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 1.58 by dl, Fri Jul 23 13:07:43 2010 UTC vs.
Revision 1.144 by jsr166, Sun Nov 18 18:03:10 2012 UTC

#	Line 1 | Line 1
1		/*
2		* Written by Doug Lea with assistance from members of JCP JSR-166
3		* Expert Group and released to the public domain, as explained at
4	<	* http://creativecommons.org/licenses/publicdomain
4	>	* http://creativecommons.org/publicdomain/zero/1.0/
5		*/
6
7		package jsr166y;
8
9	–	import java.util.concurrent.*;
10	–
9		import java.util.ArrayList;
10		import java.util.Arrays;
11		import java.util.Collection;
12		import java.util.Collections;
13		import java.util.List;
14	<	import java.util.concurrent.locks.LockSupport;
15	<	import java.util.concurrent.locks.ReentrantLock;
16	<	import java.util.concurrent.atomic.AtomicInteger;
17	<	import java.util.concurrent.CountDownLatch;
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
22		/**
23		* An {@link ExecutorService} for running {@link ForkJoinTask}s.
#	Line 27 | Line 28 | import java.util.concurrent.CountDownLat
28		* <p>A {@code ForkJoinPool} differs from other kinds of {@link
29		* ExecutorService} mainly by virtue of employing
30		* <em>work-stealing</em>: all threads in the pool attempt to find and
31	<	* execute subtasks created by other active tasks (eventually blocking
32	<	* waiting for work if none exist). This enables efficient processing
33	<	* when most tasks spawn other subtasks (as do most {@code
34	<	* ForkJoinTask}s). When setting <em>asyncMode</em> to true in
35	<	* constructors, {@code ForkJoinPool}s may also be appropriate for use
36	<	* with event-style tasks that are never joined.
31	>	* execute tasks submitted to the pool and/or created by other active
32	>	* tasks (eventually blocking waiting for work if none exist). This
33	>	* enables efficient processing when most tasks spawn other subtasks
34	>	* (as do most {@code ForkJoinTask}s), as well as when many small
35	>	* tasks are submitted to the pool from external clients.  Especially
36	>	* when setting <em>asyncMode</em> to true in constructors, {@code
37	>	* ForkJoinPool}s may also be appropriate for use with event-style
38	>	* tasks that are never joined.
39	>	*
40	>	* <p>A static {@link #commonPool} is available and appropriate for
41	>	* most applications. The common pool is used by any ForkJoinTask that
42	>	* is not explicitly submitted to a specified pool. Using the common
43	>	* pool normally reduces resource usage (its threads are slowly
44	>	* reclaimed during periods of non-use, and reinstated upon subsequent
45	>	* use).
46		*
47	<	* <p>A {@code ForkJoinPool} is constructed with a given target
48	<	* parallelism level; by default, equal to the number of available
49	<	* processors. The pool attempts to maintain enough active (or
50	<	* available) threads by dynamically adding, suspending, or resuming
51	<	* internal worker threads, even if some tasks are stalled waiting to
52	<	* join others. However, no such adjustments are guaranteed in the
53	<	* face of blocked IO or other unmanaged synchronization. The nested
54	<	* {@link ManagedBlocker} interface enables extension of the kinds of
47	>	* <p>For applications that require separate or custom pools, a {@code
48	>	* ForkJoinPool} may be constructed with a given target parallelism
49	>	* level; by default, equal to the number of available processors. The
50	>	* pool attempts to maintain enough active (or available) threads by
51	>	* dynamically adding, suspending, or resuming internal worker
52	>	* threads, even if some tasks are stalled waiting to join
53	>	* others. However, no such adjustments are guaranteed in the face of
54	>	* blocked IO or other unmanaged synchronization. The nested {@link
55	>	* ManagedBlocker} interface enables extension of the kinds of
56		* synchronization accommodated.
57		*
58		* <p>In addition to execution and lifecycle control methods, this
#	Line 51 | Line 62 | import java.util.concurrent.CountDownLat
62		* {@link #toString} returns indications of pool state in a
63		* convenient form for informal monitoring.
64		*
65	<	* <p> As is the case with other ExecutorServices, there are three
66	<	* main task execution methods summarized in the follwoing
67	<	* table. These are designed to be used by clients not already engaged
68	<	* in fork/join computations in the current pool.  The main forms of
69	<	* these methods accept instances of {@code ForkJoinTask}, but
70	<	* overloaded forms also allow mixed execution of plain {@code
65	>	* <p>As is the case with other ExecutorServices, there are three
66	>	* main task execution methods summarized in the following table.
67	>	* These are designed to be used primarily by clients not already
68	>	* engaged in fork/join computations in the current pool.  The main
69	>	* forms of these methods accept instances of {@code ForkJoinTask},
70	>	* but overloaded forms also allow mixed execution of plain {@code
71		* Runnable}- or {@code Callable}- based activities as well.  However,
72	<	* tasks that are already executing in a pool should normally
73	<	* <em>NOT</em> use these pool execution methods, but instead use the
74	<	* within-computation forms listed in the table.
72	>	* tasks that are already executing in a pool should normally instead
73	>	* use the within-computation forms listed in the table unless using
74	>	* async event-style tasks that are not usually joined, in which case
75	>	* there is little difference among choice of methods.
76		*
77		* <table BORDER CELLPADDING=3 CELLSPACING=1>
78		*  <tr>
#	Line 69 | Line 81 | import java.util.concurrent.CountDownLat
81		*    <td ALIGN=CENTER> <b>Call from within fork/join computations</b></td>
82		*  </tr>
83		*  <tr>
84	<	*    <td> <b>Arange async execution</td>
84	>	*    <td> <b>Arrange async execution</td>
85		*    <td> {@link #execute(ForkJoinTask)}</td>
86		*    <td> {@link ForkJoinTask#fork}</td>
87		*  </tr>
#	Line 84 | Line 96 | import java.util.concurrent.CountDownLat
96		*    <td> {@link ForkJoinTask#fork} (ForkJoinTasks <em>are</em> Futures)</td>
97		*  </tr>
98		* </table>
87	–	*
88	–	* <p><b>Sample Usage.</b> Normally a single {@code ForkJoinPool} is
89	–	* used for all parallel task execution in a program or subsystem.
90	–	* Otherwise, use would not usually outweigh the construction and
91	–	* bookkeeping overhead of creating a large set of threads. For
92	–	* example, a common pool could be used for the {@code SortTasks}
93	–	* illustrated in {@link RecursiveAction}. Because {@code
94	–	* ForkJoinPool} uses threads in {@linkplain java.lang.Thread#isDaemon
95	–	* daemon} mode, there is typically no need to explicitly {@link
96	–	* #shutdown} such a pool upon program exit.
99		*
100	<	* <pre>
101	<	* static final ForkJoinPool mainPool = new ForkJoinPool();
102	<	* ...
103	<	* public void sort(long[] array) {
104	<	*   mainPool.invoke(new SortTask(array, 0, array.length));
105	<	* }
106	<	* </pre>
100	>	* <p>The common pool is by default constructed with default
101	>	* parameters, but these may be controlled by setting three {@link
102	>	* System#getProperty properties} with prefix {@code
103	>	* java.util.concurrent.ForkJoinPool.common}: {@code parallelism} --
104	>	* an integer greater than zero, {@code threadFactory} -- the class
105	>	* name of a {@link ForkJoinWorkerThreadFactory}, and {@code
106	>	* exceptionHandler} -- the class name of a {@link
107	>	* java.lang.Thread.UncaughtExceptionHandler
108	>	* Thread.UncaughtExceptionHandler}. Upon any error in establishing
109	>	* these settings, default parameters are used.
110		*
111		* <p><b>Implementation notes</b>: This implementation restricts the
112		* maximum number of running threads to 32767. Attempts to create
#	Line 110 | Line 115 | import java.util.concurrent.CountDownLat
115		*
116		* <p>This implementation rejects submitted tasks (that is, by throwing
117		* {@link RejectedExecutionException}) only when the pool is shut down
118	<	* or internal resources have been exhuasted.
118	>	* or internal resources have been exhausted.
119		*
120		* @since 1.7
121		* @author Doug Lea
#	Line 120 | Line 125 | public class ForkJoinPool extends Abstra
125		/*
126		* Implementation Overview
127		*
128	<	* This class provides the central bookkeeping and control for a
129	<	* set of worker threads: Submissions from non-FJ threads enter
130	<	* into a submission queue. Workers take these tasks and typically
131	<	* split them into subtasks that may be stolen by other workers.
132	<	* The main work-stealing mechanics implemented in class
133	<	* ForkJoinWorkerThread give first priority to processing tasks
134	<	* from their own queues (LIFO or FIFO, depending on mode), then
135	<	* to randomized FIFO steals of tasks in other worker queues, and
136	<	* lastly to new submissions. These mechanics do not consider
137	<	* affinities, loads, cache localities, etc, so rarely provide the
138	<	* best possible performance on a given machine, but portably
139	<	* provide good throughput by averaging over these factors.
140	<	* (Further, even if we did try to use such information, we do not
141	<	* usually have a basis for exploiting it. For example, some sets
142	<	* of tasks profit from cache affinities, but others are harmed by
143	<	* cache pollution effects.)
144	<	*
145	<	* Beyond work-stealing support and essential bookkeeping, the
146	<	* main responsibility of this framework is to arrange tactics for
147	<	* when one worker is waiting to join a task stolen (or always
148	<	* held by) another.  Becauae we are multiplexing many tasks on to
149	<	* a pool of workers, we can't just let them block (as in
150	<	* Thread.join).  We also cannot just reassign the joiner's
151	<	* run-time stack with another and replace it later, which would
152	<	* be a form of "continuation", that even if possible is not
153	<	* necessarily a good idea. Given that the creation costs of most
154	<	* threads on most systems mainly surrounds setting up runtime
155	<	* stacks, thread creation and switching is usually not much more
156	<	* expensive than stack creation and switching, and is more
157	<	* flexible). Instead we combine two tactics:
158	<	*
159	<	*   1. Arranging for the joiner to execute some task that it
160	<	*      would be running if the steal had not occurred.  Method
161	<	*      ForkJoinWorkerThread.helpJoinTask tracks joining->stealing
162	<	*      links to try to find such a task.
163	<	*
164	<	*   2. Unless there are already enough live threads, creating or
165	<	*      or re-activating a spare thread to compensate for the
166	<	*      (blocked) joiner until it unblocks.  Spares then suspend
167	<	*      at their next opportunity or eventually die if unused for
168	<	*      too long.  See below and the internal documentation
169	<	*      for tryAwaitJoin for more details about compensation
170	<	*      rules.
171	<	*
172	<	* Because the determining existence of conservatively safe
173	<	* helping targets, the availability of already-created spares,
174	<	* and the apparent need to create new spares are all racy and
175	<	* require heuristic guidance, joins (in
176	<	* ForkJoinWorkerThread.joinTask) interleave these options until
177	<	* successful.  Creating a new spare always succeeds, but also
178	<	* increases application footprint, so we try to avoid it, within
179	<	* reason.
128	>	* This class and its nested classes provide the main
129	>	* functionality and control for a set of worker threads:
130	>	* Submissions from non-FJ threads enter into submission queues.
131	>	* Workers take these tasks and typically split them into subtasks
132	>	* that may be stolen by other workers.  Preference rules give
133	>	* first priority to processing tasks from their own queues (LIFO
134	>	* or FIFO, depending on mode), then to randomized FIFO steals of
135	>	* tasks in other queues.
136	>	*
137	>	* WorkQueues
138	>	* ==========
139	>	*
140	>	* Most operations occur within work-stealing queues (in nested
141	>	* class WorkQueue).  These are special forms of Deques that
142	>	* support only three of the four possible end-operations -- push,
143	>	* pop, and poll (aka steal), under the further constraints that
144	>	* push and pop are called only from the owning thread (or, as
145	>	* extended here, under a lock), while poll may be called from
146	>	* other threads.  (If you are unfamiliar with them, you probably
147	>	* want to read Herlihy and Shavit's book "The Art of
148	>	* Multiprocessor programming", chapter 16 describing these in
149	>	* more detail before proceeding.)  The main work-stealing queue
150	>	* design is roughly similar to those in the papers "Dynamic
151	>	* Circular Work-Stealing Deque" by Chase and Lev, SPAA 2005
152	>	* (http://research.sun.com/scalable/pubs/index.html) and
153	>	* "Idempotent work stealing" by Michael, Saraswat, and Vechev,
154	>	* PPoPP 2009 (http://portal.acm.org/citation.cfm?id=1504186).
155	>	* The main differences ultimately stem from GC requirements that
156	>	* we null out taken slots as soon as we can, to maintain as small
157	>	* a footprint as possible even in programs generating huge
158	>	* numbers of tasks. To accomplish this, we shift the CAS
159	>	* arbitrating pop vs poll (steal) from being on the indices
160	>	* ("base" and "top") to the slots themselves.  So, both a
161	>	* successful pop and poll mainly entail a CAS of a slot from
162	>	* non-null to null.  Because we rely on CASes of references, we
163	>	* do not need tag bits on base or top.  They are simple ints as
164	>	* used in any circular array-based queue (see for example
165	>	* ArrayDeque).  Updates to the indices must still be ordered in a
166	>	* way that guarantees that top == base means the queue is empty,
167	>	* but otherwise may err on the side of possibly making the queue
168	>	* appear nonempty when a push, pop, or poll have not fully
169	>	* committed. Note that this means that the poll operation,
170	>	* considered individually, is not wait-free. One thief cannot
171	>	* successfully continue until another in-progress one (or, if
172	>	* previously empty, a push) completes.  However, in the
173	>	* aggregate, we ensure at least probabilistic non-blockingness.
174	>	* If an attempted steal fails, a thief always chooses a different
175	>	* random victim target to try next. So, in order for one thief to
176	>	* progress, it suffices for any in-progress poll or new push on
177	>	* any empty queue to complete. (This is why we normally use
178	>	* method pollAt and its variants that try once at the apparent
179	>	* base index, else consider alternative actions, rather than
180	>	* method poll.)
181	>	*
182	>	* This approach also enables support of a user mode in which local
183	>	* task processing is in FIFO, not LIFO order, simply by using
184	>	* poll rather than pop.  This can be useful in message-passing
185	>	* frameworks in which tasks are never joined.  However neither
186	>	* mode considers affinities, loads, cache localities, etc, so
187	>	* rarely provide the best possible performance on a given
188	>	* machine, but portably provide good throughput by averaging over
189	>	* these factors.  (Further, even if we did try to use such
190	>	* information, we do not usually have a basis for exploiting it.
191	>	* For example, some sets of tasks profit from cache affinities,
192	>	* but others are harmed by cache pollution effects.)
193	>	*
194	>	* WorkQueues are also used in a similar way for tasks submitted
195	>	* to the pool. We cannot mix these tasks in the same queues used
196	>	* for work-stealing (this would contaminate lifo/fifo
197	>	* processing). Instead, we randomly associate submission queues
198	>	* with submitting threads, using a form of hashing.  The
199	>	* ThreadLocal Submitter class contains a value initially used as
200	>	* a hash code for choosing existing queues, but may be randomly
201	>	* repositioned upon contention with other submitters.  In
202	>	* essence, submitters act like workers except that they are
203	>	* restricted to executing local tasks that they submitted (or in
204	>	* the case of CountedCompleters, others with the same root task).
205	>	* However, because most shared/external queue operations are more
206	>	* expensive than internal, and because, at steady state, external
207	>	* submitters will compete for CPU with workers, ForkJoinTask.join
208	>	* and related methods disable them from repeatedly helping to
209	>	* process tasks if all workers are active.  Insertion of tasks in
210	>	* shared mode requires a lock (mainly to protect in the case of
211	>	* resizing) but we use only a simple spinlock (using bits in
212	>	* field qlock), because submitters encountering a busy queue move
213	>	* on to try or create other queues -- they block only when
214	>	* creating and registering new queues.
215		*
216	<	* The ManagedBlocker extension API can't use option (1) so uses a
217	<	* special version of (2) in method awaitBlocker.
216	>	* Management
217	>	* ==========
218		*
219		* The main throughput advantages of work-stealing stem from
220	<	* decentralized control -- workers mostly steal tasks from each
221	<	* other. We do not want to negate this by creating bottlenecks
222	<	* implementing other management responsibilities. So we use a
223	<	* collection of techniques that avoid, reduce, or cope well with
224	<	* contention. These entail several instances of bit-packing into
225	<	* CASable fields to maintain only the minimally required
226	<	* atomicity. To enable such packing, we restrict maximum
227	<	* parallelism to (1<<15)-1 (enabling twice this (to accommodate
228	<	* unbalanced increments and decrements) to fit into a 16 bit
229	<	* field, which is far in excess of normal operating range.  Even
230	<	* though updates to some of these bookkeeping fields do sometimes
231	<	* contend with each other, they don't normally cache-contend with
232	<	* updates to others enough to warrant memory padding or
233	<	* isolation. So they are all held as fields of ForkJoinPool
234	<	* objects.  The main capabilities are as follows:
235	<	*
236	<	* 1. Creating and removing workers. Workers are recorded in the
237	<	* "workers" array. This is an array as opposed to some other data
238	<	* structure to support index-based random steals by workers.
239	<	* Updates to the array recording new workers and unrecording
240	<	* terminated ones are protected from each other by a lock
241	<	* (workerLock) but the array is otherwise concurrently readable,
242	<	* and accessed directly by workers. To simplify index-based
243	<	* operations, the array size is always a power of two, and all
244	<	* readers must tolerate null slots. Currently, all worker thread
245	<	* creation is on-demand, triggered by task submissions,
246	<	* replacement of terminated workers, and/or compensation for
247	<	* blocked workers. However, all other support code is set up to
248	<	* work with other policies.
249	<	*
250	<	* 2. Bookkeeping for dynamically adding and removing workers. We
251	<	* aim to approximately maintain the given level of parallelism.
252	<	* When some workers are known to be blocked (on joins or via
253	<	* ManagedBlocker), we may create or resume others to take their
254	<	* place until they unblock (see below). Implementing this
255	<	* requires counts of the number of "running" threads (i.e., those
256	<	* that are neither blocked nor artifically suspended) as well as
257	<	* the total number.  These two values are packed into one field,
258	<	* "workerCounts" because we need accurate snapshots when deciding
259	<	* to create, resume or suspend.  Note however that the
260	<	* correspondance of these counts to reality is not guaranteed. In
261	<	* particular updates for unblocked threads may lag until they
262	<	* actually wake up.
263	<	*
264	<	* 3. Maintaining global run state. The run state of the pool
265	<	* consists of a runLevel (SHUTDOWN, TERMINATING, etc) similar to
266	<	* those in other Executor implementations, as well as a count of
267	<	* "active" workers -- those that are, or soon will be, or
268	<	* recently were executing tasks. The runLevel and active count
269	<	* are packed together in order to correctly trigger shutdown and
270	<	* termination. Without care, active counts can be subject to very
271	<	* high contention.  We substantially reduce this contention by
272	<	* relaxing update rules.  A worker must claim active status
273	<	* prospectively, by activating if it sees that a submitted or
274	<	* stealable task exists (it may find after activating that the
275	<	* task no longer exists). It stays active while processing this
276	<	* task (if it exists) and any other local subtasks it produces,
277	<	* until it cannot find any other tasks. It then tries
278	<	* inactivating (see method preStep), but upon update contention
279	<	* instead scans for more tasks, later retrying inactivation if it
280	<	* doesn't find any.
281	<	*
282	<	* 4. Managing idle workers waiting for tasks. We cannot let
283	<	* workers spin indefinitely scanning for tasks when none are
284	<	* available. On the other hand, we must quickly prod them into
285	<	* action when new tasks are submitted or generated.  We
286	<	* park/unpark these idle workers using an event-count scheme.
287	<	* Field eventCount is incremented upon events that may enable
288	<	* workers that previously could not find a task to now find one:
289	<	* Submission of a new task to the pool, or another worker pushing
290	<	* a task onto a previously empty queue.  (We also use this
291	<	* mechanism for termination and reconfiguration actions that
292	<	* require wakeups of idle workers).  Each worker maintains its
293	<	* last known event count, and blocks when a scan for work did not
294	<	* find a task AND its lastEventCount matches the current
295	<	* eventCount. Waiting idle workers are recorded in a variant of
296	<	* Treiber stack headed by field eventWaiters which, when nonzero,
297	<	* encodes the thread index and count awaited for by the worker
298	<	* thread most recently calling eventSync. This thread in turn has
299	<	* a record (field nextEventWaiter) for the next waiting worker.
300	<	* In addition to allowing simpler decisions about need for
301	<	* wakeup, the event count bits in eventWaiters serve the role of
302	<	* tags to avoid ABA errors in Treiber stacks.  To reduce delays
303	<	* in task diffusion, workers not otherwise occupied may invoke
304	<	* method releaseWaiters, that removes and signals (unparks)
305	<	* workers not waiting on current count. To minimize task
306	<	* production stalls associate with signalling, any worker pushing
307	<	* a task on an empty queue invokes the weaker method signalWork,
308	<	* that only releases idle workers until it detects interference
309	<	* by other threads trying to release, and lets them take
310	<	* over. The net effect is a tree-like diffusion of signals, where
311	<	* released threads (and possibly others) help with unparks.  To
312	<	* further reduce contention effects a bit, failed CASes to
313	<	* increment field eventCount are tolerated without retries.
314	<	* Conceptually they are merged into the same event, which is OK
315	<	* when their only purpose is to enable workers to scan for work.
316	<	*
317	<	* 5. Managing suspension of extra workers. When a worker is about
318	<	* to block waiting for a join (or via ManagedBlockers), we may
319	<	* create a new thread to maintain parallelism level, or at least
320	<	* avoid starvation. Usually, extra threads are needed for only
321	<	* very short periods, yet join dependencies are such that we
322	<	* sometimes need them in bursts. Rather than create new threads
323	<	* each time this happens, we suspend no-longer-needed extra ones
324	<	* as "spares". For most purposes, we don't distinguish "extra"
325	<	* spare threads from normal "core" threads: On each call to
326	<	* preStep (the only point at which we can do this) a worker
327	<	* checks to see if there are now too many running workers, and if
328	<	* so, suspends itself.  Methods tryAwaitJoin and awaitBlocker
329	<	* look for suspended threads to resume before considering
330	<	* creating a new replacement. We don't need a special data
331	<	* structure to maintain spares; simply scanning the workers array
332	<	* looking for worker.isSuspended() is fine because the calling
333	<	* thread is otherwise not doing anything useful anyway; we are at
334	<	* least as happy if after locating a spare, the caller doesn't
335	<	* actually block because the join is ready before we try to
336	<	* adjust and compensate.  Note that this is intrinsically racy.
337	<	* One thread may become a spare at about the same time as another
338	<	* is needlessly being created. We counteract this and related
339	<	* slop in part by requiring resumed spares to immediately recheck
340	<	* (in preStep) to see whether they they should re-suspend. The
341	<	* only effective difference between "extra" and "core" threads is
342	<	* that we allow the "extra" ones to time out and die if they are
343	<	* not resumed within a keep-alive interval of a few seconds. This
344	<	* is implemented mainly within ForkJoinWorkerThread, but requires
345	<	* some coordination (isTrimmed() -- meaning killed while
346	<	* suspended) to correctly maintain pool counts.
347	<	*
348	<	* 6. Deciding when to create new workers. The main dynamic
349	<	* control in this class is deciding when to create extra threads,
350	<	* in methods awaitJoin and awaitBlocker. We always need to create
351	<	* one when the number of running threads would become zero and
352	<	* all workers are busy. However, this is not easy to detect
353	<	* reliably in the presence of transients so we use retries and
354	<	* allow slack (in tryAwaitJoin) to reduce false alarms.  These
355	<	* effectively reduce churn at the price of systematically
356	<	* undershooting target parallelism when many threads are blocked.
357	<	* However, biasing toward undeshooting partially compensates for
358	<	* the above mechanics to suspend extra threads, that normally
359	<	* lead to overshoot because we can only suspend workers
360	<	* in-between top-level actions. It also better copes with the
361	<	* fact that some of the methods in this class tend to never
362	<	* become compiled (but are interpreted), so some components of
363	<	* the entire set of controls might execute many times faster than
364	<	* others. And similarly for cases where the apparent lack of work
365	<	* is just due to GC stalls and other transient system activity.
366	<	*
367	<	* Beware that there is a lot of representation-level coupling
368	<	* among classes ForkJoinPool, ForkJoinWorkerThread, and
369	<	* ForkJoinTask.  For example, direct access to "workers" array by
370	<	* workers, and direct access to ForkJoinTask.status by both
371	<	* ForkJoinPool and ForkJoinWorkerThread.  There is little point
220	>	* decentralized control -- workers mostly take tasks from
221	>	* themselves or each other. We cannot negate this in the
222	>	* implementation of other management responsibilities. The main
223	>	* tactic for avoiding bottlenecks is packing nearly all
224	>	* essentially atomic control state into two volatile variables
225	>	* that are by far most often read (not written) as status and
226	>	* consistency checks.
227	>	*
228	>	* Field "ctl" contains 64 bits holding all the information needed
229	>	* to atomically decide to add, inactivate, enqueue (on an event
230	>	* queue), dequeue, and/or re-activate workers.  To enable this
231	>	* packing, we restrict maximum parallelism to (1<<15)-1 (which is
232	>	* far in excess of normal operating range) to allow ids, counts,
233	>	* and their negations (used for thresholding) to fit into 16bit
234	>	* fields.
235	>	*
236	>	* Field "plock" is a form of sequence lock with a saturating
237	>	* shutdown bit (similarly for per-queue "qlocks"), mainly
238	>	* protecting updates to the workQueues array, as well as to
239	>	* enable shutdown.  When used as a lock, it is normally only very
240	>	* briefly held, so is nearly always available after at most a
241	>	* brief spin, but we use a monitor-based backup strategy to
242	>	* blocking when needed.
243	>	*
244	>	* Recording WorkQueues.  WorkQueues are recorded in the
245	>	* "workQueues" array that is created upon first use and expanded
246	>	* if necessary.  Updates to the array while recording new workers
247	>	* and unrecording terminated ones are protected from each other
248	>	* by a lock but the array is otherwise concurrently readable, and
249	>	* accessed directly.  To simplify index-based operations, the
250	>	* array size is always a power of two, and all readers must
251	>	* tolerate null slots. Worker queues are at odd indices Shared
252	>	* (submission) queues are at even indices, up to a maximum of 64
253	>	* slots, to limit growth even if array needs to expand to add
254	>	* more workers. Grouping them together in this way simplifies and
255	>	* speeds up task scanning.
256	>	*
257	>	* All worker thread creation is on-demand, triggered by task
258	>	* submissions, replacement of terminated workers, and/or
259	>	* compensation for blocked workers. However, all other support
260	>	* code is set up to work with other policies.  To ensure that we
261	>	* do not hold on to worker references that would prevent GC, ALL
262	>	* accesses to workQueues are via indices into the workQueues
263	>	* array (which is one source of some of the messy code
264	>	* constructions here). In essence, the workQueues array serves as
265	>	* a weak reference mechanism. Thus for example the wait queue
266	>	* field of ctl stores indices, not references.  Access to the
267	>	* workQueues in associated methods (for example signalWork) must
268	>	* both index-check and null-check the IDs. All such accesses
269	>	* ignore bad IDs by returning out early from what they are doing,
270	>	* since this can only be associated with termination, in which
271	>	* case it is OK to give up.  All uses of the workQueues array
272	>	* also check that it is non-null (even if previously
273	>	* non-null). This allows nulling during termination, which is
274	>	* currently not necessary, but remains an option for
275	>	* resource-revocation-based shutdown schemes. It also helps
276	>	* reduce JIT issuance of uncommon-trap code, which tends to
277	>	* unnecessarily complicate control flow in some methods.
278	>	*
279	>	* Event Queuing. Unlike HPC work-stealing frameworks, we cannot
280	>	* let workers spin indefinitely scanning for tasks when none can
281	>	* be found immediately, and we cannot start/resume workers unless
282	>	* there appear to be tasks available.  On the other hand, we must
283	>	* quickly prod them into action when new tasks are submitted or
284	>	* generated. In many usages, ramp-up time to activate workers is
285	>	* the main limiting factor in overall performance (this is
286	>	* compounded at program start-up by JIT compilation and
287	>	* allocation). So we try to streamline this as much as possible.
288	>	* We park/unpark workers after placing in an event wait queue
289	>	* when they cannot find work. This "queue" is actually a simple
290	>	* Treiber stack, headed by the "id" field of ctl, plus a 15bit
291	>	* counter value (that reflects the number of times a worker has
292	>	* been inactivated) to avoid ABA effects (we need only as many
293	>	* version numbers as worker threads). Successors are held in
294	>	* field WorkQueue.nextWait.  Queuing deals with several intrinsic
295	>	* races, mainly that a task-producing thread can miss seeing (and
296	>	* signalling) another thread that gave up looking for work but
297	>	* has not yet entered the wait queue. We solve this by requiring
298	>	* a full sweep of all workers (via repeated calls to method
299	>	* scan()) both before and after a newly waiting worker is added
300	>	* to the wait queue. During a rescan, the worker might release
301	>	* some other queued worker rather than itself, which has the same
302	>	* net effect. Because enqueued workers may actually be rescanning
303	>	* rather than waiting, we set and clear the "parker" field of
304	>	* WorkQueues to reduce unnecessary calls to unpark.  (This
305	>	* requires a secondary recheck to avoid missed signals.)  Note
306	>	* the unusual conventions about Thread.interrupts surrounding
307	>	* parking and other blocking: Because interrupts are used solely
308	>	* to alert threads to check termination, which is checked anyway
309	>	* upon blocking, we clear status (using Thread.interrupted)
310	>	* before any call to park, so that park does not immediately
311	>	* return due to status being set via some other unrelated call to
312	>	* interrupt in user code.
313	>	*
314	>	* Signalling.  We create or wake up workers only when there
315	>	* appears to be at least one task they might be able to find and
316	>	* execute. However, many other threads may notice the same task
317	>	* and each signal to wake up a thread that might take it. So in
318	>	* general, pools will be over-signalled.  When a submission is
319	>	* added or another worker adds a task to a queue that is
320	>	* apparently empty, they signal waiting workers (or trigger
321	>	* creation of new ones if fewer than the given parallelism level
322	>	* -- see signalWork).  These primary signals are buttressed by
323	>	* signals whenever other threads scan for work or do not have a
324	>	* task to process. On most platforms, signalling (unpark)
325	>	* overhead time is noticeably long, and the time between
326	>	* signalling a thread and it actually making progress can be very
327	>	* noticeably long, so it is worth offloading these delays from
328	>	* critical paths as much as possible.
329	>	*
330	>	* Trimming workers. To release resources after periods of lack of
331	>	* use, a worker starting to wait when the pool is quiescent will
332	>	* time out and terminate if the pool has remained quiescent for a
333	>	* given period -- a short period if there are more threads than
334	>	* parallelism, longer as the number of threads decreases. This
335	>	* will slowly propagate, eventually terminating all workers after
336	>	* periods of non-use.
337	>	*
338	>	* Shutdown and Termination. A call to shutdownNow atomically sets
339	>	* a plock bit and then (non-atomically) sets each worker's
340	>	* qlock status, cancels all unprocessed tasks, and wakes up
341	>	* all waiting workers.  Detecting whether termination should
342	>	* commence after a non-abrupt shutdown() call requires more work
343	>	* and bookkeeping. We need consensus about quiescence (i.e., that
344	>	* there is no more work). The active count provides a primary
345	>	* indication but non-abrupt shutdown still requires a rechecking
346	>	* scan for any workers that are inactive but not queued.
347	>	*
348	>	* Joining Tasks
349	>	* =============
350	>	*
351	>	* Any of several actions may be taken when one worker is waiting
352	>	* to join a task stolen (or always held) by another.  Because we
353	>	* are multiplexing many tasks on to a pool of workers, we can't
354	>	* just let them block (as in Thread.join).  We also cannot just
355	>	* reassign the joiner's run-time stack with another and replace
356	>	* it later, which would be a form of "continuation", that even if
357	>	* possible is not necessarily a good idea since we sometimes need
358	>	* both an unblocked task and its continuation to progress.
359	>	* Instead we combine two tactics:
360	>	*
361	>	*   Helping: Arranging for the joiner to execute some task that it
362	>	*      would be running if the steal had not occurred.
363	>	*
364	>	*   Compensating: Unless there are already enough live threads,
365	>	*      method tryCompensate() may create or re-activate a spare
366	>	*      thread to compensate for blocked joiners until they unblock.
367	>	*
368	>	* A third form (implemented in tryRemoveAndExec) amounts to
369	>	* helping a hypothetical compensator: If we can readily tell that
370	>	* a possible action of a compensator is to steal and execute the
371	>	* task being joined, the joining thread can do so directly,
372	>	* without the need for a compensation thread (although at the
373	>	* expense of larger run-time stacks, but the tradeoff is
374	>	* typically worthwhile).
375	>	*
376	>	* The ManagedBlocker extension API can't use helping so relies
377	>	* only on compensation in method awaitBlocker.
378	>	*
379	>	* The algorithm in tryHelpStealer entails a form of "linear"
380	>	* helping: Each worker records (in field currentSteal) the most
381	>	* recent task it stole from some other worker. Plus, it records
382	>	* (in field currentJoin) the task it is currently actively
383	>	* joining. Method tryHelpStealer uses these markers to try to
384	>	* find a worker to help (i.e., steal back a task from and execute
385	>	* it) that could hasten completion of the actively joined task.
386	>	* In essence, the joiner executes a task that would be on its own
387	>	* local deque had the to-be-joined task not been stolen. This may
388	>	* be seen as a conservative variant of the approach in Wagner &
389	>	* Calder "Leapfrogging: a portable technique for implementing
390	>	* efficient futures" SIGPLAN Notices, 1993
391	>	* (http://portal.acm.org/citation.cfm?id=155354). It differs in
392	>	* that: (1) We only maintain dependency links across workers upon
393	>	* steals, rather than use per-task bookkeeping.  This sometimes
394	>	* requires a linear scan of workQueues array to locate stealers,
395	>	* but often doesn't because stealers leave hints (that may become
396	>	* stale/wrong) of where to locate them.  A stealHint is only a
397	>	* hint because a worker might have had multiple steals and the
398	>	* hint records only one of them (usually the most current).
399	>	* Hinting isolates cost to when it is needed, rather than adding
400	>	* to per-task overhead.  (2) It is "shallow", ignoring nesting
401	>	* and potentially cyclic mutual steals.  (3) It is intentionally
402	>	* racy: field currentJoin is updated only while actively joining,
403	>	* which means that we miss links in the chain during long-lived
404	>	* tasks, GC stalls etc (which is OK since blocking in such cases
405	>	* is usually a good idea).  (4) We bound the number of attempts
406	>	* to find work (see MAX_HELP) and fall back to suspending the
407	>	* worker and if necessary replacing it with another.
408	>	*
409	>	* Helping actions for CountedCompleters are much simpler: Method
410	>	* helpComplete can take and execute any task with the same root
411	>	* as the task being waited on. However, this still entails some
412	>	* traversal of completer chains, so is less efficient than using
413	>	* CountedCompleters without explicit joins.
414	>	*
415	>	* It is impossible to keep exactly the target parallelism number
416	>	* of threads running at any given time.  Determining the
417	>	* existence of conservatively safe helping targets, the
418	>	* availability of already-created spares, and the apparent need
419	>	* to create new spares are all racy, so we rely on multiple
420	>	* retries of each.  Compensation in the apparent absence of
421	>	* helping opportunities is challenging to control on JVMs, where
422	>	* GC and other activities can stall progress of tasks that in
423	>	* turn stall out many other dependent tasks, without us being
424	>	* able to determine whether they will ever require compensation.
425	>	* Even though work-stealing otherwise encounters little
426	>	* degradation in the presence of more threads than cores,
427	>	* aggressively adding new threads in such cases entails risk of
428	>	* unwanted positive feedback control loops in which more threads
429	>	* cause more dependent stalls (as well as delayed progress of
430	>	* unblocked threads to the point that we know they are available)
431	>	* leading to more situations requiring more threads, and so
432	>	* on. This aspect of control can be seen as an (analytically
433	>	* intractable) game with an opponent that may choose the worst
434	>	* (for us) active thread to stall at any time.  We take several
435	>	* precautions to bound losses (and thus bound gains), mainly in
436	>	* methods tryCompensate and awaitJoin.
437	>	*
438	>	* Common Pool
439	>	* ===========
440	>	*
441	>	* The static commonPool always exists after static
442	>	* initialization.  Since it (or any other created pool) need
443	>	* never be used, we minimize initial construction overhead and
444	>	* footprint to the setup of about a dozen fields, with no nested
445	>	* allocation. Most bootstrapping occurs within method
446	>	* fullExternalPush during the first submission to the pool.
447	>	*
448	>	* When external threads submit to the common pool, they can
449	>	* perform some subtask processing (see externalHelpJoin and
450	>	* related methods).  We do not need to record whether these
451	>	* submissions are to the common pool -- if not, externalHelpJoin
452	>	* returns quickly (at the most helping to signal some common pool
453	>	* workers). These submitters would otherwise be blocked waiting
454	>	* for completion, so the extra effort (with liberally sprinkled
455	>	* task status checks) in inapplicable cases amounts to an odd
456	>	* form of limited spin-wait before blocking in ForkJoinTask.join.
457	>	*
458	>	* Style notes
459	>	* ===========
460	>	*
461	>	* There is a lot of representation-level coupling among classes
462	>	* ForkJoinPool, ForkJoinWorkerThread, and ForkJoinTask.  The
463	>	* fields of WorkQueue maintain data structures managed by
464	>	* ForkJoinPool, so are directly accessed.  There is little point
465		* trying to reduce this, since any associated future changes in
466		* representations will need to be accompanied by algorithmic
467	<	* changes anyway.
468	<	*
469	<	* Style notes: There are lots of inline assignments (of form
470	<	* "while ((local = field) != 0)") which are usually the simplest
471	<	* way to ensure read orderings. Also several occurrences of the
472	<	* unusual "do {} while(!cas...)" which is the simplest way to
473	<	* force an update of a CAS'ed variable. There are also other
474	<	* coding oddities that help some methods perform reasonably even
475	<	* when interpreted (not compiled), at the expense of messiness.
476	<	*
477	<	* The order of declarations in this file is: (1) statics (2)
478	<	* fields (along with constants used when unpacking some of them)
479	<	* (3) internal control methods (4) callbacks and other support
480	<	* for ForkJoinTask and ForkJoinWorkerThread classes, (5) exported
481	<	* methods (plus a few little helpers).
467	>	* changes anyway. Several methods intrinsically sprawl because
468	>	* they must accumulate sets of consistent reads of volatiles held
469	>	* in local variables.  Methods signalWork() and scan() are the
470	>	* main bottlenecks, so are especially heavily
471	>	* micro-optimized/mangled.  There are lots of inline assignments
472	>	* (of form "while ((local = field) != 0)") which are usually the
473	>	* simplest way to ensure the required read orderings (which are
474	>	* sometimes critical). This leads to a "C"-like style of listing
475	>	* declarations of these locals at the heads of methods or blocks.
476	>	* There are several occurrences of the unusual "do {} while
477	>	* (!cas...)"  which is the simplest way to force an update of a
478	>	* CAS'ed variable. There are also other coding oddities (including
479	>	* several unnecessary-looking hoisted null checks) that help
480	>	* some methods perform reasonably even when interpreted (not
481	>	* compiled).
482	>	*
483	>	* The order of declarations in this file is:
484	>	* (1) Static utility functions
485	>	* (2) Nested (static) classes
486	>	* (3) Static fields
487	>	* (4) Fields, along with constants used when unpacking some of them
488	>	* (5) Internal control methods
489	>	* (6) Callbacks and other support for ForkJoinTask methods
490	>	* (7) Exported methods
491	>	* (8) Static block initializing statics in minimally dependent order
492	>	*/
493	>
494	>	// Static utilities
495	>
496	>	/**
497	>	* If there is a security manager, makes sure caller has
498	>	* permission to modify threads.
499		*/
500	+	private static void checkPermission() {
501	+	SecurityManager security = System.getSecurityManager();
502	+	if (security != null)
503	+	security.checkPermission(modifyThreadPermission);
504	+	}
505	+
506	+	// Nested classes
507
508		/**
509		* Factory for creating new {@link ForkJoinWorkerThread}s.
#	Line 376 | Line 533 | public class ForkJoinPool extends Abstra
533		}
534
535		/**
536	<	* Creates a new ForkJoinWorkerThread. This factory is used unless
537	<	* overridden in ForkJoinPool constructors.
536	>	* Class for artificial tasks that are used to replace the target
537	>	* of local joins if they are removed from an interior queue slot
538	>	* in WorkQueue.tryRemoveAndExec. We don't need the proxy to
539	>	* actually do anything beyond having a unique identity.
540	>	*/
541	>	static final class EmptyTask extends ForkJoinTask<Void> {
542	>	private static final long serialVersionUID = -7721805057305804111L;
543	>	EmptyTask() { status = ForkJoinTask.NORMAL; } // force done
544	>	public final Void getRawResult() { return null; }
545	>	public final void setRawResult(Void x) {}
546	>	public final boolean exec() { return true; }
547	>	}
548	>
549	>	/**
550	>	* Queues supporting work-stealing as well as external task
551	>	* submission. See above for main rationale and algorithms.
552	>	* Implementation relies heavily on "Unsafe" intrinsics
553	>	* and selective use of "volatile":
554	>	*
555	>	* Field "base" is the index (mod array.length) of the least valid
556	>	* queue slot, which is always the next position to steal (poll)
557	>	* from if nonempty. Reads and writes require volatile orderings
558	>	* but not CAS, because updates are only performed after slot
559	>	* CASes.
560	>	*
561	>	* Field "top" is the index (mod array.length) of the next queue
562	>	* slot to push to or pop from. It is written only by owner thread
563	>	* for push, or under lock for external/shared push, and accessed
564	>	* by other threads only after reading (volatile) base.  Both top
565	>	* and base are allowed to wrap around on overflow, but (top -
566	>	* base) (or more commonly -(base - top) to force volatile read of
567	>	* base before top) still estimates size. The lock ("qlock") is
568	>	* forced to -1 on termination, causing all further lock attempts
569	>	* to fail. (Note: we don't need CAS for termination state because
570	>	* upon pool shutdown, all shared-queues will stop being used
571	>	* anyway.)  Nearly all lock bodies are set up so that exceptions
572	>	* within lock bodies are "impossible" (modulo JVM errors that
573	>	* would cause failure anyway.)
574	>	*
575	>	* The array slots are read and written using the emulation of
576	>	* volatiles/atomics provided by Unsafe. Insertions must in
577	>	* general use putOrderedObject as a form of releasing store to
578	>	* ensure that all writes to the task object are ordered before
579	>	* its publication in the queue.  All removals entail a CAS to
580	>	* null.  The array is always a power of two. To ensure safety of
581	>	* Unsafe array operations, all accesses perform explicit null
582	>	* checks and implicit bounds checks via power-of-two masking.
583	>	*
584	>	* In addition to basic queuing support, this class contains
585	>	* fields described elsewhere to control execution. It turns out
586	>	* to work better memory-layout-wise to include them in this class
587	>	* rather than a separate class.
588	>	*
589	>	* Performance on most platforms is very sensitive to placement of
590	>	* instances of both WorkQueues and their arrays -- we absolutely
591	>	* do not want multiple WorkQueue instances or multiple queue
592	>	* arrays sharing cache lines. (It would be best for queue objects
593	>	* and their arrays to share, but there is nothing available to
594	>	* help arrange that).  Unfortunately, because they are recorded
595	>	* in a common array, WorkQueue instances are often moved to be
596	>	* adjacent by garbage collectors. To reduce impact, we use field
597	>	* padding that works OK on common platforms; this effectively
598	>	* trades off slightly slower average field access for the sake of
599	>	* avoiding really bad worst-case access. (Until better JVM
600	>	* support is in place, this padding is dependent on transient
601	>	* properties of JVM field layout rules.)  We also take care in
602	>	* allocating, sizing and resizing the array. Non-shared queue
603	>	* arrays are initialized by workers before use. Others are
604	>	* allocated on first use.
605		*/
606	<	public static final ForkJoinWorkerThreadFactory
607	<	defaultForkJoinWorkerThreadFactory =
608	<	new DefaultForkJoinWorkerThreadFactory();
606	>	static final class WorkQueue {
607	>	/**
608	>	* Capacity of work-stealing queue array upon initialization.
609	>	* Must be a power of two; at least 4, but should be larger to
610	>	* reduce or eliminate cacheline sharing among queues.
611	>	* Currently, it is much larger, as a partial workaround for
612	>	* the fact that JVMs often place arrays in locations that
613	>	* share GC bookkeeping (especially cardmarks) such that
614	>	* per-write accesses encounter serious memory contention.
615	>	*/
616	>	static final int INITIAL_QUEUE_CAPACITY = 1 << 13;
617
618	<	/**
619	<	* Permission required for callers of methods that may start or
620	<	* kill threads.
621	<	*/
622	<	private static final RuntimePermission modifyThreadPermission =
623	<	new RuntimePermission("modifyThread");
618	>	/**
619	>	* Maximum size for queue arrays. Must be a power of two less
620	>	* than or equal to 1 << (31 - width of array entry) to ensure
621	>	* lack of wraparound of index calculations, but defined to a
622	>	* value a bit less than this to help users trap runaway
623	>	* programs before saturating systems.
624	>	*/
625	>	static final int MAXIMUM_QUEUE_CAPACITY = 1 << 26; // 64M
626
627	<	/**
628	<	* If there is a security manager, makes sure caller has
629	<	* permission to modify threads.
630	<	*/
631	<	private static void checkPermission() {
632	<	SecurityManager security = System.getSecurityManager();
633	<	if (security != null)
634	<	security.checkPermission(modifyThreadPermission);
627	>	int seed;                  // for random scanning; initialize nonzero
628	>	volatile int eventCount;   // encoded inactivation count; < 0 if inactive
629	>	int nextWait;              // encoded record of next event waiter
630	>	final int mode;            // lifo, fifo, or shared
631	>	int nsteals;               // cumulative number of steals
632	>	int poolIndex;             // index of this queue in pool (or 0)
633	>	int stealHint;             // index of most recent known stealer
634	>	volatile int qlock;        // 1: locked, -1: terminate; else 0
635	>	volatile int base;         // index of next slot for poll
636	>	int top;                   // index of next slot for push
637	>	ForkJoinTask<?>[] array;   // the elements (initially unallocated)
638	>	final ForkJoinPool pool;   // the containing pool (may be null)
639	>	final ForkJoinWorkerThread owner; // owning thread or null if shared
640	>	volatile Thread parker;    // == owner during call to park; else null
641	>	volatile ForkJoinTask<?> currentJoin;  // task being joined in awaitJoin
642	>	ForkJoinTask<?> currentSteal; // current non-local task being executed
643	>	// Heuristic padding to ameliorate unfortunate memory placements
644	>	Object p00, p01, p02, p03, p04, p05, p06, p07;
645	>	Object p08, p09, p0a, p0b, p0c, p0d, p0e;
646	>
647	>	WorkQueue(ForkJoinPool pool, ForkJoinWorkerThread owner, int mode) {
648	>	this.mode = mode;
649	>	this.pool = pool;
650	>	this.owner = owner;
651	>	// Place indices in the center of array (that is not yet allocated)
652	>	base = top = INITIAL_QUEUE_CAPACITY >>> 1;
653	>	}
654	>
655	>	/**
656	>	* Pushes a task. Call only by owner in unshared queues.
657	>	* Cases needing resizing or rejection are relayed to fullPush
658	>	* (that also handles shared queues).
659	>	*
660	>	* @param task the task. Caller must ensure non-null.
661	>	* @throw RejectedExecutionException if array cannot be resized
662	>	*/
663	>	final void push(ForkJoinTask<?> task) {
664	>	ForkJoinPool p; ForkJoinTask<?>[] a;
665	>	int s = top, n;
666	>	if ((a = array) != null && a.length > (n = s + 1 - base)) {
667	>	U.putOrderedObject
668	>	(a, (((a.length - 1) & s) << ASHIFT) + ABASE, task);
669	>	top = s + 1;
670	>	if (n <= 1 && (p = pool) != null)
671	>	p.signalWork(this, 1);
672	>	}
673	>	else
674	>	fullPush(task, true);
675	>	}
676	>
677	>	/**
678	>	* Pushes a task if lock is free and array is either big
679	>	* enough or can be resized to be big enough. Note: a
680	>	* specialization of a common fast path of this method is in
681	>	* ForkJoinPool.externalPush. When called from a FJWT queue,
682	>	* this can fail only if the pool has been shut down or
683	>	* an out of memory error.
684	>	*
685	>	* @param task the task. Caller must ensure non-null.
686	>	* @param owned if true, throw RJE on failure
687	>	*/
688	>	final boolean fullPush(ForkJoinTask<?> task, boolean owned) {
689	>	ForkJoinPool p; ForkJoinTask<?>[] a;
690	>	if (owned) {
691	>	if (qlock < 0) // must be shutting down
692	>	throw new RejectedExecutionException();
693	>	}
694	>	else if (!U.compareAndSwapInt(this, QLOCK, 0, 1))
695	>	return false;
696	>	try {
697	>	int s = top, oldLen, len;
698	>	if ((a = array) == null)
699	>	a = array = new ForkJoinTask<?>[len=INITIAL_QUEUE_CAPACITY];
700	>	else if ((oldLen = a.length) > s + 1 - base)
701	>	len = oldLen;
702	>	else if ((len = oldLen << 1) > MAXIMUM_QUEUE_CAPACITY)
703	>	throw new RejectedExecutionException("Capacity exceeded");
704	>	else {
705	>	int oldMask, b;
706	>	ForkJoinTask<?>[] oldA = a;
707	>	a = array = new ForkJoinTask<?>[len];
708	>	if ((oldMask = oldLen - 1) >= 0 && s - (b = base) > 0) {
709	>	int mask = len - 1;
710	>	do {
711	>	ForkJoinTask<?> x;
712	>	int oldj = ((b & oldMask) << ASHIFT) + ABASE;
713	>	int j    = ((b &    mask) << ASHIFT) + ABASE;
714	>	x = (ForkJoinTask<?>)
715	>	U.getObjectVolatile(oldA, oldj);
716	>	if (x != null &&
717	>	U.compareAndSwapObject(oldA, oldj, x, null))
718	>	U.putObjectVolatile(a, j, x);
719	>	} while (++b != s);
720	>	}
721	>	}
722	>	U.putOrderedObject
723	>	(a, (((len - 1) & s) << ASHIFT) + ABASE, task);
724	>	top = s + 1;
725	>	} finally {
726	>	if (!owned)
727	>	qlock = 0;
728	>	}
729	>	if ((p = pool) != null)
730	>	p.signalWork(this, 1);
731	>	return true;
732	>	}
733	>
734	>	/**
735	>	* Takes next task, if one exists, in LIFO order.  Call only
736	>	* by owner in unshared queues.
737	>	*/
738	>	final ForkJoinTask<?> pop() {
739	>	ForkJoinTask<?>[] a; ForkJoinTask<?> t; int m;
740	>	if ((a = array) != null && (m = a.length - 1) >= 0) {
741	>	for (int s; (s = top - 1) - base >= 0;) {
742	>	long j = ((m & s) << ASHIFT) + ABASE;
743	>	if ((t = (ForkJoinTask<?>)U.getObject(a, j)) == null)
744	>	break;
745	>	if (U.compareAndSwapObject(a, j, t, null)) {
746	>	top = s;
747	>	return t;
748	>	}
749	>	}
750	>	}
751	>	return null;
752	>	}
753	>
754	>	/**
755	>	* Takes a task in FIFO order if b is base of queue and a task
756	>	* can be claimed without contention. Specialized versions
757	>	* appear in ForkJoinPool methods scan and tryHelpStealer.
758	>	*/
759	>	final ForkJoinTask<?> pollAt(int b) {
760	>	ForkJoinTask<?> t; ForkJoinTask<?>[] a;
761	>	if ((a = array) != null) {
762	>	int j = (((a.length - 1) & b) << ASHIFT) + ABASE;
763	>	if ((t = (ForkJoinTask<?>)U.getObjectVolatile(a, j)) != null &&
764	>	base == b &&
765	>	U.compareAndSwapObject(a, j, t, null)) {
766	>	base = b + 1;
767	>	return t;
768	>	}
769	>	}
770	>	return null;
771	>	}
772	>
773	>	/**
774	>	* Takes next task, if one exists, in FIFO order.
775	>	*/
776	>	final ForkJoinTask<?> poll() {
777	>	ForkJoinTask<?>[] a; int b; ForkJoinTask<?> t;
778	>	while ((b = base) - top < 0 && (a = array) != null) {
779	>	int j = (((a.length - 1) & b) << ASHIFT) + ABASE;
780	>	t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
781	>	if (t != null) {
782	>	if (base == b &&
783	>	U.compareAndSwapObject(a, j, t, null)) {
784	>	base = b + 1;
785	>	return t;
786	>	}
787	>	}
788	>	else if (base == b) {
789	>	if (b + 1 == top)
790	>	break;
791	>	Thread.yield(); // wait for lagging update (very rare)
792	>	}
793	>	}
794	>	return null;
795	>	}
796	>
797	>	/**
798	>	* Takes next task, if one exists, in order specified by mode.
799	>	*/
800	>	final ForkJoinTask<?> nextLocalTask() {
801	>	return mode == 0 ? pop() : poll();
802	>	}
803	>
804	>	/**
805	>	* Returns next task, if one exists, in order specified by mode.
806	>	*/
807	>	final ForkJoinTask<?> peek() {
808	>	ForkJoinTask<?>[] a = array; int m;
809	>	if (a == null || (m = a.length - 1) < 0)
810	>	return null;
811	>	int i = mode == 0 ? top - 1 : base;
812	>	int j = ((i & m) << ASHIFT) + ABASE;
813	>	return (ForkJoinTask<?>)U.getObjectVolatile(a, j);
814	>	}
815	>
816	>	/**
817	>	* Pops the given task only if it is at the current top.
818	>	* (A shared version is available only via FJP.tryExternalUnpush)
819	>	*/
820	>	final boolean tryUnpush(ForkJoinTask<?> t) {
821	>	ForkJoinTask<?>[] a; int s;
822	>	if ((a = array) != null && (s = top) != base &&
823	>	U.compareAndSwapObject
824	>	(a, (((a.length - 1) & --s) << ASHIFT) + ABASE, t, null)) {
825	>	top = s;
826	>	return true;
827	>	}
828	>	return false;
829	>	}
830	>
831	>	/**
832	>	* Removes and cancels all known tasks, ignoring any exceptions.
833	>	*/
834	>	final void cancelAll() {
835	>	ForkJoinTask.cancelIgnoringExceptions(currentJoin);
836	>	ForkJoinTask.cancelIgnoringExceptions(currentSteal);
837	>	for (ForkJoinTask<?> t; (t = poll()) != null; )
838	>	ForkJoinTask.cancelIgnoringExceptions(t);
839	>	}
840	>
841	>	/**
842	>	* Computes next value for random probes.  Scans don't require
843	>	* a very high quality generator, but also not a crummy one.
844	>	* Marsaglia xor-shift is cheap and works well enough.  Note:
845	>	* This is manually inlined in its usages in ForkJoinPool to
846	>	* avoid writes inside busy scan loops.
847	>	*/
848	>	final int nextSeed() {
849	>	int r = seed;
850	>	r ^= r << 13;
851	>	r ^= r >>> 17;
852	>	return seed = r ^= r << 5;
853	>	}
854	>
855	>	/**
856	>	* Provides a more accurate estimate of size than (top - base)
857	>	* by ordering reads and checking whether a near-empty queue
858	>	* has at least one unclaimed task.
859	>	*/
860	>	final int queueSize() {
861	>	ForkJoinTask<?>[] a; int k, s, n;
862	>	return ((n = base - (s = top)) < 0 &&
863	>	(n != -1 ||
864	>	((a = array) != null && (k = a.length) > 0 &&
865	>	U.getObject
866	>	(a, (long)((((k - 1) & (s - 1)) << ASHIFT) + ABASE)) != null))) ?
867	>	-n : 0;
868	>	}
869	>
870	>	// Specialized execution methods
871	>
872	>	/**
873	>	* Pops and runs tasks until empty.
874	>	*/
875	>	private void popAndExecAll() {
876	>	// A bit faster than repeated pop calls
877	>	ForkJoinTask<?>[] a; int m, s; long j; ForkJoinTask<?> t;
878	>	while ((a = array) != null && (m = a.length - 1) >= 0 &&
879	>	(s = top - 1) - base >= 0 &&
880	>	(t = ((ForkJoinTask<?>)
881	>	U.getObject(a, j = ((m & s) << ASHIFT) + ABASE)))
882	>	!= null) {
883	>	if (U.compareAndSwapObject(a, j, t, null)) {
884	>	top = s;
885	>	t.doExec();
886	>	}
887	>	}
888	>	}
889	>
890	>	/**
891	>	* Polls and runs tasks until empty.
892	>	*/
893	>	private void pollAndExecAll() {
894	>	for (ForkJoinTask<?> t; (t = poll()) != null;)
895	>	t.doExec();
896	>	}
897	>
898	>	/**
899	>	* If present, removes from queue and executes the given task,
900	>	* or any other cancelled task. Returns (true) on any CAS
901	>	* or consistency check failure so caller can retry.
902	>	*
903	>	* @return false if no progress can be made, else true;
904	>	*/
905	>	final boolean tryRemoveAndExec(ForkJoinTask<?> task) {
906	>	boolean stat = true, removed = false, empty = true;
907	>	ForkJoinTask<?>[] a; int m, s, b, n;
908	>	if ((a = array) != null && (m = a.length - 1) >= 0 &&
909	>	(n = (s = top) - (b = base)) > 0) {
910	>	for (ForkJoinTask<?> t;;) {           // traverse from s to b
911	>	int j = ((--s & m) << ASHIFT) + ABASE;
912	>	t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
913	>	if (t == null)                    // inconsistent length
914	>	break;
915	>	else if (t == task) {
916	>	if (s + 1 == top) {           // pop
917	>	if (!U.compareAndSwapObject(a, j, task, null))
918	>	break;
919	>	top = s;
920	>	removed = true;
921	>	}
922	>	else if (base == b)           // replace with proxy
923	>	removed = U.compareAndSwapObject(a, j, task,
924	>	new EmptyTask());
925	>	break;
926	>	}
927	>	else if (t.status >= 0)
928	>	empty = false;
929	>	else if (s + 1 == top) {          // pop and throw away
930	>	if (U.compareAndSwapObject(a, j, t, null))
931	>	top = s;
932	>	break;
933	>	}
934	>	if (--n == 0) {
935	>	if (!empty && base == b)
936	>	stat = false;
937	>	break;
938	>	}
939	>	}
940	>	}
941	>	if (removed)
942	>	task.doExec();
943	>	return stat;
944	>	}
945	>
946	>	/**
947	>	* Polls for and executes the given task or any other task in
948	>	* its CountedCompleter computation
949	>	*/
950	>	final boolean pollAndExecCC(ForkJoinTask<?> root) {
951	>	ForkJoinTask<?>[] a; int b; Object o;
952	>	outer: while ((b = base) - top < 0 && (a = array) != null) {
953	>	long j = (((a.length - 1) & b) << ASHIFT) + ABASE;
954	>	if ((o = U.getObject(a, j)) == null ||
955	>	!(o instanceof CountedCompleter))
956	>	break;
957	>	for (CountedCompleter<?> t = (CountedCompleter<?>)o, r = t;;) {
958	>	if (r == root) {
959	>	if (base == b &&
960	>	U.compareAndSwapObject(a, j, t, null)) {
961	>	base = b + 1;
962	>	t.doExec();
963	>	return true;
964	>	}
965	>	else
966	>	break; // restart
967	>	}
968	>	if ((r = r.completer) == null)
969	>	break outer; // not part of root computation
970	>	}
971	>	}
972	>	return false;
973	>	}
974	>
975	>	/**
976	>	* Executes a top-level task and any local tasks remaining
977	>	* after execution.
978	>	*/
979	>	final void runTask(ForkJoinTask<?> t) {
980	>	if (t != null) {
981	>	(currentSteal = t).doExec();
982	>	currentSteal = null;
983	>	if (++nsteals < 0) {     // spill on overflow
984	>	ForkJoinPool p;
985	>	if ((p = pool) != null)
986	>	p.collectStealCount(this);
987	>	}
988	>	if (top != base) {       // process remaining local tasks
989	>	if (mode == 0)
990	>	popAndExecAll();
991	>	else
992	>	pollAndExecAll();
993	>	}
994	>	}
995	>	}
996	>
997	>	/**
998	>	* Executes a non-top-level (stolen) task.
999	>	*/
1000	>	final void runSubtask(ForkJoinTask<?> t) {
1001	>	if (t != null) {
1002	>	ForkJoinTask<?> ps = currentSteal;
1003	>	(currentSteal = t).doExec();
1004	>	currentSteal = ps;
1005	>	}
1006	>	}
1007	>
1008	>	/**
1009	>	* Returns true if owned and not known to be blocked.
1010	>	*/
1011	>	final boolean isApparentlyUnblocked() {
1012	>	Thread wt; Thread.State s;
1013	>	return (eventCount >= 0 &&
1014	>	(wt = owner) != null &&
1015	>	(s = wt.getState()) != Thread.State.BLOCKED &&
1016	>	s != Thread.State.WAITING &&
1017	>	s != Thread.State.TIMED_WAITING);
1018	>	}
1019	>
1020	>	/**
1021	>	* If this owned and is not already interrupted, try to
1022	>	* interrupt and/or unpark, ignoring exceptions.
1023	>	*/
1024	>	final void interruptOwner() {
1025	>	Thread wt, p;
1026	>	if ((wt = owner) != null && !wt.isInterrupted()) {
1027	>	try {
1028	>	wt.interrupt();
1029	>	} catch (SecurityException ignore) {
1030	>	}
1031	>	}
1032	>	if ((p = parker) != null)
1033	>	U.unpark(p);
1034	>	}
1035	>
1036	>	// Unsafe mechanics
1037	>	private static final sun.misc.Unsafe U;
1038	>	private static final long QLOCK;
1039	>	private static final int ABASE;
1040	>	private static final int ASHIFT;
1041	>	static {
1042	>	int s;
1043	>	try {
1044	>	U = getUnsafe();
1045	>	Class<?> k = WorkQueue.class;
1046	>	Class<?> ak = ForkJoinTask[].class;
1047	>	QLOCK = U.objectFieldOffset
1048	>	(k.getDeclaredField("qlock"));
1049	>	ABASE = U.arrayBaseOffset(ak);
1050	>	s = U.arrayIndexScale(ak);
1051	>	} catch (Exception e) {
1052	>	throw new Error(e);
1053	>	}
1054	>	if ((s & (s-1)) != 0)
1055	>	throw new Error("data type scale not a power of two");
1056	>	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
1057	>	}
1058		}
1059
1060		/**
1061	<	* Generator for assigning sequence numbers as pool names.
1061	>	* Per-thread records for threads that submit to pools. Currently
1062	>	* holds only pseudo-random seed / index that is used to choose
1063	>	* submission queues in method externalPush. In the future, this may
1064	>	* also incorporate a means to implement different task rejection
1065	>	* and resubmission policies.
1066	>	*
1067	>	* Seeds for submitters and workers/workQueues work in basically
1068	>	* the same way but are initialized and updated using slightly
1069	>	* different mechanics. Both are initialized using the same
1070	>	* approach as in class ThreadLocal, where successive values are
1071	>	* unlikely to collide with previous values. Seeds are then
1072	>	* randomly modified upon collisions using xorshifts, which
1073	>	* requires a non-zero seed.
1074		*/
1075	<	private static final AtomicInteger poolNumberGenerator =
1076	<	new AtomicInteger();
1075	>	static final class Submitter {
1076	>	int seed;
1077	>	Submitter(int s) { seed = s; }
1078	>	}
1079
1080	<	/**
1081	<	* Absolute bound for parallelism level. Twice this number must
1082	<	* fit into a 16bit field to enable word-packing for some counts.
1083	<	*/
1084	<	private static final int MAX_THREADS = 0x7fff;
1080	>	/** Property prefix for constructing common pool */
1081	>	private static final String propPrefix =
1082	>	"java.util.concurrent.ForkJoinPool.common.";
1083	>
1084	>	// static fields (initialized in static initializer below)
1085
1086		/**
1087	<	* Array holding all worker threads in the pool.  Array size must
1088	<	* be a power of two.  Updates and replacements are protected by
418	<	* workerLock, but the array is always kept in a consistent enough
419	<	* state to be randomly accessed without locking by workers
420	<	* performing work-stealing, as well as other traversal-based
421	<	* methods in this class. All readers must tolerate that some
422	<	* array slots may be null.
1087	>	* Creates a new ForkJoinWorkerThread. This factory is used unless
1088	>	* overridden in ForkJoinPool constructors.
1089		*/
1090	<	volatile ForkJoinWorkerThread[] workers;
1090	>	public static final ForkJoinWorkerThreadFactory
1091	>	defaultForkJoinWorkerThreadFactory;
1092
1093		/**
1094	<	* Queue for external submissions.
1094	>	* Common (static) pool. Non-null for public use unless a static
1095	>	* construction exception, but internal usages null-check on use
1096	>	* to paranoically avoid potential initialization circularities
1097	>	* as well as to simplify generated code.
1098		*/
1099	<	private final LinkedTransferQueue<ForkJoinTask<?>> submissionQueue;
1099	>	static final ForkJoinPool commonPool;
1100
1101		/**
1102	<	* Lock protecting updates to workers array.
1102	>	* Permission required for callers of methods that may start or
1103	>	* kill threads.
1104		*/
1105	<	private final ReentrantLock workerLock;
1105	>	private static final RuntimePermission modifyThreadPermission;
1106
1107		/**
1108	<	* Latch released upon termination.
1108	>	* Per-thread submission bookkeeping. Shared across all pools
1109	>	* to reduce ThreadLocal pollution and because random motion
1110	>	* to avoid contention in one pool is likely to hold for others.
1111	>	* Lazily initialized on first submission (but null-checked
1112	>	* in other contexts to avoid unnecessary initialization).
1113		*/
1114	<	private final Phaser termination;
1114	>	static final ThreadLocal<Submitter> submitters;
1115
1116		/**
1117	<	* Creation factory for worker threads.
1117	>	* Common pool parallelism. Must equal commonPool.parallelism.
1118		*/
1119	<	private final ForkJoinWorkerThreadFactory factory;
1119	>	static final int commonPoolParallelism;
1120
1121		/**
1122	<	* Sum of per-thread steal counts, updated only when threads are
448	<	* idle or terminating.
1122	>	* Sequence number for creating workerNamePrefix.
1123		*/
1124	<	private volatile long stealCount;
1124	>	private static int poolNumberSequence;
1125
1126		/**
1127	<	* Encoded record of top of treiber stack of threads waiting for
1128	<	* events. The top 32 bits contain the count being waited for. The
455	<	* bottom word contains one plus the pool index of waiting worker
456	<	* thread.
1127	>	* Return the next sequence number. We don't expect this to
1128	>	* ever contend so use simple builtin sync.
1129		*/
1130	<	private volatile long eventWaiters;
1130	>	private static final synchronized int nextPoolId() {
1131	>	return ++poolNumberSequence;
1132	>	}
1133
1134	<	private static final int  EVENT_COUNT_SHIFT = 32;
461	<	private static final long WAITER_ID_MASK = (1L << EVENT_COUNT_SHIFT)-1L;
1134	>	// static constants
1135
1136		/**
1137	<	* A counter for events that may wake up worker threads:
1138	<	*   - Submission of a new task to the pool
1139	<	*   - A worker pushing a task on an empty queue
1140	<	*   - termination and reconfiguration
1137	>	* Initial timeout value (in nanoseconds) for the thread
1138	>	* triggering quiescence to park waiting for new work. On timeout,
1139	>	* the thread will instead try to shrink the number of
1140	>	* workers. The value should be large enough to avoid overly
1141	>	* aggressive shrinkage during most transient stalls (long GCs
1142	>	* etc).
1143		*/
1144	<	private volatile int eventCount;
1144	>	private static final long IDLE_TIMEOUT      = 2000L * 1000L * 1000L; // 2sec
1145
1146		/**
1147	<	* Lifecycle control. The low word contains the number of workers
473	<	* that are (probably) executing tasks. This value is atomically
474	<	* incremented before a worker gets a task to run, and decremented
475	<	* when worker has no tasks and cannot find any.  Bits 16-18
476	<	* contain runLevel value. When all are zero, the pool is
477	<	* running. Level transitions are monotonic (running -> shutdown
478	<	* -> terminating -> terminated) so each transition adds a bit.
479	<	* These are bundled together to ensure consistent read for
480	<	* termination checks (i.e., that runLevel is at least SHUTDOWN
481	<	* and active threads is zero).
1147	>	* Timeout value when there are more threads than parallelism level
1148		*/
1149	<	private volatile int runState;
484	<
485	<	// Note: The order among run level values matters.
486	<	private static final int RUNLEVEL_SHIFT     = 16;
487	<	private static final int SHUTDOWN           = 1 << RUNLEVEL_SHIFT;
488	<	private static final int TERMINATING        = 1 << (RUNLEVEL_SHIFT + 1);
489	<	private static final int TERMINATED         = 1 << (RUNLEVEL_SHIFT + 2);
490	<	private static final int ACTIVE_COUNT_MASK  = (1 << RUNLEVEL_SHIFT) - 1;
491	<	private static final int ONE_ACTIVE         = 1; // active update delta
1149	>	private static final long FAST_IDLE_TIMEOUT =  200L * 1000L * 1000L;
1150
1151		/**
1152	<	* Holds number of total (i.e., created and not yet terminated)
1153	<	* and running (i.e., not blocked on joins or other managed sync)
1154	<	* threads, packed together to ensure consistent snapshot when
1155	<	* making decisions about creating and suspending spare
1156	<	* threads. Updated only by CAS. Note that adding a new worker
1157	<	* requires incrementing both counts, since workers start off in
500	<	* running state.  This field is also used for memory-fencing
501	<	* configuration parameters.
1152	>	* The maximum stolen->joining link depth allowed in method
1153	>	* tryHelpStealer.  Must be a power of two.  Depths for legitimate
1154	>	* chains are unbounded, but we use a fixed constant to avoid
1155	>	* (otherwise unchecked) cycles and to bound staleness of
1156	>	* traversal parameters at the expense of sometimes blocking when
1157	>	* we could be helping.
1158		*/
1159	<	private volatile int workerCounts;
504	<
505	<	private static final int TOTAL_COUNT_SHIFT  = 16;
506	<	private static final int RUNNING_COUNT_MASK = (1 << TOTAL_COUNT_SHIFT) - 1;
507	<	private static final int ONE_RUNNING        = 1;
508	<	private static final int ONE_TOTAL          = 1 << TOTAL_COUNT_SHIFT;
1159	>	private static final int MAX_HELP = 64;
1160
1161		/**
1162	<	* The target parallelism level.
1163	<	* Accessed directly by ForkJoinWorkerThreads.
1162	>	* Increment for seed generators. See class ThreadLocal for
1163	>	* explanation.
1164		*/
1165	<	final int parallelism;
1165	>	private static final int SEED_INCREMENT = 0x61c88647;
1166
1167		/**
1168	<	* True if use local fifo, not default lifo, for local polling
1169	<	* Read by, and replicated by ForkJoinWorkerThreads
1168	>	* Bits and masks for control variables
1169	>	*
1170	>	* Field ctl is a long packed with:
1171	>	* AC: Number of active running workers minus target parallelism (16 bits)
1172	>	* TC: Number of total workers minus target parallelism (16 bits)
1173	>	* ST: true if pool is terminating (1 bit)
1174	>	* EC: the wait count of top waiting thread (15 bits)
1175	>	* ID: poolIndex of top of Treiber stack of waiters (16 bits)
1176	>	*
1177	>	* When convenient, we can extract the upper 32 bits of counts and
1178	>	* the lower 32 bits of queue state, u = (int)(ctl >>> 32) and e =
1179	>	* (int)ctl.  The ec field is never accessed alone, but always
1180	>	* together with id and st. The offsets of counts by the target
1181	>	* parallelism and the positionings of fields makes it possible to
1182	>	* perform the most common checks via sign tests of fields: When
1183	>	* ac is negative, there are not enough active workers, when tc is
1184	>	* negative, there are not enough total workers, and when e is
1185	>	* negative, the pool is terminating.  To deal with these possibly
1186	>	* negative fields, we use casts in and out of "short" and/or
1187	>	* signed shifts to maintain signedness.
1188	>	*
1189	>	* When a thread is queued (inactivated), its eventCount field is
1190	>	* set negative, which is the only way to tell if a worker is
1191	>	* prevented from executing tasks, even though it must continue to
1192	>	* scan for them to avoid queuing races. Note however that
1193	>	* eventCount updates lag releases so usage requires care.
1194	>	*
1195	>	* Field plock is an int packed with:
1196	>	* SHUTDOWN: true if shutdown is enabled (1 bit)
1197	>	* SEQ:  a sequence lock, with PL_LOCK bit set if locked (30 bits)
1198	>	* SIGNAL: set when threads may be waiting on the lock (1 bit)
1199	>	*
1200	>	* The sequence number enables simple consistency checks:
1201	>	* Staleness of read-only operations on the workQueues array can
1202	>	* be checked by comparing plock before vs after the reads.
1203		*/
520	–	final boolean locallyFifo;
1204
1205	<	/**
1206	<	* The uncaught exception handler used when any worker abruptly
1207	<	* terminates.
1208	<	*/
1209	<	private final Thread.UncaughtExceptionHandler ueh;
1205	>	// bit positions/shifts for fields
1206	>	private static final int  AC_SHIFT   = 48;
1207	>	private static final int  TC_SHIFT   = 32;
1208	>	private static final int  ST_SHIFT   = 31;
1209	>	private static final int  EC_SHIFT   = 16;
1210
1211	<	/**
1212	<	* Pool number, just for assigning useful names to worker threads
1213	<	*/
1214	<	private final int poolNumber;
1211	>	// bounds
1212	>	private static final int  SMASK      = 0xffff;  // short bits
1213	>	private static final int  MAX_CAP    = 0x7fff;  // max #workers - 1
1214	>	private static final int  EVENMASK   = 0xfffe;  // even short bits
1215	>	private static final int  SQMASK     = 0x007e;  // max 64 (even) slots
1216	>	private static final int  SHORT_SIGN = 1 << 15;
1217	>	private static final int  INT_SIGN   = 1 << 31;
1218
1219	<	// Utilities for CASing fields. Note that several of these
1220	<	// are manually inlined by callers
1219	>	// masks
1220	>	private static final long STOP_BIT   = 0x0001L << ST_SHIFT;
1221	>	private static final long AC_MASK    = ((long)SMASK) << AC_SHIFT;
1222	>	private static final long TC_MASK    = ((long)SMASK) << TC_SHIFT;
1223
1224	<	/**
1225	<	* Increments running count.  Also used by ForkJoinTask.
1226	<	*/
539	<	final void incrementRunningCount() {
540	<	int c;
541	<	do {} while (!UNSAFE.compareAndSwapInt(this, workerCountsOffset,
542	<	c = workerCounts,
543	<	c + ONE_RUNNING));
544	<	}
1224	>	// units for incrementing and decrementing
1225	>	private static final long TC_UNIT    = 1L << TC_SHIFT;
1226	>	private static final long AC_UNIT    = 1L << AC_SHIFT;
1227
1228	<	/**
1229	<	* Tries to decrement running count unless already zero
1230	<	*/
1231	<	final boolean tryDecrementRunningCount() {
1232	<	int wc = workerCounts;
1233	<	if ((wc & RUNNING_COUNT_MASK) == 0)
1234	<	return false;
553	<	return UNSAFE.compareAndSwapInt(this, workerCountsOffset,
554	<	wc, wc - ONE_RUNNING);
555	<	}
1228	>	// masks and units for dealing with u = (int)(ctl >>> 32)
1229	>	private static final int  UAC_SHIFT  = AC_SHIFT - 32;
1230	>	private static final int  UTC_SHIFT  = TC_SHIFT - 32;
1231	>	private static final int  UAC_MASK   = SMASK << UAC_SHIFT;
1232	>	private static final int  UTC_MASK   = SMASK << UTC_SHIFT;
1233	>	private static final int  UAC_UNIT   = 1 << UAC_SHIFT;
1234	>	private static final int  UTC_UNIT   = 1 << UTC_SHIFT;
1235
1236	<	/**
1237	<	* Tries to increment running count
1238	<	*/
560	<	final boolean tryIncrementRunningCount() {
561	<	int wc;
562	<	return UNSAFE.compareAndSwapInt(this, workerCountsOffset,
563	<	wc = workerCounts, wc + ONE_RUNNING);
564	<	}
1236	>	// masks and units for dealing with e = (int)ctl
1237	>	private static final int E_MASK      = 0x7fffffff; // no STOP_BIT
1238	>	private static final int E_SEQ       = 1 << EC_SHIFT;
1239
1240	<	/**
1241	<	* Tries incrementing active count; fails on contention.
1242	<	* Called by workers before executing tasks.
1243	<	*
1244	<	* @return true on success
571	<	*/
572	<	final boolean tryIncrementActiveCount() {
573	<	int c;
574	<	return UNSAFE.compareAndSwapInt(this, runStateOffset,
575	<	c = runState, c + ONE_ACTIVE);
576	<	}
1240	>	// plock bits
1241	>	private static final int SHUTDOWN    = 1 << 31;
1242	>	private static final int PL_LOCK     = 2;
1243	>	private static final int PL_SIGNAL   = 1;
1244	>	private static final int PL_SPINS    = 1 << 8;
1245
1246	<	/**
1247	<	* Tries decrementing active count; fails on contention.
1248	<	* Called when workers cannot find tasks to run.
1249	<	*/
582	<	final boolean tryDecrementActiveCount() {
583	<	int c;
584	<	return UNSAFE.compareAndSwapInt(this, runStateOffset,
585	<	c = runState, c - ONE_ACTIVE);
586	<	}
1246	>	// access mode for WorkQueue
1247	>	static final int LIFO_QUEUE          =  0;
1248	>	static final int FIFO_QUEUE          =  1;
1249	>	static final int SHARED_QUEUE        = -1;
1250
1251	<	/**
1252	<	* Advances to at least the given level. Returns true if not
1253	<	* already in at least the given level.
1251	>	// Instance fields
1252	>
1253	>	/*
1254	>	* Field layout order in this class tends to matter more than one
1255	>	* would like. Runtime layout order is only loosely related to
1256	>	* declaration order and may differ across JVMs, but the following
1257	>	* empirically works OK on current JVMs.
1258	>	*/
1259	>	volatile long stealCount;                  // collects worker counts
1260	>	volatile long ctl;                         // main pool control
1261	>	final int parallelism;                     // parallelism level
1262	>	final int localMode;                       // per-worker scheduling mode
1263	>	volatile int indexSeed;                    // worker/submitter index seed
1264	>	volatile int plock;                        // shutdown status and seqLock
1265	>	WorkQueue[] workQueues;                    // main registry
1266	>	final ForkJoinWorkerThreadFactory factory; // factory for new workers
1267	>	final Thread.UncaughtExceptionHandler ueh; // per-worker UEH
1268	>	final String workerNamePrefix;             // to create worker name string
1269	>
1270	>	/*
1271	>	* Acquires the plock lock to protect worker array and related
1272	>	* updates. This method is called only if an initial CAS on plock
1273	>	* fails. This acts as a spinLock for normal cases, but falls back
1274	>	* to builtin monitor to block when (rarely) needed. This would be
1275	>	* a terrible idea for a highly contended lock, but works fine as
1276	>	* a more conservative alternative to a pure spinlock.  See
1277	>	* internal ConcurrentHashMap documentation for further
1278	>	* explanation of nearly the same construction.
1279		*/
1280	<	private boolean advanceRunLevel(int level) {
1280	>	private int acquirePlock() {
1281	>	int spins = PL_SPINS, r = 0, ps, nps;
1282		for (;;) {
1283	<	int s = runState;
1284	<	if ((s & level) != 0)
1285	<	return false;
1286	<	if (UNSAFE.compareAndSwapInt(this, runStateOffset, s, s | level))
1287	<	return true;
1283	>	if (((ps = plock) & PL_LOCK) == 0 &&
1284	>	U.compareAndSwapInt(this, PLOCK, ps, nps = ps + PL_LOCK))
1285	>	return nps;
1286	>	else if (r == 0)
1287	>	r = ThreadLocalRandom.current().nextInt(); // randomize spins
1288	>	else if (spins >= 0) {
1289	>	r ^= r << 1; r ^= r >>> 3; r ^= r << 10; // xorshift
1290	>	if (r >= 0)
1291	>	--spins;
1292	>	}
1293	>	else if (U.compareAndSwapInt(this, PLOCK, ps, ps | PL_SIGNAL)) {
1294	>	synchronized (this) {
1295	>	if ((plock & PL_SIGNAL) != 0) {
1296	>	try {
1297	>	wait();
1298	>	} catch (InterruptedException ie) {
1299	>	try {
1300	>	Thread.currentThread().interrupt();
1301	>	} catch (SecurityException ignore) {
1302	>	}
1303	>	}
1304	>	}
1305	>	else
1306	>	notifyAll();
1307	>	}
1308	>	}
1309		}
1310		}
1311
602	–	// workers array maintenance
603	–
1312		/**
1313	<	* Records and returns a workers array index for new worker.
1313	>	* Unlocks and signals any thread waiting for plock. Called only
1314	>	* when CAS of seq value for unlock fails.
1315		*/
1316	<	private int recordWorker(ForkJoinWorkerThread w) {
1317	<	// Try using slot totalCount-1. If not available, scan and/or resize
1318	<	int k = (workerCounts >>> TOTAL_COUNT_SHIFT) - 1;
1319	<	final ReentrantLock lock = this.workerLock;
1320	<	lock.lock();
1316	>	private void releasePlock(int ps) {
1317	>	plock = ps;
1318	>	synchronized (this) { notifyAll(); }
1319	>	}
1320	>
1321	>	//  Registering and deregistering workers
1322	>
1323	>	/**
1324	>	* Callback from ForkJoinWorkerThread constructor to establish its
1325	>	* poolIndex and record its WorkQueue. To avoid scanning bias due
1326	>	* to packing entries in front of the workQueues array, we treat
1327	>	* the array as a simple power-of-two hash table using per-thread
1328	>	* seed as hash, expanding as needed.
1329	>	*
1330	>	* @param w the worker's queue
1331	>	*/
1332	>	final void registerWorker(WorkQueue w) {
1333	>	int s, ps; // generate a rarely colliding candidate index seed
1334	>	do {} while (!U.compareAndSwapInt(this, INDEXSEED,
1335	>	s = indexSeed, s += SEED_INCREMENT) ||
1336	>	s == 0); // skip 0
1337	>	if (((ps = plock) & PL_LOCK) != 0 ||
1338	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1339	>	ps = acquirePlock();
1340	>	int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1341		try {
1342	<	ForkJoinWorkerThread[] ws = workers;
1343	<	int nws = ws.length;
1344	<	if (k < 0 || k >= nws || ws[k] != null) {
1345	<	for (k = 0; k < nws && ws[k] != null; ++k)
1346	<	;
1347	<	if (k == nws)
1348	<	ws = Arrays.copyOf(ws, nws << 1);
1342	>	WorkQueue[] ws;
1343	>	if (w != null && (ws = workQueues) != null) {
1344	>	w.seed = s;
1345	>	int n = ws.length, m = n - 1;
1346	>	int r = (s << 1) | 1;               // use odd-numbered indices
1347	>	if (ws[r &= m] != null) {           // collision
1348	>	int probes = 0;                 // step by approx half size
1349	>	int step = (n <= 4) ? 2 : ((n >>> 1) & EVENMASK) + 2;
1350	>	while (ws[r = (r + step) & m] != null) {
1351	>	if (++probes >= n) {
1352	>	workQueues = ws = Arrays.copyOf(ws, n <<= 1);
1353	>	m = n - 1;
1354	>	probes = 0;
1355	>	}
1356	>	}
1357	>	}
1358	>	w.eventCount = w.poolIndex = r;     // establish before recording
1359	>	ws[r] = w;
1360		}
621	–	ws[k] = w;
622	–	workers = ws; // volatile array write ensures slot visibility
1361		} finally {
1362	<	lock.unlock();
1362	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1363	>	releasePlock(nps);
1364		}
626	–	return k;
1365		}
1366
1367		/**
1368	<	* Nulls out record of worker in workers array
1369	<	*/
1370	<	private void forgetWorker(ForkJoinWorkerThread w) {
1371	<	int idx = w.poolIndex;
1372	<	// Locking helps method recordWorker avoid unecessary expansion
1373	<	final ReentrantLock lock = this.workerLock;
1374	<	lock.lock();
1375	<	try {
1376	<	ForkJoinWorkerThread[] ws = workers;
1377	<	if (idx >= 0 && idx < ws.length && ws[idx] == w) // verify
1378	<	ws[idx] = null;
1379	<	} finally {
1380	<	lock.unlock();
1368	>	* Final callback from terminating worker, as well as upon failure
1369	>	* to construct or start a worker.  Removes record of worker from
1370	>	* array, and adjusts counts. If pool is shutting down, tries to
1371	>	* complete termination.
1372	>	*
1373	>	* @param wt the worker thread or null if construction failed
1374	>	* @param ex the exception causing failure, or null if none
1375	>	*/
1376	>	final void deregisterWorker(ForkJoinWorkerThread wt, Throwable ex) {
1377	>	WorkQueue w = null;
1378	>	if (wt != null && (w = wt.workQueue) != null) {
1379	>	int ps;
1380	>	collectStealCount(w);
1381	>	w.qlock = -1;                // ensure set
1382	>	if (((ps = plock) & PL_LOCK) != 0 ||
1383	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1384	>	ps = acquirePlock();
1385	>	int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1386	>	try {
1387	>	int idx = w.poolIndex;
1388	>	WorkQueue[] ws = workQueues;
1389	>	if (ws != null && idx >= 0 && idx < ws.length && ws[idx] == w)
1390	>	ws[idx] = null;
1391	>	} finally {
1392	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1393	>	releasePlock(nps);
1394	>	}
1395		}
644	–	}
1396
1397	<	// adding and removing workers
1397	>	long c;                             // adjust ctl counts
1398	>	do {} while (!U.compareAndSwapLong
1399	>	(this, CTL, c = ctl, (((c - AC_UNIT) & AC_MASK) |
1400	>	((c - TC_UNIT) & TC_MASK) |
1401	>	(c & ~(AC_MASK|TC_MASK)))));
1402	>
1403	>	if (!tryTerminate(false, false) && w != null) {
1404	>	w.cancelAll();                  // cancel remaining tasks
1405	>	if (w.array != null)            // suppress signal if never ran
1406	>	signalWork(null, 1);        // wake up or create replacement
1407	>	if (ex == null)                 // help clean refs on way out
1408	>	ForkJoinTask.helpExpungeStaleExceptions();
1409	>	}
1410	>
1411	>	if (ex != null)                     // rethrow
1412	>	ForkJoinTask.rethrow(ex);
1413	>	}
1414
1415		/**
1416	<	* Tries to create and add new worker. Assumes that worker counts
1417	<	* are already updated to accommodate the worker, so adjusts on
1418	<	* failure.
1419	<	*
1420	<	* @return new worker or null if creation failed
1421	<	*/
1422	<	private ForkJoinWorkerThread addWorker() {
1423	<	ForkJoinWorkerThread w = null;
1424	<	try {
1425	<	w = factory.newThread(this);
1426	<	} finally { // Adjust on either null or exceptional factory return
1427	<	if (w == null) {
1428	<	onWorkerCreationFailure();
1429	<	return null;
1416	>	* Collect worker steal count into total. Called on termination
1417	>	* and upon int overflow of local count. (There is a possible race
1418	>	* in the latter case vs any caller of getStealCount, which can
1419	>	* make its results less accurate than usual.)
1420	>	*/
1421	>	final void collectStealCount(WorkQueue w) {
1422	>	if (w != null) {
1423	>	long sc;
1424	>	int ns = w.nsteals;
1425	>	w.nsteals = 0; // handle overflow
1426	>	long steals = (ns >= 0) ? ns : 1L + (long)(Integer.MAX_VALUE);
1427	>	do {} while (!U.compareAndSwapLong(this, STEALCOUNT,
1428	>	sc = stealCount, sc + steals));
1429	>	}
1430	>	}
1431	>
1432	>	// Submissions
1433	>
1434	>	/**
1435	>	* Unless shutting down, adds the given task to a submission queue
1436	>	* at submitter's current queue index (modulo submission
1437	>	* range). Only the most common path is directly handled in this
1438	>	* method. All others are relayed to fullExternalPush.
1439	>	*
1440	>	* @param task the task. Caller must ensure non-null.
1441	>	*/
1442	>	final void externalPush(ForkJoinTask<?> task) {
1443	>	WorkQueue[] ws; WorkQueue q; Submitter z; int m; ForkJoinTask<?>[] a;
1444	>	if ((z = submitters.get()) != null && plock > 0 &&
1445	>	(ws = workQueues) != null && (m = (ws.length - 1)) >= 0 &&
1446	>	(q = ws[m & z.seed & SQMASK]) != null &&
1447	>	U.compareAndSwapInt(q, QLOCK, 0, 1)) { // lock
1448	>	int s = q.top, n;
1449	>	if ((a = q.array) != null && a.length > (n = s + 1 - q.base)) {
1450	>	U.putObject(a, (long)(((a.length - 1) & s) << ASHIFT) + ABASE,
1451	>	task);
1452	>	q.top = s + 1;                     // push on to deque
1453	>	q.qlock = 0;
1454	>	if (n <= 1)
1455	>	signalWork(q, 1);
1456	>	return;
1457		}
1458	+	q.qlock = 0;
1459		}
1460	<	w.start(recordWorker(w), ueh);
666	<	return w;
1460	>	fullExternalPush(task);
1461		}
1462
1463		/**
1464	<	* Adjusts counts upon failure to create worker
1465	<	*/
1466	<	private void onWorkerCreationFailure() {
1467	<	for (;;) {
1468	<	int wc = workerCounts;
1469	<	if ((wc >>> TOTAL_COUNT_SHIFT) == 0)
1470	<	Thread.yield(); // wait for other counts to settle
1471	<	else if (UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
1472	<	wc - (ONE_RUNNING|ONE_TOTAL)))
1473	<	break;
1464	>	* Full version of externalPush. This method is called, among
1465	>	* other times, upon the first submission of the first task to the
1466	>	* pool, so must perform secondary initialization: creating
1467	>	* workQueue array and setting plock to a valid value. It also
1468	>	* detects first submission by an external thread by looking up
1469	>	* its ThreadLocal, and creates a new shared queue if the one at
1470	>	* index if empty or contended. The lock bodies must be
1471	>	* exception-free (so no try/finally) so we optimistically
1472	>	* allocate new queues/arrays outside the locks and throw them
1473	>	* away if (very rarely) not needed. Note that the plock seq value
1474	>	* can eventually wrap around zero, but if so harmlessly fails to
1475	>	* reinitialize.
1476	>	*/
1477	>	private void fullExternalPush(ForkJoinTask<?> task) {
1478	>	for (Submitter z = null;;) {
1479	>	WorkQueue[] ws; WorkQueue q; int ps, m, r, s;
1480	>	if ((ps = plock) < 0)
1481	>	throw new RejectedExecutionException();
1482	>	else if ((ws = workQueues) == null || (m = ws.length - 1) < 0) {
1483	>	int n = parallelism - 1; n |= n >>> 1; n |= n >>> 2;
1484	>	n |= n >>> 4; n |= n >>> 8; n |= n >>> 16;
1485	>	WorkQueue[] nws = new WorkQueue[(n + 1) << 1]; // power of two
1486	>	if ((ps & PL_LOCK) != 0 ||
1487	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1488	>	ps = acquirePlock();
1489	>	if ((ws = workQueues) == null)
1490	>	workQueues = nws;
1491	>	int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1492	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1493	>	releasePlock(nps);
1494	>	}
1495	>	else if (z == null && (z = submitters.get()) == null) {
1496	>	if (U.compareAndSwapInt(this, INDEXSEED,
1497	>	s = indexSeed, s += SEED_INCREMENT) &&
1498	>	s != 0) // skip 0
1499	>	submitters.set(z = new Submitter(s));
1500	>	}
1501	>	else {
1502	>	int k = (r = z.seed) & m & SQMASK;
1503	>	if ((q = ws[k]) == null && (ps & PL_LOCK) == 0) {
1504	>	(q = new WorkQueue(this, null, SHARED_QUEUE)).poolIndex = k;
1505	>	if (((ps = plock) & PL_LOCK) != 0 ||
1506	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1507	>	ps = acquirePlock();
1508	>	WorkQueue w = null;
1509	>	if ((ws = workQueues) != null && k < ws.length &&
1510	>	(w = ws[k]) == null)
1511	>	ws[k] = q;
1512	>	else
1513	>	q = w;
1514	>	int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1515	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1516	>	releasePlock(nps);
1517	>	}
1518	>	if (q != null && q.qlock == 0 && q.fullPush(task, false))
1519	>	return;
1520	>	r ^= r << 13;                // same xorshift as WorkQueues
1521	>	r ^= r >>> 17;
1522	>	z.seed = r ^= r << 5;        // move to a different index
1523	>	}
1524		}
681	–	tryTerminate(false); // in case of failure during shutdown
1525		}
1526
1527	+	// Maintaining ctl counts
1528	+
1529		/**
1530	<	* Creates and/or resumes enough workers to establish target
686	<	* parallelism, giving up if terminating or addWorker fails
687	<	*
688	<	* TODO: recast this to support lazier creation and automated
689	<	* parallelism maintenance
1530	>	* Increments active count; mainly called upon return from blocking.
1531		*/
1532	<	private void ensureEnoughWorkers() {
1533	<	for (;;) {
1534	<	int pc = parallelism;
1535	<	int wc = workerCounts;
1536	<	int rc = wc & RUNNING_COUNT_MASK;
1537	<	int tc = wc >>> TOTAL_COUNT_SHIFT;
1538	<	if (tc < pc) {
1539	<	if (runState == TERMINATING ||
1540	<	(UNSAFE.compareAndSwapInt
1541	<	(this, workerCountsOffset,
1542	<	wc, wc + (ONE_RUNNING|ONE_TOTAL)) &&
1543	<	addWorker() == null))
1532	>	final void incrementActiveCount() {
1533	>	long c;
1534	>	do {} while (!U.compareAndSwapLong(this, CTL, c = ctl, c + AC_UNIT));
1535	>	}
1536	>
1537	>	/**
1538	>	* Tries to create (at most one) or activate (possibly several)
1539	>	* workers if too few are active. On contention failure, continues
1540	>	* until at least one worker is signalled or the given queue is
1541	>	* empty or all workers are active.
1542	>	*
1543	>	* @param q if non-null, the queue holding tasks to be signalled
1544	>	* @param signals the target number of signals.
1545	>	*/
1546	>	final void signalWork(WorkQueue q, int signals) {
1547	>	long c; int e, u, i; WorkQueue[] ws; WorkQueue w; Thread p;
1548	>	while ((u = (int)((c = ctl) >>> 32)) < 0) {
1549	>	if ((e = (int)c) > 0) {
1550	>	if ((ws = workQueues) != null && ws.length > (i = e & SMASK) &&
1551	>	(w = ws[i]) != null && w.eventCount == (e | INT_SIGN)) {
1552	>	long nc = (((long)(w.nextWait & E_MASK)) |
1553	>	((long)(u + UAC_UNIT) << 32));
1554	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1555	>	w.eventCount = (e + E_SEQ) & E_MASK;
1556	>	if ((p = w.parker) != null)
1557	>	U.unpark(p);
1558	>	if (--signals <= 0)
1559	>	break;
1560	>	}
1561	>	else
1562	>	signals = 1;
1563	>	if ((q != null && q.queueSize() == 0))
1564	>	break;
1565	>	}
1566	>	else
1567		break;
1568		}
1569	<	else if (tc > pc && rc < pc &&
1570	<	tc > (runState & ACTIVE_COUNT_MASK)) {
1571	<	ForkJoinWorkerThread spare = null;
1572	<	ForkJoinWorkerThread[] ws = workers;
1573	<	int nws = ws.length;
1574	<	for (int i = 0; i < nws; ++i) {
1575	<	ForkJoinWorkerThread w = ws[i];
1576	<	if (w != null && w.isSuspended()) {
1577	<	if ((workerCounts & RUNNING_COUNT_MASK) > pc ||
1578	<	runState == TERMINATING)
1579	<	return;
1580	<	if (w.tryResumeSpare())
1581	<	incrementRunningCount();
1582	<	break;
1569	>	else if (e == 0 && (u & SHORT_SIGN) != 0) {
1570	>	long nc = (long)(((u + UTC_UNIT) & UTC_MASK) |
1571	>	((u + UAC_UNIT) & UAC_MASK)) << 32;
1572	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1573	>	ForkJoinWorkerThread wt = null;
1574	>	Throwable ex = null;
1575	>	boolean started = false;
1576	>	try {
1577	>	ForkJoinWorkerThreadFactory fac;
1578	>	if ((fac = factory) != null &&
1579	>	(wt = fac.newThread(this)) != null) {
1580	>	wt.start();
1581	>	started = true;
1582	>	}
1583	>	} catch (Throwable rex) {
1584	>	ex = rex;
1585		}
1586	+	if (!started)
1587	+	deregisterWorker(wt, ex); // adjust counts on failure
1588	+	break;
1589		}
1590		}
1591		else
#	Line 724 | Line 1593 | public class ForkJoinPool extends Abstra
1593		}
1594		}
1595
1596	<	/**
728	<	* Final callback from terminating worker.  Removes record of
729	<	* worker from array, and adjusts counts. If pool is shutting
730	<	* down, tries to complete terminatation, else possibly replaces
731	<	* the worker.
732	<	*
733	<	* @param w the worker
734	<	*/
735	<	final void workerTerminated(ForkJoinWorkerThread w) {
736	<	if (w.active) { // force inactive
737	<	w.active = false;
738	<	do {} while (!tryDecrementActiveCount());
739	<	}
740	<	forgetWorker(w);
741	<
742	<	// Decrement total count, and if was running, running count
743	<	// Spin (waiting for other updates) if either would be negative
744	<	int nr = w.isTrimmed() ? 0 : ONE_RUNNING;
745	<	int unit = ONE_TOTAL + nr;
746	<	for (;;) {
747	<	int wc = workerCounts;
748	<	int rc = wc & RUNNING_COUNT_MASK;
749	<	if (rc - nr < 0 || (wc >>> TOTAL_COUNT_SHIFT) == 0)
750	<	Thread.yield(); // back off if waiting for other updates
751	<	else if (UNSAFE.compareAndSwapInt(this, workerCountsOffset,
752	<	wc, wc - unit))
753	<	break;
754	<	}
1596	>	// Scanning for tasks
1597
1598	<	accumulateStealCount(w); // collect final count
1599	<	if (!tryTerminate(false))
1600	<	ensureEnoughWorkers();
1598	>	/**
1599	>	* Top-level runloop for workers, called by ForkJoinWorkerThread.run.
1600	>	*/
1601	>	final void runWorker(WorkQueue w) {
1602	>	// initialize queue array in this thread
1603	>	w.array = new ForkJoinTask<?>[WorkQueue.INITIAL_QUEUE_CAPACITY];
1604	>	do { w.runTask(scan(w)); } while (w.qlock >= 0);
1605		}
1606
761	–	// Waiting for and signalling events
762	–
1607		/**
1608	<	* Releases workers blocked on a count not equal to current count.
1609	<	* @return true if any released
1610	<	*/
1611	<	private void releaseWaiters() {
1612	<	long top;
1613	<	while ((top = eventWaiters) != 0L) {
1614	<	ForkJoinWorkerThread[] ws = workers;
1615	<	int n = ws.length;
1616	<	for (;;) {
1617	<	int i = ((int)(top & WAITER_ID_MASK)) - 1;
1618	<	if (i < 0 || (int)(top >>> EVENT_COUNT_SHIFT) == eventCount)
1619	<	return;
1620	<	ForkJoinWorkerThread w;
1621	<	if (i < n && (w = ws[i]) != null &&
1622	<	UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
1623	<	top, w.nextWaiter)) {
1624	<	LockSupport.unpark(w);
1625	<	top = eventWaiters;
1608	>	* Scans for and, if found, returns one task, else possibly
1609	>	* inactivates the worker. This method operates on single reads of
1610	>	* volatile state and is designed to be re-invoked continuously,
1611	>	* in part because it returns upon detecting inconsistencies,
1612	>	* contention, or state changes that indicate possible success on
1613	>	* re-invocation.
1614	>	*
1615	>	* The scan searches for tasks across a random permutation of
1616	>	* queues (starting at a random index and stepping by a random
1617	>	* relative prime, checking each at least once).  The scan
1618	>	* terminates upon either finding a non-empty queue, or completing
1619	>	* the sweep. If the worker is not inactivated, it takes and
1620	>	* returns a task from this queue. Otherwise, if not activated, it
1621	>	* signals workers (that may include itself) and returns so caller
1622	>	* can retry. Also returns for trtry if the worker array may have
1623	>	* changed during an empty scan.  On failure to find a task, we
1624	>	* take one of the following actions, after which the caller will
1625	>	* retry calling this method unless terminated.
1626	>	*
1627	>	* * If pool is terminating, terminate the worker.
1628	>	*
1629	>	* * If not already enqueued, try to inactivate and enqueue the
1630	>	* worker on wait queue. Or, if inactivating has caused the pool
1631	>	* to be quiescent, relay to idleAwaitWork to check for
1632	>	* termination and possibly shrink pool.
1633	>	*
1634	>	* * If already enqueued and none of the above apply, possibly
1635	>	* (with 1/2 probability) park awaiting signal, else lingering to
1636	>	* help scan and signal.
1637	>	*
1638	>	* @param w the worker (via its WorkQueue)
1639	>	* @return a task or null if none found
1640	>	*/
1641	>	private final ForkJoinTask<?> scan(WorkQueue w) {
1642	>	WorkQueue[] ws; WorkQueue q;           // first update random seed
1643	>	int r = w.seed; r ^= r << 13; r ^= r >>> 17; w.seed = r ^= r << 5;
1644	>	int ps = plock, m;                     // volatile read order matters
1645	>	if ((ws = workQueues) != null && (m = ws.length - 1) > 0) {
1646	>	int ec = w.eventCount;             // ec is negative if inactive
1647	>	int step = (r >>> 16) | 1;         // relatively prime
1648	>	for (int j = (m + 1) << 2;  ; --j, r += step) {
1649	>	ForkJoinTask<?> t; ForkJoinTask<?>[] a; int b, n;
1650	>	if ((q = ws[r & m]) != null && (b = q.base) - q.top < 0 &&
1651	>	(a = q.array) != null) {   // probably nonempty
1652	>	int i = (((a.length - 1) & b) << ASHIFT) + ABASE;
1653	>	t = (ForkJoinTask<?>)U.getObjectVolatile(a, i);
1654	>	if (q.base == b && ec >= 0 && t != null &&
1655	>	U.compareAndSwapObject(a, i, t, null)) {
1656	>	if ((n = q.top - (q.base = b + 1)) > 0)
1657	>	signalWork(q, n);
1658	>	return t;              // taken
1659	>	}
1660	>	if (j < m || (ec < 0 && (ec = w.eventCount) < 0)) {
1661	>	if ((n = q.queueSize() - 1) > 0)
1662	>	signalWork(q, n);
1663	>	break;                 // let caller retry after signal
1664	>	}
1665	>	}
1666	>	else if (j < 0) {              // end of scan
1667	>	long c = ctl; int e;
1668	>	if (plock != ps)           // incomplete sweep
1669	>	break;
1670	>	if ((e = (int)c) < 0)      // pool is terminating
1671	>	w.qlock = -1;
1672	>	else if (ec >= 0) {        // try to enqueue/inactivate
1673	>	long nc = ((long)ec |
1674	>	((c - AC_UNIT) & (AC_MASK|TC_MASK)));
1675	>	w.nextWait = e;
1676	>	w.eventCount = ec | INT_SIGN; // mark as inactive
1677	>	if (ctl != c ||
1678	>	!U.compareAndSwapLong(this, CTL, c, nc))
1679	>	w.eventCount = ec; // unmark on CAS failure
1680	>	else if ((int)(c >> AC_SHIFT) == 1 - parallelism)
1681	>	idleAwaitWork(w, nc, c);  // quiescent
1682	>	}
1683	>	else if (w.seed >= 0 && w.eventCount < 0) {
1684	>	Thread wt = Thread.currentThread();
1685	>	Thread.interrupted();  // clear status
1686	>	U.putObject(wt, PARKBLOCKER, this);
1687	>	w.parker = wt;         // emulate LockSupport.park
1688	>	if (w.eventCount < 0)  // recheck
1689	>	U.park(false, 0L);
1690	>	w.parker = null;
1691	>	U.putObject(wt, PARKBLOCKER, null);
1692	>	}
1693	>	break;
1694		}
783	–	else
784	–	break;      // possibly stale; reread
1695		}
1696		}
1697	+	return null;
1698		}
1699
1700		/**
1701	<	* Ensures eventCount on exit is different (mod 2^32) than on
1702	<	* entry and wakes up all waiters
1703	<	*/
1704	<	private void signalEvent() {
1705	<	int c;
1706	<	do {} while (!UNSAFE.compareAndSwapInt(this, eventCountOffset,
1707	<	c = eventCount, c+1));
1708	<	releaseWaiters();
1701	>	* If inactivating worker w has caused the pool to become
1702	>	* quiescent, checks for pool termination, and, so long as this is
1703	>	* not the only worker, waits for event for up to a given
1704	>	* duration.  On timeout, if ctl has not changed, terminates the
1705	>	* worker, which will in turn wake up another worker to possibly
1706	>	* repeat this process.
1707	>	*
1708	>	* @param w the calling worker
1709	>	* @param currentCtl the ctl value triggering possible quiescence
1710	>	* @param prevCtl the ctl value to restore if thread is terminated
1711	>	*/
1712	>	private void idleAwaitWork(WorkQueue w, long currentCtl, long prevCtl) {
1713	>	if (w.eventCount < 0 &&
1714	>	(this == commonPool || !tryTerminate(false, false)) &&
1715	>	(int)prevCtl != 0) {
1716	>	int dc = -(short)(currentCtl >>> TC_SHIFT);
1717	>	long parkTime = dc < 0 ? FAST_IDLE_TIMEOUT: (dc + 1) * IDLE_TIMEOUT;
1718	>	long deadline = System.nanoTime() + parkTime - 100000L; // 1ms slop
1719	>	Thread wt = Thread.currentThread();
1720	>	while (ctl == currentCtl) {
1721	>	Thread.interrupted();  // timed variant of version in scan()
1722	>	U.putObject(wt, PARKBLOCKER, this);
1723	>	w.parker = wt;
1724	>	if (ctl == currentCtl)
1725	>	U.park(false, parkTime);
1726	>	w.parker = null;
1727	>	U.putObject(wt, PARKBLOCKER, null);
1728	>	if (ctl != currentCtl)
1729	>	break;
1730	>	if (deadline - System.nanoTime() <= 0L &&
1731	>	U.compareAndSwapLong(this, CTL, currentCtl, prevCtl)) {
1732	>	w.eventCount = (w.eventCount + E_SEQ) | E_MASK;
1733	>	w.qlock = -1;   // shrink
1734	>	break;
1735	>	}
1736	>	}
1737	>	}
1738		}
1739
1740		/**
1741	<	* Advances eventCount and releases waiters until interference by
1742	<	* other releasing threads is detected.
1741	>	* Scans through queues looking for work while joining a task;
1742	>	* if any are present, signals.
1743	>	*
1744	>	* @param task to return early if done
1745	>	* @param origin an index to start scan
1746		*/
1747	<	final void signalWork() {
1748	<	int c;
1749	<	UNSAFE.compareAndSwapInt(this, eventCountOffset, c=eventCount, c+1);
1750	<	long top;
1751	<	while ((top = eventWaiters) != 0L) {
1752	<	int ec = eventCount;
1753	<	ForkJoinWorkerThread[] ws = workers;
1754	<	int n = ws.length;
1755	<	for (;;) {
1756	<	int i = ((int)(top & WAITER_ID_MASK)) - 1;
1757	<	if (i < 0 || (int)(top >>> EVENT_COUNT_SHIFT) == ec)
1758	<	return;
816	<	ForkJoinWorkerThread w;
817	<	if (i < n && (w = ws[i]) != null &&
818	<	UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
819	<	top, top = w.nextWaiter)) {
820	<	LockSupport.unpark(w);
821	<	if (top != eventWaiters) // let someone else take over
822	<	return;
1747	>	final int helpSignal(ForkJoinTask<?> task, int origin) {
1748	>	WorkQueue[] ws; WorkQueue q; int m, n, s;
1749	>	if (task != null && (ws = workQueues) != null &&
1750	>	(m = ws.length - 1) >= 0) {
1751	>	for (int i = 0; i <= m; ++i) {
1752	>	if ((s = task.status) < 0)
1753	>	return s;
1754	>	if ((q = ws[(i + origin) & m]) != null &&
1755	>	(n = q.queueSize()) > 0) {
1756	>	signalWork(q, n);
1757	>	if ((int)(ctl >> AC_SHIFT) >= 0)
1758	>	break;
1759		}
824	–	else
825	–	break;      // possibly stale; reread
1760		}
1761		}
1762	+	return 0;
1763		}
1764
1765		/**
1766	<	* If worker is inactive, blocks until terminating or event count
1767	<	* advances from last value held by worker; in any case helps
1768	<	* release others.
1769	<	*
1770	<	* @param w the calling worker thread
1771	<	* @param retries the number of scans by caller failing to find work
1772	<	* @return false if now too many threads running
1773	<	*/
1774	<	private boolean eventSync(ForkJoinWorkerThread w, int retries) {
1775	<	int wec = w.lastEventCount;
1776	<	if (retries > 1) { // can only block after 2nd miss
1777	<	long nextTop = (((long)wec << EVENT_COUNT_SHIFT) |
1778	<	((long)(w.poolIndex + 1)));
1779	<	long top;
1780	<	while ((runState < SHUTDOWN || !tryTerminate(false)) &&
1781	<	(((int)(top = eventWaiters) & WAITER_ID_MASK) == 0 ||
1782	<	(int)(top >>> EVENT_COUNT_SHIFT) == wec) &&
1783	<	eventCount == wec) {
1784	<	if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
1785	<	w.nextWaiter = top, nextTop)) {
1786	<	accumulateStealCount(w); // transfer steals while idle
1787	<	Thread.interrupted();    // clear/ignore interrupt
1788	<	while (eventCount == wec)
1789	<	w.doPark();
1790	<	break;
1766	>	* Tries to locate and execute tasks for a stealer of the given
1767	>	* task, or in turn one of its stealers, Traces currentSteal ->
1768	>	* currentJoin links looking for a thread working on a descendant
1769	>	* of the given task and with a non-empty queue to steal back and
1770	>	* execute tasks from. The first call to this method upon a
1771	>	* waiting join will often entail scanning/search, (which is OK
1772	>	* because the joiner has nothing better to do), but this method
1773	>	* leaves hints in workers to speed up subsequent calls. The
1774	>	* implementation is very branchy to cope with potential
1775	>	* inconsistencies or loops encountering chains that are stale,
1776	>	* unknown, or so long that they are likely cyclic.
1777	>	*
1778	>	* @param joiner the joining worker
1779	>	* @param task the task to join
1780	>	* @return 0 if no progress can be made, negative if task
1781	>	* known complete, else positive
1782	>	*/
1783	>	private int tryHelpStealer(WorkQueue joiner, ForkJoinTask<?> task) {
1784	>	int stat = 0, steps = 0;                    // bound to avoid cycles
1785	>	if (joiner != null && task != null) {       // hoist null checks
1786	>	restart: for (;;) {
1787	>	ForkJoinTask<?> subtask = task;     // current target
1788	>	for (WorkQueue j = joiner, v;;) {   // v is stealer of subtask
1789	>	WorkQueue[] ws; int m, s, h;
1790	>	if ((s = task.status) < 0) {
1791	>	stat = s;
1792	>	break restart;
1793	>	}
1794	>	if ((ws = workQueues) == null || (m = ws.length - 1) <= 0)
1795	>	break restart;              // shutting down
1796	>	if ((v = ws[h = (j.stealHint | 1) & m]) == null ||
1797	>	v.currentSteal != subtask) {
1798	>	for (int origin = h;;) {    // find stealer
1799	>	if (((h = (h + 2) & m) & 15) == 1 &&
1800	>	(subtask.status < 0 || j.currentJoin != subtask))
1801	>	continue restart;   // occasional staleness check
1802	>	if ((v = ws[h]) != null &&
1803	>	v.currentSteal == subtask) {
1804	>	j.stealHint = h;    // save hint
1805	>	break;
1806	>	}
1807	>	if (h == origin)
1808	>	break restart;      // cannot find stealer
1809	>	}
1810	>	}
1811	>	for (;;) { // help stealer or descend to its stealer
1812	>	ForkJoinTask[] a;  int b;
1813	>	if (subtask.status < 0)     // surround probes with
1814	>	continue restart;       //   consistency checks
1815	>	if ((b = v.base) - v.top < 0 && (a = v.array) != null) {
1816	>	int i = (((a.length - 1) & b) << ASHIFT) + ABASE;
1817	>	ForkJoinTask<?> t =
1818	>	(ForkJoinTask<?>)U.getObjectVolatile(a, i);
1819	>	if (subtask.status < 0 || j.currentJoin != subtask ||
1820	>	v.currentSteal != subtask)
1821	>	continue restart;   // stale
1822	>	stat = 1;               // apparent progress
1823	>	if (t != null && v.base == b &&
1824	>	U.compareAndSwapObject(a, i, t, null)) {
1825	>	v.base = b + 1;     // help stealer
1826	>	joiner.runSubtask(t);
1827	>	}
1828	>	else if (v.base == b && ++steps == MAX_HELP)
1829	>	break restart;      // v apparently stalled
1830	>	}
1831	>	else {                      // empty -- try to descend
1832	>	ForkJoinTask<?> next = v.currentJoin;
1833	>	if (subtask.status < 0 || j.currentJoin != subtask ||
1834	>	v.currentSteal != subtask)
1835	>	continue restart;   // stale
1836	>	else if (next == null || ++steps == MAX_HELP)
1837	>	break restart;      // dead-end or maybe cyclic
1838	>	else {
1839	>	subtask = next;
1840	>	j = v;
1841	>	break;
1842	>	}
1843	>	}
1844	>	}
1845		}
1846		}
858	–	wec = eventCount;
859	–	}
860	–	releaseWaiters();
861	–	int wc = workerCounts;
862	–	if ((wc & RUNNING_COUNT_MASK) <= parallelism) {
863	–	w.lastEventCount = wec;
864	–	return true;
1847		}
1848	<	if (wec != w.lastEventCount) // back up if may re-wait
867	<	w.lastEventCount = wec - (wc >>> TOTAL_COUNT_SHIFT);
868	<	return false;
1848	>	return stat;
1849		}
1850
1851		/**
1852	<	* Callback from workers invoked upon each top-level action (i.e.,
1853	<	* stealing a task or taking a submission and running
1854	<	* it). Performs one or both of the following:
1855	<	*
1856	<	* * If the worker cannot find work, updates its active status to
1857	<	* inactive and updates activeCount unless there is contention, in
1858	<	* which case it may try again (either in this or a subsequent
1859	<	* call).  Additionally, awaits the next task event and/or helps
1860	<	* wake up other releasable waiters.
1861	<	*
1862	<	* * If there are too many running threads, suspends this worker
1863	<	* (first forcing inactivation if necessary).  If it is not
1864	<	* resumed before a keepAlive elapses, the worker may be "trimmed"
1865	<	* -- killed while suspended within suspendAsSpare. Otherwise,
1866	<	* upon resume it rechecks to make sure that it is still needed.
1867	<	*
1868	<	* @param w the worker
1869	<	* @param retries the number of scans by caller failing to find work
1870	<	* find any (in which case it may block waiting for work).
1871	<	*/
1872	<	final void preStep(ForkJoinWorkerThread w, int retries) {
893	<	boolean active = w.active;
894	<	boolean inactivate = active && retries != 0;
895	<	for (;;) {
896	<	int rs, wc;
897	<	if (inactivate &&
898	<	UNSAFE.compareAndSwapInt(this, runStateOffset,
899	<	rs = runState, rs - ONE_ACTIVE))
900	<	inactivate = active = w.active = false;
901	<	if (((wc = workerCounts) & RUNNING_COUNT_MASK) <= parallelism) {
902	<	if (active || eventSync(w, retries))
1852	>	* Analog of tryHelpStealer for CountedCompleters. Tries to steal
1853	>	* and run tasks within the target's computation
1854	>	*
1855	>	* @param task the task to join
1856	>	* @param mode if shared, exit upon completing any task
1857	>	* if all workers are active
1858	>	*
1859	>	*/
1860	>	private int helpComplete(ForkJoinTask<?> task, int mode) {
1861	>	WorkQueue[] ws; WorkQueue q; int m, n, s;
1862	>	if (task != null && (ws = workQueues) != null &&
1863	>	(m = ws.length - 1) >= 0) {
1864	>	for (int j = 1, origin = j;;) {
1865	>	if ((s = task.status) < 0)
1866	>	return s;
1867	>	if ((q = ws[j & m]) != null && q.pollAndExecCC(task)) {
1868	>	origin = j;
1869	>	if (mode == SHARED_QUEUE && (int)(ctl >> AC_SHIFT) >= 0)
1870	>	break;
1871	>	}
1872	>	else if ((j = (j + 2) & m) == origin)
1873		break;
1874		}
905	–	else if (!(inactivate |= active) &&  // must inactivate to suspend
906	–	UNSAFE.compareAndSwapInt(this, workerCountsOffset,
907	–	wc, wc - ONE_RUNNING) &&
908	–	!w.suspendAsSpare())             // false if trimmed
909	–	break;
1875		}
1876	+	return 0;
1877		}
1878
1879		/**
1880	<	* Awaits join of the given task if enough threads, or can resume
1881	<	* or create a spare. Fails (in which case the given task might
1882	<	* not be done) upon contention or lack of decision about
1883	<	* blocking. Returns void because caller must check
1884	<	* task status on return anyway.
1885	<	*
1886	<	* We allow blocking if:
1887	<	*
1888	<	* 1. There would still be at least as many running threads as
1889	<	*    parallelism level if this thread blocks.
1890	<	*
1891	<	* 2. A spare is resumed to replace this worker. We tolerate
1892	<	*    slop in the decision to replace if a spare is found without
1893	<	*    first decrementing run count.  This may release too many,
1894	<	*    but if so, the superfluous ones will re-suspend via
1895	<	*    preStep().
1896	<	*
1897	<	* 3. After #spares repeated checks, there are no fewer than #spare
1898	<	*    threads not running. We allow this slack to avoid hysteresis
933	<	*    and as a hedge against lag/uncertainty of running count
934	<	*    estimates when signalling or unblocking stalls.
935	<	*
936	<	* 4. All existing workers are busy (as rechecked via repeated
937	<	*    retries by caller) and a new spare is created.
938	<	*
939	<	* If none of the above hold, we try to escape out by
940	<	* re-incrementing count and returning to caller, which can retry
941	<	* later.
942	<	*
943	<	* @param joinMe the task to join
944	<	* @param retries if negative, then serve only as a precheck
945	<	*   that the thread can be replaced by a spare. Otherwise,
946	<	*   the number of repeated calls to this method returning busy
947	<	* @return true if the call must be retried because there
948	<	*   none of the blocking checks hold
949	<	*/
950	<	final boolean tryAwaitJoin(ForkJoinTask<?> joinMe, int retries) {
951	<	if (joinMe.status < 0) // precheck to prime loop
952	<	return false;
953	<	int pc = parallelism;
954	<	boolean running = true; // false when running count decremented
955	<	outer:for (;;) {
956	<	int wc = workerCounts;
957	<	int rc = wc & RUNNING_COUNT_MASK;
958	<	int tc = wc >>> TOTAL_COUNT_SHIFT;
959	<	if (running) { // replace with spare or decrement count
960	<	if (rc <= pc && tc > pc &&
961	<	(retries > 0 || tc > (runState & ACTIVE_COUNT_MASK))) {
962	<	ForkJoinWorkerThread[] ws = workers;
963	<	int nws = ws.length;
964	<	for (int i = 0; i < nws; ++i) { // search for spare
965	<	ForkJoinWorkerThread w = ws[i];
966	<	if (w != null) {
967	<	if (joinMe.status < 0)
968	<	return false;
969	<	if (w.isSuspended()) {
970	<	if ((workerCounts & RUNNING_COUNT_MASK)>=pc &&
971	<	w.tryResumeSpare()) {
972	<	running = false;
973	<	break outer;
974	<	}
975	<	continue outer; // rescan
976	<	}
977	<	}
978	<	}
1880	>	* Tries to decrement active count (sometimes implicitly) and
1881	>	* possibly release or create a compensating worker in preparation
1882	>	* for blocking. Fails on contention or termination. Otherwise,
1883	>	* adds a new thread if no idle workers are available and pool
1884	>	* may become starved.
1885	>	*/
1886	>	final boolean tryCompensate() {
1887	>	int pc = parallelism, e, u, i, tc; long c;
1888	>	WorkQueue[] ws; WorkQueue w; Thread p;
1889	>	if ((e = (int)(c = ctl)) >= 0 && (ws = workQueues) != null) {
1890	>	if (e != 0 && (i = e & SMASK) < ws.length &&
1891	>	(w = ws[i]) != null && w.eventCount == (e | INT_SIGN)) {
1892	>	long nc = ((long)(w.nextWait & E_MASK) |
1893	>	(c & (AC_MASK|TC_MASK)));
1894	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1895	>	w.eventCount = (e + E_SEQ) & E_MASK;
1896	>	if ((p = w.parker) != null)
1897	>	U.unpark(p);
1898	>	return true;   // replace with idle worker
1899		}
980	–	if (retries < 0 || // < 0 means replacement check only
981	–	rc == 0 || joinMe.status < 0 || workerCounts != wc ||
982	–	!UNSAFE.compareAndSwapInt(this, workerCountsOffset,
983	–	wc, wc - ONE_RUNNING))
984	–	return false; // done or inconsistent or contended
985	–	running = false;
986	–	if (rc > pc)
987	–	break;
1900		}
1901	<	else { // allow blocking if enough threads
1902	<	if (rc >= pc || joinMe.status < 0)
1903	<	break;
1904	<	int sc = tc - pc + 1; // = spare threads, plus the one to add
1905	<	if (retries > sc) {
1906	<	if (rc > 0 && rc >= pc - sc) // allow slack
1907	<	break;
1908	<	if (tc < MAX_THREADS &&
1909	<	tc == (runState & ACTIVE_COUNT_MASK) &&
1910	<	workerCounts == wc &&
1911	<	UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
1912	<	wc+(ONE_RUNNING|ONE_TOTAL))) {
1913	<	addWorker();
1914	<	break;
1901	>	else if ((short)((u = (int)(c >>> 32)) >>> UTC_SHIFT) >= 0 &&
1902	>	(u >> UAC_SHIFT) + pc > 1) {
1903	>	long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK);
1904	>	if (U.compareAndSwapLong(this, CTL, c, nc))
1905	>	return true;    // no compensation
1906	>	}
1907	>	else if ((tc = u + pc) < MAX_CAP) {
1908	>	long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK);
1909	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1910	>	Throwable ex = null;
1911	>	ForkJoinWorkerThread wt = null;
1912	>	try {
1913	>	ForkJoinWorkerThreadFactory fac;
1914	>	if ((fac = factory) != null &&
1915	>	(wt = fac.newThread(this)) != null) {
1916	>	wt.start();
1917	>	return true;
1918	>	}
1919	>	} catch (Throwable rex) {
1920	>	ex = rex;
1921		}
1922	<	}
1005	<	if (workerCounts == wc &&        // back out to allow rescan
1006	<	UNSAFE.compareAndSwapInt (this, workerCountsOffset,
1007	<	wc, wc + ONE_RUNNING)) {
1008	<	releaseWaiters();            // help others progress
1009	<	return true;                 // let caller retry
1922	>	deregisterWorker(wt, ex); // adjust counts etc
1923		}
1924		}
1925		}
1013	–	// arrive here if can block
1014	–	joinMe.internalAwaitDone();
1015	–	int c;                      // to inline incrementRunningCount
1016	–	do {} while (!UNSAFE.compareAndSwapInt
1017	–	(this, workerCountsOffset,
1018	–	c = workerCounts, c + ONE_RUNNING));
1926		return false;
1927		}
1928
1929		/**
1930	<	* Same idea as (and shares many code snippets with) tryAwaitJoin,
1931	<	* but self-contained because there are no caller retries.
1932	<	* TODO: Rework to use simpler API.
1930	>	* Helps and/or blocks until the given task is done.
1931	>	*
1932	>	* @param joiner the joining worker
1933	>	* @param task the task
1934	>	* @return task status on exit
1935		*/
1936	<	final void awaitBlocker(ManagedBlocker blocker)
1937	<	throws InterruptedException {
1938	<	boolean done;
1939	<	if (done = blocker.isReleasable())
1940	<	return;
1941	<	int pc = parallelism;
1942	<	int retries = 0;
1943	<	boolean running = true; // false when running count decremented
1944	<	outer:for (;;) {
1945	<	int wc = workerCounts;
1946	<	int rc = wc & RUNNING_COUNT_MASK;
1947	<	int tc = wc >>> TOTAL_COUNT_SHIFT;
1948	<	if (running) {
1949	<	if (rc <= pc && tc > pc &&
1950	<	(retries > 0 || tc > (runState & ACTIVE_COUNT_MASK))) {
1951	<	ForkJoinWorkerThread[] ws = workers;
1952	<	int nws = ws.length;
1953	<	for (int i = 0; i < nws; ++i) {
1954	<	ForkJoinWorkerThread w = ws[i];
1955	<	if (w != null) {
1956	<	if (done = blocker.isReleasable())
1957	<	return;
1049	<	if (w.isSuspended()) {
1050	<	if ((workerCounts & RUNNING_COUNT_MASK)>=pc &&
1051	<	w.tryResumeSpare()) {
1052	<	running = false;
1053	<	break outer;
1936	>	final int awaitJoin(WorkQueue joiner, ForkJoinTask<?> task) {
1937	>	int s = 0;
1938	>	if (joiner != null && task != null && (s = task.status) >= 0) {
1939	>	ForkJoinTask<?> prevJoin = joiner.currentJoin;
1940	>	joiner.currentJoin = task;
1941	>	do {} while ((s = task.status) >= 0 &&
1942	>	joiner.queueSize() > 0 &&
1943	>	joiner.tryRemoveAndExec(task)); // process local tasks
1944	>	if (s >= 0 && (s = task.status) >= 0 &&
1945	>	(s = helpSignal(task, joiner.poolIndex)) >= 0 &&
1946	>	(task instanceof CountedCompleter))
1947	>	s = helpComplete(task, LIFO_QUEUE);
1948	>	while (s >= 0 && (s = task.status) >= 0) {
1949	>	if ((joiner.queueSize() > 0 ||           // try helping
1950	>	(s = tryHelpStealer(joiner, task)) == 0) &&
1951	>	(s = task.status) >= 0 && tryCompensate()) {
1952	>	if (task.trySetSignal() && (s = task.status) >= 0) {
1953	>	synchronized (task) {
1954	>	if (task.status >= 0) {
1955	>	try {                // see ForkJoinTask
1956	>	task.wait();     //  for explanation
1957	>	} catch (InterruptedException ie) {
1958		}
1055	–	continue outer; // rescan
1959		}
1960	+	else
1961	+	task.notifyAll();
1962		}
1963		}
1964	+	long c;                          // re-activate
1965	+	do {} while (!U.compareAndSwapLong
1966	+	(this, CTL, c = ctl, c + AC_UNIT));
1967		}
1060	–	if (done = blocker.isReleasable())
1061	–	return;
1062	–	if (rc == 0 || workerCounts != wc ||
1063	–	!UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1064	–	wc, wc - ONE_RUNNING))
1065	–	continue;
1066	–	running = false;
1067	–	if (rc > pc)
1068	–	break;
1069	–	}
1070	–	else {
1071	–	if (rc >= pc || (done = blocker.isReleasable()))
1072	–	break;
1073	–	int sc = tc - pc + 1;
1074	–	if (retries++ > sc) {
1075	–	if (rc > 0 && rc >= pc - sc)
1076	–	break;
1077	–	if (tc < MAX_THREADS &&
1078	–	tc == (runState & ACTIVE_COUNT_MASK) &&
1079	–	workerCounts == wc &&
1080	–	UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
1081	–	wc+(ONE_RUNNING|ONE_TOTAL))) {
1082	–	addWorker();
1083	–	break;
1084	–	}
1085	–	}
1086	–	Thread.yield();
1968		}
1969	+	joiner.currentJoin = prevJoin;
1970		}
1971	<
1972	<	try {
1091	<	if (!done)
1092	<	do {} while (!blocker.isReleasable() && !blocker.block());
1093	<	} finally {
1094	<	if (!running) {
1095	<	int c;
1096	<	do {} while (!UNSAFE.compareAndSwapInt
1097	<	(this, workerCountsOffset,
1098	<	c = workerCounts, c + ONE_RUNNING));
1099	<	}
1100	<	}
1101	<	}
1971	>	return s;
1972	>	}
1973
1974		/**
1975	<	* Possibly initiates and/or completes termination.
1975	>	* Stripped-down variant of awaitJoin used by timed joins. Tries
1976	>	* to help join only while there is continuous progress. (Caller
1977	>	* will then enter a timed wait.)
1978		*
1979	<	* @param now if true, unconditionally terminate, else only
1980	<	* if shutdown and empty queue and no active workers
1108	<	* @return true if now terminating or terminated
1979	>	* @param joiner the joining worker
1980	>	* @param task the task
1981		*/
1982	<	private boolean tryTerminate(boolean now) {
1983	<	if (now)
1984	<	advanceRunLevel(SHUTDOWN); // ensure at least SHUTDOWN
1985	<	else if (runState < SHUTDOWN ||
1986	<	!submissionQueue.isEmpty() ||
1987	<	(runState & ACTIVE_COUNT_MASK) != 0)
1988	<	return false;
1989	<
1990	<	if (advanceRunLevel(TERMINATING))
1991	<	startTerminating();
1992	<
1993	<	// Finish now if all threads terminated; else in some subsequent call
1994	<	if ((workerCounts >>> TOTAL_COUNT_SHIFT) == 0) {
1995	<	advanceRunLevel(TERMINATED);
1996	<	termination.arrive();
1982	>	final void helpJoinOnce(WorkQueue joiner, ForkJoinTask<?> task) {
1983	>	int s;
1984	>	if (joiner != null && task != null && (s = task.status) >= 0) {
1985	>	ForkJoinTask<?> prevJoin = joiner.currentJoin;
1986	>	joiner.currentJoin = task;
1987	>	do {} while ((s = task.status) >= 0 &&
1988	>	joiner.queueSize() > 0 &&
1989	>	joiner.tryRemoveAndExec(task));
1990	>	if (s >= 0 && (s = task.status) >= 0 &&
1991	>	(s = helpSignal(task, joiner.poolIndex)) >= 0 &&
1992	>	(task instanceof CountedCompleter))
1993	>	s = helpComplete(task, LIFO_QUEUE);
1994	>	if (s >= 0 && joiner.queueSize() == 0) {
1995	>	do {} while (task.status >= 0 &&
1996	>	tryHelpStealer(joiner, task) > 0);
1997	>	}
1998	>	joiner.currentJoin = prevJoin;
1999		}
1126	–	return true;
2000		}
2001
2002		/**
2003	<	* Actions on transition to TERMINATING
2003	>	* Returns a (probably) non-empty steal queue, if one is found
2004	>	* during a random, then cyclic scan, else null.  This method must
2005	>	* be retried by caller if, by the time it tries to use the queue,
2006	>	* it is empty.
2007	>	* @param r a (random) seed for scanning
2008		*/
2009	<	private void startTerminating() {
2010	<	for (int i = 0; i < 2; ++i) { // twice to mop up newly created workers
2011	<	cancelSubmissions();
2012	<	shutdownWorkers();
2013	<	cancelWorkerTasks();
2014	<	signalEvent();
2015	<	interruptWorkers();
2009	>	private WorkQueue findNonEmptyStealQueue(int r) {
2010	>	int step = (r >>> 16) | 1;
2011	>	for (WorkQueue[] ws;;) {
2012	>	int ps = plock, m;
2013	>	if ((ws = workQueues) == null || (m = ws.length - 1) < 1)
2014	>	return null;
2015	>	for (int j = (m + 1) << 2; ; r += step) {
2016	>	WorkQueue q = ws[((r << 1) | 1) & m];
2017	>	if (q != null && q.queueSize() > 0)
2018	>	return q;
2019	>	else if (--j < 0) {
2020	>	if (plock == ps)
2021	>	return null;
2022	>	break;
2023	>	}
2024	>	}
2025		}
2026		}
2027
2028		/**
2029	<	* Clear out and cancel submissions, ignoring exceptions
2030	<	*/
2031	<	private void cancelSubmissions() {
2032	<	ForkJoinTask<?> task;
2033	<	while ((task = submissionQueue.poll()) != null) {
2034	<	try {
2035	<	task.cancel(false);
2036	<	} catch (Throwable ignore) {
2029	>	* Runs tasks until {@code isQuiescent()}. We piggyback on
2030	>	* active count ctl maintenance, but rather than blocking
2031	>	* when tasks cannot be found, we rescan until all others cannot
2032	>	* find tasks either.
2033	>	*/
2034	>	final void helpQuiescePool(WorkQueue w) {
2035	>	for (boolean active = true;;) {
2036	>	ForkJoinTask<?> localTask; // exhaust local queue
2037	>	while ((localTask = w.nextLocalTask()) != null)
2038	>	localTask.doExec();
2039	>	// Similar to loop in scan(), but ignoring submissions
2040	>	WorkQueue q = findNonEmptyStealQueue(w.nextSeed());
2041	>	if (q != null) {
2042	>	ForkJoinTask<?> t; int b;
2043	>	if (!active) {      // re-establish active count
2044	>	long c;
2045	>	active = true;
2046	>	do {} while (!U.compareAndSwapLong
2047	>	(this, CTL, c = ctl, c + AC_UNIT));
2048	>	}
2049	>	if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null)
2050	>	w.runSubtask(t);
2051	>	}
2052	>	else {
2053	>	long c;
2054	>	if (active) {       // decrement active count without queuing
2055	>	active = false;
2056	>	do {} while (!U.compareAndSwapLong
2057	>	(this, CTL, c = ctl, c -= AC_UNIT));
2058	>	}
2059	>	else
2060	>	c = ctl;        // re-increment on exit
2061	>	if ((int)(c >> AC_SHIFT) + parallelism == 0) {
2062	>	do {} while (!U.compareAndSwapLong
2063	>	(this, CTL, c = ctl, c + AC_UNIT));
2064	>	break;
2065	>	}
2066		}
2067		}
2068		}
2069
2070		/**
2071	<	* Sets all worker run states to at least shutdown,
2072	<	* also resuming suspended workers
2071	>	* Gets and removes a local or stolen task for the given worker.
2072	>	*
2073	>	* @return a task, if available
2074		*/
2075	<	private void shutdownWorkers() {
2076	<	ForkJoinWorkerThread[] ws = workers;
2077	<	int nws = ws.length;
2078	<	for (int i = 0; i < nws; ++i) {
2079	<	ForkJoinWorkerThread w = ws[i];
2080	<	if (w != null)
2081	<	w.shutdown();
2075	>	final ForkJoinTask<?> nextTaskFor(WorkQueue w) {
2076	>	for (ForkJoinTask<?> t;;) {
2077	>	WorkQueue q; int b;
2078	>	if ((t = w.nextLocalTask()) != null)
2079	>	return t;
2080	>	if ((q = findNonEmptyStealQueue(w.nextSeed())) == null)
2081	>	return null;
2082	>	if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null)
2083	>	return t;
2084		}
2085		}
2086
2087		/**
2088	<	* Clears out and cancels all locally queued tasks
2088	>	* Returns a cheap heuristic guide for task partitioning when
2089	>	* programmers, frameworks, tools, or languages have little or no
2090	>	* idea about task granularity.  In essence by offering this
2091	>	* method, we ask users only about tradeoffs in overhead vs
2092	>	* expected throughput and its variance, rather than how finely to
2093	>	* partition tasks.
2094	>	*
2095	>	* In a steady state strict (tree-structured) computation, each
2096	>	* thread makes available for stealing enough tasks for other
2097	>	* threads to remain active. Inductively, if all threads play by
2098	>	* the same rules, each thread should make available only a
2099	>	* constant number of tasks.
2100	>	*
2101	>	* The minimum useful constant is just 1. But using a value of 1
2102	>	* would require immediate replenishment upon each steal to
2103	>	* maintain enough tasks, which is infeasible.  Further,
2104	>	* partitionings/granularities of offered tasks should minimize
2105	>	* steal rates, which in general means that threads nearer the top
2106	>	* of computation tree should generate more than those nearer the
2107	>	* bottom. In perfect steady state, each thread is at
2108	>	* approximately the same level of computation tree. However,
2109	>	* producing extra tasks amortizes the uncertainty of progress and
2110	>	* diffusion assumptions.
2111	>	*
2112	>	* So, users will want to use values larger, but not much larger
2113	>	* than 1 to both smooth over transient shortages and hedge
2114	>	* against uneven progress; as traded off against the cost of
2115	>	* extra task overhead. We leave the user to pick a threshold
2116	>	* value to compare with the results of this call to guide
2117	>	* decisions, but recommend values such as 3.
2118	>	*
2119	>	* When all threads are active, it is on average OK to estimate
2120	>	* surplus strictly locally. In steady-state, if one thread is
2121	>	* maintaining say 2 surplus tasks, then so are others. So we can
2122	>	* just use estimated queue length.  However, this strategy alone
2123	>	* leads to serious mis-estimates in some non-steady-state
2124	>	* conditions (ramp-up, ramp-down, other stalls). We can detect
2125	>	* many of these by further considering the number of "idle"
2126	>	* threads, that are known to have zero queued tasks, so
2127	>	* compensate by a factor of (#idle/#active) threads.
2128	>	*
2129	>	* Note: The approximation of #busy workers as #active workers is
2130	>	* not very good under current signalling scheme, and should be
2131	>	* improved.
2132	>	*/
2133	>	static int getSurplusQueuedTaskCount() {
2134	>	Thread t; ForkJoinWorkerThread wt; ForkJoinPool pool; WorkQueue q;
2135	>	if (((t = Thread.currentThread()) instanceof ForkJoinWorkerThread)) {
2136	>	int b = (q = (wt = (ForkJoinWorkerThread)t).workQueue).base;
2137	>	int p = (pool = wt.pool).parallelism;
2138	>	int a = (int)(pool.ctl >> AC_SHIFT) + p;
2139	>	return q.top - b - (a > (p >>>= 1) ? 0 :
2140	>	a > (p >>>= 1) ? 1 :
2141	>	a > (p >>>= 1) ? 2 :
2142	>	a > (p >>>= 1) ? 4 :
2143	>	8);
2144	>	}
2145	>	return 0;
2146	>	}
2147	>
2148	>	//  Termination
2149	>
2150	>	/**
2151	>	* Possibly initiates and/or completes termination.  The caller
2152	>	* triggering termination runs three passes through workQueues:
2153	>	* (0) Setting termination status, followed by wakeups of queued
2154	>	* workers; (1) cancelling all tasks; (2) interrupting lagging
2155	>	* threads (likely in external tasks, but possibly also blocked in
2156	>	* joins).  Each pass repeats previous steps because of potential
2157	>	* lagging thread creation.
2158	>	*
2159	>	* @param now if true, unconditionally terminate, else only
2160	>	* if no work and no active workers
2161	>	* @param enable if true, enable shutdown when next possible
2162	>	* @return true if now terminating or terminated
2163		*/
2164	<	private void cancelWorkerTasks() {
2165	<	ForkJoinWorkerThread[] ws = workers;
2166	<	int nws = ws.length;
2167	<	for (int i = 0; i < nws; ++i) {
2168	<	ForkJoinWorkerThread w = ws[i];
2169	<	if (w != null)
2170	<	w.cancelTasks();
2164	>	private boolean tryTerminate(boolean now, boolean enable) {
2165	>	if (this == commonPool)                     // cannot shut down
2166	>	return false;
2167	>	for (long c;;) {
2168	>	if (((c = ctl) & STOP_BIT) != 0) {      // already terminating
2169	>	if ((short)(c >>> TC_SHIFT) == -parallelism) {
2170	>	synchronized (this) {
2171	>	notifyAll();                // signal when 0 workers
2172	>	}
2173	>	}
2174	>	return true;
2175	>	}
2176	>	if (plock >= 0) {                       // not yet enabled
2177	>	int ps;
2178	>	if (!enable)
2179	>	return false;
2180	>	if (((ps = plock) & PL_LOCK) != 0 ||
2181	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
2182	>	ps = acquirePlock();
2183	>	int nps = SHUTDOWN;
2184	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
2185	>	releasePlock(nps);
2186	>	}
2187	>	if (!now) {                             // check if idle & no tasks
2188	>	if ((int)(c >> AC_SHIFT) != -parallelism ||
2189	>	hasQueuedSubmissions())
2190	>	return false;
2191	>	// Check for unqueued inactive workers. One pass suffices.
2192	>	WorkQueue[] ws = workQueues; WorkQueue w;
2193	>	if (ws != null) {
2194	>	for (int i = 1; i < ws.length; i += 2) {
2195	>	if ((w = ws[i]) != null && w.eventCount >= 0)
2196	>	return false;
2197	>	}
2198	>	}
2199	>	}
2200	>	if (U.compareAndSwapLong(this, CTL, c, c | STOP_BIT)) {
2201	>	for (int pass = 0; pass < 3; ++pass) {
2202	>	WorkQueue[] ws = workQueues;
2203	>	if (ws != null) {
2204	>	WorkQueue w;
2205	>	int n = ws.length;
2206	>	for (int i = 0; i < n; ++i) {
2207	>	if ((w = ws[i]) != null) {
2208	>	w.qlock = -1;
2209	>	if (pass > 0) {
2210	>	w.cancelAll();
2211	>	if (pass > 1)
2212	>	w.interruptOwner();
2213	>	}
2214	>	}
2215	>	}
2216	>	// Wake up workers parked on event queue
2217	>	int i, e; long cc; Thread p;
2218	>	while ((e = (int)(cc = ctl) & E_MASK) != 0 &&
2219	>	(i = e & SMASK) < n &&
2220	>	(w = ws[i]) != null) {
2221	>	long nc = ((long)(w.nextWait & E_MASK) |
2222	>	((cc + AC_UNIT) & AC_MASK) |
2223	>	(cc & (TC_MASK|STOP_BIT)));
2224	>	if (w.eventCount == (e | INT_SIGN) &&
2225	>	U.compareAndSwapLong(this, CTL, cc, nc)) {
2226	>	w.eventCount = (e + E_SEQ) & E_MASK;
2227	>	w.qlock = -1;
2228	>	if ((p = w.parker) != null)
2229	>	U.unpark(p);
2230	>	}
2231	>	}
2232	>	}
2233	>	}
2234	>	}
2235		}
2236		}
2237
2238	+	// external operations on common pool
2239	+
2240		/**
2241	<	* Unsticks all workers blocked on joins etc
2241	>	* Returns common pool queue for a thread that has submitted at
2242	>	* least one task.
2243		*/
2244	<	private void interruptWorkers() {
2245	<	ForkJoinWorkerThread[] ws = workers;
2246	<	int nws = ws.length;
2247	<	for (int i = 0; i < nws; ++i) {
2248	<	ForkJoinWorkerThread w = ws[i];
2249	<	if (w != null && !w.isTerminated()) {
2250	<	try {
2251	<	w.interrupt();
2252	<	} catch (SecurityException ignore) {
2253	<	}
2244	>	static WorkQueue commonSubmitterQueue() {
2245	>	ForkJoinPool p; WorkQueue[] ws; int m; Submitter z;
2246	>	return ((z = submitters.get()) != null &&
2247	>	(p = commonPool) != null &&
2248	>	(ws = p.workQueues) != null &&
2249	>	(m = ws.length - 1) >= 0) ?
2250	>	ws[m & z.seed & SQMASK] : null;
2251	>	}
2252	>
2253	>	/**
2254	>	* Tries to pop the given task from submitter's queue in common pool.
2255	>	*/
2256	>	static boolean tryExternalUnpush(ForkJoinTask<?> t) {
2257	>	ForkJoinPool p; WorkQueue[] ws; WorkQueue q; Submitter z;
2258	>	ForkJoinTask<?>[] a;  int m, s; long j;
2259	>	if ((z = submitters.get()) != null &&
2260	>	(p = commonPool) != null &&
2261	>	(ws = p.workQueues) != null &&
2262	>	(m = ws.length - 1) >= 0 &&
2263	>	(q = ws[m & z.seed & SQMASK]) != null &&
2264	>	(s = q.top) != q.base &&
2265	>	(a = q.array) != null &&
2266	>	U.getObjectVolatile
2267	>	(a, j = (((a.length - 1) & (s - 1)) << ASHIFT) + ABASE) == t &&
2268	>	U.compareAndSwapInt(q, QLOCK, 0, 1)) {
2269	>	if (q.array == a && q.top == s && // recheck
2270	>	U.compareAndSwapObject(a, j, t, null)) {
2271	>	q.top = s - 1;
2272	>	q.qlock = 0;
2273	>	return true;
2274		}
2275	+	q.qlock = 0;
2276		}
2277	+	return false;
2278		}
2279
1199	–	// misc support for ForkJoinWorkerThread
1200	–
2280		/**
2281	<	* Returns pool number
2282	<	*/
2283	<	final int getPoolNumber() {
2284	<	return poolNumber;
2281	>	* Tries to pop and run local tasks within the same computation
2282	>	* as the given root. On failure, tries to help complete from
2283	>	* other queues via helpComplete.
2284	>	*/
2285	>	private void externalHelpComplete(WorkQueue q, ForkJoinTask<?> root) {
2286	>	ForkJoinTask<?>[] a; int m;
2287	>	if (q != null && (a = q.array) != null && (m = (a.length - 1)) >= 0 &&
2288	>	root != null && root.status >= 0) {
2289	>	for (;;) {
2290	>	int s; Object o; CountedCompleter<?> task = null;
2291	>	if ((s = q.top) - q.base > 0) {
2292	>	long j = ((m & (s - 1)) << ASHIFT) + ABASE;
2293	>	if ((o = U.getObject(a, j)) != null &&
2294	>	(o instanceof CountedCompleter)) {
2295	>	CountedCompleter<?> t = (CountedCompleter<?>)o, r = t;
2296	>	do {
2297	>	if (r == root) {
2298	>	if (U.compareAndSwapInt(q, QLOCK, 0, 1)) {
2299	>	if (q.array == a && q.top == s &&
2300	>	U.compareAndSwapObject(a, j, t, null)) {
2301	>	q.top = s - 1;
2302	>	task = t;
2303	>	}
2304	>	q.qlock = 0;
2305	>	}
2306	>	break;
2307	>	}
2308	>	} while ((r = r.completer) != null);
2309	>	}
2310	>	}
2311	>	if (task != null)
2312	>	task.doExec();
2313	>	if (root.status < 0 || (int)(ctl >> AC_SHIFT) >= 0)
2314	>	break;
2315	>	if (task == null) {
2316	>	if (helpSignal(root, q.poolIndex) >= 0)
2317	>	helpComplete(root, SHARED_QUEUE);
2318	>	break;
2319	>	}
2320	>	}
2321	>	}
2322		}
2323
2324		/**
2325	<	* Accumulates steal count from a worker, clearing
2326	<	* the worker's value
2325	>	* Tries to help execute or signal availability of the given task
2326	>	* from submitter's queue in common pool.
2327		*/
2328	<	final void accumulateStealCount(ForkJoinWorkerThread w) {
2329	<	int sc = w.stealCount;
2330	<	if (sc != 0) {
2331	<	long c;
2332	<	w.stealCount = 0;
2333	<	do {} while (!UNSAFE.compareAndSwapLong(this, stealCountOffset,
2334	<	c = stealCount, c + sc));
2328	>	static void externalHelpJoin(ForkJoinTask<?> t) {
2329	>	// Some hard-to-avoid overlap with tryExternalUnpush
2330	>	ForkJoinPool p; WorkQueue[] ws; WorkQueue q, w; Submitter z;
2331	>	ForkJoinTask<?>[] a;  int m, s, n; long j;
2332	>	if (t != null && t.status >= 0 &&
2333	>	(z = submitters.get()) != null &&
2334	>	(p = commonPool) != null &&
2335	>	(ws = p.workQueues) != null &&
2336	>	(m = ws.length - 1) >= 0 &&
2337	>	(q = ws[m & z.seed & SQMASK]) != null &&
2338	>	(a = q.array) != null) {
2339	>	if ((s = q.top) != q.base &&
2340	>	U.getObjectVolatile
2341	>	(a, j = (((a.length - 1) & (s - 1)) << ASHIFT) + ABASE) == t &&
2342	>	U.compareAndSwapInt(q, QLOCK, 0, 1)) {
2343	>	if (q.array == a && q.top == s &&
2344	>	U.compareAndSwapObject(a, j, t, null)) {
2345	>	q.top = s - 1;
2346	>	q.qlock = 0;
2347	>	t.doExec();
2348	>	}
2349	>	else
2350	>	q.qlock = 0;
2351	>	}
2352	>	if (t.status >= 0) {
2353	>	if (t instanceof CountedCompleter)
2354	>	p.externalHelpComplete(q, t);
2355	>	else
2356	>	p.helpSignal(t, q.poolIndex);
2357	>	}
2358		}
2359		}
2360
2361		/**
2362	<	* Returns the approximate (non-atomic) number of idle threads per
1224	<	* active thread.
2362	>	* Restricted version of helpQuiescePool for external callers
2363		*/
2364	<	final int idlePerActive() {
2365	<	int pc = parallelism; // use parallelism, not rc
2366	<	int ac = runState;    // no mask -- artifically boosts during shutdown
2367	<	// Use exact results for small values, saturate past 4
2368	<	return pc <= ac? 0 : pc >>> 1 <= ac? 1 : pc >>> 2 <= ac? 3 : pc >>> 3;
2364	>	static void externalHelpQuiescePool() {
2365	>	ForkJoinPool p; ForkJoinTask<?> t; WorkQueue q; int b;
2366	>	int r = ThreadLocalRandom.current().nextInt();
2367	>	if ((p = commonPool) != null &&
2368	>	(q = p.findNonEmptyStealQueue(r)) != null &&
2369	>	(b = q.base) - q.top < 0 &&
2370	>	(t = q.pollAt(b)) != null)
2371	>	t.doExec();
2372		}
2373
2374	<	// Public and protected methods
2374	>	// Exported methods
2375
2376		// Constructors
2377
#	Line 1275 | Line 2416 | public class ForkJoinPool extends Abstra
2416		* use {@link java.lang.Runtime#availableProcessors}.
2417		* @param factory the factory for creating new threads. For default value,
2418		* use {@link #defaultForkJoinWorkerThreadFactory}.
2419	<	* @param handler the handler for internal worker threads that
2420	<	* terminate due to unrecoverable errors encountered while executing
2421	<	* tasks. For default value, use <code>null</code>.
2422	<	* @param asyncMode if true,
2419	>	* @param handler the handler for internal worker threads that
2420	>	* terminate due to unrecoverable errors encountered while executing
2421	>	* tasks. For default value, use {@code null}.
2422	>	* @param asyncMode if true,
2423		* establishes local first-in-first-out scheduling mode for forked
2424		* tasks that are never joined. This mode may be more appropriate
2425		* than default locally stack-based mode in applications in which
2426		* worker threads only process event-style asynchronous tasks.
2427	<	* For default value, use <code>false</code>.
2427	>	* For default value, use {@code false}.
2428		* @throws IllegalArgumentException if parallelism less than or
2429		*         equal to zero, or greater than implementation limit
2430		* @throws NullPointerException if the factory is null
#	Line 1292 | Line 2433 | public class ForkJoinPool extends Abstra
2433		*         because it does not hold {@link
2434		*         java.lang.RuntimePermission}{@code ("modifyThread")}
2435		*/
2436	<	public ForkJoinPool(int parallelism,
2436	>	public ForkJoinPool(int parallelism,
2437		ForkJoinWorkerThreadFactory factory,
2438		Thread.UncaughtExceptionHandler handler,
2439		boolean asyncMode) {
2440		checkPermission();
2441		if (factory == null)
2442		throw new NullPointerException();
2443	<	if (parallelism <= 0 || parallelism > MAX_THREADS)
2443	>	if (parallelism <= 0 || parallelism > MAX_CAP)
2444		throw new IllegalArgumentException();
2445		this.parallelism = parallelism;
2446		this.factory = factory;
2447		this.ueh = handler;
2448	<	this.locallyFifo = asyncMode;
2449	<	int arraySize = initialArraySizeFor(parallelism);
2450	<	this.workers = new ForkJoinWorkerThread[arraySize];
2451	<	this.submissionQueue = new LinkedTransferQueue<ForkJoinTask<?>>();
2452	<	this.workerLock = new ReentrantLock();
2453	<	this.termination = new Phaser(1);
2454	<	this.poolNumber = poolNumberGenerator.incrementAndGet();
2448	>	this.localMode = asyncMode ? FIFO_QUEUE : LIFO_QUEUE;
2449	>	long np = (long)(-parallelism); // offset ctl counts
2450	>	this.ctl = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK);
2451	>	int pn = nextPoolId();
2452	>	StringBuilder sb = new StringBuilder("ForkJoinPool-");
2453	>	sb.append(Integer.toString(pn));
2454	>	sb.append("-worker-");
2455	>	this.workerNamePrefix = sb.toString();
2456		}
2457
2458		/**
2459	<	* Returns initial power of two size for workers array.
2460	<	* @param pc the initial parallelism level
2461	<	*/
2462	<	private static int initialArraySizeFor(int pc) {
2463	<	// See Hackers Delight, sec 3.2. We know MAX_THREADS < (1 >>> 16)
2464	<	int size = pc < MAX_THREADS ? pc + 1 : MAX_THREADS;
2465	<	size |= size >>> 1;
2466	<	size |= size >>> 2;
2467	<	size |= size >>> 4;
2468	<	size |= size >>> 8;
2469	<	return size + 1;
2459	>	* Constructor for common pool, suitable only for static initialization.
2460	>	* Basically the same as above, but uses smallest possible initial footprint.
2461	>	*/
2462	>	ForkJoinPool(int parallelism, long ctl,
2463	>	ForkJoinWorkerThreadFactory factory,
2464	>	Thread.UncaughtExceptionHandler handler) {
2465	>	this.parallelism = parallelism;
2466	>	this.ctl = ctl;
2467	>	this.factory = factory;
2468	>	this.ueh = handler;
2469	>	this.localMode = LIFO_QUEUE;
2470	>	this.workerNamePrefix = "ForkJoinPool.commonPool-worker-";
2471		}
2472
1330	–	// Execution methods
1331	–
2473		/**
2474	<	* Common code for execute, invoke and submit
2474	>	* Returns the common pool instance.
2475	>	*
2476	>	* @return the common pool instance
2477		*/
2478	<	private <T> void doSubmit(ForkJoinTask<T> task) {
2479	<	if (task == null)
1337	<	throw new NullPointerException();
1338	<	if (runState >= SHUTDOWN)
1339	<	throw new RejectedExecutionException();
1340	<	submissionQueue.offer(task);
1341	<	signalEvent();
1342	<	ensureEnoughWorkers();
2478	>	public static ForkJoinPool commonPool() {
2479	>	return commonPool; // cannot be null (if so, a static init error)
2480		}
2481
2482	+	// Execution methods
2483	+
2484		/**
2485		* Performs the given task, returning its result upon completion.
2486	<	* If the caller is already engaged in a fork/join computation in
2487	<	* the current pool, this method is equivalent in effect to
2488	<	* {@link ForkJoinTask#invoke}.
2486	>	* If the computation encounters an unchecked Exception or Error,
2487	>	* it is rethrown as the outcome of this invocation.  Rethrown
2488	>	* exceptions behave in the same way as regular exceptions, but,
2489	>	* when possible, contain stack traces (as displayed for example
2490	>	* using {@code ex.printStackTrace()}) of both the current thread
2491	>	* as well as the thread actually encountering the exception;
2492	>	* minimally only the latter.
2493		*
2494		* @param task the task
2495		* @return the task's result
#	Line 1355 | Line 2498 | public class ForkJoinPool extends Abstra
2498		*         scheduled for execution
2499		*/
2500		public <T> T invoke(ForkJoinTask<T> task) {
2501	<	doSubmit(task);
2501	>	if (task == null)
2502	>	throw new NullPointerException();
2503	>	externalPush(task);
2504		return task.join();
2505		}
2506
2507		/**
2508		* Arranges for (asynchronous) execution of the given task.
1364	–	* If the caller is already engaged in a fork/join computation in
1365	–	* the current pool, this method is equivalent in effect to
1366	–	* {@link ForkJoinTask#fork}.
2509		*
2510		* @param task the task
2511		* @throws NullPointerException if the task is null
#	Line 1371 | Line 2513 | public class ForkJoinPool extends Abstra
2513		*         scheduled for execution
2514		*/
2515		public void execute(ForkJoinTask<?> task) {
2516	<	doSubmit(task);
2516	>	if (task == null)
2517	>	throw new NullPointerException();
2518	>	externalPush(task);
2519		}
2520
2521		// AbstractExecutorService methods
#	Line 1382 | Line 2526 | public class ForkJoinPool extends Abstra
2526		*         scheduled for execution
2527		*/
2528		public void execute(Runnable task) {
2529	+	if (task == null)
2530	+	throw new NullPointerException();
2531		ForkJoinTask<?> job;
2532		if (task instanceof ForkJoinTask<?>) // avoid re-wrap
2533		job = (ForkJoinTask<?>) task;
2534		else
2535	<	job = ForkJoinTask.adapt(task, null);
2536	<	doSubmit(job);
2535	>	job = new ForkJoinTask.AdaptedRunnableAction(task);
2536	>	externalPush(job);
2537		}
2538
2539		/**
2540		* Submits a ForkJoinTask for execution.
1395	–	* If the caller is already engaged in a fork/join computation in
1396	–	* the current pool, this method is equivalent in effect to
1397	–	* {@link ForkJoinTask#fork}.
2541		*
2542		* @param task the task to submit
2543		* @return the task
#	Line 1403 | Line 2546 | public class ForkJoinPool extends Abstra
2546		*         scheduled for execution
2547		*/
2548		public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) {
2549	<	doSubmit(task);
2549	>	if (task == null)
2550	>	throw new NullPointerException();
2551	>	externalPush(task);
2552		return task;
2553		}
2554
#	Line 1413 | Line 2558 | public class ForkJoinPool extends Abstra
2558		*         scheduled for execution
2559		*/
2560		public <T> ForkJoinTask<T> submit(Callable<T> task) {
2561	<	ForkJoinTask<T> job = ForkJoinTask.adapt(task);
2562	<	doSubmit(job);
2561	>	ForkJoinTask<T> job = new ForkJoinTask.AdaptedCallable<T>(task);
2562	>	externalPush(job);
2563		return job;
2564		}
2565
#	Line 1424 | Line 2569 | public class ForkJoinPool extends Abstra
2569		*         scheduled for execution
2570		*/
2571		public <T> ForkJoinTask<T> submit(Runnable task, T result) {
2572	<	ForkJoinTask<T> job = ForkJoinTask.adapt(task, result);
2573	<	doSubmit(job);
2572	>	ForkJoinTask<T> job = new ForkJoinTask.AdaptedRunnable<T>(task, result);
2573	>	externalPush(job);
2574		return job;
2575		}
2576
#	Line 1435 | Line 2580 | public class ForkJoinPool extends Abstra
2580		*         scheduled for execution
2581		*/
2582		public ForkJoinTask<?> submit(Runnable task) {
2583	+	if (task == null)
2584	+	throw new NullPointerException();
2585		ForkJoinTask<?> job;
2586		if (task instanceof ForkJoinTask<?>) // avoid re-wrap
2587		job = (ForkJoinTask<?>) task;
2588		else
2589	<	job = ForkJoinTask.adapt(task, null);
2590	<	doSubmit(job);
2589	>	job = new ForkJoinTask.AdaptedRunnableAction(task);
2590	>	externalPush(job);
2591		return job;
2592		}
2593
#	Line 1449 | Line 2596 | public class ForkJoinPool extends Abstra
2596		* @throws RejectedExecutionException {@inheritDoc}
2597		*/
2598		public <T> List<Future<T>> invokeAll(Collection<? extends Callable<T>> tasks) {
2599	<	ArrayList<ForkJoinTask<T>> forkJoinTasks =
2600	<	new ArrayList<ForkJoinTask<T>>(tasks.size());
2601	<	for (Callable<T> task : tasks)
2602	<	forkJoinTasks.add(ForkJoinTask.adapt(task));
2603	<	invoke(new InvokeAll<T>(forkJoinTasks));
2604	<
2599	>	// In previous versions of this class, this method constructed
2600	>	// a task to run ForkJoinTask.invokeAll, but now external
2601	>	// invocation of multiple tasks is at least as efficient.
2602	>	List<ForkJoinTask<T>> fs = new ArrayList<ForkJoinTask<T>>(tasks.size());
2603	>	// Workaround needed because method wasn't declared with
2604	>	// wildcards in return type but should have been.
2605		@SuppressWarnings({"unchecked", "rawtypes"})
2606	<	List<Future<T>> futures = (List<Future<T>>) (List) forkJoinTasks;
1460	<	return futures;
1461	<	}
2606	>	List<Future<T>> futures = (List<Future<T>>) (List) fs;
2607
2608	<	static final class InvokeAll<T> extends RecursiveAction {
2609	<	final ArrayList<ForkJoinTask<T>> tasks;
2610	<	InvokeAll(ArrayList<ForkJoinTask<T>> tasks) { this.tasks = tasks; }
2611	<	public void compute() {
2612	<	try { invokeAll(tasks); }
2613	<	catch (Exception ignore) {}
2608	>	boolean done = false;
2609	>	try {
2610	>	for (Callable<T> t : tasks) {
2611	>	ForkJoinTask<T> f = new ForkJoinTask.AdaptedCallable<T>(t);
2612	>	externalPush(f);
2613	>	fs.add(f);
2614	>	}
2615	>	for (ForkJoinTask<T> f : fs)
2616	>	f.quietlyJoin();
2617	>	done = true;
2618	>	return futures;
2619	>	} finally {
2620	>	if (!done)
2621	>	for (ForkJoinTask<T> f : fs)
2622	>	f.cancel(false);
2623		}
1470	–	private static final long serialVersionUID = -7914297376763021607L;
2624		}
2625
2626		/**
#	Line 1499 | Line 2652 | public class ForkJoinPool extends Abstra
2652		}
2653
2654		/**
2655	+	* Returns the targeted parallelism level of the common pool.
2656	+	*
2657	+	* @return the targeted parallelism level of the common pool
2658	+	*/
2659	+	public static int getCommonPoolParallelism() {
2660	+	return commonPoolParallelism;
2661	+	}
2662	+
2663	+	/**
2664		* Returns the number of worker threads that have started but not
2665	<	* yet terminated.  This result returned by this method may differ
2665	>	* yet terminated.  The result returned by this method may differ
2666		* from {@link #getParallelism} when threads are created to
2667		* maintain parallelism when others are cooperatively blocked.
2668		*
2669		* @return the number of worker threads
2670		*/
2671		public int getPoolSize() {
2672	<	return workerCounts >>> TOTAL_COUNT_SHIFT;
2672	>	return parallelism + (short)(ctl >>> TC_SHIFT);
2673		}
2674
2675		/**
#	Line 1517 | Line 2679 | public class ForkJoinPool extends Abstra
2679		* @return {@code true} if this pool uses async mode
2680		*/
2681		public boolean getAsyncMode() {
2682	<	return locallyFifo;
2682	>	return localMode != 0;
2683		}
2684
2685		/**
#	Line 1529 | Line 2691 | public class ForkJoinPool extends Abstra
2691		* @return the number of worker threads
2692		*/
2693		public int getRunningThreadCount() {
2694	<	return workerCounts & RUNNING_COUNT_MASK;
2694	>	int rc = 0;
2695	>	WorkQueue[] ws; WorkQueue w;
2696	>	if ((ws = workQueues) != null) {
2697	>	for (int i = 1; i < ws.length; i += 2) {
2698	>	if ((w = ws[i]) != null && w.isApparentlyUnblocked())
2699	>	++rc;
2700	>	}
2701	>	}
2702	>	return rc;
2703		}
2704
2705		/**
#	Line 1540 | Line 2710 | public class ForkJoinPool extends Abstra
2710		* @return the number of active threads
2711		*/
2712		public int getActiveThreadCount() {
2713	<	return runState & ACTIVE_COUNT_MASK;
2713	>	int r = parallelism + (int)(ctl >> AC_SHIFT);
2714	>	return (r <= 0) ? 0 : r; // suppress momentarily negative values
2715		}
2716
2717		/**
#	Line 1555 | Line 2726 | public class ForkJoinPool extends Abstra
2726		* @return {@code true} if all threads are currently idle
2727		*/
2728		public boolean isQuiescent() {
2729	<	return (runState & ACTIVE_COUNT_MASK) == 0;
2729	>	return (int)(ctl >> AC_SHIFT) + parallelism == 0;
2730		}
2731
2732		/**
#	Line 1570 | Line 2741 | public class ForkJoinPool extends Abstra
2741		* @return the number of steals
2742		*/
2743		public long getStealCount() {
2744	<	return stealCount;
2744	>	long count = stealCount;
2745	>	WorkQueue[] ws; WorkQueue w;
2746	>	if ((ws = workQueues) != null) {
2747	>	for (int i = 1; i < ws.length; i += 2) {
2748	>	if ((w = ws[i]) != null)
2749	>	count += w.nsteals;
2750	>	}
2751	>	}
2752	>	return count;
2753		}
2754
2755		/**
#	Line 1585 | Line 2764 | public class ForkJoinPool extends Abstra
2764		*/
2765		public long getQueuedTaskCount() {
2766		long count = 0;
2767	<	ForkJoinWorkerThread[] ws = workers;
2768	<	int nws = ws.length;
2769	<	for (int i = 0; i < nws; ++i) {
2770	<	ForkJoinWorkerThread w = ws[i];
2771	<	if (w != null)
2772	<	count += w.getQueueSize();
2767	>	WorkQueue[] ws; WorkQueue w;
2768	>	if ((ws = workQueues) != null) {
2769	>	for (int i = 1; i < ws.length; i += 2) {
2770	>	if ((w = ws[i]) != null)
2771	>	count += w.queueSize();
2772	>	}
2773		}
2774		return count;
2775		}
2776
2777		/**
2778		* Returns an estimate of the number of tasks submitted to this
2779	<	* pool that have not yet begun executing.  This method takes time
2780	<	* proportional to the number of submissions.
2779	>	* pool that have not yet begun executing.  This method may take
2780	>	* time proportional to the number of submissions.
2781		*
2782		* @return the number of queued submissions
2783		*/
2784		public int getQueuedSubmissionCount() {
2785	<	return submissionQueue.size();
2785	>	int count = 0;
2786	>	WorkQueue[] ws; WorkQueue w;
2787	>	if ((ws = workQueues) != null) {
2788	>	for (int i = 0; i < ws.length; i += 2) {
2789	>	if ((w = ws[i]) != null)
2790	>	count += w.queueSize();
2791	>	}
2792	>	}
2793	>	return count;
2794		}
2795
2796		/**
#	Line 1613 | Line 2800 | public class ForkJoinPool extends Abstra
2800		* @return {@code true} if there are any queued submissions
2801		*/
2802		public boolean hasQueuedSubmissions() {
2803	<	return !submissionQueue.isEmpty();
2803	>	WorkQueue[] ws; WorkQueue w;
2804	>	if ((ws = workQueues) != null) {
2805	>	for (int i = 0; i < ws.length; i += 2) {
2806	>	if ((w = ws[i]) != null && w.queueSize() != 0)
2807	>	return true;
2808	>	}
2809	>	}
2810	>	return false;
2811		}
2812
2813		/**
#	Line 1624 | Line 2818 | public class ForkJoinPool extends Abstra
2818		* @return the next submission, or {@code null} if none
2819		*/
2820		protected ForkJoinTask<?> pollSubmission() {
2821	<	return submissionQueue.poll();
2821	>	WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2822	>	if ((ws = workQueues) != null) {
2823	>	for (int i = 0; i < ws.length; i += 2) {
2824	>	if ((w = ws[i]) != null && (t = w.poll()) != null)
2825	>	return t;
2826	>	}
2827	>	}
2828	>	return null;
2829		}
2830
2831		/**
#	Line 1645 | Line 2846 | public class ForkJoinPool extends Abstra
2846		* @return the number of elements transferred
2847		*/
2848		protected int drainTasksTo(Collection<? super ForkJoinTask<?>> c) {
1648	–	int n = submissionQueue.drainTo(c);
1649	–	ForkJoinWorkerThread[] ws = workers;
1650	–	int nws = ws.length;
1651	–	for (int i = 0; i < nws; ++i) {
1652	–	ForkJoinWorkerThread w = ws[i];
1653	–	if (w != null)
1654	–	n += w.drainTasksTo(c);
1655	–	}
1656	–	return n;
1657	–	}
1658	–
1659	–	/**
1660	–	* Returns count of total parks by existing workers.
1661	–	* Used during development only since not meaningful to users.
1662	–	*/
1663	–	private int collectParkCount() {
2849		int count = 0;
2850	<	ForkJoinWorkerThread[] ws = workers;
2851	<	int nws = ws.length;
2852	<	for (int i = 0; i < nws; ++i) {
2853	<	ForkJoinWorkerThread w = ws[i];
2854	<	if (w != null)
2855	<	count += w.parkCount;
2850	>	WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2851	>	if ((ws = workQueues) != null) {
2852	>	for (int i = 0; i < ws.length; ++i) {
2853	>	if ((w = ws[i]) != null) {
2854	>	while ((t = w.poll()) != null) {
2855	>	c.add(t);
2856	>	++count;
2857	>	}
2858	>	}
2859	>	}
2860		}
2861		return count;
2862		}
#	Line 1680 | Line 2869 | public class ForkJoinPool extends Abstra
2869		* @return a string identifying this pool, as well as its state
2870		*/
2871		public String toString() {
2872	<	long st = getStealCount();
2873	<	long qt = getQueuedTaskCount();
2874	<	long qs = getQueuedSubmissionCount();
2875	<	int wc = workerCounts;
2876	<	int tc = wc >>> TOTAL_COUNT_SHIFT;
2877	<	int rc = wc & RUNNING_COUNT_MASK;
2872	>	// Use a single pass through workQueues to collect counts
2873	>	long qt = 0L, qs = 0L; int rc = 0;
2874	>	long st = stealCount;
2875	>	long c = ctl;
2876	>	WorkQueue[] ws; WorkQueue w;
2877	>	if ((ws = workQueues) != null) {
2878	>	for (int i = 0; i < ws.length; ++i) {
2879	>	if ((w = ws[i]) != null) {
2880	>	int size = w.queueSize();
2881	>	if ((i & 1) == 0)
2882	>	qs += size;
2883	>	else {
2884	>	qt += size;
2885	>	st += w.nsteals;
2886	>	if (w.isApparentlyUnblocked())
2887	>	++rc;
2888	>	}
2889	>	}
2890	>	}
2891	>	}
2892		int pc = parallelism;
2893	<	int rs = runState;
2894	<	int ac = rs & ACTIVE_COUNT_MASK;
2895	<	//        int pk = collectParkCount();
2893	>	int tc = pc + (short)(c >>> TC_SHIFT);
2894	>	int ac = pc + (int)(c >> AC_SHIFT);
2895	>	if (ac < 0) // ignore transient negative
2896	>	ac = 0;
2897	>	String level;
2898	>	if ((c & STOP_BIT) != 0)
2899	>	level = (tc == 0) ? "Terminated" : "Terminating";
2900	>	else
2901	>	level = plock < 0 ? "Shutting down" : "Running";
2902		return super.toString() +
2903	<	"[" + runLevelToString(rs) +
2903	>	"[" + level +
2904		", parallelism = " + pc +
2905		", size = " + tc +
2906		", active = " + ac +
#	Line 1699 | Line 2908 | public class ForkJoinPool extends Abstra
2908		", steals = " + st +
2909		", tasks = " + qt +
2910		", submissions = " + qs +
1702	–	//            ", parks = " + pk +
2911		"]";
2912		}
2913
1706	–	private static String runLevelToString(int s) {
1707	–	return ((s & TERMINATED) != 0 ? "Terminated" :
1708	–	((s & TERMINATING) != 0 ? "Terminating" :
1709	–	((s & SHUTDOWN) != 0 ? "Shutting down" :
1710	–	"Running")));
1711	–	}
1712	–
2914		/**
2915	<	* Initiates an orderly shutdown in which previously submitted
2916	<	* tasks are executed, but no new tasks will be accepted.
2917	<	* Invocation has no additional effect if already shut down.
2918	<	* Tasks that are in the process of being submitted concurrently
2919	<	* during the course of this method may or may not be rejected.
2915	>	* Possibly initiates an orderly shutdown in which previously
2916	>	* submitted tasks are executed, but no new tasks will be
2917	>	* accepted. Invocation has no effect on execution state if this
2918	>	* is the {@link #commonPool}, and no additional effect if
2919	>	* already shut down.  Tasks that are in the process of being
2920	>	* submitted concurrently during the course of this method may or
2921	>	* may not be rejected.
2922		*
2923		* @throws SecurityException if a security manager exists and
2924		*         the caller is not permitted to modify threads
#	Line 1724 | Line 2927 | public class ForkJoinPool extends Abstra
2927		*/
2928		public void shutdown() {
2929		checkPermission();
2930	<	advanceRunLevel(SHUTDOWN);
1728	<	tryTerminate(false);
2930	>	tryTerminate(false, true);
2931		}
2932
2933		/**
2934	<	* Attempts to cancel and/or stop all tasks, and reject all
2935	<	* subsequently submitted tasks.  Tasks that are in the process of
2936	<	* being submitted or executed concurrently during the course of
2937	<	* this method may or may not be rejected. This method cancels
2938	<	* both existing and unexecuted tasks, in order to permit
2939	<	* termination in the presence of task dependencies. So the method
2940	<	* always returns an empty list (unlike the case for some other
2941	<	* Executors).
2934	>	* Possibly attempts to cancel and/or stop all tasks, and reject
2935	>	* all subsequently submitted tasks.  Invocation has no effect on
2936	>	* execution state if this is the {@link #commonPool}, and no
2937	>	* additional effect if already shut down. Otherwise, tasks that
2938	>	* are in the process of being submitted or executed concurrently
2939	>	* during the course of this method may or may not be
2940	>	* rejected. This method cancels both existing and unexecuted
2941	>	* tasks, in order to permit termination in the presence of task
2942	>	* dependencies. So the method always returns an empty list
2943	>	* (unlike the case for some other Executors).
2944		*
2945		* @return an empty list
2946		* @throws SecurityException if a security manager exists and
#	Line 1746 | Line 2950 | public class ForkJoinPool extends Abstra
2950		*/
2951		public List<Runnable> shutdownNow() {
2952		checkPermission();
2953	<	tryTerminate(true);
2953	>	tryTerminate(true, true);
2954		return Collections.emptyList();
2955		}
2956
#	Line 1756 | Line 2960 | public class ForkJoinPool extends Abstra
2960		* @return {@code true} if all tasks have completed following shut down
2961		*/
2962		public boolean isTerminated() {
2963	<	return runState >= TERMINATED;
2963	>	long c = ctl;
2964	>	return ((c & STOP_BIT) != 0L &&
2965	>	(short)(c >>> TC_SHIFT) == -parallelism);
2966		}
2967
2968		/**
#	Line 1764 | Line 2970 | public class ForkJoinPool extends Abstra
2970		* commenced but not yet completed.  This method may be useful for
2971		* debugging. A return of {@code true} reported a sufficient
2972		* period after shutdown may indicate that submitted tasks have
2973	<	* ignored or suppressed interruption, causing this executor not
2974	<	* to properly terminate.
2973	>	* ignored or suppressed interruption, or are waiting for IO,
2974	>	* causing this executor not to properly terminate. (See the
2975	>	* advisory notes for class {@link ForkJoinTask} stating that
2976	>	* tasks should not normally entail blocking operations.  But if
2977	>	* they do, they must abort them on interrupt.)
2978		*
2979		* @return {@code true} if terminating but not yet terminated
2980		*/
2981		public boolean isTerminating() {
2982	<	return (runState & (TERMINATING|TERMINATED)) == TERMINATING;
2982	>	long c = ctl;
2983	>	return ((c & STOP_BIT) != 0L &&
2984	>	(short)(c >>> TC_SHIFT) != -parallelism);
2985		}
2986
2987		/**
#	Line 1779 | Line 2990 | public class ForkJoinPool extends Abstra
2990		* @return {@code true} if this pool has been shut down
2991		*/
2992		public boolean isShutdown() {
2993	<	return runState >= SHUTDOWN;
2993	>	return plock < 0;
2994		}
2995
2996		/**
2997	<	* Blocks until all tasks have completed execution after a shutdown
2998	<	* request, or the timeout occurs, or the current thread is
2999	<	* interrupted, whichever happens first.
2997	>	* Blocks until all tasks have completed execution after a
2998	>	* shutdown request, or the timeout occurs, or the current thread
2999	>	* is interrupted, whichever happens first. Note that the {@link
3000	>	* #commonPool()} never terminates until program shutdown so
3001	>	* this method will always time out.
3002		*
3003		* @param timeout the maximum time to wait
3004		* @param unit the time unit of the timeout argument
#	Line 1795 | Line 3008 | public class ForkJoinPool extends Abstra
3008		*/
3009		public boolean awaitTermination(long timeout, TimeUnit unit)
3010		throws InterruptedException {
3011	<	try {
3012	<	return termination.awaitAdvanceInterruptibly(0, timeout, unit) > 0;
3013	<	} catch(TimeoutException ex) {
3014	<	return false;
3011	>	long nanos = unit.toNanos(timeout);
3012	>	if (isTerminated())
3013	>	return true;
3014	>	long startTime = System.nanoTime();
3015	>	boolean terminated = false;
3016	>	synchronized (this) {
3017	>	for (long waitTime = nanos, millis = 0L;;) {
3018	>	if (terminated = isTerminated() ||
3019	>	waitTime <= 0L ||
3020	>	(millis = unit.toMillis(waitTime)) <= 0L)
3021	>	break;
3022	>	wait(millis);
3023	>	waitTime = nanos - (System.nanoTime() - startTime);
3024	>	}
3025		}
3026	+	return terminated;
3027		}
3028
3029		/**
3030		* Interface for extending managed parallelism for tasks running
3031		* in {@link ForkJoinPool}s.
3032		*
3033	<	* <p>A {@code ManagedBlocker} provides two methods.
3034	<	* Method {@code isReleasable} must return {@code true} if
3035	<	* blocking is not necessary. Method {@code block} blocks the
3036	<	* current thread if necessary (perhaps internally invoking
3037	<	* {@code isReleasable} before actually blocking).
3033	>	* <p>A {@code ManagedBlocker} provides two methods.  Method
3034	>	* {@code isReleasable} must return {@code true} if blocking is
3035	>	* not necessary. Method {@code block} blocks the current thread
3036	>	* if necessary (perhaps internally invoking {@code isReleasable}
3037	>	* before actually blocking). These actions are performed by any
3038	>	* thread invoking {@link ForkJoinPool#managedBlock}.  The
3039	>	* unusual methods in this API accommodate synchronizers that may,
3040	>	* but don't usually, block for long periods. Similarly, they
3041	>	* allow more efficient internal handling of cases in which
3042	>	* additional workers may be, but usually are not, needed to
3043	>	* ensure sufficient parallelism.  Toward this end,
3044	>	* implementations of method {@code isReleasable} must be amenable
3045	>	* to repeated invocation.
3046		*
3047		* <p>For example, here is a ManagedBlocker based on a
3048		* ReentrantLock:
#	Line 1828 | Line 3060 | public class ForkJoinPool extends Abstra
3060		*     return hasLock || (hasLock = lock.tryLock());
3061		*   }
3062		* }}</pre>
3063	+	*
3064	+	* <p>Here is a class that possibly blocks waiting for an
3065	+	* item on a given queue:
3066	+	*  <pre> {@code
3067	+	* class QueueTaker<E> implements ManagedBlocker {
3068	+	*   final BlockingQueue<E> queue;
3069	+	*   volatile E item = null;
3070	+	*   QueueTaker(BlockingQueue<E> q) { this.queue = q; }
3071	+	*   public boolean block() throws InterruptedException {
3072	+	*     if (item == null)
3073	+	*       item = queue.take();
3074	+	*     return true;
3075	+	*   }
3076	+	*   public boolean isReleasable() {
3077	+	*     return item != null || (item = queue.poll()) != null;
3078	+	*   }
3079	+	*   public E getItem() { // call after pool.managedBlock completes
3080	+	*     return item;
3081	+	*   }
3082	+	* }}</pre>
3083		*/
3084		public static interface ManagedBlocker {
3085		/**
#	Line 1870 | Line 3122 | public class ForkJoinPool extends Abstra
3122		public static void managedBlock(ManagedBlocker blocker)
3123		throws InterruptedException {
3124		Thread t = Thread.currentThread();
3125	<	if (t instanceof ForkJoinWorkerThread)
3126	<	((ForkJoinWorkerThread) t).pool.awaitBlocker(blocker);
3125	>	if (t instanceof ForkJoinWorkerThread) {
3126	>	ForkJoinPool p = ((ForkJoinWorkerThread)t).pool;
3127	>	while (!blocker.isReleasable()) { // variant of helpSignal
3128	>	WorkQueue[] ws; WorkQueue q; int m, n;
3129	>	if ((ws = p.workQueues) != null && (m = ws.length - 1) >= 0) {
3130	>	for (int i = 0; i <= m; ++i) {
3131	>	if (blocker.isReleasable())
3132	>	return;
3133	>	if ((q = ws[i]) != null && (n = q.queueSize()) > 0) {
3134	>	p.signalWork(q, n);
3135	>	if ((int)(p.ctl >> AC_SHIFT) >= 0)
3136	>	break;
3137	>	}
3138	>	}
3139	>	}
3140	>	if (p.tryCompensate()) {
3141	>	try {
3142	>	do {} while (!blocker.isReleasable() &&
3143	>	!blocker.block());
3144	>	} finally {
3145	>	p.incrementActiveCount();
3146	>	}
3147	>	break;
3148	>	}
3149	>	}
3150	>	}
3151		else {
3152	<	do {} while (!blocker.isReleasable() && !blocker.block());
3152	>	do {} while (!blocker.isReleasable() &&
3153	>	!blocker.block());
3154		}
3155		}
3156
#	Line 1882 | Line 3159 | public class ForkJoinPool extends Abstra
3159		// implement RunnableFuture.
3160
3161		protected <T> RunnableFuture<T> newTaskFor(Runnable runnable, T value) {
3162	<	return (RunnableFuture<T>) ForkJoinTask.adapt(runnable, value);
3162	>	return new ForkJoinTask.AdaptedRunnable<T>(runnable, value);
3163		}
3164
3165		protected <T> RunnableFuture<T> newTaskFor(Callable<T> callable) {
3166	<	return (RunnableFuture<T>) ForkJoinTask.adapt(callable);
3166	>	return new ForkJoinTask.AdaptedCallable<T>(callable);
3167		}
3168
3169		// Unsafe mechanics
3170	<
3171	<	private static final sun.misc.Unsafe UNSAFE = getUnsafe();
3172	<	private static final long workerCountsOffset =
3173	<	objectFieldOffset("workerCounts", ForkJoinPool.class);
3174	<	private static final long runStateOffset =
3175	<	objectFieldOffset("runState", ForkJoinPool.class);
3176	<	private static final long eventCountOffset =
3177	<	objectFieldOffset("eventCount", ForkJoinPool.class);
3178	<	private static final long eventWaitersOffset =
3179	<	objectFieldOffset("eventWaiters",ForkJoinPool.class);
3180	<	private static final long stealCountOffset =
3181	<	objectFieldOffset("stealCount",ForkJoinPool.class);
3182	<
3183	<	private static long objectFieldOffset(String field, Class<?> klazz) {
3170	>	private static final sun.misc.Unsafe U;
3171	>	private static final long CTL;
3172	>	private static final long PARKBLOCKER;
3173	>	private static final int ABASE;
3174	>	private static final int ASHIFT;
3175	>	private static final long STEALCOUNT;
3176	>	private static final long PLOCK;
3177	>	private static final long INDEXSEED;
3178	>	private static final long QLOCK;
3179	>
3180	>	static {
3181	>	// Establish common pool parameters
3182	>	// TBD: limit or report ignored exceptions?
3183	>
3184	>	int par = 0;
3185	>	ForkJoinWorkerThreadFactory fac = null;
3186	>	Thread.UncaughtExceptionHandler handler = null;
3187		try {
3188	<	return UNSAFE.objectFieldOffset(klazz.getDeclaredField(field));
3189	<	} catch (NoSuchFieldException e) {
3190	<	// Convert Exception to corresponding Error
3191	<	NoSuchFieldError error = new NoSuchFieldError(field);
3192	<	error.initCause(e);
3193	<	throw error;
3188	>	String pp = System.getProperty(propPrefix + "parallelism");
3189	>	String hp = System.getProperty(propPrefix + "exceptionHandler");
3190	>	String fp = System.getProperty(propPrefix + "threadFactory");
3191	>	if (fp != null)
3192	>	fac = ((ForkJoinWorkerThreadFactory)ClassLoader.
3193	>	getSystemClassLoader().loadClass(fp).newInstance());
3194	>	if (hp != null)
3195	>	handler = ((Thread.UncaughtExceptionHandler)ClassLoader.
3196	>	getSystemClassLoader().loadClass(hp).newInstance());
3197	>	if (pp != null)
3198	>	par = Integer.parseInt(pp);
3199	>	} catch (Exception ignore) {
3200		}
3201	+
3202	+	int s; // initialize field offsets for CAS etc
3203	+	try {
3204	+	U = getUnsafe();
3205	+	Class<?> k = ForkJoinPool.class;
3206	+	CTL = U.objectFieldOffset
3207	+	(k.getDeclaredField("ctl"));
3208	+	STEALCOUNT = U.objectFieldOffset
3209	+	(k.getDeclaredField("stealCount"));
3210	+	PLOCK = U.objectFieldOffset
3211	+	(k.getDeclaredField("plock"));
3212	+	INDEXSEED = U.objectFieldOffset
3213	+	(k.getDeclaredField("indexSeed"));
3214	+	Class<?> tk = Thread.class;
3215	+	PARKBLOCKER = U.objectFieldOffset
3216	+	(tk.getDeclaredField("parkBlocker"));
3217	+	Class<?> wk = WorkQueue.class;
3218	+	QLOCK = U.objectFieldOffset
3219	+	(wk.getDeclaredField("qlock"));
3220	+	Class<?> ak = ForkJoinTask[].class;
3221	+	ABASE = U.arrayBaseOffset(ak);
3222	+	s = U.arrayIndexScale(ak);
3223	+	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
3224	+	} catch (Exception e) {
3225	+	throw new Error(e);
3226	+	}
3227	+	if ((s & (s-1)) != 0)
3228	+	throw new Error("data type scale not a power of two");
3229	+
3230	+	/*
3231	+	* For extra caution, computations to set up pool state are
3232	+	* here; the constructor just assigns these values to fields.
3233	+	*/
3234	+	ForkJoinWorkerThreadFactory defaultFac =
3235	+	defaultForkJoinWorkerThreadFactory =
3236	+	new DefaultForkJoinWorkerThreadFactory();
3237	+	if (fac == null)
3238	+	fac = defaultFac;
3239	+	if (par <= 0)
3240	+	par = Runtime.getRuntime().availableProcessors();
3241	+	if (par > MAX_CAP)
3242	+	par = MAX_CAP;
3243	+	long np = (long)(-par); // precompute initial ctl value
3244	+	long ct = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK);
3245	+
3246	+	commonPoolParallelism = par;
3247	+	commonPool = new ForkJoinPool(par, ct, fac, handler);
3248	+	modifyThreadPermission = new RuntimePermission("modifyThread");
3249	+	submitters = new ThreadLocal<Submitter>();
3250		}
3251
3252		/**
#	Line 1941 | Line 3276 | public class ForkJoinPool extends Abstra
3276		}
3277		}
3278		}
3279	+
3280		}

Diff Legend

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