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 1.59 by dl, Fri Jul 23 14:09:17 2010 UTC vs.
Revision 1.165 by dl, Thu Dec 20 17:14:39 2012 UTC

#	Line 1 | Line 1
1		/*
2		* Written by Doug Lea with assistance from members of JCP JSR-166
3		* Expert Group and released to the public domain, as explained at
4	<	* http://creativecommons.org/licenses/publicdomain
4	>	* http://creativecommons.org/publicdomain/zero/1.0/
5		*/
6
7		package jsr166y;
8
9	–	import java.util.concurrent.*;
10	–
9		import java.util.ArrayList;
10		import java.util.Arrays;
11		import java.util.Collection;
12		import java.util.Collections;
13		import java.util.List;
14	<	import java.util.concurrent.locks.LockSupport;
15	<	import java.util.concurrent.locks.ReentrantLock;
16	<	import java.util.concurrent.atomic.AtomicInteger;
17	<	import java.util.concurrent.CountDownLatch;
14	>	import java.util.concurrent.AbstractExecutorService;
15	>	import java.util.concurrent.Callable;
16	>	import java.util.concurrent.ExecutorService;
17	>	import java.util.concurrent.Future;
18	>	import java.util.concurrent.RejectedExecutionException;
19	>	import java.util.concurrent.RunnableFuture;
20	>	import java.util.concurrent.TimeUnit;
21
22		/**
23		* An {@link ExecutorService} for running {@link ForkJoinTask}s.
#	Line 27 | Line 28 | import java.util.concurrent.CountDownLat
28		* <p>A {@code ForkJoinPool} differs from other kinds of {@link
29		* ExecutorService} mainly by virtue of employing
30		* <em>work-stealing</em>: all threads in the pool attempt to find and
31	<	* execute subtasks created by other active tasks (eventually blocking
32	<	* waiting for work if none exist). This enables efficient processing
33	<	* when most tasks spawn other subtasks (as do most {@code
34	<	* ForkJoinTask}s). When setting <em>asyncMode</em> to true in
35	<	* constructors, {@code ForkJoinPool}s may also be appropriate for use
36	<	* with event-style tasks that are never joined.
31	>	* execute tasks submitted to the pool and/or created by other active
32	>	* tasks (eventually blocking waiting for work if none exist). This
33	>	* enables efficient processing when most tasks spawn other subtasks
34	>	* (as do most {@code ForkJoinTask}s), as well as when many small
35	>	* tasks are submitted to the pool from external clients.  Especially
36	>	* when setting <em>asyncMode</em> to true in constructors, {@code
37	>	* ForkJoinPool}s may also be appropriate for use with event-style
38	>	* tasks that are never joined.
39	>	*
40	>	* <p>A static {@link #commonPool()} is available and appropriate for
41	>	* most applications. The common pool is used by any ForkJoinTask that
42	>	* is not explicitly submitted to a specified pool. Using the common
43	>	* pool normally reduces resource usage (its threads are slowly
44	>	* reclaimed during periods of non-use, and reinstated upon subsequent
45	>	* use).
46		*
47	<	* <p>A {@code ForkJoinPool} is constructed with a given target
48	<	* parallelism level; by default, equal to the number of available
49	<	* processors. The pool attempts to maintain enough active (or
50	<	* available) threads by dynamically adding, suspending, or resuming
51	<	* internal worker threads, even if some tasks are stalled waiting to
52	<	* join others. However, no such adjustments are guaranteed in the
53	<	* face of blocked IO or other unmanaged synchronization. The nested
54	<	* {@link ManagedBlocker} interface enables extension of the kinds of
47	>	* <p>For applications that require separate or custom pools, a {@code
48	>	* ForkJoinPool} may be constructed with a given target parallelism
49	>	* level; by default, equal to the number of available processors. The
50	>	* pool attempts to maintain enough active (or available) threads by
51	>	* dynamically adding, suspending, or resuming internal worker
52	>	* threads, even if some tasks are stalled waiting to join
53	>	* others. However, no such adjustments are guaranteed in the face of
54	>	* blocked I/O or other unmanaged synchronization. The nested {@link
55	>	* ManagedBlocker} interface enables extension of the kinds of
56		* synchronization accommodated.
57		*
58		* <p>In addition to execution and lifecycle control methods, this
#	Line 51 | Line 62 | import java.util.concurrent.CountDownLat
62		* {@link #toString} returns indications of pool state in a
63		* convenient form for informal monitoring.
64		*
65	<	* <p> As is the case with other ExecutorServices, there are three
66	<	* main task execution methods summarized in the follwoing
67	<	* table. These are designed to be used by clients not already engaged
68	<	* in fork/join computations in the current pool.  The main forms of
69	<	* these methods accept instances of {@code ForkJoinTask}, but
70	<	* overloaded forms also allow mixed execution of plain {@code
65	>	* <p>As is the case with other ExecutorServices, there are three
66	>	* main task execution methods summarized in the following table.
67	>	* These are designed to be used primarily by clients not already
68	>	* engaged in fork/join computations in the current pool.  The main
69	>	* forms of these methods accept instances of {@code ForkJoinTask},
70	>	* but overloaded forms also allow mixed execution of plain {@code
71		* Runnable}- or {@code Callable}- based activities as well.  However,
72	<	* tasks that are already executing in a pool should normally
73	<	* <em>NOT</em> use these pool execution methods, but instead use the
74	<	* within-computation forms listed in the table.
72	>	* tasks that are already executing in a pool should normally instead
73	>	* use the within-computation forms listed in the table unless using
74	>	* async event-style tasks that are not usually joined, in which case
75	>	* there is little difference among choice of methods.
76		*
77		* <table BORDER CELLPADDING=3 CELLSPACING=1>
78		*  <tr>
#	Line 69 | Line 81 | import java.util.concurrent.CountDownLat
81		*    <td ALIGN=CENTER> <b>Call from within fork/join computations</b></td>
82		*  </tr>
83		*  <tr>
84	<	*    <td> <b>Arange async execution</td>
84	>	*    <td> <b>Arrange async execution</td>
85		*    <td> {@link #execute(ForkJoinTask)}</td>
86		*    <td> {@link ForkJoinTask#fork}</td>
87		*  </tr>
#	Line 85 | Line 97 | import java.util.concurrent.CountDownLat
97		*  </tr>
98		* </table>
99		*
100	<	* <p><b>Sample Usage.</b> Normally a single {@code ForkJoinPool} is
101	<	* used for all parallel task execution in a program or subsystem.
102	<	* Otherwise, use would not usually outweigh the construction and
103	<	* bookkeeping overhead of creating a large set of threads. For
104	<	* example, a common pool could be used for the {@code SortTasks}
105	<	* illustrated in {@link RecursiveAction}. Because {@code
106	<	* ForkJoinPool} uses threads in {@linkplain java.lang.Thread#isDaemon
107	<	* daemon} mode, there is typically no need to explicitly {@link
108	<	* #shutdown} such a pool upon program exit.
109	<	*
98	<	* <pre>
99	<	* static final ForkJoinPool mainPool = new ForkJoinPool();
100	<	* ...
101	<	* public void sort(long[] array) {
102	<	*   mainPool.invoke(new SortTask(array, 0, array.length));
103	<	* }
104	<	* </pre>
100	>	* <p>The common pool is by default constructed with default
101	>	* parameters, but these may be controlled by setting three {@link
102	>	* System#getProperty system properties} with prefix {@code
103	>	* java.util.concurrent.ForkJoinPool.common}: {@code parallelism} --
104	>	* an integer greater than zero, {@code threadFactory} -- the class
105	>	* name of a {@link ForkJoinWorkerThreadFactory}, and {@code
106	>	* exceptionHandler} -- the class name of a {@link
107	>	* java.lang.Thread.UncaughtExceptionHandler
108	>	* Thread.UncaughtExceptionHandler}. Upon any error in establishing
109	>	* these settings, default parameters are used.
110		*
111		* <p><b>Implementation notes</b>: This implementation restricts the
112		* maximum number of running threads to 32767. Attempts to create
#	Line 110 | Line 115 | import java.util.concurrent.CountDownLat
115		*
116		* <p>This implementation rejects submitted tasks (that is, by throwing
117		* {@link RejectedExecutionException}) only when the pool is shut down
118	<	* or internal resources have been exhuasted.
118	>	* or internal resources have been exhausted.
119		*
120		* @since 1.7
121		* @author Doug Lea
#	Line 120 | Line 125 | public class ForkJoinPool extends Abstra
125		/*
126		* Implementation Overview
127		*
128	<	* This class provides the central bookkeeping and control for a
129	<	* set of worker threads: Submissions from non-FJ threads enter
130	<	* into a submission queue. Workers take these tasks and typically
131	<	* split them into subtasks that may be stolen by other workers.
132	<	* The main work-stealing mechanics implemented in class
133	<	* ForkJoinWorkerThread give first priority to processing tasks
134	<	* from their own queues (LIFO or FIFO, depending on mode), then
135	<	* to randomized FIFO steals of tasks in other worker queues, and
136	<	* lastly to new submissions. These mechanics do not consider
137	<	* affinities, loads, cache localities, etc, so rarely provide the
138	<	* best possible performance on a given machine, but portably
139	<	* provide good throughput by averaging over these factors.
140	<	* (Further, even if we did try to use such information, we do not
141	<	* usually have a basis for exploiting it. For example, some sets
142	<	* of tasks profit from cache affinities, but others are harmed by
143	<	* cache pollution effects.)
144	<	*
145	<	* Beyond work-stealing support and essential bookkeeping, the
146	<	* main responsibility of this framework is to arrange tactics for
147	<	* when one worker is waiting to join a task stolen (or always
148	<	* held by) another.  Becauae we are multiplexing many tasks on to
149	<	* a pool of workers, we can't just let them block (as in
150	<	* Thread.join).  We also cannot just reassign the joiner's
151	<	* run-time stack with another and replace it later, which would
152	<	* be a form of "continuation", that even if possible is not
153	<	* necessarily a good idea. Given that the creation costs of most
154	<	* threads on most systems mainly surrounds setting up runtime
155	<	* stacks, thread creation and switching is usually not much more
156	<	* expensive than stack creation and switching, and is more
157	<	* flexible). Instead we combine two tactics:
158	<	*
159	<	*   1. Arranging for the joiner to execute some task that it
160	<	*      would be running if the steal had not occurred.  Method
161	<	*      ForkJoinWorkerThread.helpJoinTask tracks joining->stealing
162	<	*      links to try to find such a task.
163	<	*
164	<	*   2. Unless there are already enough live threads, creating or
165	<	*      or re-activating a spare thread to compensate for the
166	<	*      (blocked) joiner until it unblocks.  Spares then suspend
167	<	*      at their next opportunity or eventually die if unused for
168	<	*      too long.  See below and the internal documentation
169	<	*      for tryAwaitJoin for more details about compensation
170	<	*      rules.
171	<	*
172	<	* Because the determining existence of conservatively safe
173	<	* helping targets, the availability of already-created spares,
174	<	* and the apparent need to create new spares are all racy and
175	<	* require heuristic guidance, joins (in
176	<	* ForkJoinWorkerThread.joinTask) interleave these options until
177	<	* successful.  Creating a new spare always succeeds, but also
178	<	* increases application footprint, so we try to avoid it, within
179	<	* reason.
128	>	* This class and its nested classes provide the main
129	>	* functionality and control for a set of worker threads:
130	>	* Submissions from non-FJ threads enter into submission queues.
131	>	* Workers take these tasks and typically split them into subtasks
132	>	* that may be stolen by other workers.  Preference rules give
133	>	* first priority to processing tasks from their own queues (LIFO
134	>	* or FIFO, depending on mode), then to randomized FIFO steals of
135	>	* tasks in other queues.
136	>	*
137	>	* WorkQueues
138	>	* ==========
139	>	*
140	>	* Most operations occur within work-stealing queues (in nested
141	>	* class WorkQueue).  These are special forms of Deques that
142	>	* support only three of the four possible end-operations -- push,
143	>	* pop, and poll (aka steal), under the further constraints that
144	>	* push and pop are called only from the owning thread (or, as
145	>	* extended here, under a lock), while poll may be called from
146	>	* other threads.  (If you are unfamiliar with them, you probably
147	>	* want to read Herlihy and Shavit's book "The Art of
148	>	* Multiprocessor programming", chapter 16 describing these in
149	>	* more detail before proceeding.)  The main work-stealing queue
150	>	* design is roughly similar to those in the papers "Dynamic
151	>	* Circular Work-Stealing Deque" by Chase and Lev, SPAA 2005
152	>	* (http://research.sun.com/scalable/pubs/index.html) and
153	>	* "Idempotent work stealing" by Michael, Saraswat, and Vechev,
154	>	* PPoPP 2009 (http://portal.acm.org/citation.cfm?id=1504186).
155	>	* The main differences ultimately stem from GC requirements that
156	>	* we null out taken slots as soon as we can, to maintain as small
157	>	* a footprint as possible even in programs generating huge
158	>	* numbers of tasks. To accomplish this, we shift the CAS
159	>	* arbitrating pop vs poll (steal) from being on the indices
160	>	* ("base" and "top") to the slots themselves.  So, both a
161	>	* successful pop and poll mainly entail a CAS of a slot from
162	>	* non-null to null.  Because we rely on CASes of references, we
163	>	* do not need tag bits on base or top.  They are simple ints as
164	>	* used in any circular array-based queue (see for example
165	>	* ArrayDeque).  Updates to the indices must still be ordered in a
166	>	* way that guarantees that top == base means the queue is empty,
167	>	* but otherwise may err on the side of possibly making the queue
168	>	* appear nonempty when a push, pop, or poll have not fully
169	>	* committed. Note that this means that the poll operation,
170	>	* considered individually, is not wait-free. One thief cannot
171	>	* successfully continue until another in-progress one (or, if
172	>	* previously empty, a push) completes.  However, in the
173	>	* aggregate, we ensure at least probabilistic non-blockingness.
174	>	* If an attempted steal fails, a thief always chooses a different
175	>	* random victim target to try next. So, in order for one thief to
176	>	* progress, it suffices for any in-progress poll or new push on
177	>	* any empty queue to complete. (This is why we normally use
178	>	* method pollAt and its variants that try once at the apparent
179	>	* base index, else consider alternative actions, rather than
180	>	* method poll.)
181	>	*
182	>	* This approach also enables support of a user mode in which local
183	>	* task processing is in FIFO, not LIFO order, simply by using
184	>	* poll rather than pop.  This can be useful in message-passing
185	>	* frameworks in which tasks are never joined.  However neither
186	>	* mode considers affinities, loads, cache localities, etc, so
187	>	* rarely provide the best possible performance on a given
188	>	* machine, but portably provide good throughput by averaging over
189	>	* these factors.  (Further, even if we did try to use such
190	>	* information, we do not usually have a basis for exploiting it.
191	>	* For example, some sets of tasks profit from cache affinities,
192	>	* but others are harmed by cache pollution effects.)
193	>	*
194	>	* WorkQueues are also used in a similar way for tasks submitted
195	>	* to the pool. We cannot mix these tasks in the same queues used
196	>	* for work-stealing (this would contaminate lifo/fifo
197	>	* processing). Instead, we randomly associate submission queues
198	>	* with submitting threads, using a form of hashing.  The
199	>	* ThreadLocal Submitter class contains a value initially used as
200	>	* a hash code for choosing existing queues, but may be randomly
201	>	* repositioned upon contention with other submitters.  In
202	>	* essence, submitters act like workers except that they are
203	>	* restricted to executing local tasks that they submitted (or in
204	>	* the case of CountedCompleters, others with the same root task).
205	>	* However, because most shared/external queue operations are more
206	>	* expensive than internal, and because, at steady state, external
207	>	* submitters will compete for CPU with workers, ForkJoinTask.join
208	>	* and related methods disable them from repeatedly helping to
209	>	* process tasks if all workers are active.  Insertion of tasks in
210	>	* shared mode requires a lock (mainly to protect in the case of
211	>	* resizing) but we use only a simple spinlock (using bits in
212	>	* field qlock), because submitters encountering a busy queue move
213	>	* on to try or create other queues -- they block only when
214	>	* creating and registering new queues.
215		*
216	<	* The ManagedBlocker extension API can't use option (1) so uses a
217	<	* special version of (2) in method awaitBlocker.
216	>	* Management
217	>	* ==========
218		*
219		* The main throughput advantages of work-stealing stem from
220	<	* decentralized control -- workers mostly steal tasks from each
221	<	* other. We do not want to negate this by creating bottlenecks
222	<	* implementing other management responsibilities. So we use a
223	<	* collection of techniques that avoid, reduce, or cope well with
224	<	* contention. These entail several instances of bit-packing into
225	<	* CASable fields to maintain only the minimally required
226	<	* atomicity. To enable such packing, we restrict maximum
227	<	* parallelism to (1<<15)-1 (enabling twice this (to accommodate
228	<	* unbalanced increments and decrements) to fit into a 16 bit
229	<	* field, which is far in excess of normal operating range.  Even
230	<	* though updates to some of these bookkeeping fields do sometimes
231	<	* contend with each other, they don't normally cache-contend with
232	<	* updates to others enough to warrant memory padding or
233	<	* isolation. So they are all held as fields of ForkJoinPool
234	<	* objects.  The main capabilities are as follows:
235	<	*
236	<	* 1. Creating and removing workers. Workers are recorded in the
237	<	* "workers" array. This is an array as opposed to some other data
238	<	* structure to support index-based random steals by workers.
239	<	* Updates to the array recording new workers and unrecording
240	<	* terminated ones are protected from each other by a lock
241	<	* (workerLock) but the array is otherwise concurrently readable,
242	<	* and accessed directly by workers. To simplify index-based
243	<	* operations, the array size is always a power of two, and all
244	<	* readers must tolerate null slots. Currently, all worker thread
245	<	* creation is on-demand, triggered by task submissions,
246	<	* replacement of terminated workers, and/or compensation for
247	<	* blocked workers. However, all other support code is set up to
248	<	* work with other policies.
249	<	*
250	<	* 2. Bookkeeping for dynamically adding and removing workers. We
251	<	* aim to approximately maintain the given level of parallelism.
252	<	* When some workers are known to be blocked (on joins or via
253	<	* ManagedBlocker), we may create or resume others to take their
254	<	* place until they unblock (see below). Implementing this
255	<	* requires counts of the number of "running" threads (i.e., those
256	<	* that are neither blocked nor artifically suspended) as well as
257	<	* the total number.  These two values are packed into one field,
258	<	* "workerCounts" because we need accurate snapshots when deciding
259	<	* to create, resume or suspend.  Note however that the
260	<	* correspondance of these counts to reality is not guaranteed. In
261	<	* particular updates for unblocked threads may lag until they
262	<	* actually wake up.
263	<	*
264	<	* 3. Maintaining global run state. The run state of the pool
265	<	* consists of a runLevel (SHUTDOWN, TERMINATING, etc) similar to
266	<	* those in other Executor implementations, as well as a count of
267	<	* "active" workers -- those that are, or soon will be, or
268	<	* recently were executing tasks. The runLevel and active count
269	<	* are packed together in order to correctly trigger shutdown and
270	<	* termination. Without care, active counts can be subject to very
271	<	* high contention.  We substantially reduce this contention by
272	<	* relaxing update rules.  A worker must claim active status
273	<	* prospectively, by activating if it sees that a submitted or
274	<	* stealable task exists (it may find after activating that the
275	<	* task no longer exists). It stays active while processing this
276	<	* task (if it exists) and any other local subtasks it produces,
277	<	* until it cannot find any other tasks. It then tries
278	<	* inactivating (see method preStep), but upon update contention
279	<	* instead scans for more tasks, later retrying inactivation if it
280	<	* doesn't find any.
281	<	*
282	<	* 4. Managing idle workers waiting for tasks. We cannot let
283	<	* workers spin indefinitely scanning for tasks when none are
284	<	* available. On the other hand, we must quickly prod them into
285	<	* action when new tasks are submitted or generated.  We
286	<	* park/unpark these idle workers using an event-count scheme.
287	<	* Field eventCount is incremented upon events that may enable
288	<	* workers that previously could not find a task to now find one:
289	<	* Submission of a new task to the pool, or another worker pushing
290	<	* a task onto a previously empty queue.  (We also use this
291	<	* mechanism for termination and reconfiguration actions that
292	<	* require wakeups of idle workers).  Each worker maintains its
293	<	* last known event count, and blocks when a scan for work did not
294	<	* find a task AND its lastEventCount matches the current
295	<	* eventCount. Waiting idle workers are recorded in a variant of
296	<	* Treiber stack headed by field eventWaiters which, when nonzero,
297	<	* encodes the thread index and count awaited for by the worker
298	<	* thread most recently calling eventSync. This thread in turn has
299	<	* a record (field nextEventWaiter) for the next waiting worker.
300	<	* In addition to allowing simpler decisions about need for
301	<	* wakeup, the event count bits in eventWaiters serve the role of
302	<	* tags to avoid ABA errors in Treiber stacks.  To reduce delays
303	<	* in task diffusion, workers not otherwise occupied may invoke
304	<	* method releaseWaiters, that removes and signals (unparks)
305	<	* workers not waiting on current count. To minimize task
306	<	* production stalls associate with signalling, any worker pushing
307	<	* a task on an empty queue invokes the weaker method signalWork,
308	<	* that only releases idle workers until it detects interference
309	<	* by other threads trying to release, and lets them take
310	<	* over. The net effect is a tree-like diffusion of signals, where
311	<	* released threads (and possibly others) help with unparks.  To
312	<	* further reduce contention effects a bit, failed CASes to
313	<	* increment field eventCount are tolerated without retries.
314	<	* Conceptually they are merged into the same event, which is OK
315	<	* when their only purpose is to enable workers to scan for work.
316	<	*
317	<	* 5. Managing suspension of extra workers. When a worker is about
318	<	* to block waiting for a join (or via ManagedBlockers), we may
319	<	* create a new thread to maintain parallelism level, or at least
320	<	* avoid starvation. Usually, extra threads are needed for only
321	<	* very short periods, yet join dependencies are such that we
322	<	* sometimes need them in bursts. Rather than create new threads
323	<	* each time this happens, we suspend no-longer-needed extra ones
324	<	* as "spares". For most purposes, we don't distinguish "extra"
325	<	* spare threads from normal "core" threads: On each call to
326	<	* preStep (the only point at which we can do this) a worker
327	<	* checks to see if there are now too many running workers, and if
328	<	* so, suspends itself.  Methods tryAwaitJoin and awaitBlocker
329	<	* look for suspended threads to resume before considering
330	<	* creating a new replacement. We don't need a special data
331	<	* structure to maintain spares; simply scanning the workers array
332	<	* looking for worker.isSuspended() is fine because the calling
333	<	* thread is otherwise not doing anything useful anyway; we are at
334	<	* least as happy if after locating a spare, the caller doesn't
335	<	* actually block because the join is ready before we try to
336	<	* adjust and compensate.  Note that this is intrinsically racy.
337	<	* One thread may become a spare at about the same time as another
338	<	* is needlessly being created. We counteract this and related
339	<	* slop in part by requiring resumed spares to immediately recheck
340	<	* (in preStep) to see whether they they should re-suspend. The
341	<	* only effective difference between "extra" and "core" threads is
342	<	* that we allow the "extra" ones to time out and die if they are
343	<	* not resumed within a keep-alive interval of a few seconds. This
344	<	* is implemented mainly within ForkJoinWorkerThread, but requires
345	<	* some coordination (isTrimmed() -- meaning killed while
346	<	* suspended) to correctly maintain pool counts.
347	<	*
348	<	* 6. Deciding when to create new workers. The main dynamic
349	<	* control in this class is deciding when to create extra threads,
350	<	* in methods awaitJoin and awaitBlocker. We always need to create
351	<	* one when the number of running threads would become zero and
352	<	* all workers are busy. However, this is not easy to detect
353	<	* reliably in the presence of transients so we use retries and
354	<	* allow slack (in tryAwaitJoin) to reduce false alarms.  These
355	<	* effectively reduce churn at the price of systematically
356	<	* undershooting target parallelism when many threads are blocked.
357	<	* However, biasing toward undeshooting partially compensates for
358	<	* the above mechanics to suspend extra threads, that normally
359	<	* lead to overshoot because we can only suspend workers
360	<	* in-between top-level actions. It also better copes with the
361	<	* fact that some of the methods in this class tend to never
362	<	* become compiled (but are interpreted), so some components of
363	<	* the entire set of controls might execute many times faster than
364	<	* others. And similarly for cases where the apparent lack of work
365	<	* is just due to GC stalls and other transient system activity.
366	<	*
367	<	* Beware that there is a lot of representation-level coupling
368	<	* among classes ForkJoinPool, ForkJoinWorkerThread, and
369	<	* ForkJoinTask.  For example, direct access to "workers" array by
370	<	* workers, and direct access to ForkJoinTask.status by both
371	<	* ForkJoinPool and ForkJoinWorkerThread.  There is little point
220	>	* decentralized control -- workers mostly take tasks from
221	>	* themselves or each other. We cannot negate this in the
222	>	* implementation of other management responsibilities. The main
223	>	* tactic for avoiding bottlenecks is packing nearly all
224	>	* essentially atomic control state into two volatile variables
225	>	* that are by far most often read (not written) as status and
226	>	* consistency checks.
227	>	*
228	>	* Field "ctl" contains 64 bits holding all the information needed
229	>	* to atomically decide to add, inactivate, enqueue (on an event
230	>	* queue), dequeue, and/or re-activate workers.  To enable this
231	>	* packing, we restrict maximum parallelism to (1<<15)-1 (which is
232	>	* far in excess of normal operating range) to allow ids, counts,
233	>	* and their negations (used for thresholding) to fit into 16bit
234	>	* fields.
235	>	*
236	>	* Field "plock" is a form of sequence lock with a saturating
237	>	* shutdown bit (similarly for per-queue "qlocks"), mainly
238	>	* protecting updates to the workQueues array, as well as to
239	>	* enable shutdown.  When used as a lock, it is normally only very
240	>	* briefly held, so is nearly always available after at most a
241	>	* brief spin, but we use a monitor-based backup strategy to
242	>	* block when needed.
243	>	*
244	>	* Recording WorkQueues.  WorkQueues are recorded in the
245	>	* "workQueues" array that is created upon first use and expanded
246	>	* if necessary.  Updates to the array while recording new workers
247	>	* and unrecording terminated ones are protected from each other
248	>	* by a lock but the array is otherwise concurrently readable, and
249	>	* accessed directly.  To simplify index-based operations, the
250	>	* array size is always a power of two, and all readers must
251	>	* tolerate null slots. Worker queues are at odd indices. Shared
252	>	* (submission) queues are at even indices, up to a maximum of 64
253	>	* slots, to limit growth even if array needs to expand to add
254	>	* more workers. Grouping them together in this way simplifies and
255	>	* speeds up task scanning.
256	>	*
257	>	* All worker thread creation is on-demand, triggered by task
258	>	* submissions, replacement of terminated workers, and/or
259	>	* compensation for blocked workers. However, all other support
260	>	* code is set up to work with other policies.  To ensure that we
261	>	* do not hold on to worker references that would prevent GC, ALL
262	>	* accesses to workQueues are via indices into the workQueues
263	>	* array (which is one source of some of the messy code
264	>	* constructions here). In essence, the workQueues array serves as
265	>	* a weak reference mechanism. Thus for example the wait queue
266	>	* field of ctl stores indices, not references.  Access to the
267	>	* workQueues in associated methods (for example signalWork) must
268	>	* both index-check and null-check the IDs. All such accesses
269	>	* ignore bad IDs by returning out early from what they are doing,
270	>	* since this can only be associated with termination, in which
271	>	* case it is OK to give up.  All uses of the workQueues array
272	>	* also check that it is non-null (even if previously
273	>	* non-null). This allows nulling during termination, which is
274	>	* currently not necessary, but remains an option for
275	>	* resource-revocation-based shutdown schemes. It also helps
276	>	* reduce JIT issuance of uncommon-trap code, which tends to
277	>	* unnecessarily complicate control flow in some methods.
278	>	*
279	>	* Event Queuing. Unlike HPC work-stealing frameworks, we cannot
280	>	* let workers spin indefinitely scanning for tasks when none can
281	>	* be found immediately, and we cannot start/resume workers unless
282	>	* there appear to be tasks available.  On the other hand, we must
283	>	* quickly prod them into action when new tasks are submitted or
284	>	* generated. In many usages, ramp-up time to activate workers is
285	>	* the main limiting factor in overall performance (this is
286	>	* compounded at program start-up by JIT compilation and
287	>	* allocation). So we try to streamline this as much as possible.
288	>	* We park/unpark workers after placing in an event wait queue
289	>	* when they cannot find work. This "queue" is actually a simple
290	>	* Treiber stack, headed by the "id" field of ctl, plus a 15bit
291	>	* counter value (that reflects the number of times a worker has
292	>	* been inactivated) to avoid ABA effects (we need only as many
293	>	* version numbers as worker threads). Successors are held in
294	>	* field WorkQueue.nextWait.  Queuing deals with several intrinsic
295	>	* races, mainly that a task-producing thread can miss seeing (and
296	>	* signalling) another thread that gave up looking for work but
297	>	* has not yet entered the wait queue. We solve this by requiring
298	>	* a full sweep of all workers (via repeated calls to method
299	>	* scan()) both before and after a newly waiting worker is added
300	>	* to the wait queue. During a rescan, the worker might release
301	>	* some other queued worker rather than itself, which has the same
302	>	* net effect. Because enqueued workers may actually be rescanning
303	>	* rather than waiting, we set and clear the "parker" field of
304	>	* WorkQueues to reduce unnecessary calls to unpark.  (This
305	>	* requires a secondary recheck to avoid missed signals.)  Note
306	>	* the unusual conventions about Thread.interrupts surrounding
307	>	* parking and other blocking: Because interrupts are used solely
308	>	* to alert threads to check termination, which is checked anyway
309	>	* upon blocking, we clear status (using Thread.interrupted)
310	>	* before any call to park, so that park does not immediately
311	>	* return due to status being set via some other unrelated call to
312	>	* interrupt in user code.
313	>	*
314	>	* Signalling.  We create or wake up workers only when there
315	>	* appears to be at least one task they might be able to find and
316	>	* execute. However, many other threads may notice the same task
317	>	* and each signal to wake up a thread that might take it. So in
318	>	* general, pools will be over-signalled.  When a submission is
319	>	* added or another worker adds a task to a queue that has fewer
320	>	* than two tasks, they signal waiting workers (or trigger
321	>	* creation of new ones if fewer than the given parallelism level
322	>	* -- signalWork), and may leave a hint to the unparked worker to
323	>	* help signal others upon wakeup).  These primary signals are
324	>	* buttressed by others (see method helpSignal) whenever other
325	>	* threads scan for work or do not have a task to process.  On
326	>	* most platforms, signalling (unpark) overhead time is noticeably
327	>	* long, and the time between signalling a thread and it actually
328	>	* making progress can be very noticeably long, so it is worth
329	>	* offloading these delays from critical paths as much as
330	>	* possible.
331	>	*
332	>	* Trimming workers. To release resources after periods of lack of
333	>	* use, a worker starting to wait when the pool is quiescent will
334	>	* time out and terminate if the pool has remained quiescent for a
335	>	* given period -- a short period if there are more threads than
336	>	* parallelism, longer as the number of threads decreases. This
337	>	* will slowly propagate, eventually terminating all workers after
338	>	* periods of non-use.
339	>	*
340	>	* Shutdown and Termination. A call to shutdownNow atomically sets
341	>	* a plock bit and then (non-atomically) sets each worker's
342	>	* qlock status, cancels all unprocessed tasks, and wakes up
343	>	* all waiting workers.  Detecting whether termination should
344	>	* commence after a non-abrupt shutdown() call requires more work
345	>	* and bookkeeping. We need consensus about quiescence (i.e., that
346	>	* there is no more work). The active count provides a primary
347	>	* indication but non-abrupt shutdown still requires a rechecking
348	>	* scan for any workers that are inactive but not queued.
349	>	*
350	>	* Joining Tasks
351	>	* =============
352	>	*
353	>	* Any of several actions may be taken when one worker is waiting
354	>	* to join a task stolen (or always held) by another.  Because we
355	>	* are multiplexing many tasks on to a pool of workers, we can't
356	>	* just let them block (as in Thread.join).  We also cannot just
357	>	* reassign the joiner's run-time stack with another and replace
358	>	* it later, which would be a form of "continuation", that even if
359	>	* possible is not necessarily a good idea since we sometimes need
360	>	* both an unblocked task and its continuation to progress.
361	>	* Instead we combine two tactics:
362	>	*
363	>	*   Helping: Arranging for the joiner to execute some task that it
364	>	*      would be running if the steal had not occurred.
365	>	*
366	>	*   Compensating: Unless there are already enough live threads,
367	>	*      method tryCompensate() may create or re-activate a spare
368	>	*      thread to compensate for blocked joiners until they unblock.
369	>	*
370	>	* A third form (implemented in tryRemoveAndExec) amounts to
371	>	* helping a hypothetical compensator: If we can readily tell that
372	>	* a possible action of a compensator is to steal and execute the
373	>	* task being joined, the joining thread can do so directly,
374	>	* without the need for a compensation thread (although at the
375	>	* expense of larger run-time stacks, but the tradeoff is
376	>	* typically worthwhile).
377	>	*
378	>	* The ManagedBlocker extension API can't use helping so relies
379	>	* only on compensation in method awaitBlocker.
380	>	*
381	>	* The algorithm in tryHelpStealer entails a form of "linear"
382	>	* helping: Each worker records (in field currentSteal) the most
383	>	* recent task it stole from some other worker. Plus, it records
384	>	* (in field currentJoin) the task it is currently actively
385	>	* joining. Method tryHelpStealer uses these markers to try to
386	>	* find a worker to help (i.e., steal back a task from and execute
387	>	* it) that could hasten completion of the actively joined task.
388	>	* In essence, the joiner executes a task that would be on its own
389	>	* local deque had the to-be-joined task not been stolen. This may
390	>	* be seen as a conservative variant of the approach in Wagner &
391	>	* Calder "Leapfrogging: a portable technique for implementing
392	>	* efficient futures" SIGPLAN Notices, 1993
393	>	* (http://portal.acm.org/citation.cfm?id=155354). It differs in
394	>	* that: (1) We only maintain dependency links across workers upon
395	>	* steals, rather than use per-task bookkeeping.  This sometimes
396	>	* requires a linear scan of workQueues array to locate stealers,
397	>	* but often doesn't because stealers leave hints (that may become
398	>	* stale/wrong) of where to locate them.  It is only a hint
399	>	* because a worker might have had multiple steals and the hint
400	>	* records only one of them (usually the most current).  Hinting
401	>	* isolates cost to when it is needed, rather than adding to
402	>	* per-task overhead.  (2) It is "shallow", ignoring nesting and
403	>	* potentially cyclic mutual steals.  (3) It is intentionally
404	>	* racy: field currentJoin is updated only while actively joining,
405	>	* which means that we miss links in the chain during long-lived
406	>	* tasks, GC stalls etc (which is OK since blocking in such cases
407	>	* is usually a good idea).  (4) We bound the number of attempts
408	>	* to find work (see MAX_HELP) and fall back to suspending the
409	>	* worker and if necessary replacing it with another.
410	>	*
411	>	* Helping actions for CountedCompleters are much simpler: Method
412	>	* helpComplete can take and execute any task with the same root
413	>	* as the task being waited on. However, this still entails some
414	>	* traversal of completer chains, so is less efficient than using
415	>	* CountedCompleters without explicit joins.
416	>	*
417	>	* It is impossible to keep exactly the target parallelism number
418	>	* of threads running at any given time.  Determining the
419	>	* existence of conservatively safe helping targets, the
420	>	* availability of already-created spares, and the apparent need
421	>	* to create new spares are all racy, so we rely on multiple
422	>	* retries of each.  Compensation in the apparent absence of
423	>	* helping opportunities is challenging to control on JVMs, where
424	>	* GC and other activities can stall progress of tasks that in
425	>	* turn stall out many other dependent tasks, without us being
426	>	* able to determine whether they will ever require compensation.
427	>	* Even though work-stealing otherwise encounters little
428	>	* degradation in the presence of more threads than cores,
429	>	* aggressively adding new threads in such cases entails risk of
430	>	* unwanted positive feedback control loops in which more threads
431	>	* cause more dependent stalls (as well as delayed progress of
432	>	* unblocked threads to the point that we know they are available)
433	>	* leading to more situations requiring more threads, and so
434	>	* on. This aspect of control can be seen as an (analytically
435	>	* intractable) game with an opponent that may choose the worst
436	>	* (for us) active thread to stall at any time.  We take several
437	>	* precautions to bound losses (and thus bound gains), mainly in
438	>	* methods tryCompensate and awaitJoin.
439	>	*
440	>	* Common Pool
441	>	* ===========
442	>	*
443	>	* The static commonPool always exists after static
444	>	* initialization.  Since it (or any other created pool) need
445	>	* never be used, we minimize initial construction overhead and
446	>	* footprint to the setup of about a dozen fields, with no nested
447	>	* allocation. Most bootstrapping occurs within method
448	>	* fullExternalPush during the first submission to the pool.
449	>	*
450	>	* When external threads submit to the common pool, they can
451	>	* perform some subtask processing (see externalHelpJoin and
452	>	* related methods).  We do not need to record whether these
453	>	* submissions are to the common pool -- if not, externalHelpJoin
454	>	* returns quickly (at the most helping to signal some common pool
455	>	* workers). These submitters would otherwise be blocked waiting
456	>	* for completion, so the extra effort (with liberally sprinkled
457	>	* task status checks) in inapplicable cases amounts to an odd
458	>	* form of limited spin-wait before blocking in ForkJoinTask.join.
459	>	*
460	>	* Style notes
461	>	* ===========
462	>	*
463	>	* There is a lot of representation-level coupling among classes
464	>	* ForkJoinPool, ForkJoinWorkerThread, and ForkJoinTask.  The
465	>	* fields of WorkQueue maintain data structures managed by
466	>	* ForkJoinPool, so are directly accessed.  There is little point
467		* trying to reduce this, since any associated future changes in
468		* representations will need to be accompanied by algorithmic
469	<	* changes anyway.
470	<	*
471	<	* Style notes: There are lots of inline assignments (of form
472	<	* "while ((local = field) != 0)") which are usually the simplest
473	<	* way to ensure read orderings. Also several occurrences of the
474	<	* unusual "do {} while(!cas...)" which is the simplest way to
475	<	* force an update of a CAS'ed variable. There are also other
476	<	* coding oddities that help some methods perform reasonably even
477	<	* when interpreted (not compiled), at the expense of messiness.
478	<	*
479	<	* The order of declarations in this file is: (1) statics (2)
480	<	* fields (along with constants used when unpacking some of them)
481	<	* (3) internal control methods (4) callbacks and other support
482	<	* for ForkJoinTask and ForkJoinWorkerThread classes, (5) exported
483	<	* methods (plus a few little helpers).
469	>	* changes anyway. Several methods intrinsically sprawl because
470	>	* they must accumulate sets of consistent reads of volatiles held
471	>	* in local variables.  Methods signalWork() and scan() are the
472	>	* main bottlenecks, so are especially heavily
473	>	* micro-optimized/mangled.  There are lots of inline assignments
474	>	* (of form "while ((local = field) != 0)") which are usually the
475	>	* simplest way to ensure the required read orderings (which are
476	>	* sometimes critical). This leads to a "C"-like style of listing
477	>	* declarations of these locals at the heads of methods or blocks.
478	>	* There are several occurrences of the unusual "do {} while
479	>	* (!cas...)"  which is the simplest way to force an update of a
480	>	* CAS'ed variable. There are also other coding oddities (including
481	>	* several unnecessary-looking hoisted null checks) that help
482	>	* some methods perform reasonably even when interpreted (not
483	>	* compiled).
484	>	*
485	>	* The order of declarations in this file is:
486	>	* (1) Static utility functions
487	>	* (2) Nested (static) classes
488	>	* (3) Static fields
489	>	* (4) Fields, along with constants used when unpacking some of them
490	>	* (5) Internal control methods
491	>	* (6) Callbacks and other support for ForkJoinTask methods
492	>	* (7) Exported methods
493	>	* (8) Static block initializing statics in minimally dependent order
494	>	*/
495	>
496	>	// Static utilities
497	>
498	>	/**
499	>	* If there is a security manager, makes sure caller has
500	>	* permission to modify threads.
501		*/
502	+	private static void checkPermission() {
503	+	SecurityManager security = System.getSecurityManager();
504	+	if (security != null)
505	+	security.checkPermission(modifyThreadPermission);
506	+	}
507	+
508	+	// Nested classes
509
510		/**
511		* Factory for creating new {@link ForkJoinWorkerThread}s.
#	Line 368 | Line 527 | public class ForkJoinPool extends Abstra
527		* Default ForkJoinWorkerThreadFactory implementation; creates a
528		* new ForkJoinWorkerThread.
529		*/
530	<	static class DefaultForkJoinWorkerThreadFactory
530	>	static final class DefaultForkJoinWorkerThreadFactory
531		implements ForkJoinWorkerThreadFactory {
532	<	public ForkJoinWorkerThread newThread(ForkJoinPool pool) {
532	>	public final ForkJoinWorkerThread newThread(ForkJoinPool pool) {
533		return new ForkJoinWorkerThread(pool);
534		}
535		}
536
537		/**
538	+	* Per-thread records for threads that submit to pools. Currently
539	+	* holds only pseudo-random seed / index that is used to choose
540	+	* submission queues in method externalPush. In the future, this may
541	+	* also incorporate a means to implement different task rejection
542	+	* and resubmission policies.
543	+	*
544	+	* Seeds for submitters and workers/workQueues work in basically
545	+	* the same way but are initialized and updated using slightly
546	+	* different mechanics. Both are initialized using the same
547	+	* approach as in class ThreadLocal, where successive values are
548	+	* unlikely to collide with previous values. Seeds are then
549	+	* randomly modified upon collisions using xorshifts, which
550	+	* requires a non-zero seed.
551	+	*/
552	+	static final class Submitter {
553	+	int seed;
554	+	Submitter(int s) { seed = s; }
555	+	}
556	+
557	+	/**
558	+	* Class for artificial tasks that are used to replace the target
559	+	* of local joins if they are removed from an interior queue slot
560	+	* in WorkQueue.tryRemoveAndExec. We don't need the proxy to
561	+	* actually do anything beyond having a unique identity.
562	+	*/
563	+	static final class EmptyTask extends ForkJoinTask<Void> {
564	+	private static final long serialVersionUID = -7721805057305804111L;
565	+	EmptyTask() { status = ForkJoinTask.NORMAL; } // force done
566	+	public final Void getRawResult() { return null; }
567	+	public final void setRawResult(Void x) {}
568	+	public final boolean exec() { return true; }
569	+	}
570	+
571	+	/**
572	+	* Queues supporting work-stealing as well as external task
573	+	* submission. See above for main rationale and algorithms.
574	+	* Implementation relies heavily on "Unsafe" intrinsics
575	+	* and selective use of "volatile":
576	+	*
577	+	* Field "base" is the index (mod array.length) of the least valid
578	+	* queue slot, which is always the next position to steal (poll)
579	+	* from if nonempty. Reads and writes require volatile orderings
580	+	* but not CAS, because updates are only performed after slot
581	+	* CASes.
582	+	*
583	+	* Field "top" is the index (mod array.length) of the next queue
584	+	* slot to push to or pop from. It is written only by owner thread
585	+	* for push, or under lock for external/shared push, and accessed
586	+	* by other threads only after reading (volatile) base.  Both top
587	+	* and base are allowed to wrap around on overflow, but (top -
588	+	* base) (or more commonly -(base - top) to force volatile read of
589	+	* base before top) still estimates size. The lock ("qlock") is
590	+	* forced to -1 on termination, causing all further lock attempts
591	+	* to fail. (Note: we don't need CAS for termination state because
592	+	* upon pool shutdown, all shared-queues will stop being used
593	+	* anyway.)  Nearly all lock bodies are set up so that exceptions
594	+	* within lock bodies are "impossible" (modulo JVM errors that
595	+	* would cause failure anyway.)
596	+	*
597	+	* The array slots are read and written using the emulation of
598	+	* volatiles/atomics provided by Unsafe. Insertions must in
599	+	* general use putOrderedObject as a form of releasing store to
600	+	* ensure that all writes to the task object are ordered before
601	+	* its publication in the queue.  All removals entail a CAS to
602	+	* null.  The array is always a power of two. To ensure safety of
603	+	* Unsafe array operations, all accesses perform explicit null
604	+	* checks and implicit bounds checks via power-of-two masking.
605	+	*
606	+	* In addition to basic queuing support, this class contains
607	+	* fields described elsewhere to control execution. It turns out
608	+	* to work better memory-layout-wise to include them in this class
609	+	* rather than a separate class.
610	+	*
611	+	* Performance on most platforms is very sensitive to placement of
612	+	* instances of both WorkQueues and their arrays -- we absolutely
613	+	* do not want multiple WorkQueue instances or multiple queue
614	+	* arrays sharing cache lines. (It would be best for queue objects
615	+	* and their arrays to share, but there is nothing available to
616	+	* help arrange that).  Unfortunately, because they are recorded
617	+	* in a common array, WorkQueue instances are often moved to be
618	+	* adjacent by garbage collectors. To reduce impact, we use field
619	+	* padding that works OK on common platforms; this effectively
620	+	* trades off slightly slower average field access for the sake of
621	+	* avoiding really bad worst-case access. (Until better JVM
622	+	* support is in place, this padding is dependent on transient
623	+	* properties of JVM field layout rules.) We also take care in
624	+	* allocating, sizing and resizing the array. Non-shared queue
625	+	* arrays are initialized by workers before use. Others are
626	+	* allocated on first use.
627	+	*/
628	+	static final class WorkQueue {
629	+	/**
630	+	* Capacity of work-stealing queue array upon initialization.
631	+	* Must be a power of two; at least 4, but should be larger to
632	+	* reduce or eliminate cacheline sharing among queues.
633	+	* Currently, it is much larger, as a partial workaround for
634	+	* the fact that JVMs often place arrays in locations that
635	+	* share GC bookkeeping (especially cardmarks) such that
636	+	* per-write accesses encounter serious memory contention.
637	+	*/
638	+	static final int INITIAL_QUEUE_CAPACITY = 1 << 13;
639	+
640	+	/**
641	+	* Maximum size for queue arrays. Must be a power of two less
642	+	* than or equal to 1 << (31 - width of array entry) to ensure
643	+	* lack of wraparound of index calculations, but defined to a
644	+	* value a bit less than this to help users trap runaway
645	+	* programs before saturating systems.
646	+	*/
647	+	static final int MAXIMUM_QUEUE_CAPACITY = 1 << 26; // 64M
648	+
649	+	// Heuristic padding to ameliorate unfortunate memory placements
650	+	volatile long pad00, pad01, pad02, pad03, pad04, pad05, pad06;
651	+
652	+	int seed;                  // for random scanning; initialize nonzero
653	+	volatile int eventCount;   // encoded inactivation count; < 0 if inactive
654	+	int nextWait;              // encoded record of next event waiter
655	+	int hint;                  // steal or signal hint (index)
656	+	int poolIndex;             // index of this queue in pool (or 0)
657	+	final int mode;            // 0: lifo, > 0: fifo, < 0: shared
658	+	int nsteals;               // number of steals
659	+	volatile int qlock;        // 1: locked, -1: terminate; else 0
660	+	volatile int base;         // index of next slot for poll
661	+	int top;                   // index of next slot for push
662	+	ForkJoinTask<?>[] array;   // the elements (initially unallocated)
663	+	final ForkJoinPool pool;   // the containing pool (may be null)
664	+	final ForkJoinWorkerThread owner; // owning thread or null if shared
665	+	volatile Thread parker;    // == owner during call to park; else null
666	+	volatile ForkJoinTask<?> currentJoin;  // task being joined in awaitJoin
667	+	ForkJoinTask<?> currentSteal; // current non-local task being executed
668	+
669	+	volatile Object pad10, pad11, pad12, pad13, pad14, pad15, pad16, pad17;
670	+	volatile Object pad18, pad19, pad1a, pad1b, pad1c, pad1d;
671	+
672	+	WorkQueue(ForkJoinPool pool, ForkJoinWorkerThread owner, int mode,
673	+	int seed) {
674	+	this.pool = pool;
675	+	this.owner = owner;
676	+	this.mode = mode;
677	+	this.seed = seed;
678	+	// Place indices in the center of array (that is not yet allocated)
679	+	base = top = INITIAL_QUEUE_CAPACITY >>> 1;
680	+	}
681	+
682	+	/**
683	+	* Returns the approximate number of tasks in the queue.
684	+	*/
685	+	final int queueSize() {
686	+	int n = base - top;       // non-owner callers must read base first
687	+	return (n >= 0) ? 0 : -n; // ignore transient negative
688	+	}
689	+
690	+	/**
691	+	* Provides a more accurate estimate of whether this queue has
692	+	* any tasks than does queueSize, by checking whether a
693	+	* near-empty queue has at least one unclaimed task.
694	+	*/
695	+	final boolean isEmpty() {
696	+	ForkJoinTask<?>[] a; int m, s;
697	+	int n = base - (s = top);
698	+	return (n >= 0 ||
699	+	(n == -1 &&
700	+	((a = array) == null ||
701	+	(m = a.length - 1) < 0 ||
702	+	U.getObject
703	+	(a, (long)((m & (s - 1)) << ASHIFT) + ABASE) == null)));
704	+	}
705	+
706	+	/**
707	+	* Pushes a task. Call only by owner in unshared queues.  (The
708	+	* shared-queue version is embedded in method externalPush.)
709	+	*
710	+	* @param task the task. Caller must ensure non-null.
711	+	* @throw RejectedExecutionException if array cannot be resized
712	+	*/
713	+	final void push(ForkJoinTask<?> task) {
714	+	ForkJoinTask<?>[] a; ForkJoinPool p;
715	+	int s = top, m, n;
716	+	if ((a = array) != null) {    // ignore if queue removed
717	+	int j = (((m = a.length - 1) & s) << ASHIFT) + ABASE;
718	+	U.putOrderedObject(a, j, task);
719	+	if ((n = (top = s + 1) - base) <= 2) {
720	+	if ((p = pool) != null)
721	+	p.signalWork(this);
722	+	}
723	+	else if (n >= m)
724	+	growArray();
725	+	}
726	+	}
727	+
728	+	/**
729	+	* Initializes or doubles the capacity of array. Call either
730	+	* by owner or with lock held -- it is OK for base, but not
731	+	* top, to move while resizings are in progress.
732	+	*/
733	+	final ForkJoinTask<?>[] growArray() {
734	+	ForkJoinTask<?>[] oldA = array;
735	+	int size = oldA != null ? oldA.length << 1 : INITIAL_QUEUE_CAPACITY;
736	+	if (size > MAXIMUM_QUEUE_CAPACITY)
737	+	throw new RejectedExecutionException("Queue capacity exceeded");
738	+	int oldMask, t, b;
739	+	ForkJoinTask<?>[] a = array = new ForkJoinTask<?>[size];
740	+	if (oldA != null && (oldMask = oldA.length - 1) >= 0 &&
741	+	(t = top) - (b = base) > 0) {
742	+	int mask = size - 1;
743	+	do {
744	+	ForkJoinTask<?> x;
745	+	int oldj = ((b & oldMask) << ASHIFT) + ABASE;
746	+	int j    = ((b &    mask) << ASHIFT) + ABASE;
747	+	x = (ForkJoinTask<?>)U.getObjectVolatile(oldA, oldj);
748	+	if (x != null &&
749	+	U.compareAndSwapObject(oldA, oldj, x, null))
750	+	U.putObjectVolatile(a, j, x);
751	+	} while (++b != t);
752	+	}
753	+	return a;
754	+	}
755	+
756	+	/**
757	+	* Takes next task, if one exists, in LIFO order.  Call only
758	+	* by owner in unshared queues.
759	+	*/
760	+	final ForkJoinTask<?> pop() {
761	+	ForkJoinTask<?>[] a; ForkJoinTask<?> t; int m;
762	+	if ((a = array) != null && (m = a.length - 1) >= 0) {
763	+	for (int s; (s = top - 1) - base >= 0;) {
764	+	long j = ((m & s) << ASHIFT) + ABASE;
765	+	if ((t = (ForkJoinTask<?>)U.getObject(a, j)) == null)
766	+	break;
767	+	if (U.compareAndSwapObject(a, j, t, null)) {
768	+	top = s;
769	+	return t;
770	+	}
771	+	}
772	+	}
773	+	return null;
774	+	}
775	+
776	+	/**
777	+	* Takes a task in FIFO order if b is base of queue and a task
778	+	* can be claimed without contention. Specialized versions
779	+	* appear in ForkJoinPool methods scan and tryHelpStealer.
780	+	*/
781	+	final ForkJoinTask<?> pollAt(int b) {
782	+	ForkJoinTask<?> t; ForkJoinTask<?>[] a;
783	+	if ((a = array) != null) {
784	+	int j = (((a.length - 1) & b) << ASHIFT) + ABASE;
785	+	if ((t = (ForkJoinTask<?>)U.getObjectVolatile(a, j)) != null &&
786	+	base == b &&
787	+	U.compareAndSwapObject(a, j, t, null)) {
788	+	base = b + 1;
789	+	return t;
790	+	}
791	+	}
792	+	return null;
793	+	}
794	+
795	+	/**
796	+	* Takes next task, if one exists, in FIFO order.
797	+	*/
798	+	final ForkJoinTask<?> poll() {
799	+	ForkJoinTask<?>[] a; int b; ForkJoinTask<?> t;
800	+	while ((b = base) - top < 0 && (a = array) != null) {
801	+	int j = (((a.length - 1) & b) << ASHIFT) + ABASE;
802	+	t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
803	+	if (t != null) {
804	+	if (base == b &&
805	+	U.compareAndSwapObject(a, j, t, null)) {
806	+	base = b + 1;
807	+	return t;
808	+	}
809	+	}
810	+	else if (base == b) {
811	+	if (b + 1 == top)
812	+	break;
813	+	Thread.yield(); // wait for lagging update (very rare)
814	+	}
815	+	}
816	+	return null;
817	+	}
818	+
819	+	/**
820	+	* Takes next task, if one exists, in order specified by mode.
821	+	*/
822	+	final ForkJoinTask<?> nextLocalTask() {
823	+	return mode == 0 ? pop() : poll();
824	+	}
825	+
826	+	/**
827	+	* Returns next task, if one exists, in order specified by mode.
828	+	*/
829	+	final ForkJoinTask<?> peek() {
830	+	ForkJoinTask<?>[] a = array; int m;
831	+	if (a == null || (m = a.length - 1) < 0)
832	+	return null;
833	+	int i = mode == 0 ? top - 1 : base;
834	+	int j = ((i & m) << ASHIFT) + ABASE;
835	+	return (ForkJoinTask<?>)U.getObjectVolatile(a, j);
836	+	}
837	+
838	+	/**
839	+	* Pops the given task only if it is at the current top.
840	+	* (A shared version is available only via FJP.tryExternalUnpush)
841	+	*/
842	+	final boolean tryUnpush(ForkJoinTask<?> t) {
843	+	ForkJoinTask<?>[] a; int s;
844	+	if ((a = array) != null && (s = top) != base &&
845	+	U.compareAndSwapObject
846	+	(a, (((a.length - 1) & --s) << ASHIFT) + ABASE, t, null)) {
847	+	top = s;
848	+	return true;
849	+	}
850	+	return false;
851	+	}
852	+
853	+	/**
854	+	* Removes and cancels all known tasks, ignoring any exceptions.
855	+	*/
856	+	final void cancelAll() {
857	+	ForkJoinTask.cancelIgnoringExceptions(currentJoin);
858	+	ForkJoinTask.cancelIgnoringExceptions(currentSteal);
859	+	for (ForkJoinTask<?> t; (t = poll()) != null; )
860	+	ForkJoinTask.cancelIgnoringExceptions(t);
861	+	}
862	+
863	+	/**
864	+	* Computes next value for random probes.  Scans don't require
865	+	* a very high quality generator, but also not a crummy one.
866	+	* Marsaglia xor-shift is cheap and works well enough.  Note:
867	+	* This is manually inlined in its usages in ForkJoinPool to
868	+	* avoid writes inside busy scan loops.
869	+	*/
870	+	final int nextSeed() {
871	+	int r = seed;
872	+	r ^= r << 13;
873	+	r ^= r >>> 17;
874	+	return seed = r ^= r << 5;
875	+	}
876	+
877	+	// Specialized execution methods
878	+
879	+	/**
880	+	* Pops and runs tasks until empty.
881	+	*/
882	+	private void popAndExecAll() {
883	+	// A bit faster than repeated pop calls
884	+	ForkJoinTask<?>[] a; int m, s; long j; ForkJoinTask<?> t;
885	+	while ((a = array) != null && (m = a.length - 1) >= 0 &&
886	+	(s = top - 1) - base >= 0 &&
887	+	(t = ((ForkJoinTask<?>)
888	+	U.getObject(a, j = ((m & s) << ASHIFT) + ABASE)))
889	+	!= null) {
890	+	if (U.compareAndSwapObject(a, j, t, null)) {
891	+	top = s;
892	+	t.doExec();
893	+	}
894	+	}
895	+	}
896	+
897	+	/**
898	+	* Polls and runs tasks until empty.
899	+	*/
900	+	private void pollAndExecAll() {
901	+	for (ForkJoinTask<?> t; (t = poll()) != null;)
902	+	t.doExec();
903	+	}
904	+
905	+	/**
906	+	* If present, removes from queue and executes the given task,
907	+	* or any other cancelled task. Returns (true) on any CAS
908	+	* or consistency check failure so caller can retry.
909	+	*
910	+	* @return false if no progress can be made, else true;
911	+	*/
912	+	final boolean tryRemoveAndExec(ForkJoinTask<?> task) {
913	+	boolean stat = true, removed = false, empty = true;
914	+	ForkJoinTask<?>[] a; int m, s, b, n;
915	+	if ((a = array) != null && (m = a.length - 1) >= 0 &&
916	+	(n = (s = top) - (b = base)) > 0) {
917	+	for (ForkJoinTask<?> t;;) {           // traverse from s to b
918	+	int j = ((--s & m) << ASHIFT) + ABASE;
919	+	t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
920	+	if (t == null)                    // inconsistent length
921	+	break;
922	+	else if (t == task) {
923	+	if (s + 1 == top) {           // pop
924	+	if (!U.compareAndSwapObject(a, j, task, null))
925	+	break;
926	+	top = s;
927	+	removed = true;
928	+	}
929	+	else if (base == b)           // replace with proxy
930	+	removed = U.compareAndSwapObject(a, j, task,
931	+	new EmptyTask());
932	+	break;
933	+	}
934	+	else if (t.status >= 0)
935	+	empty = false;
936	+	else if (s + 1 == top) {          // pop and throw away
937	+	if (U.compareAndSwapObject(a, j, t, null))
938	+	top = s;
939	+	break;
940	+	}
941	+	if (--n == 0) {
942	+	if (!empty && base == b)
943	+	stat = false;
944	+	break;
945	+	}
946	+	}
947	+	}
948	+	if (removed)
949	+	task.doExec();
950	+	return stat;
951	+	}
952	+
953	+	/**
954	+	* Polls for and executes the given task or any other task in
955	+	* its CountedCompleter computation
956	+	*/
957	+	final boolean pollAndExecCC(ForkJoinTask<?> root) {
958	+	ForkJoinTask<?>[] a; int b; Object o;
959	+	outer: while ((b = base) - top < 0 && (a = array) != null) {
960	+	long j = (((a.length - 1) & b) << ASHIFT) + ABASE;
961	+	if ((o = U.getObject(a, j)) == null ||
962	+	!(o instanceof CountedCompleter))
963	+	break;
964	+	for (CountedCompleter<?> t = (CountedCompleter<?>)o, r = t;;) {
965	+	if (r == root) {
966	+	if (base == b &&
967	+	U.compareAndSwapObject(a, j, t, null)) {
968	+	base = b + 1;
969	+	t.doExec();
970	+	return true;
971	+	}
972	+	else
973	+	break; // restart
974	+	}
975	+	if ((r = r.completer) == null)
976	+	break outer; // not part of root computation
977	+	}
978	+	}
979	+	return false;
980	+	}
981	+
982	+	/**
983	+	* Executes a top-level task and any local tasks remaining
984	+	* after execution.
985	+	*/
986	+	final void runTask(ForkJoinTask<?> t) {
987	+	if (t != null) {
988	+	(currentSteal = t).doExec();
989	+	currentSteal = null;
990	+	++nsteals;
991	+	if (base - top < 0) {       // process remaining local tasks
992	+	if (mode == 0)
993	+	popAndExecAll();
994	+	else
995	+	pollAndExecAll();
996	+	}
997	+	}
998	+	}
999	+
1000	+	/**
1001	+	* Executes a non-top-level (stolen) task.
1002	+	*/
1003	+	final void runSubtask(ForkJoinTask<?> t) {
1004	+	if (t != null) {
1005	+	ForkJoinTask<?> ps = currentSteal;
1006	+	(currentSteal = t).doExec();
1007	+	currentSteal = ps;
1008	+	}
1009	+	}
1010	+
1011	+	/**
1012	+	* Returns true if owned and not known to be blocked.
1013	+	*/
1014	+	final boolean isApparentlyUnblocked() {
1015	+	Thread wt; Thread.State s;
1016	+	return (eventCount >= 0 &&
1017	+	(wt = owner) != null &&
1018	+	(s = wt.getState()) != Thread.State.BLOCKED &&
1019	+	s != Thread.State.WAITING &&
1020	+	s != Thread.State.TIMED_WAITING);
1021	+	}
1022	+
1023	+	// Unsafe mechanics
1024	+	private static final sun.misc.Unsafe U;
1025	+	private static final long QLOCK;
1026	+	private static final int ABASE;
1027	+	private static final int ASHIFT;
1028	+	static {
1029	+	int s;
1030	+	try {
1031	+	U = getUnsafe();
1032	+	Class<?> k = WorkQueue.class;
1033	+	Class<?> ak = ForkJoinTask[].class;
1034	+	QLOCK = U.objectFieldOffset
1035	+	(k.getDeclaredField("qlock"));
1036	+	ABASE = U.arrayBaseOffset(ak);
1037	+	s = U.arrayIndexScale(ak);
1038	+	} catch (Exception e) {
1039	+	throw new Error(e);
1040	+	}
1041	+	if ((s & (s-1)) != 0)
1042	+	throw new Error("data type scale not a power of two");
1043	+	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
1044	+	}
1045	+	}
1046	+
1047	+	// static fields (initialized in static initializer below)
1048	+
1049	+	/**
1050		* Creates a new ForkJoinWorkerThread. This factory is used unless
1051		* overridden in ForkJoinPool constructors.
1052		*/
1053		public static final ForkJoinWorkerThreadFactory
1054	<	defaultForkJoinWorkerThreadFactory =
384	<	new DefaultForkJoinWorkerThreadFactory();
1054	>	defaultForkJoinWorkerThreadFactory;
1055
1056		/**
1057	<	* Permission required for callers of methods that may start or
1058	<	* kill threads.
1057	>	* Per-thread submission bookkeeping. Shared across all pools
1058	>	* to reduce ThreadLocal pollution and because random motion
1059	>	* to avoid contention in one pool is likely to hold for others.
1060	>	* Lazily initialized on first submission (but null-checked
1061	>	* in other contexts to avoid unnecessary initialization).
1062		*/
1063	<	private static final RuntimePermission modifyThreadPermission =
391	<	new RuntimePermission("modifyThread");
1063	>	static final ThreadLocal<Submitter> submitters;
1064
1065		/**
1066	<	* If there is a security manager, makes sure caller has
1067	<	* permission to modify threads.
1066	>	* Permission required for callers of methods that may start or
1067	>	* kill threads.
1068		*/
1069	<	private static void checkPermission() {
398	<	SecurityManager security = System.getSecurityManager();
399	<	if (security != null)
400	<	security.checkPermission(modifyThreadPermission);
401	<	}
1069	>	private static final RuntimePermission modifyThreadPermission;
1070
1071		/**
1072	<	* Generator for assigning sequence numbers as pool names.
1072	>	* Common (static) pool. Non-null for public use unless a static
1073	>	* construction exception, but internal usages null-check on use
1074	>	* to paranoically avoid potential initialization circularities
1075	>	* as well as to simplify generated code.
1076		*/
1077	<	private static final AtomicInteger poolNumberGenerator =
407	<	new AtomicInteger();
1077	>	static final ForkJoinPool commonPool;
1078
1079		/**
1080	<	* Absolute bound for parallelism level. Twice this number must
411	<	* fit into a 16bit field to enable word-packing for some counts.
1080	>	* Common pool parallelism. Must equal commonPool.parallelism.
1081		*/
1082	<	private static final int MAX_THREADS = 0x7fff;
1082	>	static final int commonPoolParallelism;
1083
1084		/**
1085	<	* Array holding all worker threads in the pool.  Array size must
417	<	* be a power of two.  Updates and replacements are protected by
418	<	* workerLock, but the array is always kept in a consistent enough
419	<	* state to be randomly accessed without locking by workers
420	<	* performing work-stealing, as well as other traversal-based
421	<	* methods in this class. All readers must tolerate that some
422	<	* array slots may be null.
1085	>	* Sequence number for creating workerNamePrefix.
1086		*/
1087	<	volatile ForkJoinWorkerThread[] workers;
1087	>	private static int poolNumberSequence;
1088
1089		/**
1090	<	* Queue for external submissions.
1090	>	* Return the next sequence number. We don't expect this to
1091	>	* ever contend so use simple builtin sync.
1092		*/
1093	<	private final LinkedTransferQueue<ForkJoinTask<?>> submissionQueue;
1093	>	private static final synchronized int nextPoolId() {
1094	>	return ++poolNumberSequence;
1095	>	}
1096
1097	<	/**
432	<	* Lock protecting updates to workers array.
433	<	*/
434	<	private final ReentrantLock workerLock;
1097	>	// static constants
1098
1099		/**
1100	<	* Latch released upon termination.
1100	>	* Initial timeout value (in nanoseconds) for the thread
1101	>	* triggering quiescence to park waiting for new work. On timeout,
1102	>	* the thread will instead try to shrink the number of
1103	>	* workers. The value should be large enough to avoid overly
1104	>	* aggressive shrinkage during most transient stalls (long GCs
1105	>	* etc).
1106		*/
1107	<	private final Phaser termination;
1107	>	private static final long IDLE_TIMEOUT      = 2000L * 1000L * 1000L; // 2sec
1108
1109		/**
1110	<	* Creation factory for worker threads.
1110	>	* Timeout value when there are more threads than parallelism level
1111		*/
1112	<	private final ForkJoinWorkerThreadFactory factory;
1112	>	private static final long FAST_IDLE_TIMEOUT =  200L * 1000L * 1000L;
1113
1114		/**
1115	<	* Sum of per-thread steal counts, updated only when threads are
448	<	* idle or terminating.
1115	>	* Tolerance for idle timeouts, to cope with timer undershoots
1116		*/
1117	<	private volatile long stealCount;
1117	>	private static final long TIMEOUT_SLOP = 2000000L;
1118
1119		/**
1120	<	* Encoded record of top of treiber stack of threads waiting for
1121	<	* events. The top 32 bits contain the count being waited for. The
1122	<	* bottom word contains one plus the pool index of waiting worker
1123	<	* thread.
1120	>	* The maximum stolen->joining link depth allowed in method
1121	>	* tryHelpStealer.  Must be a power of two.  Depths for legitimate
1122	>	* chains are unbounded, but we use a fixed constant to avoid
1123	>	* (otherwise unchecked) cycles and to bound staleness of
1124	>	* traversal parameters at the expense of sometimes blocking when
1125	>	* we could be helping.
1126		*/
1127	<	private volatile long eventWaiters;
459	<
460	<	private static final int  EVENT_COUNT_SHIFT = 32;
461	<	private static final long WAITER_ID_MASK = (1L << EVENT_COUNT_SHIFT)-1L;
1127	>	private static final int MAX_HELP = 64;
1128
1129		/**
1130	<	* A counter for events that may wake up worker threads:
1131	<	*   - Submission of a new task to the pool
466	<	*   - A worker pushing a task on an empty queue
467	<	*   - termination and reconfiguration
1130	>	* Increment for seed generators. See class ThreadLocal for
1131	>	* explanation.
1132		*/
1133	<	private volatile int eventCount;
1133	>	private static final int SEED_INCREMENT = 0x61c88647;
1134
1135		/**
1136	<	* Lifecycle control. The low word contains the number of workers
1137	<	* that are (probably) executing tasks. This value is atomically
1138	<	* incremented before a worker gets a task to run, and decremented
1139	<	* when worker has no tasks and cannot find any.  Bits 16-18
1140	<	* contain runLevel value. When all are zero, the pool is
1141	<	* running. Level transitions are monotonic (running -> shutdown
1142	<	* -> terminating -> terminated) so each transition adds a bit.
1143	<	* These are bundled together to ensure consistent read for
1144	<	* termination checks (i.e., that runLevel is at least SHUTDOWN
1145	<	* and active threads is zero).
1136	>	* Bits and masks for control variables
1137	>	*
1138	>	* Field ctl is a long packed with:
1139	>	* AC: Number of active running workers minus target parallelism (16 bits)
1140	>	* TC: Number of total workers minus target parallelism (16 bits)
1141	>	* ST: true if pool is terminating (1 bit)
1142	>	* EC: the wait count of top waiting thread (15 bits)
1143	>	* ID: poolIndex of top of Treiber stack of waiters (16 bits)
1144	>	*
1145	>	* When convenient, we can extract the upper 32 bits of counts and
1146	>	* the lower 32 bits of queue state, u = (int)(ctl >>> 32) and e =
1147	>	* (int)ctl.  The ec field is never accessed alone, but always
1148	>	* together with id and st. The offsets of counts by the target
1149	>	* parallelism and the positionings of fields makes it possible to
1150	>	* perform the most common checks via sign tests of fields: When
1151	>	* ac is negative, there are not enough active workers, when tc is
1152	>	* negative, there are not enough total workers, and when e is
1153	>	* negative, the pool is terminating.  To deal with these possibly
1154	>	* negative fields, we use casts in and out of "short" and/or
1155	>	* signed shifts to maintain signedness.
1156	>	*
1157	>	* When a thread is queued (inactivated), its eventCount field is
1158	>	* set negative, which is the only way to tell if a worker is
1159	>	* prevented from executing tasks, even though it must continue to
1160	>	* scan for them to avoid queuing races. Note however that
1161	>	* eventCount updates lag releases so usage requires care.
1162	>	*
1163	>	* Field plock is an int packed with:
1164	>	* SHUTDOWN: true if shutdown is enabled (1 bit)
1165	>	* SEQ:  a sequence lock, with PL_LOCK bit set if locked (30 bits)
1166	>	* SIGNAL: set when threads may be waiting on the lock (1 bit)
1167	>	*
1168	>	* The sequence number enables simple consistency checks:
1169	>	* Staleness of read-only operations on the workQueues array can
1170	>	* be checked by comparing plock before vs after the reads.
1171		*/
483	–	private volatile int runState;
1172
1173	<	// Note: The order among run level values matters.
1174	<	private static final int RUNLEVEL_SHIFT     = 16;
1175	<	private static final int SHUTDOWN           = 1 << RUNLEVEL_SHIFT;
1176	<	private static final int TERMINATING        = 1 << (RUNLEVEL_SHIFT + 1);
1177	<	private static final int TERMINATED         = 1 << (RUNLEVEL_SHIFT + 2);
490	<	private static final int ACTIVE_COUNT_MASK  = (1 << RUNLEVEL_SHIFT) - 1;
491	<	private static final int ONE_ACTIVE         = 1; // active update delta
1173	>	// bit positions/shifts for fields
1174	>	private static final int  AC_SHIFT   = 48;
1175	>	private static final int  TC_SHIFT   = 32;
1176	>	private static final int  ST_SHIFT   = 31;
1177	>	private static final int  EC_SHIFT   = 16;
1178
1179	<	/**
1180	<	* Holds number of total (i.e., created and not yet terminated)
1181	<	* and running (i.e., not blocked on joins or other managed sync)
1182	<	* threads, packed together to ensure consistent snapshot when
1183	<	* making decisions about creating and suspending spare
1184	<	* threads. Updated only by CAS. Note that adding a new worker
1185	<	* requires incrementing both counts, since workers start off in
500	<	* running state.  This field is also used for memory-fencing
501	<	* configuration parameters.
502	<	*/
503	<	private volatile int workerCounts;
1179	>	// bounds
1180	>	private static final int  SMASK      = 0xffff;  // short bits
1181	>	private static final int  MAX_CAP    = 0x7fff;  // max #workers - 1
1182	>	private static final int  EVENMASK   = 0xfffe;  // even short bits
1183	>	private static final int  SQMASK     = 0x007e;  // max 64 (even) slots
1184	>	private static final int  SHORT_SIGN = 1 << 15;
1185	>	private static final int  INT_SIGN   = 1 << 31;
1186
1187	<	private static final int TOTAL_COUNT_SHIFT  = 16;
1188	<	private static final int RUNNING_COUNT_MASK = (1 << TOTAL_COUNT_SHIFT) - 1;
1189	<	private static final int ONE_RUNNING        = 1;
1190	<	private static final int ONE_TOTAL          = 1 << TOTAL_COUNT_SHIFT;
1187	>	// masks
1188	>	private static final long STOP_BIT   = 0x0001L << ST_SHIFT;
1189	>	private static final long AC_MASK    = ((long)SMASK) << AC_SHIFT;
1190	>	private static final long TC_MASK    = ((long)SMASK) << TC_SHIFT;
1191
1192	<	/**
1193	<	* The target parallelism level.
1194	<	* Accessed directly by ForkJoinWorkerThreads.
513	<	*/
514	<	final int parallelism;
1192	>	// units for incrementing and decrementing
1193	>	private static final long TC_UNIT    = 1L << TC_SHIFT;
1194	>	private static final long AC_UNIT    = 1L << AC_SHIFT;
1195
1196	<	/**
1197	<	* True if use local fifo, not default lifo, for local polling
1198	<	* Read by, and replicated by ForkJoinWorkerThreads
1199	<	*/
1200	<	final boolean locallyFifo;
1196	>	// masks and units for dealing with u = (int)(ctl >>> 32)
1197	>	private static final int  UAC_SHIFT  = AC_SHIFT - 32;
1198	>	private static final int  UTC_SHIFT  = TC_SHIFT - 32;
1199	>	private static final int  UAC_MASK   = SMASK << UAC_SHIFT;
1200	>	private static final int  UTC_MASK   = SMASK << UTC_SHIFT;
1201	>	private static final int  UAC_UNIT   = 1 << UAC_SHIFT;
1202	>	private static final int  UTC_UNIT   = 1 << UTC_SHIFT;
1203
1204	<	/**
1205	<	* The uncaught exception handler used when any worker abruptly
1206	<	* terminates.
525	<	*/
526	<	private final Thread.UncaughtExceptionHandler ueh;
1204	>	// masks and units for dealing with e = (int)ctl
1205	>	private static final int E_MASK      = 0x7fffffff; // no STOP_BIT
1206	>	private static final int E_SEQ       = 1 << EC_SHIFT;
1207
1208	<	/**
1209	<	* Pool number, just for assigning useful names to worker threads
1210	<	*/
1211	<	private final int poolNumber;
1208	>	// plock bits
1209	>	private static final int SHUTDOWN    = 1 << 31;
1210	>	private static final int PL_LOCK     = 2;
1211	>	private static final int PL_SIGNAL   = 1;
1212	>	private static final int PL_SPINS    = 1 << 8;
1213
1214	<	// Utilities for CASing fields. Note that several of these
1215	<	// are manually inlined by callers
1214	>	// access mode for WorkQueue
1215	>	static final int LIFO_QUEUE          =  0;
1216	>	static final int FIFO_QUEUE          =  1;
1217	>	static final int SHARED_QUEUE        = -1;
1218
1219	<	/**
1220	<	* Increments running count.  Also used by ForkJoinTask.
1221	<	*/
539	<	final void incrementRunningCount() {
540	<	int c;
541	<	do {} while (!UNSAFE.compareAndSwapInt(this, workerCountsOffset,
542	<	c = workerCounts,
543	<	c + ONE_RUNNING));
544	<	}
1219	>	// bounds for #steps in scan loop -- must be power 2 minus 1
1220	>	private static final int MIN_SCAN    = 0x1ff;   // cover estimation slop
1221	>	private static final int MAX_SCAN    = 0x1ffff; // 4 * max workers
1222
1223	<	/**
547	<	* Tries to decrement running count unless already zero
548	<	*/
549	<	final boolean tryDecrementRunningCount() {
550	<	int wc = workerCounts;
551	<	if ((wc & RUNNING_COUNT_MASK) == 0)
552	<	return false;
553	<	return UNSAFE.compareAndSwapInt(this, workerCountsOffset,
554	<	wc, wc - ONE_RUNNING);
555	<	}
1223	>	// Instance fields
1224
1225	<	/**
1226	<	* Tries to increment running count
1227	<	*/
1228	<	final boolean tryIncrementRunningCount() {
1229	<	int wc;
1230	<	return UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1231	<	wc = workerCounts, wc + ONE_RUNNING);
1232	<	}
1225	>	/*
1226	>	* Field layout of this class tends to matter more than one would
1227	>	* like. Runtime layout order is only loosely related to
1228	>	* declaration order and may differ across JVMs, but the following
1229	>	* empirically works OK on current JVMs.
1230	>	*/
1231	>
1232	>	// Heuristic padding to ameliorate unfortunate memory placements
1233	>	volatile long pad00, pad01, pad02, pad03, pad04, pad05, pad06;
1234	>
1235	>	volatile long stealCount;                  // collects worker counts
1236	>	volatile long ctl;                         // main pool control
1237	>	volatile int plock;                        // shutdown status and seqLock
1238	>	volatile int indexSeed;                    // worker/submitter index seed
1239	>	final int config;                          // mode and parallelism level
1240	>	WorkQueue[] workQueues;                    // main registry
1241	>	final ForkJoinWorkerThreadFactory factory;
1242	>	final Thread.UncaughtExceptionHandler ueh; // per-worker UEH
1243	>	final String workerNamePrefix;             // to create worker name string
1244
1245	<	/**
1246	<	* Tries incrementing active count; fails on contention.
1247	<	* Called by workers before executing tasks.
1248	<	*
1249	<	* @return true on success
1245	>	volatile Object pad10, pad11, pad12, pad13, pad14, pad15, pad16, pad17;
1246	>	volatile Object pad18, pad19, pad1a, pad1b;
1247	>
1248	>	/*
1249	>	* Acquires the plock lock to protect worker array and related
1250	>	* updates. This method is called only if an initial CAS on plock
1251	>	* fails. This acts as a spinLock for normal cases, but falls back
1252	>	* to builtin monitor to block when (rarely) needed. This would be
1253	>	* a terrible idea for a highly contended lock, but works fine as
1254	>	* a more conservative alternative to a pure spinlock.
1255		*/
1256	<	final boolean tryIncrementActiveCount() {
1257	<	int c;
1258	<	return UNSAFE.compareAndSwapInt(this, runStateOffset,
1259	<	c = runState, c + ONE_ACTIVE);
1256	>	private int acquirePlock() {
1257	>	int spins = PL_SPINS, r = 0, ps, nps;
1258	>	for (;;) {
1259	>	if (((ps = plock) & PL_LOCK) == 0 &&
1260	>	U.compareAndSwapInt(this, PLOCK, ps, nps = ps + PL_LOCK))
1261	>	return nps;
1262	>	else if (r == 0) { // randomize spins if possible
1263	>	Thread t = Thread.currentThread(); WorkQueue w; Submitter z;
1264	>	if ((t instanceof ForkJoinWorkerThread) &&
1265	>	(w = ((ForkJoinWorkerThread)t).workQueue) != null)
1266	>	r = w.seed;
1267	>	else if ((z = submitters.get()) != null)
1268	>	r = z.seed;
1269	>	else
1270	>	r = 1;
1271	>	}
1272	>	else if (spins >= 0) {
1273	>	r ^= r << 1; r ^= r >>> 3; r ^= r << 10; // xorshift
1274	>	if (r >= 0)
1275	>	--spins;
1276	>	}
1277	>	else if (U.compareAndSwapInt(this, PLOCK, ps, ps | PL_SIGNAL)) {
1278	>	synchronized (this) {
1279	>	if ((plock & PL_SIGNAL) != 0) {
1280	>	try {
1281	>	wait();
1282	>	} catch (InterruptedException ie) {
1283	>	try {
1284	>	Thread.currentThread().interrupt();
1285	>	} catch (SecurityException ignore) {
1286	>	}
1287	>	}
1288	>	}
1289	>	else
1290	>	notifyAll();
1291	>	}
1292	>	}
1293	>	}
1294		}
1295
1296		/**
1297	<	* Tries decrementing active count; fails on contention.
1298	<	* Called when workers cannot find tasks to run.
1297	>	* Unlocks and signals any thread waiting for plock. Called only
1298	>	* when CAS of seq value for unlock fails.
1299		*/
1300	<	final boolean tryDecrementActiveCount() {
1301	<	int c;
1302	<	return UNSAFE.compareAndSwapInt(this, runStateOffset,
585	<	c = runState, c - ONE_ACTIVE);
1300	>	private void releasePlock(int ps) {
1301	>	plock = ps;
1302	>	synchronized (this) { notifyAll(); }
1303		}
1304
1305		/**
1306	<	* Advances to at least the given level. Returns true if not
1307	<	* already in at least the given level.
1308	<	*/
1309	<	private boolean advanceRunLevel(int level) {
1310	<	for (;;) {
1311	<	int s = runState;
1312	<	if ((s & level) != 0)
1313	<	return false;
1314	<	if (UNSAFE.compareAndSwapInt(this, runStateOffset, s, s | level))
1315	<	return true;
1306	>	* Tries to create and start one worker if fewer than target
1307	>	* parallelism level exist. Adjusts counts etc on failure.
1308	>	*/
1309	>	private void tryAddWorker() {
1310	>	long c; int u;
1311	>	while ((u = (int)((c = ctl) >>> 32)) < 0 &&
1312	>	(u & SHORT_SIGN) != 0 && (int)c == 0) {
1313	>	long nc = (long)(((u + UTC_UNIT) & UTC_MASK) |
1314	>	((u + UAC_UNIT) & UAC_MASK)) << 32;
1315	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1316	>	ForkJoinWorkerThreadFactory fac;
1317	>	Throwable ex = null;
1318	>	ForkJoinWorkerThread wt = null;
1319	>	try {
1320	>	if ((fac = factory) != null &&
1321	>	(wt = fac.newThread(this)) != null) {
1322	>	wt.start();
1323	>	break;
1324	>	}
1325	>	} catch (Throwable e) {
1326	>	ex = e;
1327	>	}
1328	>	deregisterWorker(wt, ex);
1329	>	break;
1330	>	}
1331		}
1332		}
1333
1334	<	// workers array maintenance
1334	>	//  Registering and deregistering workers
1335
1336		/**
1337	<	* Records and returns a workers array index for new worker.
1338	<	*/
1339	<	private int recordWorker(ForkJoinWorkerThread w) {
1340	<	// Try using slot totalCount-1. If not available, scan and/or resize
1341	<	int k = (workerCounts >>> TOTAL_COUNT_SHIFT) - 1;
1342	<	final ReentrantLock lock = this.workerLock;
1343	<	lock.lock();
1337	>	* Callback from ForkJoinWorkerThread to establish and record its
1338	>	* WorkQueue. To avoid scanning bias due to packing entries in
1339	>	* front of the workQueues array, we treat the array as a simple
1340	>	* power-of-two hash table using per-thread seed as hash,
1341	>	* expanding as needed.
1342	>	*
1343	>	* @param wt the worker thread
1344	>	* @return the worker's queue
1345	>	*/
1346	>	final WorkQueue registerWorker(ForkJoinWorkerThread wt) {
1347	>	Thread.UncaughtExceptionHandler handler; WorkQueue[] ws; int s, ps;
1348	>	wt.setDaemon(true);
1349	>	if ((handler = ueh) != null)
1350	>	wt.setUncaughtExceptionHandler(handler);
1351	>	do {} while (!U.compareAndSwapInt(this, INDEXSEED, s = indexSeed,
1352	>	s += SEED_INCREMENT) ||
1353	>	s == 0); // skip 0
1354	>	WorkQueue w = new WorkQueue(this, wt, config >>> 16, s);
1355	>	if (((ps = plock) & PL_LOCK) != 0 ||
1356	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1357	>	ps = acquirePlock();
1358	>	int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1359		try {
1360	<	ForkJoinWorkerThread[] ws = workers;
1361	<	int nws = ws.length;
1362	<	if (k < 0 || k >= nws || ws[k] != null) {
1363	<	for (k = 0; k < nws && ws[k] != null; ++k)
1364	<	;
1365	<	if (k == nws)
1366	<	ws = Arrays.copyOf(ws, nws << 1);
1360	>	if ((ws = workQueues) != null) {    // skip if shutting down
1361	>	int n = ws.length, m = n - 1;
1362	>	int r = (s << 1) | 1;           // use odd-numbered indices
1363	>	if (ws[r &= m] != null) {       // collision
1364	>	int probes = 0;             // step by approx half size
1365	>	int step = (n <= 4) ? 2 : ((n >>> 1) & EVENMASK) + 2;
1366	>	while (ws[r = (r + step) & m] != null) {
1367	>	if (++probes >= n) {
1368	>	workQueues = ws = Arrays.copyOf(ws, n <<= 1);
1369	>	m = n - 1;
1370	>	probes = 0;
1371	>	}
1372	>	}
1373	>	}
1374	>	w.eventCount = w.poolIndex = r; // volatile write orders
1375	>	ws[r] = w;
1376		}
621	–	ws[k] = w;
622	–	workers = ws; // volatile array write ensures slot visibility
1377		} finally {
1378	<	lock.unlock();
1378	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1379	>	releasePlock(nps);
1380		}
1381	<	return k;
1381	>	wt.setName(workerNamePrefix.concat(Integer.toString(w.poolIndex)));
1382	>	return w;
1383		}
1384
1385		/**
1386	<	* Nulls out record of worker in workers array
1387	<	*/
1388	<	private void forgetWorker(ForkJoinWorkerThread w) {
1389	<	int idx = w.poolIndex;
1390	<	// Locking helps method recordWorker avoid unecessary expansion
1391	<	final ReentrantLock lock = this.workerLock;
1392	<	lock.lock();
1393	<	try {
1394	<	ForkJoinWorkerThread[] ws = workers;
1395	<	if (idx >= 0 && idx < ws.length && ws[idx] == w) // verify
1396	<	ws[idx] = null;
1397	<	} finally {
1398	<	lock.unlock();
1386	>	* Final callback from terminating worker, as well as upon failure
1387	>	* to construct or start a worker.  Removes record of worker from
1388	>	* array, and adjusts counts. If pool is shutting down, tries to
1389	>	* complete termination.
1390	>	*
1391	>	* @param wt the worker thread or null if construction failed
1392	>	* @param ex the exception causing failure, or null if none
1393	>	*/
1394	>	final void deregisterWorker(ForkJoinWorkerThread wt, Throwable ex) {
1395	>	WorkQueue w = null;
1396	>	if (wt != null && (w = wt.workQueue) != null) {
1397	>	int ps;
1398	>	w.qlock = -1;                // ensure set
1399	>	long ns = w.nsteals, sc;     // collect steal count
1400	>	do {} while (!U.compareAndSwapLong(this, STEALCOUNT,
1401	>	sc = stealCount, sc + ns));
1402	>	if (((ps = plock) & PL_LOCK) != 0 ||
1403	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1404	>	ps = acquirePlock();
1405	>	int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1406	>	try {
1407	>	int idx = w.poolIndex;
1408	>	WorkQueue[] ws = workQueues;
1409	>	if (ws != null && idx >= 0 && idx < ws.length && ws[idx] == w)
1410	>	ws[idx] = null;
1411	>	} finally {
1412	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1413	>	releasePlock(nps);
1414	>	}
1415		}
644	–	}
1416
1417	<	// adding and removing workers
1418	<
1419	<	/**
1420	<	* Tries to create and add new worker. Assumes that worker counts
1421	<	* are already updated to accommodate the worker, so adjusts on
1422	<	* failure.
1423	<	*
1424	<	* @return new worker or null if creation failed
1425	<	*/
1426	<	private ForkJoinWorkerThread addWorker() {
1427	<	ForkJoinWorkerThread w = null;
1428	<	try {
1429	<	w = factory.newThread(this);
1430	<	} finally { // Adjust on either null or exceptional factory return
1431	<	if (w == null) {
1432	<	onWorkerCreationFailure();
1433	<	return null;
1417	>	long c;                          // adjust ctl counts
1418	>	do {} while (!U.compareAndSwapLong
1419	>	(this, CTL, c = ctl, (((c - AC_UNIT) & AC_MASK) |
1420	>	((c - TC_UNIT) & TC_MASK) |
1421	>	(c & ~(AC_MASK|TC_MASK)))));
1422	>
1423	>	if (!tryTerminate(false, false) && w != null && w.array != null) {
1424	>	w.cancelAll();               // cancel remaining tasks
1425	>	WorkQueue[] ws; WorkQueue v; Thread p; int u, i, e;
1426	>	while ((u = (int)((c = ctl) >>> 32)) < 0 && (e = (int)c) >= 0) {
1427	>	if (e > 0) {             // activate or create replacement
1428	>	if ((ws = workQueues) == null ||
1429	>	(i = e & SMASK) >= ws.length ||
1430	>	(v = ws[i]) == null)
1431	>	break;
1432	>	long nc = (((long)(v.nextWait & E_MASK)) |
1433	>	((long)(u + UAC_UNIT) << 32));
1434	>	if (v.eventCount != (e | INT_SIGN))
1435	>	break;
1436	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1437	>	v.eventCount = (e + E_SEQ) & E_MASK;
1438	>	if ((p = v.parker) != null)
1439	>	U.unpark(p);
1440	>	break;
1441	>	}
1442	>	}
1443	>	else {
1444	>	if ((short)u < 0)
1445	>	tryAddWorker();
1446	>	break;
1447	>	}
1448		}
1449		}
1450	<	w.start(recordWorker(w), ueh);
1451	<	return w;
1452	<	}
1453	<
1454	<	/**
1455	<	* Adjusts counts upon failure to create worker
1456	<	*/
1457	<	private void onWorkerCreationFailure() {
1458	<	for (;;) {
1459	<	int wc = workerCounts;
1460	<	if ((wc >>> TOTAL_COUNT_SHIFT) == 0)
1461	<	Thread.yield(); // wait for other counts to settle
1462	<	else if (UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
1463	<	wc - (ONE_RUNNING|ONE_TOTAL)))
1464	<	break;
1450	>	if (ex == null)                     // help clean refs on way out
1451	>	ForkJoinTask.helpExpungeStaleExceptions();
1452	>	else                                // rethrow
1453	>	ForkJoinTask.rethrow(ex);
1454	>	}
1455	>
1456	>	// Submissions
1457	>
1458	>	/**
1459	>	* Unless shutting down, adds the given task to a submission queue
1460	>	* at submitter's current queue index (modulo submission
1461	>	* range). Only the most common path is directly handled in this
1462	>	* method. All others are relayed to fullExternalPush.
1463	>	*
1464	>	* @param task the task. Caller must ensure non-null.
1465	>	*/
1466	>	final void externalPush(ForkJoinTask<?> task) {
1467	>	WorkQueue[] ws; WorkQueue q; Submitter z; int m; ForkJoinTask<?>[] a;
1468	>	if ((z = submitters.get()) != null && plock > 0 &&
1469	>	(ws = workQueues) != null && (m = (ws.length - 1)) >= 0 &&
1470	>	(q = ws[m & z.seed & SQMASK]) != null &&
1471	>	U.compareAndSwapInt(q, QLOCK, 0, 1)) { // lock
1472	>	int b = q.base, s = q.top, n, an;
1473	>	if ((a = q.array) != null && (an = a.length) > (n = s + 1 - b)) {
1474	>	int j = (((an - 1) & s) << ASHIFT) + ABASE;
1475	>	U.putOrderedObject(a, j, task);
1476	>	q.top = s + 1;                     // push on to deque
1477	>	q.qlock = 0;
1478	>	if (n <= 2)
1479	>	signalWork(q);
1480	>	return;
1481	>	}
1482	>	q.qlock = 0;
1483		}
1484	<	tryTerminate(false); // in case of failure during shutdown
1484	>	fullExternalPush(task);
1485		}
1486
1487		/**
1488	<	* Creates and/or resumes enough workers to establish target
1489	<	* parallelism, giving up if terminating or addWorker fails
1490	<	*
1491	<	* TODO: recast this to support lazier creation and automated
1492	<	* parallelism maintenance
1493	<	*/
1494	<	private void ensureEnoughWorkers() {
1495	<	while ((runState & TERMINATING) == 0) {
1496	<	int pc = parallelism;
1497	<	int wc = workerCounts;
1498	<	int rc = wc & RUNNING_COUNT_MASK;
1499	<	int tc = wc >>> TOTAL_COUNT_SHIFT;
1500	<	if (tc < pc) {
1501	<	if (UNSAFE.compareAndSwapInt
1502	<	(this, workerCountsOffset,
1503	<	wc, wc + (ONE_RUNNING|ONE_TOTAL)) &&
1504	<	addWorker() == null)
1505	<	break;
1488	>	* Full version of externalPush. This method is called, among
1489	>	* other times, upon the first submission of the first task to the
1490	>	* pool, so must perform secondary initialization.  It also
1491	>	* detects first submission by an external thread by looking up
1492	>	* its ThreadLocal, and creates a new shared queue if the one at
1493	>	* index if empty or contended. The plock lock body must be
1494	>	* exception-free (so no try/finally) so we optimistically
1495	>	* allocate new queues outside the lock and throw them away if
1496	>	* (very rarely) not needed.
1497	>	*
1498	>	* Secondary initialization occurs when plock is zero, to create
1499	>	* workQueue array and set plock to a valid value.  This lock body
1500	>	* must also be exception-free. Because the plock seq value can
1501	>	* eventually wrap around zero, this method harmlessly fails to
1502	>	* reinitialize if workQueues exists, while still advancing plock.
1503	>	*/
1504	>	private void fullExternalPush(ForkJoinTask<?> task) {
1505	>	int r = 0; // random index seed
1506	>	for (Submitter z = submitters.get();;) {
1507	>	WorkQueue[] ws; WorkQueue q; int ps, m, k;
1508	>	if (z == null) {
1509	>	if (U.compareAndSwapInt(this, INDEXSEED, r = indexSeed,
1510	>	r += SEED_INCREMENT) && r != 0)
1511	>	submitters.set(z = new Submitter(r));
1512		}
1513	<	else if (tc > pc && rc < pc &&
1514	<	tc > (runState & ACTIVE_COUNT_MASK)) {
1515	<	ForkJoinWorkerThread spare = null;
1516	<	ForkJoinWorkerThread[] ws = workers;
1517	<	int nws = ws.length;
1518	<	for (int i = 0; i < nws; ++i) {
1519	<	ForkJoinWorkerThread w = ws[i];
1520	<	if (w != null && w.isSuspended()) {
1521	<	if ((workerCounts & RUNNING_COUNT_MASK) > pc)
1522	<	return;
1523	<	if (w.tryResumeSpare())
1524	<	incrementRunningCount();
1525	<	break;
1513	>	else if (r == 0) {                  // move to a different index
1514	>	r = z.seed;
1515	>	r ^= r << 13;                   // same xorshift as WorkQueues
1516	>	r ^= r >>> 17;
1517	>	z.seed = r ^ (r << 5);
1518	>	}
1519	>	else if ((ps = plock) < 0)
1520	>	throw new RejectedExecutionException();
1521	>	else if (ps == 0 || (ws = workQueues) == null ||
1522	>	(m = ws.length - 1) < 0) { // initialize workQueues
1523	>	int p = config & SMASK;         // find power of two table size
1524	>	int n = (p > 1) ? p - 1 : 1;    // ensure at least 2 slots
1525	>	n |= n >>> 1; n |= n >>> 2;  n |= n >>> 4;
1526	>	n |= n >>> 8; n |= n >>> 16; n = (n + 1) << 1;
1527	>	WorkQueue[] nws = ((ws = workQueues) == null || ws.length == 0 ?
1528	>	new WorkQueue[n] : null);
1529	>	if (((ps = plock) & PL_LOCK) != 0 ||
1530	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1531	>	ps = acquirePlock();
1532	>	if (((ws = workQueues) == null || ws.length == 0) && nws != null)
1533	>	workQueues = nws;
1534	>	int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1535	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1536	>	releasePlock(nps);
1537	>	}
1538	>	else if ((q = ws[k = r & m & SQMASK]) != null) {
1539	>	if (q.qlock == 0 && U.compareAndSwapInt(q, QLOCK, 0, 1)) {
1540	>	ForkJoinTask<?>[] a = q.array;
1541	>	int s = q.top;
1542	>	boolean submitted = false;
1543	>	try {                      // locked version of push
1544	>	if ((a != null && a.length > s + 1 - q.base) ||
1545	>	(a = q.growArray()) != null) {   // must presize
1546	>	int j = (((a.length - 1) & s) << ASHIFT) + ABASE;
1547	>	U.putOrderedObject(a, j, task);
1548	>	q.top = s + 1;
1549	>	submitted = true;
1550	>	}
1551	>	} finally {
1552	>	q.qlock = 0;  // unlock
1553	>	}
1554	>	if (submitted) {
1555	>	signalWork(q);
1556	>	return;
1557		}
1558		}
1559	+	r = 0; // move on failure
1560	+	}
1561	+	else if (((ps = plock) & PL_LOCK) == 0) { // create new queue
1562	+	q = new WorkQueue(this, null, SHARED_QUEUE, r);
1563	+	if (((ps = plock) & PL_LOCK) != 0 ||
1564	+	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1565	+	ps = acquirePlock();
1566	+	if ((ws = workQueues) != null && k < ws.length && ws[k] == null)
1567	+	ws[k] = q;
1568	+	int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1569	+	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1570	+	releasePlock(nps);
1571		}
1572		else
1573	<	break;
1573	>	r = 0; // try elsewhere while lock held
1574		}
1575		}
1576
1577	<	/**
726	<	* Final callback from terminating worker.  Removes record of
727	<	* worker from array, and adjusts counts. If pool is shutting
728	<	* down, tries to complete terminatation, else possibly replaces
729	<	* the worker.
730	<	*
731	<	* @param w the worker
732	<	*/
733	<	final void workerTerminated(ForkJoinWorkerThread w) {
734	<	if (w.active) { // force inactive
735	<	w.active = false;
736	<	do {} while (!tryDecrementActiveCount());
737	<	}
738	<	forgetWorker(w);
739	<
740	<	// Decrement total count, and if was running, running count
741	<	// Spin (waiting for other updates) if either would be negative
742	<	int nr = w.isTrimmed() ? 0 : ONE_RUNNING;
743	<	int unit = ONE_TOTAL + nr;
744	<	for (;;) {
745	<	int wc = workerCounts;
746	<	int rc = wc & RUNNING_COUNT_MASK;
747	<	if (rc - nr < 0 || (wc >>> TOTAL_COUNT_SHIFT) == 0)
748	<	Thread.yield(); // back off if waiting for other updates
749	<	else if (UNSAFE.compareAndSwapInt(this, workerCountsOffset,
750	<	wc, wc - unit))
751	<	break;
752	<	}
1577	>	// Maintaining ctl counts
1578
1579	<	accumulateStealCount(w); // collect final count
1580	<	if (!tryTerminate(false))
1581	<	ensureEnoughWorkers();
1579	>	/**
1580	>	* Increments active count; mainly called upon return from blocking.
1581	>	*/
1582	>	final void incrementActiveCount() {
1583	>	long c;
1584	>	do {} while (!U.compareAndSwapLong(this, CTL, c = ctl, c + AC_UNIT));
1585		}
1586
759	–	// Waiting for and signalling events
760	–
1587		/**
1588	<	* Releases workers blocked on a count not equal to current count.
1589	<	* @return true if any released
1588	>	* Tries to create or activate a worker if too few are active.
1589	>	*
1590	>	* @param q the (non-null) queue holding tasks to be signalled
1591		*/
1592	<	private void releaseWaiters() {
1593	<	long top;
1594	<	while ((top = eventWaiters) != 0L) {
1595	<	ForkJoinWorkerThread[] ws = workers;
1596	<	int n = ws.length;
1597	<	for (;;) {
1598	<	int i = ((int)(top & WAITER_ID_MASK)) - 1;
1599	<	if (i < 0 || (int)(top >>> EVENT_COUNT_SHIFT) == eventCount)
1600	<	return;
1601	<	ForkJoinWorkerThread w;
1602	<	if (i < n && (w = ws[i]) != null &&
1603	<	UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
1604	<	top, w.nextWaiter)) {
1605	<	LockSupport.unpark(w);
1606	<	top = eventWaiters;
1592	>	final void signalWork(WorkQueue q) {
1593	>	int hint = q.poolIndex;
1594	>	long c; int e, u, i, n; WorkQueue[] ws; WorkQueue w; Thread p;
1595	>	while ((u = (int)((c = ctl) >>> 32)) < 0) {
1596	>	if ((e = (int)c) > 0) {
1597	>	if ((ws = workQueues) != null && ws.length > (i = e & SMASK) &&
1598	>	(w = ws[i]) != null && w.eventCount == (e | INT_SIGN)) {
1599	>	long nc = (((long)(w.nextWait & E_MASK)) |
1600	>	((long)(u + UAC_UNIT) << 32));
1601	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1602	>	w.hint = hint;
1603	>	w.eventCount = (e + E_SEQ) & E_MASK;
1604	>	if ((p = w.parker) != null)
1605	>	U.unpark(p);
1606	>	break;
1607	>	}
1608	>	if (q.top - q.base <= 0)
1609	>	break;
1610		}
1611		else
1612	<	break;      // possibly stale; reread
1612	>	break;
1613	>	}
1614	>	else {
1615	>	if ((short)u < 0)
1616	>	tryAddWorker();
1617	>	break;
1618		}
1619		}
1620		}
1621
1622	<	/**
788	<	* Ensures eventCount on exit is different (mod 2^32) than on
789	<	* entry and wakes up all waiters
790	<	*/
791	<	private void signalEvent() {
792	<	int c;
793	<	do {} while (!UNSAFE.compareAndSwapInt(this, eventCountOffset,
794	<	c = eventCount, c+1));
795	<	releaseWaiters();
796	<	}
1622	>	// Scanning for tasks
1623
1624		/**
1625	<	* Advances eventCount and releases waiters until interference by
800	<	* other releasing threads is detected.
1625	>	* Top-level runloop for workers, called by ForkJoinWorkerThread.run.
1626		*/
1627	<	final void signalWork() {
1628	<	int c;
1629	<	UNSAFE.compareAndSwapInt(this, eventCountOffset, c=eventCount, c+1);
1630	<	long top;
1631	<	while ((top = eventWaiters) != 0L) {
1632	<	int ec = eventCount;
1633	<	ForkJoinWorkerThread[] ws = workers;
1634	<	int n = ws.length;
1635	<	for (;;) {
1636	<	int i = ((int)(top & WAITER_ID_MASK)) - 1;
1637	<	if (i < 0 || (int)(top >>> EVENT_COUNT_SHIFT) == ec)
1638	<	return;
1639	<	ForkJoinWorkerThread w;
1640	<	if (i < n && (w = ws[i]) != null &&
1641	<	UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
1642	<	top, top = w.nextWaiter)) {
1643	<	LockSupport.unpark(w);
1644	<	if (top != eventWaiters) // let someone else take over
1645	<	return;
1627	>	final void runWorker(WorkQueue w) {
1628	>	w.growArray(); // allocate queue
1629	>	do { w.runTask(scan(w)); } while (w.qlock >= 0);
1630	>	}
1631	>
1632	>	/**
1633	>	* Scans for and, if found, returns one task, else possibly
1634	>	* inactivates the worker. This method operates on single reads of
1635	>	* volatile state and is designed to be re-invoked continuously,
1636	>	* in part because it returns upon detecting inconsistencies,
1637	>	* contention, or state changes that indicate possible success on
1638	>	* re-invocation.
1639	>	*
1640	>	* The scan searches for tasks across queues (starting at a random
1641	>	* index, and relying on registerWorker to irregularly scatter
1642	>	* them within array to avoid bias), checking each at least twice.
1643	>	* The scan terminates upon either finding a non-empty queue, or
1644	>	* completing the sweep. If the worker is not inactivated, it
1645	>	* takes and returns a task from this queue. Otherwise, if not
1646	>	* activated, it signals workers (that may include itself) and
1647	>	* returns so caller can retry. Also returns for true if the
1648	>	* worker array may have changed during an empty scan.  On failure
1649	>	* to find a task, we take one of the following actions, after
1650	>	* which the caller will retry calling this method unless
1651	>	* terminated.
1652	>	*
1653	>	* * If pool is terminating, terminate the worker.
1654	>	*
1655	>	* * If not already enqueued, try to inactivate and enqueue the
1656	>	* worker on wait queue. Or, if inactivating has caused the pool
1657	>	* to be quiescent, relay to idleAwaitWork to possibly shrink
1658	>	* pool.
1659	>	*
1660	>	* * If already enqueued and none of the above apply, possibly
1661	>	* park awaiting signal, else lingering to help scan and signal.
1662	>	*
1663	>	* * If a non-empty queue discovered or left as a hint,
1664	>	* help wake up other workers before return
1665	>	*
1666	>	* @param w the worker (via its WorkQueue)
1667	>	* @return a task or null if none found
1668	>	*/
1669	>	private final ForkJoinTask<?> scan(WorkQueue w) {
1670	>	WorkQueue[] ws; int m;
1671	>	int ps = plock;                          // read plock before ws
1672	>	if (w != null && (ws = workQueues) != null && (m = ws.length - 1) >= 0) {
1673	>	int ec = w.eventCount;               // ec is negative if inactive
1674	>	int r = w.seed; r ^= r << 13; r ^= r >>> 17; w.seed = r ^= r << 5;
1675	>	w.hint = -1;                         // update seed and clear hint
1676	>	int j = ((m + m + 1) | MIN_SCAN) & MAX_SCAN;
1677	>	do {
1678	>	WorkQueue q; ForkJoinTask<?>[] a; int b;
1679	>	if ((q = ws[(r + j) & m]) != null && (b = q.base) - q.top < 0 &&
1680	>	(a = q.array) != null) {     // probably nonempty
1681	>	int i = (((a.length - 1) & b) << ASHIFT) + ABASE;
1682	>	ForkJoinTask<?> t = (ForkJoinTask<?>)
1683	>	U.getObjectVolatile(a, i);
1684	>	if (q.base == b && ec >= 0 && t != null &&
1685	>	U.compareAndSwapObject(a, i, t, null)) {
1686	>	if ((q.base = b + 1) - q.top < 0)
1687	>	signalWork(q);
1688	>	return t;                // taken
1689	>	}
1690	>	else if ((ec < 0 || j < m) && (int)(ctl >> AC_SHIFT) <= 0) {
1691	>	w.hint = (r + j) & m;    // help signal below
1692	>	break;                   // cannot take
1693	>	}
1694	>	}
1695	>	} while (--j >= 0);
1696	>
1697	>	int h, e, ns; long c, sc; WorkQueue q;
1698	>	if ((ns = w.nsteals) != 0) {
1699	>	if (U.compareAndSwapLong(this, STEALCOUNT,
1700	>	sc = stealCount, sc + ns))
1701	>	w.nsteals = 0;               // collect steals and rescan
1702	>	}
1703	>	else if (plock != ps)                // consistency check
1704	>	;                                // skip
1705	>	else if ((e = (int)(c = ctl)) < 0)
1706	>	w.qlock = -1;                    // pool is terminating
1707	>	else {
1708	>	if ((h = w.hint) < 0) {
1709	>	if (ec >= 0) {               // try to enqueue/inactivate
1710	>	long nc = (((long)ec |
1711	>	((c - AC_UNIT) & (AC_MASK|TC_MASK))));
1712	>	w.nextWait = e;          // link and mark inactive
1713	>	w.eventCount = ec | INT_SIGN;
1714	>	if (ctl != c || !U.compareAndSwapLong(this, CTL, c, nc))
1715	>	w.eventCount = ec;   // unmark on CAS failure
1716	>	else if ((int)(c >> AC_SHIFT) == 1 - (config & SMASK))
1717	>	idleAwaitWork(w, nc, c);
1718	>	}
1719	>	else if (w.eventCount < 0 && ctl == c) {
1720	>	Thread wt = Thread.currentThread();
1721	>	Thread.interrupted();    // clear status
1722	>	U.putObject(wt, PARKBLOCKER, this);
1723	>	w.parker = wt;           // emulate LockSupport.park
1724	>	if (w.eventCount < 0)    // recheck
1725	>	U.park(false, 0L);   // block
1726	>	w.parker = null;
1727	>	U.putObject(wt, PARKBLOCKER, null);
1728	>	}
1729	>	}
1730	>	if ((h >= 0 || (h = w.hint) >= 0) &&
1731	>	(ws = workQueues) != null && h < ws.length &&
1732	>	(q = ws[h]) != null) {      // signal others before retry
1733	>	WorkQueue v; Thread p; int u, i, s;
1734	>	for (int n = (config & SMASK) - 1;;) {
1735	>	int idleCount = (w.eventCount < 0) ? 0 : -1;
1736	>	if (((s = idleCount - q.base + q.top) <= n &&
1737	>	(n = s) <= 0) ||
1738	>	(u = (int)((c = ctl) >>> 32)) >= 0 ||
1739	>	(e = (int)c) <= 0 || m < (i = e & SMASK) ||
1740	>	(v = ws[i]) == null)
1741	>	break;
1742	>	long nc = (((long)(v.nextWait & E_MASK)) |
1743	>	((long)(u + UAC_UNIT) << 32));
1744	>	if (v.eventCount != (e | INT_SIGN) ||
1745	>	!U.compareAndSwapLong(this, CTL, c, nc))
1746	>	break;
1747	>	v.hint = h;
1748	>	v.eventCount = (e + E_SEQ) & E_MASK;
1749	>	if ((p = v.parker) != null)
1750	>	U.unpark(p);
1751	>	if (--n <= 0)
1752	>	break;
1753	>	}
1754		}
822	–	else
823	–	break;      // possibly stale; reread
1755		}
1756		}
1757	+	return null;
1758		}
1759
1760		/**
1761	<	* If worker is inactive, blocks until terminating or event count
1762	<	* advances from last value held by worker; in any case helps
1763	<	* release others.
1764	<	*
1765	<	* @param w the calling worker thread
1766	<	* @param retries the number of scans by caller failing to find work
1767	<	* @return false if now too many threads running
1768	<	*/
1769	<	private boolean eventSync(ForkJoinWorkerThread w, int retries) {
1770	<	int wec = w.lastEventCount;
1771	<	if (retries > 1) { // can only block after 2nd miss
1772	<	long nextTop = (((long)wec << EVENT_COUNT_SHIFT) |
1773	<	((long)(w.poolIndex + 1)));
1774	<	long top;
1775	<	while ((runState < SHUTDOWN || !tryTerminate(false)) &&
1776	<	(((int)(top = eventWaiters) & WAITER_ID_MASK) == 0 ||
1777	<	(int)(top >>> EVENT_COUNT_SHIFT) == wec) &&
1778	<	eventCount == wec) {
1779	<	if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
1780	<	w.nextWaiter = top, nextTop)) {
1781	<	accumulateStealCount(w); // transfer steals while idle
1782	<	Thread.interrupted();    // clear/ignore interrupt
1783	<	while (eventCount == wec)
1784	<	w.doPark();
1761	>	* If inactivating worker w has caused the pool to become
1762	>	* quiescent, checks for pool termination, and, so long as this is
1763	>	* not the only worker, waits for event for up to a given
1764	>	* duration.  On timeout, if ctl has not changed, terminates the
1765	>	* worker, which will in turn wake up another worker to possibly
1766	>	* repeat this process.
1767	>	*
1768	>	* @param w the calling worker
1769	>	* @param currentCtl the ctl value triggering possible quiescence
1770	>	* @param prevCtl the ctl value to restore if thread is terminated
1771	>	*/
1772	>	private void idleAwaitWork(WorkQueue w, long currentCtl, long prevCtl) {
1773	>	if (w != null && w.eventCount < 0 &&
1774	>	!tryTerminate(false, false) && (int)prevCtl != 0 &&
1775	>	ctl == currentCtl) {
1776	>	int dc = -(short)(currentCtl >>> TC_SHIFT);
1777	>	long parkTime = dc < 0 ? FAST_IDLE_TIMEOUT: (dc + 1) * IDLE_TIMEOUT;
1778	>	long deadline = System.nanoTime() + parkTime - TIMEOUT_SLOP;
1779	>	Thread wt = Thread.currentThread();
1780	>	while (ctl == currentCtl) {
1781	>	Thread.interrupted();  // timed variant of version in scan()
1782	>	U.putObject(wt, PARKBLOCKER, this);
1783	>	w.parker = wt;
1784	>	if (ctl == currentCtl)
1785	>	U.park(false, parkTime);
1786	>	w.parker = null;
1787	>	U.putObject(wt, PARKBLOCKER, null);
1788	>	if (ctl != currentCtl)
1789	>	break;
1790	>	if (deadline - System.nanoTime() <= 0L &&
1791	>	U.compareAndSwapLong(this, CTL, currentCtl, prevCtl)) {
1792	>	w.eventCount = (w.eventCount + E_SEQ) | E_MASK;
1793	>	w.hint = -1;
1794	>	w.qlock = -1;   // shrink
1795		break;
1796		}
1797		}
856	–	wec = eventCount;
857	–	}
858	–	releaseWaiters();
859	–	int wc = workerCounts;
860	–	if ((wc & RUNNING_COUNT_MASK) <= parallelism) {
861	–	w.lastEventCount = wec;
862	–	return true;
1798		}
864	–	if (wec != w.lastEventCount) // back up if may re-wait
865	–	w.lastEventCount = wec - (wc >>> TOTAL_COUNT_SHIFT);
866	–	return false;
1799		}
1800
1801		/**
1802	<	* Callback from workers invoked upon each top-level action (i.e.,
1803	<	* stealing a task or taking a submission and running
1804	<	* it). Performs one or both of the following:
1805	<	*
1806	<	* * If the worker cannot find work, updates its active status to
1807	<	* inactive and updates activeCount unless there is contention, in
1808	<	* which case it may try again (either in this or a subsequent
1809	<	* call).  Additionally, awaits the next task event and/or helps
1810	<	* wake up other releasable waiters.
1811	<	*
1812	<	* * If there are too many running threads, suspends this worker
1813	<	* (first forcing inactivation if necessary).  If it is not
1814	<	* resumed before a keepAlive elapses, the worker may be "trimmed"
1815	<	* -- killed while suspended within suspendAsSpare. Otherwise,
1816	<	* upon resume it rechecks to make sure that it is still needed.
1817	<	*
1818	<	* @param w the worker
1819	<	* @param retries the number of scans by caller failing to find work
1820	<	* find any (in which case it may block waiting for work).
1821	<	*/
1822	<	final void preStep(ForkJoinWorkerThread w, int retries) {
1823	<	boolean active = w.active;
1824	<	boolean inactivate = active && retries != 0;
1825	<	for (;;) {
1826	<	int rs, wc;
1827	<	if (inactivate &&
1828	<	UNSAFE.compareAndSwapInt(this, runStateOffset,
1829	<	rs = runState, rs - ONE_ACTIVE))
1830	<	inactivate = active = w.active = false;
1831	<	if (((wc = workerCounts) & RUNNING_COUNT_MASK) <= parallelism) {
1832	<	if (active || eventSync(w, retries))
1833	<	break;
1802	>	* Scans through queues looking for work while joining a task; if
1803	>	* any present, signals. May return early if more signalling is
1804	>	* detectably unneeded.
1805	>	*
1806	>	* @param task return early if done
1807	>	* @param origin an index to start scan
1808	>	*/
1809	>	private void helpSignal(ForkJoinTask<?> task, int origin) {
1810	>	WorkQueue[] ws; WorkQueue w; Thread p; long c; int m, u, e, i, s;
1811	>	if (task != null && task.status >= 0 &&
1812	>	(u = (int)(ctl >>> 32)) < 0 && (u >> UAC_SHIFT) < 0 &&
1813	>	(ws = workQueues) != null && (m = ws.length - 1) >= 0) {
1814	>	outer: for (int k = origin, j = m; j >= 0; --j) {
1815	>	WorkQueue q = ws[k++ & m];
1816	>	for (int n = m;;) { // limit to at most m signals
1817	>	if (task.status < 0)
1818	>	break outer;
1819	>	if (q == null ||
1820	>	((s = -q.base + q.top) <= n && (n = s) <= 0))
1821	>	break;
1822	>	if ((u = (int)((c = ctl) >>> 32)) >= 0 ||
1823	>	(e = (int)c) <= 0 || m < (i = e & SMASK) ||
1824	>	(w = ws[i]) == null)
1825	>	break outer;
1826	>	long nc = (((long)(w.nextWait & E_MASK)) |
1827	>	((long)(u + UAC_UNIT) << 32));
1828	>	if (w.eventCount != (e | INT_SIGN))
1829	>	break outer;
1830	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1831	>	w.eventCount = (e + E_SEQ) & E_MASK;
1832	>	if ((p = w.parker) != null)
1833	>	U.unpark(p);
1834	>	if (--n <= 0)
1835	>	break;
1836	>	}
1837	>	}
1838		}
903	–	else if (!(inactivate |= active) &&  // must inactivate to suspend
904	–	UNSAFE.compareAndSwapInt(this, workerCountsOffset,
905	–	wc, wc - ONE_RUNNING) &&
906	–	!w.suspendAsSpare())             // false if trimmed
907	–	break;
1839		}
1840		}
1841
1842		/**
1843	<	* Awaits join of the given task if enough threads, or can resume
1844	<	* or create a spare. Fails (in which case the given task might
1845	<	* not be done) upon contention or lack of decision about
1846	<	* blocking. Returns void because caller must check
1847	<	* task status on return anyway.
1848	<	*
1849	<	* We allow blocking if:
1850	<	*
1851	<	* 1. There would still be at least as many running threads as
1852	<	*    parallelism level if this thread blocks.
1853	<	*
1854	<	* 2. A spare is resumed to replace this worker. We tolerate
1855	<	*    slop in the decision to replace if a spare is found without
1856	<	*    first decrementing run count.  This may release too many,
1857	<	*    but if so, the superfluous ones will re-suspend via
1858	<	*    preStep().
1859	<	*
1860	<	* 3. After #spares repeated checks, there are no fewer than #spare
1861	<	*    threads not running. We allow this slack to avoid hysteresis
1862	<	*    and as a hedge against lag/uncertainty of running count
1863	<	*    estimates when signalling or unblocking stalls.
1864	<	*
1865	<	* 4. All existing workers are busy (as rechecked via repeated
1866	<	*    retries by caller) and a new spare is created.
1867	<	*
1868	<	* If none of the above hold, we try to escape out by
1869	<	* re-incrementing count and returning to caller, which can retry
1870	<	* later.
1871	<	*
1872	<	* @param joinMe the task to join
1873	<	* @param retries if negative, then serve only as a precheck
1874	<	*   that the thread can be replaced by a spare. Otherwise,
1875	<	*   the number of repeated calls to this method returning busy
1876	<	* @return true if the call must be retried because there
1877	<	*   none of the blocking checks hold
1878	<	*/
1879	<	final boolean tryAwaitJoin(ForkJoinTask<?> joinMe, int retries) {
1880	<	if (joinMe.status < 0) // precheck for cancellation
1881	<	return false;
1882	<	if ((runState & TERMINATING) != 0) { // shutting down
1883	<	joinMe.cancelIgnoringExceptions();
1884	<	return false;
1885	<	}
1886	<
1887	<	int pc = parallelism;
1888	<	boolean running = true; // false when running count decremented
1889	<	outer:for (;;) {
1890	<	int wc = workerCounts;
1891	<	int rc = wc & RUNNING_COUNT_MASK;
1892	<	int tc = wc >>> TOTAL_COUNT_SHIFT;
1893	<	if (running) { // replace with spare or decrement count
1894	<	if (rc <= pc && tc > pc &&
1895	<	(retries > 0 || tc > (runState & ACTIVE_COUNT_MASK))) {
1896	<	ForkJoinWorkerThread[] ws = workers;
1897	<	int nws = ws.length;
1898	<	for (int i = 0; i < nws; ++i) { // search for spare
1899	<	ForkJoinWorkerThread w = ws[i];
1900	<	if (w != null) {
1901	<	if (joinMe.status < 0)
1902	<	return false;
1903	<	if (w.isSuspended()) {
1904	<	if ((workerCounts & RUNNING_COUNT_MASK)>=pc &&
1905	<	w.tryResumeSpare()) {
1906	<	running = false;
1907	<	break outer;
1908	<	}
1909	<	continue outer; // rescan
1843	>	* Tries to locate and execute tasks for a stealer of the given
1844	>	* task, or in turn one of its stealers, Traces currentSteal ->
1845	>	* currentJoin links looking for a thread working on a descendant
1846	>	* of the given task and with a non-empty queue to steal back and
1847	>	* execute tasks from. The first call to this method upon a
1848	>	* waiting join will often entail scanning/search, (which is OK
1849	>	* because the joiner has nothing better to do), but this method
1850	>	* leaves hints in workers to speed up subsequent calls. The
1851	>	* implementation is very branchy to cope with potential
1852	>	* inconsistencies or loops encountering chains that are stale,
1853	>	* unknown, or so long that they are likely cyclic.
1854	>	*
1855	>	* @param joiner the joining worker
1856	>	* @param task the task to join
1857	>	* @return 0 if no progress can be made, negative if task
1858	>	* known complete, else positive
1859	>	*/
1860	>	private int tryHelpStealer(WorkQueue joiner, ForkJoinTask<?> task) {
1861	>	int stat = 0, steps = 0;                    // bound to avoid cycles
1862	>	if (joiner != null && task != null) {       // hoist null checks
1863	>	restart: for (;;) {
1864	>	ForkJoinTask<?> subtask = task;     // current target
1865	>	for (WorkQueue j = joiner, v;;) {   // v is stealer of subtask
1866	>	WorkQueue[] ws; int m, s, h;
1867	>	if ((s = task.status) < 0) {
1868	>	stat = s;
1869	>	break restart;
1870	>	}
1871	>	if ((ws = workQueues) == null || (m = ws.length - 1) <= 0)
1872	>	break restart;              // shutting down
1873	>	if ((v = ws[h = (j.hint | 1) & m]) == null ||
1874	>	v.currentSteal != subtask) {
1875	>	for (int origin = h;;) {    // find stealer
1876	>	if (((h = (h + 2) & m) & 15) == 1 &&
1877	>	(subtask.status < 0 || j.currentJoin != subtask))
1878	>	continue restart;   // occasional staleness check
1879	>	if ((v = ws[h]) != null &&
1880	>	v.currentSteal == subtask) {
1881	>	j.hint = h;        // save hint
1882	>	break;
1883	>	}
1884	>	if (h == origin)
1885	>	break restart;      // cannot find stealer
1886	>	}
1887	>	}
1888	>	for (;;) { // help stealer or descend to its stealer
1889	>	ForkJoinTask[] a;  int b;
1890	>	if (subtask.status < 0)     // surround probes with
1891	>	continue restart;       //   consistency checks
1892	>	if ((b = v.base) - v.top < 0 && (a = v.array) != null) {
1893	>	int i = (((a.length - 1) & b) << ASHIFT) + ABASE;
1894	>	ForkJoinTask<?> t =
1895	>	(ForkJoinTask<?>)U.getObjectVolatile(a, i);
1896	>	if (subtask.status < 0 || j.currentJoin != subtask ||
1897	>	v.currentSteal != subtask)
1898	>	continue restart;   // stale
1899	>	stat = 1;               // apparent progress
1900	>	if (t != null && v.base == b &&
1901	>	U.compareAndSwapObject(a, i, t, null)) {
1902	>	v.base = b + 1;     // help stealer
1903	>	joiner.runSubtask(t);
1904	>	}
1905	>	else if (v.base == b && ++steps == MAX_HELP)
1906	>	break restart;      // v apparently stalled
1907	>	}
1908	>	else {                      // empty -- try to descend
1909	>	ForkJoinTask<?> next = v.currentJoin;
1910	>	if (subtask.status < 0 || j.currentJoin != subtask ||
1911	>	v.currentSteal != subtask)
1912	>	continue restart;   // stale
1913	>	else if (next == null || ++steps == MAX_HELP)
1914	>	break restart;      // dead-end or maybe cyclic
1915	>	else {
1916	>	subtask = next;
1917	>	j = v;
1918	>	break;
1919		}
1920		}
1921		}
1922		}
983	–	if (retries < 0 || // < 0 means replacement check only
984	–	rc == 0 || joinMe.status < 0 || workerCounts != wc ||
985	–	!UNSAFE.compareAndSwapInt(this, workerCountsOffset,
986	–	wc, wc - ONE_RUNNING))
987	–	return false; // done or inconsistent or contended
988	–	running = false;
989	–	if (rc > pc)
990	–	break;
1923		}
1924	<	else { // allow blocking if enough threads
1925	<	if (rc >= pc || joinMe.status < 0)
1926	<	break;
1927	<	int sc = tc - pc + 1; // = spare threads, plus the one to add
1928	<	if (retries > sc) {
1929	<	if (rc > 0 && rc >= pc - sc) // allow slack
1930	<	break;
1931	<	if (tc < MAX_THREADS &&
1932	<	tc == (runState & ACTIVE_COUNT_MASK) &&
1933	<	workerCounts == wc &&
1934	<	UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
1935	<	wc+(ONE_RUNNING|ONE_TOTAL))) {
1936	<	addWorker();
1924	>	}
1925	>	return stat;
1926	>	}
1927	>
1928	>	/**
1929	>	* Analog of tryHelpStealer for CountedCompleters. Tries to steal
1930	>	* and run tasks within the target's computation.
1931	>	*
1932	>	* @param task the task to join
1933	>	* @param mode if shared, exit upon completing any task
1934	>	* if all workers are active
1935	>	*
1936	>	*/
1937	>	private int helpComplete(ForkJoinTask<?> task, int mode) {
1938	>	WorkQueue[] ws; WorkQueue q; int m, n, s, u;
1939	>	if (task != null && (ws = workQueues) != null &&
1940	>	(m = ws.length - 1) >= 0) {
1941	>	for (int j = 1, origin = j;;) {
1942	>	if ((s = task.status) < 0)
1943	>	return s;
1944	>	if ((q = ws[j & m]) != null && q.pollAndExecCC(task)) {
1945	>	origin = j;
1946	>	if (mode == SHARED_QUEUE &&
1947	>	((u = (int)(ctl >>> 32)) >= 0 || (u >> UAC_SHIFT) >= 0))
1948		break;
1006	–	}
1949		}
1950	<	if (workerCounts == wc &&        // back out to allow rescan
1951	<	UNSAFE.compareAndSwapInt (this, workerCountsOffset,
1952	<	wc, wc + ONE_RUNNING)) {
1953	<	releaseWaiters();            // help others progress
1954	<	return true;                 // let caller retry
1950	>	else if ((j = (j + 2) & m) == origin)
1951	>	break;
1952	>	}
1953	>	}
1954	>	return 0;
1955	>	}
1956	>
1957	>	/**
1958	>	* Tries to decrement active count (sometimes implicitly) and
1959	>	* possibly release or create a compensating worker in preparation
1960	>	* for blocking. Fails on contention or termination. Otherwise,
1961	>	* adds a new thread if no idle workers are available and pool
1962	>	* may become starved.
1963	>	*/
1964	>	final boolean tryCompensate() {
1965	>	int pc = config & SMASK, e, i, tc; long c;
1966	>	WorkQueue[] ws; WorkQueue w; Thread p;
1967	>	if ((ws = workQueues) != null && (e = (int)(c = ctl)) >= 0) {
1968	>	if (e != 0 && (i = e & SMASK) < ws.length &&
1969	>	(w = ws[i]) != null && w.eventCount == (e | INT_SIGN)) {
1970	>	long nc = ((long)(w.nextWait & E_MASK) |
1971	>	(c & (AC_MASK|TC_MASK)));
1972	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1973	>	w.eventCount = (e + E_SEQ) & E_MASK;
1974	>	if ((p = w.parker) != null)
1975	>	U.unpark(p);
1976	>	return true;   // replace with idle worker
1977	>	}
1978	>	}
1979	>	else if ((tc = (short)(c >>> TC_SHIFT)) >= 0 &&
1980	>	(int)(c >> AC_SHIFT) + pc > 1) {
1981	>	long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK);
1982	>	if (U.compareAndSwapLong(this, CTL, c, nc))
1983	>	return true;   // no compensation
1984	>	}
1985	>	else if (tc + pc < MAX_CAP) {
1986	>	long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK);
1987	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1988	>	ForkJoinWorkerThreadFactory fac;
1989	>	Throwable ex = null;
1990	>	ForkJoinWorkerThread wt = null;
1991	>	try {
1992	>	if ((fac = factory) != null &&
1993	>	(wt = fac.newThread(this)) != null) {
1994	>	wt.start();
1995	>	return true;
1996	>	}
1997	>	} catch (Throwable rex) {
1998	>	ex = rex;
1999	>	}
2000	>	deregisterWorker(wt, ex); // clean up and return false
2001		}
2002		}
2003		}
1016	–	// arrive here if can block
1017	–	joinMe.internalAwaitDone();
1018	–	int c;                      // to inline incrementRunningCount
1019	–	do {} while (!UNSAFE.compareAndSwapInt
1020	–	(this, workerCountsOffset,
1021	–	c = workerCounts, c + ONE_RUNNING));
2004		return false;
2005		}
2006
2007		/**
2008	<	* Same idea as (and shares many code snippets with) tryAwaitJoin,
2009	<	* but self-contained because there are no caller retries.
2010	<	* TODO: Rework to use simpler API.
2008	>	* Helps and/or blocks until the given task is done.
2009	>	*
2010	>	* @param joiner the joining worker
2011	>	* @param task the task
2012	>	* @return task status on exit
2013		*/
2014	<	final void awaitBlocker(ManagedBlocker blocker)
2015	<	throws InterruptedException {
2016	<	boolean done;
2017	<	if (done = blocker.isReleasable())
2018	<	return;
2019	<	int pc = parallelism;
2020	<	int retries = 0;
2021	<	boolean running = true; // false when running count decremented
2022	<	outer:for (;;) {
2023	<	int wc = workerCounts;
2024	<	int rc = wc & RUNNING_COUNT_MASK;
2025	<	int tc = wc >>> TOTAL_COUNT_SHIFT;
2026	<	if (running) {
2027	<	if (rc <= pc && tc > pc &&
2028	<	(retries > 0 || tc > (runState & ACTIVE_COUNT_MASK))) {
2029	<	ForkJoinWorkerThread[] ws = workers;
2030	<	int nws = ws.length;
2031	<	for (int i = 0; i < nws; ++i) {
2032	<	ForkJoinWorkerThread w = ws[i];
2033	<	if (w != null) {
2034	<	if (done = blocker.isReleasable())
2035	<	return;
2036	<	if (w.isSuspended()) {
2037	<	if ((workerCounts & RUNNING_COUNT_MASK)>=pc &&
2038	<	w.tryResumeSpare()) {
2039	<	running = false;
1056	<	break outer;
2014	>	final int awaitJoin(WorkQueue joiner, ForkJoinTask<?> task) {
2015	>	int s = 0;
2016	>	if (joiner != null && task != null && (s = task.status) >= 0) {
2017	>	ForkJoinTask<?> prevJoin = joiner.currentJoin;
2018	>	joiner.currentJoin = task;
2019	>	do {} while ((s = task.status) >= 0 && !joiner.isEmpty() &&
2020	>	joiner.tryRemoveAndExec(task)); // process local tasks
2021	>	if (s >= 0 && (s = task.status) >= 0) {
2022	>	helpSignal(task, joiner.poolIndex);
2023	>	if ((s = task.status) >= 0 &&
2024	>	(task instanceof CountedCompleter))
2025	>	s = helpComplete(task, LIFO_QUEUE);
2026	>	}
2027	>	while (s >= 0 && (s = task.status) >= 0) {
2028	>	if ((!joiner.isEmpty() ||           // try helping
2029	>	(s = tryHelpStealer(joiner, task)) == 0) &&
2030	>	(s = task.status) >= 0) {
2031	>	helpSignal(task, joiner.poolIndex);
2032	>	if ((s = task.status) >= 0 && tryCompensate()) {
2033	>	if (task.trySetSignal() && (s = task.status) >= 0) {
2034	>	synchronized (task) {
2035	>	if (task.status >= 0) {
2036	>	try {                // see ForkJoinTask
2037	>	task.wait();     //  for explanation
2038	>	} catch (InterruptedException ie) {
2039	>	}
2040		}
2041	<	continue outer; // rescan
2041	>	else
2042	>	task.notifyAll();
2043		}
2044		}
2045	+	long c;                          // re-activate
2046	+	do {} while (!U.compareAndSwapLong
2047	+	(this, CTL, c = ctl, c + AC_UNIT));
2048		}
2049		}
1063	–	if (done = blocker.isReleasable())
1064	–	return;
1065	–	if (rc == 0 || workerCounts != wc ||
1066	–	!UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1067	–	wc, wc - ONE_RUNNING))
1068	–	continue;
1069	–	running = false;
1070	–	if (rc > pc)
1071	–	break;
1072	–	}
1073	–	else {
1074	–	if (rc >= pc || (done = blocker.isReleasable()))
1075	–	break;
1076	–	int sc = tc - pc + 1;
1077	–	if (retries++ > sc) {
1078	–	if (rc > 0 && rc >= pc - sc)
1079	–	break;
1080	–	if (tc < MAX_THREADS &&
1081	–	tc == (runState & ACTIVE_COUNT_MASK) &&
1082	–	workerCounts == wc &&
1083	–	UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
1084	–	wc+(ONE_RUNNING|ONE_TOTAL))) {
1085	–	addWorker();
1086	–	break;
1087	–	}
1088	–	}
1089	–	Thread.yield();
1090	–	}
1091	–	}
1092	–
1093	–	try {
1094	–	if (!done)
1095	–	do {} while (!blocker.isReleasable() && !blocker.block());
1096	–	} finally {
1097	–	if (!running) {
1098	–	int c;
1099	–	do {} while (!UNSAFE.compareAndSwapInt
1100	–	(this, workerCountsOffset,
1101	–	c = workerCounts, c + ONE_RUNNING));
2050		}
2051	+	joiner.currentJoin = prevJoin;
2052		}
2053	+	return s;
2054		}
2055
2056		/**
2057	<	* Possibly initiates and/or completes termination.
2057	>	* Stripped-down variant of awaitJoin used by timed joins. Tries
2058	>	* to help join only while there is continuous progress. (Caller
2059	>	* will then enter a timed wait.)
2060		*
2061	<	* @param now if true, unconditionally terminate, else only
2062	<	* if shutdown and empty queue and no active workers
1111	<	* @return true if now terminating or terminated
2061	>	* @param joiner the joining worker
2062	>	* @param task the task
2063		*/
2064	<	private boolean tryTerminate(boolean now) {
2065	<	if (now)
2066	<	advanceRunLevel(SHUTDOWN); // ensure at least SHUTDOWN
2067	<	else if (runState < SHUTDOWN ||
2068	<	!submissionQueue.isEmpty() ||
2069	<	(runState & ACTIVE_COUNT_MASK) != 0)
2070	<	return false;
2071	<
2072	<	if (advanceRunLevel(TERMINATING))
2073	<	startTerminating();
2074	<
2075	<	// Finish now if all threads terminated; else in some subsequent call
2076	<	if ((workerCounts >>> TOTAL_COUNT_SHIFT) == 0) {
2077	<	advanceRunLevel(TERMINATED);
2078	<	termination.arrive();
2064	>	final void helpJoinOnce(WorkQueue joiner, ForkJoinTask<?> task) {
2065	>	int s;
2066	>	if (joiner != null && task != null && (s = task.status) >= 0) {
2067	>	ForkJoinTask<?> prevJoin = joiner.currentJoin;
2068	>	joiner.currentJoin = task;
2069	>	do {} while ((s = task.status) >= 0 && !joiner.isEmpty() &&
2070	>	joiner.tryRemoveAndExec(task));
2071	>	if (s >= 0 && (s = task.status) >= 0) {
2072	>	helpSignal(task, joiner.poolIndex);
2073	>	if ((s = task.status) >= 0 &&
2074	>	(task instanceof CountedCompleter))
2075	>	s = helpComplete(task, LIFO_QUEUE);
2076	>	}
2077	>	if (s >= 0 && joiner.isEmpty()) {
2078	>	do {} while (task.status >= 0 &&
2079	>	tryHelpStealer(joiner, task) > 0);
2080	>	}
2081	>	joiner.currentJoin = prevJoin;
2082		}
1129	–	return true;
2083		}
2084
2085		/**
2086	<	* Actions on transition to TERMINATING
2086	>	* Returns a (probably) non-empty steal queue, if one is found
2087	>	* during a scan, else null.  This method must be retried by
2088	>	* caller if, by the time it tries to use the queue, it is empty.
2089	>	* @param r a (random) seed for scanning
2090		*/
2091	<	private void startTerminating() {
2092	<	for (int i = 0; i < 2; ++i) { // twice to mop up newly created workers
2093	<	cancelSubmissions();
2094	<	shutdownWorkers();
2095	<	cancelWorkerTasks();
2096	<	signalEvent();
2097	<	interruptWorkers();
2091	>	private WorkQueue findNonEmptyStealQueue(int r) {
2092	>	for (;;) {
2093	>	int ps = plock, m; WorkQueue[] ws; WorkQueue q;
2094	>	if ((ws = workQueues) != null && (m = ws.length - 1) >= 0) {
2095	>	for (int j = (m + 1) << 2; j >= 0; --j) {
2096	>	if ((q = ws[(((r + j) << 1) | 1) & m]) != null &&
2097	>	q.base - q.top < 0)
2098	>	return q;
2099	>	}
2100	>	}
2101	>	if (plock == ps)
2102	>	return null;
2103		}
2104		}
2105
2106		/**
2107	<	* Clear out and cancel submissions, ignoring exceptions
2108	<	*/
2109	<	private void cancelSubmissions() {
2110	<	ForkJoinTask<?> task;
2111	<	while ((task = submissionQueue.poll()) != null) {
2112	<	try {
2113	<	task.cancel(false);
2114	<	} catch (Throwable ignore) {
2107	>	* Runs tasks until {@code isQuiescent()}. We piggyback on
2108	>	* active count ctl maintenance, but rather than blocking
2109	>	* when tasks cannot be found, we rescan until all others cannot
2110	>	* find tasks either.
2111	>	*/
2112	>	final void helpQuiescePool(WorkQueue w) {
2113	>	for (boolean active = true;;) {
2114	>	long c; WorkQueue q; ForkJoinTask<?> t; int b;
2115	>	while ((t = w.nextLocalTask()) != null) {
2116	>	if (w.base - w.top < 0)
2117	>	signalWork(w);
2118	>	t.doExec();
2119		}
2120	+	if ((q = findNonEmptyStealQueue(w.nextSeed())) != null) {
2121	+	if (!active) {      // re-establish active count
2122	+	active = true;
2123	+	do {} while (!U.compareAndSwapLong
2124	+	(this, CTL, c = ctl, c + AC_UNIT));
2125	+	}
2126	+	if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null) {
2127	+	if (q.base - q.top < 0)
2128	+	signalWork(q);
2129	+	w.runSubtask(t);
2130	+	}
2131	+	}
2132	+	else if (active) {       // decrement active count without queuing
2133	+	long nc = (c = ctl) - AC_UNIT;
2134	+	if ((int)(nc >> AC_SHIFT) + (config & SMASK) == 0)
2135	+	return;          // bypass decrement-then-increment
2136	+	if (U.compareAndSwapLong(this, CTL, c, nc))
2137	+	active = false;
2138	+	}
2139	+	else if ((int)((c = ctl) >> AC_SHIFT) + (config & SMASK) == 0 &&
2140	+	U.compareAndSwapLong(this, CTL, c, c + AC_UNIT))
2141	+	return;
2142		}
2143		}
2144
2145		/**
2146	<	* Sets all worker run states to at least shutdown,
2147	<	* also resuming suspended workers
2146	>	* Gets and removes a local or stolen task for the given worker.
2147	>	*
2148	>	* @return a task, if available
2149		*/
2150	<	private void shutdownWorkers() {
2151	<	ForkJoinWorkerThread[] ws = workers;
2152	<	int nws = ws.length;
2153	<	for (int i = 0; i < nws; ++i) {
2154	<	ForkJoinWorkerThread w = ws[i];
2155	<	if (w != null)
2156	<	w.shutdown();
2150	>	final ForkJoinTask<?> nextTaskFor(WorkQueue w) {
2151	>	for (ForkJoinTask<?> t;;) {
2152	>	WorkQueue q; int b;
2153	>	if ((t = w.nextLocalTask()) != null)
2154	>	return t;
2155	>	if ((q = findNonEmptyStealQueue(w.nextSeed())) == null)
2156	>	return null;
2157	>	if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null) {
2158	>	if (q.base - q.top < 0)
2159	>	signalWork(q);
2160	>	return t;
2161	>	}
2162		}
2163		}
2164
2165		/**
2166	<	* Clears out and cancels all locally queued tasks
2166	>	* Returns a cheap heuristic guide for task partitioning when
2167	>	* programmers, frameworks, tools, or languages have little or no
2168	>	* idea about task granularity.  In essence by offering this
2169	>	* method, we ask users only about tradeoffs in overhead vs
2170	>	* expected throughput and its variance, rather than how finely to
2171	>	* partition tasks.
2172	>	*
2173	>	* In a steady state strict (tree-structured) computation, each
2174	>	* thread makes available for stealing enough tasks for other
2175	>	* threads to remain active. Inductively, if all threads play by
2176	>	* the same rules, each thread should make available only a
2177	>	* constant number of tasks.
2178	>	*
2179	>	* The minimum useful constant is just 1. But using a value of 1
2180	>	* would require immediate replenishment upon each steal to
2181	>	* maintain enough tasks, which is infeasible.  Further,
2182	>	* partitionings/granularities of offered tasks should minimize
2183	>	* steal rates, which in general means that threads nearer the top
2184	>	* of computation tree should generate more than those nearer the
2185	>	* bottom. In perfect steady state, each thread is at
2186	>	* approximately the same level of computation tree. However,
2187	>	* producing extra tasks amortizes the uncertainty of progress and
2188	>	* diffusion assumptions.
2189	>	*
2190	>	* So, users will want to use values larger, but not much larger
2191	>	* than 1 to both smooth over transient shortages and hedge
2192	>	* against uneven progress; as traded off against the cost of
2193	>	* extra task overhead. We leave the user to pick a threshold
2194	>	* value to compare with the results of this call to guide
2195	>	* decisions, but recommend values such as 3.
2196	>	*
2197	>	* When all threads are active, it is on average OK to estimate
2198	>	* surplus strictly locally. In steady-state, if one thread is
2199	>	* maintaining say 2 surplus tasks, then so are others. So we can
2200	>	* just use estimated queue length.  However, this strategy alone
2201	>	* leads to serious mis-estimates in some non-steady-state
2202	>	* conditions (ramp-up, ramp-down, other stalls). We can detect
2203	>	* many of these by further considering the number of "idle"
2204	>	* threads, that are known to have zero queued tasks, so
2205	>	* compensate by a factor of (#idle/#active) threads.
2206	>	*
2207	>	* Note: The approximation of #busy workers as #active workers is
2208	>	* not very good under current signalling scheme, and should be
2209	>	* improved.
2210	>	*/
2211	>	static int getSurplusQueuedTaskCount() {
2212	>	Thread t; ForkJoinWorkerThread wt; ForkJoinPool pool; WorkQueue q;
2213	>	if (((t = Thread.currentThread()) instanceof ForkJoinWorkerThread)) {
2214	>	int p = (pool = (wt = (ForkJoinWorkerThread)t).pool).config & SMASK;
2215	>	int n = (q = wt.workQueue).top - q.base;
2216	>	int a = (int)(pool.ctl >> AC_SHIFT) + p;
2217	>	return n - (a > (p >>>= 1) ? 0 :
2218	>	a > (p >>>= 1) ? 1 :
2219	>	a > (p >>>= 1) ? 2 :
2220	>	a > (p >>>= 1) ? 4 :
2221	>	8);
2222	>	}
2223	>	return 0;
2224	>	}
2225	>
2226	>	//  Termination
2227	>
2228	>	/**
2229	>	* Possibly initiates and/or completes termination.  The caller
2230	>	* triggering termination runs three passes through workQueues:
2231	>	* (0) Setting termination status, followed by wakeups of queued
2232	>	* workers; (1) cancelling all tasks; (2) interrupting lagging
2233	>	* threads (likely in external tasks, but possibly also blocked in
2234	>	* joins).  Each pass repeats previous steps because of potential
2235	>	* lagging thread creation.
2236	>	*
2237	>	* @param now if true, unconditionally terminate, else only
2238	>	* if no work and no active workers
2239	>	* @param enable if true, enable shutdown when next possible
2240	>	* @return true if now terminating or terminated
2241		*/
2242	<	private void cancelWorkerTasks() {
2243	<	ForkJoinWorkerThread[] ws = workers;
2244	<	int nws = ws.length;
2245	<	for (int i = 0; i < nws; ++i) {
2246	<	ForkJoinWorkerThread w = ws[i];
2247	<	if (w != null)
2248	<	w.cancelTasks();
2242	>	private boolean tryTerminate(boolean now, boolean enable) {
2243	>	int ps;
2244	>	if (this == commonPool)                    // cannot shut down
2245	>	return false;
2246	>	if ((ps = plock) >= 0) {                   // enable by setting plock
2247	>	if (!enable)
2248	>	return false;
2249	>	if ((ps & PL_LOCK) != 0 ||
2250	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
2251	>	ps = acquirePlock();
2252	>	int nps = ((ps + PL_LOCK) & ~SHUTDOWN) | SHUTDOWN;
2253	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
2254	>	releasePlock(nps);
2255	>	}
2256	>	for (long c;;) {
2257	>	if (((c = ctl) & STOP_BIT) != 0) {     // already terminating
2258	>	if ((short)(c >>> TC_SHIFT) == -(config & SMASK)) {
2259	>	synchronized (this) {
2260	>	notifyAll();               // signal when 0 workers
2261	>	}
2262	>	}
2263	>	return true;
2264	>	}
2265	>	if (!now) {                            // check if idle & no tasks
2266	>	WorkQueue[] ws; WorkQueue w;
2267	>	if ((int)(c >> AC_SHIFT) != -(config & SMASK))
2268	>	return false;
2269	>	if ((ws = workQueues) != null) {
2270	>	for (int i = 0; i < ws.length; ++i) {
2271	>	if ((w = ws[i]) != null) {
2272	>	if (!w.isEmpty()) {    // signal unprocessed tasks
2273	>	signalWork(w);
2274	>	return false;
2275	>	}
2276	>	if ((i & 1) != 0 && w.eventCount >= 0)
2277	>	return false;      // unqueued inactive worker
2278	>	}
2279	>	}
2280	>	}
2281	>	}
2282	>	if (U.compareAndSwapLong(this, CTL, c, c | STOP_BIT)) {
2283	>	for (int pass = 0; pass < 3; ++pass) {
2284	>	WorkQueue[] ws; WorkQueue w; Thread wt;
2285	>	if ((ws = workQueues) != null) {
2286	>	int n = ws.length;
2287	>	for (int i = 0; i < n; ++i) {
2288	>	if ((w = ws[i]) != null) {
2289	>	w.qlock = -1;
2290	>	if (pass > 0) {
2291	>	w.cancelAll();
2292	>	if (pass > 1 && (wt = w.owner) != null) {
2293	>	if (!wt.isInterrupted()) {
2294	>	try {
2295	>	wt.interrupt();
2296	>	} catch (Throwable ignore) {
2297	>	}
2298	>	}
2299	>	U.unpark(wt);
2300	>	}
2301	>	}
2302	>	}
2303	>	}
2304	>	// Wake up workers parked on event queue
2305	>	int i, e; long cc; Thread p;
2306	>	while ((e = (int)(cc = ctl) & E_MASK) != 0 &&
2307	>	(i = e & SMASK) < n && i >= 0 &&
2308	>	(w = ws[i]) != null) {
2309	>	long nc = ((long)(w.nextWait & E_MASK) |
2310	>	((cc + AC_UNIT) & AC_MASK) |
2311	>	(cc & (TC_MASK|STOP_BIT)));
2312	>	if (w.eventCount == (e | INT_SIGN) &&
2313	>	U.compareAndSwapLong(this, CTL, cc, nc)) {
2314	>	w.eventCount = (e + E_SEQ) & E_MASK;
2315	>	w.qlock = -1;
2316	>	if ((p = w.parker) != null)
2317	>	U.unpark(p);
2318	>	}
2319	>	}
2320	>	}
2321	>	}
2322	>	}
2323		}
2324		}
2325
2326	+	// external operations on common pool
2327	+
2328		/**
2329	<	* Unsticks all workers blocked on joins etc
2329	>	* Returns common pool queue for a thread that has submitted at
2330	>	* least one task.
2331		*/
2332	<	private void interruptWorkers() {
2333	<	ForkJoinWorkerThread[] ws = workers;
2334	<	int nws = ws.length;
2335	<	for (int i = 0; i < nws; ++i) {
2336	<	ForkJoinWorkerThread w = ws[i];
2337	<	if (w != null && !w.isTerminated()) {
2338	<	try {
2339	<	w.interrupt();
2340	<	} catch (SecurityException ignore) {
2332	>	static WorkQueue commonSubmitterQueue() {
2333	>	ForkJoinPool p; WorkQueue[] ws; int m; Submitter z;
2334	>	return ((z = submitters.get()) != null &&
2335	>	(p = commonPool) != null &&
2336	>	(ws = p.workQueues) != null &&
2337	>	(m = ws.length - 1) >= 0) ?
2338	>	ws[m & z.seed & SQMASK] : null;
2339	>	}
2340	>
2341	>	/**
2342	>	* Tries to pop the given task from submitter's queue in common pool.
2343	>	*/
2344	>	static boolean tryExternalUnpush(ForkJoinTask<?> t) {
2345	>	ForkJoinPool p; WorkQueue[] ws; WorkQueue q; Submitter z;
2346	>	ForkJoinTask<?>[] a;  int m, s;
2347	>	if (t != null &&
2348	>	(z = submitters.get()) != null &&
2349	>	(p = commonPool) != null &&
2350	>	(ws = p.workQueues) != null &&
2351	>	(m = ws.length - 1) >= 0 &&
2352	>	(q = ws[m & z.seed & SQMASK]) != null &&
2353	>	(s = q.top) != q.base &&
2354	>	(a = q.array) != null) {
2355	>	long j = (((a.length - 1) & (s - 1)) << ASHIFT) + ABASE;
2356	>	if (U.getObject(a, j) == t &&
2357	>	U.compareAndSwapInt(q, QLOCK, 0, 1)) {
2358	>	if (q.array == a && q.top == s && // recheck
2359	>	U.compareAndSwapObject(a, j, t, null)) {
2360	>	q.top = s - 1;
2361	>	q.qlock = 0;
2362	>	return true;
2363		}
2364	+	q.qlock = 0;
2365		}
2366		}
2367	+	return false;
2368		}
2369
1202	–	// misc support for ForkJoinWorkerThread
1203	–
2370		/**
2371	<	* Returns pool number
2372	<	*/
2373	<	final int getPoolNumber() {
2374	<	return poolNumber;
2371	>	* Tries to pop and run local tasks within the same computation
2372	>	* as the given root. On failure, tries to help complete from
2373	>	* other queues via helpComplete.
2374	>	*/
2375	>	private void externalHelpComplete(WorkQueue q, ForkJoinTask<?> root) {
2376	>	ForkJoinTask<?>[] a; int m;
2377	>	if (q != null && (a = q.array) != null && (m = (a.length - 1)) >= 0 &&
2378	>	root != null && root.status >= 0) {
2379	>	for (;;) {
2380	>	int s, u; Object o; CountedCompleter<?> task = null;
2381	>	if ((s = q.top) - q.base > 0) {
2382	>	long j = ((m & (s - 1)) << ASHIFT) + ABASE;
2383	>	if ((o = U.getObject(a, j)) != null &&
2384	>	(o instanceof CountedCompleter)) {
2385	>	CountedCompleter<?> t = (CountedCompleter<?>)o, r = t;
2386	>	do {
2387	>	if (r == root) {
2388	>	if (U.compareAndSwapInt(q, QLOCK, 0, 1)) {
2389	>	if (q.array == a && q.top == s &&
2390	>	U.compareAndSwapObject(a, j, t, null)) {
2391	>	q.top = s - 1;
2392	>	task = t;
2393	>	}
2394	>	q.qlock = 0;
2395	>	}
2396	>	break;
2397	>	}
2398	>	} while ((r = r.completer) != null);
2399	>	}
2400	>	}
2401	>	if (task != null)
2402	>	task.doExec();
2403	>	if (root.status < 0 ||
2404	>	(u = (int)(ctl >>> 32)) >= 0 || (u >> UAC_SHIFT) >= 0)
2405	>	break;
2406	>	if (task == null) {
2407	>	helpSignal(root, q.poolIndex);
2408	>	if (root.status >= 0)
2409	>	helpComplete(root, SHARED_QUEUE);
2410	>	break;
2411	>	}
2412	>	}
2413	>	}
2414		}
2415
2416		/**
2417	<	* Accumulates steal count from a worker, clearing
2418	<	* the worker's value
2417	>	* Tries to help execute or signal availability of the given task
2418	>	* from submitter's queue in common pool.
2419		*/
2420	<	final void accumulateStealCount(ForkJoinWorkerThread w) {
2421	<	int sc = w.stealCount;
2422	<	if (sc != 0) {
2423	<	long c;
2424	<	w.stealCount = 0;
2425	<	do {} while (!UNSAFE.compareAndSwapLong(this, stealCountOffset,
2426	<	c = stealCount, c + sc));
2420	>	static void externalHelpJoin(ForkJoinTask<?> t) {
2421	>	// Some hard-to-avoid overlap with tryExternalUnpush
2422	>	ForkJoinPool p; WorkQueue[] ws; WorkQueue q, w; Submitter z;
2423	>	ForkJoinTask<?>[] a;  int m, s, n;
2424	>	if (t != null &&
2425	>	(z = submitters.get()) != null &&
2426	>	(p = commonPool) != null &&
2427	>	(ws = p.workQueues) != null &&
2428	>	(m = ws.length - 1) >= 0 &&
2429	>	(q = ws[m & z.seed & SQMASK]) != null &&
2430	>	(a = q.array) != null) {
2431	>	int am = a.length - 1;
2432	>	if ((s = q.top) != q.base) {
2433	>	long j = ((am & (s - 1)) << ASHIFT) + ABASE;
2434	>	if (U.getObject(a, j) == t &&
2435	>	U.compareAndSwapInt(q, QLOCK, 0, 1)) {
2436	>	if (q.array == a && q.top == s &&
2437	>	U.compareAndSwapObject(a, j, t, null)) {
2438	>	q.top = s - 1;
2439	>	q.qlock = 0;
2440	>	t.doExec();
2441	>	}
2442	>	else
2443	>	q.qlock = 0;
2444	>	}
2445	>	}
2446	>	if (t.status >= 0) {
2447	>	if (t instanceof CountedCompleter)
2448	>	p.externalHelpComplete(q, t);
2449	>	else
2450	>	p.helpSignal(t, q.poolIndex);
2451	>	}
2452		}
2453		}
2454
2455		/**
2456	<	* Returns the approximate (non-atomic) number of idle threads per
1227	<	* active thread.
2456	>	* Restricted version of helpQuiescePool for external callers
2457		*/
2458	<	final int idlePerActive() {
2459	<	int pc = parallelism; // use parallelism, not rc
2460	<	int ac = runState;    // no mask -- artifically boosts during shutdown
2461	<	// Use exact results for small values, saturate past 4
2462	<	return pc <= ac? 0 : pc >>> 1 <= ac? 1 : pc >>> 2 <= ac? 3 : pc >>> 3;
2458	>	static void externalHelpQuiescePool() {
2459	>	ForkJoinPool p; ForkJoinTask<?> t; WorkQueue q; int b;
2460	>	if ((p = commonPool) != null &&
2461	>	(q = p.findNonEmptyStealQueue(1)) != null &&
2462	>	(b = q.base) - q.top < 0 &&
2463	>	(t = q.pollAt(b)) != null) {
2464	>	if (q.base - q.top < 0)
2465	>	p.signalWork(q);
2466	>	t.doExec();
2467	>	}
2468		}
2469
2470	<	// Public and protected methods
2470	>	// Exported methods
2471
2472		// Constructors
2473
#	Line 1280 | Line 2514 | public class ForkJoinPool extends Abstra
2514		* use {@link #defaultForkJoinWorkerThreadFactory}.
2515		* @param handler the handler for internal worker threads that
2516		* terminate due to unrecoverable errors encountered while executing
2517	<	* tasks. For default value, use <code>null</code>.
2517	>	* tasks. For default value, use {@code null}.
2518		* @param asyncMode if true,
2519		* establishes local first-in-first-out scheduling mode for forked
2520		* tasks that are never joined. This mode may be more appropriate
2521		* than default locally stack-based mode in applications in which
2522		* worker threads only process event-style asynchronous tasks.
2523	<	* For default value, use <code>false</code>.
2523	>	* For default value, use {@code false}.
2524		* @throws IllegalArgumentException if parallelism less than or
2525		*         equal to zero, or greater than implementation limit
2526		* @throws NullPointerException if the factory is null
#	Line 1302 | Line 2536 | public class ForkJoinPool extends Abstra
2536		checkPermission();
2537		if (factory == null)
2538		throw new NullPointerException();
2539	<	if (parallelism <= 0 || parallelism > MAX_THREADS)
2539	>	if (parallelism <= 0 || parallelism > MAX_CAP)
2540		throw new IllegalArgumentException();
1307	–	this.parallelism = parallelism;
2541		this.factory = factory;
2542		this.ueh = handler;
2543	<	this.locallyFifo = asyncMode;
2544	<	int arraySize = initialArraySizeFor(parallelism);
2545	<	this.workers = new ForkJoinWorkerThread[arraySize];
2546	<	this.submissionQueue = new LinkedTransferQueue<ForkJoinTask<?>>();
2547	<	this.workerLock = new ReentrantLock();
2548	<	this.termination = new Phaser(1);
2549	<	this.poolNumber = poolNumberGenerator.incrementAndGet();
2543	>	this.config = parallelism | (asyncMode ? (FIFO_QUEUE << 16) : 0);
2544	>	long np = (long)(-parallelism); // offset ctl counts
2545	>	this.ctl = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK);
2546	>	int pn = nextPoolId();
2547	>	StringBuilder sb = new StringBuilder("ForkJoinPool-");
2548	>	sb.append(Integer.toString(pn));
2549	>	sb.append("-worker-");
2550	>	this.workerNamePrefix = sb.toString();
2551		}
2552
2553		/**
2554	<	* Returns initial power of two size for workers array.
2555	<	* @param pc the initial parallelism level
2556	<	*/
2557	<	private static int initialArraySizeFor(int pc) {
2558	<	// See Hackers Delight, sec 3.2. We know MAX_THREADS < (1 >>> 16)
2559	<	int size = pc < MAX_THREADS ? pc + 1 : MAX_THREADS;
2560	<	size |= size >>> 1;
2561	<	size |= size >>> 2;
2562	<	size |= size >>> 4;
2563	<	size |= size >>> 8;
2564	<	return size + 1;
2554	>	* Constructor for common pool, suitable only for static initialization.
2555	>	* Basically the same as above, but uses smallest possible initial footprint.
2556	>	*/
2557	>	ForkJoinPool(int parallelism, long ctl,
2558	>	ForkJoinWorkerThreadFactory factory,
2559	>	Thread.UncaughtExceptionHandler handler) {
2560	>	this.config = parallelism;
2561	>	this.ctl = ctl;
2562	>	this.factory = factory;
2563	>	this.ueh = handler;
2564	>	this.workerNamePrefix = "ForkJoinPool.commonPool-worker-";
2565		}
2566
1333	–	// Execution methods
1334	–
2567		/**
2568	<	* Common code for execute, invoke and submit
2568	>	* Returns the common pool instance. This pool is statically
2569	>	* constructed; its run state is unaffected by attempts to
2570	>	* {@link #shutdown} or {@link #shutdownNow}.
2571	>	*
2572	>	* @return the common pool instance
2573		*/
2574	<	private <T> void doSubmit(ForkJoinTask<T> task) {
2575	<	if (task == null)
2576	<	throw new NullPointerException();
1341	<	if (runState >= SHUTDOWN)
1342	<	throw new RejectedExecutionException();
1343	<	submissionQueue.offer(task);
1344	<	signalEvent();
1345	<	ensureEnoughWorkers();
2574	>	public static ForkJoinPool commonPool() {
2575	>	// assert commonPool != null : "static init error";
2576	>	return commonPool;
2577		}
2578
2579	+	// Execution methods
2580	+
2581		/**
2582		* Performs the given task, returning its result upon completion.
2583	<	* If the caller is already engaged in a fork/join computation in
2584	<	* the current pool, this method is equivalent in effect to
2585	<	* {@link ForkJoinTask#invoke}.
2583	>	* If the computation encounters an unchecked Exception or Error,
2584	>	* it is rethrown as the outcome of this invocation.  Rethrown
2585	>	* exceptions behave in the same way as regular exceptions, but,
2586	>	* when possible, contain stack traces (as displayed for example
2587	>	* using {@code ex.printStackTrace()}) of both the current thread
2588	>	* as well as the thread actually encountering the exception;
2589	>	* minimally only the latter.
2590		*
2591		* @param task the task
2592		* @return the task's result
#	Line 1358 | Line 2595 | public class ForkJoinPool extends Abstra
2595		*         scheduled for execution
2596		*/
2597		public <T> T invoke(ForkJoinTask<T> task) {
2598	<	doSubmit(task);
2598	>	if (task == null)
2599	>	throw new NullPointerException();
2600	>	externalPush(task);
2601		return task.join();
2602		}
2603
2604		/**
2605		* Arranges for (asynchronous) execution of the given task.
1367	–	* If the caller is already engaged in a fork/join computation in
1368	–	* the current pool, this method is equivalent in effect to
1369	–	* {@link ForkJoinTask#fork}.
2606		*
2607		* @param task the task
2608		* @throws NullPointerException if the task is null
#	Line 1374 | Line 2610 | public class ForkJoinPool extends Abstra
2610		*         scheduled for execution
2611		*/
2612		public void execute(ForkJoinTask<?> task) {
2613	<	doSubmit(task);
2613	>	if (task == null)
2614	>	throw new NullPointerException();
2615	>	externalPush(task);
2616		}
2617
2618		// AbstractExecutorService methods
#	Line 1385 | Line 2623 | public class ForkJoinPool extends Abstra
2623		*         scheduled for execution
2624		*/
2625		public void execute(Runnable task) {
2626	+	if (task == null)
2627	+	throw new NullPointerException();
2628		ForkJoinTask<?> job;
2629		if (task instanceof ForkJoinTask<?>) // avoid re-wrap
2630		job = (ForkJoinTask<?>) task;
2631		else
2632	<	job = ForkJoinTask.adapt(task, null);
2633	<	doSubmit(job);
2632	>	job = new ForkJoinTask.AdaptedRunnableAction(task);
2633	>	externalPush(job);
2634		}
2635
2636		/**
2637		* Submits a ForkJoinTask for execution.
1398	–	* If the caller is already engaged in a fork/join computation in
1399	–	* the current pool, this method is equivalent in effect to
1400	–	* {@link ForkJoinTask#fork}.
2638		*
2639		* @param task the task to submit
2640		* @return the task
#	Line 1406 | Line 2643 | public class ForkJoinPool extends Abstra
2643		*         scheduled for execution
2644		*/
2645		public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) {
2646	<	doSubmit(task);
2646	>	if (task == null)
2647	>	throw new NullPointerException();
2648	>	externalPush(task);
2649		return task;
2650		}
2651
#	Line 1416 | Line 2655 | public class ForkJoinPool extends Abstra
2655		*         scheduled for execution
2656		*/
2657		public <T> ForkJoinTask<T> submit(Callable<T> task) {
2658	<	ForkJoinTask<T> job = ForkJoinTask.adapt(task);
2659	<	doSubmit(job);
2658	>	ForkJoinTask<T> job = new ForkJoinTask.AdaptedCallable<T>(task);
2659	>	externalPush(job);
2660		return job;
2661		}
2662
#	Line 1427 | Line 2666 | public class ForkJoinPool extends Abstra
2666		*         scheduled for execution
2667		*/
2668		public <T> ForkJoinTask<T> submit(Runnable task, T result) {
2669	<	ForkJoinTask<T> job = ForkJoinTask.adapt(task, result);
2670	<	doSubmit(job);
2669	>	ForkJoinTask<T> job = new ForkJoinTask.AdaptedRunnable<T>(task, result);
2670	>	externalPush(job);
2671		return job;
2672		}
2673
#	Line 1438 | Line 2677 | public class ForkJoinPool extends Abstra
2677		*         scheduled for execution
2678		*/
2679		public ForkJoinTask<?> submit(Runnable task) {
2680	+	if (task == null)
2681	+	throw new NullPointerException();
2682		ForkJoinTask<?> job;
2683		if (task instanceof ForkJoinTask<?>) // avoid re-wrap
2684		job = (ForkJoinTask<?>) task;
2685		else
2686	<	job = ForkJoinTask.adapt(task, null);
2687	<	doSubmit(job);
2686	>	job = new ForkJoinTask.AdaptedRunnableAction(task);
2687	>	externalPush(job);
2688		return job;
2689		}
2690
#	Line 1452 | Line 2693 | public class ForkJoinPool extends Abstra
2693		* @throws RejectedExecutionException {@inheritDoc}
2694		*/
2695		public <T> List<Future<T>> invokeAll(Collection<? extends Callable<T>> tasks) {
2696	<	ArrayList<ForkJoinTask<T>> forkJoinTasks =
2697	<	new ArrayList<ForkJoinTask<T>>(tasks.size());
2698	<	for (Callable<T> task : tasks)
2699	<	forkJoinTasks.add(ForkJoinTask.adapt(task));
2700	<	invoke(new InvokeAll<T>(forkJoinTasks));
2701	<
2696	>	// In previous versions of this class, this method constructed
2697	>	// a task to run ForkJoinTask.invokeAll, but now external
2698	>	// invocation of multiple tasks is at least as efficient.
2699	>	List<ForkJoinTask<T>> fs = new ArrayList<ForkJoinTask<T>>(tasks.size());
2700	>	// Workaround needed because method wasn't declared with
2701	>	// wildcards in return type but should have been.
2702		@SuppressWarnings({"unchecked", "rawtypes"})
2703	<	List<Future<T>> futures = (List<Future<T>>) (List) forkJoinTasks;
1463	<	return futures;
1464	<	}
2703	>	List<Future<T>> futures = (List<Future<T>>) (List) fs;
2704
2705	<	static final class InvokeAll<T> extends RecursiveAction {
2706	<	final ArrayList<ForkJoinTask<T>> tasks;
2707	<	InvokeAll(ArrayList<ForkJoinTask<T>> tasks) { this.tasks = tasks; }
2708	<	public void compute() {
2709	<	try { invokeAll(tasks); }
2710	<	catch (Exception ignore) {}
2705	>	boolean done = false;
2706	>	try {
2707	>	for (Callable<T> t : tasks) {
2708	>	ForkJoinTask<T> f = new ForkJoinTask.AdaptedCallable<T>(t);
2709	>	externalPush(f);
2710	>	fs.add(f);
2711	>	}
2712	>	for (ForkJoinTask<T> f : fs)
2713	>	f.quietlyJoin();
2714	>	done = true;
2715	>	return futures;
2716	>	} finally {
2717	>	if (!done)
2718	>	for (ForkJoinTask<T> f : fs)
2719	>	f.cancel(false);
2720		}
1473	–	private static final long serialVersionUID = -7914297376763021607L;
2721		}
2722
2723		/**
#	Line 1498 | Line 2745 | public class ForkJoinPool extends Abstra
2745		* @return the targeted parallelism level of this pool
2746		*/
2747		public int getParallelism() {
2748	<	return parallelism;
2748	>	return config & SMASK;
2749	>	}
2750	>
2751	>	/**
2752	>	* Returns the targeted parallelism level of the common pool.
2753	>	*
2754	>	* @return the targeted parallelism level of the common pool
2755	>	*/
2756	>	public static int getCommonPoolParallelism() {
2757	>	return commonPoolParallelism;
2758		}
2759
2760		/**
2761		* Returns the number of worker threads that have started but not
2762	<	* yet terminated.  This result returned by this method may differ
2762	>	* yet terminated.  The result returned by this method may differ
2763		* from {@link #getParallelism} when threads are created to
2764		* maintain parallelism when others are cooperatively blocked.
2765		*
2766		* @return the number of worker threads
2767		*/
2768		public int getPoolSize() {
2769	<	return workerCounts >>> TOTAL_COUNT_SHIFT;
2769	>	return (config & SMASK) + (short)(ctl >>> TC_SHIFT);
2770		}
2771
2772		/**
#	Line 1520 | Line 2776 | public class ForkJoinPool extends Abstra
2776		* @return {@code true} if this pool uses async mode
2777		*/
2778		public boolean getAsyncMode() {
2779	<	return locallyFifo;
2779	>	return (config >>> 16) == FIFO_QUEUE;
2780		}
2781
2782		/**
#	Line 1532 | Line 2788 | public class ForkJoinPool extends Abstra
2788		* @return the number of worker threads
2789		*/
2790		public int getRunningThreadCount() {
2791	<	return workerCounts & RUNNING_COUNT_MASK;
2791	>	int rc = 0;
2792	>	WorkQueue[] ws; WorkQueue w;
2793	>	if ((ws = workQueues) != null) {
2794	>	for (int i = 1; i < ws.length; i += 2) {
2795	>	if ((w = ws[i]) != null && w.isApparentlyUnblocked())
2796	>	++rc;
2797	>	}
2798	>	}
2799	>	return rc;
2800		}
2801
2802		/**
#	Line 1543 | Line 2807 | public class ForkJoinPool extends Abstra
2807		* @return the number of active threads
2808		*/
2809		public int getActiveThreadCount() {
2810	<	return runState & ACTIVE_COUNT_MASK;
2810	>	int r = (config & SMASK) + (int)(ctl >> AC_SHIFT);
2811	>	return (r <= 0) ? 0 : r; // suppress momentarily negative values
2812		}
2813
2814		/**
#	Line 1558 | Line 2823 | public class ForkJoinPool extends Abstra
2823		* @return {@code true} if all threads are currently idle
2824		*/
2825		public boolean isQuiescent() {
2826	<	return (runState & ACTIVE_COUNT_MASK) == 0;
2826	>	return (int)(ctl >> AC_SHIFT) + (config & SMASK) == 0;
2827		}
2828
2829		/**
#	Line 1573 | Line 2838 | public class ForkJoinPool extends Abstra
2838		* @return the number of steals
2839		*/
2840		public long getStealCount() {
2841	<	return stealCount;
2841	>	long count = stealCount;
2842	>	WorkQueue[] ws; WorkQueue w;
2843	>	if ((ws = workQueues) != null) {
2844	>	for (int i = 1; i < ws.length; i += 2) {
2845	>	if ((w = ws[i]) != null)
2846	>	count += w.nsteals;
2847	>	}
2848	>	}
2849	>	return count;
2850		}
2851
2852		/**
#	Line 1588 | Line 2861 | public class ForkJoinPool extends Abstra
2861		*/
2862		public long getQueuedTaskCount() {
2863		long count = 0;
2864	<	ForkJoinWorkerThread[] ws = workers;
2865	<	int nws = ws.length;
2866	<	for (int i = 0; i < nws; ++i) {
2867	<	ForkJoinWorkerThread w = ws[i];
2868	<	if (w != null)
2869	<	count += w.getQueueSize();
2864	>	WorkQueue[] ws; WorkQueue w;
2865	>	if ((ws = workQueues) != null) {
2866	>	for (int i = 1; i < ws.length; i += 2) {
2867	>	if ((w = ws[i]) != null)
2868	>	count += w.queueSize();
2869	>	}
2870		}
2871		return count;
2872		}
2873
2874		/**
2875		* Returns an estimate of the number of tasks submitted to this
2876	<	* pool that have not yet begun executing.  This method takes time
2877	<	* proportional to the number of submissions.
2876	>	* pool that have not yet begun executing.  This method may take
2877	>	* time proportional to the number of submissions.
2878		*
2879		* @return the number of queued submissions
2880		*/
2881		public int getQueuedSubmissionCount() {
2882	<	return submissionQueue.size();
2882	>	int count = 0;
2883	>	WorkQueue[] ws; WorkQueue w;
2884	>	if ((ws = workQueues) != null) {
2885	>	for (int i = 0; i < ws.length; i += 2) {
2886	>	if ((w = ws[i]) != null)
2887	>	count += w.queueSize();
2888	>	}
2889	>	}
2890	>	return count;
2891		}
2892
2893		/**
#	Line 1616 | Line 2897 | public class ForkJoinPool extends Abstra
2897		* @return {@code true} if there are any queued submissions
2898		*/
2899		public boolean hasQueuedSubmissions() {
2900	<	return !submissionQueue.isEmpty();
2900	>	WorkQueue[] ws; WorkQueue w;
2901	>	if ((ws = workQueues) != null) {
2902	>	for (int i = 0; i < ws.length; i += 2) {
2903	>	if ((w = ws[i]) != null && !w.isEmpty())
2904	>	return true;
2905	>	}
2906	>	}
2907	>	return false;
2908		}
2909
2910		/**
#	Line 1627 | Line 2915 | public class ForkJoinPool extends Abstra
2915		* @return the next submission, or {@code null} if none
2916		*/
2917		protected ForkJoinTask<?> pollSubmission() {
2918	<	return submissionQueue.poll();
2918	>	WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2919	>	if ((ws = workQueues) != null) {
2920	>	for (int i = 0; i < ws.length; i += 2) {
2921	>	if ((w = ws[i]) != null && (t = w.poll()) != null)
2922	>	return t;
2923	>	}
2924	>	}
2925	>	return null;
2926		}
2927
2928		/**
#	Line 1648 | Line 2943 | public class ForkJoinPool extends Abstra
2943		* @return the number of elements transferred
2944		*/
2945		protected int drainTasksTo(Collection<? super ForkJoinTask<?>> c) {
1651	–	int n = submissionQueue.drainTo(c);
1652	–	ForkJoinWorkerThread[] ws = workers;
1653	–	int nws = ws.length;
1654	–	for (int i = 0; i < nws; ++i) {
1655	–	ForkJoinWorkerThread w = ws[i];
1656	–	if (w != null)
1657	–	n += w.drainTasksTo(c);
1658	–	}
1659	–	return n;
1660	–	}
1661	–
1662	–	/**
1663	–	* Returns count of total parks by existing workers.
1664	–	* Used during development only since not meaningful to users.
1665	–	*/
1666	–	private int collectParkCount() {
2946		int count = 0;
2947	<	ForkJoinWorkerThread[] ws = workers;
2948	<	int nws = ws.length;
2949	<	for (int i = 0; i < nws; ++i) {
2950	<	ForkJoinWorkerThread w = ws[i];
2951	<	if (w != null)
2952	<	count += w.parkCount;
2947	>	WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2948	>	if ((ws = workQueues) != null) {
2949	>	for (int i = 0; i < ws.length; ++i) {
2950	>	if ((w = ws[i]) != null) {
2951	>	while ((t = w.poll()) != null) {
2952	>	c.add(t);
2953	>	++count;
2954	>	}
2955	>	}
2956	>	}
2957		}
2958		return count;
2959		}
#	Line 1683 | Line 2966 | public class ForkJoinPool extends Abstra
2966		* @return a string identifying this pool, as well as its state
2967		*/
2968		public String toString() {
2969	<	long st = getStealCount();
2970	<	long qt = getQueuedTaskCount();
2971	<	long qs = getQueuedSubmissionCount();
2972	<	int wc = workerCounts;
2973	<	int tc = wc >>> TOTAL_COUNT_SHIFT;
2974	<	int rc = wc & RUNNING_COUNT_MASK;
2975	<	int pc = parallelism;
2976	<	int rs = runState;
2977	<	int ac = rs & ACTIVE_COUNT_MASK;
2978	<	//        int pk = collectParkCount();
2969	>	// Use a single pass through workQueues to collect counts
2970	>	long qt = 0L, qs = 0L; int rc = 0;
2971	>	long st = stealCount;
2972	>	long c = ctl;
2973	>	WorkQueue[] ws; WorkQueue w;
2974	>	if ((ws = workQueues) != null) {
2975	>	for (int i = 0; i < ws.length; ++i) {
2976	>	if ((w = ws[i]) != null) {
2977	>	int size = w.queueSize();
2978	>	if ((i & 1) == 0)
2979	>	qs += size;
2980	>	else {
2981	>	qt += size;
2982	>	st += w.nsteals;
2983	>	if (w.isApparentlyUnblocked())
2984	>	++rc;
2985	>	}
2986	>	}
2987	>	}
2988	>	}
2989	>	int pc = (config & SMASK);
2990	>	int tc = pc + (short)(c >>> TC_SHIFT);
2991	>	int ac = pc + (int)(c >> AC_SHIFT);
2992	>	if (ac < 0) // ignore transient negative
2993	>	ac = 0;
2994	>	String level;
2995	>	if ((c & STOP_BIT) != 0)
2996	>	level = (tc == 0) ? "Terminated" : "Terminating";
2997	>	else
2998	>	level = plock < 0 ? "Shutting down" : "Running";
2999		return super.toString() +
3000	<	"[" + runLevelToString(rs) +
3000	>	"[" + level +
3001		", parallelism = " + pc +
3002		", size = " + tc +
3003		", active = " + ac +
#	Line 1702 | Line 3005 | public class ForkJoinPool extends Abstra
3005		", steals = " + st +
3006		", tasks = " + qt +
3007		", submissions = " + qs +
1705	–	//            ", parks = " + pk +
3008		"]";
3009		}
3010
1709	–	private static String runLevelToString(int s) {
1710	–	return ((s & TERMINATED) != 0 ? "Terminated" :
1711	–	((s & TERMINATING) != 0 ? "Terminating" :
1712	–	((s & SHUTDOWN) != 0 ? "Shutting down" :
1713	–	"Running")));
1714	–	}
1715	–
3011		/**
3012	<	* Initiates an orderly shutdown in which previously submitted
3013	<	* tasks are executed, but no new tasks will be accepted.
3014	<	* Invocation has no additional effect if already shut down.
3015	<	* Tasks that are in the process of being submitted concurrently
3016	<	* during the course of this method may or may not be rejected.
3012	>	* Possibly initiates an orderly shutdown in which previously
3013	>	* submitted tasks are executed, but no new tasks will be
3014	>	* accepted. Invocation has no effect on execution state if this
3015	>	* is the {@link #commonPool}, and no additional effect if
3016	>	* already shut down.  Tasks that are in the process of being
3017	>	* submitted concurrently during the course of this method may or
3018	>	* may not be rejected.
3019		*
3020		* @throws SecurityException if a security manager exists and
3021		*         the caller is not permitted to modify threads
#	Line 1727 | Line 3024 | public class ForkJoinPool extends Abstra
3024		*/
3025		public void shutdown() {
3026		checkPermission();
3027	<	advanceRunLevel(SHUTDOWN);
1731	<	tryTerminate(false);
3027	>	tryTerminate(false, true);
3028		}
3029
3030		/**
3031	<	* Attempts to cancel and/or stop all tasks, and reject all
3032	<	* subsequently submitted tasks.  Tasks that are in the process of
3033	<	* being submitted or executed concurrently during the course of
3034	<	* this method may or may not be rejected. This method cancels
3035	<	* both existing and unexecuted tasks, in order to permit
3036	<	* termination in the presence of task dependencies. So the method
3037	<	* always returns an empty list (unlike the case for some other
3038	<	* Executors).
3031	>	* Possibly attempts to cancel and/or stop all tasks, and reject
3032	>	* all subsequently submitted tasks.  Invocation has no effect on
3033	>	* execution state if this is the {@link #commonPool}, and no
3034	>	* additional effect if already shut down. Otherwise, tasks that
3035	>	* are in the process of being submitted or executed concurrently
3036	>	* during the course of this method may or may not be
3037	>	* rejected. This method cancels both existing and unexecuted
3038	>	* tasks, in order to permit termination in the presence of task
3039	>	* dependencies. So the method always returns an empty list
3040	>	* (unlike the case for some other Executors).
3041		*
3042		* @return an empty list
3043		* @throws SecurityException if a security manager exists and
#	Line 1749 | Line 3047 | public class ForkJoinPool extends Abstra
3047		*/
3048		public List<Runnable> shutdownNow() {
3049		checkPermission();
3050	<	tryTerminate(true);
3050	>	tryTerminate(true, true);
3051		return Collections.emptyList();
3052		}
3053
#	Line 1759 | Line 3057 | public class ForkJoinPool extends Abstra
3057		* @return {@code true} if all tasks have completed following shut down
3058		*/
3059		public boolean isTerminated() {
3060	<	return runState >= TERMINATED;
3060	>	long c = ctl;
3061	>	return ((c & STOP_BIT) != 0L &&
3062	>	(short)(c >>> TC_SHIFT) == -(config & SMASK));
3063		}
3064
3065		/**
#	Line 1767 | Line 3067 | public class ForkJoinPool extends Abstra
3067		* commenced but not yet completed.  This method may be useful for
3068		* debugging. A return of {@code true} reported a sufficient
3069		* period after shutdown may indicate that submitted tasks have
3070	<	* ignored or suppressed interruption, causing this executor not
3071	<	* to properly terminate.
3070	>	* ignored or suppressed interruption, or are waiting for I/O,
3071	>	* causing this executor not to properly terminate. (See the
3072	>	* advisory notes for class {@link ForkJoinTask} stating that
3073	>	* tasks should not normally entail blocking operations.  But if
3074	>	* they do, they must abort them on interrupt.)
3075		*
3076		* @return {@code true} if terminating but not yet terminated
3077		*/
3078		public boolean isTerminating() {
3079	<	return (runState & (TERMINATING|TERMINATED)) == TERMINATING;
3079	>	long c = ctl;
3080	>	return ((c & STOP_BIT) != 0L &&
3081	>	(short)(c >>> TC_SHIFT) != -(config & SMASK));
3082		}
3083
3084		/**
#	Line 1782 | Line 3087 | public class ForkJoinPool extends Abstra
3087		* @return {@code true} if this pool has been shut down
3088		*/
3089		public boolean isShutdown() {
3090	<	return runState >= SHUTDOWN;
3090	>	return plock < 0;
3091		}
3092
3093		/**
3094	<	* Blocks until all tasks have completed execution after a shutdown
3095	<	* request, or the timeout occurs, or the current thread is
3096	<	* interrupted, whichever happens first.
3094	>	* Blocks until all tasks have completed execution after a
3095	>	* shutdown request, or the timeout occurs, or the current thread
3096	>	* is interrupted, whichever happens first. Note that the {@link
3097	>	* #commonPool()} never terminates until program shutdown so
3098	>	* this method will always time out.
3099		*
3100		* @param timeout the maximum time to wait
3101		* @param unit the time unit of the timeout argument
#	Line 1798 | Line 3105 | public class ForkJoinPool extends Abstra
3105		*/
3106		public boolean awaitTermination(long timeout, TimeUnit unit)
3107		throws InterruptedException {
3108	<	try {
3109	<	return termination.awaitAdvanceInterruptibly(0, timeout, unit) > 0;
3110	<	} catch(TimeoutException ex) {
3111	<	return false;
3108	>	long nanos = unit.toNanos(timeout);
3109	>	if (isTerminated())
3110	>	return true;
3111	>	long startTime = System.nanoTime();
3112	>	boolean terminated = false;
3113	>	synchronized (this) {
3114	>	for (long waitTime = nanos, millis = 0L;;) {
3115	>	if (terminated = isTerminated() ||
3116	>	waitTime <= 0L ||
3117	>	(millis = unit.toMillis(waitTime)) <= 0L)
3118	>	break;
3119	>	wait(millis);
3120	>	waitTime = nanos - (System.nanoTime() - startTime);
3121	>	}
3122		}
3123	+	return terminated;
3124		}
3125
3126		/**
3127		* Interface for extending managed parallelism for tasks running
3128		* in {@link ForkJoinPool}s.
3129		*
3130	<	* <p>A {@code ManagedBlocker} provides two methods.
3131	<	* Method {@code isReleasable} must return {@code true} if
3132	<	* blocking is not necessary. Method {@code block} blocks the
3133	<	* current thread if necessary (perhaps internally invoking
3134	<	* {@code isReleasable} before actually blocking).
3130	>	* <p>A {@code ManagedBlocker} provides two methods.  Method
3131	>	* {@code isReleasable} must return {@code true} if blocking is
3132	>	* not necessary. Method {@code block} blocks the current thread
3133	>	* if necessary (perhaps internally invoking {@code isReleasable}
3134	>	* before actually blocking). These actions are performed by any
3135	>	* thread invoking {@link ForkJoinPool#managedBlock}.  The
3136	>	* unusual methods in this API accommodate synchronizers that may,
3137	>	* but don't usually, block for long periods. Similarly, they
3138	>	* allow more efficient internal handling of cases in which
3139	>	* additional workers may be, but usually are not, needed to
3140	>	* ensure sufficient parallelism.  Toward this end,
3141	>	* implementations of method {@code isReleasable} must be amenable
3142	>	* to repeated invocation.
3143		*
3144		* <p>For example, here is a ManagedBlocker based on a
3145		* ReentrantLock:
#	Line 1831 | Line 3157 | public class ForkJoinPool extends Abstra
3157		*     return hasLock || (hasLock = lock.tryLock());
3158		*   }
3159		* }}</pre>
3160	+	*
3161	+	* <p>Here is a class that possibly blocks waiting for an
3162	+	* item on a given queue:
3163	+	*  <pre> {@code
3164	+	* class QueueTaker<E> implements ManagedBlocker {
3165	+	*   final BlockingQueue<E> queue;
3166	+	*   volatile E item = null;
3167	+	*   QueueTaker(BlockingQueue<E> q) { this.queue = q; }
3168	+	*   public boolean block() throws InterruptedException {
3169	+	*     if (item == null)
3170	+	*       item = queue.take();
3171	+	*     return true;
3172	+	*   }
3173	+	*   public boolean isReleasable() {
3174	+	*     return item != null || (item = queue.poll()) != null;
3175	+	*   }
3176	+	*   public E getItem() { // call after pool.managedBlock completes
3177	+	*     return item;
3178	+	*   }
3179	+	* }}</pre>
3180		*/
3181		public static interface ManagedBlocker {
3182		/**
#	Line 1873 | Line 3219 | public class ForkJoinPool extends Abstra
3219		public static void managedBlock(ManagedBlocker blocker)
3220		throws InterruptedException {
3221		Thread t = Thread.currentThread();
3222	<	if (t instanceof ForkJoinWorkerThread)
3223	<	((ForkJoinWorkerThread) t).pool.awaitBlocker(blocker);
3222	>	if (t instanceof ForkJoinWorkerThread) {
3223	>	ForkJoinPool p = ((ForkJoinWorkerThread)t).pool;
3224	>	while (!blocker.isReleasable()) { // variant of helpSignal
3225	>	WorkQueue[] ws; WorkQueue q; int m, u;
3226	>	if ((ws = p.workQueues) != null && (m = ws.length - 1) >= 0) {
3227	>	for (int i = 0; i <= m; ++i) {
3228	>	if (blocker.isReleasable())
3229	>	return;
3230	>	if ((q = ws[i]) != null && q.base - q.top < 0) {
3231	>	p.signalWork(q);
3232	>	if ((u = (int)(p.ctl >>> 32)) >= 0 ||
3233	>	(u >> UAC_SHIFT) >= 0)
3234	>	break;
3235	>	}
3236	>	}
3237	>	}
3238	>	if (p.tryCompensate()) {
3239	>	try {
3240	>	do {} while (!blocker.isReleasable() &&
3241	>	!blocker.block());
3242	>	} finally {
3243	>	p.incrementActiveCount();
3244	>	}
3245	>	break;
3246	>	}
3247	>	}
3248	>	}
3249		else {
3250	<	do {} while (!blocker.isReleasable() && !blocker.block());
3250	>	do {} while (!blocker.isReleasable() &&
3251	>	!blocker.block());
3252		}
3253		}
3254
#	Line 1885 | Line 3257 | public class ForkJoinPool extends Abstra
3257		// implement RunnableFuture.
3258
3259		protected <T> RunnableFuture<T> newTaskFor(Runnable runnable, T value) {
3260	<	return (RunnableFuture<T>) ForkJoinTask.adapt(runnable, value);
3260	>	return new ForkJoinTask.AdaptedRunnable<T>(runnable, value);
3261		}
3262
3263		protected <T> RunnableFuture<T> newTaskFor(Callable<T> callable) {
3264	<	return (RunnableFuture<T>) ForkJoinTask.adapt(callable);
3264	>	return new ForkJoinTask.AdaptedCallable<T>(callable);
3265		}
3266
3267		// Unsafe mechanics
3268	+	private static final sun.misc.Unsafe U;
3269	+	private static final long CTL;
3270	+	private static final long PARKBLOCKER;
3271	+	private static final int ABASE;
3272	+	private static final int ASHIFT;
3273	+	private static final long STEALCOUNT;
3274	+	private static final long PLOCK;
3275	+	private static final long INDEXSEED;
3276	+	private static final long QLOCK;
3277
3278	<	private static final sun.misc.Unsafe UNSAFE = getUnsafe();
3279	<	private static final long workerCountsOffset =
1899	<	objectFieldOffset("workerCounts", ForkJoinPool.class);
1900	<	private static final long runStateOffset =
1901	<	objectFieldOffset("runState", ForkJoinPool.class);
1902	<	private static final long eventCountOffset =
1903	<	objectFieldOffset("eventCount", ForkJoinPool.class);
1904	<	private static final long eventWaitersOffset =
1905	<	objectFieldOffset("eventWaiters",ForkJoinPool.class);
1906	<	private static final long stealCountOffset =
1907	<	objectFieldOffset("stealCount",ForkJoinPool.class);
1908	<
1909	<	private static long objectFieldOffset(String field, Class<?> klazz) {
3278	>	static {
3279	>	int s; // initialize field offsets for CAS etc
3280		try {
3281	<	return UNSAFE.objectFieldOffset(klazz.getDeclaredField(field));
3282	<	} catch (NoSuchFieldException e) {
3283	<	// Convert Exception to corresponding Error
3284	<	NoSuchFieldError error = new NoSuchFieldError(field);
3285	<	error.initCause(e);
3286	<	throw error;
3287	<	}
3281	>	U = getUnsafe();
3282	>	Class<?> k = ForkJoinPool.class;
3283	>	CTL = U.objectFieldOffset
3284	>	(k.getDeclaredField("ctl"));
3285	>	STEALCOUNT = U.objectFieldOffset
3286	>	(k.getDeclaredField("stealCount"));
3287	>	PLOCK = U.objectFieldOffset
3288	>	(k.getDeclaredField("plock"));
3289	>	INDEXSEED = U.objectFieldOffset
3290	>	(k.getDeclaredField("indexSeed"));
3291	>	Class<?> tk = Thread.class;
3292	>	PARKBLOCKER = U.objectFieldOffset
3293	>	(tk.getDeclaredField("parkBlocker"));
3294	>	Class<?> wk = WorkQueue.class;
3295	>	QLOCK = U.objectFieldOffset
3296	>	(wk.getDeclaredField("qlock"));
3297	>	Class<?> ak = ForkJoinTask[].class;
3298	>	ABASE = U.arrayBaseOffset(ak);
3299	>	s = U.arrayIndexScale(ak);
3300	>	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
3301	>	} catch (Exception e) {
3302	>	throw new Error(e);
3303	>	}
3304	>	if ((s & (s-1)) != 0)
3305	>	throw new Error("data type scale not a power of two");
3306	>
3307	>	submitters = new ThreadLocal<Submitter>();
3308	>	ForkJoinWorkerThreadFactory fac = defaultForkJoinWorkerThreadFactory =
3309	>	new DefaultForkJoinWorkerThreadFactory();
3310	>	modifyThreadPermission = new RuntimePermission("modifyThread");
3311	>
3312	>	/*
3313	>	* Establish common pool parameters.  For extra caution,
3314	>	* computations to set up common pool state are here; the
3315	>	* constructor just assigns these values to fields.
3316	>	*/
3317	>
3318	>	int par = 0;
3319	>	Thread.UncaughtExceptionHandler handler = null;
3320	>	try {  // TBD: limit or report ignored exceptions?
3321	>	String pp = System.getProperty
3322	>	("java.util.concurrent.ForkJoinPool.common.parallelism");
3323	>	String hp = System.getProperty
3324	>	("java.util.concurrent.ForkJoinPool.common.exceptionHandler");
3325	>	String fp = System.getProperty
3326	>	("java.util.concurrent.ForkJoinPool.common.threadFactory");
3327	>	if (fp != null)
3328	>	fac = ((ForkJoinWorkerThreadFactory)ClassLoader.
3329	>	getSystemClassLoader().loadClass(fp).newInstance());
3330	>	if (hp != null)
3331	>	handler = ((Thread.UncaughtExceptionHandler)ClassLoader.
3332	>	getSystemClassLoader().loadClass(hp).newInstance());
3333	>	if (pp != null)
3334	>	par = Integer.parseInt(pp);
3335	>	} catch (Exception ignore) {
3336	>	}
3337	>
3338	>	if (par <= 0)
3339	>	par = Runtime.getRuntime().availableProcessors();
3340	>	if (par > MAX_CAP)
3341	>	par = MAX_CAP;
3342	>	commonPoolParallelism = par;
3343	>	long np = (long)(-par); // precompute initial ctl value
3344	>	long ct = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK);
3345	>
3346	>	commonPool = new ForkJoinPool(par, ct, fac, handler);
3347		}
3348
3349	+
3350		/**
3351		* Returns a sun.misc.Unsafe.  Suitable for use in a 3rd party package.
3352		* Replace with a simple call to Unsafe.getUnsafe when integrating
#	Line 1944 | Line 3374 | public class ForkJoinPool extends Abstra
3374		}
3375		}
3376		}
3377	+
3378		}

Diff Legend

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