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 1.91 by dl, Tue Feb 22 00:39:31 2011 UTC vs.
Revision 1.148 by jsr166, Tue Nov 20 06:18:39 2012 UTC

#	Line 1 | Line 1
1		/*
2		* Written by Doug Lea with assistance from members of JCP JSR-166
3		* Expert Group and released to the public domain, as explained at
4	<	* http://creativecommons.org/licenses/publicdomain
4	>	* http://creativecommons.org/publicdomain/zero/1.0/
5		*/
6
7		package jsr166y;
#	Line 11 | Line 11 | import java.util.Arrays;
11		import java.util.Collection;
12		import java.util.Collections;
13		import java.util.List;
14	–	import java.util.Random;
14		import java.util.concurrent.AbstractExecutorService;
15		import java.util.concurrent.Callable;
16		import java.util.concurrent.ExecutorService;
#	Line 19 | Line 18 | import java.util.concurrent.Future;
18		import java.util.concurrent.RejectedExecutionException;
19		import java.util.concurrent.RunnableFuture;
20		import java.util.concurrent.TimeUnit;
22	–	import java.util.concurrent.TimeoutException;
23	–	import java.util.concurrent.atomic.AtomicInteger;
24	–	import java.util.concurrent.locks.LockSupport;
25	–	import java.util.concurrent.locks.ReentrantLock;
26	–	import java.util.concurrent.locks.Condition;
21
22		/**
23		* An {@link ExecutorService} for running {@link ForkJoinTask}s.
#	Line 34 | Line 28 | import java.util.concurrent.locks.Condit
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 58 | Line 62 | import java.util.concurrent.locks.Condit
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 92 | Line 97 | import java.util.concurrent.locks.Condit
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	<	*
105	<	* <pre>
106	<	* static final ForkJoinPool mainPool = new ForkJoinPool();
107	<	* ...
108	<	* public void sort(long[] array) {
109	<	*   mainPool.invoke(new SortTask(array, 0, array.length));
110	<	* }
111	<	* </pre>
100	>	* <p>The common pool is by default constructed with default
101	>	* parameters, but these may be controlled by setting three {@link
102	>	* System#getProperty properties} with prefix {@code
103	>	* java.util.concurrent.ForkJoinPool.common}: {@code parallelism} --
104	>	* an integer greater than zero, {@code threadFactory} -- the class
105	>	* name of a {@link ForkJoinWorkerThreadFactory}, and {@code
106	>	* exceptionHandler} -- the class name of a {@link
107	>	* java.lang.Thread.UncaughtExceptionHandler
108	>	* Thread.UncaughtExceptionHandler}. Upon any error in establishing
109	>	* these settings, default parameters are used.
110		*
111		* <p><b>Implementation notes</b>: This implementation restricts the
112		* maximum number of running threads to 32767. Attempts to create
#	Line 127 | Line 125 | public class ForkJoinPool extends Abstra
125		/*
126		* Implementation Overview
127		*
128	<	* This class provides the central bookkeeping and control for a
129	<	* set of worker threads: Submissions from non-FJ threads enter
130	<	* into a submission queue. Workers take these tasks and typically
131	<	* split them into subtasks that may be stolen by other workers.
132	<	* Preference rules give first priority to processing tasks from
133	<	* their own queues (LIFO or FIFO, depending on mode), then to
134	<	* randomized FIFO steals of tasks in other worker queues, and
135	<	* lastly to new submissions.
128	>	* This class and its nested classes provide the main
129	>	* functionality and control for a set of worker threads:
130	>	* Submissions from non-FJ threads enter into submission queues.
131	>	* Workers take these tasks and typically split them into subtasks
132	>	* that may be stolen by other workers.  Preference rules give
133	>	* first priority to processing tasks from their own queues (LIFO
134	>	* or FIFO, depending on mode), then to randomized FIFO steals of
135	>	* tasks in other queues.
136	>	*
137	>	* WorkQueues
138	>	* ==========
139	>	*
140	>	* Most operations occur within work-stealing queues (in nested
141	>	* class WorkQueue).  These are special forms of Deques that
142	>	* support only three of the four possible end-operations -- push,
143	>	* pop, and poll (aka steal), under the further constraints that
144	>	* push and pop are called only from the owning thread (or, as
145	>	* extended here, under a lock), while poll may be called from
146	>	* other threads.  (If you are unfamiliar with them, you probably
147	>	* want to read Herlihy and Shavit's book "The Art of
148	>	* Multiprocessor programming", chapter 16 describing these in
149	>	* more detail before proceeding.)  The main work-stealing queue
150	>	* design is roughly similar to those in the papers "Dynamic
151	>	* Circular Work-Stealing Deque" by Chase and Lev, SPAA 2005
152	>	* (http://research.sun.com/scalable/pubs/index.html) and
153	>	* "Idempotent work stealing" by Michael, Saraswat, and Vechev,
154	>	* PPoPP 2009 (http://portal.acm.org/citation.cfm?id=1504186).
155	>	* The main differences ultimately stem from GC requirements that
156	>	* we null out taken slots as soon as we can, to maintain as small
157	>	* a footprint as possible even in programs generating huge
158	>	* numbers of tasks. To accomplish this, we shift the CAS
159	>	* arbitrating pop vs poll (steal) from being on the indices
160	>	* ("base" and "top") to the slots themselves.  So, both a
161	>	* successful pop and poll mainly entail a CAS of a slot from
162	>	* non-null to null.  Because we rely on CASes of references, we
163	>	* do not need tag bits on base or top.  They are simple ints as
164	>	* used in any circular array-based queue (see for example
165	>	* ArrayDeque).  Updates to the indices must still be ordered in a
166	>	* way that guarantees that top == base means the queue is empty,
167	>	* but otherwise may err on the side of possibly making the queue
168	>	* appear nonempty when a push, pop, or poll have not fully
169	>	* committed. Note that this means that the poll operation,
170	>	* considered individually, is not wait-free. One thief cannot
171	>	* successfully continue until another in-progress one (or, if
172	>	* previously empty, a push) completes.  However, in the
173	>	* aggregate, we ensure at least probabilistic non-blockingness.
174	>	* If an attempted steal fails, a thief always chooses a different
175	>	* random victim target to try next. So, in order for one thief to
176	>	* progress, it suffices for any in-progress poll or new push on
177	>	* any empty queue to complete. (This is why we normally use
178	>	* method pollAt and its variants that try once at the apparent
179	>	* base index, else consider alternative actions, rather than
180	>	* method poll.)
181	>	*
182	>	* This approach also enables support of a user mode in which local
183	>	* task processing is in FIFO, not LIFO order, simply by using
184	>	* poll rather than pop.  This can be useful in message-passing
185	>	* frameworks in which tasks are never joined.  However neither
186	>	* mode considers affinities, loads, cache localities, etc, so
187	>	* rarely provide the best possible performance on a given
188	>	* machine, but portably provide good throughput by averaging over
189	>	* these factors.  (Further, even if we did try to use such
190	>	* information, we do not usually have a basis for exploiting it.
191	>	* For example, some sets of tasks profit from cache affinities,
192	>	* but others are harmed by cache pollution effects.)
193	>	*
194	>	* WorkQueues are also used in a similar way for tasks submitted
195	>	* to the pool. We cannot mix these tasks in the same queues used
196	>	* for work-stealing (this would contaminate lifo/fifo
197	>	* processing). Instead, we randomly associate submission queues
198	>	* with submitting threads, using a form of hashing.  The
199	>	* ThreadLocal Submitter class contains a value initially used as
200	>	* a hash code for choosing existing queues, but may be randomly
201	>	* repositioned upon contention with other submitters.  In
202	>	* essence, submitters act like workers except that they are
203	>	* restricted to executing local tasks that they submitted (or in
204	>	* the case of CountedCompleters, others with the same root task).
205	>	* However, because most shared/external queue operations are more
206	>	* expensive than internal, and because, at steady state, external
207	>	* submitters will compete for CPU with workers, ForkJoinTask.join
208	>	* and related methods disable them from repeatedly helping to
209	>	* process tasks if all workers are active.  Insertion of tasks in
210	>	* shared mode requires a lock (mainly to protect in the case of
211	>	* resizing) but we use only a simple spinlock (using bits in
212	>	* field qlock), because submitters encountering a busy queue move
213	>	* on to try or create other queues -- they block only when
214	>	* creating and registering new queues.
215	>	*
216	>	* Management
217	>	* ==========
218		*
219		* The main throughput advantages of work-stealing stem from
220		* decentralized control -- workers mostly take tasks from
221		* themselves or each other. We cannot negate this in the
222		* implementation of other management responsibilities. The main
223		* tactic for avoiding bottlenecks is packing nearly all
224	<	* essentially atomic control state into a single 64bit volatile
225	<	* variable ("ctl"). This variable is read on the order of 10-100
226	<	* times as often as it is modified (always via CAS). (There is
227	<	* some additional control state, for example variable "shutdown"
228	<	* for which we can cope with uncoordinated updates.)  This
229	<	* streamlines synchronization and control at the expense of messy
230	<	* constructions needed to repack status bits upon updates.
231	<	* Updates tend not to contend with each other except during
232	<	* bursts while submitted tasks begin or end.  In some cases when
233	<	* they do contend, threads can instead do something else
234	<	* (usually, scan for tesks) until contention subsides.
235	<	*
236	<	* To enable packing, we restrict maximum parallelism to (1<<15)-1
237	<	* (which is far in excess of normal operating range) to allow
238	<	* ids, counts, and their negations (used for thresholding) to fit
239	<	* into 16bit fields.
240	<	*
241	<	* Recording Workers.  Workers are recorded in the "workers" array
242	<	* that is created upon pool construction and expanded if (rarely)
243	<	* necessary.  This is an array as opposed to some other data
244	<	* structure to support index-based random steals by workers.
245	<	* Updates to the array recording new workers and unrecording
246	<	* terminated ones are protected from each other by a seqLock
247	<	* (scanGuard) but the array is otherwise concurrently readable,
248	<	* and accessed directly by workers. To simplify index-based
249	<	* operations, the array size is always a power of two, and all
250	<	* readers must tolerate null slots. To avoid flailing during
251	<	* start-up, the array is presized to hold twice #parallelism
252	<	* workers (which is unlikely to need further resizing during
253	<	* execution). But to avoid dealing with so many null slots,
254	<	* variable scanGuard includes a mask for the nearest power of two
255	<	* that contains all current workers.  All worker thread creation
256	<	* is on-demand, triggered by task submissions, replacement of
257	<	* terminated workers, and/or compensation for blocked
258	<	* workers. However, all other support code is set up to work with
259	<	* other policies.  To ensure that we do not hold on to worker
260	<	* references that would prevent GC, ALL accesses to workers are
261	<	* via indices into the workers array (which is one source of some
262	<	* of the messy code constructions here). In essence, the workers
263	<	* array serves as a weak reference mechanism. Thus for example
264	<	* the wait queue field of ctl stores worker indices, not worker
265	<	* references.  Access to the workers in associated methods (for
266	<	* example signalWork) must both index-check and null-check the
267	<	* IDs. All such accesses ignore bad IDs by returning out early
268	<	* from what they are doing, since this can only be associated
269	<	* with termination, in which case it is OK to give up.
270	<	*
271	<	* All uses of the workers array, as well as queue arrays, check
272	<	* that the array is non-null (even if previously non-null). This
273	<	* allows nulling during termination, which is currently not
274	<	* necessary, but remains an option for resource-revocation-based
275	<	* shutdown schemes.
276	<	*
277	<	* Wait Queuing. Unlike HPC work-stealing frameworks, we cannot
278	<	* let workers spin indefinitely scanning for tasks when none are
279	<	* can be immediately found, and we cannot start/resume workers
280	<	* unless there appear to be tasks available.  On the other hand,
281	<	* we must quickly prod them into action when new tasks are
282	<	* submitted or generated.  We park/unpark workers after placing
283	<	* in an event wait queue when they cannot find work. This "queue"
284	<	* is actually a simple Treiber stack, headed by the "id" field of
285	<	* ctl, plus a 15bit counter value to both wake up waiters (by
286	<	* advancing their count) and avoid ABA effects. Successors are
287	<	* held in worker field "nextWait".  Queuing deals with several
288	<	* intrinsic races, mainly that a task-producing thread can miss
289	<	* seeing (and signalling) another thread that gave up looking for
290	<	* work but has not yet entered the wait queue. We solve this by
291	<	* requiring a full sweep of all workers both before (in scan())
292	<	* and after (in awaitWork()) a newly waiting worker is added to
293	<	* the wait queue. During a rescan, the worker might release some
294	<	* other queued worker rather than itself, which has the same net
295	<	* effect.
224	>	* essentially atomic control state into two volatile variables
225	>	* that are by far most often read (not written) as status and
226	>	* consistency checks.
227	>	*
228	>	* Field "ctl" contains 64 bits holding all the information needed
229	>	* to atomically decide to add, inactivate, enqueue (on an event
230	>	* queue), dequeue, and/or re-activate workers.  To enable this
231	>	* packing, we restrict maximum parallelism to (1<<15)-1 (which is
232	>	* far in excess of normal operating range) to allow ids, counts,
233	>	* and their negations (used for thresholding) to fit into 16bit
234	>	* fields.
235	>	*
236	>	* Field "plock" is a form of sequence lock with a saturating
237	>	* shutdown bit (similarly for per-queue "qlocks"), mainly
238	>	* protecting updates to the workQueues array, as well as to
239	>	* enable shutdown.  When used as a lock, it is normally only very
240	>	* briefly held, so is nearly always available after at most a
241	>	* brief spin, but we use a monitor-based backup strategy to
242	>	* block when needed.
243	>	*
244	>	* Recording WorkQueues.  WorkQueues are recorded in the
245	>	* "workQueues" array that is created upon first use and expanded
246	>	* if necessary.  Updates to the array while recording new workers
247	>	* and unrecording terminated ones are protected from each other
248	>	* by a lock but the array is otherwise concurrently readable, and
249	>	* accessed directly.  To simplify index-based operations, the
250	>	* array size is always a power of two, and all readers must
251	>	* tolerate null slots. Worker queues are at odd indices. Shared
252	>	* (submission) queues are at even indices, up to a maximum of 64
253	>	* slots, to limit growth even if array needs to expand to add
254	>	* more workers. Grouping them together in this way simplifies and
255	>	* speeds up task scanning.
256	>	*
257	>	* All worker thread creation is on-demand, triggered by task
258	>	* submissions, replacement of terminated workers, and/or
259	>	* compensation for blocked workers. However, all other support
260	>	* code is set up to work with other policies.  To ensure that we
261	>	* do not hold on to worker references that would prevent GC, ALL
262	>	* accesses to workQueues are via indices into the workQueues
263	>	* array (which is one source of some of the messy code
264	>	* constructions here). In essence, the workQueues array serves as
265	>	* a weak reference mechanism. Thus for example the wait queue
266	>	* field of ctl stores indices, not references.  Access to the
267	>	* workQueues in associated methods (for example signalWork) must
268	>	* both index-check and null-check the IDs. All such accesses
269	>	* ignore bad IDs by returning out early from what they are doing,
270	>	* since this can only be associated with termination, in which
271	>	* case it is OK to give up.  All uses of the workQueues array
272	>	* also check that it is non-null (even if previously
273	>	* non-null). This allows nulling during termination, which is
274	>	* currently not necessary, but remains an option for
275	>	* resource-revocation-based shutdown schemes. It also helps
276	>	* reduce JIT issuance of uncommon-trap code, which tends to
277	>	* unnecessarily complicate control flow in some methods.
278	>	*
279	>	* Event Queuing. Unlike HPC work-stealing frameworks, we cannot
280	>	* let workers spin indefinitely scanning for tasks when none can
281	>	* be found immediately, and we cannot start/resume workers unless
282	>	* there appear to be tasks available.  On the other hand, we must
283	>	* quickly prod them into action when new tasks are submitted or
284	>	* generated. In many usages, ramp-up time to activate workers is
285	>	* the main limiting factor in overall performance (this is
286	>	* compounded at program start-up by JIT compilation and
287	>	* allocation). So we try to streamline this as much as possible.
288	>	* We park/unpark workers after placing in an event wait queue
289	>	* when they cannot find work. This "queue" is actually a simple
290	>	* Treiber stack, headed by the "id" field of ctl, plus a 15bit
291	>	* counter value (that reflects the number of times a worker has
292	>	* been inactivated) to avoid ABA effects (we need only as many
293	>	* version numbers as worker threads). Successors are held in
294	>	* field WorkQueue.nextWait.  Queuing deals with several intrinsic
295	>	* races, mainly that a task-producing thread can miss seeing (and
296	>	* signalling) another thread that gave up looking for work but
297	>	* has not yet entered the wait queue. We solve this by requiring
298	>	* a full sweep of all workers (via repeated calls to method
299	>	* scan()) both before and after a newly waiting worker is added
300	>	* to the wait queue. During a rescan, the worker might release
301	>	* some other queued worker rather than itself, which has the same
302	>	* net effect. Because enqueued workers may actually be rescanning
303	>	* rather than waiting, we set and clear the "parker" field of
304	>	* WorkQueues to reduce unnecessary calls to unpark.  (This
305	>	* requires a secondary recheck to avoid missed signals.)  Note
306	>	* the unusual conventions about Thread.interrupts surrounding
307	>	* parking and other blocking: Because interrupts are used solely
308	>	* to alert threads to check termination, which is checked anyway
309	>	* upon blocking, we clear status (using Thread.interrupted)
310	>	* before any call to park, so that park does not immediately
311	>	* return due to status being set via some other unrelated call to
312	>	* interrupt in user code.
313		*
314		* Signalling.  We create or wake up workers only when there
315		* appears to be at least one task they might be able to find and
316	<	* execute.  When a submission is added or another worker adds a
317	<	* task to a queue that previously had two or fewer tasks, they
318	<	* signal waiting workers (or trigger creation of new ones if
319	<	* fewer than the given parallelism level -- see signalWork).
320	<	* These primary signals are buttressed by signals during rescans
321	<	* as well as those performed when a worker steals a task and
322	<	* notices that there are more tasks too; together these cover the
323	<	* signals needed in cases when more than two tasks are pushed
324	<	* but untaken.
316	>	* execute. However, many other threads may notice the same task
317	>	* and each signal to wake up a thread that might take it. So in
318	>	* general, pools will be over-signalled.  When a submission is
319	>	* added or another worker adds a task to a queue that is
320	>	* apparently empty, they signal waiting workers (or trigger
321	>	* creation of new ones if fewer than the given parallelism
322	>	* level).  These primary signals are buttressed by signals
323	>	* whenever other threads scan for work or do not have a task to
324	>	* process (including the case of leaving a hint to unparked
325	>	* threads to help signal others upon wakeup).  On most platforms,
326	>	* signalling (unpark) overhead time is noticeably long, and the
327	>	* time between signalling a thread and it actually making
328	>	* progress can be very noticeably long, so it is worth offloading
329	>	* these delays from critical paths as much as possible.
330		*
331		* Trimming workers. To release resources after periods of lack of
332		* use, a worker starting to wait when the pool is quiescent will
333	<	* time out and terminate if the pool has remained quiescent for
334	<	* SHRINK_RATE nanosecs.
333	>	* time out and terminate if the pool has remained quiescent for a
334	>	* given period -- a short period if there are more threads than
335	>	* parallelism, longer as the number of threads decreases. This
336	>	* will slowly propagate, eventually terminating all workers after
337	>	* periods of non-use.
338		*
339	<	* Submissions. External submissions are maintained in an
340	<	* array-based queue that is structured identically to
341	<	* ForkJoinWorkerThread queues (which see) except for the use of
342	<	* submissionLock in method addSubmission. Unlike worker queues,
343	<	* multiple external threads can add new submissions.
344	<	*
345	<	* Compensation. Beyond work-stealing support and lifecycle
346	<	* control, the main responsibility of this framework is to take
347	<	* actions when one worker is waiting to join a task stolen (or
348	<	* always held by) another.  Because we are multiplexing many
349	<	* tasks on to a pool of workers, we can't just let them block (as
350	<	* in Thread.join).  We also cannot just reassign the joiner's
351	<	* run-time stack with another and replace it later, which would
352	<	* be a form of "continuation", that even if possible is not
353	<	* necessarily a good idea since we sometimes need both an
354	<	* unblocked task and its continuation to progress. Instead we
355	<	* combine two tactics:
339	>	* Shutdown and Termination. A call to shutdownNow atomically sets
340	>	* a plock bit and then (non-atomically) sets each worker's
341	>	* qlock status, cancels all unprocessed tasks, and wakes up
342	>	* all waiting workers.  Detecting whether termination should
343	>	* commence after a non-abrupt shutdown() call requires more work
344	>	* and bookkeeping. We need consensus about quiescence (i.e., that
345	>	* there is no more work). The active count provides a primary
346	>	* indication but non-abrupt shutdown still requires a rechecking
347	>	* scan for any workers that are inactive but not queued.
348	>	*
349	>	* Joining Tasks
350	>	* =============
351	>	*
352	>	* Any of several actions may be taken when one worker is waiting
353	>	* to join a task stolen (or always held) by another.  Because we
354	>	* are multiplexing many tasks on to a pool of workers, we can't
355	>	* just let them block (as in Thread.join).  We also cannot just
356	>	* reassign the joiner's run-time stack with another and replace
357	>	* it later, which would be a form of "continuation", that even if
358	>	* possible is not necessarily a good idea since we sometimes need
359	>	* both an unblocked task and its continuation to progress.
360	>	* Instead we combine two tactics:
361		*
362		*   Helping: Arranging for the joiner to execute some task that it
363	<	*      would be running if the steal had not occurred.  Method
254	<	*      ForkJoinWorkerThread.joinTask tracks joining->stealing
255	<	*      links to try to find such a task.
363	>	*      would be running if the steal had not occurred.
364		*
365		*   Compensating: Unless there are already enough live threads,
366	<	*      method tryPreBlock() may create or re-activate a spare
367	<	*      thread to compensate for blocked joiners until they
368	<	*      unblock.
366	>	*      method tryCompensate() may create or re-activate a spare
367	>	*      thread to compensate for blocked joiners until they unblock.
368	>	*
369	>	* A third form (implemented in tryRemoveAndExec) amounts to
370	>	* helping a hypothetical compensator: If we can readily tell that
371	>	* a possible action of a compensator is to steal and execute the
372	>	* task being joined, the joining thread can do so directly,
373	>	* without the need for a compensation thread (although at the
374	>	* expense of larger run-time stacks, but the tradeoff is
375	>	* typically worthwhile).
376		*
377		* The ManagedBlocker extension API can't use helping so relies
378		* only on compensation in method awaitBlocker.
379		*
380	+	* The algorithm in tryHelpStealer entails a form of "linear"
381	+	* helping: Each worker records (in field currentSteal) the most
382	+	* recent task it stole from some other worker. Plus, it records
383	+	* (in field currentJoin) the task it is currently actively
384	+	* joining. Method tryHelpStealer uses these markers to try to
385	+	* find a worker to help (i.e., steal back a task from and execute
386	+	* it) that could hasten completion of the actively joined task.
387	+	* In essence, the joiner executes a task that would be on its own
388	+	* local deque had the to-be-joined task not been stolen. This may
389	+	* be seen as a conservative variant of the approach in Wagner &
390	+	* Calder "Leapfrogging: a portable technique for implementing
391	+	* efficient futures" SIGPLAN Notices, 1993
392	+	* (http://portal.acm.org/citation.cfm?id=155354). It differs in
393	+	* that: (1) We only maintain dependency links across workers upon
394	+	* steals, rather than use per-task bookkeeping.  This sometimes
395	+	* requires a linear scan of workQueues array to locate stealers,
396	+	* but often doesn't because stealers leave hints (that may become
397	+	* stale/wrong) of where to locate them.  It is only a hint
398	+	* because a worker might have had multiple steals and the hint
399	+	* records only one of them (usually the most current).  Hinting
400	+	* isolates cost to when it is needed, rather than adding to
401	+	* per-task overhead.  (2) It is "shallow", ignoring nesting and
402	+	* potentially cyclic mutual steals.  (3) It is intentionally
403	+	* racy: field currentJoin is updated only while actively joining,
404	+	* which means that we miss links in the chain during long-lived
405	+	* tasks, GC stalls etc (which is OK since blocking in such cases
406	+	* is usually a good idea).  (4) We bound the number of attempts
407	+	* to find work (see MAX_HELP) and fall back to suspending the
408	+	* worker and if necessary replacing it with another.
409	+	*
410	+	* Helping actions for CountedCompleters are much simpler: Method
411	+	* helpComplete can take and execute any task with the same root
412	+	* as the task being waited on. However, this still entails some
413	+	* traversal of completer chains, so is less efficient than using
414	+	* CountedCompleters without explicit joins.
415	+	*
416		* It is impossible to keep exactly the target parallelism number
417		* of threads running at any given time.  Determining the
418		* existence of conservatively safe helping targets, the
419		* availability of already-created spares, and the apparent need
420	<	* to create new spares are all racy and require heuristic
421	<	* guidance, so we rely on multiple retries of each.  Currently,
422	<	* in keeping with on-demand signalling policy, we compensate only
423	<	* if blocking would leave less than one active (non-waiting,
424	<	* non-blocked) worker. Additionally, to avoid some false alarms
425	<	* due to GC, lagging counters, system activity, etc, compensated
426	<	* blocking for joins is only attempted after a number of rechecks
427	<	* proportional to the current apparent deficit (where retries are
428	<	* interspersed with Thread.yield, for good citizenship).  The
429	<	* variable blockedCount, incremented before blocking and
430	<	* decremented after, is sometimes needed to distinguish cases of
431	<	* waiting for work vs blocking on joins or other managed sync,
432	<	* but both the cases are equivalent for most pool control, so we
433	<	* can update non-atomically. (Additionally, contention on
434	<	* blockedCount alleviates some contention on ctl).
435	<	*
436	<	* Shutdown and Termination. A call to shutdownNow atomically sets
437	<	* the ctl stop bit and then (non-atomically) sets each workers
438	<	* "terminate" status, cancels all unprocessed tasks, and wakes up
439	<	* all waiting workers.  Detecting whether termination should
440	<	* commence after a non-abrupt shutdown() call requires more work
441	<	* and bookkeeping. We need consensus about quiesence (i.e., that
442	<	* there is no more work) which is reflected in active counts so
443	<	* long as there are no current blockers, as well as possible
444	<	* re-evaluations during independent changes in blocking or
445	<	* quiescing workers.
446	<	*
447	<	* Style notes: There is a lot of representation-level coupling
448	<	* among classes ForkJoinPool, ForkJoinWorkerThread, and
449	<	* ForkJoinTask.  Most fields of ForkJoinWorkerThread maintain
450	<	* data structures managed by ForkJoinPool, so are directly
451	<	* accessed.  Conversely we allow access to "workers" array by
452	<	* workers, and direct access to ForkJoinTask.status by both
453	<	* ForkJoinPool and ForkJoinWorkerThread.  There is little point
420	>	* to create new spares are all racy, so we rely on multiple
421	>	* retries of each.  Compensation in the apparent absence of
422	>	* helping opportunities is challenging to control on JVMs, where
423	>	* GC and other activities can stall progress of tasks that in
424	>	* turn stall out many other dependent tasks, without us being
425	>	* able to determine whether they will ever require compensation.
426	>	* Even though work-stealing otherwise encounters little
427	>	* degradation in the presence of more threads than cores,
428	>	* aggressively adding new threads in such cases entails risk of
429	>	* unwanted positive feedback control loops in which more threads
430	>	* cause more dependent stalls (as well as delayed progress of
431	>	* unblocked threads to the point that we know they are available)
432	>	* leading to more situations requiring more threads, and so
433	>	* on. This aspect of control can be seen as an (analytically
434	>	* intractable) game with an opponent that may choose the worst
435	>	* (for us) active thread to stall at any time.  We take several
436	>	* precautions to bound losses (and thus bound gains), mainly in
437	>	* methods tryCompensate and awaitJoin.
438	>	*
439	>	* Common Pool
440	>	* ===========
441	>	*
442	>	* The static commonPool always exists after static
443	>	* initialization.  Since it (or any other created pool) need
444	>	* never be used, we minimize initial construction overhead and
445	>	* footprint to the setup of about a dozen fields, with no nested
446	>	* allocation. Most bootstrapping occurs within method
447	>	* fullExternalPush during the first submission to the pool.
448	>	*
449	>	* When external threads submit to the common pool, they can
450	>	* perform some subtask processing (see externalHelpJoin and
451	>	* related methods).  We do not need to record whether these
452	>	* submissions are to the common pool -- if not, externalHelpJoin
453	>	* returns quickly (at the most helping to signal some common pool
454	>	* workers). These submitters would otherwise be blocked waiting
455	>	* for completion, so the extra effort (with liberally sprinkled
456	>	* task status checks) in inapplicable cases amounts to an odd
457	>	* form of limited spin-wait before blocking in ForkJoinTask.join.
458	>	*
459	>	* Style notes
460	>	* ===========
461	>	*
462	>	* There is a lot of representation-level coupling among classes
463	>	* ForkJoinPool, ForkJoinWorkerThread, and ForkJoinTask.  The
464	>	* fields of WorkQueue maintain data structures managed by
465	>	* ForkJoinPool, so are directly accessed.  There is little point
466		* trying to reduce this, since any associated future changes in
467		* representations will need to be accompanied by algorithmic
468	<	* changes anyway. All together, these low-level implementation
469	<	* choices produce as much as a factor of 4 performance
470	<	* improvement compared to naive implementations, and enable the
471	<	* processing of billions of tasks per second, at the expense of
472	<	* some ugliness.
473	<	*
474	<	* Methods signalWork() and scan() are the main bottlenecks so are
475	<	* especially heavily micro-optimized/mangled.  There are lots of
476	<	* inline assignments (of form "while ((local = field) != 0)")
477	<	* which are usually the simplest way to ensure the required read
478	<	* orderings (which are sometimes critical). This leads to a
479	<	* "C"-like style of listing declarations of these locals at the
480	<	* heads of methods or blocks.  There are several occurrences of
481	<	* the unusual "do {} while (!cas...)"  which is the simplest way
482	<	* to force an update of a CAS'ed variable. There are also other
483	<	* coding oddities that help some methods perform reasonably even
484	<	* when interpreted (not compiled).
485	<	*
486	<	* The order of declarations in this file is: (1) declarations of
487	<	* statics (2) fields (along with constants used when unpacking
488	<	* some of them), listed in an order that tends to reduce
489	<	* contention among them a bit under most JVMs.  (3) internal
490	<	* control methods (4) callbacks and other support for
491	<	* ForkJoinTask and ForkJoinWorkerThread classes, (5) exported
492	<	* methods (plus a few little helpers). (6) static block
330	<	* initializing all statics in a minimally dependent order.
468	>	* changes anyway. Several methods intrinsically sprawl because
469	>	* they must accumulate sets of consistent reads of volatiles held
470	>	* in local variables.  Methods signalWork() and scan() are the
471	>	* main bottlenecks, so are especially heavily
472	>	* micro-optimized/mangled.  There are lots of inline assignments
473	>	* (of form "while ((local = field) != 0)") which are usually the
474	>	* simplest way to ensure the required read orderings (which are
475	>	* sometimes critical). This leads to a "C"-like style of listing
476	>	* declarations of these locals at the heads of methods or blocks.
477	>	* There are several occurrences of the unusual "do {} while
478	>	* (!cas...)"  which is the simplest way to force an update of a
479	>	* CAS'ed variable. There are also other coding oddities (including
480	>	* several unnecessary-looking hoisted null checks) that help
481	>	* some methods perform reasonably even when interpreted (not
482	>	* compiled).
483	>	*
484	>	* The order of declarations in this file is:
485	>	* (1) Static utility functions
486	>	* (2) Nested (static) classes
487	>	* (3) Static fields
488	>	* (4) Fields, along with constants used when unpacking some of them
489	>	* (5) Internal control methods
490	>	* (6) Callbacks and other support for ForkJoinTask methods
491	>	* (7) Exported methods
492	>	* (8) Static block initializing statics in minimally dependent order
493		*/
494
495	+	// Static utilities
496	+
497	+	/**
498	+	* If there is a security manager, makes sure caller has
499	+	* permission to modify threads.
500	+	*/
501	+	private static void checkPermission() {
502	+	SecurityManager security = System.getSecurityManager();
503	+	if (security != null)
504	+	security.checkPermission(modifyThreadPermission);
505	+	}
506	+
507	+	// Nested classes
508	+
509		/**
510		* Factory for creating new {@link ForkJoinWorkerThread}s.
511		* A {@code ForkJoinWorkerThreadFactory} must be defined and used
#	Line 350 | Line 526 | public class ForkJoinPool extends Abstra
526		* Default ForkJoinWorkerThreadFactory implementation; creates a
527		* new ForkJoinWorkerThread.
528		*/
529	<	static class DefaultForkJoinWorkerThreadFactory
529	>	static final class DefaultForkJoinWorkerThreadFactory
530		implements ForkJoinWorkerThreadFactory {
531	<	public ForkJoinWorkerThread newThread(ForkJoinPool pool) {
531	>	public final ForkJoinWorkerThread newThread(ForkJoinPool pool) {
532		return new ForkJoinWorkerThread(pool);
533		}
534		}
535
536		/**
537	<	* Creates a new ForkJoinWorkerThread. This factory is used unless
538	<	* overridden in ForkJoinPool constructors.
537	>	* Class for artificial tasks that are used to replace the target
538	>	* of local joins if they are removed from an interior queue slot
539	>	* in WorkQueue.tryRemoveAndExec. We don't need the proxy to
540	>	* actually do anything beyond having a unique identity.
541	>	*/
542	>	static final class EmptyTask extends ForkJoinTask<Void> {
543	>	private static final long serialVersionUID = -7721805057305804111L;
544	>	EmptyTask() { status = ForkJoinTask.NORMAL; } // force done
545	>	public final Void getRawResult() { return null; }
546	>	public final void setRawResult(Void x) {}
547	>	public final boolean exec() { return true; }
548	>	}
549	>
550	>	/**
551	>	* Queues supporting work-stealing as well as external task
552	>	* submission. See above for main rationale and algorithms.
553	>	* Implementation relies heavily on "Unsafe" intrinsics
554	>	* and selective use of "volatile":
555	>	*
556	>	* Field "base" is the index (mod array.length) of the least valid
557	>	* queue slot, which is always the next position to steal (poll)
558	>	* from if nonempty. Reads and writes require volatile orderings
559	>	* but not CAS, because updates are only performed after slot
560	>	* CASes.
561	>	*
562	>	* Field "top" is the index (mod array.length) of the next queue
563	>	* slot to push to or pop from. It is written only by owner thread
564	>	* for push, or under lock for external/shared push, and accessed
565	>	* by other threads only after reading (volatile) base.  Both top
566	>	* and base are allowed to wrap around on overflow, but (top -
567	>	* base) (or more commonly -(base - top) to force volatile read of
568	>	* base before top) still estimates size. The lock ("qlock") is
569	>	* forced to -1 on termination, causing all further lock attempts
570	>	* to fail. (Note: we don't need CAS for termination state because
571	>	* upon pool shutdown, all shared-queues will stop being used
572	>	* anyway.)  Nearly all lock bodies are set up so that exceptions
573	>	* within lock bodies are "impossible" (modulo JVM errors that
574	>	* would cause failure anyway.)
575	>	*
576	>	* The array slots are read and written using the emulation of
577	>	* volatiles/atomics provided by Unsafe. Insertions must in
578	>	* general use putOrderedObject as a form of releasing store to
579	>	* ensure that all writes to the task object are ordered before
580	>	* its publication in the queue.  All removals entail a CAS to
581	>	* null.  The array is always a power of two. To ensure safety of
582	>	* Unsafe array operations, all accesses perform explicit null
583	>	* checks and implicit bounds checks via power-of-two masking.
584	>	*
585	>	* In addition to basic queuing support, this class contains
586	>	* fields described elsewhere to control execution. It turns out
587	>	* to work better memory-layout-wise to include them in this class
588	>	* rather than a separate class.
589	>	*
590	>	* Performance on most platforms is very sensitive to placement of
591	>	* instances of both WorkQueues and their arrays -- we absolutely
592	>	* do not want multiple WorkQueue instances or multiple queue
593	>	* arrays sharing cache lines. (It would be best for queue objects
594	>	* and their arrays to share, but there is nothing available to
595	>	* help arrange that).  Unfortunately, because they are recorded
596	>	* in a common array, WorkQueue instances are often moved to be
597	>	* adjacent by garbage collectors. To reduce impact, we use field
598	>	* padding that works OK on common platforms; this effectively
599	>	* trades off slightly slower average field access for the sake of
600	>	* avoiding really bad worst-case access. (Until better JVM
601	>	* support is in place, this padding is dependent on transient
602	>	* properties of JVM field layout rules.)
603		*/
604	<	public static final ForkJoinWorkerThreadFactory
605	<	defaultForkJoinWorkerThreadFactory;
604	>	static final class WorkQueue {
605	>	/**
606	>	* Capacity of work-stealing queue array upon initialization.
607	>	* Must be a power of two; at least 4, but should be larger to
608	>	* reduce or eliminate cacheline sharing among queues.
609	>	* Currently, it is much larger, as a partial workaround for
610	>	* the fact that JVMs often place arrays in locations that
611	>	* share GC bookkeeping (especially cardmarks) such that
612	>	* per-write accesses encounter serious memory contention.
613	>	*/
614	>	static final int INITIAL_QUEUE_CAPACITY = 1 << 13;
615
616	<	/**
617	<	* Permission required for callers of methods that may start or
618	<	* kill threads.
619	<	*/
620	<	private static final RuntimePermission modifyThreadPermission;
616	>	/**
617	>	* Maximum size for queue arrays. Must be a power of two less
618	>	* than or equal to 1 << (31 - width of array entry) to ensure
619	>	* lack of wraparound of index calculations, but defined to a
620	>	* value a bit less than this to help users trap runaway
621	>	* programs before saturating systems.
622	>	*/
623	>	static final int MAXIMUM_QUEUE_CAPACITY = 1 << 26; // 64M
624
625	<	/**
626	<	* If there is a security manager, makes sure caller has
627	<	* permission to modify threads.
628	<	*/
629	<	private static void checkPermission() {
630	<	SecurityManager security = System.getSecurityManager();
631	<	if (security != null)
632	<	security.checkPermission(modifyThreadPermission);
625	>	int seed;                  // for random scanning; initialize nonzero
626	>	volatile int eventCount;   // encoded inactivation count; < 0 if inactive
627	>	int nextWait;              // encoded record of next event waiter
628	>	int hint;                  // steal or signal hint (index)
629	>	int poolIndex;             // index of this queue in pool (or 0)
630	>	final int mode;            // 0: lifo, > 0: fifo, < 0: shared
631	>	int nsteals;               // number of steals
632	>	volatile int qlock;        // 1: locked, -1: terminate; else 0
633	>	volatile int base;         // index of next slot for poll
634	>	int top;                   // index of next slot for push
635	>	ForkJoinTask<?>[] array;   // the elements (initially unallocated)
636	>	final ForkJoinPool pool;   // the containing pool (may be null)
637	>	final ForkJoinWorkerThread owner; // owning thread or null if shared
638	>	volatile Thread parker;    // == owner during call to park; else null
639	>	volatile ForkJoinTask<?> currentJoin;  // task being joined in awaitJoin
640	>	ForkJoinTask<?> currentSteal; // current non-local task being executed
641	>
642	>	// Heuristic padding to ameliorate unfortunate memory placements
643	>	Object p00, p01, p02, p03, p04, p05, p06, p07;
644	>	Object p08, p09, p0a, p0b, p0c;
645	>
646	>	WorkQueue(ForkJoinPool pool, ForkJoinWorkerThread owner, int mode,
647	>	int seed) {
648	>	this.array = new ForkJoinTask<?>[WorkQueue.INITIAL_QUEUE_CAPACITY];
649	>	this.pool = pool;
650	>	this.owner = owner;
651	>	this.mode = mode;
652	>	this.seed = seed;
653	>	// Place indices in the center of array
654	>	base = top = INITIAL_QUEUE_CAPACITY >>> 1;
655	>	}
656	>
657	>	/**
658	>	* Pushes a task. Call only by owner in unshared queues.
659	>	* Cases needing resizing or rejection are relayed to fullPush
660	>	* (that also handles shared queues).
661	>	*
662	>	* @param task the task. Caller must ensure non-null.
663	>	* @throw RejectedExecutionException if array cannot be resized
664	>	*/
665	>	final void push(ForkJoinTask<?> task) {
666	>	ForkJoinTask<?>[] a; ForkJoinPool p;
667	>	int s = top, m, n;
668	>	if ((a = array) != null) {    // ignore if queue removed
669	>	U.putOrderedObject
670	>	(a, (((m = a.length - 1) & s) << ASHIFT) + ABASE, task);
671	>	if ((n = (top = s + 1) - base) <= 1) {
672	>	if ((p = pool) != null)
673	>	p.signalWork(this, 0);
674	>	}
675	>	else if (n >= m)
676	>	growArray();
677	>	}
678	>	}
679	>
680	>	/**
681	>	* Pushes a task if lock is free and array is either big
682	>	* enough or can be resized to be big enough.
683	>	*
684	>	* @param task the task. Caller must ensure non-null.
685	>	* @return true if submitted
686	>	*/
687	>	final boolean trySharedPush(ForkJoinTask<?> task) {
688	>	boolean submitted = false;
689	>	if (qlock == 0 && U.compareAndSwapInt(this, QLOCK, 0, 1)) {
690	>	ForkJoinTask<?>[] a = array;  ForkJoinPool p;
691	>	int s = top;
692	>	try {
693	>	if ((a != null && a.length > s + 1 - base) ||
694	>	(a = growArray()) != null) {   // must presize
695	>	int j = (((a.length - 1) & s) << ASHIFT) + ABASE;
696	>	U.putOrderedObject(a, j, task);
697	>	top = s + 1;
698	>	submitted = true;
699	>	}
700	>	} finally {
701	>	qlock = 0;                         // unlock
702	>	}
703	>	if (submitted && (p = pool) != null)
704	>	p.signalWork(this, 0);
705	>	}
706	>	return submitted;
707	>	}
708	>
709	>	/**
710	>	* Initializes or doubles the capacity of array. Call either
711	>	* by owner or with lock held -- it is OK for base, but not
712	>	* top, to move while resizings are in progress.
713	>	*/
714	>	final ForkJoinTask<?>[] growArray() {
715	>	ForkJoinTask<?>[] oldA = array;
716	>	int size = oldA != null ? oldA.length << 1 : INITIAL_QUEUE_CAPACITY;
717	>	if (size > MAXIMUM_QUEUE_CAPACITY)
718	>	throw new RejectedExecutionException("Queue capacity exceeded");
719	>	int oldMask, t, b;
720	>	ForkJoinTask<?>[] a = array = new ForkJoinTask<?>[size];
721	>	if (oldA != null && (oldMask = oldA.length - 1) >= 0 &&
722	>	(t = top) - (b = base) > 0) {
723	>	int mask = size - 1;
724	>	do {
725	>	ForkJoinTask<?> x;
726	>	int oldj = ((b & oldMask) << ASHIFT) + ABASE;
727	>	int j    = ((b &    mask) << ASHIFT) + ABASE;
728	>	x = (ForkJoinTask<?>)U.getObjectVolatile(oldA, oldj);
729	>	if (x != null &&
730	>	U.compareAndSwapObject(oldA, oldj, x, null))
731	>	U.putObjectVolatile(a, j, x);
732	>	} while (++b != t);
733	>	}
734	>	return a;
735	>	}
736	>
737	>	/**
738	>	* Takes next task, if one exists, in LIFO order.  Call only
739	>	* by owner in unshared queues.
740	>	*/
741	>	final ForkJoinTask<?> pop() {
742	>	ForkJoinTask<?>[] a; ForkJoinTask<?> t; int m;
743	>	if ((a = array) != null && (m = a.length - 1) >= 0) {
744	>	for (int s; (s = top - 1) - base >= 0;) {
745	>	long j = ((m & s) << ASHIFT) + ABASE;
746	>	if ((t = (ForkJoinTask<?>)U.getObject(a, j)) == null)
747	>	break;
748	>	if (U.compareAndSwapObject(a, j, t, null)) {
749	>	top = s;
750	>	return t;
751	>	}
752	>	}
753	>	}
754	>	return null;
755	>	}
756	>
757	>	/**
758	>	* Takes a task in FIFO order if b is base of queue and a task
759	>	* can be claimed without contention. Specialized versions
760	>	* appear in ForkJoinPool methods scan and tryHelpStealer.
761	>	*/
762	>	final ForkJoinTask<?> pollAt(int b) {
763	>	ForkJoinTask<?> t; ForkJoinTask<?>[] a;
764	>	if ((a = array) != null) {
765	>	int j = (((a.length - 1) & b) << ASHIFT) + ABASE;
766	>	if ((t = (ForkJoinTask<?>)U.getObjectVolatile(a, j)) != null &&
767	>	base == b &&
768	>	U.compareAndSwapObject(a, j, t, null)) {
769	>	base = b + 1;
770	>	return t;
771	>	}
772	>	}
773	>	return null;
774	>	}
775	>
776	>	/**
777	>	* Takes next task, if one exists, in FIFO order.
778	>	*/
779	>	final ForkJoinTask<?> poll() {
780	>	ForkJoinTask<?>[] a; int b; ForkJoinTask<?> t;
781	>	while ((b = base) - top < 0 && (a = array) != null) {
782	>	int j = (((a.length - 1) & b) << ASHIFT) + ABASE;
783	>	t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
784	>	if (t != null) {
785	>	if (base == b &&
786	>	U.compareAndSwapObject(a, j, t, null)) {
787	>	base = b + 1;
788	>	return t;
789	>	}
790	>	}
791	>	else if (base == b) {
792	>	if (b + 1 == top)
793	>	break;
794	>	Thread.yield(); // wait for lagging update (very rare)
795	>	}
796	>	}
797	>	return null;
798	>	}
799	>
800	>	/**
801	>	* Takes next task, if one exists, in order specified by mode.
802	>	*/
803	>	final ForkJoinTask<?> nextLocalTask() {
804	>	return mode == 0 ? pop() : poll();
805	>	}
806	>
807	>	/**
808	>	* Returns next task, if one exists, in order specified by mode.
809	>	*/
810	>	final ForkJoinTask<?> peek() {
811	>	ForkJoinTask<?>[] a = array; int m;
812	>	if (a == null || (m = a.length - 1) < 0)
813	>	return null;
814	>	int i = mode == 0 ? top - 1 : base;
815	>	int j = ((i & m) << ASHIFT) + ABASE;
816	>	return (ForkJoinTask<?>)U.getObjectVolatile(a, j);
817	>	}
818	>
819	>	/**
820	>	* Pops the given task only if it is at the current top.
821	>	* (A shared version is available only via FJP.tryExternalUnpush)
822	>	*/
823	>	final boolean tryUnpush(ForkJoinTask<?> t) {
824	>	ForkJoinTask<?>[] a; int s;
825	>	if ((a = array) != null && (s = top) != base &&
826	>	U.compareAndSwapObject
827	>	(a, (((a.length - 1) & --s) << ASHIFT) + ABASE, t, null)) {
828	>	top = s;
829	>	return true;
830	>	}
831	>	return false;
832	>	}
833	>
834	>	/**
835	>	* Removes and cancels all known tasks, ignoring any exceptions.
836	>	*/
837	>	final void cancelAll() {
838	>	ForkJoinTask.cancelIgnoringExceptions(currentJoin);
839	>	ForkJoinTask.cancelIgnoringExceptions(currentSteal);
840	>	for (ForkJoinTask<?> t; (t = poll()) != null; )
841	>	ForkJoinTask.cancelIgnoringExceptions(t);
842	>	}
843	>
844	>	/**
845	>	* Computes next value for random probes.  Scans don't require
846	>	* a very high quality generator, but also not a crummy one.
847	>	* Marsaglia xor-shift is cheap and works well enough.  Note:
848	>	* This is manually inlined in its usages in ForkJoinPool to
849	>	* avoid writes inside busy scan loops.
850	>	*/
851	>	final int nextSeed() {
852	>	int r = seed;
853	>	r ^= r << 13;
854	>	r ^= r >>> 17;
855	>	return seed = r ^= r << 5;
856	>	}
857	>
858	>	/**
859	>	* Provides a more accurate estimate of size than (top - base)
860	>	* by ordering reads and checking whether a near-empty queue
861	>	* has at least one unclaimed task.
862	>	*/
863	>	final int queueSize() {
864	>	ForkJoinTask<?>[] a; int k, s, n;
865	>	return ((n = base - (s = top)) < 0 &&
866	>	(n != -1 ||
867	>	((a = array) != null && (k = a.length) > 0 &&
868	>	U.getObject
869	>	(a, (long)((((k - 1) & (s - 1)) << ASHIFT) + ABASE)) != null))) ?
870	>	-n : 0;
871	>	}
872	>
873	>	// Specialized execution methods
874	>
875	>	/**
876	>	* Pops and runs tasks until empty.
877	>	*/
878	>	private void popAndExecAll() {
879	>	// A bit faster than repeated pop calls
880	>	ForkJoinTask<?>[] a; int m, s; long j; ForkJoinTask<?> t;
881	>	while ((a = array) != null && (m = a.length - 1) >= 0 &&
882	>	(s = top - 1) - base >= 0 &&
883	>	(t = ((ForkJoinTask<?>)
884	>	U.getObject(a, j = ((m & s) << ASHIFT) + ABASE)))
885	>	!= null) {
886	>	if (U.compareAndSwapObject(a, j, t, null)) {
887	>	top = s;
888	>	t.doExec();
889	>	}
890	>	}
891	>	}
892	>
893	>	/**
894	>	* Polls and runs tasks until empty.
895	>	*/
896	>	private void pollAndExecAll() {
897	>	for (ForkJoinTask<?> t; (t = poll()) != null;)
898	>	t.doExec();
899	>	}
900	>
901	>	/**
902	>	* If present, removes from queue and executes the given task,
903	>	* or any other cancelled task. Returns (true) on any CAS
904	>	* or consistency check failure so caller can retry.
905	>	*
906	>	* @return false if no progress can be made, else true;
907	>	*/
908	>	final boolean tryRemoveAndExec(ForkJoinTask<?> task) {
909	>	boolean stat = true, removed = false, empty = true;
910	>	ForkJoinTask<?>[] a; int m, s, b, n;
911	>	if ((a = array) != null && (m = a.length - 1) >= 0 &&
912	>	(n = (s = top) - (b = base)) > 0) {
913	>	for (ForkJoinTask<?> t;;) {           // traverse from s to b
914	>	int j = ((--s & m) << ASHIFT) + ABASE;
915	>	t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
916	>	if (t == null)                    // inconsistent length
917	>	break;
918	>	else if (t == task) {
919	>	if (s + 1 == top) {           // pop
920	>	if (!U.compareAndSwapObject(a, j, task, null))
921	>	break;
922	>	top = s;
923	>	removed = true;
924	>	}
925	>	else if (base == b)           // replace with proxy
926	>	removed = U.compareAndSwapObject(a, j, task,
927	>	new EmptyTask());
928	>	break;
929	>	}
930	>	else if (t.status >= 0)
931	>	empty = false;
932	>	else if (s + 1 == top) {          // pop and throw away
933	>	if (U.compareAndSwapObject(a, j, t, null))
934	>	top = s;
935	>	break;
936	>	}
937	>	if (--n == 0) {
938	>	if (!empty && base == b)
939	>	stat = false;
940	>	break;
941	>	}
942	>	}
943	>	}
944	>	if (removed)
945	>	task.doExec();
946	>	return stat;
947	>	}
948	>
949	>	/**
950	>	* Polls for and executes the given task or any other task in
951	>	* its CountedCompleter computation
952	>	*/
953	>	final boolean pollAndExecCC(ForkJoinTask<?> root) {
954	>	ForkJoinTask<?>[] a; int b; Object o;
955	>	outer: while ((b = base) - top < 0 && (a = array) != null) {
956	>	long j = (((a.length - 1) & b) << ASHIFT) + ABASE;
957	>	if ((o = U.getObject(a, j)) == null ||
958	>	!(o instanceof CountedCompleter))
959	>	break;
960	>	for (CountedCompleter<?> t = (CountedCompleter<?>)o, r = t;;) {
961	>	if (r == root) {
962	>	if (base == b &&
963	>	U.compareAndSwapObject(a, j, t, null)) {
964	>	base = b + 1;
965	>	t.doExec();
966	>	return true;
967	>	}
968	>	else
969	>	break; // restart
970	>	}
971	>	if ((r = r.completer) == null)
972	>	break outer; // not part of root computation
973	>	}
974	>	}
975	>	return false;
976	>	}
977	>
978	>	/**
979	>	* Executes a top-level task and any local tasks remaining
980	>	* after execution.
981	>	*/
982	>	final void runTask(ForkJoinTask<?> t) {
983	>	if (t != null) {
984	>	(currentSteal = t).doExec();
985	>	currentSteal = null;
986	>	++nsteals;
987	>	if (top != base) {       // process remaining local tasks
988	>	if (mode == 0)
989	>	popAndExecAll();
990	>	else
991	>	pollAndExecAll();
992	>	}
993	>	}
994	>	}
995	>
996	>	/**
997	>	* Executes a non-top-level (stolen) task.
998	>	*/
999	>	final void runSubtask(ForkJoinTask<?> t) {
1000	>	if (t != null) {
1001	>	ForkJoinTask<?> ps = currentSteal;
1002	>	(currentSteal = t).doExec();
1003	>	currentSteal = ps;
1004	>	}
1005	>	}
1006	>
1007	>	/**
1008	>	* Returns true if owned and not known to be blocked.
1009	>	*/
1010	>	final boolean isApparentlyUnblocked() {
1011	>	Thread wt; Thread.State s;
1012	>	return (eventCount >= 0 &&
1013	>	(wt = owner) != null &&
1014	>	(s = wt.getState()) != Thread.State.BLOCKED &&
1015	>	s != Thread.State.WAITING &&
1016	>	s != Thread.State.TIMED_WAITING);
1017	>	}
1018	>
1019	>	/**
1020	>	* If this owned and is not already interrupted, try to
1021	>	* interrupt and/or unpark, ignoring exceptions.
1022	>	*/
1023	>	final void interruptOwner() {
1024	>	Thread wt, p;
1025	>	if ((wt = owner) != null && !wt.isInterrupted()) {
1026	>	try {
1027	>	wt.interrupt();
1028	>	} catch (SecurityException ignore) {
1029	>	}
1030	>	}
1031	>	if ((p = parker) != null)
1032	>	U.unpark(p);
1033	>	}
1034	>
1035	>	// Unsafe mechanics
1036	>	private static final sun.misc.Unsafe U;
1037	>	private static final long QLOCK;
1038	>	private static final int ABASE;
1039	>	private static final int ASHIFT;
1040	>	static {
1041	>	int s;
1042	>	try {
1043	>	U = getUnsafe();
1044	>	Class<?> k = WorkQueue.class;
1045	>	Class<?> ak = ForkJoinTask[].class;
1046	>	QLOCK = U.objectFieldOffset
1047	>	(k.getDeclaredField("qlock"));
1048	>	ABASE = U.arrayBaseOffset(ak);
1049	>	s = U.arrayIndexScale(ak);
1050	>	} catch (Exception e) {
1051	>	throw new Error(e);
1052	>	}
1053	>	if ((s & (s-1)) != 0)
1054	>	throw new Error("data type scale not a power of two");
1055	>	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
1056	>	}
1057		}
1058
1059	+	// static fields (initialized in static initializer below)
1060	+
1061		/**
1062	<	* Generator for assigning sequence numbers as pool names.
1062	>	* Creates a new ForkJoinWorkerThread. This factory is used unless
1063	>	* overridden in ForkJoinPool constructors.
1064		*/
1065	<	private static final AtomicInteger poolNumberGenerator;
1065	>	public static final ForkJoinWorkerThreadFactory
1066	>	defaultForkJoinWorkerThreadFactory;
1067
1068		/**
1069	<	* Generator for initial random seeds for worker victim
1070	<	* selection. This is used only to create initial seeds. Random
1071	<	* steals use a cheaper xorshift generator per steal attempt. We
1072	<	* don't expect much contention on seedGenerator, so just use a
1073	<	* plain Random.
1069	>	* Per-thread records for threads that submit to pools. Currently
1070	>	* holds only pseudo-random seed / index that is used to choose
1071	>	* submission queues in method externalPush. In the future, this may
1072	>	* also incorporate a means to implement different task rejection
1073	>	* and resubmission policies.
1074	>	*
1075	>	* Seeds for submitters and workers/workQueues work in basically
1076	>	* the same way but are initialized and updated using slightly
1077	>	* different mechanics. Both are initialized using the same
1078	>	* approach as in class ThreadLocal, where successive values are
1079	>	* unlikely to collide with previous values. Seeds are then
1080	>	* randomly modified upon collisions using xorshifts, which
1081	>	* requires a non-zero seed.
1082		*/
1083	<	static final Random workerSeedGenerator;
1083	>	static final class Submitter {
1084	>	int seed;
1085	>	Submitter(int s) { seed = s; }
1086	>	}
1087
1088		/**
1089	<	* Array holding all worker threads in the pool.  Initialized upon
1090	<	* construction. Array size must be a power of two.  Updates and
1091	<	* replacements are protected by scanGuard, but the array is
1092	<	* always kept in a consistent enough state to be randomly
1093	<	* accessed without locking by workers performing work-stealing,
403	<	* as well as other traversal-based methods in this class, so long
404	<	* as reads memory-acquire by first reading ctl. All readers must
405	<	* tolerate that some array slots may be null.
1089	>	* Per-thread submission bookkeeping. Shared across all pools
1090	>	* to reduce ThreadLocal pollution and because random motion
1091	>	* to avoid contention in one pool is likely to hold for others.
1092	>	* Lazily initialized on first submission (but null-checked
1093	>	* in other contexts to avoid unnecessary initialization).
1094		*/
1095	<	ForkJoinWorkerThread[] workers;
1095	>	static final ThreadLocal<Submitter> submitters;
1096
1097		/**
1098	<	* Initial size for submission queue array. Must be a power of
1099	<	* two.  In many applications, these always stay small so we use a
1100	<	* small initial cap.
1098	>	* Common (static) pool. Non-null for public use unless a static
1099	>	* construction exception, but internal usages null-check on use
1100	>	* to paranoically avoid potential initialization circularities
1101	>	* as well as to simplify generated code.
1102		*/
1103	<	private static final int INITIAL_QUEUE_CAPACITY = 8;
1103	>	static final ForkJoinPool commonPool;
1104
1105		/**
1106	<	* Maximum size for submission queue array. Must be a power of two
1107	<	* less than or equal to 1 << (31 - width of array entry) to
419	<	* ensure lack of index wraparound, but is capped at a lower
420	<	* value to help users trap runaway computations.
1106	>	* Permission required for callers of methods that may start or
1107	>	* kill threads.
1108		*/
1109	<	private static final int MAXIMUM_QUEUE_CAPACITY = 1 << 24; // 16M
1109	>	private static final RuntimePermission modifyThreadPermission;
1110
1111		/**
1112	<	* Array serving as submission queue. Initialized upon construction.
1112	>	* Common pool parallelism. Must equal commonPool.parallelism.
1113		*/
1114	<	private ForkJoinTask<?>[] submissionQueue;
1114	>	static final int commonPoolParallelism;
1115
1116		/**
1117	<	* Lock protecting submissions array for addSubmission
1117	>	* Sequence number for creating workerNamePrefix.
1118		*/
1119	<	private final ReentrantLock submissionLock;
1119	>	private static int poolNumberSequence;
1120
1121		/**
1122	<	* Condition for awaitTermination, using submissionLock for
1123	<	* convenience.
1122	>	* Return the next sequence number. We don't expect this to
1123	>	* ever contend so use simple builtin sync.
1124		*/
1125	<	private final Condition termination;
1125	>	private static final synchronized int nextPoolId() {
1126	>	return ++poolNumberSequence;
1127	>	}
1128	>
1129	>	// static constants
1130
1131		/**
1132	<	* Creation factory for worker threads.
1132	>	* Initial timeout value (in nanoseconds) for the thread
1133	>	* triggering quiescence to park waiting for new work. On timeout,
1134	>	* the thread will instead try to shrink the number of
1135	>	* workers. The value should be large enough to avoid overly
1136	>	* aggressive shrinkage during most transient stalls (long GCs
1137	>	* etc).
1138		*/
1139	<	private final ForkJoinWorkerThreadFactory factory;
1139	>	private static final long IDLE_TIMEOUT      = 2000L * 1000L * 1000L; // 2sec
1140
1141		/**
1142	<	* The uncaught exception handler used when any worker abruptly
447	<	* terminates.
1142	>	* Timeout value when there are more threads than parallelism level
1143		*/
1144	<	final Thread.UncaughtExceptionHandler ueh;
1144	>	private static final long FAST_IDLE_TIMEOUT =  200L * 1000L * 1000L;
1145
1146		/**
1147	<	* Prefix for assigning names to worker threads
1147	>	* The maximum stolen->joining link depth allowed in method
1148	>	* tryHelpStealer.  Must be a power of two.  Depths for legitimate
1149	>	* chains are unbounded, but we use a fixed constant to avoid
1150	>	* (otherwise unchecked) cycles and to bound staleness of
1151	>	* traversal parameters at the expense of sometimes blocking when
1152	>	* we could be helping.
1153		*/
1154	<	private final String workerNamePrefix;
1154	>	private static final int MAX_HELP = 64;
1155
1156		/**
1157	<	* Sum of per-thread steal counts, updated only when threads are
1158	<	* idle or terminating.
1157	>	* Increment for seed generators. See class ThreadLocal for
1158	>	* explanation.
1159		*/
1160	<	private volatile long stealCount;
1160	>	private static final int SEED_INCREMENT = 0x61c88647;
1161
1162		/**
1163	<	* Main pool control -- a long packed with:
1163	>	* Bits and masks for control variables
1164	>	*
1165	>	* Field ctl is a long packed with:
1166		* AC: Number of active running workers minus target parallelism (16 bits)
1167	<	* TC: Number of total workers minus target parallelism (16bits)
1167	>	* TC: Number of total workers minus target parallelism (16 bits)
1168		* ST: true if pool is terminating (1 bit)
1169		* EC: the wait count of top waiting thread (15 bits)
1170	<	* ID: ~poolIndex of top of Treiber stack of waiting threads (16 bits)
1170	>	* ID: poolIndex of top of Treiber stack of waiters (16 bits)
1171		*
1172		* When convenient, we can extract the upper 32 bits of counts and
1173		* the lower 32 bits of queue state, u = (int)(ctl >>> 32) and e =
#	Line 474 | Line 1176 | public class ForkJoinPool extends Abstra
1176		* parallelism and the positionings of fields makes it possible to
1177		* perform the most common checks via sign tests of fields: When
1178		* ac is negative, there are not enough active workers, when tc is
1179	<	* negative, there are not enough total workers, when id is
478	<	* negative, there is at least one waiting worker, and when e is
1179	>	* negative, there are not enough total workers, and when e is
1180		* negative, the pool is terminating.  To deal with these possibly
1181		* negative fields, we use casts in and out of "short" and/or
1182	<	* signed shifts to maintain signedness.  Note: AC_SHIFT is
1183	<	* redundantly declared in ForkJoinWorkerThread in order to
1184	<	* integrate a surplus-threads check.
1182	>	* signed shifts to maintain signedness.
1183	>	*
1184	>	* When a thread is queued (inactivated), its eventCount field is
1185	>	* set negative, which is the only way to tell if a worker is
1186	>	* prevented from executing tasks, even though it must continue to
1187	>	* scan for them to avoid queuing races. Note however that
1188	>	* eventCount updates lag releases so usage requires care.
1189	>	*
1190	>	* Field plock is an int packed with:
1191	>	* SHUTDOWN: true if shutdown is enabled (1 bit)
1192	>	* SEQ:  a sequence lock, with PL_LOCK bit set if locked (30 bits)
1193	>	* SIGNAL: set when threads may be waiting on the lock (1 bit)
1194	>	*
1195	>	* The sequence number enables simple consistency checks:
1196	>	* Staleness of read-only operations on the workQueues array can
1197	>	* be checked by comparing plock before vs after the reads.
1198		*/
485	–	volatile long ctl;
1199
1200		// bit positions/shifts for fields
1201		private static final int  AC_SHIFT   = 48;
#	Line 491 | Line 1204 | public class ForkJoinPool extends Abstra
1204		private static final int  EC_SHIFT   = 16;
1205
1206		// bounds
1207	<	private static final int  MAX_ID     = 0x7fff;  // max poolIndex
1208	<	private static final int  SMASK      = 0xffff;  // mask short bits
1207	>	private static final int  SMASK      = 0xffff;  // short bits
1208	>	private static final int  MAX_CAP    = 0x7fff;  // max #workers - 1
1209	>	private static final int  EVENMASK   = 0xfffe;  // even short bits
1210	>	private static final int  SQMASK     = 0x007e;  // max 64 (even) slots
1211		private static final int  SHORT_SIGN = 1 << 15;
1212		private static final int  INT_SIGN   = 1 << 31;
1213
#	Line 514 | Line 1229 | public class ForkJoinPool extends Abstra
1229		private static final int  UTC_UNIT   = 1 << UTC_SHIFT;
1230
1231		// masks and units for dealing with e = (int)ctl
1232	<	private static final int  E_MASK     = 0x7fffffff; // no STOP_BIT
1233	<	private static final int  EC_UNIT    = 1 << EC_SHIFT;
1232	>	private static final int E_MASK      = 0x7fffffff; // no STOP_BIT
1233	>	private static final int E_SEQ       = 1 << EC_SHIFT;
1234
1235	<	/**
1236	<	* The target parallelism level.
1237	<	*/
1238	<	final int parallelism;
1235	>	// plock bits
1236	>	private static final int SHUTDOWN    = 1 << 31;
1237	>	private static final int PL_LOCK     = 2;
1238	>	private static final int PL_SIGNAL   = 1;
1239	>	private static final int PL_SPINS    = 1 << 8;
1240	>
1241	>	// access mode for WorkQueue
1242	>	static final int LIFO_QUEUE          =  0;
1243	>	static final int FIFO_QUEUE          =  1;
1244	>	static final int SHARED_QUEUE        = -1;
1245	>
1246	>	// bounds for #steps in scan loop -- must be power 2 minus 1
1247	>	private static final int MIN_SCAN    = 0x1ff;   // cover estimation slop
1248	>	private static final int MAX_SCAN    = 0x1ffff; // 4 * max workers
1249
1250	<	/**
526	<	* Index (mod submission queue length) of next element to take
527	<	* from submission queue.
528	<	*/
529	<	volatile int queueBase;
1250	>	// Instance fields
1251
1252	<	/**
1253	<	* Index (mod submission queue length) of next element to add
1254	<	* in submission queue.
1255	<	*/
1256	<	int queueTop;
1252	>	/*
1253	>	* Field layout of this class tends to matter more than one would
1254	>	* like. Runtime layout order is only loosely related to
1255	>	* declaration order and may differ across JVMs, but the following
1256	>	* empirically works OK on current JVMs.
1257	>	*/
1258	>	volatile long stealCount;                  // collects worker counts
1259	>	volatile long ctl;                         // main pool control
1260	>	volatile int plock;                        // shutdown status and seqLock
1261	>	volatile int indexSeed;                    // worker/submitter index seed
1262	>	final int config;                          // mode and parallelism level
1263	>	WorkQueue[] workQueues;                    // main registry
1264	>	final ForkJoinWorkerThreadFactory factory;
1265	>	final Thread.UncaughtExceptionHandler ueh; // per-worker UEH
1266	>	final String workerNamePrefix;             // to create worker name string
1267
1268	<	/**
1269	<	* True when shutdown() has been called.
1270	<	*/
1271	<	volatile boolean shutdown;
1268	>	/*
1269	>	* Acquires the plock lock to protect worker array and related
1270	>	* updates. This method is called only if an initial CAS on plock
1271	>	* fails. This acts as a spinLock for normal cases, but falls back
1272	>	* to builtin monitor to block when (rarely) needed. This would be
1273	>	* a terrible idea for a highly contended lock, but works fine as
1274	>	* a more conservative alternative to a pure spinlock.  See
1275	>	* internal ConcurrentHashMap documentation for further
1276	>	* explanation of nearly the same construction.
1277	>	*/
1278	>	private int acquirePlock() {
1279	>	int spins = PL_SPINS, r = 0, ps, nps;
1280	>	for (;;) {
1281	>	if (((ps = plock) & PL_LOCK) == 0 &&
1282	>	U.compareAndSwapInt(this, PLOCK, ps, nps = ps + PL_LOCK))
1283	>	return nps;
1284	>	else if (r == 0) { // randomize spins if possible
1285	>	Thread t = Thread.currentThread(); WorkQueue w; Submitter z;
1286	>	if ((t instanceof ForkJoinWorkerThread) &&
1287	>	(w = ((ForkJoinWorkerThread)t).workQueue) != null)
1288	>	r = w.seed;
1289	>	else if ((z = submitters.get()) != null)
1290	>	r = z.seed;
1291	>	else
1292	>	r = 1;
1293	>	}
1294	>	else if (spins >= 0) {
1295	>	r ^= r << 1; r ^= r >>> 3; r ^= r << 10; // xorshift
1296	>	if (r >= 0)
1297	>	--spins;
1298	>	}
1299	>	else if (U.compareAndSwapInt(this, PLOCK, ps, ps | PL_SIGNAL)) {
1300	>	synchronized (this) {
1301	>	if ((plock & PL_SIGNAL) != 0) {
1302	>	try {
1303	>	wait();
1304	>	} catch (InterruptedException ie) {
1305	>	try {
1306	>	Thread.currentThread().interrupt();
1307	>	} catch (SecurityException ignore) {
1308	>	}
1309	>	}
1310	>	}
1311	>	else
1312	>	notifyAll();
1313	>	}
1314	>	}
1315	>	}
1316	>	}
1317
1318		/**
1319	<	* True if use local fifo, not default lifo, for local polling
1320	<	* Read by, and replicated by ForkJoinWorkerThreads
1319	>	* Unlocks and signals any thread waiting for plock. Called only
1320	>	* when CAS of seq value for unlock fails.
1321		*/
1322	<	final boolean locallyFifo;
1322	>	private void releasePlock(int ps) {
1323	>	plock = ps;
1324	>	synchronized (this) { notifyAll(); }
1325	>	}
1326
1327		/**
1328	<	* The number of threads in ForkJoinWorkerThreads.helpQuiescePool.
550	<	* When non-zero, suppresses automatic shutdown when active
551	<	* counts become zero.
1328	>	* Tries to create and start a worker; adjusts counts etc on failure
1329		*/
1330	<	volatile int quiescerCount;
1330	>	private void addWorker() {
1331	>	ForkJoinWorkerThread wt = null;
1332	>	try {
1333	>	(wt = factory.newThread(this)).start();
1334	>	} catch (Throwable ex) {
1335	>	deregisterWorker(wt, ex); // adjust on failure
1336	>	}
1337	>	}
1338
1339		/**
1340	<	* The number of threads blocked in join.
1341	<	*/
1342	<	volatile int blockedCount;
1340	>	* Performs secondary initialization, called when plock is zero.
1341	>	* Creates workQueue array and sets plock to a valid value.  The
1342	>	* lock body must be exception-free (so no try/finally) so we
1343	>	* optimistically allocate new array outside the lock and throw
1344	>	* away if (very rarely) not needed. (A similar tactic is used in
1345	>	* fullExternalPush.)  Because the plock seq value can eventually
1346	>	* wrap around zero, this method harmlessly fails to reinitialize
1347	>	* if workQueues exists, while still advancing plock.
1348	>	*/
1349	>	private void initWorkQueuesArray() {
1350	>	WorkQueue[] ws; int ps;
1351	>	int p = config & SMASK;        // find power of two table size
1352	>	int n = (p > 1) ? p - 1 : 1;   // ensure at least 2 slots
1353	>	n |= n >>> 1; n |= n >>> 2; n |= n >>> 4; n |= n >>> 8; n |= n >>> 16;
1354	>	WorkQueue[] nws = new WorkQueue[(n + 1) << 1];
1355	>	if (((ps = plock) & PL_LOCK) != 0 ||
1356	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1357	>	ps = acquirePlock();
1358	>	if ((ws = workQueues) == null || ws.length == 0)
1359	>	workQueues = nws;
1360	>	int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1361	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1362	>	releasePlock(nps);
1363	>	long c; int u;
1364	>	if ((u = (int)((c = ctl) >>> 32)) < 0 && (int)c == 0) {
1365	>	long nc = (long)(((u + UTC_UNIT) & UTC_MASK) |
1366	>	((u + UAC_UNIT) & UAC_MASK)) << 32;
1367	>	if (U.compareAndSwapLong(this, CTL, c, nc))
1368	>	addWorker();
1369	>	}
1370
1371	<	/**
561	<	* Counter for worker Thread names (unrelated to their poolIndex)
562	<	*/
563	<	private volatile int nextWorkerNumber;
1371	>	}
1372
1373	<	/**
566	<	* The index for the next created worker. Accessed under scanGuard.
567	<	*/
568	<	private int nextWorkerIndex;
1373	>	//  Registering and deregistering workers
1374
1375		/**
1376	<	* SeqLock and index masking for for updates to workers array.
1377	<	* Locked when SG_UNIT is set. Unlocking clears bit by adding
1378	<	* SG_UNIT. Staleness of read-only operations can be checked by
1379	<	* comparing scanGuard to value before the reads. The low 16 bits
1380	<	* (i.e, anding with SMASK) hold (the smallest power of two
1381	<	* covering all worker indices, minus one, and is used to avoid
1382	<	* dealing with large numbers of null slots when the workers array
1383	<	* is overallocated.
1384	<	*/
1385	<	volatile int scanGuard;
1386	<
1387	<	private static final int SG_UNIT = 1 << 16;
1376	>	* Callback from ForkJoinWorkerThread to establish and record its
1377	>	* WorkQueue. To avoid scanning bias due to packing entries in
1378	>	* front of the workQueues array, we treat the array as a simple
1379	>	* power-of-two hash table using per-thread seed as hash,
1380	>	* expanding as needed.
1381	>	*
1382	>	* @param wt the worker thread
1383	>	*/
1384	>	final void registerWorker(ForkJoinWorkerThread wt) {
1385	>	if (wt != null && wt.workQueue == null) {
1386	>	int s, ps;    // generate a rarely colliding candidate index seed
1387	>	do {} while (!U.compareAndSwapInt(this, INDEXSEED, s = indexSeed,
1388	>	s += SEED_INCREMENT) ||
1389	>	s == 0); // skip 0
1390	>	WorkQueue w = new WorkQueue(this, wt, config >>> 16, s);
1391	>	if (((ps = plock) & PL_LOCK) != 0 ||
1392	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1393	>	ps = acquirePlock();
1394	>	int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1395	>	try {
1396	>	WorkQueue[] ws;
1397	>	if ((ws = workQueues) != null && wt.workQueue == null) {
1398	>	int n = ws.length, m = n - 1;
1399	>	int r = (s << 1) | 1;           // use odd-numbered indices
1400	>	if (ws[r &= m] != null) {       // collision
1401	>	int probes = 0;             // step by approx half size
1402	>	int step = (n <= 4) ? 2 : ((n >>> 1) & EVENMASK) + 2;
1403	>	while (ws[r = (r + step) & m] != null) {
1404	>	if (++probes >= n) {
1405	>	workQueues = ws = Arrays.copyOf(ws, n <<= 1);
1406	>	m = n - 1;
1407	>	probes = 0;
1408	>	}
1409	>	}
1410	>	}
1411	>	w.eventCount = w.poolIndex = r; // volatile write orders
1412	>	wt.workQueue = ws[r] = w;
1413	>	}
1414	>	} finally {
1415	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1416	>	releasePlock(nps);
1417	>	}
1418	>	}
1419	>	}
1420
1421		/**
1422	<	* The wakeup interval (in nanoseconds) for a worker waiting for a
1423	<	* task when the pool is quiescent to instead try to shrink the
1424	<	* number of workers.  The exact value does not matter too
1425	<	* much. It must be short enough to release resources during
1426	<	* sustained periods of idleness, but not so short that threads
1427	<	* are continually re-created.
1428	<	*/
1429	<	private static final long SHRINK_RATE =
1430	<	4L * 1000L * 1000L * 1000L; // 4 seconds
1422	>	* Final callback from terminating worker, as well as upon failure
1423	>	* to construct or start a worker.  Removes record of worker from
1424	>	* array, and adjusts counts. If pool is shutting down, tries to
1425	>	* complete termination.
1426	>	*
1427	>	* @param wt the worker thread or null if construction failed
1428	>	* @param ex the exception causing failure, or null if none
1429	>	*/
1430	>	final void deregisterWorker(ForkJoinWorkerThread wt, Throwable ex) {
1431	>	WorkQueue w = null;
1432	>	if (wt != null && (w = wt.workQueue) != null) {
1433	>	int ps;
1434	>	w.qlock = -1;                // ensure set
1435	>	long ns = w.nsteals, sc;     // collect steal count
1436	>	do {} while (!U.compareAndSwapLong(this, STEALCOUNT,
1437	>	sc = stealCount, sc + ns));
1438	>	if (((ps = plock) & PL_LOCK) != 0 ||
1439	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1440	>	ps = acquirePlock();
1441	>	int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1442	>	try {
1443	>	int idx = w.poolIndex;
1444	>	WorkQueue[] ws = workQueues;
1445	>	if (ws != null && idx >= 0 && idx < ws.length && ws[idx] == w)
1446	>	ws[idx] = null;
1447	>	} finally {
1448	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1449	>	releasePlock(nps);
1450	>	}
1451	>	}
1452
1453	<	/**
1454	<	* Top-level loop for worker threads: On each step: if the
1455	<	* previous step swept through all queues and found no tasks, or
1456	<	* there are excess threads, then possibly blocks. Otherwise,
1457	<	* scans for and, if found, executes a task. Returns when pool
1458	<	* and/or worker terminate.
1459	<	*
1460	<	* @param w the worker
1461	<	*/
1462	<	final void work(ForkJoinWorkerThread w) {
1463	<	boolean swept = false;                // true on empty scans
1464	<	long c;
607	<	while (!w.terminate && (int)(c = ctl) >= 0) {
608	<	int a;                            // active count
609	<	if (!swept && (a = (int)(c >> AC_SHIFT)) <= 0)
610	<	swept = scan(w, a);
611	<	else if (tryAwaitWork(w, c))
612	<	swept = false;
1453	>	long c;                             // adjust ctl counts
1454	>	do {} while (!U.compareAndSwapLong
1455	>	(this, CTL, c = ctl, (((c - AC_UNIT) & AC_MASK) |
1456	>	((c - TC_UNIT) & TC_MASK) |
1457	>	(c & ~(AC_MASK|TC_MASK)))));
1458	>
1459	>	if (!tryTerminate(false, false) && w != null) {
1460	>	w.cancelAll();                  // cancel remaining tasks
1461	>	if (w.array != null)            // suppress signal if never ran
1462	>	helpSignal(null, 0);        // wake up or create replacement
1463	>	if (ex == null)                 // help clean refs on way out
1464	>	ForkJoinTask.helpExpungeStaleExceptions();
1465		}
1466	+
1467	+	if (ex != null)                     // rethrow
1468	+	ForkJoinTask.rethrow(ex);
1469		}
1470
1471	<	// Signalling
1471	>	// Submissions
1472
1473		/**
1474	<	* Wakes up or creates a worker.
1475	<	*/
1476	<	final void signalWork() {
1477	<	/*
1478	<	* The while condition is true if: (there is are too few total
1479	<	* workers OR there is at least one waiter) AND (there are too
1480	<	* few active workers OR the pool is terminating).  The value
1481	<	* of e distinguishes the remaining cases: zero (no waiters)
1482	<	* for create, negative if terminating (in which case do
1483	<	* nothing), else release a waiter. The secondary checks for
1484	<	* release (non-null array etc) can fail if the pool begins
1485	<	* terminating after the test, and don't impose any added cost
1486	<	* because JVMs must perform null and bounds checks anyway.
1487	<	*/
1488	<	long c; int e, u;
1489	<	while ((((e = (int)(c = ctl)) | (u = (int)(c >>> 32))) &
1490	<	(INT_SIGN|SHORT_SIGN)) == (INT_SIGN|SHORT_SIGN) && e >= 0) {
1491	<	if (e > 0) {                         // release a waiting worker
1492	<	int i; ForkJoinWorkerThread w; ForkJoinWorkerThread[] ws;
1493	<	if ((ws = workers) == null ||
1494	<	(i = ~e & SMASK) >= ws.length ||
1495	<	(w = ws[i]) == null)
1496	<	break;
1497	<	long nc = (((long)(w.nextWait & E_MASK)) |
1498	<	((long)(u + UAC_UNIT) << 32));
1499	<	if (w.eventCount == e &&
1500	<	UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
1501	<	w.eventCount = (e + EC_UNIT) & E_MASK;
1502	<	if (w.parked)
1503	<	UNSAFE.unpark(w);
1504	<	break;
1505	<	}
1474	>	* Unless shutting down, adds the given task to a submission queue
1475	>	* at submitter's current queue index (modulo submission
1476	>	* range). Only the most common path is directly handled in this
1477	>	* method. All others are relayed to fullExternalPush.
1478	>	*
1479	>	* @param task the task. Caller must ensure non-null.
1480	>	*/
1481	>	final void externalPush(ForkJoinTask<?> task) {
1482	>	WorkQueue[] ws; WorkQueue q; Submitter z; int m; ForkJoinTask<?>[] a;
1483	>	if ((z = submitters.get()) != null && plock > 0 &&
1484	>	(ws = workQueues) != null && (m = (ws.length - 1)) >= 0 &&
1485	>	(q = ws[m & z.seed & SQMASK]) != null &&
1486	>	U.compareAndSwapInt(q, QLOCK, 0, 1)) { // lock
1487	>	int b = q.base, s = q.top, n, an;
1488	>	if ((a = q.array) != null && (an = a.length) > (n = s + 1 - b)) {
1489	>	U.putObject(a, (long)(((an - 1) & s) << ASHIFT) + ABASE, task);
1490	>	q.top = s + 1;                     // push on to deque
1491	>	q.qlock = 0;
1492	>	if (n <= 2)
1493	>	signalWork(q, 0);
1494	>	return;
1495	>	}
1496	>	q.qlock = 0;
1497	>	}
1498	>	fullExternalPush(task);
1499	>	}
1500	>
1501	>	/**
1502	>	* Full version of externalPush. This method is called, among
1503	>	* other times, upon the first submission of the first task to the
1504	>	* pool, so must perform secondary initialization (via
1505	>	* initWorkQueuesArray). It also detects first submission by an
1506	>	* external thread by looking up its ThreadLocal, and creates a
1507	>	* new shared queue if the one at index if empty or contended. The
1508	>	* lock body must be exception-free (so no try/finally) so we
1509	>	* optimistically allocate new queues outside the lock and throw
1510	>	* them away if (very rarely) not needed.
1511	>	*/
1512	>	private void fullExternalPush(ForkJoinTask<?> task) {
1513	>	int r = 0;
1514	>	for (Submitter z = submitters.get();;) {
1515	>	WorkQueue[] ws; WorkQueue q; int ps, m, k;
1516	>	if (z == null) {
1517	>	if (U.compareAndSwapInt(this, INDEXSEED, r = indexSeed,
1518	>	r += SEED_INCREMENT) && r != 0)
1519	>	submitters.set(z = new Submitter(r));
1520	>	}
1521	>	else if (r == 0) {               // move to a different index
1522	>	r = z.seed;
1523	>	r ^= r << 13;                // same xorshift as WorkQueues
1524	>	r ^= r >>> 17;
1525	>	z.seed = r ^ (r << 5);
1526	>	}
1527	>	else if ((ps = plock) < 0)
1528	>	throw new RejectedExecutionException();
1529	>	else if (ps == 0 || (ws = workQueues) == null ||
1530	>	(m = ws.length - 1) < 0)
1531	>	initWorkQueuesArray();
1532	>	else if ((q = ws[k = r & m & SQMASK]) != null) {
1533	>	if (q.trySharedPush(task))
1534	>	return;
1535	>	else
1536	>	r = 0; // move on contention
1537		}
1538	<	else if (UNSAFE.compareAndSwapLong
1539	<	(this, ctlOffset, c,
1540	<	(long)(((u + UTC_UNIT) & UTC_MASK) |
1541	<	((u + UAC_UNIT) & UAC_MASK)) << 32)) {
1542	<	addWorker();
1543	<	break;
1538	>	else if (((ps = plock) & PL_LOCK) == 0) { // create new queue
1539	>	q = new WorkQueue(this, null, SHARED_QUEUE, r);
1540	>	if (((ps = plock) & PL_LOCK) != 0 ||
1541	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1542	>	ps = acquirePlock();
1543	>	if ((ws = workQueues) != null && k < ws.length && ws[k] == null)
1544	>	ws[k] = q;
1545	>	int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1546	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1547	>	releasePlock(nps);
1548		}
1549	+	else
1550	+	r = 0; // try elsewhere while lock held
1551		}
1552		}
1553
1554	+	// Maintaining ctl counts
1555	+
1556		/**
1557	<	* Variant of signalWork to help release waiters on rescans.
664	<	* Tries once to release a waiter if active count < 0.
665	<	*
666	<	* @return false if failed due to contention, else true
1557	>	* Increments active count; mainly called upon return from blocking.
1558		*/
1559	<	private boolean tryReleaseWaiter() {
1560	<	long c; int e, i; ForkJoinWorkerThread w; ForkJoinWorkerThread[] ws;
1561	<	if ((e = (int)(c = ctl)) > 0 &&
671	<	(int)(c >> AC_SHIFT) < 0 &&
672	<	(ws = workers) != null &&
673	<	(i = ~e & SMASK) < ws.length &&
674	<	(w = ws[i]) != null) {
675	<	long nc = ((long)(w.nextWait & E_MASK) |
676	<	((c + AC_UNIT) & (AC_MASK|TC_MASK)));
677	<	if (w.eventCount != e ||
678	<	!UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc))
679	<	return false;
680	<	w.eventCount = (e + EC_UNIT) & E_MASK;
681	<	if (w.parked)
682	<	UNSAFE.unpark(w);
683	<	}
684	<	return true;
1559	>	final void incrementActiveCount() {
1560	>	long c;
1561	>	do {} while (!U.compareAndSwapLong(this, CTL, c = ctl, c + AC_UNIT));
1562		}
1563
687	–	// Scanning for tasks
688	–
1564		/**
1565	<	* Scans for and, if found, executes one task. Scans start at a
1566	<	* random index of workers array, and randomly select the first
1567	<	* (2*#workers)-1 probes, and then, if all empty, resort to 2
1568	<	* circular sweeps, which is necessary to check quiescence. and
1569	<	* taking a submission only if no stealable tasks were found.  The
1570	<	* steal code inside the loop is a specialized form of
1571	<	* ForkJoinWorkerThread.deqTask, followed bookkeeping to support
1572	<	* helpJoinTask and signal propagation. The code for submission
1573	<	* queues is almost identical. On each steal, the worker completes
1574	<	* not only the task, but also all local tasks that this task may
1575	<	* have generated. On detecting staleness or contention when
1576	<	* trying to take a task, this method returns without finishing
1577	<	* sweep, which allows global state rechecks before retry.
1578	<	*
1579	<	* @param w the worker
1580	<	* @param a the number of active workers
1581	<	* @return true if swept all queues without finding a task
1582	<	*/
1583	<	private boolean scan(ForkJoinWorkerThread w, int a) {
1584	<	int g = scanGuard; // mask 0 avoids useless scans if only one active
1585	<	int m = parallelism == 1 - a? 0 : g & SMASK;
1586	<	ForkJoinWorkerThread[] ws = workers;
1587	<	if (ws == null || ws.length <= m)         // staleness check
1588	<	return false;
1589	<	for (int r = w.seed, k = r, j = -(m + m); j <= m + m; ++j) {
1590	<	ForkJoinTask<?> t; ForkJoinTask<?>[] q; int b, i;
1591	<	ForkJoinWorkerThread v = ws[k & m];
1592	<	if (v != null && (b = v.queueBase) != v.queueTop &&
1593	<	(q = v.queue) != null && (i = (q.length - 1) & b) >= 0) {
1594	<	long u = (i << ASHIFT) + ABASE;
720	<	if ((t = q[i]) != null && v.queueBase == b &&
721	<	UNSAFE.compareAndSwapObject(q, u, t, null)) {
722	<	int d = (v.queueBase = b + 1) - v.queueTop;
723	<	v.stealHint = w.poolIndex;
724	<	if (d != 0)
725	<	signalWork();             // propagate if nonempty
726	<	w.execTask(t);
1565	>	* Tries to create (at most one) or activate (possibly several)
1566	>	* workers if too few are active. On contention failure, continues
1567	>	* until at least one worker is signalled or the given queue is
1568	>	* empty or all workers are active.
1569	>	*
1570	>	* @param q if non-null, the queue holding tasks to be signalled
1571	>	* @param signals the target number of signals (at least one --
1572	>	* if argument is zero also sets signallee hint if parked).
1573	>	*/
1574	>	final void signalWork(WorkQueue q, int signals) {
1575	>	long c; int e, u, i, s; WorkQueue[] ws; WorkQueue w; Thread p;
1576	>	while ((u = (int)((c = ctl) >>> 32)) < 0) {
1577	>	if ((e = (int)c) > 0) {
1578	>	if ((ws = workQueues) != null && ws.length > (i = e & SMASK) &&
1579	>	(w = ws[i]) != null && w.eventCount == (e | INT_SIGN)) {
1580	>	long nc = (((long)(w.nextWait & E_MASK)) |
1581	>	((long)(u + UAC_UNIT) << 32));
1582	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1583	>	w.eventCount = (e + E_SEQ) & E_MASK;
1584	>	if ((p = w.parker) != null) {
1585	>	if (q != null && signals == 0)
1586	>	w.hint = q.poolIndex;
1587	>	U.unpark(p);
1588	>	}
1589	>	if (--signals <= 0)
1590	>	break;
1591	>	}
1592	>	if (q != null && (s = q.queueSize()) <= signals &&
1593	>	(signals = s) <= 0)
1594	>	break;
1595		}
1596	<	r ^= r << 13; r ^= r >>> 17; w.seed = r ^ (r << 5);
1597	<	return false;                     // store next seed
730	<	}
731	<	else if (j < 0) {                     // xorshift
732	<	r ^= r << 13; r ^= r >>> 17; k = r ^= r << 5;
1596	>	else
1597	>	break;
1598		}
1599	<	else
1600	<	++k;
1601	<	}
1602	<	if (scanGuard != g)                       // staleness check
1603	<	return false;
1604	<	else {                                    // try to take submission
740	<	ForkJoinTask<?> t; ForkJoinTask<?>[] q; int b, i;
741	<	if ((b = queueBase) != queueTop &&
742	<	(q = submissionQueue) != null &&
743	<	(i = (q.length - 1) & b) >= 0) {
744	<	long u = (i << ASHIFT) + ABASE;
745	<	if ((t = q[i]) != null && queueBase == b &&
746	<	UNSAFE.compareAndSwapObject(q, u, t, null)) {
747	<	queueBase = b + 1;
748	<	w.execTask(t);
1599	>	else if (e == 0 && (u & SHORT_SIGN) != 0) {
1600	>	long nc = (long)(((u + UTC_UNIT) & UTC_MASK) |
1601	>	((u + UAC_UNIT) & UAC_MASK)) << 32;
1602	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1603	>	addWorker();
1604	>	break;
1605		}
750	–	return false;
1606		}
1607	<	return true;                         // all queues empty
1607	>	else
1608	>	break;
1609		}
1610		}
1611
1612	+	// Scanning for tasks
1613	+
1614		/**
1615	<	* Tries to enqueue worker in wait queue and await change in
758	<	* worker's eventCount.  Before blocking, rescans queues to avoid
759	<	* missed signals.  If the pool is quiescent, possibly terminates
760	<	* worker upon exit.
761	<	*
762	<	* @param w the calling worker
763	<	* @param c the ctl value on entry
764	<	* @return true if waited or another thread was released upon enq
1615	>	* Top-level runloop for workers, called by ForkJoinWorkerThread.run.
1616		*/
1617	<	private boolean tryAwaitWork(ForkJoinWorkerThread w, long c) {
1618	<	int v = w.eventCount;
1619	<	w.nextWait = (int)c;                       // w's successor record
1620	<	long nc = (long)(v & E_MASK) | ((c - AC_UNIT) & (AC_MASK|TC_MASK));
1621	<	if (ctl != c || !UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
1622	<	long d = ctl; // return true if lost to a deq, to force rescan
1623	<	return (int)d != (int)c && ((d - c) & AC_MASK) >= 0L;
1624	<	}
1625	<	boolean rescanned = false;
1626	<	for (int sc;;) {
1627	<	if (w.eventCount != v)
1628	<	return true;
1629	<	if ((sc = w.stealCount) != 0) {
1630	<	long s = stealCount;               // accumulate stealCount
1631	<	if (UNSAFE.compareAndSwapLong(this, stealCountOffset, s, s+sc))
1632	<	w.stealCount = 0;
1633	<	}
1634	<	else if (!rescanned) {
1635	<	int g = scanGuard, m = g & SMASK;
1636	<	ForkJoinWorkerThread[] ws = workers;
1637	<	if (ws != null && m < ws.length) {
1638	<	rescanned = true;
1639	<	for (int i = 0; i <= m; ++i) {
1640	<	ForkJoinWorkerThread u = ws[i];
1641	<	if (u != null) {
1642	<	if (u.queueBase != u.queueTop &&
1643	<	!tryReleaseWaiter())
1644	<	rescanned = false; // contended
1645	<	if (w.eventCount != v)
1646	<	return true;
1647	<	}
1617	>	final void runWorker(WorkQueue w) {
1618	>	if (w != null) // skip on initialization failure
1619	>	do { w.runTask(scan(w)); } while (w.qlock >= 0);
1620	>	}
1621	>
1622	>	/**
1623	>	* Scans for and, if found, returns one task, else possibly
1624	>	* inactivates the worker. This method operates on single reads of
1625	>	* volatile state and is designed to be re-invoked continuously,
1626	>	* in part because it returns upon detecting inconsistencies,
1627	>	* contention, or state changes that indicate possible success on
1628	>	* re-invocation.
1629	>	*
1630	>	* The scan searches for tasks across queues (starting at a random
1631	>	* index, and relying on registerWorker to irregularly scatter
1632	>	* them within array to avoid bias), checking each at least twice.
1633	>	* The scan terminates upon either finding a non-empty queue, or
1634	>	* completing the sweep. If the worker is not inactivated, it
1635	>	* takes and returns a task from this queue. Otherwise, if not
1636	>	* activated, it signals workers (that may include itself) and
1637	>	* returns so caller can retry. Also returns for true if the
1638	>	* worker array may have changed during an empty scan.  On failure
1639	>	* to find a task, we take one of the following actions, after
1640	>	* which the caller will retry calling this method unless
1641	>	* terminated.
1642	>	*
1643	>	* * If pool is terminating, terminate the worker.
1644	>	*
1645	>	* * If not already enqueued, try to inactivate and enqueue the
1646	>	* worker on wait queue. Or, if inactivating has caused the pool
1647	>	* to be quiescent, relay to idleAwaitWork to check for
1648	>	* termination and possibly shrink pool.
1649	>	*
1650	>	* * If already enqueued and none of the above apply, possibly
1651	>	* (with 1/2 probability) park awaiting signal, else lingering to
1652	>	* help scan and signal.
1653	>	*
1654	>	* @param w the worker (via its WorkQueue)
1655	>	* @return a task or null if none found
1656	>	*/
1657	>	private final ForkJoinTask<?> scan(WorkQueue w) {
1658	>	WorkQueue[] ws; int m, hint;
1659	>	int ps = plock;                          // read plock before ws
1660	>	if (w != null && (ws = workQueues) != null && (m = ws.length - 1) >= 0) {
1661	>	int ec = w.eventCount;               // ec is negative if inactive
1662	>	int r = w.seed; r ^= r << 13; r ^= r >>> 17; w.seed = r ^= r << 5;
1663	>	for (int j = ((m + m + 1) | MIN_SCAN) & MAX_SCAN; ; --j) {
1664	>	WorkQueue q; ForkJoinTask<?>[] a; int b;
1665	>	if ((q = ws[(r + j) & m]) != null && (b = q.base) - q.top < 0 &&
1666	>	(a = q.array) != null) {     // probably nonempty
1667	>	int i = (((a.length - 1) & b) << ASHIFT) + ABASE;
1668	>	ForkJoinTask<?> t = (ForkJoinTask<?>)
1669	>	U.getObjectVolatile(a, i);
1670	>	if (q.base == b && ec >= 0 && t != null &&
1671	>	U.compareAndSwapObject(a, i, t, null)) {
1672	>	if ((q.base = b + 1) - q.top < 0)
1673	>	signalWork(q, 0);
1674	>	return t;                // taken
1675	>	}
1676	>	else if (ec < 0 || j < m) {  // cannot take or cannot rescan
1677	>	w.hint = q.poolIndex;    // use hint below
1678	>	break;                   // let caller retry after signal
1679		}
1680		}
1681	<	if (scanGuard != g ||              // stale
1682	<	(queueBase != queueTop && !tryReleaseWaiter()))
1683	<	rescanned = false;
1684	<	if (!rescanned)
1685	<	Thread.yield();                // reduce contention
1686	<	else
1687	<	Thread.interrupted();          // clear before park
1688	<	}
1689	<	else if (parallelism + (int)(ctl >> AC_SHIFT) == 0 &&
1690	<	blockedCount == 0 && quiescerCount == 0)
1691	<	idleAwaitWork(w, v);               // quiescent -- maybe shrink
1692	<	else {
1693	<	w.parked = true;                   // must recheck
1694	<	if (w.eventCount != v) {
1695	<	w.parked = false;
1696	<	return true;
1681	>	else if (j < 0) { // end of scan; in loop to simplify code
1682	>	long c, sc; int e, ns;
1683	>	if ((ns = w.nsteals) != 0) {
1684	>	if (U.compareAndSwapLong(this, STEALCOUNT,
1685	>	sc = stealCount, sc + ns))
1686	>	w.nsteals = 0;       // collect steals
1687	>	}
1688	>	else if (plock != ps)        // ws may have changed
1689	>	break;
1690	>	else if ((e = (int)(c = ctl)) < 0)
1691	>	w.qlock = -1;            // pool is terminating
1692	>	else if (ec >= 0) {          // try to enqueue/inactivate
1693	>	long nc = ((long)ec |
1694	>	((c - AC_UNIT) & (AC_MASK|TC_MASK)));
1695	>	w.nextWait = e;          // link and mark inactive
1696	>	w.hint = -1;             // use hint if set while parked
1697	>	w.eventCount = ec | INT_SIGN;
1698	>	if (ctl != c ||
1699	>	!U.compareAndSwapLong(this, CTL, c, nc))
1700	>	w.eventCount = ec;  // unmark on CAS failure
1701	>	else if ((int)(c >> AC_SHIFT) == 1 - (config & SMASK))
1702	>	idleAwaitWork(w, nc, c);
1703	>	}
1704	>	else if (w.eventCount < 0) { // block
1705	>	Thread wt = Thread.currentThread();
1706	>	Thread.interrupted();    // clear status
1707	>	U.putObject(wt, PARKBLOCKER, this);
1708	>	w.parker = wt;           // emulate LockSupport.park
1709	>	if (w.eventCount < 0)    // recheck
1710	>	U.park(false, 0L);
1711	>	w.parker = null;
1712	>	U.putObject(wt, PARKBLOCKER, null);
1713	>	}
1714	>	break;
1715		}
1716	<	LockSupport.park(this);
1717	<	rescanned = w.parked = false;
1716	>	}
1717	>	if ((hint = w.hint) >= 0) {          // help signal
1718	>	WorkQueue[] vs; WorkQueue v; int k;
1719	>	w.hint = -1;                     // suppress resignal
1720	>	if ((vs = workQueues) != null && hint < vs.length &&
1721	>	(v = vs[hint]) != null && (k = v.base - v.top) < -1)
1722	>	signalWork(v, 1 - k);
1723		}
1724		}
1725	+	return null;
1726		}
1727
1728		/**
1729	<	* If pool is quiescent, checks for termination, and waits for
1730	<	* event signal for up to SHRINK_RATE nanosecs. On timeout, if ctl
1731	<	* has not changed, terminates the worker. Upon its termination
1732	<	* (see deregisterWorker), it may wake up another worker to
1733	<	* possibly repeat this process.
1729	>	* If inactivating worker w has caused the pool to become
1730	>	* quiescent, checks for pool termination, and, so long as this is
1731	>	* not the only worker, waits for event for up to a given
1732	>	* duration.  On timeout, if ctl has not changed, terminates the
1733	>	* worker, which will in turn wake up another worker to possibly
1734	>	* repeat this process.
1735		*
1736		* @param w the calling worker
1737	<	* @param v the eventCount w must wait until changed
1737	>	* @param currentCtl the ctl value triggering possible quiescence
1738	>	* @param prevCtl the ctl value to restore if thread is terminated
1739		*/
1740	<	private void idleAwaitWork(ForkJoinWorkerThread w, int v) {
1741	<	ForkJoinTask.helpExpungeStaleExceptions(); // help clean weak refs
1742	<	if (shutdown)
1743	<	tryTerminate(false);
1744	<	long c = ctl;
1745	<	long nc = (((c & (AC_MASK|TC_MASK)) + AC_UNIT) |
1746	<	(long)(w.nextWait & E_MASK)); // ctl value to release w
1747	<	if (w.eventCount == v &&
1748	<	parallelism + (int)(c >> AC_SHIFT) == 0 &&
1749	<	blockedCount == 0 && quiescerCount == 0) {
1750	<	long startTime = System.nanoTime();
1751	<	Thread.interrupted();
1752	<	if (w.eventCount == v) {
1753	<	w.parked = true;
1754	<	if (w.eventCount == v)
1755	<	LockSupport.parkNanos(this, SHRINK_RATE);
1756	<	w.parked = false;
1757	<	if (w.eventCount == v && ctl == c &&
1758	<	System.nanoTime() - startTime >= SHRINK_RATE &&
1759	<	UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
1760	<	w.terminate = true;
1761	<	w.eventCount = ((int)c + EC_UNIT) & E_MASK;
1740	>	private void idleAwaitWork(WorkQueue w, long currentCtl, long prevCtl) {
1741	>	if (w != null && w.eventCount < 0 &&
1742	>	!tryTerminate(false, false) && (int)prevCtl != 0) {
1743	>	int dc = -(short)(currentCtl >>> TC_SHIFT);
1744	>	long parkTime = dc < 0 ? FAST_IDLE_TIMEOUT: (dc + 1) * IDLE_TIMEOUT;
1745	>	long deadline = System.nanoTime() + parkTime - 100000L; // 1ms slop
1746	>	Thread wt = Thread.currentThread();
1747	>	while (ctl == currentCtl) {
1748	>	Thread.interrupted();  // timed variant of version in scan()
1749	>	U.putObject(wt, PARKBLOCKER, this);
1750	>	w.parker = wt;
1751	>	if (ctl == currentCtl)
1752	>	U.park(false, parkTime);
1753	>	w.parker = null;
1754	>	U.putObject(wt, PARKBLOCKER, null);
1755	>	if (ctl != currentCtl)
1756	>	break;
1757	>	if (deadline - System.nanoTime() <= 0L &&
1758	>	U.compareAndSwapLong(this, CTL, currentCtl, prevCtl)) {
1759	>	w.eventCount = (w.eventCount + E_SEQ) | E_MASK;
1760	>	w.qlock = -1;   // shrink
1761	>	w.hint = -1;    // suppress helping
1762	>	break;
1763		}
1764		}
1765		}
1766		}
1767
859	–	// Submissions
860	–
1768		/**
1769	<	* Enqueues the given task in the submissionQueue.  Same idea as
1770	<	* ForkJoinWorkerThread.pushTask except for use of submissionLock.
1771	<	*
1772	<	* @param t the task
1773	<	*/
1774	<	private void addSubmission(ForkJoinTask<?> t) {
1775	<	final ReentrantLock lock = this.submissionLock;
1776	<	lock.lock();
1777	<	try {
1778	<	ForkJoinTask<?>[] q; int s, m;
1779	<	if ((q = submissionQueue) != null) {    // ignore if queue removed
1780	<	long u = (((s = queueTop) & (m = q.length-1)) << ASHIFT)+ABASE;
1781	<	UNSAFE.putOrderedObject(q, u, t);
1782	<	queueTop = s + 1;
1783	<	if (s - queueBase == m)
1784	<	growSubmissionQueue();
1769	>	* Scans through queues looking for work (optionally, while
1770	>	* joining a task); if any are present, signals. May return early
1771	>	* if more signalling is detectably unneeded.
1772	>	*
1773	>	* @param task if non-null, return early if done
1774	>	* @param origin an index to start scan
1775	>	*/
1776	>	final int helpSignal(ForkJoinTask<?> task, int origin) {
1777	>	WorkQueue[] ws; WorkQueue q; int m, n, s, u;
1778	>	if ((ws = workQueues) != null && (m = ws.length - 1) >= 0) {
1779	>	for (int i = 0; i <= m; ++i) {
1780	>	if (task != null && (s = task.status) < 0)
1781	>	return s;
1782	>	if ((q = ws[(i + origin) & m]) != null &&
1783	>	(n = q.queueSize()) > 0) {
1784	>	signalWork(q, n);
1785	>	if ((u = (int)(ctl >>> 32)) >= 0 || (u >> UAC_SHIFT) >= 0)
1786	>	break;
1787	>	}
1788		}
879	–	} finally {
880	–	lock.unlock();
1789		}
1790	<	signalWork();
1790	>	return 0;
1791		}
1792
885	–	//  (pollSubmission is defined below with exported methods)
886	–
1793		/**
1794	<	* Creates or doubles submissionQueue array.
1795	<	* Basically identical to ForkJoinWorkerThread version
1796	<	*/
1797	<	private void growSubmissionQueue() {
1798	<	ForkJoinTask<?>[] oldQ = submissionQueue;
1799	<	int size = oldQ != null ? oldQ.length << 1 : INITIAL_QUEUE_CAPACITY;
1800	<	if (size > MAXIMUM_QUEUE_CAPACITY)
1801	<	throw new RejectedExecutionException("Queue capacity exceeded");
1802	<	if (size < INITIAL_QUEUE_CAPACITY)
1803	<	size = INITIAL_QUEUE_CAPACITY;
1804	<	ForkJoinTask<?>[] q = submissionQueue = new ForkJoinTask<?>[size];
1805	<	int mask = size - 1;
1806	<	int top = queueTop;
1807	<	int oldMask;
1808	<	if (oldQ != null && (oldMask = oldQ.length - 1) >= 0) {
1809	<	for (int b = queueBase; b != top; ++b) {
1810	<	long u = ((b & oldMask) << ASHIFT) + ABASE;
1811	<	Object x = UNSAFE.getObjectVolatile(oldQ, u);
1812	<	if (x != null && UNSAFE.compareAndSwapObject(oldQ, u, x, null))
1813	<	UNSAFE.putObjectVolatile
1814	<	(q, ((b & mask) << ASHIFT) + ABASE, x);
1815	<	}
1816	<	}
1817	<	}
1818	<
1819	<	// Blocking support
1820	<
1821	<	/**
1822	<	* Tries to increment blockedCount, decrement active count
1823	<	* (sometimes implicitly) and possibly release or create a
1824	<	* compensating worker in preparation for blocking. Fails
1825	<	* on contention or termination.
1826	<	*
1827	<	* @return true if the caller can block, else should recheck and retry
1828	<	*/
1829	<	private boolean tryPreBlock() {
1830	<	int b = blockedCount;
1831	<	if (UNSAFE.compareAndSwapInt(this, blockedCountOffset, b, b + 1)) {
1832	<	int pc = parallelism;
1833	<	do {
1834	<	ForkJoinWorkerThread[] ws; ForkJoinWorkerThread w;
1835	<	int e, ac, tc, rc, i;
1836	<	long c = ctl;
1837	<	int u = (int)(c >>> 32);
1838	<	if ((e = (int)c) < 0) {
1839	<	// skip -- terminating
1840	<	}
1841	<	else if ((ac = (u >> UAC_SHIFT)) <= 0 && e != 0 &&
1842	<	(ws = workers) != null &&
1843	<	(i = ~e & SMASK) < ws.length &&
1844	<	(w = ws[i]) != null) {
1845	<	long nc = ((long)(w.nextWait & E_MASK) |
1846	<	(c & (AC_MASK|TC_MASK)));
1847	<	if (w.eventCount == e &&
1848	<	UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
1849	<	w.eventCount = (e + EC_UNIT) & E_MASK;
1850	<	if (w.parked)
1851	<	UNSAFE.unpark(w);
1852	<	return true;             // release an idle worker
1853	<	}
1854	<	}
1855	<	else if ((tc = (short)(u >>> UTC_SHIFT)) >= 0 && ac + pc > 1) {
1856	<	long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK);
1857	<	if (UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc))
1858	<	return true;             // no compensation needed
1859	<	}
1860	<	else if (tc + pc < MAX_ID) {
1861	<	long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK);
1862	<	if (UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
1863	<	addWorker();
1864	<	return true;            // create a replacement
1865	<	}
1866	<	}
1867	<	// try to back out on any failure and let caller retry
1868	<	} while (!UNSAFE.compareAndSwapInt(this, blockedCountOffset,
1869	<	b = blockedCount, b - 1));
1794	>	* Tries to locate and execute tasks for a stealer of the given
1795	>	* task, or in turn one of its stealers, Traces currentSteal ->
1796	>	* currentJoin links looking for a thread working on a descendant
1797	>	* of the given task and with a non-empty queue to steal back and
1798	>	* execute tasks from. The first call to this method upon a
1799	>	* waiting join will often entail scanning/search, (which is OK
1800	>	* because the joiner has nothing better to do), but this method
1801	>	* leaves hints in workers to speed up subsequent calls. The
1802	>	* implementation is very branchy to cope with potential
1803	>	* inconsistencies or loops encountering chains that are stale,
1804	>	* unknown, or so long that they are likely cyclic.
1805	>	*
1806	>	* @param joiner the joining worker
1807	>	* @param task the task to join
1808	>	* @return 0 if no progress can be made, negative if task
1809	>	* known complete, else positive
1810	>	*/
1811	>	private int tryHelpStealer(WorkQueue joiner, ForkJoinTask<?> task) {
1812	>	int stat = 0, steps = 0;                    // bound to avoid cycles
1813	>	if (joiner != null && task != null) {       // hoist null checks
1814	>	restart: for (;;) {
1815	>	ForkJoinTask<?> subtask = task;     // current target
1816	>	for (WorkQueue j = joiner, v;;) {   // v is stealer of subtask
1817	>	WorkQueue[] ws; int m, s, h;
1818	>	if ((s = task.status) < 0) {
1819	>	stat = s;
1820	>	break restart;
1821	>	}
1822	>	if ((ws = workQueues) == null || (m = ws.length - 1) <= 0)
1823	>	break restart;              // shutting down
1824	>	if ((v = ws[h = (j.hint | 1) & m]) == null ||
1825	>	v.currentSteal != subtask) {
1826	>	for (int origin = h;;) {    // find stealer
1827	>	if (((h = (h + 2) & m) & 15) == 1 &&
1828	>	(subtask.status < 0 || j.currentJoin != subtask))
1829	>	continue restart;   // occasional staleness check
1830	>	if ((v = ws[h]) != null &&
1831	>	v.currentSteal == subtask) {
1832	>	j.hint = h;        // save hint
1833	>	break;
1834	>	}
1835	>	if (h == origin)
1836	>	break restart;      // cannot find stealer
1837	>	}
1838	>	}
1839	>	for (;;) { // help stealer or descend to its stealer
1840	>	ForkJoinTask[] a;  int b;
1841	>	if (subtask.status < 0)     // surround probes with
1842	>	continue restart;       //   consistency checks
1843	>	if ((b = v.base) - v.top < 0 && (a = v.array) != null) {
1844	>	int i = (((a.length - 1) & b) << ASHIFT) + ABASE;
1845	>	ForkJoinTask<?> t =
1846	>	(ForkJoinTask<?>)U.getObjectVolatile(a, i);
1847	>	if (subtask.status < 0 || j.currentJoin != subtask ||
1848	>	v.currentSteal != subtask)
1849	>	continue restart;   // stale
1850	>	stat = 1;               // apparent progress
1851	>	if (t != null && v.base == b &&
1852	>	U.compareAndSwapObject(a, i, t, null)) {
1853	>	v.base = b + 1;     // help stealer
1854	>	joiner.runSubtask(t);
1855	>	}
1856	>	else if (v.base == b && ++steps == MAX_HELP)
1857	>	break restart;      // v apparently stalled
1858	>	}
1859	>	else {                      // empty -- try to descend
1860	>	ForkJoinTask<?> next = v.currentJoin;
1861	>	if (subtask.status < 0 || j.currentJoin != subtask ||
1862	>	v.currentSteal != subtask)
1863	>	continue restart;   // stale
1864	>	else if (next == null || ++steps == MAX_HELP)
1865	>	break restart;      // dead-end or maybe cyclic
1866	>	else {
1867	>	subtask = next;
1868	>	j = v;
1869	>	break;
1870	>	}
1871	>	}
1872	>	}
1873	>	}
1874	>	}
1875		}
1876	<	return false;
1876	>	return stat;
1877		}
1878
1879		/**
1880	<	* Decrements blockedCount and increments active count
1881	<	*/
1882	<	private void postBlock() {
1883	<	long c;
1884	<	do {} while (!UNSAFE.compareAndSwapLong(this, ctlOffset,  // no mask
1885	<	c = ctl, c + AC_UNIT));
975	<	int b;
976	<	do {} while(!UNSAFE.compareAndSwapInt(this, blockedCountOffset,
977	<	b = blockedCount, b - 1));
978	<	}
979	<
980	<	/**
981	<	* Possibly blocks waiting for the given task to complete, or
982	<	* cancels the task if terminating.  Fails to wait if contended.
1880	>	* Analog of tryHelpStealer for CountedCompleters. Tries to steal
1881	>	* and run tasks within the target's computation.
1882	>	*
1883	>	* @param task the task to join
1884	>	* @param mode if shared, exit upon completing any task
1885	>	* if all workers are active
1886		*
984	–	* @param joinMe the task
1887		*/
1888	<	final void tryAwaitJoin(ForkJoinTask<?> joinMe) {
1889	<	int s;
1890	<	Thread.interrupted(); // clear interrupts before checking termination
1891	<	if (joinMe.status >= 0) {
1892	<	if (tryPreBlock()) {
1893	<	joinMe.tryAwaitDone(0L);
1894	<	postBlock();
1888	>	private int helpComplete(ForkJoinTask<?> task, int mode) {
1889	>	WorkQueue[] ws; WorkQueue q; int m, n, s, u;
1890	>	if (task != null && (ws = workQueues) != null &&
1891	>	(m = ws.length - 1) >= 0) {
1892	>	for (int j = 1, origin = j;;) {
1893	>	if ((s = task.status) < 0)
1894	>	return s;
1895	>	if ((q = ws[j & m]) != null && q.pollAndExecCC(task)) {
1896	>	origin = j;
1897	>	if (mode == SHARED_QUEUE &&
1898	>	((u = (int)(ctl >>> 32)) >= 0 || (u >> UAC_SHIFT) >= 0))
1899	>	break;
1900	>	}
1901	>	else if ((j = (j + 2) & m) == origin)
1902	>	break;
1903		}
994	–	if ((ctl & STOP_BIT) != 0L)
995	–	joinMe.cancelIgnoringExceptions();
1904		}
1905	+	return 0;
1906		}
1907
1908		/**
1909	<	* Possibly blocks the given worker waiting for joinMe to
1910	<	* complete or timeout
1911	<	*
1912	<	* @param joinMe the task
1913	<	* @param millis the wait time for underlying Object.wait
1914	<	*/
1915	<	final void timedAwaitJoin(ForkJoinTask<?> joinMe, long nanos) {
1916	<	while (joinMe.status >= 0) {
1917	<	Thread.interrupted();
1918	<	if ((ctl & STOP_BIT) != 0L) {
1919	<	joinMe.cancelIgnoringExceptions();
1920	<	break;
1909	>	* Tries to decrement active count (sometimes implicitly) and
1910	>	* possibly release or create a compensating worker in preparation
1911	>	* for blocking. Fails on contention or termination. Otherwise,
1912	>	* adds a new thread if no idle workers are available and pool
1913	>	* may become starved.
1914	>	*/
1915	>	final boolean tryCompensate() {
1916	>	int pc = config & SMASK, e, i, tc; long c;
1917	>	WorkQueue[] ws; WorkQueue w; Thread p;
1918	>	if ((ws = workQueues) != null && (e = (int)(c = ctl)) >= 0) {
1919	>	if (e != 0 && (i = e & SMASK) < ws.length &&
1920	>	(w = ws[i]) != null && w.eventCount == (e | INT_SIGN)) {
1921	>	long nc = ((long)(w.nextWait & E_MASK) |
1922	>	(c & (AC_MASK|TC_MASK)));
1923	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1924	>	w.eventCount = (e + E_SEQ) & E_MASK;
1925	>	if ((p = w.parker) != null)
1926	>	U.unpark(p);
1927	>	return true;   // replace with idle worker
1928	>	}
1929		}
1930	<	if (tryPreBlock()) {
1931	<	long last = System.nanoTime();
1932	<	while (joinMe.status >= 0) {
1933	<	long millis = TimeUnit.NANOSECONDS.toMillis(nanos);
1934	<	if (millis <= 0)
1935	<	break;
1936	<	joinMe.tryAwaitDone(millis);
1937	<	if (joinMe.status < 0)
1938	<	break;
1939	<	if ((ctl & STOP_BIT) != 0L) {
1940	<	joinMe.cancelIgnoringExceptions();
1024	<	break;
1025	<	}
1026	<	long now = System.nanoTime();
1027	<	nanos -= now - last;
1028	<	last = now;
1930	>	else if ((tc = (short)(c >>> TC_SHIFT)) >= 0 &&
1931	>	(int)(c >> AC_SHIFT) + pc > 1) {
1932	>	long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK);
1933	>	if (U.compareAndSwapLong(this, CTL, c, nc))
1934	>	return true;   // no compensation
1935	>	}
1936	>	else if (tc + pc < MAX_CAP) {
1937	>	long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK);
1938	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1939	>	addWorker();
1940	>	return true;
1941		}
1030	–	postBlock();
1031	–	break;
1942		}
1943		}
1944	+	return false;
1945		}
1946
1947		/**
1948	<	* If necessary, compensates for blocker, and blocks
1948	>	* Helps and/or blocks until the given task is done.
1949	>	*
1950	>	* @param joiner the joining worker
1951	>	* @param task the task
1952	>	* @return task status on exit
1953		*/
1954	<	private void awaitBlocker(ManagedBlocker blocker)
1955	<	throws InterruptedException {
1956	<	while (!blocker.isReleasable()) {
1957	<	if (tryPreBlock()) {
1958	<	try {
1959	<	do {} while (!blocker.isReleasable() && !blocker.block());
1960	<	} finally {
1961	<	postBlock();
1954	>	final int awaitJoin(WorkQueue joiner, ForkJoinTask<?> task) {
1955	>	int s = 0;
1956	>	if (joiner != null && task != null && (s = task.status) >= 0) {
1957	>	ForkJoinTask<?> prevJoin = joiner.currentJoin;
1958	>	joiner.currentJoin = task;
1959	>	do {} while ((s = task.status) >= 0 &&
1960	>	joiner.queueSize() > 0 &&
1961	>	joiner.tryRemoveAndExec(task)); // process local tasks
1962	>	if (s >= 0 && (s = task.status) >= 0 &&
1963	>	(s = helpSignal(task, joiner.poolIndex)) >= 0 &&
1964	>	(task instanceof CountedCompleter))
1965	>	s = helpComplete(task, LIFO_QUEUE);
1966	>	int k = 0; // to perform pre-block yield for politeness
1967	>	while (s >= 0 && (s = task.status) >= 0) {
1968	>	if ((joiner.queueSize() > 0 ||           // try helping
1969	>	(s = tryHelpStealer(joiner, task)) == 0) &&
1970	>	(s = task.status) >= 0) {
1971	>	if (k < 3) {
1972	>	if (++k < 3)
1973	>	s = helpSignal(task, joiner.poolIndex);
1974	>	else
1975	>	Thread.yield();
1976	>	}
1977	>	else if (!tryCompensate())
1978	>	k = 0;
1979	>	else {
1980	>	if (task.trySetSignal() && (s = task.status) >= 0) {
1981	>	synchronized (task) {
1982	>	if (task.status >= 0) {
1983	>	try {                // see ForkJoinTask
1984	>	task.wait();     //  for explanation
1985	>	} catch (InterruptedException ie) {
1986	>	}
1987	>	}
1988	>	else
1989	>	task.notifyAll();
1990	>	}
1991	>	}
1992	>	long c;                          // re-activate
1993	>	do {} while (!U.compareAndSwapLong
1994	>	(this, CTL, c = ctl, c + AC_UNIT));
1995	>	}
1996		}
1048	–	break;
1997		}
1998	+	joiner.currentJoin = prevJoin;
1999		}
2000	+	return s;
2001		}
2002
1053	–	// Creating, registering and deregistring workers
1054	–
2003		/**
2004	<	* Tries to create and start a worker; minimally rolls back counts
2005	<	* on failure.
2004	>	* Stripped-down variant of awaitJoin used by timed joins. Tries
2005	>	* to help join only while there is continuous progress. (Caller
2006	>	* will then enter a timed wait.)
2007	>	*
2008	>	* @param joiner the joining worker
2009	>	* @param task the task
2010		*/
2011	<	private void addWorker() {
2012	<	Throwable ex = null;
2013	<	ForkJoinWorkerThread t = null;
2014	<	try {
2015	<	t = factory.newThread(this);
2016	<	} catch (Throwable e) {
2017	<	ex = e;
2018	<	}
2019	<	if (t == null) {  // null or exceptional factory return
2020	<	long c;       // adjust counts
2021	<	do {} while (!UNSAFE.compareAndSwapLong
2022	<	(this, ctlOffset, c = ctl,
2023	<	(((c - AC_UNIT) & AC_MASK) |
2024	<	((c - TC_UNIT) & TC_MASK) |
2025	<	(c & ~(AC_MASK|TC_MASK)))));
2026	<	// Propagate exception if originating from an external caller
2027	<	if (!tryTerminate(false) && ex != null &&
1076	<	!(Thread.currentThread() instanceof ForkJoinWorkerThread))
1077	<	UNSAFE.throwException(ex);
2011	>	final void helpJoinOnce(WorkQueue joiner, ForkJoinTask<?> task) {
2012	>	int s;
2013	>	if (joiner != null && task != null && (s = task.status) >= 0) {
2014	>	ForkJoinTask<?> prevJoin = joiner.currentJoin;
2015	>	joiner.currentJoin = task;
2016	>	do {} while ((s = task.status) >= 0 &&
2017	>	joiner.queueSize() > 0 &&
2018	>	joiner.tryRemoveAndExec(task));
2019	>	if (s >= 0 && (s = task.status) >= 0 &&
2020	>	(s = helpSignal(task, joiner.poolIndex)) >= 0 &&
2021	>	(task instanceof CountedCompleter))
2022	>	s = helpComplete(task, LIFO_QUEUE);
2023	>	if (s >= 0 && joiner.queueSize() == 0) {
2024	>	do {} while (task.status >= 0 &&
2025	>	tryHelpStealer(joiner, task) > 0);
2026	>	}
2027	>	joiner.currentJoin = prevJoin;
2028		}
1079	–	else
1080	–	t.start();
2029		}
2030
2031		/**
2032	<	* Callback from ForkJoinWorkerThread constructor to assign a
2033	<	* public name
2034	<	*/
2035	<	final String nextWorkerName() {
2036	<	for (int n;;) {
2037	<	if (UNSAFE.compareAndSwapInt(this, nextWorkerNumberOffset,
2038	<	n = nextWorkerNumber, ++n))
2039	<	return workerNamePrefix + n;
2032	>	* Returns a (probably) non-empty steal queue, if one is found
2033	>	* during a random, then cyclic scan, else null.  This method must
2034	>	* be retried by caller if, by the time it tries to use the queue,
2035	>	* it is empty.
2036	>	* @param r a (random) seed for scanning
2037	>	*/
2038	>	private WorkQueue findNonEmptyStealQueue(int r) {
2039	>	for (WorkQueue[] ws;;) {
2040	>	int ps = plock, m, n;
2041	>	if ((ws = workQueues) == null || (m = ws.length - 1) < 1)
2042	>	return null;
2043	>	for (int j = (m + 1) << 2; ;) {
2044	>	WorkQueue q = ws[(((r + j) << 1) | 1) & m];
2045	>	if (q != null && (n = q.queueSize()) > 0) {
2046	>	if (n > 1)
2047	>	signalWork(q, 0);
2048	>	return q;
2049	>	}
2050	>	else if (--j < 0) {
2051	>	if (plock == ps)
2052	>	return null;
2053	>	break;
2054	>	}
2055	>	}
2056		}
2057		}
2058
2059		/**
2060	<	* Callback from ForkJoinWorkerThread constructor to
2061	<	* determine its poolIndex and record in workers array.
2062	<	*
2063	<	* @param w the worker
2064	<	* @return the worker's pool index
2065	<	*/
2066	<	final int registerWorker(ForkJoinWorkerThread w) {
2067	<	/*
2068	<	* In the typical case, a new worker acquires the lock, uses
2069	<	* next available index and returns quickly.  Since we should
2070	<	* not block callers (ultimately from signalWork or
2071	<	* tryPreBlock) waiting for the lock needed to do this, we
2072	<	* instead help release other workers while waiting for the
2073	<	* lock.
2074	<	*/
2075	<	for (int g;;) {
2076	<	ForkJoinWorkerThread[] ws;
2077	<	if (((g = scanGuard) & SG_UNIT) == 0 &&
2078	<	UNSAFE.compareAndSwapInt(this, scanGuardOffset,
1115	<	g, g | SG_UNIT)) {
1116	<	int k = nextWorkerIndex;
1117	<	try {
1118	<	if ((ws = workers) != null) { // ignore on shutdown
1119	<	int n = ws.length;
1120	<	if (k < 0 || k >= n || ws[k] != null) {
1121	<	for (k = 0; k < n && ws[k] != null; ++k)
1122	<	;
1123	<	if (k == n)
1124	<	ws = workers = Arrays.copyOf(ws, n << 1);
1125	<	}
1126	<	ws[k] = w;
1127	<	nextWorkerIndex = k + 1;
1128	<	int m = g & SMASK;
1129	<	g = k >= m? ((m << 1) + 1) & SMASK : g + (SG_UNIT<<1);
1130	<	}
1131	<	} finally {
1132	<	scanGuard = g;
2060	>	* Runs tasks until {@code isQuiescent()}. We piggyback on
2061	>	* active count ctl maintenance, but rather than blocking
2062	>	* when tasks cannot be found, we rescan until all others cannot
2063	>	* find tasks either.
2064	>	*/
2065	>	final void helpQuiescePool(WorkQueue w) {
2066	>	for (boolean active = true;;) {
2067	>	ForkJoinTask<?> localTask; // exhaust local queue
2068	>	while ((localTask = w.nextLocalTask()) != null)
2069	>	localTask.doExec();
2070	>	// Similar to loop in scan(), but ignoring submissions
2071	>	WorkQueue q = findNonEmptyStealQueue(w.nextSeed());
2072	>	if (q != null) {
2073	>	ForkJoinTask<?> t; int b;
2074	>	if (!active) {      // re-establish active count
2075	>	long c;
2076	>	active = true;
2077	>	do {} while (!U.compareAndSwapLong
2078	>	(this, CTL, c = ctl, c + AC_UNIT));
2079		}
2080	<	return k;
2080	>	if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null)
2081	>	w.runSubtask(t);
2082		}
2083	<	else if ((ws = workers) != null) { // help release others
2084	<	for (ForkJoinWorkerThread u : ws) {
2085	<	if (u != null && u.queueBase != u.queueTop) {
2086	<	if (tryReleaseWaiter())
2087	<	break;
2088	<	}
2083	>	else {
2084	>	long c;
2085	>	if (active) {       // decrement active count without queuing
2086	>	active = false;
2087	>	do {} while (!U.compareAndSwapLong
2088	>	(this, CTL, c = ctl, c -= AC_UNIT));
2089	>	}
2090	>	else
2091	>	c = ctl;        // re-increment on exit
2092	>	if ((int)(c >> AC_SHIFT) + (config & SMASK) == 0) {
2093	>	do {} while (!U.compareAndSwapLong
2094	>	(this, CTL, c = ctl, c + AC_UNIT));
2095	>	break;
2096		}
2097		}
2098		}
2099		}
2100
2101		/**
2102	<	* Final callback from terminating worker.  Removes record of
2103	<	* worker from array, and adjusts counts. If pool is shutting
2104	<	* down, tries to complete termination.
2105	<	*
2106	<	* @param w the worker
2107	<	*/
2108	<	final void deregisterWorker(ForkJoinWorkerThread w, Throwable ex) {
2109	<	int idx = w.poolIndex;
2110	<	int sc = w.stealCount;
2111	<	int steps = 0;
2112	<	// Remove from array, adjust worker counts and collect steal count.
2113	<	// We can intermix failed removes or adjusts with steal updates
2114	<	do {
1161	<	long s, c;
1162	<	int g;
1163	<	if (steps == 0 && ((g = scanGuard) & SG_UNIT) == 0 &&
1164	<	UNSAFE.compareAndSwapInt(this, scanGuardOffset,
1165	<	g, g |= SG_UNIT)) {
1166	<	ForkJoinWorkerThread[] ws = workers;
1167	<	if (ws != null && idx >= 0 &&
1168	<	idx < ws.length && ws[idx] == w)
1169	<	ws[idx] = null;    // verify
1170	<	nextWorkerIndex = idx;
1171	<	scanGuard = g + SG_UNIT;
1172	<	steps = 1;
1173	<	}
1174	<	if (steps == 1 &&
1175	<	UNSAFE.compareAndSwapLong(this, ctlOffset, c = ctl,
1176	<	(((c - AC_UNIT) & AC_MASK) |
1177	<	((c - TC_UNIT) & TC_MASK) |
1178	<	(c & ~(AC_MASK|TC_MASK)))))
1179	<	steps = 2;
1180	<	if (sc != 0 &&
1181	<	UNSAFE.compareAndSwapLong(this, stealCountOffset,
1182	<	s = stealCount, s + sc))
1183	<	sc = 0;
1184	<	} while (steps != 2 || sc != 0);
1185	<	if (!tryTerminate(false)) {
1186	<	if (ex != null)   // possibly replace if died abnormally
1187	<	signalWork();
1188	<	else
1189	<	tryReleaseWaiter();
2102	>	* Gets and removes a local or stolen task for the given worker.
2103	>	*
2104	>	* @return a task, if available
2105	>	*/
2106	>	final ForkJoinTask<?> nextTaskFor(WorkQueue w) {
2107	>	for (ForkJoinTask<?> t;;) {
2108	>	WorkQueue q; int b;
2109	>	if ((t = w.nextLocalTask()) != null)
2110	>	return t;
2111	>	if ((q = findNonEmptyStealQueue(w.nextSeed())) == null)
2112	>	return null;
2113	>	if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null)
2114	>	return t;
2115		}
2116		}
2117
1193	–	// Shutdown and termination
1194	–
2118		/**
2119	<	* Possibly initiates and/or completes termination.
2119	>	* Returns a cheap heuristic guide for task partitioning when
2120	>	* programmers, frameworks, tools, or languages have little or no
2121	>	* idea about task granularity.  In essence by offering this
2122	>	* method, we ask users only about tradeoffs in overhead vs
2123	>	* expected throughput and its variance, rather than how finely to
2124	>	* partition tasks.
2125	>	*
2126	>	* In a steady state strict (tree-structured) computation, each
2127	>	* thread makes available for stealing enough tasks for other
2128	>	* threads to remain active. Inductively, if all threads play by
2129	>	* the same rules, each thread should make available only a
2130	>	* constant number of tasks.
2131	>	*
2132	>	* The minimum useful constant is just 1. But using a value of 1
2133	>	* would require immediate replenishment upon each steal to
2134	>	* maintain enough tasks, which is infeasible.  Further,
2135	>	* partitionings/granularities of offered tasks should minimize
2136	>	* steal rates, which in general means that threads nearer the top
2137	>	* of computation tree should generate more than those nearer the
2138	>	* bottom. In perfect steady state, each thread is at
2139	>	* approximately the same level of computation tree. However,
2140	>	* producing extra tasks amortizes the uncertainty of progress and
2141	>	* diffusion assumptions.
2142	>	*
2143	>	* So, users will want to use values larger, but not much larger
2144	>	* than 1 to both smooth over transient shortages and hedge
2145	>	* against uneven progress; as traded off against the cost of
2146	>	* extra task overhead. We leave the user to pick a threshold
2147	>	* value to compare with the results of this call to guide
2148	>	* decisions, but recommend values such as 3.
2149	>	*
2150	>	* When all threads are active, it is on average OK to estimate
2151	>	* surplus strictly locally. In steady-state, if one thread is
2152	>	* maintaining say 2 surplus tasks, then so are others. So we can
2153	>	* just use estimated queue length.  However, this strategy alone
2154	>	* leads to serious mis-estimates in some non-steady-state
2155	>	* conditions (ramp-up, ramp-down, other stalls). We can detect
2156	>	* many of these by further considering the number of "idle"
2157	>	* threads, that are known to have zero queued tasks, so
2158	>	* compensate by a factor of (#idle/#active) threads.
2159	>	*
2160	>	* Note: The approximation of #busy workers as #active workers is
2161	>	* not very good under current signalling scheme, and should be
2162	>	* improved.
2163	>	*/
2164	>	static int getSurplusQueuedTaskCount() {
2165	>	Thread t; ForkJoinWorkerThread wt; ForkJoinPool pool; WorkQueue q;
2166	>	if (((t = Thread.currentThread()) instanceof ForkJoinWorkerThread)) {
2167	>	int p = (pool = (wt = (ForkJoinWorkerThread)t).pool).config & SMASK;
2168	>	int n = (q = wt.workQueue).top - q.base;
2169	>	int a = (int)(pool.ctl >> AC_SHIFT) + p;
2170	>	return n - (a > (p >>>= 1) ? 0 :
2171	>	a > (p >>>= 1) ? 1 :
2172	>	a > (p >>>= 1) ? 2 :
2173	>	a > (p >>>= 1) ? 4 :
2174	>	8);
2175	>	}
2176	>	return 0;
2177	>	}
2178	>
2179	>	//  Termination
2180	>
2181	>	/**
2182	>	* Possibly initiates and/or completes termination.  The caller
2183	>	* triggering termination runs three passes through workQueues:
2184	>	* (0) Setting termination status, followed by wakeups of queued
2185	>	* workers; (1) cancelling all tasks; (2) interrupting lagging
2186	>	* threads (likely in external tasks, but possibly also blocked in
2187	>	* joins).  Each pass repeats previous steps because of potential
2188	>	* lagging thread creation.
2189		*
2190		* @param now if true, unconditionally terminate, else only
2191	<	* if shutdown and empty queue and no active workers
2191	>	* if no work and no active workers
2192	>	* @param enable if true, enable shutdown when next possible
2193		* @return true if now terminating or terminated
2194		*/
2195	<	private boolean tryTerminate(boolean now) {
2196	<	long c;
2197	<	while (((c = ctl) & STOP_BIT) == 0) {
2198	<	if (!now) {
2199	<	if ((int)(c >> AC_SHIFT) != -parallelism)
2195	>	private boolean tryTerminate(boolean now, boolean enable) {
2196	>	if (this == commonPool)                     // cannot shut down
2197	>	return false;
2198	>	for (long c;;) {
2199	>	if (((c = ctl) & STOP_BIT) != 0) {      // already terminating
2200	>	if ((short)(c >>> TC_SHIFT) == -(config & SMASK)) {
2201	>	synchronized (this) {
2202	>	notifyAll();                // signal when 0 workers
2203	>	}
2204	>	}
2205	>	return true;
2206	>	}
2207	>	if (plock >= 0) {                       // not yet enabled
2208	>	int ps;
2209	>	if (!enable)
2210		return false;
2211	<	if (!shutdown || blockedCount != 0 || quiescerCount != 0 ||
2212	<	queueTop - queueBase > 0) {
2213	<	if (ctl == c) // staleness check
2214	<	return false;
2215	<	continue;
2216	<	}
2217	<	}
2218	<	if (UNSAFE.compareAndSwapLong(this, ctlOffset, c, c | STOP_BIT))
2219	<	startTerminating();
2220	<	}
2221	<	if ((short)(c >>> TC_SHIFT) == -parallelism) {
2222	<	submissionLock.lock();
2223	<	termination.signalAll();
2224	<	submissionLock.unlock();
2225	<	}
2226	<	return true;
2227	<	}
2228	<
2229	<	/**
2230	<	* Runs up to three passes through workers: (0) Setting
2231	<	* termination status for each worker, followed by wakeups up
2232	<	* queued workers (1) helping cancel tasks (2) interrupting
2233	<	* lagging threads (likely in external tasks, but possibly also
2234	<	* blocked in joins).  Each pass repeats previous steps because of
2235	<	* potential lagging thread creation.
2236	<	*/
2237	<	private void startTerminating() {
2238	<	cancelSubmissions();
2239	<	for (int pass = 0; pass < 3; ++pass) {
2240	<	ForkJoinWorkerThread[] ws = workers;
2241	<	if (ws != null) {
2242	<	for (ForkJoinWorkerThread w : ws) {
2243	<	if (w != null) {
1241	<	w.terminate = true;
1242	<	if (pass > 0) {
1243	<	w.cancelTasks();
1244	<	if (pass > 1 && !w.isInterrupted()) {
1245	<	try {
1246	<	w.interrupt();
1247	<	} catch (SecurityException ignore) {
2211	>	if (((ps = plock) & PL_LOCK) != 0 ||
2212	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
2213	>	ps = acquirePlock();
2214	>	int nps = SHUTDOWN;
2215	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
2216	>	releasePlock(nps);
2217	>	}
2218	>	if (!now) {                             // check if idle & no tasks
2219	>	if ((int)(c >> AC_SHIFT) != -(config & SMASK) ||
2220	>	hasQueuedSubmissions())
2221	>	return false;
2222	>	// Check for unqueued inactive workers. One pass suffices.
2223	>	WorkQueue[] ws = workQueues; WorkQueue w;
2224	>	if (ws != null) {
2225	>	for (int i = 1; i < ws.length; i += 2) {
2226	>	if ((w = ws[i]) != null && w.eventCount >= 0)
2227	>	return false;
2228	>	}
2229	>	}
2230	>	}
2231	>	if (U.compareAndSwapLong(this, CTL, c, c | STOP_BIT)) {
2232	>	for (int pass = 0; pass < 3; ++pass) {
2233	>	WorkQueue[] ws = workQueues;
2234	>	if (ws != null) {
2235	>	WorkQueue w;
2236	>	int n = ws.length;
2237	>	for (int i = 0; i < n; ++i) {
2238	>	if ((w = ws[i]) != null) {
2239	>	w.qlock = -1;
2240	>	if (pass > 0) {
2241	>	w.cancelAll();
2242	>	if (pass > 1)
2243	>	w.interruptOwner();
2244		}
2245		}
2246		}
2247	+	// Wake up workers parked on event queue
2248	+	int i, e; long cc; Thread p;
2249	+	while ((e = (int)(cc = ctl) & E_MASK) != 0 &&
2250	+	(i = e & SMASK) < n &&
2251	+	(w = ws[i]) != null) {
2252	+	long nc = ((long)(w.nextWait & E_MASK) |
2253	+	((cc + AC_UNIT) & AC_MASK) |
2254	+	(cc & (TC_MASK|STOP_BIT)));
2255	+	if (w.eventCount == (e | INT_SIGN) &&
2256	+	U.compareAndSwapLong(this, CTL, cc, nc)) {
2257	+	w.eventCount = (e + E_SEQ) & E_MASK;
2258	+	w.qlock = -1;
2259	+	if ((p = w.parker) != null)
2260	+	U.unpark(p);
2261	+	}
2262	+	}
2263		}
2264		}
1253	–	terminateWaiters();
2265		}
2266		}
2267		}
2268
2269	+	// external operations on common pool
2270	+
2271		/**
2272	<	* Polls and cancels all submissions. Called only during termination.
2272	>	* Returns common pool queue for a thread that has submitted at
2273	>	* least one task.
2274		*/
2275	<	private void cancelSubmissions() {
2276	<	while (queueBase != queueTop) {
2277	<	ForkJoinTask<?> task = pollSubmission();
2278	<	if (task != null) {
2279	<	try {
2280	<	task.cancel(false);
2281	<	} catch (Throwable ignore) {
2282	<	}
2275	>	static WorkQueue commonSubmitterQueue() {
2276	>	ForkJoinPool p; WorkQueue[] ws; int m; Submitter z;
2277	>	return ((z = submitters.get()) != null &&
2278	>	(p = commonPool) != null &&
2279	>	(ws = p.workQueues) != null &&
2280	>	(m = ws.length - 1) >= 0) ?
2281	>	ws[m & z.seed & SQMASK] : null;
2282	>	}
2283	>
2284	>	/**
2285	>	* Tries to pop the given task from submitter's queue in common pool.
2286	>	*/
2287	>	static boolean tryExternalUnpush(ForkJoinTask<?> t) {
2288	>	ForkJoinPool p; WorkQueue[] ws; WorkQueue q; Submitter z;
2289	>	ForkJoinTask<?>[] a;  int m, s; long j;
2290	>	if ((z = submitters.get()) != null &&
2291	>	(p = commonPool) != null &&
2292	>	(ws = p.workQueues) != null &&
2293	>	(m = ws.length - 1) >= 0 &&
2294	>	(q = ws[m & z.seed & SQMASK]) != null &&
2295	>	(s = q.top) != q.base &&
2296	>	(a = q.array) != null &&
2297	>	U.getObjectVolatile
2298	>	(a, j = (((a.length - 1) & (s - 1)) << ASHIFT) + ABASE) == t &&
2299	>	U.compareAndSwapInt(q, QLOCK, 0, 1)) {
2300	>	if (q.array == a && q.top == s && // recheck
2301	>	U.compareAndSwapObject(a, j, t, null)) {
2302	>	q.top = s - 1;
2303	>	q.qlock = 0;
2304	>	return true;
2305		}
2306	+	q.qlock = 0;
2307		}
2308	+	return false;
2309		}
2310
2311		/**
2312	<	* Tries to set the termination status of waiting workers, and
2313	<	* then wake them up (after which they will terminate).
2314	<	*/
2315	<	private void terminateWaiters() {
2316	<	ForkJoinWorkerThread[] ws = workers;
2317	<	if (ws != null) {
2318	<	ForkJoinWorkerThread w; long c; int i, e;
2319	<	int n = ws.length;
2320	<	while ((i = ~(e = (int)(c = ctl)) & SMASK) < n &&
2321	<	(w = ws[i]) != null && w.eventCount == (e & E_MASK)) {
2322	<	if (UNSAFE.compareAndSwapLong(this, ctlOffset, c,
2323	<	(long)(w.nextWait & E_MASK) |
2324	<	((c + AC_UNIT) & AC_MASK) |
2325	<	(c & (TC_MASK|STOP_BIT)))) {
2326	<	w.terminate = true;
2327	<	w.eventCount = e + EC_UNIT;
2328	<	if (w.parked)
2329	<	UNSAFE.unpark(w);
2312	>	* Tries to pop and run local tasks within the same computation
2313	>	* as the given root. On failure, tries to help complete from
2314	>	* other queues via helpComplete.
2315	>	*/
2316	>	private void externalHelpComplete(WorkQueue q, ForkJoinTask<?> root) {
2317	>	ForkJoinTask<?>[] a; int m;
2318	>	if (q != null && (a = q.array) != null && (m = (a.length - 1)) >= 0 &&
2319	>	root != null && root.status >= 0) {
2320	>	for (;;) {
2321	>	int s, u; Object o; CountedCompleter<?> task = null;
2322	>	if ((s = q.top) - q.base > 0) {
2323	>	long j = ((m & (s - 1)) << ASHIFT) + ABASE;
2324	>	if ((o = U.getObject(a, j)) != null &&
2325	>	(o instanceof CountedCompleter)) {
2326	>	CountedCompleter<?> t = (CountedCompleter<?>)o, r = t;
2327	>	do {
2328	>	if (r == root) {
2329	>	if (U.compareAndSwapInt(q, QLOCK, 0, 1)) {
2330	>	if (q.array == a && q.top == s &&
2331	>	U.compareAndSwapObject(a, j, t, null)) {
2332	>	q.top = s - 1;
2333	>	task = t;
2334	>	}
2335	>	q.qlock = 0;
2336	>	}
2337	>	break;
2338	>	}
2339	>	} while ((r = r.completer) != null);
2340	>	}
2341	>	}
2342	>	if (task != null)
2343	>	task.doExec();
2344	>	if (root.status < 0 ||
2345	>	(u = (int)(ctl >>> 32)) >= 0 || (u >> UAC_SHIFT) >= 0)
2346	>	break;
2347	>	if (task == null) {
2348	>	if (helpSignal(root, q.poolIndex) >= 0)
2349	>	helpComplete(root, SHARED_QUEUE);
2350	>	break;
2351		}
2352		}
2353		}
2354		}
2355
1297	–	// misc ForkJoinWorkerThread support
1298	–
2356		/**
2357	<	* Increment or decrement quiescerCount. Needed only to prevent
2358	<	* triggering shutdown if a worker is transiently inactive while
1302	<	* checking quiescence.
1303	<	*
1304	<	* @param delta 1 for increment, -1 for decrement
2357	>	* Tries to help execute or signal availability of the given task
2358	>	* from submitter's queue in common pool.
2359		*/
2360	<	final void addQuiescerCount(int delta) {
2361	<	int c;
2362	<	do {} while(!UNSAFE.compareAndSwapInt(this, quiescerCountOffset,
2363	<	c = quiescerCount, c + delta));
2360	>	static void externalHelpJoin(ForkJoinTask<?> t) {
2361	>	// Some hard-to-avoid overlap with tryExternalUnpush
2362	>	ForkJoinPool p; WorkQueue[] ws; WorkQueue q, w; Submitter z;
2363	>	ForkJoinTask<?>[] a;  int m, s, n; long j;
2364	>	if (t != null &&
2365	>	(z = submitters.get()) != null &&
2366	>	(p = commonPool) != null &&
2367	>	(ws = p.workQueues) != null &&
2368	>	(m = ws.length - 1) >= 0 &&
2369	>	(q = ws[m & z.seed & SQMASK]) != null &&
2370	>	(a = q.array) != null &&
2371	>	t.status >= 0) {
2372	>	if ((s = q.top) != q.base &&
2373	>	U.getObjectVolatile
2374	>	(a, j = (((a.length - 1) & (s - 1)) << ASHIFT) + ABASE) == t &&
2375	>	U.compareAndSwapInt(q, QLOCK, 0, 1)) {
2376	>	if (q.array == a && q.top == s &&
2377	>	U.compareAndSwapObject(a, j, t, null)) {
2378	>	q.top = s - 1;
2379	>	q.qlock = 0;
2380	>	t.doExec();
2381	>	}
2382	>	else
2383	>	q.qlock = 0;
2384	>	}
2385	>	if (t.status >= 0) {
2386	>	if (t instanceof CountedCompleter)
2387	>	p.externalHelpComplete(q, t);
2388	>	else
2389	>	p.helpSignal(t, q.poolIndex);
2390	>	}
2391	>	}
2392		}
2393
2394		/**
2395	<	* Directly increment or decrement active count without
1314	<	* queuing. This method is used to transiently assert inactivation
1315	<	* while checking quiescence.
1316	<	*
1317	<	* @param delta 1 for increment, -1 for decrement
2395	>	* Restricted version of helpQuiescePool for external callers
2396		*/
2397	<	final void addActiveCount(int delta) {
2398	<	long d = delta < 0 ? -AC_UNIT : AC_UNIT;
2399	<	long c;
2400	<	do {} while (!UNSAFE.compareAndSwapLong(this, ctlOffset, c = ctl,
2401	<	((c + d) & AC_MASK) |
2402	<	(c & ~AC_MASK)));
2403	<	}
1326	<
1327	<	/**
1328	<	* Returns the approximate (non-atomic) number of idle threads per
1329	<	* active thread.
1330	<	*/
1331	<	final int idlePerActive() {
1332	<	// Approximate at powers of two for small values, saturate past 4
1333	<	int p = parallelism;
1334	<	int a = p + (int)(ctl >> AC_SHIFT);
1335	<	return (a > (p >>>= 1) ? 0 :
1336	<	a > (p >>>= 1) ? 1 :
1337	<	a > (p >>>= 1) ? 2 :
1338	<	a > (p >>>= 1) ? 4 :
1339	<	8);
2397	>	static void externalHelpQuiescePool() {
2398	>	ForkJoinPool p; ForkJoinTask<?> t; WorkQueue q; int b;
2399	>	if ((p = commonPool) != null &&
2400	>	(q = p.findNonEmptyStealQueue(1)) != null &&
2401	>	(b = q.base) - q.top < 0 &&
2402	>	(t = q.pollAt(b)) != null)
2403	>	t.doExec();
2404		}
2405
2406		// Exported methods
#	Line 1408 | Line 2472 | public class ForkJoinPool extends Abstra
2472		checkPermission();
2473		if (factory == null)
2474		throw new NullPointerException();
2475	<	if (parallelism <= 0 || parallelism > MAX_ID)
2475	>	if (parallelism <= 0 || parallelism > MAX_CAP)
2476		throw new IllegalArgumentException();
1413	–	this.parallelism = parallelism;
2477		this.factory = factory;
2478		this.ueh = handler;
2479	<	this.locallyFifo = asyncMode;
2479	>	this.config = parallelism | (asyncMode ? (FIFO_QUEUE << 16) : 0);
2480		long np = (long)(-parallelism); // offset ctl counts
2481		this.ctl = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK);
2482	<	this.submissionQueue = new ForkJoinTask<?>[INITIAL_QUEUE_CAPACITY];
1420	<	// initialize workers array with room for 2*parallelism if possible
1421	<	int n = parallelism << 1;
1422	<	if (n >= MAX_ID)
1423	<	n = MAX_ID;
1424	<	else { // See Hackers Delight, sec 3.2, where n < (1 << 16)
1425	<	n |= n >>> 1; n |= n >>> 2; n |= n >>> 4; n |= n >>> 8;
1426	<	}
1427	<	workers = new ForkJoinWorkerThread[n + 1];
1428	<	this.submissionLock = new ReentrantLock();
1429	<	this.termination = submissionLock.newCondition();
2482	>	int pn = nextPoolId();
2483		StringBuilder sb = new StringBuilder("ForkJoinPool-");
2484	<	sb.append(poolNumberGenerator.incrementAndGet());
2484	>	sb.append(Integer.toString(pn));
2485		sb.append("-worker-");
2486		this.workerNamePrefix = sb.toString();
2487		}
2488
2489	+	/**
2490	+	* Constructor for common pool, suitable only for static initialization.
2491	+	* Basically the same as above, but uses smallest possible initial footprint.
2492	+	*/
2493	+	ForkJoinPool(int parallelism, long ctl,
2494	+	ForkJoinWorkerThreadFactory factory,
2495	+	Thread.UncaughtExceptionHandler handler) {
2496	+	this.config = parallelism;
2497	+	this.ctl = ctl;
2498	+	this.factory = factory;
2499	+	this.ueh = handler;
2500	+	this.workerNamePrefix = "ForkJoinPool.commonPool-worker-";
2501	+	}
2502	+
2503	+	/**
2504	+	* Returns the common pool instance.
2505	+	*
2506	+	* @return the common pool instance
2507	+	*/
2508	+	public static ForkJoinPool commonPool() {
2509	+	// assert commonPool != null : "static init error";
2510	+	return commonPool;
2511	+	}
2512	+
2513		// Execution methods
2514
2515		/**
#	Line 1452 | Line 2529 | public class ForkJoinPool extends Abstra
2529		*         scheduled for execution
2530		*/
2531		public <T> T invoke(ForkJoinTask<T> task) {
1455	–	Thread t = Thread.currentThread();
2532		if (task == null)
2533		throw new NullPointerException();
2534	<	if (shutdown)
2535	<	throw new RejectedExecutionException();
1460	<	if ((t instanceof ForkJoinWorkerThread) &&
1461	<	((ForkJoinWorkerThread)t).pool == this)
1462	<	return task.invoke();  // bypass submit if in same pool
1463	<	else {
1464	<	addSubmission(task);
1465	<	return task.join();
1466	<	}
1467	<	}
1468	<
1469	<	/**
1470	<	* Unless terminating, forks task if within an ongoing FJ
1471	<	* computation in the current pool, else submits as external task.
1472	<	*/
1473	<	private <T> void forkOrSubmit(ForkJoinTask<T> task) {
1474	<	ForkJoinWorkerThread w;
1475	<	Thread t = Thread.currentThread();
1476	<	if (shutdown)
1477	<	throw new RejectedExecutionException();
1478	<	if ((t instanceof ForkJoinWorkerThread) &&
1479	<	(w = (ForkJoinWorkerThread)t).pool == this)
1480	<	w.pushTask(task);
1481	<	else
1482	<	addSubmission(task);
2534	>	externalPush(task);
2535	>	return task.join();
2536		}
2537
2538		/**
#	Line 1493 | Line 2546 | public class ForkJoinPool extends Abstra
2546		public void execute(ForkJoinTask<?> task) {
2547		if (task == null)
2548		throw new NullPointerException();
2549	<	forkOrSubmit(task);
2549	>	externalPush(task);
2550		}
2551
2552		// AbstractExecutorService methods
#	Line 1510 | Line 2563 | public class ForkJoinPool extends Abstra
2563		if (task instanceof ForkJoinTask<?>) // avoid re-wrap
2564		job = (ForkJoinTask<?>) task;
2565		else
2566	<	job = ForkJoinTask.adapt(task, null);
2567	<	forkOrSubmit(job);
2566	>	job = new ForkJoinTask.AdaptedRunnableAction(task);
2567	>	externalPush(job);
2568		}
2569
2570		/**
#	Line 1526 | Line 2579 | public class ForkJoinPool extends Abstra
2579		public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) {
2580		if (task == null)
2581		throw new NullPointerException();
2582	<	forkOrSubmit(task);
2582	>	externalPush(task);
2583		return task;
2584		}
2585
#	Line 1536 | Line 2589 | public class ForkJoinPool extends Abstra
2589		*         scheduled for execution
2590		*/
2591		public <T> ForkJoinTask<T> submit(Callable<T> task) {
2592	<	if (task == null)
2593	<	throw new NullPointerException();
1541	<	ForkJoinTask<T> job = ForkJoinTask.adapt(task);
1542	<	forkOrSubmit(job);
2592	>	ForkJoinTask<T> job = new ForkJoinTask.AdaptedCallable<T>(task);
2593	>	externalPush(job);
2594		return job;
2595		}
2596
#	Line 1549 | Line 2600 | public class ForkJoinPool extends Abstra
2600		*         scheduled for execution
2601		*/
2602		public <T> ForkJoinTask<T> submit(Runnable task, T result) {
2603	<	if (task == null)
2604	<	throw new NullPointerException();
1554	<	ForkJoinTask<T> job = ForkJoinTask.adapt(task, result);
1555	<	forkOrSubmit(job);
2603	>	ForkJoinTask<T> job = new ForkJoinTask.AdaptedRunnable<T>(task, result);
2604	>	externalPush(job);
2605		return job;
2606		}
2607
#	Line 1568 | Line 2617 | public class ForkJoinPool extends Abstra
2617		if (task instanceof ForkJoinTask<?>) // avoid re-wrap
2618		job = (ForkJoinTask<?>) task;
2619		else
2620	<	job = ForkJoinTask.adapt(task, null);
2621	<	forkOrSubmit(job);
2620	>	job = new ForkJoinTask.AdaptedRunnableAction(task);
2621	>	externalPush(job);
2622		return job;
2623		}
2624
#	Line 1578 | Line 2627 | public class ForkJoinPool extends Abstra
2627		* @throws RejectedExecutionException {@inheritDoc}
2628		*/
2629		public <T> List<Future<T>> invokeAll(Collection<? extends Callable<T>> tasks) {
2630	<	ArrayList<ForkJoinTask<T>> forkJoinTasks =
2631	<	new ArrayList<ForkJoinTask<T>>(tasks.size());
2632	<	for (Callable<T> task : tasks)
2633	<	forkJoinTasks.add(ForkJoinTask.adapt(task));
2634	<	invoke(new InvokeAll<T>(forkJoinTasks));
2635	<
2630	>	// In previous versions of this class, this method constructed
2631	>	// a task to run ForkJoinTask.invokeAll, but now external
2632	>	// invocation of multiple tasks is at least as efficient.
2633	>	List<ForkJoinTask<T>> fs = new ArrayList<ForkJoinTask<T>>(tasks.size());
2634	>	// Workaround needed because method wasn't declared with
2635	>	// wildcards in return type but should have been.
2636		@SuppressWarnings({"unchecked", "rawtypes"})
2637	<	List<Future<T>> futures = (List<Future<T>>) (List) forkJoinTasks;
1589	<	return futures;
1590	<	}
2637	>	List<Future<T>> futures = (List<Future<T>>) (List) fs;
2638
2639	<	static final class InvokeAll<T> extends RecursiveAction {
2640	<	final ArrayList<ForkJoinTask<T>> tasks;
2641	<	InvokeAll(ArrayList<ForkJoinTask<T>> tasks) { this.tasks = tasks; }
2642	<	public void compute() {
2643	<	try { invokeAll(tasks); }
2644	<	catch (Exception ignore) {}
2639	>	boolean done = false;
2640	>	try {
2641	>	for (Callable<T> t : tasks) {
2642	>	ForkJoinTask<T> f = new ForkJoinTask.AdaptedCallable<T>(t);
2643	>	externalPush(f);
2644	>	fs.add(f);
2645	>	}
2646	>	for (ForkJoinTask<T> f : fs)
2647	>	f.quietlyJoin();
2648	>	done = true;
2649	>	return futures;
2650	>	} finally {
2651	>	if (!done)
2652	>	for (ForkJoinTask<T> f : fs)
2653	>	f.cancel(false);
2654		}
1599	–	private static final long serialVersionUID = -7914297376763021607L;
2655		}
2656
2657		/**
#	Line 1624 | Line 2679 | public class ForkJoinPool extends Abstra
2679		* @return the targeted parallelism level of this pool
2680		*/
2681		public int getParallelism() {
2682	<	return parallelism;
2682	>	return config & SMASK;
2683	>	}
2684	>
2685	>	/**
2686	>	* Returns the targeted parallelism level of the common pool.
2687	>	*
2688	>	* @return the targeted parallelism level of the common pool
2689	>	*/
2690	>	public static int getCommonPoolParallelism() {
2691	>	return commonPoolParallelism;
2692		}
2693
2694		/**
#	Line 1636 | Line 2700 | public class ForkJoinPool extends Abstra
2700		* @return the number of worker threads
2701		*/
2702		public int getPoolSize() {
2703	<	return parallelism + (short)(ctl >>> TC_SHIFT);
2703	>	return (config & SMASK) + (short)(ctl >>> TC_SHIFT);
2704		}
2705
2706		/**
#	Line 1646 | Line 2710 | public class ForkJoinPool extends Abstra
2710		* @return {@code true} if this pool uses async mode
2711		*/
2712		public boolean getAsyncMode() {
2713	<	return locallyFifo;
2713	>	return (config >>> 16) == FIFO_QUEUE;
2714		}
2715
2716		/**
#	Line 1658 | Line 2722 | public class ForkJoinPool extends Abstra
2722		* @return the number of worker threads
2723		*/
2724		public int getRunningThreadCount() {
2725	<	int r = parallelism + (int)(ctl >> AC_SHIFT);
2726	<	return r <= 0? 0 : r; // suppress momentarily negative values
2725	>	int rc = 0;
2726	>	WorkQueue[] ws; WorkQueue w;
2727	>	if ((ws = workQueues) != null) {
2728	>	for (int i = 1; i < ws.length; i += 2) {
2729	>	if ((w = ws[i]) != null && w.isApparentlyUnblocked())
2730	>	++rc;
2731	>	}
2732	>	}
2733	>	return rc;
2734		}
2735
2736		/**
#	Line 1670 | Line 2741 | public class ForkJoinPool extends Abstra
2741		* @return the number of active threads
2742		*/
2743		public int getActiveThreadCount() {
2744	<	int r = parallelism + (int)(ctl >> AC_SHIFT) + blockedCount;
2745	<	return r <= 0? 0 : r; // suppress momentarily negative values
2744	>	int r = (config & SMASK) + (int)(ctl >> AC_SHIFT);
2745	>	return (r <= 0) ? 0 : r; // suppress momentarily negative values
2746		}
2747
2748		/**
#	Line 1686 | Line 2757 | public class ForkJoinPool extends Abstra
2757		* @return {@code true} if all threads are currently idle
2758		*/
2759		public boolean isQuiescent() {
2760	<	return parallelism + (int)(ctl >> AC_SHIFT) + blockedCount == 0;
2760	>	return (int)(ctl >> AC_SHIFT) + (config & SMASK) == 0;
2761		}
2762
2763		/**
#	Line 1701 | Line 2772 | public class ForkJoinPool extends Abstra
2772		* @return the number of steals
2773		*/
2774		public long getStealCount() {
2775	<	return stealCount;
2775	>	long count = stealCount;
2776	>	WorkQueue[] ws; WorkQueue w;
2777	>	if ((ws = workQueues) != null) {
2778	>	for (int i = 1; i < ws.length; i += 2) {
2779	>	if ((w = ws[i]) != null)
2780	>	count += w.nsteals;
2781	>	}
2782	>	}
2783	>	return count;
2784		}
2785
2786		/**
#	Line 1716 | Line 2795 | public class ForkJoinPool extends Abstra
2795		*/
2796		public long getQueuedTaskCount() {
2797		long count = 0;
2798	<	ForkJoinWorkerThread[] ws;
2799	<	if ((short)(ctl >>> TC_SHIFT) > -parallelism &&
2800	<	(ws = workers) != null) {
2801	<	for (ForkJoinWorkerThread w : ws)
2802	<	if (w != null)
2803	<	count -= w.queueBase - w.queueTop; // must read base first
2798	>	WorkQueue[] ws; WorkQueue w;
2799	>	if ((ws = workQueues) != null) {
2800	>	for (int i = 1; i < ws.length; i += 2) {
2801	>	if ((w = ws[i]) != null)
2802	>	count += w.queueSize();
2803	>	}
2804		}
2805		return count;
2806		}
2807
2808		/**
2809		* Returns an estimate of the number of tasks submitted to this
2810	<	* pool that have not yet begun executing.  This meThod may take
2810	>	* pool that have not yet begun executing.  This method may take
2811		* time proportional to the number of submissions.
2812		*
2813		* @return the number of queued submissions
2814		*/
2815		public int getQueuedSubmissionCount() {
2816	<	return -queueBase + queueTop;
2816	>	int count = 0;
2817	>	WorkQueue[] ws; WorkQueue w;
2818	>	if ((ws = workQueues) != null) {
2819	>	for (int i = 0; i < ws.length; i += 2) {
2820	>	if ((w = ws[i]) != null)
2821	>	count += w.queueSize();
2822	>	}
2823	>	}
2824	>	return count;
2825		}
2826
2827		/**
#	Line 1744 | Line 2831 | public class ForkJoinPool extends Abstra
2831		* @return {@code true} if there are any queued submissions
2832		*/
2833		public boolean hasQueuedSubmissions() {
2834	<	return queueBase != queueTop;
2834	>	WorkQueue[] ws; WorkQueue w;
2835	>	if ((ws = workQueues) != null) {
2836	>	for (int i = 0; i < ws.length; i += 2) {
2837	>	if ((w = ws[i]) != null && w.queueSize() != 0)
2838	>	return true;
2839	>	}
2840	>	}
2841	>	return false;
2842		}
2843
2844		/**
#	Line 1755 | Line 2849 | public class ForkJoinPool extends Abstra
2849		* @return the next submission, or {@code null} if none
2850		*/
2851		protected ForkJoinTask<?> pollSubmission() {
2852	<	ForkJoinTask<?> t; ForkJoinTask<?>[] q; int b, i;
2853	<	while ((b = queueBase) != queueTop &&
2854	<	(q = submissionQueue) != null &&
2855	<	(i = (q.length - 1) & b) >= 0) {
2856	<	long u = (i << ASHIFT) + ABASE;
1763	<	if ((t = q[i]) != null &&
1764	<	queueBase == b &&
1765	<	UNSAFE.compareAndSwapObject(q, u, t, null)) {
1766	<	queueBase = b + 1;
1767	<	return t;
2852	>	WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2853	>	if ((ws = workQueues) != null) {
2854	>	for (int i = 0; i < ws.length; i += 2) {
2855	>	if ((w = ws[i]) != null && (t = w.poll()) != null)
2856	>	return t;
2857		}
2858		}
2859		return null;
#	Line 1789 | Line 2878 | public class ForkJoinPool extends Abstra
2878		*/
2879		protected int drainTasksTo(Collection<? super ForkJoinTask<?>> c) {
2880		int count = 0;
2881	<	while (queueBase != queueTop) {
2882	<	ForkJoinTask<?> t = pollSubmission();
2883	<	if (t != null) {
2884	<	c.add(t);
2885	<	++count;
2881	>	WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2882	>	if ((ws = workQueues) != null) {
2883	>	for (int i = 0; i < ws.length; ++i) {
2884	>	if ((w = ws[i]) != null) {
2885	>	while ((t = w.poll()) != null) {
2886	>	c.add(t);
2887	>	++count;
2888	>	}
2889	>	}
2890		}
2891		}
1799	–	ForkJoinWorkerThread[] ws;
1800	–	if ((short)(ctl >>> TC_SHIFT) > -parallelism &&
1801	–	(ws = workers) != null) {
1802	–	for (ForkJoinWorkerThread w : ws)
1803	–	if (w != null)
1804	–	count += w.drainTasksTo(c);
1805	–	}
2892		return count;
2893		}
2894
#	Line 1814 | Line 2900 | public class ForkJoinPool extends Abstra
2900		* @return a string identifying this pool, as well as its state
2901		*/
2902		public String toString() {
2903	<	long st = getStealCount();
2904	<	long qt = getQueuedTaskCount();
2905	<	long qs = getQueuedSubmissionCount();
1820	<	int pc = parallelism;
2903	>	// Use a single pass through workQueues to collect counts
2904	>	long qt = 0L, qs = 0L; int rc = 0;
2905	>	long st = stealCount;
2906		long c = ctl;
2907	+	WorkQueue[] ws; WorkQueue w;
2908	+	if ((ws = workQueues) != null) {
2909	+	for (int i = 0; i < ws.length; ++i) {
2910	+	if ((w = ws[i]) != null) {
2911	+	int size = w.queueSize();
2912	+	if ((i & 1) == 0)
2913	+	qs += size;
2914	+	else {
2915	+	qt += size;
2916	+	st += w.nsteals;
2917	+	if (w.isApparentlyUnblocked())
2918	+	++rc;
2919	+	}
2920	+	}
2921	+	}
2922	+	}
2923	+	int pc = (config & SMASK);
2924		int tc = pc + (short)(c >>> TC_SHIFT);
2925	<	int rc = pc + (int)(c >> AC_SHIFT);
2926	<	if (rc < 0) // ignore transient negative
2927	<	rc = 0;
1826	<	int ac = rc + blockedCount;
2925	>	int ac = pc + (int)(c >> AC_SHIFT);
2926	>	if (ac < 0) // ignore transient negative
2927	>	ac = 0;
2928		String level;
2929		if ((c & STOP_BIT) != 0)
2930	<	level = (tc == 0)? "Terminated" : "Terminating";
2930	>	level = (tc == 0) ? "Terminated" : "Terminating";
2931		else
2932	<	level = shutdown? "Shutting down" : "Running";
2932	>	level = plock < 0 ? "Shutting down" : "Running";
2933		return super.toString() +
2934		"[" + level +
2935		", parallelism = " + pc +
#	Line 1842 | Line 2943 | public class ForkJoinPool extends Abstra
2943		}
2944
2945		/**
2946	<	* Initiates an orderly shutdown in which previously submitted
2947	<	* tasks are executed, but no new tasks will be accepted.
2948	<	* Invocation has no additional effect if already shut down.
2949	<	* Tasks that are in the process of being submitted concurrently
2950	<	* during the course of this method may or may not be rejected.
2946	>	* Possibly initiates an orderly shutdown in which previously
2947	>	* submitted tasks are executed, but no new tasks will be
2948	>	* accepted. Invocation has no effect on execution state if this
2949	>	* is the {@link #commonPool}, and no additional effect if
2950	>	* already shut down.  Tasks that are in the process of being
2951	>	* submitted concurrently during the course of this method may or
2952	>	* may not be rejected.
2953		*
2954		* @throws SecurityException if a security manager exists and
2955		*         the caller is not permitted to modify threads
#	Line 1855 | Line 2958 | public class ForkJoinPool extends Abstra
2958		*/
2959		public void shutdown() {
2960		checkPermission();
2961	<	shutdown = true;
1859	<	tryTerminate(false);
2961	>	tryTerminate(false, true);
2962		}
2963
2964		/**
2965	<	* Attempts to cancel and/or stop all tasks, and reject all
2966	<	* subsequently submitted tasks.  Tasks that are in the process of
2967	<	* being submitted or executed concurrently during the course of
2968	<	* this method may or may not be rejected. This method cancels
2969	<	* both existing and unexecuted tasks, in order to permit
2970	<	* termination in the presence of task dependencies. So the method
2971	<	* always returns an empty list (unlike the case for some other
2972	<	* Executors).
2965	>	* Possibly attempts to cancel and/or stop all tasks, and reject
2966	>	* all subsequently submitted tasks.  Invocation has no effect on
2967	>	* execution state if this is the {@link #commonPool}, and no
2968	>	* additional effect if already shut down. Otherwise, tasks that
2969	>	* are in the process of being submitted or executed concurrently
2970	>	* during the course of this method may or may not be
2971	>	* rejected. This method cancels both existing and unexecuted
2972	>	* tasks, in order to permit termination in the presence of task
2973	>	* dependencies. So the method always returns an empty list
2974	>	* (unlike the case for some other Executors).
2975		*
2976		* @return an empty list
2977		* @throws SecurityException if a security manager exists and
#	Line 1877 | Line 2981 | public class ForkJoinPool extends Abstra
2981		*/
2982		public List<Runnable> shutdownNow() {
2983		checkPermission();
2984	<	shutdown = true;
1881	<	tryTerminate(true);
2984	>	tryTerminate(true, true);
2985		return Collections.emptyList();
2986		}
2987
#	Line 1890 | Line 2993 | public class ForkJoinPool extends Abstra
2993		public boolean isTerminated() {
2994		long c = ctl;
2995		return ((c & STOP_BIT) != 0L &&
2996	<	(short)(c >>> TC_SHIFT) == -parallelism);
2996	>	(short)(c >>> TC_SHIFT) == -(config & SMASK));
2997		}
2998
2999		/**
#	Line 1909 | Line 3012 | public class ForkJoinPool extends Abstra
3012		public boolean isTerminating() {
3013		long c = ctl;
3014		return ((c & STOP_BIT) != 0L &&
3015	<	(short)(c >>> TC_SHIFT) != -parallelism);
1913	<	}
1914	<
1915	<	/**
1916	<	* Returns true if terminating or terminated. Used by ForkJoinWorkerThread.
1917	<	*/
1918	<	final boolean isAtLeastTerminating() {
1919	<	return (ctl & STOP_BIT) != 0L;
3015	>	(short)(c >>> TC_SHIFT) != -(config & SMASK));
3016		}
3017
3018		/**
#	Line 1925 | Line 3021 | public class ForkJoinPool extends Abstra
3021		* @return {@code true} if this pool has been shut down
3022		*/
3023		public boolean isShutdown() {
3024	<	return shutdown;
3024	>	return plock < 0;
3025		}
3026
3027		/**
3028	<	* Blocks until all tasks have completed execution after a shutdown
3029	<	* request, or the timeout occurs, or the current thread is
3030	<	* interrupted, whichever happens first.
3028	>	* Blocks until all tasks have completed execution after a
3029	>	* shutdown request, or the timeout occurs, or the current thread
3030	>	* is interrupted, whichever happens first. Note that the {@link
3031	>	* #commonPool()} never terminates until program shutdown so
3032	>	* this method will always time out.
3033		*
3034		* @param timeout the maximum time to wait
3035		* @param unit the time unit of the timeout argument
#	Line 1942 | Line 3040 | public class ForkJoinPool extends Abstra
3040		public boolean awaitTermination(long timeout, TimeUnit unit)
3041		throws InterruptedException {
3042		long nanos = unit.toNanos(timeout);
3043	<	final ReentrantLock lock = this.submissionLock;
3044	<	lock.lock();
3045	<	try {
3046	<	for (;;) {
3047	<	if (isTerminated())
3048	<	return true;
3049	<	if (nanos <= 0)
3050	<	return false;
3051	<	nanos = termination.awaitNanos(nanos);
3043	>	if (isTerminated())
3044	>	return true;
3045	>	long startTime = System.nanoTime();
3046	>	boolean terminated = false;
3047	>	synchronized (this) {
3048	>	for (long waitTime = nanos, millis = 0L;;) {
3049	>	if (terminated = isTerminated() ||
3050	>	waitTime <= 0L ||
3051	>	(millis = unit.toMillis(waitTime)) <= 0L)
3052	>	break;
3053	>	wait(millis);
3054	>	waitTime = nanos - (System.nanoTime() - startTime);
3055		}
1955	–	} finally {
1956	–	lock.unlock();
3056		}
3057	+	return terminated;
3058		}
3059
3060		/**
#	Line 1965 | Line 3065 | public class ForkJoinPool extends Abstra
3065		* {@code isReleasable} must return {@code true} if blocking is
3066		* not necessary. Method {@code block} blocks the current thread
3067		* if necessary (perhaps internally invoking {@code isReleasable}
3068	<	* before actually blocking). The unusual methods in this API
3069	<	* accommodate synchronizers that may, but don't usually, block
3070	<	* for long periods. Similarly, they allow more efficient internal
3071	<	* handling of cases in which additional workers may be, but
3072	<	* usually are not, needed to ensure sufficient parallelism.
3073	<	* Toward this end, implementations of method {@code isReleasable}
3074	<	* must be amenable to repeated invocation.
3068	>	* before actually blocking). These actions are performed by any
3069	>	* thread invoking {@link ForkJoinPool#managedBlock}.  The
3070	>	* unusual methods in this API accommodate synchronizers that may,
3071	>	* but don't usually, block for long periods. Similarly, they
3072	>	* allow more efficient internal handling of cases in which
3073	>	* additional workers may be, but usually are not, needed to
3074	>	* ensure sufficient parallelism.  Toward this end,
3075	>	* implementations of method {@code isReleasable} must be amenable
3076	>	* to repeated invocation.
3077		*
3078		* <p>For example, here is a ManagedBlocker based on a
3079		* ReentrantLock:
#	Line 2052 | Line 3154 | public class ForkJoinPool extends Abstra
3154		throws InterruptedException {
3155		Thread t = Thread.currentThread();
3156		if (t instanceof ForkJoinWorkerThread) {
3157	<	ForkJoinWorkerThread w = (ForkJoinWorkerThread) t;
3158	<	w.pool.awaitBlocker(blocker);
3157	>	ForkJoinPool p = ((ForkJoinWorkerThread)t).pool;
3158	>	while (!blocker.isReleasable()) { // variant of helpSignal
3159	>	WorkQueue[] ws; WorkQueue q; int m, n, u;
3160	>	if ((ws = p.workQueues) != null && (m = ws.length - 1) >= 0) {
3161	>	for (int i = 0; i <= m; ++i) {
3162	>	if (blocker.isReleasable())
3163	>	return;
3164	>	if ((q = ws[i]) != null && (n = q.queueSize()) > 0) {
3165	>	p.signalWork(q, n);
3166	>	if ((u = (int)(p.ctl >>> 32)) >= 0 ||
3167	>	(u >> UAC_SHIFT) >= 0)
3168	>	break;
3169	>	}
3170	>	}
3171	>	}
3172	>	if (p.tryCompensate()) {
3173	>	try {
3174	>	do {} while (!blocker.isReleasable() &&
3175	>	!blocker.block());
3176	>	} finally {
3177	>	p.incrementActiveCount();
3178	>	}
3179	>	break;
3180	>	}
3181	>	}
3182		}
3183		else {
3184	<	do {} while (!blocker.isReleasable() && !blocker.block());
3184	>	do {} while (!blocker.isReleasable() &&
3185	>	!blocker.block());
3186		}
3187		}
3188
#	Line 2065 | Line 3191 | public class ForkJoinPool extends Abstra
3191		// implement RunnableFuture.
3192
3193		protected <T> RunnableFuture<T> newTaskFor(Runnable runnable, T value) {
3194	<	return (RunnableFuture<T>) ForkJoinTask.adapt(runnable, value);
3194	>	return new ForkJoinTask.AdaptedRunnable<T>(runnable, value);
3195		}
3196
3197		protected <T> RunnableFuture<T> newTaskFor(Callable<T> callable) {
3198	<	return (RunnableFuture<T>) ForkJoinTask.adapt(callable);
3198	>	return new ForkJoinTask.AdaptedCallable<T>(callable);
3199		}
3200
3201		// Unsafe mechanics
3202	<	private static final sun.misc.Unsafe UNSAFE;
3203	<	private static final long ctlOffset;
3204	<	private static final long stealCountOffset;
3205	<	private static final long blockedCountOffset;
2080	<	private static final long quiescerCountOffset;
2081	<	private static final long scanGuardOffset;
2082	<	private static final long nextWorkerNumberOffset;
2083	<	private static final long ABASE;
3202	>	private static final sun.misc.Unsafe U;
3203	>	private static final long CTL;
3204	>	private static final long PARKBLOCKER;
3205	>	private static final int ABASE;
3206		private static final int ASHIFT;
3207	+	private static final long STEALCOUNT;
3208	+	private static final long PLOCK;
3209	+	private static final long INDEXSEED;
3210	+	private static final long QLOCK;
3211
3212		static {
3213	<	poolNumberGenerator = new AtomicInteger();
2088	<	workerSeedGenerator = new Random();
2089	<	modifyThreadPermission = new RuntimePermission("modifyThread");
2090	<	defaultForkJoinWorkerThreadFactory =
2091	<	new DefaultForkJoinWorkerThreadFactory();
2092	<	int s;
3213	>	int s; // initialize field offsets for CAS etc
3214		try {
3215	<	UNSAFE = getUnsafe();
3216	<	Class k = ForkJoinPool.class;
3217	<	ctlOffset = UNSAFE.objectFieldOffset
3215	>	U = getUnsafe();
3216	>	Class<?> k = ForkJoinPool.class;
3217	>	CTL = U.objectFieldOffset
3218		(k.getDeclaredField("ctl"));
3219	<	stealCountOffset = UNSAFE.objectFieldOffset
3219	>	STEALCOUNT = U.objectFieldOffset
3220		(k.getDeclaredField("stealCount"));
3221	<	blockedCountOffset = UNSAFE.objectFieldOffset
3222	<	(k.getDeclaredField("blockedCount"));
3223	<	quiescerCountOffset = UNSAFE.objectFieldOffset
3224	<	(k.getDeclaredField("quiescerCount"));
3225	<	scanGuardOffset = UNSAFE.objectFieldOffset
3226	<	(k.getDeclaredField("scanGuard"));
3227	<	nextWorkerNumberOffset = UNSAFE.objectFieldOffset
3228	<	(k.getDeclaredField("nextWorkerNumber"));
3229	<	Class a = ForkJoinTask[].class;
3230	<	ABASE = UNSAFE.arrayBaseOffset(a);
3231	<	s = UNSAFE.arrayIndexScale(a);
3221	>	PLOCK = U.objectFieldOffset
3222	>	(k.getDeclaredField("plock"));
3223	>	INDEXSEED = U.objectFieldOffset
3224	>	(k.getDeclaredField("indexSeed"));
3225	>	Class<?> tk = Thread.class;
3226	>	PARKBLOCKER = U.objectFieldOffset
3227	>	(tk.getDeclaredField("parkBlocker"));
3228	>	Class<?> wk = WorkQueue.class;
3229	>	QLOCK = U.objectFieldOffset
3230	>	(wk.getDeclaredField("qlock"));
3231	>	Class<?> ak = ForkJoinTask[].class;
3232	>	ABASE = U.arrayBaseOffset(ak);
3233	>	s = U.arrayIndexScale(ak);
3234	>	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
3235		} catch (Exception e) {
3236		throw new Error(e);
3237		}
3238		if ((s & (s-1)) != 0)
3239		throw new Error("data type scale not a power of two");
3240	<	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
3240	>
3241	>	submitters = new ThreadLocal<Submitter>();
3242	>	ForkJoinWorkerThreadFactory fac = defaultForkJoinWorkerThreadFactory =
3243	>	new DefaultForkJoinWorkerThreadFactory();
3244	>	/*
3245	>	* Establish common pool parameters.  For extra caution,
3246	>	* computations to set up common pool state are here; the
3247	>	* constructor just assigns these values to fields.
3248	>	*/
3249	>
3250	>	int par = 0;
3251	>	Thread.UncaughtExceptionHandler handler = null;
3252	>	try {  // TBD: limit or report ignored exceptions?
3253	>	String pp = System.getProperty
3254	>	("java.util.concurrent.ForkJoinPool.common.parallelism");
3255	>	String hp = System.getProperty
3256	>	("java.util.concurrent.ForkJoinPool.common.exceptionHandler");
3257	>	String fp = System.getProperty
3258	>	("java.util.concurrent.ForkJoinPool.common.threadFactory");
3259	>	if (fp != null)
3260	>	fac = ((ForkJoinWorkerThreadFactory)ClassLoader.
3261	>	getSystemClassLoader().loadClass(fp).newInstance());
3262	>	if (hp != null)
3263	>	handler = ((Thread.UncaughtExceptionHandler)ClassLoader.
3264	>	getSystemClassLoader().loadClass(hp).newInstance());
3265	>	if (pp != null)
3266	>	par = Integer.parseInt(pp);
3267	>	} catch (Exception ignore) {
3268	>	}
3269	>
3270	>	if (par <= 0)
3271	>	par = Runtime.getRuntime().availableProcessors();
3272	>	if (par > MAX_CAP)
3273	>	par = MAX_CAP;
3274	>	commonPoolParallelism = par;
3275	>	long np = (long)(-par); // precompute initial ctl value
3276	>	long ct = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK);
3277	>
3278	>	commonPool = new ForkJoinPool(par, ct, fac, handler);
3279	>	modifyThreadPermission = new RuntimePermission("modifyThread");
3280		}
3281
3282		/**
#	Line 2143 | Line 3306 | public class ForkJoinPool extends Abstra
3306		}
3307		}
3308		}
3309	+
3310		}

Diff Legend

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