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 1.61 by dl, Wed Aug 11 18:45:12 2010 UTC vs.
Revision 1.150 by jsr166, Wed Nov 21 22:15:41 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 IO 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 following
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 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 take actions when
147	<	* one worker is waiting to join a task stolen (or always held by)
148	<	* another.  Becauae we are multiplexing many tasks on to a pool
149	<	* of workers, we can't just let them block (as in Thread.join).
150	<	* We also cannot just reassign the joiner's run-time stack with
151	<	* another and replace it later, which would be a form of
152	<	* "continuation", that even if possible is not necessarily a good
153	<	* idea. Given that the creation costs of most threads on most
154	<	* systems mainly surrounds setting up runtime stacks, thread
155	<	* creation and switching is usually not much more expensive than
156	<	* stack creation and switching, and is more flexible). Instead we
157	<	* combine two tactics:
128	>	* This class and its nested classes provide the main
129	>	* functionality and control for a set of worker threads:
130	>	* Submissions from non-FJ threads enter into submission queues.
131	>	* Workers take these tasks and typically split them into subtasks
132	>	* that may be stolen by other workers.  Preference rules give
133	>	* first priority to processing tasks from their own queues (LIFO
134	>	* or FIFO, depending on mode), then to randomized FIFO steals of
135	>	* tasks in other queues.
136	>	*
137	>	* WorkQueues
138	>	* ==========
139	>	*
140	>	* Most operations occur within work-stealing queues (in nested
141	>	* class WorkQueue).  These are special forms of Deques that
142	>	* support only three of the four possible end-operations -- push,
143	>	* pop, and poll (aka steal), under the further constraints that
144	>	* push and pop are called only from the owning thread (or, as
145	>	* extended here, under a lock), while poll may be called from
146	>	* other threads.  (If you are unfamiliar with them, you probably
147	>	* want to read Herlihy and Shavit's book "The Art of
148	>	* Multiprocessor programming", chapter 16 describing these in
149	>	* more detail before proceeding.)  The main work-stealing queue
150	>	* design is roughly similar to those in the papers "Dynamic
151	>	* Circular Work-Stealing Deque" by Chase and Lev, SPAA 2005
152	>	* (http://research.sun.com/scalable/pubs/index.html) and
153	>	* "Idempotent work stealing" by Michael, Saraswat, and Vechev,
154	>	* PPoPP 2009 (http://portal.acm.org/citation.cfm?id=1504186).
155	>	* The main differences ultimately stem from GC requirements that
156	>	* we null out taken slots as soon as we can, to maintain as small
157	>	* a footprint as possible even in programs generating huge
158	>	* numbers of tasks. To accomplish this, we shift the CAS
159	>	* arbitrating pop vs poll (steal) from being on the indices
160	>	* ("base" and "top") to the slots themselves.  So, both a
161	>	* successful pop and poll mainly entail a CAS of a slot from
162	>	* non-null to null.  Because we rely on CASes of references, we
163	>	* do not need tag bits on base or top.  They are simple ints as
164	>	* used in any circular array-based queue (see for example
165	>	* ArrayDeque).  Updates to the indices must still be ordered in a
166	>	* way that guarantees that top == base means the queue is empty,
167	>	* but otherwise may err on the side of possibly making the queue
168	>	* appear nonempty when a push, pop, or poll have not fully
169	>	* committed. Note that this means that the poll operation,
170	>	* considered individually, is not wait-free. One thief cannot
171	>	* successfully continue until another in-progress one (or, if
172	>	* previously empty, a push) completes.  However, in the
173	>	* aggregate, we ensure at least probabilistic non-blockingness.
174	>	* If an attempted steal fails, a thief always chooses a different
175	>	* random victim target to try next. So, in order for one thief to
176	>	* progress, it suffices for any in-progress poll or new push on
177	>	* any empty queue to complete. (This is why we normally use
178	>	* method pollAt and its variants that try once at the apparent
179	>	* base index, else consider alternative actions, rather than
180	>	* method poll.)
181	>	*
182	>	* This approach also enables support of a user mode in which local
183	>	* task processing is in FIFO, not LIFO order, simply by using
184	>	* poll rather than pop.  This can be useful in message-passing
185	>	* frameworks in which tasks are never joined.  However neither
186	>	* mode considers affinities, loads, cache localities, etc, so
187	>	* rarely provide the best possible performance on a given
188	>	* machine, but portably provide good throughput by averaging over
189	>	* these factors.  (Further, even if we did try to use such
190	>	* information, we do not usually have a basis for exploiting it.
191	>	* For example, some sets of tasks profit from cache affinities,
192	>	* but others are harmed by cache pollution effects.)
193	>	*
194	>	* WorkQueues are also used in a similar way for tasks submitted
195	>	* to the pool. We cannot mix these tasks in the same queues used
196	>	* for work-stealing (this would contaminate lifo/fifo
197	>	* processing). Instead, we randomly associate submission queues
198	>	* with submitting threads, using a form of hashing.  The
199	>	* ThreadLocal Submitter class contains a value initially used as
200	>	* a hash code for choosing existing queues, but may be randomly
201	>	* repositioned upon contention with other submitters.  In
202	>	* essence, submitters act like workers except that they are
203	>	* restricted to executing local tasks that they submitted (or in
204	>	* the case of CountedCompleters, others with the same root task).
205	>	* However, because most shared/external queue operations are more
206	>	* expensive than internal, and because, at steady state, external
207	>	* submitters will compete for CPU with workers, ForkJoinTask.join
208	>	* and related methods disable them from repeatedly helping to
209	>	* process tasks if all workers are active.  Insertion of tasks in
210	>	* shared mode requires a lock (mainly to protect in the case of
211	>	* resizing) but we use only a simple spinlock (using bits in
212	>	* field qlock), because submitters encountering a busy queue move
213	>	* on to try or create other queues -- they block only when
214	>	* creating and registering new queues.
215	>	*
216	>	* Management
217	>	* ==========
218	>	*
219	>	* The main throughput advantages of work-stealing stem from
220	>	* decentralized control -- workers mostly 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.  Method
156	<	*      ForkJoinWorkerThread.helpJoinTask tracks joining->stealing
157	<	*      links to try to find such a task.
364	>	*      would be running if the steal had not occurred.
365		*
366		*   Compensating: Unless there are already enough live threads,
367	<	*      method helpMaintainParallelism() may create or or
368	<	*      re-activate a spare thread to compensate for blocked
369	<	*      joiners until they unblock.
370	<	*
371	<	* Because the determining existence of conservatively safe
372	<	* helping targets, the availability of already-created spares,
373	<	* and the apparent need to create new spares are all racy and
374	<	* require heuristic guidance, we rely on multiple retries of
375	<	* each. Further, because it is impossible to keep exactly the
376	<	* target (parallelism) number of threads running at any given
170	<	* time, we allow compensation during joins to fail, and enlist
171	<	* all other threads to help out whenever they are not otherwise
172	<	* occupied (i.e., mainly in method preStep).
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 main throughput advantages of work-stealing stem from
382	<	* decentralized control -- workers mostly steal tasks from each
383	<	* other. We do not want to negate this by creating bottlenecks
384	<	* implementing other management responsibilities. So we use a
385	<	* collection of techniques that avoid, reduce, or cope well with
386	<	* contention. These entail several instances of bit-packing into
387	<	* CASable fields to maintain only the minimally required
388	<	* atomicity. To enable such packing, we restrict maximum
389	<	* parallelism to (1<<15)-1 (enabling twice this (to accommodate
390	<	* unbalanced increments and decrements) to fit into a 16 bit
391	<	* field, which is far in excess of normal operating range.  Even
392	<	* though updates to some of these bookkeeping fields do sometimes
393	<	* contend with each other, they don't normally cache-contend with
394	<	* updates to others enough to warrant memory padding or
395	<	* isolation. So they are all held as fields of ForkJoinPool
396	<	* objects.  The main capabilities are as follows:
397	<	*
398	<	* 1. Creating and removing workers. Workers are recorded in the
399	<	* "workers" array. This is an array as opposed to some other data
400	<	* structure to support index-based random steals by workers.
401	<	* Updates to the array recording new workers and unrecording
402	<	* terminated ones are protected from each other by a lock
403	<	* (workerLock) but the array is otherwise concurrently readable,
404	<	* and accessed directly by workers. To simplify index-based
405	<	* operations, the array size is always a power of two, and all
406	<	* readers must tolerate null slots. Currently, all worker thread
407	<	* creation is on-demand, triggered by task submissions,
408	<	* replacement of terminated workers, and/or compensation for
409	<	* blocked workers. However, all other support code is set up to
410	<	* work with other policies.
411	<	*
412	<	* To ensure that we do not hold on to worker references that
413	<	* would prevent GC, ALL accesses to workers are via indices into
414	<	* the workers array (which is one source of some of the unusual
415	<	* code constructions here). In essence, the workers array serves
416	<	* as a WeakReference mechanism. Thus for example the event queue
417	<	* stores worker indices, not worker references. Access to the
418	<	* workers in associated methods (for example releaseEventWaiters)
419	<	* must both index-check and null-check the IDs. All such accesses
420	<	* ignore bad IDs by returning out early from what they are doing,
421	<	* since this can only be associated with shutdown, in which case
422	<	* it is OK to give up. On termination, we just clobber these
423	<	* data structures without trying to use them.
424	<	*
425	<	* 2. Bookkeeping for dynamically adding and removing workers. We
426	<	* aim to approximately maintain the given level of parallelism.
427	<	* When some workers are known to be blocked (on joins or via
428	<	* ManagedBlocker), we may create or resume others to take their
429	<	* place until they unblock (see below). Implementing this
430	<	* requires counts of the number of "running" threads (i.e., those
431	<	* that are neither blocked nor artifically suspended) as well as
432	<	* the total number.  These two values are packed into one field,
433	<	* "workerCounts" because we need accurate snapshots when deciding
434	<	* to create, resume or suspend.  Note however that the
435	<	* correspondance of these counts to reality is not guaranteed. In
436	<	* particular updates for unblocked threads may lag until they
437	<	* actually wake up.
438	<	*
439	<	* 3. Maintaining global run state. The run state of the pool
440	<	* consists of a runLevel (SHUTDOWN, TERMINATING, etc) similar to
441	<	* those in other Executor implementations, as well as a count of
442	<	* "active" workers -- those that are, or soon will be, or
443	<	* recently were executing tasks. The runLevel and active count
444	<	* are packed together in order to correctly trigger shutdown and
445	<	* termination. Without care, active counts can be subject to very
446	<	* high contention.  We substantially reduce this contention by
447	<	* relaxing update rules.  A worker must claim active status
448	<	* prospectively, by activating if it sees that a submitted or
449	<	* stealable task exists (it may find after activating that the
450	<	* task no longer exists). It stays active while processing this
451	<	* task (if it exists) and any other local subtasks it produces,
452	<	* until it cannot find any other tasks. It then tries
453	<	* inactivating (see method preStep), but upon update contention
454	<	* instead scans for more tasks, later retrying inactivation if it
455	<	* doesn't find any.
456	<	*
457	<	* 4. Managing idle workers waiting for tasks. We cannot let
458	<	* workers spin indefinitely scanning for tasks when none are
459	<	* available. On the other hand, we must quickly prod them into
460	<	* action when new tasks are submitted or generated.  We
461	<	* park/unpark these idle workers using an event-count scheme.
462	<	* Field eventCount is incremented upon events that may enable
463	<	* workers that previously could not find a task to now find one:
464	<	* Submission of a new task to the pool, or another worker pushing
465	<	* a task onto a previously empty queue.  (We also use this
466	<	* mechanism for termination actions that require wakeups of idle
263	<	* workers).  Each worker maintains its last known event count,
264	<	* and blocks when a scan for work did not find a task AND its
265	<	* lastEventCount matches the current eventCount. Waiting idle
266	<	* workers are recorded in a variant of Treiber stack headed by
267	<	* field eventWaiters which, when nonzero, encodes the thread
268	<	* index and count awaited for by the worker thread most recently
269	<	* calling eventSync. This thread in turn has a record (field
270	<	* nextEventWaiter) for the next waiting worker.  In addition to
271	<	* allowing simpler decisions about need for wakeup, the event
272	<	* count bits in eventWaiters serve the role of tags to avoid ABA
273	<	* errors in Treiber stacks.  To reduce delays in task diffusion,
274	<	* workers not otherwise occupied may invoke method
275	<	* releaseEventWaiters, that removes and signals (unparks) workers
276	<	* not waiting on current count. To reduce stalls, To minimize
277	<	* task production stalls associate with signalling, any worker
278	<	* pushing a task on an empty queue invokes the weaker method
279	<	* signalWork, that only releases idle workers until it detects
280	<	* interference by other threads trying to release, and lets them
281	<	* take over.  The net effect is a tree-like diffusion of signals,
282	<	* where released threads (and possibly others) help with unparks.
283	<	* To further reduce contention effects a bit, failed CASes to
284	<	* increment field eventCount are tolerated without retries.
285	<	* Conceptually they are merged into the same event, which is OK
286	<	* when their only purpose is to enable workers to scan for work.
287	<	*
288	<	* 5. Managing suspension of extra workers. When a worker is about
289	<	* to block waiting for a join (or via ManagedBlockers), we may
290	<	* create a new thread to maintain parallelism level, or at least
291	<	* avoid starvation. Usually, extra threads are needed for only
292	<	* very short periods, yet join dependencies are such that we
293	<	* sometimes need them in bursts. Rather than create new threads
294	<	* each time this happens, we suspend no-longer-needed extra ones
295	<	* as "spares". For most purposes, we don't distinguish "extra"
296	<	* spare threads from normal "core" threads: On each call to
297	<	* preStep (the only point at which we can do this) a worker
298	<	* checks to see if there are now too many running workers, and if
299	<	* so, suspends itself.  Method helpMaintainParallelism looks for
300	<	* suspended threads to resume before considering creating a new
301	<	* replacement. The spares themselves are encoded on another
302	<	* variant of a Treiber Stack, headed at field "spareWaiters".
303	<	* Note that the use of spares is intrinsically racy.  One thread
304	<	* may become a spare at about the same time as another is
305	<	* needlessly being created. We counteract this and related slop
306	<	* in part by requiring resumed spares to immediately recheck (in
307	<	* preStep) to see whether they they should re-suspend.  To avoid
308	<	* long-term build-up of spares, the oldest spare (see
309	<	* ForkJoinWorkerThread.suspendAsSpare) occasionally wakes up if
310	<	* not signalled and calls tryTrimSpare, which uses two different
311	<	* thresholds: Always killing if the number of spares is greater
312	<	* that 25% of total, and killing others only at a slower rate
313	<	* (UNUSED_SPARE_TRIM_RATE_NANOS).
314	<	*
315	<	* 6. Deciding when to create new workers. The main dynamic
316	<	* control in this class is deciding when to create extra threads
317	<	* in method helpMaintainParallelism. We would like to keep
318	<	* exactly #parallelism threads running, which is an impossble
319	<	* task. We always need to create one when the number of running
320	<	* threads would become zero and all workers are busy. Beyond
321	<	* this, we must rely on heuristics that work well in the the
322	<	* presence of transients phenomena such as GC stalls, dynamic
323	<	* compilation, and wake-up lags. These transients are extremely
324	<	* common -- we are normally trying to fully saturate the CPUs on
325	<	* a machine, so almost any activity other than running tasks
326	<	* impedes accuracy. Our main defense is to allow some slack in
327	<	* creation thresholds, using rules that reflect the fact that the
328	<	* more threads we have running, the more likely that we are
329	<	* underestimating the number running threads. The rules also
330	<	* better cope with the fact that some of the methods in this
331	<	* class tend to never become compiled (but are interpreted), so
332	<	* some components of the entire set of controls might execute 100
333	<	* times faster than others. And similarly for cases where the
334	<	* apparent lack of work is just due to GC stalls and other
335	<	* transient system activity.
336	<	*
337	<	* Beware that there is a lot of representation-level coupling
338	<	* among classes ForkJoinPool, ForkJoinWorkerThread, and
339	<	* ForkJoinTask.  For example, direct access to "workers" array by
340	<	* workers, and direct access to ForkJoinTask.status by both
341	<	* ForkJoinPool and ForkJoinWorkerThread.  There is little point
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 the required read orderings (which are sometimes
474	<	* critical). Also several occurrences of the unusual "do {}
475	<	* while(!cas...)" which is the simplest way to force an update of
476	<	* a CAS'ed variable. There are also other coding oddities that
477	<	* help some methods perform reasonably even when interpreted (not
478	<	* compiled), at the expense of some messy constructions that
479	<	* reduce byte code counts.
480	<	*
481	<	* The order of declarations in this file is: (1) statics (2)
482	<	* fields (along with constants used when unpacking some of them)
483	<	* (3) internal control methods (4) callbacks and other support
484	<	* for ForkJoinTask and ForkJoinWorkerThread classes, (5) exported
485	<	* 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 380 | Line 527 | public class ForkJoinPool extends Abstra
527		* Default ForkJoinWorkerThreadFactory implementation; creates a
528		* new ForkJoinWorkerThread.
529		*/
530	<	static class DefaultForkJoinWorkerThreadFactory
530	>	static final class DefaultForkJoinWorkerThreadFactory
531		implements ForkJoinWorkerThreadFactory {
532	<	public ForkJoinWorkerThread newThread(ForkJoinPool pool) {
532	>	public final ForkJoinWorkerThread newThread(ForkJoinPool pool) {
533		return new ForkJoinWorkerThread(pool);
534		}
535		}
536
537		/**
538	<	* Creates a new ForkJoinWorkerThread. This factory is used unless
539	<	* overridden in ForkJoinPool constructors.
538	>	* Per-thread records for threads that submit to pools. Currently
539	>	* holds only pseudo-random seed / index that is used to choose
540	>	* submission queues in method externalPush. In the future, this may
541	>	* also incorporate a means to implement different task rejection
542	>	* and resubmission policies.
543	>	*
544	>	* Seeds for submitters and workers/workQueues work in basically
545	>	* the same way but are initialized and updated using slightly
546	>	* different mechanics. Both are initialized using the same
547	>	* approach as in class ThreadLocal, where successive values are
548	>	* unlikely to collide with previous values. Seeds are then
549	>	* randomly modified upon collisions using xorshifts, which
550	>	* requires a non-zero seed.
551	>	*/
552	>	static final class Submitter {
553	>	int seed;
554	>	Submitter(int s) { seed = s; }
555	>	}
556	>
557	>	/**
558	>	* Class for artificial tasks that are used to replace the target
559	>	* of local joins if they are removed from an interior queue slot
560	>	* in WorkQueue.tryRemoveAndExec. We don't need the proxy to
561	>	* actually do anything beyond having a unique identity.
562	>	*/
563	>	static final class EmptyTask extends ForkJoinTask<Void> {
564	>	private static final long serialVersionUID = -7721805057305804111L;
565	>	EmptyTask() { status = ForkJoinTask.NORMAL; } // force done
566	>	public final Void getRawResult() { return null; }
567	>	public final void setRawResult(Void x) {}
568	>	public final boolean exec() { return true; }
569	>	}
570	>
571	>	/**
572	>	* Queues supporting work-stealing as well as external task
573	>	* submission. See above for main rationale and algorithms.
574	>	* Implementation relies heavily on "Unsafe" intrinsics
575	>	* and selective use of "volatile":
576	>	*
577	>	* Field "base" is the index (mod array.length) of the least valid
578	>	* queue slot, which is always the next position to steal (poll)
579	>	* from if nonempty. Reads and writes require volatile orderings
580	>	* but not CAS, because updates are only performed after slot
581	>	* CASes.
582	>	*
583	>	* Field "top" is the index (mod array.length) of the next queue
584	>	* slot to push to or pop from. It is written only by owner thread
585	>	* for push, or under lock for external/shared push, and accessed
586	>	* by other threads only after reading (volatile) base.  Both top
587	>	* and base are allowed to wrap around on overflow, but (top -
588	>	* base) (or more commonly -(base - top) to force volatile read of
589	>	* base before top) still estimates size. The lock ("qlock") is
590	>	* forced to -1 on termination, causing all further lock attempts
591	>	* to fail. (Note: we don't need CAS for termination state because
592	>	* upon pool shutdown, all shared-queues will stop being used
593	>	* anyway.)  Nearly all lock bodies are set up so that exceptions
594	>	* within lock bodies are "impossible" (modulo JVM errors that
595	>	* would cause failure anyway.)
596	>	*
597	>	* The array slots are read and written using the emulation of
598	>	* volatiles/atomics provided by Unsafe. Insertions must in
599	>	* general use putOrderedObject as a form of releasing store to
600	>	* ensure that all writes to the task object are ordered before
601	>	* its publication in the queue.  All removals entail a CAS to
602	>	* null.  The array is always a power of two. To ensure safety of
603	>	* Unsafe array operations, all accesses perform explicit null
604	>	* checks and implicit bounds checks via power-of-two masking.
605	>	*
606	>	* In addition to basic queuing support, this class contains
607	>	* fields described elsewhere to control execution. It turns out
608	>	* to work better memory-layout-wise to include them in this class
609	>	* rather than a separate class.
610	>	*
611	>	* Performance on most platforms is very sensitive to placement of
612	>	* instances of both WorkQueues and their arrays -- we absolutely
613	>	* do not want multiple WorkQueue instances or multiple queue
614	>	* arrays sharing cache lines. (It would be best for queue objects
615	>	* and their arrays to share, but there is nothing available to
616	>	* help arrange that).  Unfortunately, because they are recorded
617	>	* in a common array, WorkQueue instances are often moved to be
618	>	* adjacent by garbage collectors. To reduce impact, we use field
619	>	* padding that works OK on common platforms; this effectively
620	>	* trades off slightly slower average field access for the sake of
621	>	* avoiding really bad worst-case access. (Until better JVM
622	>	* support is in place, this padding is dependent on transient
623	>	* properties of JVM field layout rules.) We also take care in
624	>	* allocating, sizing and resizing the array. Non-shared queue
625	>	* arrays are initialized by workers before use. Others are
626	>	* allocated on first use.
627		*/
628	<	public static final ForkJoinWorkerThreadFactory
629	<	defaultForkJoinWorkerThreadFactory =
630	<	new DefaultForkJoinWorkerThreadFactory();
628	>	static final class WorkQueue {
629	>	/**
630	>	* Capacity of work-stealing queue array upon initialization.
631	>	* Must be a power of two; at least 4, but should be larger to
632	>	* reduce or eliminate cacheline sharing among queues.
633	>	* Currently, it is much larger, as a partial workaround for
634	>	* the fact that JVMs often place arrays in locations that
635	>	* share GC bookkeeping (especially cardmarks) such that
636	>	* per-write accesses encounter serious memory contention.
637	>	*/
638	>	static final int INITIAL_QUEUE_CAPACITY = 1 << 13;
639
640	<	/**
641	<	* Permission required for callers of methods that may start or
642	<	* kill threads.
643	<	*/
644	<	private static final RuntimePermission modifyThreadPermission =
645	<	new RuntimePermission("modifyThread");
640	>	/**
641	>	* Maximum size for queue arrays. Must be a power of two less
642	>	* than or equal to 1 << (31 - width of array entry) to ensure
643	>	* lack of wraparound of index calculations, but defined to a
644	>	* value a bit less than this to help users trap runaway
645	>	* programs before saturating systems.
646	>	*/
647	>	static final int MAXIMUM_QUEUE_CAPACITY = 1 << 26; // 64M
648
649	<	/**
650	<	* If there is a security manager, makes sure caller has
651	<	* permission to modify threads.
652	<	*/
653	<	private static void checkPermission() {
654	<	SecurityManager security = System.getSecurityManager();
655	<	if (security != null)
656	<	security.checkPermission(modifyThreadPermission);
649	>	// Heuristic padding to ameliorate unfortunate memory placements
650	>	volatile long pad00, pad01, pad02, pad03, pad04, pad05, pad06;
651	>
652	>	int seed;                  // for random scanning; initialize nonzero
653	>	volatile int eventCount;   // encoded inactivation count; < 0 if inactive
654	>	int nextWait;              // encoded record of next event waiter
655	>	int hint;                  // steal or signal hint (index)
656	>	int poolIndex;             // index of this queue in pool (or 0)
657	>	final int mode;            // 0: lifo, > 0: fifo, < 0: shared
658	>	int nsteals;               // number of steals
659	>	volatile int qlock;        // 1: locked, -1: terminate; else 0
660	>	volatile int base;         // index of next slot for poll
661	>	int top;                   // index of next slot for push
662	>	ForkJoinTask<?>[] array;   // the elements (initially unallocated)
663	>	final ForkJoinPool pool;   // the containing pool (may be null)
664	>	final ForkJoinWorkerThread owner; // owning thread or null if shared
665	>	volatile Thread parker;    // == owner during call to park; else null
666	>	volatile ForkJoinTask<?> currentJoin;  // task being joined in awaitJoin
667	>	ForkJoinTask<?> currentSteal; // current non-local task being executed
668	>
669	>	volatile Object pad10, pad11, pad12, pad13, pad14, pad15, pad16, pad17;
670	>	volatile Object pad18, pad19, pad1a, pad1b, pad1c, pad1d;
671	>
672	>	WorkQueue(ForkJoinPool pool, ForkJoinWorkerThread owner, int mode,
673	>	int seed) {
674	>	this.pool = pool;
675	>	this.owner = owner;
676	>	this.mode = mode;
677	>	this.seed = seed;
678	>	// Place indices in the center of array (that is not yet allocated)
679	>	base = top = INITIAL_QUEUE_CAPACITY >>> 1;
680	>	}
681	>
682	>	/**
683	>	* Returns the approximate number of tasks in the queue.
684	>	*/
685	>	final int queueSize() {
686	>	int n = base - top;       // non-owner callers must read base first
687	>	return (n >= 0) ? 0 : -n; // ignore transient negative
688	>	}
689	>
690	>	/**
691	>	* Provides a more accurate estimate of whether this queue has
692	>	* any tasks than does queueSize, by checking whether a
693	>	* near-empty queue has at least one unclaimed task.
694	>	*/
695	>	final boolean isEmpty() {
696	>	ForkJoinTask<?>[] a; int m, s;
697	>	int n = base - (s = top);
698	>	return (n >= 0 ||
699	>	(n == -1 &&
700	>	((a = array) == null ||
701	>	(m = a.length - 1) < 0 ||
702	>	U.getObject
703	>	(a, (long)((m & (s - 1)) << ASHIFT) + ABASE) == null)));
704	>	}
705	>
706	>	/**
707	>	* Pushes a task. Call only by owner in unshared queues.  (The
708	>	* shared-queue version is embedded in method externalPush.)
709	>	*
710	>	* @param task the task. Caller must ensure non-null.
711	>	* @throw RejectedExecutionException if array cannot be resized
712	>	*/
713	>	final void push(ForkJoinTask<?> task) {
714	>	ForkJoinTask<?>[] a; ForkJoinPool p;
715	>	int s = top, m, n;
716	>	if ((a = array) != null) {    // ignore if queue removed
717	>	int j = (((m = a.length - 1) & s) << ASHIFT) + ABASE;
718	>	U.putOrderedObject(a, j, task);
719	>	if ((n = (top = s + 1) - base) <= 2) {
720	>	if ((p = pool) != null)
721	>	p.signalWork(this);
722	>	}
723	>	else if (n >= m)
724	>	growArray();
725	>	}
726	>	}
727	>
728	>	/**
729	>	* Initializes or doubles the capacity of array. Call either
730	>	* by owner or with lock held -- it is OK for base, but not
731	>	* top, to move while resizings are in progress.
732	>	*/
733	>	final ForkJoinTask<?>[] growArray() {
734	>	ForkJoinTask<?>[] oldA = array;
735	>	int size = oldA != null ? oldA.length << 1 : INITIAL_QUEUE_CAPACITY;
736	>	if (size > MAXIMUM_QUEUE_CAPACITY)
737	>	throw new RejectedExecutionException("Queue capacity exceeded");
738	>	int oldMask, t, b;
739	>	ForkJoinTask<?>[] a = array = new ForkJoinTask<?>[size];
740	>	if (oldA != null && (oldMask = oldA.length - 1) >= 0 &&
741	>	(t = top) - (b = base) > 0) {
742	>	int mask = size - 1;
743	>	do {
744	>	ForkJoinTask<?> x;
745	>	int oldj = ((b & oldMask) << ASHIFT) + ABASE;
746	>	int j    = ((b &    mask) << ASHIFT) + ABASE;
747	>	x = (ForkJoinTask<?>)U.getObjectVolatile(oldA, oldj);
748	>	if (x != null &&
749	>	U.compareAndSwapObject(oldA, oldj, x, null))
750	>	U.putObjectVolatile(a, j, x);
751	>	} while (++b != t);
752	>	}
753	>	return a;
754	>	}
755	>
756	>	/**
757	>	* Takes next task, if one exists, in LIFO order.  Call only
758	>	* by owner in unshared queues.
759	>	*/
760	>	final ForkJoinTask<?> pop() {
761	>	ForkJoinTask<?>[] a; ForkJoinTask<?> t; int m;
762	>	if ((a = array) != null && (m = a.length - 1) >= 0) {
763	>	for (int s; (s = top - 1) - base >= 0;) {
764	>	long j = ((m & s) << ASHIFT) + ABASE;
765	>	if ((t = (ForkJoinTask<?>)U.getObject(a, j)) == null)
766	>	break;
767	>	if (U.compareAndSwapObject(a, j, t, null)) {
768	>	top = s;
769	>	return t;
770	>	}
771	>	}
772	>	}
773	>	return null;
774	>	}
775	>
776	>	/**
777	>	* Takes a task in FIFO order if b is base of queue and a task
778	>	* can be claimed without contention. Specialized versions
779	>	* appear in ForkJoinPool methods scan and tryHelpStealer.
780	>	*/
781	>	final ForkJoinTask<?> pollAt(int b) {
782	>	ForkJoinTask<?> t; ForkJoinTask<?>[] a;
783	>	if ((a = array) != null) {
784	>	int j = (((a.length - 1) & b) << ASHIFT) + ABASE;
785	>	if ((t = (ForkJoinTask<?>)U.getObjectVolatile(a, j)) != null &&
786	>	base == b &&
787	>	U.compareAndSwapObject(a, j, t, null)) {
788	>	base = b + 1;
789	>	return t;
790	>	}
791	>	}
792	>	return null;
793	>	}
794	>
795	>	/**
796	>	* Takes next task, if one exists, in FIFO order.
797	>	*/
798	>	final ForkJoinTask<?> poll() {
799	>	ForkJoinTask<?>[] a; int b; ForkJoinTask<?> t;
800	>	while ((b = base) - top < 0 && (a = array) != null) {
801	>	int j = (((a.length - 1) & b) << ASHIFT) + ABASE;
802	>	t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
803	>	if (t != null) {
804	>	if (base == b &&
805	>	U.compareAndSwapObject(a, j, t, null)) {
806	>	base = b + 1;
807	>	return t;
808	>	}
809	>	}
810	>	else if (base == b) {
811	>	if (b + 1 == top)
812	>	break;
813	>	Thread.yield(); // wait for lagging update (very rare)
814	>	}
815	>	}
816	>	return null;
817	>	}
818	>
819	>	/**
820	>	* Takes next task, if one exists, in order specified by mode.
821	>	*/
822	>	final ForkJoinTask<?> nextLocalTask() {
823	>	return mode == 0 ? pop() : poll();
824	>	}
825	>
826	>	/**
827	>	* Returns next task, if one exists, in order specified by mode.
828	>	*/
829	>	final ForkJoinTask<?> peek() {
830	>	ForkJoinTask<?>[] a = array; int m;
831	>	if (a == null || (m = a.length - 1) < 0)
832	>	return null;
833	>	int i = mode == 0 ? top - 1 : base;
834	>	int j = ((i & m) << ASHIFT) + ABASE;
835	>	return (ForkJoinTask<?>)U.getObjectVolatile(a, j);
836	>	}
837	>
838	>	/**
839	>	* Pops the given task only if it is at the current top.
840	>	* (A shared version is available only via FJP.tryExternalUnpush)
841	>	*/
842	>	final boolean tryUnpush(ForkJoinTask<?> t) {
843	>	ForkJoinTask<?>[] a; int s;
844	>	if ((a = array) != null && (s = top) != base &&
845	>	U.compareAndSwapObject
846	>	(a, (((a.length - 1) & --s) << ASHIFT) + ABASE, t, null)) {
847	>	top = s;
848	>	return true;
849	>	}
850	>	return false;
851	>	}
852	>
853	>	/**
854	>	* Removes and cancels all known tasks, ignoring any exceptions.
855	>	*/
856	>	final void cancelAll() {
857	>	ForkJoinTask.cancelIgnoringExceptions(currentJoin);
858	>	ForkJoinTask.cancelIgnoringExceptions(currentSteal);
859	>	for (ForkJoinTask<?> t; (t = poll()) != null; )
860	>	ForkJoinTask.cancelIgnoringExceptions(t);
861	>	}
862	>
863	>	/**
864	>	* Computes next value for random probes.  Scans don't require
865	>	* a very high quality generator, but also not a crummy one.
866	>	* Marsaglia xor-shift is cheap and works well enough.  Note:
867	>	* This is manually inlined in its usages in ForkJoinPool to
868	>	* avoid writes inside busy scan loops.
869	>	*/
870	>	final int nextSeed() {
871	>	int r = seed;
872	>	r ^= r << 13;
873	>	r ^= r >>> 17;
874	>	return seed = r ^= r << 5;
875	>	}
876	>
877	>	// Specialized execution methods
878	>
879	>	/**
880	>	* Pops and runs tasks until empty.
881	>	*/
882	>	private void popAndExecAll() {
883	>	// A bit faster than repeated pop calls
884	>	ForkJoinTask<?>[] a; int m, s; long j; ForkJoinTask<?> t;
885	>	while ((a = array) != null && (m = a.length - 1) >= 0 &&
886	>	(s = top - 1) - base >= 0 &&
887	>	(t = ((ForkJoinTask<?>)
888	>	U.getObject(a, j = ((m & s) << ASHIFT) + ABASE)))
889	>	!= null) {
890	>	if (U.compareAndSwapObject(a, j, t, null)) {
891	>	top = s;
892	>	t.doExec();
893	>	}
894	>	}
895	>	}
896	>
897	>	/**
898	>	* Polls and runs tasks until empty.
899	>	*/
900	>	private void pollAndExecAll() {
901	>	for (ForkJoinTask<?> t; (t = poll()) != null;)
902	>	t.doExec();
903	>	}
904	>
905	>	/**
906	>	* If present, removes from queue and executes the given task,
907	>	* or any other cancelled task. Returns (true) on any CAS
908	>	* or consistency check failure so caller can retry.
909	>	*
910	>	* @return false if no progress can be made, else true;
911	>	*/
912	>	final boolean tryRemoveAndExec(ForkJoinTask<?> task) {
913	>	boolean stat = true, removed = false, empty = true;
914	>	ForkJoinTask<?>[] a; int m, s, b, n;
915	>	if ((a = array) != null && (m = a.length - 1) >= 0 &&
916	>	(n = (s = top) - (b = base)) > 0) {
917	>	for (ForkJoinTask<?> t;;) {           // traverse from s to b
918	>	int j = ((--s & m) << ASHIFT) + ABASE;
919	>	t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
920	>	if (t == null)                    // inconsistent length
921	>	break;
922	>	else if (t == task) {
923	>	if (s + 1 == top) {           // pop
924	>	if (!U.compareAndSwapObject(a, j, task, null))
925	>	break;
926	>	top = s;
927	>	removed = true;
928	>	}
929	>	else if (base == b)           // replace with proxy
930	>	removed = U.compareAndSwapObject(a, j, task,
931	>	new EmptyTask());
932	>	break;
933	>	}
934	>	else if (t.status >= 0)
935	>	empty = false;
936	>	else if (s + 1 == top) {          // pop and throw away
937	>	if (U.compareAndSwapObject(a, j, t, null))
938	>	top = s;
939	>	break;
940	>	}
941	>	if (--n == 0) {
942	>	if (!empty && base == b)
943	>	stat = false;
944	>	break;
945	>	}
946	>	}
947	>	}
948	>	if (removed)
949	>	task.doExec();
950	>	return stat;
951	>	}
952	>
953	>	/**
954	>	* Polls for and executes the given task or any other task in
955	>	* its CountedCompleter computation
956	>	*/
957	>	final boolean pollAndExecCC(ForkJoinTask<?> root) {
958	>	ForkJoinTask<?>[] a; int b; Object o;
959	>	outer: while ((b = base) - top < 0 && (a = array) != null) {
960	>	long j = (((a.length - 1) & b) << ASHIFT) + ABASE;
961	>	if ((o = U.getObject(a, j)) == null ||
962	>	!(o instanceof CountedCompleter))
963	>	break;
964	>	for (CountedCompleter<?> t = (CountedCompleter<?>)o, r = t;;) {
965	>	if (r == root) {
966	>	if (base == b &&
967	>	U.compareAndSwapObject(a, j, t, null)) {
968	>	base = b + 1;
969	>	t.doExec();
970	>	return true;
971	>	}
972	>	else
973	>	break; // restart
974	>	}
975	>	if ((r = r.completer) == null)
976	>	break outer; // not part of root computation
977	>	}
978	>	}
979	>	return false;
980	>	}
981	>
982	>	/**
983	>	* Executes a top-level task and any local tasks remaining
984	>	* after execution.
985	>	*/
986	>	final void runTask(ForkJoinTask<?> t) {
987	>	if (t != null) {
988	>	(currentSteal = t).doExec();
989	>	currentSteal = null;
990	>	if (base - top < 0) {       // process remaining local tasks
991	>	if (mode == 0)
992	>	popAndExecAll();
993	>	else
994	>	pollAndExecAll();
995	>	}
996	>	++nsteals;
997	>	hint = -1;
998	>	}
999	>	}
1000	>
1001	>	/**
1002	>	* Executes a non-top-level (stolen) task.
1003	>	*/
1004	>	final void runSubtask(ForkJoinTask<?> t) {
1005	>	if (t != null) {
1006	>	ForkJoinTask<?> ps = currentSteal;
1007	>	(currentSteal = t).doExec();
1008	>	currentSteal = ps;
1009	>	}
1010	>	}
1011	>
1012	>	/**
1013	>	* Returns true if owned and not known to be blocked.
1014	>	*/
1015	>	final boolean isApparentlyUnblocked() {
1016	>	Thread wt; Thread.State s;
1017	>	return (eventCount >= 0 &&
1018	>	(wt = owner) != null &&
1019	>	(s = wt.getState()) != Thread.State.BLOCKED &&
1020	>	s != Thread.State.WAITING &&
1021	>	s != Thread.State.TIMED_WAITING);
1022	>	}
1023	>
1024	>	/**
1025	>	* If this owned and is not already interrupted, try to
1026	>	* interrupt and/or unpark, ignoring exceptions.
1027	>	*/
1028	>	final void interruptOwner() {
1029	>	Thread wt, p;
1030	>	if ((wt = owner) != null && !wt.isInterrupted()) {
1031	>	try {
1032	>	wt.interrupt();
1033	>	} catch (SecurityException ignore) {
1034	>	}
1035	>	}
1036	>	if ((p = parker) != null)
1037	>	U.unpark(p);
1038	>	}
1039	>
1040	>	// Unsafe mechanics
1041	>	private static final sun.misc.Unsafe U;
1042	>	private static final long QLOCK;
1043	>	private static final int ABASE;
1044	>	private static final int ASHIFT;
1045	>	static {
1046	>	int s;
1047	>	try {
1048	>	U = getUnsafe();
1049	>	Class<?> k = WorkQueue.class;
1050	>	Class<?> ak = ForkJoinTask[].class;
1051	>	QLOCK = U.objectFieldOffset
1052	>	(k.getDeclaredField("qlock"));
1053	>	ABASE = U.arrayBaseOffset(ak);
1054	>	s = U.arrayIndexScale(ak);
1055	>	} catch (Exception e) {
1056	>	throw new Error(e);
1057	>	}
1058	>	if ((s & (s-1)) != 0)
1059	>	throw new Error("data type scale not a power of two");
1060	>	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
1061	>	}
1062		}
1063
1064	<	/**
416	<	* Generator for assigning sequence numbers as pool names.
417	<	*/
418	<	private static final AtomicInteger poolNumberGenerator =
419	<	new AtomicInteger();
1064	>	// static fields (initialized in static initializer below)
1065
1066		/**
1067	<	* Absolute bound for parallelism level. Twice this number plus
1068	<	* one (i.e., 0xfff) must fit into a 16bit field to enable
424	<	* word-packing for some counts and indices.
1067	>	* Creates a new ForkJoinWorkerThread. This factory is used unless
1068	>	* overridden in ForkJoinPool constructors.
1069		*/
1070	<	private static final int MAX_WORKERS   = 0x7fff;
1070	>	public static final ForkJoinWorkerThreadFactory
1071	>	defaultForkJoinWorkerThreadFactory;
1072
1073		/**
1074	<	* Array holding all worker threads in the pool.  Array size must
1075	<	* be a power of two.  Updates and replacements are protected by
1076	<	* workerLock, but the array is always kept in a consistent enough
1077	<	* state to be randomly accessed without locking by workers
1078	<	* performing work-stealing, as well as other traversal-based
434	<	* methods in this class. All readers must tolerate that some
435	<	* array slots may be null.
1074	>	* Per-thread submission bookkeeping. Shared across all pools
1075	>	* to reduce ThreadLocal pollution and because random motion
1076	>	* to avoid contention in one pool is likely to hold for others.
1077	>	* Lazily initialized on first submission (but null-checked
1078	>	* in other contexts to avoid unnecessary initialization).
1079		*/
1080	<	volatile ForkJoinWorkerThread[] workers;
1080	>	static final ThreadLocal<Submitter> submitters;
1081
1082		/**
1083	<	* Queue for external submissions.
1083	>	* Permission required for callers of methods that may start or
1084	>	* kill threads.
1085		*/
1086	<	private final LinkedTransferQueue<ForkJoinTask<?>> submissionQueue;
1086	>	private static final RuntimePermission modifyThreadPermission;
1087
1088		/**
1089	<	* Lock protecting updates to workers array.
1089	>	* Common (static) pool. Non-null for public use unless a static
1090	>	* construction exception, but internal usages null-check on use
1091	>	* to paranoically avoid potential initialization circularities
1092	>	* as well as to simplify generated code.
1093		*/
1094	<	private final ReentrantLock workerLock;
1094	>	static final ForkJoinPool commonPool;
1095
1096		/**
1097	<	* Latch released upon termination.
1097	>	* Common pool parallelism. Must equal commonPool.parallelism.
1098		*/
1099	<	private final Phaser termination;
1099	>	static final int commonPoolParallelism;
1100
1101		/**
1102	<	* Creation factory for worker threads.
1102	>	* Sequence number for creating workerNamePrefix.
1103		*/
1104	<	private final ForkJoinWorkerThreadFactory factory;
1104	>	private static int poolNumberSequence;
1105
1106		/**
1107	<	* Sum of per-thread steal counts, updated only when threads are
1108	<	* idle or terminating.
1107	>	* Return the next sequence number. We don't expect this to
1108	>	* ever contend so use simple builtin sync.
1109		*/
1110	<	private volatile long stealCount;
1110	>	private static final synchronized int nextPoolId() {
1111	>	return ++poolNumberSequence;
1112	>	}
1113
1114	<	/**
466	<	* The last nanoTime that a spare thread was trimmed
467	<	*/
468	<	private volatile long trimTime;
1114	>	// static constants
1115
1116		/**
1117	<	* The rate at which to trim unused spares
1117	>	* Initial timeout value (in nanoseconds) for the thread
1118	>	* triggering quiescence to park waiting for new work. On timeout,
1119	>	* the thread will instead try to shrink the number of
1120	>	* workers. The value should be large enough to avoid overly
1121	>	* aggressive shrinkage during most transient stalls (long GCs
1122	>	* etc).
1123		*/
1124	<	static final long UNUSED_SPARE_TRIM_RATE_NANOS =
474	<	1000L * 1000L * 1000L; // 1 sec
1124	>	private static final long IDLE_TIMEOUT      = 2000L * 1000L * 1000L; // 2sec
1125
1126		/**
1127	<	* Encoded record of top of treiber stack of threads waiting for
478	<	* events. The top 32 bits contain the count being waited for. The
479	<	* bottom 16 bits contains one plus the pool index of waiting
480	<	* worker thread. (Bits 16-31 are unused.)
1127	>	* Timeout value when there are more threads than parallelism level
1128		*/
1129	<	private volatile long eventWaiters;
483	<
484	<	private static final int  EVENT_COUNT_SHIFT = 32;
485	<	private static final long WAITER_ID_MASK    = (1L << 16) - 1L;
1129	>	private static final long FAST_IDLE_TIMEOUT =  200L * 1000L * 1000L;
1130
1131		/**
1132	<	* A counter for events that may wake up worker threads:
1133	<	*   - Submission of a new task to the pool
1134	<	*   - A worker pushing a task on an empty queue
1135	<	*   - termination
1132	>	* The maximum stolen->joining link depth allowed in method
1133	>	* tryHelpStealer.  Must be a power of two.  Depths for legitimate
1134	>	* chains are unbounded, but we use a fixed constant to avoid
1135	>	* (otherwise unchecked) cycles and to bound staleness of
1136	>	* traversal parameters at the expense of sometimes blocking when
1137	>	* we could be helping.
1138		*/
1139	<	private volatile int eventCount;
1139	>	private static final int MAX_HELP = 64;
1140
1141		/**
1142	<	* Encoded record of top of treiber stack of spare threads waiting
1143	<	* for resumption. The top 16 bits contain an arbitrary count to
498	<	* avoid ABA effects. The bottom 16bits contains one plus the pool
499	<	* index of waiting worker thread.
1142	>	* Increment for seed generators. See class ThreadLocal for
1143	>	* explanation.
1144		*/
1145	<	private volatile int spareWaiters;
502	<
503	<	private static final int SPARE_COUNT_SHIFT = 16;
504	<	private static final int SPARE_ID_MASK     = (1 << 16) - 1;
1145	>	private static final int SEED_INCREMENT = 0x61c88647;
1146
1147		/**
1148	<	* Lifecycle control. The low word contains the number of workers
1149	<	* that are (probably) executing tasks. This value is atomically
1150	<	* incremented before a worker gets a task to run, and decremented
1151	<	* when worker has no tasks and cannot find any.  Bits 16-18
1152	<	* contain runLevel value. When all are zero, the pool is
1153	<	* running. Level transitions are monotonic (running -> shutdown
1154	<	* -> terminating -> terminated) so each transition adds a bit.
1155	<	* These are bundled together to ensure consistent read for
1156	<	* termination checks (i.e., that runLevel is at least SHUTDOWN
1157	<	* and active threads is zero).
1148	>	* Bits and masks for control variables
1149	>	*
1150	>	* Field ctl is a long packed with:
1151	>	* AC: Number of active running workers minus target parallelism (16 bits)
1152	>	* TC: Number of total workers minus target parallelism (16 bits)
1153	>	* ST: true if pool is terminating (1 bit)
1154	>	* EC: the wait count of top waiting thread (15 bits)
1155	>	* ID: poolIndex of top of Treiber stack of waiters (16 bits)
1156	>	*
1157	>	* When convenient, we can extract the upper 32 bits of counts and
1158	>	* the lower 32 bits of queue state, u = (int)(ctl >>> 32) and e =
1159	>	* (int)ctl.  The ec field is never accessed alone, but always
1160	>	* together with id and st. The offsets of counts by the target
1161	>	* parallelism and the positionings of fields makes it possible to
1162	>	* perform the most common checks via sign tests of fields: When
1163	>	* ac is negative, there are not enough active workers, when tc is
1164	>	* negative, there are not enough total workers, and when e is
1165	>	* negative, the pool is terminating.  To deal with these possibly
1166	>	* negative fields, we use casts in and out of "short" and/or
1167	>	* signed shifts to maintain signedness.
1168	>	*
1169	>	* When a thread is queued (inactivated), its eventCount field is
1170	>	* set negative, which is the only way to tell if a worker is
1171	>	* prevented from executing tasks, even though it must continue to
1172	>	* scan for them to avoid queuing races. Note however that
1173	>	* eventCount updates lag releases so usage requires care.
1174	>	*
1175	>	* Field plock is an int packed with:
1176	>	* SHUTDOWN: true if shutdown is enabled (1 bit)
1177	>	* SEQ:  a sequence lock, with PL_LOCK bit set if locked (30 bits)
1178	>	* SIGNAL: set when threads may be waiting on the lock (1 bit)
1179	>	*
1180	>	* The sequence number enables simple consistency checks:
1181	>	* Staleness of read-only operations on the workQueues array can
1182	>	* be checked by comparing plock before vs after the reads.
1183		*/
518	–	private volatile int runState;
1184
1185	<	// Note: The order among run level values matters.
1186	<	private static final int RUNLEVEL_SHIFT     = 16;
1187	<	private static final int SHUTDOWN           = 1 << RUNLEVEL_SHIFT;
1188	<	private static final int TERMINATING        = 1 << (RUNLEVEL_SHIFT + 1);
1189	<	private static final int TERMINATED         = 1 << (RUNLEVEL_SHIFT + 2);
525	<	private static final int ACTIVE_COUNT_MASK  = (1 << RUNLEVEL_SHIFT) - 1;
526	<	private static final int ONE_ACTIVE         = 1; // active update delta
1185	>	// bit positions/shifts for fields
1186	>	private static final int  AC_SHIFT   = 48;
1187	>	private static final int  TC_SHIFT   = 32;
1188	>	private static final int  ST_SHIFT   = 31;
1189	>	private static final int  EC_SHIFT   = 16;
1190
1191	<	/**
1192	<	* Holds number of total (i.e., created and not yet terminated)
1193	<	* and running (i.e., not blocked on joins or other managed sync)
1194	<	* threads, packed together to ensure consistent snapshot when
1195	<	* making decisions about creating and suspending spare
1196	<	* threads. Updated only by CAS. Note that adding a new worker
1197	<	* requires incrementing both counts, since workers start off in
535	<	* running state.
536	<	*/
537	<	private volatile int workerCounts;
1191	>	// bounds
1192	>	private static final int  SMASK      = 0xffff;  // short bits
1193	>	private static final int  MAX_CAP    = 0x7fff;  // max #workers - 1
1194	>	private static final int  EVENMASK   = 0xfffe;  // even short bits
1195	>	private static final int  SQMASK     = 0x007e;  // max 64 (even) slots
1196	>	private static final int  SHORT_SIGN = 1 << 15;
1197	>	private static final int  INT_SIGN   = 1 << 31;
1198
1199	<	private static final int TOTAL_COUNT_SHIFT  = 16;
1200	<	private static final int RUNNING_COUNT_MASK = (1 << TOTAL_COUNT_SHIFT) - 1;
1201	<	private static final int ONE_RUNNING        = 1;
1202	<	private static final int ONE_TOTAL          = 1 << TOTAL_COUNT_SHIFT;
1199	>	// masks
1200	>	private static final long STOP_BIT   = 0x0001L << ST_SHIFT;
1201	>	private static final long AC_MASK    = ((long)SMASK) << AC_SHIFT;
1202	>	private static final long TC_MASK    = ((long)SMASK) << TC_SHIFT;
1203
1204	<	/**
1205	<	* The target parallelism level.
1206	<	* Accessed directly by ForkJoinWorkerThreads.
547	<	*/
548	<	final int parallelism;
1204	>	// units for incrementing and decrementing
1205	>	private static final long TC_UNIT    = 1L << TC_SHIFT;
1206	>	private static final long AC_UNIT    = 1L << AC_SHIFT;
1207
1208	<	/**
1209	<	* True if use local fifo, not default lifo, for local polling
1210	<	* Read by, and replicated by ForkJoinWorkerThreads
1211	<	*/
1212	<	final boolean locallyFifo;
1208	>	// masks and units for dealing with u = (int)(ctl >>> 32)
1209	>	private static final int  UAC_SHIFT  = AC_SHIFT - 32;
1210	>	private static final int  UTC_SHIFT  = TC_SHIFT - 32;
1211	>	private static final int  UAC_MASK   = SMASK << UAC_SHIFT;
1212	>	private static final int  UTC_MASK   = SMASK << UTC_SHIFT;
1213	>	private static final int  UAC_UNIT   = 1 << UAC_SHIFT;
1214	>	private static final int  UTC_UNIT   = 1 << UTC_SHIFT;
1215
1216	<	/**
1217	<	* The uncaught exception handler used when any worker abruptly
1218	<	* terminates.
559	<	*/
560	<	private final Thread.UncaughtExceptionHandler ueh;
1216	>	// masks and units for dealing with e = (int)ctl
1217	>	private static final int E_MASK      = 0x7fffffff; // no STOP_BIT
1218	>	private static final int E_SEQ       = 1 << EC_SHIFT;
1219
1220	<	/**
1221	<	* Pool number, just for assigning useful names to worker threads
1222	<	*/
1223	<	private final int poolNumber;
1220	>	// plock bits
1221	>	private static final int SHUTDOWN    = 1 << 31;
1222	>	private static final int PL_LOCK     = 2;
1223	>	private static final int PL_SIGNAL   = 1;
1224	>	private static final int PL_SPINS    = 1 << 8;
1225
1226	+	// access mode for WorkQueue
1227	+	static final int LIFO_QUEUE          =  0;
1228	+	static final int FIFO_QUEUE          =  1;
1229	+	static final int SHARED_QUEUE        = -1;
1230
1231	<	// Utilities for CASing fields. Note that several of these
1232	<	// are manually inlined by callers
1231	>	// bounds for #steps in scan loop -- must be power 2 minus 1
1232	>	private static final int MIN_SCAN    = 0x1ff;   // cover estimation slop
1233	>	private static final int MAX_SCAN    = 0x1ffff; // 4 * max workers
1234
1235	<	/**
572	<	* Increments running count part of workerCounts
573	<	*/
574	<	final void incrementRunningCount() {
575	<	int c;
576	<	do {} while (!UNSAFE.compareAndSwapInt(this, workerCountsOffset,
577	<	c = workerCounts,
578	<	c + ONE_RUNNING));
579	<	}
1235	>	// Instance fields
1236
1237	<	/**
1238	<	* Tries to decrement running count unless already zero
1239	<	*/
1240	<	final boolean tryDecrementRunningCount() {
1241	<	int wc = workerCounts;
1242	<	if ((wc & RUNNING_COUNT_MASK) == 0)
1243	<	return false;
1244	<	return UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1245	<	wc, wc - ONE_RUNNING);
1246	<	}
1237	>	/*
1238	>	* Field layout of this class tends to matter more than one would
1239	>	* like. Runtime layout order is only loosely related to
1240	>	* declaration order and may differ across JVMs, but the following
1241	>	* empirically works OK on current JVMs.
1242	>	*/
1243	>
1244	>	// Heuristic padding to ameliorate unfortunate memory placements
1245	>	volatile long pad00, pad01, pad02, pad03, pad04, pad05, pad06;
1246	>
1247	>	volatile long stealCount;                  // collects worker counts
1248	>	volatile long ctl;                         // main pool control
1249	>	volatile int plock;                        // shutdown status and seqLock
1250	>	volatile int indexSeed;                    // worker/submitter index seed
1251	>	final int config;                          // mode and parallelism level
1252	>	WorkQueue[] workQueues;                    // main registry
1253	>	final ForkJoinWorkerThreadFactory factory;
1254	>	final Thread.UncaughtExceptionHandler ueh; // per-worker UEH
1255	>	final String workerNamePrefix;             // to create worker name string
1256
1257	<	/**
1258	<	* Forces decrement of encoded workerCounts, awaiting nonzero if
1259	<	* (rarely) necessary when other count updates lag.
1260	<	*
1261	<	* @param dr -- either zero or ONE_RUNNING
1262	<	* @param dt == either zero or ONE_TOTAL
1257	>	volatile Object pad10, pad11, pad12, pad13, pad14, pad15, pad16, pad17;
1258	>	volatile Object pad18, pad19, pad1a, pad1b;
1259	>
1260	>	/*
1261	>	* Acquires the plock lock to protect worker array and related
1262	>	* updates. This method is called only if an initial CAS on plock
1263	>	* fails. This acts as a spinLock for normal cases, but falls back
1264	>	* to builtin monitor to block when (rarely) needed. This would be
1265	>	* a terrible idea for a highly contended lock, but works fine as
1266	>	* a more conservative alternative to a pure spinlock.  See
1267	>	* internal ConcurrentHashMap documentation for further
1268	>	* explanation of nearly the same construction.
1269		*/
1270	<	private void decrementWorkerCounts(int dr, int dt) {
1270	>	private int acquirePlock() {
1271	>	int spins = PL_SPINS, r = 0, ps, nps;
1272		for (;;) {
1273	<	int wc = workerCounts;
1274	<	if (wc == 0 && (runState & TERMINATED) != 0)
1275	<	return; // lagging termination on a backout
1276	<	if ((wc & RUNNING_COUNT_MASK)  - dr < 0 ||
1277	<	(wc >>> TOTAL_COUNT_SHIFT) - dt < 0)
1278	<	Thread.yield();
1279	<	if (UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1280	<	wc, wc - (dr + dt)))
1281	<	return;
1273	>	if (((ps = plock) & PL_LOCK) == 0 &&
1274	>	U.compareAndSwapInt(this, PLOCK, ps, nps = ps + PL_LOCK))
1275	>	return nps;
1276	>	else if (r == 0) { // randomize spins if possible
1277	>	Thread t = Thread.currentThread(); WorkQueue w; Submitter z;
1278	>	if ((t instanceof ForkJoinWorkerThread) &&
1279	>	(w = ((ForkJoinWorkerThread)t).workQueue) != null)
1280	>	r = w.seed;
1281	>	else if ((z = submitters.get()) != null)
1282	>	r = z.seed;
1283	>	else
1284	>	r = 1;
1285	>	}
1286	>	else if (spins >= 0) {
1287	>	r ^= r << 1; r ^= r >>> 3; r ^= r << 10; // xorshift
1288	>	if (r >= 0)
1289	>	--spins;
1290	>	}
1291	>	else if (U.compareAndSwapInt(this, PLOCK, ps, ps | PL_SIGNAL)) {
1292	>	synchronized (this) {
1293	>	if ((plock & PL_SIGNAL) != 0) {
1294	>	try {
1295	>	wait();
1296	>	} catch (InterruptedException ie) {
1297	>	try {
1298	>	Thread.currentThread().interrupt();
1299	>	} catch (SecurityException ignore) {
1300	>	}
1301	>	}
1302	>	}
1303	>	else
1304	>	notifyAll();
1305	>	}
1306	>	}
1307		}
1308		}
1309
1310		/**
1311	<	* Increments event count
1311	>	* Unlocks and signals any thread waiting for plock. Called only
1312	>	* when CAS of seq value for unlock fails.
1313		*/
1314	<	private void advanceEventCount() {
1315	<	int c;
1316	<	do {} while(!UNSAFE.compareAndSwapInt(this, eventCountOffset,
619	<	c = eventCount, c+1));
1314	>	private void releasePlock(int ps) {
1315	>	plock = ps;
1316	>	synchronized (this) { notifyAll(); }
1317		}
1318
1319		/**
1320	<	* Tries incrementing active count; fails on contention.
1321	<	* Called by workers before executing tasks.
1322	<	*
1323	<	* @return true on success
1324	<	*/
1325	<	final boolean tryIncrementActiveCount() {
1326	<	int c;
1327	<	return UNSAFE.compareAndSwapInt(this, runStateOffset,
1328	<	c = runState, c + ONE_ACTIVE);
1329	<	}
1330	<
1331	<	/**
1332	<	* Tries decrementing active count; fails on contention.
1333	<	* Called when workers cannot find tasks to run.
1334	<	*/
1335	<	final boolean tryDecrementActiveCount() {
1336	<	int c;
1337	<	return UNSAFE.compareAndSwapInt(this, runStateOffset,
1338	<	c = runState, c - ONE_ACTIVE);
1320	>	* Performs secondary initialization, called when plock is zero.
1321	>	* Creates workQueue array and sets plock to a valid value.  The
1322	>	* lock body must be exception-free (so no try/finally) so we
1323	>	* optimistically allocate new array outside the lock and throw
1324	>	* away if (very rarely) not needed. (A similar tactic is used in
1325	>	* fullExternalPush.)  Because the plock seq value can eventually
1326	>	* wrap around zero, this method harmlessly fails to reinitialize
1327	>	* if workQueues exists, while still advancing plock.
1328	>	*
1329	>	* Additonally tries to create the first worker.
1330	>	*/
1331	>	private void initWorkers() {
1332	>	WorkQueue[] ws, nws; int ps;
1333	>	int p = config & SMASK;        // find power of two table size
1334	>	int n = (p > 1) ? p - 1 : 1;   // ensure at least 2 slots
1335	>	n |= n >>> 1; n |= n >>> 2; n |= n >>> 4; n |= n >>> 8; n |= n >>> 16;
1336	>	n = (n + 1) << 1;
1337	>	if ((ws = workQueues) == null || ws.length == 0)
1338	>	nws = new WorkQueue[n];
1339	>	else
1340	>	nws = null;
1341	>	if (((ps = plock) & PL_LOCK) != 0 ||
1342	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1343	>	ps = acquirePlock();
1344	>	if (((ws = workQueues) == null || ws.length == 0) && nws != null)
1345	>	workQueues = nws;
1346	>	int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1347	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1348	>	releasePlock(nps);
1349	>	tryAddWorker();
1350		}
1351
1352		/**
1353	<	* Advances to at least the given level. Returns true if not
1354	<	* already in at least the given level.
1353	>	* Tries to create and start one worker. Adjusts counts etc on
1354	>	* failure.
1355		*/
1356	<	private boolean advanceRunLevel(int level) {
1357	<	for (;;) {
1358	<	int s = runState;
1359	<	if ((s & level) != 0)
1360	<	return false;
1361	<	if (UNSAFE.compareAndSwapInt(this, runStateOffset, s, s | level))
1362	<	return true;
1356	>	private void tryAddWorker() {
1357	>	long c; int u;
1358	>	while ((u = (int)((c = ctl) >>> 32)) < 0 &&
1359	>	(u & SHORT_SIGN) != 0 && (int)c == 0) {
1360	>	long nc = (long)(((u + UTC_UNIT) & UTC_MASK) |
1361	>	((u + UAC_UNIT) & UAC_MASK)) << 32;
1362	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1363	>	ForkJoinWorkerThreadFactory fac;
1364	>	Throwable ex = null;
1365	>	ForkJoinWorkerThread wt = null;
1366	>	try {
1367	>	if ((fac = factory) != null &&
1368	>	(wt = fac.newThread(this)) != null) {
1369	>	wt.start();
1370	>	break;
1371	>	}
1372	>	} catch (Throwable e) {
1373	>	ex = e;
1374	>	}
1375	>	deregisterWorker(wt, ex);
1376	>	break;
1377	>	}
1378		}
1379		}
1380
1381	<	// workers array maintenance
1381	>	//  Registering and deregistering workers
1382
1383		/**
1384	<	* Records and returns a workers array index for new worker.
1385	<	*/
1386	<	private int recordWorker(ForkJoinWorkerThread w) {
1387	<	// Try using slot totalCount-1. If not available, scan and/or resize
1388	<	int k = (workerCounts >>> TOTAL_COUNT_SHIFT) - 1;
1389	<	final ReentrantLock lock = this.workerLock;
1390	<	lock.lock();
1384	>	* Callback from ForkJoinWorkerThread to establish and record its
1385	>	* WorkQueue. To avoid scanning bias due to packing entries in
1386	>	* front of the workQueues array, we treat the array as a simple
1387	>	* power-of-two hash table using per-thread seed as hash,
1388	>	* expanding as needed.
1389	>	*
1390	>	* @param wt the worker thread
1391	>	* @return the worker's queue
1392	>	*/
1393	>	final WorkQueue registerWorker(ForkJoinWorkerThread wt) {
1394	>	Thread.UncaughtExceptionHandler handler; WorkQueue[] ws; int s, ps;
1395	>	wt.setDaemon(true);
1396	>	if ((handler = ueh) != null)
1397	>	wt.setUncaughtExceptionHandler(handler);
1398	>	do {} while (!U.compareAndSwapInt(this, INDEXSEED, s = indexSeed,
1399	>	s += SEED_INCREMENT) ||
1400	>	s == 0); // skip 0
1401	>	WorkQueue w = new WorkQueue(this, wt, config >>> 16, s);
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	<	ForkJoinWorkerThread[] ws = workers;
1408	<	int n = ws.length;
1409	<	if (k < 0 || k >= n || ws[k] != null) {
1410	<	for (k = 0; k < n && ws[k] != null; ++k)
1411	<	;
1412	<	if (k == n)
1413	<	ws = Arrays.copyOf(ws, n << 1);
1407	>	if ((ws = workQueues) != null) {    // skip if shutting down
1408	>	int n = ws.length, m = n - 1;
1409	>	int r = (s << 1) | 1;           // use odd-numbered indices
1410	>	if (ws[r &= m] != null) {       // collision
1411	>	int probes = 0;             // step by approx half size
1412	>	int step = (n <= 4) ? 2 : ((n >>> 1) & EVENMASK) + 2;
1413	>	while (ws[r = (r + step) & m] != null) {
1414	>	if (++probes >= n) {
1415	>	workQueues = ws = Arrays.copyOf(ws, n <<= 1);
1416	>	m = n - 1;
1417	>	probes = 0;
1418	>	}
1419	>	}
1420	>	}
1421	>	w.eventCount = w.poolIndex = r; // volatile write orders
1422	>	ws[r] = w;
1423		}
677	–	ws[k] = w;
678	–	workers = ws; // volatile array write ensures slot visibility
1424		} finally {
1425	<	lock.unlock();
1425	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1426	>	releasePlock(nps);
1427		}
1428	<	return k;
1428	>	wt.setName(workerNamePrefix.concat(Integer.toString(w.poolIndex)));
1429	>	return w;
1430		}
1431
1432		/**
1433	<	* Nulls out record of worker in workers array
1434	<	*/
1435	<	private void forgetWorker(ForkJoinWorkerThread w) {
1436	<	int idx = w.poolIndex;
1437	<	// Locking helps method recordWorker avoid unecessary expansion
1438	<	final ReentrantLock lock = this.workerLock;
1439	<	lock.lock();
1440	<	try {
1441	<	ForkJoinWorkerThread[] ws = workers;
1442	<	if (idx >= 0 && idx < ws.length && ws[idx] == w) // verify
1443	<	ws[idx] = null;
1444	<	} finally {
1445	<	lock.unlock();
1433	>	* Final callback from terminating worker, as well as upon failure
1434	>	* to construct or start a worker.  Removes record of worker from
1435	>	* array, and adjusts counts. If pool is shutting down, tries to
1436	>	* complete termination.
1437	>	*
1438	>	* @param wt the worker thread or null if construction failed
1439	>	* @param ex the exception causing failure, or null if none
1440	>	*/
1441	>	final void deregisterWorker(ForkJoinWorkerThread wt, Throwable ex) {
1442	>	WorkQueue w = null;
1443	>	if (wt != null && (w = wt.workQueue) != null) {
1444	>	int ps;
1445	>	w.qlock = -1;                // ensure set
1446	>	long ns = w.nsteals, sc;     // collect steal count
1447	>	do {} while (!U.compareAndSwapLong(this, STEALCOUNT,
1448	>	sc = stealCount, sc + ns));
1449	>	if (((ps = plock) & PL_LOCK) != 0 ||
1450	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1451	>	ps = acquirePlock();
1452	>	int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1453	>	try {
1454	>	int idx = w.poolIndex;
1455	>	WorkQueue[] ws = workQueues;
1456	>	if (ws != null && idx >= 0 && idx < ws.length && ws[idx] == w)
1457	>	ws[idx] = null;
1458	>	} finally {
1459	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1460	>	releasePlock(nps);
1461	>	}
1462		}
700	–	}
1463
1464	<	// adding and removing workers
1464	>	long c;                             // adjust ctl counts
1465	>	do {} while (!U.compareAndSwapLong
1466	>	(this, CTL, c = ctl, (((c - AC_UNIT) & AC_MASK) |
1467	>	((c - TC_UNIT) & TC_MASK) |
1468	>	(c & ~(AC_MASK|TC_MASK)))));
1469	>
1470	>	if (!tryTerminate(false, false) && w != null) {
1471	>	w.cancelAll();                  // cancel remaining tasks
1472	>	if (w.array != null)            // suppress signal if never ran
1473	>	tryAddWorker();             // create replacement
1474	>	if (ex == null)                 // help clean refs on way out
1475	>	ForkJoinTask.helpExpungeStaleExceptions();
1476	>	}
1477	>
1478	>	if (ex != null)                     // rethrow
1479	>	ForkJoinTask.rethrow(ex);
1480	>	}
1481	>
1482	>	// Submissions
1483	>
1484	>	/**
1485	>	* Unless shutting down, adds the given task to a submission queue
1486	>	* at submitter's current queue index (modulo submission
1487	>	* range). Only the most common path is directly handled in this
1488	>	* method. All others are relayed to fullExternalPush.
1489	>	*
1490	>	* @param task the task. Caller must ensure non-null.
1491	>	*/
1492	>	final void externalPush(ForkJoinTask<?> task) {
1493	>	WorkQueue[] ws; WorkQueue q; Submitter z; int m; ForkJoinTask<?>[] a;
1494	>	if ((z = submitters.get()) != null && plock > 0 &&
1495	>	(ws = workQueues) != null && (m = (ws.length - 1)) >= 0 &&
1496	>	(q = ws[m & z.seed & SQMASK]) != null &&
1497	>	U.compareAndSwapInt(q, QLOCK, 0, 1)) { // lock
1498	>	int b = q.base, s = q.top, n, an;
1499	>	if ((a = q.array) != null && (an = a.length) > (n = s + 1 - b)) {
1500	>	int j = (((an - 1) & s) << ASHIFT) + ABASE;
1501	>	U.putOrderedObject(a, j, task);
1502	>	q.top = s + 1;                     // push on to deque
1503	>	q.qlock = 0;
1504	>	if (n <= 2)
1505	>	signalWork(q);
1506	>	return;
1507	>	}
1508	>	q.qlock = 0;
1509	>	}
1510	>	fullExternalPush(task);
1511	>	}
1512
1513		/**
1514	<	* Tries to create and add new worker. Assumes that worker counts
1515	<	* are already updated to accommodate the worker, so adjusts on
1516	<	* failure.
1517	<	*/
1518	<	private void addWorker() {
1519	<	ForkJoinWorkerThread w = null;
1520	<	try {
1521	<	w = factory.newThread(this);
1522	<	} finally { // Adjust on either null or exceptional factory return
1523	<	if (w == null) {
1524	<	decrementWorkerCounts(ONE_RUNNING, ONE_TOTAL);
1525	<	tryTerminate(false); // in case of failure during shutdown
1514	>	* Full version of externalPush. This method is called, among
1515	>	* other times, upon the first submission of the first task to the
1516	>	* pool, so must perform secondary initialization (via
1517	>	* initWorkers). It also detects first submission by an external
1518	>	* thread by looking up its ThreadLocal, and creates a new shared
1519	>	* queue if the one at index if empty or contended. The plock lock
1520	>	* body must be exception-free (so no try/finally) so we
1521	>	* optimistically allocate new queues outside the lock and throw
1522	>	* them away if (very rarely) not needed.
1523	>	*/
1524	>	private void fullExternalPush(ForkJoinTask<?> task) {
1525	>	int r = 0; // random index seed
1526	>	for (Submitter z = submitters.get();;) {
1527	>	WorkQueue[] ws; WorkQueue q; int ps, m, k;
1528	>	if (z == null) {
1529	>	if (U.compareAndSwapInt(this, INDEXSEED, r = indexSeed,
1530	>	r += SEED_INCREMENT) && r != 0)
1531	>	submitters.set(z = new Submitter(r));
1532	>	}
1533	>	else if (r == 0) {               // move to a different index
1534	>	r = z.seed;
1535	>	r ^= r << 13;                // same xorshift as WorkQueues
1536	>	r ^= r >>> 17;
1537	>	z.seed = r ^ (r << 5);
1538	>	}
1539	>	else if ((ps = plock) < 0)
1540	>	throw new RejectedExecutionException();
1541	>	else if (ps == 0 || (ws = workQueues) == null ||
1542	>	(m = ws.length - 1) < 0)
1543	>	initWorkers();
1544	>	else if ((q = ws[k = r & m & SQMASK]) != null) {
1545	>	if (q.qlock == 0 && U.compareAndSwapInt(q, QLOCK, 0, 1)) {
1546	>	ForkJoinTask<?>[] a = q.array;
1547	>	int s = q.top;
1548	>	boolean submitted = false;
1549	>	try {                      // locked version of push
1550	>	if ((a != null && a.length > s + 1 - q.base) ||
1551	>	(a = q.growArray()) != null) {   // must presize
1552	>	int j = (((a.length - 1) & s) << ASHIFT) + ABASE;
1553	>	U.putOrderedObject(a, j, task);
1554	>	q.top = s + 1;
1555	>	submitted = true;
1556	>	}
1557	>	} finally {
1558	>	q.qlock = 0;  // unlock
1559	>	}
1560	>	if (submitted) {
1561	>	signalWork(q);
1562	>	return;
1563	>	}
1564	>	}
1565	>	r = 0; // move on failure
1566		}
1567	+	else if (((ps = plock) & PL_LOCK) == 0) { // create new queue
1568	+	q = new WorkQueue(this, null, SHARED_QUEUE, r);
1569	+	if (((ps = plock) & PL_LOCK) != 0 ||
1570	+	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1571	+	ps = acquirePlock();
1572	+	if ((ws = workQueues) != null && k < ws.length && ws[k] == null)
1573	+	ws[k] = q;
1574	+	int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1575	+	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1576	+	releasePlock(nps);
1577	+	}
1578	+	else
1579	+	r = 0; // try elsewhere while lock held
1580		}
719	–	if (w != null)
720	–	w.start(recordWorker(w), ueh);
1581		}
1582
1583	+	// Maintaining ctl counts
1584	+
1585		/**
1586	<	* Final callback from terminating worker.  Removes record of
725	<	* worker from array, and adjusts counts. If pool is shutting
726	<	* down, tries to complete terminatation.
727	<	*
728	<	* @param w the worker
1586	>	* Increments active count; mainly called upon return from blocking.
1587		*/
1588	<	final void workerTerminated(ForkJoinWorkerThread w) {
1589	<	forgetWorker(w);
1590	<	decrementWorkerCounts(w.isTrimmed()? 0 : ONE_RUNNING, ONE_TOTAL);
733	<	while (w.stealCount != 0) // collect final count
734	<	tryAccumulateStealCount(w);
735	<	tryTerminate(false);
1588	>	final void incrementActiveCount() {
1589	>	long c;
1590	>	do {} while (!U.compareAndSwapLong(this, CTL, c = ctl, c + AC_UNIT));
1591		}
1592
738	–	// Waiting for and signalling events
739	–
1593		/**
1594	<	* Releases workers blocked on a count not equal to current count.
742	<	* Normally called after precheck that eventWaiters isn't zero to
743	<	* avoid wasted array checks.
1594	>	* Tries to create or activate a worker if too few are active.
1595		*
1596	<	* @param signalling true if caller is a signalling worker so can
746	<	* exit upon (conservatively) detected contention by other threads
747	<	* who will continue to release
1596	>	* @param q the (non-null) queue holding tasks to be signalled
1597		*/
1598	<	private void releaseEventWaiters(boolean signalling) {
1599	<	ForkJoinWorkerThread[] ws = workers;
1600	<	int n = ws.length;
1601	<	long h; // head of stack
1602	<	ForkJoinWorkerThread w; int id, ec;
1603	<	while ((id = ((int)((h = eventWaiters) & WAITER_ID_MASK)) - 1) >= 0 &&
1604	<	(int)(h >>> EVENT_COUNT_SHIFT) != (ec = eventCount) &&
1605	<	id < n && (w = ws[id]) != null) {
1606	<	if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
1607	<	h, h = w.nextWaiter))
1608	<	LockSupport.unpark(w);
1609	<	if (signalling && (eventCount != ec || eventWaiters != h))
1598	>	final void signalWork(WorkQueue q) {
1599	>	int hint = q.poolIndex;
1600	>	long c; int e, u, i, n; WorkQueue[] ws; WorkQueue w; Thread p;
1601	>	while ((u = (int)((c = ctl) >>> 32)) < 0) {
1602	>	if ((e = (int)c) > 0) {
1603	>	if ((ws = workQueues) != null && ws.length > (i = e & SMASK) &&
1604	>	(w = ws[i]) != null && w.eventCount == (e | INT_SIGN)) {
1605	>	long nc = (((long)(w.nextWait & E_MASK)) |
1606	>	((long)(u + UAC_UNIT) << 32));
1607	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1608	>	w.hint = hint;
1609	>	w.eventCount = (e + E_SEQ) & E_MASK;
1610	>	if ((p = w.parker) != null)
1611	>	U.unpark(p);
1612	>	break;
1613	>	}
1614	>	if (q.top - q.base <= 0)
1615	>	break;
1616	>	}
1617	>	else
1618	>	break;
1619	>	}
1620	>	else {
1621	>	if ((short)u < 0)
1622	>	tryAddWorker();
1623		break;
1624	+	}
1625		}
1626		}
1627
1628	+	// Scanning for tasks
1629	+
1630		/**
1631	<	* Tries to advance eventCount and releases waiters. Called only
767	<	* from workers.
1631	>	* Top-level runloop for workers, called by ForkJoinWorkerThread.run.
1632		*/
1633	<	final void signalWork() {
1634	<	int c; // try to increment event count -- CAS failure OK
1635	<	UNSAFE.compareAndSwapInt(this, eventCountOffset, c = eventCount, c+1);
1636	<	if (eventWaiters != 0L)
1637	<	releaseEventWaiters(true);
1633	>	final void runWorker(WorkQueue w) {
1634	>	w.growArray(); // allocate queue
1635	>	do { w.runTask(scan(w)); } while (w.qlock >= 0);
1636	>	}
1637	>
1638	>	/**
1639	>	* Scans for and, if found, returns one task, else possibly
1640	>	* inactivates the worker. This method operates on single reads of
1641	>	* volatile state and is designed to be re-invoked continuously,
1642	>	* in part because it returns upon detecting inconsistencies,
1643	>	* contention, or state changes that indicate possible success on
1644	>	* re-invocation.
1645	>	*
1646	>	* The scan searches for tasks across queues (starting at a random
1647	>	* index, and relying on registerWorker to irregularly scatter
1648	>	* them within array to avoid bias), checking each at least twice.
1649	>	* The scan terminates upon either finding a non-empty queue, or
1650	>	* completing the sweep. If the worker is not inactivated, it
1651	>	* takes and returns a task from this queue. Otherwise, if not
1652	>	* activated, it signals workers (that may include itself) and
1653	>	* returns so caller can retry. Also returns for true if the
1654	>	* worker array may have changed during an empty scan.  On failure
1655	>	* to find a task, we take one of the following actions, after
1656	>	* which the caller will retry calling this method unless
1657	>	* terminated.
1658	>	*
1659	>	* * If pool is terminating, terminate the worker.
1660	>	*
1661	>	* * If not already enqueued, try to inactivate and enqueue the
1662	>	* worker on wait queue. Or, if inactivating has caused the pool
1663	>	* to be quiescent, relay to idleAwaitWork to check for
1664	>	* termination and possibly shrink pool.
1665	>	*
1666	>	* * If already enqueued and none of the above apply, possibly
1667	>	* (with 1/2 probability) park awaiting signal, else lingering to
1668	>	* help scan and signal.
1669	>	*
1670	>	* @param w the worker (via its WorkQueue)
1671	>	* @return a task or null if none found
1672	>	*/
1673	>	private final ForkJoinTask<?> scan(WorkQueue w) {
1674	>	WorkQueue[] ws; int m;
1675	>	int ps = plock;                          // read plock before ws
1676	>	if (w != null && (ws = workQueues) != null && (m = ws.length - 1) >= 0) {
1677	>	int ec = w.eventCount;               // ec is negative if inactive
1678	>	int r = w.seed; r ^= r << 13; r ^= r >>> 17; w.seed = r ^= r << 5;
1679	>	int j = ((m + m + 1) | MIN_SCAN) & MAX_SCAN;
1680	>	do {
1681	>	WorkQueue q; ForkJoinTask<?>[] a; int b;
1682	>	if ((q = ws[(r + j) & m]) != null && (b = q.base) - q.top < 0 &&
1683	>	(a = q.array) != null) {     // probably nonempty
1684	>	int i = (((a.length - 1) & b) << ASHIFT) + ABASE;
1685	>	ForkJoinTask<?> t = (ForkJoinTask<?>)
1686	>	U.getObjectVolatile(a, i);
1687	>	if (q.base == b && ec >= 0 && t != null &&
1688	>	U.compareAndSwapObject(a, i, t, null)) {
1689	>	if ((q.base = b + 1) - q.top < 0)
1690	>	signalWork(q);
1691	>	return t;                // taken
1692	>	}
1693	>	else if ((ec < 0 || j < m) && (int)(ctl >> AC_SHIFT) <= 0) {
1694	>	w.hint = (r + j) & m;    // help signal below
1695	>	break;                   // cannot take
1696	>	}
1697	>	}
1698	>	} while (--j >= 0);
1699	>
1700	>	long c, sc; int e, ns, h;
1701	>	if ((h = w.hint) < 0) {
1702	>	if ((ns = w.nsteals) != 0) {
1703	>	if (U.compareAndSwapLong(this, STEALCOUNT,
1704	>	sc = stealCount, sc + ns))
1705	>	w.nsteals = 0;           // collect steals
1706	>	}
1707	>	else if (plock != ps)            // consistency check
1708	>	;                            // skip
1709	>	else if ((e = (int)(c = ctl)) < 0)
1710	>	w.qlock = -1;                // pool is terminating
1711	>	else if (ec >= 0) {              // try to enqueue/inactivate
1712	>	long nc = ((long)ec | ((c - AC_UNIT) & (AC_MASK|TC_MASK)));
1713	>	w.nextWait = e;              // link and mark inactive
1714	>	w.eventCount = ec | INT_SIGN;
1715	>	if (ctl != c || !U.compareAndSwapLong(this, CTL, c, nc))
1716	>	w.eventCount = ec;       // unmark on CAS failure
1717	>	else if ((int)(c >> AC_SHIFT) == 1 - (config & SMASK))
1718	>	idleAwaitWork(w, nc, c);
1719	>	}
1720	>	else if (w.eventCount < 0) {     // block
1721	>	Thread wt = Thread.currentThread();
1722	>	Thread.interrupted();        // clear status
1723	>	U.putObject(wt, PARKBLOCKER, this);
1724	>	w.parker = wt;               // emulate LockSupport.park
1725	>	if (w.eventCount < 0)        // recheck
1726	>	U.park(false, 0L);
1727	>	w.parker = null;
1728	>	U.putObject(wt, PARKBLOCKER, null);
1729	>	}
1730	>	}
1731	>	if (h >= 0 || (h = w.hint) >= 0) {   // signal others before retry
1732	>	w.hint = -1;                     // reset
1733	>	helpSignal(null, h, true);
1734	>	}
1735	>	}
1736	>	return null;
1737		}
1738
1739		/**
1740	<	* Blocks worker until terminating or event count
1741	<	* advances from last value held by worker
1742	<	*
1743	<	* @param w the calling worker thread
1744	<	*/
1745	<	private void eventSync(ForkJoinWorkerThread w) {
1746	<	int wec = w.lastEventCount;
1747	<	long nh = (((long)wec) << EVENT_COUNT_SHIFT) | ((long)(w.poolIndex+1));
1748	<	long h;
1749	<	while ((runState < SHUTDOWN || !tryTerminate(false)) &&
1750	<	((h = eventWaiters) == 0L ||
1751	<	(int)(h >>> EVENT_COUNT_SHIFT) == wec) &&
1752	<	eventCount == wec) {
1753	<	if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
1754	<	w.nextWaiter = h, nh)) {
1755	<	while (runState < TERMINATING && eventCount == wec) {
1756	<	if (!tryAccumulateStealCount(w))  // transfer while idle
1757	<	continue;
1758	<	Thread.interrupted();             // clear/ignore interrupt
1759	<	if (eventCount != wec)
1760	<	break;
1761	<	LockSupport.park(w);
1740	>	* If inactivating worker w has caused the pool to become
1741	>	* quiescent, checks for pool termination, and, so long as this is
1742	>	* not the only worker, waits for event for up to a given
1743	>	* duration.  On timeout, if ctl has not changed, terminates the
1744	>	* worker, which will in turn wake up another worker to possibly
1745	>	* repeat this process.
1746	>	*
1747	>	* @param w the calling worker
1748	>	* @param currentCtl the ctl value triggering possible quiescence
1749	>	* @param prevCtl the ctl value to restore if thread is terminated
1750	>	*/
1751	>	private void idleAwaitWork(WorkQueue w, long currentCtl, long prevCtl) {
1752	>	if (w != null && w.eventCount < 0 &&
1753	>	!tryTerminate(false, false) && (int)prevCtl != 0) {
1754	>	int dc = -(short)(currentCtl >>> TC_SHIFT);
1755	>	long parkTime = dc < 0 ? FAST_IDLE_TIMEOUT: (dc + 1) * IDLE_TIMEOUT;
1756	>	long deadline = System.nanoTime() + parkTime - 100000L; // 1ms slop
1757	>	Thread wt = Thread.currentThread();
1758	>	while (ctl == currentCtl) {
1759	>	Thread.interrupted();  // timed variant of version in scan()
1760	>	U.putObject(wt, PARKBLOCKER, this);
1761	>	w.parker = wt;
1762	>	if (ctl == currentCtl)
1763	>	U.park(false, parkTime);
1764	>	w.parker = null;
1765	>	U.putObject(wt, PARKBLOCKER, null);
1766	>	if (ctl != currentCtl)
1767	>	break;
1768	>	if (deadline - System.nanoTime() <= 0L &&
1769	>	U.compareAndSwapLong(this, CTL, currentCtl, prevCtl)) {
1770	>	w.eventCount = (w.eventCount + E_SEQ) | E_MASK;
1771	>	w.qlock = -1;   // shrink
1772	>	break;
1773		}
800	–	break;
1774		}
1775		}
803	–	w.lastEventCount = eventCount;
1776		}
1777
1778	<	// Maintaining spares
1778	>	/**
1779	>	* Scans through queues looking for work (optionally, while
1780	>	* joining a task); if any present, signals. May return early if
1781	>	* more signalling is detectably unneeded.
1782	>	*
1783	>	* @param task if non-null, return early if done
1784	>	* @param origin an index to start scan
1785	>	* @param once if only the origin should be checked
1786	>	*/
1787	>	private void helpSignal(ForkJoinTask<?> task, int origin, boolean once) {
1788	>	WorkQueue[] ws; WorkQueue w; Thread p; long c; int m, u, e, i, s;
1789	>	if ((u = (int)(ctl >>> 32)) < 0 && (u >> UAC_SHIFT) < 0 &&
1790	>	(ws = workQueues) != null && (m = ws.length - 1) >= 0) {
1791	>	outer: for (int k = origin, j = once ? 0 : m; j >= 0; --j) {
1792	>	WorkQueue q = ws[k++ & m];
1793	>	for (int n = m;;) { // limit to at most m signals
1794	>	if (task != null && task.status < 0)
1795	>	break outer;
1796	>	if (q == null ||
1797	>	((s = (task == null ? -1 : 0) - q.base + q.top) <= n &&
1798	>	(n = s) <= 0))
1799	>	break;
1800	>	if ((u = (int)((c = ctl) >>> 32)) >= 0 ||
1801	>	(e = (int)c) <= 0 || m < (i = e & SMASK) ||
1802	>	(w = ws[i]) == null)
1803	>	break outer;
1804	>	long nc = (((long)(w.nextWait & E_MASK)) |
1805	>	((long)(u + UAC_UNIT) << 32));
1806	>	if (w.eventCount == (e | INT_SIGN) &&
1807	>	U.compareAndSwapLong(this, CTL, c, nc)) {
1808	>	w.eventCount = (e + E_SEQ) & E_MASK;
1809	>	if ((p = w.parker) != null)
1810	>	U.unpark(p);
1811	>	if (--n <= 0)
1812	>	break;
1813	>	}
1814	>	}
1815	>	}
1816	>	}
1817	>	}
1818
1819		/**
1820	<	* Pushes worker onto the spare stack
1821	<	*/
1822	<	final void pushSpare(ForkJoinWorkerThread w) {
1823	<	int ns = (++w.spareCount << SPARE_COUNT_SHIFT) | (w.poolIndex+1);
1824	<	do {} while (!UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
1825	<	w.nextSpare = spareWaiters,ns));
1826	<	}
1827	<
1828	<	/**
1829	<	* Tries (once) to resume a spare if running count is less than
1830	<	* target parallelism. Fails on contention or stale workers.
1831	<	*/
1832	<	private void tryResumeSpare() {
1833	<	int sw, id;
1834	<	ForkJoinWorkerThread w;
1835	<	ForkJoinWorkerThread[] ws;
1836	<	if ((id = ((sw = spareWaiters) & SPARE_ID_MASK) - 1) >= 0 &&
1837	<	id < (ws = workers).length && (w = ws[id]) != null &&
1838	<	(workerCounts & RUNNING_COUNT_MASK) < parallelism &&
1839	<	eventWaiters == 0L &&
1840	<	spareWaiters == sw &&
1841	<	UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
1842	<	sw, w.nextSpare) &&
1843	<	w.tryUnsuspend()) {
1844	<	int c; // try increment; if contended, finish after unpark
1845	<	boolean inc = UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1846	<	c = workerCounts,
1847	<	c + ONE_RUNNING);
1848	<	LockSupport.unpark(w);
1849	<	if (!inc) {
1850	<	do {} while(!UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1851	<	c = workerCounts,
1852	<	c + ONE_RUNNING));
1820	>	* Tries to locate and execute tasks for a stealer of the given
1821	>	* task, or in turn one of its stealers, Traces currentSteal ->
1822	>	* currentJoin links looking for a thread working on a descendant
1823	>	* of the given task and with a non-empty queue to steal back and
1824	>	* execute tasks from. The first call to this method upon a
1825	>	* waiting join will often entail scanning/search, (which is OK
1826	>	* because the joiner has nothing better to do), but this method
1827	>	* leaves hints in workers to speed up subsequent calls. The
1828	>	* implementation is very branchy to cope with potential
1829	>	* inconsistencies or loops encountering chains that are stale,
1830	>	* unknown, or so long that they are likely cyclic.
1831	>	*
1832	>	* @param joiner the joining worker
1833	>	* @param task the task to join
1834	>	* @return 0 if no progress can be made, negative if task
1835	>	* known complete, else positive
1836	>	*/
1837	>	private int tryHelpStealer(WorkQueue joiner, ForkJoinTask<?> task) {
1838	>	int stat = 0, steps = 0;                    // bound to avoid cycles
1839	>	if (joiner != null && task != null) {       // hoist null checks
1840	>	restart: for (;;) {
1841	>	ForkJoinTask<?> subtask = task;     // current target
1842	>	for (WorkQueue j = joiner, v;;) {   // v is stealer of subtask
1843	>	WorkQueue[] ws; int m, s, h;
1844	>	if ((s = task.status) < 0) {
1845	>	stat = s;
1846	>	break restart;
1847	>	}
1848	>	if ((ws = workQueues) == null || (m = ws.length - 1) <= 0)
1849	>	break restart;              // shutting down
1850	>	if ((v = ws[h = (j.hint | 1) & m]) == null ||
1851	>	v.currentSteal != subtask) {
1852	>	for (int origin = h;;) {    // find stealer
1853	>	if (((h = (h + 2) & m) & 15) == 1 &&
1854	>	(subtask.status < 0 || j.currentJoin != subtask))
1855	>	continue restart;   // occasional staleness check
1856	>	if ((v = ws[h]) != null &&
1857	>	v.currentSteal == subtask) {
1858	>	j.hint = h;        // save hint
1859	>	break;
1860	>	}
1861	>	if (h == origin)
1862	>	break restart;      // cannot find stealer
1863	>	}
1864	>	}
1865	>	for (;;) { // help stealer or descend to its stealer
1866	>	ForkJoinTask[] a;  int b;
1867	>	if (subtask.status < 0)     // surround probes with
1868	>	continue restart;       //   consistency checks
1869	>	if ((b = v.base) - v.top < 0 && (a = v.array) != null) {
1870	>	int i = (((a.length - 1) & b) << ASHIFT) + ABASE;
1871	>	ForkJoinTask<?> t =
1872	>	(ForkJoinTask<?>)U.getObjectVolatile(a, i);
1873	>	if (subtask.status < 0 || j.currentJoin != subtask ||
1874	>	v.currentSteal != subtask)
1875	>	continue restart;   // stale
1876	>	stat = 1;               // apparent progress
1877	>	if (t != null && v.base == b &&
1878	>	U.compareAndSwapObject(a, i, t, null)) {
1879	>	v.base = b + 1;     // help stealer
1880	>	joiner.runSubtask(t);
1881	>	}
1882	>	else if (v.base == b && ++steps == MAX_HELP)
1883	>	break restart;      // v apparently stalled
1884	>	}
1885	>	else {                      // empty -- try to descend
1886	>	ForkJoinTask<?> next = v.currentJoin;
1887	>	if (subtask.status < 0 || j.currentJoin != subtask ||
1888	>	v.currentSteal != subtask)
1889	>	continue restart;   // stale
1890	>	else if (next == null || ++steps == MAX_HELP)
1891	>	break restart;      // dead-end or maybe cyclic
1892	>	else {
1893	>	subtask = next;
1894	>	j = v;
1895	>	break;
1896	>	}
1897	>	}
1898	>	}
1899	>	}
1900		}
1901		}
1902	+	return stat;
1903		}
1904
1905		/**
1906	<	* Callback from oldest spare occasionally waking up.  Tries
1907	<	* (once) to shutdown a spare if more than 25% spare overage, or
1908	<	* if UNUSED_SPARE_TRIM_RATE_NANOS have elapsed and there are at
1909	<	* least #parallelism running threads. Note that we don't need CAS
1910	<	* or locks here because the method is called only from the oldest
1911	<	* suspended spare occasionally waking (and even misfires are OK).
1912	<	*
1913	<	* @param now the wake up nanoTime of caller
1914	<	*/
1915	<	final void tryTrimSpare(long now) {
1916	<	long lastTrim = trimTime;
1917	<	trimTime = now;
1918	<	helpMaintainParallelism(); // first, help wake up any needed spares
1919	<	int sw, id;
1920	<	ForkJoinWorkerThread w;
1921	<	ForkJoinWorkerThread[] ws;
1922	<	int pc = parallelism;
1923	<	int wc = workerCounts;
1924	<	if ((wc & RUNNING_COUNT_MASK) >= pc &&
1925	<	(((wc >>> TOTAL_COUNT_SHIFT) - pc) > (pc >>> 2) + 1 ||// approx 25%
1926	<	now - lastTrim >= UNUSED_SPARE_TRIM_RATE_NANOS) &&
1927	<	(id = ((sw = spareWaiters) & SPARE_ID_MASK) - 1) >= 0 &&
869	<	id < (ws = workers).length && (w = ws[id]) != null &&
870	<	UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
871	<	sw, w.nextSpare))
872	<	w.shutdown(false);
873	<	}
874	<
875	<	/**
876	<	* Does at most one of:
877	<	*
878	<	* 1. Help wake up existing workers waiting for work via
879	<	*    releaseEventWaiters. (If any exist, then it probably doesn't
880	<	*    matter right now if under target parallelism level.)
881	<	*
882	<	* 2. If below parallelism level and a spare exists, try (once)
883	<	*    to resume it via tryResumeSpare.
884	<	*
885	<	* 3. If neither of the above, tries (once) to add a new
886	<	*    worker if either there are not enough total, or if all
887	<	*    existing workers are busy, there are either no running
888	<	*    workers or the deficit is at least twice the surplus.
889	<	*/
890	<	private void helpMaintainParallelism() {
891	<	// uglified to work better when not compiled
892	<	int pc, wc, rc, tc, rs; long h;
893	<	if ((h = eventWaiters) != 0L) {
894	<	if ((int)(h >>> EVENT_COUNT_SHIFT) != eventCount)
895	<	releaseEventWaiters(false); // avoid useless call
896	<	}
897	<	else if ((pc = parallelism) >
898	<	(rc = ((wc = workerCounts) & RUNNING_COUNT_MASK))) {
899	<	if (spareWaiters != 0)
900	<	tryResumeSpare();
901	<	else if ((rs = runState) < TERMINATING &&
902	<	((tc = wc >>> TOTAL_COUNT_SHIFT) < pc ||
903	<	(tc == (rs & ACTIVE_COUNT_MASK) && // all busy
904	<	(rc == 0 ||                       // must add
905	<	rc < pc - ((tc - pc) << 1)) &&   // within slack
906	<	tc < MAX_WORKERS && runState == rs)) && // recheck busy
907	<	workerCounts == wc &&
908	<	UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
909	<	wc + (ONE_RUNNING|ONE_TOTAL)))
910	<	addWorker();
911	<	}
912	<	}
913	<
914	<	/**
915	<	* Callback from workers invoked upon each top-level action (i.e.,
916	<	* stealing a task or taking a submission and running
917	<	* it). Performs one or more of the following:
918	<	*
919	<	* 1. If the worker cannot find work (misses > 0), updates its
920	<	*    active status to inactive and updates activeCount unless
921	<	*    this is the first miss and there is contention, in which
922	<	*    case it may try again (either in this or a subsequent
923	<	*    call).
924	<	*
925	<	* 2. If there are at least 2 misses, awaits the next task event
926	<	*    via eventSync
927	<	*
928	<	* 3. If there are too many running threads, suspends this worker
929	<	*    (first forcing inactivation if necessary).  If it is not
930	<	*    needed, it may be killed while suspended via
931	<	*    tryTrimSpare. Otherwise, upon resume it rechecks to make
932	<	*    sure that it is still needed.
933	<	*
934	<	* 4. Helps release and/or reactivate other workers via
935	<	*    helpMaintainParallelism
936	<	*
937	<	* @param w the worker
938	<	* @param misses the number of scans by caller failing to find work
939	<	* (saturating at 2 just to avoid wraparound)
940	<	*/
941	<	final void preStep(ForkJoinWorkerThread w, int misses) {
942	<	boolean active = w.active;
943	<	int pc = parallelism;
944	<	for (;;) {
945	<	int wc = workerCounts;
946	<	int rc = wc & RUNNING_COUNT_MASK;
947	<	if (active && (misses > 0 || rc > pc)) {
948	<	int rs;                      // try inactivate
949	<	if (UNSAFE.compareAndSwapInt(this, runStateOffset,
950	<	rs = runState, rs - ONE_ACTIVE))
951	<	active = w.active = false;
952	<	else if (misses > 1 || rc > pc ||
953	<	(rs & ACTIVE_COUNT_MASK) >= pc)
954	<	continue;                // force inactivate
955	<	}
956	<	if (misses > 1) {
957	<	misses = 0;                  // don't re-sync
958	<	eventSync(w);                // continue loop to recheck rc
959	<	}
960	<	else if (rc > pc) {
961	<	if (workerCounts == wc &&   // try to suspend as spare
962	<	UNSAFE.compareAndSwapInt(this, workerCountsOffset,
963	<	wc, wc - ONE_RUNNING) &&
964	<	!w.suspendAsSpare())    // false if killed
1906	>	* Analog of tryHelpStealer for CountedCompleters. Tries to steal
1907	>	* and run tasks within the target's computation.
1908	>	*
1909	>	* @param task the task to join
1910	>	* @param mode if shared, exit upon completing any task
1911	>	* if all workers are active
1912	>	*
1913	>	*/
1914	>	private int helpComplete(ForkJoinTask<?> task, int mode) {
1915	>	WorkQueue[] ws; WorkQueue q; int m, n, s, u;
1916	>	if (task != null && (ws = workQueues) != null &&
1917	>	(m = ws.length - 1) >= 0) {
1918	>	for (int j = 1, origin = j;;) {
1919	>	if ((s = task.status) < 0)
1920	>	return s;
1921	>	if ((q = ws[j & m]) != null && q.pollAndExecCC(task)) {
1922	>	origin = j;
1923	>	if (mode == SHARED_QUEUE &&
1924	>	((u = (int)(ctl >>> 32)) >= 0 || (u >> UAC_SHIFT) >= 0))
1925	>	break;
1926	>	}
1927	>	else if ((j = (j + 2) & m) == origin)
1928		break;
1929		}
967	–	else {
968	–	if (rc < pc || eventWaiters != 0L)
969	–	helpMaintainParallelism();
970	–	break;
971	–	}
1930		}
1931	+	return 0;
1932		}
1933
1934		/**
1935	<	* Helps and/or blocks awaiting join of the given task.
1936	<	* Alternates between helpJoinTask() and helpMaintainParallelism()
1937	<	* as many times as there is a deficit in running count (or longer
1938	<	* if running count would become zero), then blocks if task still
1939	<	* not done.
1940	<	*
1941	<	* @param joinMe the task to join
1942	<	*/
1943	<	final void awaitJoin(ForkJoinTask<?> joinMe, ForkJoinWorkerThread worker) {
1944	<	int threshold = parallelism;         // descend blocking thresholds
1945	<	while (joinMe.status >= 0) {
1946	<	boolean block; int wc;
1947	<	worker.helpJoinTask(joinMe);
1948	<	if (joinMe.status < 0)
1949	<	break;
1950	<	if (((wc = workerCounts) & RUNNING_COUNT_MASK) <= threshold) {
1951	<	if (threshold > 0)
1952	<	--threshold;
1953	<	else
1954	<	advanceEventCount(); // force release
996	<	block = false;
1935	>	* Tries to decrement active count (sometimes implicitly) and
1936	>	* possibly release or create a compensating worker in preparation
1937	>	* for blocking. Fails on contention or termination. Otherwise,
1938	>	* adds a new thread if no idle workers are available and pool
1939	>	* may become starved.
1940	>	*/
1941	>	final boolean tryCompensate() {
1942	>	int pc = config & SMASK, e, i, tc; long c;
1943	>	WorkQueue[] ws; WorkQueue w; Thread p;
1944	>	if ((ws = workQueues) != null && (e = (int)(c = ctl)) >= 0) {
1945	>	if (e != 0 && (i = e & SMASK) < ws.length &&
1946	>	(w = ws[i]) != null && w.eventCount == (e | INT_SIGN)) {
1947	>	long nc = ((long)(w.nextWait & E_MASK) |
1948	>	(c & (AC_MASK|TC_MASK)));
1949	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1950	>	w.eventCount = (e + E_SEQ) & E_MASK;
1951	>	if ((p = w.parker) != null)
1952	>	U.unpark(p);
1953	>	return true;   // replace with idle worker
1954	>	}
1955		}
1956	<	else
1957	<	block = UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1958	<	wc, wc - ONE_RUNNING);
1959	<	helpMaintainParallelism();
1960	<	if (block) {
1961	<	int c;
1962	<	joinMe.internalAwaitDone();
1963	<	do {} while (!UNSAFE.compareAndSwapInt
1964	<	(this, workerCountsOffset,
1965	<	c = workerCounts, c + ONE_RUNNING));
1966	<	break;
1956	>	else if ((tc = (short)(c >>> TC_SHIFT)) >= 0 &&
1957	>	(int)(c >> AC_SHIFT) + pc > 1) {
1958	>	long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK);
1959	>	if (U.compareAndSwapLong(this, CTL, c, nc))
1960	>	return true;   // no compensation
1961	>	}
1962	>	else if (tc + pc < MAX_CAP) {
1963	>	long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK);
1964	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1965	>	ForkJoinWorkerThreadFactory fac;
1966	>	Throwable ex = null;
1967	>	ForkJoinWorkerThread wt = null;
1968	>	try {
1969	>	if ((fac = factory) != null &&
1970	>	(wt = fac.newThread(this)) != null) {
1971	>	wt.start();
1972	>	return true;
1973	>	}
1974	>	} catch (Throwable rex) {
1975	>	ex = rex;
1976	>	}
1977	>	deregisterWorker(wt, ex); // clean up and return false
1978	>	}
1979		}
1980		}
1981	+	return false;
1982		}
1983
1984		/**
1985	<	* Same idea as awaitJoin, but no helping
1985	>	* Helps and/or blocks until the given task is done.
1986	>	*
1987	>	* @param joiner the joining worker
1988	>	* @param task the task
1989	>	* @return task status on exit
1990		*/
1991	<	final void awaitBlocker(ManagedBlocker blocker)
1992	<	throws InterruptedException {
1993	<	int threshold = parallelism;
1994	<	while (!blocker.isReleasable()) {
1995	<	boolean block; int wc;
1996	<	if (((wc = workerCounts) & RUNNING_COUNT_MASK) <= threshold) {
1997	<	if (threshold > 0)
1998	<	--threshold;
1999	<	else
2000	<	advanceEventCount();
2001	<	block = false;
1991	>	final int awaitJoin(WorkQueue joiner, ForkJoinTask<?> task) {
1992	>	int s = 0;
1993	>	if (joiner != null && task != null && (s = task.status) >= 0) {
1994	>	ForkJoinTask<?> prevJoin = joiner.currentJoin;
1995	>	joiner.currentJoin = task;
1996	>	do {} while ((s = task.status) >= 0 && !joiner.isEmpty() &&
1997	>	joiner.tryRemoveAndExec(task)); // process local tasks
1998	>	if (s >= 0 && (s = task.status) >= 0) {
1999	>	helpSignal(task, joiner.poolIndex, false);
2000	>	if ((s = task.status) >= 0 &&
2001	>	(task instanceof CountedCompleter))
2002	>	s = helpComplete(task, LIFO_QUEUE);
2003		}
2004	<	else
2005	<	block = UNSAFE.compareAndSwapInt(this, workerCountsOffset,
2006	<	wc, wc - ONE_RUNNING);
2007	<	helpMaintainParallelism();
2008	<	if (block) {
2009	<	try {
2010	<	do {} while (!blocker.isReleasable() && !blocker.block());
2011	<	} finally {
2012	<	int c;
2013	<	do {} while (!UNSAFE.compareAndSwapInt
2014	<	(this, workerCountsOffset,
2015	<	c = workerCounts, c + ONE_RUNNING));
2004	>	while (s >= 0 && (s = task.status) >= 0) {
2005	>	if ((!joiner.isEmpty() ||           // try helping
2006	>	(s = tryHelpStealer(joiner, task)) == 0) &&
2007	>	(s = task.status) >= 0) {
2008	>	helpSignal(task, joiner.poolIndex, false);
2009	>	if ((s = task.status) >= 0 && tryCompensate()) {
2010	>	if (task.trySetSignal() && (s = task.status) >= 0) {
2011	>	synchronized (task) {
2012	>	if (task.status >= 0) {
2013	>	try {                // see ForkJoinTask
2014	>	task.wait();     //  for explanation
2015	>	} catch (InterruptedException ie) {
2016	>	}
2017	>	}
2018	>	else
2019	>	task.notifyAll();
2020	>	}
2021	>	}
2022	>	long c;                          // re-activate
2023	>	do {} while (!U.compareAndSwapLong
2024	>	(this, CTL, c = ctl, c + AC_UNIT));
2025	>	}
2026		}
1041	–	break;
2027		}
2028	+	joiner.currentJoin = prevJoin;
2029		}
2030	+	return s;
2031		}
2032
2033		/**
2034	<	* Possibly initiates and/or completes termination.
2034	>	* Stripped-down variant of awaitJoin used by timed joins. Tries
2035	>	* to help join only while there is continuous progress. (Caller
2036	>	* will then enter a timed wait.)
2037		*
2038	<	* @param now if true, unconditionally terminate, else only
2039	<	* if shutdown and empty queue and no active workers
1051	<	* @return true if now terminating or terminated
2038	>	* @param joiner the joining worker
2039	>	* @param task the task
2040		*/
2041	<	private boolean tryTerminate(boolean now) {
2042	<	if (now)
2043	<	advanceRunLevel(SHUTDOWN); // ensure at least SHUTDOWN
2044	<	else if (runState < SHUTDOWN ||
2045	<	!submissionQueue.isEmpty() ||
2046	<	(runState & ACTIVE_COUNT_MASK) != 0)
2047	<	return false;
2041	>	final void helpJoinOnce(WorkQueue joiner, ForkJoinTask<?> task) {
2042	>	int s;
2043	>	if (joiner != null && task != null && (s = task.status) >= 0) {
2044	>	ForkJoinTask<?> prevJoin = joiner.currentJoin;
2045	>	joiner.currentJoin = task;
2046	>	do {} while ((s = task.status) >= 0 && !joiner.isEmpty() &&
2047	>	joiner.tryRemoveAndExec(task));
2048	>	if (s >= 0 && (s = task.status) >= 0) {
2049	>	helpSignal(task, joiner.poolIndex, false);
2050	>	if ((s = task.status) >= 0 &&
2051	>	(task instanceof CountedCompleter))
2052	>	s = helpComplete(task, LIFO_QUEUE);
2053	>	}
2054	>	if (s >= 0 && joiner.isEmpty()) {
2055	>	do {} while (task.status >= 0 &&
2056	>	tryHelpStealer(joiner, task) > 0);
2057	>	}
2058	>	joiner.currentJoin = prevJoin;
2059	>	}
2060	>	}
2061
2062	<	if (advanceRunLevel(TERMINATING))
2063	<	startTerminating();
2062	>	/**
2063	>	* Returns a (probably) non-empty steal queue, if one is found
2064	>	* during a random, then cyclic scan, else null.  This method must
2065	>	* be retried by caller if, by the time it tries to use the queue,
2066	>	* it is empty.
2067	>	* @param r a (random) seed for scanning
2068	>	*/
2069	>	private WorkQueue findNonEmptyStealQueue(int r) {
2070	>	for (WorkQueue[] ws;;) {
2071	>	int ps = plock, m, n;
2072	>	if ((ws = workQueues) == null || (m = ws.length - 1) < 1)
2073	>	return null;
2074	>	for (int j = (m + 1) << 2; ;) {
2075	>	WorkQueue q = ws[(((r + j) << 1) | 1) & m];
2076	>	if (q != null && (n = q.base - q.top) < 0) {
2077	>	if (n < -1)
2078	>	signalWork(q);
2079	>	return q;
2080	>	}
2081	>	else if (--j < 0) {
2082	>	if (plock == ps)
2083	>	return null;
2084	>	break;
2085	>	}
2086	>	}
2087	>	}
2088	>	}
2089
2090	<	// Finish now if all threads terminated; else in some subsequent call
2091	<	if ((workerCounts >>> TOTAL_COUNT_SHIFT) == 0) {
2092	<	advanceRunLevel(TERMINATED);
2093	<	termination.arrive();
2094	<	}
2095	<	return true;
2090	>	/**
2091	>	* Runs tasks until {@code isQuiescent()}. We piggyback on
2092	>	* active count ctl maintenance, but rather than blocking
2093	>	* when tasks cannot be found, we rescan until all others cannot
2094	>	* find tasks either.
2095	>	*/
2096	>	final void helpQuiescePool(WorkQueue w) {
2097	>	for (boolean active = true;;) {
2098	>	ForkJoinTask<?> localTask; // exhaust local queue
2099	>	while ((localTask = w.nextLocalTask()) != null)
2100	>	localTask.doExec();
2101	>	// Similar to loop in scan(), but ignoring submissions
2102	>	WorkQueue q = findNonEmptyStealQueue(w.nextSeed());
2103	>	if (q != null) {
2104	>	ForkJoinTask<?> t; int b;
2105	>	if (!active) {      // re-establish active count
2106	>	long c;
2107	>	active = true;
2108	>	do {} while (!U.compareAndSwapLong
2109	>	(this, CTL, c = ctl, c + AC_UNIT));
2110	>	}
2111	>	if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null)
2112	>	w.runSubtask(t);
2113	>	}
2114	>	else {
2115	>	long c;
2116	>	if (active) {       // decrement active count without queuing
2117	>	active = false;
2118	>	do {} while (!U.compareAndSwapLong
2119	>	(this, CTL, c = ctl, c -= AC_UNIT));
2120	>	}
2121	>	else
2122	>	c = ctl;        // re-increment on exit
2123	>	if ((int)(c >> AC_SHIFT) + (config & SMASK) == 0) {
2124	>	do {} while (!U.compareAndSwapLong
2125	>	(this, CTL, c = ctl, c + AC_UNIT));
2126	>	break;
2127	>	}
2128	>	}
2129	>	}
2130		}
2131
2132		/**
2133	<	* Actions on transition to TERMINATING
2134	<	*
2135	<	* Runs up to four passes through workers: (0) shutting down each
2136	<	* quietly (without waking up if parked) to quickly spread
2137	<	* notifications without unnecessary bouncing around event queues
2138	<	* etc (1) wake up and help cancel tasks (2) interrupt (3) mop up
2139	<	* races with interrupted workers
2140	<	*/
2141	<	private void startTerminating() {
2142	<	cancelSubmissions();
2143	<	for (int passes = 0; passes < 4 && workerCounts != 0; ++passes) {
2144	<	advanceEventCount();
2145	<	eventWaiters = 0L; // clobber lists
2146	<	spareWaiters = 0;
2147	<	ForkJoinWorkerThread[] ws = workers;
2148	<	int n = ws.length;
2149	<	for (int i = 0; i < n; ++i) {
2150	<	ForkJoinWorkerThread w = ws[i];
2151	<	if (w != null) {
2152	<	w.shutdown(true);
2153	<	if (passes > 0 && !w.isTerminated()) {
2154	<	w.cancelTasks();
2155	<	LockSupport.unpark(w);
2156	<	if (passes > 1) {
2157	<	try {
2158	<	w.interrupt();
2159	<	} catch (SecurityException ignore) {
2133	>	* Gets and removes a local or stolen task for the given worker.
2134	>	*
2135	>	* @return a task, if available
2136	>	*/
2137	>	final ForkJoinTask<?> nextTaskFor(WorkQueue w) {
2138	>	for (ForkJoinTask<?> t;;) {
2139	>	WorkQueue q; int b;
2140	>	if ((t = w.nextLocalTask()) != null)
2141	>	return t;
2142	>	if ((q = findNonEmptyStealQueue(w.nextSeed())) == null)
2143	>	return null;
2144	>	if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null)
2145	>	return t;
2146	>	}
2147	>	}
2148	>
2149	>	/**
2150	>	* Returns a cheap heuristic guide for task partitioning when
2151	>	* programmers, frameworks, tools, or languages have little or no
2152	>	* idea about task granularity.  In essence by offering this
2153	>	* method, we ask users only about tradeoffs in overhead vs
2154	>	* expected throughput and its variance, rather than how finely to
2155	>	* partition tasks.
2156	>	*
2157	>	* In a steady state strict (tree-structured) computation, each
2158	>	* thread makes available for stealing enough tasks for other
2159	>	* threads to remain active. Inductively, if all threads play by
2160	>	* the same rules, each thread should make available only a
2161	>	* constant number of tasks.
2162	>	*
2163	>	* The minimum useful constant is just 1. But using a value of 1
2164	>	* would require immediate replenishment upon each steal to
2165	>	* maintain enough tasks, which is infeasible.  Further,
2166	>	* partitionings/granularities of offered tasks should minimize
2167	>	* steal rates, which in general means that threads nearer the top
2168	>	* of computation tree should generate more than those nearer the
2169	>	* bottom. In perfect steady state, each thread is at
2170	>	* approximately the same level of computation tree. However,
2171	>	* producing extra tasks amortizes the uncertainty of progress and
2172	>	* diffusion assumptions.
2173	>	*
2174	>	* So, users will want to use values larger, but not much larger
2175	>	* than 1 to both smooth over transient shortages and hedge
2176	>	* against uneven progress; as traded off against the cost of
2177	>	* extra task overhead. We leave the user to pick a threshold
2178	>	* value to compare with the results of this call to guide
2179	>	* decisions, but recommend values such as 3.
2180	>	*
2181	>	* When all threads are active, it is on average OK to estimate
2182	>	* surplus strictly locally. In steady-state, if one thread is
2183	>	* maintaining say 2 surplus tasks, then so are others. So we can
2184	>	* just use estimated queue length.  However, this strategy alone
2185	>	* leads to serious mis-estimates in some non-steady-state
2186	>	* conditions (ramp-up, ramp-down, other stalls). We can detect
2187	>	* many of these by further considering the number of "idle"
2188	>	* threads, that are known to have zero queued tasks, so
2189	>	* compensate by a factor of (#idle/#active) threads.
2190	>	*
2191	>	* Note: The approximation of #busy workers as #active workers is
2192	>	* not very good under current signalling scheme, and should be
2193	>	* improved.
2194	>	*/
2195	>	static int getSurplusQueuedTaskCount() {
2196	>	Thread t; ForkJoinWorkerThread wt; ForkJoinPool pool; WorkQueue q;
2197	>	if (((t = Thread.currentThread()) instanceof ForkJoinWorkerThread)) {
2198	>	int p = (pool = (wt = (ForkJoinWorkerThread)t).pool).config & SMASK;
2199	>	int n = (q = wt.workQueue).top - q.base;
2200	>	int a = (int)(pool.ctl >> AC_SHIFT) + p;
2201	>	return n - (a > (p >>>= 1) ? 0 :
2202	>	a > (p >>>= 1) ? 1 :
2203	>	a > (p >>>= 1) ? 2 :
2204	>	a > (p >>>= 1) ? 4 :
2205	>	8);
2206	>	}
2207	>	return 0;
2208	>	}
2209	>
2210	>	//  Termination
2211	>
2212	>	/**
2213	>	* Possibly initiates and/or completes termination.  The caller
2214	>	* triggering termination runs three passes through workQueues:
2215	>	* (0) Setting termination status, followed by wakeups of queued
2216	>	* workers; (1) cancelling all tasks; (2) interrupting lagging
2217	>	* threads (likely in external tasks, but possibly also blocked in
2218	>	* joins).  Each pass repeats previous steps because of potential
2219	>	* lagging thread creation.
2220	>	*
2221	>	* @param now if true, unconditionally terminate, else only
2222	>	* if no work and no active workers
2223	>	* @param enable if true, enable shutdown when next possible
2224	>	* @return true if now terminating or terminated
2225	>	*/
2226	>	private boolean tryTerminate(boolean now, boolean enable) {
2227	>	if (this == commonPool)                     // cannot shut down
2228	>	return false;
2229	>	for (long c;;) {
2230	>	if (((c = ctl) & STOP_BIT) != 0) {      // already terminating
2231	>	if ((short)(c >>> TC_SHIFT) == -(config & SMASK)) {
2232	>	synchronized (this) {
2233	>	notifyAll();                // signal when 0 workers
2234	>	}
2235	>	}
2236	>	return true;
2237	>	}
2238	>	if (plock >= 0) {                       // not yet enabled
2239	>	int ps;
2240	>	if (!enable)
2241	>	return false;
2242	>	if (((ps = plock) & PL_LOCK) != 0 ||
2243	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
2244	>	ps = acquirePlock();
2245	>	int nps = SHUTDOWN;
2246	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
2247	>	releasePlock(nps);
2248	>	}
2249	>	if (!now) {                             // check if idle & no tasks
2250	>	if ((int)(c >> AC_SHIFT) != -(config & SMASK) ||
2251	>	hasQueuedSubmissions())
2252	>	return false;
2253	>	// Check for unqueued inactive workers. One pass suffices.
2254	>	WorkQueue[] ws = workQueues; WorkQueue w;
2255	>	if (ws != null) {
2256	>	for (int i = 1; i < ws.length; i += 2) {
2257	>	if ((w = ws[i]) != null && w.eventCount >= 0)
2258	>	return false;
2259	>	}
2260	>	}
2261	>	}
2262	>	if (U.compareAndSwapLong(this, CTL, c, c | STOP_BIT)) {
2263	>	for (int pass = 0; pass < 3; ++pass) {
2264	>	WorkQueue[] ws = workQueues;
2265	>	if (ws != null) {
2266	>	WorkQueue w;
2267	>	int n = ws.length;
2268	>	for (int i = 0; i < n; ++i) {
2269	>	if ((w = ws[i]) != null) {
2270	>	w.qlock = -1;
2271	>	if (pass > 0) {
2272	>	w.cancelAll();
2273	>	if (pass > 1)
2274	>	w.interruptOwner();
2275	>	}
2276	>	}
2277	>	}
2278	>	// Wake up workers parked on event queue
2279	>	int i, e; long cc; Thread p;
2280	>	while ((e = (int)(cc = ctl) & E_MASK) != 0 &&
2281	>	(i = e & SMASK) < n &&
2282	>	(w = ws[i]) != null) {
2283	>	long nc = ((long)(w.nextWait & E_MASK) |
2284	>	((cc + AC_UNIT) & AC_MASK) |
2285	>	(cc & (TC_MASK|STOP_BIT)));
2286	>	if (w.eventCount == (e | INT_SIGN) &&
2287	>	U.compareAndSwapLong(this, CTL, cc, nc)) {
2288	>	w.eventCount = (e + E_SEQ) & E_MASK;
2289	>	w.qlock = -1;
2290	>	if ((p = w.parker) != null)
2291	>	U.unpark(p);
2292		}
2293		}
2294		}
#	Line 1105 | Line 2297 | public class ForkJoinPool extends Abstra
2297		}
2298		}
2299
2300	+	// external operations on common pool
2301	+
2302		/**
2303	<	* Clear out and cancel submissions, ignoring exceptions
2303	>	* Returns common pool queue for a thread that has submitted at
2304	>	* least one task.
2305		*/
2306	<	private void cancelSubmissions() {
2307	<	ForkJoinTask<?> task;
2308	<	while ((task = submissionQueue.poll()) != null) {
2309	<	try {
2310	<	task.cancel(false);
2311	<	} catch (Throwable ignore) {
2306	>	static WorkQueue commonSubmitterQueue() {
2307	>	ForkJoinPool p; WorkQueue[] ws; int m; Submitter z;
2308	>	return ((z = submitters.get()) != null &&
2309	>	(p = commonPool) != null &&
2310	>	(ws = p.workQueues) != null &&
2311	>	(m = ws.length - 1) >= 0) ?
2312	>	ws[m & z.seed & SQMASK] : null;
2313	>	}
2314	>
2315	>	/**
2316	>	* Tries to pop the given task from submitter's queue in common pool.
2317	>	*/
2318	>	static boolean tryExternalUnpush(ForkJoinTask<?> t) {
2319	>	ForkJoinPool p; WorkQueue[] ws; WorkQueue q; Submitter z;
2320	>	ForkJoinTask<?>[] a;  int m, s;
2321	>	if (t != null &&
2322	>	(z = submitters.get()) != null &&
2323	>	(p = commonPool) != null &&
2324	>	(ws = p.workQueues) != null &&
2325	>	(m = ws.length - 1) >= 0 &&
2326	>	(q = ws[m & z.seed & SQMASK]) != null &&
2327	>	(s = q.top) != q.base &&
2328	>	(a = q.array) != null) {
2329	>	long j = (((a.length - 1) & (s - 1)) << ASHIFT) + ABASE;
2330	>	if (U.getObject(a, j) == t &&
2331	>	U.compareAndSwapInt(q, QLOCK, 0, 1)) {
2332	>	if (q.array == a && q.top == s && // recheck
2333	>	U.compareAndSwapObject(a, j, t, null)) {
2334	>	q.top = s - 1;
2335	>	q.qlock = 0;
2336	>	return true;
2337	>	}
2338	>	q.qlock = 0;
2339		}
2340		}
2341	+	return false;
2342		}
2343
1121	–	// misc support for ForkJoinWorkerThread
1122	–
2344		/**
2345	<	* Returns pool number
2346	<	*/
2347	<	final int getPoolNumber() {
2348	<	return poolNumber;
2345	>	* Tries to pop and run local tasks within the same computation
2346	>	* as the given root. On failure, tries to help complete from
2347	>	* other queues via helpComplete.
2348	>	*/
2349	>	private void externalHelpComplete(WorkQueue q, ForkJoinTask<?> root) {
2350	>	ForkJoinTask<?>[] a; int m;
2351	>	if (q != null && (a = q.array) != null && (m = (a.length - 1)) >= 0 &&
2352	>	root != null && root.status >= 0) {
2353	>	for (;;) {
2354	>	int s, u; Object o; CountedCompleter<?> task = null;
2355	>	if ((s = q.top) - q.base > 0) {
2356	>	long j = ((m & (s - 1)) << ASHIFT) + ABASE;
2357	>	if ((o = U.getObject(a, j)) != null &&
2358	>	(o instanceof CountedCompleter)) {
2359	>	CountedCompleter<?> t = (CountedCompleter<?>)o, r = t;
2360	>	do {
2361	>	if (r == root) {
2362	>	if (U.compareAndSwapInt(q, QLOCK, 0, 1)) {
2363	>	if (q.array == a && q.top == s &&
2364	>	U.compareAndSwapObject(a, j, t, null)) {
2365	>	q.top = s - 1;
2366	>	task = t;
2367	>	}
2368	>	q.qlock = 0;
2369	>	}
2370	>	break;
2371	>	}
2372	>	} while ((r = r.completer) != null);
2373	>	}
2374	>	}
2375	>	if (task != null)
2376	>	task.doExec();
2377	>	if (root.status < 0 ||
2378	>	(u = (int)(ctl >>> 32)) >= 0 || (u >> UAC_SHIFT) >= 0)
2379	>	break;
2380	>	if (task == null) {
2381	>	helpSignal(root, q.poolIndex, false);
2382	>	if (root.status >= 0)
2383	>	helpComplete(root, SHARED_QUEUE);
2384	>	break;
2385	>	}
2386	>	}
2387	>	}
2388		}
2389
2390		/**
2391	<	* Tries to accumulates steal count from a worker, clearing
2392	<	* the worker's value.
1133	<	*
1134	<	* @return true if worker steal count now zero
2391	>	* Tries to help execute or signal availability of the given task
2392	>	* from submitter's queue in common pool.
2393		*/
2394	<	final boolean tryAccumulateStealCount(ForkJoinWorkerThread w) {
2395	<	int sc = w.stealCount;
2396	<	long c = stealCount;
2397	<	// CAS even if zero, for fence effects
2398	<	if (UNSAFE.compareAndSwapLong(this, stealCountOffset, c, c + sc)) {
2399	<	if (sc != 0)
2400	<	w.stealCount = 0;
2401	<	return true;
2394	>	static void externalHelpJoin(ForkJoinTask<?> t) {
2395	>	// Some hard-to-avoid overlap with tryExternalUnpush
2396	>	ForkJoinPool p; WorkQueue[] ws; WorkQueue q, w; Submitter z;
2397	>	ForkJoinTask<?>[] a;  int m, s, n;
2398	>	if (t != null &&
2399	>	(z = submitters.get()) != null &&
2400	>	(p = commonPool) != null &&
2401	>	(ws = p.workQueues) != null &&
2402	>	(m = ws.length - 1) >= 0 &&
2403	>	(q = ws[m & z.seed & SQMASK]) != null &&
2404	>	(a = q.array) != null) {
2405	>	int am = a.length - 1;
2406	>	if ((s = q.top) != q.base) {
2407	>	long j = ((am & (s - 1)) << ASHIFT) + ABASE;
2408	>	if (U.getObject(a, j) == t &&
2409	>	U.compareAndSwapInt(q, QLOCK, 0, 1)) {
2410	>	if (q.array == a && q.top == s &&
2411	>	U.compareAndSwapObject(a, j, t, null)) {
2412	>	q.top = s - 1;
2413	>	q.qlock = 0;
2414	>	t.doExec();
2415	>	}
2416	>	else
2417	>	q.qlock = 0;
2418	>	}
2419	>	}
2420	>	if (t.status >= 0) {
2421	>	if (t instanceof CountedCompleter)
2422	>	p.externalHelpComplete(q, t);
2423	>	else
2424	>	p.helpSignal(t, q.poolIndex, false);
2425	>	}
2426		}
1145	–	return sc == 0;
2427		}
2428
2429		/**
2430	<	* Returns the approximate (non-atomic) number of idle threads per
1150	<	* active thread.
2430	>	* Restricted version of helpQuiescePool for external callers
2431		*/
2432	<	final int idlePerActive() {
2433	<	int pc = parallelism; // use parallelism, not rc
2434	<	int ac = runState;    // no mask -- artifically boosts during shutdown
2435	<	// Use exact results for small values, saturate past 4
2436	<	return pc <= ac? 0 : pc >>> 1 <= ac? 1 : pc >>> 2 <= ac? 3 : pc >>> 3;
2432	>	static void externalHelpQuiescePool() {
2433	>	ForkJoinPool p; ForkJoinTask<?> t; WorkQueue q; int b;
2434	>	if ((p = commonPool) != null &&
2435	>	(q = p.findNonEmptyStealQueue(1)) != null &&
2436	>	(b = q.base) - q.top < 0 &&
2437	>	(t = q.pollAt(b)) != null)
2438	>	t.doExec();
2439		}
2440
2441	<	// Public and protected methods
2441	>	// Exported methods
2442
2443		// Constructors
2444
#	Line 1203 | Line 2485 | public class ForkJoinPool extends Abstra
2485		* use {@link #defaultForkJoinWorkerThreadFactory}.
2486		* @param handler the handler for internal worker threads that
2487		* terminate due to unrecoverable errors encountered while executing
2488	<	* tasks. For default value, use <code>null</code>.
2488	>	* tasks. For default value, use {@code null}.
2489		* @param asyncMode if true,
2490		* establishes local first-in-first-out scheduling mode for forked
2491		* tasks that are never joined. This mode may be more appropriate
2492		* than default locally stack-based mode in applications in which
2493		* worker threads only process event-style asynchronous tasks.
2494	<	* For default value, use <code>false</code>.
2494	>	* For default value, use {@code false}.
2495		* @throws IllegalArgumentException if parallelism less than or
2496		*         equal to zero, or greater than implementation limit
2497		* @throws NullPointerException if the factory is null
#	Line 1225 | Line 2507 | public class ForkJoinPool extends Abstra
2507		checkPermission();
2508		if (factory == null)
2509		throw new NullPointerException();
2510	<	if (parallelism <= 0 || parallelism > MAX_WORKERS)
2510	>	if (parallelism <= 0 || parallelism > MAX_CAP)
2511		throw new IllegalArgumentException();
1230	–	this.parallelism = parallelism;
2512		this.factory = factory;
2513		this.ueh = handler;
2514	<	this.locallyFifo = asyncMode;
2515	<	int arraySize = initialArraySizeFor(parallelism);
2516	<	this.workers = new ForkJoinWorkerThread[arraySize];
2517	<	this.submissionQueue = new LinkedTransferQueue<ForkJoinTask<?>>();
2518	<	this.workerLock = new ReentrantLock();
2519	<	this.termination = new Phaser(1);
2520	<	this.poolNumber = poolNumberGenerator.incrementAndGet();
2521	<	this.trimTime = System.nanoTime();
2514	>	this.config = parallelism | (asyncMode ? (FIFO_QUEUE << 16) : 0);
2515	>	long np = (long)(-parallelism); // offset ctl counts
2516	>	this.ctl = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK);
2517	>	int pn = nextPoolId();
2518	>	StringBuilder sb = new StringBuilder("ForkJoinPool-");
2519	>	sb.append(Integer.toString(pn));
2520	>	sb.append("-worker-");
2521	>	this.workerNamePrefix = sb.toString();
2522		}
2523
2524		/**
2525	<	* Returns initial power of two size for workers array.
2526	<	* @param pc the initial parallelism level
2527	<	*/
2528	<	private static int initialArraySizeFor(int pc) {
2529	<	// See Hackers Delight, sec 3.2. We know MAX_WORKERS < (1 >>> 16)
2530	<	int size = pc < MAX_WORKERS ? pc + 1 : MAX_WORKERS;
2531	<	size |= size >>> 1;
2532	<	size |= size >>> 2;
2533	<	size |= size >>> 4;
2534	<	size |= size >>> 8;
2535	<	return size + 1;
2525	>	* Constructor for common pool, suitable only for static initialization.
2526	>	* Basically the same as above, but uses smallest possible initial footprint.
2527	>	*/
2528	>	ForkJoinPool(int parallelism, long ctl,
2529	>	ForkJoinWorkerThreadFactory factory,
2530	>	Thread.UncaughtExceptionHandler handler) {
2531	>	this.config = parallelism;
2532	>	this.ctl = ctl;
2533	>	this.factory = factory;
2534	>	this.ueh = handler;
2535	>	this.workerNamePrefix = "ForkJoinPool.commonPool-worker-";
2536		}
2537
1257	–	// Execution methods
1258	–
2538		/**
2539	<	* Common code for execute, invoke and submit
2539	>	* Returns the common pool instance.
2540	>	*
2541	>	* @return the common pool instance
2542		*/
2543	<	private <T> void doSubmit(ForkJoinTask<T> task) {
2544	<	if (task == null)
2545	<	throw new NullPointerException();
1265	<	if (runState >= SHUTDOWN)
1266	<	throw new RejectedExecutionException();
1267	<	submissionQueue.offer(task);
1268	<	advanceEventCount();
1269	<	helpMaintainParallelism();         // start or wake up workers
2543	>	public static ForkJoinPool commonPool() {
2544	>	// assert commonPool != null : "static init error";
2545	>	return commonPool;
2546		}
2547
2548	+	// Execution methods
2549	+
2550		/**
2551		* Performs the given task, returning its result upon completion.
2552	<	* If the caller is already engaged in a fork/join computation in
2553	<	* the current pool, this method is equivalent in effect to
2554	<	* {@link ForkJoinTask#invoke}.
2552	>	* If the computation encounters an unchecked Exception or Error,
2553	>	* it is rethrown as the outcome of this invocation.  Rethrown
2554	>	* exceptions behave in the same way as regular exceptions, but,
2555	>	* when possible, contain stack traces (as displayed for example
2556	>	* using {@code ex.printStackTrace()}) of both the current thread
2557	>	* as well as the thread actually encountering the exception;
2558	>	* minimally only the latter.
2559		*
2560		* @param task the task
2561		* @return the task's result
#	Line 1282 | Line 2564 | public class ForkJoinPool extends Abstra
2564		*         scheduled for execution
2565		*/
2566		public <T> T invoke(ForkJoinTask<T> task) {
2567	<	doSubmit(task);
2567	>	if (task == null)
2568	>	throw new NullPointerException();
2569	>	externalPush(task);
2570		return task.join();
2571		}
2572
2573		/**
2574		* Arranges for (asynchronous) execution of the given task.
1291	–	* If the caller is already engaged in a fork/join computation in
1292	–	* the current pool, this method is equivalent in effect to
1293	–	* {@link ForkJoinTask#fork}.
2575		*
2576		* @param task the task
2577		* @throws NullPointerException if the task is null
#	Line 1298 | Line 2579 | public class ForkJoinPool extends Abstra
2579		*         scheduled for execution
2580		*/
2581		public void execute(ForkJoinTask<?> task) {
2582	<	doSubmit(task);
2582	>	if (task == null)
2583	>	throw new NullPointerException();
2584	>	externalPush(task);
2585		}
2586
2587		// AbstractExecutorService methods
#	Line 1309 | Line 2592 | public class ForkJoinPool extends Abstra
2592		*         scheduled for execution
2593		*/
2594		public void execute(Runnable task) {
2595	+	if (task == null)
2596	+	throw new NullPointerException();
2597		ForkJoinTask<?> job;
2598		if (task instanceof ForkJoinTask<?>) // avoid re-wrap
2599		job = (ForkJoinTask<?>) task;
2600		else
2601	<	job = ForkJoinTask.adapt(task, null);
2602	<	doSubmit(job);
2601	>	job = new ForkJoinTask.AdaptedRunnableAction(task);
2602	>	externalPush(job);
2603		}
2604
2605		/**
2606		* Submits a ForkJoinTask for execution.
1322	–	* If the caller is already engaged in a fork/join computation in
1323	–	* the current pool, this method is equivalent in effect to
1324	–	* {@link ForkJoinTask#fork}.
2607		*
2608		* @param task the task to submit
2609		* @return the task
#	Line 1330 | Line 2612 | public class ForkJoinPool extends Abstra
2612		*         scheduled for execution
2613		*/
2614		public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) {
2615	<	doSubmit(task);
2615	>	if (task == null)
2616	>	throw new NullPointerException();
2617	>	externalPush(task);
2618		return task;
2619		}
2620
#	Line 1340 | Line 2624 | public class ForkJoinPool extends Abstra
2624		*         scheduled for execution
2625		*/
2626		public <T> ForkJoinTask<T> submit(Callable<T> task) {
2627	<	ForkJoinTask<T> job = ForkJoinTask.adapt(task);
2628	<	doSubmit(job);
2627	>	ForkJoinTask<T> job = new ForkJoinTask.AdaptedCallable<T>(task);
2628	>	externalPush(job);
2629		return job;
2630		}
2631
#	Line 1351 | Line 2635 | public class ForkJoinPool extends Abstra
2635		*         scheduled for execution
2636		*/
2637		public <T> ForkJoinTask<T> submit(Runnable task, T result) {
2638	<	ForkJoinTask<T> job = ForkJoinTask.adapt(task, result);
2639	<	doSubmit(job);
2638	>	ForkJoinTask<T> job = new ForkJoinTask.AdaptedRunnable<T>(task, result);
2639	>	externalPush(job);
2640		return job;
2641		}
2642
#	Line 1362 | Line 2646 | public class ForkJoinPool extends Abstra
2646		*         scheduled for execution
2647		*/
2648		public ForkJoinTask<?> submit(Runnable task) {
2649	+	if (task == null)
2650	+	throw new NullPointerException();
2651		ForkJoinTask<?> job;
2652		if (task instanceof ForkJoinTask<?>) // avoid re-wrap
2653		job = (ForkJoinTask<?>) task;
2654		else
2655	<	job = ForkJoinTask.adapt(task, null);
2656	<	doSubmit(job);
2655	>	job = new ForkJoinTask.AdaptedRunnableAction(task);
2656	>	externalPush(job);
2657		return job;
2658		}
2659
#	Line 1376 | Line 2662 | public class ForkJoinPool extends Abstra
2662		* @throws RejectedExecutionException {@inheritDoc}
2663		*/
2664		public <T> List<Future<T>> invokeAll(Collection<? extends Callable<T>> tasks) {
2665	<	ArrayList<ForkJoinTask<T>> forkJoinTasks =
2666	<	new ArrayList<ForkJoinTask<T>>(tasks.size());
2667	<	for (Callable<T> task : tasks)
2668	<	forkJoinTasks.add(ForkJoinTask.adapt(task));
2669	<	invoke(new InvokeAll<T>(forkJoinTasks));
2670	<
2665	>	// In previous versions of this class, this method constructed
2666	>	// a task to run ForkJoinTask.invokeAll, but now external
2667	>	// invocation of multiple tasks is at least as efficient.
2668	>	List<ForkJoinTask<T>> fs = new ArrayList<ForkJoinTask<T>>(tasks.size());
2669	>	// Workaround needed because method wasn't declared with
2670	>	// wildcards in return type but should have been.
2671		@SuppressWarnings({"unchecked", "rawtypes"})
2672	<	List<Future<T>> futures = (List<Future<T>>) (List) forkJoinTasks;
1387	<	return futures;
1388	<	}
2672	>	List<Future<T>> futures = (List<Future<T>>) (List) fs;
2673
2674	<	static final class InvokeAll<T> extends RecursiveAction {
2675	<	final ArrayList<ForkJoinTask<T>> tasks;
2676	<	InvokeAll(ArrayList<ForkJoinTask<T>> tasks) { this.tasks = tasks; }
2677	<	public void compute() {
2678	<	try { invokeAll(tasks); }
2679	<	catch (Exception ignore) {}
2674	>	boolean done = false;
2675	>	try {
2676	>	for (Callable<T> t : tasks) {
2677	>	ForkJoinTask<T> f = new ForkJoinTask.AdaptedCallable<T>(t);
2678	>	externalPush(f);
2679	>	fs.add(f);
2680	>	}
2681	>	for (ForkJoinTask<T> f : fs)
2682	>	f.quietlyJoin();
2683	>	done = true;
2684	>	return futures;
2685	>	} finally {
2686	>	if (!done)
2687	>	for (ForkJoinTask<T> f : fs)
2688	>	f.cancel(false);
2689		}
1397	–	private static final long serialVersionUID = -7914297376763021607L;
2690		}
2691
2692		/**
#	Line 1422 | Line 2714 | public class ForkJoinPool extends Abstra
2714		* @return the targeted parallelism level of this pool
2715		*/
2716		public int getParallelism() {
2717	<	return parallelism;
2717	>	return config & SMASK;
2718	>	}
2719	>
2720	>	/**
2721	>	* Returns the targeted parallelism level of the common pool.
2722	>	*
2723	>	* @return the targeted parallelism level of the common pool
2724	>	*/
2725	>	public static int getCommonPoolParallelism() {
2726	>	return commonPoolParallelism;
2727		}
2728
2729		/**
2730		* Returns the number of worker threads that have started but not
2731	<	* yet terminated.  This result returned by this method may differ
2731	>	* yet terminated.  The result returned by this method may differ
2732		* from {@link #getParallelism} when threads are created to
2733		* maintain parallelism when others are cooperatively blocked.
2734		*
2735		* @return the number of worker threads
2736		*/
2737		public int getPoolSize() {
2738	<	return workerCounts >>> TOTAL_COUNT_SHIFT;
2738	>	return (config & SMASK) + (short)(ctl >>> TC_SHIFT);
2739		}
2740
2741		/**
#	Line 1444 | Line 2745 | public class ForkJoinPool extends Abstra
2745		* @return {@code true} if this pool uses async mode
2746		*/
2747		public boolean getAsyncMode() {
2748	<	return locallyFifo;
2748	>	return (config >>> 16) == FIFO_QUEUE;
2749		}
2750
2751		/**
#	Line 1456 | Line 2757 | public class ForkJoinPool extends Abstra
2757		* @return the number of worker threads
2758		*/
2759		public int getRunningThreadCount() {
2760	<	return workerCounts & RUNNING_COUNT_MASK;
2760	>	int rc = 0;
2761	>	WorkQueue[] ws; WorkQueue w;
2762	>	if ((ws = workQueues) != null) {
2763	>	for (int i = 1; i < ws.length; i += 2) {
2764	>	if ((w = ws[i]) != null && w.isApparentlyUnblocked())
2765	>	++rc;
2766	>	}
2767	>	}
2768	>	return rc;
2769		}
2770
2771		/**
#	Line 1467 | Line 2776 | public class ForkJoinPool extends Abstra
2776		* @return the number of active threads
2777		*/
2778		public int getActiveThreadCount() {
2779	<	return runState & ACTIVE_COUNT_MASK;
2779	>	int r = (config & SMASK) + (int)(ctl >> AC_SHIFT);
2780	>	return (r <= 0) ? 0 : r; // suppress momentarily negative values
2781		}
2782
2783		/**
#	Line 1482 | Line 2792 | public class ForkJoinPool extends Abstra
2792		* @return {@code true} if all threads are currently idle
2793		*/
2794		public boolean isQuiescent() {
2795	<	return (runState & ACTIVE_COUNT_MASK) == 0;
2795	>	return (int)(ctl >> AC_SHIFT) + (config & SMASK) == 0;
2796		}
2797
2798		/**
#	Line 1497 | Line 2807 | public class ForkJoinPool extends Abstra
2807		* @return the number of steals
2808		*/
2809		public long getStealCount() {
2810	<	return stealCount;
2810	>	long count = stealCount;
2811	>	WorkQueue[] ws; WorkQueue w;
2812	>	if ((ws = workQueues) != null) {
2813	>	for (int i = 1; i < ws.length; i += 2) {
2814	>	if ((w = ws[i]) != null)
2815	>	count += w.nsteals;
2816	>	}
2817	>	}
2818	>	return count;
2819		}
2820
2821		/**
#	Line 1512 | Line 2830 | public class ForkJoinPool extends Abstra
2830		*/
2831		public long getQueuedTaskCount() {
2832		long count = 0;
2833	<	ForkJoinWorkerThread[] ws = workers;
2834	<	int n = ws.length;
2835	<	for (int i = 0; i < n; ++i) {
2836	<	ForkJoinWorkerThread w = ws[i];
2837	<	if (w != null)
2838	<	count += w.getQueueSize();
2833	>	WorkQueue[] ws; WorkQueue w;
2834	>	if ((ws = workQueues) != null) {
2835	>	for (int i = 1; i < ws.length; i += 2) {
2836	>	if ((w = ws[i]) != null)
2837	>	count += w.queueSize();
2838	>	}
2839		}
2840		return count;
2841		}
2842
2843		/**
2844		* Returns an estimate of the number of tasks submitted to this
2845	<	* pool that have not yet begun executing.  This method takes time
2846	<	* proportional to the number of submissions.
2845	>	* pool that have not yet begun executing.  This method may take
2846	>	* time proportional to the number of submissions.
2847		*
2848		* @return the number of queued submissions
2849		*/
2850		public int getQueuedSubmissionCount() {
2851	<	return submissionQueue.size();
2851	>	int count = 0;
2852	>	WorkQueue[] ws; WorkQueue w;
2853	>	if ((ws = workQueues) != null) {
2854	>	for (int i = 0; i < ws.length; i += 2) {
2855	>	if ((w = ws[i]) != null)
2856	>	count += w.queueSize();
2857	>	}
2858	>	}
2859	>	return count;
2860		}
2861
2862		/**
#	Line 1540 | Line 2866 | public class ForkJoinPool extends Abstra
2866		* @return {@code true} if there are any queued submissions
2867		*/
2868		public boolean hasQueuedSubmissions() {
2869	<	return !submissionQueue.isEmpty();
2869	>	WorkQueue[] ws; WorkQueue w;
2870	>	if ((ws = workQueues) != null) {
2871	>	for (int i = 0; i < ws.length; i += 2) {
2872	>	if ((w = ws[i]) != null && !w.isEmpty())
2873	>	return true;
2874	>	}
2875	>	}
2876	>	return false;
2877		}
2878
2879		/**
#	Line 1551 | Line 2884 | public class ForkJoinPool extends Abstra
2884		* @return the next submission, or {@code null} if none
2885		*/
2886		protected ForkJoinTask<?> pollSubmission() {
2887	<	return submissionQueue.poll();
2887	>	WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2888	>	if ((ws = workQueues) != null) {
2889	>	for (int i = 0; i < ws.length; i += 2) {
2890	>	if ((w = ws[i]) != null && (t = w.poll()) != null)
2891	>	return t;
2892	>	}
2893	>	}
2894	>	return null;
2895		}
2896
2897		/**
#	Line 1572 | Line 2912 | public class ForkJoinPool extends Abstra
2912		* @return the number of elements transferred
2913		*/
2914		protected int drainTasksTo(Collection<? super ForkJoinTask<?>> c) {
2915	<	int count = submissionQueue.drainTo(c);
2916	<	ForkJoinWorkerThread[] ws = workers;
2917	<	int n = ws.length;
2918	<	for (int i = 0; i < n; ++i) {
2919	<	ForkJoinWorkerThread w = ws[i];
2920	<	if (w != null)
2921	<	count += w.drainTasksTo(c);
2915	>	int count = 0;
2916	>	WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2917	>	if ((ws = workQueues) != null) {
2918	>	for (int i = 0; i < ws.length; ++i) {
2919	>	if ((w = ws[i]) != null) {
2920	>	while ((t = w.poll()) != null) {
2921	>	c.add(t);
2922	>	++count;
2923	>	}
2924	>	}
2925	>	}
2926		}
2927		return count;
2928		}
#	Line 1591 | Line 2935 | public class ForkJoinPool extends Abstra
2935		* @return a string identifying this pool, as well as its state
2936		*/
2937		public String toString() {
2938	<	long st = getStealCount();
2939	<	long qt = getQueuedTaskCount();
2940	<	long qs = getQueuedSubmissionCount();
2941	<	int wc = workerCounts;
2942	<	int tc = wc >>> TOTAL_COUNT_SHIFT;
2943	<	int rc = wc & RUNNING_COUNT_MASK;
2944	<	int pc = parallelism;
2945	<	int rs = runState;
2946	<	int ac = rs & ACTIVE_COUNT_MASK;
2938	>	// Use a single pass through workQueues to collect counts
2939	>	long qt = 0L, qs = 0L; int rc = 0;
2940	>	long st = stealCount;
2941	>	long c = ctl;
2942	>	WorkQueue[] ws; WorkQueue w;
2943	>	if ((ws = workQueues) != null) {
2944	>	for (int i = 0; i < ws.length; ++i) {
2945	>	if ((w = ws[i]) != null) {
2946	>	int size = w.queueSize();
2947	>	if ((i & 1) == 0)
2948	>	qs += size;
2949	>	else {
2950	>	qt += size;
2951	>	st += w.nsteals;
2952	>	if (w.isApparentlyUnblocked())
2953	>	++rc;
2954	>	}
2955	>	}
2956	>	}
2957	>	}
2958	>	int pc = (config & SMASK);
2959	>	int tc = pc + (short)(c >>> TC_SHIFT);
2960	>	int ac = pc + (int)(c >> AC_SHIFT);
2961	>	if (ac < 0) // ignore transient negative
2962	>	ac = 0;
2963	>	String level;
2964	>	if ((c & STOP_BIT) != 0)
2965	>	level = (tc == 0) ? "Terminated" : "Terminating";
2966	>	else
2967	>	level = plock < 0 ? "Shutting down" : "Running";
2968		return super.toString() +
2969	<	"[" + runLevelToString(rs) +
2969	>	"[" + level +
2970		", parallelism = " + pc +
2971		", size = " + tc +
2972		", active = " + ac +
#	Line 1612 | Line 2977 | public class ForkJoinPool extends Abstra
2977		"]";
2978		}
2979
1615	–	private static String runLevelToString(int s) {
1616	–	return ((s & TERMINATED) != 0 ? "Terminated" :
1617	–	((s & TERMINATING) != 0 ? "Terminating" :
1618	–	((s & SHUTDOWN) != 0 ? "Shutting down" :
1619	–	"Running")));
1620	–	}
1621	–
2980		/**
2981	<	* Initiates an orderly shutdown in which previously submitted
2982	<	* tasks are executed, but no new tasks will be accepted.
2983	<	* Invocation has no additional effect if already shut down.
2984	<	* Tasks that are in the process of being submitted concurrently
2985	<	* during the course of this method may or may not be rejected.
2981	>	* Possibly initiates an orderly shutdown in which previously
2982	>	* submitted tasks are executed, but no new tasks will be
2983	>	* accepted. Invocation has no effect on execution state if this
2984	>	* is the {@link #commonPool}, and no additional effect if
2985	>	* already shut down.  Tasks that are in the process of being
2986	>	* submitted concurrently during the course of this method may or
2987	>	* may not be rejected.
2988		*
2989		* @throws SecurityException if a security manager exists and
2990		*         the caller is not permitted to modify threads
#	Line 1633 | Line 2993 | public class ForkJoinPool extends Abstra
2993		*/
2994		public void shutdown() {
2995		checkPermission();
2996	<	advanceRunLevel(SHUTDOWN);
1637	<	tryTerminate(false);
2996	>	tryTerminate(false, true);
2997		}
2998
2999		/**
3000	<	* Attempts to cancel and/or stop all tasks, and reject all
3001	<	* subsequently submitted tasks.  Tasks that are in the process of
3002	<	* being submitted or executed concurrently during the course of
3003	<	* this method may or may not be rejected. This method cancels
3004	<	* both existing and unexecuted tasks, in order to permit
3005	<	* termination in the presence of task dependencies. So the method
3006	<	* always returns an empty list (unlike the case for some other
3007	<	* Executors).
3000	>	* Possibly attempts to cancel and/or stop all tasks, and reject
3001	>	* all subsequently submitted tasks.  Invocation has no effect on
3002	>	* execution state if this is the {@link #commonPool}, and no
3003	>	* additional effect if already shut down. Otherwise, tasks that
3004	>	* are in the process of being submitted or executed concurrently
3005	>	* during the course of this method may or may not be
3006	>	* rejected. This method cancels both existing and unexecuted
3007	>	* tasks, in order to permit termination in the presence of task
3008	>	* dependencies. So the method always returns an empty list
3009	>	* (unlike the case for some other Executors).
3010		*
3011		* @return an empty list
3012		* @throws SecurityException if a security manager exists and
#	Line 1655 | Line 3016 | public class ForkJoinPool extends Abstra
3016		*/
3017		public List<Runnable> shutdownNow() {
3018		checkPermission();
3019	<	tryTerminate(true);
3019	>	tryTerminate(true, true);
3020		return Collections.emptyList();
3021		}
3022
#	Line 1665 | Line 3026 | public class ForkJoinPool extends Abstra
3026		* @return {@code true} if all tasks have completed following shut down
3027		*/
3028		public boolean isTerminated() {
3029	<	return runState >= TERMINATED;
3029	>	long c = ctl;
3030	>	return ((c & STOP_BIT) != 0L &&
3031	>	(short)(c >>> TC_SHIFT) == -(config & SMASK));
3032		}
3033
3034		/**
#	Line 1673 | Line 3036 | public class ForkJoinPool extends Abstra
3036		* commenced but not yet completed.  This method may be useful for
3037		* debugging. A return of {@code true} reported a sufficient
3038		* period after shutdown may indicate that submitted tasks have
3039	<	* ignored or suppressed interruption, causing this executor not
3040	<	* to properly terminate.
3039	>	* ignored or suppressed interruption, or are waiting for IO,
3040	>	* causing this executor not to properly terminate. (See the
3041	>	* advisory notes for class {@link ForkJoinTask} stating that
3042	>	* tasks should not normally entail blocking operations.  But if
3043	>	* they do, they must abort them on interrupt.)
3044		*
3045		* @return {@code true} if terminating but not yet terminated
3046		*/
3047		public boolean isTerminating() {
3048	<	return (runState & (TERMINATING|TERMINATED)) == TERMINATING;
3048	>	long c = ctl;
3049	>	return ((c & STOP_BIT) != 0L &&
3050	>	(short)(c >>> TC_SHIFT) != -(config & SMASK));
3051		}
3052
3053		/**
#	Line 1688 | Line 3056 | public class ForkJoinPool extends Abstra
3056		* @return {@code true} if this pool has been shut down
3057		*/
3058		public boolean isShutdown() {
3059	<	return runState >= SHUTDOWN;
3059	>	return plock < 0;
3060		}
3061
3062		/**
3063	<	* Blocks until all tasks have completed execution after a shutdown
3064	<	* request, or the timeout occurs, or the current thread is
3065	<	* interrupted, whichever happens first.
3063	>	* Blocks until all tasks have completed execution after a
3064	>	* shutdown request, or the timeout occurs, or the current thread
3065	>	* is interrupted, whichever happens first. Note that the {@link
3066	>	* #commonPool()} never terminates until program shutdown so
3067	>	* this method will always time out.
3068		*
3069		* @param timeout the maximum time to wait
3070		* @param unit the time unit of the timeout argument
#	Line 1704 | Line 3074 | public class ForkJoinPool extends Abstra
3074		*/
3075		public boolean awaitTermination(long timeout, TimeUnit unit)
3076		throws InterruptedException {
3077	<	try {
3078	<	return termination.awaitAdvanceInterruptibly(0, timeout, unit) > 0;
3079	<	} catch(TimeoutException ex) {
3080	<	return false;
3077	>	long nanos = unit.toNanos(timeout);
3078	>	if (isTerminated())
3079	>	return true;
3080	>	long startTime = System.nanoTime();
3081	>	boolean terminated = false;
3082	>	synchronized (this) {
3083	>	for (long waitTime = nanos, millis = 0L;;) {
3084	>	if (terminated = isTerminated() ||
3085	>	waitTime <= 0L ||
3086	>	(millis = unit.toMillis(waitTime)) <= 0L)
3087	>	break;
3088	>	wait(millis);
3089	>	waitTime = nanos - (System.nanoTime() - startTime);
3090	>	}
3091		}
3092	+	return terminated;
3093		}
3094
3095		/**
#	Line 1719 | Line 3100 | public class ForkJoinPool extends Abstra
3100		* {@code isReleasable} must return {@code true} if blocking is
3101		* not necessary. Method {@code block} blocks the current thread
3102		* if necessary (perhaps internally invoking {@code isReleasable}
3103	<	* before actually blocking). The unusual methods in this API
3104	<	* accommodate synchronizers that may, but don't usually, block
3105	<	* for long periods. Similarly, they allow more efficient internal
3106	<	* handling of cases in which additional workers may be, but
3107	<	* usually are not, needed to ensure sufficient parallelism.
3108	<	* Toward this end, implementations of method {@code isReleasable}
3109	<	* must be amenable to repeated invocation.
3103	>	* before actually blocking). These actions are performed by any
3104	>	* thread invoking {@link ForkJoinPool#managedBlock}.  The
3105	>	* unusual methods in this API accommodate synchronizers that may,
3106	>	* but don't usually, block for long periods. Similarly, they
3107	>	* allow more efficient internal handling of cases in which
3108	>	* additional workers may be, but usually are not, needed to
3109	>	* ensure sufficient parallelism.  Toward this end,
3110	>	* implementations of method {@code isReleasable} must be amenable
3111	>	* to repeated invocation.
3112		*
3113		* <p>For example, here is a ManagedBlocker based on a
3114		* ReentrantLock:
#	Line 1753 | Line 3136 | public class ForkJoinPool extends Abstra
3136		*   QueueTaker(BlockingQueue<E> q) { this.queue = q; }
3137		*   public boolean block() throws InterruptedException {
3138		*     if (item == null)
3139	<	*       item = queue.take
3139	>	*       item = queue.take();
3140		*     return true;
3141		*   }
3142		*   public boolean isReleasable() {
3143	<	*     return item != null || (item = queue.poll) != null;
3143	>	*     return item != null || (item = queue.poll()) != null;
3144		*   }
3145		*   public E getItem() { // call after pool.managedBlock completes
3146		*     return item;
#	Line 1806 | Line 3189 | public class ForkJoinPool extends Abstra
3189		throws InterruptedException {
3190		Thread t = Thread.currentThread();
3191		if (t instanceof ForkJoinWorkerThread) {
3192	<	ForkJoinWorkerThread w = (ForkJoinWorkerThread) t;
3193	<	w.pool.awaitBlocker(blocker);
3192	>	ForkJoinPool p = ((ForkJoinWorkerThread)t).pool;
3193	>	while (!blocker.isReleasable()) { // variant of helpSignal
3194	>	WorkQueue[] ws; WorkQueue q; int m, u;
3195	>	if ((ws = p.workQueues) != null && (m = ws.length - 1) >= 0) {
3196	>	for (int i = 0; i <= m; ++i) {
3197	>	if (blocker.isReleasable())
3198	>	return;
3199	>	if ((q = ws[i]) != null && q.base - q.top < 0) {
3200	>	p.signalWork(q);
3201	>	if ((u = (int)(p.ctl >>> 32)) >= 0 ||
3202	>	(u >> UAC_SHIFT) >= 0)
3203	>	break;
3204	>	}
3205	>	}
3206	>	}
3207	>	if (p.tryCompensate()) {
3208	>	try {
3209	>	do {} while (!blocker.isReleasable() &&
3210	>	!blocker.block());
3211	>	} finally {
3212	>	p.incrementActiveCount();
3213	>	}
3214	>	break;
3215	>	}
3216	>	}
3217		}
3218		else {
3219	<	do {} while (!blocker.isReleasable() && !blocker.block());
3219	>	do {} while (!blocker.isReleasable() &&
3220	>	!blocker.block());
3221		}
3222		}
3223
#	Line 1819 | Line 3226 | public class ForkJoinPool extends Abstra
3226		// implement RunnableFuture.
3227
3228		protected <T> RunnableFuture<T> newTaskFor(Runnable runnable, T value) {
3229	<	return (RunnableFuture<T>) ForkJoinTask.adapt(runnable, value);
3229	>	return new ForkJoinTask.AdaptedRunnable<T>(runnable, value);
3230		}
3231
3232		protected <T> RunnableFuture<T> newTaskFor(Callable<T> callable) {
3233	<	return (RunnableFuture<T>) ForkJoinTask.adapt(callable);
3233	>	return new ForkJoinTask.AdaptedCallable<T>(callable);
3234		}
3235
3236		// Unsafe mechanics
3237	+	private static final sun.misc.Unsafe U;
3238	+	private static final long CTL;
3239	+	private static final long PARKBLOCKER;
3240	+	private static final int ABASE;
3241	+	private static final int ASHIFT;
3242	+	private static final long STEALCOUNT;
3243	+	private static final long PLOCK;
3244	+	private static final long INDEXSEED;
3245	+	private static final long QLOCK;
3246
3247	<	private static final sun.misc.Unsafe UNSAFE = getUnsafe();
3248	<	private static final long workerCountsOffset =
1833	<	objectFieldOffset("workerCounts", ForkJoinPool.class);
1834	<	private static final long runStateOffset =
1835	<	objectFieldOffset("runState", ForkJoinPool.class);
1836	<	private static final long eventCountOffset =
1837	<	objectFieldOffset("eventCount", ForkJoinPool.class);
1838	<	private static final long eventWaitersOffset =
1839	<	objectFieldOffset("eventWaiters",ForkJoinPool.class);
1840	<	private static final long stealCountOffset =
1841	<	objectFieldOffset("stealCount",ForkJoinPool.class);
1842	<	private static final long spareWaitersOffset =
1843	<	objectFieldOffset("spareWaiters",ForkJoinPool.class);
1844	<
1845	<	private static long objectFieldOffset(String field, Class<?> klazz) {
3247	>	static {
3248	>	int s; // initialize field offsets for CAS etc
3249		try {
3250	<	return UNSAFE.objectFieldOffset(klazz.getDeclaredField(field));
3251	<	} catch (NoSuchFieldException e) {
3252	<	// Convert Exception to corresponding Error
3253	<	NoSuchFieldError error = new NoSuchFieldError(field);
3254	<	error.initCause(e);
3255	<	throw error;
3256	<	}
3250	>	U = getUnsafe();
3251	>	Class<?> k = ForkJoinPool.class;
3252	>	CTL = U.objectFieldOffset
3253	>	(k.getDeclaredField("ctl"));
3254	>	STEALCOUNT = U.objectFieldOffset
3255	>	(k.getDeclaredField("stealCount"));
3256	>	PLOCK = U.objectFieldOffset
3257	>	(k.getDeclaredField("plock"));
3258	>	INDEXSEED = U.objectFieldOffset
3259	>	(k.getDeclaredField("indexSeed"));
3260	>	Class<?> tk = Thread.class;
3261	>	PARKBLOCKER = U.objectFieldOffset
3262	>	(tk.getDeclaredField("parkBlocker"));
3263	>	Class<?> wk = WorkQueue.class;
3264	>	QLOCK = U.objectFieldOffset
3265	>	(wk.getDeclaredField("qlock"));
3266	>	Class<?> ak = ForkJoinTask[].class;
3267	>	ABASE = U.arrayBaseOffset(ak);
3268	>	s = U.arrayIndexScale(ak);
3269	>	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
3270	>	} catch (Exception e) {
3271	>	throw new Error(e);
3272	>	}
3273	>	if ((s & (s-1)) != 0)
3274	>	throw new Error("data type scale not a power of two");
3275	>
3276	>	submitters = new ThreadLocal<Submitter>();
3277	>	ForkJoinWorkerThreadFactory fac = defaultForkJoinWorkerThreadFactory =
3278	>	new DefaultForkJoinWorkerThreadFactory();
3279	>	modifyThreadPermission = new RuntimePermission("modifyThread");
3280	>
3281	>	/*
3282	>	* Establish common pool parameters.  For extra caution,
3283	>	* computations to set up common pool state are here; the
3284	>	* constructor just assigns these values to fields.
3285	>	*/
3286	>
3287	>	int par = 0;
3288	>	Thread.UncaughtExceptionHandler handler = null;
3289	>	try {  // TBD: limit or report ignored exceptions?
3290	>	String pp = System.getProperty
3291	>	("java.util.concurrent.ForkJoinPool.common.parallelism");
3292	>	String hp = System.getProperty
3293	>	("java.util.concurrent.ForkJoinPool.common.exceptionHandler");
3294	>	String fp = System.getProperty
3295	>	("java.util.concurrent.ForkJoinPool.common.threadFactory");
3296	>	if (fp != null)
3297	>	fac = ((ForkJoinWorkerThreadFactory)ClassLoader.
3298	>	getSystemClassLoader().loadClass(fp).newInstance());
3299	>	if (hp != null)
3300	>	handler = ((Thread.UncaughtExceptionHandler)ClassLoader.
3301	>	getSystemClassLoader().loadClass(hp).newInstance());
3302	>	if (pp != null)
3303	>	par = Integer.parseInt(pp);
3304	>	} catch (Exception ignore) {
3305	>	}
3306	>
3307	>	if (par <= 0)
3308	>	par = Runtime.getRuntime().availableProcessors();
3309	>	if (par > MAX_CAP)
3310	>	par = MAX_CAP;
3311	>	commonPoolParallelism = par;
3312	>	long np = (long)(-par); // precompute initial ctl value
3313	>	long ct = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK);
3314	>
3315	>	commonPool = new ForkJoinPool(par, ct, fac, handler);
3316		}
3317
3318		/**
#	Line 1880 | Line 3342 | public class ForkJoinPool extends Abstra
3342		}
3343		}
3344		}
3345	+
3346		}

Diff Legend

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