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 1.105 by jsr166, Fri Jun 10 18:10:53 2011 UTC vs.
Revision 1.145 by jsr166, Mon Nov 19 01:04:24 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.atomic.AtomicInteger;
23	–	import java.util.concurrent.locks.LockSupport;
24	–	import java.util.concurrent.locks.ReentrantLock;
25	–	import java.util.concurrent.locks.Condition;
21
22		/**
23		* An {@link ExecutorService} for running {@link ForkJoinTask}s.
#	Line 33 | Line 28 | import java.util.concurrent.locks.Condit
28		* <p>A {@code ForkJoinPool} differs from other kinds of {@link
29		* ExecutorService} mainly by virtue of employing
30		* <em>work-stealing</em>: all threads in the pool attempt to find and
31	<	* execute subtasks created by other active tasks (eventually blocking
32	<	* waiting for work if none exist). This enables efficient processing
33	<	* when most tasks spawn other subtasks (as do most {@code
34	<	* ForkJoinTask}s). When setting <em>asyncMode</em> to true in
35	<	* constructors, {@code ForkJoinPool}s may also be appropriate for use
36	<	* with event-style tasks that are never joined.
31	>	* execute tasks submitted to the pool and/or created by other active
32	>	* tasks (eventually blocking waiting for work if none exist). This
33	>	* enables efficient processing when most tasks spawn other subtasks
34	>	* (as do most {@code ForkJoinTask}s), as well as when many small
35	>	* tasks are submitted to the pool from external clients.  Especially
36	>	* when setting <em>asyncMode</em> to true in constructors, {@code
37	>	* ForkJoinPool}s may also be appropriate for use with event-style
38	>	* tasks that are never joined.
39		*
40	<	* <p>A {@code ForkJoinPool} is constructed with a given target
41	<	* parallelism level; by default, equal to the number of available
42	<	* processors. The pool attempts to maintain enough active (or
43	<	* available) threads by dynamically adding, suspending, or resuming
44	<	* internal worker threads, even if some tasks are stalled waiting to
45	<	* join others. However, no such adjustments are guaranteed in the
46	<	* face of blocked IO or other unmanaged synchronization. The nested
47	<	* {@link ManagedBlocker} interface enables extension of the kinds of
40	>	* <p>A static {@link #commonPool} is available and appropriate for
41	>	* most applications. The common pool is used by any ForkJoinTask that
42	>	* is not explicitly submitted to a specified pool. Using the common
43	>	* pool normally reduces resource usage (its threads are slowly
44	>	* reclaimed during periods of non-use, and reinstated upon subsequent
45	>	* use).
46	>	*
47	>	* <p>For applications that require separate or custom pools, a {@code
48	>	* ForkJoinPool} may be constructed with a given target parallelism
49	>	* level; by default, equal to the number of available processors. The
50	>	* pool attempts to maintain enough active (or available) threads by
51	>	* dynamically adding, suspending, or resuming internal worker
52	>	* threads, even if some tasks are stalled waiting to join
53	>	* others. However, no such adjustments are guaranteed in the face of
54	>	* blocked IO or other unmanaged synchronization. The nested {@link
55	>	* ManagedBlocker} interface enables extension of the kinds of
56		* synchronization accommodated.
57		*
58		* <p>In addition to execution and lifecycle control methods, this
#	Line 57 | 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 91 | 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	<	*
104	<	*  <pre> {@code
105	<	* static final ForkJoinPool mainPool = new ForkJoinPool();
106	<	* ...
107	<	* public void sort(long[] array) {
108	<	*   mainPool.invoke(new SortTask(array, 0, array.length));
109	<	* }}</pre>
100	>	* <p>The common pool is by default constructed with default
101	>	* parameters, but these may be controlled by setting three {@link
102	>	* System#getProperty properties} with prefix {@code
103	>	* java.util.concurrent.ForkJoinPool.common}: {@code parallelism} --
104	>	* an integer greater than zero, {@code threadFactory} -- the class
105	>	* name of a {@link ForkJoinWorkerThreadFactory}, and {@code
106	>	* exceptionHandler} -- the class name of a {@link
107	>	* java.lang.Thread.UncaughtExceptionHandler
108	>	* Thread.UncaughtExceptionHandler}. Upon any error in establishing
109	>	* these settings, default parameters are used.
110		*
111		* <p><b>Implementation notes</b>: This implementation restricts the
112		* maximum number of running threads to 32767. Attempts to create
#	Line 125 | 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	>	* blocking 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 is
320	>	* apparently empty, they signal waiting workers (or trigger
321	>	* creation of new ones if fewer than the given parallelism level
322	>	* -- see signalWork).  These primary signals are buttressed by
323	>	* signals whenever other threads scan for work or do not have a
324	>	* task to process. On most platforms, signalling (unpark)
325	>	* overhead time is noticeably long, and the time between
326	>	* signalling a thread and it actually making progress can be very
327	>	* noticeably long, so it is worth offloading these delays from
328	>	* critical paths as much as possible.
329		*
330		* Trimming workers. To release resources after periods of lack of
331		* use, a worker starting to wait when the pool is quiescent will
332	<	* time out and terminate if the pool has remained quiescent for
333	<	* SHRINK_RATE nanosecs. This will slowly propagate, eventually
334	<	* terminating all workers after long periods of non-use.
335	<	*
336	<	* Submissions. External submissions are maintained in an
337	<	* array-based queue that is structured identically to
338	<	* ForkJoinWorkerThread queues except for the use of
339	<	* submissionLock in method addSubmission. Unlike the case for
340	<	* worker queues, multiple external threads can add new
341	<	* submissions, so adding requires a lock.
342	<	*
343	<	* Compensation. Beyond work-stealing support and lifecycle
344	<	* control, the main responsibility of this framework is to take
345	<	* actions when one worker is waiting to join a task stolen (or
346	<	* always held by) another.  Because we are multiplexing many
347	<	* tasks on to a pool of workers, we can't just let them block (as
348	<	* in Thread.join).  We also cannot just reassign the joiner's
349	<	* run-time stack with another and replace it later, which would
350	<	* be a form of "continuation", that even if possible is not
351	<	* necessarily a good idea since we sometimes need both an
352	<	* unblocked task and its continuation to progress. Instead we
353	<	* combine two tactics:
332	>	* time out and terminate if the pool has remained quiescent for a
333	>	* given period -- a short period if there are more threads than
334	>	* parallelism, longer as the number of threads decreases. This
335	>	* will slowly propagate, eventually terminating all workers after
336	>	* periods of non-use.
337	>	*
338	>	* Shutdown and Termination. A call to shutdownNow atomically sets
339	>	* a plock bit and then (non-atomically) sets each worker's
340	>	* qlock status, cancels all unprocessed tasks, and wakes up
341	>	* all waiting workers.  Detecting whether termination should
342	>	* commence after a non-abrupt shutdown() call requires more work
343	>	* and bookkeeping. We need consensus about quiescence (i.e., that
344	>	* there is no more work). The active count provides a primary
345	>	* indication but non-abrupt shutdown still requires a rechecking
346	>	* scan for any workers that are inactive but not queued.
347	>	*
348	>	* Joining Tasks
349	>	* =============
350	>	*
351	>	* Any of several actions may be taken when one worker is waiting
352	>	* to join a task stolen (or always held) by another.  Because we
353	>	* are multiplexing many tasks on to a pool of workers, we can't
354	>	* just let them block (as in Thread.join).  We also cannot just
355	>	* reassign the joiner's run-time stack with another and replace
356	>	* it later, which would be a form of "continuation", that even if
357	>	* possible is not necessarily a good idea since we sometimes need
358	>	* both an unblocked task and its continuation to progress.
359	>	* Instead we combine two tactics:
360		*
361		*   Helping: Arranging for the joiner to execute some task that it
362	<	*      would be running if the steal had not occurred.  Method
258	<	*      ForkJoinWorkerThread.joinTask tracks joining->stealing
259	<	*      links to try to find such a task.
362	>	*      would be running if the steal had not occurred.
363		*
364		*   Compensating: Unless there are already enough live threads,
365	<	*      method tryPreBlock() may create or re-activate a spare
366	<	*      thread to compensate for blocked joiners until they
367	<	*      unblock.
365	>	*      method tryCompensate() may create or re-activate a spare
366	>	*      thread to compensate for blocked joiners until they unblock.
367	>	*
368	>	* A third form (implemented in tryRemoveAndExec) amounts to
369	>	* helping a hypothetical compensator: If we can readily tell that
370	>	* a possible action of a compensator is to steal and execute the
371	>	* task being joined, the joining thread can do so directly,
372	>	* without the need for a compensation thread (although at the
373	>	* expense of larger run-time stacks, but the tradeoff is
374	>	* typically worthwhile).
375		*
376		* The ManagedBlocker extension API can't use helping so relies
377		* only on compensation in method awaitBlocker.
378		*
379	+	* The algorithm in tryHelpStealer entails a form of "linear"
380	+	* helping: Each worker records (in field currentSteal) the most
381	+	* recent task it stole from some other worker. Plus, it records
382	+	* (in field currentJoin) the task it is currently actively
383	+	* joining. Method tryHelpStealer uses these markers to try to
384	+	* find a worker to help (i.e., steal back a task from and execute
385	+	* it) that could hasten completion of the actively joined task.
386	+	* In essence, the joiner executes a task that would be on its own
387	+	* local deque had the to-be-joined task not been stolen. This may
388	+	* be seen as a conservative variant of the approach in Wagner &
389	+	* Calder "Leapfrogging: a portable technique for implementing
390	+	* efficient futures" SIGPLAN Notices, 1993
391	+	* (http://portal.acm.org/citation.cfm?id=155354). It differs in
392	+	* that: (1) We only maintain dependency links across workers upon
393	+	* steals, rather than use per-task bookkeeping.  This sometimes
394	+	* requires a linear scan of workQueues array to locate stealers,
395	+	* but often doesn't because stealers leave hints (that may become
396	+	* stale/wrong) of where to locate them.  A stealHint is only a
397	+	* hint because a worker might have had multiple steals and the
398	+	* hint records only one of them (usually the most current).
399	+	* Hinting isolates cost to when it is needed, rather than adding
400	+	* to per-task overhead.  (2) It is "shallow", ignoring nesting
401	+	* and potentially cyclic mutual steals.  (3) It is intentionally
402	+	* racy: field currentJoin is updated only while actively joining,
403	+	* which means that we miss links in the chain during long-lived
404	+	* tasks, GC stalls etc (which is OK since blocking in such cases
405	+	* is usually a good idea).  (4) We bound the number of attempts
406	+	* to find work (see MAX_HELP) and fall back to suspending the
407	+	* worker and if necessary replacing it with another.
408	+	*
409	+	* Helping actions for CountedCompleters are much simpler: Method
410	+	* helpComplete can take and execute any task with the same root
411	+	* as the task being waited on. However, this still entails some
412	+	* traversal of completer chains, so is less efficient than using
413	+	* CountedCompleters without explicit joins.
414	+	*
415		* It is impossible to keep exactly the target parallelism number
416		* of threads running at any given time.  Determining the
417		* existence of conservatively safe helping targets, the
418		* availability of already-created spares, and the apparent need
419	<	* to create new spares are all racy and require heuristic
420	<	* guidance, so we rely on multiple retries of each.  Currently,
421	<	* in keeping with on-demand signalling policy, we compensate only
422	<	* if blocking would leave less than one active (non-waiting,
423	<	* non-blocked) worker. Additionally, to avoid some false alarms
424	<	* due to GC, lagging counters, system activity, etc, compensated
425	<	* blocking for joins is only attempted after rechecks stabilize
426	<	* (retries are interspersed with Thread.yield, for good
427	<	* citizenship).  The variable blockedCount, incremented before
428	<	* blocking and decremented after, is sometimes needed to
429	<	* distinguish cases of waiting for work vs blocking on joins or
430	<	* other managed sync. Both cases are equivalent for most pool
431	<	* control, so we can update non-atomically. (Additionally,
432	<	* contention on blockedCount alleviates some contention on ctl).
433	<	*
434	<	* Shutdown and Termination. A call to shutdownNow atomically sets
435	<	* the ctl stop bit and then (non-atomically) sets each workers
436	<	* "terminate" status, cancels all unprocessed tasks, and wakes up
437	<	* all waiting workers.  Detecting whether termination should
438	<	* commence after a non-abrupt shutdown() call requires more work
439	<	* and bookkeeping. We need consensus about quiescence (i.e., that
440	<	* there is no more work) which is reflected in active counts so
441	<	* long as there are no current blockers, as well as possible
442	<	* re-evaluations during independent changes in blocking or
443	<	* quiescing workers.
444	<	*
445	<	* Style notes: There is a lot of representation-level coupling
446	<	* among classes ForkJoinPool, ForkJoinWorkerThread, and
447	<	* ForkJoinTask.  Most fields of ForkJoinWorkerThread maintain
448	<	* data structures managed by ForkJoinPool, so are directly
449	<	* accessed.  Conversely we allow access to "workers" array by
450	<	* workers, and direct access to ForkJoinTask.status by both
451	<	* ForkJoinPool and ForkJoinWorkerThread.  There is little point
419	>	* to create new spares are all racy, so we rely on multiple
420	>	* retries of each.  Compensation in the apparent absence of
421	>	* helping opportunities is challenging to control on JVMs, where
422	>	* GC and other activities can stall progress of tasks that in
423	>	* turn stall out many other dependent tasks, without us being
424	>	* able to determine whether they will ever require compensation.
425	>	* Even though work-stealing otherwise encounters little
426	>	* degradation in the presence of more threads than cores,
427	>	* aggressively adding new threads in such cases entails risk of
428	>	* unwanted positive feedback control loops in which more threads
429	>	* cause more dependent stalls (as well as delayed progress of
430	>	* unblocked threads to the point that we know they are available)
431	>	* leading to more situations requiring more threads, and so
432	>	* on. This aspect of control can be seen as an (analytically
433	>	* intractable) game with an opponent that may choose the worst
434	>	* (for us) active thread to stall at any time.  We take several
435	>	* precautions to bound losses (and thus bound gains), mainly in
436	>	* methods tryCompensate and awaitJoin.
437	>	*
438	>	* Common Pool
439	>	* ===========
440	>	*
441	>	* The static commonPool always exists after static
442	>	* initialization.  Since it (or any other created pool) need
443	>	* never be used, we minimize initial construction overhead and
444	>	* footprint to the setup of about a dozen fields, with no nested
445	>	* allocation. Most bootstrapping occurs within method
446	>	* fullExternalPush during the first submission to the pool.
447	>	*
448	>	* When external threads submit to the common pool, they can
449	>	* perform some subtask processing (see externalHelpJoin and
450	>	* related methods).  We do not need to record whether these
451	>	* submissions are to the common pool -- if not, externalHelpJoin
452	>	* returns quickly (at the most helping to signal some common pool
453	>	* workers). These submitters would otherwise be blocked waiting
454	>	* for completion, so the extra effort (with liberally sprinkled
455	>	* task status checks) in inapplicable cases amounts to an odd
456	>	* form of limited spin-wait before blocking in ForkJoinTask.join.
457	>	*
458	>	* Style notes
459	>	* ===========
460	>	*
461	>	* There is a lot of representation-level coupling among classes
462	>	* ForkJoinPool, ForkJoinWorkerThread, and ForkJoinTask.  The
463	>	* fields of WorkQueue maintain data structures managed by
464	>	* ForkJoinPool, so are directly accessed.  There is little point
465		* trying to reduce this, since any associated future changes in
466		* representations will need to be accompanied by algorithmic
467	<	* changes anyway. All together, these low-level implementation
468	<	* choices produce as much as a factor of 4 performance
469	<	* improvement compared to naive implementations, and enable the
470	<	* processing of billions of tasks per second, at the expense of
471	<	* some ugliness.
472	<	*
473	<	* Methods signalWork() and scan() are the main bottlenecks so are
474	<	* especially heavily micro-optimized/mangled.  There are lots of
475	<	* inline assignments (of form "while ((local = field) != 0)")
476	<	* which are usually the simplest way to ensure the required read
477	<	* orderings (which are sometimes critical). This leads to a
478	<	* "C"-like style of listing declarations of these locals at the
479	<	* heads of methods or blocks.  There are several occurrences of
480	<	* the unusual "do {} while (!cas...)"  which is the simplest way
481	<	* to force an update of a CAS'ed variable. There are also other
482	<	* coding oddities that help some methods perform reasonably even
483	<	* when interpreted (not compiled).
484	<	*
485	<	* The order of declarations in this file is: (1) declarations of
486	<	* statics (2) fields (along with constants used when unpacking
487	<	* some of them), listed in an order that tends to reduce
488	<	* contention among them a bit under most JVMs.  (3) internal
489	<	* control methods (4) callbacks and other support for
490	<	* ForkJoinTask and ForkJoinWorkerThread classes, (5) exported
491	<	* methods (plus a few little helpers). (6) static block
333	<	* initializing all statics in a minimally dependent order.
467	>	* changes anyway. Several methods intrinsically sprawl because
468	>	* they must accumulate sets of consistent reads of volatiles held
469	>	* in local variables.  Methods signalWork() and scan() are the
470	>	* main bottlenecks, so are especially heavily
471	>	* micro-optimized/mangled.  There are lots of inline assignments
472	>	* (of form "while ((local = field) != 0)") which are usually the
473	>	* simplest way to ensure the required read orderings (which are
474	>	* sometimes critical). This leads to a "C"-like style of listing
475	>	* declarations of these locals at the heads of methods or blocks.
476	>	* There are several occurrences of the unusual "do {} while
477	>	* (!cas...)"  which is the simplest way to force an update of a
478	>	* CAS'ed variable. There are also other coding oddities (including
479	>	* several unnecessary-looking hoisted null checks) that help
480	>	* some methods perform reasonably even when interpreted (not
481	>	* compiled).
482	>	*
483	>	* The order of declarations in this file is:
484	>	* (1) Static utility functions
485	>	* (2) Nested (static) classes
486	>	* (3) Static fields
487	>	* (4) Fields, along with constants used when unpacking some of them
488	>	* (5) Internal control methods
489	>	* (6) Callbacks and other support for ForkJoinTask methods
490	>	* (7) Exported methods
491	>	* (8) Static block initializing statics in minimally dependent order
492		*/
493
494	+	// Static utilities
495	+
496	+	/**
497	+	* If there is a security manager, makes sure caller has
498	+	* permission to modify threads.
499	+	*/
500	+	private static void checkPermission() {
501	+	SecurityManager security = System.getSecurityManager();
502	+	if (security != null)
503	+	security.checkPermission(modifyThreadPermission);
504	+	}
505	+
506	+	// Nested classes
507	+
508		/**
509		* Factory for creating new {@link ForkJoinWorkerThread}s.
510		* A {@code ForkJoinWorkerThreadFactory} must be defined and used
#	Line 361 | Line 533 | public class ForkJoinPool extends Abstra
533		}
534
535		/**
536	<	* Creates a new ForkJoinWorkerThread. This factory is used unless
537	<	* overridden in ForkJoinPool constructors.
536	>	* Class for artificial tasks that are used to replace the target
537	>	* of local joins if they are removed from an interior queue slot
538	>	* in WorkQueue.tryRemoveAndExec. We don't need the proxy to
539	>	* actually do anything beyond having a unique identity.
540	>	*/
541	>	static final class EmptyTask extends ForkJoinTask<Void> {
542	>	private static final long serialVersionUID = -7721805057305804111L;
543	>	EmptyTask() { status = ForkJoinTask.NORMAL; } // force done
544	>	public final Void getRawResult() { return null; }
545	>	public final void setRawResult(Void x) {}
546	>	public final boolean exec() { return true; }
547	>	}
548	>
549	>	/**
550	>	* Queues supporting work-stealing as well as external task
551	>	* submission. See above for main rationale and algorithms.
552	>	* Implementation relies heavily on "Unsafe" intrinsics
553	>	* and selective use of "volatile":
554	>	*
555	>	* Field "base" is the index (mod array.length) of the least valid
556	>	* queue slot, which is always the next position to steal (poll)
557	>	* from if nonempty. Reads and writes require volatile orderings
558	>	* but not CAS, because updates are only performed after slot
559	>	* CASes.
560	>	*
561	>	* Field "top" is the index (mod array.length) of the next queue
562	>	* slot to push to or pop from. It is written only by owner thread
563	>	* for push, or under lock for external/shared push, and accessed
564	>	* by other threads only after reading (volatile) base.  Both top
565	>	* and base are allowed to wrap around on overflow, but (top -
566	>	* base) (or more commonly -(base - top) to force volatile read of
567	>	* base before top) still estimates size. The lock ("qlock") is
568	>	* forced to -1 on termination, causing all further lock attempts
569	>	* to fail. (Note: we don't need CAS for termination state because
570	>	* upon pool shutdown, all shared-queues will stop being used
571	>	* anyway.)  Nearly all lock bodies are set up so that exceptions
572	>	* within lock bodies are "impossible" (modulo JVM errors that
573	>	* would cause failure anyway.)
574	>	*
575	>	* The array slots are read and written using the emulation of
576	>	* volatiles/atomics provided by Unsafe. Insertions must in
577	>	* general use putOrderedObject as a form of releasing store to
578	>	* ensure that all writes to the task object are ordered before
579	>	* its publication in the queue.  All removals entail a CAS to
580	>	* null.  The array is always a power of two. To ensure safety of
581	>	* Unsafe array operations, all accesses perform explicit null
582	>	* checks and implicit bounds checks via power-of-two masking.
583	>	*
584	>	* In addition to basic queuing support, this class contains
585	>	* fields described elsewhere to control execution. It turns out
586	>	* to work better memory-layout-wise to include them in this class
587	>	* rather than a separate class.
588	>	*
589	>	* Performance on most platforms is very sensitive to placement of
590	>	* instances of both WorkQueues and their arrays -- we absolutely
591	>	* do not want multiple WorkQueue instances or multiple queue
592	>	* arrays sharing cache lines. (It would be best for queue objects
593	>	* and their arrays to share, but there is nothing available to
594	>	* help arrange that).  Unfortunately, because they are recorded
595	>	* in a common array, WorkQueue instances are often moved to be
596	>	* adjacent by garbage collectors. To reduce impact, we use field
597	>	* padding that works OK on common platforms; this effectively
598	>	* trades off slightly slower average field access for the sake of
599	>	* avoiding really bad worst-case access. (Until better JVM
600	>	* support is in place, this padding is dependent on transient
601	>	* properties of JVM field layout rules.)  We also take care in
602	>	* allocating, sizing and resizing the array. Non-shared queue
603	>	* arrays are initialized by workers before use. Others are
604	>	* allocated on first use.
605		*/
606	<	public static final ForkJoinWorkerThreadFactory
607	<	defaultForkJoinWorkerThreadFactory;
606	>	static final class WorkQueue {
607	>	/**
608	>	* Capacity of work-stealing queue array upon initialization.
609	>	* Must be a power of two; at least 4, but should be larger to
610	>	* reduce or eliminate cacheline sharing among queues.
611	>	* Currently, it is much larger, as a partial workaround for
612	>	* the fact that JVMs often place arrays in locations that
613	>	* share GC bookkeeping (especially cardmarks) such that
614	>	* per-write accesses encounter serious memory contention.
615	>	*/
616	>	static final int INITIAL_QUEUE_CAPACITY = 1 << 13;
617
618	<	/**
619	<	* Permission required for callers of methods that may start or
620	<	* kill threads.
621	<	*/
622	<	private static final RuntimePermission modifyThreadPermission;
618	>	/**
619	>	* Maximum size for queue arrays. Must be a power of two less
620	>	* than or equal to 1 << (31 - width of array entry) to ensure
621	>	* lack of wraparound of index calculations, but defined to a
622	>	* value a bit less than this to help users trap runaway
623	>	* programs before saturating systems.
624	>	*/
625	>	static final int MAXIMUM_QUEUE_CAPACITY = 1 << 26; // 64M
626
627	<	/**
628	<	* If there is a security manager, makes sure caller has
629	<	* permission to modify threads.
630	<	*/
631	<	private static void checkPermission() {
632	<	SecurityManager security = System.getSecurityManager();
633	<	if (security != null)
634	<	security.checkPermission(modifyThreadPermission);
627	>	int seed;                  // for random scanning; initialize nonzero
628	>	volatile int eventCount;   // encoded inactivation count; < 0 if inactive
629	>	int nextWait;              // encoded record of next event waiter
630	>	final int mode;            // lifo, fifo, or shared
631	>	int nsteals;               // cumulative number of steals
632	>	int poolIndex;             // index of this queue in pool (or 0)
633	>	int stealHint;             // index of most recent known stealer
634	>	volatile int qlock;        // 1: locked, -1: terminate; else 0
635	>	volatile int base;         // index of next slot for poll
636	>	int top;                   // index of next slot for push
637	>	ForkJoinTask<?>[] array;   // the elements (initially unallocated)
638	>	final ForkJoinPool pool;   // the containing pool (may be null)
639	>	final ForkJoinWorkerThread owner; // owning thread or null if shared
640	>	volatile Thread parker;    // == owner during call to park; else null
641	>	volatile ForkJoinTask<?> currentJoin;  // task being joined in awaitJoin
642	>	ForkJoinTask<?> currentSteal; // current non-local task being executed
643	>	// Heuristic padding to ameliorate unfortunate memory placements
644	>	Object p00, p01, p02, p03, p04, p05, p06, p07;
645	>	Object p08, p09, p0a, p0b, p0c, p0d, p0e;
646	>
647	>	WorkQueue(ForkJoinPool pool, ForkJoinWorkerThread owner, int mode) {
648	>	this.mode = mode;
649	>	this.pool = pool;
650	>	this.owner = owner;
651	>	// Place indices in the center of array (that is not yet allocated)
652	>	base = top = INITIAL_QUEUE_CAPACITY >>> 1;
653	>	}
654	>
655	>	/**
656	>	* Pushes a task. Call only by owner in unshared queues.
657	>	* Cases needing resizing or rejection are relayed to fullPush
658	>	* (that also handles shared queues).
659	>	*
660	>	* @param task the task. Caller must ensure non-null.
661	>	* @throw RejectedExecutionException if array cannot be resized
662	>	*/
663	>	final void push(ForkJoinTask<?> task) {
664	>	ForkJoinPool p; ForkJoinTask<?>[] a;
665	>	int s = top, n;
666	>	if ((a = array) != null && a.length > (n = s + 1 - base)) {
667	>	U.putOrderedObject
668	>	(a, (((a.length - 1) & s) << ASHIFT) + ABASE, task);
669	>	top = s + 1;
670	>	if (n <= 1 && (p = pool) != null)
671	>	p.signalWork(this, 1);
672	>	}
673	>	else
674	>	fullPush(task, true);
675	>	}
676	>
677	>	/**
678	>	* Pushes a task if lock is free and array is either big
679	>	* enough or can be resized to be big enough. Note: a
680	>	* specialization of a common fast path of this method is in
681	>	* ForkJoinPool.externalPush. When called from a FJWT queue,
682	>	* this can fail only if the pool has been shut down or
683	>	* an out of memory error.
684	>	*
685	>	* @param task the task. Caller must ensure non-null.
686	>	* @param owned if true, throw RJE on failure
687	>	*/
688	>	final boolean fullPush(ForkJoinTask<?> task, boolean owned) {
689	>	ForkJoinPool p; ForkJoinTask<?>[] a;
690	>	if (owned) {
691	>	if (qlock < 0) // must be shutting down
692	>	throw new RejectedExecutionException();
693	>	}
694	>	else if (!U.compareAndSwapInt(this, QLOCK, 0, 1))
695	>	return false;
696	>	try {
697	>	int s = top, oldLen, len;
698	>	if ((a = array) == null)
699	>	a = array = new ForkJoinTask<?>[len=INITIAL_QUEUE_CAPACITY];
700	>	else if ((oldLen = a.length) > s + 1 - base)
701	>	len = oldLen;
702	>	else if ((len = oldLen << 1) > MAXIMUM_QUEUE_CAPACITY)
703	>	throw new RejectedExecutionException("Capacity exceeded");
704	>	else {
705	>	int oldMask, b;
706	>	ForkJoinTask<?>[] oldA = a;
707	>	a = array = new ForkJoinTask<?>[len];
708	>	if ((oldMask = oldLen - 1) >= 0 && s - (b = base) > 0) {
709	>	int mask = len - 1;
710	>	do {
711	>	ForkJoinTask<?> x;
712	>	int oldj = ((b & oldMask) << ASHIFT) + ABASE;
713	>	int j    = ((b &    mask) << ASHIFT) + ABASE;
714	>	x = (ForkJoinTask<?>)
715	>	U.getObjectVolatile(oldA, oldj);
716	>	if (x != null &&
717	>	U.compareAndSwapObject(oldA, oldj, x, null))
718	>	U.putObjectVolatile(a, j, x);
719	>	} while (++b != s);
720	>	}
721	>	}
722	>	U.putOrderedObject
723	>	(a, (((len - 1) & s) << ASHIFT) + ABASE, task);
724	>	top = s + 1;
725	>	} finally {
726	>	if (!owned)
727	>	qlock = 0;
728	>	}
729	>	if ((p = pool) != null)
730	>	p.signalWork(this, 1);
731	>	return true;
732	>	}
733	>
734	>	/**
735	>	* Takes next task, if one exists, in LIFO order.  Call only
736	>	* by owner in unshared queues.
737	>	*/
738	>	final ForkJoinTask<?> pop() {
739	>	ForkJoinTask<?>[] a; ForkJoinTask<?> t; int m;
740	>	if ((a = array) != null && (m = a.length - 1) >= 0) {
741	>	for (int s; (s = top - 1) - base >= 0;) {
742	>	long j = ((m & s) << ASHIFT) + ABASE;
743	>	if ((t = (ForkJoinTask<?>)U.getObject(a, j)) == null)
744	>	break;
745	>	if (U.compareAndSwapObject(a, j, t, null)) {
746	>	top = s;
747	>	return t;
748	>	}
749	>	}
750	>	}
751	>	return null;
752	>	}
753	>
754	>	/**
755	>	* Takes a task in FIFO order if b is base of queue and a task
756	>	* can be claimed without contention. Specialized versions
757	>	* appear in ForkJoinPool methods scan and tryHelpStealer.
758	>	*/
759	>	final ForkJoinTask<?> pollAt(int b) {
760	>	ForkJoinTask<?> t; ForkJoinTask<?>[] a;
761	>	if ((a = array) != null) {
762	>	int j = (((a.length - 1) & b) << ASHIFT) + ABASE;
763	>	if ((t = (ForkJoinTask<?>)U.getObjectVolatile(a, j)) != null &&
764	>	base == b &&
765	>	U.compareAndSwapObject(a, j, t, null)) {
766	>	base = b + 1;
767	>	return t;
768	>	}
769	>	}
770	>	return null;
771	>	}
772	>
773	>	/**
774	>	* Takes next task, if one exists, in FIFO order.
775	>	*/
776	>	final ForkJoinTask<?> poll() {
777	>	ForkJoinTask<?>[] a; int b; ForkJoinTask<?> t;
778	>	while ((b = base) - top < 0 && (a = array) != null) {
779	>	int j = (((a.length - 1) & b) << ASHIFT) + ABASE;
780	>	t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
781	>	if (t != null) {
782	>	if (base == b &&
783	>	U.compareAndSwapObject(a, j, t, null)) {
784	>	base = b + 1;
785	>	return t;
786	>	}
787	>	}
788	>	else if (base == b) {
789	>	if (b + 1 == top)
790	>	break;
791	>	Thread.yield(); // wait for lagging update (very rare)
792	>	}
793	>	}
794	>	return null;
795	>	}
796	>
797	>	/**
798	>	* Takes next task, if one exists, in order specified by mode.
799	>	*/
800	>	final ForkJoinTask<?> nextLocalTask() {
801	>	return mode == 0 ? pop() : poll();
802	>	}
803	>
804	>	/**
805	>	* Returns next task, if one exists, in order specified by mode.
806	>	*/
807	>	final ForkJoinTask<?> peek() {
808	>	ForkJoinTask<?>[] a = array; int m;
809	>	if (a == null || (m = a.length - 1) < 0)
810	>	return null;
811	>	int i = mode == 0 ? top - 1 : base;
812	>	int j = ((i & m) << ASHIFT) + ABASE;
813	>	return (ForkJoinTask<?>)U.getObjectVolatile(a, j);
814	>	}
815	>
816	>	/**
817	>	* Pops the given task only if it is at the current top.
818	>	* (A shared version is available only via FJP.tryExternalUnpush)
819	>	*/
820	>	final boolean tryUnpush(ForkJoinTask<?> t) {
821	>	ForkJoinTask<?>[] a; int s;
822	>	if ((a = array) != null && (s = top) != base &&
823	>	U.compareAndSwapObject
824	>	(a, (((a.length - 1) & --s) << ASHIFT) + ABASE, t, null)) {
825	>	top = s;
826	>	return true;
827	>	}
828	>	return false;
829	>	}
830	>
831	>	/**
832	>	* Removes and cancels all known tasks, ignoring any exceptions.
833	>	*/
834	>	final void cancelAll() {
835	>	ForkJoinTask.cancelIgnoringExceptions(currentJoin);
836	>	ForkJoinTask.cancelIgnoringExceptions(currentSteal);
837	>	for (ForkJoinTask<?> t; (t = poll()) != null; )
838	>	ForkJoinTask.cancelIgnoringExceptions(t);
839	>	}
840	>
841	>	/**
842	>	* Computes next value for random probes.  Scans don't require
843	>	* a very high quality generator, but also not a crummy one.
844	>	* Marsaglia xor-shift is cheap and works well enough.  Note:
845	>	* This is manually inlined in its usages in ForkJoinPool to
846	>	* avoid writes inside busy scan loops.
847	>	*/
848	>	final int nextSeed() {
849	>	int r = seed;
850	>	r ^= r << 13;
851	>	r ^= r >>> 17;
852	>	return seed = r ^= r << 5;
853	>	}
854	>
855	>	/**
856	>	* Provides a more accurate estimate of size than (top - base)
857	>	* by ordering reads and checking whether a near-empty queue
858	>	* has at least one unclaimed task.
859	>	*/
860	>	final int queueSize() {
861	>	ForkJoinTask<?>[] a; int k, s, n;
862	>	return ((n = base - (s = top)) < 0 &&
863	>	(n != -1 ||
864	>	((a = array) != null && (k = a.length) > 0 &&
865	>	U.getObject
866	>	(a, (long)((((k - 1) & (s - 1)) << ASHIFT) + ABASE)) != null))) ?
867	>	-n : 0;
868	>	}
869	>
870	>	// Specialized execution methods
871	>
872	>	/**
873	>	* Pops and runs tasks until empty.
874	>	*/
875	>	private void popAndExecAll() {
876	>	// A bit faster than repeated pop calls
877	>	ForkJoinTask<?>[] a; int m, s; long j; ForkJoinTask<?> t;
878	>	while ((a = array) != null && (m = a.length - 1) >= 0 &&
879	>	(s = top - 1) - base >= 0 &&
880	>	(t = ((ForkJoinTask<?>)
881	>	U.getObject(a, j = ((m & s) << ASHIFT) + ABASE)))
882	>	!= null) {
883	>	if (U.compareAndSwapObject(a, j, t, null)) {
884	>	top = s;
885	>	t.doExec();
886	>	}
887	>	}
888	>	}
889	>
890	>	/**
891	>	* Polls and runs tasks until empty.
892	>	*/
893	>	private void pollAndExecAll() {
894	>	for (ForkJoinTask<?> t; (t = poll()) != null;)
895	>	t.doExec();
896	>	}
897	>
898	>	/**
899	>	* If present, removes from queue and executes the given task,
900	>	* or any other cancelled task. Returns (true) on any CAS
901	>	* or consistency check failure so caller can retry.
902	>	*
903	>	* @return false if no progress can be made, else true;
904	>	*/
905	>	final boolean tryRemoveAndExec(ForkJoinTask<?> task) {
906	>	boolean stat = true, removed = false, empty = true;
907	>	ForkJoinTask<?>[] a; int m, s, b, n;
908	>	if ((a = array) != null && (m = a.length - 1) >= 0 &&
909	>	(n = (s = top) - (b = base)) > 0) {
910	>	for (ForkJoinTask<?> t;;) {           // traverse from s to b
911	>	int j = ((--s & m) << ASHIFT) + ABASE;
912	>	t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
913	>	if (t == null)                    // inconsistent length
914	>	break;
915	>	else if (t == task) {
916	>	if (s + 1 == top) {           // pop
917	>	if (!U.compareAndSwapObject(a, j, task, null))
918	>	break;
919	>	top = s;
920	>	removed = true;
921	>	}
922	>	else if (base == b)           // replace with proxy
923	>	removed = U.compareAndSwapObject(a, j, task,
924	>	new EmptyTask());
925	>	break;
926	>	}
927	>	else if (t.status >= 0)
928	>	empty = false;
929	>	else if (s + 1 == top) {          // pop and throw away
930	>	if (U.compareAndSwapObject(a, j, t, null))
931	>	top = s;
932	>	break;
933	>	}
934	>	if (--n == 0) {
935	>	if (!empty && base == b)
936	>	stat = false;
937	>	break;
938	>	}
939	>	}
940	>	}
941	>	if (removed)
942	>	task.doExec();
943	>	return stat;
944	>	}
945	>
946	>	/**
947	>	* Polls for and executes the given task or any other task in
948	>	* its CountedCompleter computation
949	>	*/
950	>	final boolean pollAndExecCC(ForkJoinTask<?> root) {
951	>	ForkJoinTask<?>[] a; int b; Object o;
952	>	outer: while ((b = base) - top < 0 && (a = array) != null) {
953	>	long j = (((a.length - 1) & b) << ASHIFT) + ABASE;
954	>	if ((o = U.getObject(a, j)) == null ||
955	>	!(o instanceof CountedCompleter))
956	>	break;
957	>	for (CountedCompleter<?> t = (CountedCompleter<?>)o, r = t;;) {
958	>	if (r == root) {
959	>	if (base == b &&
960	>	U.compareAndSwapObject(a, j, t, null)) {
961	>	base = b + 1;
962	>	t.doExec();
963	>	return true;
964	>	}
965	>	else
966	>	break; // restart
967	>	}
968	>	if ((r = r.completer) == null)
969	>	break outer; // not part of root computation
970	>	}
971	>	}
972	>	return false;
973	>	}
974	>
975	>	/**
976	>	* Executes a top-level task and any local tasks remaining
977	>	* after execution.
978	>	*/
979	>	final void runTask(ForkJoinTask<?> t) {
980	>	if (t != null) {
981	>	(currentSteal = t).doExec();
982	>	currentSteal = null;
983	>	if (++nsteals < 0) {     // spill on overflow
984	>	ForkJoinPool p;
985	>	if ((p = pool) != null)
986	>	p.collectStealCount(this);
987	>	}
988	>	if (top != base) {       // process remaining local tasks
989	>	if (mode == 0)
990	>	popAndExecAll();
991	>	else
992	>	pollAndExecAll();
993	>	}
994	>	}
995	>	}
996	>
997	>	/**
998	>	* Executes a non-top-level (stolen) task.
999	>	*/
1000	>	final void runSubtask(ForkJoinTask<?> t) {
1001	>	if (t != null) {
1002	>	ForkJoinTask<?> ps = currentSteal;
1003	>	(currentSteal = t).doExec();
1004	>	currentSteal = ps;
1005	>	}
1006	>	}
1007	>
1008	>	/**
1009	>	* Returns true if owned and not known to be blocked.
1010	>	*/
1011	>	final boolean isApparentlyUnblocked() {
1012	>	Thread wt; Thread.State s;
1013	>	return (eventCount >= 0 &&
1014	>	(wt = owner) != null &&
1015	>	(s = wt.getState()) != Thread.State.BLOCKED &&
1016	>	s != Thread.State.WAITING &&
1017	>	s != Thread.State.TIMED_WAITING);
1018	>	}
1019	>
1020	>	/**
1021	>	* If this owned and is not already interrupted, try to
1022	>	* interrupt and/or unpark, ignoring exceptions.
1023	>	*/
1024	>	final void interruptOwner() {
1025	>	Thread wt, p;
1026	>	if ((wt = owner) != null && !wt.isInterrupted()) {
1027	>	try {
1028	>	wt.interrupt();
1029	>	} catch (SecurityException ignore) {
1030	>	}
1031	>	}
1032	>	if ((p = parker) != null)
1033	>	U.unpark(p);
1034	>	}
1035	>
1036	>	// Unsafe mechanics
1037	>	private static final sun.misc.Unsafe U;
1038	>	private static final long QLOCK;
1039	>	private static final int ABASE;
1040	>	private static final int ASHIFT;
1041	>	static {
1042	>	int s;
1043	>	try {
1044	>	U = getUnsafe();
1045	>	Class<?> k = WorkQueue.class;
1046	>	Class<?> ak = ForkJoinTask[].class;
1047	>	QLOCK = U.objectFieldOffset
1048	>	(k.getDeclaredField("qlock"));
1049	>	ABASE = U.arrayBaseOffset(ak);
1050	>	s = U.arrayIndexScale(ak);
1051	>	} catch (Exception e) {
1052	>	throw new Error(e);
1053	>	}
1054	>	if ((s & (s-1)) != 0)
1055	>	throw new Error("data type scale not a power of two");
1056	>	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
1057	>	}
1058		}
1059
1060		/**
1061	<	* Generator for assigning sequence numbers as pool names.
1062	<	*/
1063	<	private static final AtomicInteger poolNumberGenerator;
1061	>	* Per-thread records for threads that submit to pools. Currently
1062	>	* holds only pseudo-random seed / index that is used to choose
1063	>	* submission queues in method externalPush. In the future, this may
1064	>	* also incorporate a means to implement different task rejection
1065	>	* and resubmission policies.
1066	>	*
1067	>	* Seeds for submitters and workers/workQueues work in basically
1068	>	* the same way but are initialized and updated using slightly
1069	>	* different mechanics. Both are initialized using the same
1070	>	* approach as in class ThreadLocal, where successive values are
1071	>	* unlikely to collide with previous values. Seeds are then
1072	>	* randomly modified upon collisions using xorshifts, which
1073	>	* requires a non-zero seed.
1074	>	*/
1075	>	static final class Submitter {
1076	>	int seed;
1077	>	Submitter(int s) { seed = s; }
1078	>	}
1079	>
1080	>	/** Property prefix for constructing common pool */
1081	>	private static final String propPrefix =
1082	>	"java.util.concurrent.ForkJoinPool.common.";
1083	>
1084	>	// static fields (initialized in static initializer below)
1085
1086		/**
1087	<	* Generator for initial random seeds for worker victim
1088	<	* selection. This is used only to create initial seeds. Random
394	<	* steals use a cheaper xorshift generator per steal attempt. We
395	<	* don't expect much contention on seedGenerator, so just use a
396	<	* plain Random.
1087	>	* Creates a new ForkJoinWorkerThread. This factory is used unless
1088	>	* overridden in ForkJoinPool constructors.
1089		*/
1090	<	static final Random workerSeedGenerator;
1090	>	public static final ForkJoinWorkerThreadFactory
1091	>	defaultForkJoinWorkerThreadFactory;
1092
1093		/**
1094	<	* Array holding all worker threads in the pool.  Initialized upon
1095	<	* construction. Array size must be a power of two.  Updates and
1096	<	* replacements are protected by scanGuard, but the array is
1097	<	* always kept in a consistent enough state to be randomly
405	<	* accessed without locking by workers performing work-stealing,
406	<	* as well as other traversal-based methods in this class, so long
407	<	* as reads memory-acquire by first reading ctl. All readers must
408	<	* tolerate that some array slots may be null.
1094	>	* Common (static) pool. Non-null for public use unless a static
1095	>	* construction exception, but internal usages null-check on use
1096	>	* to paranoically avoid potential initialization circularities
1097	>	* as well as to simplify generated code.
1098		*/
1099	<	ForkJoinWorkerThread[] workers;
1099	>	static final ForkJoinPool commonPool;
1100
1101		/**
1102	<	* Initial size for submission queue array. Must be a power of
1103	<	* two.  In many applications, these always stay small so we use a
415	<	* small initial cap.
1102	>	* Permission required for callers of methods that may start or
1103	>	* kill threads.
1104		*/
1105	<	private static final int INITIAL_QUEUE_CAPACITY = 8;
1105	>	private static final RuntimePermission modifyThreadPermission;
1106
1107		/**
1108	<	* Maximum size for submission queue array. Must be a power of two
1109	<	* less than or equal to 1 << (31 - width of array entry) to
1110	<	* ensure lack of index wraparound, but is capped at a lower
1111	<	* value to help users trap runaway computations.
1108	>	* Per-thread submission bookkeeping. Shared across all pools
1109	>	* to reduce ThreadLocal pollution and because random motion
1110	>	* to avoid contention in one pool is likely to hold for others.
1111	>	* Lazily initialized on first submission (but null-checked
1112	>	* in other contexts to avoid unnecessary initialization).
1113		*/
1114	<	private static final int MAXIMUM_QUEUE_CAPACITY = 1 << 24; // 16M
1114	>	static final ThreadLocal<Submitter> submitters;
1115
1116		/**
1117	<	* Array serving as submission queue. Initialized upon construction.
1117	>	* Common pool parallelism. Must equal commonPool.parallelism.
1118		*/
1119	<	private ForkJoinTask<?>[] submissionQueue;
1119	>	static final int commonPoolParallelism;
1120
1121		/**
1122	<	* Lock protecting submissions array for addSubmission
1122	>	* Sequence number for creating workerNamePrefix.
1123		*/
1124	<	private final ReentrantLock submissionLock;
1124	>	private static int poolNumberSequence;
1125
1126		/**
1127	<	* Condition for awaitTermination, using submissionLock for
1128	<	* convenience.
1127	>	* Return the next sequence number. We don't expect this to
1128	>	* ever contend so use simple builtin sync.
1129		*/
1130	<	private final Condition termination;
1130	>	private static final synchronized int nextPoolId() {
1131	>	return ++poolNumberSequence;
1132	>	}
1133	>
1134	>	// static constants
1135
1136		/**
1137	<	* Creation factory for worker threads.
1137	>	* Initial timeout value (in nanoseconds) for the thread
1138	>	* triggering quiescence to park waiting for new work. On timeout,
1139	>	* the thread will instead try to shrink the number of
1140	>	* workers. The value should be large enough to avoid overly
1141	>	* aggressive shrinkage during most transient stalls (long GCs
1142	>	* etc).
1143		*/
1144	<	private final ForkJoinWorkerThreadFactory factory;
1144	>	private static final long IDLE_TIMEOUT      = 2000L * 1000L * 1000L; // 2sec
1145
1146		/**
1147	<	* The uncaught exception handler used when any worker abruptly
450	<	* terminates.
1147	>	* Timeout value when there are more threads than parallelism level
1148		*/
1149	<	final Thread.UncaughtExceptionHandler ueh;
1149	>	private static final long FAST_IDLE_TIMEOUT =  200L * 1000L * 1000L;
1150
1151		/**
1152	<	* Prefix for assigning names to worker threads
1152	>	* The maximum stolen->joining link depth allowed in method
1153	>	* tryHelpStealer.  Must be a power of two.  Depths for legitimate
1154	>	* chains are unbounded, but we use a fixed constant to avoid
1155	>	* (otherwise unchecked) cycles and to bound staleness of
1156	>	* traversal parameters at the expense of sometimes blocking when
1157	>	* we could be helping.
1158		*/
1159	<	private final String workerNamePrefix;
1159	>	private static final int MAX_HELP = 64;
1160
1161		/**
1162	<	* Sum of per-thread steal counts, updated only when threads are
1163	<	* idle or terminating.
1162	>	* Increment for seed generators. See class ThreadLocal for
1163	>	* explanation.
1164		*/
1165	<	private volatile long stealCount;
1165	>	private static final int SEED_INCREMENT = 0x61c88647;
1166
1167		/**
1168	<	* Main pool control -- a long packed with:
1168	>	* Bits and masks for control variables
1169	>	*
1170	>	* Field ctl is a long packed with:
1171		* AC: Number of active running workers minus target parallelism (16 bits)
1172	<	* TC: Number of total workers minus target parallelism (16bits)
1172	>	* TC: Number of total workers minus target parallelism (16 bits)
1173		* ST: true if pool is terminating (1 bit)
1174		* EC: the wait count of top waiting thread (15 bits)
1175	<	* ID: ~poolIndex of top of Treiber stack of waiting threads (16 bits)
1175	>	* ID: poolIndex of top of Treiber stack of waiters (16 bits)
1176		*
1177		* When convenient, we can extract the upper 32 bits of counts and
1178		* the lower 32 bits of queue state, u = (int)(ctl >>> 32) and e =
#	Line 477 | Line 1181 | public class ForkJoinPool extends Abstra
1181		* parallelism and the positionings of fields makes it possible to
1182		* perform the most common checks via sign tests of fields: When
1183		* ac is negative, there are not enough active workers, when tc is
1184	<	* negative, there are not enough total workers, when id is
481	<	* negative, there is at least one waiting worker, and when e is
1184	>	* negative, there are not enough total workers, and when e is
1185		* negative, the pool is terminating.  To deal with these possibly
1186		* negative fields, we use casts in and out of "short" and/or
1187		* signed shifts to maintain signedness.
1188	+	*
1189	+	* When a thread is queued (inactivated), its eventCount field is
1190	+	* set negative, which is the only way to tell if a worker is
1191	+	* prevented from executing tasks, even though it must continue to
1192	+	* scan for them to avoid queuing races. Note however that
1193	+	* eventCount updates lag releases so usage requires care.
1194	+	*
1195	+	* Field plock is an int packed with:
1196	+	* SHUTDOWN: true if shutdown is enabled (1 bit)
1197	+	* SEQ:  a sequence lock, with PL_LOCK bit set if locked (30 bits)
1198	+	* SIGNAL: set when threads may be waiting on the lock (1 bit)
1199	+	*
1200	+	* The sequence number enables simple consistency checks:
1201	+	* Staleness of read-only operations on the workQueues array can
1202	+	* be checked by comparing plock before vs after the reads.
1203		*/
486	–	volatile long ctl;
1204
1205		// bit positions/shifts for fields
1206		private static final int  AC_SHIFT   = 48;
#	Line 492 | Line 1209 | public class ForkJoinPool extends Abstra
1209		private static final int  EC_SHIFT   = 16;
1210
1211		// bounds
1212	<	private static final int  MAX_ID     = 0x7fff;  // max poolIndex
1213	<	private static final int  SMASK      = 0xffff;  // mask short bits
1212	>	private static final int  SMASK      = 0xffff;  // short bits
1213	>	private static final int  MAX_CAP    = 0x7fff;  // max #workers - 1
1214	>	private static final int  EVENMASK   = 0xfffe;  // even short bits
1215	>	private static final int  SQMASK     = 0x007e;  // max 64 (even) slots
1216		private static final int  SHORT_SIGN = 1 << 15;
1217		private static final int  INT_SIGN   = 1 << 31;
1218
#	Line 515 | Line 1234 | public class ForkJoinPool extends Abstra
1234		private static final int  UTC_UNIT   = 1 << UTC_SHIFT;
1235
1236		// masks and units for dealing with e = (int)ctl
1237	<	private static final int  E_MASK     = 0x7fffffff; // no STOP_BIT
1238	<	private static final int  EC_UNIT    = 1 << EC_SHIFT;
520	<
521	<	/**
522	<	* The target parallelism level.
523	<	*/
524	<	final int parallelism;
525	<
526	<	/**
527	<	* Index (mod submission queue length) of next element to take
528	<	* from submission queue. Usage is identical to that for
529	<	* per-worker queues -- see ForkJoinWorkerThread internal
530	<	* documentation.
531	<	*/
532	<	volatile int queueBase;
1237	>	private static final int E_MASK      = 0x7fffffff; // no STOP_BIT
1238	>	private static final int E_SEQ       = 1 << EC_SHIFT;
1239
1240	<	/**
1241	<	* Index (mod submission queue length) of next element to add
1242	<	* in submission queue. Usage is identical to that for
1243	<	* per-worker queues -- see ForkJoinWorkerThread internal
1244	<	* documentation.
1245	<	*/
1246	<	int queueTop;
1247	<
1248	<	/**
1249	<	* True when shutdown() has been called.
544	<	*/
545	<	volatile boolean shutdown;
546	<
547	<	/**
548	<	* True if use local fifo, not default lifo, for local polling
549	<	* Read by, and replicated by ForkJoinWorkerThreads
550	<	*/
551	<	final boolean locallyFifo;
1240	>	// plock bits
1241	>	private static final int SHUTDOWN    = 1 << 31;
1242	>	private static final int PL_LOCK     = 2;
1243	>	private static final int PL_SIGNAL   = 1;
1244	>	private static final int PL_SPINS    = 1 << 8;
1245	>
1246	>	// access mode for WorkQueue
1247	>	static final int LIFO_QUEUE          =  0;
1248	>	static final int FIFO_QUEUE          =  1;
1249	>	static final int SHARED_QUEUE        = -1;
1250
1251	<	/**
554	<	* The number of threads in ForkJoinWorkerThreads.helpQuiescePool.
555	<	* When non-zero, suppresses automatic shutdown when active
556	<	* counts become zero.
557	<	*/
558	<	volatile int quiescerCount;
1251	>	// Instance fields
1252
1253	<	/**
1254	<	* The number of threads blocked in join.
1255	<	*/
1256	<	volatile int blockedCount;
1253	>	/*
1254	>	* Field layout order in this class tends to matter more than one
1255	>	* would like. Runtime layout order is only loosely related to
1256	>	* declaration order and may differ across JVMs, but the following
1257	>	* empirically works OK on current JVMs.
1258	>	*/
1259	>	volatile long stealCount;                  // collects worker counts
1260	>	volatile long ctl;                         // main pool control
1261	>	final int parallelism;                     // parallelism level
1262	>	final int localMode;                       // per-worker scheduling mode
1263	>	volatile int indexSeed;                    // worker/submitter index seed
1264	>	volatile int plock;                        // shutdown status and seqLock
1265	>	WorkQueue[] workQueues;                    // main registry
1266	>	final ForkJoinWorkerThreadFactory factory; // factory for new workers
1267	>	final Thread.UncaughtExceptionHandler ueh; // per-worker UEH
1268	>	final String workerNamePrefix;             // to create worker name string
1269
1270	<	/**
1271	<	* Counter for worker Thread names (unrelated to their poolIndex)
1272	<	*/
1273	<	private volatile int nextWorkerNumber;
1270	>	/*
1271	>	* Acquires the plock lock to protect worker array and related
1272	>	* updates. This method is called only if an initial CAS on plock
1273	>	* fails. This acts as a spinLock for normal cases, but falls back
1274	>	* to builtin monitor to block when (rarely) needed. This would be
1275	>	* a terrible idea for a highly contended lock, but works fine as
1276	>	* a more conservative alternative to a pure spinlock.  See
1277	>	* internal ConcurrentHashMap documentation for further
1278	>	* explanation of nearly the same construction.
1279	>	*/
1280	>	private int acquirePlock() {
1281	>	int spins = PL_SPINS, r = 0, ps, nps;
1282	>	for (;;) {
1283	>	if (((ps = plock) & PL_LOCK) == 0 &&
1284	>	U.compareAndSwapInt(this, PLOCK, ps, nps = ps + PL_LOCK))
1285	>	return nps;
1286	>	else if (r == 0)
1287	>	r = ThreadLocalRandom.current().nextInt(); // randomize spins
1288	>	else if (spins >= 0) {
1289	>	r ^= r << 1; r ^= r >>> 3; r ^= r << 10; // xorshift
1290	>	if (r >= 0)
1291	>	--spins;
1292	>	}
1293	>	else if (U.compareAndSwapInt(this, PLOCK, ps, ps | PL_SIGNAL)) {
1294	>	synchronized (this) {
1295	>	if ((plock & PL_SIGNAL) != 0) {
1296	>	try {
1297	>	wait();
1298	>	} catch (InterruptedException ie) {
1299	>	try {
1300	>	Thread.currentThread().interrupt();
1301	>	} catch (SecurityException ignore) {
1302	>	}
1303	>	}
1304	>	}
1305	>	else
1306	>	notifyAll();
1307	>	}
1308	>	}
1309	>	}
1310	>	}
1311
1312		/**
1313	<	* The index for the next created worker. Accessed under scanGuard.
1313	>	* Unlocks and signals any thread waiting for plock. Called only
1314	>	* when CAS of seq value for unlock fails.
1315		*/
1316	<	private int nextWorkerIndex;
1316	>	private void releasePlock(int ps) {
1317	>	plock = ps;
1318	>	synchronized (this) { notifyAll(); }
1319	>	}
1320	>
1321	>	//  Registering and deregistering workers
1322	>
1323	>	/**
1324	>	* Callback from ForkJoinWorkerThread constructor to establish its
1325	>	* poolIndex and record its WorkQueue. To avoid scanning bias due
1326	>	* to packing entries in front of the workQueues array, we treat
1327	>	* the array as a simple power-of-two hash table using per-thread
1328	>	* seed as hash, expanding as needed.
1329	>	*
1330	>	* @param w the worker's queue
1331	>	*/
1332	>	final void registerWorker(WorkQueue w) {
1333	>	int s, ps; // generate a rarely colliding candidate index seed
1334	>	do {} while (!U.compareAndSwapInt(this, INDEXSEED,
1335	>	s = indexSeed, s += SEED_INCREMENT) ||
1336	>	s == 0); // skip 0
1337	>	if (((ps = plock) & PL_LOCK) != 0 ||
1338	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1339	>	ps = acquirePlock();
1340	>	int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1341	>	try {
1342	>	WorkQueue[] ws;
1343	>	if (w != null && (ws = workQueues) != null) {
1344	>	w.seed = s;
1345	>	int n = ws.length, m = n - 1;
1346	>	int r = (s << 1) | 1;               // use odd-numbered indices
1347	>	if (ws[r &= m] != null) {           // collision
1348	>	int probes = 0;                 // step by approx half size
1349	>	int step = (n <= 4) ? 2 : ((n >>> 1) & EVENMASK) + 2;
1350	>	while (ws[r = (r + step) & m] != null) {
1351	>	if (++probes >= n) {
1352	>	workQueues = ws = Arrays.copyOf(ws, n <<= 1);
1353	>	m = n - 1;
1354	>	probes = 0;
1355	>	}
1356	>	}
1357	>	}
1358	>	w.eventCount = w.poolIndex = r;     // establish before recording
1359	>	ws[r] = w;
1360	>	}
1361	>	} finally {
1362	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1363	>	releasePlock(nps);
1364	>	}
1365	>	}
1366
1367		/**
1368	<	* SeqLock and index masking for updates to workers array.  Locked
1369	<	* when SG_UNIT is set. Unlocking clears bit by adding
1370	<	* SG_UNIT. Staleness of read-only operations can be checked by
1371	<	* comparing scanGuard to value before the reads. The low 16 bits
1372	<	* (i.e, anding with SMASK) hold (the smallest power of two
1373	<	* covering all worker indices, minus one, and is used to avoid
1374	<	* dealing with large numbers of null slots when the workers array
1375	<	* is overallocated.
1376	<	*/
1377	<	volatile int scanGuard;
1368	>	* Final callback from terminating worker, as well as upon failure
1369	>	* to construct or start a worker.  Removes record of worker from
1370	>	* array, and adjusts counts. If pool is shutting down, tries to
1371	>	* complete termination.
1372	>	*
1373	>	* @param wt the worker thread or null if construction failed
1374	>	* @param ex the exception causing failure, or null if none
1375	>	*/
1376	>	final void deregisterWorker(ForkJoinWorkerThread wt, Throwable ex) {
1377	>	WorkQueue w = null;
1378	>	if (wt != null && (w = wt.workQueue) != null) {
1379	>	int ps;
1380	>	collectStealCount(w);
1381	>	w.qlock = -1;                // ensure set
1382	>	if (((ps = plock) & PL_LOCK) != 0 ||
1383	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1384	>	ps = acquirePlock();
1385	>	int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1386	>	try {
1387	>	int idx = w.poolIndex;
1388	>	WorkQueue[] ws = workQueues;
1389	>	if (ws != null && idx >= 0 && idx < ws.length && ws[idx] == w)
1390	>	ws[idx] = null;
1391	>	} finally {
1392	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1393	>	releasePlock(nps);
1394	>	}
1395	>	}
1396
1397	<	private static final int SG_UNIT = 1 << 16;
1397	>	long c;                             // adjust ctl counts
1398	>	do {} while (!U.compareAndSwapLong
1399	>	(this, CTL, c = ctl, (((c - AC_UNIT) & AC_MASK) |
1400	>	((c - TC_UNIT) & TC_MASK) |
1401	>	(c & ~(AC_MASK|TC_MASK)))));
1402
1403	<	/**
1404	<	* The wakeup interval (in nanoseconds) for a worker waiting for a
1405	<	* task when the pool is quiescent to instead try to shrink the
1406	<	* number of workers.  The exact value does not matter too
1407	<	* much. It must be short enough to release resources during
1408	<	* sustained periods of idleness, but not so short that threads
1409	<	* are continually re-created.
1410	<	*/
1411	<	private static final long SHRINK_RATE =
1412	<	4L * 1000L * 1000L * 1000L; // 4 seconds
1403	>	if (!tryTerminate(false, false) && w != null) {
1404	>	w.cancelAll();                  // cancel remaining tasks
1405	>	if (w.array != null)            // suppress signal if never ran
1406	>	signalWork(null, 1);        // wake up or create replacement
1407	>	if (ex == null)                 // help clean refs on way out
1408	>	ForkJoinTask.helpExpungeStaleExceptions();
1409	>	}
1410	>
1411	>	if (ex != null)                     // rethrow
1412	>	ForkJoinTask.rethrow(ex);
1413	>	}
1414
1415		/**
1416	<	* Top-level loop for worker threads: On each step: if the
1417	<	* previous step swept through all queues and found no tasks, or
1418	<	* there are excess threads, then possibly blocks. Otherwise,
1419	<	* scans for and, if found, executes a task. Returns when pool
1420	<	* and/or worker terminate.
1421	<	*
1422	<	* @param w the worker
1423	<	*/
1424	<	final void work(ForkJoinWorkerThread w) {
1425	<	boolean swept = false;                // true on empty scans
1426	<	long c;
1427	<	while (!w.terminate && (int)(c = ctl) >= 0) {
1428	<	int a;                            // active count
614	<	if (!swept && (a = (int)(c >> AC_SHIFT)) <= 0)
615	<	swept = scan(w, a);
616	<	else if (tryAwaitWork(w, c))
617	<	swept = false;
1416	>	* Collect worker steal count into total. Called on termination
1417	>	* and upon int overflow of local count. (There is a possible race
1418	>	* in the latter case vs any caller of getStealCount, which can
1419	>	* make its results less accurate than usual.)
1420	>	*/
1421	>	final void collectStealCount(WorkQueue w) {
1422	>	if (w != null) {
1423	>	long sc;
1424	>	int ns = w.nsteals;
1425	>	w.nsteals = 0; // handle overflow
1426	>	long steals = (ns >= 0) ? ns : 1L + (long)(Integer.MAX_VALUE);
1427	>	do {} while (!U.compareAndSwapLong(this, STEALCOUNT,
1428	>	sc = stealCount, sc + steals));
1429		}
1430		}
1431
1432	<	// Signalling
1432	>	// Submissions
1433
1434		/**
1435	<	* Wakes up or creates a worker.
1436	<	*/
1437	<	final void signalWork() {
1438	<	/*
1439	<	* The while condition is true if: (there is are too few total
1440	<	* workers OR there is at least one waiter) AND (there are too
1441	<	* few active workers OR the pool is terminating).  The value
1442	<	* of e distinguishes the remaining cases: zero (no waiters)
1443	<	* for create, negative if terminating (in which case do
1444	<	* nothing), else release a waiter. The secondary checks for
1445	<	* release (non-null array etc) can fail if the pool begins
1446	<	* terminating after the test, and don't impose any added cost
1447	<	* because JVMs must perform null and bounds checks anyway.
1448	<	*/
1449	<	long c; int e, u;
1450	<	while ((((e = (int)(c = ctl)) | (u = (int)(c >>> 32))) &
1451	<	(INT_SIGN|SHORT_SIGN)) == (INT_SIGN|SHORT_SIGN) && e >= 0) {
1452	<	if (e > 0) {                         // release a waiting worker
1453	<	int i; ForkJoinWorkerThread w; ForkJoinWorkerThread[] ws;
1454	<	if ((ws = workers) == null ||
1455	<	(i = ~e & SMASK) >= ws.length ||
1456	<	(w = ws[i]) == null)
1457	<	break;
1458	<	long nc = (((long)(w.nextWait & E_MASK)) |
1459	<	((long)(u + UAC_UNIT) << 32));
1460	<	if (w.eventCount == e &&
1461	<	UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
1462	<	w.eventCount = (e + EC_UNIT) & E_MASK;
1463	<	if (w.parked)
1464	<	UNSAFE.unpark(w);
1465	<	break;
1466	<	}
1435	>	* Unless shutting down, adds the given task to a submission queue
1436	>	* at submitter's current queue index (modulo submission
1437	>	* range). Only the most common path is directly handled in this
1438	>	* method. All others are relayed to fullExternalPush.
1439	>	*
1440	>	* @param task the task. Caller must ensure non-null.
1441	>	*/
1442	>	final void externalPush(ForkJoinTask<?> task) {
1443	>	WorkQueue[] ws; WorkQueue q; Submitter z; int m; ForkJoinTask<?>[] a;
1444	>	if ((z = submitters.get()) != null && plock > 0 &&
1445	>	(ws = workQueues) != null && (m = (ws.length - 1)) >= 0 &&
1446	>	(q = ws[m & z.seed & SQMASK]) != null &&
1447	>	U.compareAndSwapInt(q, QLOCK, 0, 1)) { // lock
1448	>	int s = q.top, n;
1449	>	if ((a = q.array) != null && a.length > (n = s + 1 - q.base)) {
1450	>	U.putObject(a, (long)(((a.length - 1) & s) << ASHIFT) + ABASE,
1451	>	task);
1452	>	q.top = s + 1;                     // push on to deque
1453	>	q.qlock = 0;
1454	>	if (n <= 1)
1455	>	signalWork(q, 1);
1456	>	return;
1457	>	}
1458	>	q.qlock = 0;
1459	>	}
1460	>	fullExternalPush(task);
1461	>	}
1462	>
1463	>	/**
1464	>	* Full version of externalPush. This method is called, among
1465	>	* other times, upon the first submission of the first task to the
1466	>	* pool, so must perform secondary initialization: creating
1467	>	* workQueue array and setting plock to a valid value. It also
1468	>	* detects first submission by an external thread by looking up
1469	>	* its ThreadLocal, and creates a new shared queue if the one at
1470	>	* index if empty or contended. The lock bodies must be
1471	>	* exception-free (so no try/finally) so we optimistically
1472	>	* allocate new queues/arrays outside the locks and throw them
1473	>	* away if (very rarely) not needed. Note that the plock seq value
1474	>	* can eventually wrap around zero, but if so harmlessly fails to
1475	>	* reinitialize.
1476	>	*/
1477	>	private void fullExternalPush(ForkJoinTask<?> task) {
1478	>	for (Submitter z = null;;) {
1479	>	WorkQueue[] ws; WorkQueue q; int ps, m, r, s;
1480	>	if ((ps = plock) < 0)
1481	>	throw new RejectedExecutionException();
1482	>	else if ((ws = workQueues) == null || (m = ws.length - 1) < 0) {
1483	>	int n = parallelism - 1; n |= n >>> 1; n |= n >>> 2;
1484	>	n |= n >>> 4; n |= n >>> 8; n |= n >>> 16;
1485	>	WorkQueue[] nws = new WorkQueue[(n + 1) << 1]; // power of two
1486	>	if ((ps & PL_LOCK) != 0 ||
1487	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1488	>	ps = acquirePlock();
1489	>	if ((ws = workQueues) == null)
1490	>	workQueues = nws;
1491	>	int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1492	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1493	>	releasePlock(nps);
1494	>	}
1495	>	else if (z == null && (z = submitters.get()) == null) {
1496	>	if (U.compareAndSwapInt(this, INDEXSEED,
1497	>	s = indexSeed, s += SEED_INCREMENT) &&
1498	>	s != 0) // skip 0
1499	>	submitters.set(z = new Submitter(s));
1500		}
1501	<	else if (UNSAFE.compareAndSwapLong
1502	<	(this, ctlOffset, c,
1503	<	(long)(((u + UTC_UNIT) & UTC_MASK) |
1504	<	((u + UAC_UNIT) & UAC_MASK)) << 32)) {
1505	<	addWorker();
1506	<	break;
1501	>	else {
1502	>	int k = (r = z.seed) & m & SQMASK;
1503	>	if ((q = ws[k]) == null && (ps & PL_LOCK) == 0) {
1504	>	(q = new WorkQueue(this, null, SHARED_QUEUE)).poolIndex = k;
1505	>	if (((ps = plock) & PL_LOCK) != 0 ||
1506	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1507	>	ps = acquirePlock();
1508	>	WorkQueue w = null;
1509	>	if ((ws = workQueues) != null && k < ws.length &&
1510	>	(w = ws[k]) == null)
1511	>	ws[k] = q;
1512	>	else
1513	>	q = w;
1514	>	int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1515	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1516	>	releasePlock(nps);
1517	>	}
1518	>	if (q != null && q.qlock == 0 && q.fullPush(task, false))
1519	>	return;
1520	>	r ^= r << 13;                // same xorshift as WorkQueues
1521	>	r ^= r >>> 17;
1522	>	z.seed = r ^= r << 5;        // move to a different index
1523		}
1524		}
1525		}
1526
1527	+	// Maintaining ctl counts
1528	+
1529		/**
1530	<	* Variant of signalWork to help release waiters on rescans.
669	<	* Tries once to release a waiter if active count < 0.
670	<	*
671	<	* @return false if failed due to contention, else true
1530	>	* Increments active count; mainly called upon return from blocking.
1531		*/
1532	<	private boolean tryReleaseWaiter() {
1533	<	long c; int e, i; ForkJoinWorkerThread w; ForkJoinWorkerThread[] ws;
1534	<	if ((e = (int)(c = ctl)) > 0 &&
676	<	(int)(c >> AC_SHIFT) < 0 &&
677	<	(ws = workers) != null &&
678	<	(i = ~e & SMASK) < ws.length &&
679	<	(w = ws[i]) != null) {
680	<	long nc = ((long)(w.nextWait & E_MASK) |
681	<	((c + AC_UNIT) & (AC_MASK|TC_MASK)));
682	<	if (w.eventCount != e ||
683	<	!UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc))
684	<	return false;
685	<	w.eventCount = (e + EC_UNIT) & E_MASK;
686	<	if (w.parked)
687	<	UNSAFE.unpark(w);
688	<	}
689	<	return true;
1532	>	final void incrementActiveCount() {
1533	>	long c;
1534	>	do {} while (!U.compareAndSwapLong(this, CTL, c = ctl, c + AC_UNIT));
1535		}
1536
692	–	// Scanning for tasks
693	–
1537		/**
1538	<	* Scans for and, if found, executes one task. Scans start at a
1539	<	* random index of workers array, and randomly select the first
1540	<	* (2*#workers)-1 probes, and then, if all empty, resort to 2
1541	<	* circular sweeps, which is necessary to check quiescence. and
1542	<	* taking a submission only if no stealable tasks were found.  The
1543	<	* steal code inside the loop is a specialized form of
1544	<	* ForkJoinWorkerThread.deqTask, followed bookkeeping to support
1545	<	* helpJoinTask and signal propagation. The code for submission
1546	<	* queues is almost identical. On each steal, the worker completes
1547	<	* not only the task, but also all local tasks that this task may
1548	<	* have generated. On detecting staleness or contention when
1549	<	* trying to take a task, this method returns without finishing
1550	<	* sweep, which allows global state rechecks before retry.
1551	<	*
1552	<	* @param w the worker
1553	<	* @param a the number of active workers
1554	<	* @return true if swept all queues without finding a task
1555	<	*/
1556	<	private boolean scan(ForkJoinWorkerThread w, int a) {
1557	<	int g = scanGuard; // mask 0 avoids useless scans if only one active
1558	<	int m = (parallelism == 1 - a && blockedCount == 0) ? 0 : g & SMASK;
1559	<	ForkJoinWorkerThread[] ws = workers;
1560	<	if (ws == null || ws.length <= m)         // staleness check
1561	<	return false;
1562	<	for (int r = w.seed, k = r, j = -(m + m); j <= m + m; ++j) {
1563	<	ForkJoinTask<?> t; ForkJoinTask<?>[] q; int b, i;
1564	<	ForkJoinWorkerThread v = ws[k & m];
722	<	if (v != null && (b = v.queueBase) != v.queueTop &&
723	<	(q = v.queue) != null && (i = (q.length - 1) & b) >= 0) {
724	<	long u = (i << ASHIFT) + ABASE;
725	<	if ((t = q[i]) != null && v.queueBase == b &&
726	<	UNSAFE.compareAndSwapObject(q, u, t, null)) {
727	<	int d = (v.queueBase = b + 1) - v.queueTop;
728	<	v.stealHint = w.poolIndex;
729	<	if (d != 0)
730	<	signalWork();             // propagate if nonempty
731	<	w.execTask(t);
1538	>	* Tries to create (at most one) or activate (possibly several)
1539	>	* workers if too few are active. On contention failure, continues
1540	>	* until at least one worker is signalled or the given queue is
1541	>	* empty or all workers are active.
1542	>	*
1543	>	* @param q if non-null, the queue holding tasks to be signalled
1544	>	* @param signals the target number of signals.
1545	>	*/
1546	>	final void signalWork(WorkQueue q, int signals) {
1547	>	long c; int e, u, i; WorkQueue[] ws; WorkQueue w; Thread p;
1548	>	while ((u = (int)((c = ctl) >>> 32)) < 0) {
1549	>	if ((e = (int)c) > 0) {
1550	>	if ((ws = workQueues) != null && ws.length > (i = e & SMASK) &&
1551	>	(w = ws[i]) != null && w.eventCount == (e | INT_SIGN)) {
1552	>	long nc = (((long)(w.nextWait & E_MASK)) |
1553	>	((long)(u + UAC_UNIT) << 32));
1554	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1555	>	w.eventCount = (e + E_SEQ) & E_MASK;
1556	>	if ((p = w.parker) != null)
1557	>	U.unpark(p);
1558	>	if (--signals <= 0)
1559	>	break;
1560	>	}
1561	>	else
1562	>	signals = 1;
1563	>	if ((q != null && q.queueSize() == 0))
1564	>	break;
1565		}
1566	<	r ^= r << 13; r ^= r >>> 17; w.seed = r ^ (r << 5);
1567	<	return false;                     // store next seed
735	<	}
736	<	else if (j < 0) {                     // xorshift
737	<	r ^= r << 13; r ^= r >>> 17; k = r ^= r << 5;
1566	>	else
1567	>	break;
1568		}
1569	<	else
1570	<	++k;
1571	<	}
1572	<	if (scanGuard != g)                       // staleness check
1573	<	return false;
1574	<	else {                                    // try to take submission
1575	<	ForkJoinTask<?> t; ForkJoinTask<?>[] q; int b, i;
1576	<	if ((b = queueBase) != queueTop &&
1577	<	(q = submissionQueue) != null &&
1578	<	(i = (q.length - 1) & b) >= 0) {
1579	<	long u = (i << ASHIFT) + ABASE;
1580	<	if ((t = q[i]) != null && queueBase == b &&
1581	<	UNSAFE.compareAndSwapObject(q, u, t, null)) {
1582	<	queueBase = b + 1;
1583	<	w.execTask(t);
1569	>	else if (e == 0 && (u & SHORT_SIGN) != 0) {
1570	>	long nc = (long)(((u + UTC_UNIT) & UTC_MASK) |
1571	>	((u + UAC_UNIT) & UAC_MASK)) << 32;
1572	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1573	>	ForkJoinWorkerThread wt = null;
1574	>	Throwable ex = null;
1575	>	boolean started = false;
1576	>	try {
1577	>	ForkJoinWorkerThreadFactory fac;
1578	>	if ((fac = factory) != null &&
1579	>	(wt = fac.newThread(this)) != null) {
1580	>	wt.start();
1581	>	started = true;
1582	>	}
1583	>	} catch (Throwable rex) {
1584	>	ex = rex;
1585	>	}
1586	>	if (!started)
1587	>	deregisterWorker(wt, ex); // adjust counts on failure
1588	>	break;
1589		}
755	–	return false;
1590		}
1591	<	return true;                         // all queues empty
1591	>	else
1592	>	break;
1593		}
1594		}
1595
1596	+	// Scanning for tasks
1597	+
1598		/**
1599	<	* Tries to enqueue worker w in wait queue and await change in
763	<	* worker's eventCount.  If the pool is quiescent and there is
764	<	* more than one worker, possibly terminates worker upon exit.
765	<	* Otherwise, before blocking, rescans queues to avoid missed
766	<	* signals.  Upon finding work, releases at least one worker
767	<	* (which may be the current worker). Rescans restart upon
768	<	* detected staleness or failure to release due to
769	<	* contention. Note the unusual conventions about Thread.interrupt
770	<	* here and elsewhere: Because interrupts are used solely to alert
771	<	* threads to check termination, which is checked here anyway, we
772	<	* clear status (using Thread.interrupted) before any call to
773	<	* park, so that park does not immediately return due to status
774	<	* being set via some other unrelated call to interrupt in user
775	<	* code.
776	<	*
777	<	* @param w the calling worker
778	<	* @param c the ctl value on entry
779	<	* @return true if waited or another thread was released upon enq
1599	>	* Top-level runloop for workers, called by ForkJoinWorkerThread.run.
1600		*/
1601	<	private boolean tryAwaitWork(ForkJoinWorkerThread w, long c) {
1602	<	int v = w.eventCount;
1603	<	w.nextWait = (int)c;                      // w's successor record
1604	<	long nc = (long)(v & E_MASK) | ((c - AC_UNIT) & (AC_MASK|TC_MASK));
1605	<	if (ctl != c || !UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
1606	<	long d = ctl; // return true if lost to a deq, to force scan
1607	<	return (int)d != (int)c && ((d - c) & AC_MASK) >= 0L;
1608	<	}
1609	<	for (int sc = w.stealCount; sc != 0;) {   // accumulate stealCount
1610	<	long s = stealCount;
1611	<	if (UNSAFE.compareAndSwapLong(this, stealCountOffset, s, s + sc))
1612	<	sc = w.stealCount = 0;
1613	<	else if (w.eventCount != v)
1614	<	return true;                      // update next time
1615	<	}
1616	<	if ((!shutdown || !tryTerminate(false)) &&
1617	<	(int)c != 0 && parallelism + (int)(nc >> AC_SHIFT) == 0 &&
1618	<	blockedCount == 0 && quiescerCount == 0)
1619	<	idleAwaitWork(w, nc, c, v);           // quiescent
1620	<	for (boolean rescanned = false;;) {
1621	<	if (w.eventCount != v)
1622	<	return true;
1623	<	if (!rescanned) {
1624	<	int g = scanGuard, m = g & SMASK;
1625	<	ForkJoinWorkerThread[] ws = workers;
1626	<	if (ws != null && m < ws.length) {
1627	<	rescanned = true;
1628	<	for (int i = 0; i <= m; ++i) {
1629	<	ForkJoinWorkerThread u = ws[i];
1630	<	if (u != null) {
1631	<	if (u.queueBase != u.queueTop &&
1632	<	!tryReleaseWaiter())
1633	<	rescanned = false; // contended
1634	<	if (w.eventCount != v)
1635	<	return true;
1636	<	}
1601	>	final void runWorker(WorkQueue w) {
1602	>	// initialize queue array in this thread
1603	>	w.array = new ForkJoinTask<?>[WorkQueue.INITIAL_QUEUE_CAPACITY];
1604	>	do { w.runTask(scan(w)); } while (w.qlock >= 0);
1605	>	}
1606	>
1607	>	/**
1608	>	* Scans for and, if found, returns one task, else possibly
1609	>	* inactivates the worker. This method operates on single reads of
1610	>	* volatile state and is designed to be re-invoked continuously,
1611	>	* in part because it returns upon detecting inconsistencies,
1612	>	* contention, or state changes that indicate possible success on
1613	>	* re-invocation.
1614	>	*
1615	>	* The scan searches for tasks across a random permutation of
1616	>	* queues (starting at a random index and stepping by a random
1617	>	* relative prime, checking each at least once).  The scan
1618	>	* terminates upon either finding a non-empty queue, or completing
1619	>	* the sweep. If the worker is not inactivated, it takes and
1620	>	* returns a task from this queue. Otherwise, if not activated, it
1621	>	* signals workers (that may include itself) and returns so caller
1622	>	* can retry. Also returns for trtry if the worker array may have
1623	>	* changed during an empty scan.  On failure to find a task, we
1624	>	* take one of the following actions, after which the caller will
1625	>	* retry calling this method unless terminated.
1626	>	*
1627	>	* * If pool is terminating, terminate the worker.
1628	>	*
1629	>	* * If not already enqueued, try to inactivate and enqueue the
1630	>	* worker on wait queue. Or, if inactivating has caused the pool
1631	>	* to be quiescent, relay to idleAwaitWork to check for
1632	>	* termination and possibly shrink pool.
1633	>	*
1634	>	* * If already enqueued and none of the above apply, possibly
1635	>	* (with 1/2 probability) park awaiting signal, else lingering to
1636	>	* help scan and signal.
1637	>	*
1638	>	* @param w the worker (via its WorkQueue)
1639	>	* @return a task or null if none found
1640	>	*/
1641	>	private final ForkJoinTask<?> scan(WorkQueue w) {
1642	>	WorkQueue[] ws; WorkQueue q;           // first update random seed
1643	>	int r = w.seed; r ^= r << 13; r ^= r >>> 17; w.seed = r ^= r << 5;
1644	>	int ps = plock, m;                     // volatile read order matters
1645	>	if ((ws = workQueues) != null && (m = ws.length - 1) > 0) {
1646	>	int ec = w.eventCount;             // ec is negative if inactive
1647	>	int step = (r >>> 16) | 1;         // relatively prime
1648	>	for (int j = (m + 1) << 2;  ; --j, r += step) {
1649	>	ForkJoinTask<?> t; ForkJoinTask<?>[] a; int b, n;
1650	>	if ((q = ws[r & m]) != null && (b = q.base) - q.top < 0 &&
1651	>	(a = q.array) != null) {   // probably nonempty
1652	>	int i = (((a.length - 1) & b) << ASHIFT) + ABASE;
1653	>	t = (ForkJoinTask<?>)U.getObjectVolatile(a, i);
1654	>	if (q.base == b && ec >= 0 && t != null &&
1655	>	U.compareAndSwapObject(a, i, t, null)) {
1656	>	if ((n = q.top - (q.base = b + 1)) > 0)
1657	>	signalWork(q, n);
1658	>	return t;              // taken
1659	>	}
1660	>	if (j < m || (ec < 0 && (ec = w.eventCount) < 0)) {
1661	>	if ((n = q.queueSize() - 1) > 0)
1662	>	signalWork(q, n);
1663	>	break;                 // let caller retry after signal
1664		}
1665		}
1666	<	if (scanGuard != g ||              // stale
1667	<	(queueBase != queueTop && !tryReleaseWaiter()))
1668	<	rescanned = false;
1669	<	if (!rescanned)
1670	<	Thread.yield();                // reduce contention
1671	<	else
1672	<	Thread.interrupted();          // clear before park
1673	<	}
1674	<	else {
1675	<	w.parked = true;                   // must recheck
1676	<	if (w.eventCount != v) {
1677	<	w.parked = false;
1678	<	return true;
1666	>	else if (j < 0) {              // end of scan
1667	>	long c = ctl; int e;
1668	>	if (plock != ps)           // incomplete sweep
1669	>	break;
1670	>	if ((e = (int)c) < 0)      // pool is terminating
1671	>	w.qlock = -1;
1672	>	else if (ec >= 0) {        // try to enqueue/inactivate
1673	>	long nc = ((long)ec |
1674	>	((c - AC_UNIT) & (AC_MASK|TC_MASK)));
1675	>	w.nextWait = e;
1676	>	w.eventCount = ec | INT_SIGN; // mark as inactive
1677	>	if (ctl != c ||
1678	>	!U.compareAndSwapLong(this, CTL, c, nc))
1679	>	w.eventCount = ec; // unmark on CAS failure
1680	>	else if ((int)(c >> AC_SHIFT) == 1 - parallelism)
1681	>	idleAwaitWork(w, nc, c);  // quiescent
1682	>	}
1683	>	else if (w.seed >= 0 && w.eventCount < 0) {
1684	>	Thread wt = Thread.currentThread();
1685	>	Thread.interrupted();  // clear status
1686	>	U.putObject(wt, PARKBLOCKER, this);
1687	>	w.parker = wt;         // emulate LockSupport.park
1688	>	if (w.eventCount < 0)  // recheck
1689	>	U.park(false, 0L);
1690	>	w.parker = null;
1691	>	U.putObject(wt, PARKBLOCKER, null);
1692	>	}
1693	>	break;
1694		}
833	–	LockSupport.park(this);
834	–	rescanned = w.parked = false;
1695		}
1696		}
1697	+	return null;
1698		}
1699
1700		/**
1701	<	* If inactivating worker w has caused pool to become
1702	<	* quiescent, check for pool termination, and wait for event
1703	<	* for up to SHRINK_RATE nanosecs (rescans are unnecessary in
1704	<	* this case because quiescence reflects consensus about lack
1705	<	* of work). On timeout, if ctl has not changed, terminate the
1706	<	* worker. Upon its termination (see deregisterWorker), it may
846	<	* wake up another worker to possibly repeat this process.
1701	>	* If inactivating worker w has caused the pool to become
1702	>	* quiescent, checks for pool termination, and, so long as this is
1703	>	* not the only worker, waits for event for up to a given
1704	>	* duration.  On timeout, if ctl has not changed, terminates the
1705	>	* worker, which will in turn wake up another worker to possibly
1706	>	* repeat this process.
1707		*
1708		* @param w the calling worker
1709	<	* @param currentCtl the ctl value after enqueuing w
1710	<	* @param prevCtl the ctl value if w terminated
1711	<	* @param v the eventCount w awaits change
1712	<	*/
1713	<	private void idleAwaitWork(ForkJoinWorkerThread w, long currentCtl,
1714	<	long prevCtl, int v) {
1715	<	if (w.eventCount == v) {
1716	<	if (shutdown)
1717	<	tryTerminate(false);
1718	<	ForkJoinTask.helpExpungeStaleExceptions(); // help clean weak refs
1709	>	* @param currentCtl the ctl value triggering possible quiescence
1710	>	* @param prevCtl the ctl value to restore if thread is terminated
1711	>	*/
1712	>	private void idleAwaitWork(WorkQueue w, long currentCtl, long prevCtl) {
1713	>	if (w.eventCount < 0 &&
1714	>	(this == commonPool || !tryTerminate(false, false)) &&
1715	>	(int)prevCtl != 0) {
1716	>	int dc = -(short)(currentCtl >>> TC_SHIFT);
1717	>	long parkTime = dc < 0 ? FAST_IDLE_TIMEOUT: (dc + 1) * IDLE_TIMEOUT;
1718	>	long deadline = System.nanoTime() + parkTime - 100000L; // 1ms slop
1719	>	Thread wt = Thread.currentThread();
1720		while (ctl == currentCtl) {
1721	<	long startTime = System.nanoTime();
1722	<	w.parked = true;
1723	<	if (w.eventCount == v)             // must recheck
1724	<	LockSupport.parkNanos(this, SHRINK_RATE);
1725	<	w.parked = false;
1726	<	if (w.eventCount != v)
1721	>	Thread.interrupted();  // timed variant of version in scan()
1722	>	U.putObject(wt, PARKBLOCKER, this);
1723	>	w.parker = wt;
1724	>	if (ctl == currentCtl)
1725	>	U.park(false, parkTime);
1726	>	w.parker = null;
1727	>	U.putObject(wt, PARKBLOCKER, null);
1728	>	if (ctl != currentCtl)
1729		break;
1730	<	else if (System.nanoTime() - startTime <
1731	<	SHRINK_RATE - (SHRINK_RATE / 10)) // timing slop
1732	<	Thread.interrupted();          // spurious wakeup
1733	<	else if (UNSAFE.compareAndSwapLong(this, ctlOffset,
871	<	currentCtl, prevCtl)) {
872	<	w.terminate = true;            // restore previous
873	<	w.eventCount = ((int)currentCtl + EC_UNIT) & E_MASK;
1730	>	if (deadline - System.nanoTime() <= 0L &&
1731	>	U.compareAndSwapLong(this, CTL, currentCtl, prevCtl)) {
1732	>	w.eventCount = (w.eventCount + E_SEQ) | E_MASK;
1733	>	w.qlock = -1;   // shrink
1734		break;
1735		}
1736		}
1737		}
1738		}
1739
880	–	// Submissions
881	–
1740		/**
1741	<	* Enqueues the given task in the submissionQueue.  Same idea as
1742	<	* ForkJoinWorkerThread.pushTask except for use of submissionLock.
1741	>	* Scans through queues looking for work while joining a task;
1742	>	* if any are present, signals.
1743		*
1744	<	* @param t the task
1744	>	* @param task to return early if done
1745	>	* @param origin an index to start scan
1746		*/
1747	<	private void addSubmission(ForkJoinTask<?> t) {
1748	<	final ReentrantLock lock = this.submissionLock;
1749	<	lock.lock();
1750	<	try {
1751	<	ForkJoinTask<?>[] q; int s, m;
1752	<	if ((q = submissionQueue) != null) {    // ignore if queue removed
1753	<	long u = (((s = queueTop) & (m = q.length-1)) << ASHIFT)+ABASE;
1754	<	UNSAFE.putOrderedObject(q, u, t);
1755	<	queueTop = s + 1;
1756	<	if (s - queueBase == m)
1757	<	growSubmissionQueue();
1747	>	final int helpSignal(ForkJoinTask<?> task, int origin) {
1748	>	WorkQueue[] ws; WorkQueue q; int m, n, s;
1749	>	if (task != null && (ws = workQueues) != null &&
1750	>	(m = ws.length - 1) >= 0) {
1751	>	for (int i = 0; i <= m; ++i) {
1752	>	if ((s = task.status) < 0)
1753	>	return s;
1754	>	if ((q = ws[(i + origin) & m]) != null &&
1755	>	(n = q.queueSize()) > 0) {
1756	>	signalWork(q, n);
1757	>	if ((int)(ctl >> AC_SHIFT) >= 0)
1758	>	break;
1759	>	}
1760		}
900	–	} finally {
901	–	lock.unlock();
1761		}
1762	<	signalWork();
1762	>	return 0;
1763		}
1764
906	–	//  (pollSubmission is defined below with exported methods)
907	–
1765		/**
1766	<	* Creates or doubles submissionQueue array.
1767	<	* Basically identical to ForkJoinWorkerThread version.
1768	<	*/
1769	<	private void growSubmissionQueue() {
1770	<	ForkJoinTask<?>[] oldQ = submissionQueue;
1771	<	int size = oldQ != null ? oldQ.length << 1 : INITIAL_QUEUE_CAPACITY;
1772	<	if (size > MAXIMUM_QUEUE_CAPACITY)
1773	<	throw new RejectedExecutionException("Queue capacity exceeded");
1774	<	if (size < INITIAL_QUEUE_CAPACITY)
1775	<	size = INITIAL_QUEUE_CAPACITY;
1776	<	ForkJoinTask<?>[] q = submissionQueue = new ForkJoinTask<?>[size];
1777	<	int mask = size - 1;
1778	<	int top = queueTop;
1779	<	int oldMask;
1780	<	if (oldQ != null && (oldMask = oldQ.length - 1) >= 0) {
1781	<	for (int b = queueBase; b != top; ++b) {
1782	<	long u = ((b & oldMask) << ASHIFT) + ABASE;
1783	<	Object x = UNSAFE.getObjectVolatile(oldQ, u);
1784	<	if (x != null && UNSAFE.compareAndSwapObject(oldQ, u, x, null))
1785	<	UNSAFE.putObjectVolatile
1786	<	(q, ((b & mask) << ASHIFT) + ABASE, x);
1787	<	}
1788	<	}
1789	<	}
1790	<
1791	<	// Blocking support
1792	<
1793	<	/**
1794	<	* Tries to increment blockedCount, decrement active count
1795	<	* (sometimes implicitly) and possibly release or create a
1796	<	* compensating worker in preparation for blocking. Fails
1797	<	* on contention or termination.
1798	<	*
1799	<	* @return true if the caller can block, else should recheck and retry
1800	<	*/
1801	<	private boolean tryPreBlock() {
1802	<	int b = blockedCount;
1803	<	if (UNSAFE.compareAndSwapInt(this, blockedCountOffset, b, b + 1)) {
1804	<	int pc = parallelism;
1805	<	do {
1806	<	ForkJoinWorkerThread[] ws; ForkJoinWorkerThread w;
1807	<	int e, ac, tc, i;
1808	<	long c = ctl;
1809	<	int u = (int)(c >>> 32);
1810	<	if ((e = (int)c) < 0) {
1811	<	// skip -- terminating
1812	<	}
1813	<	else if ((ac = (u >> UAC_SHIFT)) <= 0 && e != 0 &&
1814	<	(ws = workers) != null &&
1815	<	(i = ~e & SMASK) < ws.length &&
1816	<	(w = ws[i]) != null) {
1817	<	long nc = ((long)(w.nextWait & E_MASK) |
1818	<	(c & (AC_MASK|TC_MASK)));
1819	<	if (w.eventCount == e &&
1820	<	UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
1821	<	w.eventCount = (e + EC_UNIT) & E_MASK;
1822	<	if (w.parked)
1823	<	UNSAFE.unpark(w);
1824	<	return true;             // release an idle worker
1825	<	}
1826	<	}
1827	<	else if ((tc = (short)(u >>> UTC_SHIFT)) >= 0 && ac + pc > 1) {
1828	<	long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK);
1829	<	if (UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc))
1830	<	return true;             // no compensation needed
1831	<	}
1832	<	else if (tc + pc < MAX_ID) {
1833	<	long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK);
1834	<	if (UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
1835	<	addWorker();
1836	<	return true;            // create a replacement
1837	<	}
1838	<	}
1839	<	// try to back out on any failure and let caller retry
1840	<	} while (!UNSAFE.compareAndSwapInt(this, blockedCountOffset,
1841	<	b = blockedCount, b - 1));
1766	>	* Tries to locate and execute tasks for a stealer of the given
1767	>	* task, or in turn one of its stealers, Traces currentSteal ->
1768	>	* currentJoin links looking for a thread working on a descendant
1769	>	* of the given task and with a non-empty queue to steal back and
1770	>	* execute tasks from. The first call to this method upon a
1771	>	* waiting join will often entail scanning/search, (which is OK
1772	>	* because the joiner has nothing better to do), but this method
1773	>	* leaves hints in workers to speed up subsequent calls. The
1774	>	* implementation is very branchy to cope with potential
1775	>	* inconsistencies or loops encountering chains that are stale,
1776	>	* unknown, or so long that they are likely cyclic.
1777	>	*
1778	>	* @param joiner the joining worker
1779	>	* @param task the task to join
1780	>	* @return 0 if no progress can be made, negative if task
1781	>	* known complete, else positive
1782	>	*/
1783	>	private int tryHelpStealer(WorkQueue joiner, ForkJoinTask<?> task) {
1784	>	int stat = 0, steps = 0;                    // bound to avoid cycles
1785	>	if (joiner != null && task != null) {       // hoist null checks
1786	>	restart: for (;;) {
1787	>	ForkJoinTask<?> subtask = task;     // current target
1788	>	for (WorkQueue j = joiner, v;;) {   // v is stealer of subtask
1789	>	WorkQueue[] ws; int m, s, h;
1790	>	if ((s = task.status) < 0) {
1791	>	stat = s;
1792	>	break restart;
1793	>	}
1794	>	if ((ws = workQueues) == null || (m = ws.length - 1) <= 0)
1795	>	break restart;              // shutting down
1796	>	if ((v = ws[h = (j.stealHint | 1) & m]) == null ||
1797	>	v.currentSteal != subtask) {
1798	>	for (int origin = h;;) {    // find stealer
1799	>	if (((h = (h + 2) & m) & 15) == 1 &&
1800	>	(subtask.status < 0 || j.currentJoin != subtask))
1801	>	continue restart;   // occasional staleness check
1802	>	if ((v = ws[h]) != null &&
1803	>	v.currentSteal == subtask) {
1804	>	j.stealHint = h;    // save hint
1805	>	break;
1806	>	}
1807	>	if (h == origin)
1808	>	break restart;      // cannot find stealer
1809	>	}
1810	>	}
1811	>	for (;;) { // help stealer or descend to its stealer
1812	>	ForkJoinTask[] a;  int b;
1813	>	if (subtask.status < 0)     // surround probes with
1814	>	continue restart;       //   consistency checks
1815	>	if ((b = v.base) - v.top < 0 && (a = v.array) != null) {
1816	>	int i = (((a.length - 1) & b) << ASHIFT) + ABASE;
1817	>	ForkJoinTask<?> t =
1818	>	(ForkJoinTask<?>)U.getObjectVolatile(a, i);
1819	>	if (subtask.status < 0 || j.currentJoin != subtask ||
1820	>	v.currentSteal != subtask)
1821	>	continue restart;   // stale
1822	>	stat = 1;               // apparent progress
1823	>	if (t != null && v.base == b &&
1824	>	U.compareAndSwapObject(a, i, t, null)) {
1825	>	v.base = b + 1;     // help stealer
1826	>	joiner.runSubtask(t);
1827	>	}
1828	>	else if (v.base == b && ++steps == MAX_HELP)
1829	>	break restart;      // v apparently stalled
1830	>	}
1831	>	else {                      // empty -- try to descend
1832	>	ForkJoinTask<?> next = v.currentJoin;
1833	>	if (subtask.status < 0 || j.currentJoin != subtask ||
1834	>	v.currentSteal != subtask)
1835	>	continue restart;   // stale
1836	>	else if (next == null || ++steps == MAX_HELP)
1837	>	break restart;      // dead-end or maybe cyclic
1838	>	else {
1839	>	subtask = next;
1840	>	j = v;
1841	>	break;
1842	>	}
1843	>	}
1844	>	}
1845	>	}
1846	>	}
1847		}
1848	<	return false;
987	<	}
988	<
989	<	/**
990	<	* Decrements blockedCount and increments active count.
991	<	*/
992	<	private void postBlock() {
993	<	long c;
994	<	do {} while (!UNSAFE.compareAndSwapLong(this, ctlOffset,  // no mask
995	<	c = ctl, c + AC_UNIT));
996	<	int b;
997	<	do {} while (!UNSAFE.compareAndSwapInt(this, blockedCountOffset,
998	<	b = blockedCount, b - 1));
1848	>	return stat;
1849		}
1850
1851		/**
1852	<	* Possibly blocks waiting for the given task to complete, or
1853	<	* cancels the task if terminating.  Fails to wait if contended.
1852	>	* Analog of tryHelpStealer for CountedCompleters. Tries to steal
1853	>	* and run tasks within the target's computation.
1854	>	*
1855	>	* @param task the task to join
1856	>	* @param mode if shared, exit upon completing any task
1857	>	* if all workers are active
1858		*
1005	–	* @param joinMe the task
1859		*/
1860	<	final void tryAwaitJoin(ForkJoinTask<?> joinMe) {
1861	<	Thread.interrupted(); // clear interrupts before checking termination
1862	<	if (joinMe.status >= 0) {
1863	<	if (tryPreBlock()) {
1864	<	joinMe.tryAwaitDone(0L);
1865	<	postBlock();
1860	>	private int helpComplete(ForkJoinTask<?> task, int mode) {
1861	>	WorkQueue[] ws; WorkQueue q; int m, n, s;
1862	>	if (task != null && (ws = workQueues) != null &&
1863	>	(m = ws.length - 1) >= 0) {
1864	>	for (int j = 1, origin = j;;) {
1865	>	if ((s = task.status) < 0)
1866	>	return s;
1867	>	if ((q = ws[j & m]) != null && q.pollAndExecCC(task)) {
1868	>	origin = j;
1869	>	if (mode == SHARED_QUEUE && (int)(ctl >> AC_SHIFT) >= 0)
1870	>	break;
1871	>	}
1872	>	else if ((j = (j + 2) & m) == origin)
1873	>	break;
1874		}
1014	–	else if ((ctl & STOP_BIT) != 0L)
1015	–	joinMe.cancelIgnoringExceptions();
1875		}
1876	+	return 0;
1877		}
1878
1879		/**
1880	<	* Possibly blocks the given worker waiting for joinMe to
1881	<	* complete or timeout.
1882	<	*
1883	<	* @param joinMe the task
1884	<	* @param millis the wait time for underlying Object.wait
1885	<	*/
1886	<	final void timedAwaitJoin(ForkJoinTask<?> joinMe, long nanos) {
1887	<	while (joinMe.status >= 0) {
1888	<	Thread.interrupted();
1889	<	if ((ctl & STOP_BIT) != 0L) {
1890	<	joinMe.cancelIgnoringExceptions();
1891	<	break;
1880	>	* Tries to decrement active count (sometimes implicitly) and
1881	>	* possibly release or create a compensating worker in preparation
1882	>	* for blocking. Fails on contention or termination. Otherwise,
1883	>	* adds a new thread if no idle workers are available and pool
1884	>	* may become starved.
1885	>	*/
1886	>	final boolean tryCompensate() {
1887	>	int pc = parallelism, e, u, i, tc; long c;
1888	>	WorkQueue[] ws; WorkQueue w; Thread p;
1889	>	if ((e = (int)(c = ctl)) >= 0 && (ws = workQueues) != null) {
1890	>	if (e != 0 && (i = e & SMASK) < ws.length &&
1891	>	(w = ws[i]) != null && w.eventCount == (e | INT_SIGN)) {
1892	>	long nc = ((long)(w.nextWait & E_MASK) |
1893	>	(c & (AC_MASK|TC_MASK)));
1894	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1895	>	w.eventCount = (e + E_SEQ) & E_MASK;
1896	>	if ((p = w.parker) != null)
1897	>	U.unpark(p);
1898	>	return true;   // replace with idle worker
1899	>	}
1900		}
1901	<	if (tryPreBlock()) {
1902	<	long last = System.nanoTime();
1903	<	while (joinMe.status >= 0) {
1904	<	long millis = TimeUnit.NANOSECONDS.toMillis(nanos);
1905	<	if (millis <= 0)
1906	<	break;
1907	<	joinMe.tryAwaitDone(millis);
1908	<	if (joinMe.status < 0)
1909	<	break;
1910	<	if ((ctl & STOP_BIT) != 0L) {
1911	<	joinMe.cancelIgnoringExceptions();
1912	<	break;
1901	>	else if ((short)((u = (int)(c >>> 32)) >>> UTC_SHIFT) >= 0 &&
1902	>	(u >> UAC_SHIFT) + pc > 1) {
1903	>	long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK);
1904	>	if (U.compareAndSwapLong(this, CTL, c, nc))
1905	>	return true;    // no compensation
1906	>	}
1907	>	else if ((tc = u + pc) < MAX_CAP) {
1908	>	long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK);
1909	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1910	>	Throwable ex = null;
1911	>	ForkJoinWorkerThread wt = null;
1912	>	try {
1913	>	ForkJoinWorkerThreadFactory fac;
1914	>	if ((fac = factory) != null &&
1915	>	(wt = fac.newThread(this)) != null) {
1916	>	wt.start();
1917	>	return true;
1918	>	}
1919	>	} catch (Throwable rex) {
1920	>	ex = rex;
1921		}
1922	<	long now = System.nanoTime();
1047	<	nanos -= now - last;
1048	<	last = now;
1922	>	deregisterWorker(wt, ex); // adjust counts etc
1923		}
1050	–	postBlock();
1051	–	break;
1924		}
1925		}
1926	+	return false;
1927		}
1928
1929		/**
1930	<	* If necessary, compensates for blocker, and blocks.
1930	>	* Helps and/or blocks until the given task is done.
1931	>	*
1932	>	* @param joiner the joining worker
1933	>	* @param task the task
1934	>	* @return task status on exit
1935		*/
1936	<	private void awaitBlocker(ManagedBlocker blocker)
1937	<	throws InterruptedException {
1938	<	while (!blocker.isReleasable()) {
1939	<	if (tryPreBlock()) {
1940	<	try {
1941	<	do {} while (!blocker.isReleasable() && !blocker.block());
1942	<	} finally {
1943	<	postBlock();
1936	>	final int awaitJoin(WorkQueue joiner, ForkJoinTask<?> task) {
1937	>	int s = 0;
1938	>	if (joiner != null && task != null && (s = task.status) >= 0) {
1939	>	ForkJoinTask<?> prevJoin = joiner.currentJoin;
1940	>	joiner.currentJoin = task;
1941	>	do {} while ((s = task.status) >= 0 &&
1942	>	joiner.queueSize() > 0 &&
1943	>	joiner.tryRemoveAndExec(task)); // process local tasks
1944	>	if (s >= 0 && (s = task.status) >= 0 &&
1945	>	(s = helpSignal(task, joiner.poolIndex)) >= 0 &&
1946	>	(task instanceof CountedCompleter))
1947	>	s = helpComplete(task, LIFO_QUEUE);
1948	>	while (s >= 0 && (s = task.status) >= 0) {
1949	>	if ((joiner.queueSize() > 0 ||           // try helping
1950	>	(s = tryHelpStealer(joiner, task)) == 0) &&
1951	>	(s = task.status) >= 0 && tryCompensate()) {
1952	>	if (task.trySetSignal() && (s = task.status) >= 0) {
1953	>	synchronized (task) {
1954	>	if (task.status >= 0) {
1955	>	try {                // see ForkJoinTask
1956	>	task.wait();     //  for explanation
1957	>	} catch (InterruptedException ie) {
1958	>	}
1959	>	}
1960	>	else
1961	>	task.notifyAll();
1962	>	}
1963	>	}
1964	>	long c;                          // re-activate
1965	>	do {} while (!U.compareAndSwapLong
1966	>	(this, CTL, c = ctl, c + AC_UNIT));
1967		}
1068	–	break;
1968		}
1969	+	joiner.currentJoin = prevJoin;
1970		}
1971	+	return s;
1972		}
1973
1073	–	// Creating, registering and deregistring workers
1074	–
1974		/**
1975	<	* Tries to create and start a worker; minimally rolls back counts
1976	<	* on failure.
1975	>	* Stripped-down variant of awaitJoin used by timed joins. Tries
1976	>	* to help join only while there is continuous progress. (Caller
1977	>	* will then enter a timed wait.)
1978	>	*
1979	>	* @param joiner the joining worker
1980	>	* @param task the task
1981		*/
1982	<	private void addWorker() {
1983	<	Throwable ex = null;
1984	<	ForkJoinWorkerThread t = null;
1985	<	try {
1986	<	t = factory.newThread(this);
1987	<	} catch (Throwable e) {
1988	<	ex = e;
1989	<	}
1990	<	if (t == null) {  // null or exceptional factory return
1991	<	long c;       // adjust counts
1992	<	do {} while (!UNSAFE.compareAndSwapLong
1993	<	(this, ctlOffset, c = ctl,
1994	<	(((c - AC_UNIT) & AC_MASK) |
1995	<	((c - TC_UNIT) & TC_MASK) |
1996	<	(c & ~(AC_MASK|TC_MASK)))));
1997	<	// Propagate exception if originating from an external caller
1998	<	if (!tryTerminate(false) && ex != null &&
1096	<	!(Thread.currentThread() instanceof ForkJoinWorkerThread))
1097	<	UNSAFE.throwException(ex);
1982	>	final void helpJoinOnce(WorkQueue joiner, ForkJoinTask<?> task) {
1983	>	int s;
1984	>	if (joiner != null && task != null && (s = task.status) >= 0) {
1985	>	ForkJoinTask<?> prevJoin = joiner.currentJoin;
1986	>	joiner.currentJoin = task;
1987	>	do {} while ((s = task.status) >= 0 &&
1988	>	joiner.queueSize() > 0 &&
1989	>	joiner.tryRemoveAndExec(task));
1990	>	if (s >= 0 && (s = task.status) >= 0 &&
1991	>	(s = helpSignal(task, joiner.poolIndex)) >= 0 &&
1992	>	(task instanceof CountedCompleter))
1993	>	s = helpComplete(task, LIFO_QUEUE);
1994	>	if (s >= 0 && joiner.queueSize() == 0) {
1995	>	do {} while (task.status >= 0 &&
1996	>	tryHelpStealer(joiner, task) > 0);
1997	>	}
1998	>	joiner.currentJoin = prevJoin;
1999		}
1099	–	else
1100	–	t.start();
2000		}
2001
2002		/**
2003	<	* Callback from ForkJoinWorkerThread constructor to assign a
2004	<	* public name
2005	<	*/
2006	<	final String nextWorkerName() {
2007	<	for (int n;;) {
2008	<	if (UNSAFE.compareAndSwapInt(this, nextWorkerNumberOffset,
2009	<	n = nextWorkerNumber, ++n))
2010	<	return workerNamePrefix + n;
2003	>	* Returns a (probably) non-empty steal queue, if one is found
2004	>	* during a random, then cyclic scan, else null.  This method must
2005	>	* be retried by caller if, by the time it tries to use the queue,
2006	>	* it is empty.
2007	>	* @param r a (random) seed for scanning
2008	>	*/
2009	>	private WorkQueue findNonEmptyStealQueue(int r) {
2010	>	int step = (r >>> 16) | 1;
2011	>	for (WorkQueue[] ws;;) {
2012	>	int ps = plock, m;
2013	>	if ((ws = workQueues) == null || (m = ws.length - 1) < 1)
2014	>	return null;
2015	>	for (int j = (m + 1) << 2; ; r += step) {
2016	>	WorkQueue q = ws[((r << 1) | 1) & m];
2017	>	if (q != null && q.queueSize() > 0)
2018	>	return q;
2019	>	else if (--j < 0) {
2020	>	if (plock == ps)
2021	>	return null;
2022	>	break;
2023	>	}
2024	>	}
2025		}
2026		}
2027
2028		/**
2029	<	* Callback from ForkJoinWorkerThread constructor to
2030	<	* determine its poolIndex and record in workers array.
2031	<	*
2032	<	* @param w the worker
2033	<	* @return the worker's pool index
2034	<	*/
2035	<	final int registerWorker(ForkJoinWorkerThread w) {
2036	<	/*
2037	<	* In the typical case, a new worker acquires the lock, uses
2038	<	* next available index and returns quickly.  Since we should
2039	<	* not block callers (ultimately from signalWork or
2040	<	* tryPreBlock) waiting for the lock needed to do this, we
2041	<	* instead help release other workers while waiting for the
2042	<	* lock.
2043	<	*/
2044	<	for (int g;;) {
2045	<	ForkJoinWorkerThread[] ws;
2046	<	if (((g = scanGuard) & SG_UNIT) == 0 &&
2047	<	UNSAFE.compareAndSwapInt(this, scanGuardOffset,
2048	<	g, g | SG_UNIT)) {
2049	<	int k = nextWorkerIndex;
2050	<	try {
2051	<	if ((ws = workers) != null) { // ignore on shutdown
2052	<	int n = ws.length;
2053	<	if (k < 0 || k >= n || ws[k] != null) {
2054	<	for (k = 0; k < n && ws[k] != null; ++k)
2055	<	;
2056	<	if (k == n)
2057	<	ws = workers = Arrays.copyOf(ws, n << 1);
2058	<	}
2059	<	ws[k] = w;
2060	<	nextWorkerIndex = k + 1;
2061	<	int m = g & SMASK;
2062	<	g = (k > m) ? ((m << 1) + 1) & SMASK : g + (SG_UNIT<<1);
2063	<	}
2064	<	} finally {
1152	<	scanGuard = g;
1153	<	}
1154	<	return k;
1155	<	}
1156	<	else if ((ws = workers) != null) { // help release others
1157	<	for (ForkJoinWorkerThread u : ws) {
1158	<	if (u != null && u.queueBase != u.queueTop) {
1159	<	if (tryReleaseWaiter())
1160	<	break;
1161	<	}
2029	>	* Runs tasks until {@code isQuiescent()}. We piggyback on
2030	>	* active count ctl maintenance, but rather than blocking
2031	>	* when tasks cannot be found, we rescan until all others cannot
2032	>	* find tasks either.
2033	>	*/
2034	>	final void helpQuiescePool(WorkQueue w) {
2035	>	for (boolean active = true;;) {
2036	>	ForkJoinTask<?> localTask; // exhaust local queue
2037	>	while ((localTask = w.nextLocalTask()) != null)
2038	>	localTask.doExec();
2039	>	// Similar to loop in scan(), but ignoring submissions
2040	>	WorkQueue q = findNonEmptyStealQueue(w.nextSeed());
2041	>	if (q != null) {
2042	>	ForkJoinTask<?> t; int b;
2043	>	if (!active) {      // re-establish active count
2044	>	long c;
2045	>	active = true;
2046	>	do {} while (!U.compareAndSwapLong
2047	>	(this, CTL, c = ctl, c + AC_UNIT));
2048	>	}
2049	>	if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null)
2050	>	w.runSubtask(t);
2051	>	}
2052	>	else {
2053	>	long c;
2054	>	if (active) {       // decrement active count without queuing
2055	>	active = false;
2056	>	do {} while (!U.compareAndSwapLong
2057	>	(this, CTL, c = ctl, c -= AC_UNIT));
2058	>	}
2059	>	else
2060	>	c = ctl;        // re-increment on exit
2061	>	if ((int)(c >> AC_SHIFT) + parallelism == 0) {
2062	>	do {} while (!U.compareAndSwapLong
2063	>	(this, CTL, c = ctl, c + AC_UNIT));
2064	>	break;
2065		}
2066		}
2067		}
2068		}
2069
2070		/**
2071	<	* Final callback from terminating worker.  Removes record of
2072	<	* worker from array, and adjusts counts. If pool is shutting
2073	<	* down, tries to complete termination.
2074	<	*
2075	<	* @param w the worker
2076	<	*/
2077	<	final void deregisterWorker(ForkJoinWorkerThread w, Throwable ex) {
2078	<	int idx = w.poolIndex;
2079	<	int sc = w.stealCount;
2080	<	int steps = 0;
2081	<	// Remove from array, adjust worker counts and collect steal count.
2082	<	// We can intermix failed removes or adjusts with steal updates
2083	<	do {
1181	<	long s, c;
1182	<	int g;
1183	<	if (steps == 0 && ((g = scanGuard) & SG_UNIT) == 0 &&
1184	<	UNSAFE.compareAndSwapInt(this, scanGuardOffset,
1185	<	g, g |= SG_UNIT)) {
1186	<	ForkJoinWorkerThread[] ws = workers;
1187	<	if (ws != null && idx >= 0 &&
1188	<	idx < ws.length && ws[idx] == w)
1189	<	ws[idx] = null;    // verify
1190	<	nextWorkerIndex = idx;
1191	<	scanGuard = g + SG_UNIT;
1192	<	steps = 1;
1193	<	}
1194	<	if (steps == 1 &&
1195	<	UNSAFE.compareAndSwapLong(this, ctlOffset, c = ctl,
1196	<	(((c - AC_UNIT) & AC_MASK) |
1197	<	((c - TC_UNIT) & TC_MASK) |
1198	<	(c & ~(AC_MASK|TC_MASK)))))
1199	<	steps = 2;
1200	<	if (sc != 0 &&
1201	<	UNSAFE.compareAndSwapLong(this, stealCountOffset,
1202	<	s = stealCount, s + sc))
1203	<	sc = 0;
1204	<	} while (steps != 2 || sc != 0);
1205	<	if (!tryTerminate(false)) {
1206	<	if (ex != null)   // possibly replace if died abnormally
1207	<	signalWork();
1208	<	else
1209	<	tryReleaseWaiter();
2071	>	* Gets and removes a local or stolen task for the given worker.
2072	>	*
2073	>	* @return a task, if available
2074	>	*/
2075	>	final ForkJoinTask<?> nextTaskFor(WorkQueue w) {
2076	>	for (ForkJoinTask<?> t;;) {
2077	>	WorkQueue q; int b;
2078	>	if ((t = w.nextLocalTask()) != null)
2079	>	return t;
2080	>	if ((q = findNonEmptyStealQueue(w.nextSeed())) == null)
2081	>	return null;
2082	>	if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null)
2083	>	return t;
2084		}
2085		}
2086
1213	–	// Shutdown and termination
1214	–
2087		/**
2088	<	* Possibly initiates and/or completes termination.
2088	>	* Returns a cheap heuristic guide for task partitioning when
2089	>	* programmers, frameworks, tools, or languages have little or no
2090	>	* idea about task granularity.  In essence by offering this
2091	>	* method, we ask users only about tradeoffs in overhead vs
2092	>	* expected throughput and its variance, rather than how finely to
2093	>	* partition tasks.
2094	>	*
2095	>	* In a steady state strict (tree-structured) computation, each
2096	>	* thread makes available for stealing enough tasks for other
2097	>	* threads to remain active. Inductively, if all threads play by
2098	>	* the same rules, each thread should make available only a
2099	>	* constant number of tasks.
2100	>	*
2101	>	* The minimum useful constant is just 1. But using a value of 1
2102	>	* would require immediate replenishment upon each steal to
2103	>	* maintain enough tasks, which is infeasible.  Further,
2104	>	* partitionings/granularities of offered tasks should minimize
2105	>	* steal rates, which in general means that threads nearer the top
2106	>	* of computation tree should generate more than those nearer the
2107	>	* bottom. In perfect steady state, each thread is at
2108	>	* approximately the same level of computation tree. However,
2109	>	* producing extra tasks amortizes the uncertainty of progress and
2110	>	* diffusion assumptions.
2111	>	*
2112	>	* So, users will want to use values larger, but not much larger
2113	>	* than 1 to both smooth over transient shortages and hedge
2114	>	* against uneven progress; as traded off against the cost of
2115	>	* extra task overhead. We leave the user to pick a threshold
2116	>	* value to compare with the results of this call to guide
2117	>	* decisions, but recommend values such as 3.
2118	>	*
2119	>	* When all threads are active, it is on average OK to estimate
2120	>	* surplus strictly locally. In steady-state, if one thread is
2121	>	* maintaining say 2 surplus tasks, then so are others. So we can
2122	>	* just use estimated queue length.  However, this strategy alone
2123	>	* leads to serious mis-estimates in some non-steady-state
2124	>	* conditions (ramp-up, ramp-down, other stalls). We can detect
2125	>	* many of these by further considering the number of "idle"
2126	>	* threads, that are known to have zero queued tasks, so
2127	>	* compensate by a factor of (#idle/#active) threads.
2128	>	*
2129	>	* Note: The approximation of #busy workers as #active workers is
2130	>	* not very good under current signalling scheme, and should be
2131	>	* improved.
2132	>	*/
2133	>	static int getSurplusQueuedTaskCount() {
2134	>	Thread t; ForkJoinWorkerThread wt; ForkJoinPool pool; WorkQueue q;
2135	>	if (((t = Thread.currentThread()) instanceof ForkJoinWorkerThread)) {
2136	>	int b = (q = (wt = (ForkJoinWorkerThread)t).workQueue).base;
2137	>	int p = (pool = wt.pool).parallelism;
2138	>	int a = (int)(pool.ctl >> AC_SHIFT) + p;
2139	>	return q.top - b - (a > (p >>>= 1) ? 0 :
2140	>	a > (p >>>= 1) ? 1 :
2141	>	a > (p >>>= 1) ? 2 :
2142	>	a > (p >>>= 1) ? 4 :
2143	>	8);
2144	>	}
2145	>	return 0;
2146	>	}
2147	>
2148	>	//  Termination
2149	>
2150	>	/**
2151	>	* Possibly initiates and/or completes termination.  The caller
2152	>	* triggering termination runs three passes through workQueues:
2153	>	* (0) Setting termination status, followed by wakeups of queued
2154	>	* workers; (1) cancelling all tasks; (2) interrupting lagging
2155	>	* threads (likely in external tasks, but possibly also blocked in
2156	>	* joins).  Each pass repeats previous steps because of potential
2157	>	* lagging thread creation.
2158		*
2159		* @param now if true, unconditionally terminate, else only
2160	<	* if shutdown and empty queue and no active workers
2160	>	* if no work and no active workers
2161	>	* @param enable if true, enable shutdown when next possible
2162		* @return true if now terminating or terminated
2163		*/
2164	<	private boolean tryTerminate(boolean now) {
2165	<	long c;
2166	<	while (((c = ctl) & STOP_BIT) == 0) {
2167	<	if (!now) {
2168	<	if ((int)(c >> AC_SHIFT) != -parallelism)
2169	<	return false;
2170	<	if (!shutdown || blockedCount != 0 || quiescerCount != 0 ||
2171	<	queueBase != queueTop) {
2172	<	if (ctl == c) // staleness check
1231	<	return false;
1232	<	continue;
2164	>	private boolean tryTerminate(boolean now, boolean enable) {
2165	>	if (this == commonPool)                     // cannot shut down
2166	>	return false;
2167	>	for (long c;;) {
2168	>	if (((c = ctl) & STOP_BIT) != 0) {      // already terminating
2169	>	if ((short)(c >>> TC_SHIFT) == -parallelism) {
2170	>	synchronized (this) {
2171	>	notifyAll();                // signal when 0 workers
2172	>	}
2173		}
2174	+	return true;
2175		}
2176	<	if (UNSAFE.compareAndSwapLong(this, ctlOffset, c, c | STOP_BIT))
2177	<	startTerminating();
2178	<	}
2179	<	if ((short)(c >>> TC_SHIFT) == -parallelism) { // signal when 0 workers
2180	<	final ReentrantLock lock = this.submissionLock;
2181	<	lock.lock();
2182	<	try {
2183	<	termination.signalAll();
2184	<	} finally {
2185	<	lock.unlock();
2176	>	if (plock >= 0) {                       // not yet enabled
2177	>	int ps;
2178	>	if (!enable)
2179	>	return false;
2180	>	if (((ps = plock) & PL_LOCK) != 0 ||
2181	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
2182	>	ps = acquirePlock();
2183	>	int nps = SHUTDOWN;
2184	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
2185	>	releasePlock(nps);
2186	>	}
2187	>	if (!now) {                             // check if idle & no tasks
2188	>	if ((int)(c >> AC_SHIFT) != -parallelism ||
2189	>	hasQueuedSubmissions())
2190	>	return false;
2191	>	// Check for unqueued inactive workers. One pass suffices.
2192	>	WorkQueue[] ws = workQueues; WorkQueue w;
2193	>	if (ws != null) {
2194	>	for (int i = 1; i < ws.length; i += 2) {
2195	>	if ((w = ws[i]) != null && w.eventCount >= 0)
2196	>	return false;
2197	>	}
2198	>	}
2199		}
2200	<	}
2201	<	return true;
2202	<	}
2203	<
2204	<	/**
2205	<	* Runs up to three passes through workers: (0) Setting
2206	<	* termination status for each worker, followed by wakeups up to
2207	<	* queued workers; (1) helping cancel tasks; (2) interrupting
2208	<	* lagging threads (likely in external tasks, but possibly also
2209	<	* blocked in joins).  Each pass repeats previous steps because of
2210	<	* potential lagging thread creation.
2211	<	*/
2212	<	private void startTerminating() {
1259	<	cancelSubmissions();
1260	<	for (int pass = 0; pass < 3; ++pass) {
1261	<	ForkJoinWorkerThread[] ws = workers;
1262	<	if (ws != null) {
1263	<	for (ForkJoinWorkerThread w : ws) {
1264	<	if (w != null) {
1265	<	w.terminate = true;
1266	<	if (pass > 0) {
1267	<	w.cancelTasks();
1268	<	if (pass > 1 && !w.isInterrupted()) {
1269	<	try {
1270	<	w.interrupt();
1271	<	} catch (SecurityException ignore) {
2200	>	if (U.compareAndSwapLong(this, CTL, c, c | STOP_BIT)) {
2201	>	for (int pass = 0; pass < 3; ++pass) {
2202	>	WorkQueue[] ws = workQueues;
2203	>	if (ws != null) {
2204	>	WorkQueue w;
2205	>	int n = ws.length;
2206	>	for (int i = 0; i < n; ++i) {
2207	>	if ((w = ws[i]) != null) {
2208	>	w.qlock = -1;
2209	>	if (pass > 0) {
2210	>	w.cancelAll();
2211	>	if (pass > 1)
2212	>	w.interruptOwner();
2213		}
2214		}
2215		}
2216	+	// Wake up workers parked on event queue
2217	+	int i, e; long cc; Thread p;
2218	+	while ((e = (int)(cc = ctl) & E_MASK) != 0 &&
2219	+	(i = e & SMASK) < n &&
2220	+	(w = ws[i]) != null) {
2221	+	long nc = ((long)(w.nextWait & E_MASK) |
2222	+	((cc + AC_UNIT) & AC_MASK) |
2223	+	(cc & (TC_MASK|STOP_BIT)));
2224	+	if (w.eventCount == (e | INT_SIGN) &&
2225	+	U.compareAndSwapLong(this, CTL, cc, nc)) {
2226	+	w.eventCount = (e + E_SEQ) & E_MASK;
2227	+	w.qlock = -1;
2228	+	if ((p = w.parker) != null)
2229	+	U.unpark(p);
2230	+	}
2231	+	}
2232		}
2233		}
1277	–	terminateWaiters();
2234		}
2235		}
2236		}
2237
2238	+	// external operations on common pool
2239	+
2240		/**
2241	<	* Polls and cancels all submissions. Called only during termination.
2241	>	* Returns common pool queue for a thread that has submitted at
2242	>	* least one task.
2243		*/
2244	<	private void cancelSubmissions() {
2245	<	while (queueBase != queueTop) {
2246	<	ForkJoinTask<?> task = pollSubmission();
2247	<	if (task != null) {
2248	<	try {
2249	<	task.cancel(false);
2250	<	} catch (Throwable ignore) {
2251	<	}
2244	>	static WorkQueue commonSubmitterQueue() {
2245	>	ForkJoinPool p; WorkQueue[] ws; int m; Submitter z;
2246	>	return ((z = submitters.get()) != null &&
2247	>	(p = commonPool) != null &&
2248	>	(ws = p.workQueues) != null &&
2249	>	(m = ws.length - 1) >= 0) ?
2250	>	ws[m & z.seed & SQMASK] : null;
2251	>	}
2252	>
2253	>	/**
2254	>	* Tries to pop the given task from submitter's queue in common pool.
2255	>	*/
2256	>	static boolean tryExternalUnpush(ForkJoinTask<?> t) {
2257	>	ForkJoinPool p; WorkQueue[] ws; WorkQueue q; Submitter z;
2258	>	ForkJoinTask<?>[] a;  int m, s; long j;
2259	>	if ((z = submitters.get()) != null &&
2260	>	(p = commonPool) != null &&
2261	>	(ws = p.workQueues) != null &&
2262	>	(m = ws.length - 1) >= 0 &&
2263	>	(q = ws[m & z.seed & SQMASK]) != null &&
2264	>	(s = q.top) != q.base &&
2265	>	(a = q.array) != null &&
2266	>	U.getObjectVolatile
2267	>	(a, j = (((a.length - 1) & (s - 1)) << ASHIFT) + ABASE) == t &&
2268	>	U.compareAndSwapInt(q, QLOCK, 0, 1)) {
2269	>	if (q.array == a && q.top == s && // recheck
2270	>	U.compareAndSwapObject(a, j, t, null)) {
2271	>	q.top = s - 1;
2272	>	q.qlock = 0;
2273	>	return true;
2274		}
2275	+	q.qlock = 0;
2276		}
2277	+	return false;
2278		}
2279
2280		/**
2281	<	* Tries to set the termination status of waiting workers, and
2282	<	* then wakes them up (after which they will terminate).
2283	<	*/
2284	<	private void terminateWaiters() {
2285	<	ForkJoinWorkerThread[] ws = workers;
2286	<	if (ws != null) {
2287	<	ForkJoinWorkerThread w; long c; int i, e;
2288	<	int n = ws.length;
2289	<	while ((i = ~(e = (int)(c = ctl)) & SMASK) < n &&
2290	<	(w = ws[i]) != null && w.eventCount == (e & E_MASK)) {
2291	<	if (UNSAFE.compareAndSwapLong(this, ctlOffset, c,
2292	<	(long)(w.nextWait & E_MASK) |
2293	<	((c + AC_UNIT) & AC_MASK) |
2294	<	(c & (TC_MASK|STOP_BIT)))) {
2295	<	w.terminate = true;
2296	<	w.eventCount = e + EC_UNIT;
2297	<	if (w.parked)
2298	<	UNSAFE.unpark(w);
2281	>	* Tries to pop and run local tasks within the same computation
2282	>	* as the given root. On failure, tries to help complete from
2283	>	* other queues via helpComplete.
2284	>	*/
2285	>	private void externalHelpComplete(WorkQueue q, ForkJoinTask<?> root) {
2286	>	ForkJoinTask<?>[] a; int m;
2287	>	if (q != null && (a = q.array) != null && (m = (a.length - 1)) >= 0 &&
2288	>	root != null && root.status >= 0) {
2289	>	for (;;) {
2290	>	int s; Object o; CountedCompleter<?> task = null;
2291	>	if ((s = q.top) - q.base > 0) {
2292	>	long j = ((m & (s - 1)) << ASHIFT) + ABASE;
2293	>	if ((o = U.getObject(a, j)) != null &&
2294	>	(o instanceof CountedCompleter)) {
2295	>	CountedCompleter<?> t = (CountedCompleter<?>)o, r = t;
2296	>	do {
2297	>	if (r == root) {
2298	>	if (U.compareAndSwapInt(q, QLOCK, 0, 1)) {
2299	>	if (q.array == a && q.top == s &&
2300	>	U.compareAndSwapObject(a, j, t, null)) {
2301	>	q.top = s - 1;
2302	>	task = t;
2303	>	}
2304	>	q.qlock = 0;
2305	>	}
2306	>	break;
2307	>	}
2308	>	} while ((r = r.completer) != null);
2309	>	}
2310	>	}
2311	>	if (task != null)
2312	>	task.doExec();
2313	>	if (root.status < 0 || (int)(ctl >> AC_SHIFT) >= 0)
2314	>	break;
2315	>	if (task == null) {
2316	>	if (helpSignal(root, q.poolIndex) >= 0)
2317	>	helpComplete(root, SHARED_QUEUE);
2318	>	break;
2319		}
2320		}
2321		}
2322		}
2323
1321	–	// misc ForkJoinWorkerThread support
1322	–
2324		/**
2325	<	* Increment or decrement quiescerCount. Needed only to prevent
2326	<	* triggering shutdown if a worker is transiently inactive while
1326	<	* checking quiescence.
1327	<	*
1328	<	* @param delta 1 for increment, -1 for decrement
2325	>	* Tries to help execute or signal availability of the given task
2326	>	* from submitter's queue in common pool.
2327		*/
2328	<	final void addQuiescerCount(int delta) {
2329	<	int c;
2330	<	do {} while (!UNSAFE.compareAndSwapInt(this, quiescerCountOffset,
2331	<	c = quiescerCount, c + delta));
2328	>	static void externalHelpJoin(ForkJoinTask<?> t) {
2329	>	// Some hard-to-avoid overlap with tryExternalUnpush
2330	>	ForkJoinPool p; WorkQueue[] ws; WorkQueue q, w; Submitter z;
2331	>	ForkJoinTask<?>[] a;  int m, s, n; long j;
2332	>	if (t != null && t.status >= 0 &&
2333	>	(z = submitters.get()) != null &&
2334	>	(p = commonPool) != null &&
2335	>	(ws = p.workQueues) != null &&
2336	>	(m = ws.length - 1) >= 0 &&
2337	>	(q = ws[m & z.seed & SQMASK]) != null &&
2338	>	(a = q.array) != null) {
2339	>	if ((s = q.top) != q.base &&
2340	>	U.getObjectVolatile
2341	>	(a, j = (((a.length - 1) & (s - 1)) << ASHIFT) + ABASE) == t &&
2342	>	U.compareAndSwapInt(q, QLOCK, 0, 1)) {
2343	>	if (q.array == a && q.top == s &&
2344	>	U.compareAndSwapObject(a, j, t, null)) {
2345	>	q.top = s - 1;
2346	>	q.qlock = 0;
2347	>	t.doExec();
2348	>	}
2349	>	else
2350	>	q.qlock = 0;
2351	>	}
2352	>	if (t.status >= 0) {
2353	>	if (t instanceof CountedCompleter)
2354	>	p.externalHelpComplete(q, t);
2355	>	else
2356	>	p.helpSignal(t, q.poolIndex);
2357	>	}
2358	>	}
2359		}
2360
2361		/**
2362	<	* Directly increment or decrement active count without
1338	<	* queuing. This method is used to transiently assert inactivation
1339	<	* while checking quiescence.
1340	<	*
1341	<	* @param delta 1 for increment, -1 for decrement
2362	>	* Restricted version of helpQuiescePool for external callers
2363		*/
2364	<	final void addActiveCount(int delta) {
2365	<	long d = delta < 0 ? -AC_UNIT : AC_UNIT;
2366	<	long c;
2367	<	do {} while (!UNSAFE.compareAndSwapLong(this, ctlOffset, c = ctl,
2368	<	((c + d) & AC_MASK) |
2369	<	(c & ~AC_MASK)));
2370	<	}
2371	<
1351	<	/**
1352	<	* Returns the approximate (non-atomic) number of idle threads per
1353	<	* active thread.
1354	<	*/
1355	<	final int idlePerActive() {
1356	<	// Approximate at powers of two for small values, saturate past 4
1357	<	int p = parallelism;
1358	<	int a = p + (int)(ctl >> AC_SHIFT);
1359	<	return (a > (p >>>= 1) ? 0 :
1360	<	a > (p >>>= 1) ? 1 :
1361	<	a > (p >>>= 1) ? 2 :
1362	<	a > (p >>>= 1) ? 4 :
1363	<	8);
2364	>	static void externalHelpQuiescePool() {
2365	>	ForkJoinPool p; ForkJoinTask<?> t; WorkQueue q; int b;
2366	>	int r = ThreadLocalRandom.current().nextInt();
2367	>	if ((p = commonPool) != null &&
2368	>	(q = p.findNonEmptyStealQueue(r)) != null &&
2369	>	(b = q.base) - q.top < 0 &&
2370	>	(t = q.pollAt(b)) != null)
2371	>	t.doExec();
2372		}
2373
2374		// Exported methods
#	Line 1432 | Line 2440 | public class ForkJoinPool extends Abstra
2440		checkPermission();
2441		if (factory == null)
2442		throw new NullPointerException();
2443	<	if (parallelism <= 0 || parallelism > MAX_ID)
2443	>	if (parallelism <= 0 || parallelism > MAX_CAP)
2444		throw new IllegalArgumentException();
2445		this.parallelism = parallelism;
2446		this.factory = factory;
2447		this.ueh = handler;
2448	<	this.locallyFifo = asyncMode;
2448	>	this.localMode = asyncMode ? FIFO_QUEUE : LIFO_QUEUE;
2449		long np = (long)(-parallelism); // offset ctl counts
2450		this.ctl = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK);
2451	<	this.submissionQueue = new ForkJoinTask<?>[INITIAL_QUEUE_CAPACITY];
1444	<	// initialize workers array with room for 2*parallelism if possible
1445	<	int n = parallelism << 1;
1446	<	if (n >= MAX_ID)
1447	<	n = MAX_ID;
1448	<	else { // See Hackers Delight, sec 3.2, where n < (1 << 16)
1449	<	n |= n >>> 1; n |= n >>> 2; n |= n >>> 4; n |= n >>> 8;
1450	<	}
1451	<	workers = new ForkJoinWorkerThread[n + 1];
1452	<	this.submissionLock = new ReentrantLock();
1453	<	this.termination = submissionLock.newCondition();
2451	>	int pn = nextPoolId();
2452		StringBuilder sb = new StringBuilder("ForkJoinPool-");
2453	<	sb.append(poolNumberGenerator.incrementAndGet());
2453	>	sb.append(Integer.toString(pn));
2454		sb.append("-worker-");
2455		this.workerNamePrefix = sb.toString();
2456		}
2457
2458	+	/**
2459	+	* Constructor for common pool, suitable only for static initialization.
2460	+	* Basically the same as above, but uses smallest possible initial footprint.
2461	+	*/
2462	+	ForkJoinPool(int parallelism, long ctl,
2463	+	ForkJoinWorkerThreadFactory factory,
2464	+	Thread.UncaughtExceptionHandler handler) {
2465	+	this.parallelism = parallelism;
2466	+	this.ctl = ctl;
2467	+	this.factory = factory;
2468	+	this.ueh = handler;
2469	+	this.localMode = LIFO_QUEUE;
2470	+	this.workerNamePrefix = "ForkJoinPool.commonPool-worker-";
2471	+	}
2472	+
2473	+	/**
2474	+	* Returns the common pool instance.
2475	+	*
2476	+	* @return the common pool instance
2477	+	*/
2478	+	public static ForkJoinPool commonPool() {
2479	+	return commonPool; // cannot be null (if so, a static init error)
2480	+	}
2481	+
2482		// Execution methods
2483
2484		/**
#	Line 1476 | Line 2498 | public class ForkJoinPool extends Abstra
2498		*         scheduled for execution
2499		*/
2500		public <T> T invoke(ForkJoinTask<T> task) {
1479	–	Thread t = Thread.currentThread();
2501		if (task == null)
2502		throw new NullPointerException();
2503	<	if (shutdown)
2504	<	throw new RejectedExecutionException();
1484	<	if ((t instanceof ForkJoinWorkerThread) &&
1485	<	((ForkJoinWorkerThread)t).pool == this)
1486	<	return task.invoke();  // bypass submit if in same pool
1487	<	else {
1488	<	addSubmission(task);
1489	<	return task.join();
1490	<	}
1491	<	}
1492	<
1493	<	/**
1494	<	* Unless terminating, forks task if within an ongoing FJ
1495	<	* computation in the current pool, else submits as external task.
1496	<	*/
1497	<	private <T> void forkOrSubmit(ForkJoinTask<T> task) {
1498	<	ForkJoinWorkerThread w;
1499	<	Thread t = Thread.currentThread();
1500	<	if (shutdown)
1501	<	throw new RejectedExecutionException();
1502	<	if ((t instanceof ForkJoinWorkerThread) &&
1503	<	(w = (ForkJoinWorkerThread)t).pool == this)
1504	<	w.pushTask(task);
1505	<	else
1506	<	addSubmission(task);
2503	>	externalPush(task);
2504	>	return task.join();
2505		}
2506
2507		/**
#	Line 1517 | Line 2515 | public class ForkJoinPool extends Abstra
2515		public void execute(ForkJoinTask<?> task) {
2516		if (task == null)
2517		throw new NullPointerException();
2518	<	forkOrSubmit(task);
2518	>	externalPush(task);
2519		}
2520
2521		// AbstractExecutorService methods
#	Line 1534 | Line 2532 | public class ForkJoinPool extends Abstra
2532		if (task instanceof ForkJoinTask<?>) // avoid re-wrap
2533		job = (ForkJoinTask<?>) task;
2534		else
2535	<	job = ForkJoinTask.adapt(task, null);
2536	<	forkOrSubmit(job);
2535	>	job = new ForkJoinTask.AdaptedRunnableAction(task);
2536	>	externalPush(job);
2537		}
2538
2539		/**
#	Line 1550 | Line 2548 | public class ForkJoinPool extends Abstra
2548		public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) {
2549		if (task == null)
2550		throw new NullPointerException();
2551	<	forkOrSubmit(task);
2551	>	externalPush(task);
2552		return task;
2553		}
2554
#	Line 1560 | Line 2558 | public class ForkJoinPool extends Abstra
2558		*         scheduled for execution
2559		*/
2560		public <T> ForkJoinTask<T> submit(Callable<T> task) {
2561	<	if (task == null)
2562	<	throw new NullPointerException();
1565	<	ForkJoinTask<T> job = ForkJoinTask.adapt(task);
1566	<	forkOrSubmit(job);
2561	>	ForkJoinTask<T> job = new ForkJoinTask.AdaptedCallable<T>(task);
2562	>	externalPush(job);
2563		return job;
2564		}
2565
#	Line 1573 | Line 2569 | public class ForkJoinPool extends Abstra
2569		*         scheduled for execution
2570		*/
2571		public <T> ForkJoinTask<T> submit(Runnable task, T result) {
2572	<	if (task == null)
2573	<	throw new NullPointerException();
1578	<	ForkJoinTask<T> job = ForkJoinTask.adapt(task, result);
1579	<	forkOrSubmit(job);
2572	>	ForkJoinTask<T> job = new ForkJoinTask.AdaptedRunnable<T>(task, result);
2573	>	externalPush(job);
2574		return job;
2575		}
2576
#	Line 1592 | Line 2586 | public class ForkJoinPool extends Abstra
2586		if (task instanceof ForkJoinTask<?>) // avoid re-wrap
2587		job = (ForkJoinTask<?>) task;
2588		else
2589	<	job = ForkJoinTask.adapt(task, null);
2590	<	forkOrSubmit(job);
2589	>	job = new ForkJoinTask.AdaptedRunnableAction(task);
2590	>	externalPush(job);
2591		return job;
2592		}
2593
#	Line 1602 | Line 2596 | public class ForkJoinPool extends Abstra
2596		* @throws RejectedExecutionException {@inheritDoc}
2597		*/
2598		public <T> List<Future<T>> invokeAll(Collection<? extends Callable<T>> tasks) {
2599	<	ArrayList<ForkJoinTask<T>> forkJoinTasks =
2600	<	new ArrayList<ForkJoinTask<T>>(tasks.size());
2601	<	for (Callable<T> task : tasks)
2602	<	forkJoinTasks.add(ForkJoinTask.adapt(task));
2603	<	invoke(new InvokeAll<T>(forkJoinTasks));
2604	<
2599	>	// In previous versions of this class, this method constructed
2600	>	// a task to run ForkJoinTask.invokeAll, but now external
2601	>	// invocation of multiple tasks is at least as efficient.
2602	>	List<ForkJoinTask<T>> fs = new ArrayList<ForkJoinTask<T>>(tasks.size());
2603	>	// Workaround needed because method wasn't declared with
2604	>	// wildcards in return type but should have been.
2605		@SuppressWarnings({"unchecked", "rawtypes"})
2606	<	List<Future<T>> futures = (List<Future<T>>) (List) forkJoinTasks;
1613	<	return futures;
1614	<	}
2606	>	List<Future<T>> futures = (List<Future<T>>) (List) fs;
2607
2608	<	static final class InvokeAll<T> extends RecursiveAction {
2609	<	final ArrayList<ForkJoinTask<T>> tasks;
2610	<	InvokeAll(ArrayList<ForkJoinTask<T>> tasks) { this.tasks = tasks; }
2611	<	public void compute() {
2612	<	try { invokeAll(tasks); }
2613	<	catch (Exception ignore) {}
2608	>	boolean done = false;
2609	>	try {
2610	>	for (Callable<T> t : tasks) {
2611	>	ForkJoinTask<T> f = new ForkJoinTask.AdaptedCallable<T>(t);
2612	>	externalPush(f);
2613	>	fs.add(f);
2614	>	}
2615	>	for (ForkJoinTask<T> f : fs)
2616	>	f.quietlyJoin();
2617	>	done = true;
2618	>	return futures;
2619	>	} finally {
2620	>	if (!done)
2621	>	for (ForkJoinTask<T> f : fs)
2622	>	f.cancel(false);
2623		}
1623	–	private static final long serialVersionUID = -7914297376763021607L;
2624		}
2625
2626		/**
#	Line 1652 | Line 2652 | public class ForkJoinPool extends Abstra
2652		}
2653
2654		/**
2655	+	* Returns the targeted parallelism level of the common pool.
2656	+	*
2657	+	* @return the targeted parallelism level of the common pool
2658	+	*/
2659	+	public static int getCommonPoolParallelism() {
2660	+	return commonPoolParallelism;
2661	+	}
2662	+
2663	+	/**
2664		* Returns the number of worker threads that have started but not
2665		* yet terminated.  The result returned by this method may differ
2666		* from {@link #getParallelism} when threads are created to
#	Line 1670 | Line 2679 | public class ForkJoinPool extends Abstra
2679		* @return {@code true} if this pool uses async mode
2680		*/
2681		public boolean getAsyncMode() {
2682	<	return locallyFifo;
2682	>	return localMode != 0;
2683		}
2684
2685		/**
#	Line 1682 | Line 2691 | public class ForkJoinPool extends Abstra
2691		* @return the number of worker threads
2692		*/
2693		public int getRunningThreadCount() {
2694	<	int r = parallelism + (int)(ctl >> AC_SHIFT);
2695	<	return (r <= 0) ? 0 : r; // suppress momentarily negative values
2694	>	int rc = 0;
2695	>	WorkQueue[] ws; WorkQueue w;
2696	>	if ((ws = workQueues) != null) {
2697	>	for (int i = 1; i < ws.length; i += 2) {
2698	>	if ((w = ws[i]) != null && w.isApparentlyUnblocked())
2699	>	++rc;
2700	>	}
2701	>	}
2702	>	return rc;
2703		}
2704
2705		/**
#	Line 1694 | Line 2710 | public class ForkJoinPool extends Abstra
2710		* @return the number of active threads
2711		*/
2712		public int getActiveThreadCount() {
2713	<	int r = parallelism + (int)(ctl >> AC_SHIFT) + blockedCount;
2713	>	int r = parallelism + (int)(ctl >> AC_SHIFT);
2714		return (r <= 0) ? 0 : r; // suppress momentarily negative values
2715		}
2716
#	Line 1710 | Line 2726 | public class ForkJoinPool extends Abstra
2726		* @return {@code true} if all threads are currently idle
2727		*/
2728		public boolean isQuiescent() {
2729	<	return parallelism + (int)(ctl >> AC_SHIFT) + blockedCount == 0;
2729	>	return (int)(ctl >> AC_SHIFT) + parallelism == 0;
2730		}
2731
2732		/**
#	Line 1725 | Line 2741 | public class ForkJoinPool extends Abstra
2741		* @return the number of steals
2742		*/
2743		public long getStealCount() {
2744	<	return stealCount;
2744	>	long count = stealCount;
2745	>	WorkQueue[] ws; WorkQueue w;
2746	>	if ((ws = workQueues) != null) {
2747	>	for (int i = 1; i < ws.length; i += 2) {
2748	>	if ((w = ws[i]) != null)
2749	>	count += w.nsteals;
2750	>	}
2751	>	}
2752	>	return count;
2753		}
2754
2755		/**
#	Line 1740 | Line 2764 | public class ForkJoinPool extends Abstra
2764		*/
2765		public long getQueuedTaskCount() {
2766		long count = 0;
2767	<	ForkJoinWorkerThread[] ws;
2768	<	if ((short)(ctl >>> TC_SHIFT) > -parallelism &&
2769	<	(ws = workers) != null) {
2770	<	for (ForkJoinWorkerThread w : ws)
2771	<	if (w != null)
2772	<	count -= w.queueBase - w.queueTop; // must read base first
2767	>	WorkQueue[] ws; WorkQueue w;
2768	>	if ((ws = workQueues) != null) {
2769	>	for (int i = 1; i < ws.length; i += 2) {
2770	>	if ((w = ws[i]) != null)
2771	>	count += w.queueSize();
2772	>	}
2773		}
2774		return count;
2775		}
#	Line 1758 | Line 2782 | public class ForkJoinPool extends Abstra
2782		* @return the number of queued submissions
2783		*/
2784		public int getQueuedSubmissionCount() {
2785	<	return -queueBase + queueTop;
2785	>	int count = 0;
2786	>	WorkQueue[] ws; WorkQueue w;
2787	>	if ((ws = workQueues) != null) {
2788	>	for (int i = 0; i < ws.length; i += 2) {
2789	>	if ((w = ws[i]) != null)
2790	>	count += w.queueSize();
2791	>	}
2792	>	}
2793	>	return count;
2794		}
2795
2796		/**
#	Line 1768 | Line 2800 | public class ForkJoinPool extends Abstra
2800		* @return {@code true} if there are any queued submissions
2801		*/
2802		public boolean hasQueuedSubmissions() {
2803	<	return queueBase != queueTop;
2803	>	WorkQueue[] ws; WorkQueue w;
2804	>	if ((ws = workQueues) != null) {
2805	>	for (int i = 0; i < ws.length; i += 2) {
2806	>	if ((w = ws[i]) != null && w.queueSize() != 0)
2807	>	return true;
2808	>	}
2809	>	}
2810	>	return false;
2811		}
2812
2813		/**
#	Line 1779 | Line 2818 | public class ForkJoinPool extends Abstra
2818		* @return the next submission, or {@code null} if none
2819		*/
2820		protected ForkJoinTask<?> pollSubmission() {
2821	<	ForkJoinTask<?> t; ForkJoinTask<?>[] q; int b, i;
2822	<	while ((b = queueBase) != queueTop &&
2823	<	(q = submissionQueue) != null &&
2824	<	(i = (q.length - 1) & b) >= 0) {
2825	<	long u = (i << ASHIFT) + ABASE;
1787	<	if ((t = q[i]) != null &&
1788	<	queueBase == b &&
1789	<	UNSAFE.compareAndSwapObject(q, u, t, null)) {
1790	<	queueBase = b + 1;
1791	<	return t;
2821	>	WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2822	>	if ((ws = workQueues) != null) {
2823	>	for (int i = 0; i < ws.length; i += 2) {
2824	>	if ((w = ws[i]) != null && (t = w.poll()) != null)
2825	>	return t;
2826		}
2827		}
2828		return null;
#	Line 1813 | Line 2847 | public class ForkJoinPool extends Abstra
2847		*/
2848		protected int drainTasksTo(Collection<? super ForkJoinTask<?>> c) {
2849		int count = 0;
2850	<	while (queueBase != queueTop) {
2851	<	ForkJoinTask<?> t = pollSubmission();
2852	<	if (t != null) {
2853	<	c.add(t);
2854	<	++count;
2850	>	WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2851	>	if ((ws = workQueues) != null) {
2852	>	for (int i = 0; i < ws.length; ++i) {
2853	>	if ((w = ws[i]) != null) {
2854	>	while ((t = w.poll()) != null) {
2855	>	c.add(t);
2856	>	++count;
2857	>	}
2858	>	}
2859		}
2860		}
1823	–	ForkJoinWorkerThread[] ws;
1824	–	if ((short)(ctl >>> TC_SHIFT) > -parallelism &&
1825	–	(ws = workers) != null) {
1826	–	for (ForkJoinWorkerThread w : ws)
1827	–	if (w != null)
1828	–	count += w.drainTasksTo(c);
1829	–	}
2861		return count;
2862		}
2863
#	Line 1838 | Line 2869 | public class ForkJoinPool extends Abstra
2869		* @return a string identifying this pool, as well as its state
2870		*/
2871		public String toString() {
2872	<	long st = getStealCount();
2873	<	long qt = getQueuedTaskCount();
2874	<	long qs = getQueuedSubmissionCount();
1844	<	int pc = parallelism;
2872	>	// Use a single pass through workQueues to collect counts
2873	>	long qt = 0L, qs = 0L; int rc = 0;
2874	>	long st = stealCount;
2875		long c = ctl;
2876	+	WorkQueue[] ws; WorkQueue w;
2877	+	if ((ws = workQueues) != null) {
2878	+	for (int i = 0; i < ws.length; ++i) {
2879	+	if ((w = ws[i]) != null) {
2880	+	int size = w.queueSize();
2881	+	if ((i & 1) == 0)
2882	+	qs += size;
2883	+	else {
2884	+	qt += size;
2885	+	st += w.nsteals;
2886	+	if (w.isApparentlyUnblocked())
2887	+	++rc;
2888	+	}
2889	+	}
2890	+	}
2891	+	}
2892	+	int pc = parallelism;
2893		int tc = pc + (short)(c >>> TC_SHIFT);
2894	<	int rc = pc + (int)(c >> AC_SHIFT);
2895	<	if (rc < 0) // ignore transient negative
2896	<	rc = 0;
1850	<	int ac = rc + blockedCount;
2894	>	int ac = pc + (int)(c >> AC_SHIFT);
2895	>	if (ac < 0) // ignore transient negative
2896	>	ac = 0;
2897		String level;
2898		if ((c & STOP_BIT) != 0)
2899		level = (tc == 0) ? "Terminated" : "Terminating";
2900		else
2901	<	level = shutdown ? "Shutting down" : "Running";
2901	>	level = plock < 0 ? "Shutting down" : "Running";
2902		return super.toString() +
2903		"[" + level +
2904		", parallelism = " + pc +
#	Line 1866 | Line 2912 | public class ForkJoinPool extends Abstra
2912		}
2913
2914		/**
2915	<	* Initiates an orderly shutdown in which previously submitted
2916	<	* tasks are executed, but no new tasks will be accepted.
2917	<	* Invocation has no additional effect if already shut down.
2918	<	* Tasks that are in the process of being submitted concurrently
2919	<	* during the course of this method may or may not be rejected.
2915	>	* Possibly initiates an orderly shutdown in which previously
2916	>	* submitted tasks are executed, but no new tasks will be
2917	>	* accepted. Invocation has no effect on execution state if this
2918	>	* is the {@link #commonPool}, and no additional effect if
2919	>	* already shut down.  Tasks that are in the process of being
2920	>	* submitted concurrently during the course of this method may or
2921	>	* may not be rejected.
2922		*
2923		* @throws SecurityException if a security manager exists and
2924		*         the caller is not permitted to modify threads
#	Line 1879 | Line 2927 | public class ForkJoinPool extends Abstra
2927		*/
2928		public void shutdown() {
2929		checkPermission();
2930	<	shutdown = true;
1883	<	tryTerminate(false);
2930	>	tryTerminate(false, true);
2931		}
2932
2933		/**
2934	<	* Attempts to cancel and/or stop all tasks, and reject all
2935	<	* subsequently submitted tasks.  Tasks that are in the process of
2936	<	* being submitted or executed concurrently during the course of
2937	<	* this method may or may not be rejected. This method cancels
2938	<	* both existing and unexecuted tasks, in order to permit
2939	<	* termination in the presence of task dependencies. So the method
2940	<	* always returns an empty list (unlike the case for some other
2941	<	* Executors).
2934	>	* Possibly attempts to cancel and/or stop all tasks, and reject
2935	>	* all subsequently submitted tasks.  Invocation has no effect on
2936	>	* execution state if this is the {@link #commonPool}, and no
2937	>	* additional effect if already shut down. Otherwise, tasks that
2938	>	* are in the process of being submitted or executed concurrently
2939	>	* during the course of this method may or may not be
2940	>	* rejected. This method cancels both existing and unexecuted
2941	>	* tasks, in order to permit termination in the presence of task
2942	>	* dependencies. So the method always returns an empty list
2943	>	* (unlike the case for some other Executors).
2944		*
2945		* @return an empty list
2946		* @throws SecurityException if a security manager exists and
#	Line 1901 | Line 2950 | public class ForkJoinPool extends Abstra
2950		*/
2951		public List<Runnable> shutdownNow() {
2952		checkPermission();
2953	<	shutdown = true;
1905	<	tryTerminate(true);
2953	>	tryTerminate(true, true);
2954		return Collections.emptyList();
2955		}
2956
#	Line 1937 | Line 2985 | public class ForkJoinPool extends Abstra
2985		}
2986
2987		/**
1940	–	* Returns true if terminating or terminated. Used by ForkJoinWorkerThread.
1941	–	*/
1942	–	final boolean isAtLeastTerminating() {
1943	–	return (ctl & STOP_BIT) != 0L;
1944	–	}
1945	–
1946	–	/**
2988		* Returns {@code true} if this pool has been shut down.
2989		*
2990		* @return {@code true} if this pool has been shut down
2991		*/
2992		public boolean isShutdown() {
2993	<	return shutdown;
2993	>	return plock < 0;
2994		}
2995
2996		/**
2997	<	* Blocks until all tasks have completed execution after a shutdown
2998	<	* request, or the timeout occurs, or the current thread is
2999	<	* interrupted, whichever happens first.
2997	>	* Blocks until all tasks have completed execution after a
2998	>	* shutdown request, or the timeout occurs, or the current thread
2999	>	* is interrupted, whichever happens first. Note that the {@link
3000	>	* #commonPool()} never terminates until program shutdown so
3001	>	* this method will always time out.
3002		*
3003		* @param timeout the maximum time to wait
3004		* @param unit the time unit of the timeout argument
#	Line 1966 | Line 3009 | public class ForkJoinPool extends Abstra
3009		public boolean awaitTermination(long timeout, TimeUnit unit)
3010		throws InterruptedException {
3011		long nanos = unit.toNanos(timeout);
3012	<	final ReentrantLock lock = this.submissionLock;
3013	<	lock.lock();
3014	<	try {
3015	<	for (;;) {
3016	<	if (isTerminated())
3017	<	return true;
3018	<	if (nanos <= 0)
3019	<	return false;
3020	<	nanos = termination.awaitNanos(nanos);
3012	>	if (isTerminated())
3013	>	return true;
3014	>	long startTime = System.nanoTime();
3015	>	boolean terminated = false;
3016	>	synchronized (this) {
3017	>	for (long waitTime = nanos, millis = 0L;;) {
3018	>	if (terminated = isTerminated() ||
3019	>	waitTime <= 0L ||
3020	>	(millis = unit.toMillis(waitTime)) <= 0L)
3021	>	break;
3022	>	wait(millis);
3023	>	waitTime = nanos - (System.nanoTime() - startTime);
3024		}
1979	–	} finally {
1980	–	lock.unlock();
3025		}
3026	+	return terminated;
3027		}
3028
3029		/**
#	Line 2078 | Line 3123 | public class ForkJoinPool extends Abstra
3123		throws InterruptedException {
3124		Thread t = Thread.currentThread();
3125		if (t instanceof ForkJoinWorkerThread) {
3126	<	ForkJoinWorkerThread w = (ForkJoinWorkerThread) t;
3127	<	w.pool.awaitBlocker(blocker);
3126	>	ForkJoinPool p = ((ForkJoinWorkerThread)t).pool;
3127	>	while (!blocker.isReleasable()) { // variant of helpSignal
3128	>	WorkQueue[] ws; WorkQueue q; int m, n;
3129	>	if ((ws = p.workQueues) != null && (m = ws.length - 1) >= 0) {
3130	>	for (int i = 0; i <= m; ++i) {
3131	>	if (blocker.isReleasable())
3132	>	return;
3133	>	if ((q = ws[i]) != null && (n = q.queueSize()) > 0) {
3134	>	p.signalWork(q, n);
3135	>	if ((int)(p.ctl >> AC_SHIFT) >= 0)
3136	>	break;
3137	>	}
3138	>	}
3139	>	}
3140	>	if (p.tryCompensate()) {
3141	>	try {
3142	>	do {} while (!blocker.isReleasable() &&
3143	>	!blocker.block());
3144	>	} finally {
3145	>	p.incrementActiveCount();
3146	>	}
3147	>	break;
3148	>	}
3149	>	}
3150		}
3151		else {
3152	<	do {} while (!blocker.isReleasable() && !blocker.block());
3152	>	do {} while (!blocker.isReleasable() &&
3153	>	!blocker.block());
3154		}
3155		}
3156
#	Line 2091 | Line 3159 | public class ForkJoinPool extends Abstra
3159		// implement RunnableFuture.
3160
3161		protected <T> RunnableFuture<T> newTaskFor(Runnable runnable, T value) {
3162	<	return (RunnableFuture<T>) ForkJoinTask.adapt(runnable, value);
3162	>	return new ForkJoinTask.AdaptedRunnable<T>(runnable, value);
3163		}
3164
3165		protected <T> RunnableFuture<T> newTaskFor(Callable<T> callable) {
3166	<	return (RunnableFuture<T>) ForkJoinTask.adapt(callable);
3166	>	return new ForkJoinTask.AdaptedCallable<T>(callable);
3167		}
3168
3169		// Unsafe mechanics
3170	<	private static final sun.misc.Unsafe UNSAFE;
3171	<	private static final long ctlOffset;
3172	<	private static final long stealCountOffset;
3173	<	private static final long blockedCountOffset;
2106	<	private static final long quiescerCountOffset;
2107	<	private static final long scanGuardOffset;
2108	<	private static final long nextWorkerNumberOffset;
2109	<	private static final long ABASE;
3170	>	private static final sun.misc.Unsafe U;
3171	>	private static final long CTL;
3172	>	private static final long PARKBLOCKER;
3173	>	private static final int ABASE;
3174		private static final int ASHIFT;
3175	+	private static final long STEALCOUNT;
3176	+	private static final long PLOCK;
3177	+	private static final long INDEXSEED;
3178	+	private static final long QLOCK;
3179
3180		static {
3181	<	poolNumberGenerator = new AtomicInteger();
3182	<	workerSeedGenerator = new Random();
3183	<	modifyThreadPermission = new RuntimePermission("modifyThread");
3184	<	defaultForkJoinWorkerThreadFactory =
3185	<	new DefaultForkJoinWorkerThreadFactory();
3181	>	// Establish common pool parameters
3182	>	// TBD: limit or report ignored exceptions?
3183	>
3184	>	int par = 0;
3185	>	ForkJoinWorkerThreadFactory fac = null;
3186	>	Thread.UncaughtExceptionHandler handler = null;
3187	>	try {
3188	>	String pp = System.getProperty(propPrefix + "parallelism");
3189	>	String hp = System.getProperty(propPrefix + "exceptionHandler");
3190	>	String fp = System.getProperty(propPrefix + "threadFactory");
3191	>	if (fp != null)
3192	>	fac = ((ForkJoinWorkerThreadFactory)ClassLoader.
3193	>	getSystemClassLoader().loadClass(fp).newInstance());
3194	>	if (hp != null)
3195	>	handler = ((Thread.UncaughtExceptionHandler)ClassLoader.
3196	>	getSystemClassLoader().loadClass(hp).newInstance());
3197	>	if (pp != null)
3198	>	par = Integer.parseInt(pp);
3199	>	} catch (Exception ignore) {
3200	>	}
3201	>
3202	>	int s; // initialize field offsets for CAS etc
3203		try {
3204	<	UNSAFE = getUnsafe();
3204	>	U = getUnsafe();
3205		Class<?> k = ForkJoinPool.class;
3206	<	ctlOffset = UNSAFE.objectFieldOffset
3206	>	CTL = U.objectFieldOffset
3207		(k.getDeclaredField("ctl"));
3208	<	stealCountOffset = UNSAFE.objectFieldOffset
3208	>	STEALCOUNT = U.objectFieldOffset
3209		(k.getDeclaredField("stealCount"));
3210	<	blockedCountOffset = UNSAFE.objectFieldOffset
3211	<	(k.getDeclaredField("blockedCount"));
3212	<	quiescerCountOffset = UNSAFE.objectFieldOffset
3213	<	(k.getDeclaredField("quiescerCount"));
3214	<	scanGuardOffset = UNSAFE.objectFieldOffset
3215	<	(k.getDeclaredField("scanGuard"));
3216	<	nextWorkerNumberOffset = UNSAFE.objectFieldOffset
3217	<	(k.getDeclaredField("nextWorkerNumber"));
3210	>	PLOCK = U.objectFieldOffset
3211	>	(k.getDeclaredField("plock"));
3212	>	INDEXSEED = U.objectFieldOffset
3213	>	(k.getDeclaredField("indexSeed"));
3214	>	Class<?> tk = Thread.class;
3215	>	PARKBLOCKER = U.objectFieldOffset
3216	>	(tk.getDeclaredField("parkBlocker"));
3217	>	Class<?> wk = WorkQueue.class;
3218	>	QLOCK = U.objectFieldOffset
3219	>	(wk.getDeclaredField("qlock"));
3220	>	Class<?> ak = ForkJoinTask[].class;
3221	>	ABASE = U.arrayBaseOffset(ak);
3222	>	s = U.arrayIndexScale(ak);
3223	>	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
3224		} catch (Exception e) {
3225		throw new Error(e);
3226		}
2136	–	Class<?> a = ForkJoinTask[].class;
2137	–	ABASE = UNSAFE.arrayBaseOffset(a);
2138	–	int s = UNSAFE.arrayIndexScale(a);
3227		if ((s & (s-1)) != 0)
3228		throw new Error("data type scale not a power of two");
3229	<	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
3229	>
3230	>	/*
3231	>	* For extra caution, computations to set up pool state are
3232	>	* here; the constructor just assigns these values to fields.
3233	>	*/
3234	>	ForkJoinWorkerThreadFactory defaultFac =
3235	>	defaultForkJoinWorkerThreadFactory =
3236	>	new DefaultForkJoinWorkerThreadFactory();
3237	>	if (fac == null)
3238	>	fac = defaultFac;
3239	>	if (par <= 0)
3240	>	par = Runtime.getRuntime().availableProcessors();
3241	>	if (par > MAX_CAP)
3242	>	par = MAX_CAP;
3243	>	long np = (long)(-par); // precompute initial ctl value
3244	>	long ct = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK);
3245	>
3246	>	commonPoolParallelism = par;
3247	>	commonPool = new ForkJoinPool(par, ct, fac, handler);
3248	>	modifyThreadPermission = new RuntimePermission("modifyThread");
3249	>	submitters = new ThreadLocal<Submitter>();
3250		}
3251
3252		/**
#	Line 2168 | Line 3276 | public class ForkJoinPool extends Abstra
3276		}
3277		}
3278		}
3279	+
3280		}

Diff Legend

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