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 1.100 by dl, Fri Apr 1 20:20:37 2011 UTC vs.
Revision 1.164 by dl, Tue Dec 18 21:46:16 2012 UTC

#	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 I/O or other unmanaged synchronization. The nested {@link
55	>	* ManagedBlocker} interface enables extension of the kinds of
56		* synchronization accommodated.
57		*
58		* <p>In addition to execution and lifecycle control methods, this
#	Line 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 system properties} with prefix {@code
103	>	* java.util.concurrent.ForkJoinPool.common}: {@code parallelism} --
104	>	* an integer greater than zero, {@code threadFactory} -- the class
105	>	* name of a {@link ForkJoinWorkerThreadFactory}, and {@code
106	>	* exceptionHandler} -- the class name of a {@link
107	>	* java.lang.Thread.UncaughtExceptionHandler
108	>	* Thread.UncaughtExceptionHandler}. Upon any error in establishing
109	>	* these settings, default parameters are used.
110		*
111		* <p><b>Implementation notes</b>: This implementation restricts the
112		* maximum number of running threads to 32767. Attempts to create
#	Line 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 tasks) 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.
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	<	* Wait Queuing. Unlike HPC work-stealing frameworks, we cannot
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.  We park/unpark workers after placing in an event
285	<	* wait queue when they cannot find work. This "queue" is actually
286	<	* a simple Treiber stack, headed by the "id" field of ctl, plus a
287	<	* 15bit counter value to both wake up waiters (by advancing their
288	<	* count) and avoid ABA effects. Successors are held in worker
289	<	* field "nextWait".  Queuing deals with several intrinsic races,
290	<	* mainly that a task-producing thread can miss seeing (and
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 both before (in scan()) and after
299	<	* (in tryAwaitWork()) a newly waiting worker is added to the wait
300	<	* queue. During a rescan, the worker might release some other
301	<	* queued worker rather than itself, which has the same net
302	<	* effect. Because enqueued workers may actually be rescanning
303	<	* rather than waiting, we set and clear the "parked" field of
304	<	* ForkJoinWorkerThread to reduce unnecessary calls to unpark.
305	<	* (Use of the parked field requires a secondary recheck to avoid
306	<	* missed signals.)
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 has fewer
320	>	* than two tasks, they signal waiting workers (or trigger
321	>	* creation of new ones if fewer than the given parallelism level
322	>	* -- signalWork), and may leave a hint to the unparked worker to
323	>	* help signal others upon wakeup).  These primary signals are
324	>	* buttressed by others (see method helpSignal) whenever other
325	>	* threads scan for work or do not have a task to process.  On
326	>	* most platforms, signalling (unpark) overhead time is noticeably
327	>	* long, and the time between signalling a thread and it actually
328	>	* making progress can be very noticeably long, so it is worth
329	>	* offloading these delays from critical paths as much as
330	>	* possible.
331		*
332		* Trimming workers. To release resources after periods of lack of
333		* use, a worker starting to wait when the pool is quiescent will
334	<	* time out and terminate if the pool has remained quiescent for
335	<	* SHRINK_RATE nanosecs. This will slowly propagate, eventually
336	<	* terminating all workers after long periods of non-use.
337	<	*
338	<	* Submissions. External submissions are maintained in an
339	<	* array-based queue that is structured identically to
340	<	* ForkJoinWorkerThread queues except for the use of
341	<	* submissionLock in method addSubmission. Unlike the case for
342	<	* worker queues, multiple external threads can add new
343	<	* submissions, so adding requires a lock.
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:
334	>	* time out and terminate if the pool has remained quiescent for a
335	>	* given period -- a short period if there are more threads than
336	>	* parallelism, longer as the number of threads decreases. This
337	>	* will slowly propagate, eventually terminating all workers after
338	>	* periods of non-use.
339	>	*
340	>	* Shutdown and Termination. A call to shutdownNow atomically sets
341	>	* a plock bit and then (non-atomically) sets each worker's
342	>	* qlock status, cancels all unprocessed tasks, and wakes up
343	>	* all waiting workers.  Detecting whether termination should
344	>	* commence after a non-abrupt shutdown() call requires more work
345	>	* and bookkeeping. We need consensus about quiescence (i.e., that
346	>	* there is no more work). The active count provides a primary
347	>	* indication but non-abrupt shutdown still requires a rechecking
348	>	* scan for any workers that are inactive but not queued.
349	>	*
350	>	* Joining Tasks
351	>	* =============
352	>	*
353	>	* Any of several actions may be taken when one worker is waiting
354	>	* to join a task stolen (or always held) by another.  Because we
355	>	* are multiplexing many tasks on to a pool of workers, we can't
356	>	* just let them block (as in Thread.join).  We also cannot just
357	>	* reassign the joiner's run-time stack with another and replace
358	>	* it later, which would be a form of "continuation", that even if
359	>	* possible is not necessarily a good idea since we sometimes need
360	>	* both an unblocked task and its continuation to progress.
361	>	* Instead we combine two tactics:
362		*
363		*   Helping: Arranging for the joiner to execute some task that it
364	<	*      would be running if the steal had not occurred.  Method
260	<	*      ForkJoinWorkerThread.joinTask tracks joining->stealing
261	<	*      links to try to find such a task.
364	>	*      would be running if the steal had not occurred.
365		*
366		*   Compensating: Unless there are already enough live threads,
367	<	*      method tryPreBlock() may create or re-activate a spare
368	<	*      thread to compensate for blocked joiners until they
369	<	*      unblock.
367	>	*      method tryCompensate() may create or re-activate a spare
368	>	*      thread to compensate for blocked joiners until they unblock.
369	>	*
370	>	* A third form (implemented in tryRemoveAndExec) amounts to
371	>	* helping a hypothetical compensator: If we can readily tell that
372	>	* a possible action of a compensator is to steal and execute the
373	>	* task being joined, the joining thread can do so directly,
374	>	* without the need for a compensation thread (although at the
375	>	* expense of larger run-time stacks, but the tradeoff is
376	>	* typically worthwhile).
377		*
378		* The ManagedBlocker extension API can't use helping so relies
379		* only on compensation in method awaitBlocker.
380		*
381	+	* The algorithm in tryHelpStealer entails a form of "linear"
382	+	* helping: Each worker records (in field currentSteal) the most
383	+	* recent task it stole from some other worker. Plus, it records
384	+	* (in field currentJoin) the task it is currently actively
385	+	* joining. Method tryHelpStealer uses these markers to try to
386	+	* find a worker to help (i.e., steal back a task from and execute
387	+	* it) that could hasten completion of the actively joined task.
388	+	* In essence, the joiner executes a task that would be on its own
389	+	* local deque had the to-be-joined task not been stolen. This may
390	+	* be seen as a conservative variant of the approach in Wagner &
391	+	* Calder "Leapfrogging: a portable technique for implementing
392	+	* efficient futures" SIGPLAN Notices, 1993
393	+	* (http://portal.acm.org/citation.cfm?id=155354). It differs in
394	+	* that: (1) We only maintain dependency links across workers upon
395	+	* steals, rather than use per-task bookkeeping.  This sometimes
396	+	* requires a linear scan of workQueues array to locate stealers,
397	+	* but often doesn't because stealers leave hints (that may become
398	+	* stale/wrong) of where to locate them.  It is only a hint
399	+	* because a worker might have had multiple steals and the hint
400	+	* records only one of them (usually the most current).  Hinting
401	+	* isolates cost to when it is needed, rather than adding to
402	+	* per-task overhead.  (2) It is "shallow", ignoring nesting and
403	+	* potentially cyclic mutual steals.  (3) It is intentionally
404	+	* racy: field currentJoin is updated only while actively joining,
405	+	* which means that we miss links in the chain during long-lived
406	+	* tasks, GC stalls etc (which is OK since blocking in such cases
407	+	* is usually a good idea).  (4) We bound the number of attempts
408	+	* to find work (see MAX_HELP) and fall back to suspending the
409	+	* worker and if necessary replacing it with another.
410	+	*
411	+	* Helping actions for CountedCompleters are much simpler: Method
412	+	* helpComplete can take and execute any task with the same root
413	+	* as the task being waited on. However, this still entails some
414	+	* traversal of completer chains, so is less efficient than using
415	+	* CountedCompleters without explicit joins.
416	+	*
417		* It is impossible to keep exactly the target parallelism number
418		* of threads running at any given time.  Determining the
419		* existence of conservatively safe helping targets, the
420		* availability of already-created spares, and the apparent need
421	<	* to create new spares are all racy and require heuristic
422	<	* guidance, so we rely on multiple retries of each.  Currently,
423	<	* in keeping with on-demand signalling policy, we compensate only
424	<	* if blocking would leave less than one active (non-waiting,
425	<	* non-blocked) worker. Additionally, to avoid some false alarms
426	<	* due to GC, lagging counters, system activity, etc, compensated
427	<	* blocking for joins is only attempted after rechecks stabilize
428	<	* (retries are interspersed with Thread.yield, for good
429	<	* citizenship).  The variable blockedCount, incremented before
430	<	* blocking and decremented after, is sometimes needed to
431	<	* distinguish cases of waiting for work vs blocking on joins or
432	<	* other managed sync. Both cases are equivalent for most pool
433	<	* control, so we can update non-atomically. (Additionally,
434	<	* contention on 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
421	>	* to create new spares are all racy, so we rely on multiple
422	>	* retries of each.  Compensation in the apparent absence of
423	>	* helping opportunities is challenging to control on JVMs, where
424	>	* GC and other activities can stall progress of tasks that in
425	>	* turn stall out many other dependent tasks, without us being
426	>	* able to determine whether they will ever require compensation.
427	>	* Even though work-stealing otherwise encounters little
428	>	* degradation in the presence of more threads than cores,
429	>	* aggressively adding new threads in such cases entails risk of
430	>	* unwanted positive feedback control loops in which more threads
431	>	* cause more dependent stalls (as well as delayed progress of
432	>	* unblocked threads to the point that we know they are available)
433	>	* leading to more situations requiring more threads, and so
434	>	* on. This aspect of control can be seen as an (analytically
435	>	* intractable) game with an opponent that may choose the worst
436	>	* (for us) active thread to stall at any time.  We take several
437	>	* precautions to bound losses (and thus bound gains), mainly in
438	>	* methods tryCompensate and awaitJoin.
439	>	*
440	>	* Common Pool
441	>	* ===========
442	>	*
443	>	* The static commonPool always exists after static
444	>	* initialization.  Since it (or any other created pool) need
445	>	* never be used, we minimize initial construction overhead and
446	>	* footprint to the setup of about a dozen fields, with no nested
447	>	* allocation. Most bootstrapping occurs within method
448	>	* fullExternalPush during the first submission to the pool.
449	>	*
450	>	* When external threads submit to the common pool, they can
451	>	* perform some subtask processing (see externalHelpJoin and
452	>	* related methods).  We do not need to record whether these
453	>	* submissions are to the common pool -- if not, externalHelpJoin
454	>	* returns quickly (at the most helping to signal some common pool
455	>	* workers). These submitters would otherwise be blocked waiting
456	>	* for completion, so the extra effort (with liberally sprinkled
457	>	* task status checks) in inapplicable cases amounts to an odd
458	>	* form of limited spin-wait before blocking in ForkJoinTask.join.
459	>	*
460	>	* Style notes
461	>	* ===========
462	>	*
463	>	* There is a lot of representation-level coupling among classes
464	>	* ForkJoinPool, ForkJoinWorkerThread, and ForkJoinTask.  The
465	>	* fields of WorkQueue maintain data structures managed by
466	>	* ForkJoinPool, so are directly accessed.  There is little point
467		* trying to reduce this, since any associated future changes in
468		* representations will need to be accompanied by algorithmic
469	<	* changes anyway. All together, these low-level implementation
470	<	* choices produce as much as a factor of 4 performance
471	<	* improvement compared to naive implementations, and enable the
472	<	* processing of billions of tasks per second, at the expense of
473	<	* some ugliness.
474	<	*
475	<	* Methods signalWork() and scan() are the main bottlenecks so are
476	<	* especially heavily micro-optimized/mangled.  There are lots of
477	<	* inline assignments (of form "while ((local = field) != 0)")
478	<	* which are usually the simplest way to ensure the required read
479	<	* orderings (which are sometimes critical). This leads to a
480	<	* "C"-like style of listing declarations of these locals at the
481	<	* heads of methods or blocks.  There are several occurrences of
482	<	* the unusual "do {} while (!cas...)"  which is the simplest way
483	<	* to force an update of a CAS'ed variable. There are also other
484	<	* coding oddities that help some methods perform reasonably even
485	<	* when interpreted (not compiled).
486	<	*
487	<	* The order of declarations in this file is: (1) declarations of
488	<	* statics (2) fields (along with constants used when unpacking
489	<	* some of them), listed in an order that tends to reduce
490	<	* contention among them a bit under most JVMs.  (3) internal
491	<	* control methods (4) callbacks and other support for
492	<	* ForkJoinTask and ForkJoinWorkerThread classes, (5) exported
493	<	* methods (plus a few little helpers). (6) static block
335	<	* initializing all statics in a minimally dependent order.
469	>	* changes anyway. Several methods intrinsically sprawl because
470	>	* they must accumulate sets of consistent reads of volatiles held
471	>	* in local variables.  Methods signalWork() and scan() are the
472	>	* main bottlenecks, so are especially heavily
473	>	* micro-optimized/mangled.  There are lots of inline assignments
474	>	* (of form "while ((local = field) != 0)") which are usually the
475	>	* simplest way to ensure the required read orderings (which are
476	>	* sometimes critical). This leads to a "C"-like style of listing
477	>	* declarations of these locals at the heads of methods or blocks.
478	>	* There are several occurrences of the unusual "do {} while
479	>	* (!cas...)"  which is the simplest way to force an update of a
480	>	* CAS'ed variable. There are also other coding oddities (including
481	>	* several unnecessary-looking hoisted null checks) that help
482	>	* some methods perform reasonably even when interpreted (not
483	>	* compiled).
484	>	*
485	>	* The order of declarations in this file is:
486	>	* (1) Static utility functions
487	>	* (2) Nested (static) classes
488	>	* (3) Static fields
489	>	* (4) Fields, along with constants used when unpacking some of them
490	>	* (5) Internal control methods
491	>	* (6) Callbacks and other support for ForkJoinTask methods
492	>	* (7) Exported methods
493	>	* (8) Static block initializing statics in minimally dependent order
494		*/
495
496	+	// Static utilities
497	+
498	+	/**
499	+	* If there is a security manager, makes sure caller has
500	+	* permission to modify threads.
501	+	*/
502	+	private static void checkPermission() {
503	+	SecurityManager security = System.getSecurityManager();
504	+	if (security != null)
505	+	security.checkPermission(modifyThreadPermission);
506	+	}
507	+
508	+	// Nested classes
509	+
510		/**
511		* Factory for creating new {@link ForkJoinWorkerThread}s.
512		* A {@code ForkJoinWorkerThreadFactory} must be defined and used
#	Line 355 | Line 527 | public class ForkJoinPool extends Abstra
527		* Default ForkJoinWorkerThreadFactory implementation; creates a
528		* new ForkJoinWorkerThread.
529		*/
530	<	static class DefaultForkJoinWorkerThreadFactory
530	>	static final class DefaultForkJoinWorkerThreadFactory
531		implements ForkJoinWorkerThreadFactory {
532	<	public ForkJoinWorkerThread newThread(ForkJoinPool pool) {
532	>	public final ForkJoinWorkerThread newThread(ForkJoinPool pool) {
533		return new ForkJoinWorkerThread(pool);
534		}
535		}
536
537		/**
538	<	* Creates a new ForkJoinWorkerThread. This factory is used unless
539	<	* overridden in ForkJoinPool constructors.
538	>	* Per-thread records for threads that submit to pools. Currently
539	>	* holds only pseudo-random seed / index that is used to choose
540	>	* submission queues in method externalPush. In the future, this may
541	>	* also incorporate a means to implement different task rejection
542	>	* and resubmission policies.
543	>	*
544	>	* Seeds for submitters and workers/workQueues work in basically
545	>	* the same way but are initialized and updated using slightly
546	>	* different mechanics. Both are initialized using the same
547	>	* approach as in class ThreadLocal, where successive values are
548	>	* unlikely to collide with previous values. Seeds are then
549	>	* randomly modified upon collisions using xorshifts, which
550	>	* requires a non-zero seed.
551	>	*/
552	>	static final class Submitter {
553	>	int seed;
554	>	Submitter(int s) { seed = s; }
555	>	}
556	>
557	>	/**
558	>	* Class for artificial tasks that are used to replace the target
559	>	* of local joins if they are removed from an interior queue slot
560	>	* in WorkQueue.tryRemoveAndExec. We don't need the proxy to
561	>	* actually do anything beyond having a unique identity.
562	>	*/
563	>	static final class EmptyTask extends ForkJoinTask<Void> {
564	>	private static final long serialVersionUID = -7721805057305804111L;
565	>	EmptyTask() { status = ForkJoinTask.NORMAL; } // force done
566	>	public final Void getRawResult() { return null; }
567	>	public final void setRawResult(Void x) {}
568	>	public final boolean exec() { return true; }
569	>	}
570	>
571	>	/**
572	>	* Queues supporting work-stealing as well as external task
573	>	* submission. See above for main rationale and algorithms.
574	>	* Implementation relies heavily on "Unsafe" intrinsics
575	>	* and selective use of "volatile":
576	>	*
577	>	* Field "base" is the index (mod array.length) of the least valid
578	>	* queue slot, which is always the next position to steal (poll)
579	>	* from if nonempty. Reads and writes require volatile orderings
580	>	* but not CAS, because updates are only performed after slot
581	>	* CASes.
582	>	*
583	>	* Field "top" is the index (mod array.length) of the next queue
584	>	* slot to push to or pop from. It is written only by owner thread
585	>	* for push, or under lock for external/shared push, and accessed
586	>	* by other threads only after reading (volatile) base.  Both top
587	>	* and base are allowed to wrap around on overflow, but (top -
588	>	* base) (or more commonly -(base - top) to force volatile read of
589	>	* base before top) still estimates size. The lock ("qlock") is
590	>	* forced to -1 on termination, causing all further lock attempts
591	>	* to fail. (Note: we don't need CAS for termination state because
592	>	* upon pool shutdown, all shared-queues will stop being used
593	>	* anyway.)  Nearly all lock bodies are set up so that exceptions
594	>	* within lock bodies are "impossible" (modulo JVM errors that
595	>	* would cause failure anyway.)
596	>	*
597	>	* The array slots are read and written using the emulation of
598	>	* volatiles/atomics provided by Unsafe. Insertions must in
599	>	* general use putOrderedObject as a form of releasing store to
600	>	* ensure that all writes to the task object are ordered before
601	>	* its publication in the queue.  All removals entail a CAS to
602	>	* null.  The array is always a power of two. To ensure safety of
603	>	* Unsafe array operations, all accesses perform explicit null
604	>	* checks and implicit bounds checks via power-of-two masking.
605	>	*
606	>	* In addition to basic queuing support, this class contains
607	>	* fields described elsewhere to control execution. It turns out
608	>	* to work better memory-layout-wise to include them in this class
609	>	* rather than a separate class.
610	>	*
611	>	* Performance on most platforms is very sensitive to placement of
612	>	* instances of both WorkQueues and their arrays -- we absolutely
613	>	* do not want multiple WorkQueue instances or multiple queue
614	>	* arrays sharing cache lines. (It would be best for queue objects
615	>	* and their arrays to share, but there is nothing available to
616	>	* help arrange that).  Unfortunately, because they are recorded
617	>	* in a common array, WorkQueue instances are often moved to be
618	>	* adjacent by garbage collectors. To reduce impact, we use field
619	>	* padding that works OK on common platforms; this effectively
620	>	* trades off slightly slower average field access for the sake of
621	>	* avoiding really bad worst-case access. (Until better JVM
622	>	* support is in place, this padding is dependent on transient
623	>	* properties of JVM field layout rules.) We also take care in
624	>	* allocating, sizing and resizing the array. Non-shared queue
625	>	* arrays are initialized by workers before use. Others are
626	>	* allocated on first use.
627		*/
628	<	public static final ForkJoinWorkerThreadFactory
629	<	defaultForkJoinWorkerThreadFactory;
628	>	static final class WorkQueue {
629	>	/**
630	>	* Capacity of work-stealing queue array upon initialization.
631	>	* Must be a power of two; at least 4, but should be larger to
632	>	* reduce or eliminate cacheline sharing among queues.
633	>	* Currently, it is much larger, as a partial workaround for
634	>	* the fact that JVMs often place arrays in locations that
635	>	* share GC bookkeeping (especially cardmarks) such that
636	>	* per-write accesses encounter serious memory contention.
637	>	*/
638	>	static final int INITIAL_QUEUE_CAPACITY = 1 << 13;
639
640	<	/**
641	<	* Permission required for callers of methods that may start or
642	<	* kill threads.
643	<	*/
644	<	private static final RuntimePermission modifyThreadPermission;
640	>	/**
641	>	* Maximum size for queue arrays. Must be a power of two less
642	>	* than or equal to 1 << (31 - width of array entry) to ensure
643	>	* lack of wraparound of index calculations, but defined to a
644	>	* value a bit less than this to help users trap runaway
645	>	* programs before saturating systems.
646	>	*/
647	>	static final int MAXIMUM_QUEUE_CAPACITY = 1 << 26; // 64M
648
649	<	/**
650	<	* If there is a security manager, makes sure caller has
651	<	* permission to modify threads.
652	<	*/
653	<	private static void checkPermission() {
654	<	SecurityManager security = System.getSecurityManager();
655	<	if (security != null)
656	<	security.checkPermission(modifyThreadPermission);
649	>	// Heuristic padding to ameliorate unfortunate memory placements
650	>	volatile long pad00, pad01, pad02, pad03, pad04, pad05, pad06;
651	>
652	>	int seed;                  // for random scanning; initialize nonzero
653	>	volatile int eventCount;   // encoded inactivation count; < 0 if inactive
654	>	int nextWait;              // encoded record of next event waiter
655	>	int hint;                  // steal or signal hint (index)
656	>	int poolIndex;             // index of this queue in pool (or 0)
657	>	final int mode;            // 0: lifo, > 0: fifo, < 0: shared
658	>	int nsteals;               // number of steals
659	>	volatile int qlock;        // 1: locked, -1: terminate; else 0
660	>	volatile int base;         // index of next slot for poll
661	>	int top;                   // index of next slot for push
662	>	ForkJoinTask<?>[] array;   // the elements (initially unallocated)
663	>	final ForkJoinPool pool;   // the containing pool (may be null)
664	>	final ForkJoinWorkerThread owner; // owning thread or null if shared
665	>	volatile Thread parker;    // == owner during call to park; else null
666	>	volatile ForkJoinTask<?> currentJoin;  // task being joined in awaitJoin
667	>	ForkJoinTask<?> currentSteal; // current non-local task being executed
668	>
669	>	volatile Object pad10, pad11, pad12, pad13, pad14, pad15, pad16, pad17;
670	>	volatile Object pad18, pad19, pad1a, pad1b, pad1c, pad1d;
671	>
672	>	WorkQueue(ForkJoinPool pool, ForkJoinWorkerThread owner, int mode,
673	>	int seed) {
674	>	this.pool = pool;
675	>	this.owner = owner;
676	>	this.mode = mode;
677	>	this.seed = seed;
678	>	// Place indices in the center of array (that is not yet allocated)
679	>	base = top = INITIAL_QUEUE_CAPACITY >>> 1;
680	>	}
681	>
682	>	/**
683	>	* Returns the approximate number of tasks in the queue.
684	>	*/
685	>	final int queueSize() {
686	>	int n = base - top;       // non-owner callers must read base first
687	>	return (n >= 0) ? 0 : -n; // ignore transient negative
688	>	}
689	>
690	>	/**
691	>	* Provides a more accurate estimate of whether this queue has
692	>	* any tasks than does queueSize, by checking whether a
693	>	* near-empty queue has at least one unclaimed task.
694	>	*/
695	>	final boolean isEmpty() {
696	>	ForkJoinTask<?>[] a; int m, s;
697	>	int n = base - (s = top);
698	>	return (n >= 0 ||
699	>	(n == -1 &&
700	>	((a = array) == null ||
701	>	(m = a.length - 1) < 0 ||
702	>	U.getObject
703	>	(a, (long)((m & (s - 1)) << ASHIFT) + ABASE) == null)));
704	>	}
705	>
706	>	/**
707	>	* Pushes a task. Call only by owner in unshared queues.  (The
708	>	* shared-queue version is embedded in method externalPush.)
709	>	*
710	>	* @param task the task. Caller must ensure non-null.
711	>	* @throw RejectedExecutionException if array cannot be resized
712	>	*/
713	>	final void push(ForkJoinTask<?> task) {
714	>	ForkJoinTask<?>[] a; ForkJoinPool p;
715	>	int s = top, m, n;
716	>	if ((a = array) != null) {    // ignore if queue removed
717	>	int j = (((m = a.length - 1) & s) << ASHIFT) + ABASE;
718	>	U.putOrderedObject(a, j, task);
719	>	if ((n = (top = s + 1) - base) <= 2) {
720	>	if ((p = pool) != null)
721	>	p.signalWork(this);
722	>	}
723	>	else if (n >= m)
724	>	growArray();
725	>	}
726	>	}
727	>
728	>	/**
729	>	* Initializes or doubles the capacity of array. Call either
730	>	* by owner or with lock held -- it is OK for base, but not
731	>	* top, to move while resizings are in progress.
732	>	*/
733	>	final ForkJoinTask<?>[] growArray() {
734	>	ForkJoinTask<?>[] oldA = array;
735	>	int size = oldA != null ? oldA.length << 1 : INITIAL_QUEUE_CAPACITY;
736	>	if (size > MAXIMUM_QUEUE_CAPACITY)
737	>	throw new RejectedExecutionException("Queue capacity exceeded");
738	>	int oldMask, t, b;
739	>	ForkJoinTask<?>[] a = array = new ForkJoinTask<?>[size];
740	>	if (oldA != null && (oldMask = oldA.length - 1) >= 0 &&
741	>	(t = top) - (b = base) > 0) {
742	>	int mask = size - 1;
743	>	do {
744	>	ForkJoinTask<?> x;
745	>	int oldj = ((b & oldMask) << ASHIFT) + ABASE;
746	>	int j    = ((b &    mask) << ASHIFT) + ABASE;
747	>	x = (ForkJoinTask<?>)U.getObjectVolatile(oldA, oldj);
748	>	if (x != null &&
749	>	U.compareAndSwapObject(oldA, oldj, x, null))
750	>	U.putObjectVolatile(a, j, x);
751	>	} while (++b != t);
752	>	}
753	>	return a;
754	>	}
755	>
756	>	/**
757	>	* Takes next task, if one exists, in LIFO order.  Call only
758	>	* by owner in unshared queues.
759	>	*/
760	>	final ForkJoinTask<?> pop() {
761	>	ForkJoinTask<?>[] a; ForkJoinTask<?> t; int m;
762	>	if ((a = array) != null && (m = a.length - 1) >= 0) {
763	>	for (int s; (s = top - 1) - base >= 0;) {
764	>	long j = ((m & s) << ASHIFT) + ABASE;
765	>	if ((t = (ForkJoinTask<?>)U.getObject(a, j)) == null)
766	>	break;
767	>	if (U.compareAndSwapObject(a, j, t, null)) {
768	>	top = s;
769	>	return t;
770	>	}
771	>	}
772	>	}
773	>	return null;
774	>	}
775	>
776	>	/**
777	>	* Takes a task in FIFO order if b is base of queue and a task
778	>	* can be claimed without contention. Specialized versions
779	>	* appear in ForkJoinPool methods scan and tryHelpStealer.
780	>	*/
781	>	final ForkJoinTask<?> pollAt(int b) {
782	>	ForkJoinTask<?> t; ForkJoinTask<?>[] a;
783	>	if ((a = array) != null) {
784	>	int j = (((a.length - 1) & b) << ASHIFT) + ABASE;
785	>	if ((t = (ForkJoinTask<?>)U.getObjectVolatile(a, j)) != null &&
786	>	base == b &&
787	>	U.compareAndSwapObject(a, j, t, null)) {
788	>	base = b + 1;
789	>	return t;
790	>	}
791	>	}
792	>	return null;
793	>	}
794	>
795	>	/**
796	>	* Takes next task, if one exists, in FIFO order.
797	>	*/
798	>	final ForkJoinTask<?> poll() {
799	>	ForkJoinTask<?>[] a; int b; ForkJoinTask<?> t;
800	>	while ((b = base) - top < 0 && (a = array) != null) {
801	>	int j = (((a.length - 1) & b) << ASHIFT) + ABASE;
802	>	t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
803	>	if (t != null) {
804	>	if (base == b &&
805	>	U.compareAndSwapObject(a, j, t, null)) {
806	>	base = b + 1;
807	>	return t;
808	>	}
809	>	}
810	>	else if (base == b) {
811	>	if (b + 1 == top)
812	>	break;
813	>	Thread.yield(); // wait for lagging update (very rare)
814	>	}
815	>	}
816	>	return null;
817	>	}
818	>
819	>	/**
820	>	* Takes next task, if one exists, in order specified by mode.
821	>	*/
822	>	final ForkJoinTask<?> nextLocalTask() {
823	>	return mode == 0 ? pop() : poll();
824	>	}
825	>
826	>	/**
827	>	* Returns next task, if one exists, in order specified by mode.
828	>	*/
829	>	final ForkJoinTask<?> peek() {
830	>	ForkJoinTask<?>[] a = array; int m;
831	>	if (a == null || (m = a.length - 1) < 0)
832	>	return null;
833	>	int i = mode == 0 ? top - 1 : base;
834	>	int j = ((i & m) << ASHIFT) + ABASE;
835	>	return (ForkJoinTask<?>)U.getObjectVolatile(a, j);
836	>	}
837	>
838	>	/**
839	>	* Pops the given task only if it is at the current top.
840	>	* (A shared version is available only via FJP.tryExternalUnpush)
841	>	*/
842	>	final boolean tryUnpush(ForkJoinTask<?> t) {
843	>	ForkJoinTask<?>[] a; int s;
844	>	if ((a = array) != null && (s = top) != base &&
845	>	U.compareAndSwapObject
846	>	(a, (((a.length - 1) & --s) << ASHIFT) + ABASE, t, null)) {
847	>	top = s;
848	>	return true;
849	>	}
850	>	return false;
851	>	}
852	>
853	>	/**
854	>	* Removes and cancels all known tasks, ignoring any exceptions.
855	>	*/
856	>	final void cancelAll() {
857	>	ForkJoinTask.cancelIgnoringExceptions(currentJoin);
858	>	ForkJoinTask.cancelIgnoringExceptions(currentSteal);
859	>	for (ForkJoinTask<?> t; (t = poll()) != null; )
860	>	ForkJoinTask.cancelIgnoringExceptions(t);
861	>	}
862	>
863	>	/**
864	>	* Computes next value for random probes.  Scans don't require
865	>	* a very high quality generator, but also not a crummy one.
866	>	* Marsaglia xor-shift is cheap and works well enough.  Note:
867	>	* This is manually inlined in its usages in ForkJoinPool to
868	>	* avoid writes inside busy scan loops.
869	>	*/
870	>	final int nextSeed() {
871	>	int r = seed;
872	>	r ^= r << 13;
873	>	r ^= r >>> 17;
874	>	return seed = r ^= r << 5;
875	>	}
876	>
877	>	// Specialized execution methods
878	>
879	>	/**
880	>	* Pops and runs tasks until empty.
881	>	*/
882	>	private void popAndExecAll() {
883	>	// A bit faster than repeated pop calls
884	>	ForkJoinTask<?>[] a; int m, s; long j; ForkJoinTask<?> t;
885	>	while ((a = array) != null && (m = a.length - 1) >= 0 &&
886	>	(s = top - 1) - base >= 0 &&
887	>	(t = ((ForkJoinTask<?>)
888	>	U.getObject(a, j = ((m & s) << ASHIFT) + ABASE)))
889	>	!= null) {
890	>	if (U.compareAndSwapObject(a, j, t, null)) {
891	>	top = s;
892	>	t.doExec();
893	>	}
894	>	}
895	>	}
896	>
897	>	/**
898	>	* Polls and runs tasks until empty.
899	>	*/
900	>	private void pollAndExecAll() {
901	>	for (ForkJoinTask<?> t; (t = poll()) != null;)
902	>	t.doExec();
903	>	}
904	>
905	>	/**
906	>	* If present, removes from queue and executes the given task,
907	>	* or any other cancelled task. Returns (true) on any CAS
908	>	* or consistency check failure so caller can retry.
909	>	*
910	>	* @return false if no progress can be made, else true;
911	>	*/
912	>	final boolean tryRemoveAndExec(ForkJoinTask<?> task) {
913	>	boolean stat = true, removed = false, empty = true;
914	>	ForkJoinTask<?>[] a; int m, s, b, n;
915	>	if ((a = array) != null && (m = a.length - 1) >= 0 &&
916	>	(n = (s = top) - (b = base)) > 0) {
917	>	for (ForkJoinTask<?> t;;) {           // traverse from s to b
918	>	int j = ((--s & m) << ASHIFT) + ABASE;
919	>	t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
920	>	if (t == null)                    // inconsistent length
921	>	break;
922	>	else if (t == task) {
923	>	if (s + 1 == top) {           // pop
924	>	if (!U.compareAndSwapObject(a, j, task, null))
925	>	break;
926	>	top = s;
927	>	removed = true;
928	>	}
929	>	else if (base == b)           // replace with proxy
930	>	removed = U.compareAndSwapObject(a, j, task,
931	>	new EmptyTask());
932	>	break;
933	>	}
934	>	else if (t.status >= 0)
935	>	empty = false;
936	>	else if (s + 1 == top) {          // pop and throw away
937	>	if (U.compareAndSwapObject(a, j, t, null))
938	>	top = s;
939	>	break;
940	>	}
941	>	if (--n == 0) {
942	>	if (!empty && base == b)
943	>	stat = false;
944	>	break;
945	>	}
946	>	}
947	>	}
948	>	if (removed)
949	>	task.doExec();
950	>	return stat;
951	>	}
952	>
953	>	/**
954	>	* Polls for and executes the given task or any other task in
955	>	* its CountedCompleter computation
956	>	*/
957	>	final boolean pollAndExecCC(ForkJoinTask<?> root) {
958	>	ForkJoinTask<?>[] a; int b; Object o;
959	>	outer: while ((b = base) - top < 0 && (a = array) != null) {
960	>	long j = (((a.length - 1) & b) << ASHIFT) + ABASE;
961	>	if ((o = U.getObject(a, j)) == null ||
962	>	!(o instanceof CountedCompleter))
963	>	break;
964	>	for (CountedCompleter<?> t = (CountedCompleter<?>)o, r = t;;) {
965	>	if (r == root) {
966	>	if (base == b &&
967	>	U.compareAndSwapObject(a, j, t, null)) {
968	>	base = b + 1;
969	>	t.doExec();
970	>	return true;
971	>	}
972	>	else
973	>	break; // restart
974	>	}
975	>	if ((r = r.completer) == null)
976	>	break outer; // not part of root computation
977	>	}
978	>	}
979	>	return false;
980	>	}
981	>
982	>	/**
983	>	* Executes a top-level task and any local tasks remaining
984	>	* after execution.
985	>	*/
986	>	final void runTask(ForkJoinTask<?> t) {
987	>	if (t != null) {
988	>	(currentSteal = t).doExec();
989	>	currentSteal = null;
990	>	++nsteals;
991	>	if (base - top < 0) {       // process remaining local tasks
992	>	if (mode == 0)
993	>	popAndExecAll();
994	>	else
995	>	pollAndExecAll();
996	>	}
997	>	}
998	>	}
999	>
1000	>	/**
1001	>	* Executes a non-top-level (stolen) task.
1002	>	*/
1003	>	final void runSubtask(ForkJoinTask<?> t) {
1004	>	if (t != null) {
1005	>	ForkJoinTask<?> ps = currentSteal;
1006	>	(currentSteal = t).doExec();
1007	>	currentSteal = ps;
1008	>	}
1009	>	}
1010	>
1011	>	/**
1012	>	* Returns true if owned and not known to be blocked.
1013	>	*/
1014	>	final boolean isApparentlyUnblocked() {
1015	>	Thread wt; Thread.State s;
1016	>	return (eventCount >= 0 &&
1017	>	(wt = owner) != null &&
1018	>	(s = wt.getState()) != Thread.State.BLOCKED &&
1019	>	s != Thread.State.WAITING &&
1020	>	s != Thread.State.TIMED_WAITING);
1021	>	}
1022	>
1023	>	// Unsafe mechanics
1024	>	private static final sun.misc.Unsafe U;
1025	>	private static final long QLOCK;
1026	>	private static final int ABASE;
1027	>	private static final int ASHIFT;
1028	>	static {
1029	>	int s;
1030	>	try {
1031	>	U = getUnsafe();
1032	>	Class<?> k = WorkQueue.class;
1033	>	Class<?> ak = ForkJoinTask[].class;
1034	>	QLOCK = U.objectFieldOffset
1035	>	(k.getDeclaredField("qlock"));
1036	>	ABASE = U.arrayBaseOffset(ak);
1037	>	s = U.arrayIndexScale(ak);
1038	>	} catch (Exception e) {
1039	>	throw new Error(e);
1040	>	}
1041	>	if ((s & (s-1)) != 0)
1042	>	throw new Error("data type scale not a power of two");
1043	>	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
1044	>	}
1045		}
1046
1047	+	// static fields (initialized in static initializer below)
1048	+
1049		/**
1050	<	* Generator for assigning sequence numbers as pool names.
1050	>	* Creates a new ForkJoinWorkerThread. This factory is used unless
1051	>	* overridden in ForkJoinPool constructors.
1052		*/
1053	<	private static final AtomicInteger poolNumberGenerator;
1053	>	public static final ForkJoinWorkerThreadFactory
1054	>	defaultForkJoinWorkerThreadFactory;
1055
1056		/**
1057	<	* Generator for initial random seeds for worker victim
1058	<	* selection. This is used only to create initial seeds. Random
1059	<	* steals use a cheaper xorshift generator per steal attempt. We
1060	<	* don't expect much contention on seedGenerator, so just use a
1061	<	* plain Random.
1057	>	* Per-thread submission bookkeeping. Shared across all pools
1058	>	* to reduce ThreadLocal pollution and because random motion
1059	>	* to avoid contention in one pool is likely to hold for others.
1060	>	* Lazily initialized on first submission (but null-checked
1061	>	* in other contexts to avoid unnecessary initialization).
1062		*/
1063	<	static final Random workerSeedGenerator;
1063	>	static final ThreadLocal<Submitter> submitters;
1064
1065		/**
1066	<	* Array holding all worker threads in the pool.  Initialized upon
1067	<	* construction. Array size must be a power of two.  Updates and
405	<	* replacements are protected by scanGuard, but the array is
406	<	* always kept in a consistent enough state to be randomly
407	<	* accessed without locking by workers performing work-stealing,
408	<	* as well as other traversal-based methods in this class, so long
409	<	* as reads memory-acquire by first reading ctl. All readers must
410	<	* tolerate that some array slots may be null.
1066	>	* Permission required for callers of methods that may start or
1067	>	* kill threads.
1068		*/
1069	<	ForkJoinWorkerThread[] workers;
1069	>	private static final RuntimePermission modifyThreadPermission;
1070
1071		/**
1072	<	* Initial size for submission queue array. Must be a power of
1073	<	* two.  In many applications, these always stay small so we use a
1074	<	* small initial cap.
1072	>	* Common (static) pool. Non-null for public use unless a static
1073	>	* construction exception, but internal usages null-check on use
1074	>	* to paranoically avoid potential initialization circularities
1075	>	* as well as to simplify generated code.
1076		*/
1077	<	private static final int INITIAL_QUEUE_CAPACITY = 8;
1077	>	static final ForkJoinPool commonPool;
1078
1079		/**
1080	<	* Maximum size for submission queue array. Must be a power of two
423	<	* less than or equal to 1 << (31 - width of array entry) to
424	<	* ensure lack of index wraparound, but is capped at a lower
425	<	* value to help users trap runaway computations.
1080	>	* Common pool parallelism. Must equal commonPool.parallelism.
1081		*/
1082	<	private static final int MAXIMUM_QUEUE_CAPACITY = 1 << 24; // 16M
1082	>	static final int commonPoolParallelism;
1083
1084		/**
1085	<	* Array serving as submission queue. Initialized upon construction.
1085	>	* Sequence number for creating workerNamePrefix.
1086		*/
1087	<	private ForkJoinTask<?>[] submissionQueue;
1087	>	private static int poolNumberSequence;
1088
1089		/**
1090	<	* Lock protecting submissions array for addSubmission
1090	>	* Return the next sequence number. We don't expect this to
1091	>	* ever contend so use simple builtin sync.
1092		*/
1093	<	private final ReentrantLock submissionLock;
1093	>	private static final synchronized int nextPoolId() {
1094	>	return ++poolNumberSequence;
1095	>	}
1096	>
1097	>	// static constants
1098
1099		/**
1100	<	* Condition for awaitTermination, using submissionLock for
1101	<	* convenience.
1100	>	* Initial timeout value (in nanoseconds) for the thread
1101	>	* triggering quiescence to park waiting for new work. On timeout,
1102	>	* the thread will instead try to shrink the number of
1103	>	* workers. The value should be large enough to avoid overly
1104	>	* aggressive shrinkage during most transient stalls (long GCs
1105	>	* etc).
1106		*/
1107	<	private final Condition termination;
1107	>	private static final long IDLE_TIMEOUT      = 2000L * 1000L * 1000L; // 2sec
1108
1109		/**
1110	<	* Creation factory for worker threads.
1110	>	* Timeout value when there are more threads than parallelism level
1111		*/
1112	<	private final ForkJoinWorkerThreadFactory factory;
1112	>	private static final long FAST_IDLE_TIMEOUT =  200L * 1000L * 1000L;
1113
1114		/**
1115	<	* The uncaught exception handler used when any worker abruptly
452	<	* terminates.
1115	>	* Tolerance for idle timeouts, to cope with timer undershoots
1116		*/
1117	<	final Thread.UncaughtExceptionHandler ueh;
1117	>	private static final long TIMEOUT_SLOP = 2000000L;
1118
1119		/**
1120	<	* Prefix for assigning names to worker threads
1120	>	* The maximum stolen->joining link depth allowed in method
1121	>	* tryHelpStealer.  Must be a power of two.  Depths for legitimate
1122	>	* chains are unbounded, but we use a fixed constant to avoid
1123	>	* (otherwise unchecked) cycles and to bound staleness of
1124	>	* traversal parameters at the expense of sometimes blocking when
1125	>	* we could be helping.
1126		*/
1127	<	private final String workerNamePrefix;
1127	>	private static final int MAX_HELP = 64;
1128
1129		/**
1130	<	* Sum of per-thread steal counts, updated only when threads are
1131	<	* idle or terminating.
1130	>	* Increment for seed generators. See class ThreadLocal for
1131	>	* explanation.
1132		*/
1133	<	private volatile long stealCount;
1133	>	private static final int SEED_INCREMENT = 0x61c88647;
1134
1135		/**
1136	<	* Main pool control -- a long packed with:
1136	>	* Bits and masks for control variables
1137	>	*
1138	>	* Field ctl is a long packed with:
1139		* AC: Number of active running workers minus target parallelism (16 bits)
1140	<	* TC: Number of total workers minus target parallelism (16bits)
1140	>	* TC: Number of total workers minus target parallelism (16 bits)
1141		* ST: true if pool is terminating (1 bit)
1142		* EC: the wait count of top waiting thread (15 bits)
1143	<	* ID: ~poolIndex of top of Treiber stack of waiting threads (16 bits)
1143	>	* ID: poolIndex of top of Treiber stack of waiters (16 bits)
1144		*
1145		* When convenient, we can extract the upper 32 bits of counts and
1146		* the lower 32 bits of queue state, u = (int)(ctl >>> 32) and e =
#	Line 479 | Line 1149 | public class ForkJoinPool extends Abstra
1149		* parallelism and the positionings of fields makes it possible to
1150		* perform the most common checks via sign tests of fields: When
1151		* ac is negative, there are not enough active workers, when tc is
1152	<	* negative, there are not enough total workers, when id is
483	<	* negative, there is at least one waiting worker, and when e is
1152	>	* negative, there are not enough total workers, and when e is
1153		* negative, the pool is terminating.  To deal with these possibly
1154		* negative fields, we use casts in and out of "short" and/or
1155		* signed shifts to maintain signedness.
1156	+	*
1157	+	* When a thread is queued (inactivated), its eventCount field is
1158	+	* set negative, which is the only way to tell if a worker is
1159	+	* prevented from executing tasks, even though it must continue to
1160	+	* scan for them to avoid queuing races. Note however that
1161	+	* eventCount updates lag releases so usage requires care.
1162	+	*
1163	+	* Field plock is an int packed with:
1164	+	* SHUTDOWN: true if shutdown is enabled (1 bit)
1165	+	* SEQ:  a sequence lock, with PL_LOCK bit set if locked (30 bits)
1166	+	* SIGNAL: set when threads may be waiting on the lock (1 bit)
1167	+	*
1168	+	* The sequence number enables simple consistency checks:
1169	+	* Staleness of read-only operations on the workQueues array can
1170	+	* be checked by comparing plock before vs after the reads.
1171		*/
488	–	volatile long ctl;
1172
1173		// bit positions/shifts for fields
1174		private static final int  AC_SHIFT   = 48;
#	Line 494 | Line 1177 | public class ForkJoinPool extends Abstra
1177		private static final int  EC_SHIFT   = 16;
1178
1179		// bounds
1180	<	private static final int  MAX_ID     = 0x7fff;  // max poolIndex
1181	<	private static final int  SMASK      = 0xffff;  // mask short bits
1180	>	private static final int  SMASK      = 0xffff;  // short bits
1181	>	private static final int  MAX_CAP    = 0x7fff;  // max #workers - 1
1182	>	private static final int  EVENMASK   = 0xfffe;  // even short bits
1183	>	private static final int  SQMASK     = 0x007e;  // max 64 (even) slots
1184		private static final int  SHORT_SIGN = 1 << 15;
1185		private static final int  INT_SIGN   = 1 << 31;
1186
#	Line 517 | Line 1202 | public class ForkJoinPool extends Abstra
1202		private static final int  UTC_UNIT   = 1 << UTC_SHIFT;
1203
1204		// masks and units for dealing with e = (int)ctl
1205	<	private static final int  E_MASK     = 0x7fffffff; // no STOP_BIT
1206	<	private static final int  EC_UNIT    = 1 << EC_SHIFT;
1205	>	private static final int E_MASK      = 0x7fffffff; // no STOP_BIT
1206	>	private static final int E_SEQ       = 1 << EC_SHIFT;
1207
1208	<	/**
1209	<	* The target parallelism level.
1210	<	*/
1211	<	final int parallelism;
1208	>	// plock bits
1209	>	private static final int SHUTDOWN    = 1 << 31;
1210	>	private static final int PL_LOCK     = 2;
1211	>	private static final int PL_SIGNAL   = 1;
1212	>	private static final int PL_SPINS    = 1 << 8;
1213	>
1214	>	// access mode for WorkQueue
1215	>	static final int LIFO_QUEUE          =  0;
1216	>	static final int FIFO_QUEUE          =  1;
1217	>	static final int SHARED_QUEUE        = -1;
1218	>
1219	>	// bounds for #steps in scan loop -- must be power 2 minus 1
1220	>	private static final int MIN_SCAN    = 0x1ff;   // cover estimation slop
1221	>	private static final int MAX_SCAN    = 0x1ffff; // 4 * max workers
1222
1223	<	/**
529	<	* Index (mod submission queue length) of next element to take
530	<	* from submission queue. Usage is identical to that for
531	<	* per-worker queues -- see ForkJoinWorkerThread internal
532	<	* documentation.
533	<	*/
534	<	volatile int queueBase;
1223	>	// Instance fields
1224
1225	<	/**
1226	<	* Index (mod submission queue length) of next element to add
1227	<	* in submission queue. Usage is identical to that for
1228	<	* per-worker queues -- see ForkJoinWorkerThread internal
1229	<	* documentation.
1230	<	*/
1231	<	int queueTop;
1225	>	/*
1226	>	* Field layout of this class tends to matter more than one would
1227	>	* like. Runtime layout order is only loosely related to
1228	>	* declaration order and may differ across JVMs, but the following
1229	>	* empirically works OK on current JVMs.
1230	>	*/
1231	>
1232	>	// Heuristic padding to ameliorate unfortunate memory placements
1233	>	volatile long pad00, pad01, pad02, pad03, pad04, pad05, pad06;
1234	>
1235	>	volatile long stealCount;                  // collects worker counts
1236	>	volatile long ctl;                         // main pool control
1237	>	volatile int plock;                        // shutdown status and seqLock
1238	>	volatile int indexSeed;                    // worker/submitter index seed
1239	>	final int config;                          // mode and parallelism level
1240	>	WorkQueue[] workQueues;                    // main registry
1241	>	final ForkJoinWorkerThreadFactory factory;
1242	>	final Thread.UncaughtExceptionHandler ueh; // per-worker UEH
1243	>	final String workerNamePrefix;             // to create worker name string
1244
1245	<	/**
1246	<	* True when shutdown() has been called.
546	<	*/
547	<	volatile boolean shutdown;
1245	>	volatile Object pad10, pad11, pad12, pad13, pad14, pad15, pad16, pad17;
1246	>	volatile Object pad18, pad19, pad1a, pad1b;
1247
1248	<	/**
1249	<	* True if use local fifo, not default lifo, for local polling
1250	<	* Read by, and replicated by ForkJoinWorkerThreads
1251	<	*/
1252	<	final boolean locallyFifo;
1248	>	/*
1249	>	* Acquires the plock lock to protect worker array and related
1250	>	* updates. This method is called only if an initial CAS on plock
1251	>	* fails. This acts as a spinLock for normal cases, but falls back
1252	>	* to builtin monitor to block when (rarely) needed. This would be
1253	>	* a terrible idea for a highly contended lock, but works fine as
1254	>	* a more conservative alternative to a pure spinlock.
1255	>	*/
1256	>	private int acquirePlock() {
1257	>	int spins = PL_SPINS, r = 0, ps, nps;
1258	>	for (;;) {
1259	>	if (((ps = plock) & PL_LOCK) == 0 &&
1260	>	U.compareAndSwapInt(this, PLOCK, ps, nps = ps + PL_LOCK))
1261	>	return nps;
1262	>	else if (r == 0) { // randomize spins if possible
1263	>	Thread t = Thread.currentThread(); WorkQueue w; Submitter z;
1264	>	if ((t instanceof ForkJoinWorkerThread) &&
1265	>	(w = ((ForkJoinWorkerThread)t).workQueue) != null)
1266	>	r = w.seed;
1267	>	else if ((z = submitters.get()) != null)
1268	>	r = z.seed;
1269	>	else
1270	>	r = 1;
1271	>	}
1272	>	else if (spins >= 0) {
1273	>	r ^= r << 1; r ^= r >>> 3; r ^= r << 10; // xorshift
1274	>	if (r >= 0)
1275	>	--spins;
1276	>	}
1277	>	else if (U.compareAndSwapInt(this, PLOCK, ps, ps | PL_SIGNAL)) {
1278	>	synchronized (this) {
1279	>	if ((plock & PL_SIGNAL) != 0) {
1280	>	try {
1281	>	wait();
1282	>	} catch (InterruptedException ie) {
1283	>	try {
1284	>	Thread.currentThread().interrupt();
1285	>	} catch (SecurityException ignore) {
1286	>	}
1287	>	}
1288	>	}
1289	>	else
1290	>	notifyAll();
1291	>	}
1292	>	}
1293	>	}
1294	>	}
1295
1296		/**
1297	<	* The number of threads in ForkJoinWorkerThreads.helpQuiescePool.
1298	<	* When non-zero, suppresses automatic shutdown when active
558	<	* counts become zero.
1297	>	* Unlocks and signals any thread waiting for plock. Called only
1298	>	* when CAS of seq value for unlock fails.
1299		*/
1300	<	volatile int quiescerCount;
1300	>	private void releasePlock(int ps) {
1301	>	plock = ps;
1302	>	synchronized (this) { notifyAll(); }
1303	>	}
1304
1305		/**
1306	<	* The number of threads blocked in join.
1307	<	*/
1308	<	volatile int blockedCount;
1306	>	* Performs secondary initialization, called when plock is zero.
1307	>	* Creates workQueue array and sets plock to a valid value.  The
1308	>	* lock body must be exception-free (so no try/finally) so we
1309	>	* optimistically allocate new array outside the lock and throw
1310	>	* away if (very rarely) not needed. (A similar tactic is used in
1311	>	* fullExternalPush.)  Because the plock seq value can eventually
1312	>	* wrap around zero, this method harmlessly fails to reinitialize
1313	>	* if workQueues exists, while still advancing plock.
1314	>	*
1315	>	* Additionally tries to create the first worker.
1316	>	*/
1317	>	private void initWorkers() {
1318	>	WorkQueue[] ws, nws; int ps;
1319	>	int p = config & SMASK;        // find power of two table size
1320	>	int n = (p > 1) ? p - 1 : 1;   // ensure at least 2 slots
1321	>	n |= n >>> 1; n |= n >>> 2; n |= n >>> 4; n |= n >>> 8; n |= n >>> 16;
1322	>	n = (n + 1) << 1;
1323	>	if ((ws = workQueues) == null || ws.length == 0)
1324	>	nws = new WorkQueue[n];
1325	>	else
1326	>	nws = null;
1327	>	if (((ps = plock) & PL_LOCK) != 0 ||
1328	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1329	>	ps = acquirePlock();
1330	>	if (((ws = workQueues) == null || ws.length == 0) && nws != null)
1331	>	workQueues = nws;
1332	>	int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1333	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1334	>	releasePlock(nps);
1335	>	tryAddWorker();
1336	>	}
1337
1338		/**
1339	<	* Counter for worker Thread names (unrelated to their poolIndex)
1340	<	*/
1341	<	private volatile int nextWorkerNumber;
1339	>	* Tries to create and start one worker if fewer than target
1340	>	* parallelism level exist. Adjusts counts etc on failure.
1341	>	*/
1342	>	private void tryAddWorker() {
1343	>	long c; int u;
1344	>	while ((u = (int)((c = ctl) >>> 32)) < 0 &&
1345	>	(u & SHORT_SIGN) != 0 && (int)c == 0) {
1346	>	long nc = (long)(((u + UTC_UNIT) & UTC_MASK) |
1347	>	((u + UAC_UNIT) & UAC_MASK)) << 32;
1348	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1349	>	ForkJoinWorkerThreadFactory fac;
1350	>	Throwable ex = null;
1351	>	ForkJoinWorkerThread wt = null;
1352	>	try {
1353	>	if ((fac = factory) != null &&
1354	>	(wt = fac.newThread(this)) != null) {
1355	>	wt.start();
1356	>	break;
1357	>	}
1358	>	} catch (Throwable e) {
1359	>	ex = e;
1360	>	}
1361	>	deregisterWorker(wt, ex);
1362	>	break;
1363	>	}
1364	>	}
1365	>	}
1366
1367	<	/**
573	<	* The index for the next created worker. Accessed under scanGuard.
574	<	*/
575	<	private int nextWorkerIndex;
1367	>	//  Registering and deregistering workers
1368
1369		/**
1370	<	* SeqLock and index masking for updates to workers array.  Locked
1371	<	* when SG_UNIT is set. Unlocking clears bit by adding
1372	<	* SG_UNIT. Staleness of read-only operations can be checked by
1373	<	* comparing scanGuard to value before the reads. The low 16 bits
1374	<	* (i.e, anding with SMASK) hold (the smallest power of two
1375	<	* covering all worker indices, minus one, and is used to avoid
1376	<	* dealing with large numbers of null slots when the workers array
1377	<	* is overallocated.
1378	<	*/
1379	<	volatile int scanGuard;
1380	<
1381	<	private static final int SG_UNIT = 1 << 16;
1370	>	* Callback from ForkJoinWorkerThread to establish and record its
1371	>	* WorkQueue. To avoid scanning bias due to packing entries in
1372	>	* front of the workQueues array, we treat the array as a simple
1373	>	* power-of-two hash table using per-thread seed as hash,
1374	>	* expanding as needed.
1375	>	*
1376	>	* @param wt the worker thread
1377	>	* @return the worker's queue
1378	>	*/
1379	>	final WorkQueue registerWorker(ForkJoinWorkerThread wt) {
1380	>	Thread.UncaughtExceptionHandler handler; WorkQueue[] ws; int s, ps;
1381	>	wt.setDaemon(true);
1382	>	if ((handler = ueh) != null)
1383	>	wt.setUncaughtExceptionHandler(handler);
1384	>	do {} while (!U.compareAndSwapInt(this, INDEXSEED, s = indexSeed,
1385	>	s += SEED_INCREMENT) ||
1386	>	s == 0); // skip 0
1387	>	WorkQueue w = new WorkQueue(this, wt, config >>> 16, s);
1388	>	if (((ps = plock) & PL_LOCK) != 0 ||
1389	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1390	>	ps = acquirePlock();
1391	>	int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1392	>	try {
1393	>	if ((ws = workQueues) != null) {    // skip if shutting down
1394	>	int n = ws.length, m = n - 1;
1395	>	int r = (s << 1) | 1;           // use odd-numbered indices
1396	>	if (ws[r &= m] != null) {       // collision
1397	>	int probes = 0;             // step by approx half size
1398	>	int step = (n <= 4) ? 2 : ((n >>> 1) & EVENMASK) + 2;
1399	>	while (ws[r = (r + step) & m] != null) {
1400	>	if (++probes >= n) {
1401	>	workQueues = ws = Arrays.copyOf(ws, n <<= 1);
1402	>	m = n - 1;
1403	>	probes = 0;
1404	>	}
1405	>	}
1406	>	}
1407	>	w.eventCount = w.poolIndex = r; // volatile write orders
1408	>	ws[r] = w;
1409	>	}
1410	>	} finally {
1411	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1412	>	releasePlock(nps);
1413	>	}
1414	>	wt.setName(workerNamePrefix.concat(Integer.toString(w.poolIndex)));
1415	>	return w;
1416	>	}
1417
1418		/**
1419	<	* The wakeup interval (in nanoseconds) for a worker waiting for a
1420	<	* task when the pool is quiescent to instead try to shrink the
1421	<	* number of workers.  The exact value does not matter too
1422	<	* much. It must be short enough to release resources during
1423	<	* sustained periods of idleness, but not so short that threads
1424	<	* are continually re-created.
1425	<	*/
1426	<	private static final long SHRINK_RATE =
1427	<	4L * 1000L * 1000L * 1000L; // 4 seconds
1419	>	* Final callback from terminating worker, as well as upon failure
1420	>	* to construct or start a worker.  Removes record of worker from
1421	>	* array, and adjusts counts. If pool is shutting down, tries to
1422	>	* complete termination.
1423	>	*
1424	>	* @param wt the worker thread or null if construction failed
1425	>	* @param ex the exception causing failure, or null if none
1426	>	*/
1427	>	final void deregisterWorker(ForkJoinWorkerThread wt, Throwable ex) {
1428	>	WorkQueue w = null;
1429	>	if (wt != null && (w = wt.workQueue) != null) {
1430	>	int ps;
1431	>	w.qlock = -1;                // ensure set
1432	>	long ns = w.nsteals, sc;     // collect steal count
1433	>	do {} while (!U.compareAndSwapLong(this, STEALCOUNT,
1434	>	sc = stealCount, sc + ns));
1435	>	if (((ps = plock) & PL_LOCK) != 0 ||
1436	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1437	>	ps = acquirePlock();
1438	>	int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1439	>	try {
1440	>	int idx = w.poolIndex;
1441	>	WorkQueue[] ws = workQueues;
1442	>	if (ws != null && idx >= 0 && idx < ws.length && ws[idx] == w)
1443	>	ws[idx] = null;
1444	>	} finally {
1445	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1446	>	releasePlock(nps);
1447	>	}
1448	>	}
1449
1450	<	/**
1451	<	* Top-level loop for worker threads: On each step: if the
1452	<	* previous step swept through all queues and found no tasks, or
1453	<	* there are excess threads, then possibly blocks. Otherwise,
1454	<	* scans for and, if found, executes a task. Returns when pool
1455	<	* and/or worker terminate.
1456	<	*
1457	<	* @param w the worker
1458	<	*/
1459	<	final void work(ForkJoinWorkerThread w) {
1460	<	boolean swept = false;                // true on empty scans
1461	<	long c;
1462	<	while (!w.terminate && (int)(c = ctl) >= 0) {
1463	<	int a;                            // active count
1464	<	if (!swept && (a = (int)(c >> AC_SHIFT)) <= 0)
1465	<	swept = scan(w, a);
1466	<	else if (tryAwaitWork(w, c))
1467	<	swept = false;
1450	>	long c;                          // adjust ctl counts
1451	>	do {} while (!U.compareAndSwapLong
1452	>	(this, CTL, c = ctl, (((c - AC_UNIT) & AC_MASK) |
1453	>	((c - TC_UNIT) & TC_MASK) |
1454	>	(c & ~(AC_MASK|TC_MASK)))));
1455	>
1456	>	if (!tryTerminate(false, false) && w != null && w.array != null) {
1457	>	w.cancelAll();               // cancel remaining tasks
1458	>	WorkQueue[] ws; WorkQueue v; Thread p; int u, i, e;
1459	>	while ((u = (int)((c = ctl) >>> 32)) < 0 && (e = (int)c) >= 0) {
1460	>	if (e > 0) {             // activate or create replacement
1461	>	if ((ws = workQueues) == null ||
1462	>	(i = e & SMASK) >= ws.length ||
1463	>	(v = ws[i]) == null)
1464	>	break;
1465	>	long nc = (((long)(v.nextWait & E_MASK)) |
1466	>	((long)(u + UAC_UNIT) << 32));
1467	>	if (v.eventCount != (e | INT_SIGN))
1468	>	break;
1469	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1470	>	v.eventCount = (e + E_SEQ) & E_MASK;
1471	>	if ((p = v.parker) != null)
1472	>	U.unpark(p);
1473	>	break;
1474	>	}
1475	>	}
1476	>	else {
1477	>	if ((short)u < 0)
1478	>	tryAddWorker();
1479	>	break;
1480	>	}
1481	>	}
1482		}
1483	+	if (ex == null)                     // help clean refs on way out
1484	+	ForkJoinTask.helpExpungeStaleExceptions();
1485	+	else                                // rethrow
1486	+	ForkJoinTask.rethrow(ex);
1487		}
1488
1489	<	// Signalling
1489	>	// Submissions
1490
1491		/**
1492	<	* Wakes up or creates a worker.
1493	<	*/
1494	<	final void signalWork() {
1495	<	/*
1496	<	* The while condition is true if: (there is are too few total
1497	<	* workers OR there is at least one waiter) AND (there are too
1498	<	* few active workers OR the pool is terminating).  The value
1499	<	* of e distinguishes the remaining cases: zero (no waiters)
1500	<	* for create, negative if terminating (in which case do
1501	<	* nothing), else release a waiter. The secondary checks for
1502	<	* release (non-null array etc) can fail if the pool begins
1503	<	* terminating after the test, and don't impose any added cost
1504	<	* because JVMs must perform null and bounds checks anyway.
1505	<	*/
1506	<	long c; int e, u;
1507	<	while ((((e = (int)(c = ctl)) | (u = (int)(c >>> 32))) &
1508	<	(INT_SIGN|SHORT_SIGN)) == (INT_SIGN|SHORT_SIGN) && e >= 0) {
1509	<	if (e > 0) {                         // release a waiting worker
1510	<	int i; ForkJoinWorkerThread w; ForkJoinWorkerThread[] ws;
1511	<	if ((ws = workers) == null ||
1512	<	(i = ~e & SMASK) >= ws.length ||
1513	<	(w = ws[i]) == null)
1514	<	break;
1515	<	long nc = (((long)(w.nextWait & E_MASK)) |
1516	<	((long)(u + UAC_UNIT) << 32));
1517	<	if (w.eventCount == e &&
1518	<	UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
1519	<	w.eventCount = (e + EC_UNIT) & E_MASK;
1520	<	if (w.parked)
1521	<	UNSAFE.unpark(w);
1522	<	break;
1492	>	* Unless shutting down, adds the given task to a submission queue
1493	>	* at submitter's current queue index (modulo submission
1494	>	* range). Only the most common path is directly handled in this
1495	>	* method. All others are relayed to fullExternalPush.
1496	>	*
1497	>	* @param task the task. Caller must ensure non-null.
1498	>	*/
1499	>	final void externalPush(ForkJoinTask<?> task) {
1500	>	WorkQueue[] ws; WorkQueue q; Submitter z; int m; ForkJoinTask<?>[] a;
1501	>	if ((z = submitters.get()) != null && plock > 0 &&
1502	>	(ws = workQueues) != null && (m = (ws.length - 1)) >= 0 &&
1503	>	(q = ws[m & z.seed & SQMASK]) != null &&
1504	>	U.compareAndSwapInt(q, QLOCK, 0, 1)) { // lock
1505	>	int b = q.base, s = q.top, n, an;
1506	>	if ((a = q.array) != null && (an = a.length) > (n = s + 1 - b)) {
1507	>	int j = (((an - 1) & s) << ASHIFT) + ABASE;
1508	>	U.putOrderedObject(a, j, task);
1509	>	q.top = s + 1;                     // push on to deque
1510	>	q.qlock = 0;
1511	>	if (n <= 2)
1512	>	signalWork(q);
1513	>	return;
1514	>	}
1515	>	q.qlock = 0;
1516	>	}
1517	>	fullExternalPush(task);
1518	>	}
1519	>
1520	>	/**
1521	>	* Full version of externalPush. This method is called, among
1522	>	* other times, upon the first submission of the first task to the
1523	>	* pool, so must perform secondary initialization (via
1524	>	* initWorkers). It also detects first submission by an external
1525	>	* thread by looking up its ThreadLocal, and creates a new shared
1526	>	* queue if the one at index if empty or contended. The plock lock
1527	>	* body must be exception-free (so no try/finally) so we
1528	>	* optimistically allocate new queues outside the lock and throw
1529	>	* them away if (very rarely) not needed.
1530	>	*/
1531	>	private void fullExternalPush(ForkJoinTask<?> task) {
1532	>	int r = 0; // random index seed
1533	>	for (Submitter z = submitters.get();;) {
1534	>	WorkQueue[] ws; WorkQueue q; int ps, m, k;
1535	>	if (z == null) {
1536	>	if (U.compareAndSwapInt(this, INDEXSEED, r = indexSeed,
1537	>	r += SEED_INCREMENT) && r != 0)
1538	>	submitters.set(z = new Submitter(r));
1539	>	}
1540	>	else if (r == 0) {               // move to a different index
1541	>	r = z.seed;
1542	>	r ^= r << 13;                // same xorshift as WorkQueues
1543	>	r ^= r >>> 17;
1544	>	z.seed = r ^ (r << 5);
1545	>	}
1546	>	else if ((ps = plock) < 0)
1547	>	throw new RejectedExecutionException();
1548	>	else if (ps == 0 || (ws = workQueues) == null ||
1549	>	(m = ws.length - 1) < 0)
1550	>	initWorkers();
1551	>	else if ((q = ws[k = r & m & SQMASK]) != null) {
1552	>	if (q.qlock == 0 && U.compareAndSwapInt(q, QLOCK, 0, 1)) {
1553	>	ForkJoinTask<?>[] a = q.array;
1554	>	int s = q.top;
1555	>	boolean submitted = false;
1556	>	try {                      // locked version of push
1557	>	if ((a != null && a.length > s + 1 - q.base) ||
1558	>	(a = q.growArray()) != null) {   // must presize
1559	>	int j = (((a.length - 1) & s) << ASHIFT) + ABASE;
1560	>	U.putOrderedObject(a, j, task);
1561	>	q.top = s + 1;
1562	>	submitted = true;
1563	>	}
1564	>	} finally {
1565	>	q.qlock = 0;  // unlock
1566	>	}
1567	>	if (submitted) {
1568	>	signalWork(q);
1569	>	return;
1570	>	}
1571		}
1572	+	r = 0; // move on failure
1573		}
1574	<	else if (UNSAFE.compareAndSwapLong
1575	<	(this, ctlOffset, c,
1576	<	(long)(((u + UTC_UNIT) & UTC_MASK) |
1577	<	((u + UAC_UNIT) & UAC_MASK)) << 32)) {
1578	<	addWorker();
1579	<	break;
1574	>	else if (((ps = plock) & PL_LOCK) == 0) { // create new queue
1575	>	q = new WorkQueue(this, null, SHARED_QUEUE, r);
1576	>	if (((ps = plock) & PL_LOCK) != 0 ||
1577	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1578	>	ps = acquirePlock();
1579	>	if ((ws = workQueues) != null && k < ws.length && ws[k] == null)
1580	>	ws[k] = q;
1581	>	int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1582	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1583	>	releasePlock(nps);
1584		}
1585	+	else
1586	+	r = 0; // try elsewhere while lock held
1587		}
1588		}
1589
1590	+	// Maintaining ctl counts
1591	+
1592		/**
1593	<	* Variant of signalWork to help release waiters on rescans.
671	<	* Tries once to release a waiter if active count < 0.
672	<	*
673	<	* @return false if failed due to contention, else true
1593	>	* Increments active count; mainly called upon return from blocking.
1594		*/
1595	<	private boolean tryReleaseWaiter() {
1596	<	long c; int e, i; ForkJoinWorkerThread w; ForkJoinWorkerThread[] ws;
1597	<	if ((e = (int)(c = ctl)) > 0 &&
678	<	(int)(c >> AC_SHIFT) < 0 &&
679	<	(ws = workers) != null &&
680	<	(i = ~e & SMASK) < ws.length &&
681	<	(w = ws[i]) != null) {
682	<	long nc = ((long)(w.nextWait & E_MASK) |
683	<	((c + AC_UNIT) & (AC_MASK|TC_MASK)));
684	<	if (w.eventCount != e ||
685	<	!UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc))
686	<	return false;
687	<	w.eventCount = (e + EC_UNIT) & E_MASK;
688	<	if (w.parked)
689	<	UNSAFE.unpark(w);
690	<	}
691	<	return true;
1595	>	final void incrementActiveCount() {
1596	>	long c;
1597	>	do {} while (!U.compareAndSwapLong(this, CTL, c = ctl, c + AC_UNIT));
1598		}
1599
694	–	// Scanning for tasks
695	–
1600		/**
1601	<	* Scans for and, if found, executes one task. Scans start at a
1602	<	* random index of workers array, and randomly select the first
1603	<	* (2*#workers)-1 probes, and then, if all empty, resort to 2
1604	<	* circular sweeps, which is necessary to check quiescence. and
1605	<	* taking a submission only if no stealable tasks were found.  The
1606	<	* steal code inside the loop is a specialized form of
1607	<	* ForkJoinWorkerThread.deqTask, followed bookkeeping to support
1608	<	* helpJoinTask and signal propagation. The code for submission
1609	<	* queues is almost identical. On each steal, the worker completes
1610	<	* not only the task, but also all local tasks that this task may
1611	<	* have generated. On detecting staleness or contention when
1612	<	* trying to take a task, this method returns without finishing
1613	<	* sweep, which allows global state rechecks before retry.
1614	<	*
1615	<	* @param w the worker
1616	<	* @param a the number of active workers
1617	<	* @return true if swept all queues without finding a task
1618	<	*/
1619	<	private boolean scan(ForkJoinWorkerThread w, int a) {
1620	<	int g = scanGuard; // mask 0 avoids useless scans if only one active
1621	<	int m = (parallelism == 1 - a && blockedCount == 0) ? 0 : g & SMASK;
1622	<	ForkJoinWorkerThread[] ws = workers;
719	<	if (ws == null || ws.length <= m)         // staleness check
720	<	return false;
721	<	for (int r = w.seed, k = r, j = -(m + m); j <= m + m; ++j) {
722	<	ForkJoinTask<?> t; ForkJoinTask<?>[] q; int b, i;
723	<	ForkJoinWorkerThread v = ws[k & m];
724	<	if (v != null && (b = v.queueBase) != v.queueTop &&
725	<	(q = v.queue) != null && (i = (q.length - 1) & b) >= 0) {
726	<	long u = (i << ASHIFT) + ABASE;
727	<	if ((t = q[i]) != null && v.queueBase == b &&
728	<	UNSAFE.compareAndSwapObject(q, u, t, null)) {
729	<	int d = (v.queueBase = b + 1) - v.queueTop;
730	<	v.stealHint = w.poolIndex;
731	<	if (d != 0)
732	<	signalWork();             // propagate if nonempty
733	<	w.execTask(t);
1601	>	* Tries to create or activate a worker if too few are active.
1602	>	*
1603	>	* @param q the (non-null) queue holding tasks to be signalled
1604	>	*/
1605	>	final void signalWork(WorkQueue q) {
1606	>	int hint = q.poolIndex;
1607	>	long c; int e, u, i, n; WorkQueue[] ws; WorkQueue w; Thread p;
1608	>	while ((u = (int)((c = ctl) >>> 32)) < 0) {
1609	>	if ((e = (int)c) > 0) {
1610	>	if ((ws = workQueues) != null && ws.length > (i = e & SMASK) &&
1611	>	(w = ws[i]) != null && w.eventCount == (e | INT_SIGN)) {
1612	>	long nc = (((long)(w.nextWait & E_MASK)) |
1613	>	((long)(u + UAC_UNIT) << 32));
1614	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1615	>	w.hint = hint;
1616	>	w.eventCount = (e + E_SEQ) & E_MASK;
1617	>	if ((p = w.parker) != null)
1618	>	U.unpark(p);
1619	>	break;
1620	>	}
1621	>	if (q.top - q.base <= 0)
1622	>	break;
1623		}
1624	<	r ^= r << 13; r ^= r >>> 17; w.seed = r ^ (r << 5);
1625	<	return false;                     // store next seed
1624	>	else
1625	>	break;
1626		}
1627	<	else if (j < 0) {                     // xorshift
1628	<	r ^= r << 13; r ^= r >>> 17; k = r ^= r << 5;
1627	>	else {
1628	>	if ((short)u < 0)
1629	>	tryAddWorker();
1630	>	break;
1631		}
741	–	else
742	–	++k;
743	–	}
744	–	if (scanGuard != g)                       // staleness check
745	–	return false;
746	–	else {                                    // try to take submission
747	–	ForkJoinTask<?> t; ForkJoinTask<?>[] q; int b, i;
748	–	if ((b = queueBase) != queueTop &&
749	–	(q = submissionQueue) != null &&
750	–	(i = (q.length - 1) & b) >= 0) {
751	–	long u = (i << ASHIFT) + ABASE;
752	–	if ((t = q[i]) != null && queueBase == b &&
753	–	UNSAFE.compareAndSwapObject(q, u, t, null)) {
754	–	queueBase = b + 1;
755	–	w.execTask(t);
756	–	}
757	–	return false;
758	–	}
759	–	return true;                         // all queues empty
1632		}
1633		}
1634
1635	+	// Scanning for tasks
1636	+
1637		/**
1638	<	* Tries to enqueue worker w in wait queue and await change in
765	<	* worker's eventCount.  If the pool is quiescent and there is
766	<	* more than one worker, possibly terminates worker upon exit.
767	<	* Otherwise, before blocking, rescans queues to avoid missed
768	<	* signals.  Upon finding work, releases at least one worker
769	<	* (which may be the current worker). Rescans restart upon
770	<	* detected staleness or failure to release due to
771	<	* contention. Note the unusual conventions about Thread.interrupt
772	<	* here and elsewhere: Because interrupts are used solely to alert
773	<	* threads to check termination, which is checked here anyway, we
774	<	* clear status (using Thread.interrupted) before any call to
775	<	* park, so that park does not immediately return due to status
776	<	* being set via some other unrelated call to interrupt in user
777	<	* code.
778	<	*
779	<	* @param w the calling worker
780	<	* @param c the ctl value on entry
781	<	* @return true if waited or another thread was released upon enq
1638	>	* Top-level runloop for workers, called by ForkJoinWorkerThread.run.
1639		*/
1640	<	private boolean tryAwaitWork(ForkJoinWorkerThread w, long c) {
1641	<	int v = w.eventCount;
1642	<	w.nextWait = (int)c;                      // w's successor record
1643	<	long nc = (long)(v & E_MASK) | ((c - AC_UNIT) & (AC_MASK|TC_MASK));
1644	<	if (ctl != c || !UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
1645	<	long d = ctl; // return true if lost to a deq, to force scan
1646	<	return (int)d != (int)c && ((d - c) & AC_MASK) >= 0L;
1647	<	}
1648	<	for (int sc = w.stealCount; sc != 0;) {   // accumulate stealCount
1649	<	long s = stealCount;
1650	<	if (UNSAFE.compareAndSwapLong(this, stealCountOffset, s, s + sc))
1651	<	sc = w.stealCount = 0;
1652	<	else if (w.eventCount != v)
1653	<	return true;                      // update next time
1654	<	}
1655	<	if ((int)c != 0 && parallelism + (int)(nc >> AC_SHIFT) == 0 &&
1656	<	blockedCount == 0 && quiescerCount == 0)
1657	<	idleAwaitWork(w, nc, c, v);           // quiescent
1658	<	for (boolean rescanned = false;;) {
1659	<	if (w.eventCount != v)
1660	<	return true;
1661	<	if (!rescanned) {
1662	<	int g = scanGuard, m = g & SMASK;
1663	<	ForkJoinWorkerThread[] ws = workers;
1664	<	if (ws != null && m < ws.length) {
1665	<	rescanned = true;
1666	<	for (int i = 0; i <= m; ++i) {
1667	<	ForkJoinWorkerThread u = ws[i];
1668	<	if (u != null) {
1669	<	if (u.queueBase != u.queueTop &&
1670	<	!tryReleaseWaiter())
1671	<	rescanned = false; // contended
1672	<	if (w.eventCount != v)
1673	<	return true;
1674	<	}
1640	>	final void runWorker(WorkQueue w) {
1641	>	w.growArray(); // allocate queue
1642	>	do { w.runTask(scan(w)); } while (w.qlock >= 0);
1643	>	}
1644	>
1645	>	/**
1646	>	* Scans for and, if found, returns one task, else possibly
1647	>	* inactivates the worker. This method operates on single reads of
1648	>	* volatile state and is designed to be re-invoked continuously,
1649	>	* in part because it returns upon detecting inconsistencies,
1650	>	* contention, or state changes that indicate possible success on
1651	>	* re-invocation.
1652	>	*
1653	>	* The scan searches for tasks across queues (starting at a random
1654	>	* index, and relying on registerWorker to irregularly scatter
1655	>	* them within array to avoid bias), checking each at least twice.
1656	>	* The scan terminates upon either finding a non-empty queue, or
1657	>	* completing the sweep. If the worker is not inactivated, it
1658	>	* takes and returns a task from this queue. Otherwise, if not
1659	>	* activated, it signals workers (that may include itself) and
1660	>	* returns so caller can retry. Also returns for true if the
1661	>	* worker array may have changed during an empty scan.  On failure
1662	>	* to find a task, we take one of the following actions, after
1663	>	* which the caller will retry calling this method unless
1664	>	* terminated.
1665	>	*
1666	>	* * If pool is terminating, terminate the worker.
1667	>	*
1668	>	* * If not already enqueued, try to inactivate and enqueue the
1669	>	* worker on wait queue. Or, if inactivating has caused the pool
1670	>	* to be quiescent, relay to idleAwaitWork to possibly shrink
1671	>	* pool.
1672	>	*
1673	>	* * If already enqueued and none of the above apply, possibly
1674	>	* park awaiting signal, else lingering to help scan and signal.
1675	>	*
1676	>	* * If a non-empty queue discovered or left as a hint,
1677	>	* help wake up other workers before return
1678	>	*
1679	>	* @param w the worker (via its WorkQueue)
1680	>	* @return a task or null if none found
1681	>	*/
1682	>	private final ForkJoinTask<?> scan(WorkQueue w) {
1683	>	WorkQueue[] ws; int m;
1684	>	int ps = plock;                          // read plock before ws
1685	>	if (w != null && (ws = workQueues) != null && (m = ws.length - 1) >= 0) {
1686	>	int ec = w.eventCount;               // ec is negative if inactive
1687	>	int r = w.seed; r ^= r << 13; r ^= r >>> 17; w.seed = r ^= r << 5;
1688	>	w.hint = -1;                         // update seed and clear hint
1689	>	int j = ((m + m + 1) | MIN_SCAN) & MAX_SCAN;
1690	>	do {
1691	>	WorkQueue q; ForkJoinTask<?>[] a; int b;
1692	>	if ((q = ws[(r + j) & m]) != null && (b = q.base) - q.top < 0 &&
1693	>	(a = q.array) != null) {     // probably nonempty
1694	>	int i = (((a.length - 1) & b) << ASHIFT) + ABASE;
1695	>	ForkJoinTask<?> t = (ForkJoinTask<?>)
1696	>	U.getObjectVolatile(a, i);
1697	>	if (q.base == b && ec >= 0 && t != null &&
1698	>	U.compareAndSwapObject(a, i, t, null)) {
1699	>	if ((q.base = b + 1) - q.top < 0)
1700	>	signalWork(q);
1701	>	return t;                // taken
1702	>	}
1703	>	else if ((ec < 0 || j < m) && (int)(ctl >> AC_SHIFT) <= 0) {
1704	>	w.hint = (r + j) & m;    // help signal below
1705	>	break;                   // cannot take
1706		}
1707		}
1708	<	if (scanGuard != g ||              // stale
1709	<	(queueBase != queueTop && !tryReleaseWaiter()))
1710	<	rescanned = false;
1711	<	if (!rescanned)
1712	<	Thread.yield();                // reduce contention
1713	<	else
1714	<	Thread.interrupted();          // clear before park
1715	<	}
1708	>	} while (--j >= 0);
1709	>
1710	>	int h, e, ns; long c, sc; WorkQueue q;
1711	>	if ((ns = w.nsteals) != 0) {
1712	>	if (U.compareAndSwapLong(this, STEALCOUNT,
1713	>	sc = stealCount, sc + ns))
1714	>	w.nsteals = 0;               // collect steals and rescan
1715	>	}
1716	>	else if (plock != ps)                // consistency check
1717	>	;                                // skip
1718	>	else if ((e = (int)(c = ctl)) < 0)
1719	>	w.qlock = -1;                    // pool is terminating
1720		else {
1721	<	w.parked = true;                   // must recheck
1722	<	if (w.eventCount != v) {
1723	<	w.parked = false;
1724	<	return true;
1721	>	if ((h = w.hint) < 0) {
1722	>	if (ec >= 0) {               // try to enqueue/inactivate
1723	>	long nc = (((long)ec |
1724	>	((c - AC_UNIT) & (AC_MASK|TC_MASK))));
1725	>	w.nextWait = e;          // link and mark inactive
1726	>	w.eventCount = ec | INT_SIGN;
1727	>	if (ctl != c || !U.compareAndSwapLong(this, CTL, c, nc))
1728	>	w.eventCount = ec;   // unmark on CAS failure
1729	>	else if ((int)(c >> AC_SHIFT) == 1 - (config & SMASK))
1730	>	idleAwaitWork(w, nc, c);
1731	>	}
1732	>	else if (w.eventCount < 0 && !tryTerminate(false, false) &&
1733	>	ctl == c) {         // block
1734	>	Thread wt = Thread.currentThread();
1735	>	Thread.interrupted();    // clear status
1736	>	U.putObject(wt, PARKBLOCKER, this);
1737	>	w.parker = wt;           // emulate LockSupport.park
1738	>	if (w.eventCount < 0)    // recheck
1739	>	U.park(false, 0L);
1740	>	w.parker = null;
1741	>	U.putObject(wt, PARKBLOCKER, null);
1742	>	}
1743	>	}
1744	>	if ((h >= 0 || (h = w.hint) >= 0) &&
1745	>	(ws = workQueues) != null && h < ws.length &&
1746	>	(q = ws[h]) != null) {      // signal others before retry
1747	>	WorkQueue v; Thread p; int u, i, s;
1748	>	for (int n = (config & SMASK) >>> 1;;) {
1749	>	int idleCount = (w.eventCount < 0) ? 0 : -1;
1750	>	if (((s = idleCount - q.base + q.top) <= n &&
1751	>	(n = s) <= 0) ||
1752	>	(u = (int)((c = ctl) >>> 32)) >= 0 ||
1753	>	(e = (int)c) <= 0 || m < (i = e & SMASK) ||
1754	>	(v = ws[i]) == null)
1755	>	break;
1756	>	long nc = (((long)(v.nextWait & E_MASK)) |
1757	>	((long)(u + UAC_UNIT) << 32));
1758	>	if (v.eventCount != (e | INT_SIGN) ||
1759	>	!U.compareAndSwapLong(this, CTL, c, nc))
1760	>	break;
1761	>	v.hint = h;
1762	>	v.eventCount = (e + E_SEQ) & E_MASK;
1763	>	if ((p = v.parker) != null)
1764	>	U.unpark(p);
1765	>	if (--n <= 0)
1766	>	break;
1767	>	}
1768		}
834	–	LockSupport.park(this);
835	–	rescanned = w.parked = false;
1769		}
1770		}
1771	+	return null;
1772		}
1773
1774		/**
1775	<	* If inactivating worker w has caused pool to become
1776	<	* quiescent, check for pool termination, and wait for event
1777	<	* for up to SHRINK_RATE nanosecs (rescans are unnecessary in
1778	<	* this case because quiescence reflects consensus about lack
1779	<	* of work). On timeout, if ctl has not changed, terminate the
1780	<	* worker. Upon its termination (see deregisterWorker), it may
847	<	* wake up another worker to possibly repeat this process.
1775	>	* If inactivating worker w has caused the pool to become
1776	>	* quiescent, checks for pool termination, and, so long as this is
1777	>	* not the only worker, waits for event for up to a given
1778	>	* duration.  On timeout, if ctl has not changed, terminates the
1779	>	* worker, which will in turn wake up another worker to possibly
1780	>	* repeat this process.
1781		*
1782		* @param w the calling worker
1783	<	* @param currentCtl the ctl value after enqueuing w
1784	<	* @param prevCtl the ctl value if w terminated
1785	<	* @param v the eventCount w awaits change
1786	<	*/
1787	<	private void idleAwaitWork(ForkJoinWorkerThread w, long currentCtl,
1788	<	long prevCtl, int v) {
1789	<	if (w.eventCount == v) {
1790	<	if (shutdown)
1791	<	tryTerminate(false);
1792	<	ForkJoinTask.helpExpungeStaleExceptions(); // help clean weak refs
1783	>	* @param currentCtl the ctl value triggering possible quiescence
1784	>	* @param prevCtl the ctl value to restore if thread is terminated
1785	>	*/
1786	>	private void idleAwaitWork(WorkQueue w, long currentCtl, long prevCtl) {
1787	>	if (w != null && w.eventCount < 0 &&
1788	>	!tryTerminate(false, false) && (int)prevCtl != 0) {
1789	>	int dc = -(short)(currentCtl >>> TC_SHIFT);
1790	>	long parkTime = dc < 0 ? FAST_IDLE_TIMEOUT: (dc + 1) * IDLE_TIMEOUT;
1791	>	long deadline = System.nanoTime() + parkTime - TIMEOUT_SLOP;
1792	>	Thread wt = Thread.currentThread();
1793		while (ctl == currentCtl) {
1794	<	long startTime = System.nanoTime();
1795	<	w.parked = true;
1796	<	if (w.eventCount == v)             // must recheck
1797	<	LockSupport.parkNanos(this, SHRINK_RATE);
1798	<	w.parked = false;
1799	<	if (w.eventCount != v)
1794	>	Thread.interrupted();  // timed variant of version in scan()
1795	>	U.putObject(wt, PARKBLOCKER, this);
1796	>	w.parker = wt;
1797	>	if (ctl == currentCtl)
1798	>	U.park(false, parkTime);
1799	>	w.parker = null;
1800	>	U.putObject(wt, PARKBLOCKER, null);
1801	>	if (ctl != currentCtl)
1802		break;
1803	<	else if (System.nanoTime() - startTime <
1804	<	SHRINK_RATE - (SHRINK_RATE / 10)) // timing slop
1805	<	Thread.interrupted();          // spurious wakeup
1806	<	else if (UNSAFE.compareAndSwapLong(this, ctlOffset,
872	<	currentCtl, prevCtl)) {
873	<	w.terminate = true;            // restore previous
874	<	w.eventCount = ((int)currentCtl + EC_UNIT) & E_MASK;
1803	>	if (deadline - System.nanoTime() <= 0L &&
1804	>	U.compareAndSwapLong(this, CTL, currentCtl, prevCtl)) {
1805	>	w.eventCount = (w.eventCount + E_SEQ) | E_MASK;
1806	>	w.qlock = -1;   // shrink
1807		break;
1808		}
1809		}
1810		}
1811		}
1812
881	–	// Submissions
882	–
1813		/**
1814	<	* Enqueues the given task in the submissionQueue.  Same idea as
1815	<	* ForkJoinWorkerThread.pushTask except for use of submissionLock.
1816	<	*
1817	<	* @param t the task
1818	<	*/
1819	<	private void addSubmission(ForkJoinTask<?> t) {
1820	<	final ReentrantLock lock = this.submissionLock;
1821	<	lock.lock();
1822	<	try {
1823	<	ForkJoinTask<?>[] q; int s, m;
1824	<	if ((q = submissionQueue) != null) {    // ignore if queue removed
1825	<	long u = (((s = queueTop) & (m = q.length-1)) << ASHIFT)+ABASE;
1826	<	UNSAFE.putOrderedObject(q, u, t);
1827	<	queueTop = s + 1;
1828	<	if (s - queueBase == m)
1829	<	growSubmissionQueue();
1814	>	* Scans through queues looking for work while joining a task; if
1815	>	* any present, signals. May return early if more signalling is
1816	>	* detectably unneeded.
1817	>	*
1818	>	* @param task return early if done
1819	>	* @param origin an index to start scan
1820	>	*/
1821	>	private void helpSignal(ForkJoinTask<?> task, int origin) {
1822	>	WorkQueue[] ws; WorkQueue w; Thread p; long c; int m, u, e, i, s;
1823	>	if (task != null && task.status >= 0 &&
1824	>	(u = (int)(ctl >>> 32)) < 0 && (u >> UAC_SHIFT) < 0 &&
1825	>	(ws = workQueues) != null && (m = ws.length - 1) >= 0) {
1826	>	outer: for (int k = origin, j = m; j >= 0; --j) {
1827	>	WorkQueue q = ws[k++ & m];
1828	>	for (int n = m;;) { // limit to at most m signals
1829	>	if (task.status < 0)
1830	>	break outer;
1831	>	if (q == null ||
1832	>	((s = -q.base + q.top) <= n && (n = s) <= 0))
1833	>	break;
1834	>	if ((u = (int)((c = ctl) >>> 32)) >= 0 ||
1835	>	(e = (int)c) <= 0 || m < (i = e & SMASK) ||
1836	>	(w = ws[i]) == null)
1837	>	break outer;
1838	>	long nc = (((long)(w.nextWait & E_MASK)) |
1839	>	((long)(u + UAC_UNIT) << 32));
1840	>	if (w.eventCount != (e | INT_SIGN))
1841	>	break outer;
1842	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1843	>	w.eventCount = (e + E_SEQ) & E_MASK;
1844	>	if ((p = w.parker) != null)
1845	>	U.unpark(p);
1846	>	if (--n <= 0)
1847	>	break;
1848	>	}
1849	>	}
1850		}
901	–	} finally {
902	–	lock.unlock();
1851		}
904	–	signalWork();
1852		}
1853
907	–	//  (pollSubmission is defined below with exported methods)
908	–
1854		/**
1855	<	* Creates or doubles submissionQueue array.
1856	<	* Basically identical to ForkJoinWorkerThread version.
1857	<	*/
1858	<	private void growSubmissionQueue() {
1859	<	ForkJoinTask<?>[] oldQ = submissionQueue;
1860	<	int size = oldQ != null ? oldQ.length << 1 : INITIAL_QUEUE_CAPACITY;
1861	<	if (size > MAXIMUM_QUEUE_CAPACITY)
1862	<	throw new RejectedExecutionException("Queue capacity exceeded");
1863	<	if (size < INITIAL_QUEUE_CAPACITY)
1864	<	size = INITIAL_QUEUE_CAPACITY;
1865	<	ForkJoinTask<?>[] q = submissionQueue = new ForkJoinTask<?>[size];
1866	<	int mask = size - 1;
1867	<	int top = queueTop;
1868	<	int oldMask;
1869	<	if (oldQ != null && (oldMask = oldQ.length - 1) >= 0) {
1870	<	for (int b = queueBase; b != top; ++b) {
1871	<	long u = ((b & oldMask) << ASHIFT) + ABASE;
1872	<	Object x = UNSAFE.getObjectVolatile(oldQ, u);
1873	<	if (x != null && UNSAFE.compareAndSwapObject(oldQ, u, x, null))
1874	<	UNSAFE.putObjectVolatile
1875	<	(q, ((b & mask) << ASHIFT) + ABASE, x);
1855	>	* Tries to locate and execute tasks for a stealer of the given
1856	>	* task, or in turn one of its stealers, Traces currentSteal ->
1857	>	* currentJoin links looking for a thread working on a descendant
1858	>	* of the given task and with a non-empty queue to steal back and
1859	>	* execute tasks from. The first call to this method upon a
1860	>	* waiting join will often entail scanning/search, (which is OK
1861	>	* because the joiner has nothing better to do), but this method
1862	>	* leaves hints in workers to speed up subsequent calls. The
1863	>	* implementation is very branchy to cope with potential
1864	>	* inconsistencies or loops encountering chains that are stale,
1865	>	* unknown, or so long that they are likely cyclic.
1866	>	*
1867	>	* @param joiner the joining worker
1868	>	* @param task the task to join
1869	>	* @return 0 if no progress can be made, negative if task
1870	>	* known complete, else positive
1871	>	*/
1872	>	private int tryHelpStealer(WorkQueue joiner, ForkJoinTask<?> task) {
1873	>	int stat = 0, steps = 0;                    // bound to avoid cycles
1874	>	if (joiner != null && task != null) {       // hoist null checks
1875	>	restart: for (;;) {
1876	>	ForkJoinTask<?> subtask = task;     // current target
1877	>	for (WorkQueue j = joiner, v;;) {   // v is stealer of subtask
1878	>	WorkQueue[] ws; int m, s, h;
1879	>	if ((s = task.status) < 0) {
1880	>	stat = s;
1881	>	break restart;
1882	>	}
1883	>	if ((ws = workQueues) == null || (m = ws.length - 1) <= 0)
1884	>	break restart;              // shutting down
1885	>	if ((v = ws[h = (j.hint | 1) & m]) == null ||
1886	>	v.currentSteal != subtask) {
1887	>	for (int origin = h;;) {    // find stealer
1888	>	if (((h = (h + 2) & m) & 15) == 1 &&
1889	>	(subtask.status < 0 || j.currentJoin != subtask))
1890	>	continue restart;   // occasional staleness check
1891	>	if ((v = ws[h]) != null &&
1892	>	v.currentSteal == subtask) {
1893	>	j.hint = h;        // save hint
1894	>	break;
1895	>	}
1896	>	if (h == origin)
1897	>	break restart;      // cannot find stealer
1898	>	}
1899	>	}
1900	>	for (;;) { // help stealer or descend to its stealer
1901	>	ForkJoinTask[] a;  int b;
1902	>	if (subtask.status < 0)     // surround probes with
1903	>	continue restart;       //   consistency checks
1904	>	if ((b = v.base) - v.top < 0 && (a = v.array) != null) {
1905	>	int i = (((a.length - 1) & b) << ASHIFT) + ABASE;
1906	>	ForkJoinTask<?> t =
1907	>	(ForkJoinTask<?>)U.getObjectVolatile(a, i);
1908	>	if (subtask.status < 0 || j.currentJoin != subtask ||
1909	>	v.currentSteal != subtask)
1910	>	continue restart;   // stale
1911	>	stat = 1;               // apparent progress
1912	>	if (t != null && v.base == b &&
1913	>	U.compareAndSwapObject(a, i, t, null)) {
1914	>	v.base = b + 1;     // help stealer
1915	>	joiner.runSubtask(t);
1916	>	}
1917	>	else if (v.base == b && ++steps == MAX_HELP)
1918	>	break restart;      // v apparently stalled
1919	>	}
1920	>	else {                      // empty -- try to descend
1921	>	ForkJoinTask<?> next = v.currentJoin;
1922	>	if (subtask.status < 0 || j.currentJoin != subtask ||
1923	>	v.currentSteal != subtask)
1924	>	continue restart;   // stale
1925	>	else if (next == null || ++steps == MAX_HELP)
1926	>	break restart;      // dead-end or maybe cyclic
1927	>	else {
1928	>	subtask = next;
1929	>	j = v;
1930	>	break;
1931	>	}
1932	>	}
1933	>	}
1934	>	}
1935		}
1936		}
1937	+	return stat;
1938		}
1939
935	–	// Blocking support
936	–
1940		/**
1941	<	* Tries to increment blockedCount, decrement active count
1942	<	* (sometimes implicitly) and possibly release or create a
940	<	* compensating worker in preparation for blocking. Fails
941	<	* on contention or termination.
1941	>	* Analog of tryHelpStealer for CountedCompleters. Tries to steal
1942	>	* and run tasks within the target's computation.
1943		*
1944	<	* @return true if the caller can block, else should recheck and retry
1945	<	*/
1946	<	private boolean tryPreBlock() {
946	<	int b = blockedCount;
947	<	if (UNSAFE.compareAndSwapInt(this, blockedCountOffset, b, b + 1)) {
948	<	int pc = parallelism;
949	<	do {
950	<	ForkJoinWorkerThread[] ws; ForkJoinWorkerThread w;
951	<	int e, ac, tc, rc, i;
952	<	long c = ctl;
953	<	int u = (int)(c >>> 32);
954	<	if ((e = (int)c) < 0) {
955	<	// skip -- terminating
956	<	}
957	<	else if ((ac = (u >> UAC_SHIFT)) <= 0 && e != 0 &&
958	<	(ws = workers) != null &&
959	<	(i = ~e & SMASK) < ws.length &&
960	<	(w = ws[i]) != null) {
961	<	long nc = ((long)(w.nextWait & E_MASK) |
962	<	(c & (AC_MASK|TC_MASK)));
963	<	if (w.eventCount == e &&
964	<	UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
965	<	w.eventCount = (e + EC_UNIT) & E_MASK;
966	<	if (w.parked)
967	<	UNSAFE.unpark(w);
968	<	return true;             // release an idle worker
969	<	}
970	<	}
971	<	else if ((tc = (short)(u >>> UTC_SHIFT)) >= 0 && ac + pc > 1) {
972	<	long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK);
973	<	if (UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc))
974	<	return true;             // no compensation needed
975	<	}
976	<	else if (tc + pc < MAX_ID) {
977	<	long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK);
978	<	if (UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
979	<	addWorker();
980	<	return true;            // create a replacement
981	<	}
982	<	}
983	<	// try to back out on any failure and let caller retry
984	<	} while (!UNSAFE.compareAndSwapInt(this, blockedCountOffset,
985	<	b = blockedCount, b - 1));
986	<	}
987	<	return false;
988	<	}
989	<
990	<	/**
991	<	* Decrements blockedCount and increments active count
992	<	*/
993	<	private void postBlock() {
994	<	long c;
995	<	do {} while (!UNSAFE.compareAndSwapLong(this, ctlOffset,  // no mask
996	<	c = ctl, c + AC_UNIT));
997	<	int b;
998	<	do {} while(!UNSAFE.compareAndSwapInt(this, blockedCountOffset,
999	<	b = blockedCount, b - 1));
1000	<	}
1001	<
1002	<	/**
1003	<	* Possibly blocks waiting for the given task to complete, or
1004	<	* cancels the task if terminating.  Fails to wait if contended.
1944	>	* @param task the task to join
1945	>	* @param mode if shared, exit upon completing any task
1946	>	* if all workers are active
1947		*
1006	–	* @param joinMe the task
1948		*/
1949	<	final void tryAwaitJoin(ForkJoinTask<?> joinMe) {
1950	<	int s;
1951	<	Thread.interrupted(); // clear interrupts before checking termination
1952	<	if (joinMe.status >= 0) {
1953	<	if (tryPreBlock()) {
1954	<	joinMe.tryAwaitDone(0L);
1955	<	postBlock();
1949	>	private int helpComplete(ForkJoinTask<?> task, int mode) {
1950	>	WorkQueue[] ws; WorkQueue q; int m, n, s, u;
1951	>	if (task != null && (ws = workQueues) != null &&
1952	>	(m = ws.length - 1) >= 0) {
1953	>	for (int j = 1, origin = j;;) {
1954	>	if ((s = task.status) < 0)
1955	>	return s;
1956	>	if ((q = ws[j & m]) != null && q.pollAndExecCC(task)) {
1957	>	origin = j;
1958	>	if (mode == SHARED_QUEUE &&
1959	>	((u = (int)(ctl >>> 32)) >= 0 || (u >> UAC_SHIFT) >= 0))
1960	>	break;
1961	>	}
1962	>	else if ((j = (j + 2) & m) == origin)
1963	>	break;
1964		}
1016	–	else if ((ctl & STOP_BIT) != 0L)
1017	–	joinMe.cancelIgnoringExceptions();
1965		}
1966	+	return 0;
1967		}
1968
1969		/**
1970	<	* Possibly blocks the given worker waiting for joinMe to
1971	<	* complete or timeout
1972	<	*
1973	<	* @param joinMe the task
1974	<	* @param millis the wait time for underlying Object.wait
1975	<	*/
1976	<	final void timedAwaitJoin(ForkJoinTask<?> joinMe, long nanos) {
1977	<	while (joinMe.status >= 0) {
1978	<	Thread.interrupted();
1979	<	if ((ctl & STOP_BIT) != 0L) {
1980	<	joinMe.cancelIgnoringExceptions();
1981	<	break;
1970	>	* Tries to decrement active count (sometimes implicitly) and
1971	>	* possibly release or create a compensating worker in preparation
1972	>	* for blocking. Fails on contention or termination. Otherwise,
1973	>	* adds a new thread if no idle workers are available and pool
1974	>	* may become starved.
1975	>	*/
1976	>	final boolean tryCompensate() {
1977	>	int pc = config & SMASK, e, i, tc; long c;
1978	>	WorkQueue[] ws; WorkQueue w; Thread p;
1979	>	if ((ws = workQueues) != null && (e = (int)(c = ctl)) >= 0) {
1980	>	if (e != 0 && (i = e & SMASK) < ws.length &&
1981	>	(w = ws[i]) != null && w.eventCount == (e | INT_SIGN)) {
1982	>	long nc = ((long)(w.nextWait & E_MASK) |
1983	>	(c & (AC_MASK|TC_MASK)));
1984	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1985	>	w.eventCount = (e + E_SEQ) & E_MASK;
1986	>	if ((p = w.parker) != null)
1987	>	U.unpark(p);
1988	>	return true;   // replace with idle worker
1989	>	}
1990		}
1991	<	if (tryPreBlock()) {
1992	<	long last = System.nanoTime();
1993	<	while (joinMe.status >= 0) {
1994	<	long millis = TimeUnit.NANOSECONDS.toMillis(nanos);
1995	<	if (millis <= 0)
1996	<	break;
1997	<	joinMe.tryAwaitDone(millis);
1998	<	if (joinMe.status < 0)
1999	<	break;
2000	<	if ((ctl & STOP_BIT) != 0L) {
2001	<	joinMe.cancelIgnoringExceptions();
2002	<	break;
1991	>	else if ((tc = (short)(c >>> TC_SHIFT)) >= 0 &&
1992	>	(int)(c >> AC_SHIFT) + pc > 1) {
1993	>	long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK);
1994	>	if (U.compareAndSwapLong(this, CTL, c, nc))
1995	>	return true;   // no compensation
1996	>	}
1997	>	else if (tc + pc < MAX_CAP) {
1998	>	long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK);
1999	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
2000	>	ForkJoinWorkerThreadFactory fac;
2001	>	Throwable ex = null;
2002	>	ForkJoinWorkerThread wt = null;
2003	>	try {
2004	>	if ((fac = factory) != null &&
2005	>	(wt = fac.newThread(this)) != null) {
2006	>	wt.start();
2007	>	return true;
2008	>	}
2009	>	} catch (Throwable rex) {
2010	>	ex = rex;
2011		}
2012	<	long now = System.nanoTime();
1049	<	nanos -= now - last;
1050	<	last = now;
2012	>	deregisterWorker(wt, ex); // clean up and return false
2013		}
1052	–	postBlock();
1053	–	break;
2014		}
2015		}
2016	+	return false;
2017		}
2018
2019		/**
2020	<	* If necessary, compensates for blocker, and blocks
2020	>	* Helps and/or blocks until the given task is done.
2021	>	*
2022	>	* @param joiner the joining worker
2023	>	* @param task the task
2024	>	* @return task status on exit
2025		*/
2026	<	private void awaitBlocker(ManagedBlocker blocker)
2027	<	throws InterruptedException {
2028	<	while (!blocker.isReleasable()) {
2029	<	if (tryPreBlock()) {
2030	<	try {
2031	<	do {} while (!blocker.isReleasable() && !blocker.block());
2032	<	} finally {
2033	<	postBlock();
2026	>	final int awaitJoin(WorkQueue joiner, ForkJoinTask<?> task) {
2027	>	int s = 0;
2028	>	if (joiner != null && task != null && (s = task.status) >= 0) {
2029	>	ForkJoinTask<?> prevJoin = joiner.currentJoin;
2030	>	joiner.currentJoin = task;
2031	>	do {} while ((s = task.status) >= 0 && !joiner.isEmpty() &&
2032	>	joiner.tryRemoveAndExec(task)); // process local tasks
2033	>	if (s >= 0 && (s = task.status) >= 0) {
2034	>	helpSignal(task, joiner.poolIndex);
2035	>	if ((s = task.status) >= 0 &&
2036	>	(task instanceof CountedCompleter))
2037	>	s = helpComplete(task, LIFO_QUEUE);
2038	>	}
2039	>	while (s >= 0 && (s = task.status) >= 0) {
2040	>	if ((!joiner.isEmpty() ||           // try helping
2041	>	(s = tryHelpStealer(joiner, task)) == 0) &&
2042	>	(s = task.status) >= 0) {
2043	>	helpSignal(task, joiner.poolIndex);
2044	>	if ((s = task.status) >= 0 && tryCompensate()) {
2045	>	if (task.trySetSignal() && (s = task.status) >= 0) {
2046	>	synchronized (task) {
2047	>	if (task.status >= 0) {
2048	>	try {                // see ForkJoinTask
2049	>	task.wait();     //  for explanation
2050	>	} catch (InterruptedException ie) {
2051	>	}
2052	>	}
2053	>	else
2054	>	task.notifyAll();
2055	>	}
2056	>	}
2057	>	long c;                          // re-activate
2058	>	do {} while (!U.compareAndSwapLong
2059	>	(this, CTL, c = ctl, c + AC_UNIT));
2060	>	}
2061		}
1070	–	break;
2062		}
2063	+	joiner.currentJoin = prevJoin;
2064		}
2065	+	return s;
2066		}
2067
1075	–	// Creating, registering and deregistring workers
1076	–
2068		/**
2069	<	* Tries to create and start a worker; minimally rolls back counts
2070	<	* on failure.
2069	>	* Stripped-down variant of awaitJoin used by timed joins. Tries
2070	>	* to help join only while there is continuous progress. (Caller
2071	>	* will then enter a timed wait.)
2072	>	*
2073	>	* @param joiner the joining worker
2074	>	* @param task the task
2075		*/
2076	<	private void addWorker() {
2077	<	Throwable ex = null;
2078	<	ForkJoinWorkerThread t = null;
2079	<	try {
2080	<	t = factory.newThread(this);
2081	<	} catch (Throwable e) {
2082	<	ex = e;
2083	<	}
2084	<	if (t == null) {  // null or exceptional factory return
2085	<	long c;       // adjust counts
2086	<	do {} while (!UNSAFE.compareAndSwapLong
2087	<	(this, ctlOffset, c = ctl,
2088	<	(((c - AC_UNIT) & AC_MASK) |
2089	<	((c - TC_UNIT) & TC_MASK) |
2090	<	(c & ~(AC_MASK|TC_MASK)))));
2091	<	// Propagate exception if originating from an external caller
2092	<	if (!tryTerminate(false) && ex != null &&
2093	<	!(Thread.currentThread() instanceof ForkJoinWorkerThread))
1099	<	UNSAFE.throwException(ex);
2076	>	final void helpJoinOnce(WorkQueue joiner, ForkJoinTask<?> task) {
2077	>	int s;
2078	>	if (joiner != null && task != null && (s = task.status) >= 0) {
2079	>	ForkJoinTask<?> prevJoin = joiner.currentJoin;
2080	>	joiner.currentJoin = task;
2081	>	do {} while ((s = task.status) >= 0 && !joiner.isEmpty() &&
2082	>	joiner.tryRemoveAndExec(task));
2083	>	if (s >= 0 && (s = task.status) >= 0) {
2084	>	helpSignal(task, joiner.poolIndex);
2085	>	if ((s = task.status) >= 0 &&
2086	>	(task instanceof CountedCompleter))
2087	>	s = helpComplete(task, LIFO_QUEUE);
2088	>	}
2089	>	if (s >= 0 && joiner.isEmpty()) {
2090	>	do {} while (task.status >= 0 &&
2091	>	tryHelpStealer(joiner, task) > 0);
2092	>	}
2093	>	joiner.currentJoin = prevJoin;
2094		}
1101	–	else
1102	–	t.start();
2095		}
2096
2097		/**
2098	<	* Callback from ForkJoinWorkerThread constructor to assign a
2099	<	* public name
2100	<	*/
2101	<	final String nextWorkerName() {
2102	<	for (int n;;) {
2103	<	if (UNSAFE.compareAndSwapInt(this, nextWorkerNumberOffset,
2104	<	n = nextWorkerNumber, ++n))
2105	<	return workerNamePrefix + n;
2098	>	* Returns a (probably) non-empty steal queue, if one is found
2099	>	* during a random, then cyclic scan, else null.  This method must
2100	>	* be retried by caller if, by the time it tries to use the queue,
2101	>	* it is empty.
2102	>	* @param r a (random) seed for scanning
2103	>	*/
2104	>	private WorkQueue findNonEmptyStealQueue(int r) {
2105	>	for (WorkQueue[] ws;;) {
2106	>	int ps = plock, m;
2107	>	if ((ws = workQueues) == null || (m = ws.length - 1) < 1)
2108	>	return null;
2109	>	for (int j = (m + 1) << 2; ;) {
2110	>	WorkQueue q = ws[(((r + j) << 1) | 1) & m];
2111	>	if (q != null && q.base - q.top < 0)
2112	>	return q;
2113	>	else if (--j < 0) {
2114	>	if (plock == ps)
2115	>	return null;
2116	>	break;
2117	>	}
2118	>	}
2119		}
2120		}
2121
2122		/**
2123	<	* Callback from ForkJoinWorkerThread constructor to
2124	<	* determine its poolIndex and record in workers array.
2125	<	*
2126	<	* @param w the worker
2127	<	* @return the worker's pool index
2128	<	*/
2129	<	final int registerWorker(ForkJoinWorkerThread w) {
2130	<	/*
2131	<	* In the typical case, a new worker acquires the lock, uses
2132	<	* next available index and returns quickly.  Since we should
2133	<	* not block callers (ultimately from signalWork or
2134	<	* tryPreBlock) waiting for the lock needed to do this, we
2135	<	* instead help release other workers while waiting for the
2136	<	* lock.
2137	<	*/
2138	<	for (int g;;) {
2139	<	ForkJoinWorkerThread[] ws;
2140	<	if (((g = scanGuard) & SG_UNIT) == 0 &&
2141	<	UNSAFE.compareAndSwapInt(this, scanGuardOffset,
2142	<	g, g | SG_UNIT)) {
2143	<	int k = nextWorkerIndex;
2144	<	try {
2145	<	if ((ws = workers) != null) { // ignore on shutdown
2146	<	int n = ws.length;
2147	<	if (k < 0 || k >= n || ws[k] != null) {
2148	<	for (k = 0; k < n && ws[k] != null; ++k)
2149	<	;
2150	<	if (k == n)
2151	<	ws = workers = Arrays.copyOf(ws, n << 1);
2152	<	}
2153	<	ws[k] = w;
2154	<	nextWorkerIndex = k + 1;
2155	<	int m = g & SMASK;
2156	<	g = k > m? ((m << 1) + 1) & SMASK : g + (SG_UNIT<<1);
2157	<	}
2158	<	} finally {
2159	<	scanGuard = g;
2160	<	}
2161	<	return k;
1157	<	}
1158	<	else if ((ws = workers) != null) { // help release others
1159	<	for (ForkJoinWorkerThread u : ws) {
1160	<	if (u != null && u.queueBase != u.queueTop) {
1161	<	if (tryReleaseWaiter())
1162	<	break;
1163	<	}
2123	>	* Runs tasks until {@code isQuiescent()}. We piggyback on
2124	>	* active count ctl maintenance, but rather than blocking
2125	>	* when tasks cannot be found, we rescan until all others cannot
2126	>	* find tasks either.
2127	>	*/
2128	>	final void helpQuiescePool(WorkQueue w) {
2129	>	for (boolean active = true;;) {
2130	>	ForkJoinTask<?> localTask; // exhaust local queue
2131	>	while ((localTask = w.nextLocalTask()) != null)
2132	>	localTask.doExec();
2133	>	// Similar to loop in scan(), but ignoring submissions
2134	>	WorkQueue q = findNonEmptyStealQueue(w.nextSeed());
2135	>	if (q != null) {
2136	>	ForkJoinTask<?> t; int b;
2137	>	if (!active) {      // re-establish active count
2138	>	long c;
2139	>	active = true;
2140	>	do {} while (!U.compareAndSwapLong
2141	>	(this, CTL, c = ctl, c + AC_UNIT));
2142	>	}
2143	>	if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null) {
2144	>	if (q.base - q.top < 0)
2145	>	signalWork(q);
2146	>	w.runSubtask(t);
2147	>	}
2148	>	}
2149	>	else {
2150	>	long c;
2151	>	if (active) {       // decrement active count without queuing
2152	>	active = false;
2153	>	do {} while (!U.compareAndSwapLong
2154	>	(this, CTL, c = ctl, c -= AC_UNIT));
2155	>	}
2156	>	else
2157	>	c = ctl;        // re-increment on exit
2158	>	if ((int)(c >> AC_SHIFT) + (config & SMASK) == 0) {
2159	>	do {} while (!U.compareAndSwapLong
2160	>	(this, CTL, c = ctl, c + AC_UNIT));
2161	>	break;
2162		}
2163		}
2164		}
2165		}
2166
2167		/**
2168	<	* Final callback from terminating worker.  Removes record of
2169	<	* worker from array, and adjusts counts. If pool is shutting
2170	<	* down, tries to complete termination.
2171	<	*
2172	<	* @param w the worker
2173	<	*/
2174	<	final void deregisterWorker(ForkJoinWorkerThread w, Throwable ex) {
2175	<	int idx = w.poolIndex;
2176	<	int sc = w.stealCount;
2177	<	int steps = 0;
2178	<	// Remove from array, adjust worker counts and collect steal count.
2179	<	// We can intermix failed removes or adjusts with steal updates
2180	<	do {
2181	<	long s, c;
2182	<	int g;
2183	<	if (steps == 0 && ((g = scanGuard) & SG_UNIT) == 0 &&
1186	<	UNSAFE.compareAndSwapInt(this, scanGuardOffset,
1187	<	g, g |= SG_UNIT)) {
1188	<	ForkJoinWorkerThread[] ws = workers;
1189	<	if (ws != null && idx >= 0 &&
1190	<	idx < ws.length && ws[idx] == w)
1191	<	ws[idx] = null;    // verify
1192	<	nextWorkerIndex = idx;
1193	<	scanGuard = g + SG_UNIT;
1194	<	steps = 1;
1195	<	}
1196	<	if (steps == 1 &&
1197	<	UNSAFE.compareAndSwapLong(this, ctlOffset, c = ctl,
1198	<	(((c - AC_UNIT) & AC_MASK) |
1199	<	((c - TC_UNIT) & TC_MASK) |
1200	<	(c & ~(AC_MASK|TC_MASK)))))
1201	<	steps = 2;
1202	<	if (sc != 0 &&
1203	<	UNSAFE.compareAndSwapLong(this, stealCountOffset,
1204	<	s = stealCount, s + sc))
1205	<	sc = 0;
1206	<	} while (steps != 2 || sc != 0);
1207	<	if (!tryTerminate(false)) {
1208	<	if (ex != null)   // possibly replace if died abnormally
1209	<	signalWork();
1210	<	else
1211	<	tryReleaseWaiter();
2168	>	* Gets and removes a local or stolen task for the given worker.
2169	>	*
2170	>	* @return a task, if available
2171	>	*/
2172	>	final ForkJoinTask<?> nextTaskFor(WorkQueue w) {
2173	>	for (ForkJoinTask<?> t;;) {
2174	>	WorkQueue q; int b;
2175	>	if ((t = w.nextLocalTask()) != null)
2176	>	return t;
2177	>	if ((q = findNonEmptyStealQueue(w.nextSeed())) == null)
2178	>	return null;
2179	>	if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null) {
2180	>	if (q.base - q.top < 0)
2181	>	signalWork(q);
2182	>	return t;
2183	>	}
2184		}
2185		}
2186
1215	–	// Shutdown and termination
1216	–
2187		/**
2188	<	* Possibly initiates and/or completes termination.
2188	>	* Returns a cheap heuristic guide for task partitioning when
2189	>	* programmers, frameworks, tools, or languages have little or no
2190	>	* idea about task granularity.  In essence by offering this
2191	>	* method, we ask users only about tradeoffs in overhead vs
2192	>	* expected throughput and its variance, rather than how finely to
2193	>	* partition tasks.
2194	>	*
2195	>	* In a steady state strict (tree-structured) computation, each
2196	>	* thread makes available for stealing enough tasks for other
2197	>	* threads to remain active. Inductively, if all threads play by
2198	>	* the same rules, each thread should make available only a
2199	>	* constant number of tasks.
2200	>	*
2201	>	* The minimum useful constant is just 1. But using a value of 1
2202	>	* would require immediate replenishment upon each steal to
2203	>	* maintain enough tasks, which is infeasible.  Further,
2204	>	* partitionings/granularities of offered tasks should minimize
2205	>	* steal rates, which in general means that threads nearer the top
2206	>	* of computation tree should generate more than those nearer the
2207	>	* bottom. In perfect steady state, each thread is at
2208	>	* approximately the same level of computation tree. However,
2209	>	* producing extra tasks amortizes the uncertainty of progress and
2210	>	* diffusion assumptions.
2211	>	*
2212	>	* So, users will want to use values larger, but not much larger
2213	>	* than 1 to both smooth over transient shortages and hedge
2214	>	* against uneven progress; as traded off against the cost of
2215	>	* extra task overhead. We leave the user to pick a threshold
2216	>	* value to compare with the results of this call to guide
2217	>	* decisions, but recommend values such as 3.
2218	>	*
2219	>	* When all threads are active, it is on average OK to estimate
2220	>	* surplus strictly locally. In steady-state, if one thread is
2221	>	* maintaining say 2 surplus tasks, then so are others. So we can
2222	>	* just use estimated queue length.  However, this strategy alone
2223	>	* leads to serious mis-estimates in some non-steady-state
2224	>	* conditions (ramp-up, ramp-down, other stalls). We can detect
2225	>	* many of these by further considering the number of "idle"
2226	>	* threads, that are known to have zero queued tasks, so
2227	>	* compensate by a factor of (#idle/#active) threads.
2228	>	*
2229	>	* Note: The approximation of #busy workers as #active workers is
2230	>	* not very good under current signalling scheme, and should be
2231	>	* improved.
2232	>	*/
2233	>	static int getSurplusQueuedTaskCount() {
2234	>	Thread t; ForkJoinWorkerThread wt; ForkJoinPool pool; WorkQueue q;
2235	>	if (((t = Thread.currentThread()) instanceof ForkJoinWorkerThread)) {
2236	>	int p = (pool = (wt = (ForkJoinWorkerThread)t).pool).config & SMASK;
2237	>	int n = (q = wt.workQueue).top - q.base;
2238	>	int a = (int)(pool.ctl >> AC_SHIFT) + p;
2239	>	return n - (a > (p >>>= 1) ? 0 :
2240	>	a > (p >>>= 1) ? 1 :
2241	>	a > (p >>>= 1) ? 2 :
2242	>	a > (p >>>= 1) ? 4 :
2243	>	8);
2244	>	}
2245	>	return 0;
2246	>	}
2247	>
2248	>	//  Termination
2249	>
2250	>	/**
2251	>	* Possibly initiates and/or completes termination.  The caller
2252	>	* triggering termination runs three passes through workQueues:
2253	>	* (0) Setting termination status, followed by wakeups of queued
2254	>	* workers; (1) cancelling all tasks; (2) interrupting lagging
2255	>	* threads (likely in external tasks, but possibly also blocked in
2256	>	* joins).  Each pass repeats previous steps because of potential
2257	>	* lagging thread creation.
2258		*
2259		* @param now if true, unconditionally terminate, else only
2260	<	* if shutdown and empty queue and no active workers
2260	>	* if no work and no active workers
2261	>	* @param enable if true, enable shutdown when next possible
2262		* @return true if now terminating or terminated
2263		*/
2264	<	private boolean tryTerminate(boolean now) {
2265	<	long c;
2266	<	while (((c = ctl) & STOP_BIT) == 0) {
2267	<	if (!now) {
2268	<	if ((int)(c >> AC_SHIFT) != -parallelism)
2269	<	return false;
2270	<	if (!shutdown || blockedCount != 0 || quiescerCount != 0 ||
2271	<	queueBase != queueTop) {
2272	<	if (ctl == c) // staleness check
1233	<	return false;
1234	<	continue;
2264	>	private boolean tryTerminate(boolean now, boolean enable) {
2265	>	if (this == commonPool)                     // cannot shut down
2266	>	return false;
2267	>	for (long c;;) {
2268	>	if (((c = ctl) & STOP_BIT) != 0) {      // already terminating
2269	>	if ((short)(c >>> TC_SHIFT) == -(config & SMASK)) {
2270	>	synchronized (this) {
2271	>	notifyAll();                // signal when 0 workers
2272	>	}
2273		}
2274	+	return true;
2275		}
2276	<	if (UNSAFE.compareAndSwapLong(this, ctlOffset, c, c | STOP_BIT))
2277	<	startTerminating();
2278	<	}
2279	<	if ((short)(c >>> TC_SHIFT) == -parallelism) { // signal when 0 workers
2280	<	final ReentrantLock lock = this.submissionLock;
2281	<	lock.lock();
2282	<	try {
2283	<	termination.signalAll();
2284	<	} finally {
2285	<	lock.unlock();
2276	>	if (plock >= 0) {                       // not yet enabled
2277	>	int ps;
2278	>	if (!enable)
2279	>	return false;
2280	>	if (((ps = plock) & PL_LOCK) != 0 ||
2281	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
2282	>	ps = acquirePlock();
2283	>	if (!U.compareAndSwapInt(this, PLOCK, ps, SHUTDOWN))
2284	>	releasePlock(SHUTDOWN);
2285	>	}
2286	>	if (!now) {                             // check if idle & no tasks
2287	>	if ((int)(c >> AC_SHIFT) != -(config & SMASK) ||
2288	>	hasQueuedSubmissions())
2289	>	return false;
2290	>	// Check for unqueued inactive workers. One pass suffices.
2291	>	WorkQueue[] ws = workQueues; WorkQueue w;
2292	>	if (ws != null) {
2293	>	for (int i = 1; i < ws.length; i += 2) {
2294	>	if ((w = ws[i]) != null && w.eventCount >= 0)
2295	>	return false;
2296	>	}
2297	>	}
2298		}
2299	<	}
2300	<	return true;
2301	<	}
2302	<
2303	<	/**
2304	<	* Runs up to three passes through workers: (0) Setting
2305	<	* termination status for each worker, followed by wakeups up to
2306	<	* queued workers; (1) helping cancel tasks; (2) interrupting
2307	<	* lagging threads (likely in external tasks, but possibly also
2308	<	* blocked in joins).  Each pass repeats previous steps because of
2309	<	* potential lagging thread creation.
2310	<	*/
2311	<	private void startTerminating() {
2312	<	cancelSubmissions();
2313	<	for (int pass = 0; pass < 3; ++pass) {
2314	<	ForkJoinWorkerThread[] ws = workers;
2315	<	if (ws != null) {
2316	<	for (ForkJoinWorkerThread w : ws) {
2317	<	if (w != null) {
2318	<	w.terminate = true;
1268	<	if (pass > 0) {
1269	<	w.cancelTasks();
1270	<	if (pass > 1 && !w.isInterrupted()) {
1271	<	try {
1272	<	w.interrupt();
1273	<	} catch (SecurityException ignore) {
2299	>	if (U.compareAndSwapLong(this, CTL, c, c | STOP_BIT)) {
2300	>	for (int pass = 0; pass < 3; ++pass) {
2301	>	WorkQueue[] ws = workQueues;
2302	>	if (ws != null) {
2303	>	WorkQueue w; Thread wt;
2304	>	int n = ws.length;
2305	>	for (int i = 0; i < n; ++i) {
2306	>	if ((w = ws[i]) != null) {
2307	>	w.qlock = -1;
2308	>	if (pass > 0) {
2309	>	w.cancelAll();
2310	>	if (pass > 1 && (wt = w.owner) != null) {
2311	>	if (!wt.isInterrupted()) {
2312	>	try {
2313	>	wt.interrupt();
2314	>	} catch (SecurityException ignore) {
2315	>	}
2316	>	}
2317	>	U.unpark(wt);
2318	>	}
2319		}
2320		}
2321		}
2322	+	// Wake up workers parked on event queue
2323	+	int i, e; long cc; Thread p;
2324	+	while ((e = (int)(cc = ctl) & E_MASK) != 0 &&
2325	+	(i = e & SMASK) < n &&
2326	+	(w = ws[i]) != null) {
2327	+	long nc = ((long)(w.nextWait & E_MASK) |
2328	+	((cc + AC_UNIT) & AC_MASK) |
2329	+	(cc & (TC_MASK|STOP_BIT)));
2330	+	if (w.eventCount == (e | INT_SIGN) &&
2331	+	U.compareAndSwapLong(this, CTL, cc, nc)) {
2332	+	w.eventCount = (e + E_SEQ) & E_MASK;
2333	+	w.qlock = -1;
2334	+	if ((p = w.parker) != null)
2335	+	U.unpark(p);
2336	+	}
2337	+	}
2338		}
2339		}
1279	–	terminateWaiters();
2340		}
2341		}
2342		}
2343
2344	+	// external operations on common pool
2345	+
2346		/**
2347	<	* Polls and cancels all submissions. Called only during termination.
2347	>	* Returns common pool queue for a thread that has submitted at
2348	>	* least one task.
2349		*/
2350	<	private void cancelSubmissions() {
2351	<	while (queueBase != queueTop) {
2352	<	ForkJoinTask<?> task = pollSubmission();
2353	<	if (task != null) {
2354	<	try {
2355	<	task.cancel(false);
2356	<	} catch (Throwable ignore) {
2350	>	static WorkQueue commonSubmitterQueue() {
2351	>	ForkJoinPool p; WorkQueue[] ws; int m; Submitter z;
2352	>	return ((z = submitters.get()) != null &&
2353	>	(p = commonPool) != null &&
2354	>	(ws = p.workQueues) != null &&
2355	>	(m = ws.length - 1) >= 0) ?
2356	>	ws[m & z.seed & SQMASK] : null;
2357	>	}
2358	>
2359	>	/**
2360	>	* Tries to pop the given task from submitter's queue in common pool.
2361	>	*/
2362	>	static boolean tryExternalUnpush(ForkJoinTask<?> t) {
2363	>	ForkJoinPool p; WorkQueue[] ws; WorkQueue q; Submitter z;
2364	>	ForkJoinTask<?>[] a;  int m, s;
2365	>	if (t != null &&
2366	>	(z = submitters.get()) != null &&
2367	>	(p = commonPool) != null &&
2368	>	(ws = p.workQueues) != null &&
2369	>	(m = ws.length - 1) >= 0 &&
2370	>	(q = ws[m & z.seed & SQMASK]) != null &&
2371	>	(s = q.top) != q.base &&
2372	>	(a = q.array) != null) {
2373	>	long j = (((a.length - 1) & (s - 1)) << ASHIFT) + ABASE;
2374	>	if (U.getObject(a, j) == t &&
2375	>	U.compareAndSwapInt(q, QLOCK, 0, 1)) {
2376	>	if (q.array == a && q.top == s && // recheck
2377	>	U.compareAndSwapObject(a, j, t, null)) {
2378	>	q.top = s - 1;
2379	>	q.qlock = 0;
2380	>	return true;
2381		}
2382	+	q.qlock = 0;
2383		}
2384		}
2385	+	return false;
2386		}
2387
2388		/**
2389	<	* Tries to set the termination status of waiting workers, and
2390	<	* then wakes them up (after which they will terminate).
2391	<	*/
2392	<	private void terminateWaiters() {
2393	<	ForkJoinWorkerThread[] ws = workers;
2394	<	if (ws != null) {
2395	<	ForkJoinWorkerThread w; long c; int i, e;
2396	<	int n = ws.length;
2397	<	while ((i = ~(e = (int)(c = ctl)) & SMASK) < n &&
2398	<	(w = ws[i]) != null && w.eventCount == (e & E_MASK)) {
2399	<	if (UNSAFE.compareAndSwapLong(this, ctlOffset, c,
2400	<	(long)(w.nextWait & E_MASK) |
2401	<	((c + AC_UNIT) & AC_MASK) |
2402	<	(c & (TC_MASK|STOP_BIT)))) {
2403	<	w.terminate = true;
2404	<	w.eventCount = e + EC_UNIT;
2405	<	if (w.parked)
2406	<	UNSAFE.unpark(w);
2389	>	* Tries to pop and run local tasks within the same computation
2390	>	* as the given root. On failure, tries to help complete from
2391	>	* other queues via helpComplete.
2392	>	*/
2393	>	private void externalHelpComplete(WorkQueue q, ForkJoinTask<?> root) {
2394	>	ForkJoinTask<?>[] a; int m;
2395	>	if (q != null && (a = q.array) != null && (m = (a.length - 1)) >= 0 &&
2396	>	root != null && root.status >= 0) {
2397	>	for (;;) {
2398	>	int s, u; Object o; CountedCompleter<?> task = null;
2399	>	if ((s = q.top) - q.base > 0) {
2400	>	long j = ((m & (s - 1)) << ASHIFT) + ABASE;
2401	>	if ((o = U.getObject(a, j)) != null &&
2402	>	(o instanceof CountedCompleter)) {
2403	>	CountedCompleter<?> t = (CountedCompleter<?>)o, r = t;
2404	>	do {
2405	>	if (r == root) {
2406	>	if (U.compareAndSwapInt(q, QLOCK, 0, 1)) {
2407	>	if (q.array == a && q.top == s &&
2408	>	U.compareAndSwapObject(a, j, t, null)) {
2409	>	q.top = s - 1;
2410	>	task = t;
2411	>	}
2412	>	q.qlock = 0;
2413	>	}
2414	>	break;
2415	>	}
2416	>	} while ((r = r.completer) != null);
2417	>	}
2418	>	}
2419	>	if (task != null)
2420	>	task.doExec();
2421	>	if (root.status < 0 ||
2422	>	(u = (int)(ctl >>> 32)) >= 0 || (u >> UAC_SHIFT) >= 0)
2423	>	break;
2424	>	if (task == null) {
2425	>	helpSignal(root, q.poolIndex);
2426	>	if (root.status >= 0)
2427	>	helpComplete(root, SHARED_QUEUE);
2428	>	break;
2429		}
2430		}
2431		}
2432		}
2433
1323	–	// misc ForkJoinWorkerThread support
1324	–
2434		/**
2435	<	* Increment or decrement quiescerCount. Needed only to prevent
2436	<	* triggering shutdown if a worker is transiently inactive while
1328	<	* checking quiescence.
1329	<	*
1330	<	* @param delta 1 for increment, -1 for decrement
2435	>	* Tries to help execute or signal availability of the given task
2436	>	* from submitter's queue in common pool.
2437		*/
2438	<	final void addQuiescerCount(int delta) {
2439	<	int c;
2440	<	do {} while(!UNSAFE.compareAndSwapInt(this, quiescerCountOffset,
2441	<	c = quiescerCount, c + delta));
2438	>	static void externalHelpJoin(ForkJoinTask<?> t) {
2439	>	// Some hard-to-avoid overlap with tryExternalUnpush
2440	>	ForkJoinPool p; WorkQueue[] ws; WorkQueue q, w; Submitter z;
2441	>	ForkJoinTask<?>[] a;  int m, s, n;
2442	>	if (t != null &&
2443	>	(z = submitters.get()) != null &&
2444	>	(p = commonPool) != null &&
2445	>	(ws = p.workQueues) != null &&
2446	>	(m = ws.length - 1) >= 0 &&
2447	>	(q = ws[m & z.seed & SQMASK]) != null &&
2448	>	(a = q.array) != null) {
2449	>	int am = a.length - 1;
2450	>	if ((s = q.top) != q.base) {
2451	>	long j = ((am & (s - 1)) << ASHIFT) + ABASE;
2452	>	if (U.getObject(a, j) == t &&
2453	>	U.compareAndSwapInt(q, QLOCK, 0, 1)) {
2454	>	if (q.array == a && q.top == s &&
2455	>	U.compareAndSwapObject(a, j, t, null)) {
2456	>	q.top = s - 1;
2457	>	q.qlock = 0;
2458	>	t.doExec();
2459	>	}
2460	>	else
2461	>	q.qlock = 0;
2462	>	}
2463	>	}
2464	>	if (t.status >= 0) {
2465	>	if (t instanceof CountedCompleter)
2466	>	p.externalHelpComplete(q, t);
2467	>	else
2468	>	p.helpSignal(t, q.poolIndex);
2469	>	}
2470	>	}
2471		}
2472
2473		/**
2474	<	* Directly increment or decrement active count without
1340	<	* queuing. This method is used to transiently assert inactivation
1341	<	* while checking quiescence.
1342	<	*
1343	<	* @param delta 1 for increment, -1 for decrement
2474	>	* Restricted version of helpQuiescePool for external callers
2475		*/
2476	<	final void addActiveCount(int delta) {
2477	<	long d = delta < 0 ? -AC_UNIT : AC_UNIT;
2478	<	long c;
2479	<	do {} while (!UNSAFE.compareAndSwapLong(this, ctlOffset, c = ctl,
2480	<	((c + d) & AC_MASK) |
2481	<	(c & ~AC_MASK)));
2482	<	}
2483	<
2484	<	/**
2485	<	* Returns the approximate (non-atomic) number of idle threads per
1355	<	* active thread.
1356	<	*/
1357	<	final int idlePerActive() {
1358	<	// Approximate at powers of two for small values, saturate past 4
1359	<	int p = parallelism;
1360	<	int a = p + (int)(ctl >> AC_SHIFT);
1361	<	return (a > (p >>>= 1) ? 0 :
1362	<	a > (p >>>= 1) ? 1 :
1363	<	a > (p >>>= 1) ? 2 :
1364	<	a > (p >>>= 1) ? 4 :
1365	<	8);
2476	>	static void externalHelpQuiescePool() {
2477	>	ForkJoinPool p; ForkJoinTask<?> t; WorkQueue q; int b;
2478	>	if ((p = commonPool) != null &&
2479	>	(q = p.findNonEmptyStealQueue(1)) != null &&
2480	>	(b = q.base) - q.top < 0 &&
2481	>	(t = q.pollAt(b)) != null) {
2482	>	if (q.base - q.top < 0)
2483	>	p.signalWork(q);
2484	>	t.doExec();
2485	>	}
2486		}
2487
2488		// Exported methods
#	Line 1434 | Line 2554 | public class ForkJoinPool extends Abstra
2554		checkPermission();
2555		if (factory == null)
2556		throw new NullPointerException();
2557	<	if (parallelism <= 0 || parallelism > MAX_ID)
2557	>	if (parallelism <= 0 || parallelism > MAX_CAP)
2558		throw new IllegalArgumentException();
1439	–	this.parallelism = parallelism;
2559		this.factory = factory;
2560		this.ueh = handler;
2561	<	this.locallyFifo = asyncMode;
2561	>	this.config = parallelism | (asyncMode ? (FIFO_QUEUE << 16) : 0);
2562		long np = (long)(-parallelism); // offset ctl counts
2563		this.ctl = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK);
2564	<	this.submissionQueue = new ForkJoinTask<?>[INITIAL_QUEUE_CAPACITY];
1446	<	// initialize workers array with room for 2*parallelism if possible
1447	<	int n = parallelism << 1;
1448	<	if (n >= MAX_ID)
1449	<	n = MAX_ID;
1450	<	else { // See Hackers Delight, sec 3.2, where n < (1 << 16)
1451	<	n |= n >>> 1; n |= n >>> 2; n |= n >>> 4; n |= n >>> 8;
1452	<	}
1453	<	workers = new ForkJoinWorkerThread[n + 1];
1454	<	this.submissionLock = new ReentrantLock();
1455	<	this.termination = submissionLock.newCondition();
2564	>	int pn = nextPoolId();
2565		StringBuilder sb = new StringBuilder("ForkJoinPool-");
2566	<	sb.append(poolNumberGenerator.incrementAndGet());
2566	>	sb.append(Integer.toString(pn));
2567		sb.append("-worker-");
2568		this.workerNamePrefix = sb.toString();
2569		}
2570
2571	+	/**
2572	+	* Constructor for common pool, suitable only for static initialization.
2573	+	* Basically the same as above, but uses smallest possible initial footprint.
2574	+	*/
2575	+	ForkJoinPool(int parallelism, long ctl,
2576	+	ForkJoinWorkerThreadFactory factory,
2577	+	Thread.UncaughtExceptionHandler handler) {
2578	+	this.config = parallelism;
2579	+	this.ctl = ctl;
2580	+	this.factory = factory;
2581	+	this.ueh = handler;
2582	+	this.workerNamePrefix = "ForkJoinPool.commonPool-worker-";
2583	+	}
2584	+
2585	+	/**
2586	+	* Returns the common pool instance. This pool is statically
2587	+	* constructed; its run state is unaffected by attempts to
2588	+	* {@link #shutdown} or {@link #shutdownNow}.
2589	+	*
2590	+	* @return the common pool instance
2591	+	*/
2592	+	public static ForkJoinPool commonPool() {
2593	+	// assert commonPool != null : "static init error";
2594	+	return commonPool;
2595	+	}
2596	+
2597		// Execution methods
2598
2599		/**
#	Line 1478 | Line 2613 | public class ForkJoinPool extends Abstra
2613		*         scheduled for execution
2614		*/
2615		public <T> T invoke(ForkJoinTask<T> task) {
1481	–	Thread t = Thread.currentThread();
2616		if (task == null)
2617		throw new NullPointerException();
2618	<	if (shutdown)
2619	<	throw new RejectedExecutionException();
1486	<	if ((t instanceof ForkJoinWorkerThread) &&
1487	<	((ForkJoinWorkerThread)t).pool == this)
1488	<	return task.invoke();  // bypass submit if in same pool
1489	<	else {
1490	<	addSubmission(task);
1491	<	return task.join();
1492	<	}
1493	<	}
1494	<
1495	<	/**
1496	<	* Unless terminating, forks task if within an ongoing FJ
1497	<	* computation in the current pool, else submits as external task.
1498	<	*/
1499	<	private <T> void forkOrSubmit(ForkJoinTask<T> task) {
1500	<	ForkJoinWorkerThread w;
1501	<	Thread t = Thread.currentThread();
1502	<	if (shutdown)
1503	<	throw new RejectedExecutionException();
1504	<	if ((t instanceof ForkJoinWorkerThread) &&
1505	<	(w = (ForkJoinWorkerThread)t).pool == this)
1506	<	w.pushTask(task);
1507	<	else
1508	<	addSubmission(task);
2618	>	externalPush(task);
2619	>	return task.join();
2620		}
2621
2622		/**
#	Line 1519 | Line 2630 | public class ForkJoinPool extends Abstra
2630		public void execute(ForkJoinTask<?> task) {
2631		if (task == null)
2632		throw new NullPointerException();
2633	<	forkOrSubmit(task);
2633	>	externalPush(task);
2634		}
2635
2636		// AbstractExecutorService methods
#	Line 1536 | Line 2647 | public class ForkJoinPool extends Abstra
2647		if (task instanceof ForkJoinTask<?>) // avoid re-wrap
2648		job = (ForkJoinTask<?>) task;
2649		else
2650	<	job = ForkJoinTask.adapt(task, null);
2651	<	forkOrSubmit(job);
2650	>	job = new ForkJoinTask.AdaptedRunnableAction(task);
2651	>	externalPush(job);
2652		}
2653
2654		/**
#	Line 1552 | Line 2663 | public class ForkJoinPool extends Abstra
2663		public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) {
2664		if (task == null)
2665		throw new NullPointerException();
2666	<	forkOrSubmit(task);
2666	>	externalPush(task);
2667		return task;
2668		}
2669
#	Line 1562 | Line 2673 | public class ForkJoinPool extends Abstra
2673		*         scheduled for execution
2674		*/
2675		public <T> ForkJoinTask<T> submit(Callable<T> task) {
2676	<	if (task == null)
2677	<	throw new NullPointerException();
1567	<	ForkJoinTask<T> job = ForkJoinTask.adapt(task);
1568	<	forkOrSubmit(job);
2676	>	ForkJoinTask<T> job = new ForkJoinTask.AdaptedCallable<T>(task);
2677	>	externalPush(job);
2678		return job;
2679		}
2680
#	Line 1575 | Line 2684 | public class ForkJoinPool extends Abstra
2684		*         scheduled for execution
2685		*/
2686		public <T> ForkJoinTask<T> submit(Runnable task, T result) {
2687	<	if (task == null)
2688	<	throw new NullPointerException();
1580	<	ForkJoinTask<T> job = ForkJoinTask.adapt(task, result);
1581	<	forkOrSubmit(job);
2687	>	ForkJoinTask<T> job = new ForkJoinTask.AdaptedRunnable<T>(task, result);
2688	>	externalPush(job);
2689		return job;
2690		}
2691
#	Line 1594 | Line 2701 | public class ForkJoinPool extends Abstra
2701		if (task instanceof ForkJoinTask<?>) // avoid re-wrap
2702		job = (ForkJoinTask<?>) task;
2703		else
2704	<	job = ForkJoinTask.adapt(task, null);
2705	<	forkOrSubmit(job);
2704	>	job = new ForkJoinTask.AdaptedRunnableAction(task);
2705	>	externalPush(job);
2706		return job;
2707		}
2708
#	Line 1604 | Line 2711 | public class ForkJoinPool extends Abstra
2711		* @throws RejectedExecutionException {@inheritDoc}
2712		*/
2713		public <T> List<Future<T>> invokeAll(Collection<? extends Callable<T>> tasks) {
2714	<	ArrayList<ForkJoinTask<T>> forkJoinTasks =
2715	<	new ArrayList<ForkJoinTask<T>>(tasks.size());
2716	<	for (Callable<T> task : tasks)
2717	<	forkJoinTasks.add(ForkJoinTask.adapt(task));
2718	<	invoke(new InvokeAll<T>(forkJoinTasks));
2719	<
2714	>	// In previous versions of this class, this method constructed
2715	>	// a task to run ForkJoinTask.invokeAll, but now external
2716	>	// invocation of multiple tasks is at least as efficient.
2717	>	List<ForkJoinTask<T>> fs = new ArrayList<ForkJoinTask<T>>(tasks.size());
2718	>	// Workaround needed because method wasn't declared with
2719	>	// wildcards in return type but should have been.
2720		@SuppressWarnings({"unchecked", "rawtypes"})
2721	<	List<Future<T>> futures = (List<Future<T>>) (List) forkJoinTasks;
1615	<	return futures;
1616	<	}
2721	>	List<Future<T>> futures = (List<Future<T>>) (List) fs;
2722
2723	<	static final class InvokeAll<T> extends RecursiveAction {
2724	<	final ArrayList<ForkJoinTask<T>> tasks;
2725	<	InvokeAll(ArrayList<ForkJoinTask<T>> tasks) { this.tasks = tasks; }
2726	<	public void compute() {
2727	<	try { invokeAll(tasks); }
2728	<	catch (Exception ignore) {}
2723	>	boolean done = false;
2724	>	try {
2725	>	for (Callable<T> t : tasks) {
2726	>	ForkJoinTask<T> f = new ForkJoinTask.AdaptedCallable<T>(t);
2727	>	externalPush(f);
2728	>	fs.add(f);
2729	>	}
2730	>	for (ForkJoinTask<T> f : fs)
2731	>	f.quietlyJoin();
2732	>	done = true;
2733	>	return futures;
2734	>	} finally {
2735	>	if (!done)
2736	>	for (ForkJoinTask<T> f : fs)
2737	>	f.cancel(false);
2738		}
1625	–	private static final long serialVersionUID = -7914297376763021607L;
2739		}
2740
2741		/**
#	Line 1650 | Line 2763 | public class ForkJoinPool extends Abstra
2763		* @return the targeted parallelism level of this pool
2764		*/
2765		public int getParallelism() {
2766	<	return parallelism;
2766	>	return config & SMASK;
2767	>	}
2768	>
2769	>	/**
2770	>	* Returns the targeted parallelism level of the common pool.
2771	>	*
2772	>	* @return the targeted parallelism level of the common pool
2773	>	*/
2774	>	public static int getCommonPoolParallelism() {
2775	>	return commonPoolParallelism;
2776		}
2777
2778		/**
#	Line 1662 | Line 2784 | public class ForkJoinPool extends Abstra
2784		* @return the number of worker threads
2785		*/
2786		public int getPoolSize() {
2787	<	return parallelism + (short)(ctl >>> TC_SHIFT);
2787	>	return (config & SMASK) + (short)(ctl >>> TC_SHIFT);
2788		}
2789
2790		/**
#	Line 1672 | Line 2794 | public class ForkJoinPool extends Abstra
2794		* @return {@code true} if this pool uses async mode
2795		*/
2796		public boolean getAsyncMode() {
2797	<	return locallyFifo;
2797	>	return (config >>> 16) == FIFO_QUEUE;
2798		}
2799
2800		/**
#	Line 1684 | Line 2806 | public class ForkJoinPool extends Abstra
2806		* @return the number of worker threads
2807		*/
2808		public int getRunningThreadCount() {
2809	<	int r = parallelism + (int)(ctl >> AC_SHIFT);
2810	<	return r <= 0? 0 : r; // suppress momentarily negative values
2809	>	int rc = 0;
2810	>	WorkQueue[] ws; WorkQueue w;
2811	>	if ((ws = workQueues) != null) {
2812	>	for (int i = 1; i < ws.length; i += 2) {
2813	>	if ((w = ws[i]) != null && w.isApparentlyUnblocked())
2814	>	++rc;
2815	>	}
2816	>	}
2817	>	return rc;
2818		}
2819
2820		/**
#	Line 1696 | Line 2825 | public class ForkJoinPool extends Abstra
2825		* @return the number of active threads
2826		*/
2827		public int getActiveThreadCount() {
2828	<	int r = parallelism + (int)(ctl >> AC_SHIFT) + blockedCount;
2829	<	return r <= 0? 0 : r; // suppress momentarily negative values
2828	>	int r = (config & SMASK) + (int)(ctl >> AC_SHIFT);
2829	>	return (r <= 0) ? 0 : r; // suppress momentarily negative values
2830		}
2831
2832		/**
#	Line 1712 | Line 2841 | public class ForkJoinPool extends Abstra
2841		* @return {@code true} if all threads are currently idle
2842		*/
2843		public boolean isQuiescent() {
2844	<	return parallelism + (int)(ctl >> AC_SHIFT) + blockedCount == 0;
2844	>	return (int)(ctl >> AC_SHIFT) + (config & SMASK) == 0;
2845		}
2846
2847		/**
#	Line 1727 | Line 2856 | public class ForkJoinPool extends Abstra
2856		* @return the number of steals
2857		*/
2858		public long getStealCount() {
2859	<	return stealCount;
2859	>	long count = stealCount;
2860	>	WorkQueue[] ws; WorkQueue w;
2861	>	if ((ws = workQueues) != null) {
2862	>	for (int i = 1; i < ws.length; i += 2) {
2863	>	if ((w = ws[i]) != null)
2864	>	count += w.nsteals;
2865	>	}
2866	>	}
2867	>	return count;
2868		}
2869
2870		/**
#	Line 1742 | Line 2879 | public class ForkJoinPool extends Abstra
2879		*/
2880		public long getQueuedTaskCount() {
2881		long count = 0;
2882	<	ForkJoinWorkerThread[] ws;
2883	<	if ((short)(ctl >>> TC_SHIFT) > -parallelism &&
2884	<	(ws = workers) != null) {
2885	<	for (ForkJoinWorkerThread w : ws)
2886	<	if (w != null)
2887	<	count -= w.queueBase - w.queueTop; // must read base first
2882	>	WorkQueue[] ws; WorkQueue w;
2883	>	if ((ws = workQueues) != null) {
2884	>	for (int i = 1; i < ws.length; i += 2) {
2885	>	if ((w = ws[i]) != null)
2886	>	count += w.queueSize();
2887	>	}
2888		}
2889		return count;
2890		}
#	Line 1760 | Line 2897 | public class ForkJoinPool extends Abstra
2897		* @return the number of queued submissions
2898		*/
2899		public int getQueuedSubmissionCount() {
2900	<	return -queueBase + queueTop;
2900	>	int count = 0;
2901	>	WorkQueue[] ws; WorkQueue w;
2902	>	if ((ws = workQueues) != null) {
2903	>	for (int i = 0; i < ws.length; i += 2) {
2904	>	if ((w = ws[i]) != null)
2905	>	count += w.queueSize();
2906	>	}
2907	>	}
2908	>	return count;
2909		}
2910
2911		/**
#	Line 1770 | Line 2915 | public class ForkJoinPool extends Abstra
2915		* @return {@code true} if there are any queued submissions
2916		*/
2917		public boolean hasQueuedSubmissions() {
2918	<	return queueBase != queueTop;
2918	>	WorkQueue[] ws; WorkQueue w;
2919	>	if ((ws = workQueues) != null) {
2920	>	for (int i = 0; i < ws.length; i += 2) {
2921	>	if ((w = ws[i]) != null && !w.isEmpty())
2922	>	return true;
2923	>	}
2924	>	}
2925	>	return false;
2926		}
2927
2928		/**
#	Line 1781 | Line 2933 | public class ForkJoinPool extends Abstra
2933		* @return the next submission, or {@code null} if none
2934		*/
2935		protected ForkJoinTask<?> pollSubmission() {
2936	<	ForkJoinTask<?> t; ForkJoinTask<?>[] q; int b, i;
2937	<	while ((b = queueBase) != queueTop &&
2938	<	(q = submissionQueue) != null &&
2939	<	(i = (q.length - 1) & b) >= 0) {
2940	<	long u = (i << ASHIFT) + ABASE;
1789	<	if ((t = q[i]) != null &&
1790	<	queueBase == b &&
1791	<	UNSAFE.compareAndSwapObject(q, u, t, null)) {
1792	<	queueBase = b + 1;
1793	<	return t;
2936	>	WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2937	>	if ((ws = workQueues) != null) {
2938	>	for (int i = 0; i < ws.length; i += 2) {
2939	>	if ((w = ws[i]) != null && (t = w.poll()) != null)
2940	>	return t;
2941		}
2942		}
2943		return null;
#	Line 1815 | Line 2962 | public class ForkJoinPool extends Abstra
2962		*/
2963		protected int drainTasksTo(Collection<? super ForkJoinTask<?>> c) {
2964		int count = 0;
2965	<	while (queueBase != queueTop) {
2966	<	ForkJoinTask<?> t = pollSubmission();
2967	<	if (t != null) {
2968	<	c.add(t);
2969	<	++count;
2965	>	WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2966	>	if ((ws = workQueues) != null) {
2967	>	for (int i = 0; i < ws.length; ++i) {
2968	>	if ((w = ws[i]) != null) {
2969	>	while ((t = w.poll()) != null) {
2970	>	c.add(t);
2971	>	++count;
2972	>	}
2973	>	}
2974		}
2975		}
1825	–	ForkJoinWorkerThread[] ws;
1826	–	if ((short)(ctl >>> TC_SHIFT) > -parallelism &&
1827	–	(ws = workers) != null) {
1828	–	for (ForkJoinWorkerThread w : ws)
1829	–	if (w != null)
1830	–	count += w.drainTasksTo(c);
1831	–	}
2976		return count;
2977		}
2978
#	Line 1840 | Line 2984 | public class ForkJoinPool extends Abstra
2984		* @return a string identifying this pool, as well as its state
2985		*/
2986		public String toString() {
2987	<	long st = getStealCount();
2988	<	long qt = getQueuedTaskCount();
2989	<	long qs = getQueuedSubmissionCount();
1846	<	int pc = parallelism;
2987	>	// Use a single pass through workQueues to collect counts
2988	>	long qt = 0L, qs = 0L; int rc = 0;
2989	>	long st = stealCount;
2990		long c = ctl;
2991	+	WorkQueue[] ws; WorkQueue w;
2992	+	if ((ws = workQueues) != null) {
2993	+	for (int i = 0; i < ws.length; ++i) {
2994	+	if ((w = ws[i]) != null) {
2995	+	int size = w.queueSize();
2996	+	if ((i & 1) == 0)
2997	+	qs += size;
2998	+	else {
2999	+	qt += size;
3000	+	st += w.nsteals;
3001	+	if (w.isApparentlyUnblocked())
3002	+	++rc;
3003	+	}
3004	+	}
3005	+	}
3006	+	}
3007	+	int pc = (config & SMASK);
3008		int tc = pc + (short)(c >>> TC_SHIFT);
3009	<	int rc = pc + (int)(c >> AC_SHIFT);
3010	<	if (rc < 0) // ignore transient negative
3011	<	rc = 0;
1852	<	int ac = rc + blockedCount;
3009	>	int ac = pc + (int)(c >> AC_SHIFT);
3010	>	if (ac < 0) // ignore transient negative
3011	>	ac = 0;
3012		String level;
3013		if ((c & STOP_BIT) != 0)
3014	<	level = (tc == 0)? "Terminated" : "Terminating";
3014	>	level = (tc == 0) ? "Terminated" : "Terminating";
3015		else
3016	<	level = shutdown? "Shutting down" : "Running";
3016	>	level = plock < 0 ? "Shutting down" : "Running";
3017		return super.toString() +
3018		"[" + level +
3019		", parallelism = " + pc +
#	Line 1868 | Line 3027 | public class ForkJoinPool extends Abstra
3027		}
3028
3029		/**
3030	<	* Initiates an orderly shutdown in which previously submitted
3031	<	* tasks are executed, but no new tasks will be accepted.
3032	<	* Invocation has no additional effect if already shut down.
3033	<	* Tasks that are in the process of being submitted concurrently
3034	<	* during the course of this method may or may not be rejected.
3030	>	* Possibly initiates an orderly shutdown in which previously
3031	>	* submitted tasks are executed, but no new tasks will be
3032	>	* accepted. Invocation has no effect on execution state if this
3033	>	* is the {@link #commonPool}, and no additional effect if
3034	>	* already shut down.  Tasks that are in the process of being
3035	>	* submitted concurrently during the course of this method may or
3036	>	* may not be rejected.
3037		*
3038		* @throws SecurityException if a security manager exists and
3039		*         the caller is not permitted to modify threads
#	Line 1881 | Line 3042 | public class ForkJoinPool extends Abstra
3042		*/
3043		public void shutdown() {
3044		checkPermission();
3045	<	shutdown = true;
1885	<	tryTerminate(false);
3045	>	tryTerminate(false, true);
3046		}
3047
3048		/**
3049	<	* Attempts to cancel and/or stop all tasks, and reject all
3050	<	* subsequently submitted tasks.  Tasks that are in the process of
3051	<	* being submitted or executed concurrently during the course of
3052	<	* this method may or may not be rejected. This method cancels
3053	<	* both existing and unexecuted tasks, in order to permit
3054	<	* termination in the presence of task dependencies. So the method
3055	<	* always returns an empty list (unlike the case for some other
3056	<	* Executors).
3049	>	* Possibly attempts to cancel and/or stop all tasks, and reject
3050	>	* all subsequently submitted tasks.  Invocation has no effect on
3051	>	* execution state if this is the {@link #commonPool}, and no
3052	>	* additional effect if already shut down. Otherwise, tasks that
3053	>	* are in the process of being submitted or executed concurrently
3054	>	* during the course of this method may or may not be
3055	>	* rejected. This method cancels both existing and unexecuted
3056	>	* tasks, in order to permit termination in the presence of task
3057	>	* dependencies. So the method always returns an empty list
3058	>	* (unlike the case for some other Executors).
3059		*
3060		* @return an empty list
3061		* @throws SecurityException if a security manager exists and
#	Line 1903 | Line 3065 | public class ForkJoinPool extends Abstra
3065		*/
3066		public List<Runnable> shutdownNow() {
3067		checkPermission();
3068	<	shutdown = true;
1907	<	tryTerminate(true);
3068	>	tryTerminate(true, true);
3069		return Collections.emptyList();
3070		}
3071
#	Line 1916 | Line 3077 | public class ForkJoinPool extends Abstra
3077		public boolean isTerminated() {
3078		long c = ctl;
3079		return ((c & STOP_BIT) != 0L &&
3080	<	(short)(c >>> TC_SHIFT) == -parallelism);
3080	>	(short)(c >>> TC_SHIFT) == -(config & SMASK));
3081		}
3082
3083		/**
#	Line 1924 | Line 3085 | public class ForkJoinPool extends Abstra
3085		* commenced but not yet completed.  This method may be useful for
3086		* debugging. A return of {@code true} reported a sufficient
3087		* period after shutdown may indicate that submitted tasks have
3088	<	* ignored or suppressed interruption, or are waiting for IO,
3088	>	* ignored or suppressed interruption, or are waiting for I/O,
3089		* causing this executor not to properly terminate. (See the
3090		* advisory notes for class {@link ForkJoinTask} stating that
3091		* tasks should not normally entail blocking operations.  But if
#	Line 1935 | Line 3096 | public class ForkJoinPool extends Abstra
3096		public boolean isTerminating() {
3097		long c = ctl;
3098		return ((c & STOP_BIT) != 0L &&
3099	<	(short)(c >>> TC_SHIFT) != -parallelism);
1939	<	}
1940	<
1941	<	/**
1942	<	* Returns true if terminating or terminated. Used by ForkJoinWorkerThread.
1943	<	*/
1944	<	final boolean isAtLeastTerminating() {
1945	<	return (ctl & STOP_BIT) != 0L;
3099	>	(short)(c >>> TC_SHIFT) != -(config & SMASK));
3100		}
3101
3102		/**
#	Line 1951 | Line 3105 | public class ForkJoinPool extends Abstra
3105		* @return {@code true} if this pool has been shut down
3106		*/
3107		public boolean isShutdown() {
3108	<	return shutdown;
3108	>	return plock < 0;
3109		}
3110
3111		/**
3112	<	* Blocks until all tasks have completed execution after a shutdown
3113	<	* request, or the timeout occurs, or the current thread is
3114	<	* interrupted, whichever happens first.
3112	>	* Blocks until all tasks have completed execution after a
3113	>	* shutdown request, or the timeout occurs, or the current thread
3114	>	* is interrupted, whichever happens first. Note that the {@link
3115	>	* #commonPool()} never terminates until program shutdown so
3116	>	* this method will always time out.
3117		*
3118		* @param timeout the maximum time to wait
3119		* @param unit the time unit of the timeout argument
#	Line 1968 | Line 3124 | public class ForkJoinPool extends Abstra
3124		public boolean awaitTermination(long timeout, TimeUnit unit)
3125		throws InterruptedException {
3126		long nanos = unit.toNanos(timeout);
3127	<	final ReentrantLock lock = this.submissionLock;
3128	<	lock.lock();
3129	<	try {
3130	<	for (;;) {
3131	<	if (isTerminated())
3132	<	return true;
3133	<	if (nanos <= 0)
3134	<	return false;
3135	<	nanos = termination.awaitNanos(nanos);
3127	>	if (isTerminated())
3128	>	return true;
3129	>	long startTime = System.nanoTime();
3130	>	boolean terminated = false;
3131	>	synchronized (this) {
3132	>	for (long waitTime = nanos, millis = 0L;;) {
3133	>	if (terminated = isTerminated() ||
3134	>	waitTime <= 0L ||
3135	>	(millis = unit.toMillis(waitTime)) <= 0L)
3136	>	break;
3137	>	wait(millis);
3138	>	waitTime = nanos - (System.nanoTime() - startTime);
3139		}
1981	–	} finally {
1982	–	lock.unlock();
3140		}
3141	+	return terminated;
3142		}
3143
3144		/**
#	Line 2080 | Line 3238 | public class ForkJoinPool extends Abstra
3238		throws InterruptedException {
3239		Thread t = Thread.currentThread();
3240		if (t instanceof ForkJoinWorkerThread) {
3241	<	ForkJoinWorkerThread w = (ForkJoinWorkerThread) t;
3242	<	w.pool.awaitBlocker(blocker);
3241	>	ForkJoinPool p = ((ForkJoinWorkerThread)t).pool;
3242	>	while (!blocker.isReleasable()) { // variant of helpSignal
3243	>	WorkQueue[] ws; WorkQueue q; int m, u;
3244	>	if ((ws = p.workQueues) != null && (m = ws.length - 1) >= 0) {
3245	>	for (int i = 0; i <= m; ++i) {
3246	>	if (blocker.isReleasable())
3247	>	return;
3248	>	if ((q = ws[i]) != null && q.base - q.top < 0) {
3249	>	p.signalWork(q);
3250	>	if ((u = (int)(p.ctl >>> 32)) >= 0 ||
3251	>	(u >> UAC_SHIFT) >= 0)
3252	>	break;
3253	>	}
3254	>	}
3255	>	}
3256	>	if (p.tryCompensate()) {
3257	>	try {
3258	>	do {} while (!blocker.isReleasable() &&
3259	>	!blocker.block());
3260	>	} finally {
3261	>	p.incrementActiveCount();
3262	>	}
3263	>	break;
3264	>	}
3265	>	}
3266		}
3267		else {
3268	<	do {} while (!blocker.isReleasable() && !blocker.block());
3268	>	do {} while (!blocker.isReleasable() &&
3269	>	!blocker.block());
3270		}
3271		}
3272
#	Line 2093 | Line 3275 | public class ForkJoinPool extends Abstra
3275		// implement RunnableFuture.
3276
3277		protected <T> RunnableFuture<T> newTaskFor(Runnable runnable, T value) {
3278	<	return (RunnableFuture<T>) ForkJoinTask.adapt(runnable, value);
3278	>	return new ForkJoinTask.AdaptedRunnable<T>(runnable, value);
3279		}
3280
3281		protected <T> RunnableFuture<T> newTaskFor(Callable<T> callable) {
3282	<	return (RunnableFuture<T>) ForkJoinTask.adapt(callable);
3282	>	return new ForkJoinTask.AdaptedCallable<T>(callable);
3283		}
3284
3285		// Unsafe mechanics
3286	<	private static final sun.misc.Unsafe UNSAFE;
3287	<	private static final long ctlOffset;
3288	<	private static final long stealCountOffset;
3289	<	private static final long blockedCountOffset;
2108	<	private static final long quiescerCountOffset;
2109	<	private static final long scanGuardOffset;
2110	<	private static final long nextWorkerNumberOffset;
2111	<	private static final long ABASE;
3286	>	private static final sun.misc.Unsafe U;
3287	>	private static final long CTL;
3288	>	private static final long PARKBLOCKER;
3289	>	private static final int ABASE;
3290		private static final int ASHIFT;
3291	+	private static final long STEALCOUNT;
3292	+	private static final long PLOCK;
3293	+	private static final long INDEXSEED;
3294	+	private static final long QLOCK;
3295
3296		static {
3297	<	poolNumberGenerator = new AtomicInteger();
2116	<	workerSeedGenerator = new Random();
2117	<	modifyThreadPermission = new RuntimePermission("modifyThread");
2118	<	defaultForkJoinWorkerThreadFactory =
2119	<	new DefaultForkJoinWorkerThreadFactory();
2120	<	int s;
3297	>	int s; // initialize field offsets for CAS etc
3298		try {
3299	<	UNSAFE = getUnsafe();
3300	<	Class k = ForkJoinPool.class;
3301	<	ctlOffset = UNSAFE.objectFieldOffset
3299	>	U = getUnsafe();
3300	>	Class<?> k = ForkJoinPool.class;
3301	>	CTL = U.objectFieldOffset
3302		(k.getDeclaredField("ctl"));
3303	<	stealCountOffset = UNSAFE.objectFieldOffset
3303	>	STEALCOUNT = U.objectFieldOffset
3304		(k.getDeclaredField("stealCount"));
3305	<	blockedCountOffset = UNSAFE.objectFieldOffset
3306	<	(k.getDeclaredField("blockedCount"));
3307	<	quiescerCountOffset = UNSAFE.objectFieldOffset
3308	<	(k.getDeclaredField("quiescerCount"));
3309	<	scanGuardOffset = UNSAFE.objectFieldOffset
3310	<	(k.getDeclaredField("scanGuard"));
3311	<	nextWorkerNumberOffset = UNSAFE.objectFieldOffset
3312	<	(k.getDeclaredField("nextWorkerNumber"));
3313	<	Class a = ForkJoinTask[].class;
3314	<	ABASE = UNSAFE.arrayBaseOffset(a);
3315	<	s = UNSAFE.arrayIndexScale(a);
3305	>	PLOCK = U.objectFieldOffset
3306	>	(k.getDeclaredField("plock"));
3307	>	INDEXSEED = U.objectFieldOffset
3308	>	(k.getDeclaredField("indexSeed"));
3309	>	Class<?> tk = Thread.class;
3310	>	PARKBLOCKER = U.objectFieldOffset
3311	>	(tk.getDeclaredField("parkBlocker"));
3312	>	Class<?> wk = WorkQueue.class;
3313	>	QLOCK = U.objectFieldOffset
3314	>	(wk.getDeclaredField("qlock"));
3315	>	Class<?> ak = ForkJoinTask[].class;
3316	>	ABASE = U.arrayBaseOffset(ak);
3317	>	s = U.arrayIndexScale(ak);
3318	>	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
3319		} catch (Exception e) {
3320		throw new Error(e);
3321		}
3322		if ((s & (s-1)) != 0)
3323		throw new Error("data type scale not a power of two");
3324	<	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
3324	>
3325	>	submitters = new ThreadLocal<Submitter>();
3326	>	ForkJoinWorkerThreadFactory fac = defaultForkJoinWorkerThreadFactory =
3327	>	new DefaultForkJoinWorkerThreadFactory();
3328	>	modifyThreadPermission = new RuntimePermission("modifyThread");
3329	>
3330	>	/*
3331	>	* Establish common pool parameters.  For extra caution,
3332	>	* computations to set up common pool state are here; the
3333	>	* constructor just assigns these values to fields.
3334	>	*/
3335	>
3336	>	int par = 0;
3337	>	Thread.UncaughtExceptionHandler handler = null;
3338	>	try {  // TBD: limit or report ignored exceptions?
3339	>	String pp = System.getProperty
3340	>	("java.util.concurrent.ForkJoinPool.common.parallelism");
3341	>	String hp = System.getProperty
3342	>	("java.util.concurrent.ForkJoinPool.common.exceptionHandler");
3343	>	String fp = System.getProperty
3344	>	("java.util.concurrent.ForkJoinPool.common.threadFactory");
3345	>	if (fp != null)
3346	>	fac = ((ForkJoinWorkerThreadFactory)ClassLoader.
3347	>	getSystemClassLoader().loadClass(fp).newInstance());
3348	>	if (hp != null)
3349	>	handler = ((Thread.UncaughtExceptionHandler)ClassLoader.
3350	>	getSystemClassLoader().loadClass(hp).newInstance());
3351	>	if (pp != null)
3352	>	par = Integer.parseInt(pp);
3353	>	} catch (Exception ignore) {
3354	>	}
3355	>
3356	>	if (par <= 0)
3357	>	par = Runtime.getRuntime().availableProcessors();
3358	>	if (par > MAX_CAP)
3359	>	par = MAX_CAP;
3360	>	commonPoolParallelism = par;
3361	>	long np = (long)(-par); // precompute initial ctl value
3362	>	long ct = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK);
3363	>
3364	>	commonPool = new ForkJoinPool(par, ct, fac, handler);
3365		}
3366
3367		/**
#	Line 2171 | Line 3391 | public class ForkJoinPool extends Abstra
3391		}
3392		}
3393		}
3394	+
3395		}

Diff Legend

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