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 1.65 by dl, Wed Aug 18 14:05:27 2010 UTC vs.
Revision 1.118 by jsr166, Sat Jan 28 04:34:54 2012 UTC

#	Line 1 | Line 1
1		/*
2		* Written by Doug Lea with assistance from members of JCP JSR-166
3		* Expert Group and released to the public domain, as explained at
4	<	* http://creativecommons.org/licenses/publicdomain
4	>	* http://creativecommons.org/publicdomain/zero/1.0/
5		*/
6
7		package jsr166y;
8
9	–	import java.util.concurrent.*;
10	–
9		import java.util.ArrayList;
10		import java.util.Arrays;
11		import java.util.Collection;
12		import java.util.Collections;
13		import java.util.List;
14	<	import java.util.concurrent.locks.LockSupport;
15	<	import java.util.concurrent.locks.ReentrantLock;
14	>	import java.util.Random;
15	>	import java.util.concurrent.AbstractExecutorService;
16	>	import java.util.concurrent.Callable;
17	>	import java.util.concurrent.ExecutorService;
18	>	import java.util.concurrent.Future;
19	>	import java.util.concurrent.RejectedExecutionException;
20	>	import java.util.concurrent.RunnableFuture;
21	>	import java.util.concurrent.TimeUnit;
22		import java.util.concurrent.atomic.AtomicInteger;
23	<	import java.util.concurrent.CountDownLatch;
23	>	import java.util.concurrent.atomic.AtomicLong;
24	>	import java.util.concurrent.locks.ReentrantLock;
25	>	import java.util.concurrent.locks.Condition;
26
27		/**
28		* An {@link ExecutorService} for running {@link ForkJoinTask}s.
#	Line 27 | Line 33 | import java.util.concurrent.CountDownLat
33		* <p>A {@code ForkJoinPool} differs from other kinds of {@link
34		* ExecutorService} mainly by virtue of employing
35		* <em>work-stealing</em>: all threads in the pool attempt to find and
36	<	* execute subtasks created by other active tasks (eventually blocking
37	<	* waiting for work if none exist). This enables efficient processing
38	<	* when most tasks spawn other subtasks (as do most {@code
39	<	* ForkJoinTask}s). When setting <em>asyncMode</em> to true in
40	<	* constructors, {@code ForkJoinPool}s may also be appropriate for use
41	<	* with event-style tasks that are never joined.
36	>	* execute tasks submitted to the pool and/or created by other active
37	>	* tasks (eventually blocking waiting for work if none exist). This
38	>	* enables efficient processing when most tasks spawn other subtasks
39	>	* (as do most {@code ForkJoinTask}s), as well as when many small
40	>	* tasks are submitted to the pool from external clients.  Especially
41	>	* when setting <em>asyncMode</em> to true in constructors, {@code
42	>	* ForkJoinPool}s may also be appropriate for use with event-style
43	>	* tasks that are never joined.
44		*
45		* <p>A {@code ForkJoinPool} is constructed with a given target
46		* parallelism level; by default, equal to the number of available
#	Line 52 | Line 60 | import java.util.concurrent.CountDownLat
60		* convenient form for informal monitoring.
61		*
62		* <p> As is the case with other ExecutorServices, there are three
63	<	* main task execution methods summarized in the following
64	<	* table. These are designed to be used by clients not already engaged
65	<	* in fork/join computations in the current pool.  The main forms of
66	<	* these methods accept instances of {@code ForkJoinTask}, but
67	<	* overloaded forms also allow mixed execution of plain {@code
63	>	* main task execution methods summarized in the following table.
64	>	* These are designed to be used primarily by clients not already
65	>	* engaged in fork/join computations in the current pool.  The main
66	>	* forms of these methods accept instances of {@code ForkJoinTask},
67	>	* but overloaded forms also allow mixed execution of plain {@code
68		* Runnable}- or {@code Callable}- based activities as well.  However,
69	<	* tasks that are already executing in a pool should normally
70	<	* <em>NOT</em> use these pool execution methods, but instead use the
71	<	* within-computation forms listed in the table.
69	>	* tasks that are already executing in a pool should normally instead
70	>	* use the within-computation forms listed in the table unless using
71	>	* async event-style tasks that are not usually joined, in which case
72	>	* there is little difference among choice of methods.
73		*
74		* <table BORDER CELLPADDING=3 CELLSPACING=1>
75		*  <tr>
#	Line 69 | Line 78 | import java.util.concurrent.CountDownLat
78		*    <td ALIGN=CENTER> <b>Call from within fork/join computations</b></td>
79		*  </tr>
80		*  <tr>
81	<	*    <td> <b>Arange async execution</td>
81	>	*    <td> <b>Arrange async execution</td>
82		*    <td> {@link #execute(ForkJoinTask)}</td>
83		*    <td> {@link ForkJoinTask#fork}</td>
84		*  </tr>
#	Line 95 | Line 104 | import java.util.concurrent.CountDownLat
104		* daemon} mode, there is typically no need to explicitly {@link
105		* #shutdown} such a pool upon program exit.
106		*
107	<	* <pre>
107	>	*  <pre> {@code
108		* static final ForkJoinPool mainPool = new ForkJoinPool();
109		* ...
110		* public void sort(long[] array) {
111		*   mainPool.invoke(new SortTask(array, 0, array.length));
112	<	* }
104	<	* </pre>
112	>	* }}</pre>
113		*
114		* <p><b>Implementation notes</b>: This implementation restricts the
115		* maximum number of running threads to 32767. Attempts to create
#	Line 120 | Line 128 | public class ForkJoinPool extends Abstra
128		/*
129		* Implementation Overview
130		*
131	<	* This class provides the central bookkeeping and control for a
132	<	* set of worker threads: Submissions from non-FJ threads enter
133	<	* into a submission queue. Workers take these tasks and typically
134	<	* split them into subtasks that may be stolen by other workers.
135	<	* The main work-stealing mechanics implemented in class
136	<	* ForkJoinWorkerThread give first priority to processing tasks
137	<	* from their own queues (LIFO or FIFO, depending on mode), then
138	<	* to randomized FIFO steals of tasks in other worker queues, and
139	<	* lastly to new submissions. These mechanics do not consider
140	<	* affinities, loads, cache localities, etc, so rarely provide the
141	<	* best possible performance on a given machine, but portably
142	<	* provide good throughput by averaging over these factors.
143	<	* (Further, even if we did try to use such information, we do not
144	<	* usually have a basis for exploiting it. For example, some sets
145	<	* of tasks profit from cache affinities, but others are harmed by
146	<	* cache pollution effects.)
147	<	*
148	<	* Beyond work-stealing support and essential bookkeeping, the
149	<	* main responsibility of this framework is to take actions when
150	<	* one worker is waiting to join a task stolen (or always held by)
151	<	* another.  Becauae we are multiplexing many tasks on to a pool
152	<	* of workers, we can't just let them block (as in Thread.join).
153	<	* We also cannot just reassign the joiner's run-time stack with
154	<	* another and replace it later, which would be a form of
155	<	* "continuation", that even if possible is not necessarily a good
156	<	* idea. Given that the creation costs of most threads on most
157	<	* systems mainly surrounds setting up runtime stacks, thread
158	<	* creation and switching is usually not much more expensive than
159	<	* stack creation and switching, and is more flexible). Instead we
160	<	* combine two tactics:
131	>	* This class and its nested classes provide the main
132	>	* functionality and control for a set of worker threads:
133	>	* Submissions from non-FJ threads enter into submission queues.
134	>	* Workers take these tasks and typically split them into subtasks
135	>	* that may be stolen by other workers.  Preference rules give
136	>	* first priority to processing tasks from their own queues (LIFO
137	>	* or FIFO, depending on mode), then to randomized FIFO steals of
138	>	* tasks in other queues.
139	>	*
140	>	* WorkQueues
141	>	* ==========
142	>	*
143	>	* Most operations occur within work-stealing queues (in nested
144	>	* class WorkQueue).  These are special forms of Deques that
145	>	* support only three of the four possible end-operations -- push,
146	>	* pop, and poll (aka steal), under the further constraints that
147	>	* push and pop are called only from the owning thread (or, as
148	>	* extended here, under a lock), while poll may be called from
149	>	* other threads.  (If you are unfamiliar with them, you probably
150	>	* want to read Herlihy and Shavit's book "The Art of
151	>	* Multiprocessor programming", chapter 16 describing these in
152	>	* more detail before proceeding.)  The main work-stealing queue
153	>	* design is roughly similar to those in the papers "Dynamic
154	>	* Circular Work-Stealing Deque" by Chase and Lev, SPAA 2005
155	>	* (http://research.sun.com/scalable/pubs/index.html) and
156	>	* "Idempotent work stealing" by Michael, Saraswat, and Vechev,
157	>	* PPoPP 2009 (http://portal.acm.org/citation.cfm?id=1504186).
158	>	* The main differences ultimately stem from GC requirements that
159	>	* we null out taken slots as soon as we can, to maintain as small
160	>	* a footprint as possible even in programs generating huge
161	>	* numbers of tasks. To accomplish this, we shift the CAS
162	>	* arbitrating pop vs poll (steal) from being on the indices
163	>	* ("base" and "top") to the slots themselves.  So, both a
164	>	* successful pop and poll mainly entail a CAS of a slot from
165	>	* non-null to null.  Because we rely on CASes of references, we
166	>	* do not need tag bits on base or top.  They are simple ints as
167	>	* used in any circular array-based queue (see for example
168	>	* ArrayDeque).  Updates to the indices must still be ordered in a
169	>	* way that guarantees that top == base means the queue is empty,
170	>	* but otherwise may err on the side of possibly making the queue
171	>	* appear nonempty when a push, pop, or poll have not fully
172	>	* committed. Note that this means that the poll operation,
173	>	* considered individually, is not wait-free. One thief cannot
174	>	* successfully continue until another in-progress one (or, if
175	>	* previously empty, a push) completes.  However, in the
176	>	* aggregate, we ensure at least probabilistic non-blockingness.
177	>	* If an attempted steal fails, a thief always chooses a different
178	>	* random victim target to try next. So, in order for one thief to
179	>	* progress, it suffices for any in-progress poll or new push on
180	>	* any empty queue to complete.
181	>	*
182	>	* This approach also enables support of a user mode in which local
183	>	* task processing is in FIFO, not LIFO order, simply by using
184	>	* poll rather than pop.  This can be useful in message-passing
185	>	* frameworks in which tasks are never joined.  However neither
186	>	* mode considers affinities, loads, cache localities, etc, so
187	>	* rarely provide the best possible performance on a given
188	>	* machine, but portably provide good throughput by averaging over
189	>	* these factors.  (Further, even if we did try to use such
190	>	* information, we do not usually have a basis for exploiting it.
191	>	* For example, some sets of tasks profit from cache affinities,
192	>	* but others are harmed by cache pollution effects.)
193	>	*
194	>	* WorkQueues are also used in a similar way for tasks submitted
195	>	* to the pool. We cannot mix these tasks in the same queues used
196	>	* for work-stealing (this would contaminate lifo/fifo
197	>	* processing). Instead, we loosely associate submission queues
198	>	* with submitting threads, using a form of hashing.  The
199	>	* ThreadLocal Submitter class contains a value initially used as
200	>	* a hash code for choosing existing queues, but may be randomly
201	>	* repositioned upon contention with other submitters.  In
202	>	* essence, submitters act like workers except that they never
203	>	* take tasks, and they are multiplexed on to a finite number of
204	>	* shared work queues. However, classes are set up so that future
205	>	* extensions could allow submitters to optionally help perform
206	>	* tasks as well. Pool submissions from internal workers are also
207	>	* allowed, but use randomized rather than thread-hashed queue
208	>	* indices to avoid imbalance.  Insertion of tasks in shared mode
209	>	* requires a lock (mainly to protect in the case of resizing) but
210	>	* we use only a simple spinlock (using bits in field runState),
211	>	* because submitters encountering a busy queue try or create
212	>	* others so never block.
213	>	*
214	>	* Management
215	>	* ==========
216	>	*
217	>	* The main throughput advantages of work-stealing stem from
218	>	* decentralized control -- workers mostly take tasks from
219	>	* themselves or each other. We cannot negate this in the
220	>	* implementation of other management responsibilities. The main
221	>	* tactic for avoiding bottlenecks is packing nearly all
222	>	* essentially atomic control state into two volatile variables
223	>	* that are by far most often read (not written) as status and
224	>	* consistency checks.
225	>	*
226	>	* Field "ctl" contains 64 bits holding all the information needed
227	>	* to atomically decide to add, inactivate, enqueue (on an event
228	>	* queue), dequeue, and/or re-activate workers.  To enable this
229	>	* packing, we restrict maximum parallelism to (1<<15)-1 (which is
230	>	* far in excess of normal operating range) to allow ids, counts,
231	>	* and their negations (used for thresholding) to fit into 16bit
232	>	* fields.
233	>	*
234	>	* Field "runState" contains 32 bits needed to register and
235	>	* deregister WorkQueues, as well as to enable shutdown. It is
236	>	* only modified under a lock (normally briefly held, but
237	>	* occasionally protecting allocations and resizings) but even
238	>	* when locked remains available to check consistency.
239	>	*
240	>	* Recording WorkQueues.  WorkQueues are recorded in the
241	>	* "workQueues" array that is created upon pool construction and
242	>	* expanded if necessary.  Updates to the array while recording
243	>	* new workers and unrecording terminated ones are protected from
244	>	* each other by a lock but the array is otherwise concurrently
245	>	* readable, and accessed directly.  To simplify index-based
246	>	* operations, the array size is always a power of two, and all
247	>	* readers must tolerate null slots. Shared (submission) queues
248	>	* are at even indices, worker queues at odd indices. Grouping
249	>	* them together in this way simplifies and speeds up task
250	>	* scanning. To avoid flailing during start-up, the array is
251	>	* presized to hold twice #parallelism workers (which is unlikely
252	>	* to need further resizing during execution). But to avoid
253	>	* dealing with so many null slots, variable runState includes a
254	>	* mask for the nearest power of two that contains all current
255	>	* workers.  All worker thread creation is on-demand, triggered by
256	>	* task submissions, replacement of terminated workers, and/or
257	>	* compensation for blocked workers. However, all other support
258	>	* code is set up to work with other policies.  To ensure that we
259	>	* do not hold on to worker references that would prevent GC, ALL
260	>	* accesses to workQueues are via indices into the workQueues
261	>	* array (which is one source of some of the messy code
262	>	* constructions here). In essence, the workQueues array serves as
263	>	* a weak reference mechanism. Thus for example the wait queue
264	>	* field of ctl stores indices, not references.  Access to the
265	>	* workQueues in associated methods (for example signalWork) must
266	>	* both index-check and null-check the IDs. All such accesses
267	>	* ignore bad IDs by returning out early from what they are doing,
268	>	* since this can only be associated with termination, in which
269	>	* case it is OK to give up.
270	>	*
271	>	* All uses of the workQueues array check that it is non-null
272	>	* (even if previously non-null). This allows nulling during
273	>	* termination, which is currently not necessary, but remains an
274	>	* option for resource-revocation-based shutdown schemes. It also
275	>	* helps reduce JIT issuance of uncommon-trap code, which tends to
276	>	* unnecessarily complicate control flow in some methods.
277	>	*
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. 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 (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 fewer than two 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	>	* together these cover the signals needed in cases when more
321	>	* tasks are pushed but untaken, and improve performance compared
322	>	* to having one thread wake up all workers.
323	>	*
324	>	* Trimming workers. To release resources after periods of lack of
325	>	* use, a worker starting to wait when the pool is quiescent will
326	>	* time out and terminate if the pool has remained quiescent for
327	>	* SHRINK_RATE nanosecs. This will slowly propagate, eventually
328	>	* terminating all workers after long periods of non-use.
329	>	*
330	>	* Shutdown and Termination. A call to shutdownNow atomically sets
331	>	* a runState bit and then (non-atomically) sets each worker's
332	>	* runState status, cancels all unprocessed tasks, and wakes up
333	>	* all waiting workers.  Detecting whether termination should
334	>	* commence after a non-abrupt shutdown() call requires more work
335	>	* and bookkeeping. We need consensus about quiescence (i.e., that
336	>	* there is no more work). The active count provides a primary
337	>	* indication but non-abrupt shutdown still requires a rechecking
338	>	* scan for any workers that are inactive but not queued.
339	>	*
340	>	* Joining Tasks
341	>	* =============
342	>	*
343	>	* Any of several actions may be taken when one worker is waiting
344	>	* to join a task stolen (or always held) by another.  Because we
345	>	* are multiplexing many tasks on to a pool of workers, we can't
346	>	* just let them block (as in Thread.join).  We also cannot just
347	>	* reassign the joiner's run-time stack with another and replace
348	>	* it later, which would be a form of "continuation", that even if
349	>	* possible is not necessarily a good idea since we sometimes need
350	>	* both an unblocked task and its continuation to progress.
351	>	* Instead we combine two tactics:
352		*
353		*   Helping: Arranging for the joiner to execute some task that it
354	<	*      would be running if the steal had not occurred.  Method
156	<	*      ForkJoinWorkerThread.helpJoinTask tracks joining->stealing
157	<	*      links to try to find such a task.
354	>	*      would be running if the steal had not occurred.
355		*
356		*   Compensating: Unless there are already enough live threads,
357	<	*      method helpMaintainParallelism() may create or or
358	<	*      re-activate a spare thread to compensate for blocked
359	<	*      joiners until they unblock.
360	<	*
361	<	* Because the determining existence of conservatively safe
362	<	* helping targets, the availability of already-created spares,
363	<	* and the apparent need to create new spares are all racy and
364	<	* require heuristic guidance, we rely on multiple retries of
365	<	* each. Further, because it is impossible to keep exactly the
366	<	* target (parallelism) number of threads running at any given
170	<	* time, we allow compensation during joins to fail, and enlist
171	<	* all other threads to help out whenever they are not otherwise
172	<	* occupied (i.e., mainly in method preStep).
357	>	*      method tryCompensate() may create or re-activate a spare
358	>	*      thread to compensate for blocked joiners until they unblock.
359	>	*
360	>	* A third form (implemented in tryRemoveAndExec and
361	>	* tryPollForAndExec) amounts to helping a hypothetical
362	>	* compensator: If we can readily tell that a possible action of a
363	>	* compensator is to steal and execute the task being joined, the
364	>	* joining thread can do so directly, without the need for a
365	>	* compensation thread (although at the expense of larger run-time
366	>	* stacks, but the tradeoff is typically worthwhile).
367		*
368		* The ManagedBlocker extension API can't use helping so relies
369		* only on compensation in method awaitBlocker.
370		*
371	<	* The main throughput advantages of work-stealing stem from
372	<	* decentralized control -- workers mostly steal tasks from each
373	<	* other. We do not want to negate this by creating bottlenecks
374	<	* implementing other management responsibilities. So we use a
375	<	* collection of techniques that avoid, reduce, or cope well with
376	<	* contention. These entail several instances of bit-packing into
377	<	* CASable fields to maintain only the minimally required
378	<	* atomicity. To enable such packing, we restrict maximum
379	<	* parallelism to (1<<15)-1 (enabling twice this (to accommodate
380	<	* unbalanced increments and decrements) to fit into a 16 bit
381	<	* field, which is far in excess of normal operating range.  Even
382	<	* though updates to some of these bookkeeping fields do sometimes
383	<	* contend with each other, they don't normally cache-contend with
384	<	* updates to others enough to warrant memory padding or
385	<	* isolation. So they are all held as fields of ForkJoinPool
386	<	* objects.  The main capabilities are as follows:
387	<	*
388	<	* 1. Creating and removing workers. Workers are recorded in the
389	<	* "workers" array. This is an array as opposed to some other data
390	<	* structure to support index-based random steals by workers.
391	<	* Updates to the array recording new workers and unrecording
392	<	* terminated ones are protected from each other by a lock
393	<	* (workerLock) but the array is otherwise concurrently readable,
394	<	* and accessed directly by workers. To simplify index-based
395	<	* operations, the array size is always a power of two, and all
396	<	* readers must tolerate null slots. Currently, all worker thread
397	<	* creation is on-demand, triggered by task submissions,
398	<	* replacement of terminated workers, and/or compensation for
399	<	* blocked workers. However, all other support code is set up to
400	<	* work with other policies.
401	<	*
402	<	* To ensure that we do not hold on to worker references that
403	<	* would prevent GC, ALL accesses to workers are via indices into
404	<	* the workers array (which is one source of some of the unusual
405	<	* code constructions here). In essence, the workers array serves
406	<	* as a WeakReference mechanism. Thus for example the event queue
407	<	* stores worker indices, not worker references. Access to the
408	<	* workers in associated methods (for example releaseEventWaiters)
409	<	* must both index-check and null-check the IDs. All such accesses
410	<	* ignore bad IDs by returning out early from what they are doing,
411	<	* since this can only be associated with shutdown, in which case
412	<	* it is OK to give up. On termination, we just clobber these
413	<	* data structures without trying to use them.
220	<	*
221	<	* 2. Bookkeeping for dynamically adding and removing workers. We
222	<	* aim to approximately maintain the given level of parallelism.
223	<	* When some workers are known to be blocked (on joins or via
224	<	* ManagedBlocker), we may create or resume others to take their
225	<	* place until they unblock (see below). Implementing this
226	<	* requires counts of the number of "running" threads (i.e., those
227	<	* that are neither blocked nor artifically suspended) as well as
228	<	* the total number.  These two values are packed into one field,
229	<	* "workerCounts" because we need accurate snapshots when deciding
230	<	* to create, resume or suspend.  Note however that the
231	<	* correspondance of these counts to reality is not guaranteed. In
232	<	* particular updates for unblocked threads may lag until they
233	<	* actually wake up.
234	<	*
235	<	* 3. Maintaining global run state. The run state of the pool
236	<	* consists of a runLevel (SHUTDOWN, TERMINATING, etc) similar to
237	<	* those in other Executor implementations, as well as a count of
238	<	* "active" workers -- those that are, or soon will be, or
239	<	* recently were executing tasks. The runLevel and active count
240	<	* are packed together in order to correctly trigger shutdown and
241	<	* termination. Without care, active counts can be subject to very
242	<	* high contention.  We substantially reduce this contention by
243	<	* relaxing update rules.  A worker must claim active status
244	<	* prospectively, by activating if it sees that a submitted or
245	<	* stealable task exists (it may find after activating that the
246	<	* task no longer exists). It stays active while processing this
247	<	* task (if it exists) and any other local subtasks it produces,
248	<	* until it cannot find any other tasks. It then tries
249	<	* inactivating (see method preStep), but upon update contention
250	<	* instead scans for more tasks, later retrying inactivation if it
251	<	* doesn't find any.
252	<	*
253	<	* 4. Managing idle workers waiting for tasks. We cannot let
254	<	* workers spin indefinitely scanning for tasks when none are
255	<	* available. On the other hand, we must quickly prod them into
256	<	* action when new tasks are submitted or generated.  We
257	<	* park/unpark these idle workers using an event-count scheme.
258	<	* Field eventCount is incremented upon events that may enable
259	<	* workers that previously could not find a task to now find one:
260	<	* Submission of a new task to the pool, or another worker pushing
261	<	* a task onto a previously empty queue.  (We also use this
262	<	* mechanism for configuration and termination actions that
263	<	* require wakeups of idle workers).  Each worker maintains its
264	<	* last known event count, and blocks when a scan for work did not
265	<	* find a task AND its lastEventCount matches the current
266	<	* eventCount. Waiting idle workers are recorded in a variant of
267	<	* Treiber stack headed by field eventWaiters which, when nonzero,
268	<	* encodes the thread index and count awaited for by the worker
269	<	* thread most recently calling eventSync. This thread in turn has
270	<	* a record (field nextEventWaiter) for the next waiting worker.
271	<	* In addition to allowing simpler decisions about need for
272	<	* wakeup, the event count bits in eventWaiters serve the role of
273	<	* tags to avoid ABA errors in Treiber stacks. Upon any wakeup,
274	<	* released threads also try to release others (but give up upon
275	<	* contention to reduce useless flailing).  The net effect is a
276	<	* tree-like diffusion of signals, where released threads (and
277	<	* possibly others) help with unparks.  To further reduce
278	<	* contention effects a bit, failed CASes to increment field
279	<	* eventCount are tolerated without retries in signalWork.
280	<	* Conceptually they are merged into the same event, which is OK
281	<	* when their only purpose is to enable workers to scan for work.
282	<	*
283	<	* 5. Managing suspension of extra workers. When a worker is about
284	<	* to block waiting for a join (or via ManagedBlockers), we may
285	<	* create a new thread to maintain parallelism level, or at least
286	<	* avoid starvation. Usually, extra threads are needed for only
287	<	* very short periods, yet join dependencies are such that we
288	<	* sometimes need them in bursts. Rather than create new threads
289	<	* each time this happens, we suspend no-longer-needed extra ones
290	<	* as "spares". For most purposes, we don't distinguish "extra"
291	<	* spare threads from normal "core" threads: On each call to
292	<	* preStep (the only point at which we can do this) a worker
293	<	* checks to see if there are now too many running workers, and if
294	<	* so, suspends itself.  Method helpMaintainParallelism looks for
295	<	* suspended threads to resume before considering creating a new
296	<	* replacement. The spares themselves are encoded on another
297	<	* variant of a Treiber Stack, headed at field "spareWaiters".
298	<	* Note that the use of spares is intrinsically racy.  One thread
299	<	* may become a spare at about the same time as another is
300	<	* needlessly being created. We counteract this and related slop
301	<	* in part by requiring resumed spares to immediately recheck (in
302	<	* preStep) to see whether they they should re-suspend.
303	<	*
304	<	* 6. Killing off unneeded workers. The Spare and Event queues use
305	<	* similar mechanisms to shed unused workers: The oldest (first)
306	<	* waiter uses a timed rather than hard wait. When this wait times
307	<	* out without a normal wakeup, it tries to shutdown any one (for
308	<	* convenience the newest) other waiter via tryShutdownSpare or
309	<	* tryShutdownWaiter, respectively. The wakeup rates for spares
310	<	* are much shorter than for waiters. Together, they will
311	<	* eventually reduce the number of worker threads to a minimum of
312	<	* one after a long enough period without use.
313	<	*
314	<	* 7. Deciding when to create new workers. The main dynamic
315	<	* control in this class is deciding when to create extra threads
316	<	* in method helpMaintainParallelism. We would like to keep
317	<	* exactly #parallelism threads running, which is an impossble
318	<	* task. We always need to create one when the number of running
319	<	* threads would become zero and all workers are busy. Beyond
320	<	* this, we must rely on heuristics that work well in the the
321	<	* presence of transients phenomena such as GC stalls, dynamic
322	<	* compilation, and wake-up lags. These transients are extremely
323	<	* common -- we are normally trying to fully saturate the CPUs on
324	<	* a machine, so almost any activity other than running tasks
325	<	* impedes accuracy. Our main defense is to allow some slack in
326	<	* creation thresholds, using rules that reflect the fact that the
327	<	* more threads we have running, the more likely that we are
328	<	* underestimating the number running threads. (We also include
329	<	* some heuristic use of Thread.yield when all workers appear to
330	<	* be busy, to improve likelihood of counts settling.) The rules
331	<	* also better cope with the fact that some of the methods in this
332	<	* class tend to never become compiled (but are interpreted), so
333	<	* some components of the entire set of controls might execute 100
334	<	* times faster than others. And similarly for cases where the
335	<	* apparent lack of work is just due to GC stalls and other
336	<	* transient system activity.
371	>	* The algorithm in tryHelpStealer entails a form of "linear"
372	>	* helping: Each worker records (in field currentSteal) the most
373	>	* recent task it stole from some other worker. Plus, it records
374	>	* (in field currentJoin) the task it is currently actively
375	>	* joining. Method tryHelpStealer uses these markers to try to
376	>	* find a worker to help (i.e., steal back a task from and execute
377	>	* it) that could hasten completion of the actively joined task.
378	>	* In essence, the joiner executes a task that would be on its own
379	>	* local deque had the to-be-joined task not been stolen. This may
380	>	* be seen as a conservative variant of the approach in Wagner &
381	>	* Calder "Leapfrogging: a portable technique for implementing
382	>	* efficient futures" SIGPLAN Notices, 1993
383	>	* (http://portal.acm.org/citation.cfm?id=155354). It differs in
384	>	* that: (1) We only maintain dependency links across workers upon
385	>	* steals, rather than use per-task bookkeeping.  This sometimes
386	>	* requires a linear scan of workers array to locate stealers, but
387	>	* often doesn't because stealers leave hints (that may become
388	>	* stale/wrong) of where to locate them.  A stealHint is only a
389	>	* hint because a worker might have had multiple steals and the
390	>	* hint records only one of them (usually the most current).
391	>	* Hinting isolates cost to when it is needed, rather than adding
392	>	* to per-task overhead.  (2) It is "shallow", ignoring nesting
393	>	* and potentially cyclic mutual steals.  (3) It is intentionally
394	>	* racy: field currentJoin is updated only while actively joining,
395	>	* which means that we miss links in the chain during long-lived
396	>	* tasks, GC stalls etc (which is OK since blocking in such cases
397	>	* is usually a good idea).  (4) We bound the number of attempts
398	>	* to find work (see MAX_HELP_DEPTH) and fall back to suspending
399	>	* the worker and if necessary replacing it with another.
400	>	*
401	>	* It is impossible to keep exactly the target parallelism number
402	>	* of threads running at any given time.  Determining the
403	>	* existence of conservatively safe helping targets, the
404	>	* availability of already-created spares, and the apparent need
405	>	* to create new spares are all racy, so we rely on multiple
406	>	* retries of each.  Currently, in keeping with on-demand
407	>	* signalling policy, we compensate only if blocking would leave
408	>	* less than one active (non-waiting, non-blocked) worker.
409	>	* Additionally, to avoid some false alarms due to GC, lagging
410	>	* counters, system activity, etc, compensated blocking for joins
411	>	* is only attempted after rechecks stabilize in
412	>	* ForkJoinTask.awaitJoin. (Retries are interspersed with
413	>	* Thread.yield, for good citizenship.)
414		*
415	<	* Beware that there is a lot of representation-level coupling
415	>	* Style notes: There is a lot of representation-level coupling
416		* among classes ForkJoinPool, ForkJoinWorkerThread, and
417	<	* ForkJoinTask.  For example, direct access to "workers" array by
418	<	* workers, and direct access to ForkJoinTask.status by both
419	<	* ForkJoinPool and ForkJoinWorkerThread.  There is little point
420	<	* trying to reduce this, since any associated future changes in
421	<	* representations will need to be accompanied by algorithmic
422	<	* changes anyway.
423	<	*
424	<	* Style notes: There are lots of inline assignments (of form
425	<	* "while ((local = field) != 0)") which are usually the simplest
426	<	* way to ensure the required read orderings (which are sometimes
427	<	* critical). Also several occurrences of the unusual "do {}
428	<	* while(!cas...)" which is the simplest way to force an update of
429	<	* a CAS'ed variable. There are also other coding oddities that
430	<	* help some methods perform reasonably even when interpreted (not
431	<	* compiled), at the expense of some messy constructions that
432	<	* reduce byte code counts.
433	<	*
434	<	* The order of declarations in this file is: (1) statics (2)
435	<	* fields (along with constants used when unpacking some of them)
436	<	* (3) internal control methods (4) callbacks and other support
437	<	* for ForkJoinTask and ForkJoinWorkerThread classes, (5) exported
438	<	* methods (plus a few little helpers).
417	>	* ForkJoinTask.  The fields of WorkQueue maintain data structures
418	>	* managed by ForkJoinPool, so are directly accessed.  There is
419	>	* little point trying to reduce this, since any associated future
420	>	* changes in representations will need to be accompanied by
421	>	* algorithmic changes anyway. All together, these low-level
422	>	* implementation choices produce as much as a factor of 4
423	>	* performance improvement compared to naive implementations, and
424	>	* enable the processing of billions of tasks per second, at the
425	>	* expense of some ugliness.
426	>	*
427	>	* Methods signalWork() and scan() are the main bottlenecks, so are
428	>	* especially heavily micro-optimized/mangled.  There are lots of
429	>	* inline assignments (of form "while ((local = field) != 0)")
430	>	* which are usually the simplest way to ensure the required read
431	>	* orderings (which are sometimes critical). This leads to a
432	>	* "C"-like style of listing declarations of these locals at the
433	>	* heads of methods or blocks.  There are several occurrences of
434	>	* the unusual "do {} while (!cas...)"  which is the simplest way
435	>	* to force an update of a CAS'ed variable. There are also other
436	>	* coding oddities that help some methods perform reasonably even
437	>	* when interpreted (not compiled).
438	>	*
439	>	* The order of declarations in this file is:
440	>	* (1) statics
441	>	* (2) fields (along with constants used when unpacking some of
442	>	*     them), listed in an order that tends to reduce contention
443	>	*     among them a bit under most JVMs;
444	>	* (3) nested classes
445	>	* (4) internal control methods
446	>	* (5) callbacks and other support for ForkJoinTask methods
447	>	* (6) exported methods (plus a few little helpers)
448	>	* (7) static block initializing all statics in a minimally
449	>	*     dependent order.
450		*/
451
452		/**
#	Line 393 | Line 481 | public class ForkJoinPool extends Abstra
481		* overridden in ForkJoinPool constructors.
482		*/
483		public static final ForkJoinWorkerThreadFactory
484	<	defaultForkJoinWorkerThreadFactory =
397	<	new DefaultForkJoinWorkerThreadFactory();
484	>	defaultForkJoinWorkerThreadFactory;
485
486		/**
487		* Permission required for callers of methods that may start or
488		* kill threads.
489		*/
490	<	private static final RuntimePermission modifyThreadPermission =
404	<	new RuntimePermission("modifyThread");
490	>	private static final RuntimePermission modifyThreadPermission;
491
492		/**
493		* If there is a security manager, makes sure caller has
#	Line 416 | Line 502 | public class ForkJoinPool extends Abstra
502		/**
503		* Generator for assigning sequence numbers as pool names.
504		*/
505	<	private static final AtomicInteger poolNumberGenerator =
420	<	new AtomicInteger();
421	<
422	<	/**
423	<	* The wakeup interval (in nanoseconds) for the oldest worker
424	<	* worker waiting for an event invokes tryShutdownWaiter to shrink
425	<	* the number of workers.  The exact value does not matter too
426	<	* much, but should be long enough to slowly release resources
427	<	* during long periods without use without disrupting normal use.
428	<	*/
429	<	private static final long SHRINK_RATE_NANOS =
430	<	60L * 1000L * 1000L * 1000L; // one minute
505	>	private static final AtomicInteger poolNumberGenerator;
506
507		/**
508	<	* Absolute bound for parallelism level. Twice this number plus
509	<	* one (i.e., 0xfff) must fit into a 16bit field to enable
510	<	* word-packing for some counts and indices.
508	>	* Bits and masks for control variables
509	>	*
510	>	* Field ctl is a long packed with:
511	>	* AC: Number of active running workers minus target parallelism (16 bits)
512	>	* TC: Number of total workers minus target parallelism (16 bits)
513	>	* ST: true if pool is terminating (1 bit)
514	>	* EC: the wait count of top waiting thread (15 bits)
515	>	* ID: ~(poolIndex >>> 1) of top of Treiber stack of waiters (16 bits)
516	>	*
517	>	* When convenient, we can extract the upper 32 bits of counts and
518	>	* the lower 32 bits of queue state, u = (int)(ctl >>> 32) and e =
519	>	* (int)ctl.  The ec field is never accessed alone, but always
520	>	* together with id and st. The offsets of counts by the target
521	>	* parallelism and the positionings of fields makes it possible to
522	>	* perform the most common checks via sign tests of fields: When
523	>	* ac is negative, there are not enough active workers, when tc is
524	>	* negative, there are not enough total workers, when id is
525	>	* negative, there is at least one waiting worker, and when e is
526	>	* negative, the pool is terminating.  To deal with these possibly
527	>	* negative fields, we use casts in and out of "short" and/or
528	>	* signed shifts to maintain signedness.
529	>	*
530	>	* When a thread is queued (inactivated), its eventCount field is
531	>	* negative, which is the only way to tell if a worker is
532	>	* prevented from executing tasks, even though it must continue to
533	>	* scan for them to avoid queuing races.
534	>	*
535	>	* Field runState is an int packed with:
536	>	* SHUTDOWN: true if shutdown is enabled (1 bit)
537	>	* SEQ:  a sequence number updated upon (de)registering workers (15 bits)
538	>	* MASK: mask (power of 2 - 1) covering all registered poolIndexes (16 bits)
539	>	*
540	>	* The combination of mask and sequence number enables simple
541	>	* consistency checks: Staleness of read-only operations on the
542	>	* workers and queues arrays can be checked by comparing runState
543	>	* before vs after the reads. The low 16 bits (i.e, anding with
544	>	* SMASK) hold the smallest power of two covering all worker
545	>	* indices, minus one.  The mask for queues (vs workers) is twice
546	>	* this value plus 1.
547	>	*/
548	>
549	>	// bit positions/shifts for fields
550	>	private static final int  AC_SHIFT   = 48;
551	>	private static final int  TC_SHIFT   = 32;
552	>	private static final int  ST_SHIFT   = 31;
553	>	private static final int  EC_SHIFT   = 16;
554	>
555	>	// bounds
556	>	private static final int  MAX_ID     = 0x7fff;  // max poolIndex
557	>	private static final int  SMASK      = 0xffff;  // mask short bits
558	>	private static final int  SHORT_SIGN = 1 << 15;
559	>	private static final int  INT_SIGN   = 1 << 31;
560	>
561	>	// masks
562	>	private static final long STOP_BIT   = 0x0001L << ST_SHIFT;
563	>	private static final long AC_MASK    = ((long)SMASK) << AC_SHIFT;
564	>	private static final long TC_MASK    = ((long)SMASK) << TC_SHIFT;
565	>
566	>	// units for incrementing and decrementing
567	>	private static final long TC_UNIT    = 1L << TC_SHIFT;
568	>	private static final long AC_UNIT    = 1L << AC_SHIFT;
569	>
570	>	// masks and units for dealing with u = (int)(ctl >>> 32)
571	>	private static final int  UAC_SHIFT  = AC_SHIFT - 32;
572	>	private static final int  UTC_SHIFT  = TC_SHIFT - 32;
573	>	private static final int  UAC_MASK   = SMASK << UAC_SHIFT;
574	>	private static final int  UTC_MASK   = SMASK << UTC_SHIFT;
575	>	private static final int  UAC_UNIT   = 1 << UAC_SHIFT;
576	>	private static final int  UTC_UNIT   = 1 << UTC_SHIFT;
577	>
578	>	// masks and units for dealing with e = (int)ctl
579	>	private static final int E_MASK      = 0x7fffffff; // no STOP_BIT
580	>	private static final int E_SEQ       = 1 << EC_SHIFT;
581	>
582	>	// runState bits
583	>	private static final int SHUTDOWN    = 1 << 31;
584	>	private static final int RS_SEQ      = 1 << 16;
585	>	private static final int RS_SEQ_MASK = 0x7fff0000;
586	>
587	>	// access mode for WorkQueue
588	>	static final int LIFO_QUEUE          =  0;
589	>	static final int FIFO_QUEUE          =  1;
590	>	static final int SHARED_QUEUE        = -1;
591	>
592	>	/**
593	>	* The wakeup interval (in nanoseconds) for a worker waiting for a
594	>	* task when the pool is quiescent to instead try to shrink the
595	>	* number of workers.  The exact value does not matter too
596	>	* much. It must be short enough to release resources during
597	>	* sustained periods of idleness, but not so short that threads
598	>	* are continually re-created.
599	>	*/
600	>	private static final long SHRINK_RATE =
601	>	4L * 1000L * 1000L * 1000L; // 4 seconds
602	>
603	>	/**
604	>	* The timeout value for attempted shrinkage, includes
605	>	* some slop to cope with system timer imprecision.
606	>	*/
607	>	private static final long SHRINK_TIMEOUT = SHRINK_RATE - (SHRINK_RATE / 10);
608	>
609	>	/**
610	>	* The maximum stolen->joining link depth allowed in tryHelpStealer.
611	>	* Depths for legitimate chains are unbounded, but we use a fixed
612	>	* constant to avoid (otherwise unchecked) cycles and to bound
613	>	* staleness of traversal parameters at the expense of sometimes
614	>	* blocking when we could be helping.
615		*/
616	<	private static final int MAX_WORKERS   = 0x7fff;
616	>	private static final int MAX_HELP_DEPTH = 16;
617
618	<	/**
619	<	* Array holding all worker threads in the pool.  Array size must
620	<	* be a power of two.  Updates and replacements are protected by
621	<	* workerLock, but the array is always kept in a consistent enough
622	<	* state to be randomly accessed without locking by workers
623	<	* performing work-stealing, as well as other traversal-based
624	<	* methods in this class. All readers must tolerate that some
625	<	* array slots may be null.
618	>	/*
619	>	* Field layout order in this class tends to matter more than one
620	>	* would like. Runtime layout order is only loosely related to
621	>	* declaration order and may differ across JVMs, but the following
622	>	* empirically works OK on current JVMs.
623	>	*/
624	>
625	>	volatile long ctl;                       // main pool control
626	>	final int parallelism;                   // parallelism level
627	>	final int localMode;                     // per-worker scheduling mode
628	>	int nextPoolIndex;                       // hint used in registerWorker
629	>	volatile int runState;                   // shutdown status, seq, and mask
630	>	WorkQueue[] workQueues;                  // main registry
631	>	final ReentrantLock lock;                // for registration
632	>	final Condition termination;             // for awaitTermination
633	>	final ForkJoinWorkerThreadFactory factory; // factory for new workers
634	>	final Thread.UncaughtExceptionHandler ueh; // per-worker UEH
635	>	final AtomicLong stealCount;             // collect counts when terminated
636	>	final AtomicInteger nextWorkerNumber;    // to create worker name string
637	>	final String workerNamePrefix;           // Prefix for assigning worker names
638	>
639	>	/**
640	>	* Queues supporting work-stealing as well as external task
641	>	* submission. See above for main rationale and algorithms.
642	>	* Implementation relies heavily on "Unsafe" intrinsics
643	>	* and selective use of "volatile":
644	>	*
645	>	* Field "base" is the index (mod array.length) of the least valid
646	>	* queue slot, which is always the next position to steal (poll)
647	>	* from if nonempty. Reads and writes require volatile orderings
648	>	* but not CAS, because updates are only performed after slot
649	>	* CASes.
650	>	*
651	>	* Field "top" is the index (mod array.length) of the next queue
652	>	* slot to push to or pop from. It is written only by owner thread
653	>	* for push, or under lock for trySharedPush, and accessed by
654	>	* other threads only after reading (volatile) base.  Both top and
655	>	* base are allowed to wrap around on overflow, but (top - base)
656	>	* (or more commonly -(base - top) to force volatile read of base
657	>	* before top) still estimates size.
658	>	*
659	>	* The array slots are read and written using the emulation of
660	>	* volatiles/atomics provided by Unsafe. Insertions must in
661	>	* general use putOrderedObject as a form of releasing store to
662	>	* ensure that all writes to the task object are ordered before
663	>	* its publication in the queue. (Although we can avoid one case
664	>	* of this when locked in trySharedPush.) All removals entail a
665	>	* CAS to null.  The array is always a power of two. To ensure
666	>	* safety of Unsafe array operations, all accesses perform
667	>	* explicit null checks and implicit bounds checks via
668	>	* power-of-two masking.
669	>	*
670	>	* In addition to basic queuing support, this class contains
671	>	* fields described elsewhere to control execution. It turns out
672	>	* to work better memory-layout-wise to include them in this
673	>	* class rather than a separate class.
674	>	*
675	>	* Performance on most platforms is very sensitive to placement of
676	>	* instances of both WorkQueues and their arrays -- we absolutely
677	>	* do not want multiple WorkQueue instances or multiple queue
678	>	* arrays sharing cache lines. (It would be best for queue objects
679	>	* and their arrays to share, but there is nothing available to
680	>	* help arrange that).  Unfortunately, because they are recorded
681	>	* in a common array, WorkQueue instances are often moved to be
682	>	* adjacent by garbage collectors. To reduce impact, we use field
683	>	* padding that works OK on common platforms; this effectively
684	>	* trades off slightly slower average field access for the sake of
685	>	* avoiding really bad worst-case access. (Until better JVM
686	>	* support is in place, this padding is dependent on transient
687	>	* properties of JVM field layout rules.)  We also take care in
688	>	* allocating and sizing and resizing the array. Non-shared queue
689	>	* arrays are initialized (via method growArray) by workers before
690	>	* use. Others are allocated on first use.
691		*/
692	<	volatile ForkJoinWorkerThread[] workers;
692	>	static final class WorkQueue {
693	>	/**
694	>	* Capacity of work-stealing queue array upon initialization.
695	>	* Must be a power of two; at least 4, but set larger to
696	>	* reduce cacheline sharing among queues.
697	>	*/
698	>	static final int INITIAL_QUEUE_CAPACITY = 1 << 8;
699
700	<	/**
701	<	* Queue for external submissions.
702	<	*/
703	<	private final LinkedTransferQueue<ForkJoinTask<?>> submissionQueue;
700	>	/**
701	>	* Maximum size for queue arrays. Must be a power of two less
702	>	* than or equal to 1 << (31 - width of array entry) to ensure
703	>	* lack of wraparound of index calculations, but defined to a
704	>	* value a bit less than this to help users trap runaway
705	>	* programs before saturating systems.
706	>	*/
707	>	static final int MAXIMUM_QUEUE_CAPACITY = 1 << 26; // 64M
708
709	<	/**
710	<	* Lock protecting updates to workers array.
711	<	*/
712	<	private final ReentrantLock workerLock;
709	>	volatile long totalSteals; // cumulative number of steals
710	>	int seed;                  // for random scanning; initialize nonzero
711	>	volatile int eventCount;   // encoded inactivation count; < 0 if inactive
712	>	int nextWait;              // encoded record of next event waiter
713	>	int rescans;               // remaining scans until block
714	>	int nsteals;               // top-level task executions since last idle
715	>	final int mode;            // lifo, fifo, or shared
716	>	int poolIndex;             // index of this queue in pool (or 0)
717	>	int stealHint;             // index of most recent known stealer
718	>	volatile int runState;     // 1: locked, -1: terminate; else 0
719	>	volatile int base;         // index of next slot for poll
720	>	int top;                   // index of next slot for push
721	>	ForkJoinTask<?>[] array;   // the elements (initially unallocated)
722	>	final ForkJoinWorkerThread owner; // owning thread or null if shared
723	>	volatile Thread parker;    // == owner during call to park; else null
724	>	ForkJoinTask<?> currentJoin;  // task being joined in awaitJoin
725	>	ForkJoinTask<?> currentSteal; // current non-local task being executed
726	>	// Heuristic padding to ameliorate unfortunate memory placements
727	>	Object p00, p01, p02, p03, p04, p05, p06, p07, p08, p09, p0a;
728	>
729	>	WorkQueue(ForkJoinWorkerThread owner, int mode) {
730	>	this.owner = owner;
731	>	this.mode = mode;
732	>	// Place indices in the center of array (that is not yet allocated)
733	>	base = top = INITIAL_QUEUE_CAPACITY >>> 1;
734	>	}
735
736	<	/**
737	<	* Latch released upon termination.
738	<	*/
739	<	private final Phaser termination;
736	>	/**
737	>	* Returns number of tasks in the queue.
738	>	*/
739	>	final int queueSize() {
740	>	int n = base - top; // non-owner callers must read base first
741	>	return (n >= 0) ? 0 : -n;
742	>	}
743
744	<	/**
745	<	* Creation factory for worker threads.
746	<	*/
747	<	private final ForkJoinWorkerThreadFactory factory;
744	>	/**
745	>	* Pushes a task. Call only by owner in unshared queues.
746	>	*
747	>	* @param task the task. Caller must ensure non-null.
748	>	* @param p if non-null, pool to signal if necessary
749	>	* @throw RejectedExecutionException if array cannot be resized
750	>	*/
751	>	final void push(ForkJoinTask<?> task, ForkJoinPool p) {
752	>	ForkJoinTask<?>[] a;
753	>	int s = top, m, n;
754	>	if ((a = array) != null) {    // ignore if queue removed
755	>	U.putOrderedObject
756	>	(a, (((m = a.length - 1) & s) << ASHIFT) + ABASE, task);
757	>	if ((n = (top = s + 1) - base) <= 2) {
758	>	if (p != null)
759	>	p.signalWork();
760	>	}
761	>	else if (n >= m)
762	>	growArray(true);
763	>	}
764	>	}
765
766	<	/**
767	<	* Sum of per-thread steal counts, updated only when threads are
768	<	* idle or terminating.
769	<	*/
770	<	private volatile long stealCount;
766	>	/**
767	>	* Pushes a task if lock is free and array is either big
768	>	* enough or can be resized to be big enough.
769	>	*
770	>	* @param task the task. Caller must ensure non-null.
771	>	* @return true if submitted
772	>	*/
773	>	final boolean trySharedPush(ForkJoinTask<?> task) {
774	>	boolean submitted = false;
775	>	if (runState == 0 && U.compareAndSwapInt(this, RUNSTATE, 0, 1)) {
776	>	ForkJoinTask<?>[] a = array;
777	>	int s = top, n = s - base;
778	>	try {
779	>	if ((a != null && n < a.length - 1) ||
780	>	(a = growArray(false)) != null) { // must presize
781	>	int j = (((a.length - 1) & s) << ASHIFT) + ABASE;
782	>	U.putObject(a, (long)j, task);    // don't need "ordered"
783	>	top = s + 1;
784	>	submitted = true;
785	>	}
786	>	} finally {
787	>	runState = 0;                         // unlock
788	>	}
789	>	}
790	>	return submitted;
791	>	}
792
793	<	/**
794	<	* Encoded record of top of treiber stack of threads waiting for
795	<	* events. The top 32 bits contain the count being waited for. The
796	<	* bottom 16 bits contains one plus the pool index of waiting
797	<	* worker thread. (Bits 16-31 are unused.)
798	<	*/
799	<	private volatile long eventWaiters;
793	>	/**
794	>	* Takes next task, if one exists, in FIFO order.
795	>	*/
796	>	final ForkJoinTask<?> poll() {
797	>	ForkJoinTask<?>[] a; int b, i;
798	>	while ((b = base) - top < 0 && (a = array) != null &&
799	>	(i = (a.length - 1) & b) >= 0) {
800	>	int j = (i << ASHIFT) + ABASE;
801	>	ForkJoinTask<?> t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
802	>	if (t != null && base == b &&
803	>	U.compareAndSwapObject(a, j, t, null)) {
804	>	base = b + 1;
805	>	return t;
806	>	}
807	>	}
808	>	return null;
809	>	}
810
811	<	private static final int  EVENT_COUNT_SHIFT = 32;
812	<	private static final long WAITER_ID_MASK    = (1L << 16) - 1L;
811	>	/**
812	>	* Takes next task, if one exists, in LIFO order.
813	>	* Call only by owner in unshared queues.
814	>	*/
815	>	final ForkJoinTask<?> pop() {
816	>	ForkJoinTask<?> t; int m;
817	>	ForkJoinTask<?>[] a = array;
818	>	if (a != null && (m = a.length - 1) >= 0) {
819	>	for (int s; (s = top - 1) - base >= 0;) {
820	>	int j = ((m & s) << ASHIFT) + ABASE;
821	>	if ((t = (ForkJoinTask<?>)U.getObjectVolatile(a, j)) == null)
822	>	break;
823	>	if (U.compareAndSwapObject(a, j, t, null)) {
824	>	top = s;
825	>	return t;
826	>	}
827	>	}
828	>	}
829	>	return null;
830	>	}
831
832	<	/**
833	<	* A counter for events that may wake up worker threads:
834	<	*   - Submission of a new task to the pool
835	<	*   - A worker pushing a task on an empty queue
836	<	*   - termination
837	<	*/
493	<	private volatile int eventCount;
832	>	/**
833	>	* Takes next task, if one exists, in order specified by mode.
834	>	*/
835	>	final ForkJoinTask<?> nextLocalTask() {
836	>	return mode == 0 ? pop() : poll();
837	>	}
838
839	<	/**
840	<	* Encoded record of top of treiber stack of spare threads waiting
841	<	* for resumption. The top 16 bits contain an arbitrary count to
842	<	* avoid ABA effects. The bottom 16bits contains one plus the pool
843	<	* index of waiting worker thread.
844	<	*/
845	<	private volatile int spareWaiters;
839	>	/**
840	>	* Returns next task, if one exists, in order specified by mode.
841	>	*/
842	>	final ForkJoinTask<?> peek() {
843	>	ForkJoinTask<?>[] a = array; int m;
844	>	if (a == null || (m = a.length - 1) < 0)
845	>	return null;
846	>	int i = mode == 0 ? top - 1 : base;
847	>	int j = ((i & m) << ASHIFT) + ABASE;
848	>	return (ForkJoinTask<?>)U.getObjectVolatile(a, j);
849	>	}
850
851	<	private static final int SPARE_COUNT_SHIFT = 16;
852	<	private static final int SPARE_ID_MASK     = (1 << 16) - 1;
851	>	/**
852	>	* Returns task at index b if b is current base of queue.
853	>	*/
854	>	final ForkJoinTask<?> pollAt(int b) {
855	>	ForkJoinTask<?>[] a; int i;
856	>	ForkJoinTask<?> task = null;
857	>	if ((a = array) != null && (i = ((a.length - 1) & b)) >= 0) {
858	>	int j = (i << ASHIFT) + ABASE;
859	>	ForkJoinTask<?> t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
860	>	if (t != null && base == b &&
861	>	U.compareAndSwapObject(a, j, t, null)) {
862	>	base = b + 1;
863	>	task = t;
864	>	}
865	>	}
866	>	return task;
867	>	}
868
869	<	/**
870	<	* Lifecycle control. The low word contains the number of workers
871	<	* that are (probably) executing tasks. This value is atomically
872	<	* incremented before a worker gets a task to run, and decremented
873	<	* when worker has no tasks and cannot find any.  Bits 16-18
874	<	* contain runLevel value. When all are zero, the pool is
875	<	* running. Level transitions are monotonic (running -> shutdown
876	<	* -> terminating -> terminated) so each transition adds a bit.
877	<	* These are bundled together to ensure consistent read for
878	<	* termination checks (i.e., that runLevel is at least SHUTDOWN
879	<	* and active threads is zero).
880	<	*
881	<	* Notes: Most direct CASes are dependent on these bitfield
519	<	* positions.  Also, this field is non-private to enable direct
520	<	* performance-sensitive CASes in ForkJoinWorkerThread.
521	<	*/
522	<	volatile int runState;
869	>	/**
870	>	* Pops the given task only if it is at the current top.
871	>	*/
872	>	final boolean tryUnpush(ForkJoinTask<?> t) {
873	>	ForkJoinTask<?>[] a; int s;
874	>	if ((a = array) != null && (s = top) != base &&
875	>	U.compareAndSwapObject
876	>	(a, (((a.length - 1) & --s) << ASHIFT) + ABASE, t, null)) {
877	>	top = s;
878	>	return true;
879	>	}
880	>	return false;
881	>	}
882
883	<	// Note: The order among run level values matters.
884	<	private static final int RUNLEVEL_SHIFT     = 16;
885	<	private static final int SHUTDOWN           = 1 << RUNLEVEL_SHIFT;
886	<	private static final int TERMINATING        = 1 << (RUNLEVEL_SHIFT + 1);
887	<	private static final int TERMINATED         = 1 << (RUNLEVEL_SHIFT + 2);
888	<	private static final int ACTIVE_COUNT_MASK  = (1 << RUNLEVEL_SHIFT) - 1;
883	>	/**
884	>	* Polls the given task only if it is at the current base.
885	>	*/
886	>	final boolean pollFor(ForkJoinTask<?> task) {
887	>	ForkJoinTask<?>[] a; int b, i;
888	>	if ((b = base) - top < 0 && (a = array) != null &&
889	>	(i = (a.length - 1) & b) >= 0) {
890	>	int j = (i << ASHIFT) + ABASE;
891	>	if (U.getObjectVolatile(a, j) == task && base == b &&
892	>	U.compareAndSwapObject(a, j, task, null)) {
893	>	base = b + 1;
894	>	return true;
895	>	}
896	>	}
897	>	return false;
898	>	}
899
900	<	/**
901	<	* Holds number of total (i.e., created and not yet terminated)
902	<	* and running (i.e., not blocked on joins or other managed sync)
903	<	* threads, packed together to ensure consistent snapshot when
904	<	* making decisions about creating and suspending spare
905	<	* threads. Updated only by CAS. Note that adding a new worker
906	<	* requires incrementing both counts, since workers start off in
907	<	* running state.
908	<	*/
909	<	private volatile int workerCounts;
900	>	/**
901	>	* If present, removes from queue and executes the given task, or
902	>	* any other cancelled task. Returns (true) immediately on any CAS
903	>	* or consistency check failure so caller can retry.
904	>	*
905	>	* @return false if no progress can be made
906	>	*/
907	>	final boolean tryRemoveAndExec(ForkJoinTask<?> task) {
908	>	boolean removed = false, empty = true, progress = true;
909	>	ForkJoinTask<?>[] a; int m, s, b, n;
910	>	if ((a = array) != null && (m = a.length - 1) >= 0 &&
911	>	(n = (s = top) - (b = base)) > 0) {
912	>	for (ForkJoinTask<?> t;;) {           // traverse from s to b
913	>	int j = ((--s & m) << ASHIFT) + ABASE;
914	>	t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
915	>	if (t == null)                    // inconsistent length
916	>	break;
917	>	else if (t == task) {
918	>	if (s + 1 == top) {           // pop
919	>	if (!U.compareAndSwapObject(a, j, task, null))
920	>	break;
921	>	top = s;
922	>	removed = true;
923	>	}
924	>	else if (base == b)           // replace with proxy
925	>	removed = U.compareAndSwapObject(a, j, task,
926	>	new EmptyTask());
927	>	break;
928	>	}
929	>	else if (t.status >= 0)
930	>	empty = false;
931	>	else if (s + 1 == top) {          // pop and throw away
932	>	if (U.compareAndSwapObject(a, j, t, null))
933	>	top = s;
934	>	break;
935	>	}
936	>	if (--n == 0) {
937	>	if (!empty && base == b)
938	>	progress = false;
939	>	break;
940	>	}
941	>	}
942	>	}
943	>	if (removed)
944	>	task.doExec();
945	>	return progress;
946	>	}
947
948	<	private static final int TOTAL_COUNT_SHIFT  = 16;
949	<	private static final int RUNNING_COUNT_MASK = (1 << TOTAL_COUNT_SHIFT) - 1;
950	<	private static final int ONE_RUNNING        = 1;
951	<	private static final int ONE_TOTAL          = 1 << TOTAL_COUNT_SHIFT;
948	>	/**
949	>	* Initializes or doubles the capacity of array. Call either
950	>	* by owner or with lock held -- it is OK for base, but not
951	>	* top, to move while resizings are in progress.
952	>	*
953	>	* @param rejectOnFailure if true, throw exception if capacity
954	>	* exceeded (relayed ultimately to user); else return null.
955	>	*/
956	>	final ForkJoinTask<?>[] growArray(boolean rejectOnFailure) {
957	>	ForkJoinTask<?>[] oldA = array;
958	>	int size = oldA != null ? oldA.length << 1 : INITIAL_QUEUE_CAPACITY;
959	>	if (size <= MAXIMUM_QUEUE_CAPACITY) {
960	>	int oldMask, t, b;
961	>	ForkJoinTask<?>[] a = array = new ForkJoinTask<?>[size];
962	>	if (oldA != null && (oldMask = oldA.length - 1) >= 0 &&
963	>	(t = top) - (b = base) > 0) {
964	>	int mask = size - 1;
965	>	do {
966	>	ForkJoinTask<?> x;
967	>	int oldj = ((b & oldMask) << ASHIFT) + ABASE;
968	>	int j    = ((b &    mask) << ASHIFT) + ABASE;
969	>	x = (ForkJoinTask<?>)U.getObjectVolatile(oldA, oldj);
970	>	if (x != null &&
971	>	U.compareAndSwapObject(oldA, oldj, x, null))
972	>	U.putObjectVolatile(a, j, x);
973	>	} while (++b != t);
974	>	}
975	>	return a;
976	>	}
977	>	else if (!rejectOnFailure)
978	>	return null;
979	>	else
980	>	throw new RejectedExecutionException("Queue capacity exceeded");
981	>	}
982
983	<	/**
984	<	* The target parallelism level.
985	<	* Accessed directly by ForkJoinWorkerThreads.
986	<	*/
987	<	final int parallelism;
983	>	/**
984	>	* Removes and cancels all known tasks, ignoring any exceptions.
985	>	*/
986	>	final void cancelAll() {
987	>	ForkJoinTask.cancelIgnoringExceptions(currentJoin);
988	>	ForkJoinTask.cancelIgnoringExceptions(currentSteal);
989	>	for (ForkJoinTask<?> t; (t = poll()) != null; )
990	>	ForkJoinTask.cancelIgnoringExceptions(t);
991	>	}
992
993	<	/**
554	<	* True if use local fifo, not default lifo, for local polling
555	<	* Read by, and replicated by ForkJoinWorkerThreads
556	<	*/
557	<	final boolean locallyFifo;
993	>	// Execution methods
994
995	<	/**
996	<	* The uncaught exception handler used when any worker abruptly
997	<	* terminates.
998	<	*/
999	<	private final Thread.UncaughtExceptionHandler ueh;
995	>	/**
996	>	* Removes and runs tasks until empty, using local mode
997	>	* ordering.
998	>	*/
999	>	final void runLocalTasks() {
1000	>	if (base - top < 0) {
1001	>	for (ForkJoinTask<?> t; (t = nextLocalTask()) != null; )
1002	>	t.doExec();
1003	>	}
1004	>	}
1005
1006	<	/**
1007	<	* Pool number, just for assigning useful names to worker threads
1008	<	*/
1009	<	private final int poolNumber;
1006	>	/**
1007	>	* Executes a top-level task and any local tasks remaining
1008	>	* after execution.
1009	>	*
1010	>	* @return true unless terminating
1011	>	*/
1012	>	final boolean runTask(ForkJoinTask<?> t) {
1013	>	boolean alive = true;
1014	>	if (t != null) {
1015	>	currentSteal = t;
1016	>	t.doExec();
1017	>	runLocalTasks();
1018	>	++nsteals;
1019	>	currentSteal = null;
1020	>	}
1021	>	else if (runState < 0)            // terminating
1022	>	alive = false;
1023	>	return alive;
1024	>	}
1025
1026	+	/**
1027	+	* Executes a non-top-level (stolen) task.
1028	+	*/
1029	+	final void runSubtask(ForkJoinTask<?> t) {
1030	+	if (t != null) {
1031	+	ForkJoinTask<?> ps = currentSteal;
1032	+	currentSteal = t;
1033	+	t.doExec();
1034	+	currentSteal = ps;
1035	+	}
1036	+	}
1037
1038	<	// Utilities for CASing fields. Note that most of these
1039	<	// are usually manually inlined by callers
1038	>	/**
1039	>	* Computes next value for random probes.  Scans don't require
1040	>	* a very high quality generator, but also not a crummy one.
1041	>	* Marsaglia xor-shift is cheap and works well enough.  Note:
1042	>	* This is manually inlined in several usages in ForkJoinPool
1043	>	* to avoid writes inside busy scan loops.
1044	>	*/
1045	>	final int nextSeed() {
1046	>	int r = seed;
1047	>	r ^= r << 13;
1048	>	r ^= r >>> 17;
1049	>	r ^= r << 5;
1050	>	return seed = r;
1051	>	}
1052
1053	<	/**
1054	<	* Increments running count part of workerCounts
1055	<	*/
1056	<	final void incrementRunningCount() {
1057	<	int c;
1058	<	do {} while (!UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1059	<	c = workerCounts,
1060	<	c + ONE_RUNNING));
1053	>	// Unsafe mechanics
1054	>	private static final sun.misc.Unsafe U;
1055	>	private static final long RUNSTATE;
1056	>	private static final int ABASE;
1057	>	private static final int ASHIFT;
1058	>	static {
1059	>	int s;
1060	>	try {
1061	>	U = getUnsafe();
1062	>	Class<?> k = WorkQueue.class;
1063	>	Class<?> ak = ForkJoinTask[].class;
1064	>	RUNSTATE = U.objectFieldOffset
1065	>	(k.getDeclaredField("runState"));
1066	>	ABASE = U.arrayBaseOffset(ak);
1067	>	s = U.arrayIndexScale(ak);
1068	>	} catch (Exception e) {
1069	>	throw new Error(e);
1070	>	}
1071	>	if ((s & (s-1)) != 0)
1072	>	throw new Error("data type scale not a power of two");
1073	>	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
1074	>	}
1075		}
1076
1077		/**
1078	<	* Tries to decrement running count unless already zero
1078	>	* Class for artificial tasks that are used to replace the target
1079	>	* of local joins if they are removed from an interior queue slot
1080	>	* in WorkQueue.tryRemoveAndExec. We don't need the proxy to
1081	>	* actually do anything beyond having a unique identity.
1082		*/
1083	<	final boolean tryDecrementRunningCount() {
1084	<	int wc = workerCounts;
1085	<	if ((wc & RUNNING_COUNT_MASK) == 0)
1086	<	return false;
1087	<	return UNSAFE.compareAndSwapInt(this, workerCountsOffset,
592	<	wc, wc - ONE_RUNNING);
1083	>	static final class EmptyTask extends ForkJoinTask<Void> {
1084	>	EmptyTask() { status = ForkJoinTask.NORMAL; } // force done
1085	>	public Void getRawResult() { return null; }
1086	>	public void setRawResult(Void x) {}
1087	>	public boolean exec() { return true; }
1088		}
1089
1090		/**
1091	<	* Forces decrement of encoded workerCounts, awaiting nonzero if
1092	<	* (rarely) necessary when other count updates lag.
1093	<	*
1094	<	* @param dr -- either zero or ONE_RUNNING
1095	<	* @param dt == either zero or ONE_TOTAL
1091	>	* Per-thread records for (typically non-FJ) threads that submit
1092	>	* to pools. Cureently holds only psuedo-random seed / index that
1093	>	* is used to choose submission queues in method doSubmit. In the
1094	>	* future, this may incorporate a means to implement different
1095	>	* task rejection and resubmission policies.
1096		*/
1097	<	private void decrementWorkerCounts(int dr, int dt) {
1098	<	for (;;) {
1099	<	int wc = workerCounts;
1100	<	if ((wc & RUNNING_COUNT_MASK)  - dr < 0 ||
1101	<	(wc >>> TOTAL_COUNT_SHIFT) - dt < 0) {
1102	<	if ((runState & TERMINATED) != 0)
1103	<	return; // lagging termination on a backout
1104	<	Thread.yield();
1105	<	}
611	<	if (UNSAFE.compareAndSwapInt(this, workerCountsOffset,
612	<	wc, wc - (dr + dt)))
613	<	return;
1097	>	static final class Submitter {
1098	>	int seed; // seed for random submission queue selection
1099	>
1100	>	// Heuristic padding to ameliorate unfortunate memory placements
1101	>	int p0, p1, p2, p3, p4, p5, p6, p7, p8, p9, pa, pb, pc, pd, pe;
1102	>
1103	>	Submitter() {
1104	>	// Use identityHashCode, forced negative, for seed
1105	>	seed = System.identityHashCode(Thread.currentThread()) | (1 << 31);
1106		}
615	–	}
1107
1108	<	/**
1109	<	* Increments event count
1110	<	*/
1111	<	private void advanceEventCount() {
1112	<	int c;
1113	<	do {} while(!UNSAFE.compareAndSwapInt(this, eventCountOffset,
1114	<	c = eventCount, c+1));
1108	>	/**
1109	>	* Computes next value for random probes.  Like method
1110	>	* WorkQueue.nextSeed, this is manually inlined in several
1111	>	* usages to avoid writes inside busy loops.
1112	>	*/
1113	>	final int nextSeed() {
1114	>	int r = seed;
1115	>	r ^= r << 13;
1116	>	r ^= r >>> 17;
1117	>	return seed = r ^= r << 5;
1118	>	}
1119		}
1120
1121	<	/**
1122	<	* Tries incrementing active count; fails on contention.
1123	<	* Called by workers before executing tasks.
629	<	*
630	<	* @return true on success
631	<	*/
632	<	final boolean tryIncrementActiveCount() {
633	<	int c;
634	<	return UNSAFE.compareAndSwapInt(this, runStateOffset,
635	<	c = runState, c + 1);
1121	>	/** ThreadLocal class for Submitters */
1122	>	static final class ThreadSubmitter extends ThreadLocal<Submitter> {
1123	>	public Submitter initialValue() { return new Submitter(); }
1124		}
1125
1126		/**
1127	<	* Tries decrementing active count; fails on contention.
1128	<	* Called when workers cannot find tasks to run.
1127	>	* Per-thread submission bookeeping. Shared across all pools
1128	>	* to reduce ThreadLocal pollution and because random motion
1129	>	* to avoid contention in one pool is likely to hold for others.
1130		*/
1131	<	final boolean tryDecrementActiveCount() {
643	<	int c;
644	<	return UNSAFE.compareAndSwapInt(this, runStateOffset,
645	<	c = runState, c - 1);
646	<	}
1131	>	static final ThreadSubmitter submitters = new ThreadSubmitter();
1132
1133		/**
1134	<	* Advances to at least the given level. Returns true if not
650	<	* already in at least the given level.
1134	>	* Top-level runloop for workers
1135		*/
1136	<	private boolean advanceRunLevel(int level) {
1137	<	for (;;) {
1138	<	int s = runState;
1139	<	if ((s & level) != 0)
1140	<	return false;
1141	<	if (UNSAFE.compareAndSwapInt(this, runStateOffset, s, s | level))
1142	<	return true;
1143	<	}
1136	>	final void runWorker(ForkJoinWorkerThread wt) {
1137	>	// Initialize queue array and seed in this thread
1138	>	WorkQueue w = wt.workQueue;
1139	>	w.growArray(false);
1140	>	// Same initial hash as Submitters
1141	>	w.seed = System.identityHashCode(Thread.currentThread()) | (1 << 31);
1142	>
1143	>	do {} while (w.runTask(scan(w)));
1144		}
1145
1146	<	// workers array maintenance
1146	>	// Creating, registering and deregistering workers
1147
1148		/**
1149	<	* Records and returns a workers array index for new worker.
1149	>	* Tries to create and start a worker
1150		*/
1151	<	private int recordWorker(ForkJoinWorkerThread w) {
1152	<	// Try using slot totalCount-1. If not available, scan and/or resize
1153	<	int k = (workerCounts >>> TOTAL_COUNT_SHIFT) - 1;
670	<	final ReentrantLock lock = this.workerLock;
671	<	lock.lock();
1151	>	private void addWorker() {
1152	>	Throwable ex = null;
1153	>	ForkJoinWorkerThread w = null;
1154		try {
1155	<	ForkJoinWorkerThread[] ws = workers;
1156	<	int n = ws.length;
1157	<	if (k < 0 || k >= n || ws[k] != null) {
676	<	for (k = 0; k < n && ws[k] != null; ++k)
677	<	;
678	<	if (k == n)
679	<	ws = Arrays.copyOf(ws, n << 1);
1155	>	if ((w = factory.newThread(this)) != null) {
1156	>	w.start();
1157	>	return;
1158		}
1159	<	ws[k] = w;
1160	<	workers = ws; // volatile array write ensures slot visibility
683	<	} finally {
684	<	lock.unlock();
1159	>	} catch (Throwable e) {
1160	>	ex = e;
1161		}
1162	<	return k;
1162	>	deregisterWorker(w, ex);
1163		}
1164
1165		/**
1166	<	* Nulls out record of worker in workers array
1166	>	* Callback from ForkJoinWorkerThread constructor to assign a
1167	>	* public name. This must be separate from registerWorker because
1168	>	* it is called during the "super" constructor call in
1169	>	* ForkJoinWorkerThread.
1170		*/
1171	<	private void forgetWorker(ForkJoinWorkerThread w) {
1172	<	int idx = w.poolIndex;
1173	<	// Locking helps method recordWorker avoid unecessary expansion
1174	<	final ReentrantLock lock = this.workerLock;
1171	>	final String nextWorkerName() {
1172	>	return workerNamePrefix.concat
1173	>	(Integer.toString(nextWorkerNumber.addAndGet(1)));
1174	>	}
1175	>
1176	>	/**
1177	>	* Callback from ForkJoinWorkerThread constructor to establish and
1178	>	* record its WorkQueue.
1179	>	*
1180	>	* @param wt the worker thread
1181	>	*/
1182	>	final void registerWorker(ForkJoinWorkerThread wt) {
1183	>	WorkQueue w = wt.workQueue;
1184	>	ReentrantLock lock = this.lock;
1185		lock.lock();
1186		try {
1187	<	ForkJoinWorkerThread[] ws = workers;
1188	<	if (idx >= 0 && idx < ws.length && ws[idx] == w) // verify
1189	<	ws[idx] = null;
1187	>	int k = nextPoolIndex;
1188	>	WorkQueue[] ws = workQueues;
1189	>	if (ws != null) {                       // ignore on shutdown
1190	>	int n = ws.length;
1191	>	if (k < 0 || (k & 1) == 0 || k >= n || ws[k] != null) {
1192	>	for (k = 1; k < n && ws[k] != null; k += 2)
1193	>	;                           // workers are at odd indices
1194	>	if (k >= n)                     // resize
1195	>	workQueues = ws = Arrays.copyOf(ws, n << 1);
1196	>	}
1197	>	w.poolIndex = k;
1198	>	w.eventCount = ~(k >>> 1) & SMASK;  // Set up wait count
1199	>	ws[k] = w;                          // record worker
1200	>	nextPoolIndex = k + 2;
1201	>	int rs = runState;
1202	>	int m = rs & SMASK;                 // recalculate runState mask
1203	>	if (k > m)
1204	>	m = (m << 1) + 1;
1205	>	runState = (rs & SHUTDOWN) | ((rs + RS_SEQ) & RS_SEQ_MASK) | m;
1206	>	}
1207		} finally {
1208		lock.unlock();
1209		}
1210		}
1211
706	–	// adding and removing workers
707	–
1212		/**
1213	<	* Tries to create and add new worker. Assumes that worker counts
1214	<	* are already updated to accommodate the worker, so adjusts on
1215	<	* failure.
1213	>	* Final callback from terminating worker, as well as failure to
1214	>	* construct or start a worker in addWorker.  Removes record of
1215	>	* worker from array, and adjusts counts. If pool is shutting
1216	>	* down, tries to complete termination.
1217		*
1218	<	* @return the worker, or null on failure
1218	>	* @param wt the worker thread or null if addWorker failed
1219	>	* @param ex the exception causing failure, or null if none
1220		*/
1221	<	private ForkJoinWorkerThread addWorker() {
1222	<	ForkJoinWorkerThread w = null;
1223	<	try {
1224	<	w = factory.newThread(this);
1225	<	} finally { // Adjust on either null or exceptional factory return
1226	<	if (w == null) {
1227	<	decrementWorkerCounts(ONE_RUNNING, ONE_TOTAL);
1228	<	tryTerminate(false); // in case of failure during shutdown
1221	>	final void deregisterWorker(ForkJoinWorkerThread wt, Throwable ex) {
1222	>	WorkQueue w = null;
1223	>	if (wt != null && (w = wt.workQueue) != null) {
1224	>	w.runState = -1;                // ensure runState is set
1225	>	stealCount.getAndAdd(w.totalSteals + w.nsteals);
1226	>	int idx = w.poolIndex;
1227	>	ReentrantLock lock = this.lock;
1228	>	lock.lock();
1229	>	try {                           // remove record from array
1230	>	WorkQueue[] ws = workQueues;
1231	>	if (ws != null && idx >= 0 && idx < ws.length && ws[idx] == w)
1232	>	ws[nextPoolIndex = idx] = null;
1233	>	} finally {
1234	>	lock.unlock();
1235		}
1236		}
1237	<	if (w != null) {
1238	<	w.start(recordWorker(w), ueh);
1239	<	advanceEventCount();
1237	>
1238	>	long c;                             // adjust ctl counts
1239	>	do {} while (!U.compareAndSwapLong
1240	>	(this, CTL, c = ctl, (((c - AC_UNIT) & AC_MASK) |
1241	>	((c - TC_UNIT) & TC_MASK) |
1242	>	(c & ~(AC_MASK|TC_MASK)))));
1243	>
1244	>	if (!tryTerminate(false) && w != null) {
1245	>	w.cancelAll();                  // cancel remaining tasks
1246	>	if (w.array != null)            // suppress signal if never ran
1247	>	signalWork();               // wake up or create replacement
1248		}
1249	<	return w;
1249	>
1250	>	if (ex != null)                     // rethrow
1251	>	U.throwException(ex);
1252		}
1253
1254		/**
1255	<	* Final callback from terminating worker.  Removes record of
734	<	* worker from array, and adjusts counts. If pool is shutting
735	<	* down, tries to complete terminatation.
1255	>	* Tries to add and register a new queue at the given index.
1256		*
1257	<	* @param w the worker
1257	>	* @param idx the workQueues array index to register the queue
1258	>	* @return the queue, or null if could not add because could
1259	>	* not acquire lock or idx is unusable
1260		*/
1261	<	final void workerTerminated(ForkJoinWorkerThread w) {
1262	<	forgetWorker(w);
1263	<	decrementWorkerCounts(w.isTrimmed()? 0 : ONE_RUNNING, ONE_TOTAL);
1264	<	while (w.stealCount != 0) // collect final count
1265	<	tryAccumulateStealCount(w);
1266	<	tryTerminate(false);
1261	>	private WorkQueue tryAddSharedQueue(int idx) {
1262	>	WorkQueue q = null;
1263	>	ReentrantLock lock = this.lock;
1264	>	if (idx >= 0 && (idx & 1) == 0 && !lock.isLocked()) {
1265	>	// create queue outside of lock but only if apparently free
1266	>	WorkQueue nq = new WorkQueue(null, SHARED_QUEUE);
1267	>	if (lock.tryLock()) {
1268	>	try {
1269	>	WorkQueue[] ws = workQueues;
1270	>	if (ws != null && idx < ws.length) {
1271	>	if ((q = ws[idx]) == null) {
1272	>	int rs;         // update runState seq
1273	>	ws[idx] = q = nq;
1274	>	runState = (((rs = runState) & SHUTDOWN) |
1275	>	((rs + RS_SEQ) & ~SHUTDOWN));
1276	>	}
1277	>	}
1278	>	} finally {
1279	>	lock.unlock();
1280	>	}
1281	>	}
1282	>	}
1283	>	return q;
1284		}
1285
1286	<	// Waiting for and signalling events
1286	>	// Maintaining ctl counts
1287
1288		/**
1289	<	* Releases workers blocked on a count not equal to current count.
1290	<	* Normally called after precheck that eventWaiters isn't zero to
1291	<	* avoid wasted array checks. Gives up upon a change in count or
1292	<	* contention, letting other workers take over.
1293	<	*/
755	<	private void releaseEventWaiters() {
756	<	ForkJoinWorkerThread[] ws = workers;
757	<	int n = ws.length;
758	<	long h = eventWaiters;
759	<	int ec = eventCount;
760	<	ForkJoinWorkerThread w; int id;
761	<	while ((int)(h >>> EVENT_COUNT_SHIFT) != ec &&
762	<	(id = ((int)(h & WAITER_ID_MASK)) - 1) >= 0 &&
763	<	id < n && (w = ws[id]) != null &&
764	<	UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
765	<	h,  h = w.nextWaiter)) {
766	<	LockSupport.unpark(w);
767	<	if (eventWaiters != h || eventCount != ec)
768	<	break;
769	<	}
1289	>	* Increments active count; mainly called upon return from blocking.
1290	>	*/
1291	>	final void incrementActiveCount() {
1292	>	long c;
1293	>	do {} while (!U.compareAndSwapLong(this, CTL, c = ctl, c + AC_UNIT));
1294		}
1295
1296		/**
1297	<	* Tries to advance eventCount and releases waiters. Called only
774	<	* from workers.
1297	>	* Activates or creates a worker.
1298		*/
1299		final void signalWork() {
1300	<	int c; // try to increment event count -- CAS failure OK
1301	<	UNSAFE.compareAndSwapInt(this, eventCountOffset, c = eventCount, c+1);
1302	<	if (eventWaiters != 0L)
1303	<	releaseEventWaiters();
1304	<	}
1305	<
1306	<	/**
1307	<	* Adds the given worker to event queue and blocks until
1308	<	* terminating or event count advances from the workers
1309	<	* lastEventCount value
1310	<	*
1311	<	* @param w the calling worker thread
1312	<	*/
1313	<	private void eventSync(ForkJoinWorkerThread w) {
1314	<	int ec = w.lastEventCount;
1315	<	long nh = (((long)ec) << EVENT_COUNT_SHIFT) | ((long)(w.poolIndex+1));
1316	<	long h;
1317	<	while ((runState < SHUTDOWN || !tryTerminate(false)) &&
1318	<	(((int)((h = eventWaiters) & WAITER_ID_MASK)) == 0 ||
1319	<	(int)(h >>> EVENT_COUNT_SHIFT) == ec) &&
1320	<	eventCount == ec) {
1321	<	if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
799	<	w.nextWaiter = h, nh)) {
800	<	awaitEvent(w, ec);
801	<	break;
1300	>	/*
1301	>	* The while condition is true if: (there is are too few total
1302	>	* workers OR there is at least one waiter) AND (there are too
1303	>	* few active workers OR the pool is terminating).  The value
1304	>	* of e distinguishes the remaining cases: zero (no waiters)
1305	>	* for create, negative if terminating (in which case do
1306	>	* nothing), else release a waiter. The secondary checks for
1307	>	* release (non-null array etc) can fail if the pool begins
1308	>	* terminating after the test, and don't impose any added cost
1309	>	* because JVMs must perform null and bounds checks anyway.
1310	>	*/
1311	>	long c; int e, u;
1312	>	while ((((e = (int)(c = ctl)) | (u = (int)(c >>> 32))) &
1313	>	(INT_SIGN|SHORT_SIGN)) == (INT_SIGN|SHORT_SIGN)) {
1314	>	WorkQueue[] ws = workQueues; int i; WorkQueue w; Thread p;
1315	>	if (e == 0) {                    // add a new worker
1316	>	if (U.compareAndSwapLong
1317	>	(this, CTL, c, (long)(((u + UTC_UNIT) & UTC_MASK) |
1318	>	((u + UAC_UNIT) & UAC_MASK)) << 32)) {
1319	>	addWorker();
1320	>	break;
1321	>	}
1322		}
1323	<	}
1324	<	}
1325	<
1326	<	/**
1327	<	* Blocks the given worker (that has already been entered as an
1328	<	* event waiter) until terminating or event count advances from
1329	<	* the given value. The oldest (first) waiter uses a timed wait to
1330	<	* occasionally one-by-one shrink the number of workers (to a
1331	<	* minumum of one) if the pool has not been used for extended
1332	<	* periods.
813	<	*
814	<	* @param w the calling worker thread
815	<	* @param ec the count
816	<	*/
817	<	private void awaitEvent(ForkJoinWorkerThread w, int ec) {
818	<	while (eventCount == ec) {
819	<	if (tryAccumulateStealCount(w)) { // transfer while idle
820	<	boolean untimed = (w.nextWaiter != 0L ||
821	<	(workerCounts & RUNNING_COUNT_MASK) <= 1);
822	<	long startTime = untimed? 0 : System.nanoTime();
823	<	Thread.interrupted();         // clear/ignore interrupt
824	<	if (eventCount != ec || !w.isRunning() ||
825	<	runState >= TERMINATING)  // recheck after clear
1323	>	else if (e > 0 && ws != null &&
1324	>	(i = ((~e << 1) | 1) & SMASK) < ws.length &&
1325	>	(w = ws[i]) != null &&
1326	>	w.eventCount == (e | INT_SIGN)) {
1327	>	if (U.compareAndSwapLong
1328	>	(this, CTL, c, (((long)(w.nextWait & E_MASK)) |
1329	>	((long)(u + UAC_UNIT) << 32)))) {
1330	>	w.eventCount = (e + E_SEQ) & E_MASK;
1331	>	if ((p = w.parker) != null)
1332	>	U.unpark(p);         // release a waiting worker
1333		break;
827	–	if (untimed)
828	–	LockSupport.park(w);
829	–	else {
830	–	LockSupport.parkNanos(w, SHRINK_RATE_NANOS);
831	–	if (eventCount != ec || !w.isRunning() ||
832	–	runState >= TERMINATING)
833	–	break;
834	–	if (System.nanoTime() - startTime >= SHRINK_RATE_NANOS)
835	–	tryShutdownWaiter(ec);
1334		}
1335		}
1336	+	else
1337	+	break;
1338		}
1339		}
1340
1341		/**
1342	<	* Callback from the oldest waiter in awaitEvent waking up after a
1343	<	* period of non-use. Tries (once) to shutdown an event waiter (or
1344	<	* a spare, if one exists). Note that we don't need CAS or locks
1345	<	* here because the method is called only from one thread
1346	<	* occasionally waking (and even misfires are OK). Note that
1347	<	* until the shutdown worker fully terminates, workerCounts
1348	<	* will overestimate total count, which is tolerable.
1349	<	*
1350	<	* @param ec the event count waited on by caller (to abort
1351	<	* attempt if count has since changed).
1352	<	*/
1353	<	private void tryShutdownWaiter(int ec) {
1354	<	if (spareWaiters != 0) { // prefer killing spares
1355	<	tryShutdownSpare();
1356	<	return;
1357	<	}
1358	<	ForkJoinWorkerThread[] ws = workers;
1359	<	int n = ws.length;
1360	<	long h = eventWaiters;
1361	<	ForkJoinWorkerThread w; int id; long nh;
1362	<	if (runState == 0 &&
1363	<	submissionQueue.isEmpty() &&
1364	<	eventCount == ec &&
1365	<	(id = ((int)(h & WAITER_ID_MASK)) - 1) >= 0 &&
1366	<	id < n && (w = ws[id]) != null &&
1367	<	(nh = w.nextWaiter) != 0L && // keep at least one worker
1368	<	UNSAFE.compareAndSwapLong(this, eventWaitersOffset, h, nh)) {
1369	<	w.shutdown();
1370	<	LockSupport.unpark(w);
1342	>	* Tries to decrement active count (sometimes implicitly) and
1343	>	* possibly release or create a compensating worker in preparation
1344	>	* for blocking. Fails on contention or termination.
1345	>	*
1346	>	* @return true if the caller can block, else should recheck and retry
1347	>	*/
1348	>	final boolean tryCompensate() {
1349	>	WorkQueue[] ws; WorkQueue w; Thread p;
1350	>	int pc = parallelism, e, u, ac, tc, i;
1351	>	long c = ctl;
1352	>
1353	>	if ((e = (int)c) >= 0) {
1354	>	if ((ac = ((u = (int)(c >>> 32)) >> UAC_SHIFT)) <= 0 &&
1355	>	e != 0 && (ws = workQueues) != null &&
1356	>	(i = ((~e << 1) | 1) & SMASK) < ws.length &&
1357	>	(w = ws[i]) != null) {
1358	>	if (w.eventCount == (e | INT_SIGN) &&
1359	>	U.compareAndSwapLong
1360	>	(this, CTL, c, ((long)(w.nextWait & E_MASK) |
1361	>	(c & (AC_MASK|TC_MASK))))) {
1362	>	w.eventCount = (e + E_SEQ) & E_MASK;
1363	>	if ((p = w.parker) != null)
1364	>	U.unpark(p);
1365	>	return true;             // release an idle worker
1366	>	}
1367	>	}
1368	>	else if ((tc = (short)(u >>> UTC_SHIFT)) >= 0 && ac + pc > 1) {
1369	>	long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK);
1370	>	if (U.compareAndSwapLong(this, CTL, c, nc))
1371	>	return true;             // no compensation needed
1372	>	}
1373	>	else if (tc + pc < MAX_ID) {
1374	>	long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK);
1375	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1376	>	addWorker();
1377	>	return true;             // create replacement
1378	>	}
1379	>	}
1380		}
1381	<	releaseEventWaiters();
1381	>	return false;
1382		}
1383
1384	<	// Maintaining spares
876	<
877	<	/**
878	<	* Pushes worker onto the spare stack
879	<	*/
880	<	final void pushSpare(ForkJoinWorkerThread w) {
881	<	int ns = (++w.spareCount << SPARE_COUNT_SHIFT) | (w.poolIndex + 1);
882	<	do {} while (!UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
883	<	w.nextSpare = spareWaiters,ns));
884	<	}
1384	>	// Submissions
1385
1386		/**
1387	<	* Callback from oldest spare occasionally waking up.  Tries
1388	<	* (once) to shutdown a spare. Same idea as tryShutdownWaiter.
1387	>	* Unless shutting down, adds the given task to a submission queue
1388	>	* at submitter's current queue index. If no queue exists at the
1389	>	* index, one is created unless pool lock is busy.  If the queue
1390	>	* and/or lock are busy, another index is randomly chosen.
1391		*/
1392	<	final void tryShutdownSpare() {
1393	<	int sw, id;
1394	<	ForkJoinWorkerThread w;
1395	<	ForkJoinWorkerThread[] ws;
1396	<	if ((id = ((sw = spareWaiters) & SPARE_ID_MASK) - 1) >= 0 &&
1397	<	id < (ws = workers).length && (w = ws[id]) != null &&
1398	<	(workerCounts & RUNNING_COUNT_MASK) >= parallelism &&
1399	<	UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
1400	<	sw, w.nextSpare)) {
1401	<	w.shutdown();
1402	<	LockSupport.unpark(w);
1403	<	advanceEventCount();
1392	>	private void doSubmit(ForkJoinTask<?> task) {
1393	>	if (task == null)
1394	>	throw new NullPointerException();
1395	>	Submitter s = submitters.get();
1396	>	for (int r = s.seed;;) {
1397	>	WorkQueue q; int k;
1398	>	int rs = runState, m = rs & SMASK;
1399	>	WorkQueue[] ws = workQueues;
1400	>	if (rs < 0 || ws == null)   // shutting down
1401	>	throw new RejectedExecutionException();
1402	>	if (ws.length > m &&        // k must be at index
1403	>	((q = ws[k = (r << 1) & m]) != null ||
1404	>	(q = tryAddSharedQueue(k)) != null) &&
1405	>	q.trySharedPush(task)) {
1406	>	signalWork();
1407	>	return;
1408	>	}
1409	>	r ^= r << 13;               // xorshift seed to new position
1410	>	r ^= r >>> 17;
1411	>	if (((s.seed = r ^= r << 5) & m) == 0)
1412	>	Thread.yield();         // occasionally yield if busy
1413	>	}
1414	>	}
1415	>
1416	>
1417	>	// Scanning for tasks
1418	>
1419	>	/**
1420	>	* Scans for and, if found, returns one task, else possibly
1421	>	* inactivates the worker. This method operates on single reads of
1422	>	* volatile state and is designed to be re-invoked continuously in
1423	>	* part because it returns upon detecting inconsistencies,
1424	>	* contention, or state changes that indicate possible success on
1425	>	* re-invocation.
1426	>	*
1427	>	* The scan searches for tasks across queues, randomly selecting
1428	>	* the first #queues probes, favoring steals 2:1 over submissions
1429	>	* (by exploiting even/odd indexing), and then performing a
1430	>	* circular sweep of all queues.  The scan terminates upon either
1431	>	* finding a non-empty queue, or completing a full sweep. If the
1432	>	* worker is not inactivated, it takes and returns a task from
1433	>	* this queue.  On failure to find a task, we take one of the
1434	>	* following actions, after which the caller will retry calling
1435	>	* this method unless terminated.
1436	>	*
1437	>	* * If not a complete sweep, try to release a waiting worker.  If
1438	>	* the scan terminated because the worker is inactivated, then the
1439	>	* released worker will often be the calling worker, and it can
1440	>	* succeed obtaining a task on the next call. Or maybe it is
1441	>	* another worker, but with same net effect. Releasing in other
1442	>	* cases as well ensures that we have enough workers running.
1443	>	*
1444	>	* * If the caller has run a task since the last empty scan,
1445	>	* return (to allow rescan) if other workers are not also yet
1446	>	* enqueued.  Field WorkQueue.rescans counts down on each scan to
1447	>	* ensure eventual inactivation, and occasional calls to
1448	>	* Thread.yield to help avoid interference with more useful
1449	>	* activities on the system.
1450	>	*
1451	>	* * If pool is terminating, terminate the worker.
1452	>	*
1453	>	* * If not already enqueued, try to inactivate and enqueue the
1454	>	* worker on wait queue.
1455	>	*
1456	>	* * If already enqueued and none of the above apply, either park
1457	>	* awaiting signal, or if this is the most recent waiter and pool
1458	>	* is quiescent, relay to idleAwaitWork to check for termination
1459	>	* and possibly shrink pool.
1460	>	*
1461	>	* @param w the worker (via its WorkQueue)
1462	>	* @return a task or null of none found
1463	>	*/
1464	>	private final ForkJoinTask<?> scan(WorkQueue w) {
1465	>	boolean swept = false;                 // true after full empty scan
1466	>	WorkQueue[] ws;                        // volatile read order matters
1467	>	int r = w.seed, ec = w.eventCount;     // ec is negative if inactive
1468	>	int rs = runState, m = rs & SMASK;
1469	>	if ((ws = workQueues) != null && ws.length > m) {
1470	>	ForkJoinTask<?> task = null;
1471	>	for (int k = 0, j = -2 - m; ; ++j) {
1472	>	WorkQueue q; int b;
1473	>	if (j < 0) {                   // random probes while j negative
1474	>	r ^= r << 13; r ^= r >>> 17; k = (r ^= r << 5) | (j & 1);
1475	>	}                              // worker (not submit) for odd j
1476	>	else                           // cyclic scan when j >= 0
1477	>	k += (m >>> 1) | 1;        // step by half to reduce bias
1478	>
1479	>	if ((q = ws[k & m]) != null && (b = q.base) - q.top < 0) {
1480	>	if (ec >= 0)
1481	>	task = q.pollAt(b);    // steal
1482	>	break;
1483	>	}
1484	>	else if (j > m) {
1485	>	if (rs == runState)        // staleness check
1486	>	swept = true;
1487	>	break;
1488	>	}
1489	>	}
1490	>	w.seed = r;                        // save seed for next scan
1491	>	if (task != null)
1492	>	return task;
1493	>	}
1494	>
1495	>	// Decode ctl on empty scan
1496	>	long c = ctl; int e = (int)c, a = (int)(c >> AC_SHIFT), nr, ns;
1497	>	if (!swept) {                          // try to release a waiter
1498	>	WorkQueue v; Thread p;
1499	>	if (e > 0 && a < 0 && ws != null &&
1500	>	(v = ws[((~e << 1) | 1) & m]) != null &&
1501	>	v.eventCount == (e | INT_SIGN) && U.compareAndSwapLong
1502	>	(this, CTL, c, ((long)(v.nextWait & E_MASK) |
1503	>	((c + AC_UNIT) & (AC_MASK|TC_MASK))))) {
1504	>	v.eventCount = (e + E_SEQ) & E_MASK;
1505	>	if ((p = v.parker) != null)
1506	>	U.unpark(p);
1507	>	}
1508		}
1509	+	else if ((nr = w.rescans) > 0) {       // continue rescanning
1510	+	int ac = a + parallelism;
1511	+	if ((w.rescans = (ac < nr) ? ac : nr - 1) > 0 && w.seed < 0 &&
1512	+	w.eventCount == ec)
1513	+	Thread.yield();                // 1 bit randomness for yield call
1514	+	}
1515	+	else if (e < 0)                        // pool is terminating
1516	+	w.runState = -1;
1517	+	else if (ec >= 0) {                    // try to enqueue
1518	+	long nc = (long)ec | ((c - AC_UNIT) & (AC_MASK|TC_MASK));
1519	+	w.nextWait = e;
1520	+	w.eventCount = ec | INT_SIGN;      // mark as inactive
1521	+	if (!U.compareAndSwapLong(this, CTL, c, nc))
1522	+	w.eventCount = ec;             // back out on CAS failure
1523	+	else if ((ns = w.nsteals) != 0) {  // set rescans if ran task
1524	+	if (a <= 0)                    // ... unless too many active
1525	+	w.rescans = a + parallelism;
1526	+	w.nsteals = 0;
1527	+	w.totalSteals += ns;
1528	+	}
1529	+	}
1530	+	else{                                  // already queued
1531	+	if (parallelism == -a)
1532	+	idleAwaitWork(w);              // quiescent
1533	+	if (w.eventCount == ec) {
1534	+	Thread.interrupted();          // clear status
1535	+	ForkJoinWorkerThread wt = w.owner;
1536	+	U.putObject(wt, PARKBLOCKER, this);
1537	+	w.parker = wt;                 // emulate LockSupport.park
1538	+	if (w.eventCount == ec)        // recheck
1539	+	U.park(false, 0L);         // block
1540	+	w.parker = null;
1541	+	U.putObject(wt, PARKBLOCKER, null);
1542	+	}
1543	+	}
1544	+	return null;
1545		}
1546
1547		/**
1548	<	* Tries (once) to resume a spare if worker counts match
1549	<	* the given count.
1550	<	*
1551	<	* @param wc workerCounts value on invocation of this method
1552	<	*/
1553	<	private void tryResumeSpare(int wc) {
1554	<	ForkJoinWorkerThread[] ws = workers;
1555	<	int n = ws.length;
1556	<	int sw, id, rs;  ForkJoinWorkerThread w;
1557	<	if ((id = ((sw = spareWaiters) & SPARE_ID_MASK) - 1) >= 0 &&
1558	<	id < n && (w = ws[id]) != null &&
1559	<	(rs = runState) < TERMINATING &&
1560	<	eventWaiters == 0L && workerCounts == wc) {
1561	<	// In case all workers busy, heuristically back off to let settle
1562	<	Thread.yield();
1563	<	if (eventWaiters == 0L && runState == rs && // recheck
1564	<	workerCounts == wc && spareWaiters == sw &&
1565	<	UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
1566	<	sw, w.nextSpare)) {
1567	<	int c;              // increment running count before resume
1568	<	do {} while(!UNSAFE.compareAndSwapInt
1569	<	(this, workerCountsOffset,
1570	<	c = workerCounts, c + ONE_RUNNING));
1571	<	if (w.tryUnsuspend())
1572	<	LockSupport.unpark(w);
1573	<	else               // back out if w was shutdown
1574	<	decrementWorkerCounts(ONE_RUNNING, 0);
1548	>	* If inactivating worker w has caused pool to become quiescent,
1549	>	* checks for pool termination, and, so long as this is not the
1550	>	* only worker, waits for event for up to SHRINK_RATE nanosecs.
1551	>	* On timeout, if ctl has not changed, terminates the worker,
1552	>	* which will in turn wake up another worker to possibly repeat
1553	>	* this process.
1554	>	*
1555	>	* @param w the calling worker
1556	>	*/
1557	>	private void idleAwaitWork(WorkQueue w) {
1558	>	long c; int nw, ec;
1559	>	if (!tryTerminate(false) &&
1560	>	(int)((c = ctl) >> AC_SHIFT) + parallelism == 0 &&
1561	>	(ec = w.eventCount) == ((int)c | INT_SIGN) &&
1562	>	(nw = w.nextWait) != 0) {
1563	>	long nc = ((long)(nw & E_MASK) | // ctl to restore on timeout
1564	>	((c + AC_UNIT) & AC_MASK) | (c & TC_MASK));
1565	>	ForkJoinTask.helpExpungeStaleExceptions(); // help clean
1566	>	ForkJoinWorkerThread wt = w.owner;
1567	>	while (ctl == c) {
1568	>	long startTime = System.nanoTime();
1569	>	Thread.interrupted();  // timed variant of version in scan()
1570	>	U.putObject(wt, PARKBLOCKER, this);
1571	>	w.parker = wt;
1572	>	if (ctl == c)
1573	>	U.park(false, SHRINK_RATE);
1574	>	w.parker = null;
1575	>	U.putObject(wt, PARKBLOCKER, null);
1576	>	if (ctl != c)
1577	>	break;
1578	>	if (System.nanoTime() - startTime >= SHRINK_TIMEOUT &&
1579	>	U.compareAndSwapLong(this, CTL, c, nc)) {
1580	>	w.runState = -1;          // shrink
1581	>	w.eventCount = (ec + E_SEQ) | E_MASK;
1582	>	break;
1583	>	}
1584		}
1585		}
1586		}
1587
937	–	// adding workers on demand
938	–
1588		/**
1589	<	* Adds one or more workers if needed to establish target parallelism.
1590	<	* Retries upon contention.
1591	<	*/
1592	<	private void addWorkerIfBelowTarget() {
1593	<	int pc = parallelism;
1594	<	int wc;
1595	<	while (((wc = workerCounts) >>> TOTAL_COUNT_SHIFT) < pc &&
1596	<	runState < TERMINATING) {
1597	<	if (UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
1598	<	wc + (ONE_RUNNING|ONE_TOTAL))) {
1599	<	if (addWorker() == null)
1600	<	break;
1589	>	* Tries to locate and execute tasks for a stealer of the given
1590	>	* task, or in turn one of its stealers, Traces currentSteal ->
1591	>	* currentJoin links looking for a thread working on a descendant
1592	>	* of the given task and with a non-empty queue to steal back and
1593	>	* execute tasks from. The first call to this method upon a
1594	>	* waiting join will often entail scanning/search, (which is OK
1595	>	* because the joiner has nothing better to do), but this method
1596	>	* leaves hints in workers to speed up subsequent calls. The
1597	>	* implementation is very branchy to cope with potential
1598	>	* inconsistencies or loops encountering chains that are stale,
1599	>	* unknown, or of length greater than MAX_HELP_DEPTH links.  All
1600	>	* of these cases are dealt with by just retrying by caller.
1601	>	*
1602	>	* @param joiner the joining worker
1603	>	* @param task the task to join
1604	>	* @return true if found or ran a task (and so is immediately retryable)
1605	>	*/
1606	>	final boolean tryHelpStealer(WorkQueue joiner, ForkJoinTask<?> task) {
1607	>	ForkJoinTask<?> subtask;    // current target
1608	>	boolean progress = false;
1609	>	int depth = 0;              // current chain depth
1610	>	int m = runState & SMASK;
1611	>	WorkQueue[] ws = workQueues;
1612	>
1613	>	if (ws != null && ws.length > m && (subtask = task).status >= 0) {
1614	>	outer:for (WorkQueue j = joiner;;) {
1615	>	// Try to find the stealer of subtask, by first using hint
1616	>	WorkQueue stealer = null;
1617	>	WorkQueue v = ws[j.stealHint & m];
1618	>	if (v != null && v.currentSteal == subtask)
1619	>	stealer = v;
1620	>	else {
1621	>	for (int i = 1; i <= m; i += 2) {
1622	>	if ((v = ws[i]) != null && v.currentSteal == subtask) {
1623	>	stealer = v;
1624	>	j.stealHint = i; // save hint
1625	>	break;
1626	>	}
1627	>	}
1628	>	if (stealer == null)
1629	>	break;
1630	>	}
1631	>
1632	>	for (WorkQueue q = stealer;;) { // Try to help stealer
1633	>	ForkJoinTask<?> t; int b;
1634	>	if (task.status < 0)
1635	>	break outer;
1636	>	if ((b = q.base) - q.top < 0) {
1637	>	progress = true;
1638	>	if (subtask.status < 0)
1639	>	break outer;               // stale
1640	>	if ((t = q.pollAt(b)) != null) {
1641	>	stealer.stealHint = joiner.poolIndex;
1642	>	joiner.runSubtask(t);
1643	>	}
1644	>	}
1645	>	else { // empty - try to descend to find stealer's stealer
1646	>	ForkJoinTask<?> next = stealer.currentJoin;
1647	>	if (++depth == MAX_HELP_DEPTH || subtask.status < 0 ||
1648	>	next == null || next == subtask)
1649	>	break outer;  // max depth, stale, dead-end, cyclic
1650	>	subtask = next;
1651	>	j = stealer;
1652	>	break;
1653	>	}
1654	>	}
1655		}
1656		}
1657	+	return progress;
1658		}
1659
1660		/**
1661	<	* Tries (once) to add a new worker if all existing workers are
958	<	* busy, and there are either no running workers or the deficit is
959	<	* at least twice the surplus.
1661	>	* If task is at base of some steal queue, steals and executes it.
1662		*
1663	<	* @param wc workerCounts value on invocation of this method
1663	>	* @param joiner the joining worker
1664	>	* @param task the task
1665		*/
1666	<	private void tryAddWorkerIfBusy(int wc) {
1667	<	int tc, rc, rs;
1668	<	int pc = parallelism;
1669	<	if ((tc = wc >>> TOTAL_COUNT_SHIFT) < MAX_WORKERS &&
1670	<	((rc = wc & RUNNING_COUNT_MASK) == 0 ||
1671	<	rc < pc - ((tc - pc) << 1)) &&
1672	<	(rs = runState) < TERMINATING &&
1673	<	(rs & ACTIVE_COUNT_MASK) == tc) {
1674	<	// Since all workers busy, heuristically back off to let settle
1675	<	Thread.yield();
1676	<	if (eventWaiters == 0L && spareWaiters == 0 && // recheck
974	<	runState == rs && workerCounts == wc &&
975	<	UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
976	<	wc + (ONE_RUNNING|ONE_TOTAL)))
977	<	addWorker();
1666	>	final void tryPollForAndExec(WorkQueue joiner, ForkJoinTask<?> task) {
1667	>	WorkQueue[] ws;
1668	>	int m = runState & SMASK;
1669	>	if ((ws = workQueues) != null && ws.length > m) {
1670	>	for (int j = 1; j <= m && task.status >= 0; j += 2) {
1671	>	WorkQueue q = ws[j];
1672	>	if (q != null && q.pollFor(task)) {
1673	>	joiner.runSubtask(task);
1674	>	break;
1675	>	}
1676	>	}
1677		}
1678		}
1679
1680		/**
1681	<	* Does at most one of:
1682	<	*
1683	<	* 1. Help wake up existing workers waiting for work via
1684	<	*    releaseEventWaiters. (If any exist, then it doesn't
1685	<	*    matter right now if under target parallelism level.)
1686	<	*
1687	<	* 2. If a spare exists, try (once) to resume it via tryResumeSpare.
1688	<	*
1689	<	* 3. If there are not enough total workers, add some
1690	<	*    via addWorkerIfBelowTarget;
1691	<	*
1692	<	* 4. Try (once) to add a new worker if all existing workers
1693	<	*     are busy, via tryAddWorkerIfBusy
1694	<	*/
1695	<	private void helpMaintainParallelism() {
1696	<	long h; int pc, wc;
1697	<	if (((int)((h = eventWaiters) & WAITER_ID_MASK)) != 0) {
1698	<	if ((int)(h >>> EVENT_COUNT_SHIFT) != eventCount)
1699	<	releaseEventWaiters(); // avoid useless call
1700	<	}
1701	<	else if ((pc = parallelism) >
1702	<	((wc = workerCounts) & RUNNING_COUNT_MASK)) {
1703	<	if (spareWaiters != 0)
1704	<	tryResumeSpare(wc);
1705	<	else if ((wc >>> TOTAL_COUNT_SHIFT) < pc)
1007	<	addWorkerIfBelowTarget();
1008	<	else
1009	<	tryAddWorkerIfBusy(wc);
1681	>	* Returns a non-empty steal queue, if one is found during a random,
1682	>	* then cyclic scan, else null.  This method must be retried by
1683	>	* caller if, by the time it tries to use the queue, it is empty.
1684	>	*/
1685	>	private WorkQueue findNonEmptyStealQueue(WorkQueue w) {
1686	>	int r = w.seed;    // Same idea as scan(), but ignoring submissions
1687	>	for (WorkQueue[] ws;;) {
1688	>	int m = runState & SMASK;
1689	>	if ((ws = workQueues) == null)
1690	>	return null;
1691	>	if (ws.length > m) {
1692	>	WorkQueue q;
1693	>	for (int n = m << 2, k = r, j = -n;;) {
1694	>	r ^= r << 13; r ^= r >>> 17; r ^= r << 5;
1695	>	if ((q = ws[(k | 1) & m]) != null && q.base - q.top < 0) {
1696	>	w.seed = r;
1697	>	return q;
1698	>	}
1699	>	else if (j > n)
1700	>	return null;
1701	>	else
1702	>	k = (j++ < 0) ? r : k + ((m >>> 1) | 1);
1703	>
1704	>	}
1705	>	}
1706		}
1707		}
1708
1709		/**
1710	<	* Callback from workers invoked upon each top-level action (i.e.,
1711	<	* stealing a task or taking a submission and running it).
1712	<	* Performs one or more of the following:
1713	<	*
1714	<	* 1. If the worker is active, try to set its active status to
1715	<	*    inactive and update activeCount. On contention, we may try
1716	<	*    again on this or subsequent call.
1717	<	*
1718	<	* 2. Release any existing event waiters that are now relesable
1719	<	*
1720	<	* 3. If there are too many running threads, suspend this worker
1721	<	*    (first forcing inactive if necessary).  If it is not
1722	<	*    needed, it may be killed while suspended via
1723	<	*    tryShutdownSpare. Otherwise, upon resume it rechecks to make
1724	<	*    sure that it is still needed.
1725	<	*
1726	<	* 4. If more than 1 miss, await the next task event via
1727	<	*    eventSync (first forcing inactivation if necessary), upon
1728	<	*    which worker may also be killed, via tryShutdownWaiter.
1729	<	*
1730	<	* 5. Help reactivate other workers via helpMaintainParallelism
1731	<	*
1732	<	* @param w the worker
1733	<	* @param misses the number of scans by caller failing to find work
1734	<	* (saturating at 2 to avoid wraparound)
1735	<	*/
1736	<	final void preStep(ForkJoinWorkerThread w, int misses) {
1737	<	boolean active = w.active;
1738	<	int pc = parallelism;
1739	<	for (;;) {
1740	<	int rs, wc, rc, ec; long h;
1741	<	if (active && UNSAFE.compareAndSwapInt(this, runStateOffset,
1742	<	rs = runState, rs - 1))
1047	<	active = w.active = false;
1048	<	if (((int)((h = eventWaiters) & WAITER_ID_MASK)) != 0 &&
1049	<	(int)(h >>> EVENT_COUNT_SHIFT) != eventCount) {
1050	<	releaseEventWaiters();
1051	<	if (misses > 1)
1052	<	continue;                  // clear before sync below
1053	<	}
1054	<	if ((rc = ((wc = workerCounts) & RUNNING_COUNT_MASK)) > pc) {
1055	<	if (!active &&                 // must inactivate to suspend
1056	<	workerCounts == wc &&      // try to suspend as spare
1057	<	UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1058	<	wc, wc - ONE_RUNNING)) {
1059	<	w.suspendAsSpare();
1060	<	if (!w.isRunning())
1061	<	break;                 // was killed while spare
1062	<	}
1063	<	continue;
1064	<	}
1065	<	if (misses > 0) {
1066	<	if ((ec = eventCount) == w.lastEventCount && misses > 1) {
1067	<	if (!active) {             // must inactivate to sync
1068	<	eventSync(w);
1069	<	if (w.isRunning())
1070	<	misses = 1;        // don't re-sync
1071	<	else
1072	<	break;             // was killed while waiting
1073	<	}
1074	<	continue;
1710	>	* Runs tasks until {@code isQuiescent()}. We piggyback on
1711	>	* active count ctl maintenance, but rather than blocking
1712	>	* when tasks cannot be found, we rescan until all others cannot
1713	>	* find tasks either.
1714	>	*/
1715	>	final void helpQuiescePool(WorkQueue w) {
1716	>	for (boolean active = true;;) {
1717	>	w.runLocalTasks();      // exhaust local queue
1718	>	WorkQueue q = findNonEmptyStealQueue(w);
1719	>	if (q != null) {
1720	>	ForkJoinTask<?> t;
1721	>	if (!active) {      // re-establish active count
1722	>	long c;
1723	>	active = true;
1724	>	do {} while (!U.compareAndSwapLong
1725	>	(this, CTL, c = ctl, c + AC_UNIT));
1726	>	}
1727	>	if ((t = q.poll()) != null)
1728	>	w.runSubtask(t);
1729	>	}
1730	>	else {
1731	>	long c;
1732	>	if (active) {       // decrement active count without queuing
1733	>	active = false;
1734	>	do {} while (!U.compareAndSwapLong
1735	>	(this, CTL, c = ctl, c -= AC_UNIT));
1736	>	}
1737	>	else
1738	>	c = ctl;        // re-increment on exit
1739	>	if ((int)(c >> AC_SHIFT) + parallelism == 0) {
1740	>	do {} while (!U.compareAndSwapLong
1741	>	(this, CTL, c = ctl, c + AC_UNIT));
1742	>	break;
1743		}
1076	–	w.lastEventCount = ec;
1744		}
1078	–	if (rc < pc)
1079	–	helpMaintainParallelism();
1080	–	break;
1745		}
1746		}
1747
1748		/**
1749	<	* Helps and/or blocks awaiting join of the given task.
1086	<	* Alternates between helpJoinTask() and helpMaintainParallelism()
1087	<	* as many times as there is a deficit in running count (or longer
1088	<	* if running count would become zero), then blocks if task still
1089	<	* not done.
1749	>	* Gets and removes a local or stolen task for the given worker.
1750		*
1751	<	* @param joinMe the task to join
1751	>	* @return a task, if available
1752		*/
1753	<	final void awaitJoin(ForkJoinTask<?> joinMe, ForkJoinWorkerThread worker) {
1754	<	int threshold = parallelism;         // descend blocking thresholds
1755	<	while (joinMe.status >= 0) {
1756	<	boolean block; int wc;
1757	<	worker.helpJoinTask(joinMe);
1758	<	if (joinMe.status < 0)
1759	<	break;
1760	<	if (((wc = workerCounts) & RUNNING_COUNT_MASK) <= threshold) {
1761	<	if (threshold > 0)
1102	<	--threshold;
1103	<	else
1104	<	advanceEventCount(); // force release
1105	<	block = false;
1106	<	}
1107	<	else
1108	<	block = UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1109	<	wc, wc - ONE_RUNNING);
1110	<	helpMaintainParallelism();
1111	<	if (block) {
1112	<	int c;
1113	<	joinMe.internalAwaitDone();
1114	<	do {} while (!UNSAFE.compareAndSwapInt
1115	<	(this, workerCountsOffset,
1116	<	c = workerCounts, c + ONE_RUNNING));
1117	<	break;
1118	<	}
1753	>	final ForkJoinTask<?> nextTaskFor(WorkQueue w) {
1754	>	for (ForkJoinTask<?> t;;) {
1755	>	WorkQueue q;
1756	>	if ((t = w.nextLocalTask()) != null)
1757	>	return t;
1758	>	if ((q = findNonEmptyStealQueue(w)) == null)
1759	>	return null;
1760	>	if ((t = q.poll()) != null)
1761	>	return t;
1762		}
1763		}
1764
1765		/**
1766	<	* Same idea as awaitJoin, but no helping
1766	>	* Returns the approximate (non-atomic) number of idle threads per
1767	>	* active thread to offset steal queue size for method
1768	>	* ForkJoinTask.getSurplusQueuedTaskCount().
1769		*/
1770	<	final void awaitBlocker(ManagedBlocker blocker)
1771	<	throws InterruptedException {
1772	<	int threshold = parallelism;
1773	<	while (!blocker.isReleasable()) {
1774	<	boolean block; int wc;
1775	<	if (((wc = workerCounts) & RUNNING_COUNT_MASK) <= threshold) {
1776	<	if (threshold > 0)
1777	<	--threshold;
1778	<	else
1779	<	advanceEventCount();
1780	<	block = false;
1781	<	}
1782	<	else
1783	<	block = UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1784	<	wc, wc - ONE_RUNNING);
1785	<	helpMaintainParallelism();
1786	<	if (block) {
1787	<	try {
1788	<	do {} while (!blocker.isReleasable() && !blocker.block());
1789	<	} finally {
1790	<	int c;
1791	<	do {} while (!UNSAFE.compareAndSwapInt
1147	<	(this, workerCountsOffset,
1148	<	c = workerCounts, c + ONE_RUNNING));
1149	<	}
1150	<	break;
1151	<	}
1770	>	final int idlePerActive() {
1771	>	// Approximate at powers of two for small values, saturate past 4
1772	>	int p = parallelism;
1773	>	int a = p + (int)(ctl >> AC_SHIFT);
1774	>	return (a > (p >>>= 1) ? 0 :
1775	>	a > (p >>>= 1) ? 1 :
1776	>	a > (p >>>= 1) ? 2 :
1777	>	a > (p >>>= 1) ? 4 :
1778	>	8);
1779	>	}
1780	>
1781	>	// Termination
1782	>
1783	>	/**
1784	>	* Sets SHUTDOWN bit of runState under lock
1785	>	*/
1786	>	private void enableShutdown() {
1787	>	ReentrantLock lock = this.lock;
1788	>	if (runState >= 0) {
1789	>	lock.lock();                       // don't need try/finally
1790	>	runState |= SHUTDOWN;
1791	>	lock.unlock();
1792		}
1793		}
1794
1795		/**
1796	<	* Possibly initiates and/or completes termination.
1796	>	* Possibly initiates and/or completes termination.  Upon
1797	>	* termination, cancels all queued tasks and then
1798		*
1799		* @param now if true, unconditionally terminate, else only
1800	<	* if shutdown and empty queue and no active workers
1800	>	* if no work and no active workers
1801		* @return true if now terminating or terminated
1802		*/
1803		private boolean tryTerminate(boolean now) {
1804	<	if (now)
1805	<	advanceRunLevel(SHUTDOWN); // ensure at least SHUTDOWN
1806	<	else if (runState < SHUTDOWN ||
1807	<	!submissionQueue.isEmpty() ||
1808	<	(runState & ACTIVE_COUNT_MASK) != 0)
1809	<	return false;
1810	<
1811	<	if (advanceRunLevel(TERMINATING))
1812	<	startTerminating();
1813	<
1814	<	// Finish now if all threads terminated; else in some subsequent call
1815	<	if ((workerCounts >>> TOTAL_COUNT_SHIFT) == 0) {
1816	<	advanceRunLevel(TERMINATED);
1817	<	termination.arrive();
1804	>	for (long c;;) {
1805	>	if (((c = ctl) & STOP_BIT) != 0) {      // already terminating
1806	>	if ((short)(c >>> TC_SHIFT) == -parallelism) {
1807	>	ReentrantLock lock = this.lock; // signal when no workers
1808	>	lock.lock();                    // don't need try/finally
1809	>	termination.signalAll();        // signal when 0 workers
1810	>	lock.unlock();
1811	>	}
1812	>	return true;
1813	>	}
1814	>	if (!now) {
1815	>	if ((int)(c >> AC_SHIFT) != -parallelism || runState >= 0 ||
1816	>	hasQueuedSubmissions())
1817	>	return false;
1818	>	// Check for unqueued inactive workers. One pass suffices.
1819	>	WorkQueue[] ws = workQueues; WorkQueue w;
1820	>	if (ws != null) {
1821	>	int n = ws.length;
1822	>	for (int i = 1; i < n; i += 2) {
1823	>	if ((w = ws[i]) != null && w.eventCount >= 0)
1824	>	return false;
1825	>	}
1826	>	}
1827	>	}
1828	>	if (U.compareAndSwapLong(this, CTL, c, c | STOP_BIT))
1829	>	startTerminating();
1830		}
1178	–	return true;
1831		}
1832
1833		/**
1834	<	* Actions on transition to TERMINATING
1835	<	*
1836	<	* Runs up to four passes through workers: (0) shutting down each
1837	<	* (without waking up if parked) to quickly spread notifications
1838	<	* without unnecessary bouncing around event queues etc (1) wake
1839	<	* up and help cancel tasks (2) interrupt (3) mop up races with
1188	<	* interrupted workers
1834	>	* Initiates termination: Runs three passes through workQueues:
1835	>	* (0) Setting termination status, followed by wakeups of queued
1836	>	* workers; (1) cancelling all tasks; (2) interrupting lagging
1837	>	* threads (likely in external tasks, but possibly also blocked in
1838	>	* joins).  Each pass repeats previous steps because of potential
1839	>	* lagging thread creation.
1840		*/
1841		private void startTerminating() {
1842	<	cancelSubmissions();
1843	<	for (int passes = 0; passes < 4 && workerCounts != 0; ++passes) {
1844	<	advanceEventCount();
1845	<	eventWaiters = 0L; // clobber lists
1846	<	spareWaiters = 0;
1847	<	ForkJoinWorkerThread[] ws = workers;
1848	<	int n = ws.length;
1849	<	for (int i = 0; i < n; ++i) {
1850	<	ForkJoinWorkerThread w = ws[i];
1851	<	if (w != null) {
1852	<	w.shutdown();
1853	<	if (passes > 0 && !w.isTerminated()) {
1854	<	w.cancelTasks();
1855	<	LockSupport.unpark(w);
1856	<	if (passes > 1) {
1857	<	try {
1207	<	w.interrupt();
1208	<	} catch (SecurityException ignore) {
1842	>	for (int pass = 0; pass < 3; ++pass) {
1843	>	WorkQueue[] ws = workQueues;
1844	>	if (ws != null) {
1845	>	WorkQueue w; Thread wt;
1846	>	int n = ws.length;
1847	>	for (int j = 0; j < n; ++j) {
1848	>	if ((w = ws[j]) != null) {
1849	>	w.runState = -1;
1850	>	if (pass > 0) {
1851	>	w.cancelAll();
1852	>	if (pass > 1 && (wt = w.owner) != null &&
1853	>	!wt.isInterrupted()) {
1854	>	try {
1855	>	wt.interrupt();
1856	>	} catch (SecurityException ignore) {
1857	>	}
1858		}
1859		}
1860		}
1861		}
1862	+	// Wake up workers parked on event queue
1863	+	int i, e; long c; Thread p;
1864	+	while ((i = ((~(e = (int)(c = ctl)) << 1) | 1) & SMASK) < n &&
1865	+	(w = ws[i]) != null &&
1866	+	w.eventCount == (e | INT_SIGN)) {
1867	+	long nc = ((long)(w.nextWait & E_MASK) |
1868	+	((c + AC_UNIT) & AC_MASK) |
1869	+	(c & (TC_MASK|STOP_BIT)));
1870	+	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1871	+	w.eventCount = (e + E_SEQ) & E_MASK;
1872	+	if ((p = w.parker) != null)
1873	+	U.unpark(p);
1874	+	}
1875	+	}
1876		}
1877		}
1878		}
1879
1880	<	/**
1218	<	* Clear out and cancel submissions, ignoring exceptions
1219	<	*/
1220	<	private void cancelSubmissions() {
1221	<	ForkJoinTask<?> task;
1222	<	while ((task = submissionQueue.poll()) != null) {
1223	<	try {
1224	<	task.cancel(false);
1225	<	} catch (Throwable ignore) {
1226	<	}
1227	<	}
1228	<	}
1229	<
1230	<	// misc support for ForkJoinWorkerThread
1231	<
1232	<	/**
1233	<	* Returns pool number
1234	<	*/
1235	<	final int getPoolNumber() {
1236	<	return poolNumber;
1237	<	}
1238	<
1239	<	/**
1240	<	* Tries to accumulates steal count from a worker, clearing
1241	<	* the worker's value.
1242	<	*
1243	<	* @return true if worker steal count now zero
1244	<	*/
1245	<	final boolean tryAccumulateStealCount(ForkJoinWorkerThread w) {
1246	<	int sc = w.stealCount;
1247	<	long c = stealCount;
1248	<	// CAS even if zero, for fence effects
1249	<	if (UNSAFE.compareAndSwapLong(this, stealCountOffset, c, c + sc)) {
1250	<	if (sc != 0)
1251	<	w.stealCount = 0;
1252	<	return true;
1253	<	}
1254	<	return sc == 0;
1255	<	}
1256	<
1257	<	/**
1258	<	* Returns the approximate (non-atomic) number of idle threads per
1259	<	* active thread.
1260	<	*/
1261	<	final int idlePerActive() {
1262	<	int pc = parallelism; // use parallelism, not rc
1263	<	int ac = runState;    // no mask -- artifically boosts during shutdown
1264	<	// Use exact results for small values, saturate past 4
1265	<	return pc <= ac? 0 : pc >>> 1 <= ac? 1 : pc >>> 2 <= ac? 3 : pc >>> 3;
1266	<	}
1267	<
1268	<	// Public and protected methods
1880	>	// Exported methods
1881
1882		// Constructors
1883
#	Line 1312 | Line 1924 | public class ForkJoinPool extends Abstra
1924		* use {@link #defaultForkJoinWorkerThreadFactory}.
1925		* @param handler the handler for internal worker threads that
1926		* terminate due to unrecoverable errors encountered while executing
1927	<	* tasks. For default value, use <code>null</code>.
1927	>	* tasks. For default value, use {@code null}.
1928		* @param asyncMode if true,
1929		* establishes local first-in-first-out scheduling mode for forked
1930		* tasks that are never joined. This mode may be more appropriate
1931		* than default locally stack-based mode in applications in which
1932		* worker threads only process event-style asynchronous tasks.
1933	<	* For default value, use <code>false</code>.
1933	>	* For default value, use {@code false}.
1934		* @throws IllegalArgumentException if parallelism less than or
1935		*         equal to zero, or greater than implementation limit
1936		* @throws NullPointerException if the factory is null
#	Line 1334 | Line 1946 | public class ForkJoinPool extends Abstra
1946		checkPermission();
1947		if (factory == null)
1948		throw new NullPointerException();
1949	<	if (parallelism <= 0 || parallelism > MAX_WORKERS)
1949	>	if (parallelism <= 0 || parallelism > MAX_ID)
1950		throw new IllegalArgumentException();
1951		this.parallelism = parallelism;
1952		this.factory = factory;
1953		this.ueh = handler;
1954	<	this.locallyFifo = asyncMode;
1955	<	int arraySize = initialArraySizeFor(parallelism);
1956	<	this.workers = new ForkJoinWorkerThread[arraySize];
1957	<	this.submissionQueue = new LinkedTransferQueue<ForkJoinTask<?>>();
1958	<	this.workerLock = new ReentrantLock();
1959	<	this.termination = new Phaser(1);
1960	<	this.poolNumber = poolNumberGenerator.incrementAndGet();
1961	<	}
1962	<
1963	<	/**
1964	<	* Returns initial power of two size for workers array.
1965	<	* @param pc the initial parallelism level
1966	<	*/
1967	<	private static int initialArraySizeFor(int pc) {
1968	<	// See Hackers Delight, sec 3.2. We know MAX_WORKERS < (1 >>> 16)
1969	<	int size = pc < MAX_WORKERS ? pc + 1 : MAX_WORKERS;
1970	<	size |= size >>> 1;
1971	<	size |= size >>> 2;
1972	<	size |= size >>> 4;
1973	<	size |= size >>> 8;
1974	<	return size + 1;
1954	>	this.localMode = asyncMode ? FIFO_QUEUE : LIFO_QUEUE;
1955	>	this.nextPoolIndex = 1;
1956	>	long np = (long)(-parallelism); // offset ctl counts
1957	>	this.ctl = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK);
1958	>	// initialize workQueues array with room for 2*parallelism if possible
1959	>	int n = parallelism << 1;
1960	>	if (n >= MAX_ID)
1961	>	n = MAX_ID;
1962	>	else { // See Hackers Delight, sec 3.2, where n < (1 << 16)
1963	>	n |= n >>> 1; n |= n >>> 2; n |= n >>> 4; n |= n >>> 8;
1964	>	}
1965	>	this.workQueues = new WorkQueue[(n + 1) << 1];
1966	>	ReentrantLock lck = this.lock = new ReentrantLock();
1967	>	this.termination = lck.newCondition();
1968	>	this.stealCount = new AtomicLong();
1969	>	this.nextWorkerNumber = new AtomicInteger();
1970	>	StringBuilder sb = new StringBuilder("ForkJoinPool-");
1971	>	sb.append(poolNumberGenerator.incrementAndGet());
1972	>	sb.append("-worker-");
1973	>	this.workerNamePrefix = sb.toString();
1974	>	// Create initial submission queue
1975	>	WorkQueue sq = tryAddSharedQueue(0);
1976	>	if (sq != null)
1977	>	sq.growArray(false);
1978		}
1979
1980		// Execution methods
1981
1982		/**
1368	–	* Common code for execute, invoke and submit
1369	–	*/
1370	–	private <T> void doSubmit(ForkJoinTask<T> task) {
1371	–	if (task == null)
1372	–	throw new NullPointerException();
1373	–	if (runState >= SHUTDOWN)
1374	–	throw new RejectedExecutionException();
1375	–	submissionQueue.offer(task);
1376	–	advanceEventCount();
1377	–	if (eventWaiters != 0L)
1378	–	releaseEventWaiters();
1379	–	if ((workerCounts >>> TOTAL_COUNT_SHIFT) < parallelism)
1380	–	addWorkerIfBelowTarget();
1381	–	}
1382	–
1383	–	/**
1983		* Performs the given task, returning its result upon completion.
1984	+	* If the computation encounters an unchecked Exception or Error,
1985	+	* it is rethrown as the outcome of this invocation.  Rethrown
1986	+	* exceptions behave in the same way as regular exceptions, but,
1987	+	* when possible, contain stack traces (as displayed for example
1988	+	* using {@code ex.printStackTrace()}) of both the current thread
1989	+	* as well as the thread actually encountering the exception;
1990	+	* minimally only the latter.
1991		*
1992		* @param task the task
1993		* @return the task's result
#	Line 1414 | Line 2020 | public class ForkJoinPool extends Abstra
2020		*         scheduled for execution
2021		*/
2022		public void execute(Runnable task) {
2023	+	if (task == null)
2024	+	throw new NullPointerException();
2025		ForkJoinTask<?> job;
2026		if (task instanceof ForkJoinTask<?>) // avoid re-wrap
2027		job = (ForkJoinTask<?>) task;
#	Line 1442 | Line 2050 | public class ForkJoinPool extends Abstra
2050		*         scheduled for execution
2051		*/
2052		public <T> ForkJoinTask<T> submit(Callable<T> task) {
2053	+	if (task == null)
2054	+	throw new NullPointerException();
2055		ForkJoinTask<T> job = ForkJoinTask.adapt(task);
2056		doSubmit(job);
2057		return job;
#	Line 1453 | Line 2063 | public class ForkJoinPool extends Abstra
2063		*         scheduled for execution
2064		*/
2065		public <T> ForkJoinTask<T> submit(Runnable task, T result) {
2066	+	if (task == null)
2067	+	throw new NullPointerException();
2068		ForkJoinTask<T> job = ForkJoinTask.adapt(task, result);
2069		doSubmit(job);
2070		return job;
#	Line 1464 | Line 2076 | public class ForkJoinPool extends Abstra
2076		*         scheduled for execution
2077		*/
2078		public ForkJoinTask<?> submit(Runnable task) {
2079	+	if (task == null)
2080	+	throw new NullPointerException();
2081		ForkJoinTask<?> job;
2082		if (task instanceof ForkJoinTask<?>) // avoid re-wrap
2083		job = (ForkJoinTask<?>) task;
#	Line 1529 | Line 2143 | public class ForkJoinPool extends Abstra
2143
2144		/**
2145		* Returns the number of worker threads that have started but not
2146	<	* yet terminated.  This result returned by this method may differ
2146	>	* yet terminated.  The result returned by this method may differ
2147		* from {@link #getParallelism} when threads are created to
2148		* maintain parallelism when others are cooperatively blocked.
2149		*
2150		* @return the number of worker threads
2151		*/
2152		public int getPoolSize() {
2153	<	return workerCounts >>> TOTAL_COUNT_SHIFT;
2153	>	return parallelism + (short)(ctl >>> TC_SHIFT);
2154		}
2155
2156		/**
#	Line 1546 | Line 2160 | public class ForkJoinPool extends Abstra
2160		* @return {@code true} if this pool uses async mode
2161		*/
2162		public boolean getAsyncMode() {
2163	<	return locallyFifo;
2163	>	return localMode != 0;
2164		}
2165
2166		/**
#	Line 1558 | Line 2172 | public class ForkJoinPool extends Abstra
2172		* @return the number of worker threads
2173		*/
2174		public int getRunningThreadCount() {
2175	<	return workerCounts & RUNNING_COUNT_MASK;
2175	>	int rc = 0;
2176	>	WorkQueue[] ws; WorkQueue w;
2177	>	if ((ws = workQueues) != null) {
2178	>	int n = ws.length;
2179	>	for (int i = 1; i < n; i += 2) {
2180	>	Thread.State s; ForkJoinWorkerThread wt;
2181	>	if ((w = ws[i]) != null && (wt = w.owner) != null &&
2182	>	w.eventCount >= 0 &&
2183	>	(s = wt.getState()) != Thread.State.BLOCKED &&
2184	>	s != Thread.State.WAITING &&
2185	>	s != Thread.State.TIMED_WAITING)
2186	>	++rc;
2187	>	}
2188	>	}
2189	>	return rc;
2190		}
2191
2192		/**
#	Line 1569 | Line 2197 | public class ForkJoinPool extends Abstra
2197		* @return the number of active threads
2198		*/
2199		public int getActiveThreadCount() {
2200	<	return runState & ACTIVE_COUNT_MASK;
2200	>	int r = parallelism + (int)(ctl >> AC_SHIFT);
2201	>	return (r <= 0) ? 0 : r; // suppress momentarily negative values
2202		}
2203
2204		/**
#	Line 1584 | Line 2213 | public class ForkJoinPool extends Abstra
2213		* @return {@code true} if all threads are currently idle
2214		*/
2215		public boolean isQuiescent() {
2216	<	return (runState & ACTIVE_COUNT_MASK) == 0;
2216	>	return (int)(ctl >> AC_SHIFT) + parallelism == 0;
2217		}
2218
2219		/**
#	Line 1599 | Line 2228 | public class ForkJoinPool extends Abstra
2228		* @return the number of steals
2229		*/
2230		public long getStealCount() {
2231	<	return stealCount;
2231	>	long count = stealCount.get();
2232	>	WorkQueue[] ws; WorkQueue w;
2233	>	if ((ws = workQueues) != null) {
2234	>	int n = ws.length;
2235	>	for (int i = 1; i < n; i += 2) {
2236	>	if ((w = ws[i]) != null)
2237	>	count += w.totalSteals;
2238	>	}
2239	>	}
2240	>	return count;
2241		}
2242
2243		/**
#	Line 1614 | Line 2252 | public class ForkJoinPool extends Abstra
2252		*/
2253		public long getQueuedTaskCount() {
2254		long count = 0;
2255	<	ForkJoinWorkerThread[] ws = workers;
2256	<	int n = ws.length;
2257	<	for (int i = 0; i < n; ++i) {
2258	<	ForkJoinWorkerThread w = ws[i];
2259	<	if (w != null)
2260	<	count += w.getQueueSize();
2255	>	WorkQueue[] ws; WorkQueue w;
2256	>	if ((ws = workQueues) != null) {
2257	>	int n = ws.length;
2258	>	for (int i = 1; i < n; i += 2) {
2259	>	if ((w = ws[i]) != null)
2260	>	count += w.queueSize();
2261	>	}
2262		}
2263		return count;
2264		}
2265
2266		/**
2267		* Returns an estimate of the number of tasks submitted to this
2268	<	* pool that have not yet begun executing.  This method takes time
2269	<	* proportional to the number of submissions.
2268	>	* pool that have not yet begun executing.  This method may take
2269	>	* time proportional to the number of submissions.
2270		*
2271		* @return the number of queued submissions
2272		*/
2273		public int getQueuedSubmissionCount() {
2274	<	return submissionQueue.size();
2274	>	int count = 0;
2275	>	WorkQueue[] ws; WorkQueue w;
2276	>	if ((ws = workQueues) != null) {
2277	>	int n = ws.length;
2278	>	for (int i = 0; i < n; i += 2) {
2279	>	if ((w = ws[i]) != null)
2280	>	count += w.queueSize();
2281	>	}
2282	>	}
2283	>	return count;
2284		}
2285
2286		/**
#	Line 1642 | Line 2290 | public class ForkJoinPool extends Abstra
2290		* @return {@code true} if there are any queued submissions
2291		*/
2292		public boolean hasQueuedSubmissions() {
2293	<	return !submissionQueue.isEmpty();
2293	>	WorkQueue[] ws; WorkQueue w;
2294	>	if ((ws = workQueues) != null) {
2295	>	int n = ws.length;
2296	>	for (int i = 0; i < n; i += 2) {
2297	>	if ((w = ws[i]) != null && w.queueSize() != 0)
2298	>	return true;
2299	>	}
2300	>	}
2301	>	return false;
2302		}
2303
2304		/**
#	Line 1653 | Line 2309 | public class ForkJoinPool extends Abstra
2309		* @return the next submission, or {@code null} if none
2310		*/
2311		protected ForkJoinTask<?> pollSubmission() {
2312	<	return submissionQueue.poll();
2312	>	WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2313	>	if ((ws = workQueues) != null) {
2314	>	int n = ws.length;
2315	>	for (int i = 0; i < n; i += 2) {
2316	>	if ((w = ws[i]) != null && (t = w.poll()) != null)
2317	>	return t;
2318	>	}
2319	>	}
2320	>	return null;
2321		}
2322
2323		/**
#	Line 1674 | Line 2338 | public class ForkJoinPool extends Abstra
2338		* @return the number of elements transferred
2339		*/
2340		protected int drainTasksTo(Collection<? super ForkJoinTask<?>> c) {
2341	<	int count = submissionQueue.drainTo(c);
2342	<	ForkJoinWorkerThread[] ws = workers;
2343	<	int n = ws.length;
2344	<	for (int i = 0; i < n; ++i) {
2345	<	ForkJoinWorkerThread w = ws[i];
2346	<	if (w != null)
2347	<	count += w.drainTasksTo(c);
2341	>	int count = 0;
2342	>	WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2343	>	if ((ws = workQueues) != null) {
2344	>	int n = ws.length;
2345	>	for (int i = 0; i < n; ++i) {
2346	>	if ((w = ws[i]) != null) {
2347	>	while ((t = w.poll()) != null) {
2348	>	c.add(t);
2349	>	++count;
2350	>	}
2351	>	}
2352	>	}
2353		}
2354		return count;
2355		}
#	Line 1696 | Line 2365 | public class ForkJoinPool extends Abstra
2365		long st = getStealCount();
2366		long qt = getQueuedTaskCount();
2367		long qs = getQueuedSubmissionCount();
2368	<	int wc = workerCounts;
1700	<	int tc = wc >>> TOTAL_COUNT_SHIFT;
1701	<	int rc = wc & RUNNING_COUNT_MASK;
2368	>	int rc = getRunningThreadCount();
2369		int pc = parallelism;
2370	<	int rs = runState;
2371	<	int ac = rs & ACTIVE_COUNT_MASK;
2370	>	long c = ctl;
2371	>	int tc = pc + (short)(c >>> TC_SHIFT);
2372	>	int ac = pc + (int)(c >> AC_SHIFT);
2373	>	if (ac < 0) // ignore transient negative
2374	>	ac = 0;
2375	>	String level;
2376	>	if ((c & STOP_BIT) != 0)
2377	>	level = (tc == 0) ? "Terminated" : "Terminating";
2378	>	else
2379	>	level = runState < 0 ? "Shutting down" : "Running";
2380		return super.toString() +
2381	<	"[" + runLevelToString(rs) +
2381	>	"[" + level +
2382		", parallelism = " + pc +
2383		", size = " + tc +
2384		", active = " + ac +
#	Line 1714 | Line 2389 | public class ForkJoinPool extends Abstra
2389		"]";
2390		}
2391
1717	–	private static String runLevelToString(int s) {
1718	–	return ((s & TERMINATED) != 0 ? "Terminated" :
1719	–	((s & TERMINATING) != 0 ? "Terminating" :
1720	–	((s & SHUTDOWN) != 0 ? "Shutting down" :
1721	–	"Running")));
1722	–	}
1723	–
2392		/**
2393		* Initiates an orderly shutdown in which previously submitted
2394		* tasks are executed, but no new tasks will be accepted.
#	Line 1735 | Line 2403 | public class ForkJoinPool extends Abstra
2403		*/
2404		public void shutdown() {
2405		checkPermission();
2406	<	advanceRunLevel(SHUTDOWN);
2406	>	enableShutdown();
2407		tryTerminate(false);
2408		}
2409
#	Line 1757 | Line 2425 | public class ForkJoinPool extends Abstra
2425		*/
2426		public List<Runnable> shutdownNow() {
2427		checkPermission();
2428	+	enableShutdown();
2429		tryTerminate(true);
2430		return Collections.emptyList();
2431		}
#	Line 1767 | Line 2436 | public class ForkJoinPool extends Abstra
2436		* @return {@code true} if all tasks have completed following shut down
2437		*/
2438		public boolean isTerminated() {
2439	<	return runState >= TERMINATED;
2439	>	long c = ctl;
2440	>	return ((c & STOP_BIT) != 0L &&
2441	>	(short)(c >>> TC_SHIFT) == -parallelism);
2442		}
2443
2444		/**
#	Line 1775 | Line 2446 | public class ForkJoinPool extends Abstra
2446		* commenced but not yet completed.  This method may be useful for
2447		* debugging. A return of {@code true} reported a sufficient
2448		* period after shutdown may indicate that submitted tasks have
2449	<	* ignored or suppressed interruption, causing this executor not
2450	<	* to properly terminate.
2449	>	* ignored or suppressed interruption, or are waiting for IO,
2450	>	* causing this executor not to properly terminate. (See the
2451	>	* advisory notes for class {@link ForkJoinTask} stating that
2452	>	* tasks should not normally entail blocking operations.  But if
2453	>	* they do, they must abort them on interrupt.)
2454		*
2455		* @return {@code true} if terminating but not yet terminated
2456		*/
2457		public boolean isTerminating() {
2458	<	return (runState & (TERMINATING|TERMINATED)) == TERMINATING;
2458	>	long c = ctl;
2459	>	return ((c & STOP_BIT) != 0L &&
2460	>	(short)(c >>> TC_SHIFT) != -parallelism);
2461		}
2462
2463		/**
#	Line 1790 | Line 2466 | public class ForkJoinPool extends Abstra
2466		* @return {@code true} if this pool has been shut down
2467		*/
2468		public boolean isShutdown() {
2469	<	return runState >= SHUTDOWN;
2469	>	return runState < 0;
2470		}
2471
2472		/**
#	Line 1806 | Line 2482 | public class ForkJoinPool extends Abstra
2482		*/
2483		public boolean awaitTermination(long timeout, TimeUnit unit)
2484		throws InterruptedException {
2485	+	long nanos = unit.toNanos(timeout);
2486	+	final ReentrantLock lock = this.lock;
2487	+	lock.lock();
2488		try {
2489	<	return termination.awaitAdvanceInterruptibly(0, timeout, unit) > 0;
2490	<	} catch(TimeoutException ex) {
2491	<	return false;
2489	>	for (;;) {
2490	>	if (isTerminated())
2491	>	return true;
2492	>	if (nanos <= 0)
2493	>	return false;
2494	>	nanos = termination.awaitNanos(nanos);
2495	>	}
2496	>	} finally {
2497	>	lock.unlock();
2498		}
2499		}
2500
#	Line 1821 | Line 2506 | public class ForkJoinPool extends Abstra
2506		* {@code isReleasable} must return {@code true} if blocking is
2507		* not necessary. Method {@code block} blocks the current thread
2508		* if necessary (perhaps internally invoking {@code isReleasable}
2509	<	* before actually blocking). The unusual methods in this API
2510	<	* accommodate synchronizers that may, but don't usually, block
2511	<	* for long periods. Similarly, they allow more efficient internal
2512	<	* handling of cases in which additional workers may be, but
2513	<	* usually are not, needed to ensure sufficient parallelism.
2514	<	* Toward this end, implementations of method {@code isReleasable}
2515	<	* must be amenable to repeated invocation.
2509	>	* before actually blocking). These actions are performed by any
2510	>	* thread invoking {@link ForkJoinPool#managedBlock}.  The
2511	>	* unusual methods in this API accommodate synchronizers that may,
2512	>	* but don't usually, block for long periods. Similarly, they
2513	>	* allow more efficient internal handling of cases in which
2514	>	* additional workers may be, but usually are not, needed to
2515	>	* ensure sufficient parallelism.  Toward this end,
2516	>	* implementations of method {@code isReleasable} must be amenable
2517	>	* to repeated invocation.
2518		*
2519		* <p>For example, here is a ManagedBlocker based on a
2520		* ReentrantLock:
#	Line 1907 | Line 2594 | public class ForkJoinPool extends Abstra
2594		public static void managedBlock(ManagedBlocker blocker)
2595		throws InterruptedException {
2596		Thread t = Thread.currentThread();
2597	<	if (t instanceof ForkJoinWorkerThread) {
2598	<	ForkJoinWorkerThread w = (ForkJoinWorkerThread) t;
2599	<	w.pool.awaitBlocker(blocker);
2600	<	}
2601	<	else {
2602	<	do {} while (!blocker.isReleasable() && !blocker.block());
2597	>	ForkJoinPool p = ((t instanceof ForkJoinWorkerThread) ?
2598	>	((ForkJoinWorkerThread)t).pool : null);
2599	>	while (!blocker.isReleasable()) {
2600	>	if (p == null || p.tryCompensate()) {
2601	>	try {
2602	>	do {} while (!blocker.isReleasable() && !blocker.block());
2603	>	} finally {
2604	>	if (p != null)
2605	>	p.incrementActiveCount();
2606	>	}
2607	>	break;
2608	>	}
2609		}
2610		}
2611
#	Line 1929 | Line 2622 | public class ForkJoinPool extends Abstra
2622		}
2623
2624		// Unsafe mechanics
2625	<
2626	<	private static final sun.misc.Unsafe UNSAFE = getUnsafe();
2627	<	private static final long workerCountsOffset =
2628	<	objectFieldOffset("workerCounts", ForkJoinPool.class);
2629	<	private static final long runStateOffset =
2630	<	objectFieldOffset("runState", ForkJoinPool.class);
2631	<	private static final long eventCountOffset =
2632	<	objectFieldOffset("eventCount", ForkJoinPool.class);
2633	<	private static final long eventWaitersOffset =
2634	<	objectFieldOffset("eventWaiters",ForkJoinPool.class);
2635	<	private static final long stealCountOffset =
1943	<	objectFieldOffset("stealCount",ForkJoinPool.class);
1944	<	private static final long spareWaitersOffset =
1945	<	objectFieldOffset("spareWaiters",ForkJoinPool.class);
1946	<
1947	<	private static long objectFieldOffset(String field, Class<?> klazz) {
2625	>	private static final sun.misc.Unsafe U;
2626	>	private static final long CTL;
2627	>	private static final long RUNSTATE;
2628	>	private static final long PARKBLOCKER;
2629	>
2630	>	static {
2631	>	poolNumberGenerator = new AtomicInteger();
2632	>	modifyThreadPermission = new RuntimePermission("modifyThread");
2633	>	defaultForkJoinWorkerThreadFactory =
2634	>	new DefaultForkJoinWorkerThreadFactory();
2635	>	int s;
2636		try {
2637	<	return UNSAFE.objectFieldOffset(klazz.getDeclaredField(field));
2638	<	} catch (NoSuchFieldException e) {
2639	<	// Convert Exception to corresponding Error
2640	<	NoSuchFieldError error = new NoSuchFieldError(field);
2641	<	error.initCause(e);
2642	<	throw error;
2637	>	U = getUnsafe();
2638	>	Class<?> k = ForkJoinPool.class;
2639	>	Class<?> tk = Thread.class;
2640	>	CTL = U.objectFieldOffset
2641	>	(k.getDeclaredField("ctl"));
2642	>	RUNSTATE = U.objectFieldOffset
2643	>	(k.getDeclaredField("runState"));
2644	>	PARKBLOCKER = U.objectFieldOffset
2645	>	(tk.getDeclaredField("parkBlocker"));
2646	>	} catch (Exception e) {
2647	>	throw new Error(e);
2648		}
2649		}
2650
#	Line 1982 | Line 2675 | public class ForkJoinPool extends Abstra
2675		}
2676		}
2677		}
2678	+
2679		}

Diff Legend

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