r += step) {
1593	>	WorkQueue q; ForkJoinTask<?> t; ForkJoinTask<?>[] a; int b;
1594	>	if ((q = ws[r & m]) != null && (b = q.base) - q.top < 0 &&
1595	>	(a = q.array) != null) {  // probably nonempty
1596	>	int i = (((a.length - 1) & b) << ASHIFT) + ABASE;
1597	>	t = (ForkJoinTask<?>)U.getObjectVolatile(a, i);
1598	>	if (q.base == b && ec >= 0 && t != null &&
1599	>	U.compareAndSwapObject(a, i, t, null)) {
1600	>	if (q.top - (q.base = b + 1) > 0)
1601	>	signalWork();    // help pushes signal
1602	>	return t;
1603	>	}
1604	>	else if (ec < 0 || j <= m) {
1605	>	rs = 0;               // mark scan as imcomplete
1606	>	break;                // caller can retry after release
1607	>	}
1608	>	}
1609	>	if (--j < 0)
1610	>	break;
1611	>	}
1612	>
1613	>	long c = ctl; int e = (int)c, a = (int)(c >> AC_SHIFT), nr, ns;
1614	>	if (e < 0)                        // decode ctl on empty scan
1615	>	w.runState = -1;              // pool is terminating
1616	>	else if (rs == 0 || rs != runState) { // incomplete scan
1617	>	WorkQueue v; Thread p;        // try to release a waiter
1618	>	if (e > 0 && a < 0 && w.eventCount == ec &&
1619	>	(v = ws[e & m]) != null && v.eventCount == (e | INT_SIGN)) {
1620	>	long nc = ((long)(v.nextWait & E_MASK) |
1621	>	((c + AC_UNIT) & (AC_MASK|TC_MASK)));
1622	>	if (ctl == c && U.compareAndSwapLong(this, CTL, c, nc)) {
1623	>	v.eventCount = (e + E_SEQ) & E_MASK;
1624	>	if ((p = v.parker) != null)
1625	>	U.unpark(p);
1626	>	}
1627	>	}
1628	>	}
1629	>	else if (ec >= 0) {               // try to enqueue/inactivate
1630	>	long nc = (long)ec | ((c - AC_UNIT) & (AC_MASK|TC_MASK));
1631	>	w.nextWait = e;
1632	>	w.eventCount = ec | INT_SIGN; // mark as inactive
1633	>	if (ctl != c || !U.compareAndSwapLong(this, CTL, c, nc))
1634	>	w.eventCount = ec;        // unmark on CAS failure
1635	>	else {
1636	>	if ((ns = w.nsteals) != 0) {
1637	>	w.nsteals = 0;        // set rescans if ran task
1638	>	w.rescans = (a > 0) ? 0 : a + parallelism;
1639	>	w.totalSteals += ns;
1640	>	}
1641	>	if (a == 1 - parallelism) // quiescent
1642	>	idleAwaitWork(w, nc, c);
1643	>	}
1644	>	}
1645	>	else if (w.eventCount < 0) {      // already queued
1646	>	int ac = a + parallelism;
1647	>	if ((nr = w.rescans) > 0)     // continue rescanning
1648	>	w.rescans = (ac < nr) ? ac : nr - 1;
1649	>	else if (((w.seed >>> 16) & ac) == 0) { // randomize park
1650	>	Thread.interrupted();     // clear status
1651	>	Thread wt = Thread.currentThread();
1652	>	U.putObject(wt, PARKBLOCKER, this);
1653	>	w.parker = wt;            // emulate LockSupport.park
1654	>	if (w.eventCount < 0)     // recheck
1655	>	U.park(false, 0L);
1656	>	w.parker = null;
1657	>	U.putObject(wt, PARKBLOCKER, null);
1658	>	}
1659		}
1660		}
1661	+	return null;
1662		}
1663
1664		/**
1665	<	* If worker is inactive, blocks until terminating or event count
1666	<	* advances from last value held by worker; in any case helps
1667	<	* release others.
1668	<	*
1669	<	* @param w the calling worker thread
1670	<	*/
1671	<	private void eventSync(ForkJoinWorkerThread w) {
1672	<	if (!w.active) {
1673	<	int prev = w.lastEventCount;
1674	<	long nextTop = (((long)prev << EVENT_COUNT_SHIFT) |
1675	<	((long)(w.poolIndex + 1)));
1676	<	long top;
1677	<	while ((runState < SHUTDOWN || !tryTerminate(false)) &&
1678	<	(((int)(top = eventWaiters) & WAITER_INDEX_MASK) == 0 ||
1679	<	(int)(top >>> EVENT_COUNT_SHIFT) == prev) &&
1680	<	eventCount == prev) {
1681	<	if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
1682	<	w.nextWaiter = top, nextTop)) {
1683	<	accumulateStealCount(w); // transfer steals while idle
1684	<	Thread.interrupted();    // clear/ignore interrupt
1685	<	while (eventCount == prev)
1686	<	w.doPark();
1665	>	* If inactivating worker w has caused the pool to become
1666	>	* quiescent, checks for pool termination, and, so long as this is
1667	>	* not the only worker, waits for event for up to a given
1668	>	* duration.  On timeout, if ctl has not changed, terminates the
1669	>	* worker, which will in turn wake up another worker to possibly
1670	>	* repeat this process.
1671	>	*
1672	>	* @param w the calling worker
1673	>	* @param currentCtl the ctl value triggering possible quiescence
1674	>	* @param prevCtl the ctl value to restore if thread is terminated
1675	>	*/
1676	>	private void idleAwaitWork(WorkQueue w, long currentCtl, long prevCtl) {
1677	>	if (w.eventCount < 0 && !tryTerminate(false, false) &&
1678	>	(int)prevCtl != 0 && !hasQueuedSubmissions() && ctl == currentCtl) {
1679	>	int dc = -(short)(currentCtl >>> TC_SHIFT);
1680	>	long parkTime = dc < 0 ? FAST_IDLE_TIMEOUT: (dc + 1) * IDLE_TIMEOUT;
1681	>	long deadline = System.nanoTime() + parkTime - 100000L; // 1ms slop
1682	>	Thread wt = Thread.currentThread();
1683	>	while (ctl == currentCtl) {
1684	>	Thread.interrupted();  // timed variant of version in scan()
1685	>	U.putObject(wt, PARKBLOCKER, this);
1686	>	w.parker = wt;
1687	>	if (ctl == currentCtl)
1688	>	U.park(false, parkTime);
1689	>	w.parker = null;
1690	>	U.putObject(wt, PARKBLOCKER, null);
1691	>	if (ctl != currentCtl)
1692	>	break;
1693	>	if (deadline - System.nanoTime() <= 0L &&
1694	>	U.compareAndSwapLong(this, CTL, currentCtl, prevCtl)) {
1695	>	w.eventCount = (w.eventCount + E_SEQ) | E_MASK;
1696	>	w.runState = -1;   // shrink
1697		break;
1698		}
1699		}
792	–	w.lastEventCount = eventCount;
1700		}
794	–	releaseWaiters();
1701		}
1702
1703		/**
1704	<	* Callback from workers invoked upon each top-level action (i.e.,
1705	<	* stealing a task or taking a submission and running
1706	<	* it). Performs one or both of the following:
1707	<	*
1708	<	* * If the worker cannot find work, updates its active status to
1709	<	* inactive and updates activeCount unless there is contention, in
1710	<	* which case it may try again (either in this or a subsequent
1711	<	* call).  Additionally, awaits the next task event and/or helps
1712	<	* wake up other releasable waiters.
1713	<	*
1714	<	* * If there are too many running threads, suspends this worker
1715	<	* (first forcing inactivation if necessary).  If it is not
1716	<	* resumed before a keepAlive elapses, the worker may be "trimmed"
1717	<	* -- killed while suspended within suspendAsSpare. Otherwise,
1718	<	* upon resume it rechecks to make sure that it is still needed.
1719	<	*
1720	<	* @param w the worker
1721	<	* @param worked false if the worker scanned for work but didn't
1722	<	* find any (in which case it may block waiting for work).
1723	<	*/
1724	<	final void preStep(ForkJoinWorkerThread w, boolean worked) {
1725	<	boolean active = w.active;
1726	<	boolean inactivate = !worked & active;
1727	<	for (;;) {
1728	<	if (inactivate) {
1729	<	int c = runState;
1730	<	if (UNSAFE.compareAndSwapInt(this, runStateOffset,
1731	<	c, c - ONE_ACTIVE))
1732	<	inactivate = active = w.active = false;
1733	<	}
1734	<	int wc = workerCounts;
1735	<	if ((wc & RUNNING_COUNT_MASK) <= parallelism) {
1736	<	if (!worked)
1737	<	eventSync(w);
1738	<	return;
1704	>	* Tries to locate and execute tasks for a stealer of the given
1705	>	* task, or in turn one of its stealers, Traces currentSteal ->
1706	>	* currentJoin links looking for a thread working on a descendant
1707	>	* of the given task and with a non-empty queue to steal back and
1708	>	* execute tasks from. The first call to this method upon a
1709	>	* waiting join will often entail scanning/search, (which is OK
1710	>	* because the joiner has nothing better to do), but this method
1711	>	* leaves hints in workers to speed up subsequent calls. The
1712	>	* implementation is very branchy to cope with potential
1713	>	* inconsistencies or loops encountering chains that are stale,
1714	>	* unknown, or so long that they are likely cyclic.
1715	>	*
1716	>	* @param joiner the joining worker
1717	>	* @param task the task to join
1718	>	* @return 0 if no progress can be made, negative if task
1719	>	* known complete, else positive
1720	>	*/
1721	>	private int tryHelpStealer(WorkQueue joiner, ForkJoinTask<?> task) {
1722	>	int stat = 0, steps = 0;                    // bound to avoid cycles
1723	>	if (joiner != null && task != null) {       // hoist null checks
1724	>	restart: for (;;) {
1725	>	ForkJoinTask<?> subtask = task;     // current target
1726	>	for (WorkQueue j = joiner, v;;) {   // v is stealer of subtask
1727	>	WorkQueue[] ws; int m, s, h;
1728	>	if ((s = task.status) < 0) {
1729	>	stat = s;
1730	>	break restart;
1731	>	}
1732	>	if ((ws = workQueues) == null || (m = ws.length - 1) <= 0)
1733	>	break restart;              // shutting down
1734	>	if ((v = ws[h = (j.stealHint | 1) & m]) == null ||
1735	>	v.currentSteal != subtask) {
1736	>	for (int origin = h;;) {    // find stealer
1737	>	if (((h = (h + 2) & m) & 15) == 1 &&
1738	>	(subtask.status < 0 || j.currentJoin != subtask))
1739	>	continue restart;   // occasional staleness check
1740	>	if ((v = ws[h]) != null &&
1741	>	v.currentSteal == subtask) {
1742	>	j.stealHint = h;    // save hint
1743	>	break;
1744	>	}
1745	>	if (h == origin)
1746	>	break restart;      // cannot find stealer
1747	>	}
1748	>	}
1749	>	for (;;) { // help stealer or descend to its stealer
1750	>	ForkJoinTask[] a;  int b;
1751	>	if (subtask.status < 0)     // surround probes with
1752	>	continue restart;       //   consistency checks
1753	>	if ((b = v.base) - v.top < 0 && (a = v.array) != null) {
1754	>	int i = (((a.length - 1) & b) << ASHIFT) + ABASE;
1755	>	ForkJoinTask<?> t =
1756	>	(ForkJoinTask<?>)U.getObjectVolatile(a, i);
1757	>	if (subtask.status < 0 || j.currentJoin != subtask ||
1758	>	v.currentSteal != subtask)
1759	>	continue restart;   // stale
1760	>	stat = 1;               // apparent progress
1761	>	if (t != null && v.base == b &&
1762	>	U.compareAndSwapObject(a, i, t, null)) {
1763	>	v.base = b + 1;     // help stealer
1764	>	joiner.runSubtask(t);
1765	>	}
1766	>	else if (v.base == b && ++steps == MAX_HELP)
1767	>	break restart;      // v apparently stalled
1768	>	}
1769	>	else {                      // empty -- try to descend
1770	>	ForkJoinTask<?> next = v.currentJoin;
1771	>	if (subtask.status < 0 || j.currentJoin != subtask ||
1772	>	v.currentSteal != subtask)
1773	>	continue restart;   // stale
1774	>	else if (next == null || ++steps == MAX_HELP)
1775	>	break restart;      // dead-end or maybe cyclic
1776	>	else {
1777	>	subtask = next;
1778	>	j = v;
1779	>	break;
1780	>	}
1781	>	}
1782	>	}
1783	>	}
1784		}
834	–	if (!(inactivate |= active) &&  // must inactivate to suspend
835	–	UNSAFE.compareAndSwapInt(this, workerCountsOffset,
836	–	wc, wc - ONE_RUNNING) &&
837	–	!w.suspendAsSpare())        // false if trimmed
838	–	return;
1785		}
1786	+	return stat;
1787		}
1788
1789		/**
1790	<	* Adjusts counts and creates or resumes compensating threads for
1791	<	* a worker about to block on task joinMe, returning early if
1792	<	* joinMe becomes ready. First tries resuming an existing spare
1793	<	* (which usually also avoids any count adjustment), but must then
1794	<	* decrement running count to determine whether a new thread is
1795	<	* needed. See above for fuller explanation.
1796	<	*/
1797	<	final void preJoin(ForkJoinTask<?> joinMe) {
1798	<	boolean dec = false;       // true when running count decremented
1799	<	for (;;) {
1800	<	releaseWaiters();      // help other threads progress
1801	<
855	<	if (joinMe.status < 0) // surround spare search with done checks
856	<	return;
857	<	ForkJoinWorkerThread spare = null;
858	<	for (ForkJoinWorkerThread w : workers) {
859	<	if (w != null && w.isSuspended()) {
860	<	spare = w;
1790	>	* If task is at base of some steal queue, steals and executes it.
1791	>	*
1792	>	* @param joiner the joining worker
1793	>	* @param task the task
1794	>	*/
1795	>	private void tryPollForAndExec(WorkQueue joiner, ForkJoinTask<?> task) {
1796	>	WorkQueue[] ws;
1797	>	if ((ws = workQueues) != null) {
1798	>	for (int j = 1; j < ws.length && task.status >= 0; j += 2) {
1799	>	WorkQueue q = ws[j];
1800	>	if (q != null && q.pollFor(task)) {
1801	>	joiner.runSubtask(task);
1802		break;
1803		}
1804		}
1805	<	if (joinMe.status < 0)
1806	<	return;
866	<
867	<	if (spare != null && spare.tryUnsuspend()) {
868	<	if (dec || joinMe.requestSignal() < 0) {
869	<	int c;
870	<	do {} while (!UNSAFE.compareAndSwapInt(this,
871	<	workerCountsOffset,
872	<	c = workerCounts,
873	<	c + ONE_RUNNING));
874	<	} // else no net count change
875	<	LockSupport.unpark(spare);
876	<	return;
877	<	}
1805	>	}
1806	>	}
1807
1808	<	int wc = workerCounts; // decrement running count
1809	<	if (!dec && (wc & RUNNING_COUNT_MASK) != 0 &&
1810	<	(dec = UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1811	<	wc, wc -= ONE_RUNNING)) &&
1812	<	joinMe.requestSignal() < 0) { // cannot block
1813	<	int c;                        // back out
1814	<	do {} while (!UNSAFE.compareAndSwapInt(this,
1815	<	workerCountsOffset,
1816	<	c = workerCounts,
1817	<	c + ONE_RUNNING));
1818	<	return;
1808	>	/**
1809	>	* Tries to decrement active count (sometimes implicitly) and
1810	>	* possibly release or create a compensating worker in preparation
1811	>	* for blocking. Fails on contention or termination. Otherwise,
1812	>	* adds a new thread if no idle workers are available and either
1813	>	* pool would become completely starved or: (at least half
1814	>	* starved, and fewer than 50% spares exist, and there is at least
1815	>	* one task apparently available). Even though the availability
1816	>	* check requires a full scan, it is worthwhile in reducing false
1817	>	* alarms.
1818	>	*
1819	>	* @param task if non-null, a task being waited for
1820	>	* @param blocker if non-null, a blocker being waited for
1821	>	* @return true if the caller can block, else should recheck and retry
1822	>	*/
1823	>	final boolean tryCompensate(ForkJoinTask<?> task, ManagedBlocker blocker) {
1824	>	int pc = parallelism, e;
1825	>	long c = ctl;
1826	>	WorkQueue[] ws = workQueues;
1827	>	if ((e = (int)c) >= 0 && ws != null) {
1828	>	int u, a, ac, hc;
1829	>	int tc = (short)((u = (int)(c >>> 32)) >>> UTC_SHIFT) + pc;
1830	>	boolean replace = false;
1831	>	if ((a = u >> UAC_SHIFT) <= 0) {
1832	>	if ((ac = a + pc) <= 1)
1833	>	replace = true;
1834	>	else if ((e > 0 || (task != null &&
1835	>	ac <= (hc = pc >>> 1) && tc < pc + hc))) {
1836	>	WorkQueue w;
1837	>	for (int j = 0; j < ws.length; ++j) {
1838	>	if ((w = ws[j]) != null && !w.isEmpty()) {
1839	>	replace = true;
1840	>	break;   // in compensation range and tasks available
1841	>	}
1842	>	}
1843	>	}
1844		}
1845	<
1846	<	if (dec) {
1847	<	int tc = wc >>> TOTAL_COUNT_SHIFT;
1848	<	int pc = parallelism;
1849	<	int dc = pc - (wc & RUNNING_COUNT_MASK); // deficit count
1850	<	if ((dc < pc && (dc <= 0 || (dc * dc < (tc - pc) * pc) ||
1851	<	!maintainsParallelism)) ||
1852	<	tc >= maxPoolSize) // cannot add
1853	<	return;
1854	<	if (spare == null &&
1855	<	UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
1856	<	wc + (ONE_RUNNING|ONE_TOTAL))) {
1857	<	addWorker();
1858	<	return;
1845	>	if ((task == null || task.status >= 0) && // recheck need to block
1846	>	(blocker == null || !blocker.isReleasable()) && ctl == c) {
1847	>	if (!replace) {          // no compensation
1848	>	long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK);
1849	>	if (U.compareAndSwapLong(this, CTL, c, nc))
1850	>	return true;
1851	>	}
1852	>	else if (e != 0) {       // release an idle worker
1853	>	WorkQueue w; Thread p; int i;
1854	>	if ((i = e & SMASK) < ws.length && (w = ws[i]) != null) {
1855	>	long nc = ((long)(w.nextWait & E_MASK) |
1856	>	(c & (AC_MASK|TC_MASK)));
1857	>	if (w.eventCount == (e | INT_SIGN) &&
1858	>	U.compareAndSwapLong(this, CTL, c, nc)) {
1859	>	w.eventCount = (e + E_SEQ) & E_MASK;
1860	>	if ((p = w.parker) != null)
1861	>	U.unpark(p);
1862	>	return true;
1863	>	}
1864	>	}
1865	>	}
1866	>	else if (tc < MAX_CAP) { // create replacement
1867	>	long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK);
1868	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1869	>	addWorker();
1870	>	return true;
1871	>	}
1872		}
1873		}
1874		}
1875	+	return false;
1876		}
1877
1878		/**
1879	<	* Same idea as preJoin but with too many differing details to
1880	<	* integrate: There are no task-based signal counts, and only one
1881	<	* way to do the actual blocking. So for simplicity it is directly
1882	<	* incorporated into this method.
1879	>	* Helps and/or blocks until the given task is done.
1880	>	*
1881	>	* @param joiner the joining worker
1882	>	* @param task the task
1883	>	* @return task status on exit
1884		*/
1885	<	final void doBlock(ManagedBlocker blocker, boolean maintainPar)
1886	<	throws InterruptedException {
1887	<	maintainPar &= maintainsParallelism; // override
1888	<	boolean dec = false;
1889	<	boolean done = false;
1890	<	for (;;) {
1891	<	releaseWaiters();
1892	<	if (done = blocker.isReleasable())
1893	<	break;
1894	<	ForkJoinWorkerThread spare = null;
1895	<	for (ForkJoinWorkerThread w : workers) {
1896	<	if (w != null && w.isSuspended()) {
1897	<	spare = w;
1898	<	break;
1899	<	}
1900	<	}
1901	<	if (done = blocker.isReleasable())
1902	<	break;
1903	<	if (spare != null && spare.tryUnsuspend()) {
1904	<	if (dec) {
1905	<	int c;
1906	<	do {} while (!UNSAFE.compareAndSwapInt(this,
1907	<	workerCountsOffset,
1908	<	c = workerCounts,
1909	<	c + ONE_RUNNING));
1885	>	final int awaitJoin(WorkQueue joiner, ForkJoinTask<?> task) {
1886	>	int s;
1887	>	if ((s = task.status) >= 0) {
1888	>	ForkJoinTask<?> prevJoin = joiner.currentJoin;
1889	>	joiner.currentJoin = task;
1890	>	long startTime = 0L;
1891	>	for (int k = 0;;) {
1892	>	if ((s = (joiner.isEmpty() ?           // try to help
1893	>	tryHelpStealer(joiner, task) :
1894	>	joiner.tryRemoveAndExec(task))) == 0 &&
1895	>	(s = task.status) >= 0) {
1896	>	if (k == 0) {
1897	>	startTime = System.nanoTime();
1898	>	tryPollForAndExec(joiner, task); // check uncommon case
1899	>	}
1900	>	else if ((k & (MAX_HELP - 1)) == 0 &&
1901	>	System.nanoTime() - startTime >=
1902	>	COMPENSATION_DELAY &&
1903	>	tryCompensate(task, null)) {
1904	>	if (task.trySetSignal()) {
1905	>	synchronized (task) {
1906	>	if (task.status >= 0) {
1907	>	try {                // see ForkJoinTask
1908	>	task.wait();     //  for explanation
1909	>	} catch (InterruptedException ie) {
1910	>	}
1911	>	}
1912	>	else
1913	>	task.notifyAll();
1914	>	}
1915	>	}
1916	>	long c;                          // re-activate
1917	>	do {} while (!U.compareAndSwapLong
1918	>	(this, CTL, c = ctl, c + AC_UNIT));
1919	>	}
1920		}
1921	<	LockSupport.unpark(spare);
1922	<	break;
944	<	}
945	<	int wc = workerCounts;
946	<	if (!dec && (wc & RUNNING_COUNT_MASK) != 0)
947	<	dec = UNSAFE.compareAndSwapInt(this, workerCountsOffset,
948	<	wc, wc -= ONE_RUNNING);
949	<	if (dec) {
950	<	int tc = wc >>> TOTAL_COUNT_SHIFT;
951	<	int pc = parallelism;
952	<	int dc = pc - (wc & RUNNING_COUNT_MASK);
953	<	if ((dc < pc && (dc <= 0 || (dc * dc < (tc - pc) * pc) ||
954	<	!maintainPar)) ||
955	<	tc >= maxPoolSize)
956	<	break;
957	<	if (spare == null &&
958	<	UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
959	<	wc + (ONE_RUNNING|ONE_TOTAL))){
960	<	addWorker();
1921	>	if (s < 0 || (s = task.status) < 0) {
1922	>	joiner.currentJoin = prevJoin;
1923		break;
1924		}
1925	+	else if ((k++ & (MAX_HELP - 1)) == MAX_HELP >>> 1)
1926	+	Thread.yield();                     // for politeness
1927		}
1928		}
1929	<
966	<	try {
967	<	if (!done)
968	<	do {} while (!blocker.isReleasable() && !blocker.block());
969	<	} finally {
970	<	if (dec) {
971	<	int c;
972	<	do {} while (!UNSAFE.compareAndSwapInt(this,
973	<	workerCountsOffset,
974	<	c = workerCounts,
975	<	c + ONE_RUNNING));
976	<	}
977	<	}
1929	>	return s;
1930		}
1931
1932		/**
1933	<	* Possibly initiates and/or completes termination.
1933	>	* Stripped-down variant of awaitJoin used by timed joins. Tries
1934	>	* to help join only while there is continuous progress. (Caller
1935	>	* will then enter a timed wait.)
1936		*
1937	<	* @param now if true, unconditionally terminate, else only
1938	<	* if shutdown and empty queue and no active workers
1939	<	* @return true if now terminating or terminated
1937	>	* @param joiner the joining worker
1938	>	* @param task the task
1939	>	* @return task status on exit
1940		*/
1941	<	private boolean tryTerminate(boolean now) {
1942	<	if (now)
1943	<	advanceRunLevel(SHUTDOWN); // ensure at least SHUTDOWN
1944	<	else if (runState < SHUTDOWN ||
1945	<	!submissionQueue.isEmpty() ||
1946	<	(runState & ACTIVE_COUNT_MASK) != 0)
1947	<	return false;
1948	<
1949	<	if (advanceRunLevel(TERMINATING))
996	<	startTerminating();
1941	>	final int helpJoinOnce(WorkQueue joiner, ForkJoinTask<?> task) {
1942	>	int s;
1943	>	while ((s = task.status) >= 0 &&
1944	>	(joiner.isEmpty() ?
1945	>	tryHelpStealer(joiner, task) :
1946	>	joiner.tryRemoveAndExec(task)) != 0)
1947	>	;
1948	>	return s;
1949	>	}
1950
1951	<	// Finish now if all threads terminated; else in some subsequent call
1952	<	if ((workerCounts >>> TOTAL_COUNT_SHIFT) == 0) {
1953	<	advanceRunLevel(TERMINATED);
1954	<	terminationLatch.countDown();
1951	>	/**
1952	>	* Returns a (probably) non-empty steal queue, if one is found
1953	>	* during a random, then cyclic scan, else null.  This method must
1954	>	* be retried by caller if, by the time it tries to use the queue,
1955	>	* it is empty.
1956	>	*/
1957	>	private WorkQueue findNonEmptyStealQueue(WorkQueue w) {
1958	>	// Similar to loop in scan(), but ignoring submissions
1959	>	int r;
1960	>	if (w == null) // allow external callers
1961	>	r = ThreadLocalRandom.current().nextInt();
1962	>	else {
1963	>	r = w.seed; r ^= r << 13; r ^= r >>> 17; w.seed = r ^= r << 5;
1964	>	}
1965	>	int step = (r >>> 16) | 1;
1966	>	for (WorkQueue[] ws;;) {
1967	>	int rs = runState, m;
1968	>	if ((ws = workQueues) == null || (m = ws.length - 1) < 1)
1969	>	return null;
1970	>	for (int j = (m + 1) << 2; ; r += step) {
1971	>	WorkQueue q = ws[((r << 1) | 1) & m];
1972	>	if (q != null && !q.isEmpty())
1973	>	return q;
1974	>	else if (--j < 0) {
1975	>	if (runState == rs)
1976	>	return null;
1977	>	break;
1978	>	}
1979	>	}
1980		}
1003	–	return true;
1981		}
1982
1983		/**
1984	<	* Actions on transition to TERMINATING
1985	<	*/
1986	<	private void startTerminating() {
1987	<	// Clear out and cancel submissions, ignoring exceptions
1988	<	ForkJoinTask<?> task;
1989	<	while ((task = submissionQueue.poll()) != null) {
1990	<	try {
1991	<	task.cancel(false);
1992	<	} catch (Throwable ignore) {
1984	>	* Runs tasks until {@code isQuiescent()}. We piggyback on
1985	>	* active count ctl maintenance, but rather than blocking
1986	>	* when tasks cannot be found, we rescan until all others cannot
1987	>	* find tasks either.
1988	>	*/
1989	>	final void helpQuiescePool(WorkQueue w) {
1990	>	for (boolean active = true;;) {
1991	>	ForkJoinTask<?> localTask; // exhaust local queue
1992	>	while ((localTask = w.nextLocalTask()) != null)
1993	>	localTask.doExec();
1994	>	WorkQueue q = findNonEmptyStealQueue(w);
1995	>	if (q != null) {
1996	>	ForkJoinTask<?> t; int b;
1997	>	if (!active) {      // re-establish active count
1998	>	long c;
1999	>	active = true;
2000	>	do {} while (!U.compareAndSwapLong
2001	>	(this, CTL, c = ctl, c + AC_UNIT));
2002	>	}
2003	>	if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null)
2004	>	w.runSubtask(t);
2005		}
2006	<	}
2007	<	// Propagate run level
2008	<	for (ForkJoinWorkerThread w : workers) {
2009	<	if (w != null)
2010	<	w.shutdown();    // also resumes suspended workers
2011	<	}
2012	<	// Ensure no straggling local tasks
2013	<	for (ForkJoinWorkerThread w : workers) {
2014	<	if (w != null)
2015	<	w.cancelTasks();
2016	<	}
2017	<	// Wake up idle workers
2018	<	advanceEventCount();
1030	<	releaseWaiters();
1031	<	// Unstick pending joins
1032	<	for (ForkJoinWorkerThread w : workers) {
1033	<	if (w != null && !w.isTerminated()) {
1034	<	try {
1035	<	w.interrupt();
1036	<	} catch (SecurityException ignore) {
2006	>	else {
2007	>	long c;
2008	>	if (active) {       // decrement active count without queuing
2009	>	active = false;
2010	>	do {} while (!U.compareAndSwapLong
2011	>	(this, CTL, c = ctl, c -= AC_UNIT));
2012	>	}
2013	>	else
2014	>	c = ctl;        // re-increment on exit
2015	>	if ((int)(c >> AC_SHIFT) + parallelism == 0) {
2016	>	do {} while (!U.compareAndSwapLong
2017	>	(this, CTL, c = ctl, c + AC_UNIT));
2018	>	break;
2019		}
2020		}
2021		}
2022		}
2023
1042	–	// misc support for ForkJoinWorkerThread
1043	–
2024		/**
2025	<	* Returns pool number
2025	>	* Restricted version of helpQuiescePool for non-FJ callers
2026		*/
2027	<	final int getPoolNumber() {
2028	<	return poolNumber;
2027	>	static void externalHelpQuiescePool() {
2028	>	ForkJoinPool p; WorkQueue[] ws; WorkQueue w, q;
2029	>	ForkJoinTask<?> t; int b;
2030	>	int k = submitters.get().seed & SQMASK;
2031	>	if ((p = commonPool) != null &&
2032	>	(ws = p.workQueues) != null &&
2033	>	ws.length > (k &= p.submitMask) &&
2034	>	(w = ws[k]) != null &&
2035	>	(q = p.findNonEmptyStealQueue(w)) != null &&
2036	>	(b = q.base) - q.top < 0 &&
2037	>	(t = q.pollAt(b)) != null)
2038	>	t.doExec();
2039		}
2040
2041		/**
2042	<	* Accumulates steal count from a worker, clearing
2043	<	* the worker's value
2044	<	*/
2045	<	final void accumulateStealCount(ForkJoinWorkerThread w) {
2046	<	int sc = w.stealCount;
2047	<	if (sc != 0) {
2048	<	long c;
2049	<	w.stealCount = 0;
2050	<	do {} while (!UNSAFE.compareAndSwapLong(this, stealCountOffset,
2051	<	c = stealCount, c + sc));
2042	>	* Gets and removes a local or stolen task for the given worker.
2043	>	*
2044	>	* @return a task, if available
2045	>	*/
2046	>	final ForkJoinTask<?> nextTaskFor(WorkQueue w) {
2047	>	for (ForkJoinTask<?> t;;) {
2048	>	WorkQueue q; int b;
2049	>	if ((t = w.nextLocalTask()) != null)
2050	>	return t;
2051	>	if ((q = findNonEmptyStealQueue(w)) == null)
2052	>	return null;
2053	>	if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null)
2054	>	return t;
2055		}
2056		}
2057
2058		/**
2059		* Returns the approximate (non-atomic) number of idle threads per
2060	<	* active thread.
2060	>	* active thread to offset steal queue size for method
2061	>	* ForkJoinTask.getSurplusQueuedTaskCount().
2062		*/
2063		final int idlePerActive() {
2064	<	int ac = runState;    // no mask -- artifically boosts during shutdown
2065	<	int pc = parallelism; // use targeted parallelism, not rc
2066	<	// Use exact results for small values, saturate past 4
2067	<	return pc <= ac? 0 : pc >>> 1 <= ac? 1 : pc >>> 2 <= ac? 3 : pc >>> 3;
2064	>	// Approximate at powers of two for small values, saturate past 4
2065	>	int p = parallelism;
2066	>	int a = p + (int)(ctl >> AC_SHIFT);
2067	>	return (a > (p >>>= 1) ? 0 :
2068	>	a > (p >>>= 1) ? 1 :
2069	>	a > (p >>>= 1) ? 2 :
2070	>	a > (p >>>= 1) ? 4 :
2071	>	8);
2072		}
2073
2074		/**
2075	<	* Returns the approximate (non-atomic) difference between running
2076	<	* and active counts.
2075	>	* Returns approximate submission queue length for the given caller
2076	>	*/
2077	>	static int getEstimatedSubmitterQueueLength() {
2078	>	ForkJoinPool p; WorkQueue[] ws; WorkQueue q;
2079	>	int k = submitters.get().seed & SQMASK;
2080	>	return ((p = commonPool) != null &&
2081	>	p.runState >= 0 &&
2082	>	(ws = p.workQueues) != null &&
2083	>	ws.length > (k &= p.submitMask) &&
2084	>	(q = ws[k]) != null) ?
2085	>	q.queueSize() : 0;
2086	>	}
2087	>
2088	>	//  Termination
2089	>
2090	>	/**
2091	>	* Possibly initiates and/or completes termination.  The caller
2092	>	* triggering termination runs three passes through workQueues:
2093	>	* (0) Setting termination status, followed by wakeups of queued
2094	>	* workers; (1) cancelling all tasks; (2) interrupting lagging
2095	>	* threads (likely in external tasks, but possibly also blocked in
2096	>	* joins).  Each pass repeats previous steps because of potential
2097	>	* lagging thread creation.
2098	>	*
2099	>	* @param now if true, unconditionally terminate, else only
2100	>	* if no work and no active workers
2101	>	* @param enable if true, enable shutdown when next possible
2102	>	* @return true if now terminating or terminated
2103		*/
2104	<	final int inactiveCount() {
2105	<	return (workerCounts & RUNNING_COUNT_MASK) -
2106	<	(runState & ACTIVE_COUNT_MASK);
2104	>	private boolean tryTerminate(boolean now, boolean enable) {
2105	>	Mutex lock = this.lock;
2106	>	for (long c;;) {
2107	>	if (((c = ctl) & STOP_BIT) != 0) {      // already terminating
2108	>	if ((short)(c >>> TC_SHIFT) == -parallelism) {
2109	>	lock.lock();                    // don't need try/finally
2110	>	termination.signalAll();        // signal when 0 workers
2111	>	lock.unlock();
2112	>	}
2113	>	return true;
2114	>	}
2115	>	if (runState >= 0) {                    // not yet enabled
2116	>	if (!enable)
2117	>	return false;
2118	>	lock.lock();
2119	>	runState |= SHUTDOWN;
2120	>	lock.unlock();
2121	>	}
2122	>	if (!now) {                             // check if idle & no tasks
2123	>	if ((int)(c >> AC_SHIFT) != -parallelism ||
2124	>	hasQueuedSubmissions())
2125	>	return false;
2126	>	// Check for unqueued inactive workers. One pass suffices.
2127	>	WorkQueue[] ws = workQueues; WorkQueue w;
2128	>	if (ws != null) {
2129	>	for (int i = 1; i < ws.length; i += 2) {
2130	>	if ((w = ws[i]) != null && w.eventCount >= 0)
2131	>	return false;
2132	>	}
2133	>	}
2134	>	}
2135	>	if (U.compareAndSwapLong(this, CTL, c, c | STOP_BIT)) {
2136	>	for (int pass = 0; pass < 3; ++pass) {
2137	>	WorkQueue[] ws = workQueues;
2138	>	if (ws != null) {
2139	>	WorkQueue w;
2140	>	int n = ws.length;
2141	>	for (int i = 0; i < n; ++i) {
2142	>	if ((w = ws[i]) != null) {
2143	>	w.runState = -1;
2144	>	if (pass > 0) {
2145	>	w.cancelAll();
2146	>	if (pass > 1)
2147	>	w.interruptOwner();
2148	>	}
2149	>	}
2150	>	}
2151	>	// Wake up workers parked on event queue
2152	>	int i, e; long cc; Thread p;
2153	>	while ((e = (int)(cc = ctl) & E_MASK) != 0 &&
2154	>	(i = e & SMASK) < n &&
2155	>	(w = ws[i]) != null) {
2156	>	long nc = ((long)(w.nextWait & E_MASK) |
2157	>	((cc + AC_UNIT) & AC_MASK) |
2158	>	(cc & (TC_MASK|STOP_BIT)));
2159	>	if (w.eventCount == (e | INT_SIGN) &&
2160	>	U.compareAndSwapLong(this, CTL, cc, nc)) {
2161	>	w.eventCount = (e + E_SEQ) & E_MASK;
2162	>	w.runState = -1;
2163	>	if ((p = w.parker) != null)
2164	>	U.unpark(p);
2165	>	}
2166	>	}
2167	>	}
2168	>	}
2169	>	}
2170	>	}
2171		}
2172
2173	<	// Public and protected methods
2173	>	// Exported methods
2174
2175		// Constructors
2176
2177		/**
2178		* Creates a {@code ForkJoinPool} with parallelism equal to {@link
2179	<	* java.lang.Runtime#availableProcessors}, and using the {@linkplain
2180	<	* #defaultForkJoinWorkerThreadFactory default thread factory}.
2179	>	* java.lang.Runtime#availableProcessors}, using the {@linkplain
2180	>	* #defaultForkJoinWorkerThreadFactory default thread factory},
2181	>	* no UncaughtExceptionHandler, and non-async LIFO processing mode.
2182		*
2183		* @throws SecurityException if a security manager exists and
2184		*         the caller is not permitted to modify threads
#	Line 1098 | Line 2187 | public class ForkJoinPool extends Abstra
2187		*/
2188		public ForkJoinPool() {
2189		this(Runtime.getRuntime().availableProcessors(),
2190	<	defaultForkJoinWorkerThreadFactory);
2190	>	defaultForkJoinWorkerThreadFactory, null, false);
2191		}
2192
2193		/**
2194		* Creates a {@code ForkJoinPool} with the indicated parallelism
2195	<	* level and using the {@linkplain
2196	<	* #defaultForkJoinWorkerThreadFactory default thread factory}.
2195	>	* level, the {@linkplain
2196	>	* #defaultForkJoinWorkerThreadFactory default thread factory},
2197	>	* no UncaughtExceptionHandler, and non-async LIFO processing mode.
2198		*
2199		* @param parallelism the parallelism level
2200		* @throws IllegalArgumentException if parallelism less than or
#	Line 1115 | Line 2205 | public class ForkJoinPool extends Abstra
2205		*         java.lang.RuntimePermission}{@code ("modifyThread")}
2206		*/
2207		public ForkJoinPool(int parallelism) {
2208	<	this(parallelism, defaultForkJoinWorkerThreadFactory);
1119	<	}
1120	<
1121	<	/**
1122	<	* Creates a {@code ForkJoinPool} with parallelism equal to {@link
1123	<	* java.lang.Runtime#availableProcessors}, and using the given
1124	<	* thread factory.
1125	<	*
1126	<	* @param factory the factory for creating new threads
1127	<	* @throws NullPointerException if the factory is null
1128	<	* @throws SecurityException if a security manager exists and
1129	<	*         the caller is not permitted to modify threads
1130	<	*         because it does not hold {@link
1131	<	*         java.lang.RuntimePermission}{@code ("modifyThread")}
1132	<	*/
1133	<	public ForkJoinPool(ForkJoinWorkerThreadFactory factory) {
1134	<	this(Runtime.getRuntime().availableProcessors(), factory);
2208	>	this(parallelism, defaultForkJoinWorkerThreadFactory, null, false);
2209		}
2210
2211		/**
2212	<	* Creates a {@code ForkJoinPool} with the given parallelism and
1139	<	* thread factory.
2212	>	* Creates a {@code ForkJoinPool} with the given parameters.
2213		*
2214	<	* @param parallelism the parallelism level
2215	<	* @param factory the factory for creating new threads
2214	>	* @param parallelism the parallelism level. For default value,
2215	>	* use {@link java.lang.Runtime#availableProcessors}.
2216	>	* @param factory the factory for creating new threads. For default value,
2217	>	* use {@link #defaultForkJoinWorkerThreadFactory}.
2218	>	* @param handler the handler for internal worker threads that
2219	>	* terminate due to unrecoverable errors encountered while executing
2220	>	* tasks. For default value, use {@code null}.
2221	>	* @param asyncMode if true,
2222	>	* establishes local first-in-first-out scheduling mode for forked
2223	>	* tasks that are never joined. This mode may be more appropriate
2224	>	* than default locally stack-based mode in applications in which
2225	>	* worker threads only process event-style asynchronous tasks.
2226	>	* For default value, use {@code false}.
2227		* @throws IllegalArgumentException if parallelism less than or
2228		*         equal to zero, or greater than implementation limit
2229		* @throws NullPointerException if the factory is null
#	Line 1148 | Line 2232 | public class ForkJoinPool extends Abstra
2232		*         because it does not hold {@link
2233		*         java.lang.RuntimePermission}{@code ("modifyThread")}
2234		*/
2235	<	public ForkJoinPool(int parallelism, ForkJoinWorkerThreadFactory factory) {
2235	>	public ForkJoinPool(int parallelism,
2236	>	ForkJoinWorkerThreadFactory factory,
2237	>	Thread.UncaughtExceptionHandler handler,
2238	>	boolean asyncMode) {
2239		checkPermission();
2240		if (factory == null)
2241		throw new NullPointerException();
2242	<	if (parallelism <= 0 || parallelism > MAX_THREADS)
2242	>	if (parallelism <= 0 || parallelism > MAX_CAP)
2243		throw new IllegalArgumentException();
1157	–	this.poolNumber = poolNumberGenerator.incrementAndGet();
1158	–	int arraySize = initialArraySizeFor(parallelism);
2244		this.parallelism = parallelism;
2245		this.factory = factory;
2246	<	this.maxPoolSize = MAX_THREADS;
2247	<	this.maintainsParallelism = true;
2248	<	this.workers = new ForkJoinWorkerThread[arraySize];
2249	<	this.submissionQueue = new LinkedTransferQueue<ForkJoinTask<?>>();
2250	<	this.workerLock = new ReentrantLock();
2251	<	this.terminationLatch = new CountDownLatch(1);
2252	<	// Start first worker; remaining workers added upon first submission
2253	<	workerCounts = ONE_RUNNING | ONE_TOTAL;
2254	<	addWorker();
2246	>	this.ueh = handler;
2247	>	this.localMode = asyncMode ? FIFO_QUEUE : LIFO_QUEUE;
2248	>	long np = (long)(-parallelism); // offset ctl counts
2249	>	this.ctl = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK);
2250	>	// Use nearest power 2 for workQueues size. See Hackers Delight sec 3.2.
2251	>	int n = parallelism - 1;
2252	>	n |= n >>> 1; n |= n >>> 2; n |= n >>> 4; n |= n >>> 8; n |= n >>> 16;
2253	>	int size = (n + 1) << 1;        // #slots = 2*#workers
2254	>	this.submitMask = size - 1;     // room for max # of submit queues
2255	>	this.workQueues = new WorkQueue[size];
2256	>	this.termination = (this.lock = new Mutex()).newCondition();
2257	>	this.stealCount = new AtomicLong();
2258	>	this.nextWorkerNumber = new AtomicInteger();
2259	>	int pn = poolNumberGenerator.incrementAndGet();
2260	>	StringBuilder sb = new StringBuilder("ForkJoinPool-");
2261	>	sb.append(Integer.toString(pn));
2262	>	sb.append("-worker-");
2263	>	this.workerNamePrefix = sb.toString();
2264	>	lock.lock();
2265	>	this.runState = 1;              // set init flag
2266	>	lock.unlock();
2267		}
2268
2269		/**
2270	<	* Returns initial power of two size for workers array.
2271	<	* @param pc the initial parallelism level
2272	<	*/
2273	<	private static int initialArraySizeFor(int pc) {
2274	<	// See Hackers Delight, sec 3.2. We know MAX_THREADS < (1 >>> 16)
2275	<	int size = pc < MAX_THREADS ? pc + 1 : MAX_THREADS;
2276	<	size |= size >>> 1;
1180	<	size |= size >>> 2;
1181	<	size |= size >>> 4;
1182	<	size |= size >>> 8;
1183	<	return size + 1;
2270	>	* Returns the common pool instance
2271	>	*
2272	>	* @return the common pool instance
2273	>	*/
2274	>	public static ForkJoinPool commonPool() {
2275	>	ForkJoinPool p;
2276	>	return (p = commonPool) != null? p : ensureCommonPool();
2277		}
2278
2279	<	// Execution methods
2280	<
2281	<	/**
2282	<	* Common code for execute, invoke and submit
2283	<	*/
2284	<	private <T> void doSubmit(ForkJoinTask<T> task) {
2285	<	if (task == null)
2286	<	throw new NullPointerException();
2287	<	if (runState >= SHUTDOWN)
2288	<	throw new RejectedExecutionException();
2289	<	submissionQueue.offer(task);
2290	<	advanceEventCount();
2291	<	releaseWaiters();
2292	<	if ((workerCounts >>> TOTAL_COUNT_SHIFT) < parallelism)
2293	<	ensureEnoughTotalWorkers();
2279	>	private static ForkJoinPool ensureCommonPool() {
2280	>	ForkJoinPool p;
2281	>	if ((p = commonPool) == null) {
2282	>	final Mutex lock = initializationLock;
2283	>	lock.lock();
2284	>	try {
2285	>	if ((p = commonPool) == null) {
2286	>	p = commonPool = new ForkJoinPool(commonPoolParallelism,
2287	>	commonPoolFactory,
2288	>	commonPoolUEH, false);
2289	>	// use a more informative name string for workers
2290	>	p.workerNamePrefix = "ForkJoinPool.commonPool-worker-";
2291	>	}
2292	>	} finally {
2293	>	lock.unlock();
2294	>	}
2295	>	}
2296	>	return p;
2297		}
2298
2299	+	// Execution methods
2300	+
2301		/**
2302		* Performs the given task, returning its result upon completion.
2303	+	* If the computation encounters an unchecked Exception or Error,
2304	+	* it is rethrown as the outcome of this invocation.  Rethrown
2305	+	* exceptions behave in the same way as regular exceptions, but,
2306	+	* when possible, contain stack traces (as displayed for example
2307	+	* using {@code ex.printStackTrace()}) of both the current thread
2308	+	* as well as the thread actually encountering the exception;
2309	+	* minimally only the latter.
2310		*
2311		* @param task the task
2312		* @return the task's result
#	Line 1210 | Line 2315 | public class ForkJoinPool extends Abstra
2315		*         scheduled for execution
2316		*/
2317		public <T> T invoke(ForkJoinTask<T> task) {
2318	+	if (task == null)
2319	+	throw new NullPointerException();
2320		doSubmit(task);
2321		return task.join();
2322		}
#	Line 1223 | Line 2330 | public class ForkJoinPool extends Abstra
2330		*         scheduled for execution
2331		*/
2332		public void execute(ForkJoinTask<?> task) {
2333	+	if (task == null)
2334	+	throw new NullPointerException();
2335		doSubmit(task);
2336		}
2337
#	Line 1234 | Line 2343 | public class ForkJoinPool extends Abstra
2343		*         scheduled for execution
2344		*/
2345		public void execute(Runnable task) {
2346	+	if (task == null)
2347	+	throw new NullPointerException();
2348		ForkJoinTask<?> job;
2349		if (task instanceof ForkJoinTask<?>) // avoid re-wrap
2350		job = (ForkJoinTask<?>) task;
2351		else
2352	<	job = ForkJoinTask.adapt(task, null);
2352	>	job = new ForkJoinTask.AdaptedRunnableAction(task);
2353		doSubmit(job);
2354		}
2355
2356		/**
2357	+	* Submits a ForkJoinTask for execution.
2358	+	*
2359	+	* @param task the task to submit
2360	+	* @return the task
2361	+	* @throws NullPointerException if the task is null
2362	+	* @throws RejectedExecutionException if the task cannot be
2363	+	*         scheduled for execution
2364	+	*/
2365	+	public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) {
2366	+	if (task == null)
2367	+	throw new NullPointerException();
2368	+	doSubmit(task);
2369	+	return task;
2370	+	}
2371	+
2372	+	/**
2373		* @throws NullPointerException if the task is null
2374		* @throws RejectedExecutionException if the task cannot be
2375		*         scheduled for execution
2376		*/
2377		public <T> ForkJoinTask<T> submit(Callable<T> task) {
2378	<	ForkJoinTask<T> job = ForkJoinTask.adapt(task);
2378	>	ForkJoinTask<T> job = new ForkJoinTask.AdaptedCallable<T>(task);
2379		doSubmit(job);
2380		return job;
2381		}
#	Line 1259 | Line 2386 | public class ForkJoinPool extends Abstra
2386		*         scheduled for execution
2387		*/
2388		public <T> ForkJoinTask<T> submit(Runnable task, T result) {
2389	<	ForkJoinTask<T> job = ForkJoinTask.adapt(task, result);
2389	>	ForkJoinTask<T> job = new ForkJoinTask.AdaptedRunnable<T>(task, result);
2390		doSubmit(job);
2391		return job;
2392		}
#	Line 1270 | Line 2397 | public class ForkJoinPool extends Abstra
2397		*         scheduled for execution
2398		*/
2399		public ForkJoinTask<?> submit(Runnable task) {
2400	+	if (task == null)
2401	+	throw new NullPointerException();
2402		ForkJoinTask<?> job;
2403		if (task instanceof ForkJoinTask<?>) // avoid re-wrap
2404		job = (ForkJoinTask<?>) task;
2405		else
2406	<	job = ForkJoinTask.adapt(task, null);
2406	>	job = new ForkJoinTask.AdaptedRunnableAction(task);
2407		doSubmit(job);
2408		return job;
2409		}
2410
2411		/**
1283	–	* Submits a ForkJoinTask for execution.
1284	–	*
1285	–	* @param task the task to submit
1286	–	* @return the task
1287	–	* @throws NullPointerException if the task is null
1288	–	* @throws RejectedExecutionException if the task cannot be
1289	–	*         scheduled for execution
1290	–	*/
1291	–	public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) {
1292	–	doSubmit(task);
1293	–	return task;
1294	–	}
1295	–
1296	–	/**
2412		* @throws NullPointerException       {@inheritDoc}
2413		* @throws RejectedExecutionException {@inheritDoc}
2414		*/
2415		public <T> List<Future<T>> invokeAll(Collection<? extends Callable<T>> tasks) {
2416	<	ArrayList<ForkJoinTask<T>> forkJoinTasks =
2417	<	new ArrayList<ForkJoinTask<T>>(tasks.size());
2418	<	for (Callable<T> task : tasks)
2419	<	forkJoinTasks.add(ForkJoinTask.adapt(task));
2420	<	invoke(new InvokeAll<T>(forkJoinTasks));
2421	<
2416	>	// In previous versions of this class, this method constructed
2417	>	// a task to run ForkJoinTask.invokeAll, but now external
2418	>	// invocation of multiple tasks is at least as efficient.
2419	>	List<ForkJoinTask<T>> fs = new ArrayList<ForkJoinTask<T>>(tasks.size());
2420	>	// Workaround needed because method wasn't declared with
2421	>	// wildcards in return type but should have been.
2422		@SuppressWarnings({"unchecked", "rawtypes"})
2423	<	List<Future<T>> futures = (List<Future<T>>) (List) forkJoinTasks;
1309	<	return futures;
1310	<	}
2423	>	List<Future<T>> futures = (List<Future<T>>) (List) fs;
2424
2425	<	static final class InvokeAll<T> extends RecursiveAction {
2426	<	final ArrayList<ForkJoinTask<T>> tasks;
2427	<	InvokeAll(ArrayList<ForkJoinTask<T>> tasks) { this.tasks = tasks; }
2428	<	public void compute() {
2429	<	try { invokeAll(tasks); }
2430	<	catch (Exception ignore) {}
2425	>	boolean done = false;
2426	>	try {
2427	>	for (Callable<T> t : tasks) {
2428	>	ForkJoinTask<T> f = new ForkJoinTask.AdaptedCallable<T>(t);
2429	>	doSubmit(f);
2430	>	fs.add(f);
2431	>	}
2432	>	for (ForkJoinTask<T> f : fs)
2433	>	f.quietlyJoin();
2434	>	done = true;
2435	>	return futures;
2436	>	} finally {
2437	>	if (!done)
2438	>	for (ForkJoinTask<T> f : fs)
2439	>	f.cancel(false);
2440		}
1319	–	private static final long serialVersionUID = -7914297376763021607L;
2441		}
2442
2443		/**
#	Line 1335 | Line 2456 | public class ForkJoinPool extends Abstra
2456		* @return the handler, or {@code null} if none
2457		*/
2458		public Thread.UncaughtExceptionHandler getUncaughtExceptionHandler() {
1338	–	workerCountReadFence();
2459		return ueh;
2460		}
2461
2462		/**
2463	<	* Sets the handler for internal worker threads that terminate due
1344	<	* to unrecoverable errors encountered while executing tasks.
1345	<	* Unless set, the current default or ThreadGroup handler is used
1346	<	* as handler.
1347	<	*
1348	<	* @param h the new handler
1349	<	* @return the old handler, or {@code null} if none
1350	<	* @throws SecurityException if a security manager exists and
1351	<	*         the caller is not permitted to modify threads
1352	<	*         because it does not hold {@link
1353	<	*         java.lang.RuntimePermission}{@code ("modifyThread")}
1354	<	*/
1355	<	public Thread.UncaughtExceptionHandler
1356	<	setUncaughtExceptionHandler(Thread.UncaughtExceptionHandler h) {
1357	<	checkPermission();
1358	<	workerCountReadFence();
1359	<	Thread.UncaughtExceptionHandler old = ueh;
1360	<	if (h != old) {
1361	<	ueh = h;
1362	<	workerCountWriteFence();
1363	<	for (ForkJoinWorkerThread w : workers) {
1364	<	if (w != null)
1365	<	w.setUncaughtExceptionHandler(h);
1366	<	}
1367	<	}
1368	<	return old;
1369	<	}
1370	<
1371	<	/**
1372	<	* Sets the target parallelism level of this pool.
2463	>	* Returns the targeted parallelism level of this pool.
2464		*
2465	<	* @param parallelism the target parallelism
1375	<	* @throws IllegalArgumentException if parallelism less than or
1376	<	* equal to zero or greater than maximum size bounds
1377	<	* @throws SecurityException if a security manager exists and
1378	<	*         the caller is not permitted to modify threads
1379	<	*         because it does not hold {@link
1380	<	*         java.lang.RuntimePermission}{@code ("modifyThread")}
2465	>	* @return the targeted parallelism level of this pool
2466		*/
2467	<	public void setParallelism(int parallelism) {
2468	<	checkPermission();
1384	<	if (parallelism <= 0 || parallelism > maxPoolSize)
1385	<	throw new IllegalArgumentException();
1386	<	workerCountReadFence();
1387	<	int pc = this.parallelism;
1388	<	if (pc != parallelism) {
1389	<	this.parallelism = parallelism;
1390	<	workerCountWriteFence();
1391	<	// Release spares. If too many, some will die after re-suspend
1392	<	for (ForkJoinWorkerThread w : workers) {
1393	<	if (w != null && w.tryUnsuspend()) {
1394	<	updateRunningCount(1);
1395	<	LockSupport.unpark(w);
1396	<	}
1397	<	}
1398	<	ensureEnoughTotalWorkers();
1399	<	advanceEventCount();
1400	<	releaseWaiters(); // force config recheck by existing workers
1401	<	}
2467	>	public int getParallelism() {
2468	>	return parallelism;
2469		}
2470
2471		/**
2472	<	* Returns the targeted parallelism level of this pool.
2472	>	* Returns the targeted parallelism level of the common pool.
2473		*
2474	<	* @return the targeted parallelism level of this pool
2474	>	* @return the targeted parallelism level of the common pool
2475		*/
2476	<	public int getParallelism() {
2477	<	//        workerCountReadFence(); // inlined below
1411	<	int ignore = workerCounts;
1412	<	return parallelism;
2476	>	public static int getCommonPoolParallelism() {
2477	>	return commonPoolParallelism;
2478		}
2479
2480		/**
2481		* Returns the number of worker threads that have started but not
2482	<	* yet terminated.  This result returned by this method may differ
2482	>	* yet terminated.  The result returned by this method may differ
2483		* from {@link #getParallelism} when threads are created to
2484		* maintain parallelism when others are cooperatively blocked.
2485		*
2486		* @return the number of worker threads
2487		*/
2488		public int getPoolSize() {
2489	<	return workerCounts >>> TOTAL_COUNT_SHIFT;
1425	<	}
1426	<
1427	<	/**
1428	<	* Returns the maximum number of threads allowed to exist in the
1429	<	* pool. Unless set using {@link #setMaximumPoolSize}, the
1430	<	* maximum is an implementation-defined value designed only to
1431	<	* prevent runaway growth.
1432	<	*
1433	<	* @return the maximum
1434	<	*/
1435	<	public int getMaximumPoolSize() {
1436	<	workerCountReadFence();
1437	<	return maxPoolSize;
1438	<	}
1439	<
1440	<	/**
1441	<	* Sets the maximum number of threads allowed to exist in the
1442	<	* pool. The given value should normally be greater than or equal
1443	<	* to the {@link #getParallelism parallelism} level. Setting this
1444	<	* value has no effect on current pool size. It controls
1445	<	* construction of new threads.
1446	<	*
1447	<	* @throws IllegalArgumentException if negative or greater than
1448	<	* internal implementation limit
1449	<	*/
1450	<	public void setMaximumPoolSize(int newMax) {
1451	<	if (newMax < 0 || newMax > MAX_THREADS)
1452	<	throw new IllegalArgumentException();
1453	<	maxPoolSize = newMax;
1454	<	workerCountWriteFence();
1455	<	}
1456	<
1457	<	/**
1458	<	* Returns {@code true} if this pool dynamically maintains its
1459	<	* target parallelism level. If false, new threads are added only
1460	<	* to avoid possible starvation.  This setting is by default true.
1461	<	*
1462	<	* @return {@code true} if maintains parallelism
1463	<	*/
1464	<	public boolean getMaintainsParallelism() {
1465	<	workerCountReadFence();
1466	<	return maintainsParallelism;
1467	<	}
1468	<
1469	<	/**
1470	<	* Sets whether this pool dynamically maintains its target
1471	<	* parallelism level. If false, new threads are added only to
1472	<	* avoid possible starvation.
1473	<	*
1474	<	* @param enable {@code true} to maintain parallelism
1475	<	*/
1476	<	public void setMaintainsParallelism(boolean enable) {
1477	<	maintainsParallelism = enable;
1478	<	workerCountWriteFence();
1479	<	}
1480	<
1481	<	/**
1482	<	* Establishes local first-in-first-out scheduling mode for forked
1483	<	* tasks that are never joined. This mode may be more appropriate
1484	<	* than default locally stack-based mode in applications in which
1485	<	* worker threads only process asynchronous tasks.  This method is
1486	<	* designed to be invoked only when the pool is quiescent, and
1487	<	* typically only before any tasks are submitted. The effects of
1488	<	* invocations at other times may be unpredictable.
1489	<	*
1490	<	* @param async if {@code true}, use locally FIFO scheduling
1491	<	* @return the previous mode
1492	<	* @see #getAsyncMode
1493	<	*/
1494	<	public boolean setAsyncMode(boolean async) {
1495	<	workerCountReadFence();
1496	<	boolean oldMode = locallyFifo;
1497	<	if (oldMode != async) {
1498	<	locallyFifo = async;
1499	<	workerCountWriteFence();
1500	<	for (ForkJoinWorkerThread w : workers) {
1501	<	if (w != null)
1502	<	w.setAsyncMode(async);
1503	<	}
1504	<	}
1505	<	return oldMode;
2489	>	return parallelism + (short)(ctl >>> TC_SHIFT);
2490		}
2491
2492		/**
#	Line 1510 | Line 2494 | public class ForkJoinPool extends Abstra
2494		* scheduling mode for forked tasks that are never joined.
2495		*
2496		* @return {@code true} if this pool uses async mode
1513	–	* @see #setAsyncMode
2497		*/
2498		public boolean getAsyncMode() {
2499	<	workerCountReadFence();
1517	<	return locallyFifo;
2499	>	return localMode != 0;
2500		}
2501
2502		/**
#	Line 1526 | Line 2508 | public class ForkJoinPool extends Abstra
2508		* @return the number of worker threads
2509		*/
2510		public int getRunningThreadCount() {
2511	<	return workerCounts & RUNNING_COUNT_MASK;
2511	>	int rc = 0;
2512	>	WorkQueue[] ws; WorkQueue w;
2513	>	if ((ws = workQueues) != null) {
2514	>	for (int i = 1; i < ws.length; i += 2) {
2515	>	if ((w = ws[i]) != null && w.isApparentlyUnblocked())
2516	>	++rc;
2517	>	}
2518	>	}
2519	>	return rc;
2520		}
2521
2522		/**
#	Line 1537 | Line 2527 | public class ForkJoinPool extends Abstra
2527		* @return the number of active threads
2528		*/
2529		public int getActiveThreadCount() {
2530	<	return runState & ACTIVE_COUNT_MASK;
2530	>	int r = parallelism + (int)(ctl >> AC_SHIFT);
2531	>	return (r <= 0) ? 0 : r; // suppress momentarily negative values
2532		}
2533
2534		/**
#	Line 1552 | Line 2543 | public class ForkJoinPool extends Abstra
2543		* @return {@code true} if all threads are currently idle
2544		*/
2545		public boolean isQuiescent() {
2546	<	return (runState & ACTIVE_COUNT_MASK) == 0;
2546	>	return (int)(ctl >> AC_SHIFT) + parallelism == 0;
2547		}
2548
2549		/**
#	Line 1567 | Line 2558 | public class ForkJoinPool extends Abstra
2558		* @return the number of steals
2559		*/
2560		public long getStealCount() {
2561	<	return stealCount;
2561	>	long count = stealCount.get();
2562	>	WorkQueue[] ws; WorkQueue w;
2563	>	if ((ws = workQueues) != null) {
2564	>	for (int i = 1; i < ws.length; i += 2) {
2565	>	if ((w = ws[i]) != null)
2566	>	count += w.totalSteals;
2567	>	}
2568	>	}
2569	>	return count;
2570		}
2571
2572		/**
#	Line 1582 | Line 2581 | public class ForkJoinPool extends Abstra
2581		*/
2582		public long getQueuedTaskCount() {
2583		long count = 0;
2584	<	for (ForkJoinWorkerThread w : workers) {
2585	<	if (w != null)
2586	<	count += w.getQueueSize();
2584	>	WorkQueue[] ws; WorkQueue w;
2585	>	if ((ws = workQueues) != null) {
2586	>	for (int i = 1; i < ws.length; i += 2) {
2587	>	if ((w = ws[i]) != null)
2588	>	count += w.queueSize();
2589	>	}
2590		}
2591		return count;
2592		}
2593
2594		/**
2595		* Returns an estimate of the number of tasks submitted to this
2596	<	* pool that have not yet begun executing.  This method takes time
2597	<	* proportional to the number of submissions.
2596	>	* pool that have not yet begun executing.  This method may take
2597	>	* time proportional to the number of submissions.
2598		*
2599		* @return the number of queued submissions
2600		*/
2601		public int getQueuedSubmissionCount() {
2602	<	return submissionQueue.size();
2602	>	int count = 0;
2603	>	WorkQueue[] ws; WorkQueue w;
2604	>	if ((ws = workQueues) != null) {
2605	>	for (int i = 0; i < ws.length; i += 2) {
2606	>	if ((w = ws[i]) != null)
2607	>	count += w.queueSize();
2608	>	}
2609	>	}
2610	>	return count;
2611		}
2612
2613		/**
#	Line 1607 | Line 2617 | public class ForkJoinPool extends Abstra
2617		* @return {@code true} if there are any queued submissions
2618		*/
2619		public boolean hasQueuedSubmissions() {
2620	<	return !submissionQueue.isEmpty();
2620	>	WorkQueue[] ws; WorkQueue w;
2621	>	if ((ws = workQueues) != null) {
2622	>	for (int i = 0; i < ws.length; i += 2) {
2623	>	if ((w = ws[i]) != null && !w.isEmpty())
2624	>	return true;
2625	>	}
2626	>	}
2627	>	return false;
2628		}
2629
2630		/**
#	Line 1618 | Line 2635 | public class ForkJoinPool extends Abstra
2635		* @return the next submission, or {@code null} if none
2636		*/
2637		protected ForkJoinTask<?> pollSubmission() {
2638	<	return submissionQueue.poll();
2638	>	WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2639	>	if ((ws = workQueues) != null) {
2640	>	for (int i = 0; i < ws.length; i += 2) {
2641	>	if ((w = ws[i]) != null && (t = w.poll()) != null)
2642	>	return t;
2643	>	}
2644	>	}
2645	>	return null;
2646		}
2647
2648		/**
#	Line 1639 | Line 2663 | public class ForkJoinPool extends Abstra
2663		* @return the number of elements transferred
2664		*/
2665		protected int drainTasksTo(Collection<? super ForkJoinTask<?>> c) {
2666	<	int n = submissionQueue.drainTo(c);
2667	<	for (ForkJoinWorkerThread w : workers) {
2668	<	if (w != null)
2669	<	n += w.drainTasksTo(c);
2666	>	int count = 0;
2667	>	WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2668	>	if ((ws = workQueues) != null) {
2669	>	for (int i = 0; i < ws.length; ++i) {
2670	>	if ((w = ws[i]) != null) {
2671	>	while ((t = w.poll()) != null) {
2672	>	c.add(t);
2673	>	++count;
2674	>	}
2675	>	}
2676	>	}
2677		}
2678	<	return n;
2678	>	return count;
2679		}
2680
2681		/**
#	Line 1655 | Line 2686 | public class ForkJoinPool extends Abstra
2686		* @return a string identifying this pool, as well as its state
2687		*/
2688		public String toString() {
2689	<	long st = getStealCount();
2690	<	long qt = getQueuedTaskCount();
2691	<	long qs = getQueuedSubmissionCount();
2692	<	int wc = workerCounts;
2693	<	int tc = wc >>> TOTAL_COUNT_SHIFT;
2694	<	int rc = wc & RUNNING_COUNT_MASK;
2689	>	// Use a single pass through workQueues to collect counts
2690	>	long qt = 0L, qs = 0L; int rc = 0;
2691	>	long st = stealCount.get();
2692	>	long c = ctl;
2693	>	WorkQueue[] ws; WorkQueue w;
2694	>	if ((ws = workQueues) != null) {
2695	>	for (int i = 0; i < ws.length; ++i) {
2696	>	if ((w = ws[i]) != null) {
2697	>	int size = w.queueSize();
2698	>	if ((i & 1) == 0)
2699	>	qs += size;
2700	>	else {
2701	>	qt += size;
2702	>	st += w.totalSteals;
2703	>	if (w.isApparentlyUnblocked())
2704	>	++rc;
2705	>	}
2706	>	}
2707	>	}
2708	>	}
2709		int pc = parallelism;
2710	<	int rs = runState;
2711	<	int ac = rs & ACTIVE_COUNT_MASK;
2710	>	int tc = pc + (short)(c >>> TC_SHIFT);
2711	>	int ac = pc + (int)(c >> AC_SHIFT);
2712	>	if (ac < 0) // ignore transient negative
2713	>	ac = 0;
2714	>	String level;
2715	>	if ((c & STOP_BIT) != 0)
2716	>	level = (tc == 0) ? "Terminated" : "Terminating";
2717	>	else
2718	>	level = runState < 0 ? "Shutting down" : "Running";
2719		return super.toString() +
2720	<	"[" + runLevelToString(rs) +
2720	>	"[" + level +
2721		", parallelism = " + pc +
2722		", size = " + tc +
2723		", active = " + ac +
#	Line 1676 | Line 2728 | public class ForkJoinPool extends Abstra
2728		"]";
2729		}
2730
1679	–	private static String runLevelToString(int s) {
1680	–	return ((s & TERMINATED) != 0 ? "Terminated" :
1681	–	((s & TERMINATING) != 0 ? "Terminating" :
1682	–	((s & SHUTDOWN) != 0 ? "Shutting down" :
1683	–	"Running")));
1684	–	}
1685	–
2731		/**
2732	<	* Initiates an orderly shutdown in which previously submitted
2733	<	* tasks are executed, but no new tasks will be accepted.
2734	<	* Invocation has no additional effect if already shut down.
2735	<	* Tasks that are in the process of being submitted concurrently
2736	<	* during the course of this method may or may not be rejected.
2732	>	* Possibly initiates an orderly shutdown in which previously
2733	>	* submitted tasks are executed, but no new tasks will be
2734	>	* accepted. Invocation has no effect on execution state if this
2735	>	* is the {@link #commonPool}, and no additional effect if
2736	>	* already shut down.  Tasks that are in the process of being
2737	>	* submitted concurrently during the course of this method may or
2738	>	* may not be rejected.
2739		*
2740		* @throws SecurityException if a security manager exists and
2741		*         the caller is not permitted to modify threads
#	Line 1697 | Line 2744 | public class ForkJoinPool extends Abstra
2744		*/
2745		public void shutdown() {
2746		checkPermission();
2747	<	advanceRunLevel(SHUTDOWN);
2748	<	tryTerminate(false);
2747	>	if (this != commonPool)
2748	>	tryTerminate(false, true);
2749		}
2750
2751		/**
2752	<	* Attempts to cancel and/or stop all tasks, and reject all
2753	<	* subsequently submitted tasks.  Tasks that are in the process of
2754	<	* being submitted or executed concurrently during the course of
2755	<	* this method may or may not be rejected. This method cancels
2756	<	* both existing and unexecuted tasks, in order to permit
2757	<	* termination in the presence of task dependencies. So the method
2758	<	* always returns an empty list (unlike the case for some other
2759	<	* Executors).
2752	>	* Possibly attempts to cancel and/or stop all tasks, and reject
2753	>	* all subsequently submitted tasks.  Invocation has no effect on
2754	>	* execution state if this is the {@link #commonPool}, and no
2755	>	* additional effect if already shut down. Otherwise, tasks that
2756	>	* are in the process of being submitted or executed concurrently
2757	>	* during the course of this method may or may not be
2758	>	* rejected. This method cancels both existing and unexecuted
2759	>	* tasks, in order to permit termination in the presence of task
2760	>	* dependencies. So the method always returns an empty list
2761	>	* (unlike the case for some other Executors).
2762		*
2763		* @return an empty list
2764		* @throws SecurityException if a security manager exists and
#	Line 1719 | Line 2768 | public class ForkJoinPool extends Abstra
2768		*/
2769		public List<Runnable> shutdownNow() {
2770		checkPermission();
2771	<	tryTerminate(true);
2771	>	if (this != commonPool)
2772	>	tryTerminate(true, true);
2773		return Collections.emptyList();
2774		}
2775
#	Line 1729 | Line 2779 | public class ForkJoinPool extends Abstra
2779		* @return {@code true} if all tasks have completed following shut down
2780		*/
2781		public boolean isTerminated() {
2782	<	return runState >= TERMINATED;
2782	>	long c = ctl;
2783	>	return ((c & STOP_BIT) != 0L &&
2784	>	(short)(c >>> TC_SHIFT) == -parallelism);
2785		}
2786
2787		/**
#	Line 1737 | Line 2789 | public class ForkJoinPool extends Abstra
2789		* commenced but not yet completed.  This method may be useful for
2790		* debugging. A return of {@code true} reported a sufficient
2791		* period after shutdown may indicate that submitted tasks have
2792	<	* ignored or suppressed interruption, causing this executor not
2793	<	* to properly terminate.
2792	>	* ignored or suppressed interruption, or are waiting for IO,
2793	>	* causing this executor not to properly terminate. (See the
2794	>	* advisory notes for class {@link ForkJoinTask} stating that
2795	>	* tasks should not normally entail blocking operations.  But if
2796	>	* they do, they must abort them on interrupt.)
2797		*
2798		* @return {@code true} if terminating but not yet terminated
2799		*/
2800		public boolean isTerminating() {
2801	<	return (runState & (TERMINATING|TERMINATED)) == TERMINATING;
2801	>	long c = ctl;
2802	>	return ((c & STOP_BIT) != 0L &&
2803	>	(short)(c >>> TC_SHIFT) != -parallelism);
2804		}
2805
2806		/**
#	Line 1752 | Line 2809 | public class ForkJoinPool extends Abstra
2809		* @return {@code true} if this pool has been shut down
2810		*/
2811		public boolean isShutdown() {
2812	<	return runState >= SHUTDOWN;
2812	>	return runState < 0;
2813		}
2814
2815		/**
#	Line 1768 | Line 2825 | public class ForkJoinPool extends Abstra
2825		*/
2826		public boolean awaitTermination(long timeout, TimeUnit unit)
2827		throws InterruptedException {
2828	<	return terminationLatch.await(timeout, unit);
2828	>	long nanos = unit.toNanos(timeout);
2829	>	final Mutex lock = this.lock;
2830	>	lock.lock();
2831	>	try {
2832	>	for (;;) {
2833	>	if (isTerminated())
2834	>	return true;
2835	>	if (nanos <= 0)
2836	>	return false;
2837	>	nanos = termination.awaitNanos(nanos);
2838	>	}
2839	>	} finally {
2840	>	lock.unlock();
2841	>	}
2842		}
2843
2844		/**
2845		* Interface for extending managed parallelism for tasks running
2846		* in {@link ForkJoinPool}s.
2847		*
2848	<	* <p>A {@code ManagedBlocker} provides two methods.
2849	<	* Method {@code isReleasable} must return {@code true} if
2850	<	* blocking is not necessary. Method {@code block} blocks the
2851	<	* current thread if necessary (perhaps internally invoking
2852	<	* {@code isReleasable} before actually blocking).
2848	>	* <p>A {@code ManagedBlocker} provides two methods.  Method
2849	>	* {@code isReleasable} must return {@code true} if blocking is
2850	>	* not necessary. Method {@code block} blocks the current thread
2851	>	* if necessary (perhaps internally invoking {@code isReleasable}
2852	>	* before actually blocking). These actions are performed by any
2853	>	* thread invoking {@link ForkJoinPool#managedBlock}.  The
2854	>	* unusual methods in this API accommodate synchronizers that may,
2855	>	* but don't usually, block for long periods. Similarly, they
2856	>	* allow more efficient internal handling of cases in which
2857	>	* additional workers may be, but usually are not, needed to
2858	>	* ensure sufficient parallelism.  Toward this end,
2859	>	* implementations of method {@code isReleasable} must be amenable
2860	>	* to repeated invocation.
2861		*
2862		* <p>For example, here is a ManagedBlocker based on a
2863		* ReentrantLock:
#	Line 1797 | Line 2875 | public class ForkJoinPool extends Abstra
2875		*     return hasLock || (hasLock = lock.tryLock());
2876		*   }
2877		* }}</pre>
2878	+	*
2879	+	* <p>Here is a class that possibly blocks waiting for an
2880	+	* item on a given queue:
2881	+	*  <pre> {@code
2882	+	* class QueueTaker<E> implements ManagedBlocker {
2883	+	*   final BlockingQueue<E> queue;
2884	+	*   volatile E item = null;
2885	+	*   QueueTaker(BlockingQueue<E> q) { this.queue = q; }
2886	+	*   public boolean block() throws InterruptedException {
2887	+	*     if (item == null)
2888	+	*       item = queue.take();
2889	+	*     return true;
2890	+	*   }
2891	+	*   public boolean isReleasable() {
2892	+	*     return item != null || (item = queue.poll()) != null;
2893	+	*   }
2894	+	*   public E getItem() { // call after pool.managedBlock completes
2895	+	*     return item;
2896	+	*   }
2897	+	* }}</pre>
2898		*/
2899		public static interface ManagedBlocker {
2900		/**
#	Line 1820 | Line 2918 | public class ForkJoinPool extends Abstra
2918		* Blocks in accord with the given blocker.  If the current thread
2919		* is a {@link ForkJoinWorkerThread}, this method possibly
2920		* arranges for a spare thread to be activated if necessary to
2921	<	* ensure parallelism while the current thread is blocked.
1824	<	*
1825	<	* <p>If {@code maintainParallelism} is {@code true} and the pool
1826	<	* supports it ({@link #getMaintainsParallelism}), this method
1827	<	* attempts to maintain the pool's nominal parallelism. Otherwise
1828	<	* it activates a thread only if necessary to avoid complete
1829	<	* starvation. This option may be preferable when blockages use
1830	<	* timeouts, or are almost always brief.
2921	>	* ensure sufficient parallelism while the current thread is blocked.
2922		*
2923		* <p>If the caller is not a {@link ForkJoinTask}, this method is
2924		* behaviorally equivalent to
#	Line 1841 | Line 2932 | public class ForkJoinPool extends Abstra
2932		* first be expanded to ensure parallelism, and later adjusted.
2933		*
2934		* @param blocker the blocker
1844	–	* @param maintainParallelism if {@code true} and supported by
1845	–	* this pool, attempt to maintain the pool's nominal parallelism;
1846	–	* otherwise activate a thread only if necessary to avoid
1847	–	* complete starvation.
2935		* @throws InterruptedException if blocker.block did so
2936		*/
2937	<	public static void managedBlock(ManagedBlocker blocker,
1851	<	boolean maintainParallelism)
2937	>	public static void managedBlock(ManagedBlocker blocker)
2938		throws InterruptedException {
2939		Thread t = Thread.currentThread();
2940	<	if (t instanceof ForkJoinWorkerThread)
2941	<	((ForkJoinWorkerThread) t).pool.
2942	<	doBlock(blocker, maintainParallelism);
2943	<	else
2944	<	awaitBlocker(blocker);
2945	<	}
2946	<
2947	<	/**
2948	<	* Performs Non-FJ blocking
2949	<	*/
2950	<	private static void awaitBlocker(ManagedBlocker blocker)
2951	<	throws InterruptedException {
2952	<	do {} while (!blocker.isReleasable() && !blocker.block());
2940	>	ForkJoinPool p = ((t instanceof ForkJoinWorkerThread) ?
2941	>	((ForkJoinWorkerThread)t).pool : null);
2942	>	while (!blocker.isReleasable()) {
2943	>	if (p == null || p.tryCompensate(null, blocker)) {
2944	>	try {
2945	>	do {} while (!blocker.isReleasable() && !blocker.block());
2946	>	} finally {
2947	>	if (p != null)
2948	>	p.incrementActiveCount();
2949	>	}
2950	>	break;
2951	>	}
2952	>	}
2953		}
2954
2955		// AbstractExecutorService overrides.  These rely on undocumented
#	Line 1871 | Line 2957 | public class ForkJoinPool extends Abstra
2957		// implement RunnableFuture.
2958
2959		protected <T> RunnableFuture<T> newTaskFor(Runnable runnable, T value) {
2960	<	return (RunnableFuture<T>) ForkJoinTask.adapt(runnable, value);
2960	>	return new ForkJoinTask.AdaptedRunnable<T>(runnable, value);
2961		}
2962
2963		protected <T> RunnableFuture<T> newTaskFor(Callable<T> callable) {
2964	<	return (RunnableFuture<T>) ForkJoinTask.adapt(callable);
2964	>	return new ForkJoinTask.AdaptedCallable<T>(callable);
2965		}
2966
2967		// Unsafe mechanics
2968	+	private static final sun.misc.Unsafe U;
2969	+	private static final long CTL;
2970	+	private static final long PARKBLOCKER;
2971	+	private static final int ABASE;
2972	+	private static final int ASHIFT;
2973	+
2974	+	static {
2975	+	poolNumberGenerator = new AtomicInteger();
2976	+	nextSubmitterSeed = new AtomicInteger(0x55555555);
2977	+	modifyThreadPermission = new RuntimePermission("modifyThread");
2978	+	defaultForkJoinWorkerThreadFactory =
2979	+	new DefaultForkJoinWorkerThreadFactory();
2980	+	submitters = new ThreadSubmitter();
2981	+	initializationLock = new Mutex();
2982	+	int s;
2983	+	try {
2984	+	U = getUnsafe();
2985	+	Class<?> k = ForkJoinPool.class;
2986	+	Class<?> ak = ForkJoinTask[].class;
2987	+	CTL = U.objectFieldOffset
2988	+	(k.getDeclaredField("ctl"));
2989	+	Class<?> tk = Thread.class;
2990	+	PARKBLOCKER = U.objectFieldOffset
2991	+	(tk.getDeclaredField("parkBlocker"));
2992	+	ABASE = U.arrayBaseOffset(ak);
2993	+	s = U.arrayIndexScale(ak);
2994	+	} catch (Exception e) {
2995	+	throw new Error(e);
2996	+	}
2997	+	if ((s & (s-1)) != 0)
2998	+	throw new Error("data type scale not a power of two");
2999	+	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
3000
3001	<	private static final sun.misc.Unsafe UNSAFE = getUnsafe();
1884	<	private static final long workerCountsOffset =
1885	<	objectFieldOffset("workerCounts", ForkJoinPool.class);
1886	<	private static final long runStateOffset =
1887	<	objectFieldOffset("runState", ForkJoinPool.class);
1888	<	private static final long eventCountOffset =
1889	<	objectFieldOffset("eventCount", ForkJoinPool.class);
1890	<	private static final long eventWaitersOffset =
1891	<	objectFieldOffset("eventWaiters",ForkJoinPool.class);
1892	<	private static final long stealCountOffset =
1893	<	objectFieldOffset("stealCount",ForkJoinPool.class);
1894	<
1895	<
1896	<	private static long objectFieldOffset(String field, Class<?> klazz) {
3001	>	// Establish configuration for default pool
3002		try {
3003	<	return UNSAFE.objectFieldOffset(klazz.getDeclaredField(field));
3004	<	} catch (NoSuchFieldException e) {
3005	<	// Convert Exception to corresponding Error
3006	<	NoSuchFieldError error = new NoSuchFieldError(field);
3007	<	error.initCause(e);
3008	<	throw error;
3003	>	String pp = System.getProperty(propPrefix + "parallelism");
3004	>	String fp = System.getProperty(propPrefix + "threadFactory");
3005	>	String up = System.getProperty(propPrefix + "exceptionHandler");
3006	>	int par;
3007	>	if ((pp == null || (par = Integer.parseInt(pp)) <= 0))
3008	>	par = Runtime.getRuntime().availableProcessors();
3009	>	commonPoolParallelism = par;
3010	>	if (fp != null)
3011	>	commonPoolFactory = (ForkJoinWorkerThreadFactory)
3012	>	ClassLoader.getSystemClassLoader().loadClass(fp).newInstance();
3013	>	else
3014	>	commonPoolFactory = defaultForkJoinWorkerThreadFactory;
3015	>	if (up != null)
3016	>	commonPoolUEH = (Thread.UncaughtExceptionHandler)
3017	>	ClassLoader.getSystemClassLoader().loadClass(up).newInstance();
3018	>	else
3019	>	commonPoolUEH = null;
3020	>	} catch (Exception e) {
3021	>	throw new Error(e);
3022		}
3023		}
3024
#	Line 1931 | Line 3049 | public class ForkJoinPool extends Abstra
3049		}
3050		}
3051		}
3052	+
3053		}

Diff Legend

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