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 1.72 by jsr166, Tue Sep 7 06:19:05 2010 UTC vs.
Revision 1.124 by jsr166, Mon Feb 20 23:32:24 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.*;
8		import java.util.ArrayList;
9		import java.util.Arrays;
10		import java.util.Collection;
11		import java.util.Collections;
12		import java.util.List;
13	<	import java.util.concurrent.locks.LockSupport;
14	<	import java.util.concurrent.locks.ReentrantLock;
13	>	import java.util.Random;
14	>	import java.util.concurrent.AbstractExecutorService;
15	>	import java.util.concurrent.Callable;
16	>	import java.util.concurrent.ExecutorService;
17	>	import java.util.concurrent.Future;
18	>	import java.util.concurrent.RejectedExecutionException;
19	>	import java.util.concurrent.RunnableFuture;
20	>	import java.util.concurrent.TimeUnit;
21		import java.util.concurrent.atomic.AtomicInteger;
22	<	import java.util.concurrent.CountDownLatch;
22	>	import java.util.concurrent.atomic.AtomicLong;
23	>	import java.util.concurrent.locks.AbstractQueuedSynchronizer;
24	>	import java.util.concurrent.locks.Condition;
25
26		/**
27		* An {@link ExecutorService} for running {@link ForkJoinTask}s.
#	Line 26 | Line 32 | import java.util.concurrent.CountDownLat
32		* <p>A {@code ForkJoinPool} differs from other kinds of {@link
33		* ExecutorService} mainly by virtue of employing
34		* <em>work-stealing</em>: all threads in the pool attempt to find and
35	<	* execute subtasks created by other active tasks (eventually blocking
36	<	* waiting for work if none exist). This enables efficient processing
37	<	* when most tasks spawn other subtasks (as do most {@code
38	<	* ForkJoinTask}s). When setting <em>asyncMode</em> to true in
39	<	* constructors, {@code ForkJoinPool}s may also be appropriate for use
40	<	* with event-style tasks that are never joined.
35	>	* execute tasks submitted to the pool and/or created by other active
36	>	* tasks (eventually blocking waiting for work if none exist). This
37	>	* enables efficient processing when most tasks spawn other subtasks
38	>	* (as do most {@code ForkJoinTask}s), as well as when many small
39	>	* tasks are submitted to the pool from external clients.  Especially
40	>	* when setting <em>asyncMode</em> to true in constructors, {@code
41	>	* ForkJoinPool}s may also be appropriate for use with event-style
42	>	* tasks that are never joined.
43		*
44		* <p>A {@code ForkJoinPool} is constructed with a given target
45		* parallelism level; by default, equal to the number of available
#	Line 51 | Line 59 | import java.util.concurrent.CountDownLat
59		* convenient form for informal monitoring.
60		*
61		* <p> As is the case with other ExecutorServices, there are three
62	<	* main task execution methods summarized in the following
63	<	* table. These are designed to be used by clients not already engaged
64	<	* in fork/join computations in the current pool.  The main forms of
65	<	* these methods accept instances of {@code ForkJoinTask}, but
66	<	* overloaded forms also allow mixed execution of plain {@code
62	>	* main task execution methods summarized in the following table.
63	>	* These are designed to be used primarily by clients not already
64	>	* engaged in fork/join computations in the current pool.  The main
65	>	* forms of these methods accept instances of {@code ForkJoinTask},
66	>	* but overloaded forms also allow mixed execution of plain {@code
67		* Runnable}- or {@code Callable}- based activities as well.  However,
68	<	* tasks that are already executing in a pool should normally
69	<	* <em>NOT</em> use these pool execution methods, but instead use the
70	<	* within-computation forms listed in the table.
68	>	* tasks that are already executing in a pool should normally instead
69	>	* use the within-computation forms listed in the table unless using
70	>	* async event-style tasks that are not usually joined, in which case
71	>	* there is little difference among choice of methods.
72		*
73		* <table BORDER CELLPADDING=3 CELLSPACING=1>
74		*  <tr>
#	Line 94 | Line 103 | import java.util.concurrent.CountDownLat
103		* daemon} mode, there is typically no need to explicitly {@link
104		* #shutdown} such a pool upon program exit.
105		*
106	<	* <pre>
106	>	*  <pre> {@code
107		* static final ForkJoinPool mainPool = new ForkJoinPool();
108		* ...
109		* public void sort(long[] array) {
110		*   mainPool.invoke(new SortTask(array, 0, array.length));
111	<	* }
103	<	* </pre>
111	>	* }}</pre>
112		*
113		* <p><b>Implementation notes</b>: This implementation restricts the
114		* maximum number of running threads to 32767. Attempts to create
#	Line 119 | Line 127 | public class ForkJoinPool extends Abstra
127		/*
128		* Implementation Overview
129		*
130	<	* This class provides the central bookkeeping and control for a
131	<	* set of worker threads: Submissions from non-FJ threads enter
132	<	* into a submission queue. Workers take these tasks and typically
133	<	* split them into subtasks that may be stolen by other workers.
134	<	* The main work-stealing mechanics implemented in class
135	<	* ForkJoinWorkerThread give first priority to processing tasks
136	<	* from their own queues (LIFO or FIFO, depending on mode), then
137	<	* to randomized FIFO steals of tasks in other worker queues, and
138	<	* lastly to new submissions. These mechanics do not consider
139	<	* affinities, loads, cache localities, etc, so rarely provide the
140	<	* best possible performance on a given machine, but portably
141	<	* provide good throughput by averaging over these factors.
142	<	* (Further, even if we did try to use such information, we do not
143	<	* usually have a basis for exploiting it. For example, some sets
144	<	* of tasks profit from cache affinities, but others are harmed by
145	<	* cache pollution effects.)
146	<	*
147	<	* Beyond work-stealing support and essential bookkeeping, the
148	<	* main responsibility of this framework is to take actions when
149	<	* one worker is waiting to join a task stolen (or always held by)
150	<	* another.  Because we are multiplexing many tasks on to a pool
151	<	* of workers, we can't just let them block (as in Thread.join).
152	<	* We also cannot just reassign the joiner's run-time stack with
153	<	* another and replace it later, which would be a form of
154	<	* "continuation", that even if possible is not necessarily a good
155	<	* idea. Given that the creation costs of most threads on most
156	<	* systems mainly surrounds setting up runtime stacks, thread
157	<	* creation and switching is usually not much more expensive than
158	<	* stack creation and switching, and is more flexible). Instead we
159	<	* combine two tactics:
130	>	* This class and its nested classes provide the main
131	>	* functionality and control for a set of worker threads:
132	>	* Submissions from non-FJ threads enter into submission queues.
133	>	* Workers take these tasks and typically split them into subtasks
134	>	* that may be stolen by other workers.  Preference rules give
135	>	* first priority to processing tasks from their own queues (LIFO
136	>	* or FIFO, depending on mode), then to randomized FIFO steals of
137	>	* tasks in other queues.
138	>	*
139	>	* WorkQueues
140	>	* ==========
141	>	*
142	>	* Most operations occur within work-stealing queues (in nested
143	>	* class WorkQueue).  These are special forms of Deques that
144	>	* support only three of the four possible end-operations -- push,
145	>	* pop, and poll (aka steal), under the further constraints that
146	>	* push and pop are called only from the owning thread (or, as
147	>	* extended here, under a lock), while poll may be called from
148	>	* other threads.  (If you are unfamiliar with them, you probably
149	>	* want to read Herlihy and Shavit's book "The Art of
150	>	* Multiprocessor programming", chapter 16 describing these in
151	>	* more detail before proceeding.)  The main work-stealing queue
152	>	* design is roughly similar to those in the papers "Dynamic
153	>	* Circular Work-Stealing Deque" by Chase and Lev, SPAA 2005
154	>	* (http://research.sun.com/scalable/pubs/index.html) and
155	>	* "Idempotent work stealing" by Michael, Saraswat, and Vechev,
156	>	* PPoPP 2009 (http://portal.acm.org/citation.cfm?id=1504186).
157	>	* The main differences ultimately stem from GC requirements that
158	>	* we null out taken slots as soon as we can, to maintain as small
159	>	* a footprint as possible even in programs generating huge
160	>	* numbers of tasks. To accomplish this, we shift the CAS
161	>	* arbitrating pop vs poll (steal) from being on the indices
162	>	* ("base" and "top") to the slots themselves.  So, both a
163	>	* successful pop and poll mainly entail a CAS of a slot from
164	>	* non-null to null.  Because we rely on CASes of references, we
165	>	* do not need tag bits on base or top.  They are simple ints as
166	>	* used in any circular array-based queue (see for example
167	>	* ArrayDeque).  Updates to the indices must still be ordered in a
168	>	* way that guarantees that top == base means the queue is empty,
169	>	* but otherwise may err on the side of possibly making the queue
170	>	* appear nonempty when a push, pop, or poll have not fully
171	>	* committed. Note that this means that the poll operation,
172	>	* considered individually, is not wait-free. One thief cannot
173	>	* successfully continue until another in-progress one (or, if
174	>	* previously empty, a push) completes.  However, in the
175	>	* aggregate, we ensure at least probabilistic non-blockingness.
176	>	* If an attempted steal fails, a thief always chooses a different
177	>	* random victim target to try next. So, in order for one thief to
178	>	* progress, it suffices for any in-progress poll or new push on
179	>	* any empty queue to complete. (This is why we normally use
180	>	* method pollAt and its variants that try once at the apparent
181	>	* base index, else consider alternative actions, rather than
182	>	* method poll.)
183	>	*
184	>	* This approach also enables support of a user mode in which local
185	>	* task processing is in FIFO, not LIFO order, simply by using
186	>	* poll rather than pop.  This can be useful in message-passing
187	>	* frameworks in which tasks are never joined.  However neither
188	>	* mode considers affinities, loads, cache localities, etc, so
189	>	* rarely provide the best possible performance on a given
190	>	* machine, but portably provide good throughput by averaging over
191	>	* these factors.  (Further, even if we did try to use such
192	>	* information, we do not usually have a basis for exploiting it.
193	>	* For example, some sets of tasks profit from cache affinities,
194	>	* but others are harmed by cache pollution effects.)
195	>	*
196	>	* WorkQueues are also used in a similar way for tasks submitted
197	>	* to the pool. We cannot mix these tasks in the same queues used
198	>	* for work-stealing (this would contaminate lifo/fifo
199	>	* processing). Instead, we loosely associate submission queues
200	>	* with submitting threads, using a form of hashing.  The
201	>	* ThreadLocal Submitter class contains a value initially used as
202	>	* a hash code for choosing existing queues, but may be randomly
203	>	* repositioned upon contention with other submitters.  In
204	>	* essence, submitters act like workers except that they never
205	>	* take tasks, and they are multiplexed on to a finite number of
206	>	* shared work queues. However, classes are set up so that future
207	>	* extensions could allow submitters to optionally help perform
208	>	* tasks as well. Insertion of tasks in shared mode requires a
209	>	* lock (mainly to protect in the case of resizing) but we use
210	>	* only a simple spinlock (using bits in field runState), because
211	>	* submitters encountering a busy queue move on to try or create
212	>	* other queues -- they block only when creating and registering
213	>	* new queues.
214	>	*
215	>	* Management
216	>	* ==========
217	>	*
218	>	* The main throughput advantages of work-stealing stem from
219	>	* decentralized control -- workers mostly take tasks from
220	>	* themselves or each other. We cannot negate this in the
221	>	* implementation of other management responsibilities. The main
222	>	* tactic for avoiding bottlenecks is packing nearly all
223	>	* essentially atomic control state into two volatile variables
224	>	* that are by far most often read (not written) as status and
225	>	* consistency checks.
226	>	*
227	>	* Field "ctl" contains 64 bits holding all the information needed
228	>	* to atomically decide to add, inactivate, enqueue (on an event
229	>	* queue), dequeue, and/or re-activate workers.  To enable this
230	>	* packing, we restrict maximum parallelism to (1<<15)-1 (which is
231	>	* far in excess of normal operating range) to allow ids, counts,
232	>	* and their negations (used for thresholding) to fit into 16bit
233	>	* fields.
234	>	*
235	>	* Field "runState" contains 32 bits needed to register and
236	>	* deregister WorkQueues, as well as to enable shutdown. It is
237	>	* only modified under a lock (normally briefly held, but
238	>	* occasionally protecting allocations and resizings) but even
239	>	* when locked remains available to check consistency.
240	>	*
241	>	* Recording WorkQueues.  WorkQueues are recorded in the
242	>	* "workQueues" array that is created upon pool construction and
243	>	* expanded if necessary.  Updates to the array while recording
244	>	* new workers and unrecording terminated ones are protected from
245	>	* each other by a lock but the array is otherwise concurrently
246	>	* readable, and accessed directly.  To simplify index-based
247	>	* operations, the array size is always a power of two, and all
248	>	* readers must tolerate null slots. Shared (submission) queues
249	>	* are at even indices, worker queues at odd indices. Grouping
250	>	* them together in this way simplifies and speeds up task
251	>	* scanning.
252	>	*
253	>	* All worker thread creation is on-demand, triggered by task
254	>	* submissions, replacement of terminated workers, and/or
255	>	* compensation for blocked workers. However, all other support
256	>	* code is set up to work with other policies.  To ensure that we
257	>	* do not hold on to worker references that would prevent GC, ALL
258	>	* accesses to workQueues are via indices into the workQueues
259	>	* array (which is one source of some of the messy code
260	>	* constructions here). In essence, the workQueues array serves as
261	>	* a weak reference mechanism. Thus for example the wait queue
262	>	* field of ctl stores indices, not references.  Access to the
263	>	* workQueues in associated methods (for example signalWork) must
264	>	* both index-check and null-check the IDs. All such accesses
265	>	* ignore bad IDs by returning out early from what they are doing,
266	>	* since this can only be associated with termination, in which
267	>	* case it is OK to give up.  All uses of the workQueues array
268	>	* also check that it is non-null (even if previously
269	>	* non-null). This allows nulling during termination, which is
270	>	* currently not necessary, but remains an option for
271	>	* resource-revocation-based shutdown schemes. It also helps
272	>	* reduce JIT issuance of uncommon-trap code, which tends to
273	>	* unnecessarily complicate control flow in some methods.
274	>	*
275	>	* Event Queuing. Unlike HPC work-stealing frameworks, we cannot
276	>	* let workers spin indefinitely scanning for tasks when none can
277	>	* be found immediately, and we cannot start/resume workers unless
278	>	* there appear to be tasks available.  On the other hand, we must
279	>	* quickly prod them into action when new tasks are submitted or
280	>	* generated. In many usages, ramp-up time to activate workers is
281	>	* the main limiting factor in overall performance (this is
282	>	* compounded at program start-up by JIT compilation and
283	>	* allocation). So we try to streamline this as much as possible.
284	>	* We park/unpark workers after placing in an event wait queue
285	>	* when they cannot find work. This "queue" is actually a simple
286	>	* Treiber stack, headed by the "id" field of ctl, plus a 15bit
287	>	* counter value (that reflects the number of times a worker has
288	>	* been inactivated) to avoid ABA effects (we need only as many
289	>	* version numbers as worker threads). Successors are held in
290	>	* field WorkQueue.nextWait.  Queuing deals with several intrinsic
291	>	* races, mainly that a task-producing thread can miss seeing (and
292	>	* signalling) another thread that gave up looking for work but
293	>	* has not yet entered the wait queue. We solve this by requiring
294	>	* a full sweep of all workers (via repeated calls to method
295	>	* scan()) both before and after a newly waiting worker is added
296	>	* to the wait queue. During a rescan, the worker might release
297	>	* some other queued worker rather than itself, which has the same
298	>	* net effect. Because enqueued workers may actually be rescanning
299	>	* rather than waiting, we set and clear the "parker" field of
300	>	* WorkQueues to reduce unnecessary calls to unpark.  (This
301	>	* requires a secondary recheck to avoid missed signals.)  Note
302	>	* the unusual conventions about Thread.interrupts surrounding
303	>	* parking and other blocking: Because interrupts are used solely
304	>	* to alert threads to check termination, which is checked anyway
305	>	* upon blocking, we clear status (using Thread.interrupted)
306	>	* before any call to park, so that park does not immediately
307	>	* return due to status being set via some other unrelated call to
308	>	* interrupt in user code.
309	>	*
310	>	* Signalling.  We create or wake up workers only when there
311	>	* appears to be at least one task they might be able to find and
312	>	* execute.  When a submission is added or another worker adds a
313	>	* task to a queue that previously had fewer than two tasks, they
314	>	* signal waiting workers (or trigger creation of new ones if
315	>	* fewer than the given parallelism level -- see signalWork).
316	>	* These primary signals are buttressed by signals during rescans;
317	>	* together these cover the signals needed in cases when more
318	>	* tasks are pushed but untaken, and improve performance compared
319	>	* to having one thread wake up all workers.
320	>	*
321	>	* Trimming workers. To release resources after periods of lack of
322	>	* use, a worker starting to wait when the pool is quiescent will
323	>	* time out and terminate if the pool has remained quiescent for
324	>	* SHRINK_RATE nanosecs. This will slowly propagate, eventually
325	>	* terminating all workers after long periods of non-use.
326	>	*
327	>	* Shutdown and Termination. A call to shutdownNow atomically sets
328	>	* a runState bit and then (non-atomically) sets each worker's
329	>	* runState status, cancels all unprocessed tasks, and wakes up
330	>	* all waiting workers.  Detecting whether termination should
331	>	* commence after a non-abrupt shutdown() call requires more work
332	>	* and bookkeeping. We need consensus about quiescence (i.e., that
333	>	* there is no more work). The active count provides a primary
334	>	* indication but non-abrupt shutdown still requires a rechecking
335	>	* scan for any workers that are inactive but not queued.
336	>	*
337	>	* Joining Tasks
338	>	* =============
339	>	*
340	>	* Any of several actions may be taken when one worker is waiting
341	>	* to join a task stolen (or always held) by another.  Because we
342	>	* are multiplexing many tasks on to a pool of workers, we can't
343	>	* just let them block (as in Thread.join).  We also cannot just
344	>	* reassign the joiner's run-time stack with another and replace
345	>	* it later, which would be a form of "continuation", that even if
346	>	* possible is not necessarily a good idea since we sometimes need
347	>	* both an unblocked task and its continuation to progress.
348	>	* Instead we combine two tactics:
349		*
350		*   Helping: Arranging for the joiner to execute some task that it
351	<	*      would be running if the steal had not occurred.  Method
155	<	*      ForkJoinWorkerThread.helpJoinTask tracks joining->stealing
156	<	*      links to try to find such a task.
351	>	*      would be running if the steal had not occurred.
352		*
353		*   Compensating: Unless there are already enough live threads,
354	<	*      method helpMaintainParallelism() may create or
355	<	*      re-activate a spare thread to compensate for blocked
161	<	*      joiners until they unblock.
354	>	*      method tryCompensate() may create or re-activate a spare
355	>	*      thread to compensate for blocked joiners until they unblock.
356		*
357	<	* It is impossible to keep exactly the target (parallelism)
358	<	* number of threads running at any given time.  Determining
359	<	* existence of conservatively safe helping targets, the
360	<	* availability of already-created spares, and the apparent need
361	<	* to create new spares are all racy and require heuristic
362	<	* guidance, so we rely on multiple retries of each.  Compensation
363	<	* occurs in slow-motion. It is triggered only upon timeouts of
170	<	* Object.wait used for joins. This reduces poor decisions that
171	<	* would otherwise be made when threads are waiting for others
172	<	* that are stalled because of unrelated activities such as
173	<	* garbage collection.
357	>	* A third form (implemented in tryRemoveAndExec and
358	>	* tryPollForAndExec) amounts to helping a hypothetical
359	>	* compensator: If we can readily tell that a possible action of a
360	>	* compensator is to steal and execute the task being joined, the
361	>	* joining thread can do so directly, without the need for a
362	>	* compensation thread (although at the expense of larger run-time
363	>	* stacks, but the tradeoff is typically worthwhile).
364		*
365		* The ManagedBlocker extension API can't use helping so relies
366		* only on compensation in method awaitBlocker.
367		*
368	<	* The main throughput advantages of work-stealing stem from
369	<	* decentralized control -- workers mostly steal tasks from each
370	<	* other. We do not want to negate this by creating bottlenecks
371	<	* implementing other management responsibilities. So we use a
372	<	* collection of techniques that avoid, reduce, or cope well with
373	<	* contention. These entail several instances of bit-packing into
374	<	* CASable fields to maintain only the minimally required
375	<	* atomicity. To enable such packing, we restrict maximum
376	<	* parallelism to (1<<15)-1 (enabling twice this (to accommodate
377	<	* unbalanced increments and decrements) to fit into a 16 bit
378	<	* field, which is far in excess of normal operating range.  Even
379	<	* though updates to some of these bookkeeping fields do sometimes
380	<	* contend with each other, they don't normally cache-contend with
381	<	* updates to others enough to warrant memory padding or
382	<	* isolation. So they are all held as fields of ForkJoinPool
383	<	* objects.  The main capabilities are as follows:
384	<	*
385	<	* 1. Creating and removing workers. Workers are recorded in the
386	<	* "workers" array. This is an array as opposed to some other data
387	<	* structure to support index-based random steals by workers.
388	<	* Updates to the array recording new workers and unrecording
389	<	* terminated ones are protected from each other by a lock
390	<	* (workerLock) but the array is otherwise concurrently readable,
391	<	* and accessed directly by workers. To simplify index-based
392	<	* operations, the array size is always a power of two, and all
393	<	* readers must tolerate null slots. Currently, all worker thread
394	<	* creation is on-demand, triggered by task submissions,
395	<	* replacement of terminated workers, and/or compensation for
396	<	* blocked workers. However, all other support code is set up to
207	<	* work with other policies.
208	<	*
209	<	* To ensure that we do not hold on to worker references that
210	<	* would prevent GC, ALL accesses to workers are via indices into
211	<	* the workers array (which is one source of some of the unusual
212	<	* code constructions here). In essence, the workers array serves
213	<	* as a WeakReference mechanism. Thus for example the event queue
214	<	* stores worker indices, not worker references. Access to the
215	<	* workers in associated methods (for example releaseEventWaiters)
216	<	* must both index-check and null-check the IDs. All such accesses
217	<	* ignore bad IDs by returning out early from what they are doing,
218	<	* since this can only be associated with shutdown, in which case
219	<	* it is OK to give up. On termination, we just clobber these
220	<	* data structures without trying to use them.
221	<	*
222	<	* 2. Bookkeeping for dynamically adding and removing workers. We
223	<	* aim to approximately maintain the given level of parallelism.
224	<	* When some workers are known to be blocked (on joins or via
225	<	* ManagedBlocker), we may create or resume others to take their
226	<	* place until they unblock (see below). Implementing this
227	<	* requires counts of the number of "running" threads (i.e., those
228	<	* that are neither blocked nor artificially suspended) as well as
229	<	* the total number.  These two values are packed into one field,
230	<	* "workerCounts" because we need accurate snapshots when deciding
231	<	* to create, resume or suspend.  Note however that the
232	<	* correspondence of these counts to reality is not guaranteed. In
233	<	* particular updates for unblocked threads may lag until they
234	<	* actually wake up.
235	<	*
236	<	* 3. Maintaining global run state. The run state of the pool
237	<	* consists of a runLevel (SHUTDOWN, TERMINATING, etc) similar to
238	<	* those in other Executor implementations, as well as a count of
239	<	* "active" workers -- those that are, or soon will be, or
240	<	* recently were executing tasks. The runLevel and active count
241	<	* are packed together in order to correctly trigger shutdown and
242	<	* termination. Without care, active counts can be subject to very
243	<	* high contention.  We substantially reduce this contention by
244	<	* relaxing update rules.  A worker must claim active status
245	<	* prospectively, by activating if it sees that a submitted or
246	<	* stealable task exists (it may find after activating that the
247	<	* task no longer exists). It stays active while processing this
248	<	* task (if it exists) and any other local subtasks it produces,
249	<	* until it cannot find any other tasks. It then tries
250	<	* inactivating (see method preStep), but upon update contention
251	<	* instead scans for more tasks, later retrying inactivation if it
252	<	* doesn't find any.
253	<	*
254	<	* 4. Managing idle workers waiting for tasks. We cannot let
255	<	* workers spin indefinitely scanning for tasks when none are
256	<	* available. On the other hand, we must quickly prod them into
257	<	* action when new tasks are submitted or generated.  We
258	<	* park/unpark these idle workers using an event-count scheme.
259	<	* Field eventCount is incremented upon events that may enable
260	<	* workers that previously could not find a task to now find one:
261	<	* Submission of a new task to the pool, or another worker pushing
262	<	* a task onto a previously empty queue.  (We also use this
263	<	* mechanism for configuration and termination actions that
264	<	* require wakeups of idle workers).  Each worker maintains its
265	<	* last known event count, and blocks when a scan for work did not
266	<	* find a task AND its lastEventCount matches the current
267	<	* eventCount. Waiting idle workers are recorded in a variant of
268	<	* Treiber stack headed by field eventWaiters which, when nonzero,
269	<	* encodes the thread index and count awaited for by the worker
270	<	* thread most recently calling eventSync. This thread in turn has
271	<	* a record (field nextEventWaiter) for the next waiting worker.
272	<	* In addition to allowing simpler decisions about need for
273	<	* wakeup, the event count bits in eventWaiters serve the role of
274	<	* tags to avoid ABA errors in Treiber stacks. Upon any wakeup,
275	<	* released threads also try to release at most two others.  The
276	<	* net effect is a tree-like diffusion of signals, where released
277	<	* threads (and possibly others) help with unparks.  To further
278	<	* reduce contention effects a bit, failed CASes to increment
279	<	* field 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 notices
284	<	* (usually upon timeout of a wait()) that there are too few
285	<	* running threads, we may create a new thread to maintain
286	<	* parallelism level, or at least avoid starvation. Usually, extra
287	<	* threads are needed for only very short periods, yet join
288	<	* dependencies are such that we sometimes need them in
289	<	* bursts. Rather than create new threads each time this happens,
290	<	* we suspend no-longer-needed extra ones as "spares". For most
291	<	* purposes, we don't distinguish "extra" spare threads from
292	<	* normal "core" threads: On each call to preStep (the only point
293	<	* at which we can do this) a worker checks to see if there are
294	<	* now too many running workers, and if so, suspends itself.
295	<	* Method helpMaintainParallelism looks for suspended threads to
296	<	* resume before considering creating a new replacement. The
297	<	* spares themselves are encoded on another variant of a Treiber
298	<	* Stack, headed at field "spareWaiters".  Note that the use of
299	<	* spares is intrinsically racy.  One thread may become a spare at
300	<	* about the same time as another is needlessly being created. We
301	<	* counteract this and related slop in part by requiring resumed
302	<	* spares to immediately recheck (in preStep) to see whether they
303	<	* should re-suspend.
304	<	*
305	<	* 6. Killing off unneeded workers. A timeout mechanism is used to
306	<	* shed unused workers: The oldest (first) event queue waiter uses
307	<	* a timed rather than hard wait. When this wait times out without
308	<	* a normal wakeup, it tries to shutdown any one (for convenience
309	<	* the newest) other spare or event waiter via
310	<	* tryShutdownUnusedWorker. This eventually reduces the number of
311	<	* worker threads to a minimum of one after a long enough period
312	<	* 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 impossible
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
321	<	* presence of transient 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 parallelism to
326	<	* lapse for a while during joins, and use a timeout to see if,
327	<	* after the resulting settling, there is still a need for
328	<	* additional workers.  This also better copes with the fact that
329	<	* some of the methods in this class tend to never become compiled
330	<	* (but are interpreted), so some components of the entire set of
331	<	* controls might execute 100 times faster than others. And
332	<	* similarly for cases where the apparent lack of work is just due
333	<	* to GC stalls and other transient system activity.
368	>	* The algorithm in tryHelpStealer entails a form of "linear"
369	>	* helping: Each worker records (in field currentSteal) the most
370	>	* recent task it stole from some other worker. Plus, it records
371	>	* (in field currentJoin) the task it is currently actively
372	>	* joining. Method tryHelpStealer uses these markers to try to
373	>	* find a worker to help (i.e., steal back a task from and execute
374	>	* it) that could hasten completion of the actively joined task.
375	>	* In essence, the joiner executes a task that would be on its own
376	>	* local deque had the to-be-joined task not been stolen. This may
377	>	* be seen as a conservative variant of the approach in Wagner &
378	>	* Calder "Leapfrogging: a portable technique for implementing
379	>	* efficient futures" SIGPLAN Notices, 1993
380	>	* (http://portal.acm.org/citation.cfm?id=155354). It differs in
381	>	* that: (1) We only maintain dependency links across workers upon
382	>	* steals, rather than use per-task bookkeeping.  This sometimes
383	>	* requires a linear scan of workQueues array to locate stealers,
384	>	* but often doesn't because stealers leave hints (that may become
385	>	* stale/wrong) of where to locate them.  A stealHint is only a
386	>	* hint because a worker might have had multiple steals and the
387	>	* hint records only one of them (usually the most current).
388	>	* Hinting isolates cost to when it is needed, rather than adding
389	>	* to per-task overhead.  (2) It is "shallow", ignoring nesting
390	>	* and potentially cyclic mutual steals.  (3) It is intentionally
391	>	* racy: field currentJoin is updated only while actively joining,
392	>	* which means that we miss links in the chain during long-lived
393	>	* tasks, GC stalls etc (which is OK since blocking in such cases
394	>	* is usually a good idea).  (4) We bound the number of attempts
395	>	* to find work (see MAX_HELP) and fall back to suspending the
396	>	* worker and if necessary replacing it with another.
397		*
398	<	* Beware that there is a lot of representation-level coupling
398	>	* It is impossible to keep exactly the target parallelism number
399	>	* of threads running at any given time.  Determining the
400	>	* existence of conservatively safe helping targets, the
401	>	* availability of already-created spares, and the apparent need
402	>	* to create new spares are all racy, so we rely on multiple
403	>	* retries of each.  Compensation in the apparent absence of
404	>	* helping opportunities is challenging to control on JVMs, where
405	>	* GC and other activities can stall progress of tasks that in
406	>	* turn stall out many other dependent tasks, without us being
407	>	* able to determine whether they will ever require compensation.
408	>	* Even though work-stealing otherwise encounters little
409	>	* degradation in the presence of more threads than cores,
410	>	* aggressively adding new threads in such cases entails risk of
411	>	* unwanted positive feedback control loops in which more threads
412	>	* cause more dependent stalls (as well as delayed progress of
413	>	* unblocked threads to the point that we know they are available)
414	>	* leading to more situations requiring more threads, and so
415	>	* on. This aspect of control can be seen as an (analytically
416	>	* intractible) game with an opponent that may choose the worst
417	>	* (for us) active thread to stall at any time.  We take several
418	>	* precautions to bound losses (and thus bound gains), mainly in
419	>	* methods tryCompensate and awaitJoin: (1) We only try
420	>	* compensation after attempting enough helping steps (measured
421	>	* via counting and timing) that we have already consumed the
422	>	* estimated cost of creating and activating a new thread.  (2) We
423	>	* allow up to 50% of threads to be blocked before initially
424	>	* adding any others, and unless completely saturated, check that
425	>	* some work is available for a new worker before adding. Also, we
426	>	* create up to only 50% more threads until entering a mode that
427	>	* only adds a thread if all others are possibly blocked.  All
428	>	* together, this means that we might be half as fast to react,
429	>	* and create half as many threads as possible in the ideal case,
430	>	* but present vastly fewer anomalies in all other cases compared
431	>	* to both more aggressive and more conservative alternatives.
432	>	*
433	>	* Style notes: There is a lot of representation-level coupling
434		* among classes ForkJoinPool, ForkJoinWorkerThread, and
435	<	* ForkJoinTask.  For example, direct access to "workers" array by
436	<	* workers, and direct access to ForkJoinTask.status by both
437	<	* ForkJoinPool and ForkJoinWorkerThread.  There is little point
438	<	* trying to reduce this, since any associated future changes in
439	<	* representations will need to be accompanied by algorithmic
440	<	* changes anyway.
441	<	*
442	<	* Style notes: There are lots of inline assignments (of form
443	<	* "while ((local = field) != 0)") which are usually the simplest
444	<	* way to ensure the required read orderings (which are sometimes
445	<	* critical). Also several occurrences of the unusual "do {}
446	<	* while (!cas...)" which is the simplest way to force an update of
447	<	* a CAS'ed variable. There are also other coding oddities that
448	<	* help some methods perform reasonably even when interpreted (not
449	<	* compiled), at the expense of some messy constructions that
450	<	* reduce byte code counts.
451	<	*
452	<	* The order of declarations in this file is: (1) statics (2)
453	<	* fields (along with constants used when unpacking some of them)
454	<	* (3) internal control methods (4) callbacks and other support
455	<	* for ForkJoinTask and ForkJoinWorkerThread classes, (5) exported
456	<	* methods (plus a few little helpers).
435	>	* ForkJoinTask.  The fields of WorkQueue maintain data structures
436	>	* managed by ForkJoinPool, so are directly accessed.  There is
437	>	* little point trying to reduce this, since any associated future
438	>	* changes in representations will need to be accompanied by
439	>	* algorithmic changes anyway. Several methods intrinsically
440	>	* sprawl because they must accumulate sets of consistent reads of
441	>	* volatiles held in local variables.  Methods signalWork() and
442	>	* scan() are the main bottlenecks, so are especially heavily
443	>	* micro-optimized/mangled.  There are lots of inline assignments
444	>	* (of form "while ((local = field) != 0)") which are usually the
445	>	* simplest way to ensure the required read orderings (which are
446	>	* sometimes critical). This leads to a "C"-like style of listing
447	>	* declarations of these locals at the heads of methods or blocks.
448	>	* There are several occurrences of the unusual "do {} while
449	>	* (!cas...)"  which is the simplest way to force an update of a
450	>	* CAS'ed variable. There are also other coding oddities that help
451	>	* some methods perform reasonably even when interpreted (not
452	>	* compiled).
453	>	*
454	>	* The order of declarations in this file is:
455	>	* (1) Static utility functions
456	>	* (2) Nested (static) classes
457	>	* (3) Static fields
458	>	* (4) Fields, along with constants used when unpacking some of them
459	>	* (5) Internal control methods
460	>	* (6) Callbacks and other support for ForkJoinTask methods
461	>	* (7) Exported methods
462	>	* (8) Static block initializing statics in minimally dependent order
463		*/
464
465	+	// Static utilities
466	+
467	+	/**
468	+	* If there is a security manager, makes sure caller has
469	+	* permission to modify threads.
470	+	*/
471	+	private static void checkPermission() {
472	+	SecurityManager security = System.getSecurityManager();
473	+	if (security != null)
474	+	security.checkPermission(modifyThreadPermission);
475	+	}
476	+
477	+	// Nested classes
478	+
479		/**
480		* Factory for creating new {@link ForkJoinWorkerThread}s.
481		* A {@code ForkJoinWorkerThreadFactory} must be defined and used
#	Line 386 | Line 504 | public class ForkJoinPool extends Abstra
504		}
505
506		/**
507	<	* Creates a new ForkJoinWorkerThread. This factory is used unless
508	<	* overridden in ForkJoinPool constructors.
507	>	* A simple non-reentrant lock used for exclusion when managing
508	>	* queues and workers. We use a custom lock so that we can readily
509	>	* probe lock state in constructions that check among alternative
510	>	* actions. The lock is normally only very briefly held, and
511	>	* sometimes treated as a spinlock, but other usages block to
512	>	* reduce overall contention in those cases where locked code
513	>	* bodies perform allocation/resizing.
514	>	*/
515	>	static final class Mutex extends AbstractQueuedSynchronizer {
516	>	public final boolean tryAcquire(int ignore) {
517	>	return compareAndSetState(0, 1);
518	>	}
519	>	public final boolean tryRelease(int ignore) {
520	>	setState(0);
521	>	return true;
522	>	}
523	>	public final void lock() { acquire(0); }
524	>	public final void unlock() { release(0); }
525	>	public final boolean isHeldExclusively() { return getState() == 1; }
526	>	public final Condition newCondition() { return new ConditionObject(); }
527	>	}
528	>
529	>	/**
530	>	* Class for artificial tasks that are used to replace the target
531	>	* of local joins if they are removed from an interior queue slot
532	>	* in WorkQueue.tryRemoveAndExec. We don't need the proxy to
533	>	* actually do anything beyond having a unique identity.
534	>	*/
535	>	static final class EmptyTask extends ForkJoinTask<Void> {
536	>	EmptyTask() { status = ForkJoinTask.NORMAL; } // force done
537	>	public final Void getRawResult() { return null; }
538	>	public final void setRawResult(Void x) {}
539	>	public final boolean exec() { return true; }
540	>	}
541	>
542	>	/**
543	>	* Queues supporting work-stealing as well as external task
544	>	* submission. See above for main rationale and algorithms.
545	>	* Implementation relies heavily on "Unsafe" intrinsics
546	>	* and selective use of "volatile":
547	>	*
548	>	* Field "base" is the index (mod array.length) of the least valid
549	>	* queue slot, which is always the next position to steal (poll)
550	>	* from if nonempty. Reads and writes require volatile orderings
551	>	* but not CAS, because updates are only performed after slot
552	>	* CASes.
553	>	*
554	>	* Field "top" is the index (mod array.length) of the next queue
555	>	* slot to push to or pop from. It is written only by owner thread
556	>	* for push, or under lock for trySharedPush, and accessed by
557	>	* other threads only after reading (volatile) base.  Both top and
558	>	* base are allowed to wrap around on overflow, but (top - base)
559	>	* (or more commonly -(base - top) to force volatile read of base
560	>	* before top) still estimates size.
561	>	*
562	>	* The array slots are read and written using the emulation of
563	>	* volatiles/atomics provided by Unsafe. Insertions must in
564	>	* general use putOrderedObject as a form of releasing store to
565	>	* ensure that all writes to the task object are ordered before
566	>	* its publication in the queue. (Although we can avoid one case
567	>	* of this when locked in trySharedPush.) All removals entail a
568	>	* CAS to null.  The array is always a power of two. To ensure
569	>	* safety of Unsafe array operations, all accesses perform
570	>	* explicit null checks and implicit bounds checks via
571	>	* power-of-two masking.
572	>	*
573	>	* In addition to basic queuing support, this class contains
574	>	* fields described elsewhere to control execution. It turns out
575	>	* to work better memory-layout-wise to include them in this
576	>	* class rather than a separate class.
577	>	*
578	>	* Performance on most platforms is very sensitive to placement of
579	>	* instances of both WorkQueues and their arrays -- we absolutely
580	>	* do not want multiple WorkQueue instances or multiple queue
581	>	* arrays sharing cache lines. (It would be best for queue objects
582	>	* and their arrays to share, but there is nothing available to
583	>	* help arrange that).  Unfortunately, because they are recorded
584	>	* in a common array, WorkQueue instances are often moved to be
585	>	* adjacent by garbage collectors. To reduce impact, we use field
586	>	* padding that works OK on common platforms; this effectively
587	>	* trades off slightly slower average field access for the sake of
588	>	* avoiding really bad worst-case access. (Until better JVM
589	>	* support is in place, this padding is dependent on transient
590	>	* properties of JVM field layout rules.)  We also take care in
591	>	* allocating, sizing and resizing the array. Non-shared queue
592	>	* arrays are initialized (via method growArray) by workers before
593	>	* use. Others are allocated on first use.
594		*/
595	<	public static final ForkJoinWorkerThreadFactory
596	<	defaultForkJoinWorkerThreadFactory =
597	<	new DefaultForkJoinWorkerThreadFactory();
595	>	static final class WorkQueue {
596	>	/**
597	>	* Capacity of work-stealing queue array upon initialization.
598	>	* Must be a power of two; at least 4, but should be larger to
599	>	* reduce or eliminate cacheline sharing among queues.
600	>	* Currently, it is much larger, as a partial workaround for
601	>	* the fact that JVMs often place arrays in locations that
602	>	* share GC bookkeeping (especially cardmarks) such that
603	>	* per-write accesses encounter serious memory contention.
604	>	*/
605	>	static final int INITIAL_QUEUE_CAPACITY = 1 << 13;
606
607	<	/**
608	<	* Permission required for callers of methods that may start or
609	<	* kill threads.
610	<	*/
611	<	private static final RuntimePermission modifyThreadPermission =
612	<	new RuntimePermission("modifyThread");
607	>	/**
608	>	* Maximum size for queue arrays. Must be a power of two less
609	>	* than or equal to 1 << (31 - width of array entry) to ensure
610	>	* lack of wraparound of index calculations, but defined to a
611	>	* value a bit less than this to help users trap runaway
612	>	* programs before saturating systems.
613	>	*/
614	>	static final int MAXIMUM_QUEUE_CAPACITY = 1 << 26; // 64M
615
616	<	/**
617	<	* If there is a security manager, makes sure caller has
618	<	* permission to modify threads.
619	<	*/
620	<	private static void checkPermission() {
621	<	SecurityManager security = System.getSecurityManager();
622	<	if (security != null)
623	<	security.checkPermission(modifyThreadPermission);
624	<	}
616	>	volatile long totalSteals; // cumulative number of steals
617	>	int seed;                  // for random scanning; initialize nonzero
618	>	volatile int eventCount;   // encoded inactivation count; < 0 if inactive
619	>	int nextWait;              // encoded record of next event waiter
620	>	int rescans;               // remaining scans until block
621	>	int nsteals;               // top-level task executions since last idle
622	>	final int mode;            // lifo, fifo, or shared
623	>	int poolIndex;             // index of this queue in pool (or 0)
624	>	int stealHint;             // index of most recent known stealer
625	>	volatile int runState;     // 1: locked, -1: terminate; else 0
626	>	volatile int base;         // index of next slot for poll
627	>	int top;                   // index of next slot for push
628	>	ForkJoinTask<?>[] array;   // the elements (initially unallocated)
629	>	final ForkJoinPool pool;   // the containing pool (may be null)
630	>	final ForkJoinWorkerThread owner; // owning thread or null if shared
631	>	volatile Thread parker;    // == owner during call to park; else null
632	>	ForkJoinTask<?> currentJoin;  // task being joined in awaitJoin
633	>	ForkJoinTask<?> currentSteal; // current non-local task being executed
634	>	// Heuristic padding to ameliorate unfortunate memory placements
635	>	Object p00, p01, p02, p03, p04, p05, p06, p07;
636	>	Object p08, p09, p0a, p0b, p0c, p0d, p0e;
637	>
638	>	WorkQueue(ForkJoinPool pool, ForkJoinWorkerThread owner, int mode) {
639	>	this.mode = mode;
640	>	this.pool = pool;
641	>	this.owner = owner;
642	>	// Place indices in the center of array (that is not yet allocated)
643	>	base = top = INITIAL_QUEUE_CAPACITY >>> 1;
644	>	}
645
646	<	/**
647	<	* Generator for assigning sequence numbers as pool names.
648	<	*/
649	<	private static final AtomicInteger poolNumberGenerator =
650	<	new AtomicInteger();
646	>	/**
647	>	* Returns the approximate number of tasks in the queue.
648	>	*/
649	>	final int queueSize() {
650	>	int n = base - top;       // non-owner callers must read base first
651	>	return (n >= 0) ? 0 : -n; // ignore transient negative
652	>	}
653
654	<	/**
655	<	* The time to block in a join (see awaitJoin) before checking if
656	<	* a new worker should be (re)started to maintain parallelism
657	<	* level. The value should be short enough to maintain global
658	<	* responsiveness and progress but long enough to avoid
659	<	* counterproductive firings during GC stalls or unrelated system
660	<	* activity, and to not bog down systems with continual re-firings
661	<	* on GCs or legitimately long waits.
662	<	*/
663	<	private static final long JOIN_TIMEOUT_MILLIS = 250L; // 4 per second
654	>	/**
655	>	* Provides a more accurate estimate of whether this queue has
656	>	* any tasks than does queueSize, by checking whether a
657	>	* near-empty queue has at least one unclaimed task.
658	>	*/
659	>	final boolean isEmpty() {
660	>	ForkJoinTask<?>[] a; int m, s;
661	>	int n = base - (s = top);
662	>	return (n >= 0 ||
663	>	(n == -1 &&
664	>	((a = array) == null ||
665	>	(m = a.length - 1) < 0 ||
666	>	U.getObjectVolatile
667	>	(a, ((m & (s - 1)) << ASHIFT) + ABASE) == null)));
668	>	}
669
670	<	/**
671	<	* The wakeup interval (in nanoseconds) for the oldest worker
672	<	* waiting for an event invokes tryShutdownUnusedWorker to shrink
673	<	* the number of workers.  The exact value does not matter too
674	<	* much, but should be long enough to slowly release resources
675	<	* during long periods without use without disrupting normal use.
676	<	*/
677	<	private static final long SHRINK_RATE_NANOS =
678	<	30L * 1000L * 1000L * 1000L; // 2 per minute
670	>	/**
671	>	* Pushes a task. Call only by owner in unshared queues.
672	>	*
673	>	* @param task the task. Caller must ensure non-null.
674	>	* @throw RejectedExecutionException if array cannot be resized
675	>	*/
676	>	final void push(ForkJoinTask<?> task) {
677	>	ForkJoinTask<?>[] a; ForkJoinPool p;
678	>	int s = top, m, n;
679	>	if ((a = array) != null) {    // ignore if queue removed
680	>	U.putOrderedObject
681	>	(a, (((m = a.length - 1) & s) << ASHIFT) + ABASE, task);
682	>	if ((n = (top = s + 1) - base) <= 2) {
683	>	if ((p = pool) != null)
684	>	p.signalWork();
685	>	}
686	>	else if (n >= m)
687	>	growArray(true);
688	>	}
689	>	}
690
691	<	/**
692	<	* Absolute bound for parallelism level. Twice this number plus
693	<	* one (i.e., 0xfff) must fit into a 16bit field to enable
694	<	* word-packing for some counts and indices.
695	<	*/
696	<	private static final int MAX_WORKERS   = 0x7fff;
691	>	/**
692	>	* Pushes a task if lock is free and array is either big
693	>	* enough or can be resized to be big enough.
694	>	*
695	>	* @param task the task. Caller must ensure non-null.
696	>	* @return true if submitted
697	>	*/
698	>	final boolean trySharedPush(ForkJoinTask<?> task) {
699	>	boolean submitted = false;
700	>	if (runState == 0 && U.compareAndSwapInt(this, RUNSTATE, 0, 1)) {
701	>	ForkJoinTask<?>[] a = array;
702	>	int s = top;
703	>	try {
704	>	if ((a != null && a.length > s + 1 - base) ||
705	>	(a = growArray(false)) != null) { // must presize
706	>	int j = (((a.length - 1) & s) << ASHIFT) + ABASE;
707	>	U.putObject(a, (long)j, task);    // don't need "ordered"
708	>	top = s + 1;
709	>	submitted = true;
710	>	}
711	>	} finally {
712	>	runState = 0;                         // unlock
713	>	}
714	>	}
715	>	return submitted;
716	>	}
717
718	<	/**
719	<	* Array holding all worker threads in the pool.  Array size must
720	<	* be a power of two.  Updates and replacements are protected by
721	<	* workerLock, but the array is always kept in a consistent enough
722	<	* state to be randomly accessed without locking by workers
723	<	* performing work-stealing, as well as other traversal-based
724	<	* methods in this class. All readers must tolerate that some
725	<	* array slots may be null.
726	<	*/
727	<	volatile ForkJoinWorkerThread[] workers;
718	>	/**
719	>	* Takes next task, if one exists, in LIFO order.  Call only
720	>	* by owner in unshared queues. (We do not have a shared
721	>	* version of this method because it is never needed.)
722	>	*/
723	>	final ForkJoinTask<?> pop() {
724	>	ForkJoinTask<?> t; int m;
725	>	ForkJoinTask<?>[] a = array;
726	>	if (a != null && (m = a.length - 1) >= 0) {
727	>	for (int s; (s = top - 1) - base >= 0;) {
728	>	int j = ((m & s) << ASHIFT) + ABASE;
729	>	if ((t = (ForkJoinTask<?>)U.getObjectVolatile(a, j)) == null)
730	>	break;
731	>	if (U.compareAndSwapObject(a, j, t, null)) {
732	>	top = s;
733	>	return t;
734	>	}
735	>	}
736	>	}
737	>	return null;
738	>	}
739
740	<	/**
741	<	* Queue for external submissions.
742	<	*/
743	<	private final LinkedTransferQueue<ForkJoinTask<?>> submissionQueue;
740	>	/**
741	>	* Takes a task in FIFO order if b is base of queue and a task
742	>	* can be claimed without contention. Specialized versions
743	>	* appear in ForkJoinPool methods scan and tryHelpStealer.
744	>	*/
745	>	final ForkJoinTask<?> pollAt(int b) {
746	>	ForkJoinTask<?> t; ForkJoinTask<?>[] a;
747	>	if ((a = array) != null) {
748	>	int j = (((a.length - 1) & b) << ASHIFT) + ABASE;
749	>	if ((t = (ForkJoinTask<?>)U.getObjectVolatile(a, j)) != null &&
750	>	base == b &&
751	>	U.compareAndSwapObject(a, j, t, null)) {
752	>	base = b + 1;
753	>	return t;
754	>	}
755	>	}
756	>	return null;
757	>	}
758
759	<	/**
760	<	* Lock protecting updates to workers array.
761	<	*/
762	<	private final ReentrantLock workerLock;
759	>	/**
760	>	* Takes next task, if one exists, in FIFO order.
761	>	*/
762	>	final ForkJoinTask<?> poll() {
763	>	ForkJoinTask<?>[] a; int b; ForkJoinTask<?> t;
764	>	while ((b = base) - top < 0 && (a = array) != null) {
765	>	int j = (((a.length - 1) & b) << ASHIFT) + ABASE;
766	>	t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
767	>	if (t != null) {
768	>	if (base == b &&
769	>	U.compareAndSwapObject(a, j, t, null)) {
770	>	base = b + 1;
771	>	return t;
772	>	}
773	>	}
774	>	else if (base == b) {
775	>	if (b + 1 == top)
776	>	break;
777	>	Thread.yield(); // wait for lagging update
778	>	}
779	>	}
780	>	return null;
781	>	}
782
783	<	/**
784	<	* Latch released upon termination.
785	<	*/
786	<	private final Phaser termination;
783	>	/**
784	>	* Takes next task, if one exists, in order specified by mode.
785	>	*/
786	>	final ForkJoinTask<?> nextLocalTask() {
787	>	return mode == 0 ? pop() : poll();
788	>	}
789
790	<	/**
791	<	* Creation factory for worker threads.
792	<	*/
793	<	private final ForkJoinWorkerThreadFactory factory;
790	>	/**
791	>	* Returns next task, if one exists, in order specified by mode.
792	>	*/
793	>	final ForkJoinTask<?> peek() {
794	>	ForkJoinTask<?>[] a = array; int m;
795	>	if (a == null || (m = a.length - 1) < 0)
796	>	return null;
797	>	int i = mode == 0 ? top - 1 : base;
798	>	int j = ((i & m) << ASHIFT) + ABASE;
799	>	return (ForkJoinTask<?>)U.getObjectVolatile(a, j);
800	>	}
801
802	<	/**
803	<	* Sum of per-thread steal counts, updated only when threads are
804	<	* idle or terminating.
805	<	*/
806	<	private volatile long stealCount;
802	>	/**
803	>	* Pops the given task only if it is at the current top.
804	>	*/
805	>	final boolean tryUnpush(ForkJoinTask<?> t) {
806	>	ForkJoinTask<?>[] a; int s;
807	>	if ((a = array) != null && (s = top) != base &&
808	>	U.compareAndSwapObject
809	>	(a, (((a.length - 1) & --s) << ASHIFT) + ABASE, t, null)) {
810	>	top = s;
811	>	return true;
812	>	}
813	>	return false;
814	>	}
815
816	<	/**
817	<	* Encoded record of top of Treiber stack of threads waiting for
818	<	* events. The top 32 bits contain the count being waited for. The
819	<	* bottom 16 bits contains one plus the pool index of waiting
820	<	* worker thread. (Bits 16-31 are unused.)
821	<	*/
822	<	private volatile long eventWaiters;
816	>	/**
817	>	* Polls the given task only if it is at the current base.
818	>	*/
819	>	final boolean pollFor(ForkJoinTask<?> task) {
820	>	ForkJoinTask<?>[] a; int b;
821	>	if ((b = base) - top < 0 && (a = array) != null) {
822	>	int j = (((a.length - 1) & b) << ASHIFT) + ABASE;
823	>	if (U.getObjectVolatile(a, j) == task && base == b &&
824	>	U.compareAndSwapObject(a, j, task, null)) {
825	>	base = b + 1;
826	>	return true;
827	>	}
828	>	}
829	>	return false;
830	>	}
831	>
832	>	/**
833	>	* If present, removes from queue and executes the given task, or
834	>	* any other cancelled task. Returns (true) immediately on any CAS
835	>	* or consistency check failure so caller can retry.
836	>	*
837	>	* @return false if no progress can be made
838	>	*/
839	>	final boolean tryRemoveAndExec(ForkJoinTask<?> task) {
840	>	boolean removed = false, empty = true, progress = true;
841	>	ForkJoinTask<?>[] a; int m, s, b, n;
842	>	if ((a = array) != null && (m = a.length - 1) >= 0 &&
843	>	(n = (s = top) - (b = base)) > 0) {
844	>	for (ForkJoinTask<?> t;;) {           // traverse from s to b
845	>	int j = ((--s & m) << ASHIFT) + ABASE;
846	>	t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
847	>	if (t == null)                    // inconsistent length
848	>	break;
849	>	else if (t == task) {
850	>	if (s + 1 == top) {           // pop
851	>	if (!U.compareAndSwapObject(a, j, task, null))
852	>	break;
853	>	top = s;
854	>	removed = true;
855	>	}
856	>	else if (base == b)           // replace with proxy
857	>	removed = U.compareAndSwapObject(a, j, task,
858	>	new EmptyTask());
859	>	break;
860	>	}
861	>	else if (t.status >= 0)
862	>	empty = false;
863	>	else if (s + 1 == top) {          // pop and throw away
864	>	if (U.compareAndSwapObject(a, j, t, null))
865	>	top = s;
866	>	break;
867	>	}
868	>	if (--n == 0) {
869	>	if (!empty && base == b)
870	>	progress = false;
871	>	break;
872	>	}
873	>	}
874	>	}
875	>	if (removed)
876	>	task.doExec();
877	>	return progress;
878	>	}
879
880	<	private static final int  EVENT_COUNT_SHIFT = 32;
881	<	private static final long WAITER_ID_MASK    = (1L << 16) - 1L;
880	>	/**
881	>	* Initializes or doubles the capacity of array. Call either
882	>	* by owner or with lock held -- it is OK for base, but not
883	>	* top, to move while resizings are in progress.
884	>	*
885	>	* @param rejectOnFailure if true, throw exception if capacity
886	>	* exceeded (relayed ultimately to user); else return null.
887	>	*/
888	>	final ForkJoinTask<?>[] growArray(boolean rejectOnFailure) {
889	>	ForkJoinTask<?>[] oldA = array;
890	>	int size = oldA != null ? oldA.length << 1 : INITIAL_QUEUE_CAPACITY;
891	>	if (size <= MAXIMUM_QUEUE_CAPACITY) {
892	>	int oldMask, t, b;
893	>	ForkJoinTask<?>[] a = array = new ForkJoinTask<?>[size];
894	>	if (oldA != null && (oldMask = oldA.length - 1) >= 0 &&
895	>	(t = top) - (b = base) > 0) {
896	>	int mask = size - 1;
897	>	do {
898	>	ForkJoinTask<?> x;
899	>	int oldj = ((b & oldMask) << ASHIFT) + ABASE;
900	>	int j    = ((b &    mask) << ASHIFT) + ABASE;
901	>	x = (ForkJoinTask<?>)U.getObjectVolatile(oldA, oldj);
902	>	if (x != null &&
903	>	U.compareAndSwapObject(oldA, oldj, x, null))
904	>	U.putObjectVolatile(a, j, x);
905	>	} while (++b != t);
906	>	}
907	>	return a;
908	>	}
909	>	else if (!rejectOnFailure)
910	>	return null;
911	>	else
912	>	throw new RejectedExecutionException("Queue capacity exceeded");
913	>	}
914
915	<	/**
916	<	* A counter for events that may wake up worker threads:
917	<	*   - Submission of a new task to the pool
918	<	*   - A worker pushing a task on an empty queue
919	<	*   - termination
920	<	*/
921	<	private volatile int eventCount;
915	>	/**
916	>	* Removes and cancels all known tasks, ignoring any exceptions.
917	>	*/
918	>	final void cancelAll() {
919	>	ForkJoinTask.cancelIgnoringExceptions(currentJoin);
920	>	ForkJoinTask.cancelIgnoringExceptions(currentSteal);
921	>	for (ForkJoinTask<?> t; (t = poll()) != null; )
922	>	ForkJoinTask.cancelIgnoringExceptions(t);
923	>	}
924
925	<	/**
926	<	* Encoded record of top of Treiber stack of spare threads waiting
927	<	* for resumption. The top 16 bits contain an arbitrary count to
928	<	* avoid ABA effects. The bottom 16bits contains one plus the pool
929	<	* index of waiting worker thread.
930	<	*/
931	<	private volatile int spareWaiters;
925	>	/**
926	>	* Computes next value for random probes.  Scans don't require
927	>	* a very high quality generator, but also not a crummy one.
928	>	* Marsaglia xor-shift is cheap and works well enough.  Note:
929	>	* This is manually inlined in its usages in ForkJoinPool to
930	>	* avoid writes inside busy scan loops.
931	>	*/
932	>	final int nextSeed() {
933	>	int r = seed;
934	>	r ^= r << 13;
935	>	r ^= r >>> 17;
936	>	return seed = r ^= r << 5;
937	>	}
938
939	<	private static final int SPARE_COUNT_SHIFT = 16;
940	<	private static final int SPARE_ID_MASK     = (1 << 16) - 1;
939	>	// Execution methods
940	>
941	>	/**
942	>	* Removes and runs tasks until empty, using local mode
943	>	* ordering. Normally called only after checking for apparent
944	>	* non-emptiness.
945	>	*/
946	>	final void runLocalTasks() {
947	>	// hoist checks from repeated pop/poll
948	>	ForkJoinTask<?>[] a; int m;
949	>	if ((a = array) != null && (m = a.length - 1) >= 0) {
950	>	if (mode == 0) {
951	>	for (int s; (s = top - 1) - base >= 0;) {
952	>	int j = ((m & s) << ASHIFT) + ABASE;
953	>	ForkJoinTask<?> t =
954	>	(ForkJoinTask<?>)U.getObjectVolatile(a, j);
955	>	if (t != null) {
956	>	if (U.compareAndSwapObject(a, j, t, null)) {
957	>	top = s;
958	>	t.doExec();
959	>	}
960	>	}
961	>	else
962	>	break;
963	>	}
964	>	}
965	>	else {
966	>	for (int b; (b = base) - top < 0;) {
967	>	int j = ((m & b) << ASHIFT) + ABASE;
968	>	ForkJoinTask<?> t =
969	>	(ForkJoinTask<?>)U.getObjectVolatile(a, j);
970	>	if (t != null) {
971	>	if (base == b &&
972	>	U.compareAndSwapObject(a, j, t, null)) {
973	>	base = b + 1;
974	>	t.doExec();
975	>	}
976	>	} else if (base == b) {
977	>	if (b + 1 == top)
978	>	break;
979	>	Thread.yield(); // wait for lagging update
980	>	}
981	>	}
982	>	}
983	>	}
984	>	}
985	>
986	>	/**
987	>	* Executes a top-level task and any local tasks remaining
988	>	* after execution.
989	>	*
990	>	* @return true unless terminating
991	>	*/
992	>	final boolean runTask(ForkJoinTask<?> t) {
993	>	boolean alive = true;
994	>	if (t != null) {
995	>	currentSteal = t;
996	>	t.doExec();
997	>	if (top != base)        // conservative guard
998	>	runLocalTasks();
999	>	++nsteals;
1000	>	currentSteal = null;
1001	>	}
1002	>	else if (runState < 0)      // terminating
1003	>	alive = false;
1004	>	return alive;
1005	>	}
1006	>
1007	>	/**
1008	>	* Executes a non-top-level (stolen) task.
1009	>	*/
1010	>	final void runSubtask(ForkJoinTask<?> t) {
1011	>	if (t != null) {
1012	>	ForkJoinTask<?> ps = currentSteal;
1013	>	currentSteal = t;
1014	>	t.doExec();
1015	>	currentSteal = ps;
1016	>	}
1017	>	}
1018	>
1019	>	/**
1020	>	* Returns true if owned and not known to be blocked.
1021	>	*/
1022	>	final boolean isApparentlyUnblocked() {
1023	>	Thread wt; Thread.State s;
1024	>	return (eventCount >= 0 &&
1025	>	(wt = owner) != null &&
1026	>	(s = wt.getState()) != Thread.State.BLOCKED &&
1027	>	s != Thread.State.WAITING &&
1028	>	s != Thread.State.TIMED_WAITING);
1029	>	}
1030	>
1031	>	/**
1032	>	* If this owned and is not already interrupted, try to
1033	>	* interrupt and/or unpark, ignoring exceptions.
1034	>	*/
1035	>	final void interruptOwner() {
1036	>	Thread wt, p;
1037	>	if ((wt = owner) != null && !wt.isInterrupted()) {
1038	>	try {
1039	>	wt.interrupt();
1040	>	} catch (SecurityException ignore) {
1041	>	}
1042	>	}
1043	>	if ((p = parker) != null)
1044	>	U.unpark(p);
1045	>	}
1046	>
1047	>	// Unsafe mechanics
1048	>	private static final sun.misc.Unsafe U;
1049	>	private static final long RUNSTATE;
1050	>	private static final int ABASE;
1051	>	private static final int ASHIFT;
1052	>	static {
1053	>	int s;
1054	>	try {
1055	>	U = getUnsafe();
1056	>	Class<?> k = WorkQueue.class;
1057	>	Class<?> ak = ForkJoinTask[].class;
1058	>	RUNSTATE = U.objectFieldOffset
1059	>	(k.getDeclaredField("runState"));
1060	>	ABASE = U.arrayBaseOffset(ak);
1061	>	s = U.arrayIndexScale(ak);
1062	>	} catch (Exception e) {
1063	>	throw new Error(e);
1064	>	}
1065	>	if ((s & (s-1)) != 0)
1066	>	throw new Error("data type scale not a power of two");
1067	>	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
1068	>	}
1069	>	}
1070
1071		/**
1072	<	* Lifecycle control. The low word contains the number of workers
1073	<	* that are (probably) executing tasks. This value is atomically
1074	<	* incremented before a worker gets a task to run, and decremented
1075	<	* when worker has no tasks and cannot find any.  Bits 16-18
1076	<	* contain runLevel value. When all are zero, the pool is
520	<	* running. Level transitions are monotonic (running -> shutdown
521	<	* -> terminating -> terminated) so each transition adds a bit.
522	<	* These are bundled together to ensure consistent read for
523	<	* termination checks (i.e., that runLevel is at least SHUTDOWN
524	<	* and active threads is zero).
1072	>	* Per-thread records for threads that submit to pools. Currently
1073	>	* holds only pseudo-random seed / index that is used to choose
1074	>	* submission queues in method doSubmit. In the future, this may
1075	>	* also incorporate a means to implement different task rejection
1076	>	* and resubmission policies.
1077		*
1078	<	* Notes: Most direct CASes are dependent on these bitfield
1079	<	* positions.  Also, this field is non-private to enable direct
1080	<	* performance-sensitive CASes in ForkJoinWorkerThread.
1078	>	* Seeds for submitters and workers/workQueues work in basically
1079	>	* the same way but are initialized and updated using slightly
1080	>	* different mechanics. Both are initialized using the same
1081	>	* approach as in class ThreadLocal, where successive values are
1082	>	* unlikely to collide with previous values. This is done during
1083	>	* registration for workers, but requires a separate AtomicInteger
1084	>	* for submitters. Seeds are then randomly modified upon
1085	>	* collisions using xorshifts, which requires a non-zero seed.
1086		*/
1087	<	volatile int runState;
1087	>	static final class Submitter {
1088	>	int seed;
1089	>	Submitter() {
1090	>	int s = nextSubmitterSeed.getAndAdd(SEED_INCREMENT);
1091	>	seed = (s == 0) ? 1 : s; // ensure non-zero
1092	>	}
1093	>	}
1094
1095	<	// Note: The order among run level values matters.
1096	<	private static final int RUNLEVEL_SHIFT     = 16;
1097	<	private static final int SHUTDOWN           = 1 << RUNLEVEL_SHIFT;
1098	<	private static final int TERMINATING        = 1 << (RUNLEVEL_SHIFT + 1);
1099	<	private static final int TERMINATED         = 1 << (RUNLEVEL_SHIFT + 2);
1100	<	private static final int ACTIVE_COUNT_MASK  = (1 << RUNLEVEL_SHIFT) - 1;
1095	>	/** ThreadLocal class for Submitters */
1096	>	static final class ThreadSubmitter extends ThreadLocal<Submitter> {
1097	>	public Submitter initialValue() { return new Submitter(); }
1098	>	}
1099	>
1100	>	// static fields (initialized in static initializer below)
1101
1102		/**
1103	<	* Holds number of total (i.e., created and not yet terminated)
1104	<	* and running (i.e., not blocked on joins or other managed sync)
542	<	* threads, packed together to ensure consistent snapshot when
543	<	* making decisions about creating and suspending spare
544	<	* threads. Updated only by CAS. Note that adding a new worker
545	<	* requires incrementing both counts, since workers start off in
546	<	* running state.
1103	>	* Creates a new ForkJoinWorkerThread. This factory is used unless
1104	>	* overridden in ForkJoinPool constructors.
1105		*/
1106	<	private volatile int workerCounts;
1107	<
550	<	private static final int TOTAL_COUNT_SHIFT  = 16;
551	<	private static final int RUNNING_COUNT_MASK = (1 << TOTAL_COUNT_SHIFT) - 1;
552	<	private static final int ONE_RUNNING        = 1;
553	<	private static final int ONE_TOTAL          = 1 << TOTAL_COUNT_SHIFT;
1106	>	public static final ForkJoinWorkerThreadFactory
1107	>	defaultForkJoinWorkerThreadFactory;
1108
1109		/**
1110	<	* The target parallelism level.
557	<	* Accessed directly by ForkJoinWorkerThreads.
1110	>	* Generator for assigning sequence numbers as pool names.
1111		*/
1112	<	final int parallelism;
1112	>	private static final AtomicInteger poolNumberGenerator;
1113
1114		/**
1115	<	* True if use local fifo, not default lifo, for local polling
1116	<	* Read by, and replicated by ForkJoinWorkerThreads
1115	>	* Generator for initial hashes/seeds for submitters. Accessed by
1116	>	* Submitter class constructor.
1117		*/
1118	<	final boolean locallyFifo;
1118	>	static final AtomicInteger nextSubmitterSeed;
1119
1120		/**
1121	<	* The uncaught exception handler used when any worker abruptly
1122	<	* terminates.
1121	>	* Permission required for callers of methods that may start or
1122	>	* kill threads.
1123		*/
1124	<	private final Thread.UncaughtExceptionHandler ueh;
1124	>	private static final RuntimePermission modifyThreadPermission;
1125
1126		/**
1127	<	* Pool number, just for assigning useful names to worker threads
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	<	private final int poolNumber;
1131	>	private static final ThreadSubmitter submitters;
1132
1133	<	// Utilities for CASing fields. Note that most of these
579	<	// are usually manually inlined by callers
1133	>	// static constants
1134
1135		/**
1136	<	* Increments running count part of workerCounts
1136	>	* The wakeup interval (in nanoseconds) for a worker waiting for a
1137	>	* task when the pool is quiescent to instead try to shrink the
1138	>	* number of workers.  The exact value does not matter too
1139	>	* much. It must be short enough to release resources during
1140	>	* sustained periods of idleness, but not so short that threads
1141	>	* are continually re-created.
1142		*/
1143	<	final void incrementRunningCount() {
1144	<	int c;
586	<	do {} while (!UNSAFE.compareAndSwapInt(this, workerCountsOffset,
587	<	c = workerCounts,
588	<	c + ONE_RUNNING));
589	<	}
1143	>	private static final long SHRINK_RATE =
1144	>	4L * 1000L * 1000L * 1000L; // 4 seconds
1145
1146		/**
1147	<	* Tries to decrement running count unless already zero
1147	>	* The timeout value for attempted shrinkage, includes
1148	>	* some slop to cope with system timer imprecision.
1149		*/
1150	<	final boolean tryDecrementRunningCount() {
595	<	int wc = workerCounts;
596	<	if ((wc & RUNNING_COUNT_MASK) == 0)
597	<	return false;
598	<	return UNSAFE.compareAndSwapInt(this, workerCountsOffset,
599	<	wc, wc - ONE_RUNNING);
600	<	}
1150	>	private static final long SHRINK_TIMEOUT = SHRINK_RATE - (SHRINK_RATE / 10);
1151
1152		/**
1153	<	* Forces decrement of encoded workerCounts, awaiting nonzero if
1154	<	* (rarely) necessary when other count updates lag.
1155	<	*
1156	<	* @param dr -- either zero or ONE_RUNNING
1157	<	* @param dt == either zero or ONE_TOTAL
1153	>	* The maximum stolen->joining link depth allowed in method
1154	>	* tryHelpStealer.  Must be a power of two. This value also
1155	>	* controls the maximum number of times to try to help join a task
1156	>	* without any apparent progress or change in pool state before
1157	>	* giving up and blocking (see awaitJoin).  Depths for legitimate
1158	>	* chains are unbounded, but we use a fixed constant to avoid
1159	>	* (otherwise unchecked) cycles and to bound staleness of
1160	>	* traversal parameters at the expense of sometimes blocking when
1161	>	* we could be helping.
1162		*/
1163	<	private void decrementWorkerCounts(int dr, int dt) {
610	<	for (;;) {
611	<	int wc = workerCounts;
612	<	if ((wc & RUNNING_COUNT_MASK)  - dr < 0 ||
613	<	(wc >>> TOTAL_COUNT_SHIFT) - dt < 0) {
614	<	if ((runState & TERMINATED) != 0)
615	<	return; // lagging termination on a backout
616	<	Thread.yield();
617	<	}
618	<	if (UNSAFE.compareAndSwapInt(this, workerCountsOffset,
619	<	wc, wc - (dr + dt)))
620	<	return;
621	<	}
622	<	}
1163	>	private static final int MAX_HELP = 32;
1164
1165		/**
1166	<	* Tries decrementing active count; fails on contention.
1167	<	* Called when workers cannot find tasks to run.
1166	>	* Secondary time-based bound (in nanosecs) for helping attempts
1167	>	* before trying compensated blocking in awaitJoin. Used in
1168	>	* conjunction with MAX_HELP to reduce variance due to different
1169	>	* polling rates associated with different helping options. The
1170	>	* value should roughly approximate the time required to create
1171	>	* and/or activate a worker thread.
1172		*/
1173	<	final boolean tryDecrementActiveCount() {
629	<	int c;
630	<	return UNSAFE.compareAndSwapInt(this, runStateOffset,
631	<	c = runState, c - 1);
632	<	}
1173	>	private static final long COMPENSATION_DELAY = 100L * 1000L; // 0.1 millisec
1174
1175		/**
1176	<	* Advances to at least the given level. Returns true if not
1177	<	* already in at least the given level.
1176	>	* Increment for seed generators. See class ThreadLocal for
1177	>	* explanation.
1178		*/
1179	<	private boolean advanceRunLevel(int level) {
639	<	for (;;) {
640	<	int s = runState;
641	<	if ((s & level) != 0)
642	<	return false;
643	<	if (UNSAFE.compareAndSwapInt(this, runStateOffset, s, s | level))
644	<	return true;
645	<	}
646	<	}
647	<
648	<	// workers array maintenance
1179	>	private static final int SEED_INCREMENT = 0x61c88647;
1180
1181		/**
1182	<	* Records and returns a workers array index for new worker.
1182	>	* Bits and masks for control variables
1183	>	*
1184	>	* Field ctl is a long packed with:
1185	>	* AC: Number of active running workers minus target parallelism (16 bits)
1186	>	* TC: Number of total workers minus target parallelism (16 bits)
1187	>	* ST: true if pool is terminating (1 bit)
1188	>	* EC: the wait count of top waiting thread (15 bits)
1189	>	* ID: poolIndex of top of Treiber stack of waiters (16 bits)
1190	>	*
1191	>	* When convenient, we can extract the upper 32 bits of counts and
1192	>	* the lower 32 bits of queue state, u = (int)(ctl >>> 32) and e =
1193	>	* (int)ctl.  The ec field is never accessed alone, but always
1194	>	* together with id and st. The offsets of counts by the target
1195	>	* parallelism and the positionings of fields makes it possible to
1196	>	* perform the most common checks via sign tests of fields: When
1197	>	* ac is negative, there are not enough active workers, when tc is
1198	>	* negative, there are not enough total workers, and when e is
1199	>	* negative, the pool is terminating.  To deal with these possibly
1200	>	* negative fields, we use casts in and out of "short" and/or
1201	>	* signed shifts to maintain signedness.
1202	>	*
1203	>	* When a thread is queued (inactivated), its eventCount field is
1204	>	* set negative, which is the only way to tell if a worker is
1205	>	* prevented from executing tasks, even though it must continue to
1206	>	* scan for them to avoid queuing races. Note however that
1207	>	* eventCount updates lag releases so usage requires care.
1208	>	*
1209	>	* Field runState is an int packed with:
1210	>	* SHUTDOWN: true if shutdown is enabled (1 bit)
1211	>	* SEQ:  a sequence number updated upon (de)registering workers (30 bits)
1212	>	* INIT: set true after workQueues array construction (1 bit)
1213	>	*
1214	>	* The sequence number enables simple consistency checks:
1215	>	* Staleness of read-only operations on the workQueues array can
1216	>	* be checked by comparing runState before vs after the reads.
1217		*/
1218	<	private int recordWorker(ForkJoinWorkerThread w) {
1219	<	// Try using slot totalCount-1. If not available, scan and/or resize
1220	<	int k = (workerCounts >>> TOTAL_COUNT_SHIFT) - 1;
1221	<	final ReentrantLock lock = this.workerLock;
1222	<	lock.lock();
1218	>
1219	>	// bit positions/shifts for fields
1220	>	private static final int  AC_SHIFT   = 48;
1221	>	private static final int  TC_SHIFT   = 32;
1222	>	private static final int  ST_SHIFT   = 31;
1223	>	private static final int  EC_SHIFT   = 16;
1224	>
1225	>	// bounds
1226	>	private static final int  SMASK      = 0xffff;  // short bits
1227	>	private static final int  MAX_CAP    = 0x7fff;  // max #workers - 1
1228	>	private static final int  SQMASK     = 0xfffe;  // even short bits
1229	>	private static final int  SHORT_SIGN = 1 << 15;
1230	>	private static final int  INT_SIGN   = 1 << 31;
1231	>
1232	>	// masks
1233	>	private static final long STOP_BIT   = 0x0001L << ST_SHIFT;
1234	>	private static final long AC_MASK    = ((long)SMASK) << AC_SHIFT;
1235	>	private static final long TC_MASK    = ((long)SMASK) << TC_SHIFT;
1236	>
1237	>	// units for incrementing and decrementing
1238	>	private static final long TC_UNIT    = 1L << TC_SHIFT;
1239	>	private static final long AC_UNIT    = 1L << AC_SHIFT;
1240	>
1241	>	// masks and units for dealing with u = (int)(ctl >>> 32)
1242	>	private static final int  UAC_SHIFT  = AC_SHIFT - 32;
1243	>	private static final int  UTC_SHIFT  = TC_SHIFT - 32;
1244	>	private static final int  UAC_MASK   = SMASK << UAC_SHIFT;
1245	>	private static final int  UTC_MASK   = SMASK << UTC_SHIFT;
1246	>	private static final int  UAC_UNIT   = 1 << UAC_SHIFT;
1247	>	private static final int  UTC_UNIT   = 1 << UTC_SHIFT;
1248	>
1249	>	// masks and units for dealing with e = (int)ctl
1250	>	private static final int E_MASK      = 0x7fffffff; // no STOP_BIT
1251	>	private static final int E_SEQ       = 1 << EC_SHIFT;
1252	>
1253	>	// runState bits
1254	>	private static final int SHUTDOWN    = 1 << 31;
1255	>
1256	>	// access mode for WorkQueue
1257	>	static final int LIFO_QUEUE          =  0;
1258	>	static final int FIFO_QUEUE          =  1;
1259	>	static final int SHARED_QUEUE        = -1;
1260	>
1261	>	// Instance fields
1262	>
1263	>	/*
1264	>	* Field layout order in this class tends to matter more than one
1265	>	* would like. Runtime layout order is only loosely related to
1266	>	* declaration order and may differ across JVMs, but the following
1267	>	* empirically works OK on current JVMs.
1268	>	*/
1269	>
1270	>	volatile long ctl;                         // main pool control
1271	>	final int parallelism;                     // parallelism level
1272	>	final int localMode;                       // per-worker scheduling mode
1273	>	final int submitMask;                      // submit queue index bound
1274	>	int nextSeed;                              // for initializing worker seeds
1275	>	volatile int runState;                     // shutdown status and seq
1276	>	WorkQueue[] workQueues;                    // main registry
1277	>	final Mutex lock;                          // for registration
1278	>	final Condition termination;               // for awaitTermination
1279	>	final ForkJoinWorkerThreadFactory factory; // factory for new workers
1280	>	final Thread.UncaughtExceptionHandler ueh; // per-worker UEH
1281	>	final AtomicLong stealCount;               // collect counts when terminated
1282	>	final AtomicInteger nextWorkerNumber;      // to create worker name string
1283	>	final String workerNamePrefix;             // to create worker name string
1284	>
1285	>	//  Creating, registering, and deregistering workers
1286	>
1287	>	/**
1288	>	* Tries to create and start a worker
1289	>	*/
1290	>	private void addWorker() {
1291	>	Throwable ex = null;
1292	>	ForkJoinWorkerThread wt = null;
1293		try {
1294	<	ForkJoinWorkerThread[] ws = workers;
1295	<	int n = ws.length;
1296	<	if (k < 0 || k >= n || ws[k] != null) {
662	<	for (k = 0; k < n && ws[k] != null; ++k)
663	<	;
664	<	if (k == n)
665	<	ws = Arrays.copyOf(ws, n << 1);
1294	>	if ((wt = factory.newThread(this)) != null) {
1295	>	wt.start();
1296	>	return;
1297		}
1298	<	ws[k] = w;
1299	<	workers = ws; // volatile array write ensures slot visibility
669	<	} finally {
670	<	lock.unlock();
1298	>	} catch (Throwable e) {
1299	>	ex = e;
1300		}
1301	<	return k;
1301	>	deregisterWorker(wt, ex); // adjust counts etc on failure
1302	>	}
1303	>
1304	>	/**
1305	>	* Callback from ForkJoinWorkerThread constructor to assign a
1306	>	* public name. This must be separate from registerWorker because
1307	>	* it is called during the "super" constructor call in
1308	>	* ForkJoinWorkerThread.
1309	>	*/
1310	>	final String nextWorkerName() {
1311	>	return workerNamePrefix.concat
1312	>	(Integer.toString(nextWorkerNumber.addAndGet(1)));
1313		}
1314
1315		/**
1316	<	* Nulls out record of worker in workers array.
1316	>	* Callback from ForkJoinWorkerThread constructor to establish its
1317	>	* poolIndex and record its WorkQueue. To avoid scanning bias due
1318	>	* to packing entries in front of the workQueues array, we treat
1319	>	* the array as a simple power-of-two hash table using per-thread
1320	>	* seed as hash, expanding as needed.
1321	>	*
1322	>	* @param w the worker's queue
1323		*/
1324	<	private void forgetWorker(ForkJoinWorkerThread w) {
1325	<	int idx = w.poolIndex;
680	<	// Locking helps method recordWorker avoid unnecessary expansion
681	<	final ReentrantLock lock = this.workerLock;
1324	>	final void registerWorker(WorkQueue w) {
1325	>	Mutex lock = this.lock;
1326		lock.lock();
1327		try {
1328	<	ForkJoinWorkerThread[] ws = workers;
1329	<	if (idx >= 0 && idx < ws.length && ws[idx] == w) // verify
1330	<	ws[idx] = null;
1328	>	WorkQueue[] ws = workQueues;
1329	>	if (w != null && ws != null) {          // skip on shutdown/failure
1330	>	int rs, n;
1331	>	while ((n = ws.length) <            // ensure can hold total
1332	>	(parallelism + (short)(ctl >>> TC_SHIFT) << 1))
1333	>	workQueues = ws = Arrays.copyOf(ws, n << 1);
1334	>	int m = n - 1;
1335	>	int s = nextSeed += SEED_INCREMENT; // rarely-colliding sequence
1336	>	w.seed = (s == 0) ? 1 : s;          // ensure non-zero seed
1337	>	int r = (s << 1) | 1;               // use odd-numbered indices
1338	>	while (ws[r &= m] != null)          // step by approx half size
1339	>	r += ((n >>> 1) & SQMASK) + 2;
1340	>	w.eventCount = w.poolIndex = r;     // establish before recording
1341	>	ws[r] = w;                          // also update seq
1342	>	runState = ((rs = runState) & SHUTDOWN) | ((rs + 2) & ~SHUTDOWN);
1343	>	}
1344		} finally {
1345		lock.unlock();
1346		}
1347		}
1348
1349		/**
1350	<	* Final callback from terminating worker.  Removes record of
1350	>	* Final callback from terminating worker, as well as upon failure
1351	>	* to construct or start a worker in addWorker.  Removes record of
1352		* worker from array, and adjusts counts. If pool is shutting
1353		* down, tries to complete termination.
1354		*
1355	<	* @param w the worker
1355	>	* @param wt the worker thread or null if addWorker failed
1356	>	* @param ex the exception causing failure, or null if none
1357		*/
1358	<	final void workerTerminated(ForkJoinWorkerThread w) {
1359	<	forgetWorker(w);
1360	<	decrementWorkerCounts(w.isTrimmed()? 0 : ONE_RUNNING, ONE_TOTAL);
1361	<	while (w.stealCount != 0) // collect final count
1362	<	tryAccumulateStealCount(w);
1363	<	tryTerminate(false);
1364	<	}
1365	<
1366	<	// Waiting for and signalling events
1367	<
1368	<	/**
1369	<	* Releases workers blocked on a count not equal to current count.
1370	<	* Normally called after precheck that eventWaiters isn't zero to
1371	<	* avoid wasted array checks. Gives up upon a change in count or
713	<	* upon releasing two workers, letting others take over.
714	<	*/
715	<	private void releaseEventWaiters() {
716	<	ForkJoinWorkerThread[] ws = workers;
717	<	int n = ws.length;
718	<	long h = eventWaiters;
719	<	int ec = eventCount;
720	<	boolean releasedOne = false;
721	<	ForkJoinWorkerThread w; int id;
722	<	while ((id = ((int)(h & WAITER_ID_MASK)) - 1) >= 0 &&
723	<	(int)(h >>> EVENT_COUNT_SHIFT) != ec &&
724	<	id < n && (w = ws[id]) != null) {
725	<	if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
726	<	h,  w.nextWaiter)) {
727	<	LockSupport.unpark(w);
728	<	if (releasedOne) // exit on second release
729	<	break;
730	<	releasedOne = true;
1358	>	final void deregisterWorker(ForkJoinWorkerThread wt, Throwable ex) {
1359	>	Mutex lock = this.lock;
1360	>	WorkQueue w = null;
1361	>	if (wt != null && (w = wt.workQueue) != null) {
1362	>	w.runState = -1;                // ensure runState is set
1363	>	stealCount.getAndAdd(w.totalSteals + w.nsteals);
1364	>	int idx = w.poolIndex;
1365	>	lock.lock();
1366	>	try {                           // remove record from array
1367	>	WorkQueue[] ws = workQueues;
1368	>	if (ws != null && idx >= 0 && idx < ws.length && ws[idx] == w)
1369	>	ws[idx] = null;
1370	>	} finally {
1371	>	lock.unlock();
1372		}
732	–	if (eventCount != ec)
733	–	break;
734	–	h = eventWaiters;
1373		}
736	–	}
1374
1375	<	/**
1376	<	* Tries to advance eventCount and releases waiters. Called only
1377	<	* from workers.
1378	<	*/
1379	<	final void signalWork() {
1380	<	int c; // try to increment event count -- CAS failure OK
1381	<	UNSAFE.compareAndSwapInt(this, eventCountOffset, c = eventCount, c+1);
1382	<	if (eventWaiters != 0L)
1383	<	releaseEventWaiters();
1375	>	long c;                             // adjust ctl counts
1376	>	do {} while (!U.compareAndSwapLong
1377	>	(this, CTL, c = ctl, (((c - AC_UNIT) & AC_MASK) |
1378	>	((c - TC_UNIT) & TC_MASK) |
1379	>	(c & ~(AC_MASK|TC_MASK)))));
1380	>
1381	>	if (!tryTerminate(false, false) && w != null) {
1382	>	w.cancelAll();                  // cancel remaining tasks
1383	>	if (w.array != null)            // suppress signal if never ran
1384	>	signalWork();               // wake up or create replacement
1385	>	if (ex == null)                 // help clean refs on way out
1386	>	ForkJoinTask.helpExpungeStaleExceptions();
1387	>	}
1388	>
1389	>	if (ex != null)                     // rethrow
1390	>	U.throwException(ex);
1391		}
1392
749	–	/**
750	–	* Adds the given worker to event queue and blocks until
751	–	* terminating or event count advances from the given value
752	–	*
753	–	* @param w the calling worker thread
754	–	* @param ec the count
755	–	*/
756	–	private void eventSync(ForkJoinWorkerThread w, int ec) {
757	–	long nh = (((long)ec) << EVENT_COUNT_SHIFT) | ((long)(w.poolIndex+1));
758	–	long h;
759	–	while ((runState < SHUTDOWN || !tryTerminate(false)) &&
760	–	(((int)((h = eventWaiters) & WAITER_ID_MASK)) == 0 ||
761	–	(int)(h >>> EVENT_COUNT_SHIFT) == ec) &&
762	–	eventCount == ec) {
763	–	if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
764	–	w.nextWaiter = h, nh)) {
765	–	awaitEvent(w, ec);
766	–	break;
767	–	}
768	–	}
769	–	}
1393
1394	<	/**
1395	<	* Blocks the given worker (that has already been entered as an
1396	<	* event waiter) until terminating or event count advances from
1397	<	* the given value. The oldest (first) waiter uses a timed wait to
1398	<	* occasionally one-by-one shrink the number of workers (to a
1399	<	* minimum of one) if the pool has not been used for extended
1400	<	* periods.
1401	<	*
1402	<	* @param w the calling worker thread
1403	<	* @param ec the count
1404	<	*/
1405	<	private void awaitEvent(ForkJoinWorkerThread w, int ec) {
1406	<	while (eventCount == ec) {
1407	<	if (tryAccumulateStealCount(w)) { // transfer while idle
1408	<	boolean untimed = (w.nextWaiter != 0L ||
1409	<	(workerCounts & RUNNING_COUNT_MASK) <= 1);
1410	<	long startTime = untimed? 0 : System.nanoTime();
1411	<	Thread.interrupted();         // clear/ignore interrupt
1412	<	if (eventCount != ec || w.runState != 0 ||
1413	<	runState >= TERMINATING)  // recheck after clear
1414	<	break;
1415	<	if (untimed)
1416	<	LockSupport.park(w);
1417	<	else {
1418	<	LockSupport.parkNanos(w, SHRINK_RATE_NANOS);
1419	<	if (eventCount != ec || w.runState != 0 ||
1420	<	runState >= TERMINATING)
1421	<	break;
1422	<	if (System.nanoTime() - startTime >= SHRINK_RATE_NANOS)
1423	<	tryShutdownUnusedWorker(ec);
1394	>	// Submissions
1395	>
1396	>	/**
1397	>	* Unless shutting down, adds the given task to a submission queue
1398	>	* at submitter's current queue index (modulo submission
1399	>	* range). If no queue exists at the index, one is created.  If
1400	>	* the queue is busy, another index is randomly chosen. The
1401	>	* submitMask bounds the effective number of queues to the
1402	>	* (nearest poswer of two for) parallelism level.
1403	>	*
1404	>	* @param task the task. Caller must ensure non-null.
1405	>	*/
1406	>	private void doSubmit(ForkJoinTask<?> task) {
1407	>	Submitter s = submitters.get();
1408	>	for (int r = s.seed, m = submitMask;;) {
1409	>	WorkQueue[] ws; WorkQueue q;
1410	>	int k = r & m & SQMASK;          // use only even indices
1411	>	if (runState < 0 || (ws = workQueues) == null || ws.length <= k)
1412	>	throw new RejectedExecutionException(); // shutting down
1413	>	else if ((q = ws[k]) == null) {  // create new queue
1414	>	WorkQueue nq = new WorkQueue(this, null, SHARED_QUEUE);
1415	>	Mutex lock = this.lock;      // construct outside lock
1416	>	lock.lock();
1417	>	try {                        // recheck under lock
1418	>	int rs = runState;       // to update seq
1419	>	if (ws == workQueues && ws[k] == null) {
1420	>	ws[k] = nq;
1421	>	runState = ((rs & SHUTDOWN) | ((rs + 2) & ~SHUTDOWN));
1422	>	}
1423	>	} finally {
1424	>	lock.unlock();
1425		}
1426		}
1427	+	else if (q.trySharedPush(task)) {
1428	+	signalWork();
1429	+	return;
1430	+	}
1431	+	else if (m > 1) {                // move to a different index
1432	+	r ^= r << 13;                // same xorshift as WorkQueues
1433	+	r ^= r >>> 17;
1434	+	s.seed = r ^= r << 5;
1435	+	}
1436	+	else
1437	+	Thread.yield();              // yield if no alternatives
1438		}
1439		}
1440
1441	<	// Maintaining parallelism
1441	>	// Maintaining ctl counts
1442
1443		/**
1444	<	* Pushes worker onto the spare stack
1444	>	* Increments active count; mainly called upon return from blocking.
1445		*/
1446	<	final void pushSpare(ForkJoinWorkerThread w) {
1447	<	int ns = (++w.spareCount << SPARE_COUNT_SHIFT) | (w.poolIndex + 1);
1448	<	do {} while (!UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
814	<	w.nextSpare = spareWaiters,ns));
1446	>	final void incrementActiveCount() {
1447	>	long c;
1448	>	do {} while (!U.compareAndSwapLong(this, CTL, c = ctl, c + AC_UNIT));
1449		}
1450
1451		/**
1452	<	* Tries (once) to resume a spare if the number of running
819	<	* threads is less than target.
1452	>	* Tries to activate or create a worker if too few are active.
1453		*/
1454	<	private void tryResumeSpare() {
1455	<	int sw, id;
1456	<	ForkJoinWorkerThread[] ws = workers;
1457	<	int n = ws.length;
1458	<	ForkJoinWorkerThread w;
1459	<	if ((sw = spareWaiters) != 0 &&
1460	<	(id = (sw & SPARE_ID_MASK) - 1) >= 0 &&
1461	<	id < n && (w = ws[id]) != null &&
1462	<	(workerCounts & RUNNING_COUNT_MASK) < parallelism &&
1463	<	spareWaiters == sw &&
1464	<	UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
1465	<	sw, w.nextSpare)) {
1466	<	int c; // increment running count before resume
1467	<	do {} while (!UNSAFE.compareAndSwapInt
1468	<	(this, workerCountsOffset,
1469	<	c = workerCounts, c + ONE_RUNNING));
1470	<	if (w.tryUnsuspend())
1471	<	LockSupport.unpark(w);
1472	<	else   // back out if w was shutdown
1473	<	decrementWorkerCounts(ONE_RUNNING, 0);
1454	>	final void signalWork() {
1455	>	long c; int u;
1456	>	while ((u = (int)((c = ctl) >>> 32)) < 0) {     // too few active
1457	>	WorkQueue[] ws = workQueues; int e, i; WorkQueue w; Thread p;
1458	>	if ((e = (int)c) > 0) {                     // at least one waiting
1459	>	if (ws != null && (i = e & SMASK) < ws.length &&
1460	>	(w = ws[i]) != null && w.eventCount == (e | INT_SIGN)) {
1461	>	long nc = (((long)(w.nextWait & E_MASK)) |
1462	>	((long)(u + UAC_UNIT) << 32));
1463	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1464	>	w.eventCount = (e + E_SEQ) & E_MASK;
1465	>	if ((p = w.parker) != null)
1466	>	U.unpark(p);                // activate and release
1467	>	break;
1468	>	}
1469	>	}
1470	>	else
1471	>	break;
1472	>	}
1473	>	else if (e == 0 && (u & SHORT_SIGN) != 0) { // too few total
1474	>	long nc = (long)(((u + UTC_UNIT) & UTC_MASK) |
1475	>	((u + UAC_UNIT) & UAC_MASK)) << 32;
1476	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1477	>	addWorker();
1478	>	break;
1479	>	}
1480	>	}
1481	>	else
1482	>	break;
1483		}
1484		}
1485
1486	+
1487	+	// Scanning for tasks
1488	+
1489		/**
1490	<	* Tries to increase the number of running workers if below target
846	<	* parallelism: If a spare exists tries to resume it via
847	<	* tryResumeSpare.  Otherwise, if not enough total workers or all
848	<	* existing workers are busy, adds a new worker. In all cases also
849	<	* helps wake up releasable workers waiting for work.
1490	>	* Top-level runloop for workers, called by ForkJoinWorkerThread.run.
1491		*/
1492	<	private void helpMaintainParallelism() {
1493	<	int pc = parallelism;
1494	<	int wc, rs, tc;
1495	<	while (((wc = workerCounts) & RUNNING_COUNT_MASK) < pc &&
1496	<	(rs = runState) < TERMINATING) {
1497	<	if (spareWaiters != 0)
1498	<	tryResumeSpare();
1499	<	else if ((tc = wc >>> TOTAL_COUNT_SHIFT) >= MAX_WORKERS ||
1500	<	(tc >= pc && (rs & ACTIVE_COUNT_MASK) != tc))
1501	<	break;   // enough total
1502	<	else if (runState == rs && workerCounts == wc &&
1503	<	UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
1504	<	wc + (ONE_RUNNING|ONE_TOTAL))) {
1505	<	ForkJoinWorkerThread w = null;
1506	<	try {
1507	<	w = factory.newThread(this);
1508	<	} finally { // adjust on null or exceptional factory return
1509	<	if (w == null) {
1510	<	decrementWorkerCounts(ONE_RUNNING, ONE_TOTAL);
1511	<	tryTerminate(false); // handle failure during shutdown
1492	>	final void runWorker(WorkQueue w) {
1493	>	w.growArray(false);         // initialize queue array in this thread
1494	>	do {} while (w.runTask(scan(w)));
1495	>	}
1496	>
1497	>	/**
1498	>	* Scans for and, if found, returns one task, else possibly
1499	>	* inactivates the worker. This method operates on single reads of
1500	>	* volatile state and is designed to be re-invoked continuously,
1501	>	* in part because it returns upon detecting inconsistencies,
1502	>	* contention, or state changes that indicate possible success on
1503	>	* re-invocation.
1504	>	*
1505	>	* The scan searches for tasks across a random permutation of
1506	>	* queues (starting at a random index and stepping by a random
1507	>	* relative prime, checking each at least once).  The scan
1508	>	* terminates upon either finding a non-empty queue, or completing
1509	>	* the sweep. If the worker is not inactivated, it takes and
1510	>	* returns a task from this queue.  On failure to find a task, we
1511	>	* take one of the following actions, after which the caller will
1512	>	* retry calling this method unless terminated.
1513	>	*
1514	>	* * If pool is terminating, terminate the worker.
1515	>	*
1516	>	* * If not a complete sweep, try to release a waiting worker.  If
1517	>	* the scan terminated because the worker is inactivated, then the
1518	>	* released worker will often be the calling worker, and it can
1519	>	* succeed obtaining a task on the next call. Or maybe it is
1520	>	* another worker, but with same net effect. Releasing in other
1521	>	* cases as well ensures that we have enough workers running.
1522	>	*
1523	>	* * If not already enqueued, try to inactivate and enqueue the
1524	>	* worker on wait queue. Or, if inactivating has caused the pool
1525	>	* to be quiescent, relay to idleAwaitWork to check for
1526	>	* termination and possibly shrink pool.
1527	>	*
1528	>	* * If already inactive, and the caller has run a task since the
1529	>	* last empty scan, return (to allow rescan) unless others are
1530	>	* also inactivated.  Field WorkQueue.rescans counts down on each
1531	>	* scan to ensure eventual inactivation and blocking.
1532	>	*
1533	>	* * If already enqueued and none of the above apply, park
1534	>	* awaiting signal,
1535	>	*
1536	>	* @param w the worker (via its WorkQueue)
1537	>	* @return a task or null of none found
1538	>	*/
1539	>	private final ForkJoinTask<?> scan(WorkQueue w) {
1540	>	WorkQueue[] ws;                       // first update random seed
1541	>	int r = w.seed; r ^= r << 13; r ^= r >>> 17; w.seed = r ^= r << 5;
1542	>	int rs = runState, m;                 // volatile read order matters
1543	>	if ((ws = workQueues) != null && (m = ws.length - 1) > 0) {
1544	>	int ec = w.eventCount;            // ec is negative if inactive
1545	>	int step = (r >>> 16) | 1;        // relative prime
1546	>	for (int j = (m + 1) << 2; ; r += step) {
1547	>	WorkQueue q; ForkJoinTask<?> t; ForkJoinTask<?>[] a; int b;
1548	>	if ((q = ws[r & m]) != null && (b = q.base) - q.top < 0 &&
1549	>	(a = q.array) != null) {  // probably nonempty
1550	>	int i = (((a.length - 1) & b) << ASHIFT) + ABASE;
1551	>	t = (ForkJoinTask<?>)U.getObjectVolatile(a, i);
1552	>	if (q.base == b && ec >= 0 && t != null &&
1553	>	U.compareAndSwapObject(a, i, t, null)) {
1554	>	q.base = b + 1;       // specialization of pollAt
1555	>	return t;
1556	>	}
1557	>	else if ((t != null || b + 1 != q.top) &&
1558	>	(ec < 0 || j <= m)) {
1559	>	rs = 0;               // mark scan as imcomplete
1560	>	break;                // caller can retry after release
1561		}
1562		}
1563	<	if (w == null)
874	<	break;
875	<	w.start(recordWorker(w), ueh);
876	<	if ((workerCounts >>> TOTAL_COUNT_SHIFT) >= pc) {
877	<	int c; // advance event count
878	<	UNSAFE.compareAndSwapInt(this, eventCountOffset,
879	<	c = eventCount, c+1);
880	<	break; // add at most one unless total below target
881	<	}
882	<	}
883	<	}
884	<	if (eventWaiters != 0L)
885	<	releaseEventWaiters();
886	<	}
887	<
888	<	/**
889	<	* Callback from the oldest waiter in awaitEvent waking up after a
890	<	* period of non-use. If all workers are idle, tries (once) to
891	<	* shutdown an event waiter or a spare, if one exists. Note that
892	<	* we don't need CAS or locks here because the method is called
893	<	* only from one thread occasionally waking (and even misfires are
894	<	* OK). Note that until the shutdown worker fully terminates,
895	<	* workerCounts will overestimate total count, which is tolerable.
896	<	*
897	<	* @param ec the event count waited on by caller (to abort
898	<	* attempt if count has since changed).
899	<	*/
900	<	private void tryShutdownUnusedWorker(int ec) {
901	<	if (runState == 0 && eventCount == ec) { // only trigger if all idle
902	<	ForkJoinWorkerThread[] ws = workers;
903	<	int n = ws.length;
904	<	ForkJoinWorkerThread w = null;
905	<	boolean shutdown = false;
906	<	int sw;
907	<	long h;
908	<	if ((sw = spareWaiters) != 0) { // prefer killing spares
909	<	int id = (sw & SPARE_ID_MASK) - 1;
910	<	if (id >= 0 && id < n && (w = ws[id]) != null &&
911	<	UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
912	<	sw, w.nextSpare))
913	<	shutdown = true;
914	<	}
915	<	else if ((h = eventWaiters) != 0L) {
916	<	long nh;
917	<	int id = ((int)(h & WAITER_ID_MASK)) - 1;
918	<	if (id >= 0 && id < n && (w = ws[id]) != null &&
919	<	(nh = w.nextWaiter) != 0L && // keep at least one worker
920	<	UNSAFE.compareAndSwapLong(this, eventWaitersOffset, h, nh))
921	<	shutdown = true;
922	<	}
923	<	if (w != null && shutdown) {
924	<	w.shutdown();
925	<	LockSupport.unpark(w);
926	<	}
927	<	}
928	<	releaseEventWaiters(); // in case of interference
929	<	}
930	<
931	<	/**
932	<	* Callback from workers invoked upon each top-level action (i.e.,
933	<	* stealing a task or taking a submission and running it).
934	<	* Performs one or more of the following:
935	<	*
936	<	* 1. If the worker is active and either did not run a task
937	<	*    or there are too many workers, try to set its active status
938	<	*    to inactive and update activeCount. On contention, we may
939	<	*    try again in this or a subsequent call.
940	<	*
941	<	* 2. If not enough total workers, help create some.
942	<	*
943	<	* 3. If there are too many running workers, suspend this worker
944	<	*    (first forcing inactive if necessary).  If it is not needed,
945	<	*    it may be shutdown while suspended (via
946	<	*    tryShutdownUnusedWorker).  Otherwise, upon resume it
947	<	*    rechecks running thread count and need for event sync.
948	<	*
949	<	* 4. If worker did not run a task, await the next task event via
950	<	*    eventSync if necessary (first forcing inactivation), upon
951	<	*    which the worker may be shutdown via
952	<	*    tryShutdownUnusedWorker.  Otherwise, help release any
953	<	*    existing event waiters that are now releasable,
954	<	*
955	<	* @param w the worker
956	<	* @param ran true if worker ran a task since last call to this method
957	<	*/
958	<	final void preStep(ForkJoinWorkerThread w, boolean ran) {
959	<	int wec = w.lastEventCount;
960	<	boolean active = w.active;
961	<	boolean inactivate = false;
962	<	int pc = parallelism;
963	<	int rs;
964	<	while (w.runState == 0 && (rs = runState) < TERMINATING) {
965	<	if ((inactivate || (active && (rs & ACTIVE_COUNT_MASK) >= pc)) &&
966	<	UNSAFE.compareAndSwapInt(this, runStateOffset, rs, rs - 1))
967	<	inactivate = active = w.active = false;
968	<	int wc = workerCounts;
969	<	if ((wc & RUNNING_COUNT_MASK) > pc) {
970	<	if (!(inactivate |= active) && // must inactivate to suspend
971	<	workerCounts == wc &&      // try to suspend as spare
972	<	UNSAFE.compareAndSwapInt(this, workerCountsOffset,
973	<	wc, wc - ONE_RUNNING))
974	<	w.suspendAsSpare();
975	<	}
976	<	else if ((wc >>> TOTAL_COUNT_SHIFT) < pc)
977	<	helpMaintainParallelism();     // not enough workers
978	<	else if (!ran) {
979	<	long h = eventWaiters;
980	<	int ec = eventCount;
981	<	if (h != 0L && (int)(h >>> EVENT_COUNT_SHIFT) != ec)
982	<	releaseEventWaiters();     // release others before waiting
983	<	else if (ec != wec) {
984	<	w.lastEventCount = ec;     // no need to wait
1563	>	if (--j < 0)
1564		break;
1565	+	}
1566	+	long c = ctl; int e = (int)c, a = (int)(c >> AC_SHIFT), nr, ns;
1567	+	if (e < 0)                        // decode ctl on empty scan
1568	+	w.runState = -1;              // pool is terminating
1569	+	else if (rs == 0 || rs != runState) { // incomplete scan
1570	+	WorkQueue v; Thread p;        // try to release a waiter
1571	+	if (e > 0 && a < 0 && w.eventCount == ec &&
1572	+	(v = ws[e & m]) != null && v.eventCount == (e | INT_SIGN)) {
1573	+	long nc = ((long)(v.nextWait & E_MASK) |
1574	+	((c + AC_UNIT) & (AC_MASK|TC_MASK)));
1575	+	if (ctl == c && U.compareAndSwapLong(this, CTL, c, nc)) {
1576	+	v.eventCount = (e + E_SEQ) & E_MASK;
1577	+	if ((p = v.parker) != null)
1578	+	U.unpark(p);
1579	+	}
1580	+	}
1581	+	}
1582	+	else if (ec >= 0) {               // try to enqueue/inactivate
1583	+	long nc = (long)ec | ((c - AC_UNIT) & (AC_MASK|TC_MASK));
1584	+	w.nextWait = e;
1585	+	w.eventCount = ec | INT_SIGN; // mark as inactive
1586	+	if (ctl != c || !U.compareAndSwapLong(this, CTL, c, nc))
1587	+	w.eventCount = ec;        // unmark on CAS failure
1588	+	else {
1589	+	if ((ns = w.nsteals) != 0) {
1590	+	w.nsteals = 0;        // set rescans if ran task
1591	+	w.rescans = (a > 0) ? 0 : a + parallelism;
1592	+	w.totalSteals += ns;
1593	+	}
1594	+	if (a == 1 - parallelism) // quiescent
1595	+	idleAwaitWork(w, nc, c);
1596	+	}
1597	+	}
1598	+	else if (w.eventCount < 0) {      // already queued
1599	+	if ((nr = w.rescans) > 0) {   // continue rescanning
1600	+	int ac = a + parallelism;
1601	+	if (((w.rescans = (ac < nr) ? ac : nr - 1) & 3) == 0)
1602	+	Thread.yield();       // yield before block
1603	+	}
1604	+	else {
1605	+	Thread.interrupted();     // clear status
1606	+	Thread wt = Thread.currentThread();
1607	+	U.putObject(wt, PARKBLOCKER, this);
1608	+	w.parker = wt;            // emulate LockSupport.park
1609	+	if (w.eventCount < 0)     // recheck
1610	+	U.park(false, 0L);
1611	+	w.parker = null;
1612	+	U.putObject(wt, PARKBLOCKER, null);
1613		}
987	–	else if (!(inactivate |= active))
988	–	eventSync(w, wec);         // must inactivate before sync
1614		}
990	–	else
991	–	break;
1615		}
1616	+	return null;
1617		}
1618
1619		/**
1620	<	* Helps and/or blocks awaiting join of the given task.
1621	<	* See above for explanation.
1622	<	*
1623	<	* @param joinMe the task to join
1624	<	* @param worker the current worker thread
1625	<	*/
1626	<	final void awaitJoin(ForkJoinTask<?> joinMe, ForkJoinWorkerThread worker) {
1627	<	int retries = 2 + (parallelism >> 2); // #helpJoins before blocking
1628	<	while (joinMe.status >= 0) {
1629	<	int wc;
1630	<	worker.helpJoinTask(joinMe);
1631	<	if (joinMe.status < 0)
1632	<	break;
1633	<	else if (retries > 0)
1634	<	--retries;
1635	<	else if (((wc = workerCounts) & RUNNING_COUNT_MASK) != 0 &&
1636	<	UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1637	<	wc, wc - ONE_RUNNING)) {
1638	<	int stat, c; long h;
1639	<	while ((stat = joinMe.status) >= 0 &&
1640	<	(h = eventWaiters) != 0L && // help release others
1641	<	(int)(h >>> EVENT_COUNT_SHIFT) != eventCount)
1642	<	releaseEventWaiters();
1643	<	if (stat >= 0 &&
1644	<	((workerCounts & RUNNING_COUNT_MASK) == 0 ||
1645	<	(stat =
1646	<	joinMe.internalAwaitDone(JOIN_TIMEOUT_MILLIS)) >= 0))
1647	<	helpMaintainParallelism(); // timeout or no running workers
1648	<	do {} while (!UNSAFE.compareAndSwapInt
1649	<	(this, workerCountsOffset,
1650	<	c = workerCounts, c + ONE_RUNNING));
1651	<	if (stat < 0)
1652	<	break;   // else restart
1620	>	* If inactivating worker w has caused the pool to become
1621	>	* quiescent, checks for pool termination, and, so long as this is
1622	>	* not the only worker, waits for event for up to SHRINK_RATE
1623	>	* nanosecs.  On timeout, if ctl has not changed, terminates the
1624	>	* worker, which will in turn wake up another worker to possibly
1625	>	* repeat this process.
1626	>	*
1627	>	* @param w the calling worker
1628	>	* @param currentCtl the ctl value triggering possible quiescence
1629	>	* @param prevCtl the ctl value to restore if thread is terminated
1630	>	*/
1631	>	private void idleAwaitWork(WorkQueue w, long currentCtl, long prevCtl) {
1632	>	if (w.eventCount < 0 && !tryTerminate(false, false) &&
1633	>	(int)prevCtl != 0 && ctl == currentCtl) {
1634	>	Thread wt = Thread.currentThread();
1635	>	Thread.yield();            // yield before block
1636	>	while (ctl == currentCtl) {
1637	>	long startTime = System.nanoTime();
1638	>	Thread.interrupted();  // timed variant of version in scan()
1639	>	U.putObject(wt, PARKBLOCKER, this);
1640	>	w.parker = wt;
1641	>	if (ctl == currentCtl)
1642	>	U.park(false, SHRINK_RATE);
1643	>	w.parker = null;
1644	>	U.putObject(wt, PARKBLOCKER, null);
1645	>	if (ctl != currentCtl)
1646	>	break;
1647	>	if (System.nanoTime() - startTime >= SHRINK_TIMEOUT &&
1648	>	U.compareAndSwapLong(this, CTL, currentCtl, prevCtl)) {
1649	>	w.eventCount = (w.eventCount + E_SEQ) | E_MASK;
1650	>	w.runState = -1;   // shrink
1651	>	break;
1652	>	}
1653		}
1654		}
1655		}
1656
1657		/**
1658	<	* Same idea as awaitJoin, but no helping, retries, or timeouts.
1659	<	*/
1660	<	final void awaitBlocker(ManagedBlocker blocker)
1661	<	throws InterruptedException {
1662	<	while (!blocker.isReleasable()) {
1663	<	int wc = workerCounts;
1664	<	if ((wc & RUNNING_COUNT_MASK) != 0 &&
1665	<	UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1666	<	wc, wc - ONE_RUNNING)) {
1667	<	try {
1668	<	while (!blocker.isReleasable()) {
1669	<	long h = eventWaiters;
1670	<	if (h != 0L &&
1671	<	(int)(h >>> EVENT_COUNT_SHIFT) != eventCount)
1672	<	releaseEventWaiters();
1673	<	else if ((workerCounts & RUNNING_COUNT_MASK) == 0 &&
1674	<	runState < TERMINATING)
1675	<	helpMaintainParallelism();
1676	<	else if (blocker.block())
1658	>	* Tries to locate and execute tasks for a stealer of the given
1659	>	* task, or in turn one of its stealers, Traces currentSteal ->
1660	>	* currentJoin links looking for a thread working on a descendant
1661	>	* of the given task and with a non-empty queue to steal back and
1662	>	* execute tasks from. The first call to this method upon a
1663	>	* waiting join will often entail scanning/search, (which is OK
1664	>	* because the joiner has nothing better to do), but this method
1665	>	* leaves hints in workers to speed up subsequent calls. The
1666	>	* implementation is very branchy to cope with potential
1667	>	* inconsistencies or loops encountering chains that are stale,
1668	>	* unknown, or so long that they are likely cyclic.  All of these
1669	>	* cases are dealt with by just retrying by caller.
1670	>	*
1671	>	* @param joiner the joining worker
1672	>	* @param task the task to join
1673	>	* @return true if found or ran a task (and so is immediately retryable)
1674	>	*/
1675	>	private boolean tryHelpStealer(WorkQueue joiner, ForkJoinTask<?> task) {
1676	>	WorkQueue[] ws;
1677	>	int m, depth = MAX_HELP;                // remaining chain depth
1678	>	boolean progress = false;
1679	>	if ((ws = workQueues) != null && (m = ws.length - 1) > 0 &&
1680	>	task.status >= 0) {
1681	>	ForkJoinTask<?> subtask = task;     // current target
1682	>	outer: for (WorkQueue j = joiner;;) {
1683	>	WorkQueue stealer = null;       // find stealer of subtask
1684	>	WorkQueue v = ws[j.stealHint & m]; // try hint
1685	>	if (v != null && v.currentSteal == subtask)
1686	>	stealer = v;
1687	>	else {                          // scan
1688	>	for (int i = 1; i <= m; i += 2) {
1689	>	if ((v = ws[i]) != null && v.currentSteal == subtask &&
1690	>	v != joiner) {
1691	>	stealer = v;
1692	>	j.stealHint = i;    // save hint
1693		break;
1694	+	}
1695	+	}
1696	+	if (stealer == null)
1697	+	break;
1698	+	}
1699	+
1700	+	for (WorkQueue q = stealer;;) { // try to help stealer
1701	+	ForkJoinTask[] a; ForkJoinTask<?> t; int b;
1702	+	if (task.status < 0)
1703	+	break outer;
1704	+	if ((b = q.base) - q.top < 0 && (a = q.array) != null) {
1705	+	progress = true;
1706	+	int i = (((a.length - 1) & b) << ASHIFT) + ABASE;
1707	+	t = (ForkJoinTask<?>)U.getObjectVolatile(a, i);
1708	+	if (subtask.status < 0) // must recheck before taking
1709	+	break outer;
1710	+	if (t != null &&
1711	+	q.base == b &&
1712	+	U.compareAndSwapObject(a, i, t, null)) {
1713	+	q.base = b + 1;
1714	+	joiner.runSubtask(t);
1715	+	}
1716	+	else if (q.base == b)
1717	+	break outer;        // possibly stalled
1718	+	}
1719	+	else {                      // descend
1720	+	ForkJoinTask<?> next = stealer.currentJoin;
1721	+	if (--depth <= 0 || subtask.status < 0 ||
1722	+	next == null || next == subtask)
1723	+	break outer;        // stale, dead-end, or cyclic
1724	+	subtask = next;
1725	+	j = stealer;
1726	+	break;
1727		}
1055	–	} finally {
1056	–	int c;
1057	–	do {} while (!UNSAFE.compareAndSwapInt
1058	–	(this, workerCountsOffset,
1059	–	c = workerCounts, c + ONE_RUNNING));
1728		}
1061	–	break;
1729		}
1730		}
1731	+	return progress;
1732		}
1733
1734		/**
1735	<	* Possibly initiates and/or completes termination.
1735	>	* If task is at base of some steal queue, steals and executes it.
1736		*
1737	<	* @param now if true, unconditionally terminate, else only
1738	<	* if shutdown and empty queue and no active workers
1071	<	* @return true if now terminating or terminated
1737	>	* @param joiner the joining worker
1738	>	* @param task the task
1739		*/
1740	<	private boolean tryTerminate(boolean now) {
1741	<	if (now)
1742	<	advanceRunLevel(SHUTDOWN); // ensure at least SHUTDOWN
1743	<	else if (runState < SHUTDOWN ||
1744	<	!submissionQueue.isEmpty() ||
1745	<	(runState & ACTIVE_COUNT_MASK) != 0)
1746	<	return false;
1747	<
1748	<	if (advanceRunLevel(TERMINATING))
1749	<	startTerminating();
1740	>	private void tryPollForAndExec(WorkQueue joiner, ForkJoinTask<?> task) {
1741	>	WorkQueue[] ws;
1742	>	if ((ws = workQueues) != null) {
1743	>	for (int j = 1; j < ws.length && task.status >= 0; j += 2) {
1744	>	WorkQueue q = ws[j];
1745	>	if (q != null && q.pollFor(task)) {
1746	>	joiner.runSubtask(task);
1747	>	break;
1748	>	}
1749	>	}
1750	>	}
1751	>	}
1752
1753	<	// Finish now if all threads terminated; else in some subsequent call
1754	<	if ((workerCounts >>> TOTAL_COUNT_SHIFT) == 0) {
1755	<	advanceRunLevel(TERMINATED);
1756	<	termination.arrive();
1757	<	}
1758	<	return true;
1753	>	/**
1754	>	* Tries to decrement active count (sometimes implicitly) and
1755	>	* possibly release or create a compensating worker in preparation
1756	>	* for blocking. Fails on contention or termination. Otherwise,
1757	>	* adds a new thread if no idle workers are available and either
1758	>	* pool would become completely starved or: (at least half
1759	>	* starved, and fewer than 50% spares exist, and there is at least
1760	>	* one task apparently available). Even though the availablity
1761	>	* check requires a full scan, it is worthwhile in reducing false
1762	>	* alarms.
1763	>	*
1764	>	* @param task if nonnull, a task being waited for
1765	>	* @param blocker if nonnull, a blocker being waited for
1766	>	* @return true if the caller can block, else should recheck and retry
1767	>	*/
1768	>	final boolean tryCompensate(ForkJoinTask<?> task, ManagedBlocker blocker) {
1769	>	int pc = parallelism, e;
1770	>	long c = ctl;
1771	>	WorkQueue[] ws = workQueues;
1772	>	if ((e = (int)c) >= 0 && ws != null) {
1773	>	int u, a, ac, hc;
1774	>	int tc = (short)((u = (int)(c >>> 32)) >>> UTC_SHIFT) + pc;
1775	>	boolean replace = false;
1776	>	if ((a = u >> UAC_SHIFT) <= 0) {
1777	>	if ((ac = a + pc) <= 1)
1778	>	replace = true;
1779	>	else if ((e > 0 || (task != null &&
1780	>	ac <= (hc = pc >>> 1) && tc < pc + hc))) {
1781	>	WorkQueue w;
1782	>	for (int j = 0; j < ws.length; ++j) {
1783	>	if ((w = ws[j]) != null && !w.isEmpty()) {
1784	>	replace = true;
1785	>	break;   // in compensation range and tasks available
1786	>	}
1787	>	}
1788	>	}
1789	>	}
1790	>	if ((task == null || task.status >= 0) && // recheck need to block
1791	>	(blocker == null || !blocker.isReleasable()) && ctl == c) {
1792	>	if (!replace) {          // no compensation
1793	>	long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK);
1794	>	if (U.compareAndSwapLong(this, CTL, c, nc))
1795	>	return true;
1796	>	}
1797	>	else if (e != 0) {       // release an idle worker
1798	>	WorkQueue w; Thread p; int i;
1799	>	if ((i = e & SMASK) < ws.length && (w = ws[i]) != null) {
1800	>	long nc = ((long)(w.nextWait & E_MASK) |
1801	>	(c & (AC_MASK|TC_MASK)));
1802	>	if (w.eventCount == (e | INT_SIGN) &&
1803	>	U.compareAndSwapLong(this, CTL, c, nc)) {
1804	>	w.eventCount = (e + E_SEQ) & E_MASK;
1805	>	if ((p = w.parker) != null)
1806	>	U.unpark(p);
1807	>	return true;
1808	>	}
1809	>	}
1810	>	}
1811	>	else if (tc < MAX_CAP) { // create replacement
1812	>	long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK);
1813	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1814	>	addWorker();
1815	>	return true;
1816	>	}
1817	>	}
1818	>	}
1819	>	}
1820	>	return false;
1821		}
1822
1823		/**
1824	<	* Actions on transition to TERMINATING
1825	<	*
1826	<	* Runs up to four passes through workers: (0) shutting down each
1827	<	* (without waking up if parked) to quickly spread notifications
1828	<	* without unnecessary bouncing around event queues etc (1) wake
1829	<	* up and help cancel tasks (2) interrupt (3) mop up races with
1830	<	* interrupted workers
1831	<	*/
1832	<	private void startTerminating() {
1833	<	cancelSubmissions();
1834	<	for (int passes = 0; passes < 4 && workerCounts != 0; ++passes) {
1835	<	int c; // advance event count
1836	<	UNSAFE.compareAndSwapInt(this, eventCountOffset,
1837	<	c = eventCount, c+1);
1838	<	eventWaiters = 0L; // clobber lists
1839	<	spareWaiters = 0;
1840	<	for (ForkJoinWorkerThread w : workers) {
1841	<	if (w != null) {
1842	<	w.shutdown();
1843	<	if (passes > 0 && !w.isTerminated()) {
1844	<	w.cancelTasks();
1845	<	LockSupport.unpark(w);
1846	<	if (passes > 1) {
1847	<	try {
1848	<	w.interrupt();
1849	<	} catch (SecurityException ignore) {
1824	>	* Helps and/or blocks until the given task is done
1825	>	*
1826	>	* @param joiner the joining worker
1827	>	* @param task the task
1828	>	* @return task status on exit
1829	>	*/
1830	>	final int awaitJoin(WorkQueue joiner, ForkJoinTask<?> task) {
1831	>	ForkJoinTask<?> prevJoin = joiner.currentJoin;
1832	>	joiner.currentJoin = task;
1833	>	long startTime = 0L;
1834	>	for (int k = 0, s; ; ++k) {
1835	>	if ((joiner.isEmpty() ?                  // try to help
1836	>	!tryHelpStealer(joiner, task) :
1837	>	!joiner.tryRemoveAndExec(task))) {
1838	>	if (k == 0) {
1839	>	startTime = System.nanoTime();
1840	>	tryPollForAndExec(joiner, task); // check uncommon case
1841	>	}
1842	>	else if ((k & (MAX_HELP - 1)) == 0 &&
1843	>	System.nanoTime() - startTime >= COMPENSATION_DELAY &&
1844	>	tryCompensate(task, null)) {
1845	>	if (task.trySetSignal() && task.status >= 0) {
1846	>	synchronized (task) {
1847	>	if (task.status >= 0) {
1848	>	try {                // see ForkJoinTask
1849	>	task.wait();     //  for explanation
1850	>	} catch (InterruptedException ie) {
1851	>	}
1852		}
1853	+	else
1854	+	task.notifyAll();
1855		}
1856		}
1857	+	long c;                          // re-activate
1858	+	do {} while (!U.compareAndSwapLong
1859	+	(this, CTL, c = ctl, c + AC_UNIT));
1860		}
1861		}
1862	+	if ((s = task.status) < 0) {
1863	+	joiner.currentJoin = prevJoin;
1864	+	return s;
1865	+	}
1866	+	else if ((k & (MAX_HELP - 1)) == MAX_HELP >>> 1)
1867	+	Thread.yield();                     // for politeness
1868		}
1869		}
1870
1871		/**
1872	<	* Clears out and cancels submissions, ignoring exceptions.
1872	>	* Stripped-down variant of awaitJoin used by timed joins. Tries
1873	>	* to help join only while there is continuous progress. (Caller
1874	>	* will then enter a timed wait.)
1875	>	*
1876	>	* @param joiner the joining worker
1877	>	* @param task the task
1878	>	* @return task status on exit
1879		*/
1880	<	private void cancelSubmissions() {
1881	<	ForkJoinTask<?> task;
1882	<	while ((task = submissionQueue.poll()) != null) {
1883	<	try {
1884	<	task.cancel(false);
1885	<	} catch (Throwable ignore) {
1880	>	final int helpJoinOnce(WorkQueue joiner, ForkJoinTask<?> task) {
1881	>	int s;
1882	>	while ((s = task.status) >= 0 &&
1883	>	(joiner.isEmpty() ?
1884	>	tryHelpStealer(joiner, task) :
1885	>	joiner.tryRemoveAndExec(task)))
1886	>	;
1887	>	return s;
1888	>	}
1889	>
1890	>	/**
1891	>	* Returns a (probably) non-empty steal queue, if one is found
1892	>	* during a random, then cyclic scan, else null.  This method must
1893	>	* be retried by caller if, by the time it tries to use the queue,
1894	>	* it is empty.
1895	>	*/
1896	>	private WorkQueue findNonEmptyStealQueue(WorkQueue w) {
1897	>	// Similar to loop in scan(), but ignoring submissions
1898	>	int r = w.seed; r ^= r << 13; r ^= r >>> 17; w.seed = r ^= r << 5;
1899	>	int step = (r >>> 16) | 1;
1900	>	for (WorkQueue[] ws;;) {
1901	>	int rs = runState, m;
1902	>	if ((ws = workQueues) == null || (m = ws.length - 1) < 1)
1903	>	return null;
1904	>	for (int j = (m + 1) << 2; ; r += step) {
1905	>	WorkQueue q = ws[((r << 1) | 1) & m];
1906	>	if (q != null && !q.isEmpty())
1907	>	return q;
1908	>	else if (--j < 0) {
1909	>	if (runState == rs)
1910	>	return null;
1911	>	break;
1912	>	}
1913		}
1914		}
1915		}
1916
1140	–	// misc support for ForkJoinWorkerThread
1141	–
1917		/**
1918	<	* Returns pool number.
1919	<	*/
1920	<	final int getPoolNumber() {
1921	<	return poolNumber;
1918	>	* Runs tasks until {@code isQuiescent()}. We piggyback on
1919	>	* active count ctl maintenance, but rather than blocking
1920	>	* when tasks cannot be found, we rescan until all others cannot
1921	>	* find tasks either.
1922	>	*/
1923	>	final void helpQuiescePool(WorkQueue w) {
1924	>	for (boolean active = true;;) {
1925	>	if (w.base - w.top < 0)
1926	>	w.runLocalTasks();  // exhaust local queue
1927	>	WorkQueue q = findNonEmptyStealQueue(w);
1928	>	if (q != null) {
1929	>	ForkJoinTask<?> t; int b;
1930	>	if (!active) {      // re-establish active count
1931	>	long c;
1932	>	active = true;
1933	>	do {} while (!U.compareAndSwapLong
1934	>	(this, CTL, c = ctl, c + AC_UNIT));
1935	>	}
1936	>	if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null)
1937	>	w.runSubtask(t);
1938	>	}
1939	>	else {
1940	>	long c;
1941	>	if (active) {       // decrement active count without queuing
1942	>	active = false;
1943	>	do {} while (!U.compareAndSwapLong
1944	>	(this, CTL, c = ctl, c -= AC_UNIT));
1945	>	}
1946	>	else
1947	>	c = ctl;        // re-increment on exit
1948	>	if ((int)(c >> AC_SHIFT) + parallelism == 0) {
1949	>	do {} while (!U.compareAndSwapLong
1950	>	(this, CTL, c = ctl, c + AC_UNIT));
1951	>	break;
1952	>	}
1953	>	}
1954	>	}
1955		}
1956
1957		/**
1958	<	* Tries to accumulate steal count from a worker, clearing
1151	<	* the worker's value if successful.
1958	>	* Gets and removes a local or stolen task for the given worker.
1959		*
1960	<	* @return true if worker steal count now zero
1960	>	* @return a task, if available
1961		*/
1962	<	final boolean tryAccumulateStealCount(ForkJoinWorkerThread w) {
1963	<	int sc = w.stealCount;
1964	<	long c = stealCount;
1965	<	// CAS even if zero, for fence effects
1966	<	if (UNSAFE.compareAndSwapLong(this, stealCountOffset, c, c + sc)) {
1967	<	if (sc != 0)
1968	<	w.stealCount = 0;
1969	<	return true;
1962	>	final ForkJoinTask<?> nextTaskFor(WorkQueue w) {
1963	>	for (ForkJoinTask<?> t;;) {
1964	>	WorkQueue q; int b;
1965	>	if ((t = w.nextLocalTask()) != null)
1966	>	return t;
1967	>	if ((q = findNonEmptyStealQueue(w)) == null)
1968	>	return null;
1969	>	if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null)
1970	>	return t;
1971		}
1164	–	return sc == 0;
1972		}
1973
1974		/**
1975		* Returns the approximate (non-atomic) number of idle threads per
1976	<	* active thread.
1976	>	* active thread to offset steal queue size for method
1977	>	* ForkJoinTask.getSurplusQueuedTaskCount().
1978		*/
1979		final int idlePerActive() {
1980	<	int pc = parallelism; // use parallelism, not rc
1981	<	int ac = runState;    // no mask -- artificially boosts during shutdown
1982	<	// Use exact results for small values, saturate past 4
1983	<	return ((pc <= ac) ? 0 :
1984	<	(pc >>> 1 <= ac) ? 1 :
1985	<	(pc >>> 2 <= ac) ? 3 :
1986	<	pc >>> 3);
1980	>	// Approximate at powers of two for small values, saturate past 4
1981	>	int p = parallelism;
1982	>	int a = p + (int)(ctl >> AC_SHIFT);
1983	>	return (a > (p >>>= 1) ? 0 :
1984	>	a > (p >>>= 1) ? 1 :
1985	>	a > (p >>>= 1) ? 2 :
1986	>	a > (p >>>= 1) ? 4 :
1987	>	8);
1988	>	}
1989	>
1990	>	//  Termination
1991	>
1992	>	/**
1993	>	* Possibly initiates and/or completes termination.  The caller
1994	>	* triggering termination runs three passes through workQueues:
1995	>	* (0) Setting termination status, followed by wakeups of queued
1996	>	* workers; (1) cancelling all tasks; (2) interrupting lagging
1997	>	* threads (likely in external tasks, but possibly also blocked in
1998	>	* joins).  Each pass repeats previous steps because of potential
1999	>	* lagging thread creation.
2000	>	*
2001	>	* @param now if true, unconditionally terminate, else only
2002	>	* if no work and no active workers
2003	>	* @param enable if true, enable shutdown when next possible
2004	>	* @return true if now terminating or terminated
2005	>	*/
2006	>	private boolean tryTerminate(boolean now, boolean enable) {
2007	>	Mutex lock = this.lock;
2008	>	for (long c;;) {
2009	>	if (((c = ctl) & STOP_BIT) != 0) {      // already terminating
2010	>	if ((short)(c >>> TC_SHIFT) == -parallelism) {
2011	>	lock.lock();                    // don't need try/finally
2012	>	termination.signalAll();        // signal when 0 workers
2013	>	lock.unlock();
2014	>	}
2015	>	return true;
2016	>	}
2017	>	if (runState >= 0) {                    // not yet enabled
2018	>	if (!enable)
2019	>	return false;
2020	>	lock.lock();
2021	>	runState |= SHUTDOWN;
2022	>	lock.unlock();
2023	>	}
2024	>	if (!now) {                             // check if idle & no tasks
2025	>	if ((int)(c >> AC_SHIFT) != -parallelism ||
2026	>	hasQueuedSubmissions())
2027	>	return false;
2028	>	// Check for unqueued inactive workers. One pass suffices.
2029	>	WorkQueue[] ws = workQueues; WorkQueue w;
2030	>	if (ws != null) {
2031	>	for (int i = 1; i < ws.length; i += 2) {
2032	>	if ((w = ws[i]) != null && w.eventCount >= 0)
2033	>	return false;
2034	>	}
2035	>	}
2036	>	}
2037	>	if (U.compareAndSwapLong(this, CTL, c, c | STOP_BIT)) {
2038	>	for (int pass = 0; pass < 3; ++pass) {
2039	>	WorkQueue[] ws = workQueues;
2040	>	if (ws != null) {
2041	>	WorkQueue w;
2042	>	int n = ws.length;
2043	>	for (int i = 0; i < n; ++i) {
2044	>	if ((w = ws[i]) != null) {
2045	>	w.runState = -1;
2046	>	if (pass > 0) {
2047	>	w.cancelAll();
2048	>	if (pass > 1)
2049	>	w.interruptOwner();
2050	>	}
2051	>	}
2052	>	}
2053	>	// Wake up workers parked on event queue
2054	>	int i, e; long cc; Thread p;
2055	>	while ((e = (int)(cc = ctl) & E_MASK) != 0 &&
2056	>	(i = e & SMASK) < n &&
2057	>	(w = ws[i]) != null) {
2058	>	long nc = ((long)(w.nextWait & E_MASK) |
2059	>	((cc + AC_UNIT) & AC_MASK) |
2060	>	(cc & (TC_MASK|STOP_BIT)));
2061	>	if (w.eventCount == (e | INT_SIGN) &&
2062	>	U.compareAndSwapLong(this, CTL, cc, nc)) {
2063	>	w.eventCount = (e + E_SEQ) & E_MASK;
2064	>	w.runState = -1;
2065	>	if ((p = w.parker) != null)
2066	>	U.unpark(p);
2067	>	}
2068	>	}
2069	>	}
2070	>	}
2071	>	}
2072	>	}
2073		}
2074
2075	<	// Public and protected methods
2075	>	// Exported methods
2076
2077		// Constructors
2078
#	Line 1225 | Line 2119 | public class ForkJoinPool extends Abstra
2119		* use {@link #defaultForkJoinWorkerThreadFactory}.
2120		* @param handler the handler for internal worker threads that
2121		* terminate due to unrecoverable errors encountered while executing
2122	<	* tasks. For default value, use <code>null</code>.
2122	>	* tasks. For default value, use {@code null}.
2123		* @param asyncMode if true,
2124		* establishes local first-in-first-out scheduling mode for forked
2125		* tasks that are never joined. This mode may be more appropriate
2126		* than default locally stack-based mode in applications in which
2127		* worker threads only process event-style asynchronous tasks.
2128	<	* For default value, use <code>false</code>.
2128	>	* For default value, use {@code false}.
2129		* @throws IllegalArgumentException if parallelism less than or
2130		*         equal to zero, or greater than implementation limit
2131		* @throws NullPointerException if the factory is null
#	Line 1247 | Line 2141 | public class ForkJoinPool extends Abstra
2141		checkPermission();
2142		if (factory == null)
2143		throw new NullPointerException();
2144	<	if (parallelism <= 0 || parallelism > MAX_WORKERS)
2144	>	if (parallelism <= 0 || parallelism > MAX_CAP)
2145		throw new IllegalArgumentException();
2146		this.parallelism = parallelism;
2147		this.factory = factory;
2148		this.ueh = handler;
2149	<	this.locallyFifo = asyncMode;
2150	<	int arraySize = initialArraySizeFor(parallelism);
2151	<	this.workers = new ForkJoinWorkerThread[arraySize];
2152	<	this.submissionQueue = new LinkedTransferQueue<ForkJoinTask<?>>();
2153	<	this.workerLock = new ReentrantLock();
2154	<	this.termination = new Phaser(1);
2155	<	this.poolNumber = poolNumberGenerator.incrementAndGet();
2156	<	}
2157	<
2158	<	/**
2159	<	* Returns initial power of two size for workers array.
2160	<	* @param pc the initial parallelism level
2161	<	*/
2162	<	private static int initialArraySizeFor(int pc) {
2163	<	// If possible, initially allocate enough space for one spare
2164	<	int size = pc < MAX_WORKERS ? pc + 1 : MAX_WORKERS;
2165	<	// See Hackers Delight, sec 3.2. We know MAX_WORKERS < (1 >>> 16)
2166	<	size |= size >>> 1;
2167	<	size |= size >>> 2;
2168	<	size |= size >>> 4;
1275	<	size |= size >>> 8;
1276	<	return size + 1;
2149	>	this.localMode = asyncMode ? FIFO_QUEUE : LIFO_QUEUE;
2150	>	long np = (long)(-parallelism); // offset ctl counts
2151	>	this.ctl = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK);
2152	>	// Use nearest power 2 for workQueues size. See Hackers Delight sec 3.2.
2153	>	int n = parallelism - 1;
2154	>	n |= n >>> 1; n |= n >>> 2; n |= n >>> 4; n |= n >>> 8; n |= n >>> 16;
2155	>	int size = (n + 1) << 1;        // #slots = 2*#workers
2156	>	this.submitMask = size - 1;     // room for max # of submit queues
2157	>	this.workQueues = new WorkQueue[size];
2158	>	this.termination = (this.lock = new Mutex()).newCondition();
2159	>	this.stealCount = new AtomicLong();
2160	>	this.nextWorkerNumber = new AtomicInteger();
2161	>	int pn = poolNumberGenerator.incrementAndGet();
2162	>	StringBuilder sb = new StringBuilder("ForkJoinPool-");
2163	>	sb.append(Integer.toString(pn));
2164	>	sb.append("-worker-");
2165	>	this.workerNamePrefix = sb.toString();
2166	>	lock.lock();
2167	>	this.runState = 1;              // set init flag
2168	>	lock.unlock();
2169		}
2170
2171		// Execution methods
2172
2173		/**
1282	–	* Common code for execute, invoke and submit
1283	–	*/
1284	–	private <T> void doSubmit(ForkJoinTask<T> task) {
1285	–	if (task == null)
1286	–	throw new NullPointerException();
1287	–	if (runState >= SHUTDOWN)
1288	–	throw new RejectedExecutionException();
1289	–	submissionQueue.offer(task);
1290	–	int c; // try to increment event count -- CAS failure OK
1291	–	UNSAFE.compareAndSwapInt(this, eventCountOffset, c = eventCount, c+1);
1292	–	helpMaintainParallelism(); // create, start, or resume some workers
1293	–	}
1294	–
1295	–	/**
2174		* Performs the given task, returning its result upon completion.
2175	+	* If the computation encounters an unchecked Exception or Error,
2176	+	* it is rethrown as the outcome of this invocation.  Rethrown
2177	+	* exceptions behave in the same way as regular exceptions, but,
2178	+	* when possible, contain stack traces (as displayed for example
2179	+	* using {@code ex.printStackTrace()}) of both the current thread
2180	+	* as well as the thread actually encountering the exception;
2181	+	* minimally only the latter.
2182		*
2183		* @param task the task
2184		* @return the task's result
#	Line 1302 | Line 2187 | public class ForkJoinPool extends Abstra
2187		*         scheduled for execution
2188		*/
2189		public <T> T invoke(ForkJoinTask<T> task) {
2190	+	if (task == null)
2191	+	throw new NullPointerException();
2192		doSubmit(task);
2193		return task.join();
2194		}
#	Line 1315 | Line 2202 | public class ForkJoinPool extends Abstra
2202		*         scheduled for execution
2203		*/
2204		public void execute(ForkJoinTask<?> task) {
2205	+	if (task == null)
2206	+	throw new NullPointerException();
2207		doSubmit(task);
2208		}
2209
#	Line 1326 | Line 2215 | public class ForkJoinPool extends Abstra
2215		*         scheduled for execution
2216		*/
2217		public void execute(Runnable task) {
2218	+	if (task == null)
2219	+	throw new NullPointerException();
2220		ForkJoinTask<?> job;
2221		if (task instanceof ForkJoinTask<?>) // avoid re-wrap
2222		job = (ForkJoinTask<?>) task;
2223		else
2224	<	job = ForkJoinTask.adapt(task, null);
2224	>	job = new ForkJoinTask.AdaptedRunnableAction(task);
2225		doSubmit(job);
2226		}
2227
#	Line 1344 | Line 2235 | public class ForkJoinPool extends Abstra
2235		*         scheduled for execution
2236		*/
2237		public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) {
2238	+	if (task == null)
2239	+	throw new NullPointerException();
2240		doSubmit(task);
2241		return task;
2242		}
#	Line 1354 | Line 2247 | public class ForkJoinPool extends Abstra
2247		*         scheduled for execution
2248		*/
2249		public <T> ForkJoinTask<T> submit(Callable<T> task) {
2250	<	ForkJoinTask<T> job = ForkJoinTask.adapt(task);
2250	>	ForkJoinTask<T> job = new ForkJoinTask.AdaptedCallable<T>(task);
2251		doSubmit(job);
2252		return job;
2253		}
#	Line 1365 | Line 2258 | public class ForkJoinPool extends Abstra
2258		*         scheduled for execution
2259		*/
2260		public <T> ForkJoinTask<T> submit(Runnable task, T result) {
2261	<	ForkJoinTask<T> job = ForkJoinTask.adapt(task, result);
2261	>	ForkJoinTask<T> job = new ForkJoinTask.AdaptedRunnable<T>(task, result);
2262		doSubmit(job);
2263		return job;
2264		}
#	Line 1376 | Line 2269 | public class ForkJoinPool extends Abstra
2269		*         scheduled for execution
2270		*/
2271		public ForkJoinTask<?> submit(Runnable task) {
2272	+	if (task == null)
2273	+	throw new NullPointerException();
2274		ForkJoinTask<?> job;
2275		if (task instanceof ForkJoinTask<?>) // avoid re-wrap
2276		job = (ForkJoinTask<?>) task;
2277		else
2278	<	job = ForkJoinTask.adapt(task, null);
2278	>	job = new ForkJoinTask.AdaptedRunnableAction(task);
2279		doSubmit(job);
2280		return job;
2281		}
#	Line 1390 | Line 2285 | public class ForkJoinPool extends Abstra
2285		* @throws RejectedExecutionException {@inheritDoc}
2286		*/
2287		public <T> List<Future<T>> invokeAll(Collection<? extends Callable<T>> tasks) {
2288	<	ArrayList<ForkJoinTask<T>> forkJoinTasks =
2289	<	new ArrayList<ForkJoinTask<T>>(tasks.size());
2290	<	for (Callable<T> task : tasks)
2291	<	forkJoinTasks.add(ForkJoinTask.adapt(task));
2292	<	invoke(new InvokeAll<T>(forkJoinTasks));
2293	<
2288	>	// In previous versions of this class, this method constructed
2289	>	// a task to run ForkJoinTask.invokeAll, but now external
2290	>	// invocation of multiple tasks is at least as efficient.
2291	>	List<ForkJoinTask<T>> fs = new ArrayList<ForkJoinTask<T>>(tasks.size());
2292	>	// Workaround needed because method wasn't declared with
2293	>	// wildcards in return type but should have been.
2294		@SuppressWarnings({"unchecked", "rawtypes"})
2295	<	List<Future<T>> futures = (List<Future<T>>) (List) forkJoinTasks;
1401	<	return futures;
1402	<	}
2295	>	List<Future<T>> futures = (List<Future<T>>) (List) fs;
2296
2297	<	static final class InvokeAll<T> extends RecursiveAction {
2298	<	final ArrayList<ForkJoinTask<T>> tasks;
2299	<	InvokeAll(ArrayList<ForkJoinTask<T>> tasks) { this.tasks = tasks; }
2300	<	public void compute() {
2301	<	try { invokeAll(tasks); }
2302	<	catch (Exception ignore) {}
2297	>	boolean done = false;
2298	>	try {
2299	>	for (Callable<T> t : tasks) {
2300	>	ForkJoinTask<T> f = new ForkJoinTask.AdaptedCallable<T>(t);
2301	>	doSubmit(f);
2302	>	fs.add(f);
2303	>	}
2304	>	for (ForkJoinTask<T> f : fs)
2305	>	f.quietlyJoin();
2306	>	done = true;
2307	>	return futures;
2308	>	} finally {
2309	>	if (!done)
2310	>	for (ForkJoinTask<T> f : fs)
2311	>	f.cancel(false);
2312		}
1411	–	private static final long serialVersionUID = -7914297376763021607L;
2313		}
2314
2315		/**
#	Line 1441 | Line 2342 | public class ForkJoinPool extends Abstra
2342
2343		/**
2344		* Returns the number of worker threads that have started but not
2345	<	* yet terminated.  This result returned by this method may differ
2345	>	* yet terminated.  The result returned by this method may differ
2346		* from {@link #getParallelism} when threads are created to
2347		* maintain parallelism when others are cooperatively blocked.
2348		*
2349		* @return the number of worker threads
2350		*/
2351		public int getPoolSize() {
2352	<	return workerCounts >>> TOTAL_COUNT_SHIFT;
2352	>	return parallelism + (short)(ctl >>> TC_SHIFT);
2353		}
2354
2355		/**
#	Line 1458 | Line 2359 | public class ForkJoinPool extends Abstra
2359		* @return {@code true} if this pool uses async mode
2360		*/
2361		public boolean getAsyncMode() {
2362	<	return locallyFifo;
2362	>	return localMode != 0;
2363		}
2364
2365		/**
#	Line 1470 | Line 2371 | public class ForkJoinPool extends Abstra
2371		* @return the number of worker threads
2372		*/
2373		public int getRunningThreadCount() {
2374	<	return workerCounts & RUNNING_COUNT_MASK;
2374	>	int rc = 0;
2375	>	WorkQueue[] ws; WorkQueue w;
2376	>	if ((ws = workQueues) != null) {
2377	>	for (int i = 1; i < ws.length; i += 2) {
2378	>	if ((w = ws[i]) != null && w.isApparentlyUnblocked())
2379	>	++rc;
2380	>	}
2381	>	}
2382	>	return rc;
2383		}
2384
2385		/**
#	Line 1481 | Line 2390 | public class ForkJoinPool extends Abstra
2390		* @return the number of active threads
2391		*/
2392		public int getActiveThreadCount() {
2393	<	return runState & ACTIVE_COUNT_MASK;
2393	>	int r = parallelism + (int)(ctl >> AC_SHIFT);
2394	>	return (r <= 0) ? 0 : r; // suppress momentarily negative values
2395		}
2396
2397		/**
#	Line 1496 | Line 2406 | public class ForkJoinPool extends Abstra
2406		* @return {@code true} if all threads are currently idle
2407		*/
2408		public boolean isQuiescent() {
2409	<	return (runState & ACTIVE_COUNT_MASK) == 0;
2409	>	return (int)(ctl >> AC_SHIFT) + parallelism == 0;
2410		}
2411
2412		/**
#	Line 1511 | Line 2421 | public class ForkJoinPool extends Abstra
2421		* @return the number of steals
2422		*/
2423		public long getStealCount() {
2424	<	return stealCount;
2424	>	long count = stealCount.get();
2425	>	WorkQueue[] ws; WorkQueue w;
2426	>	if ((ws = workQueues) != null) {
2427	>	for (int i = 1; i < ws.length; i += 2) {
2428	>	if ((w = ws[i]) != null)
2429	>	count += w.totalSteals;
2430	>	}
2431	>	}
2432	>	return count;
2433		}
2434
2435		/**
#	Line 1526 | Line 2444 | public class ForkJoinPool extends Abstra
2444		*/
2445		public long getQueuedTaskCount() {
2446		long count = 0;
2447	<	for (ForkJoinWorkerThread w : workers)
2448	<	if (w != null)
2449	<	count += w.getQueueSize();
2447	>	WorkQueue[] ws; WorkQueue w;
2448	>	if ((ws = workQueues) != null) {
2449	>	for (int i = 1; i < ws.length; i += 2) {
2450	>	if ((w = ws[i]) != null)
2451	>	count += w.queueSize();
2452	>	}
2453	>	}
2454		return count;
2455		}
2456
2457		/**
2458		* Returns an estimate of the number of tasks submitted to this
2459	<	* pool that have not yet begun executing.  This method takes time
2460	<	* proportional to the number of submissions.
2459	>	* pool that have not yet begun executing.  This method may take
2460	>	* time proportional to the number of submissions.
2461		*
2462		* @return the number of queued submissions
2463		*/
2464		public int getQueuedSubmissionCount() {
2465	<	return submissionQueue.size();
2465	>	int count = 0;
2466	>	WorkQueue[] ws; WorkQueue w;
2467	>	if ((ws = workQueues) != null) {
2468	>	for (int i = 0; i < ws.length; i += 2) {
2469	>	if ((w = ws[i]) != null)
2470	>	count += w.queueSize();
2471	>	}
2472	>	}
2473	>	return count;
2474		}
2475
2476		/**
#	Line 1550 | Line 2480 | public class ForkJoinPool extends Abstra
2480		* @return {@code true} if there are any queued submissions
2481		*/
2482		public boolean hasQueuedSubmissions() {
2483	<	return !submissionQueue.isEmpty();
2483	>	WorkQueue[] ws; WorkQueue w;
2484	>	if ((ws = workQueues) != null) {
2485	>	for (int i = 0; i < ws.length; i += 2) {
2486	>	if ((w = ws[i]) != null && !w.isEmpty())
2487	>	return true;
2488	>	}
2489	>	}
2490	>	return false;
2491		}
2492
2493		/**
#	Line 1561 | Line 2498 | public class ForkJoinPool extends Abstra
2498		* @return the next submission, or {@code null} if none
2499		*/
2500		protected ForkJoinTask<?> pollSubmission() {
2501	<	return submissionQueue.poll();
2501	>	WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2502	>	if ((ws = workQueues) != null) {
2503	>	for (int i = 0; i < ws.length; i += 2) {
2504	>	if ((w = ws[i]) != null && (t = w.poll()) != null)
2505	>	return t;
2506	>	}
2507	>	}
2508	>	return null;
2509		}
2510
2511		/**
#	Line 1582 | Line 2526 | public class ForkJoinPool extends Abstra
2526		* @return the number of elements transferred
2527		*/
2528		protected int drainTasksTo(Collection<? super ForkJoinTask<?>> c) {
2529	<	int count = submissionQueue.drainTo(c);
2530	<	for (ForkJoinWorkerThread w : workers)
2531	<	if (w != null)
2532	<	count += w.drainTasksTo(c);
2529	>	int count = 0;
2530	>	WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2531	>	if ((ws = workQueues) != null) {
2532	>	for (int i = 0; i < ws.length; ++i) {
2533	>	if ((w = ws[i]) != null) {
2534	>	while ((t = w.poll()) != null) {
2535	>	c.add(t);
2536	>	++count;
2537	>	}
2538	>	}
2539	>	}
2540	>	}
2541		return count;
2542		}
2543
#	Line 1597 | Line 2549 | public class ForkJoinPool extends Abstra
2549		* @return a string identifying this pool, as well as its state
2550		*/
2551		public String toString() {
2552	<	long st = getStealCount();
2553	<	long qt = getQueuedTaskCount();
2554	<	long qs = getQueuedSubmissionCount();
2555	<	int wc = workerCounts;
2556	<	int tc = wc >>> TOTAL_COUNT_SHIFT;
2557	<	int rc = wc & RUNNING_COUNT_MASK;
2552	>	// Use a single pass through workQueues to collect counts
2553	>	long qt = 0L, qs = 0L; int rc = 0;
2554	>	long st = stealCount.get();
2555	>	long c = ctl;
2556	>	WorkQueue[] ws; WorkQueue w;
2557	>	if ((ws = workQueues) != null) {
2558	>	for (int i = 0; i < ws.length; ++i) {
2559	>	if ((w = ws[i]) != null) {
2560	>	int size = w.queueSize();
2561	>	if ((i & 1) == 0)
2562	>	qs += size;
2563	>	else {
2564	>	qt += size;
2565	>	st += w.totalSteals;
2566	>	if (w.isApparentlyUnblocked())
2567	>	++rc;
2568	>	}
2569	>	}
2570	>	}
2571	>	}
2572		int pc = parallelism;
2573	<	int rs = runState;
2574	<	int ac = rs & ACTIVE_COUNT_MASK;
2573	>	int tc = pc + (short)(c >>> TC_SHIFT);
2574	>	int ac = pc + (int)(c >> AC_SHIFT);
2575	>	if (ac < 0) // ignore transient negative
2576	>	ac = 0;
2577	>	String level;
2578	>	if ((c & STOP_BIT) != 0)
2579	>	level = (tc == 0) ? "Terminated" : "Terminating";
2580	>	else
2581	>	level = runState < 0 ? "Shutting down" : "Running";
2582		return super.toString() +
2583	<	"[" + runLevelToString(rs) +
2583	>	"[" + level +
2584		", parallelism = " + pc +
2585		", size = " + tc +
2586		", active = " + ac +
#	Line 1618 | Line 2591 | public class ForkJoinPool extends Abstra
2591		"]";
2592		}
2593
1621	–	private static String runLevelToString(int s) {
1622	–	return ((s & TERMINATED) != 0 ? "Terminated" :
1623	–	((s & TERMINATING) != 0 ? "Terminating" :
1624	–	((s & SHUTDOWN) != 0 ? "Shutting down" :
1625	–	"Running")));
1626	–	}
1627	–
2594		/**
2595		* Initiates an orderly shutdown in which previously submitted
2596		* tasks are executed, but no new tasks will be accepted.
#	Line 1639 | Line 2605 | public class ForkJoinPool extends Abstra
2605		*/
2606		public void shutdown() {
2607		checkPermission();
2608	<	advanceRunLevel(SHUTDOWN);
1643	<	tryTerminate(false);
2608	>	tryTerminate(false, true);
2609		}
2610
2611		/**
#	Line 1661 | Line 2626 | public class ForkJoinPool extends Abstra
2626		*/
2627		public List<Runnable> shutdownNow() {
2628		checkPermission();
2629	<	tryTerminate(true);
2629	>	tryTerminate(true, true);
2630		return Collections.emptyList();
2631		}
2632
#	Line 1671 | Line 2636 | public class ForkJoinPool extends Abstra
2636		* @return {@code true} if all tasks have completed following shut down
2637		*/
2638		public boolean isTerminated() {
2639	<	return runState >= TERMINATED;
2639	>	long c = ctl;
2640	>	return ((c & STOP_BIT) != 0L &&
2641	>	(short)(c >>> TC_SHIFT) == -parallelism);
2642		}
2643
2644		/**
#	Line 1679 | Line 2646 | public class ForkJoinPool extends Abstra
2646		* commenced but not yet completed.  This method may be useful for
2647		* debugging. A return of {@code true} reported a sufficient
2648		* period after shutdown may indicate that submitted tasks have
2649	<	* ignored or suppressed interruption, causing this executor not
2650	<	* to properly terminate.
2649	>	* ignored or suppressed interruption, or are waiting for IO,
2650	>	* causing this executor not to properly terminate. (See the
2651	>	* advisory notes for class {@link ForkJoinTask} stating that
2652	>	* tasks should not normally entail blocking operations.  But if
2653	>	* they do, they must abort them on interrupt.)
2654		*
2655		* @return {@code true} if terminating but not yet terminated
2656		*/
2657		public boolean isTerminating() {
2658	<	return (runState & (TERMINATING|TERMINATED)) == TERMINATING;
2658	>	long c = ctl;
2659	>	return ((c & STOP_BIT) != 0L &&
2660	>	(short)(c >>> TC_SHIFT) != -parallelism);
2661		}
2662
2663		/**
#	Line 1694 | Line 2666 | public class ForkJoinPool extends Abstra
2666		* @return {@code true} if this pool has been shut down
2667		*/
2668		public boolean isShutdown() {
2669	<	return runState >= SHUTDOWN;
2669	>	return runState < 0;
2670		}
2671
2672		/**
#	Line 1710 | Line 2682 | public class ForkJoinPool extends Abstra
2682		*/
2683		public boolean awaitTermination(long timeout, TimeUnit unit)
2684		throws InterruptedException {
2685	+	long nanos = unit.toNanos(timeout);
2686	+	final Mutex lock = this.lock;
2687	+	lock.lock();
2688		try {
2689	<	return termination.awaitAdvanceInterruptibly(0, timeout, unit) > 0;
2690	<	} catch (TimeoutException ex) {
2691	<	return false;
2689	>	for (;;) {
2690	>	if (isTerminated())
2691	>	return true;
2692	>	if (nanos <= 0)
2693	>	return false;
2694	>	nanos = termination.awaitNanos(nanos);
2695	>	}
2696	>	} finally {
2697	>	lock.unlock();
2698		}
2699		}
2700
#	Line 1725 | Line 2706 | public class ForkJoinPool extends Abstra
2706		* {@code isReleasable} must return {@code true} if blocking is
2707		* not necessary. Method {@code block} blocks the current thread
2708		* if necessary (perhaps internally invoking {@code isReleasable}
2709	<	* before actually blocking). The unusual methods in this API
2710	<	* accommodate synchronizers that may, but don't usually, block
2711	<	* for long periods. Similarly, they allow more efficient internal
2712	<	* handling of cases in which additional workers may be, but
2713	<	* usually are not, needed to ensure sufficient parallelism.
2714	<	* Toward this end, implementations of method {@code isReleasable}
2715	<	* must be amenable to repeated invocation.
2709	>	* before actually blocking). These actions are performed by any
2710	>	* thread invoking {@link ForkJoinPool#managedBlock}.  The
2711	>	* unusual methods in this API accommodate synchronizers that may,
2712	>	* but don't usually, block for long periods. Similarly, they
2713	>	* allow more efficient internal handling of cases in which
2714	>	* additional workers may be, but usually are not, needed to
2715	>	* ensure sufficient parallelism.  Toward this end,
2716	>	* implementations of method {@code isReleasable} must be amenable
2717	>	* to repeated invocation.
2718		*
2719		* <p>For example, here is a ManagedBlocker based on a
2720		* ReentrantLock:
#	Line 1811 | Line 2794 | public class ForkJoinPool extends Abstra
2794		public static void managedBlock(ManagedBlocker blocker)
2795		throws InterruptedException {
2796		Thread t = Thread.currentThread();
2797	<	if (t instanceof ForkJoinWorkerThread) {
2798	<	ForkJoinWorkerThread w = (ForkJoinWorkerThread) t;
2799	<	w.pool.awaitBlocker(blocker);
2800	<	}
2801	<	else {
2802	<	do {} while (!blocker.isReleasable() && !blocker.block());
2797	>	ForkJoinPool p = ((t instanceof ForkJoinWorkerThread) ?
2798	>	((ForkJoinWorkerThread)t).pool : null);
2799	>	while (!blocker.isReleasable()) {
2800	>	if (p == null || p.tryCompensate(null, blocker)) {
2801	>	try {
2802	>	do {} while (!blocker.isReleasable() && !blocker.block());
2803	>	} finally {
2804	>	if (p != null)
2805	>	p.incrementActiveCount();
2806	>	}
2807	>	break;
2808	>	}
2809		}
2810		}
2811
#	Line 1825 | Line 2814 | public class ForkJoinPool extends Abstra
2814		// implement RunnableFuture.
2815
2816		protected <T> RunnableFuture<T> newTaskFor(Runnable runnable, T value) {
2817	<	return (RunnableFuture<T>) ForkJoinTask.adapt(runnable, value);
2817	>	return new ForkJoinTask.AdaptedRunnable<T>(runnable, value);
2818		}
2819
2820		protected <T> RunnableFuture<T> newTaskFor(Callable<T> callable) {
2821	<	return (RunnableFuture<T>) ForkJoinTask.adapt(callable);
2821	>	return new ForkJoinTask.AdaptedCallable<T>(callable);
2822		}
2823
2824		// Unsafe mechanics
2825	<
2826	<	private static final sun.misc.Unsafe UNSAFE = getUnsafe();
2827	<	private static final long workerCountsOffset =
2828	<	objectFieldOffset("workerCounts", ForkJoinPool.class);
2829	<	private static final long runStateOffset =
2830	<	objectFieldOffset("runState", ForkJoinPool.class);
2831	<	private static final long eventCountOffset =
2832	<	objectFieldOffset("eventCount", ForkJoinPool.class);
2833	<	private static final long eventWaitersOffset =
2834	<	objectFieldOffset("eventWaiters",ForkJoinPool.class);
2835	<	private static final long stealCountOffset =
2836	<	objectFieldOffset("stealCount",ForkJoinPool.class);
2837	<	private static final long spareWaitersOffset =
2838	<	objectFieldOffset("spareWaiters",ForkJoinPool.class);
1850	<
1851	<	private static long objectFieldOffset(String field, Class<?> klazz) {
2825	>	private static final sun.misc.Unsafe U;
2826	>	private static final long CTL;
2827	>	private static final long PARKBLOCKER;
2828	>	private static final int ABASE;
2829	>	private static final int ASHIFT;
2830	>
2831	>	static {
2832	>	poolNumberGenerator = new AtomicInteger();
2833	>	nextSubmitterSeed = new AtomicInteger(0x55555555);
2834	>	modifyThreadPermission = new RuntimePermission("modifyThread");
2835	>	defaultForkJoinWorkerThreadFactory =
2836	>	new DefaultForkJoinWorkerThreadFactory();
2837	>	submitters = new ThreadSubmitter();
2838	>	int s;
2839		try {
2840	<	return UNSAFE.objectFieldOffset(klazz.getDeclaredField(field));
2841	<	} catch (NoSuchFieldException e) {
2842	<	// Convert Exception to corresponding Error
2843	<	NoSuchFieldError error = new NoSuchFieldError(field);
2844	<	error.initCause(e);
2845	<	throw error;
2846	<	}
2840	>	U = getUnsafe();
2841	>	Class<?> k = ForkJoinPool.class;
2842	>	Class<?> ak = ForkJoinTask[].class;
2843	>	CTL = U.objectFieldOffset
2844	>	(k.getDeclaredField("ctl"));
2845	>	Class<?> tk = Thread.class;
2846	>	PARKBLOCKER = U.objectFieldOffset
2847	>	(tk.getDeclaredField("parkBlocker"));
2848	>	ABASE = U.arrayBaseOffset(ak);
2849	>	s = U.arrayIndexScale(ak);
2850	>	} catch (Exception e) {
2851	>	throw new Error(e);
2852	>	}
2853	>	if ((s & (s-1)) != 0)
2854	>	throw new Error("data type scale not a power of two");
2855	>	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
2856		}
2857
2858		/**
#	Line 1886 | Line 2882 | public class ForkJoinPool extends Abstra
2882		}
2883		}
2884		}
2885	+
2886		}

Diff Legend

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