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 1.102 by jsr166, Thu Apr 14 01:17:58 2011 UTC vs.
Revision 1.141 by jsr166, Wed Nov 14 18:45:53 2012 UTC

#	Line 11 | Line 11 | import java.util.Arrays;
11		import java.util.Collection;
12		import java.util.Collections;
13		import java.util.List;
14	–	import java.util.Random;
14		import java.util.concurrent.AbstractExecutorService;
15		import java.util.concurrent.Callable;
16		import java.util.concurrent.ExecutorService;
#	Line 19 | Line 18 | import java.util.concurrent.Future;
18		import java.util.concurrent.RejectedExecutionException;
19		import java.util.concurrent.RunnableFuture;
20		import java.util.concurrent.TimeUnit;
22	–	import java.util.concurrent.TimeoutException;
23	–	import java.util.concurrent.atomic.AtomicInteger;
24	–	import java.util.concurrent.locks.LockSupport;
25	–	import java.util.concurrent.locks.ReentrantLock;
26	–	import java.util.concurrent.locks.Condition;
21
22		/**
23		* An {@link ExecutorService} for running {@link ForkJoinTask}s.
#	Line 34 | Line 28 | import java.util.concurrent.locks.Condit
28		* <p>A {@code ForkJoinPool} differs from other kinds of {@link
29		* ExecutorService} mainly by virtue of employing
30		* <em>work-stealing</em>: all threads in the pool attempt to find and
31	<	* execute subtasks created by other active tasks (eventually blocking
32	<	* waiting for work if none exist). This enables efficient processing
33	<	* when most tasks spawn other subtasks (as do most {@code
34	<	* ForkJoinTask}s). When setting <em>asyncMode</em> to true in
35	<	* constructors, {@code ForkJoinPool}s may also be appropriate for use
36	<	* with event-style tasks that are never joined.
31	>	* execute tasks submitted to the pool and/or created by other active
32	>	* tasks (eventually blocking waiting for work if none exist). This
33	>	* enables efficient processing when most tasks spawn other subtasks
34	>	* (as do most {@code ForkJoinTask}s), as well as when many small
35	>	* tasks are submitted to the pool from external clients.  Especially
36	>	* when setting <em>asyncMode</em> to true in constructors, {@code
37	>	* ForkJoinPool}s may also be appropriate for use with event-style
38	>	* tasks that are never joined.
39		*
40	<	* <p>A {@code ForkJoinPool} is constructed with a given target
41	<	* parallelism level; by default, equal to the number of available
42	<	* processors. The pool attempts to maintain enough active (or
43	<	* available) threads by dynamically adding, suspending, or resuming
44	<	* internal worker threads, even if some tasks are stalled waiting to
45	<	* join others. However, no such adjustments are guaranteed in the
46	<	* face of blocked IO or other unmanaged synchronization. The nested
47	<	* {@link ManagedBlocker} interface enables extension of the kinds of
40	>	* <p>A static {@link #commonPool} is available and appropriate for
41	>	* most applications. The common pool is used by any ForkJoinTask that
42	>	* is not explicitly submitted to a specified pool. Using the common
43	>	* pool normally reduces resource usage (its threads are slowly
44	>	* reclaimed during periods of non-use, and reinstated upon subsequent
45	>	* use).
46	>	*
47	>	* <p>For applications that require separate or custom pools, a {@code
48	>	* ForkJoinPool} may be constructed with a given target parallelism
49	>	* level; by default, equal to the number of available processors. The
50	>	* pool attempts to maintain enough active (or available) threads by
51	>	* dynamically adding, suspending, or resuming internal worker
52	>	* threads, even if some tasks are stalled waiting to join
53	>	* others. However, no such adjustments are guaranteed in the face of
54	>	* blocked 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 59 | Line 63 | import java.util.concurrent.locks.Condit
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
66	>	* main task execution methods summarized in the following table.
67	>	* These are designed to be used primarily by clients not already
68	>	* engaged in fork/join computations in the current pool.  The main
69	>	* forms of these methods accept instances of {@code ForkJoinTask},
70	>	* but overloaded forms also allow mixed execution of plain {@code
71		* Runnable}- or {@code Callable}- based activities as well.  However,
72	<	* tasks that are already executing in a pool should normally
73	<	* <em>NOT</em> use these pool execution methods, but instead use the
74	<	* within-computation forms listed in the table.
72	>	* tasks that are already executing in a pool should normally instead
73	>	* use the within-computation forms listed in the table unless using
74	>	* async event-style tasks that are not usually joined, in which case
75	>	* there is little difference among choice of methods.
76		*
77		* <table BORDER CELLPADDING=3 CELLSPACING=1>
78		*  <tr>
#	Line 92 | Line 97 | import java.util.concurrent.locks.Condit
97		*  </tr>
98		* </table>
99		*
100	<	* <p><b>Sample Usage.</b> Normally a single {@code ForkJoinPool} is
101	<	* used for all parallel task execution in a program or subsystem.
102	<	* Otherwise, use would not usually outweigh the construction and
103	<	* bookkeeping overhead of creating a large set of threads. For
104	<	* example, a common pool could be used for the {@code SortTasks}
105	<	* illustrated in {@link RecursiveAction}. Because {@code
106	<	* ForkJoinPool} uses threads in {@linkplain java.lang.Thread#isDaemon
107	<	* daemon} mode, there is typically no need to explicitly {@link
108	<	* #shutdown} such a pool upon program exit.
104	<	*
105	<	* <pre>
106	<	* static final ForkJoinPool mainPool = new ForkJoinPool();
107	<	* ...
108	<	* public void sort(long[] array) {
109	<	*   mainPool.invoke(new SortTask(array, 0, array.length));
110	<	* }
111	<	* </pre>
100	>	* <p>The common pool is by default constructed with default
101	>	* parameters, but these may be controlled by setting three {@link
102	>	* System#getProperty properties} with prefix {@code
103	>	* java.util.concurrent.ForkJoinPool.common}: {@code parallelism} --
104	>	* an integer greater than zero, {@code threadFactory} -- the class
105	>	* name of a {@link ForkJoinWorkerThreadFactory}, and {@code
106	>	* exceptionHandler} -- the class name of a {@link
107	>	* Thread.UncaughtExceptionHandler}. Upon any error in establishing
108	>	* these settings, default parameters are used.
109		*
110		* <p><b>Implementation notes</b>: This implementation restricts the
111		* maximum number of running threads to 32767. Attempts to create
#	Line 127 | Line 124 | public class ForkJoinPool extends Abstra
124		/*
125		* Implementation Overview
126		*
127	<	* This class provides the central bookkeeping and control for a
128	<	* set of worker threads: Submissions from non-FJ threads enter
129	<	* into a submission queue. Workers take these tasks and typically
130	<	* split them into subtasks that may be stolen by other workers.
131	<	* Preference rules give first priority to processing tasks from
132	<	* their own queues (LIFO or FIFO, depending on mode), then to
133	<	* randomized FIFO steals of tasks in other worker queues, and
134	<	* lastly to new submissions.
127	>	* This class and its nested classes provide the main
128	>	* functionality and control for a set of worker threads:
129	>	* Submissions from non-FJ threads enter into submission queues.
130	>	* Workers take these tasks and typically split them into subtasks
131	>	* that may be stolen by other workers.  Preference rules give
132	>	* first priority to processing tasks from their own queues (LIFO
133	>	* or FIFO, depending on mode), then to randomized FIFO steals of
134	>	* tasks in other queues.
135	>	*
136	>	* WorkQueues
137	>	* ==========
138	>	*
139	>	* Most operations occur within work-stealing queues (in nested
140	>	* class WorkQueue).  These are special forms of Deques that
141	>	* support only three of the four possible end-operations -- push,
142	>	* pop, and poll (aka steal), under the further constraints that
143	>	* push and pop are called only from the owning thread (or, as
144	>	* extended here, under a lock), while poll may be called from
145	>	* other threads.  (If you are unfamiliar with them, you probably
146	>	* want to read Herlihy and Shavit's book "The Art of
147	>	* Multiprocessor programming", chapter 16 describing these in
148	>	* more detail before proceeding.)  The main work-stealing queue
149	>	* design is roughly similar to those in the papers "Dynamic
150	>	* Circular Work-Stealing Deque" by Chase and Lev, SPAA 2005
151	>	* (http://research.sun.com/scalable/pubs/index.html) and
152	>	* "Idempotent work stealing" by Michael, Saraswat, and Vechev,
153	>	* PPoPP 2009 (http://portal.acm.org/citation.cfm?id=1504186).
154	>	* The main differences ultimately stem from GC requirements that
155	>	* we null out taken slots as soon as we can, to maintain as small
156	>	* a footprint as possible even in programs generating huge
157	>	* numbers of tasks. To accomplish this, we shift the CAS
158	>	* arbitrating pop vs poll (steal) from being on the indices
159	>	* ("base" and "top") to the slots themselves.  So, both a
160	>	* successful pop and poll mainly entail a CAS of a slot from
161	>	* non-null to null.  Because we rely on CASes of references, we
162	>	* do not need tag bits on base or top.  They are simple ints as
163	>	* used in any circular array-based queue (see for example
164	>	* ArrayDeque).  Updates to the indices must still be ordered in a
165	>	* way that guarantees that top == base means the queue is empty,
166	>	* but otherwise may err on the side of possibly making the queue
167	>	* appear nonempty when a push, pop, or poll have not fully
168	>	* committed. Note that this means that the poll operation,
169	>	* considered individually, is not wait-free. One thief cannot
170	>	* successfully continue until another in-progress one (or, if
171	>	* previously empty, a push) completes.  However, in the
172	>	* aggregate, we ensure at least probabilistic non-blockingness.
173	>	* If an attempted steal fails, a thief always chooses a different
174	>	* random victim target to try next. So, in order for one thief to
175	>	* progress, it suffices for any in-progress poll or new push on
176	>	* any empty queue to complete. (This is why we normally use
177	>	* method pollAt and its variants that try once at the apparent
178	>	* base index, else consider alternative actions, rather than
179	>	* method poll.)
180	>	*
181	>	* This approach also enables support of a user mode in which local
182	>	* task processing is in FIFO, not LIFO order, simply by using
183	>	* poll rather than pop.  This can be useful in message-passing
184	>	* frameworks in which tasks are never joined.  However neither
185	>	* mode considers affinities, loads, cache localities, etc, so
186	>	* rarely provide the best possible performance on a given
187	>	* machine, but portably provide good throughput by averaging over
188	>	* these factors.  (Further, even if we did try to use such
189	>	* information, we do not usually have a basis for exploiting it.
190	>	* For example, some sets of tasks profit from cache affinities,
191	>	* but others are harmed by cache pollution effects.)
192	>	*
193	>	* WorkQueues are also used in a similar way for tasks submitted
194	>	* to the pool. We cannot mix these tasks in the same queues used
195	>	* for work-stealing (this would contaminate lifo/fifo
196	>	* processing). Instead, we randomly associate submission queues
197	>	* with submitting threads, using a form of hashing.  The
198	>	* ThreadLocal Submitter class contains a value initially used as
199	>	* a hash code for choosing existing queues, but may be randomly
200	>	* repositioned upon contention with other submitters.  In
201	>	* essence, submitters act like workers except that they are
202	>	* restricted to executing local tasks that they submitted (or in
203	>	* the case of CountedCompleters, others with the same root task).
204	>	* However, because most shared/external queue operations are more
205	>	* expensive than internal, and because, at steady state, external
206	>	* submitters will compete for CPU with workers, ForkJoinTask.join
207	>	* and related methods disable them from repeatedly helping to
208	>	* process tasks if all workers are active.  Insertion of tasks in
209	>	* shared mode requires a lock (mainly to protect in the case of
210	>	* resizing) but we use only a simple spinlock (using bits in
211	>	* field qlock), because submitters encountering a busy queue move
212	>	* on to try or create other queues -- they block only when
213	>	* creating and registering new queues.
214	>	*
215	>	* Management
216	>	* ==========
217		*
218		* The main throughput advantages of work-stealing stem from
219		* decentralized control -- workers mostly take tasks from
220		* themselves or each other. We cannot negate this in the
221		* implementation of other management responsibilities. The main
222		* tactic for avoiding bottlenecks is packing nearly all
223	<	* essentially atomic control state into a single 64bit volatile
224	<	* variable ("ctl"). This variable is read on the order of 10-100
225	<	* times as often as it is modified (always via CAS). (There is
226	<	* some additional control state, for example variable "shutdown"
227	<	* for which we can cope with uncoordinated updates.)  This
228	<	* streamlines synchronization and control at the expense of messy
229	<	* constructions needed to repack status bits upon updates.
230	<	* Updates tend not to contend with each other except during
231	<	* bursts while submitted tasks begin or end.  In some cases when
232	<	* they do contend, threads can instead do something else
233	<	* (usually, scan for tasks) until contention subsides.
234	<	*
235	<	* To enable packing, we restrict maximum parallelism to (1<<15)-1
236	<	* (which is far in excess of normal operating range) to allow
237	<	* ids, counts, and their negations (used for thresholding) to fit
238	<	* into 16bit fields.
239	<	*
240	<	* Recording Workers.  Workers are recorded in the "workers" array
241	<	* that is created upon pool construction and expanded if (rarely)
242	<	* necessary.  This is an array as opposed to some other data
243	<	* structure to support index-based random steals by workers.
244	<	* Updates to the array recording new workers and unrecording
245	<	* terminated ones are protected from each other by a seqLock
246	<	* (scanGuard) but the array is otherwise concurrently readable,
247	<	* and accessed directly by workers. To simplify index-based
248	<	* operations, the array size is always a power of two, and all
249	<	* readers must tolerate null slots. To avoid flailing during
250	<	* start-up, the array is presized to hold twice #parallelism
251	<	* workers (which is unlikely to need further resizing during
252	<	* execution). But to avoid dealing with so many null slots,
253	<	* variable scanGuard includes a mask for the nearest power of two
254	<	* that contains all current workers.  All worker thread creation
255	<	* is on-demand, triggered by task submissions, replacement of
256	<	* terminated workers, and/or compensation for blocked
257	<	* workers. However, all other support code is set up to work with
258	<	* other policies.  To ensure that we do not hold on to worker
259	<	* references that would prevent GC, ALL accesses to workers are
260	<	* via indices into the workers array (which is one source of some
261	<	* of the messy code constructions here). In essence, the workers
262	<	* array serves as a weak reference mechanism. Thus for example
263	<	* the wait queue field of ctl stores worker indices, not worker
264	<	* references.  Access to the workers in associated methods (for
265	<	* example signalWork) must both index-check and null-check the
266	<	* IDs. All such accesses ignore bad IDs by returning out early
267	<	* from what they are doing, since this can only be associated
268	<	* with termination, in which case it is OK to give up.
269	<	*
270	<	* All uses of the workers array, as well as queue arrays, check
271	<	* that the array is non-null (even if previously non-null). This
272	<	* allows nulling during termination, which is currently not
273	<	* necessary, but remains an option for resource-revocation-based
274	<	* shutdown schemes.
223	>	* essentially atomic control state into two volatile variables
224	>	* that are by far most often read (not written) as status and
225	>	* consistency checks.
226	>	*
227	>	* Field "ctl" contains 64 bits holding all the information needed
228	>	* to atomically decide to add, inactivate, enqueue (on an event
229	>	* queue), dequeue, and/or re-activate workers.  To enable this
230	>	* packing, we restrict maximum parallelism to (1<<15)-1 (which is
231	>	* far in excess of normal operating range) to allow ids, counts,
232	>	* and their negations (used for thresholding) to fit into 16bit
233	>	* fields.
234	>	*
235	>	* Field "plock" is a form of sequence lock with a saturating
236	>	* shutdown bit (similarly for per-queue "qlocks"), mainly
237	>	* protecting updates to the workQueues array, as well as to
238	>	* enable shutdown.  When used as a lock, it is normally only very
239	>	* briefly held, so is nearly always available after at most a
240	>	* brief spin, but we use a monitor-based backup strategy to
241	>	* blocking when needed.
242	>	*
243	>	* Recording WorkQueues.  WorkQueues are recorded in the
244	>	* "workQueues" array that is created upon first use and expanded
245	>	* if necessary.  Updates to the array while recording new workers
246	>	* and unrecording terminated ones are protected from each other
247	>	* by a lock but the array is otherwise concurrently readable, and
248	>	* accessed directly.  To simplify index-based operations, the
249	>	* array size is always a power of two, and all readers must
250	>	* tolerate null slots. Worker queues are at odd indices Shared
251	>	* (submission) queues are at even indices, up to a maximum of 64
252	>	* slots, to limit growth even if array needs to expand to add
253	>	* more workers. Grouping them together in this way simplifies and
254	>	* speeds up task scanning.
255	>	*
256	>	* All worker thread creation is on-demand, triggered by task
257	>	* submissions, replacement of terminated workers, and/or
258	>	* compensation for blocked workers. However, all other support
259	>	* code is set up to work with other policies.  To ensure that we
260	>	* do not hold on to worker references that would prevent GC, ALL
261	>	* accesses to workQueues are via indices into the workQueues
262	>	* array (which is one source of some of the messy code
263	>	* constructions here). In essence, the workQueues array serves as
264	>	* a weak reference mechanism. Thus for example the wait queue
265	>	* field of ctl stores indices, not references.  Access to the
266	>	* workQueues in associated methods (for example signalWork) must
267	>	* both index-check and null-check the IDs. All such accesses
268	>	* ignore bad IDs by returning out early from what they are doing,
269	>	* since this can only be associated with termination, in which
270	>	* case it is OK to give up.  All uses of the workQueues array
271	>	* also check that it is non-null (even if previously
272	>	* non-null). This allows nulling during termination, which is
273	>	* currently not necessary, but remains an option for
274	>	* resource-revocation-based shutdown schemes. It also helps
275	>	* reduce JIT issuance of uncommon-trap code, which tends to
276	>	* unnecessarily complicate control flow in some methods.
277		*
278	<	* Wait Queuing. Unlike HPC work-stealing frameworks, we cannot
278	>	* Event Queuing. Unlike HPC work-stealing frameworks, we cannot
279		* let workers spin indefinitely scanning for tasks when none can
280		* be found immediately, and we cannot start/resume workers unless
281		* there appear to be tasks available.  On the other hand, we must
282		* quickly prod them into action when new tasks are submitted or
283	<	* generated.  We park/unpark workers after placing in an event
284	<	* wait queue when they cannot find work. This "queue" is actually
285	<	* a simple Treiber stack, headed by the "id" field of ctl, plus a
286	<	* 15bit counter value to both wake up waiters (by advancing their
287	<	* count) and avoid ABA effects. Successors are held in worker
288	<	* field "nextWait".  Queuing deals with several intrinsic races,
289	<	* mainly that a task-producing thread can miss seeing (and
283	>	* generated. In many usages, ramp-up time to activate workers is
284	>	* the main limiting factor in overall performance (this is
285	>	* compounded at program start-up by JIT compilation and
286	>	* allocation). So we try to streamline this as much as possible.
287	>	* We park/unpark workers after placing in an event wait queue
288	>	* when they cannot find work. This "queue" is actually a simple
289	>	* Treiber stack, headed by the "id" field of ctl, plus a 15bit
290	>	* counter value (that reflects the number of times a worker has
291	>	* been inactivated) to avoid ABA effects (we need only as many
292	>	* version numbers as worker threads). Successors are held in
293	>	* field WorkQueue.nextWait.  Queuing deals with several intrinsic
294	>	* races, mainly that a task-producing thread can miss seeing (and
295		* signalling) another thread that gave up looking for work but
296		* has not yet entered the wait queue. We solve this by requiring
297	<	* a full sweep of all workers both before (in scan()) and after
298	<	* (in tryAwaitWork()) a newly waiting worker is added to the wait
299	<	* queue. During a rescan, the worker might release some other
300	<	* queued worker rather than itself, which has the same net
301	<	* effect. Because enqueued workers may actually be rescanning
302	<	* rather than waiting, we set and clear the "parked" field of
303	<	* ForkJoinWorkerThread to reduce unnecessary calls to unpark.
304	<	* (Use of the parked field requires a secondary recheck to avoid
305	<	* missed signals.)
297	>	* a full sweep of all workers (via repeated calls to method
298	>	* scan()) both before and after a newly waiting worker is added
299	>	* to the wait queue. During a rescan, the worker might release
300	>	* some other queued worker rather than itself, which has the same
301	>	* net effect. Because enqueued workers may actually be rescanning
302	>	* rather than waiting, we set and clear the "parker" field of
303	>	* WorkQueues to reduce unnecessary calls to unpark.  (This
304	>	* requires a secondary recheck to avoid missed signals.)  Note
305	>	* the unusual conventions about Thread.interrupts surrounding
306	>	* parking and other blocking: Because interrupts are used solely
307	>	* to alert threads to check termination, which is checked anyway
308	>	* upon blocking, we clear status (using Thread.interrupted)
309	>	* before any call to park, so that park does not immediately
310	>	* return due to status being set via some other unrelated call to
311	>	* interrupt in user code.
312		*
313		* Signalling.  We create or wake up workers only when there
314		* appears to be at least one task they might be able to find and
315	<	* execute.  When a submission is added or another worker adds a
316	<	* task to a queue that previously had two or fewer tasks, they
317	<	* signal waiting workers (or trigger creation of new ones if
318	<	* fewer than the given parallelism level -- see signalWork).
319	<	* These primary signals are buttressed by signals during rescans
320	<	* as well as those performed when a worker steals a task and
321	<	* notices that there are more tasks too; together these cover the
322	<	* signals needed in cases when more than two tasks are pushed
323	<	* but untaken.
315	>	* execute. However, many other threads may notice the same task
316	>	* and each signal to wake up a thread that might take it. So in
317	>	* general, pools will be over-signalled.  When a submission is
318	>	* added or another worker adds a task to a queue that is
319	>	* apparently empty, they signal waiting workers (or trigger
320	>	* creation of new ones if fewer than the given parallelism level
321	>	* -- see signalWork).  These primary signals are buttressed by
322	>	* signals whenever other threads scan for work or do not have a
323	>	* task to process. On most platforms, signalling (unpark)
324	>	* overhead time is noticeably long, and the time between
325	>	* signalling a thread and it actually making progress can be very
326	>	* noticeably long, so it is worth offloading these delays from
327	>	* critical paths as much as possible.
328		*
329		* Trimming workers. To release resources after periods of lack of
330		* use, a worker starting to wait when the pool is quiescent will
331	<	* time out and terminate if the pool has remained quiescent for
332	<	* SHRINK_RATE nanosecs. This will slowly propagate, eventually
333	<	* terminating all workers after long periods of non-use.
334	<	*
335	<	* Submissions. External submissions are maintained in an
336	<	* array-based queue that is structured identically to
337	<	* ForkJoinWorkerThread queues except for the use of
338	<	* submissionLock in method addSubmission. Unlike the case for
339	<	* worker queues, multiple external threads can add new
340	<	* submissions, so adding requires a lock.
341	<	*
342	<	* Compensation. Beyond work-stealing support and lifecycle
343	<	* control, the main responsibility of this framework is to take
344	<	* actions when one worker is waiting to join a task stolen (or
345	<	* always held by) another.  Because we are multiplexing many
346	<	* tasks on to a pool of workers, we can't just let them block (as
347	<	* in Thread.join).  We also cannot just reassign the joiner's
348	<	* run-time stack with another and replace it later, which would
349	<	* be a form of "continuation", that even if possible is not
350	<	* necessarily a good idea since we sometimes need both an
351	<	* unblocked task and its continuation to progress. Instead we
352	<	* combine two tactics:
331	>	* time out and terminate if the pool has remained quiescent for a
332	>	* given period -- a short period if there are more threads than
333	>	* parallelism, longer as the number of threads decreases. This
334	>	* will slowly propagate, eventually terminating all workers after
335	>	* periods of non-use.
336	>	*
337	>	* Shutdown and Termination. A call to shutdownNow atomically sets
338	>	* a plock bit and then (non-atomically) sets each worker's
339	>	* qlock status, cancels all unprocessed tasks, and wakes up
340	>	* all waiting workers.  Detecting whether termination should
341	>	* commence after a non-abrupt shutdown() call requires more work
342	>	* and bookkeeping. We need consensus about quiescence (i.e., that
343	>	* there is no more work). The active count provides a primary
344	>	* indication but non-abrupt shutdown still requires a rechecking
345	>	* scan for any workers that are inactive but not queued.
346	>	*
347	>	* Joining Tasks
348	>	* =============
349	>	*
350	>	* Any of several actions may be taken when one worker is waiting
351	>	* to join a task stolen (or always held) by another.  Because we
352	>	* are multiplexing many tasks on to a pool of workers, we can't
353	>	* just let them block (as in Thread.join).  We also cannot just
354	>	* reassign the joiner's run-time stack with another and replace
355	>	* it later, which would be a form of "continuation", that even if
356	>	* possible is not necessarily a good idea since we sometimes need
357	>	* both an unblocked task and its continuation to progress.
358	>	* Instead we combine two tactics:
359		*
360		*   Helping: Arranging for the joiner to execute some task that it
361	<	*      would be running if the steal had not occurred.  Method
260	<	*      ForkJoinWorkerThread.joinTask tracks joining->stealing
261	<	*      links to try to find such a task.
361	>	*      would be running if the steal had not occurred.
362		*
363		*   Compensating: Unless there are already enough live threads,
364	<	*      method tryPreBlock() may create or re-activate a spare
365	<	*      thread to compensate for blocked joiners until they
366	<	*      unblock.
364	>	*      method tryCompensate() may create or re-activate a spare
365	>	*      thread to compensate for blocked joiners until they unblock.
366	>	*
367	>	* A third form (implemented in tryRemoveAndExec) amounts to
368	>	* helping a hypothetical compensator: If we can readily tell that
369	>	* a possible action of a compensator is to steal and execute the
370	>	* task being joined, the joining thread can do so directly,
371	>	* without the need for a compensation thread (although at the
372	>	* expense of larger run-time stacks, but the tradeoff is
373	>	* typically worthwhile).
374		*
375		* The ManagedBlocker extension API can't use helping so relies
376		* only on compensation in method awaitBlocker.
377		*
378	+	* The algorithm in tryHelpStealer entails a form of "linear"
379	+	* helping: Each worker records (in field currentSteal) the most
380	+	* recent task it stole from some other worker. Plus, it records
381	+	* (in field currentJoin) the task it is currently actively
382	+	* joining. Method tryHelpStealer uses these markers to try to
383	+	* find a worker to help (i.e., steal back a task from and execute
384	+	* it) that could hasten completion of the actively joined task.
385	+	* In essence, the joiner executes a task that would be on its own
386	+	* local deque had the to-be-joined task not been stolen. This may
387	+	* be seen as a conservative variant of the approach in Wagner &
388	+	* Calder "Leapfrogging: a portable technique for implementing
389	+	* efficient futures" SIGPLAN Notices, 1993
390	+	* (http://portal.acm.org/citation.cfm?id=155354). It differs in
391	+	* that: (1) We only maintain dependency links across workers upon
392	+	* steals, rather than use per-task bookkeeping.  This sometimes
393	+	* requires a linear scan of workQueues array to locate stealers,
394	+	* but often doesn't because stealers leave hints (that may become
395	+	* stale/wrong) of where to locate them.  A stealHint is only a
396	+	* hint because a worker might have had multiple steals and the
397	+	* hint records only one of them (usually the most current).
398	+	* Hinting isolates cost to when it is needed, rather than adding
399	+	* to per-task overhead.  (2) It is "shallow", ignoring nesting
400	+	* and potentially cyclic mutual steals.  (3) It is intentionally
401	+	* racy: field currentJoin is updated only while actively joining,
402	+	* which means that we miss links in the chain during long-lived
403	+	* tasks, GC stalls etc (which is OK since blocking in such cases
404	+	* is usually a good idea).  (4) We bound the number of attempts
405	+	* to find work (see MAX_HELP) and fall back to suspending the
406	+	* worker and if necessary replacing it with another.
407	+	*
408	+	* Helping actions for CountedCompleters are much simpler: Method
409	+	* helpComplete can take and execute any task with the same root
410	+	* as the task being waited on. However, this still entails some
411	+	* traversal of completer chains, so is less efficient than using
412	+	* CountedCompleters without explicit joins.
413	+	*
414		* It is impossible to keep exactly the target parallelism number
415		* of threads running at any given time.  Determining the
416		* existence of conservatively safe helping targets, the
417		* availability of already-created spares, and the apparent need
418	<	* to create new spares are all racy and require heuristic
419	<	* guidance, so we rely on multiple retries of each.  Currently,
420	<	* in keeping with on-demand signalling policy, we compensate only
421	<	* if blocking would leave less than one active (non-waiting,
422	<	* non-blocked) worker. Additionally, to avoid some false alarms
423	<	* due to GC, lagging counters, system activity, etc, compensated
424	<	* blocking for joins is only attempted after rechecks stabilize
425	<	* (retries are interspersed with Thread.yield, for good
426	<	* citizenship).  The variable blockedCount, incremented before
427	<	* blocking and decremented after, is sometimes needed to
428	<	* distinguish cases of waiting for work vs blocking on joins or
429	<	* other managed sync. Both cases are equivalent for most pool
430	<	* control, so we can update non-atomically. (Additionally,
431	<	* contention on blockedCount alleviates some contention on ctl).
432	<	*
433	<	* Shutdown and Termination. A call to shutdownNow atomically sets
434	<	* the ctl stop bit and then (non-atomically) sets each workers
435	<	* "terminate" status, cancels all unprocessed tasks, and wakes up
436	<	* all waiting workers.  Detecting whether termination should
437	<	* commence after a non-abrupt shutdown() call requires more work
438	<	* and bookkeeping. We need consensus about quiesence (i.e., that
439	<	* there is no more work) which is reflected in active counts so
440	<	* long as there are no current blockers, as well as possible
441	<	* re-evaluations during independent changes in blocking or
442	<	* quiescing workers.
443	<	*
444	<	* Style notes: There is a lot of representation-level coupling
445	<	* among classes ForkJoinPool, ForkJoinWorkerThread, and
446	<	* ForkJoinTask.  Most fields of ForkJoinWorkerThread maintain
447	<	* data structures managed by ForkJoinPool, so are directly
448	<	* accessed.  Conversely we allow access to "workers" array by
449	<	* workers, and direct access to ForkJoinTask.status by both
450	<	* ForkJoinPool and ForkJoinWorkerThread.  There is little point
418	>	* to create new spares are all racy, so we rely on multiple
419	>	* retries of each.  Compensation in the apparent absence of
420	>	* helping opportunities is challenging to control on JVMs, where
421	>	* GC and other activities can stall progress of tasks that in
422	>	* turn stall out many other dependent tasks, without us being
423	>	* able to determine whether they will ever require compensation.
424	>	* Even though work-stealing otherwise encounters little
425	>	* degradation in the presence of more threads than cores,
426	>	* aggressively adding new threads in such cases entails risk of
427	>	* unwanted positive feedback control loops in which more threads
428	>	* cause more dependent stalls (as well as delayed progress of
429	>	* unblocked threads to the point that we know they are available)
430	>	* leading to more situations requiring more threads, and so
431	>	* on. This aspect of control can be seen as an (analytically
432	>	* intractable) game with an opponent that may choose the worst
433	>	* (for us) active thread to stall at any time.  We take several
434	>	* precautions to bound losses (and thus bound gains), mainly in
435	>	* methods tryCompensate and awaitJoin.
436	>	*
437	>	* Common Pool
438	>	* ===========
439	>	*
440	>	* The static commonPool always exists after static
441	>	* initialization.  Since it (or any other created pool) need
442	>	* never be used, we minimize initial construction overhead and
443	>	* footprint to the setup of about a dozen fields, with no nested
444	>	* allocation. Most bootstrapping occurs within method
445	>	* fullExternalPush during the first submission to the pool.
446	>	*
447	>	* When external threads submit to the common pool, they can
448	>	* perform some subtask processing (see externalHelpJoin and
449	>	* related methods).  We do not need to record whether these
450	>	* submissions are to the common pool -- if not, externalHelpJoin
451	>	* returns quicky (at the most helping to signal some common pool
452	>	* workers). These submitters would otherwise be blocked waiting
453	>	* for completion, so the extra effort (with liberally sprinkled
454	>	* task status checks) in inapplicable cases amounts to an odd
455	>	* form of limited spin-wait before blocking in ForkJoinTask.join.
456	>	*
457	>	* Style notes
458	>	* ===========
459	>	*
460	>	* There is a lot of representation-level coupling among classes
461	>	* ForkJoinPool, ForkJoinWorkerThread, and ForkJoinTask.  The
462	>	* fields of WorkQueue maintain data structures managed by
463	>	* ForkJoinPool, so are directly accessed.  There is little point
464		* trying to reduce this, since any associated future changes in
465		* representations will need to be accompanied by algorithmic
466	<	* changes anyway. All together, these low-level implementation
467	<	* choices produce as much as a factor of 4 performance
468	<	* improvement compared to naive implementations, and enable the
469	<	* processing of billions of tasks per second, at the expense of
470	<	* some ugliness.
471	<	*
472	<	* Methods signalWork() and scan() are the main bottlenecks so are
473	<	* especially heavily micro-optimized/mangled.  There are lots of
474	<	* inline assignments (of form "while ((local = field) != 0)")
475	<	* which are usually the simplest way to ensure the required read
476	<	* orderings (which are sometimes critical). This leads to a
477	<	* "C"-like style of listing declarations of these locals at the
478	<	* heads of methods or blocks.  There are several occurrences of
479	<	* the unusual "do {} while (!cas...)"  which is the simplest way
480	<	* to force an update of a CAS'ed variable. There are also other
481	<	* coding oddities that help some methods perform reasonably even
482	<	* when interpreted (not compiled).
483	<	*
484	<	* The order of declarations in this file is: (1) declarations of
485	<	* statics (2) fields (along with constants used when unpacking
486	<	* some of them), listed in an order that tends to reduce
487	<	* contention among them a bit under most JVMs.  (3) internal
488	<	* control methods (4) callbacks and other support for
489	<	* ForkJoinTask and ForkJoinWorkerThread classes, (5) exported
490	<	* methods (plus a few little helpers). (6) static block
335	<	* initializing all statics in a minimally dependent order.
466	>	* changes anyway. Several methods intrinsically sprawl because
467	>	* they must accumulate sets of consistent reads of volatiles held
468	>	* in local variables.  Methods signalWork() and scan() are the
469	>	* main bottlenecks, so are especially heavily
470	>	* micro-optimized/mangled.  There are lots of inline assignments
471	>	* (of form "while ((local = field) != 0)") which are usually the
472	>	* simplest way to ensure the required read orderings (which are
473	>	* sometimes critical). This leads to a "C"-like style of listing
474	>	* declarations of these locals at the heads of methods or blocks.
475	>	* There are several occurrences of the unusual "do {} while
476	>	* (!cas...)"  which is the simplest way to force an update of a
477	>	* CAS'ed variable. There are also other coding oddities (including
478	>	* several unnecessary-looking hoisted null checks) that help
479	>	* some methods perform reasonably even when interpreted (not
480	>	* compiled).
481	>	*
482	>	* The order of declarations in this file is:
483	>	* (1) Static utility functions
484	>	* (2) Nested (static) classes
485	>	* (3) Static fields
486	>	* (4) Fields, along with constants used when unpacking some of them
487	>	* (5) Internal control methods
488	>	* (6) Callbacks and other support for ForkJoinTask methods
489	>	* (7) Exported methods
490	>	* (8) Static block initializing statics in minimally dependent order
491		*/
492
493	+	// Static utilities
494	+
495	+	/**
496	+	* If there is a security manager, makes sure caller has
497	+	* permission to modify threads.
498	+	*/
499	+	private static void checkPermission() {
500	+	SecurityManager security = System.getSecurityManager();
501	+	if (security != null)
502	+	security.checkPermission(modifyThreadPermission);
503	+	}
504	+
505	+	// Nested classes
506	+
507		/**
508		* Factory for creating new {@link ForkJoinWorkerThread}s.
509		* A {@code ForkJoinWorkerThreadFactory} must be defined and used
#	Line 363 | Line 532 | public class ForkJoinPool extends Abstra
532		}
533
534		/**
535	<	* Creates a new ForkJoinWorkerThread. This factory is used unless
536	<	* overridden in ForkJoinPool constructors.
535	>	* Class for artificial tasks that are used to replace the target
536	>	* of local joins if they are removed from an interior queue slot
537	>	* in WorkQueue.tryRemoveAndExec. We don't need the proxy to
538	>	* actually do anything beyond having a unique identity.
539	>	*/
540	>	static final class EmptyTask extends ForkJoinTask<Void> {
541	>	private static final long serialVersionUID = -7721805057305804111L;
542	>	EmptyTask() { status = ForkJoinTask.NORMAL; } // force done
543	>	public final Void getRawResult() { return null; }
544	>	public final void setRawResult(Void x) {}
545	>	public final boolean exec() { return true; }
546	>	}
547	>
548	>	/**
549	>	* Queues supporting work-stealing as well as external task
550	>	* submission. See above for main rationale and algorithms.
551	>	* Implementation relies heavily on "Unsafe" intrinsics
552	>	* and selective use of "volatile":
553	>	*
554	>	* Field "base" is the index (mod array.length) of the least valid
555	>	* queue slot, which is always the next position to steal (poll)
556	>	* from if nonempty. Reads and writes require volatile orderings
557	>	* but not CAS, because updates are only performed after slot
558	>	* CASes.
559	>	*
560	>	* Field "top" is the index (mod array.length) of the next queue
561	>	* slot to push to or pop from. It is written only by owner thread
562	>	* for push, or under lock for external/shared push, and accessed
563	>	* by other threads only after reading (volatile) base.  Both top
564	>	* and base are allowed to wrap around on overflow, but (top -
565	>	* base) (or more commonly -(base - top) to force volatile read of
566	>	* base before top) still estimates size. The lock ("qlock") is
567	>	* forced to -1 on termination, causing all further lock attempts
568	>	* to fail. (Note: we don't need CAS for termination state because
569	>	* upon pool shutdown, all shared-queues will stop being used
570	>	* anyway.)  Nearly all lock bodies are set up so that exceptions
571	>	* within lock bodies are "impossible" (modulo JVM errors that
572	>	* would cause failure anyway.)
573	>	*
574	>	* The array slots are read and written using the emulation of
575	>	* volatiles/atomics provided by Unsafe. Insertions must in
576	>	* general use putOrderedObject as a form of releasing store to
577	>	* ensure that all writes to the task object are ordered before
578	>	* its publication in the queue.  All removals entail a CAS to
579	>	* null.  The array is always a power of two. To ensure safety of
580	>	* Unsafe array operations, all accesses perform explicit null
581	>	* checks and implicit bounds checks via power-of-two masking.
582	>	*
583	>	* In addition to basic queuing support, this class contains
584	>	* fields described elsewhere to control execution. It turns out
585	>	* to work better memory-layout-wise to include them in this class
586	>	* rather than a separate class.
587	>	*
588	>	* Performance on most platforms is very sensitive to placement of
589	>	* instances of both WorkQueues and their arrays -- we absolutely
590	>	* do not want multiple WorkQueue instances or multiple queue
591	>	* arrays sharing cache lines. (It would be best for queue objects
592	>	* and their arrays to share, but there is nothing available to
593	>	* help arrange that).  Unfortunately, because they are recorded
594	>	* in a common array, WorkQueue instances are often moved to be
595	>	* adjacent by garbage collectors. To reduce impact, we use field
596	>	* padding that works OK on common platforms; this effectively
597	>	* trades off slightly slower average field access for the sake of
598	>	* avoiding really bad worst-case access. (Until better JVM
599	>	* support is in place, this padding is dependent on transient
600	>	* properties of JVM field layout rules.)  We also take care in
601	>	* allocating, sizing and resizing the array. Non-shared queue
602	>	* arrays are initialized by workers before use. Others are
603	>	* allocated on first use.
604		*/
605	<	public static final ForkJoinWorkerThreadFactory
606	<	defaultForkJoinWorkerThreadFactory;
605	>	static final class WorkQueue {
606	>	/**
607	>	* Capacity of work-stealing queue array upon initialization.
608	>	* Must be a power of two; at least 4, but should be larger to
609	>	* reduce or eliminate cacheline sharing among queues.
610	>	* Currently, it is much larger, as a partial workaround for
611	>	* the fact that JVMs often place arrays in locations that
612	>	* share GC bookkeeping (especially cardmarks) such that
613	>	* per-write accesses encounter serious memory contention.
614	>	*/
615	>	static final int INITIAL_QUEUE_CAPACITY = 1 << 13;
616
617	<	/**
618	<	* Permission required for callers of methods that may start or
619	<	* kill threads.
620	<	*/
621	<	private static final RuntimePermission modifyThreadPermission;
617	>	/**
618	>	* Maximum size for queue arrays. Must be a power of two less
619	>	* than or equal to 1 << (31 - width of array entry) to ensure
620	>	* lack of wraparound of index calculations, but defined to a
621	>	* value a bit less than this to help users trap runaway
622	>	* programs before saturating systems.
623	>	*/
624	>	static final int MAXIMUM_QUEUE_CAPACITY = 1 << 26; // 64M
625
626	<	/**
627	<	* If there is a security manager, makes sure caller has
628	<	* permission to modify threads.
629	<	*/
630	<	private static void checkPermission() {
631	<	SecurityManager security = System.getSecurityManager();
632	<	if (security != null)
633	<	security.checkPermission(modifyThreadPermission);
626	>	int seed;                  // for random scanning; initialize nonzero
627	>	volatile int eventCount;   // encoded inactivation count; < 0 if inactive
628	>	int nextWait;              // encoded record of next event waiter
629	>	final int mode;            // lifo, fifo, or shared
630	>	int nsteals;               // cumulative number of steals
631	>	int poolIndex;             // index of this queue in pool (or 0)
632	>	int stealHint;             // index of most recent known stealer
633	>	volatile int qlock;        // 1: locked, -1: terminate; else 0
634	>	volatile int base;         // index of next slot for poll
635	>	int top;                   // index of next slot for push
636	>	ForkJoinTask<?>[] array;   // the elements (initially unallocated)
637	>	final ForkJoinPool pool;   // the containing pool (may be null)
638	>	final ForkJoinWorkerThread owner; // owning thread or null if shared
639	>	volatile Thread parker;    // == owner during call to park; else null
640	>	volatile ForkJoinTask<?> currentJoin;  // task being joined in awaitJoin
641	>	ForkJoinTask<?> currentSteal; // current non-local task being executed
642	>	// Heuristic padding to ameliorate unfortunate memory placements
643	>	Object p00, p01, p02, p03, p04, p05, p06, p07;
644	>	Object p08, p09, p0a, p0b, p0c, p0d, p0e;
645	>
646	>	WorkQueue(ForkJoinPool pool, ForkJoinWorkerThread owner, int mode) {
647	>	this.mode = mode;
648	>	this.pool = pool;
649	>	this.owner = owner;
650	>	// Place indices in the center of array (that is not yet allocated)
651	>	base = top = INITIAL_QUEUE_CAPACITY >>> 1;
652	>	}
653	>
654	>	/**
655	>	* Pushes a task. Call only by owner in unshared queues.
656	>	* Cases needing resizing or rejection are relyaed to fullPush
657	>	* (that also handles shared queues).
658	>	*
659	>	* @param task the task. Caller must ensure non-null.
660	>	* @throw RejectedExecutionException if array cannot be resized
661	>	*/
662	>	final void push(ForkJoinTask<?> task) {
663	>	ForkJoinPool p; ForkJoinTask<?>[] a;
664	>	int s = top, n;
665	>	if ((a = array) != null && a.length > (n = s + 1 - base)) {
666	>	U.putOrderedObject
667	>	(a, (((a.length - 1) & s) << ASHIFT) + ABASE, task);
668	>	top = s + 1;
669	>	if (n <= 1 && (p = pool) != null)
670	>	p.signalWork(this, 1);
671	>	}
672	>	else
673	>	fullPush(task, true);
674	>	}
675	>
676	>	/**
677	>	* Pushes a task if lock is free and array is either big
678	>	* enough or can be resized to be big enough. Note: a
679	>	* specialization of a common fast path of this method is in
680	>	* ForkJoinPool.externalPush. When called from a FJWT queue,
681	>	* this can fail only if the pool has been shut down or
682	>	* an out of memory error.
683	>	*
684	>	* @param task the task. Caller must ensure non-null.
685	>	* @param owned if true, throw RJE on failure
686	>	*/
687	>	final boolean fullPush(ForkJoinTask<?> task, boolean owned) {
688	>	ForkJoinPool p; ForkJoinTask<?>[] a;
689	>	if (owned) {
690	>	if (qlock < 0) // must be shutting down
691	>	throw new RejectedExecutionException();
692	>	}
693	>	else if (!U.compareAndSwapInt(this, QLOCK, 0, 1))
694	>	return false;
695	>	try {
696	>	int s = top, oldLen, len;
697	>	if ((a = array) == null)
698	>	a = array = new ForkJoinTask<?>[len=INITIAL_QUEUE_CAPACITY];
699	>	else if ((oldLen = a.length) > s + 1 - base)
700	>	len = oldLen;
701	>	else if ((len = oldLen << 1) > MAXIMUM_QUEUE_CAPACITY)
702	>	throw new RejectedExecutionException("Capacity exceeded");
703	>	else {
704	>	int oldMask, b;
705	>	ForkJoinTask<?>[] oldA = a;
706	>	a = array = new ForkJoinTask<?>[len];
707	>	if ((oldMask = oldLen - 1) >= 0 && s - (b = base) > 0) {
708	>	int mask = len - 1;
709	>	do {
710	>	ForkJoinTask<?> x;
711	>	int oldj = ((b & oldMask) << ASHIFT) + ABASE;
712	>	int j    = ((b &    mask) << ASHIFT) + ABASE;
713	>	x = (ForkJoinTask<?>)
714	>	U.getObjectVolatile(oldA, oldj);
715	>	if (x != null &&
716	>	U.compareAndSwapObject(oldA, oldj, x, null))
717	>	U.putObjectVolatile(a, j, x);
718	>	} while (++b != s);
719	>	}
720	>	}
721	>	U.putOrderedObject
722	>	(a, (((len - 1) & s) << ASHIFT) + ABASE, task);
723	>	top = s + 1;
724	>	} finally {
725	>	if (!owned)
726	>	qlock = 0;
727	>	}
728	>	if ((p = pool) != null)
729	>	p.signalWork(this, 1);
730	>	return true;
731	>	}
732	>
733	>	/**
734	>	* Takes next task, if one exists, in LIFO order.  Call only
735	>	* by owner in unshared queues.
736	>	*/
737	>	final ForkJoinTask<?> pop() {
738	>	ForkJoinTask<?>[] a; ForkJoinTask<?> t; int m;
739	>	if ((a = array) != null && (m = a.length - 1) >= 0) {
740	>	for (int s; (s = top - 1) - base >= 0;) {
741	>	long j = ((m & s) << ASHIFT) + ABASE;
742	>	if ((t = (ForkJoinTask<?>)U.getObject(a, j)) == null)
743	>	break;
744	>	if (U.compareAndSwapObject(a, j, t, null)) {
745	>	top = s;
746	>	return t;
747	>	}
748	>	}
749	>	}
750	>	return null;
751	>	}
752	>
753	>	/**
754	>	* Takes a task in FIFO order if b is base of queue and a task
755	>	* can be claimed without contention. Specialized versions
756	>	* appear in ForkJoinPool methods scan and tryHelpStealer.
757	>	*/
758	>	final ForkJoinTask<?> pollAt(int b) {
759	>	ForkJoinTask<?> t; ForkJoinTask<?>[] a;
760	>	if ((a = array) != null) {
761	>	int j = (((a.length - 1) & b) << ASHIFT) + ABASE;
762	>	if ((t = (ForkJoinTask<?>)U.getObjectVolatile(a, j)) != null &&
763	>	base == b &&
764	>	U.compareAndSwapObject(a, j, t, null)) {
765	>	base = b + 1;
766	>	return t;
767	>	}
768	>	}
769	>	return null;
770	>	}
771	>
772	>	/**
773	>	* Takes next task, if one exists, in FIFO order.
774	>	*/
775	>	final ForkJoinTask<?> poll() {
776	>	ForkJoinTask<?>[] a; int b; ForkJoinTask<?> t;
777	>	while ((b = base) - top < 0 && (a = array) != null) {
778	>	int j = (((a.length - 1) & b) << ASHIFT) + ABASE;
779	>	t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
780	>	if (t != null) {
781	>	if (base == b &&
782	>	U.compareAndSwapObject(a, j, t, null)) {
783	>	base = b + 1;
784	>	return t;
785	>	}
786	>	}
787	>	else if (base == b) {
788	>	if (b + 1 == top)
789	>	break;
790	>	Thread.yield(); // wait for lagging update (very rare)
791	>	}
792	>	}
793	>	return null;
794	>	}
795	>
796	>	/**
797	>	* Takes next task, if one exists, in order specified by mode.
798	>	*/
799	>	final ForkJoinTask<?> nextLocalTask() {
800	>	return mode == 0 ? pop() : poll();
801	>	}
802	>
803	>	/**
804	>	* Returns next task, if one exists, in order specified by mode.
805	>	*/
806	>	final ForkJoinTask<?> peek() {
807	>	ForkJoinTask<?>[] a = array; int m;
808	>	if (a == null || (m = a.length - 1) < 0)
809	>	return null;
810	>	int i = mode == 0 ? top - 1 : base;
811	>	int j = ((i & m) << ASHIFT) + ABASE;
812	>	return (ForkJoinTask<?>)U.getObjectVolatile(a, j);
813	>	}
814	>
815	>	/**
816	>	* Pops the given task only if it is at the current top.
817	>	* (A shared version is available only via FJP.tryExternalUnpush)
818	>	*/
819	>	final boolean tryUnpush(ForkJoinTask<?> t) {
820	>	ForkJoinTask<?>[] a; int s;
821	>	if ((a = array) != null && (s = top) != base &&
822	>	U.compareAndSwapObject
823	>	(a, (((a.length - 1) & --s) << ASHIFT) + ABASE, t, null)) {
824	>	top = s;
825	>	return true;
826	>	}
827	>	return false;
828	>	}
829	>
830	>	/**
831	>	* Removes and cancels all known tasks, ignoring any exceptions.
832	>	*/
833	>	final void cancelAll() {
834	>	ForkJoinTask.cancelIgnoringExceptions(currentJoin);
835	>	ForkJoinTask.cancelIgnoringExceptions(currentSteal);
836	>	for (ForkJoinTask<?> t; (t = poll()) != null; )
837	>	ForkJoinTask.cancelIgnoringExceptions(t);
838	>	}
839	>
840	>	/**
841	>	* Computes next value for random probes.  Scans don't require
842	>	* a very high quality generator, but also not a crummy one.
843	>	* Marsaglia xor-shift is cheap and works well enough.  Note:
844	>	* This is manually inlined in its usages in ForkJoinPool to
845	>	* avoid writes inside busy scan loops.
846	>	*/
847	>	final int nextSeed() {
848	>	int r = seed;
849	>	r ^= r << 13;
850	>	r ^= r >>> 17;
851	>	return seed = r ^= r << 5;
852	>	}
853	>
854	>	/**
855	>	* Provides a more accurate estimate of size than (top - base)
856	>	* by ordering reads and checking whether a near-empty queue
857	>	* has at least one unclaimed task.
858	>	*/
859	>	final int queueSize() {
860	>	ForkJoinTask<?>[] a; int k, s, n;
861	>	return ((n = base - (s = top)) < 0 &&
862	>	(n != -1 ||
863	>	((a = array) != null && (k = a.length) > 0 &&
864	>	U.getObject
865	>	(a, (long)((((k - 1) & (s - 1)) << ASHIFT) + ABASE)) != null))) ?
866	>	-n : 0;
867	>	}
868	>
869	>	// Specialized execution methods
870	>
871	>	/**
872	>	* Pops and runs tasks until empty.
873	>	*/
874	>	private void popAndExecAll() {
875	>	// A bit faster than repeated pop calls
876	>	ForkJoinTask<?>[] a; int m, s; long j; ForkJoinTask<?> t;
877	>	while ((a = array) != null && (m = a.length - 1) >= 0 &&
878	>	(s = top - 1) - base >= 0 &&
879	>	(t = ((ForkJoinTask<?>)
880	>	U.getObject(a, j = ((m & s) << ASHIFT) + ABASE)))
881	>	!= null) {
882	>	if (U.compareAndSwapObject(a, j, t, null)) {
883	>	top = s;
884	>	t.doExec();
885	>	}
886	>	}
887	>	}
888	>
889	>	/**
890	>	* Polls and runs tasks until empty.
891	>	*/
892	>	private void pollAndExecAll() {
893	>	for (ForkJoinTask<?> t; (t = poll()) != null;)
894	>	t.doExec();
895	>	}
896	>
897	>	/**
898	>	* If present, removes from queue and executes the given task,
899	>	* or any other cancelled task. Returns (true) on any CAS
900	>	* or consistency check failure so caller can retry.
901	>	*
902	>	* @return false if no progress can be made, else true;
903	>	*/
904	>	final boolean tryRemoveAndExec(ForkJoinTask<?> task) {
905	>	boolean stat = true, removed = false, empty = true;
906	>	ForkJoinTask<?>[] a; int m, s, b, n;
907	>	if ((a = array) != null && (m = a.length - 1) >= 0 &&
908	>	(n = (s = top) - (b = base)) > 0) {
909	>	for (ForkJoinTask<?> t;;) {           // traverse from s to b
910	>	int j = ((--s & m) << ASHIFT) + ABASE;
911	>	t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
912	>	if (t == null)                    // inconsistent length
913	>	break;
914	>	else if (t == task) {
915	>	if (s + 1 == top) {           // pop
916	>	if (!U.compareAndSwapObject(a, j, task, null))
917	>	break;
918	>	top = s;
919	>	removed = true;
920	>	}
921	>	else if (base == b)           // replace with proxy
922	>	removed = U.compareAndSwapObject(a, j, task,
923	>	new EmptyTask());
924	>	break;
925	>	}
926	>	else if (t.status >= 0)
927	>	empty = false;
928	>	else if (s + 1 == top) {          // pop and throw away
929	>	if (U.compareAndSwapObject(a, j, t, null))
930	>	top = s;
931	>	break;
932	>	}
933	>	if (--n == 0) {
934	>	if (!empty && base == b)
935	>	stat = false;
936	>	break;
937	>	}
938	>	}
939	>	}
940	>	if (removed)
941	>	task.doExec();
942	>	return stat;
943	>	}
944	>
945	>	/**
946	>	* Polls for and executes the given task or any other task in
947	>	* its CountedCompleter computation
948	>	*/
949	>	final boolean pollAndExecCC(ForkJoinTask<?> root) {
950	>	ForkJoinTask<?>[] a; int b; Object o;
951	>	outer: while ((b = base) - top < 0 && (a = array) != null) {
952	>	long j = (((a.length - 1) & b) << ASHIFT) + ABASE;
953	>	if ((o = U.getObject(a, j)) == null ||
954	>	!(o instanceof CountedCompleter))
955	>	break;
956	>	for (CountedCompleter<?> t = (CountedCompleter<?>)o, r = t;;) {
957	>	if (r == root) {
958	>	if (base == b &&
959	>	U.compareAndSwapObject(a, j, t, null)) {
960	>	base = b + 1;
961	>	t.doExec();
962	>	return true;
963	>	}
964	>	else
965	>	break; // restart
966	>	}
967	>	if ((r = r.completer) == null)
968	>	break outer; // not part of root computation
969	>	}
970	>	}
971	>	return false;
972	>	}
973	>
974	>	/**
975	>	* Executes a top-level task and any local tasks remaining
976	>	* after execution.
977	>	*/
978	>	final void runTask(ForkJoinTask<?> t) {
979	>	if (t != null) {
980	>	(currentSteal = t).doExec();
981	>	currentSteal = null;
982	>	if (++nsteals < 0) {     // spill on overflow
983	>	ForkJoinPool p;
984	>	if ((p = pool) != null)
985	>	p.collectStealCount(this);
986	>	}
987	>	if (top != base) {       // process remaining local tasks
988	>	if (mode == 0)
989	>	popAndExecAll();
990	>	else
991	>	pollAndExecAll();
992	>	}
993	>	}
994	>	}
995	>
996	>	/**
997	>	* Executes a non-top-level (stolen) task.
998	>	*/
999	>	final void runSubtask(ForkJoinTask<?> t) {
1000	>	if (t != null) {
1001	>	ForkJoinTask<?> ps = currentSteal;
1002	>	(currentSteal = t).doExec();
1003	>	currentSteal = ps;
1004	>	}
1005	>	}
1006	>
1007	>	/**
1008	>	* Returns true if owned and not known to be blocked.
1009	>	*/
1010	>	final boolean isApparentlyUnblocked() {
1011	>	Thread wt; Thread.State s;
1012	>	return (eventCount >= 0 &&
1013	>	(wt = owner) != null &&
1014	>	(s = wt.getState()) != Thread.State.BLOCKED &&
1015	>	s != Thread.State.WAITING &&
1016	>	s != Thread.State.TIMED_WAITING);
1017	>	}
1018	>
1019	>	/**
1020	>	* If this owned and is not already interrupted, try to
1021	>	* interrupt and/or unpark, ignoring exceptions.
1022	>	*/
1023	>	final void interruptOwner() {
1024	>	Thread wt, p;
1025	>	if ((wt = owner) != null && !wt.isInterrupted()) {
1026	>	try {
1027	>	wt.interrupt();
1028	>	} catch (SecurityException ignore) {
1029	>	}
1030	>	}
1031	>	if ((p = parker) != null)
1032	>	U.unpark(p);
1033	>	}
1034	>
1035	>	// Unsafe mechanics
1036	>	private static final sun.misc.Unsafe U;
1037	>	private static final long QLOCK;
1038	>	private static final int ABASE;
1039	>	private static final int ASHIFT;
1040	>	static {
1041	>	int s;
1042	>	try {
1043	>	U = getUnsafe();
1044	>	Class<?> k = WorkQueue.class;
1045	>	Class<?> ak = ForkJoinTask[].class;
1046	>	QLOCK = U.objectFieldOffset
1047	>	(k.getDeclaredField("qlock"));
1048	>	ABASE = U.arrayBaseOffset(ak);
1049	>	s = U.arrayIndexScale(ak);
1050	>	} catch (Exception e) {
1051	>	throw new Error(e);
1052	>	}
1053	>	if ((s & (s-1)) != 0)
1054	>	throw new Error("data type scale not a power of two");
1055	>	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
1056	>	}
1057		}
1058
1059		/**
1060	<	* Generator for assigning sequence numbers as pool names.
1061	<	*/
1062	<	private static final AtomicInteger poolNumberGenerator;
1060	>	* Per-thread records for threads that submit to pools. Currently
1061	>	* holds only pseudo-random seed / index that is used to choose
1062	>	* submission queues in method externalPush. In the future, this may
1063	>	* also incorporate a means to implement different task rejection
1064	>	* and resubmission policies.
1065	>	*
1066	>	* Seeds for submitters and workers/workQueues work in basically
1067	>	* the same way but are initialized and updated using slightly
1068	>	* different mechanics. Both are initialized using the same
1069	>	* approach as in class ThreadLocal, where successive values are
1070	>	* unlikely to collide with previous values. Seeds are then
1071	>	* randomly modified upon collisions using xorshifts, which
1072	>	* requires a non-zero seed.
1073	>	*/
1074	>	static final class Submitter {
1075	>	int seed;
1076	>	Submitter(int s) { seed = s; }
1077	>	}
1078	>
1079	>	/** Property prefix for constructing common pool */
1080	>	private static final String propPrefix =
1081	>	"java.util.concurrent.ForkJoinPool.common.";
1082	>
1083	>	// static fields (initialized in static initializer below)
1084
1085		/**
1086	<	* Generator for initial random seeds for worker victim
1087	<	* selection. This is used only to create initial seeds. Random
396	<	* steals use a cheaper xorshift generator per steal attempt. We
397	<	* don't expect much contention on seedGenerator, so just use a
398	<	* plain Random.
1086	>	* Creates a new ForkJoinWorkerThread. This factory is used unless
1087	>	* overridden in ForkJoinPool constructors.
1088		*/
1089	<	static final Random workerSeedGenerator;
1089	>	public static final ForkJoinWorkerThreadFactory
1090	>	defaultForkJoinWorkerThreadFactory;
1091
1092		/**
1093	<	* Array holding all worker threads in the pool.  Initialized upon
1094	<	* construction. Array size must be a power of two.  Updates and
1095	<	* replacements are protected by scanGuard, but the array is
1096	<	* always kept in a consistent enough state to be randomly
407	<	* accessed without locking by workers performing work-stealing,
408	<	* as well as other traversal-based methods in this class, so long
409	<	* as reads memory-acquire by first reading ctl. All readers must
410	<	* tolerate that some array slots may be null.
1093	>	* Common (static) pool. Non-null for public use unless a static
1094	>	* construction exception, but internal usages null-check on use
1095	>	* to paranoically avoid potential initialization circularities
1096	>	* as well as to simplify generated code.
1097		*/
1098	<	ForkJoinWorkerThread[] workers;
1098	>	static final ForkJoinPool commonPool;
1099
1100		/**
1101	<	* Initial size for submission queue array. Must be a power of
1102	<	* two.  In many applications, these always stay small so we use a
417	<	* small initial cap.
1101	>	* Permission required for callers of methods that may start or
1102	>	* kill threads.
1103		*/
1104	<	private static final int INITIAL_QUEUE_CAPACITY = 8;
1104	>	private static final RuntimePermission modifyThreadPermission;
1105
1106		/**
1107	<	* Maximum size for submission queue array. Must be a power of two
1108	<	* less than or equal to 1 << (31 - width of array entry) to
1109	<	* ensure lack of index wraparound, but is capped at a lower
1110	<	* value to help users trap runaway computations.
1107	>	* Per-thread submission bookkeeping. Shared across all pools
1108	>	* to reduce ThreadLocal pollution and because random motion
1109	>	* to avoid contention in one pool is likely to hold for others.
1110	>	* Lazily initialized on first submission (but null-checked
1111	>	* in other contexts to avoid unnecessary initialization).
1112		*/
1113	<	private static final int MAXIMUM_QUEUE_CAPACITY = 1 << 24; // 16M
1113	>	static final ThreadLocal<Submitter> submitters;
1114
1115		/**
1116	<	* Array serving as submission queue. Initialized upon construction.
1116	>	* Common pool parallelism. Must equal commonPool.parallelism.
1117		*/
1118	<	private ForkJoinTask<?>[] submissionQueue;
1118	>	static final int commonPoolParallelism;
1119
1120		/**
1121	<	* Lock protecting submissions array for addSubmission
1121	>	* Sequence number for creating workerNamePrefix.
1122		*/
1123	<	private final ReentrantLock submissionLock;
1123	>	private static int poolNumberSequence;
1124
1125		/**
1126	<	* Condition for awaitTermination, using submissionLock for
1127	<	* convenience.
1126	>	* Return the next sequence number. We don't expect this to
1127	>	* ever contend so use simple builtin sync.
1128		*/
1129	<	private final Condition termination;
1129	>	private static final synchronized int nextPoolId() {
1130	>	return ++poolNumberSequence;
1131	>	}
1132	>
1133	>	// static constants
1134
1135		/**
1136	<	* Creation factory for worker threads.
1136	>	* Initial timeout value (in nanoseconds) for the thread
1137	>	* triggering quiescence to park waiting for new work. On timeout,
1138	>	* the thread will instead try to shrink the number of
1139	>	* workers. The value should be large enough to avoid overly
1140	>	* aggressive shrinkage during most transient stalls (long GCs
1141	>	* etc).
1142		*/
1143	<	private final ForkJoinWorkerThreadFactory factory;
1143	>	private static final long IDLE_TIMEOUT      = 2000L * 1000L * 1000L; // 2sec
1144
1145		/**
1146	<	* The uncaught exception handler used when any worker abruptly
452	<	* terminates.
1146	>	* Timeout value when there are more threads than parallelism level
1147		*/
1148	<	final Thread.UncaughtExceptionHandler ueh;
1148	>	private static final long FAST_IDLE_TIMEOUT =  200L * 1000L * 1000L;
1149
1150		/**
1151	<	* Prefix for assigning names to worker threads
1151	>	* The maximum stolen->joining link depth allowed in method
1152	>	* tryHelpStealer.  Must be a power of two.  Depths for legitimate
1153	>	* chains are unbounded, but we use a fixed constant to avoid
1154	>	* (otherwise unchecked) cycles and to bound staleness of
1155	>	* traversal parameters at the expense of sometimes blocking when
1156	>	* we could be helping.
1157		*/
1158	<	private final String workerNamePrefix;
1158	>	private static final int MAX_HELP = 64;
1159
1160		/**
1161	<	* Sum of per-thread steal counts, updated only when threads are
1162	<	* idle or terminating.
1161	>	* Increment for seed generators. See class ThreadLocal for
1162	>	* explanation.
1163		*/
1164	<	private volatile long stealCount;
1164	>	private static final int SEED_INCREMENT = 0x61c88647;
1165
1166		/**
1167	<	* Main pool control -- a long packed with:
1167	>	* Bits and masks for control variables
1168	>	*
1169	>	* Field ctl is a long packed with:
1170		* AC: Number of active running workers minus target parallelism (16 bits)
1171	<	* TC: Number of total workers minus target parallelism (16bits)
1171	>	* TC: Number of total workers minus target parallelism (16 bits)
1172		* ST: true if pool is terminating (1 bit)
1173		* EC: the wait count of top waiting thread (15 bits)
1174	<	* ID: ~poolIndex of top of Treiber stack of waiting threads (16 bits)
1174	>	* ID: poolIndex of top of Treiber stack of waiters (16 bits)
1175		*
1176		* When convenient, we can extract the upper 32 bits of counts and
1177		* the lower 32 bits of queue state, u = (int)(ctl >>> 32) and e =
#	Line 479 | Line 1180 | public class ForkJoinPool extends Abstra
1180		* parallelism and the positionings of fields makes it possible to
1181		* perform the most common checks via sign tests of fields: When
1182		* ac is negative, there are not enough active workers, when tc is
1183	<	* negative, there are not enough total workers, when id is
483	<	* negative, there is at least one waiting worker, and when e is
1183	>	* negative, there are not enough total workers, and when e is
1184		* negative, the pool is terminating.  To deal with these possibly
1185		* negative fields, we use casts in and out of "short" and/or
1186		* signed shifts to maintain signedness.
1187	+	*
1188	+	* When a thread is queued (inactivated), its eventCount field is
1189	+	* set negative, which is the only way to tell if a worker is
1190	+	* prevented from executing tasks, even though it must continue to
1191	+	* scan for them to avoid queuing races. Note however that
1192	+	* eventCount updates lag releases so usage requires care.
1193	+	*
1194	+	* Field plock is an int packed with:
1195	+	* SHUTDOWN: true if shutdown is enabled (1 bit)
1196	+	* SEQ:  a sequence lock, with PL_LOCK bit set if locked (30 bits)
1197	+	* SIGNAL: set when threads may be waiting on the lock (1 bit)
1198	+	*
1199	+	* The sequence number enables simple consistency checks:
1200	+	* Staleness of read-only operations on the workQueues array can
1201	+	* be checked by comparing plock before vs after the reads.
1202		*/
488	–	volatile long ctl;
1203
1204		// bit positions/shifts for fields
1205		private static final int  AC_SHIFT   = 48;
#	Line 494 | Line 1208 | public class ForkJoinPool extends Abstra
1208		private static final int  EC_SHIFT   = 16;
1209
1210		// bounds
1211	<	private static final int  MAX_ID     = 0x7fff;  // max poolIndex
1212	<	private static final int  SMASK      = 0xffff;  // mask short bits
1211	>	private static final int  SMASK      = 0xffff;  // short bits
1212	>	private static final int  MAX_CAP    = 0x7fff;  // max #workers - 1
1213	>	private static final int  EVENMASK   = 0xfffe;  // even short bits
1214	>	private static final int  SQMASK     = 0x007e;  // max 64 (even) slots
1215		private static final int  SHORT_SIGN = 1 << 15;
1216		private static final int  INT_SIGN   = 1 << 31;
1217
#	Line 517 | Line 1233 | public class ForkJoinPool extends Abstra
1233		private static final int  UTC_UNIT   = 1 << UTC_SHIFT;
1234
1235		// masks and units for dealing with e = (int)ctl
1236	<	private static final int  E_MASK     = 0x7fffffff; // no STOP_BIT
1237	<	private static final int  EC_UNIT    = 1 << EC_SHIFT;
1236	>	private static final int E_MASK      = 0x7fffffff; // no STOP_BIT
1237	>	private static final int E_SEQ       = 1 << EC_SHIFT;
1238
1239	<	/**
1240	<	* The target parallelism level.
1241	<	*/
1242	<	final int parallelism;
1243	<
1244	<	/**
1245	<	* Index (mod submission queue length) of next element to take
1246	<	* from submission queue. Usage is identical to that for
1247	<	* per-worker queues -- see ForkJoinWorkerThread internal
1248	<	* documentation.
533	<	*/
534	<	volatile int queueBase;
535	<
536	<	/**
537	<	* Index (mod submission queue length) of next element to add
538	<	* in submission queue. Usage is identical to that for
539	<	* per-worker queues -- see ForkJoinWorkerThread internal
540	<	* documentation.
541	<	*/
542	<	int queueTop;
543	<
544	<	/**
545	<	* True when shutdown() has been called.
546	<	*/
547	<	volatile boolean shutdown;
548	<
549	<	/**
550	<	* True if use local fifo, not default lifo, for local polling
551	<	* Read by, and replicated by ForkJoinWorkerThreads
552	<	*/
553	<	final boolean locallyFifo;
1239	>	// plock bits
1240	>	private static final int SHUTDOWN    = 1 << 31;
1241	>	private static final int PL_LOCK     = 2;
1242	>	private static final int PL_SIGNAL   = 1;
1243	>	private static final int PL_SPINS    = 1 << 8;
1244	>
1245	>	// access mode for WorkQueue
1246	>	static final int LIFO_QUEUE          =  0;
1247	>	static final int FIFO_QUEUE          =  1;
1248	>	static final int SHARED_QUEUE        = -1;
1249
1250	<	/**
556	<	* The number of threads in ForkJoinWorkerThreads.helpQuiescePool.
557	<	* When non-zero, suppresses automatic shutdown when active
558	<	* counts become zero.
559	<	*/
560	<	volatile int quiescerCount;
1250	>	// Instance fields
1251
1252	<	/**
1253	<	* The number of threads blocked in join.
1254	<	*/
1255	<	volatile int blockedCount;
1252	>	/*
1253	>	* Field layout order in this class tends to matter more than one
1254	>	* would like. Runtime layout order is only loosely related to
1255	>	* declaration order and may differ across JVMs, but the following
1256	>	* empirically works OK on current JVMs.
1257	>	*/
1258	>	volatile long stealCount;                  // collects worker counts
1259	>	volatile long ctl;                         // main pool control
1260	>	final int parallelism;                     // parallelism level
1261	>	final int localMode;                       // per-worker scheduling mode
1262	>	volatile int indexSeed;                    // worker/submitter index seed
1263	>	volatile int plock;                        // shutdown status and seqLock
1264	>	WorkQueue[] workQueues;                    // main registry
1265	>	final ForkJoinWorkerThreadFactory factory; // factory for new workers
1266	>	final Thread.UncaughtExceptionHandler ueh; // per-worker UEH
1267	>	final String workerNamePrefix;             // to create worker name string
1268
1269	<	/**
1270	<	* Counter for worker Thread names (unrelated to their poolIndex)
1271	<	*/
1272	<	private volatile int nextWorkerNumber;
1269	>	/*
1270	>	* Acquires the plock lock to protect worker array and related
1271	>	* updates. This method is called only if an initial CAS on plock
1272	>	* fails. This acts as a spinLock for normal cases, but falls back
1273	>	* to builtin monitor to block when (rarely) needed. This would be
1274	>	* a terrible idea for a highly contended lock, but works fine as
1275	>	* a more conservative alternative to a pure spinlock.  See
1276	>	* internal ConcurrentHashMap documentation for further
1277	>	* explanation of nearly the same construction.
1278	>	*/
1279	>	private int acquirePlock() {
1280	>	int spins = PL_SPINS, r = 0, ps, nps;
1281	>	for (;;) {
1282	>	if (((ps = plock) & PL_LOCK) == 0 &&
1283	>	U.compareAndSwapInt(this, PLOCK, ps, nps = ps + PL_LOCK))
1284	>	return nps;
1285	>	else if (r == 0)
1286	>	r = ThreadLocalRandom.current().nextInt(); // randomize spins
1287	>	else if (spins >= 0) {
1288	>	r ^= r << 1; r ^= r >>> 3; r ^= r << 10; // xorshift
1289	>	if (r >= 0)
1290	>	--spins;
1291	>	}
1292	>	else if (U.compareAndSwapInt(this, PLOCK, ps, ps | PL_SIGNAL)) {
1293	>	synchronized (this) {
1294	>	if ((plock & PL_SIGNAL) != 0) {
1295	>	try {
1296	>	wait();
1297	>	} catch (InterruptedException ie) {
1298	>	try {
1299	>	Thread.currentThread().interrupt();
1300	>	} catch (SecurityException ignore) {
1301	>	}
1302	>	}
1303	>	}
1304	>	else
1305	>	notifyAll();
1306	>	}
1307	>	}
1308	>	}
1309	>	}
1310
1311		/**
1312	<	* The index for the next created worker. Accessed under scanGuard.
1312	>	* Unlocks and signals any thread waiting for plock. Called only
1313	>	* when CAS of seq value for unlock fails.
1314		*/
1315	<	private int nextWorkerIndex;
1315	>	private void releasePlock(int ps) {
1316	>	plock = ps;
1317	>	synchronized (this) { notifyAll(); }
1318	>	}
1319	>
1320	>	//  Registering and deregistering workers
1321	>
1322	>	/**
1323	>	* Callback from ForkJoinWorkerThread constructor to establish its
1324	>	* poolIndex and record its WorkQueue. To avoid scanning bias due
1325	>	* to packing entries in front of the workQueues array, we treat
1326	>	* the array as a simple power-of-two hash table using per-thread
1327	>	* seed as hash, expanding as needed.
1328	>	*
1329	>	* @param w the worker's queue
1330	>	*/
1331	>	final void registerWorker(WorkQueue w) {
1332	>	int s, ps; // generate a rarely colliding candidate index seed
1333	>	do {} while (!U.compareAndSwapInt(this, INDEXSEED,
1334	>	s = indexSeed, s += SEED_INCREMENT) ||
1335	>	s == 0); // skip 0
1336	>	if (((ps = plock) & PL_LOCK) != 0 ||
1337	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1338	>	ps = acquirePlock();
1339	>	int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1340	>	try {
1341	>	WorkQueue[] ws;
1342	>	if (w != null && (ws = workQueues) != null) {
1343	>	w.seed = s;
1344	>	int n = ws.length, m = n - 1;
1345	>	int r = (s << 1) | 1;               // use odd-numbered indices
1346	>	if (ws[r &= m] != null) {           // collision
1347	>	int probes = 0;                 // step by approx half size
1348	>	int step = (n <= 4) ? 2 : ((n >>> 1) & EVENMASK) + 2;
1349	>	while (ws[r = (r + step) & m] != null) {
1350	>	if (++probes >= n) {
1351	>	workQueues = ws = Arrays.copyOf(ws, n <<= 1);
1352	>	m = n - 1;
1353	>	probes = 0;
1354	>	}
1355	>	}
1356	>	}
1357	>	w.eventCount = w.poolIndex = r;     // establish before recording
1358	>	ws[r] = w;
1359	>	}
1360	>	} finally {
1361	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1362	>	releasePlock(nps);
1363	>	}
1364	>	}
1365
1366		/**
1367	<	* SeqLock and index masking for updates to workers array.  Locked
1368	<	* when SG_UNIT is set. Unlocking clears bit by adding
1369	<	* SG_UNIT. Staleness of read-only operations can be checked by
1370	<	* comparing scanGuard to value before the reads. The low 16 bits
1371	<	* (i.e, anding with SMASK) hold (the smallest power of two
1372	<	* covering all worker indices, minus one, and is used to avoid
1373	<	* dealing with large numbers of null slots when the workers array
1374	<	* is overallocated.
1375	<	*/
1376	<	volatile int scanGuard;
1367	>	* Final callback from terminating worker, as well as upon failure
1368	>	* to construct or start a worker.  Removes record of worker from
1369	>	* array, and adjusts counts. If pool is shutting down, tries to
1370	>	* complete termination.
1371	>	*
1372	>	* @param wt the worker thread or null if construction failed
1373	>	* @param ex the exception causing failure, or null if none
1374	>	*/
1375	>	final void deregisterWorker(ForkJoinWorkerThread wt, Throwable ex) {
1376	>	WorkQueue w = null;
1377	>	if (wt != null && (w = wt.workQueue) != null) {
1378	>	int ps;
1379	>	collectStealCount(w);
1380	>	w.qlock = -1;                // ensure set
1381	>	if (((ps = plock) & PL_LOCK) != 0 ||
1382	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1383	>	ps = acquirePlock();
1384	>	int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1385	>	try {
1386	>	int idx = w.poolIndex;
1387	>	WorkQueue[] ws = workQueues;
1388	>	if (ws != null && idx >= 0 && idx < ws.length && ws[idx] == w)
1389	>	ws[idx] = null;
1390	>	} finally {
1391	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1392	>	releasePlock(nps);
1393	>	}
1394	>	}
1395
1396	<	private static final int SG_UNIT = 1 << 16;
1396	>	long c;                             // adjust ctl counts
1397	>	do {} while (!U.compareAndSwapLong
1398	>	(this, CTL, c = ctl, (((c - AC_UNIT) & AC_MASK) |
1399	>	((c - TC_UNIT) & TC_MASK) |
1400	>	(c & ~(AC_MASK|TC_MASK)))));
1401
1402	<	/**
1403	<	* The wakeup interval (in nanoseconds) for a worker waiting for a
1404	<	* task when the pool is quiescent to instead try to shrink the
1405	<	* number of workers.  The exact value does not matter too
1406	<	* much. It must be short enough to release resources during
1407	<	* sustained periods of idleness, but not so short that threads
1408	<	* are continually re-created.
1409	<	*/
1410	<	private static final long SHRINK_RATE =
1411	<	4L * 1000L * 1000L * 1000L; // 4 seconds
1402	>	if (!tryTerminate(false, false) && w != null) {
1403	>	w.cancelAll();                  // cancel remaining tasks
1404	>	if (w.array != null)            // suppress signal if never ran
1405	>	signalWork(null, 1);        // wake up or create replacement
1406	>	if (ex == null)                 // help clean refs on way out
1407	>	ForkJoinTask.helpExpungeStaleExceptions();
1408	>	}
1409	>
1410	>	if (ex != null)                     // rethrow
1411	>	ForkJoinTask.rethrow(ex);
1412	>	}
1413
1414		/**
1415	<	* Top-level loop for worker threads: On each step: if the
1416	<	* previous step swept through all queues and found no tasks, or
1417	<	* there are excess threads, then possibly blocks. Otherwise,
1418	<	* scans for and, if found, executes a task. Returns when pool
1419	<	* and/or worker terminate.
1420	<	*
1421	<	* @param w the worker
1422	<	*/
1423	<	final void work(ForkJoinWorkerThread w) {
1424	<	boolean swept = false;                // true on empty scans
1425	<	long c;
1426	<	while (!w.terminate && (int)(c = ctl) >= 0) {
1427	<	int a;                            // active count
616	<	if (!swept && (a = (int)(c >> AC_SHIFT)) <= 0)
617	<	swept = scan(w, a);
618	<	else if (tryAwaitWork(w, c))
619	<	swept = false;
1415	>	* Collect worker steal count into total. Called on termination
1416	>	* and upon int overflow of local count. (There is a possible race
1417	>	* in the latter case vs any caller of getStealCount, which can
1418	>	* make its results less accurate than usual.)
1419	>	*/
1420	>	final void collectStealCount(WorkQueue w) {
1421	>	if (w != null) {
1422	>	long sc;
1423	>	int ns = w.nsteals;
1424	>	w.nsteals = 0; // handle overflow
1425	>	long steals = (ns >= 0) ? ns : 1L + (long)(Integer.MAX_VALUE);
1426	>	do {} while (!U.compareAndSwapLong(this, STEALCOUNT,
1427	>	sc = stealCount, sc + steals));
1428		}
1429		}
1430
1431	<	// Signalling
1431	>	// Submissions
1432
1433		/**
1434	<	* Wakes up or creates a worker.
1435	<	*/
1436	<	final void signalWork() {
1437	<	/*
1438	<	* The while condition is true if: (there is are too few total
1439	<	* workers OR there is at least one waiter) AND (there are too
1440	<	* few active workers OR the pool is terminating).  The value
1441	<	* of e distinguishes the remaining cases: zero (no waiters)
1442	<	* for create, negative if terminating (in which case do
1443	<	* nothing), else release a waiter. The secondary checks for
1444	<	* release (non-null array etc) can fail if the pool begins
1445	<	* terminating after the test, and don't impose any added cost
1446	<	* because JVMs must perform null and bounds checks anyway.
1447	<	*/
1448	<	long c; int e, u;
1449	<	while ((((e = (int)(c = ctl)) | (u = (int)(c >>> 32))) &
1450	<	(INT_SIGN|SHORT_SIGN)) == (INT_SIGN|SHORT_SIGN) && e >= 0) {
1451	<	if (e > 0) {                         // release a waiting worker
1452	<	int i; ForkJoinWorkerThread w; ForkJoinWorkerThread[] ws;
1453	<	if ((ws = workers) == null ||
1454	<	(i = ~e & SMASK) >= ws.length ||
1455	<	(w = ws[i]) == null)
1456	<	break;
1457	<	long nc = (((long)(w.nextWait & E_MASK)) |
1458	<	((long)(u + UAC_UNIT) << 32));
1459	<	if (w.eventCount == e &&
1460	<	UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
1461	<	w.eventCount = (e + EC_UNIT) & E_MASK;
1462	<	if (w.parked)
1463	<	UNSAFE.unpark(w);
1464	<	break;
1465	<	}
1434	>	* Unless shutting down, adds the given task to a submission queue
1435	>	* at submitter's current queue index (modulo submission
1436	>	* range). Only the most common path is directly handled in this
1437	>	* method. All others are relayed to fullExternalPush.
1438	>	*
1439	>	* @param task the task. Caller must ensure non-null.
1440	>	*/
1441	>	final void externalPush(ForkJoinTask<?> task) {
1442	>	WorkQueue[] ws; WorkQueue q; Submitter z; int m; ForkJoinTask<?>[] a;
1443	>	if ((z = submitters.get()) != null && plock > 0 &&
1444	>	(ws = workQueues) != null && (m = (ws.length - 1)) >= 0 &&
1445	>	(q = ws[m & z.seed & SQMASK]) != null &&
1446	>	U.compareAndSwapInt(q, QLOCK, 0, 1)) { // lock
1447	>	int s = q.top, n;
1448	>	if ((a = q.array) != null && a.length > (n = s + 1 - q.base)) {
1449	>	U.putObject(a, (long)(((a.length - 1) & s) << ASHIFT) + ABASE,
1450	>	task);
1451	>	q.top = s + 1;                     // push on to deque
1452	>	q.qlock = 0;
1453	>	if (n <= 1)
1454	>	signalWork(q, 1);
1455	>	return;
1456	>	}
1457	>	q.qlock = 0;
1458	>	}
1459	>	fullExternalPush(task);
1460	>	}
1461	>
1462	>	/**
1463	>	* Full version of externalPush. This method is called, among
1464	>	* other times, upon the first submission of the first task to the
1465	>	* pool, so must perform secondary initialization: creating
1466	>	* workQueue array and setting plock to a valid value. It also
1467	>	* detects first submission by an external thread by looking up
1468	>	* its ThreadLocal, and creates a new shared queue if the one at
1469	>	* index if empty or contended. The lock bodies must be
1470	>	* exception-free (so no try/finally) so we optimistically
1471	>	* allocate new queues/arrays outside the locks and throw them
1472	>	* away if (very rarely) not needed. Note that the plock seq value
1473	>	* can eventually wrap around zero, but if so harmlessly fails to
1474	>	* reinitialize.
1475	>	*/
1476	>	private void fullExternalPush(ForkJoinTask<?> task) {
1477	>	for (Submitter z = null;;) {
1478	>	WorkQueue[] ws; WorkQueue q; int ps, m, r, s;
1479	>	if ((ps = plock) < 0)
1480	>	throw new RejectedExecutionException();
1481	>	else if ((ws = workQueues) == null || (m = ws.length - 1) < 0) {
1482	>	int n = parallelism - 1; n |= n >>> 1; n |= n >>> 2;
1483	>	n |= n >>> 4; n |= n >>> 8; n |= n >>> 16;
1484	>	WorkQueue[] nws = new WorkQueue[(n + 1) << 1]; // power of two
1485	>	if ((ps & PL_LOCK) != 0 ||
1486	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1487	>	ps = acquirePlock();
1488	>	if ((ws = workQueues) == null)
1489	>	workQueues = nws;
1490	>	int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1491	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1492	>	releasePlock(nps);
1493	>	}
1494	>	else if (z == null && (z = submitters.get()) == null) {
1495	>	if (U.compareAndSwapInt(this, INDEXSEED,
1496	>	s = indexSeed, s += SEED_INCREMENT) &&
1497	>	s != 0) // skip 0
1498	>	submitters.set(z = new Submitter(s));
1499		}
1500	<	else if (UNSAFE.compareAndSwapLong
1501	<	(this, ctlOffset, c,
1502	<	(long)(((u + UTC_UNIT) & UTC_MASK) |
1503	<	((u + UAC_UNIT) & UAC_MASK)) << 32)) {
1504	<	addWorker();
1505	<	break;
1500	>	else {
1501	>	int k = (r = z.seed) & m & SQMASK;
1502	>	if ((q = ws[k]) == null && (ps & PL_LOCK) == 0) {
1503	>	(q = new WorkQueue(this, null, SHARED_QUEUE)).poolIndex = k;
1504	>	if (((ps = plock) & PL_LOCK) != 0 ||
1505	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1506	>	ps = acquirePlock();
1507	>	WorkQueue w = null;
1508	>	if ((ws = workQueues) != null && k < ws.length &&
1509	>	(w = ws[k]) == null)
1510	>	ws[k] = q;
1511	>	else
1512	>	q = w;
1513	>	int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1514	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1515	>	releasePlock(nps);
1516	>	}
1517	>	if (q != null && q.qlock == 0 && q.fullPush(task, false))
1518	>	return;
1519	>	r ^= r << 13;                // same xorshift as WorkQueues
1520	>	r ^= r >>> 17;
1521	>	z.seed = r ^= r << 5;        // move to a different index
1522		}
1523		}
1524		}
1525
1526	+	// Maintaining ctl counts
1527	+
1528		/**
1529	<	* Variant of signalWork to help release waiters on rescans.
671	<	* Tries once to release a waiter if active count < 0.
672	<	*
673	<	* @return false if failed due to contention, else true
1529	>	* Increments active count; mainly called upon return from blocking.
1530		*/
1531	<	private boolean tryReleaseWaiter() {
1532	<	long c; int e, i; ForkJoinWorkerThread w; ForkJoinWorkerThread[] ws;
1533	<	if ((e = (int)(c = ctl)) > 0 &&
678	<	(int)(c >> AC_SHIFT) < 0 &&
679	<	(ws = workers) != null &&
680	<	(i = ~e & SMASK) < ws.length &&
681	<	(w = ws[i]) != null) {
682	<	long nc = ((long)(w.nextWait & E_MASK) |
683	<	((c + AC_UNIT) & (AC_MASK|TC_MASK)));
684	<	if (w.eventCount != e ||
685	<	!UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc))
686	<	return false;
687	<	w.eventCount = (e + EC_UNIT) & E_MASK;
688	<	if (w.parked)
689	<	UNSAFE.unpark(w);
690	<	}
691	<	return true;
1531	>	final void incrementActiveCount() {
1532	>	long c;
1533	>	do {} while (!U.compareAndSwapLong(this, CTL, c = ctl, c + AC_UNIT));
1534		}
1535
694	–	// Scanning for tasks
695	–
1536		/**
1537	<	* Scans for and, if found, executes one task. Scans start at a
1538	<	* random index of workers array, and randomly select the first
1539	<	* (2*#workers)-1 probes, and then, if all empty, resort to 2
1540	<	* circular sweeps, which is necessary to check quiescence. and
1541	<	* taking a submission only if no stealable tasks were found.  The
1542	<	* steal code inside the loop is a specialized form of
1543	<	* ForkJoinWorkerThread.deqTask, followed bookkeeping to support
1544	<	* helpJoinTask and signal propagation. The code for submission
1545	<	* queues is almost identical. On each steal, the worker completes
1546	<	* not only the task, but also all local tasks that this task may
1547	<	* have generated. On detecting staleness or contention when
1548	<	* trying to take a task, this method returns without finishing
1549	<	* sweep, which allows global state rechecks before retry.
1550	<	*
1551	<	* @param w the worker
1552	<	* @param a the number of active workers
1553	<	* @return true if swept all queues without finding a task
1554	<	*/
1555	<	private boolean scan(ForkJoinWorkerThread w, int a) {
1556	<	int g = scanGuard; // mask 0 avoids useless scans if only one active
1557	<	int m = (parallelism == 1 - a && blockedCount == 0) ? 0 : g & SMASK;
1558	<	ForkJoinWorkerThread[] ws = workers;
1559	<	if (ws == null || ws.length <= m)         // staleness check
1560	<	return false;
1561	<	for (int r = w.seed, k = r, j = -(m + m); j <= m + m; ++j) {
1562	<	ForkJoinTask<?> t; ForkJoinTask<?>[] q; int b, i;
1563	<	ForkJoinWorkerThread v = ws[k & m];
724	<	if (v != null && (b = v.queueBase) != v.queueTop &&
725	<	(q = v.queue) != null && (i = (q.length - 1) & b) >= 0) {
726	<	long u = (i << ASHIFT) + ABASE;
727	<	if ((t = q[i]) != null && v.queueBase == b &&
728	<	UNSAFE.compareAndSwapObject(q, u, t, null)) {
729	<	int d = (v.queueBase = b + 1) - v.queueTop;
730	<	v.stealHint = w.poolIndex;
731	<	if (d != 0)
732	<	signalWork();             // propagate if nonempty
733	<	w.execTask(t);
1537	>	* Tries to create (at most one) or activate (possibly several)
1538	>	* workers if too few are active. On contention failure, continues
1539	>	* until at least one worker is signalled or the given queue is
1540	>	* empty or all workers are active.
1541	>	*
1542	>	* @param q if non-null, the queue holding tasks to be signalled
1543	>	* @param signals the target number of signals.
1544	>	*/
1545	>	final void signalWork(WorkQueue q, int signals) {
1546	>	long c; int e, u, i; WorkQueue[] ws; WorkQueue w; Thread p;
1547	>	while ((u = (int)((c = ctl) >>> 32)) < 0) {
1548	>	if ((e = (int)c) > 0) {
1549	>	if ((ws = workQueues) != null && ws.length > (i = e & SMASK) &&
1550	>	(w = ws[i]) != null && w.eventCount == (e | INT_SIGN)) {
1551	>	long nc = (((long)(w.nextWait & E_MASK)) |
1552	>	((long)(u + UAC_UNIT) << 32));
1553	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1554	>	w.eventCount = (e + E_SEQ) & E_MASK;
1555	>	if ((p = w.parker) != null)
1556	>	U.unpark(p);
1557	>	if (--signals <= 0)
1558	>	break;
1559	>	}
1560	>	else
1561	>	signals = 1;
1562	>	if ((q != null && q.queueSize() == 0))
1563	>	break;
1564		}
1565	<	r ^= r << 13; r ^= r >>> 17; w.seed = r ^ (r << 5);
1566	<	return false;                     // store next seed
737	<	}
738	<	else if (j < 0) {                     // xorshift
739	<	r ^= r << 13; r ^= r >>> 17; k = r ^= r << 5;
1565	>	else
1566	>	break;
1567		}
1568	<	else
1569	<	++k;
1570	<	}
1571	<	if (scanGuard != g)                       // staleness check
1572	<	return false;
1573	<	else {                                    // try to take submission
1574	<	ForkJoinTask<?> t; ForkJoinTask<?>[] q; int b, i;
1575	<	if ((b = queueBase) != queueTop &&
1576	<	(q = submissionQueue) != null &&
1577	<	(i = (q.length - 1) & b) >= 0) {
1578	<	long u = (i << ASHIFT) + ABASE;
1579	<	if ((t = q[i]) != null && queueBase == b &&
1580	<	UNSAFE.compareAndSwapObject(q, u, t, null)) {
1581	<	queueBase = b + 1;
1582	<	w.execTask(t);
1568	>	else if (e == 0 && (u & SHORT_SIGN) != 0) {
1569	>	long nc = (long)(((u + UTC_UNIT) & UTC_MASK) |
1570	>	((u + UAC_UNIT) & UAC_MASK)) << 32;
1571	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1572	>	ForkJoinWorkerThread wt = null;
1573	>	Throwable ex = null;
1574	>	boolean started = false;
1575	>	try {
1576	>	ForkJoinWorkerThreadFactory fac;
1577	>	if ((fac = factory) != null &&
1578	>	(wt = fac.newThread(this)) != null) {
1579	>	wt.start();
1580	>	started = true;
1581	>	}
1582	>	} catch (Throwable rex) {
1583	>	ex = rex;
1584	>	}
1585	>	if (!started)
1586	>	deregisterWorker(wt, ex); // adjust counts on failure
1587	>	break;
1588		}
757	–	return false;
1589		}
1590	<	return true;                         // all queues empty
1590	>	else
1591	>	break;
1592		}
1593		}
1594
1595	+	// Scanning for tasks
1596	+
1597		/**
1598	<	* Tries to enqueue worker w in wait queue and await change in
765	<	* worker's eventCount.  If the pool is quiescent and there is
766	<	* more than one worker, possibly terminates worker upon exit.
767	<	* Otherwise, before blocking, rescans queues to avoid missed
768	<	* signals.  Upon finding work, releases at least one worker
769	<	* (which may be the current worker). Rescans restart upon
770	<	* detected staleness or failure to release due to
771	<	* contention. Note the unusual conventions about Thread.interrupt
772	<	* here and elsewhere: Because interrupts are used solely to alert
773	<	* threads to check termination, which is checked here anyway, we
774	<	* clear status (using Thread.interrupted) before any call to
775	<	* park, so that park does not immediately return due to status
776	<	* being set via some other unrelated call to interrupt in user
777	<	* code.
778	<	*
779	<	* @param w the calling worker
780	<	* @param c the ctl value on entry
781	<	* @return true if waited or another thread was released upon enq
1598	>	* Top-level runloop for workers, called by ForkJoinWorkerThread.run.
1599		*/
1600	<	private boolean tryAwaitWork(ForkJoinWorkerThread w, long c) {
1601	<	int v = w.eventCount;
1602	<	w.nextWait = (int)c;                      // w's successor record
1603	<	long nc = (long)(v & E_MASK) | ((c - AC_UNIT) & (AC_MASK|TC_MASK));
1604	<	if (ctl != c || !UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
1605	<	long d = ctl; // return true if lost to a deq, to force scan
1606	<	return (int)d != (int)c && ((d - c) & AC_MASK) >= 0L;
1607	<	}
1608	<	for (int sc = w.stealCount; sc != 0;) {   // accumulate stealCount
1609	<	long s = stealCount;
1610	<	if (UNSAFE.compareAndSwapLong(this, stealCountOffset, s, s + sc))
1611	<	sc = w.stealCount = 0;
1612	<	else if (w.eventCount != v)
1613	<	return true;                      // update next time
1614	<	}
1615	<	if ((!shutdown || !tryTerminate(false)) &&
1616	<	(int)c != 0 && parallelism + (int)(nc >> AC_SHIFT) == 0 &&
1617	<	blockedCount == 0 && quiescerCount == 0)
1618	<	idleAwaitWork(w, nc, c, v);           // quiescent
1619	<	for (boolean rescanned = false;;) {
1620	<	if (w.eventCount != v)
1621	<	return true;
1622	<	if (!rescanned) {
1623	<	int g = scanGuard, m = g & SMASK;
1624	<	ForkJoinWorkerThread[] ws = workers;
1625	<	if (ws != null && m < ws.length) {
1626	<	rescanned = true;
1627	<	for (int i = 0; i <= m; ++i) {
1628	<	ForkJoinWorkerThread u = ws[i];
1629	<	if (u != null) {
1630	<	if (u.queueBase != u.queueTop &&
1631	<	!tryReleaseWaiter())
1632	<	rescanned = false; // contended
1633	<	if (w.eventCount != v)
1634	<	return true;
1635	<	}
1600	>	final void runWorker(WorkQueue w) {
1601	>	// initialize queue array in this thread
1602	>	w.array = new ForkJoinTask<?>[WorkQueue.INITIAL_QUEUE_CAPACITY];
1603	>	do { w.runTask(scan(w)); } while (w.qlock >= 0);
1604	>	}
1605	>
1606	>	/**
1607	>	* Scans for and, if found, returns one task, else possibly
1608	>	* inactivates the worker. This method operates on single reads of
1609	>	* volatile state and is designed to be re-invoked continuously,
1610	>	* in part because it returns upon detecting inconsistencies,
1611	>	* contention, or state changes that indicate possible success on
1612	>	* re-invocation.
1613	>	*
1614	>	* The scan searches for tasks across a random permutation of
1615	>	* queues (starting at a random index and stepping by a random
1616	>	* relative prime, checking each at least once).  The scan
1617	>	* terminates upon either finding a non-empty queue, or completing
1618	>	* the sweep. If the worker is not inactivated, it takes and
1619	>	* returns a task from this queue. Otherwise, if not activated, it
1620	>	* signals workers (that may include itself) and returns so caller
1621	>	* can retry. Also returns for trtry if the worker array may have
1622	>	* changed during an empty scan.  On failure to find a task, we
1623	>	* take one of the following actions, after which the caller will
1624	>	* retry calling this method unless terminated.
1625	>	*
1626	>	* * If pool is terminating, terminate the worker.
1627	>	*
1628	>	* * If not already enqueued, try to inactivate and enqueue the
1629	>	* worker on wait queue. Or, if inactivating has caused the pool
1630	>	* to be quiescent, relay to idleAwaitWork to check for
1631	>	* termination and possibly shrink pool.
1632	>	*
1633	>	* * If already enqueued and none of the above apply, possibly
1634	>	* (with 1/2 probablility) park awaiting signal, else lingering to
1635	>	* help scan and signal.
1636	>	*
1637	>	* @param w the worker (via its WorkQueue)
1638	>	* @return a task or null if none found
1639	>	*/
1640	>	private final ForkJoinTask<?> scan(WorkQueue w) {
1641	>	WorkQueue[] ws; WorkQueue q;           // first update random seed
1642	>	int r = w.seed; r ^= r << 13; r ^= r >>> 17; w.seed = r ^= r << 5;
1643	>	int ps = plock, m;                     // volatile read order matters
1644	>	if ((ws = workQueues) != null && (m = ws.length - 1) > 0) {
1645	>	int ec = w.eventCount;             // ec is negative if inactive
1646	>	int step = (r >>> 16) | 1;         // relatively prime
1647	>	for (int j = (m + 1) << 2;  ; --j, r += step) {
1648	>	ForkJoinTask<?> t; ForkJoinTask<?>[] a; int b, n;
1649	>	if ((q = ws[r & m]) != null && (b = q.base) - q.top < 0 &&
1650	>	(a = q.array) != null) {   // probably nonempty
1651	>	int i = (((a.length - 1) & b) << ASHIFT) + ABASE;
1652	>	t = (ForkJoinTask<?>)U.getObjectVolatile(a, i);
1653	>	if (q.base == b && ec >= 0 && t != null &&
1654	>	U.compareAndSwapObject(a, i, t, null)) {
1655	>	if ((n = q.top - (q.base = b + 1)) > 0)
1656	>	signalWork(q, n);
1657	>	return t;              // taken
1658	>	}
1659	>	if (j < m || (ec < 0 && (ec = w.eventCount) < 0)) {
1660	>	if ((n = q.queueSize() - 1) > 0)
1661	>	signalWork(q, n);
1662	>	break;                 // let caller retry after signal
1663		}
1664		}
1665	<	if (scanGuard != g ||              // stale
1666	<	(queueBase != queueTop && !tryReleaseWaiter()))
1667	<	rescanned = false;
1668	<	if (!rescanned)
1669	<	Thread.yield();                // reduce contention
1670	<	else
1671	<	Thread.interrupted();          // clear before park
1672	<	}
1673	<	else {
1674	<	w.parked = true;                   // must recheck
1675	<	if (w.eventCount != v) {
1676	<	w.parked = false;
1677	<	return true;
1665	>	else if (j < 0) {              // end of scan
1666	>	long c = ctl; int e;
1667	>	if (plock != ps)           // incomplete sweep
1668	>	break;
1669	>	if ((e = (int)c) < 0)      // pool is terminating
1670	>	w.qlock = -1;
1671	>	else if (ec >= 0) {        // try to enqueue/inactivate
1672	>	long nc = ((long)ec |
1673	>	((c - AC_UNIT) & (AC_MASK|TC_MASK)));
1674	>	w.nextWait = e;
1675	>	w.eventCount = ec | INT_SIGN; // mark as inactive
1676	>	if (ctl != c ||
1677	>	!U.compareAndSwapLong(this, CTL, c, nc))
1678	>	w.eventCount = ec; // unmark on CAS failure
1679	>	else if ((int)(c >> AC_SHIFT) == 1 - parallelism)
1680	>	idleAwaitWork(w, nc, c);  // quiescent
1681	>	}
1682	>	else if (w.seed >= 0 && w.eventCount < 0) {
1683	>	Thread wt = Thread.currentThread();
1684	>	Thread.interrupted();  // clear status
1685	>	U.putObject(wt, PARKBLOCKER, this);
1686	>	w.parker = wt;         // emulate LockSupport.park
1687	>	if (w.eventCount < 0)  // recheck
1688	>	U.park(false, 0L);
1689	>	w.parker = null;
1690	>	U.putObject(wt, PARKBLOCKER, null);
1691	>	}
1692	>	break;
1693		}
835	–	LockSupport.park(this);
836	–	rescanned = w.parked = false;
1694		}
1695		}
1696	+	return null;
1697		}
1698
1699		/**
1700	<	* If inactivating worker w has caused pool to become
1701	<	* quiescent, check for pool termination, and wait for event
1702	<	* for up to SHRINK_RATE nanosecs (rescans are unnecessary in
1703	<	* this case because quiescence reflects consensus about lack
1704	<	* of work). On timeout, if ctl has not changed, terminate the
1705	<	* worker. Upon its termination (see deregisterWorker), it may
848	<	* wake up another worker to possibly repeat this process.
1700	>	* If inactivating worker w has caused the pool to become
1701	>	* quiescent, checks for pool termination, and, so long as this is
1702	>	* not the only worker, waits for event for up to a given
1703	>	* duration.  On timeout, if ctl has not changed, terminates the
1704	>	* worker, which will in turn wake up another worker to possibly
1705	>	* repeat this process.
1706		*
1707		* @param w the calling worker
1708	<	* @param currentCtl the ctl value after enqueuing w
1709	<	* @param prevCtl the ctl value if w terminated
1710	<	* @param v the eventCount w awaits change
1711	<	*/
1712	<	private void idleAwaitWork(ForkJoinWorkerThread w, long currentCtl,
1713	<	long prevCtl, int v) {
1714	<	if (w.eventCount == v) {
1715	<	if (shutdown)
1716	<	tryTerminate(false);
1717	<	ForkJoinTask.helpExpungeStaleExceptions(); // help clean weak refs
1708	>	* @param currentCtl the ctl value triggering possible quiescence
1709	>	* @param prevCtl the ctl value to restore if thread is terminated
1710	>	*/
1711	>	private void idleAwaitWork(WorkQueue w, long currentCtl, long prevCtl) {
1712	>	if (w.eventCount < 0 &&
1713	>	(this == commonPool || !tryTerminate(false, false)) &&
1714	>	(int)prevCtl != 0) {
1715	>	int dc = -(short)(currentCtl >>> TC_SHIFT);
1716	>	long parkTime = dc < 0 ? FAST_IDLE_TIMEOUT: (dc + 1) * IDLE_TIMEOUT;
1717	>	long deadline = System.nanoTime() + parkTime - 100000L; // 1ms slop
1718	>	Thread wt = Thread.currentThread();
1719		while (ctl == currentCtl) {
1720	<	long startTime = System.nanoTime();
1721	<	w.parked = true;
1722	<	if (w.eventCount == v)             // must recheck
1723	<	LockSupport.parkNanos(this, SHRINK_RATE);
1724	<	w.parked = false;
1725	<	if (w.eventCount != v)
1720	>	Thread.interrupted();  // timed variant of version in scan()
1721	>	U.putObject(wt, PARKBLOCKER, this);
1722	>	w.parker = wt;
1723	>	if (ctl == currentCtl)
1724	>	U.park(false, parkTime);
1725	>	w.parker = null;
1726	>	U.putObject(wt, PARKBLOCKER, null);
1727	>	if (ctl != currentCtl)
1728		break;
1729	<	else if (System.nanoTime() - startTime <
1730	<	SHRINK_RATE - (SHRINK_RATE / 10)) // timing slop
1731	<	Thread.interrupted();          // spurious wakeup
1732	<	else if (UNSAFE.compareAndSwapLong(this, ctlOffset,
873	<	currentCtl, prevCtl)) {
874	<	w.terminate = true;            // restore previous
875	<	w.eventCount = ((int)currentCtl + EC_UNIT) & E_MASK;
1729	>	if (deadline - System.nanoTime() <= 0L &&
1730	>	U.compareAndSwapLong(this, CTL, currentCtl, prevCtl)) {
1731	>	w.eventCount = (w.eventCount + E_SEQ) | E_MASK;
1732	>	w.qlock = -1;   // shrink
1733		break;
1734		}
1735		}
1736		}
1737		}
1738
882	–	// Submissions
883	–
1739		/**
1740	<	* Enqueues the given task in the submissionQueue.  Same idea as
1741	<	* ForkJoinWorkerThread.pushTask except for use of submissionLock.
1740	>	* Scans through queues looking for work while joining a task;
1741	>	* if any are present, signals.
1742		*
1743	<	* @param t the task
1743	>	* @param task to return early if done
1744	>	* @param origin an index to start scan
1745		*/
1746	<	private void addSubmission(ForkJoinTask<?> t) {
1747	<	final ReentrantLock lock = this.submissionLock;
1748	<	lock.lock();
1749	<	try {
1750	<	ForkJoinTask<?>[] q; int s, m;
1751	<	if ((q = submissionQueue) != null) {    // ignore if queue removed
1752	<	long u = (((s = queueTop) & (m = q.length-1)) << ASHIFT)+ABASE;
1753	<	UNSAFE.putOrderedObject(q, u, t);
1754	<	queueTop = s + 1;
1755	<	if (s - queueBase == m)
1756	<	growSubmissionQueue();
1746	>	final int helpSignal(ForkJoinTask<?> task, int origin) {
1747	>	WorkQueue[] ws; WorkQueue q; int m, n, s;
1748	>	if (task != null && (ws = workQueues) != null &&
1749	>	(m = ws.length - 1) >= 0) {
1750	>	for (int i = 0; i <= m; ++i) {
1751	>	if ((s = task.status) < 0)
1752	>	return s;
1753	>	if ((q = ws[(i + origin) & m]) != null &&
1754	>	(n = q.queueSize()) > 0) {
1755	>	signalWork(q, n);
1756	>	if ((int)(ctl >> AC_SHIFT) >= 0)
1757	>	break;
1758	>	}
1759		}
902	–	} finally {
903	–	lock.unlock();
1760		}
1761	<	signalWork();
1761	>	return 0;
1762		}
1763
908	–	//  (pollSubmission is defined below with exported methods)
909	–
1764		/**
1765	<	* Creates or doubles submissionQueue array.
1766	<	* Basically identical to ForkJoinWorkerThread version.
1767	<	*/
1768	<	private void growSubmissionQueue() {
1769	<	ForkJoinTask<?>[] oldQ = submissionQueue;
1770	<	int size = oldQ != null ? oldQ.length << 1 : INITIAL_QUEUE_CAPACITY;
1771	<	if (size > MAXIMUM_QUEUE_CAPACITY)
1772	<	throw new RejectedExecutionException("Queue capacity exceeded");
1773	<	if (size < INITIAL_QUEUE_CAPACITY)
1774	<	size = INITIAL_QUEUE_CAPACITY;
1775	<	ForkJoinTask<?>[] q = submissionQueue = new ForkJoinTask<?>[size];
1776	<	int mask = size - 1;
1777	<	int top = queueTop;
1778	<	int oldMask;
1779	<	if (oldQ != null && (oldMask = oldQ.length - 1) >= 0) {
1780	<	for (int b = queueBase; b != top; ++b) {
1781	<	long u = ((b & oldMask) << ASHIFT) + ABASE;
1782	<	Object x = UNSAFE.getObjectVolatile(oldQ, u);
1783	<	if (x != null && UNSAFE.compareAndSwapObject(oldQ, u, x, null))
1784	<	UNSAFE.putObjectVolatile
1785	<	(q, ((b & mask) << ASHIFT) + ABASE, x);
1786	<	}
1787	<	}
1788	<	}
1789	<
1790	<	// Blocking support
1791	<
1792	<	/**
1793	<	* Tries to increment blockedCount, decrement active count
1794	<	* (sometimes implicitly) and possibly release or create a
1795	<	* compensating worker in preparation for blocking. Fails
1796	<	* on contention or termination.
1797	<	*
1798	<	* @return true if the caller can block, else should recheck and retry
1799	<	*/
1800	<	private boolean tryPreBlock() {
1801	<	int b = blockedCount;
1802	<	if (UNSAFE.compareAndSwapInt(this, blockedCountOffset, b, b + 1)) {
1803	<	int pc = parallelism;
1804	<	do {
1805	<	ForkJoinWorkerThread[] ws; ForkJoinWorkerThread w;
1806	<	int e, ac, tc, rc, i;
1807	<	long c = ctl;
1808	<	int u = (int)(c >>> 32);
1809	<	if ((e = (int)c) < 0) {
1810	<	// skip -- terminating
1811	<	}
1812	<	else if ((ac = (u >> UAC_SHIFT)) <= 0 && e != 0 &&
1813	<	(ws = workers) != null &&
1814	<	(i = ~e & SMASK) < ws.length &&
1815	<	(w = ws[i]) != null) {
1816	<	long nc = ((long)(w.nextWait & E_MASK) |
1817	<	(c & (AC_MASK|TC_MASK)));
1818	<	if (w.eventCount == e &&
1819	<	UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
1820	<	w.eventCount = (e + EC_UNIT) & E_MASK;
1821	<	if (w.parked)
1822	<	UNSAFE.unpark(w);
1823	<	return true;             // release an idle worker
1824	<	}
1825	<	}
1826	<	else if ((tc = (short)(u >>> UTC_SHIFT)) >= 0 && ac + pc > 1) {
1827	<	long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK);
1828	<	if (UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc))
1829	<	return true;             // no compensation needed
1830	<	}
1831	<	else if (tc + pc < MAX_ID) {
1832	<	long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK);
1833	<	if (UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
1834	<	addWorker();
1835	<	return true;            // create a replacement
1836	<	}
1837	<	}
1838	<	// try to back out on any failure and let caller retry
1839	<	} while (!UNSAFE.compareAndSwapInt(this, blockedCountOffset,
1840	<	b = blockedCount, b - 1));
1765	>	* Tries to locate and execute tasks for a stealer of the given
1766	>	* task, or in turn one of its stealers, Traces currentSteal ->
1767	>	* currentJoin links looking for a thread working on a descendant
1768	>	* of the given task and with a non-empty queue to steal back and
1769	>	* execute tasks from. The first call to this method upon a
1770	>	* waiting join will often entail scanning/search, (which is OK
1771	>	* because the joiner has nothing better to do), but this method
1772	>	* leaves hints in workers to speed up subsequent calls. The
1773	>	* implementation is very branchy to cope with potential
1774	>	* inconsistencies or loops encountering chains that are stale,
1775	>	* unknown, or so long that they are likely cyclic.
1776	>	*
1777	>	* @param joiner the joining worker
1778	>	* @param task the task to join
1779	>	* @return 0 if no progress can be made, negative if task
1780	>	* known complete, else positive
1781	>	*/
1782	>	private int tryHelpStealer(WorkQueue joiner, ForkJoinTask<?> task) {
1783	>	int stat = 0, steps = 0;                    // bound to avoid cycles
1784	>	if (joiner != null && task != null) {       // hoist null checks
1785	>	restart: for (;;) {
1786	>	ForkJoinTask<?> subtask = task;     // current target
1787	>	for (WorkQueue j = joiner, v;;) {   // v is stealer of subtask
1788	>	WorkQueue[] ws; int m, s, h;
1789	>	if ((s = task.status) < 0) {
1790	>	stat = s;
1791	>	break restart;
1792	>	}
1793	>	if ((ws = workQueues) == null || (m = ws.length - 1) <= 0)
1794	>	break restart;              // shutting down
1795	>	if ((v = ws[h = (j.stealHint | 1) & m]) == null ||
1796	>	v.currentSteal != subtask) {
1797	>	for (int origin = h;;) {    // find stealer
1798	>	if (((h = (h + 2) & m) & 15) == 1 &&
1799	>	(subtask.status < 0 || j.currentJoin != subtask))
1800	>	continue restart;   // occasional staleness check
1801	>	if ((v = ws[h]) != null &&
1802	>	v.currentSteal == subtask) {
1803	>	j.stealHint = h;    // save hint
1804	>	break;
1805	>	}
1806	>	if (h == origin)
1807	>	break restart;      // cannot find stealer
1808	>	}
1809	>	}
1810	>	for (;;) { // help stealer or descend to its stealer
1811	>	ForkJoinTask[] a;  int b;
1812	>	if (subtask.status < 0)     // surround probes with
1813	>	continue restart;       //   consistency checks
1814	>	if ((b = v.base) - v.top < 0 && (a = v.array) != null) {
1815	>	int i = (((a.length - 1) & b) << ASHIFT) + ABASE;
1816	>	ForkJoinTask<?> t =
1817	>	(ForkJoinTask<?>)U.getObjectVolatile(a, i);
1818	>	if (subtask.status < 0 || j.currentJoin != subtask ||
1819	>	v.currentSteal != subtask)
1820	>	continue restart;   // stale
1821	>	stat = 1;               // apparent progress
1822	>	if (t != null && v.base == b &&
1823	>	U.compareAndSwapObject(a, i, t, null)) {
1824	>	v.base = b + 1;     // help stealer
1825	>	joiner.runSubtask(t);
1826	>	}
1827	>	else if (v.base == b && ++steps == MAX_HELP)
1828	>	break restart;      // v apparently stalled
1829	>	}
1830	>	else {                      // empty -- try to descend
1831	>	ForkJoinTask<?> next = v.currentJoin;
1832	>	if (subtask.status < 0 || j.currentJoin != subtask ||
1833	>	v.currentSteal != subtask)
1834	>	continue restart;   // stale
1835	>	else if (next == null || ++steps == MAX_HELP)
1836	>	break restart;      // dead-end or maybe cyclic
1837	>	else {
1838	>	subtask = next;
1839	>	j = v;
1840	>	break;
1841	>	}
1842	>	}
1843	>	}
1844	>	}
1845	>	}
1846		}
1847	<	return false;
989	<	}
990	<
991	<	/**
992	<	* Decrements blockedCount and increments active count
993	<	*/
994	<	private void postBlock() {
995	<	long c;
996	<	do {} while (!UNSAFE.compareAndSwapLong(this, ctlOffset,  // no mask
997	<	c = ctl, c + AC_UNIT));
998	<	int b;
999	<	do {} while (!UNSAFE.compareAndSwapInt(this, blockedCountOffset,
1000	<	b = blockedCount, b - 1));
1847	>	return stat;
1848		}
1849
1850		/**
1851	<	* Possibly blocks waiting for the given task to complete, or
1852	<	* cancels the task if terminating.  Fails to wait if contended.
1851	>	* Analog of tryHelpStealer for CountedCompleters. Tries to steal
1852	>	* and run tasks within the target's computation
1853	>	*
1854	>	* @param task the task to join
1855	>	* @param mode if shared, exit upon completing any task
1856	>	* if all workers are active
1857		*
1007	–	* @param joinMe the task
1858		*/
1859	<	final void tryAwaitJoin(ForkJoinTask<?> joinMe) {
1860	<	int s;
1861	<	Thread.interrupted(); // clear interrupts before checking termination
1862	<	if (joinMe.status >= 0) {
1863	<	if (tryPreBlock()) {
1864	<	joinMe.tryAwaitDone(0L);
1865	<	postBlock();
1859	>	private int helpComplete(ForkJoinTask<?> task, int mode) {
1860	>	WorkQueue[] ws; WorkQueue q; int m, n, s;
1861	>	if (task != null && (ws = workQueues) != null &&
1862	>	(m = ws.length - 1) >= 0) {
1863	>	for (int j = 1, origin = j;;) {
1864	>	if ((s = task.status) < 0)
1865	>	return s;
1866	>	if ((q = ws[j & m]) != null && q.pollAndExecCC(task)) {
1867	>	origin = j;
1868	>	if (mode == SHARED_QUEUE && (int)(ctl >> AC_SHIFT) >= 0)
1869	>	break;
1870	>	}
1871	>	else if ((j = (j + 2) & m) == origin)
1872	>	break;
1873		}
1017	–	else if ((ctl & STOP_BIT) != 0L)
1018	–	joinMe.cancelIgnoringExceptions();
1874		}
1875	+	return 0;
1876		}
1877
1878		/**
1879	<	* Possibly blocks the given worker waiting for joinMe to
1880	<	* complete or timeout
1881	<	*
1882	<	* @param joinMe the task
1883	<	* @param millis the wait time for underlying Object.wait
1884	<	*/
1885	<	final void timedAwaitJoin(ForkJoinTask<?> joinMe, long nanos) {
1886	<	while (joinMe.status >= 0) {
1887	<	Thread.interrupted();
1888	<	if ((ctl & STOP_BIT) != 0L) {
1889	<	joinMe.cancelIgnoringExceptions();
1890	<	break;
1879	>	* Tries to decrement active count (sometimes implicitly) and
1880	>	* possibly release or create a compensating worker in preparation
1881	>	* for blocking. Fails on contention or termination. Otherwise,
1882	>	* adds a new thread if no idle workers are available and pool
1883	>	* may become starved.
1884	>	*/
1885	>	final boolean tryCompensate() {
1886	>	int pc = parallelism, e, u, i, tc; long c;
1887	>	WorkQueue[] ws; WorkQueue w; Thread p;
1888	>	if ((e = (int)(c = ctl)) >= 0 && (ws = workQueues) != null) {
1889	>	if (e != 0 && (i = e & SMASK) < ws.length &&
1890	>	(w = ws[i]) != null && w.eventCount == (e | INT_SIGN)) {
1891	>	long nc = ((long)(w.nextWait & E_MASK) |
1892	>	(c & (AC_MASK|TC_MASK)));
1893	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1894	>	w.eventCount = (e + E_SEQ) & E_MASK;
1895	>	if ((p = w.parker) != null)
1896	>	U.unpark(p);
1897	>	return true;   // replace with idle worker
1898	>	}
1899		}
1900	<	if (tryPreBlock()) {
1901	<	long last = System.nanoTime();
1902	<	while (joinMe.status >= 0) {
1903	<	long millis = TimeUnit.NANOSECONDS.toMillis(nanos);
1904	<	if (millis <= 0)
1905	<	break;
1906	<	joinMe.tryAwaitDone(millis);
1907	<	if (joinMe.status < 0)
1908	<	break;
1909	<	if ((ctl & STOP_BIT) != 0L) {
1910	<	joinMe.cancelIgnoringExceptions();
1911	<	break;
1900	>	else if ((short)((u = (int)(c >>> 32)) >>> UTC_SHIFT) >= 0 &&
1901	>	(u >> UAC_SHIFT) + pc > 1) {
1902	>	long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK);
1903	>	if (U.compareAndSwapLong(this, CTL, c, nc))
1904	>	return true;    // no compensation
1905	>	}
1906	>	else if ((tc = u + pc) < MAX_CAP) {
1907	>	long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK);
1908	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1909	>	Throwable ex = null;
1910	>	ForkJoinWorkerThread wt = null;
1911	>	try {
1912	>	ForkJoinWorkerThreadFactory fac;
1913	>	if ((fac = factory) != null &&
1914	>	(wt = fac.newThread(this)) != null) {
1915	>	wt.start();
1916	>	return true;
1917	>	}
1918	>	} catch (Throwable rex) {
1919	>	ex = rex;
1920		}
1921	<	long now = System.nanoTime();
1050	<	nanos -= now - last;
1051	<	last = now;
1921	>	deregisterWorker(wt, ex); // adjust counts etc
1922		}
1053	–	postBlock();
1054	–	break;
1923		}
1924		}
1925	+	return false;
1926		}
1927
1928		/**
1929	<	* If necessary, compensates for blocker, and blocks
1929	>	* Helps and/or blocks until the given task is done.
1930	>	*
1931	>	* @param joiner the joining worker
1932	>	* @param task the task
1933	>	* @return task status on exit
1934		*/
1935	<	private void awaitBlocker(ManagedBlocker blocker)
1936	<	throws InterruptedException {
1937	<	while (!blocker.isReleasable()) {
1938	<	if (tryPreBlock()) {
1939	<	try {
1940	<	do {} while (!blocker.isReleasable() && !blocker.block());
1941	<	} finally {
1942	<	postBlock();
1935	>	final int awaitJoin(WorkQueue joiner, ForkJoinTask<?> task) {
1936	>	int s = 0;
1937	>	if (joiner != null && task != null && (s = task.status) >= 0) {
1938	>	ForkJoinTask<?> prevJoin = joiner.currentJoin;
1939	>	joiner.currentJoin = task;
1940	>	do {} while ((s = task.status) >= 0 &&
1941	>	joiner.queueSize() > 0 &&
1942	>	joiner.tryRemoveAndExec(task)); // process local tasks
1943	>	if (s >= 0 && (s = task.status) >= 0 &&
1944	>	(s = helpSignal(task, joiner.poolIndex)) >= 0 &&
1945	>	(task instanceof CountedCompleter))
1946	>	s = helpComplete(task, LIFO_QUEUE);
1947	>	while (s >= 0 && (s = task.status) >= 0) {
1948	>	if ((joiner.queueSize() > 0 ||           // try helping
1949	>	(s = tryHelpStealer(joiner, task)) == 0) &&
1950	>	(s = task.status) >= 0 && tryCompensate()) {
1951	>	if (task.trySetSignal() && (s = task.status) >= 0) {
1952	>	synchronized (task) {
1953	>	if (task.status >= 0) {
1954	>	try {                // see ForkJoinTask
1955	>	task.wait();     //  for explanation
1956	>	} catch (InterruptedException ie) {
1957	>	}
1958	>	}
1959	>	else
1960	>	task.notifyAll();
1961	>	}
1962	>	}
1963	>	long c;                          // re-activate
1964	>	do {} while (!U.compareAndSwapLong
1965	>	(this, CTL, c = ctl, c + AC_UNIT));
1966		}
1071	–	break;
1967		}
1968	+	joiner.currentJoin = prevJoin;
1969		}
1970	+	return s;
1971		}
1972
1076	–	// Creating, registering and deregistring workers
1077	–
1973		/**
1974	<	* Tries to create and start a worker; minimally rolls back counts
1975	<	* on failure.
1974	>	* Stripped-down variant of awaitJoin used by timed joins. Tries
1975	>	* to help join only while there is continuous progress. (Caller
1976	>	* will then enter a timed wait.)
1977	>	*
1978	>	* @param joiner the joining worker
1979	>	* @param task the task
1980		*/
1981	<	private void addWorker() {
1982	<	Throwable ex = null;
1983	<	ForkJoinWorkerThread t = null;
1984	<	try {
1985	<	t = factory.newThread(this);
1986	<	} catch (Throwable e) {
1987	<	ex = e;
1988	<	}
1989	<	if (t == null) {  // null or exceptional factory return
1990	<	long c;       // adjust counts
1991	<	do {} while (!UNSAFE.compareAndSwapLong
1992	<	(this, ctlOffset, c = ctl,
1993	<	(((c - AC_UNIT) & AC_MASK) |
1994	<	((c - TC_UNIT) & TC_MASK) |
1995	<	(c & ~(AC_MASK|TC_MASK)))));
1996	<	// Propagate exception if originating from an external caller
1997	<	if (!tryTerminate(false) && ex != null &&
1099	<	!(Thread.currentThread() instanceof ForkJoinWorkerThread))
1100	<	UNSAFE.throwException(ex);
1981	>	final void helpJoinOnce(WorkQueue joiner, ForkJoinTask<?> task) {
1982	>	int s;
1983	>	if (joiner != null && task != null && (s = task.status) >= 0) {
1984	>	ForkJoinTask<?> prevJoin = joiner.currentJoin;
1985	>	joiner.currentJoin = task;
1986	>	do {} while ((s = task.status) >= 0 &&
1987	>	joiner.queueSize() > 0 &&
1988	>	joiner.tryRemoveAndExec(task));
1989	>	if (s >= 0 && (s = task.status) >= 0 &&
1990	>	(s = helpSignal(task, joiner.poolIndex)) >= 0 &&
1991	>	(task instanceof CountedCompleter))
1992	>	s = helpComplete(task, LIFO_QUEUE);
1993	>	if (s >= 0 && joiner.queueSize() == 0) {
1994	>	do {} while (task.status >= 0 &&
1995	>	tryHelpStealer(joiner, task) > 0);
1996	>	}
1997	>	joiner.currentJoin = prevJoin;
1998		}
1102	–	else
1103	–	t.start();
1999		}
2000
2001		/**
2002	<	* Callback from ForkJoinWorkerThread constructor to assign a
2003	<	* public name
2004	<	*/
2005	<	final String nextWorkerName() {
2006	<	for (int n;;) {
2007	<	if (UNSAFE.compareAndSwapInt(this, nextWorkerNumberOffset,
2008	<	n = nextWorkerNumber, ++n))
2009	<	return workerNamePrefix + n;
2002	>	* Returns a (probably) non-empty steal queue, if one is found
2003	>	* during a random, then cyclic scan, else null.  This method must
2004	>	* be retried by caller if, by the time it tries to use the queue,
2005	>	* it is empty.
2006	>	* @param r a (random) seed for scanning
2007	>	*/
2008	>	private WorkQueue findNonEmptyStealQueue(int r) {
2009	>	int step = (r >>> 16) | 1;
2010	>	for (WorkQueue[] ws;;) {
2011	>	int ps = plock, m;
2012	>	if ((ws = workQueues) == null || (m = ws.length - 1) < 1)
2013	>	return null;
2014	>	for (int j = (m + 1) << 2; ; r += step) {
2015	>	WorkQueue q = ws[((r << 1) | 1) & m];
2016	>	if (q != null && q.queueSize() > 0)
2017	>	return q;
2018	>	else if (--j < 0) {
2019	>	if (plock == ps)
2020	>	return null;
2021	>	break;
2022	>	}
2023	>	}
2024		}
2025		}
2026
2027		/**
2028	<	* Callback from ForkJoinWorkerThread constructor to
2029	<	* determine its poolIndex and record in workers array.
2030	<	*
2031	<	* @param w the worker
2032	<	* @return the worker's pool index
2033	<	*/
2034	<	final int registerWorker(ForkJoinWorkerThread w) {
2035	<	/*
2036	<	* In the typical case, a new worker acquires the lock, uses
2037	<	* next available index and returns quickly.  Since we should
2038	<	* not block callers (ultimately from signalWork or
2039	<	* tryPreBlock) waiting for the lock needed to do this, we
2040	<	* instead help release other workers while waiting for the
2041	<	* lock.
2042	<	*/
2043	<	for (int g;;) {
2044	<	ForkJoinWorkerThread[] ws;
2045	<	if (((g = scanGuard) & SG_UNIT) == 0 &&
2046	<	UNSAFE.compareAndSwapInt(this, scanGuardOffset,
2047	<	g, g | SG_UNIT)) {
2048	<	int k = nextWorkerIndex;
2049	<	try {
2050	<	if ((ws = workers) != null) { // ignore on shutdown
2051	<	int n = ws.length;
2052	<	if (k < 0 || k >= n || ws[k] != null) {
2053	<	for (k = 0; k < n && ws[k] != null; ++k)
2054	<	;
2055	<	if (k == n)
2056	<	ws = workers = Arrays.copyOf(ws, n << 1);
2057	<	}
2058	<	ws[k] = w;
2059	<	nextWorkerIndex = k + 1;
2060	<	int m = g & SMASK;
2061	<	g = (k > m) ? ((m << 1) + 1) & SMASK : g + (SG_UNIT<<1);
2062	<	}
2063	<	} finally {
1155	<	scanGuard = g;
1156	<	}
1157	<	return k;
1158	<	}
1159	<	else if ((ws = workers) != null) { // help release others
1160	<	for (ForkJoinWorkerThread u : ws) {
1161	<	if (u != null && u.queueBase != u.queueTop) {
1162	<	if (tryReleaseWaiter())
1163	<	break;
1164	<	}
2028	>	* Runs tasks until {@code isQuiescent()}. We piggyback on
2029	>	* active count ctl maintenance, but rather than blocking
2030	>	* when tasks cannot be found, we rescan until all others cannot
2031	>	* find tasks either.
2032	>	*/
2033	>	final void helpQuiescePool(WorkQueue w) {
2034	>	for (boolean active = true;;) {
2035	>	ForkJoinTask<?> localTask; // exhaust local queue
2036	>	while ((localTask = w.nextLocalTask()) != null)
2037	>	localTask.doExec();
2038	>	// Similar to loop in scan(), but ignoring submissions
2039	>	WorkQueue q = findNonEmptyStealQueue(w.nextSeed());
2040	>	if (q != null) {
2041	>	ForkJoinTask<?> t; int b;
2042	>	if (!active) {      // re-establish active count
2043	>	long c;
2044	>	active = true;
2045	>	do {} while (!U.compareAndSwapLong
2046	>	(this, CTL, c = ctl, c + AC_UNIT));
2047	>	}
2048	>	if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null)
2049	>	w.runSubtask(t);
2050	>	}
2051	>	else {
2052	>	long c;
2053	>	if (active) {       // decrement active count without queuing
2054	>	active = false;
2055	>	do {} while (!U.compareAndSwapLong
2056	>	(this, CTL, c = ctl, c -= AC_UNIT));
2057	>	}
2058	>	else
2059	>	c = ctl;        // re-increment on exit
2060	>	if ((int)(c >> AC_SHIFT) + parallelism == 0) {
2061	>	do {} while (!U.compareAndSwapLong
2062	>	(this, CTL, c = ctl, c + AC_UNIT));
2063	>	break;
2064		}
2065		}
2066		}
2067		}
2068
2069		/**
2070	<	* Final callback from terminating worker.  Removes record of
2071	<	* worker from array, and adjusts counts. If pool is shutting
2072	<	* down, tries to complete termination.
2073	<	*
2074	<	* @param w the worker
2075	<	*/
2076	<	final void deregisterWorker(ForkJoinWorkerThread w, Throwable ex) {
2077	<	int idx = w.poolIndex;
2078	<	int sc = w.stealCount;
2079	<	int steps = 0;
2080	<	// Remove from array, adjust worker counts and collect steal count.
2081	<	// We can intermix failed removes or adjusts with steal updates
2082	<	do {
1184	<	long s, c;
1185	<	int g;
1186	<	if (steps == 0 && ((g = scanGuard) & SG_UNIT) == 0 &&
1187	<	UNSAFE.compareAndSwapInt(this, scanGuardOffset,
1188	<	g, g |= SG_UNIT)) {
1189	<	ForkJoinWorkerThread[] ws = workers;
1190	<	if (ws != null && idx >= 0 &&
1191	<	idx < ws.length && ws[idx] == w)
1192	<	ws[idx] = null;    // verify
1193	<	nextWorkerIndex = idx;
1194	<	scanGuard = g + SG_UNIT;
1195	<	steps = 1;
1196	<	}
1197	<	if (steps == 1 &&
1198	<	UNSAFE.compareAndSwapLong(this, ctlOffset, c = ctl,
1199	<	(((c - AC_UNIT) & AC_MASK) |
1200	<	((c - TC_UNIT) & TC_MASK) |
1201	<	(c & ~(AC_MASK|TC_MASK)))))
1202	<	steps = 2;
1203	<	if (sc != 0 &&
1204	<	UNSAFE.compareAndSwapLong(this, stealCountOffset,
1205	<	s = stealCount, s + sc))
1206	<	sc = 0;
1207	<	} while (steps != 2 || sc != 0);
1208	<	if (!tryTerminate(false)) {
1209	<	if (ex != null)   // possibly replace if died abnormally
1210	<	signalWork();
1211	<	else
1212	<	tryReleaseWaiter();
2070	>	* Gets and removes a local or stolen task for the given worker.
2071	>	*
2072	>	* @return a task, if available
2073	>	*/
2074	>	final ForkJoinTask<?> nextTaskFor(WorkQueue w) {
2075	>	for (ForkJoinTask<?> t;;) {
2076	>	WorkQueue q; int b;
2077	>	if ((t = w.nextLocalTask()) != null)
2078	>	return t;
2079	>	if ((q = findNonEmptyStealQueue(w.nextSeed())) == null)
2080	>	return null;
2081	>	if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null)
2082	>	return t;
2083		}
2084		}
2085
1216	–	// Shutdown and termination
1217	–
2086		/**
2087	<	* Possibly initiates and/or completes termination.
2087	>	* Returns a cheap heuristic guide for task partitioning when
2088	>	* programmers, frameworks, tools, or languages have little or no
2089	>	* idea about task granularity.  In essence by offering this
2090	>	* method, we ask users only about tradeoffs in overhead vs
2091	>	* expected throughput and its variance, rather than how finely to
2092	>	* partition tasks.
2093	>	*
2094	>	* In a steady state strict (tree-structured) computation, each
2095	>	* thread makes available for stealing enough tasks for other
2096	>	* threads to remain active. Inductively, if all threads play by
2097	>	* the same rules, each thread should make available only a
2098	>	* constant number of tasks.
2099	>	*
2100	>	* The minimum useful constant is just 1. But using a value of 1
2101	>	* would require immediate replenishment upon each steal to
2102	>	* maintain enough tasks, which is infeasible.  Further,
2103	>	* partitionings/granularities of offered tasks should minimize
2104	>	* steal rates, which in general means that threads nearer the top
2105	>	* of computation tree should generate more than those nearer the
2106	>	* bottom. In perfect steady state, each thread is at
2107	>	* approximately the same level of computation tree. However,
2108	>	* producing extra tasks amortizes the uncertainty of progress and
2109	>	* diffusion assumptions.
2110	>	*
2111	>	* So, users will want to use values larger, but not much larger
2112	>	* than 1 to both smooth over transient shortages and hedge
2113	>	* against uneven progress; as traded off against the cost of
2114	>	* extra task overhead. We leave the user to pick a threshold
2115	>	* value to compare with the results of this call to guide
2116	>	* decisions, but recommend values such as 3.
2117	>	*
2118	>	* When all threads are active, it is on average OK to estimate
2119	>	* surplus strictly locally. In steady-state, if one thread is
2120	>	* maintaining say 2 surplus tasks, then so are others. So we can
2121	>	* just use estimated queue length.  However, this strategy alone
2122	>	* leads to serious mis-estimates in some non-steady-state
2123	>	* conditions (ramp-up, ramp-down, other stalls). We can detect
2124	>	* many of these by further considering the number of "idle"
2125	>	* threads, that are known to have zero queued tasks, so
2126	>	* compensate by a factor of (#idle/#active) threads.
2127	>	*
2128	>	* Note: The approximation of #busy workers as #active workers is
2129	>	* not very good under current signalling scheme, and should be
2130	>	* improved.
2131	>	*/
2132	>	static int getSurplusQueuedTaskCount() {
2133	>	Thread t; ForkJoinWorkerThread wt; ForkJoinPool pool; WorkQueue q;
2134	>	if (((t = Thread.currentThread()) instanceof ForkJoinWorkerThread)) {
2135	>	int b = (q = (wt = (ForkJoinWorkerThread)t).workQueue).base;
2136	>	int p = (pool = wt.pool).parallelism;
2137	>	int a = (int)(pool.ctl >> AC_SHIFT) + p;
2138	>	return q.top - b - (a > (p >>>= 1) ? 0 :
2139	>	a > (p >>>= 1) ? 1 :
2140	>	a > (p >>>= 1) ? 2 :
2141	>	a > (p >>>= 1) ? 4 :
2142	>	8);
2143	>	}
2144	>	return 0;
2145	>	}
2146	>
2147	>	//  Termination
2148	>
2149	>	/**
2150	>	* Possibly initiates and/or completes termination.  The caller
2151	>	* triggering termination runs three passes through workQueues:
2152	>	* (0) Setting termination status, followed by wakeups of queued
2153	>	* workers; (1) cancelling all tasks; (2) interrupting lagging
2154	>	* threads (likely in external tasks, but possibly also blocked in
2155	>	* joins).  Each pass repeats previous steps because of potential
2156	>	* lagging thread creation.
2157		*
2158		* @param now if true, unconditionally terminate, else only
2159	<	* if shutdown and empty queue and no active workers
2159	>	* if no work and no active workers
2160	>	* @param enable if true, enable shutdown when next possible
2161		* @return true if now terminating or terminated
2162		*/
2163	<	private boolean tryTerminate(boolean now) {
2164	<	long c;
2165	<	while (((c = ctl) & STOP_BIT) == 0) {
2166	<	if (!now) {
2167	<	if ((int)(c >> AC_SHIFT) != -parallelism)
2168	<	return false;
2169	<	if (!shutdown || blockedCount != 0 || quiescerCount != 0 ||
2170	<	queueBase != queueTop) {
2171	<	if (ctl == c) // staleness check
1234	<	return false;
1235	<	continue;
2163	>	private boolean tryTerminate(boolean now, boolean enable) {
2164	>	if (this == commonPool)                     // cannot shut down
2165	>	return false;
2166	>	for (long c;;) {
2167	>	if (((c = ctl) & STOP_BIT) != 0) {      // already terminating
2168	>	if ((short)(c >>> TC_SHIFT) == -parallelism) {
2169	>	synchronized (this) {
2170	>	notifyAll();                // signal when 0 workers
2171	>	}
2172		}
2173	+	return true;
2174		}
2175	<	if (UNSAFE.compareAndSwapLong(this, ctlOffset, c, c | STOP_BIT))
2176	<	startTerminating();
2177	<	}
2178	<	if ((short)(c >>> TC_SHIFT) == -parallelism) { // signal when 0 workers
2179	<	final ReentrantLock lock = this.submissionLock;
2180	<	lock.lock();
2181	<	try {
2182	<	termination.signalAll();
2183	<	} finally {
2184	<	lock.unlock();
2175	>	if (plock >= 0) {                       // not yet enabled
2176	>	int ps;
2177	>	if (!enable)
2178	>	return false;
2179	>	if (((ps = plock) & PL_LOCK) != 0 ||
2180	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
2181	>	ps = acquirePlock();
2182	>	int nps = SHUTDOWN;
2183	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
2184	>	releasePlock(nps);
2185	>	}
2186	>	if (!now) {                             // check if idle & no tasks
2187	>	if ((int)(c >> AC_SHIFT) != -parallelism ||
2188	>	hasQueuedSubmissions())
2189	>	return false;
2190	>	// Check for unqueued inactive workers. One pass suffices.
2191	>	WorkQueue[] ws = workQueues; WorkQueue w;
2192	>	if (ws != null) {
2193	>	for (int i = 1; i < ws.length; i += 2) {
2194	>	if ((w = ws[i]) != null && w.eventCount >= 0)
2195	>	return false;
2196	>	}
2197	>	}
2198		}
2199	<	}
2200	<	return true;
2201	<	}
2202	<
2203	<	/**
2204	<	* Runs up to three passes through workers: (0) Setting
2205	<	* termination status for each worker, followed by wakeups up to
2206	<	* queued workers; (1) helping cancel tasks; (2) interrupting
2207	<	* lagging threads (likely in external tasks, but possibly also
2208	<	* blocked in joins).  Each pass repeats previous steps because of
2209	<	* potential lagging thread creation.
2210	<	*/
2211	<	private void startTerminating() {
1262	<	cancelSubmissions();
1263	<	for (int pass = 0; pass < 3; ++pass) {
1264	<	ForkJoinWorkerThread[] ws = workers;
1265	<	if (ws != null) {
1266	<	for (ForkJoinWorkerThread w : ws) {
1267	<	if (w != null) {
1268	<	w.terminate = true;
1269	<	if (pass > 0) {
1270	<	w.cancelTasks();
1271	<	if (pass > 1 && !w.isInterrupted()) {
1272	<	try {
1273	<	w.interrupt();
1274	<	} catch (SecurityException ignore) {
2199	>	if (U.compareAndSwapLong(this, CTL, c, c | STOP_BIT)) {
2200	>	for (int pass = 0; pass < 3; ++pass) {
2201	>	WorkQueue[] ws = workQueues;
2202	>	if (ws != null) {
2203	>	WorkQueue w;
2204	>	int n = ws.length;
2205	>	for (int i = 0; i < n; ++i) {
2206	>	if ((w = ws[i]) != null) {
2207	>	w.qlock = -1;
2208	>	if (pass > 0) {
2209	>	w.cancelAll();
2210	>	if (pass > 1)
2211	>	w.interruptOwner();
2212		}
2213		}
2214		}
2215	+	// Wake up workers parked on event queue
2216	+	int i, e; long cc; Thread p;
2217	+	while ((e = (int)(cc = ctl) & E_MASK) != 0 &&
2218	+	(i = e & SMASK) < n &&
2219	+	(w = ws[i]) != null) {
2220	+	long nc = ((long)(w.nextWait & E_MASK) |
2221	+	((cc + AC_UNIT) & AC_MASK) |
2222	+	(cc & (TC_MASK|STOP_BIT)));
2223	+	if (w.eventCount == (e | INT_SIGN) &&
2224	+	U.compareAndSwapLong(this, CTL, cc, nc)) {
2225	+	w.eventCount = (e + E_SEQ) & E_MASK;
2226	+	w.qlock = -1;
2227	+	if ((p = w.parker) != null)
2228	+	U.unpark(p);
2229	+	}
2230	+	}
2231		}
2232		}
1280	–	terminateWaiters();
2233		}
2234		}
2235		}
2236
2237	+	// external operations on common pool
2238	+
2239		/**
2240	<	* Polls and cancels all submissions. Called only during termination.
2240	>	* Returns common pool queue for a thread that has submitted at
2241	>	* least one task.
2242		*/
2243	<	private void cancelSubmissions() {
2244	<	while (queueBase != queueTop) {
2245	<	ForkJoinTask<?> task = pollSubmission();
2246	<	if (task != null) {
2247	<	try {
2248	<	task.cancel(false);
2249	<	} catch (Throwable ignore) {
2250	<	}
2243	>	static WorkQueue commonSubmitterQueue() {
2244	>	ForkJoinPool p; WorkQueue[] ws; int m; Submitter z;
2245	>	return ((z = submitters.get()) != null &&
2246	>	(p = commonPool) != null &&
2247	>	(ws = p.workQueues) != null &&
2248	>	(m = ws.length - 1) >= 0) ?
2249	>	ws[m & z.seed & SQMASK] : null;
2250	>	}
2251	>
2252	>	/**
2253	>	* Tries to pop the given task from submitter's queue in common pool.
2254	>	*/
2255	>	static boolean tryExternalUnpush(ForkJoinTask<?> t) {
2256	>	ForkJoinPool p; WorkQueue[] ws; WorkQueue q; Submitter z;
2257	>	ForkJoinTask<?>[] a;  int m, s; long j;
2258	>	if ((z = submitters.get()) != null &&
2259	>	(p = commonPool) != null &&
2260	>	(ws = p.workQueues) != null &&
2261	>	(m = ws.length - 1) >= 0 &&
2262	>	(q = ws[m & z.seed & SQMASK]) != null &&
2263	>	(s = q.top) != q.base &&
2264	>	(a = q.array) != null &&
2265	>	U.getObjectVolatile
2266	>	(a, j = (((a.length - 1) & (s - 1)) << ASHIFT) + ABASE) == t &&
2267	>	U.compareAndSwapInt(q, QLOCK, 0, 1)) {
2268	>	if (q.array == a && q.top == s && // recheck
2269	>	U.compareAndSwapObject(a, j, t, null)) {
2270	>	q.top = s - 1;
2271	>	q.qlock = 0;
2272	>	return true;
2273		}
2274	+	q.qlock = 0;
2275		}
2276	+	return false;
2277		}
2278
2279		/**
2280	<	* Tries to set the termination status of waiting workers, and
2281	<	* then wakes them up (after which they will terminate).
2282	<	*/
2283	<	private void terminateWaiters() {
2284	<	ForkJoinWorkerThread[] ws = workers;
2285	<	if (ws != null) {
2286	<	ForkJoinWorkerThread w; long c; int i, e;
2287	<	int n = ws.length;
2288	<	while ((i = ~(e = (int)(c = ctl)) & SMASK) < n &&
2289	<	(w = ws[i]) != null && w.eventCount == (e & E_MASK)) {
2290	<	if (UNSAFE.compareAndSwapLong(this, ctlOffset, c,
2291	<	(long)(w.nextWait & E_MASK) |
2292	<	((c + AC_UNIT) & AC_MASK) |
2293	<	(c & (TC_MASK|STOP_BIT)))) {
2294	<	w.terminate = true;
2295	<	w.eventCount = e + EC_UNIT;
2296	<	if (w.parked)
2297	<	UNSAFE.unpark(w);
2280	>	* Tries to pop and run local tasks within the same computation
2281	>	* as the given root. On failure, tries to help complete from
2282	>	* other queues via helpComplete.
2283	>	*/
2284	>	private void externalHelpComplete(WorkQueue q, ForkJoinTask<?> root) {
2285	>	ForkJoinTask<?>[] a; int m;
2286	>	if (q != null && (a = q.array) != null && (m = (a.length - 1)) >= 0 &&
2287	>	root != null && root.status >= 0) {
2288	>	for (;;) {
2289	>	int s; Object o; CountedCompleter<?> task = null;
2290	>	if ((s = q.top) - q.base > 0) {
2291	>	long j = ((m & (s - 1)) << ASHIFT) + ABASE;
2292	>	if ((o = U.getObject(a, j)) != null &&
2293	>	(o instanceof CountedCompleter)) {
2294	>	CountedCompleter<?> t = (CountedCompleter<?>)o, r = t;
2295	>	do {
2296	>	if (r == root) {
2297	>	if (U.compareAndSwapInt(q, QLOCK, 0, 1)) {
2298	>	if (q.array == a && q.top == s &&
2299	>	U.compareAndSwapObject(a, j, t, null)) {
2300	>	q.top = s - 1;
2301	>	task = t;
2302	>	}
2303	>	q.qlock = 0;
2304	>	}
2305	>	break;
2306	>	}
2307	>	} while ((r = r.completer) != null);
2308	>	}
2309	>	}
2310	>	if (task != null)
2311	>	task.doExec();
2312	>	if (root.status < 0 || (int)(ctl >> AC_SHIFT) >= 0)
2313	>	break;
2314	>	if (task == null) {
2315	>	if (helpSignal(root, q.poolIndex) >= 0)
2316	>	helpComplete(root, SHARED_QUEUE);
2317	>	break;
2318		}
2319		}
2320		}
2321		}
2322
1324	–	// misc ForkJoinWorkerThread support
1325	–
2323		/**
2324	<	* Increment or decrement quiescerCount. Needed only to prevent
2325	<	* triggering shutdown if a worker is transiently inactive while
1329	<	* checking quiescence.
1330	<	*
1331	<	* @param delta 1 for increment, -1 for decrement
2324	>	* Tries to help execute or signal availability of the given task
2325	>	* from submitter's queue in common pool.
2326		*/
2327	<	final void addQuiescerCount(int delta) {
2328	<	int c;
2329	<	do {} while (!UNSAFE.compareAndSwapInt(this, quiescerCountOffset,
2330	<	c = quiescerCount, c + delta));
2327	>	static void externalHelpJoin(ForkJoinTask<?> t) {
2328	>	// Some hard-to-avoid overlap with tryExternalUnpush
2329	>	ForkJoinPool p; WorkQueue[] ws; WorkQueue q, w; Submitter z;
2330	>	ForkJoinTask<?>[] a;  int m, s, n; long j;
2331	>	if (t != null && t.status >= 0 &&
2332	>	(z = submitters.get()) != null &&
2333	>	(p = commonPool) != null &&
2334	>	(ws = p.workQueues) != null &&
2335	>	(m = ws.length - 1) >= 0 &&
2336	>	(q = ws[m & z.seed & SQMASK]) != null &&
2337	>	(a = q.array) != null) {
2338	>	if ((s = q.top) != q.base &&
2339	>	U.getObjectVolatile
2340	>	(a, j = (((a.length - 1) & (s - 1)) << ASHIFT) + ABASE) == t &&
2341	>	U.compareAndSwapInt(q, QLOCK, 0, 1)) {
2342	>	if (q.array == a && q.top == s &&
2343	>	U.compareAndSwapObject(a, j, t, null)) {
2344	>	q.top = s - 1;
2345	>	q.qlock = 0;
2346	>	t.doExec();
2347	>	}
2348	>	else
2349	>	q.qlock = 0;
2350	>	}
2351	>	if (t.status >= 0) {
2352	>	if (t instanceof CountedCompleter)
2353	>	p.externalHelpComplete(q, t);
2354	>	else
2355	>	p.helpSignal(t, q.poolIndex);
2356	>	}
2357	>	}
2358		}
2359
2360		/**
2361	<	* Directly increment or decrement active count without
1341	<	* queuing. This method is used to transiently assert inactivation
1342	<	* while checking quiescence.
1343	<	*
1344	<	* @param delta 1 for increment, -1 for decrement
2361	>	* Restricted version of helpQuiescePool for external callers
2362		*/
2363	<	final void addActiveCount(int delta) {
2364	<	long d = delta < 0 ? -AC_UNIT : AC_UNIT;
2365	<	long c;
2366	<	do {} while (!UNSAFE.compareAndSwapLong(this, ctlOffset, c = ctl,
2367	<	((c + d) & AC_MASK) |
2368	<	(c & ~AC_MASK)));
2369	<	}
2370	<
1354	<	/**
1355	<	* Returns the approximate (non-atomic) number of idle threads per
1356	<	* active thread.
1357	<	*/
1358	<	final int idlePerActive() {
1359	<	// Approximate at powers of two for small values, saturate past 4
1360	<	int p = parallelism;
1361	<	int a = p + (int)(ctl >> AC_SHIFT);
1362	<	return (a > (p >>>= 1) ? 0 :
1363	<	a > (p >>>= 1) ? 1 :
1364	<	a > (p >>>= 1) ? 2 :
1365	<	a > (p >>>= 1) ? 4 :
1366	<	8);
2363	>	static void externalHelpQuiescePool() {
2364	>	ForkJoinPool p; ForkJoinTask<?> t; WorkQueue q; int b;
2365	>	int r = ThreadLocalRandom.current().nextInt();
2366	>	if ((p = commonPool) != null &&
2367	>	(q = p.findNonEmptyStealQueue(r)) != null &&
2368	>	(b = q.base) - q.top < 0 &&
2369	>	(t = q.pollAt(b)) != null)
2370	>	t.doExec();
2371		}
2372
2373		// Exported methods
#	Line 1435 | Line 2439 | public class ForkJoinPool extends Abstra
2439		checkPermission();
2440		if (factory == null)
2441		throw new NullPointerException();
2442	<	if (parallelism <= 0 || parallelism > MAX_ID)
2442	>	if (parallelism <= 0 || parallelism > MAX_CAP)
2443		throw new IllegalArgumentException();
2444		this.parallelism = parallelism;
2445		this.factory = factory;
2446		this.ueh = handler;
2447	<	this.locallyFifo = asyncMode;
2447	>	this.localMode = asyncMode ? FIFO_QUEUE : LIFO_QUEUE;
2448		long np = (long)(-parallelism); // offset ctl counts
2449		this.ctl = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK);
2450	<	this.submissionQueue = new ForkJoinTask<?>[INITIAL_QUEUE_CAPACITY];
1447	<	// initialize workers array with room for 2*parallelism if possible
1448	<	int n = parallelism << 1;
1449	<	if (n >= MAX_ID)
1450	<	n = MAX_ID;
1451	<	else { // See Hackers Delight, sec 3.2, where n < (1 << 16)
1452	<	n |= n >>> 1; n |= n >>> 2; n |= n >>> 4; n |= n >>> 8;
1453	<	}
1454	<	workers = new ForkJoinWorkerThread[n + 1];
1455	<	this.submissionLock = new ReentrantLock();
1456	<	this.termination = submissionLock.newCondition();
2450	>	int pn = nextPoolId();
2451		StringBuilder sb = new StringBuilder("ForkJoinPool-");
2452	<	sb.append(poolNumberGenerator.incrementAndGet());
2452	>	sb.append(Integer.toString(pn));
2453		sb.append("-worker-");
2454		this.workerNamePrefix = sb.toString();
2455		}
2456
2457	+	/**
2458	+	* Constructor for common pool, suitable only for static initialization.
2459	+	* Basically the same as above, but uses smallest possible initial footprint.
2460	+	*/
2461	+	ForkJoinPool(int parallelism, long ctl,
2462	+	ForkJoinWorkerThreadFactory factory,
2463	+	Thread.UncaughtExceptionHandler handler) {
2464	+	this.parallelism = parallelism;
2465	+	this.ctl = ctl;
2466	+	this.factory = factory;
2467	+	this.ueh = handler;
2468	+	this.localMode = LIFO_QUEUE;
2469	+	this.workerNamePrefix = "ForkJoinPool.commonPool-worker-";
2470	+	}
2471	+
2472	+	/**
2473	+	* Returns the common pool instance.
2474	+	*
2475	+	* @return the common pool instance
2476	+	*/
2477	+	public static ForkJoinPool commonPool() {
2478	+	return commonPool; // cannot be null (if so, a static init error)
2479	+	}
2480	+
2481		// Execution methods
2482
2483		/**
#	Line 1479 | Line 2497 | public class ForkJoinPool extends Abstra
2497		*         scheduled for execution
2498		*/
2499		public <T> T invoke(ForkJoinTask<T> task) {
1482	–	Thread t = Thread.currentThread();
2500		if (task == null)
2501		throw new NullPointerException();
2502	<	if (shutdown)
2503	<	throw new RejectedExecutionException();
1487	<	if ((t instanceof ForkJoinWorkerThread) &&
1488	<	((ForkJoinWorkerThread)t).pool == this)
1489	<	return task.invoke();  // bypass submit if in same pool
1490	<	else {
1491	<	addSubmission(task);
1492	<	return task.join();
1493	<	}
1494	<	}
1495	<
1496	<	/**
1497	<	* Unless terminating, forks task if within an ongoing FJ
1498	<	* computation in the current pool, else submits as external task.
1499	<	*/
1500	<	private <T> void forkOrSubmit(ForkJoinTask<T> task) {
1501	<	ForkJoinWorkerThread w;
1502	<	Thread t = Thread.currentThread();
1503	<	if (shutdown)
1504	<	throw new RejectedExecutionException();
1505	<	if ((t instanceof ForkJoinWorkerThread) &&
1506	<	(w = (ForkJoinWorkerThread)t).pool == this)
1507	<	w.pushTask(task);
1508	<	else
1509	<	addSubmission(task);
2502	>	externalPush(task);
2503	>	return task.join();
2504		}
2505
2506		/**
#	Line 1520 | Line 2514 | public class ForkJoinPool extends Abstra
2514		public void execute(ForkJoinTask<?> task) {
2515		if (task == null)
2516		throw new NullPointerException();
2517	<	forkOrSubmit(task);
2517	>	externalPush(task);
2518		}
2519
2520		// AbstractExecutorService methods
#	Line 1537 | Line 2531 | public class ForkJoinPool extends Abstra
2531		if (task instanceof ForkJoinTask<?>) // avoid re-wrap
2532		job = (ForkJoinTask<?>) task;
2533		else
2534	<	job = ForkJoinTask.adapt(task, null);
2535	<	forkOrSubmit(job);
2534	>	job = new ForkJoinTask.AdaptedRunnableAction(task);
2535	>	externalPush(job);
2536		}
2537
2538		/**
#	Line 1553 | Line 2547 | public class ForkJoinPool extends Abstra
2547		public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) {
2548		if (task == null)
2549		throw new NullPointerException();
2550	<	forkOrSubmit(task);
2550	>	externalPush(task);
2551		return task;
2552		}
2553
#	Line 1563 | Line 2557 | public class ForkJoinPool extends Abstra
2557		*         scheduled for execution
2558		*/
2559		public <T> ForkJoinTask<T> submit(Callable<T> task) {
2560	<	if (task == null)
2561	<	throw new NullPointerException();
1568	<	ForkJoinTask<T> job = ForkJoinTask.adapt(task);
1569	<	forkOrSubmit(job);
2560	>	ForkJoinTask<T> job = new ForkJoinTask.AdaptedCallable<T>(task);
2561	>	externalPush(job);
2562		return job;
2563		}
2564
#	Line 1576 | Line 2568 | public class ForkJoinPool extends Abstra
2568		*         scheduled for execution
2569		*/
2570		public <T> ForkJoinTask<T> submit(Runnable task, T result) {
2571	<	if (task == null)
2572	<	throw new NullPointerException();
1581	<	ForkJoinTask<T> job = ForkJoinTask.adapt(task, result);
1582	<	forkOrSubmit(job);
2571	>	ForkJoinTask<T> job = new ForkJoinTask.AdaptedRunnable<T>(task, result);
2572	>	externalPush(job);
2573		return job;
2574		}
2575
#	Line 1595 | Line 2585 | public class ForkJoinPool extends Abstra
2585		if (task instanceof ForkJoinTask<?>) // avoid re-wrap
2586		job = (ForkJoinTask<?>) task;
2587		else
2588	<	job = ForkJoinTask.adapt(task, null);
2589	<	forkOrSubmit(job);
2588	>	job = new ForkJoinTask.AdaptedRunnableAction(task);
2589	>	externalPush(job);
2590		return job;
2591		}
2592
#	Line 1605 | Line 2595 | public class ForkJoinPool extends Abstra
2595		* @throws RejectedExecutionException {@inheritDoc}
2596		*/
2597		public <T> List<Future<T>> invokeAll(Collection<? extends Callable<T>> tasks) {
2598	<	ArrayList<ForkJoinTask<T>> forkJoinTasks =
2599	<	new ArrayList<ForkJoinTask<T>>(tasks.size());
2600	<	for (Callable<T> task : tasks)
2601	<	forkJoinTasks.add(ForkJoinTask.adapt(task));
2602	<	invoke(new InvokeAll<T>(forkJoinTasks));
2603	<
2598	>	// In previous versions of this class, this method constructed
2599	>	// a task to run ForkJoinTask.invokeAll, but now external
2600	>	// invocation of multiple tasks is at least as efficient.
2601	>	List<ForkJoinTask<T>> fs = new ArrayList<ForkJoinTask<T>>(tasks.size());
2602	>	// Workaround needed because method wasn't declared with
2603	>	// wildcards in return type but should have been.
2604		@SuppressWarnings({"unchecked", "rawtypes"})
2605	<	List<Future<T>> futures = (List<Future<T>>) (List) forkJoinTasks;
1616	<	return futures;
1617	<	}
2605	>	List<Future<T>> futures = (List<Future<T>>) (List) fs;
2606
2607	<	static final class InvokeAll<T> extends RecursiveAction {
2608	<	final ArrayList<ForkJoinTask<T>> tasks;
2609	<	InvokeAll(ArrayList<ForkJoinTask<T>> tasks) { this.tasks = tasks; }
2610	<	public void compute() {
2611	<	try { invokeAll(tasks); }
2612	<	catch (Exception ignore) {}
2607	>	boolean done = false;
2608	>	try {
2609	>	for (Callable<T> t : tasks) {
2610	>	ForkJoinTask<T> f = new ForkJoinTask.AdaptedCallable<T>(t);
2611	>	externalPush(f);
2612	>	fs.add(f);
2613	>	}
2614	>	for (ForkJoinTask<T> f : fs)
2615	>	f.quietlyJoin();
2616	>	done = true;
2617	>	return futures;
2618	>	} finally {
2619	>	if (!done)
2620	>	for (ForkJoinTask<T> f : fs)
2621	>	f.cancel(false);
2622		}
1626	–	private static final long serialVersionUID = -7914297376763021607L;
2623		}
2624
2625		/**
#	Line 1655 | Line 2651 | public class ForkJoinPool extends Abstra
2651		}
2652
2653		/**
2654	+	* Returns the targeted parallelism level of the common pool.
2655	+	*
2656	+	* @return the targeted parallelism level of the common pool
2657	+	*/
2658	+	public static int getCommonPoolParallelism() {
2659	+	return commonPoolParallelism;
2660	+	}
2661	+
2662	+	/**
2663		* Returns the number of worker threads that have started but not
2664		* yet terminated.  The result returned by this method may differ
2665		* from {@link #getParallelism} when threads are created to
#	Line 1673 | Line 2678 | public class ForkJoinPool extends Abstra
2678		* @return {@code true} if this pool uses async mode
2679		*/
2680		public boolean getAsyncMode() {
2681	<	return locallyFifo;
2681	>	return localMode != 0;
2682		}
2683
2684		/**
#	Line 1685 | Line 2690 | public class ForkJoinPool extends Abstra
2690		* @return the number of worker threads
2691		*/
2692		public int getRunningThreadCount() {
2693	<	int r = parallelism + (int)(ctl >> AC_SHIFT);
2694	<	return (r <= 0) ? 0 : r; // suppress momentarily negative values
2693	>	int rc = 0;
2694	>	WorkQueue[] ws; WorkQueue w;
2695	>	if ((ws = workQueues) != null) {
2696	>	for (int i = 1; i < ws.length; i += 2) {
2697	>	if ((w = ws[i]) != null && w.isApparentlyUnblocked())
2698	>	++rc;
2699	>	}
2700	>	}
2701	>	return rc;
2702		}
2703
2704		/**
#	Line 1697 | Line 2709 | public class ForkJoinPool extends Abstra
2709		* @return the number of active threads
2710		*/
2711		public int getActiveThreadCount() {
2712	<	int r = parallelism + (int)(ctl >> AC_SHIFT) + blockedCount;
2712	>	int r = parallelism + (int)(ctl >> AC_SHIFT);
2713		return (r <= 0) ? 0 : r; // suppress momentarily negative values
2714		}
2715
#	Line 1713 | Line 2725 | public class ForkJoinPool extends Abstra
2725		* @return {@code true} if all threads are currently idle
2726		*/
2727		public boolean isQuiescent() {
2728	<	return parallelism + (int)(ctl >> AC_SHIFT) + blockedCount == 0;
2728	>	return (int)(ctl >> AC_SHIFT) + parallelism == 0;
2729		}
2730
2731		/**
#	Line 1728 | Line 2740 | public class ForkJoinPool extends Abstra
2740		* @return the number of steals
2741		*/
2742		public long getStealCount() {
2743	<	return stealCount;
2743	>	long count = stealCount;
2744	>	WorkQueue[] ws; WorkQueue w;
2745	>	if ((ws = workQueues) != null) {
2746	>	for (int i = 1; i < ws.length; i += 2) {
2747	>	if ((w = ws[i]) != null)
2748	>	count += w.nsteals;
2749	>	}
2750	>	}
2751	>	return count;
2752		}
2753
2754		/**
#	Line 1743 | Line 2763 | public class ForkJoinPool extends Abstra
2763		*/
2764		public long getQueuedTaskCount() {
2765		long count = 0;
2766	<	ForkJoinWorkerThread[] ws;
2767	<	if ((short)(ctl >>> TC_SHIFT) > -parallelism &&
2768	<	(ws = workers) != null) {
2769	<	for (ForkJoinWorkerThread w : ws)
2770	<	if (w != null)
2771	<	count -= w.queueBase - w.queueTop; // must read base first
2766	>	WorkQueue[] ws; WorkQueue w;
2767	>	if ((ws = workQueues) != null) {
2768	>	for (int i = 1; i < ws.length; i += 2) {
2769	>	if ((w = ws[i]) != null)
2770	>	count += w.queueSize();
2771	>	}
2772		}
2773		return count;
2774		}
#	Line 1761 | Line 2781 | public class ForkJoinPool extends Abstra
2781		* @return the number of queued submissions
2782		*/
2783		public int getQueuedSubmissionCount() {
2784	<	return -queueBase + queueTop;
2784	>	int count = 0;
2785	>	WorkQueue[] ws; WorkQueue w;
2786	>	if ((ws = workQueues) != null) {
2787	>	for (int i = 0; i < ws.length; i += 2) {
2788	>	if ((w = ws[i]) != null)
2789	>	count += w.queueSize();
2790	>	}
2791	>	}
2792	>	return count;
2793		}
2794
2795		/**
#	Line 1771 | Line 2799 | public class ForkJoinPool extends Abstra
2799		* @return {@code true} if there are any queued submissions
2800		*/
2801		public boolean hasQueuedSubmissions() {
2802	<	return queueBase != queueTop;
2802	>	WorkQueue[] ws; WorkQueue w;
2803	>	if ((ws = workQueues) != null) {
2804	>	for (int i = 0; i < ws.length; i += 2) {
2805	>	if ((w = ws[i]) != null && w.queueSize() != 0)
2806	>	return true;
2807	>	}
2808	>	}
2809	>	return false;
2810		}
2811
2812		/**
#	Line 1782 | Line 2817 | public class ForkJoinPool extends Abstra
2817		* @return the next submission, or {@code null} if none
2818		*/
2819		protected ForkJoinTask<?> pollSubmission() {
2820	<	ForkJoinTask<?> t; ForkJoinTask<?>[] q; int b, i;
2821	<	while ((b = queueBase) != queueTop &&
2822	<	(q = submissionQueue) != null &&
2823	<	(i = (q.length - 1) & b) >= 0) {
2824	<	long u = (i << ASHIFT) + ABASE;
1790	<	if ((t = q[i]) != null &&
1791	<	queueBase == b &&
1792	<	UNSAFE.compareAndSwapObject(q, u, t, null)) {
1793	<	queueBase = b + 1;
1794	<	return t;
2820	>	WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2821	>	if ((ws = workQueues) != null) {
2822	>	for (int i = 0; i < ws.length; i += 2) {
2823	>	if ((w = ws[i]) != null && (t = w.poll()) != null)
2824	>	return t;
2825		}
2826		}
2827		return null;
#	Line 1816 | Line 2846 | public class ForkJoinPool extends Abstra
2846		*/
2847		protected int drainTasksTo(Collection<? super ForkJoinTask<?>> c) {
2848		int count = 0;
2849	<	while (queueBase != queueTop) {
2850	<	ForkJoinTask<?> t = pollSubmission();
2851	<	if (t != null) {
2852	<	c.add(t);
2853	<	++count;
2849	>	WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2850	>	if ((ws = workQueues) != null) {
2851	>	for (int i = 0; i < ws.length; ++i) {
2852	>	if ((w = ws[i]) != null) {
2853	>	while ((t = w.poll()) != null) {
2854	>	c.add(t);
2855	>	++count;
2856	>	}
2857	>	}
2858		}
2859		}
1826	–	ForkJoinWorkerThread[] ws;
1827	–	if ((short)(ctl >>> TC_SHIFT) > -parallelism &&
1828	–	(ws = workers) != null) {
1829	–	for (ForkJoinWorkerThread w : ws)
1830	–	if (w != null)
1831	–	count += w.drainTasksTo(c);
1832	–	}
2860		return count;
2861		}
2862
#	Line 1841 | Line 2868 | public class ForkJoinPool extends Abstra
2868		* @return a string identifying this pool, as well as its state
2869		*/
2870		public String toString() {
2871	<	long st = getStealCount();
2872	<	long qt = getQueuedTaskCount();
2873	<	long qs = getQueuedSubmissionCount();
1847	<	int pc = parallelism;
2871	>	// Use a single pass through workQueues to collect counts
2872	>	long qt = 0L, qs = 0L; int rc = 0;
2873	>	long st = stealCount;
2874		long c = ctl;
2875	+	WorkQueue[] ws; WorkQueue w;
2876	+	if ((ws = workQueues) != null) {
2877	+	for (int i = 0; i < ws.length; ++i) {
2878	+	if ((w = ws[i]) != null) {
2879	+	int size = w.queueSize();
2880	+	if ((i & 1) == 0)
2881	+	qs += size;
2882	+	else {
2883	+	qt += size;
2884	+	st += w.nsteals;
2885	+	if (w.isApparentlyUnblocked())
2886	+	++rc;
2887	+	}
2888	+	}
2889	+	}
2890	+	}
2891	+	int pc = parallelism;
2892		int tc = pc + (short)(c >>> TC_SHIFT);
2893	<	int rc = pc + (int)(c >> AC_SHIFT);
2894	<	if (rc < 0) // ignore transient negative
2895	<	rc = 0;
1853	<	int ac = rc + blockedCount;
2893	>	int ac = pc + (int)(c >> AC_SHIFT);
2894	>	if (ac < 0) // ignore transient negative
2895	>	ac = 0;
2896		String level;
2897		if ((c & STOP_BIT) != 0)
2898		level = (tc == 0) ? "Terminated" : "Terminating";
2899		else
2900	<	level = shutdown ? "Shutting down" : "Running";
2900	>	level = plock < 0 ? "Shutting down" : "Running";
2901		return super.toString() +
2902		"[" + level +
2903		", parallelism = " + pc +
#	Line 1869 | Line 2911 | public class ForkJoinPool extends Abstra
2911		}
2912
2913		/**
2914	<	* Initiates an orderly shutdown in which previously submitted
2915	<	* tasks are executed, but no new tasks will be accepted.
2916	<	* Invocation has no additional effect if already shut down.
2917	<	* Tasks that are in the process of being submitted concurrently
2918	<	* during the course of this method may or may not be rejected.
2914	>	* Possibly initiates an orderly shutdown in which previously
2915	>	* submitted tasks are executed, but no new tasks will be
2916	>	* accepted. Invocation has no effect on execution state if this
2917	>	* is the {@link #commonPool}, and no additional effect if
2918	>	* already shut down.  Tasks that are in the process of being
2919	>	* submitted concurrently during the course of this method may or
2920	>	* may not be rejected.
2921		*
2922		* @throws SecurityException if a security manager exists and
2923		*         the caller is not permitted to modify threads
#	Line 1882 | Line 2926 | public class ForkJoinPool extends Abstra
2926		*/
2927		public void shutdown() {
2928		checkPermission();
2929	<	shutdown = true;
1886	<	tryTerminate(false);
2929	>	tryTerminate(false, true);
2930		}
2931
2932		/**
2933	<	* Attempts to cancel and/or stop all tasks, and reject all
2934	<	* subsequently submitted tasks.  Tasks that are in the process of
2935	<	* being submitted or executed concurrently during the course of
2936	<	* this method may or may not be rejected. This method cancels
2937	<	* both existing and unexecuted tasks, in order to permit
2938	<	* termination in the presence of task dependencies. So the method
2939	<	* always returns an empty list (unlike the case for some other
2940	<	* Executors).
2933	>	* Possibly attempts to cancel and/or stop all tasks, and reject
2934	>	* all subsequently submitted tasks.  Invocation has no effect on
2935	>	* execution state if this is the {@link #commonPool}, and no
2936	>	* additional effect if already shut down. Otherwise, tasks that
2937	>	* are in the process of being submitted or executed concurrently
2938	>	* during the course of this method may or may not be
2939	>	* rejected. This method cancels both existing and unexecuted
2940	>	* tasks, in order to permit termination in the presence of task
2941	>	* dependencies. So the method always returns an empty list
2942	>	* (unlike the case for some other Executors).
2943		*
2944		* @return an empty list
2945		* @throws SecurityException if a security manager exists and
#	Line 1904 | Line 2949 | public class ForkJoinPool extends Abstra
2949		*/
2950		public List<Runnable> shutdownNow() {
2951		checkPermission();
2952	<	shutdown = true;
1908	<	tryTerminate(true);
2952	>	tryTerminate(true, true);
2953		return Collections.emptyList();
2954		}
2955
#	Line 1940 | Line 2984 | public class ForkJoinPool extends Abstra
2984		}
2985
2986		/**
1943	–	* Returns true if terminating or terminated. Used by ForkJoinWorkerThread.
1944	–	*/
1945	–	final boolean isAtLeastTerminating() {
1946	–	return (ctl & STOP_BIT) != 0L;
1947	–	}
1948	–
1949	–	/**
2987		* Returns {@code true} if this pool has been shut down.
2988		*
2989		* @return {@code true} if this pool has been shut down
2990		*/
2991		public boolean isShutdown() {
2992	<	return shutdown;
2992	>	return plock < 0;
2993		}
2994
2995		/**
2996	<	* Blocks until all tasks have completed execution after a shutdown
2997	<	* request, or the timeout occurs, or the current thread is
2998	<	* interrupted, whichever happens first.
2996	>	* Blocks until all tasks have completed execution after a
2997	>	* shutdown request, or the timeout occurs, or the current thread
2998	>	* is interrupted, whichever happens first. Note that the {@link
2999	>	* #commonPool()} never terminates until program shutdown so
3000	>	* this method will always time out.
3001		*
3002		* @param timeout the maximum time to wait
3003		* @param unit the time unit of the timeout argument
#	Line 1969 | Line 3008 | public class ForkJoinPool extends Abstra
3008		public boolean awaitTermination(long timeout, TimeUnit unit)
3009		throws InterruptedException {
3010		long nanos = unit.toNanos(timeout);
3011	<	final ReentrantLock lock = this.submissionLock;
3012	<	lock.lock();
3013	<	try {
3014	<	for (;;) {
3015	<	if (isTerminated())
3016	<	return true;
3017	<	if (nanos <= 0)
3018	<	return false;
3019	<	nanos = termination.awaitNanos(nanos);
3011	>	if (isTerminated())
3012	>	return true;
3013	>	long startTime = System.nanoTime();
3014	>	boolean terminated = false;
3015	>	synchronized (this) {
3016	>	for (long waitTime = nanos, millis = 0L;;) {
3017	>	if (terminated = isTerminated() ||
3018	>	waitTime <= 0L ||
3019	>	(millis = unit.toMillis(waitTime)) <= 0L)
3020	>	break;
3021	>	wait(millis);
3022	>	waitTime = nanos - (System.nanoTime() - startTime);
3023		}
1982	–	} finally {
1983	–	lock.unlock();
3024		}
3025	+	return terminated;
3026		}
3027
3028		/**
#	Line 2081 | Line 3122 | public class ForkJoinPool extends Abstra
3122		throws InterruptedException {
3123		Thread t = Thread.currentThread();
3124		if (t instanceof ForkJoinWorkerThread) {
3125	<	ForkJoinWorkerThread w = (ForkJoinWorkerThread) t;
3126	<	w.pool.awaitBlocker(blocker);
3125	>	ForkJoinPool p = ((ForkJoinWorkerThread)t).pool;
3126	>	while (!blocker.isReleasable()) { // variant of helpSignal
3127	>	WorkQueue[] ws; WorkQueue q; int m, n;
3128	>	if ((ws = p.workQueues) != null && (m = ws.length - 1) >= 0) {
3129	>	for (int i = 0; i <= m; ++i) {
3130	>	if (blocker.isReleasable())
3131	>	return;
3132	>	if ((q = ws[i]) != null && (n = q.queueSize()) > 0) {
3133	>	p.signalWork(q, n);
3134	>	if ((int)(p.ctl >> AC_SHIFT) >= 0)
3135	>	break;
3136	>	}
3137	>	}
3138	>	}
3139	>	if (p.tryCompensate()) {
3140	>	try {
3141	>	do {} while (!blocker.isReleasable() &&
3142	>	!blocker.block());
3143	>	} finally {
3144	>	p.incrementActiveCount();
3145	>	}
3146	>	break;
3147	>	}
3148	>	}
3149		}
3150		else {
3151	<	do {} while (!blocker.isReleasable() && !blocker.block());
3151	>	do {} while (!blocker.isReleasable() &&
3152	>	!blocker.block());
3153		}
3154		}
3155
#	Line 2094 | Line 3158 | public class ForkJoinPool extends Abstra
3158		// implement RunnableFuture.
3159
3160		protected <T> RunnableFuture<T> newTaskFor(Runnable runnable, T value) {
3161	<	return (RunnableFuture<T>) ForkJoinTask.adapt(runnable, value);
3161	>	return new ForkJoinTask.AdaptedRunnable<T>(runnable, value);
3162		}
3163
3164		protected <T> RunnableFuture<T> newTaskFor(Callable<T> callable) {
3165	<	return (RunnableFuture<T>) ForkJoinTask.adapt(callable);
3165	>	return new ForkJoinTask.AdaptedCallable<T>(callable);
3166		}
3167
3168		// Unsafe mechanics
3169	<	private static final sun.misc.Unsafe UNSAFE;
3170	<	private static final long ctlOffset;
3171	<	private static final long stealCountOffset;
3172	<	private static final long blockedCountOffset;
2109	<	private static final long quiescerCountOffset;
2110	<	private static final long scanGuardOffset;
2111	<	private static final long nextWorkerNumberOffset;
2112	<	private static final long ABASE;
3169	>	private static final sun.misc.Unsafe U;
3170	>	private static final long CTL;
3171	>	private static final long PARKBLOCKER;
3172	>	private static final int ABASE;
3173		private static final int ASHIFT;
3174	+	private static final long STEALCOUNT;
3175	+	private static final long PLOCK;
3176	+	private static final long INDEXSEED;
3177	+	private static final long QLOCK;
3178
3179		static {
3180	<	poolNumberGenerator = new AtomicInteger();
3181	<	workerSeedGenerator = new Random();
3182	<	modifyThreadPermission = new RuntimePermission("modifyThread");
3183	<	defaultForkJoinWorkerThreadFactory =
3184	<	new DefaultForkJoinWorkerThreadFactory();
3185	<	int s;
3180	>	// Establish common pool parameters
3181	>	// TBD: limit or report ignored exceptions?
3182	>
3183	>	int par = 0;
3184	>	ForkJoinWorkerThreadFactory fac = null;
3185	>	Thread.UncaughtExceptionHandler handler = null;
3186	>	try {
3187	>	String pp = System.getProperty(propPrefix + "parallelism");
3188	>	String hp = System.getProperty(propPrefix + "exceptionHandler");
3189	>	String fp = System.getProperty(propPrefix + "threadFactory");
3190	>	if (fp != null)
3191	>	fac = ((ForkJoinWorkerThreadFactory)ClassLoader.
3192	>	getSystemClassLoader().loadClass(fp).newInstance());
3193	>	if (hp != null)
3194	>	handler = ((Thread.UncaughtExceptionHandler)ClassLoader.
3195	>	getSystemClassLoader().loadClass(hp).newInstance());
3196	>	if (pp != null)
3197	>	par = Integer.parseInt(pp);
3198	>	} catch (Exception ignore) {
3199	>	}
3200	>
3201	>	int s; // initialize field offsets for CAS etc
3202		try {
3203	<	UNSAFE = getUnsafe();
3204	<	Class k = ForkJoinPool.class;
3205	<	ctlOffset = UNSAFE.objectFieldOffset
3203	>	U = getUnsafe();
3204	>	Class<?> k = ForkJoinPool.class;
3205	>	CTL = U.objectFieldOffset
3206		(k.getDeclaredField("ctl"));
3207	<	stealCountOffset = UNSAFE.objectFieldOffset
3207	>	STEALCOUNT = U.objectFieldOffset
3208		(k.getDeclaredField("stealCount"));
3209	<	blockedCountOffset = UNSAFE.objectFieldOffset
3210	<	(k.getDeclaredField("blockedCount"));
3211	<	quiescerCountOffset = UNSAFE.objectFieldOffset
3212	<	(k.getDeclaredField("quiescerCount"));
3213	<	scanGuardOffset = UNSAFE.objectFieldOffset
3214	<	(k.getDeclaredField("scanGuard"));
3215	<	nextWorkerNumberOffset = UNSAFE.objectFieldOffset
3216	<	(k.getDeclaredField("nextWorkerNumber"));
3217	<	Class a = ForkJoinTask[].class;
3218	<	ABASE = UNSAFE.arrayBaseOffset(a);
3219	<	s = UNSAFE.arrayIndexScale(a);
3209	>	PLOCK = U.objectFieldOffset
3210	>	(k.getDeclaredField("plock"));
3211	>	INDEXSEED = U.objectFieldOffset
3212	>	(k.getDeclaredField("indexSeed"));
3213	>	Class<?> tk = Thread.class;
3214	>	PARKBLOCKER = U.objectFieldOffset
3215	>	(tk.getDeclaredField("parkBlocker"));
3216	>	Class<?> wk = WorkQueue.class;
3217	>	QLOCK = U.objectFieldOffset
3218	>	(wk.getDeclaredField("qlock"));
3219	>	Class<?> ak = ForkJoinTask[].class;
3220	>	ABASE = U.arrayBaseOffset(ak);
3221	>	s = U.arrayIndexScale(ak);
3222	>	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
3223		} catch (Exception e) {
3224		throw new Error(e);
3225		}
3226		if ((s & (s-1)) != 0)
3227		throw new Error("data type scale not a power of two");
3228	<	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
3228	>
3229	>	/*
3230	>	* For extra caution, computations to set up pool state are
3231	>	* here; the constructor just assigns these values to fields.
3232	>	*/
3233	>	ForkJoinWorkerThreadFactory defaultFac =
3234	>	defaultForkJoinWorkerThreadFactory =
3235	>	new DefaultForkJoinWorkerThreadFactory();
3236	>	if (fac == null)
3237	>	fac = defaultFac;
3238	>	if (par <= 0)
3239	>	par = Runtime.getRuntime().availableProcessors();
3240	>	if (par > MAX_CAP)
3241	>	par = MAX_CAP;
3242	>	long np = (long)(-par); // precompute initial ctl value
3243	>	long ct = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK);
3244	>
3245	>	commonPoolParallelism = par;
3246	>	commonPool = new ForkJoinPool(par, ct, fac, handler);
3247	>	modifyThreadPermission = new RuntimePermission("modifyThread");
3248	>	submitters = new ThreadLocal<Submitter>();
3249		}
3250
3251		/**
#	Line 2172 | Line 3275 | public class ForkJoinPool extends Abstra
3275		}
3276		}
3277		}
3278	+
3279		}

Diff Legend

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