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 1.72 by jsr166, Tue Sep 7 06:19:05 2010 UTC vs.
Revision 1.147 by jsr166, Tue Nov 20 05:18:42 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.*;
9		import java.util.ArrayList;
10		import java.util.Arrays;
11		import java.util.Collection;
12		import java.util.Collections;
13		import java.util.List;
14	<	import java.util.concurrent.locks.LockSupport;
15	<	import java.util.concurrent.locks.ReentrantLock;
16	<	import java.util.concurrent.atomic.AtomicInteger;
17	<	import java.util.concurrent.CountDownLatch;
14	>	import java.util.concurrent.AbstractExecutorService;
15	>	import java.util.concurrent.Callable;
16	>	import java.util.concurrent.ExecutorService;
17	>	import java.util.concurrent.Future;
18	>	import java.util.concurrent.RejectedExecutionException;
19	>	import java.util.concurrent.RunnableFuture;
20	>	import java.util.concurrent.TimeUnit;
21
22		/**
23		* An {@link ExecutorService} for running {@link ForkJoinTask}s.
#	Line 26 | Line 28 | import java.util.concurrent.CountDownLat
28		* <p>A {@code ForkJoinPool} differs from other kinds of {@link
29		* ExecutorService} mainly by virtue of employing
30		* <em>work-stealing</em>: all threads in the pool attempt to find and
31	<	* execute subtasks created by other active tasks (eventually blocking
32	<	* waiting for work if none exist). This enables efficient processing
33	<	* when most tasks spawn other subtasks (as do most {@code
34	<	* ForkJoinTask}s). When setting <em>asyncMode</em> to true in
35	<	* constructors, {@code ForkJoinPool}s may also be appropriate for use
36	<	* with event-style tasks that are never joined.
31	>	* execute tasks submitted to the pool and/or created by other active
32	>	* tasks (eventually blocking waiting for work if none exist). This
33	>	* enables efficient processing when most tasks spawn other subtasks
34	>	* (as do most {@code ForkJoinTask}s), as well as when many small
35	>	* tasks are submitted to the pool from external clients.  Especially
36	>	* when setting <em>asyncMode</em> to true in constructors, {@code
37	>	* ForkJoinPool}s may also be appropriate for use with event-style
38	>	* tasks that are never joined.
39		*
40	<	* <p>A {@code ForkJoinPool} is constructed with a given target
41	<	* parallelism level; by default, equal to the number of available
42	<	* processors. The pool attempts to maintain enough active (or
43	<	* available) threads by dynamically adding, suspending, or resuming
44	<	* internal worker threads, even if some tasks are stalled waiting to
45	<	* join others. However, no such adjustments are guaranteed in the
46	<	* face of blocked IO or other unmanaged synchronization. The nested
47	<	* {@link ManagedBlocker} interface enables extension of the kinds of
40	>	* <p>A static {@link #commonPool()} is available and appropriate for
41	>	* most applications. The common pool is used by any ForkJoinTask that
42	>	* is not explicitly submitted to a specified pool. Using the common
43	>	* pool normally reduces resource usage (its threads are slowly
44	>	* reclaimed during periods of non-use, and reinstated upon subsequent
45	>	* use).
46	>	*
47	>	* <p>For applications that require separate or custom pools, a {@code
48	>	* ForkJoinPool} may be constructed with a given target parallelism
49	>	* level; by default, equal to the number of available processors. The
50	>	* pool attempts to maintain enough active (or available) threads by
51	>	* dynamically adding, suspending, or resuming internal worker
52	>	* threads, even if some tasks are stalled waiting to join
53	>	* others. However, no such adjustments are guaranteed in the face of
54	>	* blocked IO or other unmanaged synchronization. The nested {@link
55	>	* ManagedBlocker} interface enables extension of the kinds of
56		* synchronization accommodated.
57		*
58		* <p>In addition to execution and lifecycle control methods, this
#	Line 50 | Line 62 | import java.util.concurrent.CountDownLat
62		* {@link #toString} returns indications of pool state in a
63		* convenient form for informal monitoring.
64		*
65	<	* <p> As is the case with other ExecutorServices, there are three
66	<	* main task execution methods summarized in the following
67	<	* table. These are designed to be used by clients not already engaged
68	<	* in fork/join computations in the current pool.  The main forms of
69	<	* these methods accept instances of {@code ForkJoinTask}, but
70	<	* overloaded forms also allow mixed execution of plain {@code
65	>	* <p>As is the case with other ExecutorServices, there are three
66	>	* main task execution methods summarized in the following table.
67	>	* These are designed to be used primarily by clients not already
68	>	* engaged in fork/join computations in the current pool.  The main
69	>	* forms of these methods accept instances of {@code ForkJoinTask},
70	>	* but overloaded forms also allow mixed execution of plain {@code
71		* Runnable}- or {@code Callable}- based activities as well.  However,
72	<	* tasks that are already executing in a pool should normally
73	<	* <em>NOT</em> use these pool execution methods, but instead use the
74	<	* within-computation forms listed in the table.
72	>	* tasks that are already executing in a pool should normally instead
73	>	* use the within-computation forms listed in the table unless using
74	>	* async event-style tasks that are not usually joined, in which case
75	>	* there is little difference among choice of methods.
76		*
77		* <table BORDER CELLPADDING=3 CELLSPACING=1>
78		*  <tr>
#	Line 84 | Line 97 | import java.util.concurrent.CountDownLat
97		*  </tr>
98		* </table>
99		*
100	<	* <p><b>Sample Usage.</b> Normally a single {@code ForkJoinPool} is
101	<	* used for all parallel task execution in a program or subsystem.
102	<	* Otherwise, use would not usually outweigh the construction and
103	<	* bookkeeping overhead of creating a large set of threads. For
104	<	* example, a common pool could be used for the {@code SortTasks}
105	<	* illustrated in {@link RecursiveAction}. Because {@code
106	<	* ForkJoinPool} uses threads in {@linkplain java.lang.Thread#isDaemon
107	<	* daemon} mode, there is typically no need to explicitly {@link
108	<	* #shutdown} such a pool upon program exit.
109	<	*
110	<	* <pre>
111	<	* static final ForkJoinPool mainPool = new ForkJoinPool();
112	<	* ...
113	<	* public void sort(long[] array) {
101	<	*   mainPool.invoke(new SortTask(array, 0, array.length));
102	<	* }
103	<	* </pre>
100	>	* <p>The common pool is by default constructed with default
101	>	* parameters, but these may be controlled by setting three {@link
102	>	* System#getProperty properties} with prefix {@code
103	>	* java.util.concurrent.ForkJoinPool.common}: {@code parallelism} --
104	>	* an integer greater than zero, {@code threadFactory} -- the class
105	>	* name of a {@link ForkJoinWorkerThreadFactory}, and {@code
106	>	<<<<<<< ForkJoinPool.java
107	>	* exceptionHandler} -- the class name of a {@code
108	>	=======
109	>	* exceptionHandler} -- the class name of a {@link
110	>	* java.lang.Thread.UncaughtExceptionHandler
111	>	>>>>>>> 1.111
112	>	* Thread.UncaughtExceptionHandler}. Upon any error in establishing
113	>	* these settings, default parameters are used.
114		*
115		* <p><b>Implementation notes</b>: This implementation restricts the
116		* maximum number of running threads to 32767. Attempts to create
#	Line 119 | Line 129 | public class ForkJoinPool extends Abstra
129		/*
130		* Implementation Overview
131		*
132	<	* This class provides the central bookkeeping and control for a
133	<	* set of worker threads: Submissions from non-FJ threads enter
134	<	* into a submission queue. Workers take these tasks and typically
135	<	* split them into subtasks that may be stolen by other workers.
136	<	* The main work-stealing mechanics implemented in class
137	<	* ForkJoinWorkerThread give first priority to processing tasks
138	<	* from their own queues (LIFO or FIFO, depending on mode), then
139	<	* to randomized FIFO steals of tasks in other worker queues, and
140	<	* lastly to new submissions. These mechanics do not consider
141	<	* affinities, loads, cache localities, etc, so rarely provide the
142	<	* best possible performance on a given machine, but portably
143	<	* provide good throughput by averaging over these factors.
144	<	* (Further, even if we did try to use such information, we do not
145	<	* usually have a basis for exploiting it. For example, some sets
146	<	* of tasks profit from cache affinities, but others are harmed by
147	<	* cache pollution effects.)
148	<	*
149	<	* Beyond work-stealing support and essential bookkeeping, the
150	<	* main responsibility of this framework is to take actions when
151	<	* one worker is waiting to join a task stolen (or always held by)
152	<	* another.  Because we are multiplexing many tasks on to a pool
153	<	* of workers, we can't just let them block (as in Thread.join).
154	<	* We also cannot just reassign the joiner's run-time stack with
155	<	* another and replace it later, which would be a form of
156	<	* "continuation", that even if possible is not necessarily a good
157	<	* idea. Given that the creation costs of most threads on most
158	<	* systems mainly surrounds setting up runtime stacks, thread
159	<	* creation and switching is usually not much more expensive than
160	<	* stack creation and switching, and is more flexible). Instead we
161	<	* combine two tactics:
132	>	* This class and its nested classes provide the main
133	>	* functionality and control for a set of worker threads:
134	>	* Submissions from non-FJ threads enter into submission queues.
135	>	* Workers take these tasks and typically split them into subtasks
136	>	* that may be stolen by other workers.  Preference rules give
137	>	* first priority to processing tasks from their own queues (LIFO
138	>	* or FIFO, depending on mode), then to randomized FIFO steals of
139	>	* tasks in other queues.
140	>	*
141	>	* WorkQueues
142	>	* ==========
143	>	*
144	>	* Most operations occur within work-stealing queues (in nested
145	>	* class WorkQueue).  These are special forms of Deques that
146	>	* support only three of the four possible end-operations -- push,
147	>	* pop, and poll (aka steal), under the further constraints that
148	>	* push and pop are called only from the owning thread (or, as
149	>	* extended here, under a lock), while poll may be called from
150	>	* other threads.  (If you are unfamiliar with them, you probably
151	>	* want to read Herlihy and Shavit's book "The Art of
152	>	* Multiprocessor programming", chapter 16 describing these in
153	>	* more detail before proceeding.)  The main work-stealing queue
154	>	* design is roughly similar to those in the papers "Dynamic
155	>	* Circular Work-Stealing Deque" by Chase and Lev, SPAA 2005
156	>	* (http://research.sun.com/scalable/pubs/index.html) and
157	>	* "Idempotent work stealing" by Michael, Saraswat, and Vechev,
158	>	* PPoPP 2009 (http://portal.acm.org/citation.cfm?id=1504186).
159	>	* The main differences ultimately stem from GC requirements that
160	>	* we null out taken slots as soon as we can, to maintain as small
161	>	* a footprint as possible even in programs generating huge
162	>	* numbers of tasks. To accomplish this, we shift the CAS
163	>	* arbitrating pop vs poll (steal) from being on the indices
164	>	* ("base" and "top") to the slots themselves.  So, both a
165	>	* successful pop and poll mainly entail a CAS of a slot from
166	>	* non-null to null.  Because we rely on CASes of references, we
167	>	* do not need tag bits on base or top.  They are simple ints as
168	>	* used in any circular array-based queue (see for example
169	>	* ArrayDeque).  Updates to the indices must still be ordered in a
170	>	* way that guarantees that top == base means the queue is empty,
171	>	* but otherwise may err on the side of possibly making the queue
172	>	* appear nonempty when a push, pop, or poll have not fully
173	>	* committed. Note that this means that the poll operation,
174	>	* considered individually, is not wait-free. One thief cannot
175	>	* successfully continue until another in-progress one (or, if
176	>	* previously empty, a push) completes.  However, in the
177	>	* aggregate, we ensure at least probabilistic non-blockingness.
178	>	* If an attempted steal fails, a thief always chooses a different
179	>	* random victim target to try next. So, in order for one thief to
180	>	* progress, it suffices for any in-progress poll or new push on
181	>	* any empty queue to complete. (This is why we normally use
182	>	* method pollAt and its variants that try once at the apparent
183	>	* base index, else consider alternative actions, rather than
184	>	* method poll.)
185	>	*
186	>	* This approach also enables support of a user mode in which local
187	>	* task processing is in FIFO, not LIFO order, simply by using
188	>	* poll rather than pop.  This can be useful in message-passing
189	>	* frameworks in which tasks are never joined.  However neither
190	>	* mode considers affinities, loads, cache localities, etc, so
191	>	* rarely provide the best possible performance on a given
192	>	* machine, but portably provide good throughput by averaging over
193	>	* these factors.  (Further, even if we did try to use such
194	>	* information, we do not usually have a basis for exploiting it.
195	>	* For example, some sets of tasks profit from cache affinities,
196	>	* but others are harmed by cache pollution effects.)
197	>	*
198	>	* WorkQueues are also used in a similar way for tasks submitted
199	>	* to the pool. We cannot mix these tasks in the same queues used
200	>	* for work-stealing (this would contaminate lifo/fifo
201	>	* processing). Instead, we randomly associate submission queues
202	>	* with submitting threads, using a form of hashing.  The
203	>	* ThreadLocal Submitter class contains a value initially used as
204	>	* a hash code for choosing existing queues, but may be randomly
205	>	* repositioned upon contention with other submitters.  In
206	>	* essence, submitters act like workers except that they are
207	>	* restricted to executing local tasks that they submitted (or in
208	>	* the case of CountedCompleters, others with the same root task).
209	>	* However, because most shared/external queue operations are more
210	>	* expensive than internal, and because, at steady state, external
211	>	* submitters will compete for CPU with workers, ForkJoinTask.join
212	>	* and related methods disable them from repeatedly helping to
213	>	* process tasks if all workers are active.  Insertion of tasks in
214	>	* shared mode requires a lock (mainly to protect in the case of
215	>	* resizing) but we use only a simple spinlock (using bits in
216	>	* field qlock), because submitters encountering a busy queue move
217	>	* on to try or create other queues -- they block only when
218	>	* creating and registering new queues.
219	>	*
220	>	* Management
221	>	* ==========
222	>	*
223	>	* The main throughput advantages of work-stealing stem from
224	>	* decentralized control -- workers mostly take tasks from
225	>	* themselves or each other. We cannot negate this in the
226	>	* implementation of other management responsibilities. The main
227	>	* tactic for avoiding bottlenecks is packing nearly all
228	>	* essentially atomic control state into two volatile variables
229	>	* that are by far most often read (not written) as status and
230	>	* consistency checks.
231	>	*
232	>	* Field "ctl" contains 64 bits holding all the information needed
233	>	* to atomically decide to add, inactivate, enqueue (on an event
234	>	* queue), dequeue, and/or re-activate workers.  To enable this
235	>	* packing, we restrict maximum parallelism to (1<<15)-1 (which is
236	>	* far in excess of normal operating range) to allow ids, counts,
237	>	* and their negations (used for thresholding) to fit into 16bit
238	>	* fields.
239	>	*
240	>	* Field "plock" is a form of sequence lock with a saturating
241	>	* shutdown bit (similarly for per-queue "qlocks"), mainly
242	>	* protecting updates to the workQueues array, as well as to
243	>	* enable shutdown.  When used as a lock, it is normally only very
244	>	* briefly held, so is nearly always available after at most a
245	>	* brief spin, but we use a monitor-based backup strategy to
246	>	* block when needed.
247	>	*
248	>	* Recording WorkQueues.  WorkQueues are recorded in the
249	>	* "workQueues" array that is created upon first use and expanded
250	>	* if necessary.  Updates to the array while recording new workers
251	>	* and unrecording terminated ones are protected from each other
252	>	* by a lock but the array is otherwise concurrently readable, and
253	>	* accessed directly.  To simplify index-based operations, the
254	>	* array size is always a power of two, and all readers must
255	>	* tolerate null slots. Worker queues are at odd indices. Shared
256	>	* (submission) queues are at even indices, up to a maximum of 64
257	>	* slots, to limit growth even if array needs to expand to add
258	>	* more workers. Grouping them together in this way simplifies and
259	>	* speeds up task scanning.
260	>	*
261	>	* All worker thread creation is on-demand, triggered by task
262	>	* submissions, replacement of terminated workers, and/or
263	>	* compensation for blocked workers. However, all other support
264	>	* code is set up to work with other policies.  To ensure that we
265	>	* do not hold on to worker references that would prevent GC, ALL
266	>	* accesses to workQueues are via indices into the workQueues
267	>	* array (which is one source of some of the messy code
268	>	* constructions here). In essence, the workQueues array serves as
269	>	* a weak reference mechanism. Thus for example the wait queue
270	>	* field of ctl stores indices, not references.  Access to the
271	>	* workQueues in associated methods (for example signalWork) must
272	>	* both index-check and null-check the IDs. All such accesses
273	>	* ignore bad IDs by returning out early from what they are doing,
274	>	* since this can only be associated with termination, in which
275	>	* case it is OK to give up.  All uses of the workQueues array
276	>	* also check that it is non-null (even if previously
277	>	* non-null). This allows nulling during termination, which is
278	>	* currently not necessary, but remains an option for
279	>	* resource-revocation-based shutdown schemes. It also helps
280	>	* reduce JIT issuance of uncommon-trap code, which tends to
281	>	* unnecessarily complicate control flow in some methods.
282	>	*
283	>	* Event Queuing. Unlike HPC work-stealing frameworks, we cannot
284	>	* let workers spin indefinitely scanning for tasks when none can
285	>	* be found immediately, and we cannot start/resume workers unless
286	>	* there appear to be tasks available.  On the other hand, we must
287	>	* quickly prod them into action when new tasks are submitted or
288	>	* generated. In many usages, ramp-up time to activate workers is
289	>	* the main limiting factor in overall performance (this is
290	>	* compounded at program start-up by JIT compilation and
291	>	* allocation). So we try to streamline this as much as possible.
292	>	* We park/unpark workers after placing in an event wait queue
293	>	* when they cannot find work. This "queue" is actually a simple
294	>	* Treiber stack, headed by the "id" field of ctl, plus a 15bit
295	>	* counter value (that reflects the number of times a worker has
296	>	* been inactivated) to avoid ABA effects (we need only as many
297	>	* version numbers as worker threads). Successors are held in
298	>	* field WorkQueue.nextWait.  Queuing deals with several intrinsic
299	>	* races, mainly that a task-producing thread can miss seeing (and
300	>	* signalling) another thread that gave up looking for work but
301	>	* has not yet entered the wait queue. We solve this by requiring
302	>	* a full sweep of all workers (via repeated calls to method
303	>	* scan()) both before and after a newly waiting worker is added
304	>	* to the wait queue. During a rescan, the worker might release
305	>	* some other queued worker rather than itself, which has the same
306	>	* net effect. Because enqueued workers may actually be rescanning
307	>	* rather than waiting, we set and clear the "parker" field of
308	>	* WorkQueues to reduce unnecessary calls to unpark.  (This
309	>	* requires a secondary recheck to avoid missed signals.)  Note
310	>	* the unusual conventions about Thread.interrupts surrounding
311	>	* parking and other blocking: Because interrupts are used solely
312	>	* to alert threads to check termination, which is checked anyway
313	>	* upon blocking, we clear status (using Thread.interrupted)
314	>	* before any call to park, so that park does not immediately
315	>	* return due to status being set via some other unrelated call to
316	>	* interrupt in user code.
317	>	*
318	>	* Signalling.  We create or wake up workers only when there
319	>	* appears to be at least one task they might be able to find and
320	>	* execute. However, many other threads may notice the same task
321	>	* and each signal to wake up a thread that might take it. So in
322	>	* general, pools will be over-signalled.  When a submission is
323	>	* added or another worker adds a task to a queue that is
324	>	* apparently empty, they signal waiting workers (or trigger
325	>	* creation of new ones if fewer than the given parallelism
326	>	* level).  These primary signals are buttressed by signals
327	>	* whenever other threads scan for work or do not have a task to
328	>	* process (including the case of leaving a hint to unparked
329	>	* threads to help signal others upon wakeup).  On most platforms,
330	>	* signalling (unpark) overhead time is noticeably long, and the
331	>	* time between signalling a thread and it actually making
332	>	* progress can be very noticeably long, so it is worth offloading
333	>	* these delays from critical paths as much as possible.
334	>	*
335	>	* Trimming workers. To release resources after periods of lack of
336	>	* use, a worker starting to wait when the pool is quiescent will
337	>	* time out and terminate if the pool has remained quiescent for a
338	>	* given period -- a short period if there are more threads than
339	>	* parallelism, longer as the number of threads decreases. This
340	>	* will slowly propagate, eventually terminating all workers after
341	>	* periods of non-use.
342	>	*
343	>	* Shutdown and Termination. A call to shutdownNow atomically sets
344	>	* a plock bit and then (non-atomically) sets each worker's
345	>	* qlock status, cancels all unprocessed tasks, and wakes up
346	>	* all waiting workers.  Detecting whether termination should
347	>	* commence after a non-abrupt shutdown() call requires more work
348	>	* and bookkeeping. We need consensus about quiescence (i.e., that
349	>	* there is no more work). The active count provides a primary
350	>	* indication but non-abrupt shutdown still requires a rechecking
351	>	* scan for any workers that are inactive but not queued.
352	>	*
353	>	* Joining Tasks
354	>	* =============
355	>	*
356	>	* Any of several actions may be taken when one worker is waiting
357	>	* to join a task stolen (or always held) by another.  Because we
358	>	* are multiplexing many tasks on to a pool of workers, we can't
359	>	* just let them block (as in Thread.join).  We also cannot just
360	>	* reassign the joiner's run-time stack with another and replace
361	>	* it later, which would be a form of "continuation", that even if
362	>	* possible is not necessarily a good idea since we sometimes need
363	>	* both an unblocked task and its continuation to progress.
364	>	* Instead we combine two tactics:
365		*
366		*   Helping: Arranging for the joiner to execute some task that it
367	<	*      would be running if the steal had not occurred.  Method
155	<	*      ForkJoinWorkerThread.helpJoinTask tracks joining->stealing
156	<	*      links to try to find such a task.
367	>	*      would be running if the steal had not occurred.
368		*
369		*   Compensating: Unless there are already enough live threads,
370	<	*      method helpMaintainParallelism() may create or
371	<	*      re-activate a spare thread to compensate for blocked
161	<	*      joiners until they unblock.
370	>	*      method tryCompensate() may create or re-activate a spare
371	>	*      thread to compensate for blocked joiners until they unblock.
372		*
373	<	* It is impossible to keep exactly the target (parallelism)
374	<	* number of threads running at any given time.  Determining
375	<	* existence of conservatively safe helping targets, the
376	<	* availability of already-created spares, and the apparent need
377	<	* to create new spares are all racy and require heuristic
378	<	* guidance, so we rely on multiple retries of each.  Compensation
379	<	* occurs in slow-motion. It is triggered only upon timeouts of
170	<	* Object.wait used for joins. This reduces poor decisions that
171	<	* would otherwise be made when threads are waiting for others
172	<	* that are stalled because of unrelated activities such as
173	<	* garbage collection.
373	>	* A third form (implemented in tryRemoveAndExec) amounts to
374	>	* helping a hypothetical compensator: If we can readily tell that
375	>	* a possible action of a compensator is to steal and execute the
376	>	* task being joined, the joining thread can do so directly,
377	>	* without the need for a compensation thread (although at the
378	>	* expense of larger run-time stacks, but the tradeoff is
379	>	* typically worthwhile).
380		*
381		* The ManagedBlocker extension API can't use helping so relies
382		* only on compensation in method awaitBlocker.
383		*
384	<	* The main throughput advantages of work-stealing stem from
385	<	* decentralized control -- workers mostly steal tasks from each
386	<	* other. We do not want to negate this by creating bottlenecks
387	<	* implementing other management responsibilities. So we use a
388	<	* collection of techniques that avoid, reduce, or cope well with
389	<	* contention. These entail several instances of bit-packing into
390	<	* CASable fields to maintain only the minimally required
391	<	* atomicity. To enable such packing, we restrict maximum
392	<	* parallelism to (1<<15)-1 (enabling twice this (to accommodate
393	<	* unbalanced increments and decrements) to fit into a 16 bit
394	<	* field, which is far in excess of normal operating range.  Even
395	<	* though updates to some of these bookkeeping fields do sometimes
396	<	* contend with each other, they don't normally cache-contend with
397	<	* updates to others enough to warrant memory padding or
398	<	* isolation. So they are all held as fields of ForkJoinPool
399	<	* objects.  The main capabilities are as follows:
400	<	*
401	<	* 1. Creating and removing workers. Workers are recorded in the
402	<	* "workers" array. This is an array as opposed to some other data
403	<	* structure to support index-based random steals by workers.
404	<	* Updates to the array recording new workers and unrecording
405	<	* terminated ones are protected from each other by a lock
406	<	* (workerLock) but the array is otherwise concurrently readable,
407	<	* and accessed directly by workers. To simplify index-based
408	<	* operations, the array size is always a power of two, and all
409	<	* readers must tolerate null slots. Currently, all worker thread
410	<	* creation is on-demand, triggered by task submissions,
411	<	* replacement of terminated workers, and/or compensation for
412	<	* blocked workers. However, all other support code is set up to
413	<	* work with other policies.
414	<	*
415	<	* To ensure that we do not hold on to worker references that
416	<	* would prevent GC, ALL accesses to workers are via indices into
417	<	* the workers array (which is one source of some of the unusual
418	<	* code constructions here). In essence, the workers array serves
419	<	* as a WeakReference mechanism. Thus for example the event queue
420	<	* stores worker indices, not worker references. Access to the
421	<	* workers in associated methods (for example releaseEventWaiters)
422	<	* must both index-check and null-check the IDs. All such accesses
423	<	* ignore bad IDs by returning out early from what they are doing,
424	<	* since this can only be associated with shutdown, in which case
425	<	* it is OK to give up. On termination, we just clobber these
426	<	* data structures without trying to use them.
427	<	*
428	<	* 2. Bookkeeping for dynamically adding and removing workers. We
429	<	* aim to approximately maintain the given level of parallelism.
430	<	* When some workers are known to be blocked (on joins or via
431	<	* ManagedBlocker), we may create or resume others to take their
432	<	* place until they unblock (see below). Implementing this
433	<	* requires counts of the number of "running" threads (i.e., those
434	<	* that are neither blocked nor artificially suspended) as well as
435	<	* the total number.  These two values are packed into one field,
436	<	* "workerCounts" because we need accurate snapshots when deciding
437	<	* to create, resume or suspend.  Note however that the
438	<	* correspondence of these counts to reality is not guaranteed. In
439	<	* particular updates for unblocked threads may lag until they
440	<	* actually wake up.
441	<	*
442	<	* 3. Maintaining global run state. The run state of the pool
443	<	* consists of a runLevel (SHUTDOWN, TERMINATING, etc) similar to
444	<	* those in other Executor implementations, as well as a count of
445	<	* "active" workers -- those that are, or soon will be, or
446	<	* recently were executing tasks. The runLevel and active count
447	<	* are packed together in order to correctly trigger shutdown and
448	<	* termination. Without care, active counts can be subject to very
449	<	* high contention.  We substantially reduce this contention by
450	<	* relaxing update rules.  A worker must claim active status
451	<	* prospectively, by activating if it sees that a submitted or
452	<	* stealable task exists (it may find after activating that the
453	<	* task no longer exists). It stays active while processing this
454	<	* task (if it exists) and any other local subtasks it produces,
455	<	* until it cannot find any other tasks. It then tries
456	<	* inactivating (see method preStep), but upon update contention
457	<	* instead scans for more tasks, later retrying inactivation if it
458	<	* doesn't find any.
459	<	*
460	<	* 4. Managing idle workers waiting for tasks. We cannot let
461	<	* workers spin indefinitely scanning for tasks when none are
462	<	* available. On the other hand, we must quickly prod them into
463	<	* action when new tasks are submitted or generated.  We
464	<	* park/unpark these idle workers using an event-count scheme.
465	<	* Field eventCount is incremented upon events that may enable
466	<	* workers that previously could not find a task to now find one:
467	<	* Submission of a new task to the pool, or another worker pushing
468	<	* a task onto a previously empty queue.  (We also use this
469	<	* mechanism for configuration and termination actions that
264	<	* require wakeups of idle workers).  Each worker maintains its
265	<	* last known event count, and blocks when a scan for work did not
266	<	* find a task AND its lastEventCount matches the current
267	<	* eventCount. Waiting idle workers are recorded in a variant of
268	<	* Treiber stack headed by field eventWaiters which, when nonzero,
269	<	* encodes the thread index and count awaited for by the worker
270	<	* thread most recently calling eventSync. This thread in turn has
271	<	* a record (field nextEventWaiter) for the next waiting worker.
272	<	* In addition to allowing simpler decisions about need for
273	<	* wakeup, the event count bits in eventWaiters serve the role of
274	<	* tags to avoid ABA errors in Treiber stacks. Upon any wakeup,
275	<	* released threads also try to release at most two others.  The
276	<	* net effect is a tree-like diffusion of signals, where released
277	<	* threads (and possibly others) help with unparks.  To further
278	<	* reduce contention effects a bit, failed CASes to increment
279	<	* field eventCount are tolerated without retries in signalWork.
280	<	* Conceptually they are merged into the same event, which is OK
281	<	* when their only purpose is to enable workers to scan for work.
282	<	*
283	<	* 5. Managing suspension of extra workers. When a worker notices
284	<	* (usually upon timeout of a wait()) that there are too few
285	<	* running threads, we may create a new thread to maintain
286	<	* parallelism level, or at least avoid starvation. Usually, extra
287	<	* threads are needed for only very short periods, yet join
288	<	* dependencies are such that we sometimes need them in
289	<	* bursts. Rather than create new threads each time this happens,
290	<	* we suspend no-longer-needed extra ones as "spares". For most
291	<	* purposes, we don't distinguish "extra" spare threads from
292	<	* normal "core" threads: On each call to preStep (the only point
293	<	* at which we can do this) a worker checks to see if there are
294	<	* now too many running workers, and if so, suspends itself.
295	<	* Method helpMaintainParallelism looks for suspended threads to
296	<	* resume before considering creating a new replacement. The
297	<	* spares themselves are encoded on another variant of a Treiber
298	<	* Stack, headed at field "spareWaiters".  Note that the use of
299	<	* spares is intrinsically racy.  One thread may become a spare at
300	<	* about the same time as another is needlessly being created. We
301	<	* counteract this and related slop in part by requiring resumed
302	<	* spares to immediately recheck (in preStep) to see whether they
303	<	* should re-suspend.
304	<	*
305	<	* 6. Killing off unneeded workers. A timeout mechanism is used to
306	<	* shed unused workers: The oldest (first) event queue waiter uses
307	<	* a timed rather than hard wait. When this wait times out without
308	<	* a normal wakeup, it tries to shutdown any one (for convenience
309	<	* the newest) other spare or event waiter via
310	<	* tryShutdownUnusedWorker. This eventually reduces the number of
311	<	* worker threads to a minimum of one after a long enough period
312	<	* without use.
313	<	*
314	<	* 7. Deciding when to create new workers. The main dynamic
315	<	* control in this class is deciding when to create extra threads
316	<	* in method helpMaintainParallelism. We would like to keep
317	<	* exactly #parallelism threads running, which is an impossible
318	<	* task. We always need to create one when the number of running
319	<	* threads would become zero and all workers are busy. Beyond
320	<	* this, we must rely on heuristics that work well in the
321	<	* presence of transient phenomena such as GC stalls, dynamic
322	<	* compilation, and wake-up lags. These transients are extremely
323	<	* common -- we are normally trying to fully saturate the CPUs on
324	<	* a machine, so almost any activity other than running tasks
325	<	* impedes accuracy. Our main defense is to allow parallelism to
326	<	* lapse for a while during joins, and use a timeout to see if,
327	<	* after the resulting settling, there is still a need for
328	<	* additional workers.  This also better copes with the fact that
329	<	* some of the methods in this class tend to never become compiled
330	<	* (but are interpreted), so some components of the entire set of
331	<	* controls might execute 100 times faster than others. And
332	<	* similarly for cases where the apparent lack of work is just due
333	<	* to GC stalls and other transient system activity.
334	<	*
335	<	* Beware that there is a lot of representation-level coupling
336	<	* among classes ForkJoinPool, ForkJoinWorkerThread, and
337	<	* ForkJoinTask.  For example, direct access to "workers" array by
338	<	* workers, and direct access to ForkJoinTask.status by both
339	<	* ForkJoinPool and ForkJoinWorkerThread.  There is little point
384	>	* The algorithm in tryHelpStealer entails a form of "linear"
385	>	* helping: Each worker records (in field currentSteal) the most
386	>	* recent task it stole from some other worker. Plus, it records
387	>	* (in field currentJoin) the task it is currently actively
388	>	* joining. Method tryHelpStealer uses these markers to try to
389	>	* find a worker to help (i.e., steal back a task from and execute
390	>	* it) that could hasten completion of the actively joined task.
391	>	* In essence, the joiner executes a task that would be on its own
392	>	* local deque had the to-be-joined task not been stolen. This may
393	>	* be seen as a conservative variant of the approach in Wagner &
394	>	* Calder "Leapfrogging: a portable technique for implementing
395	>	* efficient futures" SIGPLAN Notices, 1993
396	>	* (http://portal.acm.org/citation.cfm?id=155354). It differs in
397	>	* that: (1) We only maintain dependency links across workers upon
398	>	* steals, rather than use per-task bookkeeping.  This sometimes
399	>	* requires a linear scan of workQueues array to locate stealers,
400	>	* but often doesn't because stealers leave hints (that may become
401	>	* stale/wrong) of where to locate them.  It is only a hint
402	>	* because a worker might have had multiple steals and the hint
403	>	* records only one of them (usually the most current).  Hinting
404	>	* isolates cost to when it is needed, rather than adding to
405	>	* per-task overhead.  (2) It is "shallow", ignoring nesting and
406	>	* potentially cyclic mutual steals.  (3) It is intentionally
407	>	* racy: field currentJoin is updated only while actively joining,
408	>	* which means that we miss links in the chain during long-lived
409	>	* tasks, GC stalls etc (which is OK since blocking in such cases
410	>	* is usually a good idea).  (4) We bound the number of attempts
411	>	* to find work (see MAX_HELP) and fall back to suspending the
412	>	* worker and if necessary replacing it with another.
413	>	*
414	>	* Helping actions for CountedCompleters are much simpler: Method
415	>	* helpComplete can take and execute any task with the same root
416	>	* as the task being waited on. However, this still entails some
417	>	* traversal of completer chains, so is less efficient than using
418	>	* CountedCompleters without explicit joins.
419	>	*
420	>	* It is impossible to keep exactly the target parallelism number
421	>	* of threads running at any given time.  Determining the
422	>	* existence of conservatively safe helping targets, the
423	>	* availability of already-created spares, and the apparent need
424	>	* to create new spares are all racy, so we rely on multiple
425	>	* retries of each.  Compensation in the apparent absence of
426	>	* helping opportunities is challenging to control on JVMs, where
427	>	* GC and other activities can stall progress of tasks that in
428	>	* turn stall out many other dependent tasks, without us being
429	>	* able to determine whether they will ever require compensation.
430	>	* Even though work-stealing otherwise encounters little
431	>	* degradation in the presence of more threads than cores,
432	>	* aggressively adding new threads in such cases entails risk of
433	>	* unwanted positive feedback control loops in which more threads
434	>	* cause more dependent stalls (as well as delayed progress of
435	>	* unblocked threads to the point that we know they are available)
436	>	* leading to more situations requiring more threads, and so
437	>	* on. This aspect of control can be seen as an (analytically
438	>	* intractable) game with an opponent that may choose the worst
439	>	* (for us) active thread to stall at any time.  We take several
440	>	* precautions to bound losses (and thus bound gains), mainly in
441	>	* methods tryCompensate and awaitJoin.
442	>	*
443	>	* Common Pool
444	>	* ===========
445	>	*
446	>	* The static commonPool always exists after static
447	>	* initialization.  Since it (or any other created pool) need
448	>	* never be used, we minimize initial construction overhead and
449	>	* footprint to the setup of about a dozen fields, with no nested
450	>	* allocation. Most bootstrapping occurs within method
451	>	* fullExternalPush during the first submission to the pool.
452	>	*
453	>	* When external threads submit to the common pool, they can
454	>	* perform some subtask processing (see externalHelpJoin and
455	>	* related methods).  We do not need to record whether these
456	>	* submissions are to the common pool -- if not, externalHelpJoin
457	>	* returns quickly (at the most helping to signal some common pool
458	>	* workers). These submitters would otherwise be blocked waiting
459	>	* for completion, so the extra effort (with liberally sprinkled
460	>	* task status checks) in inapplicable cases amounts to an odd
461	>	* form of limited spin-wait before blocking in ForkJoinTask.join.
462	>	*
463	>	* Style notes
464	>	* ===========
465	>	*
466	>	* There is a lot of representation-level coupling among classes
467	>	* ForkJoinPool, ForkJoinWorkerThread, and ForkJoinTask.  The
468	>	* fields of WorkQueue maintain data structures managed by
469	>	* ForkJoinPool, so are directly accessed.  There is little point
470		* trying to reduce this, since any associated future changes in
471		* representations will need to be accompanied by algorithmic
472	<	* changes anyway.
473	<	*
474	<	* Style notes: There are lots of inline assignments (of form
475	<	* "while ((local = field) != 0)") which are usually the simplest
476	<	* way to ensure the required read orderings (which are sometimes
477	<	* critical). Also several occurrences of the unusual "do {}
478	<	* while (!cas...)" which is the simplest way to force an update of
479	<	* a CAS'ed variable. There are also other coding oddities that
480	<	* help some methods perform reasonably even when interpreted (not
481	<	* compiled), at the expense of some messy constructions that
482	<	* reduce byte code counts.
483	<	*
484	<	* The order of declarations in this file is: (1) statics (2)
485	<	* fields (along with constants used when unpacking some of them)
486	<	* (3) internal control methods (4) callbacks and other support
487	<	* for ForkJoinTask and ForkJoinWorkerThread classes, (5) exported
488	<	* methods (plus a few little helpers).
472	>	* changes anyway. Several methods intrinsically sprawl because
473	>	* they must accumulate sets of consistent reads of volatiles held
474	>	* in local variables.  Methods signalWork() and scan() are the
475	>	* main bottlenecks, so are especially heavily
476	>	* micro-optimized/mangled.  There are lots of inline assignments
477	>	* (of form "while ((local = field) != 0)") which are usually the
478	>	* simplest way to ensure the required read orderings (which are
479	>	* sometimes critical). This leads to a "C"-like style of listing
480	>	* declarations of these locals at the heads of methods or blocks.
481	>	* There are several occurrences of the unusual "do {} while
482	>	* (!cas...)"  which is the simplest way to force an update of a
483	>	* CAS'ed variable. There are also other coding oddities (including
484	>	* several unnecessary-looking hoisted null checks) that help
485	>	* some methods perform reasonably even when interpreted (not
486	>	* compiled).
487	>	*
488	>	* The order of declarations in this file is:
489	>	* (1) Static utility functions
490	>	* (2) Nested (static) classes
491	>	* (3) Static fields
492	>	* (4) Fields, along with constants used when unpacking some of them
493	>	* (5) Internal control methods
494	>	* (6) Callbacks and other support for ForkJoinTask methods
495	>	* (7) Exported methods
496	>	* (8) Static block initializing statics in minimally dependent order
497		*/
498
499	+	// Static utilities
500	+
501	+	/**
502	+	* If there is a security manager, makes sure caller has
503	+	* permission to modify threads.
504	+	*/
505	+	private static void checkPermission() {
506	+	SecurityManager security = System.getSecurityManager();
507	+	if (security != null)
508	+	security.checkPermission(modifyThreadPermission);
509	+	}
510	+
511	+	// Nested classes
512	+
513		/**
514		* Factory for creating new {@link ForkJoinWorkerThread}s.
515		* A {@code ForkJoinWorkerThreadFactory} must be defined and used
#	Line 378 | Line 530 | public class ForkJoinPool extends Abstra
530		* Default ForkJoinWorkerThreadFactory implementation; creates a
531		* new ForkJoinWorkerThread.
532		*/
533	<	static class DefaultForkJoinWorkerThreadFactory
533	>	static final class DefaultForkJoinWorkerThreadFactory
534		implements ForkJoinWorkerThreadFactory {
535	<	public ForkJoinWorkerThread newThread(ForkJoinPool pool) {
535	>	public final ForkJoinWorkerThread newThread(ForkJoinPool pool) {
536		return new ForkJoinWorkerThread(pool);
537		}
538		}
539
540		/**
541	<	* Creates a new ForkJoinWorkerThread. This factory is used unless
542	<	* overridden in ForkJoinPool constructors.
541	>	* Class for artificial tasks that are used to replace the target
542	>	* of local joins if they are removed from an interior queue slot
543	>	* in WorkQueue.tryRemoveAndExec. We don't need the proxy to
544	>	* actually do anything beyond having a unique identity.
545	>	*/
546	>	static final class EmptyTask extends ForkJoinTask<Void> {
547	>	private static final long serialVersionUID = -7721805057305804111L;
548	>	EmptyTask() { status = ForkJoinTask.NORMAL; } // force done
549	>	public final Void getRawResult() { return null; }
550	>	public final void setRawResult(Void x) {}
551	>	public final boolean exec() { return true; }
552	>	}
553	>
554	>	/**
555	>	* Queues supporting work-stealing as well as external task
556	>	* submission. See above for main rationale and algorithms.
557	>	* Implementation relies heavily on "Unsafe" intrinsics
558	>	* and selective use of "volatile":
559	>	*
560	>	* Field "base" is the index (mod array.length) of the least valid
561	>	* queue slot, which is always the next position to steal (poll)
562	>	* from if nonempty. Reads and writes require volatile orderings
563	>	* but not CAS, because updates are only performed after slot
564	>	* CASes.
565	>	*
566	>	* Field "top" is the index (mod array.length) of the next queue
567	>	* slot to push to or pop from. It is written only by owner thread
568	>	* for push, or under lock for external/shared push, and accessed
569	>	* by other threads only after reading (volatile) base.  Both top
570	>	* and base are allowed to wrap around on overflow, but (top -
571	>	* base) (or more commonly -(base - top) to force volatile read of
572	>	* base before top) still estimates size. The lock ("qlock") is
573	>	* forced to -1 on termination, causing all further lock attempts
574	>	* to fail. (Note: we don't need CAS for termination state because
575	>	* upon pool shutdown, all shared-queues will stop being used
576	>	* anyway.)  Nearly all lock bodies are set up so that exceptions
577	>	* within lock bodies are "impossible" (modulo JVM errors that
578	>	* would cause failure anyway.)
579	>	*
580	>	* The array slots are read and written using the emulation of
581	>	* volatiles/atomics provided by Unsafe. Insertions must in
582	>	* general use putOrderedObject as a form of releasing store to
583	>	* ensure that all writes to the task object are ordered before
584	>	* its publication in the queue.  All removals entail a CAS to
585	>	* null.  The array is always a power of two. To ensure safety of
586	>	* Unsafe array operations, all accesses perform explicit null
587	>	* checks and implicit bounds checks via power-of-two masking.
588	>	*
589	>	* In addition to basic queuing support, this class contains
590	>	* fields described elsewhere to control execution. It turns out
591	>	* to work better memory-layout-wise to include them in this class
592	>	* rather than a separate class.
593	>	*
594	>	* Performance on most platforms is very sensitive to placement of
595	>	* instances of both WorkQueues and their arrays -- we absolutely
596	>	* do not want multiple WorkQueue instances or multiple queue
597	>	* arrays sharing cache lines. (It would be best for queue objects
598	>	* and their arrays to share, but there is nothing available to
599	>	* help arrange that).  Unfortunately, because they are recorded
600	>	* in a common array, WorkQueue instances are often moved to be
601	>	* adjacent by garbage collectors. To reduce impact, we use field
602	>	* padding that works OK on common platforms; this effectively
603	>	* trades off slightly slower average field access for the sake of
604	>	* avoiding really bad worst-case access. (Until better JVM
605	>	* support is in place, this padding is dependent on transient
606	>	* properties of JVM field layout rules.)
607		*/
608	<	public static final ForkJoinWorkerThreadFactory
609	<	defaultForkJoinWorkerThreadFactory =
610	<	new DefaultForkJoinWorkerThreadFactory();
608	>	static final class WorkQueue {
609	>	/**
610	>	* Capacity of work-stealing queue array upon initialization.
611	>	* Must be a power of two; at least 4, but should be larger to
612	>	* reduce or eliminate cacheline sharing among queues.
613	>	* Currently, it is much larger, as a partial workaround for
614	>	* the fact that JVMs often place arrays in locations that
615	>	* share GC bookkeeping (especially cardmarks) such that
616	>	* per-write accesses encounter serious memory contention.
617	>	*/
618	>	static final int INITIAL_QUEUE_CAPACITY = 1 << 13;
619
620	<	/**
621	<	* Permission required for callers of methods that may start or
622	<	* kill threads.
623	<	*/
624	<	private static final RuntimePermission modifyThreadPermission =
625	<	new RuntimePermission("modifyThread");
620	>	/**
621	>	* Maximum size for queue arrays. Must be a power of two less
622	>	* than or equal to 1 << (31 - width of array entry) to ensure
623	>	* lack of wraparound of index calculations, but defined to a
624	>	* value a bit less than this to help users trap runaway
625	>	* programs before saturating systems.
626	>	*/
627	>	static final int MAXIMUM_QUEUE_CAPACITY = 1 << 26; // 64M
628
629	<	/**
630	<	* If there is a security manager, makes sure caller has
631	<	* permission to modify threads.
632	<	*/
633	<	private static void checkPermission() {
634	<	SecurityManager security = System.getSecurityManager();
635	<	if (security != null)
636	<	security.checkPermission(modifyThreadPermission);
629	>	int seed;                  // for random scanning; initialize nonzero
630	>	volatile int eventCount;   // encoded inactivation count; < 0 if inactive
631	>	int nextWait;              // encoded record of next event waiter
632	>	int hint;                  // steal or signal hint (index)
633	>	int poolIndex;             // index of this queue in pool (or 0)
634	>	final int mode;            // 0: lifo, > 0: fifo, < 0: shared
635	>	int nsteals;               // number of steals
636	>	volatile int qlock;        // 1: locked, -1: terminate; else 0
637	>	volatile int base;         // index of next slot for poll
638	>	int top;                   // index of next slot for push
639	>	ForkJoinTask<?>[] array;   // the elements (initially unallocated)
640	>	final ForkJoinPool pool;   // the containing pool (may be null)
641	>	final ForkJoinWorkerThread owner; // owning thread or null if shared
642	>	volatile Thread parker;    // == owner during call to park; else null
643	>	volatile ForkJoinTask<?> currentJoin;  // task being joined in awaitJoin
644	>	ForkJoinTask<?> currentSteal; // current non-local task being executed
645	>
646	>	// Heuristic padding to ameliorate unfortunate memory placements
647	>	Object p00, p01, p02, p03, p04, p05, p06, p07;
648	>	Object p08, p09, p0a, p0b, p0c;
649	>
650	>	WorkQueue(ForkJoinPool pool, ForkJoinWorkerThread owner, int mode,
651	>	int seed) {
652	>	this.array = new ForkJoinTask<?>[WorkQueue.INITIAL_QUEUE_CAPACITY];
653	>	this.pool = pool;
654	>	this.owner = owner;
655	>	this.mode = mode;
656	>	this.seed = seed;
657	>	// Place indices in the center of array
658	>	base = top = INITIAL_QUEUE_CAPACITY >>> 1;
659	>	}
660	>
661	>	/**
662	>	* Pushes a task. Call only by owner in unshared queues.
663	>	* Cases needing resizing or rejection are relayed to fullPush
664	>	* (that also handles shared queues).
665	>	*
666	>	* @param task the task. Caller must ensure non-null.
667	>	* @throw RejectedExecutionException if array cannot be resized
668	>	*/
669	>	final void push(ForkJoinTask<?> task) {
670	>	ForkJoinTask<?>[] a; ForkJoinPool p;
671	>	int s = top, m, n;
672	>	if ((a = array) != null) {    // ignore if queue removed
673	>	U.putOrderedObject
674	>	(a, (((m = a.length - 1) & s) << ASHIFT) + ABASE, task);
675	>	if ((n = (top = s + 1) - base) <= 1) {
676	>	if ((p = pool) != null)
677	>	p.signalWork(this, 0);
678	>	}
679	>	else if (n >= m)
680	>	growArray();
681	>	}
682	>	}
683	>
684	>	/**
685	>	* Pushes a task if lock is free and array is either big
686	>	* enough or can be resized to be big enough.
687	>	*
688	>	* @param task the task. Caller must ensure non-null.
689	>	* @return true if submitted
690	>	*/
691	>	final boolean trySharedPush(ForkJoinTask<?> task) {
692	>	boolean submitted = false;
693	>	if (qlock == 0 && U.compareAndSwapInt(this, QLOCK, 0, 1)) {
694	>	ForkJoinTask<?>[] a = array;  ForkJoinPool p;
695	>	int s = top;
696	>	try {
697	>	if ((a != null && a.length > s + 1 - base) ||
698	>	(a = growArray()) != null) {   // must presize
699	>	int j = (((a.length - 1) & s) << ASHIFT) + ABASE;
700	>	U.putOrderedObject(a, j, task);
701	>	top = s + 1;
702	>	submitted = true;
703	>	}
704	>	} finally {
705	>	qlock = 0;                         // unlock
706	>	}
707	>	if (submitted && (p = pool) != null)
708	>	p.signalWork(this, 0);
709	>	}
710	>	return submitted;
711	>	}
712	>
713	>	/**
714	>	* Initializes or doubles the capacity of array. Call either
715	>	* by owner or with lock held -- it is OK for base, but not
716	>	* top, to move while resizings are in progress.
717	>	*/
718	>	final ForkJoinTask<?>[] growArray() {
719	>	ForkJoinTask<?>[] oldA = array;
720	>	int size = oldA != null ? oldA.length << 1 : INITIAL_QUEUE_CAPACITY;
721	>	if (size > MAXIMUM_QUEUE_CAPACITY)
722	>	throw new RejectedExecutionException("Queue capacity exceeded");
723	>	int oldMask, t, b;
724	>	ForkJoinTask<?>[] a = array = new ForkJoinTask<?>[size];
725	>	if (oldA != null && (oldMask = oldA.length - 1) >= 0 &&
726	>	(t = top) - (b = base) > 0) {
727	>	int mask = size - 1;
728	>	do {
729	>	ForkJoinTask<?> x;
730	>	int oldj = ((b & oldMask) << ASHIFT) + ABASE;
731	>	int j    = ((b &    mask) << ASHIFT) + ABASE;
732	>	x = (ForkJoinTask<?>)U.getObjectVolatile(oldA, oldj);
733	>	if (x != null &&
734	>	U.compareAndSwapObject(oldA, oldj, x, null))
735	>	U.putObjectVolatile(a, j, x);
736	>	} while (++b != t);
737	>	}
738	>	return a;
739	>	}
740	>
741	>	/**
742	>	* Takes next task, if one exists, in LIFO order.  Call only
743	>	* by owner in unshared queues.
744	>	*/
745	>	final ForkJoinTask<?> pop() {
746	>	ForkJoinTask<?>[] a; ForkJoinTask<?> t; int m;
747	>	if ((a = array) != null && (m = a.length - 1) >= 0) {
748	>	for (int s; (s = top - 1) - base >= 0;) {
749	>	long j = ((m & s) << ASHIFT) + ABASE;
750	>	if ((t = (ForkJoinTask<?>)U.getObject(a, j)) == null)
751	>	break;
752	>	if (U.compareAndSwapObject(a, j, t, null)) {
753	>	top = s;
754	>	return t;
755	>	}
756	>	}
757	>	}
758	>	return null;
759	>	}
760	>
761	>	/**
762	>	* Takes a task in FIFO order if b is base of queue and a task
763	>	* can be claimed without contention. Specialized versions
764	>	* appear in ForkJoinPool methods scan and tryHelpStealer.
765	>	*/
766	>	final ForkJoinTask<?> pollAt(int b) {
767	>	ForkJoinTask<?> t; ForkJoinTask<?>[] a;
768	>	if ((a = array) != null) {
769	>	int j = (((a.length - 1) & b) << ASHIFT) + ABASE;
770	>	if ((t = (ForkJoinTask<?>)U.getObjectVolatile(a, j)) != null &&
771	>	base == b &&
772	>	U.compareAndSwapObject(a, j, t, null)) {
773	>	base = b + 1;
774	>	return t;
775	>	}
776	>	}
777	>	return null;
778	>	}
779	>
780	>	/**
781	>	* Takes next task, if one exists, in FIFO order.
782	>	*/
783	>	final ForkJoinTask<?> poll() {
784	>	ForkJoinTask<?>[] a; int b; ForkJoinTask<?> t;
785	>	while ((b = base) - top < 0 && (a = array) != null) {
786	>	int j = (((a.length - 1) & b) << ASHIFT) + ABASE;
787	>	t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
788	>	if (t != null) {
789	>	if (base == b &&
790	>	U.compareAndSwapObject(a, j, t, null)) {
791	>	base = b + 1;
792	>	return t;
793	>	}
794	>	}
795	>	else if (base == b) {
796	>	if (b + 1 == top)
797	>	break;
798	>	Thread.yield(); // wait for lagging update (very rare)
799	>	}
800	>	}
801	>	return null;
802	>	}
803	>
804	>	/**
805	>	* Takes next task, if one exists, in order specified by mode.
806	>	*/
807	>	final ForkJoinTask<?> nextLocalTask() {
808	>	return mode == 0 ? pop() : poll();
809	>	}
810	>
811	>	/**
812	>	* Returns next task, if one exists, in order specified by mode.
813	>	*/
814	>	final ForkJoinTask<?> peek() {
815	>	ForkJoinTask<?>[] a = array; int m;
816	>	if (a == null || (m = a.length - 1) < 0)
817	>	return null;
818	>	int i = mode == 0 ? top - 1 : base;
819	>	int j = ((i & m) << ASHIFT) + ABASE;
820	>	return (ForkJoinTask<?>)U.getObjectVolatile(a, j);
821	>	}
822	>
823	>	/**
824	>	* Pops the given task only if it is at the current top.
825	>	* (A shared version is available only via FJP.tryExternalUnpush)
826	>	*/
827	>	final boolean tryUnpush(ForkJoinTask<?> t) {
828	>	ForkJoinTask<?>[] a; int s;
829	>	if ((a = array) != null && (s = top) != base &&
830	>	U.compareAndSwapObject
831	>	(a, (((a.length - 1) & --s) << ASHIFT) + ABASE, t, null)) {
832	>	top = s;
833	>	return true;
834	>	}
835	>	return false;
836	>	}
837	>
838	>	/**
839	>	* Removes and cancels all known tasks, ignoring any exceptions.
840	>	*/
841	>	final void cancelAll() {
842	>	ForkJoinTask.cancelIgnoringExceptions(currentJoin);
843	>	ForkJoinTask.cancelIgnoringExceptions(currentSteal);
844	>	for (ForkJoinTask<?> t; (t = poll()) != null; )
845	>	ForkJoinTask.cancelIgnoringExceptions(t);
846	>	}
847	>
848	>	/**
849	>	* Computes next value for random probes.  Scans don't require
850	>	* a very high quality generator, but also not a crummy one.
851	>	* Marsaglia xor-shift is cheap and works well enough.  Note:
852	>	* This is manually inlined in its usages in ForkJoinPool to
853	>	* avoid writes inside busy scan loops.
854	>	*/
855	>	final int nextSeed() {
856	>	int r = seed;
857	>	r ^= r << 13;
858	>	r ^= r >>> 17;
859	>	return seed = r ^= r << 5;
860	>	}
861	>
862	>	/**
863	>	* Provides a more accurate estimate of size than (top - base)
864	>	* by ordering reads and checking whether a near-empty queue
865	>	* has at least one unclaimed task.
866	>	*/
867	>	final int queueSize() {
868	>	ForkJoinTask<?>[] a; int k, s, n;
869	>	return ((n = base - (s = top)) < 0 &&
870	>	(n != -1 ||
871	>	((a = array) != null && (k = a.length) > 0 &&
872	>	U.getObject
873	>	(a, (long)((((k - 1) & (s - 1)) << ASHIFT) + ABASE)) != null))) ?
874	>	-n : 0;
875	>	}
876	>
877	>	// Specialized execution methods
878	>
879	>	/**
880	>	* Pops and runs tasks until empty.
881	>	*/
882	>	private void popAndExecAll() {
883	>	// A bit faster than repeated pop calls
884	>	ForkJoinTask<?>[] a; int m, s; long j; ForkJoinTask<?> t;
885	>	while ((a = array) != null && (m = a.length - 1) >= 0 &&
886	>	(s = top - 1) - base >= 0 &&
887	>	(t = ((ForkJoinTask<?>)
888	>	U.getObject(a, j = ((m & s) << ASHIFT) + ABASE)))
889	>	!= null) {
890	>	if (U.compareAndSwapObject(a, j, t, null)) {
891	>	top = s;
892	>	t.doExec();
893	>	}
894	>	}
895	>	}
896	>
897	>	/**
898	>	* Polls and runs tasks until empty.
899	>	*/
900	>	private void pollAndExecAll() {
901	>	for (ForkJoinTask<?> t; (t = poll()) != null;)
902	>	t.doExec();
903	>	}
904	>
905	>	/**
906	>	* If present, removes from queue and executes the given task,
907	>	* or any other cancelled task. Returns (true) on any CAS
908	>	* or consistency check failure so caller can retry.
909	>	*
910	>	* @return false if no progress can be made, else true;
911	>	*/
912	>	final boolean tryRemoveAndExec(ForkJoinTask<?> task) {
913	>	boolean stat = true, removed = false, empty = true;
914	>	ForkJoinTask<?>[] a; int m, s, b, n;
915	>	if ((a = array) != null && (m = a.length - 1) >= 0 &&
916	>	(n = (s = top) - (b = base)) > 0) {
917	>	for (ForkJoinTask<?> t;;) {           // traverse from s to b
918	>	int j = ((--s & m) << ASHIFT) + ABASE;
919	>	t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
920	>	if (t == null)                    // inconsistent length
921	>	break;
922	>	else if (t == task) {
923	>	if (s + 1 == top) {           // pop
924	>	if (!U.compareAndSwapObject(a, j, task, null))
925	>	break;
926	>	top = s;
927	>	removed = true;
928	>	}
929	>	else if (base == b)           // replace with proxy
930	>	removed = U.compareAndSwapObject(a, j, task,
931	>	new EmptyTask());
932	>	break;
933	>	}
934	>	else if (t.status >= 0)
935	>	empty = false;
936	>	else if (s + 1 == top) {          // pop and throw away
937	>	if (U.compareAndSwapObject(a, j, t, null))
938	>	top = s;
939	>	break;
940	>	}
941	>	if (--n == 0) {
942	>	if (!empty && base == b)
943	>	stat = false;
944	>	break;
945	>	}
946	>	}
947	>	}
948	>	if (removed)
949	>	task.doExec();
950	>	return stat;
951	>	}
952	>
953	>	/**
954	>	* Polls for and executes the given task or any other task in
955	>	* its CountedCompleter computation
956	>	*/
957	>	final boolean pollAndExecCC(ForkJoinTask<?> root) {
958	>	ForkJoinTask<?>[] a; int b; Object o;
959	>	outer: while ((b = base) - top < 0 && (a = array) != null) {
960	>	long j = (((a.length - 1) & b) << ASHIFT) + ABASE;
961	>	if ((o = U.getObject(a, j)) == null ||
962	>	!(o instanceof CountedCompleter))
963	>	break;
964	>	for (CountedCompleter<?> t = (CountedCompleter<?>)o, r = t;;) {
965	>	if (r == root) {
966	>	if (base == b &&
967	>	U.compareAndSwapObject(a, j, t, null)) {
968	>	base = b + 1;
969	>	t.doExec();
970	>	return true;
971	>	}
972	>	else
973	>	break; // restart
974	>	}
975	>	if ((r = r.completer) == null)
976	>	break outer; // not part of root computation
977	>	}
978	>	}
979	>	return false;
980	>	}
981	>
982	>	/**
983	>	* Executes a top-level task and any local tasks remaining
984	>	* after execution.
985	>	*/
986	>	final void runTask(ForkJoinTask<?> t) {
987	>	if (t != null) {
988	>	(currentSteal = t).doExec();
989	>	currentSteal = null;
990	>	++nsteals;
991	>	if (top != base) {       // process remaining local tasks
992	>	if (mode == 0)
993	>	popAndExecAll();
994	>	else
995	>	pollAndExecAll();
996	>	}
997	>	}
998	>	}
999	>
1000	>	/**
1001	>	* Executes a non-top-level (stolen) task.
1002	>	*/
1003	>	final void runSubtask(ForkJoinTask<?> t) {
1004	>	if (t != null) {
1005	>	ForkJoinTask<?> ps = currentSteal;
1006	>	(currentSteal = t).doExec();
1007	>	currentSteal = ps;
1008	>	}
1009	>	}
1010	>
1011	>	/**
1012	>	* Returns true if owned and not known to be blocked.
1013	>	*/
1014	>	final boolean isApparentlyUnblocked() {
1015	>	Thread wt; Thread.State s;
1016	>	return (eventCount >= 0 &&
1017	>	(wt = owner) != null &&
1018	>	(s = wt.getState()) != Thread.State.BLOCKED &&
1019	>	s != Thread.State.WAITING &&
1020	>	s != Thread.State.TIMED_WAITING);
1021	>	}
1022	>
1023	>	/**
1024	>	* If this owned and is not already interrupted, try to
1025	>	* interrupt and/or unpark, ignoring exceptions.
1026	>	*/
1027	>	final void interruptOwner() {
1028	>	Thread wt, p;
1029	>	if ((wt = owner) != null && !wt.isInterrupted()) {
1030	>	try {
1031	>	wt.interrupt();
1032	>	} catch (SecurityException ignore) {
1033	>	}
1034	>	}
1035	>	if ((p = parker) != null)
1036	>	U.unpark(p);
1037	>	}
1038	>
1039	>	// Unsafe mechanics
1040	>	private static final sun.misc.Unsafe U;
1041	>	private static final long QLOCK;
1042	>	private static final int ABASE;
1043	>	private static final int ASHIFT;
1044	>	static {
1045	>	int s;
1046	>	try {
1047	>	U = getUnsafe();
1048	>	Class<?> k = WorkQueue.class;
1049	>	Class<?> ak = ForkJoinTask[].class;
1050	>	QLOCK = U.objectFieldOffset
1051	>	(k.getDeclaredField("qlock"));
1052	>	ABASE = U.arrayBaseOffset(ak);
1053	>	s = U.arrayIndexScale(ak);
1054	>	} catch (Exception e) {
1055	>	throw new Error(e);
1056	>	}
1057	>	if ((s & (s-1)) != 0)
1058	>	throw new Error("data type scale not a power of two");
1059	>	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
1060	>	}
1061		}
1062
1063	<	/**
414	<	* Generator for assigning sequence numbers as pool names.
415	<	*/
416	<	private static final AtomicInteger poolNumberGenerator =
417	<	new AtomicInteger();
1063	>	// static fields (initialized in static initializer below)
1064
1065		/**
1066	<	* The time to block in a join (see awaitJoin) before checking if
1067	<	* a new worker should be (re)started to maintain parallelism
422	<	* level. The value should be short enough to maintain global
423	<	* responsiveness and progress but long enough to avoid
424	<	* counterproductive firings during GC stalls or unrelated system
425	<	* activity, and to not bog down systems with continual re-firings
426	<	* on GCs or legitimately long waits.
1066	>	* Creates a new ForkJoinWorkerThread. This factory is used unless
1067	>	* overridden in ForkJoinPool constructors.
1068		*/
1069	<	private static final long JOIN_TIMEOUT_MILLIS = 250L; // 4 per second
1069	>	public static final ForkJoinWorkerThreadFactory
1070	>	defaultForkJoinWorkerThreadFactory;
1071
1072		/**
1073	<	* The wakeup interval (in nanoseconds) for the oldest worker
1074	<	* waiting for an event invokes tryShutdownUnusedWorker to shrink
1075	<	* the number of workers.  The exact value does not matter too
1076	<	* much, but should be long enough to slowly release resources
1077	<	* during long periods without use without disrupting normal use.
1073	>	* Per-thread records for threads that submit to pools. Currently
1074	>	* holds only pseudo-random seed / index that is used to choose
1075	>	* submission queues in method externalPush. In the future, this may
1076	>	* also incorporate a means to implement different task rejection
1077	>	* and resubmission policies.
1078	>	*
1079	>	* Seeds for submitters and workers/workQueues work in basically
1080	>	* the same way but are initialized and updated using slightly
1081	>	* different mechanics. Both are initialized using the same
1082	>	* approach as in class ThreadLocal, where successive values are
1083	>	* unlikely to collide with previous values. Seeds are then
1084	>	* randomly modified upon collisions using xorshifts, which
1085	>	* requires a non-zero seed.
1086		*/
1087	<	private static final long SHRINK_RATE_NANOS =
1088	<	30L * 1000L * 1000L * 1000L; // 2 per minute
1087	>	static final class Submitter {
1088	>	int seed;
1089	>	Submitter(int s) { seed = s; }
1090	>	}
1091
1092		/**
1093	<	* Absolute bound for parallelism level. Twice this number plus
1094	<	* one (i.e., 0xfff) must fit into a 16bit field to enable
1095	<	* word-packing for some counts and indices.
1093	>	* Per-thread submission bookkeeping. Shared across all pools
1094	>	* to reduce ThreadLocal pollution and because random motion
1095	>	* to avoid contention in one pool is likely to hold for others.
1096	>	* Lazily initialized on first submission (but null-checked
1097	>	* in other contexts to avoid unnecessary initialization).
1098		*/
1099	<	private static final int MAX_WORKERS   = 0x7fff;
1099	>	static final ThreadLocal<Submitter> submitters;
1100
1101		/**
1102	<	* Array holding all worker threads in the pool.  Array size must
1103	<	* be a power of two.  Updates and replacements are protected by
1104	<	* workerLock, but the array is always kept in a consistent enough
1105	<	* state to be randomly accessed without locking by workers
452	<	* performing work-stealing, as well as other traversal-based
453	<	* methods in this class. All readers must tolerate that some
454	<	* array slots may be null.
1102	>	* Common (static) pool. Non-null for public use unless a static
1103	>	* construction exception, but internal usages null-check on use
1104	>	* to paranoically avoid potential initialization circularities
1105	>	* as well as to simplify generated code.
1106		*/
1107	<	volatile ForkJoinWorkerThread[] workers;
1107	>	static final ForkJoinPool commonPool;
1108
1109		/**
1110	<	* Queue for external submissions.
1110	>	* Permission required for callers of methods that may start or
1111	>	* kill threads.
1112		*/
1113	<	private final LinkedTransferQueue<ForkJoinTask<?>> submissionQueue;
1113	>	private static final RuntimePermission modifyThreadPermission;
1114
1115		/**
1116	<	* Lock protecting updates to workers array.
1116	>	* Common pool parallelism. Must equal commonPool.parallelism.
1117		*/
1118	<	private final ReentrantLock workerLock;
1118	>	static final int commonPoolParallelism;
1119
1120		/**
1121	<	* Latch released upon termination.
1121	>	* Sequence number for creating workerNamePrefix.
1122		*/
1123	<	private final Phaser termination;
1123	>	private static int poolNumberSequence;
1124
1125		/**
1126	<	* Creation factory for worker threads.
1126	>	* Return the next sequence number. We don't expect this to
1127	>	* ever contend so use simple builtin sync.
1128		*/
1129	<	private final ForkJoinWorkerThreadFactory factory;
1129	>	private static final synchronized int nextPoolId() {
1130	>	return ++poolNumberSequence;
1131	>	}
1132	>
1133	>	// static constants
1134
1135		/**
1136	<	* Sum of per-thread steal counts, updated only when threads are
1137	<	* idle or terminating.
1136	>	* Initial timeout value (in nanoseconds) for the thread
1137	>	* triggering quiescence to park waiting for new work. On timeout,
1138	>	* the thread will instead try to shrink the number of
1139	>	* workers. The value should be large enough to avoid overly
1140	>	* aggressive shrinkage during most transient stalls (long GCs
1141	>	* etc).
1142		*/
1143	<	private volatile long stealCount;
1143	>	private static final long IDLE_TIMEOUT      = 2000L * 1000L * 1000L; // 2sec
1144
1145		/**
1146	<	* Encoded record of top of Treiber stack of threads waiting for
486	<	* events. The top 32 bits contain the count being waited for. The
487	<	* bottom 16 bits contains one plus the pool index of waiting
488	<	* worker thread. (Bits 16-31 are unused.)
1146	>	* Timeout value when there are more threads than parallelism level
1147		*/
1148	<	private volatile long eventWaiters;
491	<
492	<	private static final int  EVENT_COUNT_SHIFT = 32;
493	<	private static final long WAITER_ID_MASK    = (1L << 16) - 1L;
1148	>	private static final long FAST_IDLE_TIMEOUT =  200L * 1000L * 1000L;
1149
1150		/**
1151	<	* A counter for events that may wake up worker threads:
1152	<	*   - Submission of a new task to the pool
1153	<	*   - A worker pushing a task on an empty queue
1154	<	*   - termination
1151	>	* The maximum stolen->joining link depth allowed in method
1152	>	* tryHelpStealer.  Must be a power of two.  Depths for legitimate
1153	>	* chains are unbounded, but we use a fixed constant to avoid
1154	>	* (otherwise unchecked) cycles and to bound staleness of
1155	>	* traversal parameters at the expense of sometimes blocking when
1156	>	* we could be helping.
1157		*/
1158	<	private volatile int eventCount;
1158	>	private static final int MAX_HELP = 64;
1159
1160		/**
1161	<	* Encoded record of top of Treiber stack of spare threads waiting
1162	<	* for resumption. The top 16 bits contain an arbitrary count to
506	<	* avoid ABA effects. The bottom 16bits contains one plus the pool
507	<	* index of waiting worker thread.
1161	>	* Increment for seed generators. See class ThreadLocal for
1162	>	* explanation.
1163		*/
1164	<	private volatile int spareWaiters;
510	<
511	<	private static final int SPARE_COUNT_SHIFT = 16;
512	<	private static final int SPARE_ID_MASK     = (1 << 16) - 1;
1164	>	private static final int SEED_INCREMENT = 0x61c88647;
1165
1166		/**
1167	<	* Lifecycle control. The low word contains the number of workers
1168	<	* that are (probably) executing tasks. This value is atomically
1169	<	* incremented before a worker gets a task to run, and decremented
1170	<	* when worker has no tasks and cannot find any.  Bits 16-18
1171	<	* contain runLevel value. When all are zero, the pool is
1172	<	* running. Level transitions are monotonic (running -> shutdown
1173	<	* -> terminating -> terminated) so each transition adds a bit.
1174	<	* These are bundled together to ensure consistent read for
1175	<	* termination checks (i.e., that runLevel is at least SHUTDOWN
1176	<	* and active threads is zero).
1167	>	* Bits and masks for control variables
1168	>	*
1169	>	* Field ctl is a long packed with:
1170	>	* AC: Number of active running workers minus target parallelism (16 bits)
1171	>	* TC: Number of total workers minus target parallelism (16 bits)
1172	>	* ST: true if pool is terminating (1 bit)
1173	>	* EC: the wait count of top waiting thread (15 bits)
1174	>	* ID: poolIndex of top of Treiber stack of waiters (16 bits)
1175	>	*
1176	>	* When convenient, we can extract the upper 32 bits of counts and
1177	>	* the lower 32 bits of queue state, u = (int)(ctl >>> 32) and e =
1178	>	* (int)ctl.  The ec field is never accessed alone, but always
1179	>	* together with id and st. The offsets of counts by the target
1180	>	* parallelism and the positionings of fields makes it possible to
1181	>	* perform the most common checks via sign tests of fields: When
1182	>	* ac is negative, there are not enough active workers, when tc is
1183	>	* negative, there are not enough total workers, and when e is
1184	>	* negative, the pool is terminating.  To deal with these possibly
1185	>	* negative fields, we use casts in and out of "short" and/or
1186	>	* signed shifts to maintain signedness.
1187	>	*
1188	>	* When a thread is queued (inactivated), its eventCount field is
1189	>	* set negative, which is the only way to tell if a worker is
1190	>	* prevented from executing tasks, even though it must continue to
1191	>	* scan for them to avoid queuing races. Note however that
1192	>	* eventCount updates lag releases so usage requires care.
1193	>	*
1194	>	* Field plock is an int packed with:
1195	>	* SHUTDOWN: true if shutdown is enabled (1 bit)
1196	>	* SEQ:  a sequence lock, with PL_LOCK bit set if locked (30 bits)
1197	>	* SIGNAL: set when threads may be waiting on the lock (1 bit)
1198		*
1199	<	* Notes: Most direct CASes are dependent on these bitfield
1200	<	* positions.  Also, this field is non-private to enable direct
1201	<	* performance-sensitive CASes in ForkJoinWorkerThread.
1199	>	* The sequence number enables simple consistency checks:
1200	>	* Staleness of read-only operations on the workQueues array can
1201	>	* be checked by comparing plock before vs after the reads.
1202		*/
530	–	volatile int runState;
1203
1204	<	// Note: The order among run level values matters.
1205	<	private static final int RUNLEVEL_SHIFT     = 16;
1206	<	private static final int SHUTDOWN           = 1 << RUNLEVEL_SHIFT;
1207	<	private static final int TERMINATING        = 1 << (RUNLEVEL_SHIFT + 1);
1208	<	private static final int TERMINATED         = 1 << (RUNLEVEL_SHIFT + 2);
537	<	private static final int ACTIVE_COUNT_MASK  = (1 << RUNLEVEL_SHIFT) - 1;
1204	>	// bit positions/shifts for fields
1205	>	private static final int  AC_SHIFT   = 48;
1206	>	private static final int  TC_SHIFT   = 32;
1207	>	private static final int  ST_SHIFT   = 31;
1208	>	private static final int  EC_SHIFT   = 16;
1209
1210	<	/**
1211	<	* Holds number of total (i.e., created and not yet terminated)
1212	<	* and running (i.e., not blocked on joins or other managed sync)
1213	<	* threads, packed together to ensure consistent snapshot when
1214	<	* making decisions about creating and suspending spare
1215	<	* threads. Updated only by CAS. Note that adding a new worker
1216	<	* requires incrementing both counts, since workers start off in
546	<	* running state.
547	<	*/
548	<	private volatile int workerCounts;
1210	>	// bounds
1211	>	private static final int  SMASK      = 0xffff;  // short bits
1212	>	private static final int  MAX_CAP    = 0x7fff;  // max #workers - 1
1213	>	private static final int  EVENMASK   = 0xfffe;  // even short bits
1214	>	private static final int  SQMASK     = 0x007e;  // max 64 (even) slots
1215	>	private static final int  SHORT_SIGN = 1 << 15;
1216	>	private static final int  INT_SIGN   = 1 << 31;
1217
1218	<	private static final int TOTAL_COUNT_SHIFT  = 16;
1219	<	private static final int RUNNING_COUNT_MASK = (1 << TOTAL_COUNT_SHIFT) - 1;
1220	<	private static final int ONE_RUNNING        = 1;
1221	<	private static final int ONE_TOTAL          = 1 << TOTAL_COUNT_SHIFT;
1218	>	// masks
1219	>	private static final long STOP_BIT   = 0x0001L << ST_SHIFT;
1220	>	private static final long AC_MASK    = ((long)SMASK) << AC_SHIFT;
1221	>	private static final long TC_MASK    = ((long)SMASK) << TC_SHIFT;
1222
1223	<	/**
1224	<	* The target parallelism level.
1225	<	* Accessed directly by ForkJoinWorkerThreads.
558	<	*/
559	<	final int parallelism;
1223	>	// units for incrementing and decrementing
1224	>	private static final long TC_UNIT    = 1L << TC_SHIFT;
1225	>	private static final long AC_UNIT    = 1L << AC_SHIFT;
1226
1227	<	/**
1228	<	* True if use local fifo, not default lifo, for local polling
1229	<	* Read by, and replicated by ForkJoinWorkerThreads
1230	<	*/
1231	<	final boolean locallyFifo;
1227	>	// masks and units for dealing with u = (int)(ctl >>> 32)
1228	>	private static final int  UAC_SHIFT  = AC_SHIFT - 32;
1229	>	private static final int  UTC_SHIFT  = TC_SHIFT - 32;
1230	>	private static final int  UAC_MASK   = SMASK << UAC_SHIFT;
1231	>	private static final int  UTC_MASK   = SMASK << UTC_SHIFT;
1232	>	private static final int  UAC_UNIT   = 1 << UAC_SHIFT;
1233	>	private static final int  UTC_UNIT   = 1 << UTC_SHIFT;
1234
1235	<	/**
1236	<	* The uncaught exception handler used when any worker abruptly
1237	<	* terminates.
570	<	*/
571	<	private final Thread.UncaughtExceptionHandler ueh;
1235	>	// masks and units for dealing with e = (int)ctl
1236	>	private static final int E_MASK      = 0x7fffffff; // no STOP_BIT
1237	>	private static final int E_SEQ       = 1 << EC_SHIFT;
1238
1239	<	/**
1240	<	* Pool number, just for assigning useful names to worker threads
1241	<	*/
1242	<	private final int poolNumber;
1239	>	// plock bits
1240	>	private static final int SHUTDOWN    = 1 << 31;
1241	>	private static final int PL_LOCK     = 2;
1242	>	private static final int PL_SIGNAL   = 1;
1243	>	private static final int PL_SPINS    = 1 << 8;
1244
1245	<	// Utilities for CASing fields. Note that most of these
1246	<	// are usually manually inlined by callers
1245	>	// access mode for WorkQueue
1246	>	static final int LIFO_QUEUE          =  0;
1247	>	static final int FIFO_QUEUE          =  1;
1248	>	static final int SHARED_QUEUE        = -1;
1249
1250	<	/**
1251	<	* Increments running count part of workerCounts
1252	<	*/
584	<	final void incrementRunningCount() {
585	<	int c;
586	<	do {} while (!UNSAFE.compareAndSwapInt(this, workerCountsOffset,
587	<	c = workerCounts,
588	<	c + ONE_RUNNING));
589	<	}
1250	>	// bounds for #steps in scan loop -- must be power 2 minus 1
1251	>	private static final int MIN_SCAN    = 0x1ff;   // cover estimation slop
1252	>	private static final int MAX_SCAN    = 0x1ffff; // 4 * max workers
1253
1254	<	/**
592	<	* Tries to decrement running count unless already zero
593	<	*/
594	<	final boolean tryDecrementRunningCount() {
595	<	int wc = workerCounts;
596	<	if ((wc & RUNNING_COUNT_MASK) == 0)
597	<	return false;
598	<	return UNSAFE.compareAndSwapInt(this, workerCountsOffset,
599	<	wc, wc - ONE_RUNNING);
600	<	}
1254	>	// Instance fields
1255
1256	<	/**
1257	<	* Forces decrement of encoded workerCounts, awaiting nonzero if
1258	<	* (rarely) necessary when other count updates lag.
1259	<	*
1260	<	* @param dr -- either zero or ONE_RUNNING
1261	<	* @param dt == either zero or ONE_TOTAL
1256	>	/*
1257	>	* Field layout of this class tends to matter more than one would
1258	>	* like. Runtime layout order is only loosely related to
1259	>	* declaration order and may differ across JVMs, but the following
1260	>	* empirically works OK on current JVMs.
1261	>	*/
1262	>	volatile long stealCount;                  // collects worker counts
1263	>	volatile long ctl;                         // main pool control
1264	>	volatile int plock;                        // shutdown status and seqLock
1265	>	volatile int indexSeed;                    // worker/submitter index seed
1266	>	final int config;                          // mode and parallelism level
1267	>	WorkQueue[] workQueues;                    // main registry
1268	>	final ForkJoinWorkerThreadFactory factory;
1269	>	final Thread.UncaughtExceptionHandler ueh; // per-worker UEH
1270	>	final String workerNamePrefix;             // to create worker name string
1271	>
1272	>	/*
1273	>	* Acquires the plock lock to protect worker array and related
1274	>	* updates. This method is called only if an initial CAS on plock
1275	>	* fails. This acts as a spinLock for normal cases, but falls back
1276	>	* to builtin monitor to block when (rarely) needed. This would be
1277	>	* a terrible idea for a highly contended lock, but works fine as
1278	>	* a more conservative alternative to a pure spinlock.  See
1279	>	* internal ConcurrentHashMap documentation for further
1280	>	* explanation of nearly the same construction.
1281		*/
1282	<	private void decrementWorkerCounts(int dr, int dt) {
1282	>	private int acquirePlock() {
1283	>	int spins = PL_SPINS, r = 0, ps, nps;
1284		for (;;) {
1285	<	int wc = workerCounts;
1286	<	if ((wc & RUNNING_COUNT_MASK)  - dr < 0 ||
1287	<	(wc >>> TOTAL_COUNT_SHIFT) - dt < 0) {
1288	<	if ((runState & TERMINATED) != 0)
1289	<	return; // lagging termination on a backout
1290	<	Thread.yield();
1285	>	if (((ps = plock) & PL_LOCK) == 0 &&
1286	>	U.compareAndSwapInt(this, PLOCK, ps, nps = ps + PL_LOCK))
1287	>	return nps;
1288	>	else if (r == 0) { // randomize spins if possible
1289	>	Thread t = Thread.currentThread(); WorkQueue w; Submitter z;
1290	>	if ((t instanceof ForkJoinWorkerThread) &&
1291	>	(w = ((ForkJoinWorkerThread)t).workQueue) != null)
1292	>	r = w.seed;
1293	>	else if ((z = submitters.get()) != null)
1294	>	r = z.seed;
1295	>	else
1296	>	r = 1;
1297	>	}
1298	>	else if (spins >= 0) {
1299	>	r ^= r << 1; r ^= r >>> 3; r ^= r << 10; // xorshift
1300	>	if (r >= 0)
1301	>	--spins;
1302	>	}
1303	>	else if (U.compareAndSwapInt(this, PLOCK, ps, ps | PL_SIGNAL)) {
1304	>	synchronized (this) {
1305	>	if ((plock & PL_SIGNAL) != 0) {
1306	>	try {
1307	>	wait();
1308	>	} catch (InterruptedException ie) {
1309	>	try {
1310	>	Thread.currentThread().interrupt();
1311	>	} catch (SecurityException ignore) {
1312	>	}
1313	>	}
1314	>	}
1315	>	else
1316	>	notifyAll();
1317	>	}
1318		}
618	–	if (UNSAFE.compareAndSwapInt(this, workerCountsOffset,
619	–	wc, wc - (dr + dt)))
620	–	return;
1319		}
1320		}
1321
1322		/**
1323	<	* Tries decrementing active count; fails on contention.
1324	<	* Called when workers cannot find tasks to run.
1323	>	* Unlocks and signals any thread waiting for plock. Called only
1324	>	* when CAS of seq value for unlock fails.
1325		*/
1326	<	final boolean tryDecrementActiveCount() {
1327	<	int c;
1328	<	return UNSAFE.compareAndSwapInt(this, runStateOffset,
631	<	c = runState, c - 1);
1326	>	private void releasePlock(int ps) {
1327	>	plock = ps;
1328	>	synchronized (this) { notifyAll(); }
1329		}
1330
1331		/**
1332	<	* Advances to at least the given level. Returns true if not
636	<	* already in at least the given level.
1332	>	* Tries to create and start a worker; adjusts counts etc on failure
1333		*/
1334	<	private boolean advanceRunLevel(int level) {
1335	<	for (;;) {
1336	<	int s = runState;
1337	<	if ((s & level) != 0)
1338	<	return false;
1339	<	if (UNSAFE.compareAndSwapInt(this, runStateOffset, s, s | level))
1340	<	return true;
1334	>	private void addWorker() {
1335	>	ForkJoinWorkerThread wt = null;
1336	>	try {
1337	>	(wt = factory.newThread(this)).start();
1338	>	} catch (Throwable ex) {
1339	>	deregisterWorker(wt, ex); // adjust on failure
1340	>	}
1341	>	}
1342	>
1343	>	/**
1344	>	* Performs secondary initialization, called when plock is zero.
1345	>	* Creates workQueue array and sets plock to a valid value.  The
1346	>	* lock body must be exception-free (so no try/finally) so we
1347	>	* optimistically allocate new array outside the lock and throw
1348	>	* away if (very rarely) not needed. (A similar tactic is used in
1349	>	* fullExternalPush.)  Because the plock seq value can eventually
1350	>	* wrap around zero, this method harmlessly fails to reinitialize
1351	>	* if workQueues exists, while still advancing plock.
1352	>	*/
1353	>	private void initWorkQueuesArray() {
1354	>	WorkQueue[] ws; int ps;
1355	>	int p = config & SMASK;        // find power of two table size
1356	>	int n = (p > 1) ? p - 1 : 1;   // ensure at least 2 slots
1357	>	n |= n >>> 1; n |= n >>> 2; n |= n >>> 4; n |= n >>> 8; n |= n >>> 16;
1358	>	WorkQueue[] nws = new WorkQueue[(n + 1) << 1];
1359	>	if (((ps = plock) & PL_LOCK) != 0 ||
1360	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1361	>	ps = acquirePlock();
1362	>	if ((ws = workQueues) == null || ws.length == 0)
1363	>	workQueues = nws;
1364	>	int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1365	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1366	>	releasePlock(nps);
1367	>	long c; int u;
1368	>	if ((u = (int)((c = ctl) >>> 32)) < 0 && (int)c == 0) {
1369	>	long nc = (long)(((u + UTC_UNIT) & UTC_MASK) |
1370	>	((u + UAC_UNIT) & UAC_MASK)) << 32;
1371	>	if (U.compareAndSwapLong(this, CTL, c, nc))
1372	>	addWorker();
1373		}
1374	+
1375		}
1376
1377	<	// workers array maintenance
1377	>	//  Registering and deregistering workers
1378	>
1379	>	/**
1380	>	* Callback from ForkJoinWorkerThread to establish and record its
1381	>	* WorkQueue. To avoid scanning bias due to packing entries in
1382	>	* front of the workQueues array, we treat the array as a simple
1383	>	* power-of-two hash table using per-thread seed as hash,
1384	>	* expanding as needed.
1385	>	*
1386	>	* @param wt the worker thread
1387	>	*/
1388	>	final void registerWorker(ForkJoinWorkerThread wt) {
1389	>	if (wt != null && wt.workQueue == null) {
1390	>	int s, ps;    // generate a rarely colliding candidate index seed
1391	>	do {} while (!U.compareAndSwapInt(this, INDEXSEED, s = indexSeed,
1392	>	s += SEED_INCREMENT) ||
1393	>	s == 0); // skip 0
1394	>	WorkQueue w = new WorkQueue(this, wt, config >>> 16, s);
1395	>	if (((ps = plock) & PL_LOCK) != 0 ||
1396	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1397	>	ps = acquirePlock();
1398	>	int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1399	>	try {
1400	>	WorkQueue[] ws;
1401	>	if ((ws = workQueues) != null && wt.workQueue == null) {
1402	>	int n = ws.length, m = n - 1;
1403	>	int r = (s << 1) | 1;           // use odd-numbered indices
1404	>	if (ws[r &= m] != null) {       // collision
1405	>	int probes = 0;             // step by approx half size
1406	>	int step = (n <= 4) ? 2 : ((n >>> 1) & EVENMASK) + 2;
1407	>	while (ws[r = (r + step) & m] != null) {
1408	>	if (++probes >= n) {
1409	>	workQueues = ws = Arrays.copyOf(ws, n <<= 1);
1410	>	m = n - 1;
1411	>	probes = 0;
1412	>	}
1413	>	}
1414	>	}
1415	>	w.eventCount = w.poolIndex = r; // volatile write orders
1416	>	wt.workQueue = ws[r] = w;
1417	>	}
1418	>	} finally {
1419	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1420	>	releasePlock(nps);
1421	>	}
1422	>	}
1423	>	}
1424
1425		/**
1426	<	* Records and returns a workers array index for new worker.
1427	<	*/
1428	<	private int recordWorker(ForkJoinWorkerThread w) {
1429	<	// Try using slot totalCount-1. If not available, scan and/or resize
1430	<	int k = (workerCounts >>> TOTAL_COUNT_SHIFT) - 1;
1431	<	final ReentrantLock lock = this.workerLock;
1432	<	lock.lock();
1433	<	try {
1434	<	ForkJoinWorkerThread[] ws = workers;
1435	<	int n = ws.length;
1436	<	if (k < 0 || k >= n || ws[k] != null) {
1437	<	for (k = 0; k < n && ws[k] != null; ++k)
1438	<	;
1439	<	if (k == n)
1440	<	ws = Arrays.copyOf(ws, n << 1);
1426	>	* Final callback from terminating worker, as well as upon failure
1427	>	* to construct or start a worker.  Removes record of worker from
1428	>	* array, and adjusts counts. If pool is shutting down, tries to
1429	>	* complete termination.
1430	>	*
1431	>	* @param wt the worker thread or null if construction failed
1432	>	* @param ex the exception causing failure, or null if none
1433	>	*/
1434	>	final void deregisterWorker(ForkJoinWorkerThread wt, Throwable ex) {
1435	>	WorkQueue w = null;
1436	>	if (wt != null && (w = wt.workQueue) != null) {
1437	>	int ps;
1438	>	w.qlock = -1;                // ensure set
1439	>	long ns = w.nsteals, sc;     // collect steal count
1440	>	do {} while (!U.compareAndSwapLong(this, STEALCOUNT,
1441	>	sc = stealCount, sc + ns));
1442	>	if (((ps = plock) & PL_LOCK) != 0 ||
1443	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1444	>	ps = acquirePlock();
1445	>	int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1446	>	try {
1447	>	int idx = w.poolIndex;
1448	>	WorkQueue[] ws = workQueues;
1449	>	if (ws != null && idx >= 0 && idx < ws.length && ws[idx] == w)
1450	>	ws[idx] = null;
1451	>	} finally {
1452	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1453	>	releasePlock(nps);
1454		}
667	–	ws[k] = w;
668	–	workers = ws; // volatile array write ensures slot visibility
669	–	} finally {
670	–	lock.unlock();
1455		}
1456	<	return k;
1456	>
1457	>	long c;                             // adjust ctl counts
1458	>	do {} while (!U.compareAndSwapLong
1459	>	(this, CTL, c = ctl, (((c - AC_UNIT) & AC_MASK) |
1460	>	((c - TC_UNIT) & TC_MASK) |
1461	>	(c & ~(AC_MASK|TC_MASK)))));
1462	>
1463	>	if (!tryTerminate(false, false) && w != null) {
1464	>	w.cancelAll();                  // cancel remaining tasks
1465	>	if (w.array != null)            // suppress signal if never ran
1466	>	helpSignal(null, 0);        // wake up or create replacement
1467	>	if (ex == null)                 // help clean refs on way out
1468	>	ForkJoinTask.helpExpungeStaleExceptions();
1469	>	}
1470	>
1471	>	if (ex != null)                     // rethrow
1472	>	ForkJoinTask.rethrow(ex);
1473	>	}
1474	>
1475	>	// Submissions
1476	>
1477	>	/**
1478	>	* Unless shutting down, adds the given task to a submission queue
1479	>	* at submitter's current queue index (modulo submission
1480	>	* range). Only the most common path is directly handled in this
1481	>	* method. All others are relayed to fullExternalPush.
1482	>	*
1483	>	* @param task the task. Caller must ensure non-null.
1484	>	*/
1485	>	final void externalPush(ForkJoinTask<?> task) {
1486	>	WorkQueue[] ws; WorkQueue q; Submitter z; int m; ForkJoinTask<?>[] a;
1487	>	if ((z = submitters.get()) != null && plock > 0 &&
1488	>	(ws = workQueues) != null && (m = (ws.length - 1)) >= 0 &&
1489	>	(q = ws[m & z.seed & SQMASK]) != null &&
1490	>	U.compareAndSwapInt(q, QLOCK, 0, 1)) { // lock
1491	>	int b = q.base, s = q.top, n, an;
1492	>	if ((a = q.array) != null && (an = a.length) > (n = s + 1 - b)) {
1493	>	U.putObject(a, (long)(((an - 1) & s) << ASHIFT) + ABASE, task);
1494	>	q.top = s + 1;                     // push on to deque
1495	>	q.qlock = 0;
1496	>	if (n <= 2)
1497	>	signalWork(q, 0);
1498	>	return;
1499	>	}
1500	>	q.qlock = 0;
1501	>	}
1502	>	fullExternalPush(task);
1503		}
1504
1505		/**
1506	<	* Nulls out record of worker in workers array.
1507	<	*/
1508	<	private void forgetWorker(ForkJoinWorkerThread w) {
1509	<	int idx = w.poolIndex;
1510	<	// Locking helps method recordWorker avoid unnecessary expansion
1511	<	final ReentrantLock lock = this.workerLock;
1512	<	lock.lock();
1513	<	try {
1514	<	ForkJoinWorkerThread[] ws = workers;
1515	<	if (idx >= 0 && idx < ws.length && ws[idx] == w) // verify
1516	<	ws[idx] = null;
1517	<	} finally {
1518	<	lock.unlock();
1506	>	* Full version of externalPush. This method is called, among
1507	>	* other times, upon the first submission of the first task to the
1508	>	* pool, so must perform secondary initialization (via
1509	>	* initWorkQueuesArray). It also detects first submission by an
1510	>	* external thread by looking up its ThreadLocal, and creates a
1511	>	* new shared queue if the one at index if empty or contended. The
1512	>	* lock body must be exception-free (so no try/finally) so we
1513	>	* optimistically allocate new queues outside the lock and throw
1514	>	* them away if (very rarely) not needed.
1515	>	*/
1516	>	private void fullExternalPush(ForkJoinTask<?> task) {
1517	>	int r = 0;
1518	>	for (Submitter z = submitters.get();;) {
1519	>	WorkQueue[] ws; WorkQueue q; int ps, m, k;
1520	>	if (z == null) {
1521	>	if (U.compareAndSwapInt(this, INDEXSEED, r = indexSeed,
1522	>	r += SEED_INCREMENT) && r != 0)
1523	>	submitters.set(z = new Submitter(r));
1524	>	}
1525	>	else if (r == 0) {               // move to a different index
1526	>	r = z.seed;
1527	>	r ^= r << 13;                // same xorshift as WorkQueues
1528	>	r ^= r >>> 17;
1529	>	z.seed = r ^ (r << 5);
1530	>	}
1531	>	else if ((ps = plock) < 0)
1532	>	throw new RejectedExecutionException();
1533	>	else if (ps == 0 || (ws = workQueues) == null ||
1534	>	(m = ws.length - 1) < 0)
1535	>	initWorkQueuesArray();
1536	>	else if ((q = ws[k = r & m & SQMASK]) != null) {
1537	>	if (q.trySharedPush(task))
1538	>	return;
1539	>	else
1540	>	r = 0; // move on contention
1541	>	}
1542	>	else if (((ps = plock) & PL_LOCK) == 0) { // create new queue
1543	>	q = new WorkQueue(this, null, SHARED_QUEUE, r);
1544	>	if (((ps = plock) & PL_LOCK) != 0 ||
1545	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1546	>	ps = acquirePlock();
1547	>	if ((ws = workQueues) != null && k < ws.length && ws[k] == null)
1548	>	ws[k] = q;
1549	>	int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1550	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1551	>	releasePlock(nps);
1552	>	}
1553	>	else
1554	>	r = 0; // try elsewhere while lock held
1555		}
1556		}
1557
1558	+	// Maintaining ctl counts
1559	+
1560		/**
1561	<	* Final callback from terminating worker.  Removes record of
694	<	* worker from array, and adjusts counts. If pool is shutting
695	<	* down, tries to complete termination.
696	<	*
697	<	* @param w the worker
1561	>	* Increments active count; mainly called upon return from blocking.
1562		*/
1563	<	final void workerTerminated(ForkJoinWorkerThread w) {
1564	<	forgetWorker(w);
1565	<	decrementWorkerCounts(w.isTrimmed()? 0 : ONE_RUNNING, ONE_TOTAL);
702	<	while (w.stealCount != 0) // collect final count
703	<	tryAccumulateStealCount(w);
704	<	tryTerminate(false);
1563	>	final void incrementActiveCount() {
1564	>	long c;
1565	>	do {} while (!U.compareAndSwapLong(this, CTL, c = ctl, c + AC_UNIT));
1566		}
1567
707	–	// Waiting for and signalling events
708	–
1568		/**
1569	<	* Releases workers blocked on a count not equal to current count.
1570	<	* Normally called after precheck that eventWaiters isn't zero to
1571	<	* avoid wasted array checks. Gives up upon a change in count or
1572	<	* upon releasing two workers, letting others take over.
1569	>	* Tries to create (at most one) or activate (possibly several)
1570	>	* workers if too few are active. On contention failure, continues
1571	>	* until at least one worker is signalled or the given queue is
1572	>	* empty or all workers are active.
1573	>	*
1574	>	* @param q if non-null, the queue holding tasks to be signalled
1575	>	* @param signals the target number of signals (at least one --
1576	>	* if argument is zero also sets signallee hint if parked).
1577		*/
1578	<	private void releaseEventWaiters() {
1579	<	ForkJoinWorkerThread[] ws = workers;
1580	<	int n = ws.length;
1581	<	long h = eventWaiters;
1582	<	int ec = eventCount;
1583	<	boolean releasedOne = false;
1584	<	ForkJoinWorkerThread w; int id;
1585	<	while ((id = ((int)(h & WAITER_ID_MASK)) - 1) >= 0 &&
1586	<	(int)(h >>> EVENT_COUNT_SHIFT) != ec &&
1587	<	id < n && (w = ws[id]) != null) {
1588	<	if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
1589	<	h,  w.nextWaiter)) {
1590	<	LockSupport.unpark(w);
1591	<	if (releasedOne) // exit on second release
1578	>	final void signalWork(WorkQueue q, int signals) {
1579	>	long c; int e, u, i, s; WorkQueue[] ws; WorkQueue w; Thread p;
1580	>	while ((u = (int)((c = ctl) >>> 32)) < 0) {
1581	>	if ((e = (int)c) > 0) {
1582	>	if ((ws = workQueues) != null && ws.length > (i = e & SMASK) &&
1583	>	(w = ws[i]) != null && w.eventCount == (e | INT_SIGN)) {
1584	>	long nc = (((long)(w.nextWait & E_MASK)) |
1585	>	((long)(u + UAC_UNIT) << 32));
1586	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1587	>	w.eventCount = (e + E_SEQ) & E_MASK;
1588	>	if ((p = w.parker) != null) {
1589	>	if (q != null && signals == 0)
1590	>	w.hint = q.poolIndex;
1591	>	U.unpark(p);
1592	>	}
1593	>	if (--signals <= 0)
1594	>	break;
1595	>	}
1596	>	if (q != null && (s = q.queueSize()) <= signals &&
1597	>	(signals = s) <= 0)
1598	>	break;
1599	>	}
1600	>	else
1601		break;
730	–	releasedOne = true;
1602		}
1603	<	if (eventCount != ec)
1603	>	else if (e == 0 && (u & SHORT_SIGN) != 0) {
1604	>	long nc = (long)(((u + UTC_UNIT) & UTC_MASK) |
1605	>	((u + UAC_UNIT) & UAC_MASK)) << 32;
1606	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1607	>	addWorker();
1608	>	break;
1609	>	}
1610	>	}
1611	>	else
1612		break;
734	–	h = eventWaiters;
1613		}
1614		}
1615
1616	+	// Scanning for tasks
1617	+
1618		/**
1619	<	* Tries to advance eventCount and releases waiters. Called only
740	<	* from workers.
1619	>	* Top-level runloop for workers, called by ForkJoinWorkerThread.run.
1620		*/
1621	<	final void signalWork() {
1622	<	int c; // try to increment event count -- CAS failure OK
1623	<	UNSAFE.compareAndSwapInt(this, eventCountOffset, c = eventCount, c+1);
745	<	if (eventWaiters != 0L)
746	<	releaseEventWaiters();
1621	>	final void runWorker(WorkQueue w) {
1622	>	if (w != null) // skip on initialization failure
1623	>	do { w.runTask(scan(w)); } while (w.qlock >= 0);
1624		}
1625
1626		/**
1627	<	* Adds the given worker to event queue and blocks until
1628	<	* terminating or event count advances from the given value
1629	<	*
1630	<	* @param w the calling worker thread
1631	<	* @param ec the count
1632	<	*/
1633	<	private void eventSync(ForkJoinWorkerThread w, int ec) {
1634	<	long nh = (((long)ec) << EVENT_COUNT_SHIFT) | ((long)(w.poolIndex+1));
1635	<	long h;
1636	<	while ((runState < SHUTDOWN || !tryTerminate(false)) &&
1637	<	(((int)((h = eventWaiters) & WAITER_ID_MASK)) == 0 ||
1638	<	(int)(h >>> EVENT_COUNT_SHIFT) == ec) &&
1639	<	eventCount == ec) {
1640	<	if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
1641	<	w.nextWaiter = h, nh)) {
1642	<	awaitEvent(w, ec);
1643	<	break;
1627	>	* Scans for and, if found, returns one task, else possibly
1628	>	* inactivates the worker. This method operates on single reads of
1629	>	* volatile state and is designed to be re-invoked continuously,
1630	>	* in part because it returns upon detecting inconsistencies,
1631	>	* contention, or state changes that indicate possible success on
1632	>	* re-invocation.
1633	>	*
1634	>	* The scan searches for tasks across queues (starting at a random
1635	>	* index, and relying on registerWorker to irregularly scatter
1636	>	* them within array to avoid bias), checking each at least twice.
1637	>	* The scan terminates upon either finding a non-empty queue, or
1638	>	* completing the sweep. If the worker is not inactivated, it
1639	>	* takes and returns a task from this queue. Otherwise, if not
1640	>	* activated, it signals workers (that may include itself) and
1641	>	* returns so caller can retry. Also returns for true if the
1642	>	* worker array may have changed during an empty scan.  On failure
1643	>	* to find a task, we take one of the following actions, after
1644	>	* which the caller will retry calling this method unless
1645	>	* terminated.
1646	>	*
1647	>	* * If pool is terminating, terminate the worker.
1648	>	*
1649	>	* * If not already enqueued, try to inactivate and enqueue the
1650	>	* worker on wait queue. Or, if inactivating has caused the pool
1651	>	* to be quiescent, relay to idleAwaitWork to check for
1652	>	* termination and possibly shrink pool.
1653	>	*
1654	>	* * If already enqueued and none of the above apply, possibly
1655	>	* (with 1/2 probability) park awaiting signal, else lingering to
1656	>	* help scan and signal.
1657	>	*
1658	>	* @param w the worker (via its WorkQueue)
1659	>	* @return a task or null if none found
1660	>	*/
1661	>	private final ForkJoinTask<?> scan(WorkQueue w) {
1662	>	WorkQueue[] ws; int m, hint;
1663	>	int ps = plock;                          // read plock before ws
1664	>	if (w != null && (ws = workQueues) != null && (m = ws.length - 1) >= 0) {
1665	>	int ec = w.eventCount;               // ec is negative if inactive
1666	>	int r = w.seed; r ^= r << 13; r ^= r >>> 17; w.seed = r ^= r << 5;
1667	>	for (int j = ((m + m + 1) | MIN_SCAN) & MAX_SCAN; ; --j) {
1668	>	WorkQueue q; ForkJoinTask<?>[] a; int b;
1669	>	if ((q = ws[(r + j) & m]) != null && (b = q.base) - q.top < 0 &&
1670	>	(a = q.array) != null) {     // probably nonempty
1671	>	int i = (((a.length - 1) & b) << ASHIFT) + ABASE;
1672	>	ForkJoinTask<?> t = (ForkJoinTask<?>)
1673	>	U.getObjectVolatile(a, i);
1674	>	if (q.base == b && ec >= 0 && t != null &&
1675	>	U.compareAndSwapObject(a, i, t, null)) {
1676	>	if ((q.base = b + 1) - q.top < 0)
1677	>	signalWork(q, 0);
1678	>	return t;                // taken
1679	>	}
1680	>	else if (ec < 0 || j < m) {  // cannot take or cannot rescan
1681	>	w.hint = q.poolIndex;    // use hint below
1682	>	break;                   // let caller retry after signal
1683	>	}
1684	>	}
1685	>	else if (j < 0) { // end of scan; in loop to simplify code
1686	>	long c, sc; int e, ns;
1687	>	if ((ns = w.nsteals) != 0) {
1688	>	if (U.compareAndSwapLong(this, STEALCOUNT,
1689	>	sc = stealCount, sc + ns))
1690	>	w.nsteals = 0;       // collect steals
1691	>	}
1692	>	else if (plock != ps)        // ws may have changed
1693	>	break;
1694	>	else if ((e = (int)(c = ctl)) < 0)
1695	>	w.qlock = -1;            // pool is terminating
1696	>	else if (ec >= 0) {          // try to enqueue/inactivate
1697	>	long nc = ((long)ec |
1698	>	((c - AC_UNIT) & (AC_MASK|TC_MASK)));
1699	>	w.nextWait = e;          // link and mark inactive
1700	>	w.hint = -1;             // use hint if set while parked
1701	>	w.eventCount = ec | INT_SIGN;
1702	>	if (ctl != c ||
1703	>	!U.compareAndSwapLong(this, CTL, c, nc))
1704	>	w.eventCount = ec;  // unmark on CAS failure
1705	>	else if ((int)(c >> AC_SHIFT) == 1 - (config & SMASK))
1706	>	idleAwaitWork(w, nc, c);
1707	>	}
1708	>	else if (w.eventCount < 0) { // block
1709	>	Thread wt = Thread.currentThread();
1710	>	Thread.interrupted();    // clear status
1711	>	U.putObject(wt, PARKBLOCKER, this);
1712	>	w.parker = wt;           // emulate LockSupport.park
1713	>	if (w.eventCount < 0)    // recheck
1714	>	U.park(false, 0L);
1715	>	w.parker = null;
1716	>	U.putObject(wt, PARKBLOCKER, null);
1717	>	}
1718	>	break;
1719	>	}
1720	>	}
1721	>	if ((hint = w.hint) >= 0) {          // help signal
1722	>	WorkQueue[] vs; WorkQueue v; int k;
1723	>	w.hint = -1;                     // suppress resignal
1724	>	if ((vs = workQueues) != null && hint < vs.length &&
1725	>	(v = vs[hint]) != null && (k = v.base - v.top) < -1)
1726	>	signalWork(v, 1 - k);
1727		}
1728		}
1729	+	return null;
1730		}
1731
1732		/**
1733	<	* Blocks the given worker (that has already been entered as an
1734	<	* event waiter) until terminating or event count advances from
1735	<	* the given value. The oldest (first) waiter uses a timed wait to
1736	<	* occasionally one-by-one shrink the number of workers (to a
1737	<	* minimum of one) if the pool has not been used for extended
1738	<	* periods.
1739	<	*
1740	<	* @param w the calling worker thread
1741	<	* @param ec the count
1742	<	*/
1743	<	private void awaitEvent(ForkJoinWorkerThread w, int ec) {
1744	<	while (eventCount == ec) {
1745	<	if (tryAccumulateStealCount(w)) { // transfer while idle
1746	<	boolean untimed = (w.nextWaiter != 0L ||
1747	<	(workerCounts & RUNNING_COUNT_MASK) <= 1);
1748	<	long startTime = untimed? 0 : System.nanoTime();
1749	<	Thread.interrupted();         // clear/ignore interrupt
1750	<	if (eventCount != ec || w.runState != 0 ||
1751	<	runState >= TERMINATING)  // recheck after clear
1733	>	* If inactivating worker w has caused the pool to become
1734	>	* quiescent, checks for pool termination, and, so long as this is
1735	>	* not the only worker, waits for event for up to a given
1736	>	* duration.  On timeout, if ctl has not changed, terminates the
1737	>	* worker, which will in turn wake up another worker to possibly
1738	>	* repeat this process.
1739	>	*
1740	>	* @param w the calling worker
1741	>	* @param currentCtl the ctl value triggering possible quiescence
1742	>	* @param prevCtl the ctl value to restore if thread is terminated
1743	>	*/
1744	>	private void idleAwaitWork(WorkQueue w, long currentCtl, long prevCtl) {
1745	>	if (w != null && w.eventCount < 0 &&
1746	>	!tryTerminate(false, false) && (int)prevCtl != 0) {
1747	>	int dc = -(short)(currentCtl >>> TC_SHIFT);
1748	>	long parkTime = dc < 0 ? FAST_IDLE_TIMEOUT: (dc + 1) * IDLE_TIMEOUT;
1749	>	long deadline = System.nanoTime() + parkTime - 100000L; // 1ms slop
1750	>	Thread wt = Thread.currentThread();
1751	>	while (ctl == currentCtl) {
1752	>	Thread.interrupted();  // timed variant of version in scan()
1753	>	U.putObject(wt, PARKBLOCKER, this);
1754	>	w.parker = wt;
1755	>	if (ctl == currentCtl)
1756	>	U.park(false, parkTime);
1757	>	w.parker = null;
1758	>	U.putObject(wt, PARKBLOCKER, null);
1759	>	if (ctl != currentCtl)
1760	>	break;
1761	>	if (deadline - System.nanoTime() <= 0L &&
1762	>	U.compareAndSwapLong(this, CTL, currentCtl, prevCtl)) {
1763	>	w.eventCount = (w.eventCount + E_SEQ) | E_MASK;
1764	>	w.qlock = -1;   // shrink
1765	>	w.hint = -1;    // suppress helping
1766		break;
1767	<	if (untimed)
1768	<	LockSupport.park(w);
1769	<	else {
1770	<	LockSupport.parkNanos(w, SHRINK_RATE_NANOS);
1771	<	if (eventCount != ec || w.runState != 0 ||
1772	<	runState >= TERMINATING)
1767	>	}
1768	>	}
1769	>	}
1770	>	}
1771	>
1772	>	/**
1773	>	* Scans through queues looking for work (optionally, while
1774	>	* joining a task); if any are present, signals. May return early
1775	>	* if more signalling is detectably unneeded.
1776	>	*
1777	>	* @param task if non-null, return early if done
1778	>	* @param origin an index to start scan
1779	>	*/
1780	>	final int helpSignal(ForkJoinTask<?> task, int origin) {
1781	>	WorkQueue[] ws; WorkQueue q; int m, n, s, u;
1782	>	if ((ws = workQueues) != null && (m = ws.length - 1) >= 0) {
1783	>	for (int i = 0; i <= m; ++i) {
1784	>	if (task != null && (s = task.status) < 0)
1785	>	return s;
1786	>	if ((q = ws[(i + origin) & m]) != null &&
1787	>	(n = q.queueSize()) > 0) {
1788	>	signalWork(q, n);
1789	>	if ((u = (int)(ctl >>> 32)) >= 0 || (u >> UAC_SHIFT) >= 0)
1790		break;
799	–	if (System.nanoTime() - startTime >= SHRINK_RATE_NANOS)
800	–	tryShutdownUnusedWorker(ec);
1791		}
1792		}
1793		}
1794	+	return 0;
1795		}
1796
1797	<	// Maintaining parallelism
1797	>	/**
1798	>	* Tries to locate and execute tasks for a stealer of the given
1799	>	* task, or in turn one of its stealers, Traces currentSteal ->
1800	>	* currentJoin links looking for a thread working on a descendant
1801	>	* of the given task and with a non-empty queue to steal back and
1802	>	* execute tasks from. The first call to this method upon a
1803	>	* waiting join will often entail scanning/search, (which is OK
1804	>	* because the joiner has nothing better to do), but this method
1805	>	* leaves hints in workers to speed up subsequent calls. The
1806	>	* implementation is very branchy to cope with potential
1807	>	* inconsistencies or loops encountering chains that are stale,
1808	>	* unknown, or so long that they are likely cyclic.
1809	>	*
1810	>	* @param joiner the joining worker
1811	>	* @param task the task to join
1812	>	* @return 0 if no progress can be made, negative if task
1813	>	* known complete, else positive
1814	>	*/
1815	>	private int tryHelpStealer(WorkQueue joiner, ForkJoinTask<?> task) {
1816	>	int stat = 0, steps = 0;                    // bound to avoid cycles
1817	>	if (joiner != null && task != null) {       // hoist null checks
1818	>	restart: for (;;) {
1819	>	ForkJoinTask<?> subtask = task;     // current target
1820	>	for (WorkQueue j = joiner, v;;) {   // v is stealer of subtask
1821	>	WorkQueue[] ws; int m, s, h;
1822	>	if ((s = task.status) < 0) {
1823	>	stat = s;
1824	>	break restart;
1825	>	}
1826	>	if ((ws = workQueues) == null || (m = ws.length - 1) <= 0)
1827	>	break restart;              // shutting down
1828	>	if ((v = ws[h = (j.hint | 1) & m]) == null ||
1829	>	v.currentSteal != subtask) {
1830	>	for (int origin = h;;) {    // find stealer
1831	>	if (((h = (h + 2) & m) & 15) == 1 &&
1832	>	(subtask.status < 0 || j.currentJoin != subtask))
1833	>	continue restart;   // occasional staleness check
1834	>	if ((v = ws[h]) != null &&
1835	>	v.currentSteal == subtask) {
1836	>	j.hint = h;        // save hint
1837	>	break;
1838	>	}
1839	>	if (h == origin)
1840	>	break restart;      // cannot find stealer
1841	>	}
1842	>	}
1843	>	for (;;) { // help stealer or descend to its stealer
1844	>	ForkJoinTask[] a;  int b;
1845	>	if (subtask.status < 0)     // surround probes with
1846	>	continue restart;       //   consistency checks
1847	>	if ((b = v.base) - v.top < 0 && (a = v.array) != null) {
1848	>	int i = (((a.length - 1) & b) << ASHIFT) + ABASE;
1849	>	ForkJoinTask<?> t =
1850	>	(ForkJoinTask<?>)U.getObjectVolatile(a, i);
1851	>	if (subtask.status < 0 || j.currentJoin != subtask ||
1852	>	v.currentSteal != subtask)
1853	>	continue restart;   // stale
1854	>	stat = 1;               // apparent progress
1855	>	if (t != null && v.base == b &&
1856	>	U.compareAndSwapObject(a, i, t, null)) {
1857	>	v.base = b + 1;     // help stealer
1858	>	joiner.runSubtask(t);
1859	>	}
1860	>	else if (v.base == b && ++steps == MAX_HELP)
1861	>	break restart;      // v apparently stalled
1862	>	}
1863	>	else {                      // empty -- try to descend
1864	>	ForkJoinTask<?> next = v.currentJoin;
1865	>	if (subtask.status < 0 || j.currentJoin != subtask ||
1866	>	v.currentSteal != subtask)
1867	>	continue restart;   // stale
1868	>	else if (next == null || ++steps == MAX_HELP)
1869	>	break restart;      // dead-end or maybe cyclic
1870	>	else {
1871	>	subtask = next;
1872	>	j = v;
1873	>	break;
1874	>	}
1875	>	}
1876	>	}
1877	>	}
1878	>	}
1879	>	}
1880	>	return stat;
1881	>	}
1882
1883		/**
1884	<	* Pushes worker onto the spare stack
1884	>	* Analog of tryHelpStealer for CountedCompleters. Tries to steal
1885	>	* and run tasks within the target's computation.
1886	>	*
1887	>	* @param task the task to join
1888	>	* @param mode if shared, exit upon completing any task
1889	>	* if all workers are active
1890	>	*
1891		*/
1892	<	final void pushSpare(ForkJoinWorkerThread w) {
1893	<	int ns = (++w.spareCount << SPARE_COUNT_SHIFT) | (w.poolIndex + 1);
1894	<	do {} while (!UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
1895	<	w.nextSpare = spareWaiters,ns));
1892	>	private int helpComplete(ForkJoinTask<?> task, int mode) {
1893	>	WorkQueue[] ws; WorkQueue q; int m, n, s, u;
1894	>	if (task != null && (ws = workQueues) != null &&
1895	>	(m = ws.length - 1) >= 0) {
1896	>	for (int j = 1, origin = j;;) {
1897	>	if ((s = task.status) < 0)
1898	>	return s;
1899	>	if ((q = ws[j & m]) != null && q.pollAndExecCC(task)) {
1900	>	origin = j;
1901	>	if (mode == SHARED_QUEUE &&
1902	>	((u = (int)(ctl >>> 32)) >= 0 || (u >> UAC_SHIFT) >= 0))
1903	>	break;
1904	>	}
1905	>	else if ((j = (j + 2) & m) == origin)
1906	>	break;
1907	>	}
1908	>	}
1909	>	return 0;
1910		}
1911
1912		/**
1913	<	* Tries (once) to resume a spare if the number of running
1914	<	* threads is less than target.
1915	<	*/
1916	<	private void tryResumeSpare() {
1917	<	int sw, id;
1918	<	ForkJoinWorkerThread[] ws = workers;
1919	<	int n = ws.length;
1920	<	ForkJoinWorkerThread w;
1921	<	if ((sw = spareWaiters) != 0 &&
1922	<	(id = (sw & SPARE_ID_MASK) - 1) >= 0 &&
1923	<	id < n && (w = ws[id]) != null &&
1924	<	(workerCounts & RUNNING_COUNT_MASK) < parallelism &&
1925	<	spareWaiters == sw &&
1926	<	UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
1927	<	sw, w.nextSpare)) {
1928	<	int c; // increment running count before resume
1929	<	do {} while (!UNSAFE.compareAndSwapInt
1930	<	(this, workerCountsOffset,
1931	<	c = workerCounts, c + ONE_RUNNING));
1932	<	if (w.tryUnsuspend())
1933	<	LockSupport.unpark(w);
1934	<	else   // back out if w was shutdown
1935	<	decrementWorkerCounts(ONE_RUNNING, 0);
1913	>	* Tries to decrement active count (sometimes implicitly) and
1914	>	* possibly release or create a compensating worker in preparation
1915	>	* for blocking. Fails on contention or termination. Otherwise,
1916	>	* adds a new thread if no idle workers are available and pool
1917	>	* may become starved.
1918	>	*/
1919	>	final boolean tryCompensate() {
1920	>	int pc = config & SMASK, e, i, tc; long c;
1921	>	WorkQueue[] ws; WorkQueue w; Thread p;
1922	>	if ((ws = workQueues) != null && (e = (int)(c = ctl)) >= 0) {
1923	>	if (e != 0 && (i = e & SMASK) < ws.length &&
1924	>	(w = ws[i]) != null && w.eventCount == (e | INT_SIGN)) {
1925	>	long nc = ((long)(w.nextWait & E_MASK) |
1926	>	(c & (AC_MASK|TC_MASK)));
1927	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1928	>	w.eventCount = (e + E_SEQ) & E_MASK;
1929	>	if ((p = w.parker) != null)
1930	>	U.unpark(p);
1931	>	return true;   // replace with idle worker
1932	>	}
1933	>	}
1934	>	else if ((tc = (short)(c >>> TC_SHIFT)) >= 0 &&
1935	>	(int)(c >> AC_SHIFT) + pc > 1) {
1936	>	long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK);
1937	>	if (U.compareAndSwapLong(this, CTL, c, nc))
1938	>	return true;   // no compensation
1939	>	}
1940	>	else if (tc + pc < MAX_CAP) {
1941	>	long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK);
1942	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1943	>	addWorker();
1944	>	return true;
1945	>	}
1946	>	}
1947		}
1948	+	return false;
1949		}
1950
1951		/**
1952	<	* Tries to increase the number of running workers if below target
1953	<	* parallelism: If a spare exists tries to resume it via
1954	<	* tryResumeSpare.  Otherwise, if not enough total workers or all
1955	<	* existing workers are busy, adds a new worker. In all cases also
1956	<	* helps wake up releasable workers waiting for work.
1952	>	* Helps and/or blocks until the given task is done.
1953	>	*
1954	>	* @param joiner the joining worker
1955	>	* @param task the task
1956	>	* @return task status on exit
1957		*/
1958	<	private void helpMaintainParallelism() {
1959	<	int pc = parallelism;
1960	<	int wc, rs, tc;
1961	<	while (((wc = workerCounts) & RUNNING_COUNT_MASK) < pc &&
1962	<	(rs = runState) < TERMINATING) {
1963	<	if (spareWaiters != 0)
1964	<	tryResumeSpare();
1965	<	else if ((tc = wc >>> TOTAL_COUNT_SHIFT) >= MAX_WORKERS ||
1966	<	(tc >= pc && (rs & ACTIVE_COUNT_MASK) != tc))
1967	<	break;   // enough total
1968	<	else if (runState == rs && workerCounts == wc &&
1969	<	UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
1970	<	wc + (ONE_RUNNING|ONE_TOTAL))) {
1971	<	ForkJoinWorkerThread w = null;
1972	<	try {
1973	<	w = factory.newThread(this);
1974	<	} finally { // adjust on null or exceptional factory return
1975	<	if (w == null) {
1976	<	decrementWorkerCounts(ONE_RUNNING, ONE_TOTAL);
1977	<	tryTerminate(false); // handle failure during shutdown
1958	>	final int awaitJoin(WorkQueue joiner, ForkJoinTask<?> task) {
1959	>	int s = 0;
1960	>	if (joiner != null && task != null && (s = task.status) >= 0) {
1961	>	ForkJoinTask<?> prevJoin = joiner.currentJoin;
1962	>	joiner.currentJoin = task;
1963	>	do {} while ((s = task.status) >= 0 &&
1964	>	joiner.queueSize() > 0 &&
1965	>	joiner.tryRemoveAndExec(task)); // process local tasks
1966	>	if (s >= 0 && (s = task.status) >= 0 &&
1967	>	(s = helpSignal(task, joiner.poolIndex)) >= 0 &&
1968	>	(task instanceof CountedCompleter))
1969	>	s = helpComplete(task, LIFO_QUEUE);
1970	>	int k = 0; // to perform pre-block yield for politeness
1971	>	while (s >= 0 && (s = task.status) >= 0) {
1972	>	if ((joiner.queueSize() > 0 ||           // try helping
1973	>	(s = tryHelpStealer(joiner, task)) == 0) &&
1974	>	(s = task.status) >= 0) {
1975	>	if (k < 3) {
1976	>	if (++k < 3)
1977	>	s = helpSignal(task, joiner.poolIndex);
1978	>	else
1979	>	Thread.yield();
1980	>	}
1981	>	else if (!tryCompensate())
1982	>	k = 0;
1983	>	else {
1984	>	if (task.trySetSignal() && (s = task.status) >= 0) {
1985	>	synchronized (task) {
1986	>	if (task.status >= 0) {
1987	>	try {                // see ForkJoinTask
1988	>	task.wait();     //  for explanation
1989	>	} catch (InterruptedException ie) {
1990	>	}
1991	>	}
1992	>	else
1993	>	task.notifyAll();
1994	>	}
1995	>	}
1996	>	long c;                          // re-activate
1997	>	do {} while (!U.compareAndSwapLong
1998	>	(this, CTL, c = ctl, c + AC_UNIT));
1999		}
2000		}
873	–	if (w == null)
874	–	break;
875	–	w.start(recordWorker(w), ueh);
876	–	if ((workerCounts >>> TOTAL_COUNT_SHIFT) >= pc) {
877	–	int c; // advance event count
878	–	UNSAFE.compareAndSwapInt(this, eventCountOffset,
879	–	c = eventCount, c+1);
880	–	break; // add at most one unless total below target
881	–	}
882	–	}
883	–	}
884	–	if (eventWaiters != 0L)
885	–	releaseEventWaiters();
886	–	}
887	–
888	–	/**
889	–	* Callback from the oldest waiter in awaitEvent waking up after a
890	–	* period of non-use. If all workers are idle, tries (once) to
891	–	* shutdown an event waiter or a spare, if one exists. Note that
892	–	* we don't need CAS or locks here because the method is called
893	–	* only from one thread occasionally waking (and even misfires are
894	–	* OK). Note that until the shutdown worker fully terminates,
895	–	* workerCounts will overestimate total count, which is tolerable.
896	–	*
897	–	* @param ec the event count waited on by caller (to abort
898	–	* attempt if count has since changed).
899	–	*/
900	–	private void tryShutdownUnusedWorker(int ec) {
901	–	if (runState == 0 && eventCount == ec) { // only trigger if all idle
902	–	ForkJoinWorkerThread[] ws = workers;
903	–	int n = ws.length;
904	–	ForkJoinWorkerThread w = null;
905	–	boolean shutdown = false;
906	–	int sw;
907	–	long h;
908	–	if ((sw = spareWaiters) != 0) { // prefer killing spares
909	–	int id = (sw & SPARE_ID_MASK) - 1;
910	–	if (id >= 0 && id < n && (w = ws[id]) != null &&
911	–	UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
912	–	sw, w.nextSpare))
913	–	shutdown = true;
914	–	}
915	–	else if ((h = eventWaiters) != 0L) {
916	–	long nh;
917	–	int id = ((int)(h & WAITER_ID_MASK)) - 1;
918	–	if (id >= 0 && id < n && (w = ws[id]) != null &&
919	–	(nh = w.nextWaiter) != 0L && // keep at least one worker
920	–	UNSAFE.compareAndSwapLong(this, eventWaitersOffset, h, nh))
921	–	shutdown = true;
922	–	}
923	–	if (w != null && shutdown) {
924	–	w.shutdown();
925	–	LockSupport.unpark(w);
926	–	}
927	–	}
928	–	releaseEventWaiters(); // in case of interference
929	–	}
930	–
931	–	/**
932	–	* Callback from workers invoked upon each top-level action (i.e.,
933	–	* stealing a task or taking a submission and running it).
934	–	* Performs one or more of the following:
935	–	*
936	–	* 1. If the worker is active and either did not run a task
937	–	*    or there are too many workers, try to set its active status
938	–	*    to inactive and update activeCount. On contention, we may
939	–	*    try again in this or a subsequent call.
940	–	*
941	–	* 2. If not enough total workers, help create some.
942	–	*
943	–	* 3. If there are too many running workers, suspend this worker
944	–	*    (first forcing inactive if necessary).  If it is not needed,
945	–	*    it may be shutdown while suspended (via
946	–	*    tryShutdownUnusedWorker).  Otherwise, upon resume it
947	–	*    rechecks running thread count and need for event sync.
948	–	*
949	–	* 4. If worker did not run a task, await the next task event via
950	–	*    eventSync if necessary (first forcing inactivation), upon
951	–	*    which the worker may be shutdown via
952	–	*    tryShutdownUnusedWorker.  Otherwise, help release any
953	–	*    existing event waiters that are now releasable,
954	–	*
955	–	* @param w the worker
956	–	* @param ran true if worker ran a task since last call to this method
957	–	*/
958	–	final void preStep(ForkJoinWorkerThread w, boolean ran) {
959	–	int wec = w.lastEventCount;
960	–	boolean active = w.active;
961	–	boolean inactivate = false;
962	–	int pc = parallelism;
963	–	int rs;
964	–	while (w.runState == 0 && (rs = runState) < TERMINATING) {
965	–	if ((inactivate || (active && (rs & ACTIVE_COUNT_MASK) >= pc)) &&
966	–	UNSAFE.compareAndSwapInt(this, runStateOffset, rs, rs - 1))
967	–	inactivate = active = w.active = false;
968	–	int wc = workerCounts;
969	–	if ((wc & RUNNING_COUNT_MASK) > pc) {
970	–	if (!(inactivate |= active) && // must inactivate to suspend
971	–	workerCounts == wc &&      // try to suspend as spare
972	–	UNSAFE.compareAndSwapInt(this, workerCountsOffset,
973	–	wc, wc - ONE_RUNNING))
974	–	w.suspendAsSpare();
975	–	}
976	–	else if ((wc >>> TOTAL_COUNT_SHIFT) < pc)
977	–	helpMaintainParallelism();     // not enough workers
978	–	else if (!ran) {
979	–	long h = eventWaiters;
980	–	int ec = eventCount;
981	–	if (h != 0L && (int)(h >>> EVENT_COUNT_SHIFT) != ec)
982	–	releaseEventWaiters();     // release others before waiting
983	–	else if (ec != wec) {
984	–	w.lastEventCount = ec;     // no need to wait
985	–	break;
986	–	}
987	–	else if (!(inactivate |= active))
988	–	eventSync(w, wec);         // must inactivate before sync
2001		}
2002	<	else
991	<	break;
2002	>	joiner.currentJoin = prevJoin;
2003		}
2004	+	return s;
2005		}
2006
2007		/**
2008	<	* Helps and/or blocks awaiting join of the given task.
2009	<	* See above for explanation.
2008	>	* Stripped-down variant of awaitJoin used by timed joins. Tries
2009	>	* to help join only while there is continuous progress. (Caller
2010	>	* will then enter a timed wait.)
2011		*
2012	<	* @param joinMe the task to join
2013	<	* @param worker the current worker thread
2012	>	* @param joiner the joining worker
2013	>	* @param task the task
2014		*/
2015	<	final void awaitJoin(ForkJoinTask<?> joinMe, ForkJoinWorkerThread worker) {
2016	<	int retries = 2 + (parallelism >> 2); // #helpJoins before blocking
2017	<	while (joinMe.status >= 0) {
2018	<	int wc;
2019	<	worker.helpJoinTask(joinMe);
2020	<	if (joinMe.status < 0)
2021	<	break;
2022	<	else if (retries > 0)
2023	<	--retries;
2024	<	else if (((wc = workerCounts) & RUNNING_COUNT_MASK) != 0 &&
2025	<	UNSAFE.compareAndSwapInt(this, workerCountsOffset,
2026	<	wc, wc - ONE_RUNNING)) {
2027	<	int stat, c; long h;
2028	<	while ((stat = joinMe.status) >= 0 &&
2029	<	(h = eventWaiters) != 0L && // help release others
1017	<	(int)(h >>> EVENT_COUNT_SHIFT) != eventCount)
1018	<	releaseEventWaiters();
1019	<	if (stat >= 0 &&
1020	<	((workerCounts & RUNNING_COUNT_MASK) == 0 ||
1021	<	(stat =
1022	<	joinMe.internalAwaitDone(JOIN_TIMEOUT_MILLIS)) >= 0))
1023	<	helpMaintainParallelism(); // timeout or no running workers
1024	<	do {} while (!UNSAFE.compareAndSwapInt
1025	<	(this, workerCountsOffset,
1026	<	c = workerCounts, c + ONE_RUNNING));
1027	<	if (stat < 0)
1028	<	break;   // else restart
2015	>	final void helpJoinOnce(WorkQueue joiner, ForkJoinTask<?> task) {
2016	>	int s;
2017	>	if (joiner != null && task != null && (s = task.status) >= 0) {
2018	>	ForkJoinTask<?> prevJoin = joiner.currentJoin;
2019	>	joiner.currentJoin = task;
2020	>	do {} while ((s = task.status) >= 0 &&
2021	>	joiner.queueSize() > 0 &&
2022	>	joiner.tryRemoveAndExec(task));
2023	>	if (s >= 0 && (s = task.status) >= 0 &&
2024	>	(s = helpSignal(task, joiner.poolIndex)) >= 0 &&
2025	>	(task instanceof CountedCompleter))
2026	>	s = helpComplete(task, LIFO_QUEUE);
2027	>	if (s >= 0 && joiner.queueSize() == 0) {
2028	>	do {} while (task.status >= 0 &&
2029	>	tryHelpStealer(joiner, task) > 0);
2030		}
2031	+	joiner.currentJoin = prevJoin;
2032		}
2033		}
2034
2035		/**
2036	<	* Same idea as awaitJoin, but no helping, retries, or timeouts.
2037	<	*/
2038	<	final void awaitBlocker(ManagedBlocker blocker)
2039	<	throws InterruptedException {
2040	<	while (!blocker.isReleasable()) {
2041	<	int wc = workerCounts;
2042	<	if ((wc & RUNNING_COUNT_MASK) != 0 &&
2043	<	UNSAFE.compareAndSwapInt(this, workerCountsOffset,
2044	<	wc, wc - ONE_RUNNING)) {
2045	<	try {
2046	<	while (!blocker.isReleasable()) {
2047	<	long h = eventWaiters;
2048	<	if (h != 0L &&
2049	<	(int)(h >>> EVENT_COUNT_SHIFT) != eventCount)
2050	<	releaseEventWaiters();
2051	<	else if ((workerCounts & RUNNING_COUNT_MASK) == 0 &&
2052	<	runState < TERMINATING)
2053	<	helpMaintainParallelism();
2054	<	else if (blocker.block())
2055	<	break;
2056	<	}
2057	<	} finally {
2058	<	int c;
2059	<	do {} while (!UNSAFE.compareAndSwapInt
2060	<	(this, workerCountsOffset,
2061	<	c = workerCounts, c + ONE_RUNNING));
2036	>	* Returns a (probably) non-empty steal queue, if one is found
2037	>	* during a random, then cyclic scan, else null.  This method must
2038	>	* be retried by caller if, by the time it tries to use the queue,
2039	>	* it is empty.
2040	>	* @param r a (random) seed for scanning
2041	>	*/
2042	>	private WorkQueue findNonEmptyStealQueue(int r) {
2043	>	for (WorkQueue[] ws;;) {
2044	>	int ps = plock, m, n;
2045	>	if ((ws = workQueues) == null || (m = ws.length - 1) < 1)
2046	>	return null;
2047	>	for (int j = (m + 1) << 2; ;) {
2048	>	WorkQueue q = ws[(((r + j) << 1) | 1) & m];
2049	>	if (q != null && (n = q.queueSize()) > 0) {
2050	>	if (n > 1)
2051	>	signalWork(q, 0);
2052	>	return q;
2053	>	}
2054	>	else if (--j < 0) {
2055	>	if (plock == ps)
2056	>	return null;
2057	>	break;
2058	>	}
2059	>	}
2060	>	}
2061	>	}
2062	>
2063	>	/**
2064	>	* Runs tasks until {@code isQuiescent()}. We piggyback on
2065	>	* active count ctl maintenance, but rather than blocking
2066	>	* when tasks cannot be found, we rescan until all others cannot
2067	>	* find tasks either.
2068	>	*/
2069	>	final void helpQuiescePool(WorkQueue w) {
2070	>	for (boolean active = true;;) {
2071	>	ForkJoinTask<?> localTask; // exhaust local queue
2072	>	while ((localTask = w.nextLocalTask()) != null)
2073	>	localTask.doExec();
2074	>	// Similar to loop in scan(), but ignoring submissions
2075	>	WorkQueue q = findNonEmptyStealQueue(w.nextSeed());
2076	>	if (q != null) {
2077	>	ForkJoinTask<?> t; int b;
2078	>	if (!active) {      // re-establish active count
2079	>	long c;
2080	>	active = true;
2081	>	do {} while (!U.compareAndSwapLong
2082	>	(this, CTL, c = ctl, c + AC_UNIT));
2083	>	}
2084	>	if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null)
2085	>	w.runSubtask(t);
2086	>	}
2087	>	else {
2088	>	long c;
2089	>	if (active) {       // decrement active count without queuing
2090	>	active = false;
2091	>	do {} while (!U.compareAndSwapLong
2092	>	(this, CTL, c = ctl, c -= AC_UNIT));
2093	>	}
2094	>	else
2095	>	c = ctl;        // re-increment on exit
2096	>	if ((int)(c >> AC_SHIFT) + (config & SMASK) == 0) {
2097	>	do {} while (!U.compareAndSwapLong
2098	>	(this, CTL, c = ctl, c + AC_UNIT));
2099	>	break;
2100		}
1061	–	break;
2101		}
2102		}
2103		}
2104
2105		/**
2106	<	* Possibly initiates and/or completes termination.
2106	>	* Gets and removes a local or stolen task for the given worker.
2107	>	*
2108	>	* @return a task, if available
2109	>	*/
2110	>	final ForkJoinTask<?> nextTaskFor(WorkQueue w) {
2111	>	for (ForkJoinTask<?> t;;) {
2112	>	WorkQueue q; int b;
2113	>	if ((t = w.nextLocalTask()) != null)
2114	>	return t;
2115	>	if ((q = findNonEmptyStealQueue(w.nextSeed())) == null)
2116	>	return null;
2117	>	if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null)
2118	>	return t;
2119	>	}
2120	>	}
2121	>
2122	>	/**
2123	>	* Returns a cheap heuristic guide for task partitioning when
2124	>	* programmers, frameworks, tools, or languages have little or no
2125	>	* idea about task granularity.  In essence by offering this
2126	>	* method, we ask users only about tradeoffs in overhead vs
2127	>	* expected throughput and its variance, rather than how finely to
2128	>	* partition tasks.
2129	>	*
2130	>	* In a steady state strict (tree-structured) computation, each
2131	>	* thread makes available for stealing enough tasks for other
2132	>	* threads to remain active. Inductively, if all threads play by
2133	>	* the same rules, each thread should make available only a
2134	>	* constant number of tasks.
2135	>	*
2136	>	* The minimum useful constant is just 1. But using a value of 1
2137	>	* would require immediate replenishment upon each steal to
2138	>	* maintain enough tasks, which is infeasible.  Further,
2139	>	* partitionings/granularities of offered tasks should minimize
2140	>	* steal rates, which in general means that threads nearer the top
2141	>	* of computation tree should generate more than those nearer the
2142	>	* bottom. In perfect steady state, each thread is at
2143	>	* approximately the same level of computation tree. However,
2144	>	* producing extra tasks amortizes the uncertainty of progress and
2145	>	* diffusion assumptions.
2146	>	*
2147	>	* So, users will want to use values larger, but not much larger
2148	>	* than 1 to both smooth over transient shortages and hedge
2149	>	* against uneven progress; as traded off against the cost of
2150	>	* extra task overhead. We leave the user to pick a threshold
2151	>	* value to compare with the results of this call to guide
2152	>	* decisions, but recommend values such as 3.
2153	>	*
2154	>	* When all threads are active, it is on average OK to estimate
2155	>	* surplus strictly locally. In steady-state, if one thread is
2156	>	* maintaining say 2 surplus tasks, then so are others. So we can
2157	>	* just use estimated queue length.  However, this strategy alone
2158	>	* leads to serious mis-estimates in some non-steady-state
2159	>	* conditions (ramp-up, ramp-down, other stalls). We can detect
2160	>	* many of these by further considering the number of "idle"
2161	>	* threads, that are known to have zero queued tasks, so
2162	>	* compensate by a factor of (#idle/#active) threads.
2163	>	*
2164	>	* Note: The approximation of #busy workers as #active workers is
2165	>	* not very good under current signalling scheme, and should be
2166	>	* improved.
2167	>	*/
2168	>	static int getSurplusQueuedTaskCount() {
2169	>	Thread t; ForkJoinWorkerThread wt; ForkJoinPool pool; WorkQueue q;
2170	>	if (((t = Thread.currentThread()) instanceof ForkJoinWorkerThread)) {
2171	>	int p = (pool = (wt = (ForkJoinWorkerThread)t).pool).config & SMASK;
2172	>	int n = (q = wt.workQueue).top - q.base;
2173	>	int a = (int)(pool.ctl >> AC_SHIFT) + p;
2174	>	return n - (a > (p >>>= 1) ? 0 :
2175	>	a > (p >>>= 1) ? 1 :
2176	>	a > (p >>>= 1) ? 2 :
2177	>	a > (p >>>= 1) ? 4 :
2178	>	8);
2179	>	}
2180	>	return 0;
2181	>	}
2182	>
2183	>	//  Termination
2184	>
2185	>	/**
2186	>	* Possibly initiates and/or completes termination.  The caller
2187	>	* triggering termination runs three passes through workQueues:
2188	>	* (0) Setting termination status, followed by wakeups of queued
2189	>	* workers; (1) cancelling all tasks; (2) interrupting lagging
2190	>	* threads (likely in external tasks, but possibly also blocked in
2191	>	* joins).  Each pass repeats previous steps because of potential
2192	>	* lagging thread creation.
2193		*
2194		* @param now if true, unconditionally terminate, else only
2195	<	* if shutdown and empty queue and no active workers
2195	>	* if no work and no active workers
2196	>	* @param enable if true, enable shutdown when next possible
2197		* @return true if now terminating or terminated
2198		*/
2199	<	private boolean tryTerminate(boolean now) {
2200	<	if (now)
1075	<	advanceRunLevel(SHUTDOWN); // ensure at least SHUTDOWN
1076	<	else if (runState < SHUTDOWN ||
1077	<	!submissionQueue.isEmpty() ||
1078	<	(runState & ACTIVE_COUNT_MASK) != 0)
2199	>	private boolean tryTerminate(boolean now, boolean enable) {
2200	>	if (this == commonPool)                     // cannot shut down
2201		return false;
2202	<
2203	<	if (advanceRunLevel(TERMINATING))
2204	<	startTerminating();
2205	<
2206	<	// Finish now if all threads terminated; else in some subsequent call
2207	<	if ((workerCounts >>> TOTAL_COUNT_SHIFT) == 0) {
2208	<	advanceRunLevel(TERMINATED);
2209	<	termination.arrive();
2210	<	}
2211	<	return true;
2212	<	}
2213	<
2214	<	/**
2215	<	* Actions on transition to TERMINATING
2216	<	*
2217	<	* Runs up to four passes through workers: (0) shutting down each
2218	<	* (without waking up if parked) to quickly spread notifications
2219	<	* without unnecessary bouncing around event queues etc (1) wake
2220	<	* up and help cancel tasks (2) interrupt (3) mop up races with
2221	<	* interrupted workers
2222	<	*/
2223	<	private void startTerminating() {
2224	<	cancelSubmissions();
2225	<	for (int passes = 0; passes < 4 && workerCounts != 0; ++passes) {
2226	<	int c; // advance event count
2227	<	UNSAFE.compareAndSwapInt(this, eventCountOffset,
2228	<	c = eventCount, c+1);
2229	<	eventWaiters = 0L; // clobber lists
2230	<	spareWaiters = 0;
2231	<	for (ForkJoinWorkerThread w : workers) {
2232	<	if (w != null) {
2233	<	w.shutdown();
2234	<	if (passes > 0 && !w.isTerminated()) {
2235	<	w.cancelTasks();
2236	<	LockSupport.unpark(w);
2237	<	if (passes > 1) {
2238	<	try {
2239	<	w.interrupt();
2240	<	} catch (SecurityException ignore) {
2202	>	for (long c;;) {
2203	>	if (((c = ctl) & STOP_BIT) != 0) {      // already terminating
2204	>	if ((short)(c >>> TC_SHIFT) == -(config & SMASK)) {
2205	>	synchronized (this) {
2206	>	notifyAll();                // signal when 0 workers
2207	>	}
2208	>	}
2209	>	return true;
2210	>	}
2211	>	if (plock >= 0) {                       // not yet enabled
2212	>	int ps;
2213	>	if (!enable)
2214	>	return false;
2215	>	if (((ps = plock) & PL_LOCK) != 0 ||
2216	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
2217	>	ps = acquirePlock();
2218	>	int nps = SHUTDOWN;
2219	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
2220	>	releasePlock(nps);
2221	>	}
2222	>	if (!now) {                             // check if idle & no tasks
2223	>	if ((int)(c >> AC_SHIFT) != -(config & SMASK) ||
2224	>	hasQueuedSubmissions())
2225	>	return false;
2226	>	// Check for unqueued inactive workers. One pass suffices.
2227	>	WorkQueue[] ws = workQueues; WorkQueue w;
2228	>	if (ws != null) {
2229	>	for (int i = 1; i < ws.length; i += 2) {
2230	>	if ((w = ws[i]) != null && w.eventCount >= 0)
2231	>	return false;
2232	>	}
2233	>	}
2234	>	}
2235	>	if (U.compareAndSwapLong(this, CTL, c, c | STOP_BIT)) {
2236	>	for (int pass = 0; pass < 3; ++pass) {
2237	>	WorkQueue[] ws = workQueues;
2238	>	if (ws != null) {
2239	>	WorkQueue w;
2240	>	int n = ws.length;
2241	>	for (int i = 0; i < n; ++i) {
2242	>	if ((w = ws[i]) != null) {
2243	>	w.qlock = -1;
2244	>	if (pass > 0) {
2245	>	w.cancelAll();
2246	>	if (pass > 1)
2247	>	w.interruptOwner();
2248	>	}
2249	>	}
2250	>	}
2251	>	// Wake up workers parked on event queue
2252	>	int i, e; long cc; Thread p;
2253	>	while ((e = (int)(cc = ctl) & E_MASK) != 0 &&
2254	>	(i = e & SMASK) < n &&
2255	>	(w = ws[i]) != null) {
2256	>	long nc = ((long)(w.nextWait & E_MASK) |
2257	>	((cc + AC_UNIT) & AC_MASK) |
2258	>	(cc & (TC_MASK|STOP_BIT)));
2259	>	if (w.eventCount == (e | INT_SIGN) &&
2260	>	U.compareAndSwapLong(this, CTL, cc, nc)) {
2261	>	w.eventCount = (e + E_SEQ) & E_MASK;
2262	>	w.qlock = -1;
2263	>	if ((p = w.parker) != null)
2264	>	U.unpark(p);
2265		}
2266		}
2267		}
#	Line 1124 | Line 2270 | public class ForkJoinPool extends Abstra
2270		}
2271		}
2272
2273	+	// external operations on common pool
2274	+
2275		/**
2276	<	* Clears out and cancels submissions, ignoring exceptions.
2276	>	* Returns common pool queue for a thread that has submitted at
2277	>	* least one task.
2278		*/
2279	<	private void cancelSubmissions() {
2280	<	ForkJoinTask<?> task;
2281	<	while ((task = submissionQueue.poll()) != null) {
2282	<	try {
2283	<	task.cancel(false);
2284	<	} catch (Throwable ignore) {
2279	>	static WorkQueue commonSubmitterQueue() {
2280	>	ForkJoinPool p; WorkQueue[] ws; int m; Submitter z;
2281	>	return ((z = submitters.get()) != null &&
2282	>	(p = commonPool) != null &&
2283	>	(ws = p.workQueues) != null &&
2284	>	(m = ws.length - 1) >= 0) ?
2285	>	ws[m & z.seed & SQMASK] : null;
2286	>	}
2287	>
2288	>	/**
2289	>	* Tries to pop the given task from submitter's queue in common pool.
2290	>	*/
2291	>	static boolean tryExternalUnpush(ForkJoinTask<?> t) {
2292	>	ForkJoinPool p; WorkQueue[] ws; WorkQueue q; Submitter z;
2293	>	ForkJoinTask<?>[] a;  int m, s; long j;
2294	>	if ((z = submitters.get()) != null &&
2295	>	(p = commonPool) != null &&
2296	>	(ws = p.workQueues) != null &&
2297	>	(m = ws.length - 1) >= 0 &&
2298	>	(q = ws[m & z.seed & SQMASK]) != null &&
2299	>	(s = q.top) != q.base &&
2300	>	(a = q.array) != null &&
2301	>	U.getObjectVolatile
2302	>	(a, j = (((a.length - 1) & (s - 1)) << ASHIFT) + ABASE) == t &&
2303	>	U.compareAndSwapInt(q, QLOCK, 0, 1)) {
2304	>	if (q.array == a && q.top == s && // recheck
2305	>	U.compareAndSwapObject(a, j, t, null)) {
2306	>	q.top = s - 1;
2307	>	q.qlock = 0;
2308	>	return true;
2309		}
2310	+	q.qlock = 0;
2311		}
2312	+	return false;
2313		}
2314
1140	–	// misc support for ForkJoinWorkerThread
1141	–
2315		/**
2316	<	* Returns pool number.
2317	<	*/
2318	<	final int getPoolNumber() {
2319	<	return poolNumber;
2316	>	* Tries to pop and run local tasks within the same computation
2317	>	* as the given root. On failure, tries to help complete from
2318	>	* other queues via helpComplete.
2319	>	*/
2320	>	private void externalHelpComplete(WorkQueue q, ForkJoinTask<?> root) {
2321	>	ForkJoinTask<?>[] a; int m;
2322	>	if (q != null && (a = q.array) != null && (m = (a.length - 1)) >= 0 &&
2323	>	root != null && root.status >= 0) {
2324	>	for (;;) {
2325	>	int s, u; Object o; CountedCompleter<?> task = null;
2326	>	if ((s = q.top) - q.base > 0) {
2327	>	long j = ((m & (s - 1)) << ASHIFT) + ABASE;
2328	>	if ((o = U.getObject(a, j)) != null &&
2329	>	(o instanceof CountedCompleter)) {
2330	>	CountedCompleter<?> t = (CountedCompleter<?>)o, r = t;
2331	>	do {
2332	>	if (r == root) {
2333	>	if (U.compareAndSwapInt(q, QLOCK, 0, 1)) {
2334	>	if (q.array == a && q.top == s &&
2335	>	U.compareAndSwapObject(a, j, t, null)) {
2336	>	q.top = s - 1;
2337	>	task = t;
2338	>	}
2339	>	q.qlock = 0;
2340	>	}
2341	>	break;
2342	>	}
2343	>	} while ((r = r.completer) != null);
2344	>	}
2345	>	}
2346	>	if (task != null)
2347	>	task.doExec();
2348	>	if (root.status < 0 ||
2349	>	(u = (int)(ctl >>> 32)) >= 0 || (u >> UAC_SHIFT) >= 0)
2350	>	break;
2351	>	if (task == null) {
2352	>	if (helpSignal(root, q.poolIndex) >= 0)
2353	>	helpComplete(root, SHARED_QUEUE);
2354	>	break;
2355	>	}
2356	>	}
2357	>	}
2358		}
2359
2360		/**
2361	<	* Tries to accumulate steal count from a worker, clearing
2362	<	* the worker's value if successful.
1152	<	*
1153	<	* @return true if worker steal count now zero
2361	>	* Tries to help execute or signal availability of the given task
2362	>	* from submitter's queue in common pool.
2363		*/
2364	<	final boolean tryAccumulateStealCount(ForkJoinWorkerThread w) {
2365	<	int sc = w.stealCount;
2366	<	long c = stealCount;
2367	<	// CAS even if zero, for fence effects
2368	<	if (UNSAFE.compareAndSwapLong(this, stealCountOffset, c, c + sc)) {
2369	<	if (sc != 0)
2370	<	w.stealCount = 0;
2371	<	return true;
2364	>	static void externalHelpJoin(ForkJoinTask<?> t) {
2365	>	// Some hard-to-avoid overlap with tryExternalUnpush
2366	>	ForkJoinPool p; WorkQueue[] ws; WorkQueue q, w; Submitter z;
2367	>	ForkJoinTask<?>[] a;  int m, s, n; long j;
2368	>	if (t != null &&
2369	>	(z = submitters.get()) != null &&
2370	>	(p = commonPool) != null &&
2371	>	(ws = p.workQueues) != null &&
2372	>	(m = ws.length - 1) >= 0 &&
2373	>	(q = ws[m & z.seed & SQMASK]) != null &&
2374	>	(a = q.array) != null &&
2375	>	t.status >= 0) {
2376	>	if ((s = q.top) != q.base &&
2377	>	U.getObjectVolatile
2378	>	(a, j = (((a.length - 1) & (s - 1)) << ASHIFT) + ABASE) == t &&
2379	>	U.compareAndSwapInt(q, QLOCK, 0, 1)) {
2380	>	if (q.array == a && q.top == s &&
2381	>	U.compareAndSwapObject(a, j, t, null)) {
2382	>	q.top = s - 1;
2383	>	q.qlock = 0;
2384	>	t.doExec();
2385	>	}
2386	>	else
2387	>	q.qlock = 0;
2388	>	}
2389	>	if (t.status >= 0) {
2390	>	if (t instanceof CountedCompleter)
2391	>	p.externalHelpComplete(q, t);
2392	>	else
2393	>	p.helpSignal(t, q.poolIndex);
2394	>	}
2395		}
1164	–	return sc == 0;
2396		}
2397
2398		/**
2399	<	* Returns the approximate (non-atomic) number of idle threads per
1169	<	* active thread.
2399	>	* Restricted version of helpQuiescePool for external callers
2400		*/
2401	<	final int idlePerActive() {
2402	<	int pc = parallelism; // use parallelism, not rc
2403	<	int ac = runState;    // no mask -- artificially boosts during shutdown
2404	<	// Use exact results for small values, saturate past 4
2405	<	return ((pc <= ac) ? 0 :
2406	<	(pc >>> 1 <= ac) ? 1 :
2407	<	(pc >>> 2 <= ac) ? 3 :
1178	<	pc >>> 3);
2401	>	static void externalHelpQuiescePool() {
2402	>	ForkJoinPool p; ForkJoinTask<?> t; WorkQueue q; int b;
2403	>	if ((p = commonPool) != null &&
2404	>	(q = p.findNonEmptyStealQueue(1)) != null &&
2405	>	(b = q.base) - q.top < 0 &&
2406	>	(t = q.pollAt(b)) != null)
2407	>	t.doExec();
2408		}
2409
2410	<	// Public and protected methods
2410	>	// Exported methods
2411
2412		// Constructors
2413
#	Line 1225 | Line 2454 | public class ForkJoinPool extends Abstra
2454		* use {@link #defaultForkJoinWorkerThreadFactory}.
2455		* @param handler the handler for internal worker threads that
2456		* terminate due to unrecoverable errors encountered while executing
2457	<	* tasks. For default value, use <code>null</code>.
2457	>	* tasks. For default value, use {@code null}.
2458		* @param asyncMode if true,
2459		* establishes local first-in-first-out scheduling mode for forked
2460		* tasks that are never joined. This mode may be more appropriate
2461		* than default locally stack-based mode in applications in which
2462		* worker threads only process event-style asynchronous tasks.
2463	<	* For default value, use <code>false</code>.
2463	>	* For default value, use {@code false}.
2464		* @throws IllegalArgumentException if parallelism less than or
2465		*         equal to zero, or greater than implementation limit
2466		* @throws NullPointerException if the factory is null
#	Line 1247 | Line 2476 | public class ForkJoinPool extends Abstra
2476		checkPermission();
2477		if (factory == null)
2478		throw new NullPointerException();
2479	<	if (parallelism <= 0 || parallelism > MAX_WORKERS)
2479	>	if (parallelism <= 0 || parallelism > MAX_CAP)
2480		throw new IllegalArgumentException();
1252	–	this.parallelism = parallelism;
2481		this.factory = factory;
2482		this.ueh = handler;
2483	<	this.locallyFifo = asyncMode;
2484	<	int arraySize = initialArraySizeFor(parallelism);
2485	<	this.workers = new ForkJoinWorkerThread[arraySize];
2486	<	this.submissionQueue = new LinkedTransferQueue<ForkJoinTask<?>>();
2487	<	this.workerLock = new ReentrantLock();
2488	<	this.termination = new Phaser(1);
2489	<	this.poolNumber = poolNumberGenerator.incrementAndGet();
2483	>	this.config = parallelism | (asyncMode ? (FIFO_QUEUE << 16) : 0);
2484	>	long np = (long)(-parallelism); // offset ctl counts
2485	>	this.ctl = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK);
2486	>	int pn = nextPoolId();
2487	>	StringBuilder sb = new StringBuilder("ForkJoinPool-");
2488	>	sb.append(Integer.toString(pn));
2489	>	sb.append("-worker-");
2490	>	this.workerNamePrefix = sb.toString();
2491		}
2492
2493		/**
2494	<	* Returns initial power of two size for workers array.
2495	<	* @param pc the initial parallelism level
2496	<	*/
2497	<	private static int initialArraySizeFor(int pc) {
2498	<	// If possible, initially allocate enough space for one spare
2499	<	int size = pc < MAX_WORKERS ? pc + 1 : MAX_WORKERS;
2500	<	// See Hackers Delight, sec 3.2. We know MAX_WORKERS < (1 >>> 16)
2501	<	size |= size >>> 1;
2502	<	size |= size >>> 2;
2503	<	size |= size >>> 4;
2504	<	size |= size >>> 8;
1276	<	return size + 1;
2494	>	* Constructor for common pool, suitable only for static initialization.
2495	>	* Basically the same as above, but uses smallest possible initial footprint.
2496	>	*/
2497	>	ForkJoinPool(int parallelism, long ctl,
2498	>	ForkJoinWorkerThreadFactory factory,
2499	>	Thread.UncaughtExceptionHandler handler) {
2500	>	this.config = parallelism;
2501	>	this.ctl = ctl;
2502	>	this.factory = factory;
2503	>	this.ueh = handler;
2504	>	this.workerNamePrefix = "ForkJoinPool.commonPool-worker-";
2505		}
2506
1279	–	// Execution methods
1280	–
2507		/**
2508	<	* Common code for execute, invoke and submit
2508	>	* Returns the common pool instance.
2509	>	*
2510	>	* @return the common pool instance
2511		*/
2512	<	private <T> void doSubmit(ForkJoinTask<T> task) {
2513	<	if (task == null)
2514	<	throw new NullPointerException();
1287	<	if (runState >= SHUTDOWN)
1288	<	throw new RejectedExecutionException();
1289	<	submissionQueue.offer(task);
1290	<	int c; // try to increment event count -- CAS failure OK
1291	<	UNSAFE.compareAndSwapInt(this, eventCountOffset, c = eventCount, c+1);
1292	<	helpMaintainParallelism(); // create, start, or resume some workers
2512	>	public static ForkJoinPool commonPool() {
2513	>	// assert commonPool != null : "static init error";
2514	>	return commonPool;
2515		}
2516
2517	+	// Execution methods
2518	+
2519		/**
2520		* Performs the given task, returning its result upon completion.
2521	+	* If the computation encounters an unchecked Exception or Error,
2522	+	* it is rethrown as the outcome of this invocation.  Rethrown
2523	+	* exceptions behave in the same way as regular exceptions, but,
2524	+	* when possible, contain stack traces (as displayed for example
2525	+	* using {@code ex.printStackTrace()}) of both the current thread
2526	+	* as well as the thread actually encountering the exception;
2527	+	* minimally only the latter.
2528		*
2529		* @param task the task
2530		* @return the task's result
#	Line 1302 | Line 2533 | public class ForkJoinPool extends Abstra
2533		*         scheduled for execution
2534		*/
2535		public <T> T invoke(ForkJoinTask<T> task) {
2536	<	doSubmit(task);
2536	>	if (task == null)
2537	>	throw new NullPointerException();
2538	>	externalPush(task);
2539		return task.join();
2540		}
2541
#	Line 1315 | Line 2548 | public class ForkJoinPool extends Abstra
2548		*         scheduled for execution
2549		*/
2550		public void execute(ForkJoinTask<?> task) {
2551	<	doSubmit(task);
2551	>	if (task == null)
2552	>	throw new NullPointerException();
2553	>	externalPush(task);
2554		}
2555
2556		// AbstractExecutorService methods
#	Line 1326 | Line 2561 | public class ForkJoinPool extends Abstra
2561		*         scheduled for execution
2562		*/
2563		public void execute(Runnable task) {
2564	+	if (task == null)
2565	+	throw new NullPointerException();
2566		ForkJoinTask<?> job;
2567		if (task instanceof ForkJoinTask<?>) // avoid re-wrap
2568		job = (ForkJoinTask<?>) task;
2569		else
2570	<	job = ForkJoinTask.adapt(task, null);
2571	<	doSubmit(job);
2570	>	job = new ForkJoinTask.AdaptedRunnableAction(task);
2571	>	externalPush(job);
2572		}
2573
2574		/**
#	Line 1344 | Line 2581 | public class ForkJoinPool extends Abstra
2581		*         scheduled for execution
2582		*/
2583		public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) {
2584	<	doSubmit(task);
2584	>	if (task == null)
2585	>	throw new NullPointerException();
2586	>	externalPush(task);
2587		return task;
2588		}
2589
#	Line 1354 | Line 2593 | public class ForkJoinPool extends Abstra
2593		*         scheduled for execution
2594		*/
2595		public <T> ForkJoinTask<T> submit(Callable<T> task) {
2596	<	ForkJoinTask<T> job = ForkJoinTask.adapt(task);
2597	<	doSubmit(job);
2596	>	ForkJoinTask<T> job = new ForkJoinTask.AdaptedCallable<T>(task);
2597	>	externalPush(job);
2598		return job;
2599		}
2600
#	Line 1365 | Line 2604 | public class ForkJoinPool extends Abstra
2604		*         scheduled for execution
2605		*/
2606		public <T> ForkJoinTask<T> submit(Runnable task, T result) {
2607	<	ForkJoinTask<T> job = ForkJoinTask.adapt(task, result);
2608	<	doSubmit(job);
2607	>	ForkJoinTask<T> job = new ForkJoinTask.AdaptedRunnable<T>(task, result);
2608	>	externalPush(job);
2609		return job;
2610		}
2611
#	Line 1376 | Line 2615 | public class ForkJoinPool extends Abstra
2615		*         scheduled for execution
2616		*/
2617		public ForkJoinTask<?> submit(Runnable task) {
2618	+	if (task == null)
2619	+	throw new NullPointerException();
2620		ForkJoinTask<?> job;
2621		if (task instanceof ForkJoinTask<?>) // avoid re-wrap
2622		job = (ForkJoinTask<?>) task;
2623		else
2624	<	job = ForkJoinTask.adapt(task, null);
2625	<	doSubmit(job);
2624	>	job = new ForkJoinTask.AdaptedRunnableAction(task);
2625	>	externalPush(job);
2626		return job;
2627		}
2628
#	Line 1390 | Line 2631 | public class ForkJoinPool extends Abstra
2631		* @throws RejectedExecutionException {@inheritDoc}
2632		*/
2633		public <T> List<Future<T>> invokeAll(Collection<? extends Callable<T>> tasks) {
2634	<	ArrayList<ForkJoinTask<T>> forkJoinTasks =
2635	<	new ArrayList<ForkJoinTask<T>>(tasks.size());
2636	<	for (Callable<T> task : tasks)
2637	<	forkJoinTasks.add(ForkJoinTask.adapt(task));
2638	<	invoke(new InvokeAll<T>(forkJoinTasks));
2639	<
2634	>	// In previous versions of this class, this method constructed
2635	>	// a task to run ForkJoinTask.invokeAll, but now external
2636	>	// invocation of multiple tasks is at least as efficient.
2637	>	List<ForkJoinTask<T>> fs = new ArrayList<ForkJoinTask<T>>(tasks.size());
2638	>	// Workaround needed because method wasn't declared with
2639	>	// wildcards in return type but should have been.
2640		@SuppressWarnings({"unchecked", "rawtypes"})
2641	<	List<Future<T>> futures = (List<Future<T>>) (List) forkJoinTasks;
1401	<	return futures;
1402	<	}
2641	>	List<Future<T>> futures = (List<Future<T>>) (List) fs;
2642
2643	<	static final class InvokeAll<T> extends RecursiveAction {
2644	<	final ArrayList<ForkJoinTask<T>> tasks;
2645	<	InvokeAll(ArrayList<ForkJoinTask<T>> tasks) { this.tasks = tasks; }
2646	<	public void compute() {
2647	<	try { invokeAll(tasks); }
2648	<	catch (Exception ignore) {}
2643	>	boolean done = false;
2644	>	try {
2645	>	for (Callable<T> t : tasks) {
2646	>	ForkJoinTask<T> f = new ForkJoinTask.AdaptedCallable<T>(t);
2647	>	externalPush(f);
2648	>	fs.add(f);
2649	>	}
2650	>	for (ForkJoinTask<T> f : fs)
2651	>	f.quietlyJoin();
2652	>	done = true;
2653	>	return futures;
2654	>	} finally {
2655	>	if (!done)
2656	>	for (ForkJoinTask<T> f : fs)
2657	>	f.cancel(false);
2658		}
1411	–	private static final long serialVersionUID = -7914297376763021607L;
2659		}
2660
2661		/**
#	Line 1436 | Line 2683 | public class ForkJoinPool extends Abstra
2683		* @return the targeted parallelism level of this pool
2684		*/
2685		public int getParallelism() {
2686	<	return parallelism;
2686	>	return config & SMASK;
2687	>	}
2688	>
2689	>	/**
2690	>	* Returns the targeted parallelism level of the common pool.
2691	>	*
2692	>	* @return the targeted parallelism level of the common pool
2693	>	*/
2694	>	public static int getCommonPoolParallelism() {
2695	>	return commonPoolParallelism;
2696		}
2697
2698		/**
2699		* Returns the number of worker threads that have started but not
2700	<	* yet terminated.  This result returned by this method may differ
2700	>	* yet terminated.  The result returned by this method may differ
2701		* from {@link #getParallelism} when threads are created to
2702		* maintain parallelism when others are cooperatively blocked.
2703		*
2704		* @return the number of worker threads
2705		*/
2706		public int getPoolSize() {
2707	<	return workerCounts >>> TOTAL_COUNT_SHIFT;
2707	>	return (config & SMASK) + (short)(ctl >>> TC_SHIFT);
2708		}
2709
2710		/**
#	Line 1458 | Line 2714 | public class ForkJoinPool extends Abstra
2714		* @return {@code true} if this pool uses async mode
2715		*/
2716		public boolean getAsyncMode() {
2717	<	return locallyFifo;
2717	>	return (config >>> 16) == FIFO_QUEUE;
2718		}
2719
2720		/**
#	Line 1470 | Line 2726 | public class ForkJoinPool extends Abstra
2726		* @return the number of worker threads
2727		*/
2728		public int getRunningThreadCount() {
2729	<	return workerCounts & RUNNING_COUNT_MASK;
2729	>	int rc = 0;
2730	>	WorkQueue[] ws; WorkQueue w;
2731	>	if ((ws = workQueues) != null) {
2732	>	for (int i = 1; i < ws.length; i += 2) {
2733	>	if ((w = ws[i]) != null && w.isApparentlyUnblocked())
2734	>	++rc;
2735	>	}
2736	>	}
2737	>	return rc;
2738		}
2739
2740		/**
#	Line 1481 | Line 2745 | public class ForkJoinPool extends Abstra
2745		* @return the number of active threads
2746		*/
2747		public int getActiveThreadCount() {
2748	<	return runState & ACTIVE_COUNT_MASK;
2748	>	int r = (config & SMASK) + (int)(ctl >> AC_SHIFT);
2749	>	return (r <= 0) ? 0 : r; // suppress momentarily negative values
2750		}
2751
2752		/**
#	Line 1496 | Line 2761 | public class ForkJoinPool extends Abstra
2761		* @return {@code true} if all threads are currently idle
2762		*/
2763		public boolean isQuiescent() {
2764	<	return (runState & ACTIVE_COUNT_MASK) == 0;
2764	>	return (int)(ctl >> AC_SHIFT) + (config & SMASK) == 0;
2765		}
2766
2767		/**
#	Line 1511 | Line 2776 | public class ForkJoinPool extends Abstra
2776		* @return the number of steals
2777		*/
2778		public long getStealCount() {
2779	<	return stealCount;
2779	>	long count = stealCount;
2780	>	WorkQueue[] ws; WorkQueue w;
2781	>	if ((ws = workQueues) != null) {
2782	>	for (int i = 1; i < ws.length; i += 2) {
2783	>	if ((w = ws[i]) != null)
2784	>	count += w.nsteals;
2785	>	}
2786	>	}
2787	>	return count;
2788		}
2789
2790		/**
#	Line 1526 | Line 2799 | public class ForkJoinPool extends Abstra
2799		*/
2800		public long getQueuedTaskCount() {
2801		long count = 0;
2802	<	for (ForkJoinWorkerThread w : workers)
2803	<	if (w != null)
2804	<	count += w.getQueueSize();
2802	>	WorkQueue[] ws; WorkQueue w;
2803	>	if ((ws = workQueues) != null) {
2804	>	for (int i = 1; i < ws.length; i += 2) {
2805	>	if ((w = ws[i]) != null)
2806	>	count += w.queueSize();
2807	>	}
2808	>	}
2809		return count;
2810		}
2811
2812		/**
2813		* Returns an estimate of the number of tasks submitted to this
2814	<	* pool that have not yet begun executing.  This method takes time
2815	<	* proportional to the number of submissions.
2814	>	* pool that have not yet begun executing.  This method may take
2815	>	* time proportional to the number of submissions.
2816		*
2817		* @return the number of queued submissions
2818		*/
2819		public int getQueuedSubmissionCount() {
2820	<	return submissionQueue.size();
2820	>	int count = 0;
2821	>	WorkQueue[] ws; WorkQueue w;
2822	>	if ((ws = workQueues) != null) {
2823	>	for (int i = 0; i < ws.length; i += 2) {
2824	>	if ((w = ws[i]) != null)
2825	>	count += w.queueSize();
2826	>	}
2827	>	}
2828	>	return count;
2829		}
2830
2831		/**
#	Line 1550 | Line 2835 | public class ForkJoinPool extends Abstra
2835		* @return {@code true} if there are any queued submissions
2836		*/
2837		public boolean hasQueuedSubmissions() {
2838	<	return !submissionQueue.isEmpty();
2838	>	WorkQueue[] ws; WorkQueue w;
2839	>	if ((ws = workQueues) != null) {
2840	>	for (int i = 0; i < ws.length; i += 2) {
2841	>	if ((w = ws[i]) != null && w.queueSize() != 0)
2842	>	return true;
2843	>	}
2844	>	}
2845	>	return false;
2846		}
2847
2848		/**
#	Line 1561 | Line 2853 | public class ForkJoinPool extends Abstra
2853		* @return the next submission, or {@code null} if none
2854		*/
2855		protected ForkJoinTask<?> pollSubmission() {
2856	<	return submissionQueue.poll();
2856	>	WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2857	>	if ((ws = workQueues) != null) {
2858	>	for (int i = 0; i < ws.length; i += 2) {
2859	>	if ((w = ws[i]) != null && (t = w.poll()) != null)
2860	>	return t;
2861	>	}
2862	>	}
2863	>	return null;
2864		}
2865
2866		/**
#	Line 1582 | Line 2881 | public class ForkJoinPool extends Abstra
2881		* @return the number of elements transferred
2882		*/
2883		protected int drainTasksTo(Collection<? super ForkJoinTask<?>> c) {
2884	<	int count = submissionQueue.drainTo(c);
2885	<	for (ForkJoinWorkerThread w : workers)
2886	<	if (w != null)
2887	<	count += w.drainTasksTo(c);
2884	>	int count = 0;
2885	>	WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2886	>	if ((ws = workQueues) != null) {
2887	>	for (int i = 0; i < ws.length; ++i) {
2888	>	if ((w = ws[i]) != null) {
2889	>	while ((t = w.poll()) != null) {
2890	>	c.add(t);
2891	>	++count;
2892	>	}
2893	>	}
2894	>	}
2895	>	}
2896		return count;
2897		}
2898
#	Line 1597 | Line 2904 | public class ForkJoinPool extends Abstra
2904		* @return a string identifying this pool, as well as its state
2905		*/
2906		public String toString() {
2907	<	long st = getStealCount();
2908	<	long qt = getQueuedTaskCount();
2909	<	long qs = getQueuedSubmissionCount();
2910	<	int wc = workerCounts;
2911	<	int tc = wc >>> TOTAL_COUNT_SHIFT;
2912	<	int rc = wc & RUNNING_COUNT_MASK;
2913	<	int pc = parallelism;
2914	<	int rs = runState;
2915	<	int ac = rs & ACTIVE_COUNT_MASK;
2907	>	// Use a single pass through workQueues to collect counts
2908	>	long qt = 0L, qs = 0L; int rc = 0;
2909	>	long st = stealCount;
2910	>	long c = ctl;
2911	>	WorkQueue[] ws; WorkQueue w;
2912	>	if ((ws = workQueues) != null) {
2913	>	for (int i = 0; i < ws.length; ++i) {
2914	>	if ((w = ws[i]) != null) {
2915	>	int size = w.queueSize();
2916	>	if ((i & 1) == 0)
2917	>	qs += size;
2918	>	else {
2919	>	qt += size;
2920	>	st += w.nsteals;
2921	>	if (w.isApparentlyUnblocked())
2922	>	++rc;
2923	>	}
2924	>	}
2925	>	}
2926	>	}
2927	>	int pc = (config & SMASK);
2928	>	int tc = pc + (short)(c >>> TC_SHIFT);
2929	>	int ac = pc + (int)(c >> AC_SHIFT);
2930	>	if (ac < 0) // ignore transient negative
2931	>	ac = 0;
2932	>	String level;
2933	>	if ((c & STOP_BIT) != 0)
2934	>	level = (tc == 0) ? "Terminated" : "Terminating";
2935	>	else
2936	>	level = plock < 0 ? "Shutting down" : "Running";
2937		return super.toString() +
2938	<	"[" + runLevelToString(rs) +
2938	>	"[" + level +
2939		", parallelism = " + pc +
2940		", size = " + tc +
2941		", active = " + ac +
#	Line 1618 | Line 2946 | public class ForkJoinPool extends Abstra
2946		"]";
2947		}
2948
1621	–	private static String runLevelToString(int s) {
1622	–	return ((s & TERMINATED) != 0 ? "Terminated" :
1623	–	((s & TERMINATING) != 0 ? "Terminating" :
1624	–	((s & SHUTDOWN) != 0 ? "Shutting down" :
1625	–	"Running")));
1626	–	}
1627	–
2949		/**
2950	<	* Initiates an orderly shutdown in which previously submitted
2951	<	* tasks are executed, but no new tasks will be accepted.
2952	<	* Invocation has no additional effect if already shut down.
2953	<	* Tasks that are in the process of being submitted concurrently
2954	<	* during the course of this method may or may not be rejected.
2950	>	* Possibly initiates an orderly shutdown in which previously
2951	>	* submitted tasks are executed, but no new tasks will be
2952	>	* accepted. Invocation has no effect on execution state if this
2953	>	* is the {@link #commonPool}, and no additional effect if
2954	>	* already shut down.  Tasks that are in the process of being
2955	>	* submitted concurrently during the course of this method may or
2956	>	* may not be rejected.
2957		*
2958		* @throws SecurityException if a security manager exists and
2959		*         the caller is not permitted to modify threads
#	Line 1639 | Line 2962 | public class ForkJoinPool extends Abstra
2962		*/
2963		public void shutdown() {
2964		checkPermission();
2965	<	advanceRunLevel(SHUTDOWN);
1643	<	tryTerminate(false);
2965	>	tryTerminate(false, true);
2966		}
2967
2968		/**
2969	<	* Attempts to cancel and/or stop all tasks, and reject all
2970	<	* subsequently submitted tasks.  Tasks that are in the process of
2971	<	* being submitted or executed concurrently during the course of
2972	<	* this method may or may not be rejected. This method cancels
2973	<	* both existing and unexecuted tasks, in order to permit
2974	<	* termination in the presence of task dependencies. So the method
2975	<	* always returns an empty list (unlike the case for some other
2976	<	* Executors).
2969	>	* Possibly attempts to cancel and/or stop all tasks, and reject
2970	>	* all subsequently submitted tasks.  Invocation has no effect on
2971	>	* execution state if this is the {@link #commonPool}, and no
2972	>	* additional effect if already shut down. Otherwise, tasks that
2973	>	* are in the process of being submitted or executed concurrently
2974	>	* during the course of this method may or may not be
2975	>	* rejected. This method cancels both existing and unexecuted
2976	>	* tasks, in order to permit termination in the presence of task
2977	>	* dependencies. So the method always returns an empty list
2978	>	* (unlike the case for some other Executors).
2979		*
2980		* @return an empty list
2981		* @throws SecurityException if a security manager exists and
#	Line 1661 | Line 2985 | public class ForkJoinPool extends Abstra
2985		*/
2986		public List<Runnable> shutdownNow() {
2987		checkPermission();
2988	<	tryTerminate(true);
2988	>	tryTerminate(true, true);
2989		return Collections.emptyList();
2990		}
2991
#	Line 1671 | Line 2995 | public class ForkJoinPool extends Abstra
2995		* @return {@code true} if all tasks have completed following shut down
2996		*/
2997		public boolean isTerminated() {
2998	<	return runState >= TERMINATED;
2998	>	long c = ctl;
2999	>	return ((c & STOP_BIT) != 0L &&
3000	>	(short)(c >>> TC_SHIFT) == -(config & SMASK));
3001		}
3002
3003		/**
#	Line 1679 | Line 3005 | public class ForkJoinPool extends Abstra
3005		* commenced but not yet completed.  This method may be useful for
3006		* debugging. A return of {@code true} reported a sufficient
3007		* period after shutdown may indicate that submitted tasks have
3008	<	* ignored or suppressed interruption, causing this executor not
3009	<	* to properly terminate.
3008	>	* ignored or suppressed interruption, or are waiting for IO,
3009	>	* causing this executor not to properly terminate. (See the
3010	>	* advisory notes for class {@link ForkJoinTask} stating that
3011	>	* tasks should not normally entail blocking operations.  But if
3012	>	* they do, they must abort them on interrupt.)
3013		*
3014		* @return {@code true} if terminating but not yet terminated
3015		*/
3016		public boolean isTerminating() {
3017	<	return (runState & (TERMINATING|TERMINATED)) == TERMINATING;
3017	>	long c = ctl;
3018	>	return ((c & STOP_BIT) != 0L &&
3019	>	(short)(c >>> TC_SHIFT) != -(config & SMASK));
3020		}
3021
3022		/**
#	Line 1694 | Line 3025 | public class ForkJoinPool extends Abstra
3025		* @return {@code true} if this pool has been shut down
3026		*/
3027		public boolean isShutdown() {
3028	<	return runState >= SHUTDOWN;
3028	>	return plock < 0;
3029		}
3030
3031		/**
3032	<	* Blocks until all tasks have completed execution after a shutdown
3033	<	* request, or the timeout occurs, or the current thread is
3034	<	* interrupted, whichever happens first.
3032	>	* Blocks until all tasks have completed execution after a
3033	>	* shutdown request, or the timeout occurs, or the current thread
3034	>	* is interrupted, whichever happens first. Note that the {@link
3035	>	* #commonPool()} never terminates until program shutdown so
3036	>	* this method will always time out.
3037		*
3038		* @param timeout the maximum time to wait
3039		* @param unit the time unit of the timeout argument
#	Line 1710 | Line 3043 | public class ForkJoinPool extends Abstra
3043		*/
3044		public boolean awaitTermination(long timeout, TimeUnit unit)
3045		throws InterruptedException {
3046	<	try {
3047	<	return termination.awaitAdvanceInterruptibly(0, timeout, unit) > 0;
3048	<	} catch (TimeoutException ex) {
3049	<	return false;
3046	>	long nanos = unit.toNanos(timeout);
3047	>	if (isTerminated())
3048	>	return true;
3049	>	long startTime = System.nanoTime();
3050	>	boolean terminated = false;
3051	>	synchronized (this) {
3052	>	for (long waitTime = nanos, millis = 0L;;) {
3053	>	if (terminated = isTerminated() ||
3054	>	waitTime <= 0L ||
3055	>	(millis = unit.toMillis(waitTime)) <= 0L)
3056	>	break;
3057	>	wait(millis);
3058	>	waitTime = nanos - (System.nanoTime() - startTime);
3059	>	}
3060		}
3061	+	return terminated;
3062		}
3063
3064		/**
#	Line 1725 | Line 3069 | public class ForkJoinPool extends Abstra
3069		* {@code isReleasable} must return {@code true} if blocking is
3070		* not necessary. Method {@code block} blocks the current thread
3071		* if necessary (perhaps internally invoking {@code isReleasable}
3072	<	* before actually blocking). The unusual methods in this API
3073	<	* accommodate synchronizers that may, but don't usually, block
3074	<	* for long periods. Similarly, they allow more efficient internal
3075	<	* handling of cases in which additional workers may be, but
3076	<	* usually are not, needed to ensure sufficient parallelism.
3077	<	* Toward this end, implementations of method {@code isReleasable}
3078	<	* must be amenable to repeated invocation.
3072	>	* before actually blocking). These actions are performed by any
3073	>	* thread invoking {@link ForkJoinPool#managedBlock}.  The
3074	>	* unusual methods in this API accommodate synchronizers that may,
3075	>	* but don't usually, block for long periods. Similarly, they
3076	>	* allow more efficient internal handling of cases in which
3077	>	* additional workers may be, but usually are not, needed to
3078	>	* ensure sufficient parallelism.  Toward this end,
3079	>	* implementations of method {@code isReleasable} must be amenable
3080	>	* to repeated invocation.
3081		*
3082		* <p>For example, here is a ManagedBlocker based on a
3083		* ReentrantLock:
#	Line 1812 | Line 3158 | public class ForkJoinPool extends Abstra
3158		throws InterruptedException {
3159		Thread t = Thread.currentThread();
3160		if (t instanceof ForkJoinWorkerThread) {
3161	<	ForkJoinWorkerThread w = (ForkJoinWorkerThread) t;
3162	<	w.pool.awaitBlocker(blocker);
3161	>	ForkJoinPool p = ((ForkJoinWorkerThread)t).pool;
3162	>	while (!blocker.isReleasable()) { // variant of helpSignal
3163	>	WorkQueue[] ws; WorkQueue q; int m, n, u;
3164	>	if ((ws = p.workQueues) != null && (m = ws.length - 1) >= 0) {
3165	>	for (int i = 0; i <= m; ++i) {
3166	>	if (blocker.isReleasable())
3167	>	return;
3168	>	if ((q = ws[i]) != null && (n = q.queueSize()) > 0) {
3169	>	p.signalWork(q, n);
3170	>	if ((u = (int)(p.ctl >>> 32)) >= 0 ||
3171	>	(u >> UAC_SHIFT) >= 0)
3172	>	break;
3173	>	}
3174	>	}
3175	>	}
3176	>	if (p.tryCompensate()) {
3177	>	try {
3178	>	do {} while (!blocker.isReleasable() &&
3179	>	!blocker.block());
3180	>	} finally {
3181	>	p.incrementActiveCount();
3182	>	}
3183	>	break;
3184	>	}
3185	>	}
3186		}
3187		else {
3188	<	do {} while (!blocker.isReleasable() && !blocker.block());
3188	>	do {} while (!blocker.isReleasable() &&
3189	>	!blocker.block());
3190		}
3191		}
3192
#	Line 1825 | Line 3195 | public class ForkJoinPool extends Abstra
3195		// implement RunnableFuture.
3196
3197		protected <T> RunnableFuture<T> newTaskFor(Runnable runnable, T value) {
3198	<	return (RunnableFuture<T>) ForkJoinTask.adapt(runnable, value);
3198	>	return new ForkJoinTask.AdaptedRunnable<T>(runnable, value);
3199		}
3200
3201		protected <T> RunnableFuture<T> newTaskFor(Callable<T> callable) {
3202	<	return (RunnableFuture<T>) ForkJoinTask.adapt(callable);
3202	>	return new ForkJoinTask.AdaptedCallable<T>(callable);
3203		}
3204
3205		// Unsafe mechanics
3206	+	private static final sun.misc.Unsafe U;
3207	+	private static final long CTL;
3208	+	private static final long PARKBLOCKER;
3209	+	private static final int ABASE;
3210	+	private static final int ASHIFT;
3211	+	private static final long STEALCOUNT;
3212	+	private static final long PLOCK;
3213	+	private static final long INDEXSEED;
3214	+	private static final long QLOCK;
3215
3216	<	private static final sun.misc.Unsafe UNSAFE = getUnsafe();
3217	<	private static final long workerCountsOffset =
1839	<	objectFieldOffset("workerCounts", ForkJoinPool.class);
1840	<	private static final long runStateOffset =
1841	<	objectFieldOffset("runState", ForkJoinPool.class);
1842	<	private static final long eventCountOffset =
1843	<	objectFieldOffset("eventCount", ForkJoinPool.class);
1844	<	private static final long eventWaitersOffset =
1845	<	objectFieldOffset("eventWaiters",ForkJoinPool.class);
1846	<	private static final long stealCountOffset =
1847	<	objectFieldOffset("stealCount",ForkJoinPool.class);
1848	<	private static final long spareWaitersOffset =
1849	<	objectFieldOffset("spareWaiters",ForkJoinPool.class);
1850	<
1851	<	private static long objectFieldOffset(String field, Class<?> klazz) {
3216	>	static {
3217	>	int s; // initialize field offsets for CAS etc
3218		try {
3219	<	return UNSAFE.objectFieldOffset(klazz.getDeclaredField(field));
3220	<	} catch (NoSuchFieldException e) {
3221	<	// Convert Exception to corresponding Error
3222	<	NoSuchFieldError error = new NoSuchFieldError(field);
3223	<	error.initCause(e);
3224	<	throw error;
3225	<	}
3219	>	U = getUnsafe();
3220	>	Class<?> k = ForkJoinPool.class;
3221	>	CTL = U.objectFieldOffset
3222	>	(k.getDeclaredField("ctl"));
3223	>	STEALCOUNT = U.objectFieldOffset
3224	>	(k.getDeclaredField("stealCount"));
3225	>	PLOCK = U.objectFieldOffset
3226	>	(k.getDeclaredField("plock"));
3227	>	INDEXSEED = U.objectFieldOffset
3228	>	(k.getDeclaredField("indexSeed"));
3229	>	Class<?> tk = Thread.class;
3230	>	PARKBLOCKER = U.objectFieldOffset
3231	>	(tk.getDeclaredField("parkBlocker"));
3232	>	Class<?> wk = WorkQueue.class;
3233	>	QLOCK = U.objectFieldOffset
3234	>	(wk.getDeclaredField("qlock"));
3235	>	Class<?> ak = ForkJoinTask[].class;
3236	>	ABASE = U.arrayBaseOffset(ak);
3237	>	s = U.arrayIndexScale(ak);
3238	>	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
3239	>	} catch (Exception e) {
3240	>	throw new Error(e);
3241	>	}
3242	>	if ((s & (s-1)) != 0)
3243	>	throw new Error("data type scale not a power of two");
3244	>
3245	>	submitters = new ThreadLocal<Submitter>();
3246	>	ForkJoinWorkerThreadFactory fac = defaultForkJoinWorkerThreadFactory =
3247	>	new DefaultForkJoinWorkerThreadFactory();
3248	>	/*
3249	>	* Establish common pool parameters.  For extra caution,
3250	>	* computations to set up common pool state are here; the
3251	>	* constructor just assigns these values to fields.
3252	>	*/
3253	>
3254	>	int par = 0;
3255	>	Thread.UncaughtExceptionHandler handler = null;
3256	>	try {  // TBD: limit or report ignored exceptions?
3257	>	String pp = System.getProperty
3258	>	("java.util.concurrent.ForkJoinPool.common.parallelism");
3259	>	String hp = System.getProperty
3260	>	("java.util.concurrent.ForkJoinPool.common.exceptionHandler");
3261	>	String fp = System.getProperty
3262	>	("java.util.concurrent.ForkJoinPool.common.threadFactory");
3263	>	if (fp != null)
3264	>	fac = ((ForkJoinWorkerThreadFactory)ClassLoader.
3265	>	getSystemClassLoader().loadClass(fp).newInstance());
3266	>	if (hp != null)
3267	>	handler = ((Thread.UncaughtExceptionHandler)ClassLoader.
3268	>	getSystemClassLoader().loadClass(hp).newInstance());
3269	>	if (pp != null)
3270	>	par = Integer.parseInt(pp);
3271	>	} catch (Exception ignore) {
3272	>	}
3273	>
3274	>	if (par <= 0)
3275	>	par = Runtime.getRuntime().availableProcessors();
3276	>	if (par > MAX_CAP)
3277	>	par = MAX_CAP;
3278	>	commonPoolParallelism = par;
3279	>	long np = (long)(-par); // precompute initial ctl value
3280	>	long ct = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK);
3281	>
3282	>	commonPool = new ForkJoinPool(par, ct, fac, handler);
3283	>	modifyThreadPermission = new RuntimePermission("modifyThread");
3284		}
3285
3286		/**
#	Line 1886 | Line 3310 | public class ForkJoinPool extends Abstra
3310		}
3311		}
3312		}
3313	+
3314		}

Diff Legend

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