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 1.68 by jsr166, Wed Sep 1 06:40:12 2010 UTC vs.
Revision 1.139 by dl, Wed Oct 31 12:49:24 2012 UTC

#	Line 1 | Line 1
1		/*
2		* Written by Doug Lea with assistance from members of JCP JSR-166
3		* Expert Group and released to the public domain, as explained at
4	<	* http://creativecommons.org/licenses/publicdomain
4	>	* http://creativecommons.org/publicdomain/zero/1.0/
5		*/
6
7		package jsr166y;
8
9	–	import java.util.concurrent.*;
10	–
9		import java.util.ArrayList;
10		import java.util.Arrays;
11		import java.util.Collection;
12		import java.util.Collections;
13		import java.util.List;
14	<	import java.util.concurrent.locks.LockSupport;
15	<	import java.util.concurrent.locks.ReentrantLock;
14	>	import java.util.Random;
15	>	import java.util.concurrent.AbstractExecutorService;
16	>	import java.util.concurrent.Callable;
17	>	import java.util.concurrent.ExecutorService;
18	>	import java.util.concurrent.Future;
19	>	import java.util.concurrent.RejectedExecutionException;
20	>	import java.util.concurrent.RunnableFuture;
21	>	import java.util.concurrent.TimeUnit;
22		import java.util.concurrent.atomic.AtomicInteger;
23	<	import java.util.concurrent.CountDownLatch;
23	>	import java.util.concurrent.atomic.AtomicLong;
24	>	import java.util.concurrent.locks.AbstractQueuedSynchronizer;
25	>	import java.util.concurrent.locks.Condition;
26
27		/**
28		* An {@link ExecutorService} for running {@link ForkJoinTask}s.
#	Line 27 | Line 33 | import java.util.concurrent.CountDownLat
33		* <p>A {@code ForkJoinPool} differs from other kinds of {@link
34		* ExecutorService} mainly by virtue of employing
35		* <em>work-stealing</em>: all threads in the pool attempt to find and
36	<	* execute subtasks created by other active tasks (eventually blocking
37	<	* waiting for work if none exist). This enables efficient processing
38	<	* when most tasks spawn other subtasks (as do most {@code
39	<	* ForkJoinTask}s). When setting <em>asyncMode</em> to true in
40	<	* constructors, {@code ForkJoinPool}s may also be appropriate for use
41	<	* with event-style tasks that are never joined.
36	>	* execute tasks submitted to the pool and/or created by other active
37	>	* tasks (eventually blocking waiting for work if none exist). This
38	>	* enables efficient processing when most tasks spawn other subtasks
39	>	* (as do most {@code ForkJoinTask}s), as well as when many small
40	>	* tasks are submitted to the pool from external clients.  Especially
41	>	* when setting <em>asyncMode</em> to true in constructors, {@code
42	>	* ForkJoinPool}s may also be appropriate for use with event-style
43	>	* tasks that are never joined.
44	>	*
45	>	* <p>A static {@link #commonPool} is available and appropriate for
46	>	* most applications. The common pool is used by any ForkJoinTask that
47	>	* is not explicitly submitted to a specified pool. Using the common
48	>	* pool normally reduces resource usage (its threads are slowly
49	>	* reclaimed during periods of non-use, and reinstated upon subsequent
50	>	* use).  The common pool is by default constructed with default
51	>	* parameters, but these may be controlled by setting any or all of
52	>	* the three properties {@code
53	>	* java.util.concurrent.ForkJoinPool.common.{parallelism,
54	>	* threadFactory, exceptionHandler}}.
55		*
56	<	* <p>A {@code ForkJoinPool} is constructed with a given target
57	<	* parallelism level; by default, equal to the number of available
58	<	* processors. The pool attempts to maintain enough active (or
59	<	* available) threads by dynamically adding, suspending, or resuming
60	<	* internal worker threads, even if some tasks are stalled waiting to
61	<	* join others. However, no such adjustments are guaranteed in the
62	<	* face of blocked IO or other unmanaged synchronization. The nested
63	<	* {@link ManagedBlocker} interface enables extension of the kinds of
56	>	* <p>For applications that require separate or custom pools, a {@code
57	>	* ForkJoinPool} may be constructed with a given target parallelism
58	>	* level; by default, equal to the number of available processors. The
59	>	* pool attempts to maintain enough active (or available) threads by
60	>	* dynamically adding, suspending, or resuming internal worker
61	>	* threads, even if some tasks are stalled waiting to join
62	>	* others. However, no such adjustments are guaranteed in the face of
63	>	* blocked IO or other unmanaged synchronization. The nested {@link
64	>	* ManagedBlocker} interface enables extension of the kinds of
65		* synchronization accommodated.
66		*
67		* <p>In addition to execution and lifecycle control methods, this
#	Line 52 | Line 72 | import java.util.concurrent.CountDownLat
72		* convenient form for informal monitoring.
73		*
74		* <p> As is the case with other ExecutorServices, there are three
75	<	* main task execution methods summarized in the following
76	<	* table. These are designed to be used by clients not already engaged
77	<	* in fork/join computations in the current pool.  The main forms of
78	<	* these methods accept instances of {@code ForkJoinTask}, but
79	<	* overloaded forms also allow mixed execution of plain {@code
75	>	* main task execution methods summarized in the following table.
76	>	* These are designed to be used primarily by clients not already
77	>	* engaged in fork/join computations in the current pool.  The main
78	>	* forms of these methods accept instances of {@code ForkJoinTask},
79	>	* but overloaded forms also allow mixed execution of plain {@code
80		* Runnable}- or {@code Callable}- based activities as well.  However,
81	<	* tasks that are already executing in a pool should normally
82	<	* <em>NOT</em> use these pool execution methods, but instead use the
83	<	* within-computation forms listed in the table.
81	>	* tasks that are already executing in a pool should normally instead
82	>	* use the within-computation forms listed in the table unless using
83	>	* async event-style tasks that are not usually joined, in which case
84	>	* there is little difference among choice of methods.
85		*
86		* <table BORDER CELLPADDING=3 CELLSPACING=1>
87		*  <tr>
#	Line 85 | Line 106 | import java.util.concurrent.CountDownLat
106		*  </tr>
107		* </table>
108		*
88	–	* <p><b>Sample Usage.</b> Normally a single {@code ForkJoinPool} is
89	–	* used for all parallel task execution in a program or subsystem.
90	–	* Otherwise, use would not usually outweigh the construction and
91	–	* bookkeeping overhead of creating a large set of threads. For
92	–	* example, a common pool could be used for the {@code SortTasks}
93	–	* illustrated in {@link RecursiveAction}. Because {@code
94	–	* ForkJoinPool} uses threads in {@linkplain java.lang.Thread#isDaemon
95	–	* daemon} mode, there is typically no need to explicitly {@link
96	–	* #shutdown} such a pool upon program exit.
97	–	*
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>
105	–	*
109		* <p><b>Implementation notes</b>: This implementation restricts the
110		* maximum number of running threads to 32767. Attempts to create
111		* pools with greater than the maximum number result in
#	Line 120 | Line 123 | public class ForkJoinPool extends Abstra
123		/*
124		* Implementation Overview
125		*
126	<	* This class provides the central bookkeeping and control for a
127	<	* set of worker threads: Submissions from non-FJ threads enter
128	<	* into a submission queue. Workers take these tasks and typically
129	<	* split them into subtasks that may be stolen by other workers.
130	<	* The main work-stealing mechanics implemented in class
131	<	* ForkJoinWorkerThread give first priority to processing tasks
132	<	* from their own queues (LIFO or FIFO, depending on mode), then
133	<	* to randomized FIFO steals of tasks in other worker queues, and
134	<	* lastly to new submissions. These mechanics do not consider
135	<	* affinities, loads, cache localities, etc, so rarely provide the
136	<	* best possible performance on a given machine, but portably
137	<	* provide good throughput by averaging over these factors.
138	<	* (Further, even if we did try to use such information, we do not
139	<	* usually have a basis for exploiting it. For example, some sets
140	<	* of tasks profit from cache affinities, but others are harmed by
141	<	* cache pollution effects.)
142	<	*
143	<	* Beyond work-stealing support and essential bookkeeping, the
144	<	* main responsibility of this framework is to take actions when
145	<	* one worker is waiting to join a task stolen (or always held by)
146	<	* another.  Because we are multiplexing many tasks on to a pool
147	<	* of workers, we can't just let them block (as in Thread.join).
148	<	* We also cannot just reassign the joiner's run-time stack with
149	<	* another and replace it later, which would be a form of
150	<	* "continuation", that even if possible is not necessarily a good
151	<	* idea. Given that the creation costs of most threads on most
152	<	* systems mainly surrounds setting up runtime stacks, thread
153	<	* creation and switching is usually not much more expensive than
154	<	* stack creation and switching, and is more flexible). Instead we
155	<	* combine two tactics:
126	>	* This class and its nested classes provide the main
127	>	* functionality and control for a set of worker threads:
128	>	* Submissions from non-FJ threads enter into submission queues.
129	>	* Workers take these tasks and typically split them into subtasks
130	>	* that may be stolen by other workers.  Preference rules give
131	>	* first priority to processing tasks from their own queues (LIFO
132	>	* or FIFO, depending on mode), then to randomized FIFO steals of
133	>	* tasks in other queues.
134	>	*
135	>	* WorkQueues
136	>	* ==========
137	>	*
138	>	* Most operations occur within work-stealing queues (in nested
139	>	* class WorkQueue).  These are special forms of Deques that
140	>	* support only three of the four possible end-operations -- push,
141	>	* pop, and poll (aka steal), under the further constraints that
142	>	* push and pop are called only from the owning thread (or, as
143	>	* extended here, under a lock), while poll may be called from
144	>	* other threads.  (If you are unfamiliar with them, you probably
145	>	* want to read Herlihy and Shavit's book "The Art of
146	>	* Multiprocessor programming", chapter 16 describing these in
147	>	* more detail before proceeding.)  The main work-stealing queue
148	>	* design is roughly similar to those in the papers "Dynamic
149	>	* Circular Work-Stealing Deque" by Chase and Lev, SPAA 2005
150	>	* (http://research.sun.com/scalable/pubs/index.html) and
151	>	* "Idempotent work stealing" by Michael, Saraswat, and Vechev,
152	>	* PPoPP 2009 (http://portal.acm.org/citation.cfm?id=1504186).
153	>	* The main differences ultimately stem from GC requirements that
154	>	* we null out taken slots as soon as we can, to maintain as small
155	>	* a footprint as possible even in programs generating huge
156	>	* numbers of tasks. To accomplish this, we shift the CAS
157	>	* arbitrating pop vs poll (steal) from being on the indices
158	>	* ("base" and "top") to the slots themselves.  So, both a
159	>	* successful pop and poll mainly entail a CAS of a slot from
160	>	* non-null to null.  Because we rely on CASes of references, we
161	>	* do not need tag bits on base or top.  They are simple ints as
162	>	* used in any circular array-based queue (see for example
163	>	* ArrayDeque).  Updates to the indices must still be ordered in a
164	>	* way that guarantees that top == base means the queue is empty,
165	>	* but otherwise may err on the side of possibly making the queue
166	>	* appear nonempty when a push, pop, or poll have not fully
167	>	* committed. Note that this means that the poll operation,
168	>	* considered individually, is not wait-free. One thief cannot
169	>	* successfully continue until another in-progress one (or, if
170	>	* previously empty, a push) completes.  However, in the
171	>	* aggregate, we ensure at least probabilistic non-blockingness.
172	>	* If an attempted steal fails, a thief always chooses a different
173	>	* random victim target to try next. So, in order for one thief to
174	>	* progress, it suffices for any in-progress poll or new push on
175	>	* any empty queue to complete. (This is why we normally use
176	>	* method pollAt and its variants that try once at the apparent
177	>	* base index, else consider alternative actions, rather than
178	>	* method poll.)
179	>	*
180	>	* This approach also enables support of a user mode in which local
181	>	* task processing is in FIFO, not LIFO order, simply by using
182	>	* poll rather than pop.  This can be useful in message-passing
183	>	* frameworks in which tasks are never joined.  However neither
184	>	* mode considers affinities, loads, cache localities, etc, so
185	>	* rarely provide the best possible performance on a given
186	>	* machine, but portably provide good throughput by averaging over
187	>	* these factors.  (Further, even if we did try to use such
188	>	* information, we do not usually have a basis for exploiting it.
189	>	* For example, some sets of tasks profit from cache affinities,
190	>	* but others are harmed by cache pollution effects.)
191	>	*
192	>	* WorkQueues are also used in a similar way for tasks submitted
193	>	* to the pool. We cannot mix these tasks in the same queues used
194	>	* for work-stealing (this would contaminate lifo/fifo
195	>	* processing). Instead, we loosely associate submission queues
196	>	* with submitting threads, using a form of hashing.  The
197	>	* ThreadLocal Submitter class contains a value initially used as
198	>	* a hash code for choosing existing queues, but may be randomly
199	>	* repositioned upon contention with other submitters.  In
200	>	* essence, submitters act like workers except that they never
201	>	* take tasks, and they are multiplexed on to a finite number of
202	>	* shared work queues. However, classes are set up so that future
203	>	* extensions could allow submitters to optionally help perform
204	>	* tasks as well. Insertion of tasks in shared mode requires a
205	>	* lock (mainly to protect in the case of resizing) but we use
206	>	* only a simple spinlock (using bits in field runState), because
207	>	* submitters encountering a busy queue move on to try or create
208	>	* other queues -- they block only when creating and registering
209	>	* new queues.
210	>	*
211	>	* Management
212	>	* ==========
213	>	*
214	>	* The main throughput advantages of work-stealing stem from
215	>	* decentralized control -- workers mostly take tasks from
216	>	* themselves or each other. We cannot negate this in the
217	>	* implementation of other management responsibilities. The main
218	>	* tactic for avoiding bottlenecks is packing nearly all
219	>	* essentially atomic control state into two volatile variables
220	>	* that are by far most often read (not written) as status and
221	>	* consistency checks.
222	>	*
223	>	* Field "ctl" contains 64 bits holding all the information needed
224	>	* to atomically decide to add, inactivate, enqueue (on an event
225	>	* queue), dequeue, and/or re-activate workers.  To enable this
226	>	* packing, we restrict maximum parallelism to (1<<15)-1 (which is
227	>	* far in excess of normal operating range) to allow ids, counts,
228	>	* and their negations (used for thresholding) to fit into 16bit
229	>	* fields.
230	>	*
231	>	* Field "runState" contains 32 bits needed to register and
232	>	* deregister WorkQueues, as well as to enable shutdown. It is
233	>	* only modified under a lock (normally briefly held, but
234	>	* occasionally protecting allocations and resizings) but even
235	>	* when locked remains available to check consistency.
236	>	*
237	>	* Recording WorkQueues.  WorkQueues are recorded in the
238	>	* "workQueues" array that is created upon first use and expanded
239	>	* if necessary.  Updates to the array while recording new workers
240	>	* and unrecording terminated ones are protected from each other
241	>	* by a lock but the array is otherwise concurrently readable, and
242	>	* accessed directly.  To simplify index-based operations, the
243	>	* array size is always a power of two, and all readers must
244	>	* tolerate null slots. Shared (submission) queues are at even
245	>	* indices, worker queues at odd indices. Grouping them together
246	>	* in this way simplifies and speeds up task scanning.
247	>	*
248	>	* All worker thread creation is on-demand, triggered by task
249	>	* submissions, replacement of terminated workers, and/or
250	>	* compensation for blocked workers. However, all other support
251	>	* code is set up to work with other policies.  To ensure that we
252	>	* do not hold on to worker references that would prevent GC, ALL
253	>	* accesses to workQueues are via indices into the workQueues
254	>	* array (which is one source of some of the messy code
255	>	* constructions here). In essence, the workQueues array serves as
256	>	* a weak reference mechanism. Thus for example the wait queue
257	>	* field of ctl stores indices, not references.  Access to the
258	>	* workQueues in associated methods (for example signalWork) must
259	>	* both index-check and null-check the IDs. All such accesses
260	>	* ignore bad IDs by returning out early from what they are doing,
261	>	* since this can only be associated with termination, in which
262	>	* case it is OK to give up.  All uses of the workQueues array
263	>	* also check that it is non-null (even if previously
264	>	* non-null). This allows nulling during termination, which is
265	>	* currently not necessary, but remains an option for
266	>	* resource-revocation-based shutdown schemes. It also helps
267	>	* reduce JIT issuance of uncommon-trap code, which tends to
268	>	* unnecessarily complicate control flow in some methods.
269	>	*
270	>	* Event Queuing. Unlike HPC work-stealing frameworks, we cannot
271	>	* let workers spin indefinitely scanning for tasks when none can
272	>	* be found immediately, and we cannot start/resume workers unless
273	>	* there appear to be tasks available.  On the other hand, we must
274	>	* quickly prod them into action when new tasks are submitted or
275	>	* generated. In many usages, ramp-up time to activate workers is
276	>	* the main limiting factor in overall performance (this is
277	>	* compounded at program start-up by JIT compilation and
278	>	* allocation). So we try to streamline this as much as possible.
279	>	* We park/unpark workers after placing in an event wait queue
280	>	* when they cannot find work. This "queue" is actually a simple
281	>	* Treiber stack, headed by the "id" field of ctl, plus a 15bit
282	>	* counter value (that reflects the number of times a worker has
283	>	* been inactivated) to avoid ABA effects (we need only as many
284	>	* version numbers as worker threads). Successors are held in
285	>	* field WorkQueue.nextWait.  Queuing deals with several intrinsic
286	>	* races, mainly that a task-producing thread can miss seeing (and
287	>	* signalling) another thread that gave up looking for work but
288	>	* has not yet entered the wait queue. We solve this by requiring
289	>	* a full sweep of all workers (via repeated calls to method
290	>	* scan()) both before and after a newly waiting worker is added
291	>	* to the wait queue. During a rescan, the worker might release
292	>	* some other queued worker rather than itself, which has the same
293	>	* net effect. Because enqueued workers may actually be rescanning
294	>	* rather than waiting, we set and clear the "parker" field of
295	>	* WorkQueues to reduce unnecessary calls to unpark.  (This
296	>	* requires a secondary recheck to avoid missed signals.)  Note
297	>	* the unusual conventions about Thread.interrupts surrounding
298	>	* parking and other blocking: Because interrupts are used solely
299	>	* to alert threads to check termination, which is checked anyway
300	>	* upon blocking, we clear status (using Thread.interrupted)
301	>	* before any call to park, so that park does not immediately
302	>	* return due to status being set via some other unrelated call to
303	>	* interrupt in user code.
304	>	*
305	>	* Signalling.  We create or wake up workers only when there
306	>	* appears to be at least one task they might be able to find and
307	>	* execute.  When a submission is added or another worker adds a
308	>	* task to a queue that previously had fewer than two tasks, they
309	>	* signal waiting workers (or trigger creation of new ones if
310	>	* fewer than the given parallelism level -- see signalWork).
311	>	* These primary signals are buttressed by signals during rescans;
312	>	* together these cover the signals needed in cases when more
313	>	* tasks are pushed but untaken, and improve performance compared
314	>	* to having one thread wake up all workers.
315	>	*
316	>	* Trimming workers. To release resources after periods of lack of
317	>	* use, a worker starting to wait when the pool is quiescent will
318	>	* time out and terminate if the pool has remained quiescent for a
319	>	* given period -- a short period if there are more threads than
320	>	* parallelism, longer as the number of threads decreases. This
321	>	* will slowly propagate, eventually terminating all workers after
322	>	* periods of non-use.
323	>	*
324	>	* Shutdown and Termination. A call to shutdownNow atomically sets
325	>	* a runState bit and then (non-atomically) sets each worker's
326	>	* runState status, cancels all unprocessed tasks, and wakes up
327	>	* all waiting workers.  Detecting whether termination should
328	>	* commence after a non-abrupt shutdown() call requires more work
329	>	* and bookkeeping. We need consensus about quiescence (i.e., that
330	>	* there is no more work). The active count provides a primary
331	>	* indication but non-abrupt shutdown still requires a rechecking
332	>	* scan for any workers that are inactive but not queued.
333	>	*
334	>	* Joining Tasks
335	>	* =============
336	>	*
337	>	* Any of several actions may be taken when one worker is waiting
338	>	* to join a task stolen (or always held) by another.  Because we
339	>	* are multiplexing many tasks on to a pool of workers, we can't
340	>	* just let them block (as in Thread.join).  We also cannot just
341	>	* reassign the joiner's run-time stack with another and replace
342	>	* it later, which would be a form of "continuation", that even if
343	>	* possible is not necessarily a good idea since we sometimes need
344	>	* both an unblocked task and its continuation to progress.
345	>	* Instead we combine two tactics:
346		*
347		*   Helping: Arranging for the joiner to execute some task that it
348	<	*      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.
348	>	*      would be running if the steal had not occurred.
349		*
350		*   Compensating: Unless there are already enough live threads,
351	<	*      method helpMaintainParallelism() may create or
352	<	*      re-activate a spare thread to compensate for blocked
162	<	*      joiners until they unblock.
351	>	*      method tryCompensate() may create or re-activate a spare
352	>	*      thread to compensate for blocked joiners until they unblock.
353		*
354	<	* It is impossible to keep exactly the target (parallelism)
355	<	* number of threads running at any given time.  Determining
356	<	* existence of conservatively safe helping targets, the
357	<	* availability of already-created spares, and the apparent need
358	<	* to create new spares are all racy and require heuristic
359	<	* guidance, so we rely on multiple retries of each.  Compensation
360	<	* occurs in slow-motion. It is triggered only upon timeouts of
171	<	* Object.wait used for joins. This reduces poor decisions that
172	<	* would otherwise be made when threads are waiting for others
173	<	* that are stalled because of unrelated activities such as
174	<	* garbage collection.
354	>	* A third form (implemented in tryRemoveAndExec and
355	>	* tryPollForAndExec) amounts to helping a hypothetical
356	>	* compensator: If we can readily tell that a possible action of a
357	>	* compensator is to steal and execute the task being joined, the
358	>	* joining thread can do so directly, without the need for a
359	>	* compensation thread (although at the expense of larger run-time
360	>	* stacks, but the tradeoff is typically worthwhile).
361		*
362		* The ManagedBlocker extension API can't use helping so relies
363		* only on compensation in method awaitBlocker.
364		*
365	<	* The main throughput advantages of work-stealing stem from
366	<	* decentralized control -- workers mostly steal tasks from each
367	<	* other. We do not want to negate this by creating bottlenecks
368	<	* implementing other management responsibilities. So we use a
369	<	* collection of techniques that avoid, reduce, or cope well with
370	<	* contention. These entail several instances of bit-packing into
371	<	* CASable fields to maintain only the minimally required
372	<	* atomicity. To enable such packing, we restrict maximum
373	<	* parallelism to (1<<15)-1 (enabling twice this (to accommodate
374	<	* unbalanced increments and decrements) to fit into a 16 bit
375	<	* field, which is far in excess of normal operating range.  Even
376	<	* though updates to some of these bookkeeping fields do sometimes
377	<	* contend with each other, they don't normally cache-contend with
378	<	* updates to others enough to warrant memory padding or
379	<	* isolation. So they are all held as fields of ForkJoinPool
380	<	* objects.  The main capabilities are as follows:
381	<	*
382	<	* 1. Creating and removing workers. Workers are recorded in the
383	<	* "workers" array. This is an array as opposed to some other data
384	<	* structure to support index-based random steals by workers.
385	<	* Updates to the array recording new workers and unrecording
386	<	* terminated ones are protected from each other by a lock
387	<	* (workerLock) but the array is otherwise concurrently readable,
388	<	* and accessed directly by workers. To simplify index-based
389	<	* operations, the array size is always a power of two, and all
390	<	* readers must tolerate null slots. Currently, all worker thread
391	<	* creation is on-demand, triggered by task submissions,
392	<	* replacement of terminated workers, and/or compensation for
393	<	* blocked workers. However, all other support code is set up to
394	<	* work with other policies.
395	<	*
396	<	* To ensure that we do not hold on to worker references that
397	<	* would prevent GC, ALL accesses to workers are via indices into
398	<	* the workers array (which is one source of some of the unusual
399	<	* code constructions here). In essence, the workers array serves
400	<	* as a WeakReference mechanism. Thus for example the event queue
401	<	* stores worker indices, not worker references. Access to the
402	<	* workers in associated methods (for example releaseEventWaiters)
403	<	* must both index-check and null-check the IDs. All such accesses
404	<	* ignore bad IDs by returning out early from what they are doing,
405	<	* since this can only be associated with shutdown, in which case
406	<	* it is OK to give up. On termination, we just clobber these
407	<	* data structures without trying to use them.
408	<	*
409	<	* 2. Bookkeeping for dynamically adding and removing workers. We
410	<	* aim to approximately maintain the given level of parallelism.
411	<	* When some workers are known to be blocked (on joins or via
412	<	* ManagedBlocker), we may create or resume others to take their
413	<	* place until they unblock (see below). Implementing this
414	<	* requires counts of the number of "running" threads (i.e., those
415	<	* that are neither blocked nor artificially suspended) as well as
416	<	* the total number.  These two values are packed into one field,
417	<	* "workerCounts" because we need accurate snapshots when deciding
418	<	* to create, resume or suspend.  Note however that the
419	<	* correspondence of these counts to reality is not guaranteed. In
420	<	* particular updates for unblocked threads may lag until they
421	<	* actually wake up.
422	<	*
423	<	* 3. Maintaining global run state. The run state of the pool
424	<	* consists of a runLevel (SHUTDOWN, TERMINATING, etc) similar to
425	<	* those in other Executor implementations, as well as a count of
426	<	* "active" workers -- those that are, or soon will be, or
427	<	* recently were executing tasks. The runLevel and active count
428	<	* are packed together in order to correctly trigger shutdown and
243	<	* termination. Without care, active counts can be subject to very
244	<	* high contention.  We substantially reduce this contention by
245	<	* relaxing update rules.  A worker must claim active status
246	<	* prospectively, by activating if it sees that a submitted or
247	<	* stealable task exists (it may find after activating that the
248	<	* task no longer exists). It stays active while processing this
249	<	* task (if it exists) and any other local subtasks it produces,
250	<	* until it cannot find any other tasks. It then tries
251	<	* inactivating (see method preStep), but upon update contention
252	<	* instead scans for more tasks, later retrying inactivation if it
253	<	* doesn't find any.
254	<	*
255	<	* 4. Managing idle workers waiting for tasks. We cannot let
256	<	* workers spin indefinitely scanning for tasks when none are
257	<	* available. On the other hand, we must quickly prod them into
258	<	* action when new tasks are submitted or generated.  We
259	<	* park/unpark these idle workers using an event-count scheme.
260	<	* Field eventCount is incremented upon events that may enable
261	<	* workers that previously could not find a task to now find one:
262	<	* Submission of a new task to the pool, or another worker pushing
263	<	* a task onto a previously empty queue.  (We also use this
264	<	* mechanism for configuration and termination actions that
265	<	* require wakeups of idle workers).  Each worker maintains its
266	<	* last known event count, and blocks when a scan for work did not
267	<	* find a task AND its lastEventCount matches the current
268	<	* eventCount. Waiting idle workers are recorded in a variant of
269	<	* Treiber stack headed by field eventWaiters which, when nonzero,
270	<	* encodes the thread index and count awaited for by the worker
271	<	* thread most recently calling eventSync. This thread in turn has
272	<	* a record (field nextEventWaiter) for the next waiting worker.
273	<	* In addition to allowing simpler decisions about need for
274	<	* wakeup, the event count bits in eventWaiters serve the role of
275	<	* tags to avoid ABA errors in Treiber stacks. Upon any wakeup,
276	<	* released threads also try to release at most two others.  The
277	<	* net effect is a tree-like diffusion of signals, where released
278	<	* threads (and possibly others) help with unparks.  To further
279	<	* reduce contention effects a bit, failed CASes to increment
280	<	* field eventCount are tolerated without retries in signalWork.
281	<	* Conceptually they are merged into the same event, which is OK
282	<	* when their only purpose is to enable workers to scan for work.
283	<	*
284	<	* 5. Managing suspension of extra workers. When a worker notices
285	<	* (usually upon timeout of a wait()) that there are too few
286	<	* running threads, we may create a new thread to maintain
287	<	* parallelism level, or at least avoid starvation. Usually, extra
288	<	* threads are needed for only very short periods, yet join
289	<	* dependencies are such that we sometimes need them in
290	<	* bursts. Rather than create new threads each time this happens,
291	<	* we suspend no-longer-needed extra ones as "spares". For most
292	<	* purposes, we don't distinguish "extra" spare threads from
293	<	* normal "core" threads: On each call to preStep (the only point
294	<	* at which we can do this) a worker checks to see if there are
295	<	* now too many running workers, and if so, suspends itself.
296	<	* Method helpMaintainParallelism looks for suspended threads to
297	<	* resume before considering creating a new replacement. The
298	<	* spares themselves are encoded on another variant of a Treiber
299	<	* Stack, headed at field "spareWaiters".  Note that the use of
300	<	* spares is intrinsically racy.  One thread may become a spare at
301	<	* about the same time as another is needlessly being created. We
302	<	* counteract this and related slop in part by requiring resumed
303	<	* spares to immediately recheck (in preStep) to see whether they
304	<	* they should re-suspend.
305	<	*
306	<	* 6. Killing off unneeded workers. A timeout mechanism is used to
307	<	* shed unused workers: The oldest (first) event queue waiter uses
308	<	* a timed rather than hard wait. When this wait times out without
309	<	* a normal wakeup, it tries to shutdown any one (for convenience
310	<	* the newest) other spare or event waiter via
311	<	* tryShutdownUnusedWorker. This eventually reduces the number of
312	<	* worker threads to a minimum of one after a long enough period
313	<	* without use.
314	<	*
315	<	* 7. 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 impossible
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
322	<	* presence of transient 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 parallelism to
327	<	* lapse for a while during joins, and use a timeout to see if,
328	<	* after the resulting settling, there is still a need for
329	<	* additional workers.  This also better copes with the fact that
330	<	* some of the methods in this class tend to never become compiled
331	<	* (but are interpreted), so some components of the entire set of
332	<	* controls might execute 100 times faster than others. And
333	<	* similarly for cases where the apparent lack of work is just due
334	<	* to GC stalls and other transient system activity.
365	>	* The algorithm in tryHelpStealer entails a form of "linear"
366	>	* helping: Each worker records (in field currentSteal) the most
367	>	* recent task it stole from some other worker. Plus, it records
368	>	* (in field currentJoin) the task it is currently actively
369	>	* joining. Method tryHelpStealer uses these markers to try to
370	>	* find a worker to help (i.e., steal back a task from and execute
371	>	* it) that could hasten completion of the actively joined task.
372	>	* In essence, the joiner executes a task that would be on its own
373	>	* local deque had the to-be-joined task not been stolen. This may
374	>	* be seen as a conservative variant of the approach in Wagner &
375	>	* Calder "Leapfrogging: a portable technique for implementing
376	>	* efficient futures" SIGPLAN Notices, 1993
377	>	* (http://portal.acm.org/citation.cfm?id=155354). It differs in
378	>	* that: (1) We only maintain dependency links across workers upon
379	>	* steals, rather than use per-task bookkeeping.  This sometimes
380	>	* requires a linear scan of workQueues array to locate stealers,
381	>	* but often doesn't because stealers leave hints (that may become
382	>	* stale/wrong) of where to locate them.  A stealHint is only a
383	>	* hint because a worker might have had multiple steals and the
384	>	* hint records only one of them (usually the most current).
385	>	* Hinting isolates cost to when it is needed, rather than adding
386	>	* to per-task overhead.  (2) It is "shallow", ignoring nesting
387	>	* and potentially cyclic mutual steals.  (3) It is intentionally
388	>	* racy: field currentJoin is updated only while actively joining,
389	>	* which means that we miss links in the chain during long-lived
390	>	* tasks, GC stalls etc (which is OK since blocking in such cases
391	>	* is usually a good idea).  (4) We bound the number of attempts
392	>	* to find work (see MAX_HELP) and fall back to suspending the
393	>	* worker and if necessary replacing it with another.
394	>	*
395	>	* It is impossible to keep exactly the target parallelism number
396	>	* of threads running at any given time.  Determining the
397	>	* existence of conservatively safe helping targets, the
398	>	* availability of already-created spares, and the apparent need
399	>	* to create new spares are all racy, so we rely on multiple
400	>	* retries of each.  Compensation in the apparent absence of
401	>	* helping opportunities is challenging to control on JVMs, where
402	>	* GC and other activities can stall progress of tasks that in
403	>	* turn stall out many other dependent tasks, without us being
404	>	* able to determine whether they will ever require compensation.
405	>	* Even though work-stealing otherwise encounters little
406	>	* degradation in the presence of more threads than cores,
407	>	* aggressively adding new threads in such cases entails risk of
408	>	* unwanted positive feedback control loops in which more threads
409	>	* cause more dependent stalls (as well as delayed progress of
410	>	* unblocked threads to the point that we know they are available)
411	>	* leading to more situations requiring more threads, and so
412	>	* on. This aspect of control can be seen as an (analytically
413	>	* intractable) game with an opponent that may choose the worst
414	>	* (for us) active thread to stall at any time.  We take several
415	>	* precautions to bound losses (and thus bound gains), mainly in
416	>	* methods tryCompensate and awaitJoin: (1) We only try
417	>	* compensation after attempting enough helping steps (measured
418	>	* via counting and timing) that we have already consumed the
419	>	* estimated cost of creating and activating a new thread.  (2) We
420	>	* allow up to 50% of threads to be blocked before initially
421	>	* adding any others, and unless completely saturated, check that
422	>	* some work is available for a new worker before adding. Also, we
423	>	* create up to only 50% more threads until entering a mode that
424	>	* only adds a thread if all others are possibly blocked.  All
425	>	* together, this means that we might be half as fast to react,
426	>	* and create half as many threads as possible in the ideal case,
427	>	* but present vastly fewer anomalies in all other cases compared
428	>	* to both more aggressive and more conservative alternatives.
429		*
430	<	* Beware that there is a lot of representation-level coupling
430	>	* Style notes: There is a lot of representation-level coupling
431		* among classes ForkJoinPool, ForkJoinWorkerThread, and
432	<	* ForkJoinTask.  For example, direct access to "workers" array by
433	<	* workers, and direct access to ForkJoinTask.status by both
434	<	* ForkJoinPool and ForkJoinWorkerThread.  There is little point
435	<	* trying to reduce this, since any associated future changes in
436	<	* representations will need to be accompanied by algorithmic
437	<	* changes anyway.
438	<	*
439	<	* Style notes: There are lots of inline assignments (of form
440	<	* "while ((local = field) != 0)") which are usually the simplest
441	<	* way to ensure the required read orderings (which are sometimes
442	<	* critical). Also several occurrences of the unusual "do {}
443	<	* while(!cas...)" which is the simplest way to force an update of
444	<	* a CAS'ed variable. There are also other coding oddities that
445	<	* help some methods perform reasonably even when interpreted (not
446	<	* compiled), at the expense of some messy constructions that
447	<	* reduce byte code counts.
448	<	*
449	<	* The order of declarations in this file is: (1) statics (2)
450	<	* fields (along with constants used when unpacking some of them)
451	<	* (3) internal control methods (4) callbacks and other support
452	<	* for ForkJoinTask and ForkJoinWorkerThread classes, (5) exported
453	<	* methods (plus a few little helpers).
432	>	* ForkJoinTask.  The fields of WorkQueue maintain data structures
433	>	* managed by ForkJoinPool, so are directly accessed.  There is
434	>	* little point trying to reduce this, since any associated future
435	>	* changes in representations will need to be accompanied by
436	>	* algorithmic changes anyway. Several methods intrinsically
437	>	* sprawl because they must accumulate sets of consistent reads of
438	>	* volatiles held in local variables.  Methods signalWork() and
439	>	* scan() are the main bottlenecks, so are especially heavily
440	>	* micro-optimized/mangled.  There are lots of inline assignments
441	>	* (of form "while ((local = field) != 0)") which are usually the
442	>	* simplest way to ensure the required read orderings (which are
443	>	* sometimes critical). This leads to a "C"-like style of listing
444	>	* declarations of these locals at the heads of methods or blocks.
445	>	* There are several occurrences of the unusual "do {} while
446	>	* (!cas...)"  which is the simplest way to force an update of a
447	>	* CAS'ed variable. There are also other coding oddities that help
448	>	* some methods perform reasonably even when interpreted (not
449	>	* compiled).
450	>	*
451	>	* The order of declarations in this file is:
452	>	* (1) Static utility functions
453	>	* (2) Nested (static) classes
454	>	* (3) Static fields
455	>	* (4) Fields, along with constants used when unpacking some of them
456	>	* (5) Internal control methods
457	>	* (6) Callbacks and other support for ForkJoinTask methods
458	>	* (7) Exported methods
459	>	* (8) Static block initializing statics in minimally dependent order
460		*/
461
462	+	// Static utilities
463	+
464	+	/**
465	+	* If there is a security manager, makes sure caller has
466	+	* permission to modify threads.
467	+	*/
468	+	private static void checkPermission() {
469	+	SecurityManager security = System.getSecurityManager();
470	+	if (security != null)
471	+	security.checkPermission(modifyThreadPermission);
472	+	}
473	+
474	+	// Nested classes
475	+
476		/**
477		* Factory for creating new {@link ForkJoinWorkerThread}s.
478		* A {@code ForkJoinWorkerThreadFactory} must be defined and used
#	Line 387 | Line 501 | public class ForkJoinPool extends Abstra
501		}
502
503		/**
504	<	* Creates a new ForkJoinWorkerThread. This factory is used unless
505	<	* overridden in ForkJoinPool constructors.
504	>	* Class for artificial tasks that are used to replace the target
505	>	* of local joins if they are removed from an interior queue slot
506	>	* in WorkQueue.tryRemoveAndExec. We don't need the proxy to
507	>	* actually do anything beyond having a unique identity.
508	>	*/
509	>	static final class EmptyTask extends ForkJoinTask<Void> {
510	>	EmptyTask() { status = ForkJoinTask.NORMAL; } // force done
511	>	public final Void getRawResult() { return null; }
512	>	public final void setRawResult(Void x) {}
513	>	public final boolean exec() { return true; }
514	>	}
515	>
516	>	/**
517	>	* Queues supporting work-stealing as well as external task
518	>	* submission. See above for main rationale and algorithms.
519	>	* Implementation relies heavily on "Unsafe" intrinsics
520	>	* and selective use of "volatile":
521	>	*
522	>	* Field "base" is the index (mod array.length) of the least valid
523	>	* queue slot, which is always the next position to steal (poll)
524	>	* from if nonempty. Reads and writes require volatile orderings
525	>	* but not CAS, because updates are only performed after slot
526	>	* CASes.
527	>	*
528	>	* Field "top" is the index (mod array.length) of the next queue
529	>	* slot to push to or pop from. It is written only by owner thread
530	>	* for push, or under lock for trySharedPush, and accessed by
531	>	* other threads only after reading (volatile) base.  Both top and
532	>	* base are allowed to wrap around on overflow, but (top - base)
533	>	* (or more commonly -(base - top) to force volatile read of base
534	>	* before top) still estimates size.
535	>	*
536	>	* The array slots are read and written using the emulation of
537	>	* volatiles/atomics provided by Unsafe. Insertions must in
538	>	* general use putOrderedObject as a form of releasing store to
539	>	* ensure that all writes to the task object are ordered before
540	>	* its publication in the queue. (Although we can avoid one case
541	>	* of this when locked in trySharedPush.) All removals entail a
542	>	* CAS to null.  The array is always a power of two. To ensure
543	>	* safety of Unsafe array operations, all accesses perform
544	>	* explicit null checks and implicit bounds checks via
545	>	* power-of-two masking.
546	>	*
547	>	* In addition to basic queuing support, this class contains
548	>	* fields described elsewhere to control execution. It turns out
549	>	* to work better memory-layout-wise to include them in this
550	>	* class rather than a separate class.
551	>	*
552	>	* Performance on most platforms is very sensitive to placement of
553	>	* instances of both WorkQueues and their arrays -- we absolutely
554	>	* do not want multiple WorkQueue instances or multiple queue
555	>	* arrays sharing cache lines. (It would be best for queue objects
556	>	* and their arrays to share, but there is nothing available to
557	>	* help arrange that).  Unfortunately, because they are recorded
558	>	* in a common array, WorkQueue instances are often moved to be
559	>	* adjacent by garbage collectors. To reduce impact, we use field
560	>	* padding that works OK on common platforms; this effectively
561	>	* trades off slightly slower average field access for the sake of
562	>	* avoiding really bad worst-case access. (Until better JVM
563	>	* support is in place, this padding is dependent on transient
564	>	* properties of JVM field layout rules.)  We also take care in
565	>	* allocating, sizing and resizing the array. Non-shared queue
566	>	* arrays are initialized (via method growArray) by workers before
567	>	* use. Others are allocated on first use.
568		*/
569	<	public static final ForkJoinWorkerThreadFactory
570	<	defaultForkJoinWorkerThreadFactory =
571	<	new DefaultForkJoinWorkerThreadFactory();
569	>	static final class WorkQueue {
570	>	/**
571	>	* Capacity of work-stealing queue array upon initialization.
572	>	* Must be a power of two; at least 4, but should be larger to
573	>	* reduce or eliminate cacheline sharing among queues.
574	>	* Currently, it is much larger, as a partial workaround for
575	>	* the fact that JVMs often place arrays in locations that
576	>	* share GC bookkeeping (especially cardmarks) such that
577	>	* per-write accesses encounter serious memory contention.
578	>	*/
579	>	static final int INITIAL_QUEUE_CAPACITY = 1 << 13;
580
581	<	/**
582	<	* Permission required for callers of methods that may start or
583	<	* kill threads.
584	<	*/
585	<	private static final RuntimePermission modifyThreadPermission =
586	<	new RuntimePermission("modifyThread");
581	>	/**
582	>	* Maximum size for queue arrays. Must be a power of two less
583	>	* than or equal to 1 << (31 - width of array entry) to ensure
584	>	* lack of wraparound of index calculations, but defined to a
585	>	* value a bit less than this to help users trap runaway
586	>	* programs before saturating systems.
587	>	*/
588	>	static final int MAXIMUM_QUEUE_CAPACITY = 1 << 26; // 64M
589
590	<	/**
591	<	* If there is a security manager, makes sure caller has
592	<	* permission to modify threads.
593	<	*/
594	<	private static void checkPermission() {
595	<	SecurityManager security = System.getSecurityManager();
596	<	if (security != null)
597	<	security.checkPermission(modifyThreadPermission);
590	>	volatile long totalSteals; // cumulative number of steals
591	>	int seed;                  // for random scanning; initialize nonzero
592	>	volatile int eventCount;   // encoded inactivation count; < 0 if inactive
593	>	int nextWait;              // encoded record of next event waiter
594	>	int rescans;               // remaining scans until block
595	>	int nsteals;               // top-level task executions since last idle
596	>	final int mode;            // lifo, fifo, or shared
597	>	int poolIndex;             // index of this queue in pool (or 0)
598	>	int stealHint;             // index of most recent known stealer
599	>	volatile int runState;     // 1: locked, -1: terminate; else 0
600	>	volatile int base;         // index of next slot for poll
601	>	int top;                   // index of next slot for push
602	>	ForkJoinTask<?>[] array;   // the elements (initially unallocated)
603	>	final ForkJoinPool pool;   // the containing pool (may be null)
604	>	final ForkJoinWorkerThread owner; // owning thread or null if shared
605	>	volatile Thread parker;    // == owner during call to park; else null
606	>	volatile ForkJoinTask<?> currentJoin;  // task being joined in awaitJoin
607	>	ForkJoinTask<?> currentSteal; // current non-local task being executed
608	>	// Heuristic padding to ameliorate unfortunate memory placements
609	>	Object p00, p01, p02, p03, p04, p05, p06, p07;
610	>	Object p08, p09, p0a, p0b, p0c, p0d, p0e;
611	>
612	>	WorkQueue(ForkJoinPool pool, ForkJoinWorkerThread owner, int mode) {
613	>	this.mode = mode;
614	>	this.pool = pool;
615	>	this.owner = owner;
616	>	// Place indices in the center of array (that is not yet allocated)
617	>	base = top = INITIAL_QUEUE_CAPACITY >>> 1;
618	>	}
619	>
620	>	/**
621	>	* Returns the approximate number of tasks in the queue.
622	>	*/
623	>	final int queueSize() {
624	>	int n = base - top;       // non-owner callers must read base first
625	>	return (n >= 0) ? 0 : -n; // ignore transient negative
626	>	}
627	>
628	>	/**
629	>	* Provides a more accurate estimate of whether this queue has
630	>	* any tasks than does queueSize, by checking whether a
631	>	* near-empty queue has at least one unclaimed task.
632	>	*/
633	>	final boolean isEmpty() {
634	>	ForkJoinTask<?>[] a; int m, s;
635	>	int n = base - (s = top);
636	>	return (n >= 0 ||
637	>	(n == -1 &&
638	>	((a = array) == null ||
639	>	(m = a.length - 1) < 0 ||
640	>	U.getObjectVolatile
641	>	(a, ((m & (s - 1)) << ASHIFT) + ABASE) == null)));
642	>	}
643	>
644	>	/**
645	>	* Pushes a task. Call only by owner in unshared queues.
646	>	*
647	>	* @param task the task. Caller must ensure non-null.
648	>	* @throw RejectedExecutionException if array cannot be resized
649	>	*/
650	>	final void push(ForkJoinTask<?> task) {
651	>	ForkJoinTask<?>[] a; ForkJoinPool p;
652	>	int s = top, m, n;
653	>	if ((a = array) != null) {    // ignore if queue removed
654	>	U.putOrderedObject
655	>	(a, (((m = a.length - 1) & s) << ASHIFT) + ABASE, task);
656	>	if ((n = (top = s + 1) - base) <= 2) {
657	>	if ((p = pool) != null)
658	>	p.signalWork();
659	>	}
660	>	else if (n >= m)
661	>	growArray(true);
662	>	}
663	>	}
664	>
665	>	/**
666	>	* Pushes a task if lock is free and array is either big
667	>	* enough or can be resized to be big enough.
668	>	*
669	>	* @param task the task. Caller must ensure non-null.
670	>	* @return true if submitted
671	>	*/
672	>	final boolean trySharedPush(ForkJoinTask<?> task) {
673	>	boolean submitted = false;
674	>	if (runState == 0 && U.compareAndSwapInt(this, RUNSTATE, 0, 1)) {
675	>	ForkJoinTask<?>[] a = array;
676	>	int s = top;
677	>	try {
678	>	if ((a != null && a.length > s + 1 - base) ||
679	>	(a = growArray(false)) != null) { // must presize
680	>	int j = (((a.length - 1) & s) << ASHIFT) + ABASE;
681	>	U.putObject(a, (long)j, task);    // don't need "ordered"
682	>	top = s + 1;
683	>	submitted = true;
684	>	}
685	>	} finally {
686	>	runState = 0;                         // unlock
687	>	}
688	>	}
689	>	return submitted;
690	>	}
691	>
692	>	/**
693	>	* Takes next task, if one exists, in LIFO order.  Call only
694	>	* by owner in unshared queues.
695	>	*/
696	>	final ForkJoinTask<?> pop() {
697	>	ForkJoinTask<?>[] a; ForkJoinTask<?> t; int m;
698	>	if ((a = array) != null && (m = a.length - 1) >= 0) {
699	>	for (int s; (s = top - 1) - base >= 0;) {
700	>	long j = ((m & s) << ASHIFT) + ABASE;
701	>	if ((t = (ForkJoinTask<?>)U.getObject(a, j)) == null)
702	>	break;
703	>	if (U.compareAndSwapObject(a, j, t, null)) {
704	>	top = s;
705	>	return t;
706	>	}
707	>	}
708	>	}
709	>	return null;
710	>	}
711	>
712	>	final ForkJoinTask<?> sharedPop() {
713	>	ForkJoinTask<?> task = null;
714	>	if (runState == 0 && U.compareAndSwapInt(this, RUNSTATE, 0, 1)) {
715	>	try {
716	>	ForkJoinTask<?>[] a; int m;
717	>	if ((a = array) != null && (m = a.length - 1) >= 0) {
718	>	for (int s; (s = top - 1) - base >= 0;) {
719	>	long j = ((m & s) << ASHIFT) + ABASE;
720	>	ForkJoinTask<?> t =
721	>	(ForkJoinTask<?>)U.getObject(a, j);
722	>	if (t == null)
723	>	break;
724	>	if (U.compareAndSwapObject(a, j, t, null)) {
725	>	top = s;
726	>	task = t;
727	>	break;
728	>	}
729	>	}
730	>	}
731	>	} finally {
732	>	runState = 0;
733	>	}
734	>	}
735	>	return task;
736	>	}
737	>
738	>
739	>	/**
740	>	* Takes a task in FIFO order if b is base of queue and a task
741	>	* can be claimed without contention. Specialized versions
742	>	* appear in ForkJoinPool methods scan and tryHelpStealer.
743	>	*/
744	>	final ForkJoinTask<?> pollAt(int b) {
745	>	ForkJoinTask<?> t; ForkJoinTask<?>[] a;
746	>	if ((a = array) != null) {
747	>	int j = (((a.length - 1) & b) << ASHIFT) + ABASE;
748	>	if ((t = (ForkJoinTask<?>)U.getObjectVolatile(a, j)) != null &&
749	>	base == b &&
750	>	U.compareAndSwapObject(a, j, t, null)) {
751	>	base = b + 1;
752	>	return t;
753	>	}
754	>	}
755	>	return null;
756	>	}
757	>
758	>	/**
759	>	* Takes next task, if one exists, in FIFO order.
760	>	*/
761	>	final ForkJoinTask<?> poll() {
762	>	ForkJoinTask<?>[] a; int b; ForkJoinTask<?> t;
763	>	while ((b = base) - top < 0 && (a = array) != null) {
764	>	int j = (((a.length - 1) & b) << ASHIFT) + ABASE;
765	>	t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
766	>	if (t != null) {
767	>	if (base == b &&
768	>	U.compareAndSwapObject(a, j, t, null)) {
769	>	base = b + 1;
770	>	return t;
771	>	}
772	>	}
773	>	else if (base == b) {
774	>	if (b + 1 == top)
775	>	break;
776	>	Thread.yield(); // wait for lagging update
777	>	}
778	>	}
779	>	return null;
780	>	}
781	>
782	>	/**
783	>	* Takes next task, if one exists, in order specified by mode.
784	>	*/
785	>	final ForkJoinTask<?> nextLocalTask() {
786	>	return mode == 0 ? pop() : poll();
787	>	}
788	>
789	>	/**
790	>	* Returns next task, if one exists, in order specified by mode.
791	>	*/
792	>	final ForkJoinTask<?> peek() {
793	>	ForkJoinTask<?>[] a = array; int m;
794	>	if (a == null || (m = a.length - 1) < 0)
795	>	return null;
796	>	int i = mode == 0 ? top - 1 : base;
797	>	int j = ((i & m) << ASHIFT) + ABASE;
798	>	return (ForkJoinTask<?>)U.getObjectVolatile(a, j);
799	>	}
800	>
801	>	/**
802	>	* Pops the given task only if it is at the current top.
803	>	*/
804	>	final boolean tryUnpush(ForkJoinTask<?> t) {
805	>	ForkJoinTask<?>[] a; int s;
806	>	if ((a = array) != null && (s = top) != base &&
807	>	U.compareAndSwapObject
808	>	(a, (((a.length - 1) & --s) << ASHIFT) + ABASE, t, null)) {
809	>	top = s;
810	>	return true;
811	>	}
812	>	return false;
813	>	}
814	>
815	>	/**
816	>	* Version of tryUnpush for shared queues; called by non-FJ
817	>	* submitters after prechecking that task probably exists.
818	>	*/
819	>	final boolean trySharedUnpush(ForkJoinTask<?> t) {
820	>	boolean success = false;
821	>	if (runState == 0 && U.compareAndSwapInt(this, RUNSTATE, 0, 1)) {
822	>	try {
823	>	ForkJoinTask<?>[] a; int s;
824	>	if ((a = array) != null && (s = top) != base &&
825	>	U.compareAndSwapObject
826	>	(a, (((a.length - 1) & --s) << ASHIFT) + ABASE, t, null)) {
827	>	top = s;
828	>	success = true;
829	>	}
830	>	} finally {
831	>	runState = 0;                         // unlock
832	>	}
833	>	}
834	>	return success;
835	>	}
836	>
837	>	/**
838	>	* Polls the given task only if it is at the current base.
839	>	*/
840	>	final boolean pollFor(ForkJoinTask<?> task) {
841	>	ForkJoinTask<?>[] a; int b;
842	>	if ((b = base) - top < 0 && (a = array) != null) {
843	>	int j = (((a.length - 1) & b) << ASHIFT) + ABASE;
844	>	if (U.getObjectVolatile(a, j) == task && base == b &&
845	>	U.compareAndSwapObject(a, j, task, null)) {
846	>	base = b + 1;
847	>	return true;
848	>	}
849	>	}
850	>	return false;
851	>	}
852	>
853	>	/**
854	>	* Initializes or doubles the capacity of array. Call either
855	>	* by owner or with lock held -- it is OK for base, but not
856	>	* top, to move while resizings are in progress.
857	>	*
858	>	* @param rejectOnFailure if true, throw exception if capacity
859	>	* exceeded (relayed ultimately to user); else return null.
860	>	*/
861	>	final ForkJoinTask<?>[] growArray(boolean rejectOnFailure) {
862	>	ForkJoinTask<?>[] oldA = array;
863	>	int size = oldA != null ? oldA.length << 1 : INITIAL_QUEUE_CAPACITY;
864	>	if (size <= MAXIMUM_QUEUE_CAPACITY) {
865	>	int oldMask, t, b;
866	>	ForkJoinTask<?>[] a = array = new ForkJoinTask<?>[size];
867	>	if (oldA != null && (oldMask = oldA.length - 1) >= 0 &&
868	>	(t = top) - (b = base) > 0) {
869	>	int mask = size - 1;
870	>	do {
871	>	ForkJoinTask<?> x;
872	>	int oldj = ((b & oldMask) << ASHIFT) + ABASE;
873	>	int j    = ((b &    mask) << ASHIFT) + ABASE;
874	>	x = (ForkJoinTask<?>)U.getObjectVolatile(oldA, oldj);
875	>	if (x != null &&
876	>	U.compareAndSwapObject(oldA, oldj, x, null))
877	>	U.putObjectVolatile(a, j, x);
878	>	} while (++b != t);
879	>	}
880	>	return a;
881	>	}
882	>	else if (!rejectOnFailure)
883	>	return null;
884	>	else
885	>	throw new RejectedExecutionException("Queue capacity exceeded");
886	>	}
887	>
888	>	/**
889	>	* Removes and cancels all known tasks, ignoring any exceptions.
890	>	*/
891	>	final void cancelAll() {
892	>	ForkJoinTask.cancelIgnoringExceptions(currentJoin);
893	>	ForkJoinTask.cancelIgnoringExceptions(currentSteal);
894	>	for (ForkJoinTask<?> t; (t = poll()) != null; )
895	>	ForkJoinTask.cancelIgnoringExceptions(t);
896	>	}
897	>
898	>	/**
899	>	* Computes next value for random probes.  Scans don't require
900	>	* a very high quality generator, but also not a crummy one.
901	>	* Marsaglia xor-shift is cheap and works well enough.  Note:
902	>	* This is manually inlined in its usages in ForkJoinPool to
903	>	* avoid writes inside busy scan loops.
904	>	*/
905	>	final int nextSeed() {
906	>	int r = seed;
907	>	r ^= r << 13;
908	>	r ^= r >>> 17;
909	>	return seed = r ^= r << 5;
910	>	}
911	>
912	>	// Specialized execution methods
913	>
914	>	/**
915	>	* Pops and runs tasks until empty.
916	>	*/
917	>	private void popAndExecAll() {
918	>	// A bit faster than repeated pop calls
919	>	ForkJoinTask<?>[] a; int m, s; long j; ForkJoinTask<?> t;
920	>	while ((a = array) != null && (m = a.length - 1) >= 0 &&
921	>	(s = top - 1) - base >= 0 &&
922	>	(t = ((ForkJoinTask<?>)
923	>	U.getObject(a, j = ((m & s) << ASHIFT) + ABASE)))
924	>	!= null) {
925	>	if (U.compareAndSwapObject(a, j, t, null)) {
926	>	top = s;
927	>	t.doExec();
928	>	}
929	>	}
930	>	}
931	>
932	>	/**
933	>	* Polls and runs tasks until empty.
934	>	*/
935	>	private void pollAndExecAll() {
936	>	for (ForkJoinTask<?> t; (t = poll()) != null;)
937	>	t.doExec();
938	>	}
939	>
940	>	/**
941	>	* If present, removes from queue and executes the given task, or
942	>	* any other cancelled task. Returns (true) immediately on any CAS
943	>	* or consistency check failure so caller can retry.
944	>	*
945	>	* @return 0 if no progress can be made, else positive
946	>	* (this unusual convention simplifies use with tryHelpStealer.)
947	>	*/
948	>	final int tryRemoveAndExec(ForkJoinTask<?> task) {
949	>	int stat = 1;
950	>	boolean removed = false, empty = true;
951	>	ForkJoinTask<?>[] a; int m, s, b, n;
952	>	if ((a = array) != null && (m = a.length - 1) >= 0 &&
953	>	(n = (s = top) - (b = base)) > 0) {
954	>	for (ForkJoinTask<?> t;;) {           // traverse from s to b
955	>	int j = ((--s & m) << ASHIFT) + ABASE;
956	>	t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
957	>	if (t == null)                    // inconsistent length
958	>	break;
959	>	else if (t == task) {
960	>	if (s + 1 == top) {           // pop
961	>	if (!U.compareAndSwapObject(a, j, task, null))
962	>	break;
963	>	top = s;
964	>	removed = true;
965	>	}
966	>	else if (base == b)           // replace with proxy
967	>	removed = U.compareAndSwapObject(a, j, task,
968	>	new EmptyTask());
969	>	break;
970	>	}
971	>	else if (t.status >= 0)
972	>	empty = false;
973	>	else if (s + 1 == top) {          // pop and throw away
974	>	if (U.compareAndSwapObject(a, j, t, null))
975	>	top = s;
976	>	break;
977	>	}
978	>	if (--n == 0) {
979	>	if (!empty && base == b)
980	>	stat = 0;
981	>	break;
982	>	}
983	>	}
984	>	}
985	>	if (removed)
986	>	task.doExec();
987	>	return stat;
988	>	}
989	>
990	>	/**
991	>	* Version of shared pop that takes top element only if it
992	>	* its root is the given CountedCompleter.
993	>	*/
994	>	final CountedCompleter<?> sharedPopCC(CountedCompleter<?> root) {
995	>	CountedCompleter<?> task = null;
996	>	if (runState == 0 && U.compareAndSwapInt(this, RUNSTATE, 0, 1)) {
997	>	try {
998	>	ForkJoinTask<?>[] a; int m;
999	>	if ((a = array) != null && (m = a.length - 1) >= 0) {
1000	>	outer:for (int s; (s = top - 1) - base >= 0;) {
1001	>	long j = ((m & s) << ASHIFT) + ABASE;
1002	>	ForkJoinTask<?> t =
1003	>	(ForkJoinTask<?>)U.getObject(a, j);
1004	>	if (t == null || !(t instanceof CountedCompleter))
1005	>	break;
1006	>	CountedCompleter<?> cc = (CountedCompleter<?>)t;
1007	>	for (CountedCompleter<?> q = cc, p;;) {
1008	>	if (q == root) {
1009	>	if (U.compareAndSwapObject(a, j, cc, null)) {
1010	>	top = s;
1011	>	task = cc;
1012	>	break outer;
1013	>	}
1014	>	break;
1015	>	}
1016	>	if ((p = q.completer) == null)
1017	>	break outer;
1018	>	q = p;
1019	>	}
1020	>	}
1021	>	}
1022	>	} finally {
1023	>	runState = 0;
1024	>	}
1025	>	}
1026	>	return task;
1027	>	}
1028	>
1029	>	/**
1030	>	* Executes a top-level task and any local tasks remaining
1031	>	* after execution.
1032	>	*/
1033	>	final void runTask(ForkJoinTask<?> t) {
1034	>	if (t != null) {
1035	>	currentSteal = t;
1036	>	t.doExec();
1037	>	if (top != base) {       // process remaining local tasks
1038	>	if (mode == 0)
1039	>	popAndExecAll();
1040	>	else
1041	>	pollAndExecAll();
1042	>	}
1043	>	++nsteals;
1044	>	currentSteal = null;
1045	>	}
1046	>	}
1047	>
1048	>	/**
1049	>	* Executes a non-top-level (stolen) task.
1050	>	*/
1051	>	final void runSubtask(ForkJoinTask<?> t) {
1052	>	if (t != null) {
1053	>	ForkJoinTask<?> ps = currentSteal;
1054	>	currentSteal = t;
1055	>	t.doExec();
1056	>	currentSteal = ps;
1057	>	}
1058	>	}
1059	>
1060	>	/**
1061	>	* Returns true if owned and not known to be blocked.
1062	>	*/
1063	>	final boolean isApparentlyUnblocked() {
1064	>	Thread wt; Thread.State s;
1065	>	return (eventCount >= 0 &&
1066	>	(wt = owner) != null &&
1067	>	(s = wt.getState()) != Thread.State.BLOCKED &&
1068	>	s != Thread.State.WAITING &&
1069	>	s != Thread.State.TIMED_WAITING);
1070	>	}
1071	>
1072	>	/**
1073	>	* If this owned and is not already interrupted, try to
1074	>	* interrupt and/or unpark, ignoring exceptions.
1075	>	*/
1076	>	final void interruptOwner() {
1077	>	Thread wt, p;
1078	>	if ((wt = owner) != null && !wt.isInterrupted()) {
1079	>	try {
1080	>	wt.interrupt();
1081	>	} catch (SecurityException ignore) {
1082	>	}
1083	>	}
1084	>	if ((p = parker) != null)
1085	>	U.unpark(p);
1086	>	}
1087	>
1088	>	// Unsafe mechanics
1089	>	private static final sun.misc.Unsafe U;
1090	>	private static final long RUNSTATE;
1091	>	private static final int ABASE;
1092	>	private static final int ASHIFT;
1093	>	static {
1094	>	int s;
1095	>	try {
1096	>	U = getUnsafe();
1097	>	Class<?> k = WorkQueue.class;
1098	>	Class<?> ak = ForkJoinTask[].class;
1099	>	RUNSTATE = U.objectFieldOffset
1100	>	(k.getDeclaredField("runState"));
1101	>	ABASE = U.arrayBaseOffset(ak);
1102	>	s = U.arrayIndexScale(ak);
1103	>	} catch (Exception e) {
1104	>	throw new Error(e);
1105	>	}
1106	>	if ((s & (s-1)) != 0)
1107	>	throw new Error("data type scale not a power of two");
1108	>	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
1109	>	}
1110		}
1111
1112		/**
1113	<	* Generator for assigning sequence numbers as pool names.
1113	>	* Per-thread records for threads that submit to pools. Currently
1114	>	* holds only pseudo-random seed / index that is used to choose
1115	>	* submission queues in method doSubmit. In the future, this may
1116	>	* also incorporate a means to implement different task rejection
1117	>	* and resubmission policies.
1118	>	*
1119	>	* Seeds for submitters and workers/workQueues work in basically
1120	>	* the same way but are initialized and updated using slightly
1121	>	* different mechanics. Both are initialized using the same
1122	>	* approach as in class ThreadLocal, where successive values are
1123	>	* unlikely to collide with previous values. This is done during
1124	>	* registration for workers, but requires a separate AtomicInteger
1125	>	* for submitters. Seeds are then randomly modified upon
1126	>	* collisions using xorshifts, which requires a non-zero seed.
1127		*/
1128	<	private static final AtomicInteger poolNumberGenerator =
1129	<	new AtomicInteger();
1128	>	static final class Submitter {
1129	>	int seed;
1130	>	Submitter() {
1131	>	int s = nextSubmitterSeed.getAndAdd(SEED_INCREMENT);
1132	>	seed = (s == 0) ? 1 : s; // ensure non-zero
1133	>	}
1134	>	}
1135	>
1136	>	/** ThreadLocal class for Submitters */
1137	>	static final class ThreadSubmitter extends ThreadLocal<Submitter> {
1138	>	public Submitter initialValue() { return new Submitter(); }
1139	>	}
1140	>
1141	>	// static fields (initialized in static initializer below)
1142
1143		/**
1144	<	* The time to block in a join (see awaitJoin) before checking if
1145	<	* a new worker should be (re)started to maintain parallelism
423	<	* level. The value should be short enough to maintain global
424	<	* responsiveness and progress but long enough to avoid
425	<	* counterproductive firings during GC stalls or unrelated system
426	<	* activity, and to not bog down systems with continual re-firings
427	<	* on GCs or legitimately long waits.
1144	>	* Creates a new ForkJoinWorkerThread. This factory is used unless
1145	>	* overridden in ForkJoinPool constructors.
1146		*/
1147	<	private static final long JOIN_TIMEOUT_MILLIS = 250L; // 4 per second
1147	>	public static final ForkJoinWorkerThreadFactory
1148	>	defaultForkJoinWorkerThreadFactory;
1149	>
1150	>	/** Property prefix for constructing common pool */
1151	>	private static final String propPrefix =
1152	>	"java.util.concurrent.ForkJoinPool.common.";
1153
1154		/**
1155	<	* The wakeup interval (in nanoseconds) for the oldest worker
1156	<	* worker waiting for an event invokes tryShutdownUnusedWorker to shrink
1157	<	* the number of workers.  The exact value does not matter too
435	<	* much, but should be long enough to slowly release resources
436	<	* during long periods without use without disrupting normal use.
1155	>	* Common (static) pool. Non-null for public use unless a static
1156	>	* construction exception, but internal usages must null-check on
1157	>	* use.
1158		*/
1159	<	private static final long SHRINK_RATE_NANOS =
439	<	30L * 1000L * 1000L * 1000L; // 2 per minute
1159	>	static final ForkJoinPool commonPool;
1160
1161		/**
1162	<	* Absolute bound for parallelism level. Twice this number plus
443	<	* one (i.e., 0xfff) must fit into a 16bit field to enable
444	<	* word-packing for some counts and indices.
1162	>	* Common pool parallelism. Must equal commonPool.parallelism.
1163		*/
1164	<	private static final int MAX_WORKERS   = 0x7fff;
1164	>	static final int commonPoolParallelism;
1165
1166		/**
1167	<	* Array holding all worker threads in the pool.  Array size must
450	<	* be a power of two.  Updates and replacements are protected by
451	<	* workerLock, but the array is always kept in a consistent enough
452	<	* state to be randomly accessed without locking by workers
453	<	* performing work-stealing, as well as other traversal-based
454	<	* methods in this class. All readers must tolerate that some
455	<	* array slots may be null.
1167	>	* Generator for assigning sequence numbers as pool names.
1168		*/
1169	<	volatile ForkJoinWorkerThread[] workers;
1169	>	private static final AtomicInteger poolNumberGenerator;
1170
1171		/**
1172	<	* Queue for external submissions.
1172	>	* Generator for initial hashes/seeds for submitters. Accessed by
1173	>	* Submitter class constructor.
1174		*/
1175	<	private final LinkedTransferQueue<ForkJoinTask<?>> submissionQueue;
1175	>	static final AtomicInteger nextSubmitterSeed;
1176
1177		/**
1178	<	* Lock protecting updates to workers array.
1178	>	* Permission required for callers of methods that may start or
1179	>	* kill threads.
1180		*/
1181	<	private final ReentrantLock workerLock;
1181	>	private static final RuntimePermission modifyThreadPermission;
1182
1183		/**
1184	<	* Latch released upon termination.
1184	>	* Per-thread submission bookkeeping. Shared across all pools
1185	>	* to reduce ThreadLocal pollution and because random motion
1186	>	* to avoid contention in one pool is likely to hold for others.
1187		*/
1188	<	private final Phaser termination;
1188	>	private static final ThreadSubmitter submitters;
1189	>
1190	>	// static constants
1191
1192		/**
1193	<	* Creation factory for worker threads.
1193	>	* Initial timeout value (in nanoseconds) for the thread triggering
1194	>	* quiescence to park waiting for new work. On timeout, the thread
1195	>	* will instead try to shrink the number of workers.
1196		*/
1197	<	private final ForkJoinWorkerThreadFactory factory;
1197	>	private static final long IDLE_TIMEOUT      = 1000L * 1000L * 1000L; // 1sec
1198
1199		/**
1200	<	* Sum of per-thread steal counts, updated only when threads are
481	<	* idle or terminating.
1200	>	* Timeout value when there are more threads than parallelism level
1201		*/
1202	<	private volatile long stealCount;
1202	>	private static final long FAST_IDLE_TIMEOUT =  100L * 1000L * 1000L;
1203
1204		/**
1205	<	* Encoded record of top of Treiber stack of threads waiting for
1206	<	* events. The top 32 bits contain the count being waited for. The
1207	<	* bottom 16 bits contains one plus the pool index of waiting
1208	<	* worker thread. (Bits 16-31 are unused.)
1205	>	* The maximum stolen->joining link depth allowed in method
1206	>	* tryHelpStealer.  Must be a power of two. This value also
1207	>	* controls the maximum number of times to try to help join a task
1208	>	* without any apparent progress or change in pool state before
1209	>	* giving up and blocking (see awaitJoin).  Depths for legitimate
1210	>	* chains are unbounded, but we use a fixed constant to avoid
1211	>	* (otherwise unchecked) cycles and to bound staleness of
1212	>	* traversal parameters at the expense of sometimes blocking when
1213	>	* we could be helping.
1214		*/
1215	<	private volatile long eventWaiters;
492	<
493	<	private static final int  EVENT_COUNT_SHIFT = 32;
494	<	private static final long WAITER_ID_MASK    = (1L << 16) - 1L;
1215	>	private static final int MAX_HELP = 64;
1216
1217		/**
1218	<	* A counter for events that may wake up worker threads:
1219	<	*   - Submission of a new task to the pool
1220	<	*   - A worker pushing a task on an empty queue
1221	<	*   - termination
1218	>	* Secondary time-based bound (in nanosecs) for helping attempts
1219	>	* before trying compensated blocking in awaitJoin. Used in
1220	>	* conjunction with MAX_HELP to reduce variance due to different
1221	>	* polling rates associated with different helping options. The
1222	>	* value should roughly approximate the time required to create
1223	>	* and/or activate a worker thread.
1224		*/
1225	<	private volatile int eventCount;
1225	>	private static final long COMPENSATION_DELAY = 1L << 18; // ~0.25 millisec
1226
1227		/**
1228	<	* Encoded record of top of Treiber stack of spare threads waiting
1229	<	* for resumption. The top 16 bits contain an arbitrary count to
507	<	* avoid ABA effects. The bottom 16bits contains one plus the pool
508	<	* index of waiting worker thread.
1228	>	* Increment for seed generators. See class ThreadLocal for
1229	>	* explanation.
1230		*/
1231	<	private volatile int spareWaiters;
511	<
512	<	private static final int SPARE_COUNT_SHIFT = 16;
513	<	private static final int SPARE_ID_MASK     = (1 << 16) - 1;
1231	>	private static final int SEED_INCREMENT = 0x61c88647;
1232
1233		/**
1234	<	* Lifecycle control. The low word contains the number of workers
1235	<	* that are (probably) executing tasks. This value is atomically
1236	<	* incremented before a worker gets a task to run, and decremented
1237	<	* when worker has no tasks and cannot find any.  Bits 16-18
1238	<	* contain runLevel value. When all are zero, the pool is
1239	<	* running. Level transitions are monotonic (running -> shutdown
1240	<	* -> terminating -> terminated) so each transition adds a bit.
1241	<	* These are bundled together to ensure consistent read for
1242	<	* termination checks (i.e., that runLevel is at least SHUTDOWN
1243	<	* and active threads is zero).
1234	>	* Bits and masks for control variables
1235	>	*
1236	>	* Field ctl is a long packed with:
1237	>	* AC: Number of active running workers minus target parallelism (16 bits)
1238	>	* TC: Number of total workers minus target parallelism (16 bits)
1239	>	* ST: true if pool is terminating (1 bit)
1240	>	* EC: the wait count of top waiting thread (15 bits)
1241	>	* ID: poolIndex of top of Treiber stack of waiters (16 bits)
1242	>	*
1243	>	* When convenient, we can extract the upper 32 bits of counts and
1244	>	* the lower 32 bits of queue state, u = (int)(ctl >>> 32) and e =
1245	>	* (int)ctl.  The ec field is never accessed alone, but always
1246	>	* together with id and st. The offsets of counts by the target
1247	>	* parallelism and the positionings of fields makes it possible to
1248	>	* perform the most common checks via sign tests of fields: When
1249	>	* ac is negative, there are not enough active workers, when tc is
1250	>	* negative, there are not enough total workers, and when e is
1251	>	* negative, the pool is terminating.  To deal with these possibly
1252	>	* negative fields, we use casts in and out of "short" and/or
1253	>	* signed shifts to maintain signedness.
1254	>	*
1255	>	* When a thread is queued (inactivated), its eventCount field is
1256	>	* set negative, which is the only way to tell if a worker is
1257	>	* prevented from executing tasks, even though it must continue to
1258	>	* scan for them to avoid queuing races. Note however that
1259	>	* eventCount updates lag releases so usage requires care.
1260		*
1261	<	* Notes: Most direct CASes are dependent on these bitfield
1262	<	* positions.  Also, this field is non-private to enable direct
1263	<	* performance-sensitive CASes in ForkJoinWorkerThread.
1261	>	* Field runState is an int packed with:
1262	>	* SHUTDOWN: true if shutdown is enabled (1 bit)
1263	>	* SEQ:  a sequence number updated upon (de)registering workers (30 bits)
1264	>	* INIT: set true after workQueues array construction (1 bit)
1265	>	*
1266	>	* The sequence number enables simple consistency checks:
1267	>	* Staleness of read-only operations on the workQueues array can
1268	>	* be checked by comparing runState before vs after the reads.
1269		*/
531	–	volatile int runState;
1270
1271	<	// Note: The order among run level values matters.
1272	<	private static final int RUNLEVEL_SHIFT     = 16;
1273	<	private static final int SHUTDOWN           = 1 << RUNLEVEL_SHIFT;
1274	<	private static final int TERMINATING        = 1 << (RUNLEVEL_SHIFT + 1);
1275	<	private static final int TERMINATED         = 1 << (RUNLEVEL_SHIFT + 2);
538	<	private static final int ACTIVE_COUNT_MASK  = (1 << RUNLEVEL_SHIFT) - 1;
1271	>	// bit positions/shifts for fields
1272	>	private static final int  AC_SHIFT   = 48;
1273	>	private static final int  TC_SHIFT   = 32;
1274	>	private static final int  ST_SHIFT   = 31;
1275	>	private static final int  EC_SHIFT   = 16;
1276
1277	<	/**
1278	<	* Holds number of total (i.e., created and not yet terminated)
1279	<	* and running (i.e., not blocked on joins or other managed sync)
1280	<	* threads, packed together to ensure consistent snapshot when
1281	<	* making decisions about creating and suspending spare
1282	<	* threads. Updated only by CAS. Note that adding a new worker
546	<	* requires incrementing both counts, since workers start off in
547	<	* running state.
548	<	*/
549	<	private volatile int workerCounts;
1277	>	// bounds
1278	>	private static final int  SMASK      = 0xffff;  // short bits
1279	>	private static final int  MAX_CAP    = 0x7fff;  // max #workers - 1
1280	>	private static final int  SQMASK     = 0xfffe;  // even short bits
1281	>	private static final int  SHORT_SIGN = 1 << 15;
1282	>	private static final int  INT_SIGN   = 1 << 31;
1283
1284	<	private static final int TOTAL_COUNT_SHIFT  = 16;
1285	<	private static final int RUNNING_COUNT_MASK = (1 << TOTAL_COUNT_SHIFT) - 1;
1286	<	private static final int ONE_RUNNING        = 1;
1287	<	private static final int ONE_TOTAL          = 1 << TOTAL_COUNT_SHIFT;
1284	>	// masks
1285	>	private static final long STOP_BIT   = 0x0001L << ST_SHIFT;
1286	>	private static final long AC_MASK    = ((long)SMASK) << AC_SHIFT;
1287	>	private static final long TC_MASK    = ((long)SMASK) << TC_SHIFT;
1288
1289	<	/**
1290	<	* The target parallelism level.
1291	<	* Accessed directly by ForkJoinWorkerThreads.
559	<	*/
560	<	final int parallelism;
1289	>	// units for incrementing and decrementing
1290	>	private static final long TC_UNIT    = 1L << TC_SHIFT;
1291	>	private static final long AC_UNIT    = 1L << AC_SHIFT;
1292
1293	<	/**
1294	<	* True if use local fifo, not default lifo, for local polling
1295	<	* Read by, and replicated by ForkJoinWorkerThreads
1296	<	*/
1297	<	final boolean locallyFifo;
1293	>	// masks and units for dealing with u = (int)(ctl >>> 32)
1294	>	private static final int  UAC_SHIFT  = AC_SHIFT - 32;
1295	>	private static final int  UTC_SHIFT  = TC_SHIFT - 32;
1296	>	private static final int  UAC_MASK   = SMASK << UAC_SHIFT;
1297	>	private static final int  UTC_MASK   = SMASK << UTC_SHIFT;
1298	>	private static final int  UAC_UNIT   = 1 << UAC_SHIFT;
1299	>	private static final int  UTC_UNIT   = 1 << UTC_SHIFT;
1300
1301	<	/**
1302	<	* The uncaught exception handler used when any worker abruptly
1303	<	* terminates.
571	<	*/
572	<	private final Thread.UncaughtExceptionHandler ueh;
1301	>	// masks and units for dealing with e = (int)ctl
1302	>	private static final int E_MASK      = 0x7fffffff; // no STOP_BIT
1303	>	private static final int E_SEQ       = 1 << EC_SHIFT;
1304
1305	<	/**
1306	<	* Pool number, just for assigning useful names to worker threads
576	<	*/
577	<	private final int poolNumber;
1305	>	// runState bits
1306	>	private static final int SHUTDOWN    = 1 << 31;
1307
1308	<	// Utilities for CASing fields. Note that most of these
1309	<	// are usually manually inlined by callers
1308	>	// access mode for WorkQueue
1309	>	static final int LIFO_QUEUE          =  0;
1310	>	static final int FIFO_QUEUE          =  1;
1311	>	static final int SHARED_QUEUE        = -1;
1312
1313	<	/**
583	<	* Increments running count part of workerCounts
584	<	*/
585	<	final void incrementRunningCount() {
586	<	int c;
587	<	do {} while (!UNSAFE.compareAndSwapInt(this, workerCountsOffset,
588	<	c = workerCounts,
589	<	c + ONE_RUNNING));
590	<	}
1313	>	// Instance fields
1314
1315	<	/**
1316	<	* Tries to decrement running count unless already zero
1317	<	*/
1318	<	final boolean tryDecrementRunningCount() {
1319	<	int wc = workerCounts;
1320	<	if ((wc & RUNNING_COUNT_MASK) == 0)
1321	<	return false;
1322	<	return UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1323	<	wc, wc - ONE_RUNNING);
1324	<	}
1315	>	/*
1316	>	* Field layout order in this class tends to matter more than one
1317	>	* would like. Runtime layout order is only loosely related to
1318	>	* declaration order and may differ across JVMs, but the following
1319	>	* empirically works OK on current JVMs.
1320	>	*/
1321	>
1322	>	volatile long stealCount;                  // collects worker counts
1323	>	volatile long ctl;                         // main pool control
1324	>	final int parallelism;                     // parallelism level
1325	>	final int localMode;                       // per-worker scheduling mode
1326	>	volatile int nextWorkerNumber;             // to create worker name string
1327	>	final int submitMask;                      // submit queue index bound
1328	>	int nextSeed;                              // for initializing worker seeds
1329	>	volatile int mainLock;                     // spinlock for array updates
1330	>	volatile int runState;                     // shutdown status and seq
1331	>	WorkQueue[] workQueues;                    // main registry
1332	>	final ForkJoinWorkerThreadFactory factory; // factory for new workers
1333	>	final Thread.UncaughtExceptionHandler ueh; // per-worker UEH
1334	>	final String workerNamePrefix;             // to create worker name string
1335
1336	<	/**
1337	<	* Forces decrement of encoded workerCounts, awaiting nonzero if
1338	<	* (rarely) necessary when other count updates lag.
1339	<	*
1340	<	* @param dr -- either zero or ONE_RUNNING
1341	<	* @param dt == either zero or ONE_TOTAL
1342	<	*/
1343	<	private void decrementWorkerCounts(int dr, int dt) {
1344	<	for (;;) {
1345	<	int wc = workerCounts;
1346	<	if ((wc & RUNNING_COUNT_MASK)  - dr < 0 ||
1347	<	(wc >>> TOTAL_COUNT_SHIFT) - dt < 0) {
1348	<	if ((runState & TERMINATED) != 0)
1349	<	return; // lagging termination on a backout
1350	<	Thread.yield();
1336	>	/*
1337	>	* Mechanics for main lock protecting worker array updates.  Uses
1338	>	* the same strategy as ConcurrentHashMap bins -- a spinLock for
1339	>	* normal cases, but falling back to builtin lock when (rarely)
1340	>	* needed.  See internal ConcurrentHashMap documentation for
1341	>	* explanation.
1342	>	*/
1343	>
1344	>	static final int LOCK_WAITING = 2; // bit to indicate need for signal
1345	>	static final int MAX_LOCK_SPINS = 1 << 8;
1346	>
1347	>	private void tryAwaitMainLock() {
1348	>	int spins = MAX_LOCK_SPINS, r = 0, h;
1349	>	while (((h = mainLock) & 1) != 0) {
1350	>	if (r == 0)
1351	>	r = ThreadLocalRandom.current().nextInt(); // randomize spins
1352	>	else if (spins >= 0) {
1353	>	r ^= r << 1; r ^= r >>> 3; r ^= r << 10; // xorshift
1354	>	if (r >= 0)
1355	>	--spins;
1356	>	}
1357	>	else if (U.compareAndSwapInt(this, MAINLOCK, h, h | LOCK_WAITING)) {
1358	>	synchronized (this) {
1359	>	if ((mainLock & LOCK_WAITING) != 0) {
1360	>	try {
1361	>	wait();
1362	>	} catch (InterruptedException ie) {
1363	>	try {
1364	>	Thread.currentThread().interrupt();
1365	>	} catch (SecurityException ignore) {
1366	>	}
1367	>	}
1368	>	}
1369	>	else
1370	>	notifyAll(); // possibly won race vs signaller
1371	>	}
1372	>	break;
1373		}
619	–	if (UNSAFE.compareAndSwapInt(this, workerCountsOffset,
620	–	wc, wc - (dr + dt)))
621	–	return;
1374		}
1375		}
1376
1377	<	/**
626	<	* Tries decrementing active count; fails on contention.
627	<	* Called when workers cannot find tasks to run.
628	<	*/
629	<	final boolean tryDecrementActiveCount() {
630	<	int c;
631	<	return UNSAFE.compareAndSwapInt(this, runStateOffset,
632	<	c = runState, c - 1);
633	<	}
1377	>	//  Creating, registering, and deregistering workers
1378
1379		/**
1380	<	* Advances to at least the given level. Returns true if not
637	<	* already in at least the given level.
1380	>	* Tries to create and start a worker
1381		*/
1382	<	private boolean advanceRunLevel(int level) {
1383	<	for (;;) {
1384	<	int s = runState;
1385	<	if ((s & level) != 0)
1386	<	return false;
1387	<	if (UNSAFE.compareAndSwapInt(this, runStateOffset, s, s | level))
1388	<	return true;
1382	>	private void addWorker() {
1383	>	Throwable ex = null;
1384	>	ForkJoinWorkerThread wt = null;
1385	>	try {
1386	>	if ((wt = factory.newThread(this)) != null) {
1387	>	wt.start();
1388	>	return;
1389	>	}
1390	>	} catch (Throwable e) {
1391	>	ex = e;
1392		}
1393	+	deregisterWorker(wt, ex); // adjust counts etc on failure
1394		}
1395
649	–	// workers array maintenance
650	–
1396		/**
1397	<	* Records and returns a workers array index for new worker.
1397	>	* Callback from ForkJoinWorkerThread constructor to assign a
1398	>	* public name. This must be separate from registerWorker because
1399	>	* it is called during the "super" constructor call in
1400	>	* ForkJoinWorkerThread.
1401		*/
1402	<	private int recordWorker(ForkJoinWorkerThread w) {
1403	<	// Try using slot totalCount-1. If not available, scan and/or resize
1404	<	int k = (workerCounts >>> TOTAL_COUNT_SHIFT) - 1;
1405	<	final ReentrantLock lock = this.workerLock;
1406	<	lock.lock();
659	<	try {
660	<	ForkJoinWorkerThread[] ws = workers;
661	<	int n = ws.length;
662	<	if (k < 0 || k >= n || ws[k] != null) {
663	<	for (k = 0; k < n && ws[k] != null; ++k)
664	<	;
665	<	if (k == n)
666	<	ws = Arrays.copyOf(ws, n << 1);
667	<	}
668	<	ws[k] = w;
669	<	workers = ws; // volatile array write ensures slot visibility
670	<	} finally {
671	<	lock.unlock();
672	<	}
673	<	return k;
1402	>	final String nextWorkerName() {
1403	>	int n;
1404	>	do {} while (!U.compareAndSwapInt(this, NEXTWORKERNUMBER,
1405	>	n = nextWorkerNumber, ++n));
1406	>	return workerNamePrefix.concat(Integer.toString(n));
1407		}
1408
1409		/**
1410	<	* Nulls out record of worker in workers array
1410	>	* Callback from ForkJoinWorkerThread constructor to establish its
1411	>	* poolIndex and record its WorkQueue. To avoid scanning bias due
1412	>	* to packing entries in front of the workQueues array, we treat
1413	>	* the array as a simple power-of-two hash table using per-thread
1414	>	* seed as hash, expanding as needed.
1415	>	*
1416	>	* @param w the worker's queue
1417		*/
1418	<	private void forgetWorker(ForkJoinWorkerThread w) {
1419	<	int idx = w.poolIndex;
1420	<	// Locking helps method recordWorker avoid unnecessary expansion
682	<	final ReentrantLock lock = this.workerLock;
683	<	lock.lock();
1418	>	final void registerWorker(WorkQueue w) {
1419	>	while (!U.compareAndSwapInt(this, MAINLOCK, 0, 1))
1420	>	tryAwaitMainLock();
1421		try {
1422	<	ForkJoinWorkerThread[] ws = workers;
1423	<	if (idx >= 0 && idx < ws.length && ws[idx] == w) // verify
1424	<	ws[idx] = null;
1422	>	WorkQueue[] ws;
1423	>	if ((ws = workQueues) == null)
1424	>	ws = workQueues = new WorkQueue[submitMask + 1];
1425	>	if (w != null) {
1426	>	int rs, n =  ws.length, m = n - 1;
1427	>	int s = nextSeed += SEED_INCREMENT; // rarely-colliding sequence
1428	>	w.seed = (s == 0) ? 1 : s;          // ensure non-zero seed
1429	>	int r = (s << 1) | 1;               // use odd-numbered indices
1430	>	if (ws[r &= m] != null) {           // collision
1431	>	int probes = 0;                 // step by approx half size
1432	>	int step = (n <= 4) ? 2 : ((n >>> 1) & SQMASK) + 2;
1433	>	while (ws[r = (r + step) & m] != null) {
1434	>	if (++probes >= n) {
1435	>	workQueues = ws = Arrays.copyOf(ws, n <<= 1);
1436	>	m = n - 1;
1437	>	probes = 0;
1438	>	}
1439	>	}
1440	>	}
1441	>	w.eventCount = w.poolIndex = r;     // establish before recording
1442	>	ws[r] = w;                          // also update seq
1443	>	runState = ((rs = runState) & SHUTDOWN) | ((rs + 2) & ~SHUTDOWN);
1444	>	}
1445		} finally {
1446	<	lock.unlock();
1446	>	if (!U.compareAndSwapInt(this, MAINLOCK, 1, 0)) {
1447	>	mainLock = 0;
1448	>	synchronized (this) { notifyAll(); };
1449	>	}
1450		}
1451		}
1452
1453		/**
1454	<	* Final callback from terminating worker.  Removes record of
1454	>	* Final callback from terminating worker, as well as upon failure
1455	>	* to construct or start a worker in addWorker.  Removes record of
1456		* worker from array, and adjusts counts. If pool is shutting
1457		* down, tries to complete termination.
1458		*
1459	<	* @param w the worker
1459	>	* @param wt the worker thread or null if addWorker failed
1460	>	* @param ex the exception causing failure, or null if none
1461		*/
1462	<	final void workerTerminated(ForkJoinWorkerThread w) {
1463	<	forgetWorker(w);
1464	<	decrementWorkerCounts(w.isTrimmed()? 0 : ONE_RUNNING, ONE_TOTAL);
1465	<	while (w.stealCount != 0) // collect final count
1466	<	tryAccumulateStealCount(w);
1467	<	tryTerminate(false);
1468	<	}
1469	<
1470	<	// Waiting for and signalling events
1471	<
1472	<	/**
1473	<	* Releases workers blocked on a count not equal to current count.
1474	<	* Normally called after precheck that eventWaiters isn't zero to
1475	<	* avoid wasted array checks. Gives up upon a change in count or
1476	<	* upon releasing two workers, letting others take over.
1477	<	*/
1478	<	private void releaseEventWaiters() {
1479	<	ForkJoinWorkerThread[] ws = workers;
1480	<	int n = ws.length;
719	<	long h = eventWaiters;
720	<	int ec = eventCount;
721	<	boolean releasedOne = false;
722	<	ForkJoinWorkerThread w; int id;
723	<	while ((id = ((int)(h & WAITER_ID_MASK)) - 1) >= 0 &&
724	<	(int)(h >>> EVENT_COUNT_SHIFT) != ec &&
725	<	id < n && (w = ws[id]) != null) {
726	<	if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
727	<	h,  w.nextWaiter)) {
728	<	LockSupport.unpark(w);
729	<	if (releasedOne) // exit on second release
730	<	break;
731	<	releasedOne = true;
1462	>	final void deregisterWorker(ForkJoinWorkerThread wt, Throwable ex) {
1463	>	WorkQueue w = null;
1464	>	if (wt != null && (w = wt.workQueue) != null) {
1465	>	w.runState = -1;                // ensure runState is set
1466	>	long steals = w.totalSteals + w.nsteals, sc;
1467	>	do {} while (!U.compareAndSwapLong(this, STEALCOUNT,
1468	>	sc = stealCount, sc + steals));
1469	>	int idx = w.poolIndex;
1470	>	while (!U.compareAndSwapInt(this, MAINLOCK, 0, 1))
1471	>	tryAwaitMainLock();
1472	>	try {
1473	>	WorkQueue[] ws = workQueues;
1474	>	if (ws != null && idx >= 0 && idx < ws.length && ws[idx] == w)
1475	>	ws[idx] = null;
1476	>	} finally {
1477	>	if (!U.compareAndSwapInt(this, MAINLOCK, 1, 0)) {
1478	>	mainLock = 0;
1479	>	synchronized (this) { notifyAll(); };
1480	>	}
1481		}
1482	<	if (eventCount != ec)
1483	<	break;
1484	<	h = eventWaiters;
1482	>	}
1483	>
1484	>	long c;                             // adjust ctl counts
1485	>	do {} while (!U.compareAndSwapLong
1486	>	(this, CTL, c = ctl, (((c - AC_UNIT) & AC_MASK) |
1487	>	((c - TC_UNIT) & TC_MASK) |
1488	>	(c & ~(AC_MASK|TC_MASK)))));
1489	>
1490	>	if (!tryTerminate(false, false) && w != null) {
1491	>	w.cancelAll();                  // cancel remaining tasks
1492	>	if (w.array != null)            // suppress signal if never ran
1493	>	signalWork();               // wake up or create replacement
1494	>	if (ex == null)                 // help clean refs on way out
1495	>	ForkJoinTask.helpExpungeStaleExceptions();
1496	>	}
1497	>
1498	>	if (ex != null)                     // rethrow
1499	>	ForkJoinTask.rethrow(ex);
1500	>	}
1501	>
1502	>	// Submissions
1503	>
1504	>	/**
1505	>	* Unless shutting down, adds the given task to a submission queue
1506	>	* at submitter's current queue index (modulo submission
1507	>	* range). If no queue exists at the index, one is created.  If
1508	>	* the queue is busy, another index is randomly chosen. The
1509	>	* submitMask bounds the effective number of queues to the
1510	>	* (nearest power of two for) parallelism level.
1511	>	*
1512	>	* @param task the task. Caller must ensure non-null.
1513	>	*/
1514	>	private void doSubmit(ForkJoinTask<?> task) {
1515	>	Submitter s = submitters.get();
1516	>	for (int r = s.seed, m = submitMask;;) {
1517	>	WorkQueue[] ws; WorkQueue q;
1518	>	int k = r & m & SQMASK;          // use only even indices
1519	>	if (runState < 0)
1520	>	throw new RejectedExecutionException(); // shutting down
1521	>	else if ((ws = workQueues) == null || ws.length <= k) {
1522	>	while (!U.compareAndSwapInt(this, MAINLOCK, 0, 1))
1523	>	tryAwaitMainLock();
1524	>	try {
1525	>	if (workQueues == null)
1526	>	workQueues = new WorkQueue[submitMask + 1];
1527	>	} finally {
1528	>	if (!U.compareAndSwapInt(this, MAINLOCK, 1, 0)) {
1529	>	mainLock = 0;
1530	>	synchronized (this) { notifyAll(); };
1531	>	}
1532	>	}
1533	>	}
1534	>	else if ((q = ws[k]) == null) {  // create new queue
1535	>	WorkQueue nq = new WorkQueue(this, null, SHARED_QUEUE);
1536	>	while (!U.compareAndSwapInt(this, MAINLOCK, 0, 1))
1537	>	tryAwaitMainLock();
1538	>	try {
1539	>	int rs = runState;       // to update seq
1540	>	if (ws == workQueues && ws[k] == null) {
1541	>	ws[k] = nq;
1542	>	runState = ((rs & SHUTDOWN) | ((rs + 2) & ~SHUTDOWN));
1543	>	}
1544	>	} finally {
1545	>	if (!U.compareAndSwapInt(this, MAINLOCK, 1, 0)) {
1546	>	mainLock = 0;
1547	>	synchronized (this) { notifyAll(); };
1548	>	}
1549	>	}
1550	>	}
1551	>	else if (q.trySharedPush(task)) {
1552	>	signalWork();
1553	>	return;
1554	>	}
1555	>	else if (m > 1) {                // move to a different index
1556	>	r ^= r << 13;                // same xorshift as WorkQueues
1557	>	r ^= r >>> 17;
1558	>	s.seed = r ^= r << 5;
1559	>	}
1560	>	else
1561	>	Thread.yield();              // yield if no alternatives
1562		}
1563		}
1564
1565		/**
1566	<	* Tries to advance eventCount and releases waiters. Called only
741	<	* from workers.
1566	>	* Submits the given (non-null) task to the common pool, if possible.
1567		*/
1568	<	final void signalWork() {
1569	<	int c; // try to increment event count -- CAS failure OK
1570	<	UNSAFE.compareAndSwapInt(this, eventCountOffset, c = eventCount, c+1);
1571	<	if (eventWaiters != 0L)
1572	<	releaseEventWaiters();
1568	>	static void submitToCommonPool(ForkJoinTask<?> task) {
1569	>	ForkJoinPool p;
1570	>	if ((p = commonPool) == null)
1571	>	throw new RejectedExecutionException("Common Pool Unavailable");
1572	>	p.doSubmit(task);
1573		}
1574
1575		/**
1576	<	* Adds the given worker to event queue and blocks until
1577	<	* terminating or event count advances from the given value
1578	<	*
1579	<	* @param w the calling worker thread
1580	<	* @param ec the count
1581	<	*/
1582	<	private void eventSync(ForkJoinWorkerThread w, int ec) {
1583	<	long nh = (((long)ec) << EVENT_COUNT_SHIFT) | ((long)(w.poolIndex+1));
1584	<	long h;
1585	<	while ((runState < SHUTDOWN || !tryTerminate(false)) &&
1586	<	(((int)((h = eventWaiters) & WAITER_ID_MASK)) == 0 ||
1587	<	(int)(h >>> EVENT_COUNT_SHIFT) == ec) &&
1588	<	eventCount == ec) {
1589	<	if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
1590	<	w.nextWaiter = h, nh)) {
1591	<	awaitEvent(w, ec);
1592	<	break;
1576	>	* Returns true if the given task was submitted to common pool
1577	>	* and has not yet commenced execution, and is available for
1578	>	* removal according to execution policies; if so removing the
1579	>	* submission from the pool.
1580	>	*
1581	>	* @param task the task
1582	>	* @return true if successful
1583	>	*/
1584	>	static boolean tryUnsubmitFromCommonPool(ForkJoinTask<?> task) {
1585	>	// If not oversaturating platform, peek, looking for task and
1586	>	// eligibility before using trySharedUnpush to actually take
1587	>	// it under lock
1588	>	ForkJoinPool p; WorkQueue[] ws; WorkQueue w, q;
1589	>	ForkJoinTask<?>[] a; int ac, s, m;
1590	>	if ((p = commonPool) != null && (ws = p.workQueues) != null) {
1591	>	int k = submitters.get().seed & p.submitMask & SQMASK;
1592	>	if ((m = ws.length - 1) >= k && (q = ws[k]) != null &&
1593	>	(ac = (int)(p.ctl >> AC_SHIFT)) <= 0) {
1594	>	if (ac == 0) { // double check if all workers active
1595	>	for (int i = 1; i <= m; i += 2) {
1596	>	if ((w = ws[i]) != null && w.parker != null) {
1597	>	ac = -1;
1598	>	break;
1599	>	}
1600	>	}
1601	>	}
1602	>	return (ac < 0 && (a = q.array) != null &&
1603	>	(s = q.top - 1) - q.base >= 0 &&
1604	>	s >= 0 && s < a.length &&
1605	>	a[s] == task &&
1606	>	q.trySharedUnpush(task));
1607		}
1608		}
1609	+	return false;
1610		}
1611
1612		/**
1613	<	* Blocks the given worker (that has already been entered as an
1614	<	* event waiter) until terminating or event count advances from
1615	<	* the given value. The oldest (first) waiter uses a timed wait to
1616	<	* occasionally one-by-one shrink the number of workers (to a
1617	<	* minimum of one) if the pool has not been used for extended
1618	<	* periods.
1619	<	*
1620	<	* @param w the calling worker thread
1621	<	* @param ec the count
1622	<	*/
1623	<	private void awaitEvent(ForkJoinWorkerThread w, int ec) {
1624	<	while (eventCount == ec) {
1625	<	if (tryAccumulateStealCount(w)) { // transfer while idle
1626	<	boolean untimed = (w.nextWaiter != 0L ||
1627	<	(workerCounts & RUNNING_COUNT_MASK) <= 1);
1628	<	long startTime = untimed? 0 : System.nanoTime();
1629	<	Thread.interrupted();         // clear/ignore interrupt
1630	<	if (eventCount != ec || w.runState != 0 ||
791	<	runState >= TERMINATING)  // recheck after clear
792	<	break;
793	<	if (untimed)
794	<	LockSupport.park(w);
795	<	else {
796	<	LockSupport.parkNanos(w, SHRINK_RATE_NANOS);
797	<	if (eventCount != ec || w.runState != 0 ||
798	<	runState >= TERMINATING)
799	<	break;
800	<	if (System.nanoTime() - startTime >= SHRINK_RATE_NANOS)
801	<	tryShutdownUnusedWorker(ec);
1613	>	* Tries to pop and run a task within same computation from common pool
1614	>	*/
1615	>	static void popAndExecCCFromCommonPool(CountedCompleter<?> cc) {
1616	>	ForkJoinPool p; WorkQueue[] ws; WorkQueue q, w; int m, ac;
1617	>	CountedCompleter<?> par, task;
1618	>	if ((p = commonPool) != null && (ws = p.workQueues) != null) {
1619	>	while ((par = cc.completer) != null) // find root
1620	>	cc = par;
1621	>	int k = submitters.get().seed & p.submitMask & SQMASK;
1622	>	if ((m = ws.length - 1) >= k && (q = ws[k]) != null &&
1623	>	(ac = (int)(p.ctl >> AC_SHIFT)) <= 0) {
1624	>	if (ac == 0) {
1625	>	for (int i = 1; i <= m; i += 2) {
1626	>	if ((w = ws[i]) != null && w.parker != null) {
1627	>	ac = -1;
1628	>	break;
1629	>	}
1630	>	}
1631		}
1632	+	if (ac < 0 && q.top - q.base > 0 &&
1633	+	(task = q.sharedPopCC(cc)) != null)
1634	+	task.exec();
1635		}
1636		}
1637		}
1638
1639	<	// Maintaining parallelism
1639	>	// Maintaining ctl counts
1640
1641		/**
1642	<	* Pushes worker onto the spare stack
1642	>	* Increments active count; mainly called upon return from blocking.
1643		*/
1644	<	final void pushSpare(ForkJoinWorkerThread w) {
1645	<	int ns = (++w.spareCount << SPARE_COUNT_SHIFT) | (w.poolIndex + 1);
1646	<	do {} while (!UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
815	<	w.nextSpare = spareWaiters,ns));
1644	>	final void incrementActiveCount() {
1645	>	long c;
1646	>	do {} while (!U.compareAndSwapLong(this, CTL, c = ctl, c + AC_UNIT));
1647		}
1648
1649		/**
1650	<	* Tries (once) to resume a spare if the number of running
820	<	* threads is less than target.
1650	>	* Tries to create one or activate one or more workers if too few are active.
1651		*/
1652	<	private void tryResumeSpare() {
1653	<	int sw, id;
1654	<	ForkJoinWorkerThread[] ws = workers;
1655	<	int n = ws.length;
1656	<	ForkJoinWorkerThread w;
1657	<	if ((sw = spareWaiters) != 0 &&
1658	<	(id = (sw & SPARE_ID_MASK) - 1) >= 0 &&
1659	<	id < n && (w = ws[id]) != null &&
1660	<	(workerCounts & RUNNING_COUNT_MASK) < parallelism &&
1661	<	spareWaiters == sw &&
1662	<	UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
1663	<	sw, w.nextSpare)) {
1664	<	int c; // increment running count before resume
1665	<	do {} while(!UNSAFE.compareAndSwapInt
1666	<	(this, workerCountsOffset,
1667	<	c = workerCounts, c + ONE_RUNNING));
1668	<	if (w.tryUnsuspend())
1669	<	LockSupport.unpark(w);
1670	<	else   // back out if w was shutdown
1671	<	decrementWorkerCounts(ONE_RUNNING, 0);
1652	>	final void signalWork() {
1653	>	long c; int u;
1654	>	while ((u = (int)((c = ctl) >>> 32)) < 0) {     // too few active
1655	>	WorkQueue[] ws = workQueues; int e, i; WorkQueue w; Thread p;
1656	>	if ((e = (int)c) > 0) {                     // at least one waiting
1657	>	if (ws != null && (i = e & SMASK) < ws.length &&
1658	>	(w = ws[i]) != null && w.eventCount == (e | INT_SIGN)) {
1659	>	long nc = (((long)(w.nextWait & E_MASK)) |
1660	>	((long)(u + UAC_UNIT) << 32));
1661	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1662	>	w.eventCount = (e + E_SEQ) & E_MASK;
1663	>	if ((p = w.parker) != null)
1664	>	U.unpark(p);                // activate and release
1665	>	break;
1666	>	}
1667	>	}
1668	>	else
1669	>	break;
1670	>	}
1671	>	else if (e == 0 && (u & SHORT_SIGN) != 0) { // too few total
1672	>	long nc = (long)(((u + UTC_UNIT) & UTC_MASK) |
1673	>	((u + UAC_UNIT) & UAC_MASK)) << 32;
1674	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1675	>	addWorker();
1676	>	break;
1677	>	}
1678	>	}
1679	>	else
1680	>	break;
1681		}
1682		}
1683
1684	+	// Scanning for tasks
1685	+
1686		/**
1687	<	* Tries to increase the number of running workers if below target
847	<	* parallelism: If a spare exists tries to resume it via
848	<	* tryResumeSpare.  Otherwise, if not enough total workers or all
849	<	* existing workers are busy, adds a new worker. In all cases also
850	<	* helps wake up releasable workers waiting for work.
1687	>	* Top-level runloop for workers, called by ForkJoinWorkerThread.run.
1688		*/
1689	<	private void helpMaintainParallelism() {
1690	<	int pc = parallelism;
1691	<	int wc, rs, tc;
1692	<	while (((wc = workerCounts) & RUNNING_COUNT_MASK) < pc &&
1693	<	(rs = runState) < TERMINATING) {
1694	<	if (spareWaiters != 0)
1695	<	tryResumeSpare();
1696	<	else if ((tc = wc >>> TOTAL_COUNT_SHIFT) >= MAX_WORKERS ||
1697	<	(tc >= pc && (rs & ACTIVE_COUNT_MASK) != tc))
1698	<	break;   // enough total
1699	<	else if (runState == rs && workerCounts == wc &&
1700	<	UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
1701	<	wc + (ONE_RUNNING|ONE_TOTAL))) {
1702	<	ForkJoinWorkerThread w = null;
1703	<	try {
1704	<	w = factory.newThread(this);
1705	<	} finally { // adjust on null or exceptional factory return
1706	<	if (w == null) {
1707	<	decrementWorkerCounts(ONE_RUNNING, ONE_TOTAL);
1708	<	tryTerminate(false); // handle failure during shutdown
1689	>	final void runWorker(WorkQueue w) {
1690	>	w.growArray(false);         // initialize queue array in this thread
1691	>	do { w.runTask(scan(w)); } while (w.runState >= 0);
1692	>	}
1693	>
1694	>	/**
1695	>	* Scans for and, if found, returns one task, else possibly
1696	>	* inactivates the worker. This method operates on single reads of
1697	>	* volatile state and is designed to be re-invoked continuously,
1698	>	* in part because it returns upon detecting inconsistencies,
1699	>	* contention, or state changes that indicate possible success on
1700	>	* re-invocation.
1701	>	*
1702	>	* The scan searches for tasks across a random permutation of
1703	>	* queues (starting at a random index and stepping by a random
1704	>	* relative prime, checking each at least once).  The scan
1705	>	* terminates upon either finding a non-empty queue, or completing
1706	>	* the sweep. If the worker is not inactivated, it takes and
1707	>	* returns a task from this queue.  On failure to find a task, we
1708	>	* take one of the following actions, after which the caller will
1709	>	* retry calling this method unless terminated.
1710	>	*
1711	>	* * If pool is terminating, terminate the worker.
1712	>	*
1713	>	* * If not a complete sweep, try to release a waiting worker.  If
1714	>	* the scan terminated because the worker is inactivated, then the
1715	>	* released worker will often be the calling worker, and it can
1716	>	* succeed obtaining a task on the next call. Or maybe it is
1717	>	* another worker, but with same net effect. Releasing in other
1718	>	* cases as well ensures that we have enough workers running.
1719	>	*
1720	>	* * If not already enqueued, try to inactivate and enqueue the
1721	>	* worker on wait queue. Or, if inactivating has caused the pool
1722	>	* to be quiescent, relay to idleAwaitWork to check for
1723	>	* termination and possibly shrink pool.
1724	>	*
1725	>	* * If already inactive, and the caller has run a task since the
1726	>	* last empty scan, return (to allow rescan) unless others are
1727	>	* also inactivated.  Field WorkQueue.rescans counts down on each
1728	>	* scan to ensure eventual inactivation and blocking.
1729	>	*
1730	>	* * If already enqueued and none of the above apply, park
1731	>	* awaiting signal,
1732	>	*
1733	>	* @param w the worker (via its WorkQueue)
1734	>	* @return a task or null if none found
1735	>	*/
1736	>	private final ForkJoinTask<?> scan(WorkQueue w) {
1737	>	WorkQueue[] ws;                       // first update random seed
1738	>	int r = w.seed; r ^= r << 13; r ^= r >>> 17; w.seed = r ^= r << 5;
1739	>	int rs = runState, m;                 // volatile read order matters
1740	>	if ((ws = workQueues) != null && (m = ws.length - 1) > 0) {
1741	>	int ec = w.eventCount;            // ec is negative if inactive
1742	>	int step = (r >>> 16) | 1;        // relative prime
1743	>	for (int j = (m + 1) << 2; ; r += step) {
1744	>	WorkQueue q; ForkJoinTask<?> t; ForkJoinTask<?>[] a; int b;
1745	>	if ((q = ws[r & m]) != null && (b = q.base) - q.top < 0 &&
1746	>	(a = q.array) != null) {  // probably nonempty
1747	>	int i = (((a.length - 1) & b) << ASHIFT) + ABASE;
1748	>	t = (ForkJoinTask<?>)U.getObjectVolatile(a, i);
1749	>	if (q.base == b && ec >= 0 && t != null &&
1750	>	U.compareAndSwapObject(a, i, t, null)) {
1751	>	if (q.top - (q.base = b + 1) > 0)
1752	>	signalWork();    // help pushes signal
1753	>	return t;
1754	>	}
1755	>	else if (ec < 0 || j <= m) {
1756	>	rs = 0;               // mark scan as imcomplete
1757	>	break;                // caller can retry after release
1758		}
1759		}
1760	<	if (w == null)
875	<	break;
876	<	w.start(recordWorker(w), ueh);
877	<	if ((workerCounts >>> TOTAL_COUNT_SHIFT) >= pc) {
878	<	int c; // advance event count
879	<	UNSAFE.compareAndSwapInt(this, eventCountOffset,
880	<	c = eventCount, c+1);
881	<	break; // add at most one unless total below target
882	<	}
883	<	}
884	<	}
885	<	if (eventWaiters != 0L)
886	<	releaseEventWaiters();
887	<	}
888	<
889	<	/**
890	<	* Callback from the oldest waiter in awaitEvent waking up after a
891	<	* period of non-use. If all workers are idle, tries (once) to
892	<	* shutdown an event waiter or a spare, if one exists. Note that
893	<	* we don't need CAS or locks here because the method is called
894	<	* only from one thread occasionally waking (and even misfires are
895	<	* OK). Note that until the shutdown worker fully terminates,
896	<	* workerCounts will overestimate total count, which is tolerable.
897	<	*
898	<	* @param ec the event count waited on by caller (to abort
899	<	* attempt if count has since changed).
900	<	*/
901	<	private void tryShutdownUnusedWorker(int ec) {
902	<	if (runState == 0 && eventCount == ec) { // only trigger if all idle
903	<	ForkJoinWorkerThread[] ws = workers;
904	<	int n = ws.length;
905	<	ForkJoinWorkerThread w = null;
906	<	boolean shutdown = false;
907	<	int sw;
908	<	long h;
909	<	if ((sw = spareWaiters) != 0) { // prefer killing spares
910	<	int id = (sw & SPARE_ID_MASK) - 1;
911	<	if (id >= 0 && id < n && (w = ws[id]) != null &&
912	<	UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
913	<	sw, w.nextSpare))
914	<	shutdown = true;
915	<	}
916	<	else if ((h = eventWaiters) != 0L) {
917	<	long nh;
918	<	int id = ((int)(h & WAITER_ID_MASK)) - 1;
919	<	if (id >= 0 && id < n && (w = ws[id]) != null &&
920	<	(nh = w.nextWaiter) != 0L && // keep at least one worker
921	<	UNSAFE.compareAndSwapLong(this, eventWaitersOffset, h, nh))
922	<	shutdown = true;
923	<	}
924	<	if (w != null && shutdown) {
925	<	w.shutdown();
926	<	LockSupport.unpark(w);
927	<	}
928	<	}
929	<	releaseEventWaiters(); // in case of interference
930	<	}
931	<
932	<	/**
933	<	* Callback from workers invoked upon each top-level action (i.e.,
934	<	* stealing a task or taking a submission and running it).
935	<	* Performs one or more of the following:
936	<	*
937	<	* 1. If the worker is active and either did not run a task
938	<	*    or there are too many workers, try to set its active status
939	<	*    to inactive and update activeCount. On contention, we may
940	<	*    try again in this or a subsequent call.
941	<	*
942	<	* 2. If not enough total workers, help create some.
943	<	*
944	<	* 3. If there are too many running workers, suspend this worker
945	<	*    (first forcing inactive if necessary).  If it is not needed,
946	<	*    it may be shutdown while suspended (via
947	<	*    tryShutdownUnusedWorker).  Otherwise, upon resume it
948	<	*    rechecks running thread count and need for event sync.
949	<	*
950	<	* 4. If worker did not run a task, await the next task event via
951	<	*    eventSync if necessary (first forcing inactivation), upon
952	<	*    which the worker may be shutdown via
953	<	*    tryShutdownUnusedWorker.  Otherwise, help release any
954	<	*    existing event waiters that are now releasable,
955	<	*
956	<	* @param w the worker
957	<	* @param ran true if worker ran a task since last call to this method
958	<	*/
959	<	final void preStep(ForkJoinWorkerThread w, boolean ran) {
960	<	int wec = w.lastEventCount;
961	<	boolean active = w.active;
962	<	boolean inactivate = false;
963	<	int pc = parallelism;
964	<	int rs;
965	<	while (w.runState == 0 && (rs = runState) < TERMINATING) {
966	<	if ((inactivate || (active && (rs & ACTIVE_COUNT_MASK) >= pc)) &&
967	<	UNSAFE.compareAndSwapInt(this, runStateOffset, rs, rs - 1))
968	<	inactivate = active = w.active = false;
969	<	int wc = workerCounts;
970	<	if ((wc & RUNNING_COUNT_MASK) > pc) {
971	<	if (!(inactivate |= active) && // must inactivate to suspend
972	<	workerCounts == wc &&      // try to suspend as spare
973	<	UNSAFE.compareAndSwapInt(this, workerCountsOffset,
974	<	wc, wc - ONE_RUNNING))
975	<	w.suspendAsSpare();
976	<	}
977	<	else if ((wc >>> TOTAL_COUNT_SHIFT) < pc)
978	<	helpMaintainParallelism();     // not enough workers
979	<	else if (!ran) {
980	<	long h = eventWaiters;
981	<	int ec = eventCount;
982	<	if (h != 0L && (int)(h >>> EVENT_COUNT_SHIFT) != ec)
983	<	releaseEventWaiters();     // release others before waiting
984	<	else if (ec != wec) {
985	<	w.lastEventCount = ec;     // no need to wait
1760	>	if (--j < 0)
1761		break;
1762	+	}
1763	+
1764	+	long c = ctl; int e = (int)c, a = (int)(c >> AC_SHIFT), nr, ns;
1765	+	if (e < 0)                        // decode ctl on empty scan
1766	+	w.runState = -1;              // pool is terminating
1767	+	else if (rs == 0 || rs != runState) { // incomplete scan
1768	+	WorkQueue v; Thread p;        // try to release a waiter
1769	+	if (e > 0 && a < 0 && w.eventCount == ec &&
1770	+	(v = ws[e & m]) != null && v.eventCount == (e | INT_SIGN)) {
1771	+	long nc = ((long)(v.nextWait & E_MASK) |
1772	+	((c + AC_UNIT) & (AC_MASK|TC_MASK)));
1773	+	if (ctl == c && U.compareAndSwapLong(this, CTL, c, nc)) {
1774	+	v.eventCount = (e + E_SEQ) & E_MASK;
1775	+	if ((p = v.parker) != null)
1776	+	U.unpark(p);
1777	+	}
1778	+	}
1779	+	}
1780	+	else if (ec >= 0) {               // try to enqueue/inactivate
1781	+	long nc = (long)ec | ((c - AC_UNIT) & (AC_MASK|TC_MASK));
1782	+	w.nextWait = e;
1783	+	w.eventCount = ec | INT_SIGN; // mark as inactive
1784	+	if (ctl != c || !U.compareAndSwapLong(this, CTL, c, nc))
1785	+	w.eventCount = ec;        // unmark on CAS failure
1786	+	else {
1787	+	if ((ns = w.nsteals) != 0) {
1788	+	w.nsteals = 0;        // set rescans if ran task
1789	+	w.rescans = (a > 0) ? 0 : a + parallelism;
1790	+	w.totalSteals += ns;
1791	+	}
1792	+	if (a == 1 - parallelism) // quiescent
1793	+	idleAwaitWork(w, nc, c);
1794	+	}
1795	+	}
1796	+	else if (w.eventCount < 0) {      // already queued
1797	+	int ac = a + parallelism;
1798	+	if ((nr = w.rescans) > 0)     // continue rescanning
1799	+	w.rescans = (ac < nr) ? ac : nr - 1;
1800	+	else if (((w.seed >>> 16) & ac) == 0) { // randomize park
1801	+	Thread.interrupted();     // clear status
1802	+	Thread wt = Thread.currentThread();
1803	+	U.putObject(wt, PARKBLOCKER, this);
1804	+	w.parker = wt;            // emulate LockSupport.park
1805	+	if (w.eventCount < 0)     // recheck
1806	+	U.park(false, 0L);
1807	+	w.parker = null;
1808	+	U.putObject(wt, PARKBLOCKER, null);
1809		}
988	–	else if (!(inactivate |= active))
989	–	eventSync(w, wec);         // must inactivate before sync
1810		}
991	–	else
992	–	break;
1811		}
1812	+	return null;
1813		}
1814
1815		/**
1816	<	* Helps and/or blocks awaiting join of the given task.
1817	<	* See above for explanation.
1818	<	*
1819	<	* @param joinMe the task to join
1820	<	* @param worker the current worker thread
1821	<	*/
1822	<	final void awaitJoin(ForkJoinTask<?> joinMe, ForkJoinWorkerThread worker) {
1823	<	int retries = 2 + (parallelism >> 2); // #helpJoins before blocking
1824	<	while (joinMe.status >= 0) {
1825	<	int wc;
1826	<	worker.helpJoinTask(joinMe);
1827	<	if (joinMe.status < 0)
1828	<	break;
1829	<	else if (retries > 0)
1830	<	--retries;
1831	<	else if (((wc = workerCounts) & RUNNING_COUNT_MASK) != 0 &&
1832	<	UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1833	<	wc, wc - ONE_RUNNING)) {
1834	<	int stat, c; long h;
1835	<	while ((stat = joinMe.status) >= 0 &&
1836	<	(h = eventWaiters) != 0L && // help release others
1837	<	(int)(h >>> EVENT_COUNT_SHIFT) != eventCount)
1838	<	releaseEventWaiters();
1839	<	if (stat >= 0 &&
1840	<	((workerCounts & RUNNING_COUNT_MASK) == 0 ||
1841	<	(stat =
1842	<	joinMe.internalAwaitDone(JOIN_TIMEOUT_MILLIS)) >= 0))
1843	<	helpMaintainParallelism(); // timeout or no running workers
1844	<	do {} while (!UNSAFE.compareAndSwapInt
1845	<	(this, workerCountsOffset,
1846	<	c = workerCounts, c + ONE_RUNNING));
1847	<	if (stat < 0)
1848	<	break;   // else restart
1816	>	* If inactivating worker w has caused the pool to become
1817	>	* quiescent, checks for pool termination, and, so long as this is
1818	>	* not the only worker, waits for event for up to a given
1819	>	* duration.  On timeout, if ctl has not changed, terminates the
1820	>	* worker, which will in turn wake up another worker to possibly
1821	>	* repeat this process.
1822	>	*
1823	>	* @param w the calling worker
1824	>	* @param currentCtl the ctl value triggering possible quiescence
1825	>	* @param prevCtl the ctl value to restore if thread is terminated
1826	>	*/
1827	>	private void idleAwaitWork(WorkQueue w, long currentCtl, long prevCtl) {
1828	>	if (w.eventCount < 0 && !tryTerminate(false, false) &&
1829	>	(int)prevCtl != 0 && !hasQueuedSubmissions() && ctl == currentCtl) {
1830	>	int dc = -(short)(currentCtl >>> TC_SHIFT);
1831	>	long parkTime = dc < 0 ? FAST_IDLE_TIMEOUT: (dc + 1) * IDLE_TIMEOUT;
1832	>	long deadline = System.nanoTime() + parkTime - 100000L; // 1ms slop
1833	>	Thread wt = Thread.currentThread();
1834	>	while (ctl == currentCtl) {
1835	>	Thread.interrupted();  // timed variant of version in scan()
1836	>	U.putObject(wt, PARKBLOCKER, this);
1837	>	w.parker = wt;
1838	>	if (ctl == currentCtl)
1839	>	U.park(false, parkTime);
1840	>	w.parker = null;
1841	>	U.putObject(wt, PARKBLOCKER, null);
1842	>	if (ctl != currentCtl)
1843	>	break;
1844	>	if (deadline - System.nanoTime() <= 0L &&
1845	>	U.compareAndSwapLong(this, CTL, currentCtl, prevCtl)) {
1846	>	w.eventCount = (w.eventCount + E_SEQ) | E_MASK;
1847	>	w.runState = -1;   // shrink
1848	>	break;
1849	>	}
1850		}
1851		}
1852		}
1853
1854		/**
1855	<	* Same idea as awaitJoin, but no helping, retries, or timeouts.
1856	<	*/
1857	<	final void awaitBlocker(ManagedBlocker blocker)
1858	<	throws InterruptedException {
1859	<	while (!blocker.isReleasable()) {
1860	<	int wc = workerCounts;
1861	<	if ((wc & RUNNING_COUNT_MASK) != 0 &&
1862	<	UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1863	<	wc, wc - ONE_RUNNING)) {
1864	<	try {
1865	<	while (!blocker.isReleasable()) {
1866	<	long h = eventWaiters;
1867	<	if (h != 0L &&
1868	<	(int)(h >>> EVENT_COUNT_SHIFT) != eventCount)
1869	<	releaseEventWaiters();
1870	<	else if ((workerCounts & RUNNING_COUNT_MASK) == 0 &&
1871	<	runState < TERMINATING)
1872	<	helpMaintainParallelism();
1873	<	else if (blocker.block())
1874	<	break;
1855	>	* Tries to locate and execute tasks for a stealer of the given
1856	>	* task, or in turn one of its stealers, Traces currentSteal ->
1857	>	* currentJoin links looking for a thread working on a descendant
1858	>	* of the given task and with a non-empty queue to steal back and
1859	>	* execute tasks from. The first call to this method upon a
1860	>	* waiting join will often entail scanning/search, (which is OK
1861	>	* because the joiner has nothing better to do), but this method
1862	>	* leaves hints in workers to speed up subsequent calls. The
1863	>	* implementation is very branchy to cope with potential
1864	>	* inconsistencies or loops encountering chains that are stale,
1865	>	* unknown, or so long that they are likely cyclic.
1866	>	*
1867	>	* @param joiner the joining worker
1868	>	* @param task the task to join
1869	>	* @return 0 if no progress can be made, negative if task
1870	>	* known complete, else positive
1871	>	*/
1872	>	private int tryHelpStealer(WorkQueue joiner, ForkJoinTask<?> task) {
1873	>	int stat = 0, steps = 0;                    // bound to avoid cycles
1874	>	if (joiner != null && task != null) {       // hoist null checks
1875	>	restart: for (;;) {
1876	>	ForkJoinTask<?> subtask = task;     // current target
1877	>	for (WorkQueue j = joiner, v;;) {   // v is stealer of subtask
1878	>	WorkQueue[] ws; int m, s, h;
1879	>	if ((s = task.status) < 0) {
1880	>	stat = s;
1881	>	break restart;
1882	>	}
1883	>	if ((ws = workQueues) == null || (m = ws.length - 1) <= 0)
1884	>	break restart;              // shutting down
1885	>	if ((v = ws[h = (j.stealHint | 1) & m]) == null ||
1886	>	v.currentSteal != subtask) {
1887	>	for (int origin = h;;) {    // find stealer
1888	>	if (((h = (h + 2) & m) & 15) == 1 &&
1889	>	(subtask.status < 0 || j.currentJoin != subtask))
1890	>	continue restart;   // occasional staleness check
1891	>	if ((v = ws[h]) != null &&
1892	>	v.currentSteal == subtask) {
1893	>	j.stealHint = h;    // save hint
1894	>	break;
1895	>	}
1896	>	if (h == origin)
1897	>	break restart;      // cannot find stealer
1898	>	}
1899	>	}
1900	>	for (;;) { // help stealer or descend to its stealer
1901	>	ForkJoinTask[] a;  int b;
1902	>	if (subtask.status < 0)     // surround probes with
1903	>	continue restart;       //   consistency checks
1904	>	if ((b = v.base) - v.top < 0 && (a = v.array) != null) {
1905	>	int i = (((a.length - 1) & b) << ASHIFT) + ABASE;
1906	>	ForkJoinTask<?> t =
1907	>	(ForkJoinTask<?>)U.getObjectVolatile(a, i);
1908	>	if (subtask.status < 0 || j.currentJoin != subtask ||
1909	>	v.currentSteal != subtask)
1910	>	continue restart;   // stale
1911	>	stat = 1;               // apparent progress
1912	>	if (t != null && v.base == b &&
1913	>	U.compareAndSwapObject(a, i, t, null)) {
1914	>	v.base = b + 1;     // help stealer
1915	>	joiner.runSubtask(t);
1916	>	}
1917	>	else if (v.base == b && ++steps == MAX_HELP)
1918	>	break restart;      // v apparently stalled
1919	>	}
1920	>	else {                      // empty -- try to descend
1921	>	ForkJoinTask<?> next = v.currentJoin;
1922	>	if (subtask.status < 0 || j.currentJoin != subtask ||
1923	>	v.currentSteal != subtask)
1924	>	continue restart;   // stale
1925	>	else if (next == null || ++steps == MAX_HELP)
1926	>	break restart;      // dead-end or maybe cyclic
1927	>	else {
1928	>	subtask = next;
1929	>	j = v;
1930	>	break;
1931	>	}
1932	>	}
1933		}
1056	–	} finally {
1057	–	int c;
1058	–	do {} while (!UNSAFE.compareAndSwapInt
1059	–	(this, workerCountsOffset,
1060	–	c = workerCounts, c + ONE_RUNNING));
1934		}
1062	–	break;
1935		}
1936		}
1937	+	return stat;
1938		}
1939
1940		/**
1941	<	* Possibly initiates and/or completes termination.
1941	>	* If task is at base of some steal queue, steals and executes it.
1942		*
1943	<	* @param now if true, unconditionally terminate, else only
1944	<	* if shutdown and empty queue and no active workers
1072	<	* @return true if now terminating or terminated
1943	>	* @param joiner the joining worker
1944	>	* @param task the task
1945		*/
1946	<	private boolean tryTerminate(boolean now) {
1947	<	if (now)
1948	<	advanceRunLevel(SHUTDOWN); // ensure at least SHUTDOWN
1949	<	else if (runState < SHUTDOWN ||
1950	<	!submissionQueue.isEmpty() ||
1951	<	(runState & ACTIVE_COUNT_MASK) != 0)
1952	<	return false;
1953	<
1954	<	if (advanceRunLevel(TERMINATING))
1955	<	startTerminating();
1946	>	private void tryPollForAndExec(WorkQueue joiner, ForkJoinTask<?> task) {
1947	>	WorkQueue[] ws;
1948	>	if ((ws = workQueues) != null) {
1949	>	for (int j = 1; j < ws.length && task.status >= 0; j += 2) {
1950	>	WorkQueue q = ws[j];
1951	>	if (q != null && q.pollFor(task)) {
1952	>	joiner.runSubtask(task);
1953	>	break;
1954	>	}
1955	>	}
1956	>	}
1957	>	}
1958
1959	<	// Finish now if all threads terminated; else in some subsequent call
1960	<	if ((workerCounts >>> TOTAL_COUNT_SHIFT) == 0) {
1961	<	advanceRunLevel(TERMINATED);
1962	<	termination.arrive();
1963	<	}
1964	<	return true;
1959	>	/**
1960	>	* Tries to decrement active count (sometimes implicitly) and
1961	>	* possibly release or create a compensating worker in preparation
1962	>	* for blocking. Fails on contention or termination. Otherwise,
1963	>	* adds a new thread if no idle workers are available and either
1964	>	* pool would become completely starved or: (at least half
1965	>	* starved, and fewer than 50% spares exist, and there is at least
1966	>	* one task apparently available). Even though the availability
1967	>	* check requires a full scan, it is worthwhile in reducing false
1968	>	* alarms.
1969	>	*
1970	>	* @param task if non-null, a task being waited for
1971	>	* @param blocker if non-null, a blocker being waited for
1972	>	* @return true if the caller can block, else should recheck and retry
1973	>	*/
1974	>	final boolean tryCompensate(ForkJoinTask<?> task, ManagedBlocker blocker) {
1975	>	int pc = parallelism, e;
1976	>	long c = ctl;
1977	>	WorkQueue[] ws = workQueues;
1978	>	if ((e = (int)c) >= 0 && ws != null) {
1979	>	int u, a, ac, hc;
1980	>	int tc = (short)((u = (int)(c >>> 32)) >>> UTC_SHIFT) + pc;
1981	>	boolean replace = false;
1982	>	if ((a = u >> UAC_SHIFT) <= 0) {
1983	>	if ((ac = a + pc) <= 1)
1984	>	replace = true;
1985	>	else if ((e > 0 || (task != null &&
1986	>	ac <= (hc = pc >>> 1) && tc < pc + hc))) {
1987	>	WorkQueue w;
1988	>	for (int j = 0; j < ws.length; ++j) {
1989	>	if ((w = ws[j]) != null && !w.isEmpty()) {
1990	>	replace = true;
1991	>	break;   // in compensation range and tasks available
1992	>	}
1993	>	}
1994	>	}
1995	>	}
1996	>	if ((task == null || task.status >= 0) && // recheck need to block
1997	>	(blocker == null || !blocker.isReleasable()) && ctl == c) {
1998	>	if (!replace) {          // no compensation
1999	>	long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK);
2000	>	if (U.compareAndSwapLong(this, CTL, c, nc))
2001	>	return true;
2002	>	}
2003	>	else if (e != 0) {       // release an idle worker
2004	>	WorkQueue w; Thread p; int i;
2005	>	if ((i = e & SMASK) < ws.length && (w = ws[i]) != null) {
2006	>	long nc = ((long)(w.nextWait & E_MASK) |
2007	>	(c & (AC_MASK|TC_MASK)));
2008	>	if (w.eventCount == (e | INT_SIGN) &&
2009	>	U.compareAndSwapLong(this, CTL, c, nc)) {
2010	>	w.eventCount = (e + E_SEQ) & E_MASK;
2011	>	if ((p = w.parker) != null)
2012	>	U.unpark(p);
2013	>	return true;
2014	>	}
2015	>	}
2016	>	}
2017	>	else if (tc < MAX_CAP) { // create replacement
2018	>	long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK);
2019	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
2020	>	addWorker();
2021	>	return true;
2022	>	}
2023	>	}
2024	>	}
2025	>	}
2026	>	return false;
2027		}
2028
2029		/**
2030	<	* Actions on transition to TERMINATING
2031	<	*
2032	<	* Runs up to four passes through workers: (0) shutting down each
2033	<	* (without waking up if parked) to quickly spread notifications
2034	<	* without unnecessary bouncing around event queues etc (1) wake
2035	<	* up and help cancel tasks (2) interrupt (3) mop up races with
2036	<	* interrupted workers
2037	<	*/
2038	<	private void startTerminating() {
2039	<	cancelSubmissions();
2040	<	for (int passes = 0; passes < 4 && workerCounts != 0; ++passes) {
2041	<	int c; // advance event count
2042	<	UNSAFE.compareAndSwapInt(this, eventCountOffset,
2043	<	c = eventCount, c+1);
2044	<	eventWaiters = 0L; // clobber lists
2045	<	spareWaiters = 0;
2046	<	ForkJoinWorkerThread[] ws = workers;
2047	<	int n = ws.length;
2048	<	for (int i = 0; i < n; ++i) {
2049	<	ForkJoinWorkerThread w = ws[i];
2050	<	if (w != null) {
2051	<	w.shutdown();
2052	<	if (passes > 0 && !w.isTerminated()) {
2053	<	w.cancelTasks();
2054	<	LockSupport.unpark(w);
2055	<	if (passes > 1) {
2056	<	try {
2057	<	w.interrupt();
2058	<	} catch (SecurityException ignore) {
2030	>	* Helps and/or blocks until the given task is done.
2031	>	*
2032	>	* @param joiner the joining worker
2033	>	* @param task the task
2034	>	* @return task status on exit
2035	>	*/
2036	>	final int awaitJoin(WorkQueue joiner, ForkJoinTask<?> task) {
2037	>	int s;
2038	>	if ((s = task.status) >= 0) {
2039	>	ForkJoinTask<?> prevJoin = joiner.currentJoin;
2040	>	joiner.currentJoin = task;
2041	>	long startTime = 0L;
2042	>	for (int k = 0;;) {
2043	>	if ((s = (joiner.isEmpty() ?           // try to help
2044	>	tryHelpStealer(joiner, task) :
2045	>	joiner.tryRemoveAndExec(task))) == 0 &&
2046	>	(s = task.status) >= 0) {
2047	>	if (k == 0) {
2048	>	startTime = System.nanoTime();
2049	>	tryPollForAndExec(joiner, task); // check uncommon case
2050	>	}
2051	>	else if ((k & (MAX_HELP - 1)) == 0 &&
2052	>	System.nanoTime() - startTime >=
2053	>	COMPENSATION_DELAY &&
2054	>	tryCompensate(task, null)) {
2055	>	if (task.trySetSignal()) {
2056	>	synchronized (task) {
2057	>	if (task.status >= 0) {
2058	>	try {                // see ForkJoinTask
2059	>	task.wait();     //  for explanation
2060	>	} catch (InterruptedException ie) {
2061	>	}
2062	>	}
2063	>	else
2064	>	task.notifyAll();
2065		}
2066		}
2067	+	long c;                          // re-activate
2068	+	do {} while (!U.compareAndSwapLong
2069	+	(this, CTL, c = ctl, c + AC_UNIT));
2070		}
2071		}
2072	+	if (s < 0 || (s = task.status) < 0) {
2073	+	joiner.currentJoin = prevJoin;
2074	+	break;
2075	+	}
2076	+	else if ((k++ & (MAX_HELP - 1)) == MAX_HELP >>> 1)
2077	+	Thread.yield();                     // for politeness
2078		}
2079		}
2080	+	return s;
2081		}
2082
2083		/**
2084	<	* Clear out and cancel submissions, ignoring exceptions
2084	>	* Stripped-down variant of awaitJoin used by timed joins. Tries
2085	>	* to help join only while there is continuous progress. (Caller
2086	>	* will then enter a timed wait.)
2087	>	*
2088	>	* @param joiner the joining worker
2089	>	* @param task the task
2090	>	* @return task status on exit
2091		*/
2092	<	private void cancelSubmissions() {
2093	<	ForkJoinTask<?> task;
2094	<	while ((task = submissionQueue.poll()) != null) {
2095	<	try {
2096	<	task.cancel(false);
2097	<	} catch (Throwable ignore) {
2092	>	final int helpJoinOnce(WorkQueue joiner, ForkJoinTask<?> task) {
2093	>	int s;
2094	>	while ((s = task.status) >= 0 &&
2095	>	(joiner.isEmpty() ?
2096	>	tryHelpStealer(joiner, task) :
2097	>	joiner.tryRemoveAndExec(task)) != 0)
2098	>	;
2099	>	return s;
2100	>	}
2101	>
2102	>	/**
2103	>	* Returns a (probably) non-empty steal queue, if one is found
2104	>	* during a random, then cyclic scan, else null.  This method must
2105	>	* be retried by caller if, by the time it tries to use the queue,
2106	>	* it is empty.
2107	>	*/
2108	>	private WorkQueue findNonEmptyStealQueue(WorkQueue w) {
2109	>	// Similar to loop in scan(), but ignoring submissions
2110	>	int r = w.seed; r ^= r << 13; r ^= r >>> 17; w.seed = r ^= r << 5;
2111	>	int step = (r >>> 16) | 1;
2112	>	for (WorkQueue[] ws;;) {
2113	>	int rs = runState, m;
2114	>	if ((ws = workQueues) == null || (m = ws.length - 1) < 1)
2115	>	return null;
2116	>	for (int j = (m + 1) << 2; ; r += step) {
2117	>	WorkQueue q = ws[((r << 1) | 1) & m];
2118	>	if (q != null && !q.isEmpty())
2119	>	return q;
2120	>	else if (--j < 0) {
2121	>	if (runState == rs)
2122	>	return null;
2123	>	break;
2124	>	}
2125		}
2126		}
2127		}
2128
2129	<	// misc support for ForkJoinWorkerThread
2129	>	/**
2130	>	* Runs tasks until {@code isQuiescent()}. We piggyback on
2131	>	* active count ctl maintenance, but rather than blocking
2132	>	* when tasks cannot be found, we rescan until all others cannot
2133	>	* find tasks either.
2134	>	*/
2135	>	final void helpQuiescePool(WorkQueue w) {
2136	>	for (boolean active = true;;) {
2137	>	ForkJoinTask<?> localTask; // exhaust local queue
2138	>	while ((localTask = w.nextLocalTask()) != null)
2139	>	localTask.doExec();
2140	>	WorkQueue q = findNonEmptyStealQueue(w);
2141	>	if (q != null) {
2142	>	ForkJoinTask<?> t; int b;
2143	>	if (!active) {      // re-establish active count
2144	>	long c;
2145	>	active = true;
2146	>	do {} while (!U.compareAndSwapLong
2147	>	(this, CTL, c = ctl, c + AC_UNIT));
2148	>	}
2149	>	if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null)
2150	>	w.runSubtask(t);
2151	>	}
2152	>	else {
2153	>	long c;
2154	>	if (active) {       // decrement active count without queuing
2155	>	active = false;
2156	>	do {} while (!U.compareAndSwapLong
2157	>	(this, CTL, c = ctl, c -= AC_UNIT));
2158	>	}
2159	>	else
2160	>	c = ctl;        // re-increment on exit
2161	>	if ((int)(c >> AC_SHIFT) + parallelism == 0) {
2162	>	do {} while (!U.compareAndSwapLong
2163	>	(this, CTL, c = ctl, c + AC_UNIT));
2164	>	break;
2165	>	}
2166	>	}
2167	>	}
2168	>	}
2169
2170		/**
2171	<	* Returns pool number
2171	>	* Restricted version of helpQuiescePool for non-FJ callers
2172		*/
2173	<	final int getPoolNumber() {
2174	<	return poolNumber;
2173	>	static void externalHelpQuiescePool() {
2174	>	ForkJoinPool p; WorkQueue[] ws; WorkQueue q, sq;
2175	>	ForkJoinTask<?>[] a; int b;
2176	>	ForkJoinTask<?> t = null;
2177	>	int k = submitters.get().seed & SQMASK;
2178	>	if ((p = commonPool) != null &&
2179	>	(ws = p.workQueues) != null &&
2180	>	ws.length > (k &= p.submitMask) &&
2181	>	(q = ws[k]) != null) {
2182	>	while (q.top - q.base > 0) {
2183	>	if ((t = q.sharedPop()) != null)
2184	>	break;
2185	>	}
2186	>	if (t == null && (sq = p.findNonEmptyStealQueue(q)) != null &&
2187	>	(b = sq.base) - sq.top < 0)
2188	>	t = sq.pollAt(b);
2189	>	if (t != null)
2190	>	t.doExec();
2191	>	}
2192		}
2193
2194		/**
2195	<	* Tries to accumulates steal count from a worker, clearing
1155	<	* the worker's value.
2195	>	* Gets and removes a local or stolen task for the given worker.
2196		*
2197	<	* @return true if worker steal count now zero
2197	>	* @return a task, if available
2198		*/
2199	<	final boolean tryAccumulateStealCount(ForkJoinWorkerThread w) {
2200	<	int sc = w.stealCount;
2201	<	long c = stealCount;
2202	<	// CAS even if zero, for fence effects
2203	<	if (UNSAFE.compareAndSwapLong(this, stealCountOffset, c, c + sc)) {
2204	<	if (sc != 0)
2205	<	w.stealCount = 0;
2206	<	return true;
2199	>	final ForkJoinTask<?> nextTaskFor(WorkQueue w) {
2200	>	for (ForkJoinTask<?> t;;) {
2201	>	WorkQueue q; int b;
2202	>	if ((t = w.nextLocalTask()) != null)
2203	>	return t;
2204	>	if ((q = findNonEmptyStealQueue(w)) == null)
2205	>	return null;
2206	>	if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null)
2207	>	return t;
2208		}
1168	–	return sc == 0;
2209		}
2210
2211		/**
2212		* Returns the approximate (non-atomic) number of idle threads per
2213	<	* active thread.
2213	>	* active thread to offset steal queue size for method
2214	>	* ForkJoinTask.getSurplusQueuedTaskCount().
2215		*/
2216		final int idlePerActive() {
2217	<	int pc = parallelism; // use parallelism, not rc
2218	<	int ac = runState;    // no mask -- artificially boosts during shutdown
2219	<	// Use exact results for small values, saturate past 4
2220	<	return pc <= ac? 0 : pc >>> 1 <= ac? 1 : pc >>> 2 <= ac? 3 : pc >>> 3;
2217	>	// Approximate at powers of two for small values, saturate past 4
2218	>	int p = parallelism;
2219	>	int a = p + (int)(ctl >> AC_SHIFT);
2220	>	return (a > (p >>>= 1) ? 0 :
2221	>	a > (p >>>= 1) ? 1 :
2222	>	a > (p >>>= 1) ? 2 :
2223	>	a > (p >>>= 1) ? 4 :
2224	>	8);
2225	>	}
2226	>
2227	>	/**
2228	>	* Returns approximate submission queue length for the given caller
2229	>	*/
2230	>	static int getEstimatedSubmitterQueueLength() {
2231	>	ForkJoinPool p; WorkQueue[] ws; WorkQueue q;
2232	>	int k = submitters.get().seed & SQMASK;
2233	>	return ((p = commonPool) != null && (ws = p.workQueues) != null &&
2234	>	ws.length > (k &= p.submitMask) &&
2235	>	(q = ws[k]) != null) ?
2236	>	q.queueSize() : 0;
2237	>	}
2238	>
2239	>	//  Termination
2240	>
2241	>	/**
2242	>	* Possibly initiates and/or completes termination.  The caller
2243	>	* triggering termination runs three passes through workQueues:
2244	>	* (0) Setting termination status, followed by wakeups of queued
2245	>	* workers; (1) cancelling all tasks; (2) interrupting lagging
2246	>	* threads (likely in external tasks, but possibly also blocked in
2247	>	* joins).  Each pass repeats previous steps because of potential
2248	>	* lagging thread creation.
2249	>	*
2250	>	* @param now if true, unconditionally terminate, else only
2251	>	* if no work and no active workers
2252	>	* @param enable if true, enable shutdown when next possible
2253	>	* @return true if now terminating or terminated
2254	>	*/
2255	>	private boolean tryTerminate(boolean now, boolean enable) {
2256	>	for (long c;;) {
2257	>	if (((c = ctl) & STOP_BIT) != 0) {      // already terminating
2258	>	if ((short)(c >>> TC_SHIFT) == -parallelism) {
2259	>	synchronized (this) {
2260	>	notifyAll();                // signal when 0 workers
2261	>	}
2262	>	}
2263	>	return true;
2264	>	}
2265	>	if (runState >= 0) {                    // not yet enabled
2266	>	if (!enable)
2267	>	return false;
2268	>	while (!U.compareAndSwapInt(this, MAINLOCK, 0, 1))
2269	>	tryAwaitMainLock();
2270	>	try {
2271	>	runState |= SHUTDOWN;
2272	>	} finally {
2273	>	if (!U.compareAndSwapInt(this, MAINLOCK, 1, 0)) {
2274	>	mainLock = 0;
2275	>	synchronized (this) { notifyAll(); };
2276	>	}
2277	>	}
2278	>	}
2279	>	if (!now) {                             // check if idle & no tasks
2280	>	if ((int)(c >> AC_SHIFT) != -parallelism ||
2281	>	hasQueuedSubmissions())
2282	>	return false;
2283	>	// Check for unqueued inactive workers. One pass suffices.
2284	>	WorkQueue[] ws = workQueues; WorkQueue w;
2285	>	if (ws != null) {
2286	>	for (int i = 1; i < ws.length; i += 2) {
2287	>	if ((w = ws[i]) != null && w.eventCount >= 0)
2288	>	return false;
2289	>	}
2290	>	}
2291	>	}
2292	>	if (U.compareAndSwapLong(this, CTL, c, c | STOP_BIT)) {
2293	>	for (int pass = 0; pass < 3; ++pass) {
2294	>	WorkQueue[] ws = workQueues;
2295	>	if (ws != null) {
2296	>	WorkQueue w;
2297	>	int n = ws.length;
2298	>	for (int i = 0; i < n; ++i) {
2299	>	if ((w = ws[i]) != null) {
2300	>	w.runState = -1;
2301	>	if (pass > 0) {
2302	>	w.cancelAll();
2303	>	if (pass > 1)
2304	>	w.interruptOwner();
2305	>	}
2306	>	}
2307	>	}
2308	>	// Wake up workers parked on event queue
2309	>	int i, e; long cc; Thread p;
2310	>	while ((e = (int)(cc = ctl) & E_MASK) != 0 &&
2311	>	(i = e & SMASK) < n &&
2312	>	(w = ws[i]) != null) {
2313	>	long nc = ((long)(w.nextWait & E_MASK) |
2314	>	((cc + AC_UNIT) & AC_MASK) |
2315	>	(cc & (TC_MASK|STOP_BIT)));
2316	>	if (w.eventCount == (e | INT_SIGN) &&
2317	>	U.compareAndSwapLong(this, CTL, cc, nc)) {
2318	>	w.eventCount = (e + E_SEQ) & E_MASK;
2319	>	w.runState = -1;
2320	>	if ((p = w.parker) != null)
2321	>	U.unpark(p);
2322	>	}
2323	>	}
2324	>	}
2325	>	}
2326	>	}
2327	>	}
2328		}
2329
2330	<	// Public and protected methods
2330	>	// Exported methods
2331
2332		// Constructors
2333
#	Line 1226 | Line 2374 | public class ForkJoinPool extends Abstra
2374		* use {@link #defaultForkJoinWorkerThreadFactory}.
2375		* @param handler the handler for internal worker threads that
2376		* terminate due to unrecoverable errors encountered while executing
2377	<	* tasks. For default value, use <code>null</code>.
2377	>	* tasks. For default value, use {@code null}.
2378		* @param asyncMode if true,
2379		* establishes local first-in-first-out scheduling mode for forked
2380		* tasks that are never joined. This mode may be more appropriate
2381		* than default locally stack-based mode in applications in which
2382		* worker threads only process event-style asynchronous tasks.
2383	<	* For default value, use <code>false</code>.
2383	>	* For default value, use {@code false}.
2384		* @throws IllegalArgumentException if parallelism less than or
2385		*         equal to zero, or greater than implementation limit
2386		* @throws NullPointerException if the factory is null
#	Line 1248 | Line 2396 | public class ForkJoinPool extends Abstra
2396		checkPermission();
2397		if (factory == null)
2398		throw new NullPointerException();
2399	<	if (parallelism <= 0 || parallelism > MAX_WORKERS)
2399	>	if (parallelism <= 0 || parallelism > MAX_CAP)
2400		throw new IllegalArgumentException();
2401		this.parallelism = parallelism;
2402		this.factory = factory;
2403		this.ueh = handler;
2404	<	this.locallyFifo = asyncMode;
2405	<	int arraySize = initialArraySizeFor(parallelism);
2406	<	this.workers = new ForkJoinWorkerThread[arraySize];
2407	<	this.submissionQueue = new LinkedTransferQueue<ForkJoinTask<?>>();
2408	<	this.workerLock = new ReentrantLock();
2409	<	this.termination = new Phaser(1);
2410	<	this.poolNumber = poolNumberGenerator.incrementAndGet();
2404	>	this.localMode = asyncMode ? FIFO_QUEUE : LIFO_QUEUE;
2405	>	long np = (long)(-parallelism); // offset ctl counts
2406	>	this.ctl = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK);
2407	>	// Use nearest power 2 for workQueues size. See Hackers Delight sec 3.2.
2408	>	int n = parallelism - 1;
2409	>	n |= n >>> 1; n |= n >>> 2; n |= n >>> 4; n |= n >>> 8; n |= n >>> 16;
2410	>	this.submitMask = ((n + 1) << 1) - 1;
2411	>	int pn = poolNumberGenerator.incrementAndGet();
2412	>	StringBuilder sb = new StringBuilder("ForkJoinPool-");
2413	>	sb.append(Integer.toString(pn));
2414	>	sb.append("-worker-");
2415	>	this.workerNamePrefix = sb.toString();
2416	>	this.runState = 1;              // set init flag
2417		}
2418
2419		/**
2420	<	* Returns initial power of two size for workers array.
2421	<	* @param pc the initial parallelism level
2422	<	*/
2423	<	private static int initialArraySizeFor(int pc) {
2424	<	// If possible, initially allocate enough space for one spare
2425	<	int size = pc < MAX_WORKERS ? pc + 1 : MAX_WORKERS;
2426	<	// See Hackers Delight, sec 3.2. We know MAX_WORKERS < (1 >>> 16)
2427	<	size |= size >>> 1;
2428	<	size |= size >>> 2;
2429	<	size |= size >>> 4;
2430	<	size |= size >>> 8;
2431	<	return size + 1;
2420	>	* Constructor for common pool, suitable only for static initialization.
2421	>	* Basically the same as above, but uses smallest possible initial footprint.
2422	>	*/
2423	>	ForkJoinPool(int parallelism, int submitMask,
2424	>	ForkJoinWorkerThreadFactory factory,
2425	>	Thread.UncaughtExceptionHandler handler) {
2426	>	this.factory = factory;
2427	>	this.ueh = handler;
2428	>	this.submitMask = submitMask;
2429	>	this.parallelism = parallelism;
2430	>	long np = (long)(-parallelism);
2431	>	this.ctl = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK);
2432	>	this.localMode = LIFO_QUEUE;
2433	>	this.workerNamePrefix = "ForkJoinPool.commonPool-worker-";
2434	>	this.runState = 1;
2435		}
2436
1280	–	// Execution methods
1281	–
2437		/**
2438	<	* Common code for execute, invoke and submit
2438	>	* Returns the common pool instance.
2439	>	*
2440	>	* @return the common pool instance
2441		*/
2442	<	private <T> void doSubmit(ForkJoinTask<T> task) {
2443	<	if (task == null)
2444	<	throw new NullPointerException();
2445	<	if (runState >= SHUTDOWN)
2446	<	throw new RejectedExecutionException();
1290	<	submissionQueue.offer(task);
1291	<	int c; // try to increment event count -- CAS failure OK
1292	<	UNSAFE.compareAndSwapInt(this, eventCountOffset, c = eventCount, c+1);
1293	<	helpMaintainParallelism(); // create, start, or resume some workers
2442	>	public static ForkJoinPool commonPool() {
2443	>	ForkJoinPool p;
2444	>	if ((p = commonPool) == null)
2445	>	throw new Error("Common Pool Unavailable");
2446	>	return p;
2447		}
2448
2449	+	// Execution methods
2450	+
2451		/**
2452		* Performs the given task, returning its result upon completion.
2453	+	* If the computation encounters an unchecked Exception or Error,
2454	+	* it is rethrown as the outcome of this invocation.  Rethrown
2455	+	* exceptions behave in the same way as regular exceptions, but,
2456	+	* when possible, contain stack traces (as displayed for example
2457	+	* using {@code ex.printStackTrace()}) of both the current thread
2458	+	* as well as the thread actually encountering the exception;
2459	+	* minimally only the latter.
2460		*
2461		* @param task the task
2462		* @return the task's result
#	Line 1303 | Line 2465 | public class ForkJoinPool extends Abstra
2465		*         scheduled for execution
2466		*/
2467		public <T> T invoke(ForkJoinTask<T> task) {
2468	+	if (task == null)
2469	+	throw new NullPointerException();
2470		doSubmit(task);
2471		return task.join();
2472		}
#	Line 1316 | Line 2480 | public class ForkJoinPool extends Abstra
2480		*         scheduled for execution
2481		*/
2482		public void execute(ForkJoinTask<?> task) {
2483	+	if (task == null)
2484	+	throw new NullPointerException();
2485		doSubmit(task);
2486		}
2487
#	Line 1327 | Line 2493 | public class ForkJoinPool extends Abstra
2493		*         scheduled for execution
2494		*/
2495		public void execute(Runnable task) {
2496	+	if (task == null)
2497	+	throw new NullPointerException();
2498		ForkJoinTask<?> job;
2499		if (task instanceof ForkJoinTask<?>) // avoid re-wrap
2500		job = (ForkJoinTask<?>) task;
2501		else
2502	<	job = ForkJoinTask.adapt(task, null);
2502	>	job = new ForkJoinTask.AdaptedRunnableAction(task);
2503		doSubmit(job);
2504		}
2505
#	Line 1345 | Line 2513 | public class ForkJoinPool extends Abstra
2513		*         scheduled for execution
2514		*/
2515		public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) {
2516	+	if (task == null)
2517	+	throw new NullPointerException();
2518		doSubmit(task);
2519		return task;
2520		}
#	Line 1355 | Line 2525 | public class ForkJoinPool extends Abstra
2525		*         scheduled for execution
2526		*/
2527		public <T> ForkJoinTask<T> submit(Callable<T> task) {
2528	<	ForkJoinTask<T> job = ForkJoinTask.adapt(task);
2528	>	ForkJoinTask<T> job = new ForkJoinTask.AdaptedCallable<T>(task);
2529		doSubmit(job);
2530		return job;
2531		}
#	Line 1366 | Line 2536 | public class ForkJoinPool extends Abstra
2536		*         scheduled for execution
2537		*/
2538		public <T> ForkJoinTask<T> submit(Runnable task, T result) {
2539	<	ForkJoinTask<T> job = ForkJoinTask.adapt(task, result);
2539	>	ForkJoinTask<T> job = new ForkJoinTask.AdaptedRunnable<T>(task, result);
2540		doSubmit(job);
2541		return job;
2542		}
#	Line 1377 | Line 2547 | public class ForkJoinPool extends Abstra
2547		*         scheduled for execution
2548		*/
2549		public ForkJoinTask<?> submit(Runnable task) {
2550	+	if (task == null)
2551	+	throw new NullPointerException();
2552		ForkJoinTask<?> job;
2553		if (task instanceof ForkJoinTask<?>) // avoid re-wrap
2554		job = (ForkJoinTask<?>) task;
2555		else
2556	<	job = ForkJoinTask.adapt(task, null);
2556	>	job = new ForkJoinTask.AdaptedRunnableAction(task);
2557		doSubmit(job);
2558		return job;
2559		}
#	Line 1391 | Line 2563 | public class ForkJoinPool extends Abstra
2563		* @throws RejectedExecutionException {@inheritDoc}
2564		*/
2565		public <T> List<Future<T>> invokeAll(Collection<? extends Callable<T>> tasks) {
2566	<	ArrayList<ForkJoinTask<T>> forkJoinTasks =
2567	<	new ArrayList<ForkJoinTask<T>>(tasks.size());
2568	<	for (Callable<T> task : tasks)
2569	<	forkJoinTasks.add(ForkJoinTask.adapt(task));
2570	<	invoke(new InvokeAll<T>(forkJoinTasks));
2571	<
2566	>	// In previous versions of this class, this method constructed
2567	>	// a task to run ForkJoinTask.invokeAll, but now external
2568	>	// invocation of multiple tasks is at least as efficient.
2569	>	List<ForkJoinTask<T>> fs = new ArrayList<ForkJoinTask<T>>(tasks.size());
2570	>	// Workaround needed because method wasn't declared with
2571	>	// wildcards in return type but should have been.
2572		@SuppressWarnings({"unchecked", "rawtypes"})
2573	<	List<Future<T>> futures = (List<Future<T>>) (List) forkJoinTasks;
1402	<	return futures;
1403	<	}
2573	>	List<Future<T>> futures = (List<Future<T>>) (List) fs;
2574
2575	<	static final class InvokeAll<T> extends RecursiveAction {
2576	<	final ArrayList<ForkJoinTask<T>> tasks;
2577	<	InvokeAll(ArrayList<ForkJoinTask<T>> tasks) { this.tasks = tasks; }
2578	<	public void compute() {
2579	<	try { invokeAll(tasks); }
2580	<	catch (Exception ignore) {}
2575	>	boolean done = false;
2576	>	try {
2577	>	for (Callable<T> t : tasks) {
2578	>	ForkJoinTask<T> f = new ForkJoinTask.AdaptedCallable<T>(t);
2579	>	doSubmit(f);
2580	>	fs.add(f);
2581	>	}
2582	>	for (ForkJoinTask<T> f : fs)
2583	>	f.quietlyJoin();
2584	>	done = true;
2585	>	return futures;
2586	>	} finally {
2587	>	if (!done)
2588	>	for (ForkJoinTask<T> f : fs)
2589	>	f.cancel(false);
2590		}
1412	–	private static final long serialVersionUID = -7914297376763021607L;
2591		}
2592
2593		/**
#	Line 1441 | Line 2619 | public class ForkJoinPool extends Abstra
2619		}
2620
2621		/**
2622	+	* Returns the targeted parallelism level of the common pool.
2623	+	*
2624	+	* @return the targeted parallelism level of the common pool
2625	+	*/
2626	+	public static int getCommonPoolParallelism() {
2627	+	return commonPoolParallelism;
2628	+	}
2629	+
2630	+	/**
2631		* Returns the number of worker threads that have started but not
2632	<	* yet terminated.  This result returned by this method may differ
2632	>	* yet terminated.  The result returned by this method may differ
2633		* from {@link #getParallelism} when threads are created to
2634		* maintain parallelism when others are cooperatively blocked.
2635		*
2636		* @return the number of worker threads
2637		*/
2638		public int getPoolSize() {
2639	<	return workerCounts >>> TOTAL_COUNT_SHIFT;
2639	>	return parallelism + (short)(ctl >>> TC_SHIFT);
2640		}
2641
2642		/**
#	Line 1459 | Line 2646 | public class ForkJoinPool extends Abstra
2646		* @return {@code true} if this pool uses async mode
2647		*/
2648		public boolean getAsyncMode() {
2649	<	return locallyFifo;
2649	>	return localMode != 0;
2650		}
2651
2652		/**
#	Line 1471 | Line 2658 | public class ForkJoinPool extends Abstra
2658		* @return the number of worker threads
2659		*/
2660		public int getRunningThreadCount() {
2661	<	return workerCounts & RUNNING_COUNT_MASK;
2661	>	int rc = 0;
2662	>	WorkQueue[] ws; WorkQueue w;
2663	>	if ((ws = workQueues) != null) {
2664	>	for (int i = 1; i < ws.length; i += 2) {
2665	>	if ((w = ws[i]) != null && w.isApparentlyUnblocked())
2666	>	++rc;
2667	>	}
2668	>	}
2669	>	return rc;
2670		}
2671
2672		/**
#	Line 1482 | Line 2677 | public class ForkJoinPool extends Abstra
2677		* @return the number of active threads
2678		*/
2679		public int getActiveThreadCount() {
2680	<	return runState & ACTIVE_COUNT_MASK;
2680	>	int r = parallelism + (int)(ctl >> AC_SHIFT);
2681	>	return (r <= 0) ? 0 : r; // suppress momentarily negative values
2682		}
2683
2684		/**
#	Line 1497 | Line 2693 | public class ForkJoinPool extends Abstra
2693		* @return {@code true} if all threads are currently idle
2694		*/
2695		public boolean isQuiescent() {
2696	<	return (runState & ACTIVE_COUNT_MASK) == 0;
2696	>	return (int)(ctl >> AC_SHIFT) + parallelism == 0;
2697		}
2698
2699		/**
#	Line 1512 | Line 2708 | public class ForkJoinPool extends Abstra
2708		* @return the number of steals
2709		*/
2710		public long getStealCount() {
2711	<	return stealCount;
2711	>	long count = stealCount;
2712	>	WorkQueue[] ws; WorkQueue w;
2713	>	if ((ws = workQueues) != null) {
2714	>	for (int i = 1; i < ws.length; i += 2) {
2715	>	if ((w = ws[i]) != null)
2716	>	count += w.totalSteals;
2717	>	}
2718	>	}
2719	>	return count;
2720		}
2721
2722		/**
#	Line 1527 | Line 2731 | public class ForkJoinPool extends Abstra
2731		*/
2732		public long getQueuedTaskCount() {
2733		long count = 0;
2734	<	ForkJoinWorkerThread[] ws = workers;
2735	<	int n = ws.length;
2736	<	for (int i = 0; i < n; ++i) {
2737	<	ForkJoinWorkerThread w = ws[i];
2738	<	if (w != null)
2739	<	count += w.getQueueSize();
2734	>	WorkQueue[] ws; WorkQueue w;
2735	>	if ((ws = workQueues) != null) {
2736	>	for (int i = 1; i < ws.length; i += 2) {
2737	>	if ((w = ws[i]) != null)
2738	>	count += w.queueSize();
2739	>	}
2740		}
2741		return count;
2742		}
2743
2744		/**
2745		* Returns an estimate of the number of tasks submitted to this
2746	<	* pool that have not yet begun executing.  This method takes time
2747	<	* proportional to the number of submissions.
2746	>	* pool that have not yet begun executing.  This method may take
2747	>	* time proportional to the number of submissions.
2748		*
2749		* @return the number of queued submissions
2750		*/
2751		public int getQueuedSubmissionCount() {
2752	<	return submissionQueue.size();
2752	>	int count = 0;
2753	>	WorkQueue[] ws; WorkQueue w;
2754	>	if ((ws = workQueues) != null) {
2755	>	for (int i = 0; i < ws.length; i += 2) {
2756	>	if ((w = ws[i]) != null)
2757	>	count += w.queueSize();
2758	>	}
2759	>	}
2760	>	return count;
2761		}
2762
2763		/**
#	Line 1555 | Line 2767 | public class ForkJoinPool extends Abstra
2767		* @return {@code true} if there are any queued submissions
2768		*/
2769		public boolean hasQueuedSubmissions() {
2770	<	return !submissionQueue.isEmpty();
2770	>	WorkQueue[] ws; WorkQueue w;
2771	>	if ((ws = workQueues) != null) {
2772	>	for (int i = 0; i < ws.length; i += 2) {
2773	>	if ((w = ws[i]) != null && !w.isEmpty())
2774	>	return true;
2775	>	}
2776	>	}
2777	>	return false;
2778		}
2779
2780		/**
#	Line 1566 | Line 2785 | public class ForkJoinPool extends Abstra
2785		* @return the next submission, or {@code null} if none
2786		*/
2787		protected ForkJoinTask<?> pollSubmission() {
2788	<	return submissionQueue.poll();
2788	>	WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2789	>	if ((ws = workQueues) != null) {
2790	>	for (int i = 0; i < ws.length; i += 2) {
2791	>	if ((w = ws[i]) != null && (t = w.poll()) != null)
2792	>	return t;
2793	>	}
2794	>	}
2795	>	return null;
2796		}
2797
2798		/**
#	Line 1587 | Line 2813 | public class ForkJoinPool extends Abstra
2813		* @return the number of elements transferred
2814		*/
2815		protected int drainTasksTo(Collection<? super ForkJoinTask<?>> c) {
2816	<	int count = submissionQueue.drainTo(c);
2817	<	ForkJoinWorkerThread[] ws = workers;
2818	<	int n = ws.length;
2819	<	for (int i = 0; i < n; ++i) {
2820	<	ForkJoinWorkerThread w = ws[i];
2821	<	if (w != null)
2822	<	count += w.drainTasksTo(c);
2816	>	int count = 0;
2817	>	WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2818	>	if ((ws = workQueues) != null) {
2819	>	for (int i = 0; i < ws.length; ++i) {
2820	>	if ((w = ws[i]) != null) {
2821	>	while ((t = w.poll()) != null) {
2822	>	c.add(t);
2823	>	++count;
2824	>	}
2825	>	}
2826	>	}
2827		}
2828		return count;
2829		}
#	Line 1606 | Line 2836 | public class ForkJoinPool extends Abstra
2836		* @return a string identifying this pool, as well as its state
2837		*/
2838		public String toString() {
2839	<	long st = getStealCount();
2840	<	long qt = getQueuedTaskCount();
2841	<	long qs = getQueuedSubmissionCount();
2842	<	int wc = workerCounts;
2843	<	int tc = wc >>> TOTAL_COUNT_SHIFT;
2844	<	int rc = wc & RUNNING_COUNT_MASK;
2839	>	// Use a single pass through workQueues to collect counts
2840	>	long qt = 0L, qs = 0L; int rc = 0;
2841	>	long st = stealCount;
2842	>	long c = ctl;
2843	>	WorkQueue[] ws; WorkQueue w;
2844	>	if ((ws = workQueues) != null) {
2845	>	for (int i = 0; i < ws.length; ++i) {
2846	>	if ((w = ws[i]) != null) {
2847	>	int size = w.queueSize();
2848	>	if ((i & 1) == 0)
2849	>	qs += size;
2850	>	else {
2851	>	qt += size;
2852	>	st += w.totalSteals;
2853	>	if (w.isApparentlyUnblocked())
2854	>	++rc;
2855	>	}
2856	>	}
2857	>	}
2858	>	}
2859		int pc = parallelism;
2860	<	int rs = runState;
2861	<	int ac = rs & ACTIVE_COUNT_MASK;
2860	>	int tc = pc + (short)(c >>> TC_SHIFT);
2861	>	int ac = pc + (int)(c >> AC_SHIFT);
2862	>	if (ac < 0) // ignore transient negative
2863	>	ac = 0;
2864	>	String level;
2865	>	if ((c & STOP_BIT) != 0)
2866	>	level = (tc == 0) ? "Terminated" : "Terminating";
2867	>	else
2868	>	level = runState < 0 ? "Shutting down" : "Running";
2869		return super.toString() +
2870	<	"[" + runLevelToString(rs) +
2870	>	"[" + level +
2871		", parallelism = " + pc +
2872		", size = " + tc +
2873		", active = " + ac +
#	Line 1627 | Line 2878 | public class ForkJoinPool extends Abstra
2878		"]";
2879		}
2880
1630	–	private static String runLevelToString(int s) {
1631	–	return ((s & TERMINATED) != 0 ? "Terminated" :
1632	–	((s & TERMINATING) != 0 ? "Terminating" :
1633	–	((s & SHUTDOWN) != 0 ? "Shutting down" :
1634	–	"Running")));
1635	–	}
1636	–
2881		/**
2882	<	* Initiates an orderly shutdown in which previously submitted
2883	<	* tasks are executed, but no new tasks will be accepted.
2884	<	* Invocation has no additional effect if already shut down.
2885	<	* Tasks that are in the process of being submitted concurrently
2886	<	* during the course of this method may or may not be rejected.
2882	>	* Possibly initiates an orderly shutdown in which previously
2883	>	* submitted tasks are executed, but no new tasks will be
2884	>	* accepted. Invocation has no effect on execution state if this
2885	>	* is the {@link #commonPool}, and no additional effect if
2886	>	* already shut down.  Tasks that are in the process of being
2887	>	* submitted concurrently during the course of this method may or
2888	>	* may not be rejected.
2889		*
2890		* @throws SecurityException if a security manager exists and
2891		*         the caller is not permitted to modify threads
#	Line 1648 | Line 2894 | public class ForkJoinPool extends Abstra
2894		*/
2895		public void shutdown() {
2896		checkPermission();
2897	<	advanceRunLevel(SHUTDOWN);
2898	<	tryTerminate(false);
2897	>	if (this != commonPool)
2898	>	tryTerminate(false, true);
2899		}
2900
2901		/**
2902	<	* Attempts to cancel and/or stop all tasks, and reject all
2903	<	* subsequently submitted tasks.  Tasks that are in the process of
2904	<	* being submitted or executed concurrently during the course of
2905	<	* this method may or may not be rejected. This method cancels
2906	<	* both existing and unexecuted tasks, in order to permit
2907	<	* termination in the presence of task dependencies. So the method
2908	<	* always returns an empty list (unlike the case for some other
2909	<	* Executors).
2902	>	* Possibly attempts to cancel and/or stop all tasks, and reject
2903	>	* all subsequently submitted tasks.  Invocation has no effect on
2904	>	* execution state if this is the {@link #commonPool}, and no
2905	>	* additional effect if already shut down. Otherwise, tasks that
2906	>	* are in the process of being submitted or executed concurrently
2907	>	* during the course of this method may or may not be
2908	>	* rejected. This method cancels both existing and unexecuted
2909	>	* tasks, in order to permit termination in the presence of task
2910	>	* dependencies. So the method always returns an empty list
2911	>	* (unlike the case for some other Executors).
2912		*
2913		* @return an empty list
2914		* @throws SecurityException if a security manager exists and
#	Line 1670 | Line 2918 | public class ForkJoinPool extends Abstra
2918		*/
2919		public List<Runnable> shutdownNow() {
2920		checkPermission();
2921	<	tryTerminate(true);
2921	>	if (this != commonPool)
2922	>	tryTerminate(true, true);
2923		return Collections.emptyList();
2924		}
2925
#	Line 1680 | Line 2929 | public class ForkJoinPool extends Abstra
2929		* @return {@code true} if all tasks have completed following shut down
2930		*/
2931		public boolean isTerminated() {
2932	<	return runState >= TERMINATED;
2932	>	long c = ctl;
2933	>	return ((c & STOP_BIT) != 0L &&
2934	>	(short)(c >>> TC_SHIFT) == -parallelism);
2935		}
2936
2937		/**
#	Line 1688 | Line 2939 | public class ForkJoinPool extends Abstra
2939		* commenced but not yet completed.  This method may be useful for
2940		* debugging. A return of {@code true} reported a sufficient
2941		* period after shutdown may indicate that submitted tasks have
2942	<	* ignored or suppressed interruption, causing this executor not
2943	<	* to properly terminate.
2942	>	* ignored or suppressed interruption, or are waiting for IO,
2943	>	* causing this executor not to properly terminate. (See the
2944	>	* advisory notes for class {@link ForkJoinTask} stating that
2945	>	* tasks should not normally entail blocking operations.  But if
2946	>	* they do, they must abort them on interrupt.)
2947		*
2948		* @return {@code true} if terminating but not yet terminated
2949		*/
2950		public boolean isTerminating() {
2951	<	return (runState & (TERMINATING|TERMINATED)) == TERMINATING;
2951	>	long c = ctl;
2952	>	return ((c & STOP_BIT) != 0L &&
2953	>	(short)(c >>> TC_SHIFT) != -parallelism);
2954		}
2955
2956		/**
#	Line 1703 | Line 2959 | public class ForkJoinPool extends Abstra
2959		* @return {@code true} if this pool has been shut down
2960		*/
2961		public boolean isShutdown() {
2962	<	return runState >= SHUTDOWN;
2962	>	return runState < 0;
2963		}
2964
2965		/**
#	Line 1719 | Line 2975 | public class ForkJoinPool extends Abstra
2975		*/
2976		public boolean awaitTermination(long timeout, TimeUnit unit)
2977		throws InterruptedException {
2978	<	try {
2979	<	return termination.awaitAdvanceInterruptibly(0, timeout, unit) > 0;
2980	<	} catch(TimeoutException ex) {
2981	<	return false;
2978	>	long nanos = unit.toNanos(timeout);
2979	>	if (isTerminated())
2980	>	return true;
2981	>	long startTime = System.nanoTime();
2982	>	boolean terminated = false;
2983	>	synchronized (this) {
2984	>	for (long waitTime = nanos, millis = 0L;;) {
2985	>	if (terminated = isTerminated() ||
2986	>	waitTime <= 0L ||
2987	>	(millis = unit.toMillis(waitTime)) <= 0L)
2988	>	break;
2989	>	wait(millis);
2990	>	waitTime = nanos - (System.nanoTime() - startTime);
2991	>	}
2992		}
2993	+	return terminated;
2994		}
2995
2996		/**
#	Line 1734 | Line 3001 | public class ForkJoinPool extends Abstra
3001		* {@code isReleasable} must return {@code true} if blocking is
3002		* not necessary. Method {@code block} blocks the current thread
3003		* if necessary (perhaps internally invoking {@code isReleasable}
3004	<	* before actually blocking). The unusual methods in this API
3005	<	* accommodate synchronizers that may, but don't usually, block
3006	<	* for long periods. Similarly, they allow more efficient internal
3007	<	* handling of cases in which additional workers may be, but
3008	<	* usually are not, needed to ensure sufficient parallelism.
3009	<	* Toward this end, implementations of method {@code isReleasable}
3010	<	* must be amenable to repeated invocation.
3004	>	* before actually blocking). These actions are performed by any
3005	>	* thread invoking {@link ForkJoinPool#managedBlock}.  The
3006	>	* unusual methods in this API accommodate synchronizers that may,
3007	>	* but don't usually, block for long periods. Similarly, they
3008	>	* allow more efficient internal handling of cases in which
3009	>	* additional workers may be, but usually are not, needed to
3010	>	* ensure sufficient parallelism.  Toward this end,
3011	>	* implementations of method {@code isReleasable} must be amenable
3012	>	* to repeated invocation.
3013		*
3014		* <p>For example, here is a ManagedBlocker based on a
3015		* ReentrantLock:
#	Line 1820 | Line 3089 | public class ForkJoinPool extends Abstra
3089		public static void managedBlock(ManagedBlocker blocker)
3090		throws InterruptedException {
3091		Thread t = Thread.currentThread();
3092	<	if (t instanceof ForkJoinWorkerThread) {
3093	<	ForkJoinWorkerThread w = (ForkJoinWorkerThread) t;
3094	<	w.pool.awaitBlocker(blocker);
3095	<	}
3096	<	else {
3097	<	do {} while (!blocker.isReleasable() && !blocker.block());
3092	>	ForkJoinPool p = ((t instanceof ForkJoinWorkerThread) ?
3093	>	((ForkJoinWorkerThread)t).pool : null);
3094	>	while (!blocker.isReleasable()) {
3095	>	if (p == null || p.tryCompensate(null, blocker)) {
3096	>	try {
3097	>	do {} while (!blocker.isReleasable() && !blocker.block());
3098	>	} finally {
3099	>	if (p != null)
3100	>	p.incrementActiveCount();
3101	>	}
3102	>	break;
3103	>	}
3104		}
3105		}
3106
#	Line 1834 | Line 3109 | public class ForkJoinPool extends Abstra
3109		// implement RunnableFuture.
3110
3111		protected <T> RunnableFuture<T> newTaskFor(Runnable runnable, T value) {
3112	<	return (RunnableFuture<T>) ForkJoinTask.adapt(runnable, value);
3112	>	return new ForkJoinTask.AdaptedRunnable<T>(runnable, value);
3113		}
3114
3115		protected <T> RunnableFuture<T> newTaskFor(Callable<T> callable) {
3116	<	return (RunnableFuture<T>) ForkJoinTask.adapt(callable);
3116	>	return new ForkJoinTask.AdaptedCallable<T>(callable);
3117		}
3118
3119		// Unsafe mechanics
3120	<
3121	<	private static final sun.misc.Unsafe UNSAFE = getUnsafe();
3122	<	private static final long workerCountsOffset =
3123	<	objectFieldOffset("workerCounts", ForkJoinPool.class);
3124	<	private static final long runStateOffset =
3125	<	objectFieldOffset("runState", ForkJoinPool.class);
3126	<	private static final long eventCountOffset =
3127	<	objectFieldOffset("eventCount", ForkJoinPool.class);
3128	<	private static final long eventWaitersOffset =
3129	<	objectFieldOffset("eventWaiters",ForkJoinPool.class);
3130	<	private static final long stealCountOffset =
3131	<	objectFieldOffset("stealCount",ForkJoinPool.class);
3132	<	private static final long spareWaitersOffset =
3133	<	objectFieldOffset("spareWaiters",ForkJoinPool.class);
3134	<
3135	<	private static long objectFieldOffset(String field, Class<?> klazz) {
3120	>	private static final sun.misc.Unsafe U;
3121	>	private static final long CTL;
3122	>	private static final long PARKBLOCKER;
3123	>	private static final int ABASE;
3124	>	private static final int ASHIFT;
3125	>	private static final long NEXTWORKERNUMBER;
3126	>	private static final long STEALCOUNT;
3127	>	private static final long MAINLOCK;
3128	>
3129	>	static {
3130	>	poolNumberGenerator = new AtomicInteger();
3131	>	nextSubmitterSeed = new AtomicInteger(0x55555555);
3132	>	modifyThreadPermission = new RuntimePermission("modifyThread");
3133	>	defaultForkJoinWorkerThreadFactory =
3134	>	new DefaultForkJoinWorkerThreadFactory();
3135	>	submitters = new ThreadSubmitter();
3136	>	int s;
3137		try {
3138	<	return UNSAFE.objectFieldOffset(klazz.getDeclaredField(field));
3139	<	} catch (NoSuchFieldException e) {
3140	<	// Convert Exception to corresponding Error
3141	<	NoSuchFieldError error = new NoSuchFieldError(field);
3142	<	error.initCause(e);
3143	<	throw error;
3138	>	U = getUnsafe();
3139	>	Class<?> k = ForkJoinPool.class;
3140	>	Class<?> ak = ForkJoinTask[].class;
3141	>	CTL = U.objectFieldOffset
3142	>	(k.getDeclaredField("ctl"));
3143	>	NEXTWORKERNUMBER = U.objectFieldOffset
3144	>	(k.getDeclaredField("nextWorkerNumber"));
3145	>	STEALCOUNT = U.objectFieldOffset
3146	>	(k.getDeclaredField("stealCount"));
3147	>	MAINLOCK = U.objectFieldOffset
3148	>	(k.getDeclaredField("mainLock"));
3149	>	Class<?> tk = Thread.class;
3150	>	PARKBLOCKER = U.objectFieldOffset
3151	>	(tk.getDeclaredField("parkBlocker"));
3152	>	ABASE = U.arrayBaseOffset(ak);
3153	>	s = U.arrayIndexScale(ak);
3154	>	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
3155	>	} catch (Exception e) {
3156	>	throw new Error(e);
3157	>	}
3158	>	if ((s & (s-1)) != 0)
3159	>	throw new Error("data type scale not a power of two");
3160	>	try { // Establish common pool
3161	>	String pp = System.getProperty(propPrefix + "parallelism");
3162	>	String fp = System.getProperty(propPrefix + "threadFactory");
3163	>	String up = System.getProperty(propPrefix + "exceptionHandler");
3164	>	ForkJoinWorkerThreadFactory fac = (fp == null) ?
3165	>	defaultForkJoinWorkerThreadFactory :
3166	>	((ForkJoinWorkerThreadFactory)ClassLoader.
3167	>	getSystemClassLoader().loadClass(fp).newInstance());
3168	>	Thread.UncaughtExceptionHandler ueh = (up == null) ? null :
3169	>	((Thread.UncaughtExceptionHandler)ClassLoader.
3170	>	getSystemClassLoader().loadClass(up).newInstance());
3171	>	int par;
3172	>	if ((pp == null || (par = Integer.parseInt(pp)) <= 0))
3173	>	par = Runtime.getRuntime().availableProcessors();
3174	>	if (par > MAX_CAP)
3175	>	par = MAX_CAP;
3176	>	commonPoolParallelism = par;
3177	>	int n = par - 1; // precompute submit mask
3178	>	n |= n >>> 1; n |= n >>> 2; n |= n >>> 4;
3179	>	n |= n >>> 8; n |= n >>> 16;
3180	>	int mask = ((n + 1) << 1) - 1;
3181	>	commonPool = new ForkJoinPool(par, mask, fac, ueh);
3182	>	} catch (Exception e) {
3183	>	throw new Error(e);
3184		}
3185		}
3186
#	Line 1895 | Line 3211 | public class ForkJoinPool extends Abstra
3211		}
3212		}
3213		}
3214	+
3215		}

Diff Legend

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