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 1.98 by dl, Mon Mar 21 23:29:03 2011 UTC vs.
Revision 1.116 by dl, Fri Jan 27 17:27:28 2012 UTC

#	Line 19 | Line 19 | import java.util.concurrent.Future;
19		import java.util.concurrent.RejectedExecutionException;
20		import java.util.concurrent.RunnableFuture;
21		import java.util.concurrent.TimeUnit;
22	–	import java.util.concurrent.TimeoutException;
22		import java.util.concurrent.atomic.AtomicInteger;
23	<	import java.util.concurrent.locks.LockSupport;
23	>	import java.util.concurrent.atomic.AtomicLong;
24		import java.util.concurrent.locks.ReentrantLock;
25		import java.util.concurrent.locks.Condition;
26
#	Line 34 | Line 33 | import java.util.concurrent.locks.Condit
33		* <p>A {@code ForkJoinPool} differs from other kinds of {@link
34		* ExecutorService} mainly by virtue of employing
35		* <em>work-stealing</em>: all threads in the pool attempt to find and
36	<	* execute subtasks created by other active tasks (eventually blocking
37	<	* waiting for work if none exist). This enables efficient processing
38	<	* when most tasks spawn other subtasks (as do most {@code
39	<	* ForkJoinTask}s). When setting <em>asyncMode</em> to true in
40	<	* constructors, {@code ForkJoinPool}s may also be appropriate for use
41	<	* with event-style tasks that are never joined.
36	>	* execute tasks submitted to the pool and/or created by other active
37	>	* tasks (eventually blocking waiting for work if none exist). This
38	>	* enables efficient processing when most tasks spawn other subtasks
39	>	* (as do most {@code ForkJoinTask}s), as well as when many small
40	>	* tasks are submitted to the pool from external clients.  Especially
41	>	* when setting <em>asyncMode</em> to true in constructors, {@code
42	>	* ForkJoinPool}s may also be appropriate for use with event-style
43	>	* tasks that are never joined.
44		*
45		* <p>A {@code ForkJoinPool} is constructed with a given target
46		* parallelism level; by default, equal to the number of available
#	Line 60 | Line 61 | import java.util.concurrent.locks.Condit
61		*
62		* <p> As is the case with other ExecutorServices, there are three
63		* main task execution methods summarized in the following
64	<	* table. These are designed to be used by clients not already engaged
65	<	* in fork/join computations in the current pool.  The main forms of
66	<	* these methods accept instances of {@code ForkJoinTask}, but
67	<	* overloaded forms also allow mixed execution of plain {@code
68	<	* Runnable}- or {@code Callable}- based activities as well.  However,
69	<	* tasks that are already executing in a pool should normally
70	<	* <em>NOT</em> use these pool execution methods, but instead use the
71	<	* within-computation forms listed in the table.
64	>	* table. These are designed to be used primarily by clients not
65	>	* already engaged in fork/join computations in the current pool.  The
66	>	* main forms of these methods accept instances of {@code
67	>	* ForkJoinTask}, but overloaded forms also allow mixed execution of
68	>	* plain {@code Runnable}- or {@code Callable}- based activities as
69	>	* well.  However, tasks that are already executing in a pool should
70	>	* normally instead use the within-computation forms listed in the
71	>	* table unless using async event-style tasks that are not usually
72	>	* joined, in which case there is little difference among choice of
73	>	* methods.
74		*
75		* <table BORDER CELLPADDING=3 CELLSPACING=1>
76		*  <tr>
#	Line 102 | Line 105 | import java.util.concurrent.locks.Condit
105		* daemon} mode, there is typically no need to explicitly {@link
106		* #shutdown} such a pool upon program exit.
107		*
108	<	* <pre>
108	>	*  <pre> {@code
109		* static final ForkJoinPool mainPool = new ForkJoinPool();
110		* ...
111		* public void sort(long[] array) {
112		*   mainPool.invoke(new SortTask(array, 0, array.length));
113	<	* }
111	<	* </pre>
113	>	* }}</pre>
114		*
115		* <p><b>Implementation notes</b>: This implementation restricts the
116		* maximum number of running threads to 32767. Attempts to create
#	Line 127 | Line 129 | public class ForkJoinPool extends Abstra
129		/*
130		* Implementation Overview
131		*
132	<	* This class provides the central bookkeeping and control for a
133	<	* set of worker threads: Submissions from non-FJ threads enter
134	<	* into a submission queue. Workers take these tasks and typically
135	<	* split them into subtasks that may be stolen by other workers.
136	<	* Preference rules give first priority to processing tasks from
137	<	* their own queues (LIFO or FIFO, depending on mode), then to
138	<	* randomized FIFO steals of tasks in other worker queues, and
139	<	* lastly to new submissions.
132	>	* This class and its nested classes provide the main
133	>	* functionality and control for a set of worker threads:
134	>	* Submissions from non-FJ threads enter into submission
135	>	* queues. Workers take these tasks and typically split them into
136	>	* subtasks that may be stolen by other workers.  Preference rules
137	>	* give first priority to processing tasks from their own queues
138	>	* (LIFO or FIFO, depending on mode), then to randomized FIFO
139	>	* steals of tasks in other queues.
140	>	*
141	>	* WorkQueues.
142	>	* ==========
143	>	*
144	>	* Most operations occur within work-stealing queues (in nested
145	>	* class WorkQueue).  These are special forms of Deques that
146	>	* support only three of the four possible end-operations -- push,
147	>	* pop, and poll (aka steal), under the further constraints that
148	>	* push and pop are called only from the owning thread (or, as
149	>	* extended here, under a lock), while poll may be called from
150	>	* other threads.  (If you are unfamiliar with them, you probably
151	>	* want to read Herlihy and Shavit's book "The Art of
152	>	* Multiprocessor programming", chapter 16 describing these in
153	>	* more detail before proceeding.)  The main work-stealing queue
154	>	* design is roughly similar to those in the papers "Dynamic
155	>	* Circular Work-Stealing Deque" by Chase and Lev, SPAA 2005
156	>	* (http://research.sun.com/scalable/pubs/index.html) and
157	>	* "Idempotent work stealing" by Michael, Saraswat, and Vechev,
158	>	* PPoPP 2009 (http://portal.acm.org/citation.cfm?id=1504186).
159	>	* The main differences ultimately stem from gc requirements that
160	>	* we null out taken slots as soon as we can, to maintain as small
161	>	* a footprint as possible even in programs generating huge
162	>	* numbers of tasks. To accomplish this, we shift the CAS
163	>	* arbitrating pop vs poll (steal) from being on the indices
164	>	* ("base" and "top") to the slots themselves.  So, both a
165	>	* successful pop and poll mainly entail a CAS of a slot from
166	>	* non-null to null.  Because we rely on CASes of references, we
167	>	* do not need tag bits on base or top.  They are simple ints as
168	>	* used in any circular array-based queue (see for example
169	>	* ArrayDeque).  Updates to the indices must still be ordered in a
170	>	* way that guarantees that top == base means the queue is empty,
171	>	* but otherwise may err on the side of possibly making the queue
172	>	* appear nonempty when a push, pop, or poll have not fully
173	>	* committed. Note that this means that the poll operation,
174	>	* considered individually, is not wait-free. One thief cannot
175	>	* successfully continue until another in-progress one (or, if
176	>	* previously empty, a push) completes.  However, in the
177	>	* aggregate, we ensure at least probabilistic non-blockingness.
178	>	* If an attempted steal fails, a thief always chooses a different
179	>	* random victim target to try next. So, in order for one thief to
180	>	* progress, it suffices for any in-progress poll or new push on
181	>	* any empty queue to complete.
182	>	*
183	>	* This approach also enables support of a user mode in which local
184	>	* task processing is in FIFO, not LIFO order, simply by using
185	>	* poll rather than pop.  This can be useful in message-passing
186	>	* frameworks in which tasks are never joined.  However neither
187	>	* mode considers affinities, loads, cache localities, etc, so
188	>	* rarely provide the best possible performance on a given
189	>	* machine, but portably provide good throughput by averaging over
190	>	* these factors.  (Further, even if we did try to use such
191	>	* information, we do not usually have a basis for exploiting
192	>	* it. For example, some sets of tasks profit from cache
193	>	* affinities, but others are harmed by cache pollution effects.)
194	>	*
195	>	* WorkQueues are also used in a similar way for tasks submitted
196	>	* to the pool. We cannot mix these tasks in the same queues used
197	>	* for work-stealing (this would contaminate lifo/fifo
198	>	* processing). Instead, we loosely associate submission queues
199	>	* with submitting threads, using a form of hashing.  The
200	>	* ThreadLocal Submitter class contains a value initially used as
201	>	* a hash code for choosing existing queues, but may be randomly
202	>	* repositioned upon contention with other submitters.  In
203	>	* essence, submitters act like workers except that they never
204	>	* take tasks, and they are multiplexed on to a finite number of
205	>	* shared work queues. However, classes are set up so that future
206	>	* extensions could allow submitters to optionally help perform
207	>	* tasks as well. Pool submissions from internal workers are also
208	>	* allowed, but use randomized rather than thread-hashed queue
209	>	* indices to avoid imbalance.  Insertion of tasks in shared mode
210	>	* requires a lock (mainly to protect in the case of resizing) but
211	>	* we use only a simple spinlock (using bits in field runState),
212	>	* because submitters encountering a busy queue try or create
213	>	* others so never block.
214	>	*
215	>	* Management.
216	>	* ==========
217		*
218		* The main throughput advantages of work-stealing stem from
219		* decentralized control -- workers mostly take tasks from
220		* themselves or each other. We cannot negate this in the
221		* implementation of other management responsibilities. The main
222		* tactic for avoiding bottlenecks is packing nearly all
223	<	* essentially atomic control state into a single 64bit volatile
224	<	* variable ("ctl"). This variable is read on the order of 10-100
225	<	* times as often as it is modified (always via CAS). (There is
226	<	* some additional control state, for example variable "shutdown"
227	<	* for which we can cope with uncoordinated updates.)  This
228	<	* streamlines synchronization and control at the expense of messy
229	<	* constructions needed to repack status bits upon updates.
230	<	* Updates tend not to contend with each other except during
231	<	* bursts while submitted tasks begin or end.  In some cases when
232	<	* they do contend, threads can instead do something else
233	<	* (usually, scan for tasks) until contention subsides.
234	<	*
235	<	* To enable packing, we restrict maximum parallelism to (1<<15)-1
236	<	* (which is far in excess of normal operating range) to allow
237	<	* ids, counts, and their negations (used for thresholding) to fit
238	<	* into 16bit fields.
239	<	*
240	<	* Recording Workers.  Workers are recorded in the "workers" array
241	<	* that is created upon pool construction and expanded if (rarely)
242	<	* necessary.  This is an array as opposed to some other data
243	<	* structure to support index-based random steals by workers.
244	<	* Updates to the array recording new workers and unrecording
245	<	* terminated ones are protected from each other by a seqLock
246	<	* (scanGuard) but the array is otherwise concurrently readable,
168	<	* and accessed directly by workers. To simplify index-based
223	>	* essentially atomic control state into two volatile variables
224	>	* that are by far most often read (not written) as status and
225	>	* consistency checks
226	>	*
227	>	* Field "ctl" contains 64 bits holding all the information needed
228	>	* to atomically decide to add, inactivate, enqueue (on an event
229	>	* queue), dequeue, and/or re-activate workers.  To enable this
230	>	* packing, we restrict maximum parallelism to (1<<15)-1 (which is
231	>	* far in excess of normal operating range) to allow ids, counts,
232	>	* and their negations (used for thresholding) to fit into 16bit
233	>	* fields.
234	>	*
235	>	* Field "runState" contains 32 bits needed to register and
236	>	* deregister WorkQueues, as well as to enable shutdown. It is
237	>	* only modified under a lock (normally briefly held, but
238	>	* occasionally protecting allocations and resizings) but even
239	>	* when locked remains available to check consistency.
240	>	*
241	>	* Recording WorkQueues.  WorkQueues are recorded in the
242	>	* "workQueues" array that is created upon pool construction and
243	>	* expanded if necessary.  Updates to the array while recording
244	>	* new workers and unrecording terminated ones are protected from
245	>	* each other by a lock but the array is otherwise concurrently
246	>	* readable, and accessed directly.  To simplify index-based
247		* operations, the array size is always a power of two, and all
248	<	* readers must tolerate null slots. To avoid flailing during
249	<	* start-up, the array is presized to hold twice #parallelism
250	<	* workers (which is unlikely to need further resizing during
251	<	* execution). But to avoid dealing with so many null slots,
252	<	* variable scanGuard includes a mask for the nearest power of two
253	<	* that contains all current workers.  All worker thread creation
254	<	* is on-demand, triggered by task submissions, replacement of
255	<	* terminated workers, and/or compensation for blocked
256	<	* workers. However, all other support code is set up to work with
257	<	* other policies.  To ensure that we do not hold on to worker
258	<	* references that would prevent GC, ALL accesses to workers are
259	<	* via indices into the workers array (which is one source of some
260	<	* of the messy code constructions here). In essence, the workers
261	<	* array serves as a weak reference mechanism. Thus for example
262	<	* the wait queue field of ctl stores worker indices, not worker
263	<	* references.  Access to the workers in associated methods (for
264	<	* example signalWork) must both index-check and null-check the
265	<	* IDs. All such accesses ignore bad IDs by returning out early
266	<	* from what they are doing, since this can only be associated
267	<	* with termination, in which case it is OK to give up.
268	<	*
269	<	* All uses of the workers array, as well as queue arrays, check
270	<	* that the array is non-null (even if previously non-null). This
271	<	* allows nulling during termination, which is currently not
272	<	* necessary, but remains an option for resource-revocation-based
273	<	* shutdown schemes.
248	>	* readers must tolerate null slots. Shared (submission) queues
249	>	* are at even indices, worker queues at odd indices. Grouping
250	>	* them together in this way simplifies and speeds up task
251	>	* scanning. To avoid flailing during start-up, the array is
252	>	* presized to hold twice #parallelism workers (which is unlikely
253	>	* to need further resizing during execution). But to avoid
254	>	* dealing with so many null slots, variable runState includes a
255	>	* mask for the nearest power of two that contains all current
256	>	* workers.  All worker thread creation is on-demand, triggered by
257	>	* task submissions, replacement of terminated workers, and/or
258	>	* compensation for blocked workers. However, all other support
259	>	* code is set up to work with other policies.  To ensure that we
260	>	* do not hold on to worker references that would prevent GC, ALL
261	>	* accesses to workQueues are via indices into the workQueues
262	>	* array (which is one source of some of the messy code
263	>	* constructions here). In essence, the workQueues array serves as
264	>	* a weak reference mechanism. Thus for example the wait queue
265	>	* field of ctl stores indices, not references.  Access to the
266	>	* workQueues in associated methods (for example signalWork) must
267	>	* both index-check and null-check the IDs. All such accesses
268	>	* ignore bad IDs by returning out early from what they are doing,
269	>	* since this can only be associated with termination, in which
270	>	* case it is OK to give up.
271	>	*
272	>	* All uses of the workQueues array check that it is non-null
273	>	* (even if previously non-null). This allows nulling during
274	>	* termination, which is currently not necessary, but remains an
275	>	* option for resource-revocation-based shutdown schemes. It also
276	>	* helps reduce JIT issuance of uncommon-trap code, which tends to
277	>	* unnecessarily complicate control flow in some methods.
278		*
279	<	* Wait Queuing. Unlike HPC work-stealing frameworks, we cannot
279	>	* Event Queuing. Unlike HPC work-stealing frameworks, we cannot
280		* let workers spin indefinitely scanning for tasks when none can
281		* be found immediately, and we cannot start/resume workers unless
282		* there appear to be tasks available.  On the other hand, we must
283		* quickly prod them into action when new tasks are submitted or
284	<	* generated.  We park/unpark workers after placing in an event
285	<	* wait queue when they cannot find work. This "queue" is actually
286	<	* a simple Treiber stack, headed by the "id" field of ctl, plus a
287	<	* 15bit counter value to both wake up waiters (by advancing their
288	<	* count) and avoid ABA effects. Successors are held in worker
289	<	* field "nextWait".  Queuing deals with several intrinsic races,
290	<	* mainly that a task-producing thread can miss seeing (and
284	>	* generated. In many usages, ramp-up time to activate workers is
285	>	* the main limiting factor in overall performance (this is
286	>	* compounded at program start-up by JIT compilation and
287	>	* allocation). So we try to streamline this as much as possible.
288	>	* We park/unpark workers after placing in an event wait queue
289	>	* when they cannot find work. This "queue" is actually a simple
290	>	* Treiber stack, headed by the "id" field of ctl, plus a 15bit
291	>	* counter value (that reflects the number of times a worker has
292	>	* been inactivated) to avoid ABA effects (we need only as many
293	>	* version numbers as worker threads). Successors are held in
294	>	* field WorkQueue.nextWait.  Queuing deals with several intrinsic
295	>	* races, mainly that a task-producing thread can miss seeing (and
296		* signalling) another thread that gave up looking for work but
297		* has not yet entered the wait queue. We solve this by requiring
298	<	* a full sweep of all workers both before (in scan()) and after
299	<	* (in tryAwaitWork()) a newly waiting worker is added to the wait
300	<	* queue. During a rescan, the worker might release some other
301	<	* queued worker rather than itself, which has the same net
302	<	* effect. Because enqueued workers may actually be rescanning
303	<	* rather than waiting, we set and clear the "parked" field of
304	<	* ForkJoinWorkerThread to reduce unnecessary calls to unpark.
305	<	* (Use of the parked field requires a secondary recheck to avoid
306	<	* missed signals.)
298	>	* a full sweep of all workers (via repeated calls to method
299	>	* scan()) both before and after a newly waiting worker is added
300	>	* to the wait queue. During a rescan, the worker might release
301	>	* some other queued worker rather than itself, which has the same
302	>	* net effect. Because enqueued workers may actually be rescanning
303	>	* rather than waiting, we set and clear the "parker" field of
304	>	* Workqueues to reduce unnecessary calls to unpark.  (This
305	>	* requires a secondary recheck to avoid missed signals.)  Note
306	>	* the unusual conventions about Thread.interrupts surrounding
307	>	* parking and other blocking: Because interrupts are used solely
308	>	* to alert threads to check termination, which is checked anyway
309	>	* upon blocking, we clear status (using Thread.interrupted)
310	>	* before any call to park, so that park does not immediately
311	>	* return due to status being set via some other unrelated call to
312	>	* interrupt in user code.
313		*
314		* Signalling.  We create or wake up workers only when there
315		* appears to be at least one task they might be able to find and
316		* execute.  When a submission is added or another worker adds a
317	<	* task to a queue that previously had two or fewer tasks, they
317	>	* task to a queue that previously had fewer than two tasks, they
318		* signal waiting workers (or trigger creation of new ones if
319		* fewer than the given parallelism level -- see signalWork).
320	<	* These primary signals are buttressed by signals during rescans
321	<	* as well as those performed when a worker steals a task and
322	<	* notices that there are more tasks too; together these cover the
323	<	* signals needed in cases when more than two tasks are pushed
231	<	* but untaken.
320	>	* These primary signals are buttressed by signals during rescans;
321	>	* together these cover the signals needed in cases when more
322	>	* tasks are pushed but untaken, and improve performance compared
323	>	* to having one thread wake up all workers.
324		*
325		* Trimming workers. To release resources after periods of lack of
326		* use, a worker starting to wait when the pool is quiescent will
#	Line 236 | Line 328 | public class ForkJoinPool extends Abstra
328		* SHRINK_RATE nanosecs. This will slowly propagate, eventually
329		* terminating all workers after long periods of non-use.
330		*
331	<	* Submissions. External submissions are maintained in an
332	<	* array-based queue that is structured identically to
333	<	* ForkJoinWorkerThread queues except for the use of
334	<	* submissionLock in method addSubmission. Unlike the case for
335	<	* worker queues, multiple external threads can add new
336	<	* submissions, so adding requires a lock.
337	<	*
338	<	* Compensation. Beyond work-stealing support and lifecycle
339	<	* control, the main responsibility of this framework is to take
340	<	* actions when one worker is waiting to join a task stolen (or
341	<	* always held by) another.  Because we are multiplexing many
342	<	* tasks on to a pool of workers, we can't just let them block (as
343	<	* in Thread.join).  We also cannot just reassign the joiner's
344	<	* run-time stack with another and replace it later, which would
345	<	* be a form of "continuation", that even if possible is not
346	<	* necessarily a good idea since we sometimes need both an
347	<	* unblocked task and its continuation to progress. Instead we
348	<	* combine two tactics:
331	>	* Shutdown and Termination. A call to shutdownNow atomically sets
332	>	* a runState bit and then (non-atomically) sets each workers
333	>	* runState status, cancels all unprocessed tasks, and wakes up
334	>	* all waiting workers.  Detecting whether termination should
335	>	* commence after a non-abrupt shutdown() call requires more work
336	>	* and bookkeeping. We need consensus about quiescence (i.e., that
337	>	* there is no more work). The active count provides a primary
338	>	* indication but non-abrupt shutdown still requires a rechecking
339	>	* scan for any workers that are inactive but not queued.
340	>	*
341	>	* Joining Tasks.
342	>	* ==============
343	>	*
344	>	* Any of several actions may be taken when one worker is waiting
345	>	* to join a task stolen (or always held by) another.  Because we
346	>	* are multiplexing many tasks on to a pool of workers, we can't
347	>	* just let them block (as in Thread.join).  We also cannot just
348	>	* reassign the joiner's run-time stack with another and replace
349	>	* it later, which would be a form of "continuation", that even if
350	>	* possible is not necessarily a good idea since we sometimes need
351	>	* both an unblocked task and its continuation to
352	>	* progress. Instead we combine two tactics:
353		*
354		*   Helping: Arranging for the joiner to execute some task that it
355	<	*      would be running if the steal had not occurred.  Method
260	<	*      ForkJoinWorkerThread.joinTask tracks joining->stealing
261	<	*      links to try to find such a task.
355	>	*      would be running if the steal had not occurred.
356		*
357		*   Compensating: Unless there are already enough live threads,
358	<	*      method tryPreBlock() may create or re-activate a spare
359	<	*      thread to compensate for blocked joiners until they
360	<	*      unblock.
358	>	*      method tryCompensate() may create or re-activate a spare
359	>	*      thread to compensate for blocked joiners until they unblock.
360	>	*
361	>	* A third form (implemented in tryRemoveAndExec and
362	>	* tryPollForAndExec) amounts to helping a hypothetical
363	>	* compensator: If we can readily tell that a possible action of a
364	>	* compensator is to steal and execute the task being joined, the
365	>	* joining thread can do so directly, without the need for a
366	>	* compensation thread (although at the expense of larger run-time
367	>	* stacks, but the tradeoff is typically worthwhile).
368		*
369		* The ManagedBlocker extension API can't use helping so relies
370		* only on compensation in method awaitBlocker.
371		*
372	+	* The algorithm in tryHelpStealer entails a form of "linear"
373	+	* helping: Each worker records (in field currentSteal) the most
374	+	* recent task it stole from some other worker. Plus, it records
375	+	* (in field currentJoin) the task it is currently actively
376	+	* joining. Method tryHelpStealer uses these markers to try to
377	+	* find a worker to help (i.e., steal back a task from and execute
378	+	* it) that could hasten completion of the actively joined task.
379	+	* In essence, the joiner executes a task that would be on its own
380	+	* local deque had the to-be-joined task not been stolen. This may
381	+	* be seen as a conservative variant of the approach in Wagner &
382	+	* Calder "Leapfrogging: a portable technique for implementing
383	+	* efficient futures" SIGPLAN Notices, 1993
384	+	* (http://portal.acm.org/citation.cfm?id=155354). It differs in
385	+	* that: (1) We only maintain dependency links across workers upon
386	+	* steals, rather than use per-task bookkeeping.  This sometimes
387	+	* requires a linear scan of workers array to locate stealers, but
388	+	* often doesn't because stealers leave hints (that may become
389	+	* stale/wrong) of where to locate them.  A stealHint is only a
390	+	* hint because a worker might have had multiple steals and the
391	+	* hint records only one of them (usually the most current).
392	+	* Hinting isolates cost to when it is needed, rather than adding
393	+	* to per-task overhead.  (2) It is "shallow", ignoring nesting
394	+	* and potentially cyclic mutual steals.  (3) It is intentionally
395	+	* racy: field currentJoin is updated only while actively joining,
396	+	* which means that we miss links in the chain during long-lived
397	+	* tasks, GC stalls etc (which is OK since blocking in such cases
398	+	* is usually a good idea).  (4) We bound the number of attempts
399	+	* to find work (see MAX_HELP_DEPTH) and fall back to suspending
400	+	* the worker and if necessary replacing it with another.
401	+	*
402		* It is impossible to keep exactly the target parallelism number
403		* of threads running at any given time.  Determining the
404		* existence of conservatively safe helping targets, the
405		* availability of already-created spares, and the apparent need
406	<	* to create new spares are all racy and require heuristic
407	<	* guidance, so we rely on multiple retries of each.  Currently,
408	<	* in keeping with on-demand signalling policy, we compensate only
409	<	* if blocking would leave less than one active (non-waiting,
410	<	* non-blocked) worker. Additionally, to avoid some false alarms
411	<	* due to GC, lagging counters, system activity, etc, compensated
412	<	* blocking for joins is only attempted after rechecks stabilize
413	<	* (retries are interspersed with Thread.yield, for good
414	<	* citizenship).  The variable blockedCount, incremented before
284	<	* blocking and decremented after, is sometimes needed to
285	<	* distinguish cases of waiting for work vs blocking on joins or
286	<	* other managed sync. Both cases are equivalent for most pool
287	<	* control, so we can update non-atomically. (Additionally,
288	<	* contention on blockedCount alleviates some contention on ctl).
289	<	*
290	<	* Shutdown and Termination. A call to shutdownNow atomically sets
291	<	* the ctl stop bit and then (non-atomically) sets each workers
292	<	* "terminate" status, cancels all unprocessed tasks, and wakes up
293	<	* all waiting workers.  Detecting whether termination should
294	<	* commence after a non-abrupt shutdown() call requires more work
295	<	* and bookkeeping. We need consensus about quiesence (i.e., that
296	<	* there is no more work) which is reflected in active counts so
297	<	* long as there are no current blockers, as well as possible
298	<	* re-evaluations during independent changes in blocking or
299	<	* quiescing workers.
406	>	* to create new spares are all racy, so we rely on multiple
407	>	* retries of each.  Currently, in keeping with on-demand
408	>	* signalling policy, we compensate only if blocking would leave
409	>	* less than one active (non-waiting, non-blocked) worker.
410	>	* Additionally, to avoid some false alarms due to GC, lagging
411	>	* counters, system activity, etc, compensated blocking for joins
412	>	* is only attempted after rechecks stabilize in
413	>	* ForkJoinTask.awaitJoin. (Retries are interspersed with
414	>	* Thread.yield, for good citizenship.)
415		*
416		* Style notes: There is a lot of representation-level coupling
417		* among classes ForkJoinPool, ForkJoinWorkerThread, and
418	<	* ForkJoinTask.  Most fields of ForkJoinWorkerThread maintain
419	<	* data structures managed by ForkJoinPool, so are directly
420	<	* accessed.  Conversely we allow access to "workers" array by
421	<	* workers, and direct access to ForkJoinTask.status by both
422	<	* ForkJoinPool and ForkJoinWorkerThread.  There is little point
423	<	* trying to reduce this, since any associated future changes in
424	<	* representations will need to be accompanied by algorithmic
425	<	* changes anyway. All together, these low-level implementation
426	<	* choices produce as much as a factor of 4 performance
312	<	* improvement compared to naive implementations, and enable the
313	<	* processing of billions of tasks per second, at the expense of
314	<	* some ugliness.
418	>	* ForkJoinTask.  The fields of WorkQueue maintain data structures
419	>	* managed by ForkJoinPool, so are directly accessed.  There is
420	>	* little point trying to reduce this, since any associated future
421	>	* changes in representations will need to be accompanied by
422	>	* algorithmic changes anyway. All together, these low-level
423	>	* implementation choices produce as much as a factor of 4
424	>	* performance improvement compared to naive implementations, and
425	>	* enable the processing of billions of tasks per second, at the
426	>	* expense of some ugliness.
427		*
428		* Methods signalWork() and scan() are the main bottlenecks so are
429		* especially heavily micro-optimized/mangled.  There are lots of
#	Line 328 | Line 440 | public class ForkJoinPool extends Abstra
440		* The order of declarations in this file is: (1) declarations of
441		* statics (2) fields (along with constants used when unpacking
442		* some of them), listed in an order that tends to reduce
443	<	* contention among them a bit under most JVMs.  (3) internal
444	<	* control methods (4) callbacks and other support for
445	<	* ForkJoinTask and ForkJoinWorkerThread classes, (5) exported
446	<	* methods (plus a few little helpers). (6) static block
447	<	* initializing all statics in a minimally dependent order.
443	>	* contention among them a bit under most JVMs; (3) nested
444	>	* classes; (4) internal control methods; (5) callbacks and other
445	>	* support for ForkJoinTask methods; (6) exported methods (plus a
446	>	* few little helpers); (7) static block initializing all statics
447	>	* in a minimally dependent order.
448		*/
449
450		/**
#	Line 391 | Line 503 | public class ForkJoinPool extends Abstra
503		private static final AtomicInteger poolNumberGenerator;
504
505		/**
506	<	* Generator for initial random seeds for worker victim
507	<	* selection. This is used only to create initial seeds. Random
508	<	* steals use a cheaper xorshift generator per steal attempt. We
397	<	* don't expect much contention on seedGenerator, so just use a
398	<	* plain Random.
399	<	*/
400	<	static final Random workerSeedGenerator;
401	<
402	<	/**
403	<	* Array holding all worker threads in the pool.  Initialized upon
404	<	* construction. Array size must be a power of two.  Updates and
405	<	* replacements are protected by scanGuard, but the array is
406	<	* always kept in a consistent enough state to be randomly
407	<	* accessed without locking by workers performing work-stealing,
408	<	* as well as other traversal-based methods in this class, so long
409	<	* as reads memory-acquire by first reading ctl. All readers must
410	<	* tolerate that some array slots may be null.
411	<	*/
412	<	ForkJoinWorkerThread[] workers;
413	<
414	<	/**
415	<	* Initial size for submission queue array. Must be a power of
416	<	* two.  In many applications, these always stay small so we use a
417	<	* small initial cap.
418	<	*/
419	<	private static final int INITIAL_QUEUE_CAPACITY = 8;
420	<
421	<	/**
422	<	* Maximum size for submission queue array. Must be a power of two
423	<	* less than or equal to 1 << (31 - width of array entry) to
424	<	* ensure lack of index wraparound, but is capped at a lower
425	<	* value to help users trap runaway computations.
426	<	*/
427	<	private static final int MAXIMUM_QUEUE_CAPACITY = 1 << 24; // 16M
428	<
429	<	/**
430	<	* Array serving as submission queue. Initialized upon construction.
431	<	*/
432	<	private ForkJoinTask<?>[] submissionQueue;
433	<
434	<	/**
435	<	* Lock protecting submissions array for addSubmission
436	<	*/
437	<	private final ReentrantLock submissionLock;
438	<
439	<	/**
440	<	* Condition for awaitTermination, using submissionLock for
441	<	* convenience.
442	<	*/
443	<	private final Condition termination;
444	<
445	<	/**
446	<	* Creation factory for worker threads.
447	<	*/
448	<	private final ForkJoinWorkerThreadFactory factory;
449	<
450	<	/**
451	<	* The uncaught exception handler used when any worker abruptly
452	<	* terminates.
453	<	*/
454	<	final Thread.UncaughtExceptionHandler ueh;
455	<
456	<	/**
457	<	* Prefix for assigning names to worker threads
458	<	*/
459	<	private final String workerNamePrefix;
460	<
461	<	/**
462	<	* Sum of per-thread steal counts, updated only when threads are
463	<	* idle or terminating.
464	<	*/
465	<	private volatile long stealCount;
466	<
467	<	/**
468	<	* Main pool control -- a long packed with:
506	>	* Bits and masks for control variables
507	>	*
508	>	* Field ctl is a long packed with:
509		* AC: Number of active running workers minus target parallelism (16 bits)
510	<	* TC: Number of total workers minus target parallelism (16bits)
510	>	* TC: Number of total workers minus target parallelism (16 bits)
511		* ST: true if pool is terminating (1 bit)
512		* EC: the wait count of top waiting thread (15 bits)
513	<	* ID: ~poolIndex of top of Treiber stack of waiting threads (16 bits)
513	>	* ID: ~(poolIndex >>> 1) of top of Treiber stack of waiters (16 bits)
514		*
515		* When convenient, we can extract the upper 32 bits of counts and
516		* the lower 32 bits of queue state, u = (int)(ctl >>> 32) and e =
#	Line 484 | Line 524 | public class ForkJoinPool extends Abstra
524		* negative, the pool is terminating.  To deal with these possibly
525		* negative fields, we use casts in and out of "short" and/or
526		* signed shifts to maintain signedness.
527	+	*
528	+	* When a thread is queued (inactivated), its eventCount field is
529	+	* negative, which is the only way to tell if a worker is
530	+	* prevented from executing tasks, even though it must continue to
531	+	* scan for them to avoid queuing races.
532	+	*
533	+	* Field runState is an int packed with:
534	+	* SHUTDOWN: true if shutdown is enabled (1 bit)
535	+	* SEQ:  a sequence number updated upon (de)registering workers (15 bits)
536	+	* MASK: mask (power of 2 - 1) covering all registered poolIndexes (16 bits)
537	+	*
538	+	* The combination of mask and sequence number enables simple
539	+	* consistency checks: Staleness of read-only operations on the
540	+	* workers and queues arrays can be checked by comparing runState
541	+	* before vs after the reads. The low 16 bits (i.e, anding with
542	+	* SMASK) hold (the smallest power of two covering all worker
543	+	* indices, minus one.  The mask for queues (vs workers) is twice
544	+	* this value plus 1.
545		*/
488	–	volatile long ctl;
546
547		// bit positions/shifts for fields
548		private static final int  AC_SHIFT   = 48;
#	Line 517 | Line 574 | public class ForkJoinPool extends Abstra
574		private static final int  UTC_UNIT   = 1 << UTC_SHIFT;
575
576		// masks and units for dealing with e = (int)ctl
577	<	private static final int  E_MASK     = 0x7fffffff; // no STOP_BIT
578	<	private static final int  EC_UNIT    = 1 << EC_SHIFT;
577	>	private static final int E_MASK      = 0x7fffffff; // no STOP_BIT
578	>	private static final int E_SEQ       = 1 << EC_SHIFT;
579	>
580	>	// runState bits
581	>	private static final int SHUTDOWN    = 1 << 31;
582	>	private static final int RS_SEQ      = 1 << 16;
583	>	private static final int RS_SEQ_MASK = 0x7fff0000;
584	>
585	>	// access mode for WorkQueue
586	>	static final int LIFO_QUEUE          =  0;
587	>	static final int FIFO_QUEUE          =  1;
588	>	static final int SHARED_QUEUE        = -1;
589
590		/**
591	<	* The target parallelism level.
591	>	* The wakeup interval (in nanoseconds) for a worker waiting for a
592	>	* task when the pool is quiescent to instead try to shrink the
593	>	* number of workers.  The exact value does not matter too
594	>	* much. It must be short enough to release resources during
595	>	* sustained periods of idleness, but not so short that threads
596	>	* are continually re-created.
597		*/
598	<	final int parallelism;
598	>	private static final long SHRINK_RATE =
599	>	4L * 1000L * 1000L * 1000L; // 4 seconds
600
601		/**
602	<	* Index (mod submission queue length) of next element to take
603	<	* from submission queue. Usage is identical to that for
531	<	* per-worker queues -- see ForkJoinWorkerThread internal
532	<	* documentation.
602	>	* The timeout value for attempted shrinkage, includes
603	>	* some slop to cope with system timer imprecision.
604		*/
605	<	volatile int queueBase;
605	>	private static final long SHRINK_TIMEOUT = SHRINK_RATE - (SHRINK_RATE / 10);
606
607		/**
608	<	* Index (mod submission queue length) of next element to add
609	<	* in submission queue. Usage is identical to that for
610	<	* per-worker queues -- see ForkJoinWorkerThread internal
611	<	* documentation.
608	>	* The maximum stolen->joining link depth allowed in tryHelpStealer.
609	>	* Depths for legitimate chains are unbounded, but we use a fixed
610	>	* constant to avoid (otherwise unchecked) cycles and to bound
611	>	* staleness of traversal parameters at the expense of sometimes
612	>	* blocking when we could be helping.
613		*/
614	<	int queueTop;
614	>	private static final int MAX_HELP_DEPTH = 16;
615
616	<	/**
617	<	* True when shutdown() has been called.
616	>	/*
617	>	* Field layout order in this class tends to matter more than one
618	>	* would like. Runtime layout order is only loosely related to
619	>	* declaration order and may differ across JVMs, but the following
620	>	* empirically works OK on current JVMs.
621	>	*/
622	>
623	>	volatile long ctl;                       // main pool control
624	>	final int parallelism;                   // parallelism level
625	>	final int localMode;                     // per-worker scheduling mode
626	>	int nextPoolIndex;                       // hint used in registerWorker
627	>	volatile int runState;                   // shutdown status, seq, and mask
628	>	WorkQueue[] workQueues;                  // main registry
629	>	final ReentrantLock lock;                // for registration
630	>	final Condition termination;             // for awaitTermination
631	>	final ForkJoinWorkerThreadFactory factory; // factory for new workers
632	>	final Thread.UncaughtExceptionHandler ueh; // per-worker UEH
633	>	final AtomicLong stealCount;             // collect counts when terminated
634	>	final AtomicInteger nextWorkerNumber;    // to create worker name string
635	>	final String workerNamePrefix;           // Prefix for assigning worker names
636	>
637	>	/**
638	>	* Queues supporting work-stealing as well as external task
639	>	* submission. See above for main rationale and algorithms.
640	>	* Implementation relies heavily on "Unsafe" intrinsics
641	>	* and selective use of "volatile":
642	>	*
643	>	* Field "base" is the index (mod array.length) of the least valid
644	>	* queue slot, which is always the next position to steal (poll)
645	>	* from if nonempty. Reads and writes require volatile orderings
646	>	* but not CAS, because updates are only performed after slot
647	>	* CASes.
648	>	*
649	>	* Field "top" is the index (mod array.length) of the next queue
650	>	* slot to push to or pop from. It is written only by owner thread
651	>	* for push, or under lock for trySharedPush, and accessed by
652	>	* other threads only after reading (volatile) base.  Both top and
653	>	* base are allowed to wrap around on overflow, but (top - base)
654	>	* (or more commonly -(base - top) to force volatile read of base
655	>	* before top) still estimates size.
656	>	*
657	>	* The array slots are read and written using the emulation of
658	>	* volatiles/atomics provided by Unsafe. Insertions must in
659	>	* general use putOrderedObject as a form of releasing store to
660	>	* ensure that all writes to the task object are ordered before
661	>	* its publication in the queue. (Although we can avoid one case
662	>	* of this when locked in trySharedPush.) All removals entail a
663	>	* CAS to null.  The array is always a power of two. To ensure
664	>	* safety of Unsafe array operations, all accesses perform
665	>	* explicit null checks and implicit bounds checks via
666	>	* power-of-two masking.
667	>	*
668	>	* In addition to basic queuing support, this class contains
669	>	* fields described elsewhere to control execution. It turns out
670	>	* to work better memory-layout-wise to include them in this
671	>	* class rather than a separate class.
672	>	*
673	>	* Performance on most platforms is very sensitive to placement of
674	>	* instances of both WorkQueues and their arrays -- we absolutely
675	>	* do not want multiple WorkQueue instances or multiple queue
676	>	* arrays sharing cache lines. (It would be best for queue objects
677	>	* and their arrays to share, but there is nothing available to
678	>	* help arrange that).  Unfortunately, because they are recorded
679	>	* in a common array, WorkQueue instances are often moved to be
680	>	* adjacent by garbage collectors. To reduce impact, we use field
681	>	* padding that works OK on common platforms; this effectively
682	>	* trades off slightly slower average field access for the sake of
683	>	* avoiding really bad worst-case access. (Until better JVM
684	>	* support is in place, this padding is dependent on transient
685	>	* properties of JVM field layout rules.)  We also take care in
686	>	* allocating and sizing and resizing the array. Non-shared queue
687	>	* arrays are initialized (via method growArray) by workers before
688	>	* use. Others are allocated on first use.
689		*/
690	<	volatile boolean shutdown;
690	>	static final class WorkQueue {
691	>	/**
692	>	* Capacity of work-stealing queue array upon initialization.
693	>	* Must be a power of two; at least 4, but set larger to
694	>	* reduce cacheline sharing among queues.
695	>	*/
696	>	static final int INITIAL_QUEUE_CAPACITY = 1 << 8;
697	>
698	>	/**
699	>	* Maximum size for queue arrays. Must be a power of two less
700	>	* than or equal to 1 << (31 - width of array entry) to ensure
701	>	* lack of wraparound of index calculations, but defined to a
702	>	* value a bit less than this to help users trap runaway
703	>	* programs before saturating systems.
704	>	*/
705	>	static final int MAXIMUM_QUEUE_CAPACITY = 1 << 26; // 64M
706	>
707	>	volatile long totalSteals; // cumulative number of steals
708	>	int seed;                  // for random scanning; initialize nonzero
709	>	volatile int eventCount;   // encoded inactivation count; < 0 if inactive
710	>	int nextWait;              // encoded record of next event waiter
711	>	int rescans;               // remaining scans until block
712	>	int nsteals;               // top-level task executions since last idle
713	>	final int mode;            // lifo, fifo, or shared
714	>	int poolIndex;             // index of this queue in pool (or 0)
715	>	int stealHint;             // index of most recent known stealer
716	>	volatile int runState;     // 1: locked, -1: terminate; else 0
717	>	volatile int base;         // index of next slot for poll
718	>	int top;                   // index of next slot for push
719	>	ForkJoinTask<?>[] array;   // the elements (initially unallocated)
720	>	final ForkJoinWorkerThread owner; // owning thread or null if shared
721	>	volatile Thread parker;    // == owner during call to park; else null
722	>	ForkJoinTask<?> currentJoin;  // task being joined in awaitJoin
723	>	ForkJoinTask<?> currentSteal; // current non-local task being executed
724	>	// Heuristic padding to ameliorate unfortunate memory placements
725	>	Object p00, p01, p02, p03, p04, p05, p06, p07, p08, p09, p0a;
726	>
727	>	WorkQueue(ForkJoinWorkerThread owner, int mode) {
728	>	this.owner = owner;
729	>	this.mode = mode;
730	>	// Place indices in the center of array (that is not yet allocated)
731	>	base = top = INITIAL_QUEUE_CAPACITY >>> 1;
732	>	}
733	>
734	>	/**
735	>	* Returns number of tasks in the queue
736	>	*/
737	>	final int queueSize() {
738	>	int n = base - top; // non-owner callers must read base first
739	>	return (n >= 0) ? 0 : -n;
740	>	}
741	>
742	>	/**
743	>	* Pushes a task. Call only by owner in unshared queues.
744	>	*
745	>	* @param task the task. Caller must ensure non-null.
746	>	* @param p, if non-null, pool to signal if necessary
747	>	* @throw RejectedExecutionException if array cannot
748	>	* be resized
749	>	*/
750	>	final void push(ForkJoinTask<?> task, ForkJoinPool p) {
751	>	ForkJoinTask<?>[] a;
752	>	int s = top, m, n;
753	>	if ((a = array) != null) {    // ignore if queue removed
754	>	U.putOrderedObject
755	>	(a, (((m = a.length - 1) & s) << ASHIFT) + ABASE, task);
756	>	if ((n = (top = s + 1) - base) <= 2) {
757	>	if (p != null)
758	>	p.signalWork();
759	>	}
760	>	else if (n >= m)
761	>	growArray(true);
762	>	}
763	>	}
764	>
765	>	/**
766	>	* Pushes a task if lock is free and array is either big
767	>	* enough or can be resized to be big enough.
768	>	*
769	>	* @param task the task. Caller must ensure non-null.
770	>	* @return true if submitted
771	>	*/
772	>	final boolean trySharedPush(ForkJoinTask<?> task) {
773	>	boolean submitted = false;
774	>	if (runState == 0 && U.compareAndSwapInt(this, RUNSTATE, 0, 1)) {
775	>	ForkJoinTask<?>[] a = array;
776	>	int s = top, n = s - base;
777	>	try {
778	>	if ((a != null && n < a.length - 1) ||
779	>	(a = growArray(false)) != null) { // must presize
780	>	int j = (((a.length - 1) & s) << ASHIFT) + ABASE;
781	>	U.putObject(a, (long)j, task);    // don't need "ordered"
782	>	top = s + 1;
783	>	submitted = true;
784	>	}
785	>	} finally {
786	>	runState = 0;                         // unlock
787	>	}
788	>	}
789	>	return submitted;
790	>	}
791	>
792	>	/**
793	>	* Takes next task, if one exists, in FIFO order.
794	>	*/
795	>	final ForkJoinTask<?> poll() {
796	>	ForkJoinTask<?>[] a; int b, i;
797	>	while ((b = base) - top < 0 && (a = array) != null &&
798	>	(i = (a.length - 1) & b) >= 0) {
799	>	int j = (i << ASHIFT) + ABASE;
800	>	ForkJoinTask<?> t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
801	>	if (t != null && base == b &&
802	>	U.compareAndSwapObject(a, j, t, null)) {
803	>	base = b + 1;
804	>	return t;
805	>	}
806	>	}
807	>	return null;
808	>	}
809	>
810	>	/**
811	>	* Takes next task, if one exists, in LIFO order.
812	>	* Call only by owner in unshared queues.
813	>	*/
814	>	final ForkJoinTask<?> pop() {
815	>	ForkJoinTask<?> t; int m;
816	>	ForkJoinTask<?>[] a = array;
817	>	if (a != null && (m = a.length - 1) >= 0) {
818	>	for (int s; (s = top - 1) - base >= 0;) {
819	>	int j = ((m & s) << ASHIFT) + ABASE;
820	>	if ((t = (ForkJoinTask<?>)U.getObjectVolatile(a, j)) == null)
821	>	break;
822	>	if (U.compareAndSwapObject(a, j, t, null)) {
823	>	top = s;
824	>	return t;
825	>	}
826	>	}
827	>	}
828	>	return null;
829	>	}
830	>
831	>	/**
832	>	* Takes next task, if one exists, in order specified by mode.
833	>	*/
834	>	final ForkJoinTask<?> nextLocalTask() {
835	>	return mode == 0 ? pop() : poll();
836	>	}
837	>
838	>	/**
839	>	* Returns next task, if one exists, in order specified by mode.
840	>	*/
841	>	final ForkJoinTask<?> peek() {
842	>	ForkJoinTask<?>[] a = array; int m;
843	>	if (a == null || (m = a.length - 1) < 0)
844	>	return null;
845	>	int i = mode == 0 ? top - 1 : base;
846	>	int j = ((i & m) << ASHIFT) + ABASE;
847	>	return (ForkJoinTask<?>)U.getObjectVolatile(a, j);
848	>	}
849	>
850	>	/**
851	>	* Returns task at index b if b is current base of queue.
852	>	*/
853	>	final ForkJoinTask<?> pollAt(int b) {
854	>	ForkJoinTask<?>[] a; int i;
855	>	ForkJoinTask<?> task = null;
856	>	if ((a = array) != null && (i = ((a.length - 1) & b)) >= 0) {
857	>	int j = (i << ASHIFT) + ABASE;
858	>	ForkJoinTask<?> t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
859	>	if (t != null && base == b &&
860	>	U.compareAndSwapObject(a, j, t, null)) {
861	>	base = b + 1;
862	>	task = t;
863	>	}
864	>	}
865	>	return task;
866	>	}
867	>
868	>	/**
869	>	* Pops the given task only if it is at the current top.
870	>	*/
871	>	final boolean tryUnpush(ForkJoinTask<?> t) {
872	>	ForkJoinTask<?>[] a; int s;
873	>	if ((a = array) != null && (s = top) != base &&
874	>	U.compareAndSwapObject
875	>	(a, (((a.length - 1) & --s) << ASHIFT) + ABASE, t, null)) {
876	>	top = s;
877	>	return true;
878	>	}
879	>	return false;
880	>	}
881	>
882	>	/**
883	>	* Polls the given task only if it is at the current base.
884	>	*/
885	>	final boolean pollFor(ForkJoinTask<?> task) {
886	>	ForkJoinTask<?>[] a; int b, i;
887	>	if ((b = base) - top < 0 && (a = array) != null &&
888	>	(i = (a.length - 1) & b) >= 0) {
889	>	int j = (i << ASHIFT) + ABASE;
890	>	if (U.getObjectVolatile(a, j) == task && base == b &&
891	>	U.compareAndSwapObject(a, j, task, null)) {
892	>	base = b + 1;
893	>	return true;
894	>	}
895	>	}
896	>	return false;
897	>	}
898	>
899	>	/**
900	>	* If present, removes from queue and executes the given task, or
901	>	* any other cancelled task. Returns (true) immediately on any CAS
902	>	* or consistency check failure so caller can retry.
903	>	*
904	>	* @return false if no progress can be made
905	>	*/
906	>	final boolean tryRemoveAndExec(ForkJoinTask<?> task) {
907	>	boolean removed = false, empty = true, progress = true;
908	>	ForkJoinTask<?>[] a; int m, s, b, n;
909	>	if ((a = array) != null && (m = a.length - 1) >= 0 &&
910	>	(n = (s = top) - (b = base)) > 0) {
911	>	for (ForkJoinTask<?> t;;) {           // traverse from s to b
912	>	int j = ((--s & m) << ASHIFT) + ABASE;
913	>	t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
914	>	if (t == null)                    // inconsistent length
915	>	break;
916	>	else if (t == task) {
917	>	if (s + 1 == top) {           // pop
918	>	if (!U.compareAndSwapObject(a, j, task, null))
919	>	break;
920	>	top = s;
921	>	removed = true;
922	>	}
923	>	else if (base == b)           // replace with proxy
924	>	removed = U.compareAndSwapObject(a, j, task,
925	>	new EmptyTask());
926	>	break;
927	>	}
928	>	else if (t.status >= 0)
929	>	empty = false;
930	>	else if (s + 1 == top) {          // pop and throw away
931	>	if (U.compareAndSwapObject(a, j, t, null))
932	>	top = s;
933	>	break;
934	>	}
935	>	if (--n == 0) {
936	>	if (!empty && base == b)
937	>	progress = false;
938	>	break;
939	>	}
940	>	}
941	>	}
942	>	if (removed)
943	>	task.doExec();
944	>	return progress;
945	>	}
946	>
947	>	/**
948	>	* Initializes or doubles the capacity of array. Call either
949	>	* by owner or with lock held -- it is OK for base, but not
950	>	* top, to move while resizings are in progress.
951	>	*
952	>	* @param rejectOnFailure if true, throw exception if capacity
953	>	* exceeded (relayed ultimately to user); else return null.
954	>	*/
955	>	final ForkJoinTask<?>[] growArray(boolean rejectOnFailure) {
956	>	ForkJoinTask<?>[] oldA = array;
957	>	int size = oldA != null ? oldA.length << 1 : INITIAL_QUEUE_CAPACITY;
958	>	if (size <= MAXIMUM_QUEUE_CAPACITY) {
959	>	int oldMask, t, b;
960	>	ForkJoinTask<?>[] a = array = new ForkJoinTask<?>[size];
961	>	if (oldA != null && (oldMask = oldA.length - 1) >= 0 &&
962	>	(t = top) - (b = base) > 0) {
963	>	int mask = size - 1;
964	>	do {
965	>	ForkJoinTask<?> x;
966	>	int oldj = ((b & oldMask) << ASHIFT) + ABASE;
967	>	int j    = ((b &    mask) << ASHIFT) + ABASE;
968	>	x = (ForkJoinTask<?>)U.getObjectVolatile(oldA, oldj);
969	>	if (x != null &&
970	>	U.compareAndSwapObject(oldA, oldj, x, null))
971	>	U.putObjectVolatile(a, j, x);
972	>	} while (++b != t);
973	>	}
974	>	return a;
975	>	}
976	>	else if (!rejectOnFailure)
977	>	return null;
978	>	else
979	>	throw new RejectedExecutionException("Queue capacity exceeded");
980	>	}
981	>
982	>	/**
983	>	* Removes and cancels all known tasks, ignoring any exceptions
984	>	*/
985	>	final void cancelAll() {
986	>	ForkJoinTask.cancelIgnoringExceptions(currentJoin);
987	>	ForkJoinTask.cancelIgnoringExceptions(currentSteal);
988	>	for (ForkJoinTask<?> t; (t = poll()) != null; )
989	>	ForkJoinTask.cancelIgnoringExceptions(t);
990	>	}
991	>
992	>	// Execution methods
993	>
994	>	/**
995	>	* Removes and runs tasks until empty, using local mode
996	>	* ordering.
997	>	*/
998	>	final void runLocalTasks() {
999	>	if (base - top < 0) {
1000	>	for (ForkJoinTask<?> t; (t = nextLocalTask()) != null; )
1001	>	t.doExec();
1002	>	}
1003	>	}
1004	>
1005	>	/**
1006	>	* Executes a top-level task and any local tasks remaining
1007	>	* after execution.
1008	>	*
1009	>	* @return true unless terminating
1010	>	*/
1011	>	final boolean runTask(ForkJoinTask<?> t) {
1012	>	boolean alive = true;
1013	>	if (t != null) {
1014	>	currentSteal = t;
1015	>	t.doExec();
1016	>	runLocalTasks();
1017	>	++nsteals;
1018	>	currentSteal = null;
1019	>	}
1020	>	else if (runState < 0)            // terminating
1021	>	alive = false;
1022	>	return alive;
1023	>	}
1024	>
1025	>	/**
1026	>	* Executes a non-top-level (stolen) task
1027	>	*/
1028	>	final void runSubtask(ForkJoinTask<?> t) {
1029	>	if (t != null) {
1030	>	ForkJoinTask<?> ps = currentSteal;
1031	>	currentSteal = t;
1032	>	t.doExec();
1033	>	currentSteal = ps;
1034	>	}
1035	>	}
1036	>
1037	>	/**
1038	>	* Computes next value for random probes.  Scans don't require
1039	>	* a very high quality generator, but also not a crummy one.
1040	>	* Marsaglia xor-shift is cheap and works well enough.  Note:
1041	>	* This is manually inlined in several usages in ForkJoinPool
1042	>	* to avoid writes inside busy scan loops.
1043	>	*/
1044	>	final int nextSeed() {
1045	>	int r = seed;
1046	>	r ^= r << 13;
1047	>	r ^= r >>> 17;
1048	>	r ^= r << 5;
1049	>	return seed = r;
1050	>	}
1051	>
1052	>	// Unsafe mechanics
1053	>	private static final sun.misc.Unsafe U;
1054	>	private static final long RUNSTATE;
1055	>	private static final int ABASE;
1056	>	private static final int ASHIFT;
1057	>	static {
1058	>	int s;
1059	>	try {
1060	>	U = getUnsafe();
1061	>	Class<?> k = WorkQueue.class;
1062	>	Class<?> ak = ForkJoinTask[].class;
1063	>	RUNSTATE = U.objectFieldOffset
1064	>	(k.getDeclaredField("runState"));
1065	>	ABASE = U.arrayBaseOffset(ak);
1066	>	s = U.arrayIndexScale(ak);
1067	>	} catch (Exception e) {
1068	>	throw new Error(e);
1069	>	}
1070	>	if ((s & (s-1)) != 0)
1071	>	throw new Error("data type scale not a power of two");
1072	>	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
1073	>	}
1074	>	}
1075
1076		/**
1077	<	* True if use local fifo, not default lifo, for local polling
1078	<	* Read by, and replicated by ForkJoinWorkerThreads
1077	>	* Class for artificial tasks that are used to replace the target
1078	>	* of local joins if they are removed from an interior queue slot
1079	>	* in WorkQueue.tryRemoveAndExec. We don't need the proxy to
1080	>	* actually do anything beyond having a unique identity.
1081		*/
1082	<	final boolean locallyFifo;
1082	>	static final class EmptyTask extends ForkJoinTask<Void> {
1083	>	EmptyTask() { status = ForkJoinTask.NORMAL; } // force done
1084	>	public Void getRawResult() { return null; }
1085	>	public void setRawResult(Void x) {}
1086	>	public boolean exec() { return true; }
1087	>	}
1088
1089		/**
1090	<	* The number of threads in ForkJoinWorkerThreads.helpQuiescePool.
1091	<	* When non-zero, suppresses automatic shutdown when active
1092	<	* counts become zero.
1090	>	<<<<<<< ForkJoinPool.java
1091	>	* Per-thread records for (typically non-FJ) threads that submit
1092	>	* to pools. Cureently holds only psuedo-random seed / index that
1093	>	* is used to chose submission queues in method doSubmit. In the
1094	>	* future, this may incorporate a means to implement different
1095	>	* task rejection and resubmission policies.
1096		*/
1097	<	volatile int quiescerCount;
1097	>	static final class Submitter {
1098	>	int seed; // seed for random submission queue selection
1099	>
1100	>	// Heuristic padding to ameliorate unfortunate memory placements
1101	>	int p0, p1, p2, p3, p4, p5, p6, p7, p8, p9, pa, pb, pc, pd, pe;
1102	>
1103	>	Submitter() {
1104	>	// Use identityHashCode, forced negative, for seed
1105	>	seed = System.identityHashCode(Thread.currentThread()) | (1 << 31);
1106	>	}
1107	>
1108	>	/**
1109	>	* Computes next value for random probes.  Like method
1110	>	* WorkQueue.nextSeed, this is manually inlined in several
1111	>	* usages to avoid writes inside busy loops.
1112	>	*/
1113	>	final int nextSeed() {
1114	>	int r = seed;
1115	>	r ^= r << 13;
1116	>	r ^= r >>> 17;
1117	>	return seed = r ^= r << 5;
1118	>	}
1119	>	}
1120	>
1121	>	/** ThreadLocal class for Submitters */
1122	>	static final class ThreadSubmitter extends ThreadLocal<Submitter> {
1123	>	public Submitter initialValue() { return new Submitter(); }
1124	>	}
1125
1126		/**
1127	<	* The number of threads blocked in join.
1127	>	* Per-thread submission bookeeping. Shared across all pools
1128	>	* to reduce ThreadLocal pollution and because random motion
1129	>	* to avoid contention in one pool is likely to hold for others.
1130		*/
1131	<	volatile int blockedCount;
1131	>	static final ThreadSubmitter submitters = new ThreadSubmitter();
1132
1133		/**
1134	<	* Counter for worker Thread names (unrelated to their poolIndex)
1134	>	* Top-level runloop for workers
1135		*/
1136	<	private volatile int nextWorkerNumber;
1136	>	final void runWorker(ForkJoinWorkerThread wt) {
1137	>	// Initialize queue array and seed in this thread
1138	>	WorkQueue w = wt.workQueue;
1139	>	w.growArray(false);
1140	>	// Same initial hash as Submitters
1141	>	w.seed = System.identityHashCode(Thread.currentThread()) | (1 << 31);
1142	>
1143	>	do {} while (w.runTask(scan(w)));
1144	>	}
1145	>
1146	>	// Creating, registering and deregistering workers
1147
1148		/**
1149	<	* The index for the next created worker. Accessed under scanGuard.
1149	>	* Tries to create and start a worker
1150		*/
1151	<	private int nextWorkerIndex;
1151	>	private void addWorker() {
1152	>	Throwable ex = null;
1153	>	ForkJoinWorkerThread w = null;
1154	>	try {
1155	>	if ((w = factory.newThread(this)) != null) {
1156	>	w.start();
1157	>	return;
1158	>	}
1159	>	} catch (Throwable e) {
1160	>	ex = e;
1161	>	}
1162	>	deregisterWorker(w, ex);
1163	>	}
1164
1165		/**
1166	<	* SeqLock and index masking for updates to workers array.  Locked
1167	<	* when SG_UNIT is set. Unlocking clears bit by adding
1168	<	* SG_UNIT. Staleness of read-only operations can be checked by
1169	<	* comparing scanGuard to value before the reads. The low 16 bits
582	<	* (i.e, anding with SMASK) hold (the smallest power of two
583	<	* covering all worker indices, minus one, and is used to avoid
584	<	* dealing with large numbers of null slots when the workers array
585	<	* is overallocated.
1166	>	* Callback from ForkJoinWorkerThread constructor to assign a
1167	>	* public name. This must be separate from registerWorker because
1168	>	* it is called during the "super" constructor call in
1169	>	* ForkJoinWorkerThread.
1170		*/
1171	<	volatile int scanGuard;
1171	>	final String nextWorkerName() {
1172	>	return workerNamePrefix.concat
1173	>	(Integer.toString(nextWorkerNumber.addAndGet(1)));
1174	>	}
1175
1176	<	private static final int SG_UNIT = 1 << 16;
1176	>	/**
1177	>	* Callback from ForkJoinWorkerThread constructor to establish and
1178	>	* record its WorkQueue
1179	>	*
1180	>	* @param wt the worker thread
1181	>	*/
1182	>	final void registerWorker(ForkJoinWorkerThread wt) {
1183	>	WorkQueue w = wt.workQueue;
1184	>	ReentrantLock lock = this.lock;
1185	>	lock.lock();
1186	>	try {
1187	>	int k = nextPoolIndex;
1188	>	WorkQueue[] ws = workQueues;
1189	>	if (ws != null) {                       // ignore on shutdown
1190	>	int n = ws.length;
1191	>	if (k < 0 || (k & 1) == 0 || k >= n || ws[k] != null) {
1192	>	for (k = 1; k < n && ws[k] != null; k += 2)
1193	>	;                           // workers are at odd indices
1194	>	if (k >= n)                     // resize
1195	>	workQueues = ws = Arrays.copyOf(ws, n << 1);
1196	>	}
1197	>	w.poolIndex = k;
1198	>	w.eventCount = ~(k >>> 1) & SMASK;  // Set up wait count
1199	>	ws[k] = w;                          // record worker
1200	>	nextPoolIndex = k + 2;
1201	>	int rs = runState;
1202	>	int m = rs & SMASK;                 // recalculate runState mask
1203	>	if (k > m)
1204	>	m = (m << 1) + 1;
1205	>	runState = (rs & SHUTDOWN) | ((rs + RS_SEQ) & RS_SEQ_MASK) | m;
1206	>	}
1207	>	} finally {
1208	>	lock.unlock();
1209	>	}
1210	>	}
1211
1212		/**
1213	<	* The wakeup interval (in nanoseconds) for a worker waiting for a
1214	<	* task when the pool is quiescent to instead try to shrink the
1215	<	* number of workers.  The exact value does not matter too
1216	<	* much. It must be short enough to release resources during
1217	<	* sustained periods of idleness, but not so short that threads
1218	<	* are continually re-created.
1213	>	* Final callback from terminating worker, as well as failure to
1214	>	* construct or start a worker in addWorker.  Removes record of
1215	>	* worker from array, and adjusts counts. If pool is shutting
1216	>	* down, tries to complete termination.
1217	>	*
1218	>	* @param wt the worker thread or null if addWorker failed
1219	>	* @param ex the exception causing failure, or null if none
1220		*/
1221	<	private static final long SHRINK_RATE =
1222	<	4L * 1000L * 1000L * 1000L; // 4 seconds
1221	>	final void deregisterWorker(ForkJoinWorkerThread wt, Throwable ex) {
1222	>	WorkQueue w = null;
1223	>	if (wt != null && (w = wt.workQueue) != null) {
1224	>	w.runState = -1;                // ensure runState is set
1225	>	stealCount.getAndAdd(w.totalSteals + w.nsteals);
1226	>	int idx = w.poolIndex;
1227	>	ReentrantLock lock = this.lock;
1228	>	lock.lock();
1229	>	try {                           // remove record from array
1230	>	WorkQueue[] ws = workQueues;
1231	>	if (ws != null && idx >= 0 && idx < ws.length && ws[idx] == w)
1232	>	ws[nextPoolIndex = idx] = null;
1233	>	} finally {
1234	>	lock.unlock();
1235	>	}
1236	>	}
1237	>
1238	>	long c;                             // adjust ctl counts
1239	>	do {} while (!U.compareAndSwapLong
1240	>	(this, CTL, c = ctl, (((c - AC_UNIT) & AC_MASK) |
1241	>	((c - TC_UNIT) & TC_MASK) |
1242	>	(c & ~(AC_MASK|TC_MASK)))));
1243	>
1244	>	if (!tryTerminate(false) && w != null) {
1245	>	w.cancelAll();                  // cancel remaining tasks
1246	>	if (w.array != null)            // suppress signal if never ran
1247	>	signalWork();               // wake up or create replacement
1248	>	}
1249	>
1250	>	if (ex != null)                     // rethrow
1251	>	U.throwException(ex);
1252	>	}
1253
1254		/**
1255	<	* Top-level loop for worker threads: On each step: if the
604	<	* previous step swept through all queues and found no tasks, or
605	<	* there are excess threads, then possibly blocks. Otherwise,
606	<	* scans for and, if found, executes a task. Returns when pool
607	<	* and/or worker terminate.
1255	>	* Tries to add and register a new queue at the given index.
1256		*
1257	<	* @param w the worker
1257	>	* @param idx the workQueues array index to register the queue
1258	>	* @return the queue, or null if could not add because could
1259	>	* not acquire lock or idx is unusable
1260		*/
1261	<	final void work(ForkJoinWorkerThread w) {
1262	<	boolean swept = false;                // true on empty scans
1263	<	long c;
1264	<	while (!w.terminate && (int)(c = ctl) >= 0) {
1265	<	int a;                            // active count
1266	<	if (!swept && (a = (int)(c >> AC_SHIFT)) <= 0)
1267	<	swept = scan(w, a);
1268	<	else if (tryAwaitWork(w, c))
1269	<	swept = false;
1261	>	private WorkQueue tryAddSharedQueue(int idx) {
1262	>	WorkQueue q = null;
1263	>	ReentrantLock lock = this.lock;
1264	>	if (idx >= 0 && (idx & 1) == 0 && !lock.isLocked()) {
1265	>	// create queue outside of lock but only if apparently free
1266	>	WorkQueue nq = new WorkQueue(null, SHARED_QUEUE);
1267	>	if (lock.tryLock()) {
1268	>	try {
1269	>	WorkQueue[] ws = workQueues;
1270	>	if (ws != null && idx < ws.length) {
1271	>	if ((q = ws[idx]) == null) {
1272	>	int rs;         // update runState seq
1273	>	ws[idx] = q = nq;
1274	>	runState = (((rs = runState) & SHUTDOWN) |
1275	>	((rs + RS_SEQ) & ~SHUTDOWN));
1276	>	}
1277	>	}
1278	>	} finally {
1279	>	lock.unlock();
1280	>	}
1281	>	}
1282		}
1283	+	return q;
1284		}
1285
1286	<	// Signalling
1286	>	// Maintaining ctl counts
1287	>
1288	>	/**
1289	>	* Increments active count; mainly called upon return from blocking
1290	>	*/
1291	>	final void incrementActiveCount() {
1292	>	long c;
1293	>	do {} while (!U.compareAndSwapLong(this, CTL, c = ctl, c + AC_UNIT));
1294	>	}
1295
1296		/**
1297	<	* Wakes up or creates a worker.
1297	>	* Activates or creates a worker
1298		*/
1299		final void signalWork() {
1300		/*
#	Line 639 | Line 1310 | public class ForkJoinPool extends Abstra
1310		*/
1311		long c; int e, u;
1312		while ((((e = (int)(c = ctl)) | (u = (int)(c >>> 32))) &
1313	<	(INT_SIGN|SHORT_SIGN)) == (INT_SIGN|SHORT_SIGN) && e >= 0) {
1314	<	if (e > 0) {                         // release a waiting worker
1315	<	int i; ForkJoinWorkerThread w; ForkJoinWorkerThread[] ws;
1316	<	if ((ws = workers) == null ||
1317	<	(i = ~e & SMASK) >= ws.length ||
1318	<	(w = ws[i]) == null)
1313	>	(INT_SIGN|SHORT_SIGN)) == (INT_SIGN|SHORT_SIGN)) {
1314	>	WorkQueue[] ws = workQueues; int i; WorkQueue w; Thread p;
1315	>	if (e == 0) {                    // add a new worker
1316	>	if (U.compareAndSwapLong
1317	>	(this, CTL, c, (long)(((u + UTC_UNIT) & UTC_MASK) |
1318	>	((u + UAC_UNIT) & UAC_MASK)) << 32)) {
1319	>	addWorker();
1320		break;
1321	<	long nc = (((long)(w.nextWait & E_MASK)) |
1322	<	((long)(u + UAC_UNIT) << 32));
1323	<	if (w.eventCount == e &&
1324	<	UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
1325	<	w.eventCount = (e + EC_UNIT) & E_MASK;
1326	<	if (w.parked)
1327	<	UNSAFE.unpark(w);
1321	>	}
1322	>	}
1323	>	else if (e > 0 && ws != null &&
1324	>	(i = ((~e << 1) | 1) & SMASK) < ws.length &&
1325	>	(w = ws[i]) != null &&
1326	>	w.eventCount == (e | INT_SIGN)) {
1327	>	if (U.compareAndSwapLong
1328	>	(this, CTL, c, (((long)(w.nextWait & E_MASK)) |
1329	>	((long)(u + UAC_UNIT) << 32)))) {
1330	>	w.eventCount = (e + E_SEQ) & E_MASK;
1331	>	if ((p = w.parker) != null)
1332	>	U.unpark(p);         // release a waiting worker
1333		break;
1334		}
1335		}
1336	<	else if (UNSAFE.compareAndSwapLong
660	<	(this, ctlOffset, c,
661	<	(long)(((u + UTC_UNIT) & UTC_MASK) |
662	<	((u + UAC_UNIT) & UAC_MASK)) << 32)) {
663	<	addWorker();
1336	>	else
1337		break;
665	–	}
1338		}
1339		}
1340
1341		/**
1342	<	* Variant of signalWork to help release waiters on rescans.
1343	<	* Tries once to release a waiter if active count < 0.
1342	>	* Tries to decrement active count (sometimes implicitly) and
1343	>	* possibly release or create a compensating worker in preparation
1344	>	* for blocking. Fails on contention or termination.
1345		*
1346	<	* @return false if failed due to contention, else true
1346	>	* @return true if the caller can block, else should recheck and retry
1347		*/
1348	<	private boolean tryReleaseWaiter() {
1349	<	long c; int e, i; ForkJoinWorkerThread w; ForkJoinWorkerThread[] ws;
1350	<	if ((e = (int)(c = ctl)) > 0 &&
1351	<	(int)(c >> AC_SHIFT) < 0 &&
679	<	(ws = workers) != null &&
680	<	(i = ~e & SMASK) < ws.length &&
681	<	(w = ws[i]) != null) {
682	<	long nc = ((long)(w.nextWait & E_MASK) |
683	<	((c + AC_UNIT) & (AC_MASK|TC_MASK)));
684	<	if (w.eventCount != e ||
685	<	!UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc))
686	<	return false;
687	<	w.eventCount = (e + EC_UNIT) & E_MASK;
688	<	if (w.parked)
689	<	UNSAFE.unpark(w);
690	<	}
691	<	return true;
692	<	}
693	<
694	<	// Scanning for tasks
1348	>	final boolean tryCompensate() {
1349	>	WorkQueue[] ws; WorkQueue w; Thread p;
1350	>	int pc = parallelism, e, u, ac, tc, i;
1351	>	long c = ctl;
1352
1353	<	/**
1354	<	* Scans for and, if found, executes one task. Scans start at a
1355	<	* random index of workers array, and randomly select the first
1356	<	* (2*#workers)-1 probes, and then, if all empty, resort to 2
1357	<	* circular sweeps, which is necessary to check quiescence. and
1358	<	* taking a submission only if no stealable tasks were found.  The
1359	<	* steal code inside the loop is a specialized form of
1360	<	* ForkJoinWorkerThread.deqTask, followed bookkeeping to support
1361	<	* helpJoinTask and signal propagation. The code for submission
1362	<	* queues is almost identical. On each steal, the worker completes
1363	<	* not only the task, but also all local tasks that this task may
1364	<	* have generated. On detecting staleness or contention when
1365	<	* trying to take a task, this method returns without finishing
709	<	* sweep, which allows global state rechecks before retry.
710	<	*
711	<	* @param w the worker
712	<	* @param a the number of active workers
713	<	* @return true if swept all queues without finding a task
714	<	*/
715	<	private boolean scan(ForkJoinWorkerThread w, int a) {
716	<	int g = scanGuard; // mask 0 avoids useless scans if only one active
717	<	int m = (parallelism == 1 - a && blockedCount == 0) ? 0 : g & SMASK;
718	<	ForkJoinWorkerThread[] ws = workers;
719	<	if (ws == null || ws.length <= m)         // staleness check
720	<	return false;
721	<	for (int r = w.seed, k = r, j = -(m + m); j <= m + m; ++j) {
722	<	ForkJoinTask<?> t; ForkJoinTask<?>[] q; int b, i;
723	<	ForkJoinWorkerThread v = ws[k & m];
724	<	if (v != null && (b = v.queueBase) != v.queueTop &&
725	<	(q = v.queue) != null && (i = (q.length - 1) & b) >= 0) {
726	<	long u = (i << ASHIFT) + ABASE;
727	<	if ((t = q[i]) != null && v.queueBase == b &&
728	<	UNSAFE.compareAndSwapObject(q, u, t, null)) {
729	<	int d = (v.queueBase = b + 1) - v.queueTop;
730	<	v.stealHint = w.poolIndex;
731	<	if (d != 0)
732	<	signalWork();             // propagate if nonempty
733	<	w.execTask(t);
1353	>	if ((e = (int)c) >= 0) {
1354	>	if ((ac = ((u = (int)(c >>> 32)) >> UAC_SHIFT)) <= 0 &&
1355	>	e != 0 && (ws = workQueues) != null &&
1356	>	(i = ((~e << 1) | 1) & SMASK) < ws.length &&
1357	>	(w = ws[i]) != null) {
1358	>	if (w.eventCount == (e | INT_SIGN) &&
1359	>	U.compareAndSwapLong
1360	>	(this, CTL, c, ((long)(w.nextWait & E_MASK) |
1361	>	(c & (AC_MASK|TC_MASK))))) {
1362	>	w.eventCount = (e + E_SEQ) & E_MASK;
1363	>	if ((p = w.parker) != null)
1364	>	U.unpark(p);
1365	>	return true;             // release an idle worker
1366		}
735	–	r ^= r << 13; r ^= r >>> 17; w.seed = r ^ (r << 5);
736	–	return false;                     // store next seed
1367		}
1368	<	else if (j < 0) {                     // xorshift
1369	<	r ^= r << 13; r ^= r >>> 17; k = r ^= r << 5;
1368	>	else if ((tc = (short)(u >>> UTC_SHIFT)) >= 0 && ac + pc > 1) {
1369	>	long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK);
1370	>	if (U.compareAndSwapLong(this, CTL, c, nc))
1371	>	return true;             // no compensation needed
1372	>	}
1373	>	else if (tc + pc < MAX_ID) {
1374	>	long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK);
1375	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1376	>	addWorker();
1377	>	return true;             // create replacement
1378	>	}
1379		}
741	–	else
742	–	++k;
743	–	}
744	–	if (scanGuard != g)                       // staleness check
745	–	return false;
746	–	else {                                    // try to take submission
747	–	ForkJoinTask<?> t; ForkJoinTask<?>[] q; int b, i;
748	–	if ((b = queueBase) != queueTop &&
749	–	(q = submissionQueue) != null &&
750	–	(i = (q.length - 1) & b) >= 0) {
751	–	long u = (i << ASHIFT) + ABASE;
752	–	if ((t = q[i]) != null && queueBase == b &&
753	–	UNSAFE.compareAndSwapObject(q, u, t, null)) {
754	–	queueBase = b + 1;
755	–	w.execTask(t);
756	–	}
757	–	return false;
758	–	}
759	–	return true;                         // all queues empty
1380		}
1381	+	return false;
1382		}
1383
1384	+	// Submissions
1385	+
1386		/**
1387	<	* Tries to enqueue worker w in wait queue and await change in
1388	<	* worker's eventCount.  If the pool is quiescent, possibly
1389	<	* terminates worker upon exit.  Otherwise, before blocking,
1390	<	* rescans queues to avoid missed signals.  Upon finding work,
768	<	* releases at least one worker (which may be the current
769	<	* worker). Rescans restart upon detected staleness or failure to
770	<	* release due to contention. Note the unusual conventions about
771	<	* Thread.interrupt here and elsewhere: Because interrupts are
772	<	* used solely to alert threads to check termination, which is
773	<	* checked here anyway, we clear status (using Thread.interrupted)
774	<	* before any call to park, so that park does not immediately
775	<	* return due to status being set via some other unrelated call to
776	<	* interrupt in user code.
777	<	*
778	<	* @param w the calling worker
779	<	* @param c the ctl value on entry
780	<	* @return true if waited or another thread was released upon enq
1387	>	* Unless shutting down, adds the given task to a submission queue
1388	>	* at submitter's current queue index. If no queue exists at the
1389	>	* index, one is created unless pool lock is busy.  If the queue
1390	>	* and/or lock are busy, another index is randomly chosen.
1391		*/
1392	<	private boolean tryAwaitWork(ForkJoinWorkerThread w, long c) {
1393	<	int v = w.eventCount;
1394	<	w.nextWait = (int)c;                      // w's successor record
1395	<	long nc = (long)(v & E_MASK) | ((c - AC_UNIT) & (AC_MASK|TC_MASK));
1396	<	if (ctl != c || !UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
1397	<	long d = ctl; // return true if lost to a deq, to force scan
1398	<	return (int)d != (int)c && ((d - c) & AC_MASK) >= 0L;
1399	<	}
1400	<	for (int sc = w.stealCount; sc != 0;) {   // accumulate stealCount
1401	<	long s = stealCount;
1402	<	if (UNSAFE.compareAndSwapLong(this, stealCountOffset, s, s + sc))
1403	<	sc = w.stealCount = 0;
1404	<	else if (w.eventCount != v)
1405	<	return true;                      // update next time
1406	<	}
1407	<	if (parallelism + (int)(nc >> AC_SHIFT) == 0 &&
798	<	blockedCount == 0 && quiescerCount == 0)
799	<	idleAwaitWork(w, nc, c, v);           // quiescent
800	<	for (boolean rescanned = false;;) {
801	<	if (w.eventCount != v)
802	<	return true;
803	<	if (!rescanned) {
804	<	int g = scanGuard, m = g & SMASK;
805	<	ForkJoinWorkerThread[] ws = workers;
806	<	if (ws != null && m < ws.length) {
807	<	rescanned = true;
808	<	for (int i = 0; i <= m; ++i) {
809	<	ForkJoinWorkerThread u = ws[i];
810	<	if (u != null) {
811	<	if (u.queueBase != u.queueTop &&
812	<	!tryReleaseWaiter())
813	<	rescanned = false; // contended
814	<	if (w.eventCount != v)
815	<	return true;
816	<	}
817	<	}
818	<	}
819	<	if (scanGuard != g ||              // stale
820	<	(queueBase != queueTop && !tryReleaseWaiter()))
821	<	rescanned = false;
822	<	if (!rescanned)
823	<	Thread.yield();                // reduce contention
824	<	else
825	<	Thread.interrupted();          // clear before park
826	<	}
827	<	else {
828	<	w.parked = true;                   // must recheck
829	<	if (w.eventCount != v) {
830	<	w.parked = false;
831	<	return true;
832	<	}
833	<	LockSupport.park(this);
834	<	rescanned = w.parked = false;
1392	>	private void doSubmit(ForkJoinTask<?> task) {
1393	>	if (task == null)
1394	>	throw new NullPointerException();
1395	>	Submitter s = submitters.get();
1396	>	for (int r = s.seed;;) {
1397	>	WorkQueue q; int k;
1398	>	int rs = runState, m = rs & SMASK;
1399	>	WorkQueue[] ws = workQueues;
1400	>	if (rs < 0 || ws == null)   // shutting down
1401	>	throw new RejectedExecutionException();
1402	>	if (ws.length > m &&        // k must be at index
1403	>	((q = ws[k = (r << 1) & m]) != null ||
1404	>	(q = tryAddSharedQueue(k)) != null) &&
1405	>	q.trySharedPush(task)) {
1406	>	signalWork();
1407	>	return;
1408		}
1409	+	r ^= r << 13;               // xorshift seed to new position
1410	+	r ^= r >>> 17;
1411	+	if (((s.seed = r ^= r << 5) & m) == 0)
1412	+	Thread.yield();         // occasionally yield if busy
1413		}
1414		}
1415
1416	+
1417	+	// Scanning for tasks
1418	+
1419		/**
1420	<	* If inactivating worker w has caused pool to become
1421	<	* quiescent, check for pool termination, and wait for event
1422	<	* for up to SHRINK_RATE nanosecs (rescans are unnecessary in
1423	<	* this case because quiescence reflects consensus about lack
1424	<	* of work). On timeout, if ctl has not changed, terminate the
1425	<	* worker. Upon its termination (see deregisterWorker), it may
1426	<	* wake up another worker to possibly repeat this process.
1427	<	*
1428	<	* @param w the calling worker
1429	<	* @param currentCtl the ctl value after enqueuing w
1430	<	* @param prevCtl the ctl value if w terminated
1431	<	* @param v the eventCount w awaits change
1432	<	*/
1433	<	private void idleAwaitWork(ForkJoinWorkerThread w, long currentCtl,
1434	<	long prevCtl, int v) {
1435	<	if (w.eventCount == v) {
1436	<	if (shutdown)
1437	<	tryTerminate(false);
1438	<	ForkJoinTask.helpExpungeStaleExceptions(); // help clean weak refs
1439	<	while (ctl == currentCtl) {
1440	<	long startTime = System.nanoTime();
1441	<	w.parked = true;
1442	<	if (w.eventCount == v)             // must recheck
1443	<	LockSupport.parkNanos(this, SHRINK_RATE);
1444	<	w.parked = false;
1445	<	if (w.eventCount != v)
1420	>	* Scans for and, if found, returns one task, else possibly
1421	>	* inactivates the worker. This method operates on single reads of
1422	>	* volatile state and is designed to be re-invoked continuously in
1423	>	* part because it returns upon detecting inconsistencies,
1424	>	* contention, or state changes that indicate possible success on
1425	>	* re-invocation.
1426	>	*
1427	>	* The scan searches for tasks across queues, randomly selecting
1428	>	* the first #queues probes, favoring steals 2:1 over submissions
1429	>	* (by exploiting even/odd indexing), and then performing a
1430	>	* circular sweep of all queues.  The scan terminates upon either
1431	>	* finding a non-empty queue, or completing a full sweep. If the
1432	>	* worker is not inactivated, it takes and returns a task from
1433	>	* this queue.  On failure to find a task, we take one of the
1434	>	* following actions, after which the caller will retry calling
1435	>	* this method unless terminated.
1436	>	*
1437	>	* * If not a complete sweep, try to release a waiting worker.  If
1438	>	* the scan terminated because the worker is inactivated, then the
1439	>	* released worker will often be the calling worker, and it can
1440	>	* succeed obtaining a task on the next call. Or maybe it is
1441	>	* another worker, but with same net effect. Releasing in other
1442	>	* cases as well ensures that we have enough workers running.
1443	>	*
1444	>	* * If the caller has run a task since the the last empty scan,
1445	>	* return (to allow rescan) if other workers are not also yet
1446	>	* enqueued.  Field WorkQueue.rescans counts down on each scan to
1447	>	* ensure eventual inactivation, and occasional calls to
1448	>	* Thread.yield to help avoid interference with more useful
1449	>	* activities on the system.
1450	>	*
1451	>	* * If pool is terminating, terminate the worker
1452	>	*
1453	>	* * If not already enqueued, try to inactivate and enqueue the
1454	>	* worker on wait queue.
1455	>	*
1456	>	* * If already enqueued and none of the above apply, either park
1457	>	* awaiting signal, or if this is the most recent waiter and pool
1458	>	* is quiescent, relay to idleAwaitWork to check for termination
1459	>	* and possibly shrink pool.
1460	>	*
1461	>	* @param w the worker (via its WorkQueue)
1462	>	* @return a task or null of none found
1463	>	*/
1464	>	private final ForkJoinTask<?> scan(WorkQueue w) {
1465	>	boolean swept = false;                 // true after full empty scan
1466	>	WorkQueue[] ws;                        // volatile read order matters
1467	>	int r = w.seed, ec = w.eventCount;     // ec is negative if inactive
1468	>	int rs = runState, m = rs & SMASK;
1469	>	if ((ws = workQueues) != null && ws.length > m) {
1470	>	ForkJoinTask<?> task = null;
1471	>	for (int k = 0, j = -2 - m; ; ++j) {
1472	>	WorkQueue q; int b;
1473	>	if (j < 0) {                    // random probes while j negative
1474	>	r ^= r << 13; r ^= r >>> 17; k = (r ^= r << 5) | (j & 1);
1475	>	}                               // worker (not submit) for odd j
1476	>	else                            // cyclic scan when j >= 0
1477	>	k += (m >>> 1) | 1;         // step by half to reduce bias
1478	>
1479	>	if ((q = ws[k & m]) != null && (b = q.base) - q.top < 0) {
1480	>	if (ec >= 0)
1481	>	task = q.pollAt(b);     // steal
1482		break;
1483	<	else if (System.nanoTime() - startTime < SHRINK_RATE)
1484	<	Thread.interrupted();          // spurious wakeup
1485	<	else if (UNSAFE.compareAndSwapLong(this, ctlOffset,
1486	<	currentCtl, prevCtl)) {
871	<	w.terminate = true;            // restore previous
872	<	w.eventCount = ((int)currentCtl + EC_UNIT) & E_MASK;
1483	>	}
1484	>	else if (j > m) {
1485	>	if (rs == runState)        // staleness check
1486	>	swept = true;
1487		break;
1488		}
1489		}
1490	<	}
1491	<	}
1492	<
1493	<	// Submissions
1494	<
1495	<	/**
1496	<	* Enqueues the given task in the submissionQueue.  Same idea as
1497	<	* ForkJoinWorkerThread.pushTask except for use of submissionLock.
1498	<	*
1499	<	* @param t the task
1500	<	*/
1501	<	private void addSubmission(ForkJoinTask<?> t) {
1502	<	final ReentrantLock lock = this.submissionLock;
1503	<	lock.lock();
1504	<	try {
1505	<	ForkJoinTask<?>[] q; int s, m;
1506	<	if ((q = submissionQueue) != null) {    // ignore if queue removed
1507	<	long u = (((s = queueTop) & (m = q.length-1)) << ASHIFT)+ABASE;
1508	<	UNSAFE.putOrderedObject(q, u, t);
1509	<	queueTop = s + 1;
1510	<	if (s - queueBase == m)
1511	<	growSubmissionQueue();
1490	>	w.seed = r;                        // save seed for next scan
1491	>	if (task != null)
1492	>	return task;
1493	>	}
1494	>
1495	>	// Decode ctl on empty scan
1496	>	long c = ctl; int e = (int)c, a = (int)(c >> AC_SHIFT), nr, ns;
1497	>	if (!swept) {                          // try to release a waiter
1498	>	WorkQueue v; Thread p;
1499	>	if (e > 0 && a < 0 && ws != null &&
1500	>	(v = ws[((~e << 1) | 1) & m]) != null &&
1501	>	v.eventCount == (e | INT_SIGN) && U.compareAndSwapLong
1502	>	(this, CTL, c, ((long)(v.nextWait & E_MASK) |
1503	>	((c + AC_UNIT) & (AC_MASK|TC_MASK))))) {
1504	>	v.eventCount = (e + E_SEQ) & E_MASK;
1505	>	if ((p = v.parker) != null)
1506	>	U.unpark(p);
1507	>	}
1508	>	}
1509	>	else if ((nr = w.rescans) > 0) {       // continue rescanning
1510	>	int ac = a + parallelism;
1511	>	if ((w.rescans = (ac < nr) ? ac : nr - 1) > 0 && w.seed < 0 &&
1512	>	w.eventCount == ec)
1513	>	Thread.yield();                // 1 bit randomness for yield call
1514	>	}
1515	>	else if (e < 0)                        // pool is terminating
1516	>	w.runState = -1;
1517	>	else if (ec >= 0) {                    // try to enqueue
1518	>	long nc = (long)ec | ((c - AC_UNIT) & (AC_MASK|TC_MASK));
1519	>	w.nextWait = e;
1520	>	w.eventCount = ec | INT_SIGN;      // mark as inactive
1521	>	if (!U.compareAndSwapLong(this, CTL, c, nc))
1522	>	w.eventCount = ec;             // back out on CAS failure
1523	>	else if ((ns = w.nsteals) != 0) {  // set rescans if ran task
1524	>	if (a <= 0)                    // ... unless too many active
1525	>	w.rescans = a + parallelism;
1526	>	w.nsteals = 0;
1527	>	w.totalSteals += ns;
1528	>	}
1529	>	}
1530	>	else{                                  // already queued
1531	>	if (parallelism == -a)
1532	>	idleAwaitWork(w);              // quiescent
1533	>	if (w.eventCount == ec) {
1534	>	Thread.interrupted();          // clear status
1535	>	ForkJoinWorkerThread wt = w.owner;
1536	>	U.putObject(wt, PARKBLOCKER, this);
1537	>	w.parker = wt;                 // emulate LockSupport.park
1538	>	if (w.eventCount == ec)        // recheck
1539	>	U.park(false, 0L);         // block
1540	>	w.parker = null;
1541	>	U.putObject(wt, PARKBLOCKER, null);
1542		}
899	–	} finally {
900	–	lock.unlock();
1543		}
1544	<	signalWork();
1544	>	return null;
1545		}
1546
905	–	//  (pollSubmission is defined below with exported methods)
906	–
1547		/**
1548	<	* Creates or doubles submissionQueue array.
1549	<	* Basically identical to ForkJoinWorkerThread version.
1548	>	* If inactivating worker w has caused pool to become quiescent,
1549	>	* check for pool termination, and, so long as this is not the
1550	>	* only worker, wait for event for up to SHRINK_RATE nanosecs On
1551	>	* timeout, if ctl has not changed, terminate the worker, which
1552	>	* will in turn wake up another worker to possibly repeat this
1553	>	* process.
1554	>	*
1555	>	* @param w the calling worker
1556		*/
1557	<	private void growSubmissionQueue() {
1558	<	ForkJoinTask<?>[] oldQ = submissionQueue;
1559	<	int size = oldQ != null ? oldQ.length << 1 : INITIAL_QUEUE_CAPACITY;
1560	<	if (size > MAXIMUM_QUEUE_CAPACITY)
1561	<	throw new RejectedExecutionException("Queue capacity exceeded");
1562	<	if (size < INITIAL_QUEUE_CAPACITY)
1563	<	size = INITIAL_QUEUE_CAPACITY;
1564	<	ForkJoinTask<?>[] q = submissionQueue = new ForkJoinTask<?>[size];
1565	<	int mask = size - 1;
1566	<	int top = queueTop;
1567	<	int oldMask;
1568	<	if (oldQ != null && (oldMask = oldQ.length - 1) >= 0) {
1569	<	for (int b = queueBase; b != top; ++b) {
1570	<	long u = ((b & oldMask) << ASHIFT) + ABASE;
1571	<	Object x = UNSAFE.getObjectVolatile(oldQ, u);
1572	<	if (x != null && UNSAFE.compareAndSwapObject(oldQ, u, x, null))
1573	<	UNSAFE.putObjectVolatile
1574	<	(q, ((b & mask) << ASHIFT) + ABASE, x);
1557	>	private void idleAwaitWork(WorkQueue w) {
1558	>	long c; int nw, ec;
1559	>	if (!tryTerminate(false) &&
1560	>	(int)((c = ctl) >> AC_SHIFT) + parallelism == 0 &&
1561	>	(ec = w.eventCount) == ((int)c | INT_SIGN) &&
1562	>	(nw = w.nextWait) != 0) {
1563	>	long nc = ((long)(nw & E_MASK) | // ctl to restore on timeout
1564	>	((c + AC_UNIT) & AC_MASK) | (c & TC_MASK));
1565	>	ForkJoinTask.helpExpungeStaleExceptions(); // help clean
1566	>	ForkJoinWorkerThread wt = w.owner;
1567	>	while (ctl == c) {
1568	>	long startTime = System.nanoTime();
1569	>	Thread.interrupted();  // timed variant of version in scan()
1570	>	U.putObject(wt, PARKBLOCKER, this);
1571	>	w.parker = wt;
1572	>	if (ctl == c)
1573	>	U.park(false, SHRINK_RATE);
1574	>	w.parker = null;
1575	>	U.putObject(wt, PARKBLOCKER, null);
1576	>	if (ctl != c)
1577	>	break;
1578	>	if (System.nanoTime() - startTime >= SHRINK_TIMEOUT &&
1579	>	U.compareAndSwapLong(this, CTL, c, nc)) {
1580	>	w.runState = -1;          // shrink
1581	>	w.eventCount = (ec + E_SEQ) | E_MASK;
1582	>	break;
1583	>	}
1584		}
1585		}
1586		}
1587
933	–	// Blocking support
934	–
1588		/**
1589	<	* Tries to increment blockedCount, decrement active count
1590	<	* (sometimes implicitly) and possibly release or create a
1591	<	* compensating worker in preparation for blocking. Fails
1592	<	* on contention or termination.
1593	<	*
1594	<	* @return true if the caller can block, else should recheck and retry
1595	<	*/
1596	<	private boolean tryPreBlock() {
1597	<	int b = blockedCount;
1598	<	if (UNSAFE.compareAndSwapInt(this, blockedCountOffset, b, b + 1)) {
1599	<	int pc = parallelism;
1600	<	do {
1601	<	ForkJoinWorkerThread[] ws; ForkJoinWorkerThread w;
1602	<	int e, ac, tc, rc, i;
1603	<	long c = ctl;
1604	<	int u = (int)(c >>> 32);
1605	<	if ((e = (int)c) < 0) {
1606	<	// skip -- terminating
1607	<	}
1608	<	else if ((ac = (u >> UAC_SHIFT)) <= 0 && e != 0 &&
1609	<	(ws = workers) != null &&
1610	<	(i = ~e & SMASK) < ws.length &&
1611	<	(w = ws[i]) != null) {
1612	<	long nc = ((long)(w.nextWait & E_MASK) |
1613	<	(c & (AC_MASK|TC_MASK)));
1614	<	if (w.eventCount == e &&
1615	<	UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
1616	<	w.eventCount = (e + EC_UNIT) & E_MASK;
1617	<	if (w.parked)
1618	<	UNSAFE.unpark(w);
1619	<	return true;             // release an idle worker
1589	>	* Tries to locate and execute tasks for a stealer of the given
1590	>	* task, or in turn one of its stealers, Traces currentSteal ->
1591	>	* currentJoin links looking for a thread working on a descendant
1592	>	* of the given task and with a non-empty queue to steal back and
1593	>	* execute tasks from. The first call to this method upon a
1594	>	* waiting join will often entail scanning/search, (which is OK
1595	>	* because the joiner has nothing better to do), but this method
1596	>	* leaves hints in workers to speed up subsequent calls. The
1597	>	* implementation is very branchy to cope with potential
1598	>	* inconsistencies or loops encountering chains that are stale,
1599	>	* unknown, or of length greater than MAX_HELP_DEPTH links.  All
1600	>	* of these cases are dealt with by just retrying by caller.
1601	>	*
1602	>	* @param joiner the joining worker
1603	>	* @param task the task to join
1604	>	* @return true if found or ran a task (and so is immediately retryable)
1605	>	*/
1606	>	final boolean tryHelpStealer(WorkQueue joiner, ForkJoinTask<?> task) {
1607	>	ForkJoinTask<?> subtask;    // current target
1608	>	boolean progress = false;
1609	>	int depth = 0;              // current chain depth
1610	>	int m = runState & SMASK;
1611	>	WorkQueue[] ws = workQueues;
1612	>
1613	>	if (ws != null && ws.length > m && (subtask = task).status >= 0) {
1614	>	outer:for (WorkQueue j = joiner;;) {
1615	>	// Try to find the stealer of subtask, by first using hint
1616	>	WorkQueue stealer = null;
1617	>	WorkQueue v = ws[j.stealHint & m];
1618	>	if (v != null && v.currentSteal == subtask)
1619	>	stealer = v;
1620	>	else {
1621	>	for (int i = 1; i <= m; i += 2) {
1622	>	if ((v = ws[i]) != null && v.currentSteal == subtask) {
1623	>	stealer = v;
1624	>	j.stealHint = i; // save hint
1625	>	break;
1626	>	}
1627		}
1628	+	if (stealer == null)
1629	+	break;
1630		}
1631	<	else if ((tc = (short)(u >>> UTC_SHIFT)) >= 0 && ac + pc > 1) {
1632	<	long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK);
1633	<	if (UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc))
1634	<	return true;             // no compensation needed
1635	<	}
1636	<	else if (tc + pc < MAX_ID) {
1637	<	long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK);
1638	<	if (UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
1639	<	addWorker();
1640	<	return true;            // create a replacement
1631	>
1632	>	for (WorkQueue q = stealer;;) { // Try to help stealer
1633	>	ForkJoinTask<?> t; int b;
1634	>	if (task.status < 0)
1635	>	break outer;
1636	>	if ((b = q.base) - q.top < 0) {
1637	>	progress = true;
1638	>	if (subtask.status < 0)
1639	>	break outer;               // stale
1640	>	if ((t = q.pollAt(b)) != null) {
1641	>	stealer.stealHint = joiner.poolIndex;
1642	>	joiner.runSubtask(t);
1643	>	}
1644	>	}
1645	>	else { // empty - try to descend to find stealer's stealer
1646	>	ForkJoinTask<?> next = stealer.currentJoin;
1647	>	if (++depth == MAX_HELP_DEPTH || subtask.status < 0 ||
1648	>	next == null || next == subtask)
1649	>	break outer;  // max depth, stale, dead-end, cyclic
1650	>	subtask = next;
1651	>	j = stealer;
1652	>	break;
1653		}
1654		}
1655	<	// try to back out on any failure and let caller retry
982	<	} while (!UNSAFE.compareAndSwapInt(this, blockedCountOffset,
983	<	b = blockedCount, b - 1));
1655	>	}
1656		}
1657	<	return false;
986	<	}
987	<
988	<	/**
989	<	* Decrements blockedCount and increments active count
990	<	*/
991	<	private void postBlock() {
992	<	long c;
993	<	do {} while (!UNSAFE.compareAndSwapLong(this, ctlOffset,  // no mask
994	<	c = ctl, c + AC_UNIT));
995	<	int b;
996	<	do {} while(!UNSAFE.compareAndSwapInt(this, blockedCountOffset,
997	<	b = blockedCount, b - 1));
1657	>	return progress;
1658		}
1659
1660		/**
1661	<	* Possibly blocks waiting for the given task to complete, or
1002	<	* cancels the task if terminating.  Fails to wait if contended.
1661	>	* If task is at base of some steal queue, steals and executes it.
1662		*
1663	<	* @param joinMe the task
1663	>	* @param joiner the joining worker
1664	>	* @param task the task
1665		*/
1666	<	final void tryAwaitJoin(ForkJoinTask<?> joinMe) {
1667	<	int s;
1668	<	Thread.interrupted(); // clear interrupts before checking termination
1669	<	if (joinMe.status >= 0) {
1670	<	if (tryPreBlock()) {
1671	<	joinMe.tryAwaitDone(0L);
1672	<	postBlock();
1666	>	final void tryPollForAndExec(WorkQueue joiner, ForkJoinTask<?> task) {
1667	>	WorkQueue[] ws;
1668	>	int m = runState & SMASK;
1669	>	if ((ws = workQueues) != null && ws.length > m) {
1670	>	for (int j = 1; j <= m && task.status >= 0; j += 2) {
1671	>	WorkQueue q = ws[j];
1672	>	if (q != null && q.pollFor(task)) {
1673	>	joiner.runSubtask(task);
1674	>	break;
1675	>	}
1676		}
1014	–	else if ((ctl & STOP_BIT) != 0L)
1015	–	joinMe.cancelIgnoringExceptions();
1677		}
1678		}
1679
1680		/**
1681	<	* Possibly blocks the given worker waiting for joinMe to
1682	<	* complete or timeout
1683	<	*
1684	<	* @param joinMe the task
1685	<	* @param millis the wait time for underlying Object.wait
1686	<	*/
1687	<	final void timedAwaitJoin(ForkJoinTask<?> joinMe, long nanos) {
1688	<	while (joinMe.status >= 0) {
1689	<	Thread.interrupted();
1690	<	if ((ctl & STOP_BIT) != 0L) {
1691	<	joinMe.cancelIgnoringExceptions();
1692	<	break;
1693	<	}
1694	<	if (tryPreBlock()) {
1695	<	long last = System.nanoTime();
1696	<	while (joinMe.status >= 0) {
1697	<	long millis = TimeUnit.NANOSECONDS.toMillis(nanos);
1037	<	if (millis <= 0)
1038	<	break;
1039	<	joinMe.tryAwaitDone(millis);
1040	<	if (joinMe.status < 0)
1041	<	break;
1042	<	if ((ctl & STOP_BIT) != 0L) {
1043	<	joinMe.cancelIgnoringExceptions();
1044	<	break;
1681	>	* Returns a non-empty steal queue, if one is found during a random,
1682	>	* then cyclic scan, else null.  This method must be retried by
1683	>	* caller if, by the time it tries to use the queue, it is empty.
1684	>	*/
1685	>	private WorkQueue findNonEmptyStealQueue(WorkQueue w) {
1686	>	int r = w.seed;    // Same idea as scan(), but ignoring submissions
1687	>	for (WorkQueue[] ws;;) {
1688	>	int m = runState & SMASK;
1689	>	if ((ws = workQueues) == null)
1690	>	return null;
1691	>	if (ws.length > m) {
1692	>	WorkQueue q;
1693	>	for (int n = m << 2, k = r, j = -n;;) {
1694	>	r ^= r << 13; r ^= r >>> 17; r ^= r << 5;
1695	>	if ((q = ws[(k | 1) & m]) != null && q.base - q.top < 0) {
1696	>	w.seed = r;
1697	>	return q;
1698		}
1699	<	long now = System.nanoTime();
1700	<	nanos -= now - last;
1701	<	last = now;
1699	>	else if (j > n)
1700	>	return null;
1701	>	else
1702	>	k = (j++ < 0) ? r : k + ((m >>> 1) | 1);
1703	>
1704		}
1050	–	postBlock();
1051	–	break;
1705		}
1706		}
1707		}
1708
1709		/**
1710	<	* If necessary, compensates for blocker, and blocks
1711	<	*/
1712	<	private void awaitBlocker(ManagedBlocker blocker)
1713	<	throws InterruptedException {
1714	<	while (!blocker.isReleasable()) {
1715	<	if (tryPreBlock()) {
1716	<	try {
1717	<	do {} while (!blocker.isReleasable() && !blocker.block());
1718	<	} finally {
1719	<	postBlock();
1710	>	* Runs tasks until {@code isQuiescent()}. We piggyback on
1711	>	* active count ctl maintenance, but rather than blocking
1712	>	* when tasks cannot be found, we rescan until all others cannot
1713	>	* find tasks either.
1714	>	*/
1715	>	final void helpQuiescePool(WorkQueue w) {
1716	>	for (boolean active = true;;) {
1717	>	w.runLocalTasks();      // exhaust local queue
1718	>	WorkQueue q = findNonEmptyStealQueue(w);
1719	>	if (q != null) {
1720	>	ForkJoinTask<?> t;
1721	>	if (!active) {      // re-establish active count
1722	>	long c;
1723	>	active = true;
1724	>	do {} while (!U.compareAndSwapLong
1725	>	(this, CTL, c = ctl, c + AC_UNIT));
1726	>	}
1727	>	if ((t = q.poll()) != null)
1728	>	w.runSubtask(t);
1729	>	}
1730	>	else {
1731	>	long c;
1732	>	if (active) {       // decrement active count without queuing
1733	>	active = false;
1734	>	do {} while (!U.compareAndSwapLong
1735	>	(this, CTL, c = ctl, c -= AC_UNIT));
1736	>	}
1737	>	else
1738	>	c = ctl;        // re-increment on exit
1739	>	if ((int)(c >> AC_SHIFT) + parallelism == 0) {
1740	>	do {} while (!U.compareAndSwapLong
1741	>	(this, CTL, c = ctl, c + AC_UNIT));
1742	>	break;
1743		}
1068	–	break;
1744		}
1745		}
1746		}
1747
1073	–	// Creating, registering and deregistring workers
1074	–
1748		/**
1749	<	* Tries to create and start a worker; minimally rolls back counts
1750	<	* on failure.
1749	>	* Gets and removes a local or stolen task for the given worker
1750	>	*
1751	>	* @return a task, if available
1752		*/
1753	<	private void addWorker() {
1754	<	Throwable ex = null;
1755	<	ForkJoinWorkerThread t = null;
1756	<	try {
1757	<	t = factory.newThread(this);
1758	<	} catch (Throwable e) {
1759	<	ex = e;
1760	<	}
1761	<	if (t == null) {  // null or exceptional factory return
1088	<	long c;       // adjust counts
1089	<	do {} while (!UNSAFE.compareAndSwapLong
1090	<	(this, ctlOffset, c = ctl,
1091	<	(((c - AC_UNIT) & AC_MASK) |
1092	<	((c - TC_UNIT) & TC_MASK) |
1093	<	(c & ~(AC_MASK|TC_MASK)))));
1094	<	// Propagate exception if originating from an external caller
1095	<	if (!tryTerminate(false) && ex != null &&
1096	<	!(Thread.currentThread() instanceof ForkJoinWorkerThread))
1097	<	UNSAFE.throwException(ex);
1753	>	final ForkJoinTask<?> nextTaskFor(WorkQueue w) {
1754	>	for (ForkJoinTask<?> t;;) {
1755	>	WorkQueue q;
1756	>	if ((t = w.nextLocalTask()) != null)
1757	>	return t;
1758	>	if ((q = findNonEmptyStealQueue(w)) == null)
1759	>	return null;
1760	>	if ((t = q.poll()) != null)
1761	>	return t;
1762		}
1099	–	else
1100	–	t.start();
1763		}
1764
1765		/**
1766	<	* Callback from ForkJoinWorkerThread constructor to assign a
1767	<	* public name
1766	>	* Returns the approximate (non-atomic) number of idle threads per
1767	>	* active thread to offset steal queue size for method
1768	>	* ForkJoinTask.getSurplusQueuedTaskCount().
1769		*/
1770	<	final String nextWorkerName() {
1771	<	for (int n;;) {
1772	<	if (UNSAFE.compareAndSwapInt(this, nextWorkerNumberOffset,
1773	<	n = nextWorkerNumber, ++n))
1774	<	return workerNamePrefix + n;
1775	<	}
1770	>	final int idlePerActive() {
1771	>	// Approximate at powers of two for small values, saturate past 4
1772	>	int p = parallelism;
1773	>	int a = p + (int)(ctl >> AC_SHIFT);
1774	>	return (a > (p >>>= 1) ? 0 :
1775	>	a > (p >>>= 1) ? 1 :
1776	>	a > (p >>>= 1) ? 2 :
1777	>	a > (p >>>= 1) ? 4 :
1778	>	8);
1779		}
1780
1781	<	/**
1116	<	* Callback from ForkJoinWorkerThread constructor to
1117	<	* determine its poolIndex and record in workers array.
1118	<	*
1119	<	* @param w the worker
1120	<	* @return the worker's pool index
1121	<	*/
1122	<	final int registerWorker(ForkJoinWorkerThread w) {
1123	<	/*
1124	<	* In the typical case, a new worker acquires the lock, uses
1125	<	* next available index and returns quickly.  Since we should
1126	<	* not block callers (ultimately from signalWork or
1127	<	* tryPreBlock) waiting for the lock needed to do this, we
1128	<	* instead help release other workers while waiting for the
1129	<	* lock.
1130	<	*/
1131	<	for (int g;;) {
1132	<	ForkJoinWorkerThread[] ws;
1133	<	if (((g = scanGuard) & SG_UNIT) == 0 &&
1134	<	UNSAFE.compareAndSwapInt(this, scanGuardOffset,
1135	<	g, g | SG_UNIT)) {
1136	<	int k = nextWorkerIndex;
1137	<	try {
1138	<	if ((ws = workers) != null) { // ignore on shutdown
1139	<	int n = ws.length;
1140	<	if (k < 0 || k >= n || ws[k] != null) {
1141	<	for (k = 0; k < n && ws[k] != null; ++k)
1142	<	;
1143	<	if (k == n)
1144	<	ws = workers = Arrays.copyOf(ws, n << 1);
1145	<	}
1146	<	ws[k] = w;
1147	<	nextWorkerIndex = k + 1;
1148	<	int m = g & SMASK;
1149	<	g = k > m? ((m << 1) + 1) & SMASK : g + (SG_UNIT<<1);
1150	<	}
1151	<	} finally {
1152	<	scanGuard = g;
1153	<	}
1154	<	return k;
1155	<	}
1156	<	else if ((ws = workers) != null) { // help release others
1157	<	for (ForkJoinWorkerThread u : ws) {
1158	<	if (u != null && u.queueBase != u.queueTop) {
1159	<	if (tryReleaseWaiter())
1160	<	break;
1161	<	}
1162	<	}
1163	<	}
1164	<	}
1165	<	}
1781	>	// Termination
1782
1783		/**
1784	<	* Final callback from terminating worker.  Removes record of
1169	<	* worker from array, and adjusts counts. If pool is shutting
1170	<	* down, tries to complete termination.
1171	<	*
1172	<	* @param w the worker
1784	>	* Sets SHUTDOWN bit of runState under lock
1785		*/
1786	<	final void deregisterWorker(ForkJoinWorkerThread w, Throwable ex) {
1787	<	int idx = w.poolIndex;
1788	<	int sc = w.stealCount;
1789	<	int steps = 0;
1790	<	// Remove from array, adjust worker counts and collect steal count.
1791	<	// We can intermix failed removes or adjusts with steal updates
1180	<	do {
1181	<	long s, c;
1182	<	int g;
1183	<	if (steps == 0 && ((g = scanGuard) & SG_UNIT) == 0 &&
1184	<	UNSAFE.compareAndSwapInt(this, scanGuardOffset,
1185	<	g, g |= SG_UNIT)) {
1186	<	ForkJoinWorkerThread[] ws = workers;
1187	<	if (ws != null && idx >= 0 &&
1188	<	idx < ws.length && ws[idx] == w)
1189	<	ws[idx] = null;    // verify
1190	<	nextWorkerIndex = idx;
1191	<	scanGuard = g + SG_UNIT;
1192	<	steps = 1;
1193	<	}
1194	<	if (steps == 1 &&
1195	<	UNSAFE.compareAndSwapLong(this, ctlOffset, c = ctl,
1196	<	(((c - AC_UNIT) & AC_MASK) |
1197	<	((c - TC_UNIT) & TC_MASK) |
1198	<	(c & ~(AC_MASK|TC_MASK)))))
1199	<	steps = 2;
1200	<	if (sc != 0 &&
1201	<	UNSAFE.compareAndSwapLong(this, stealCountOffset,
1202	<	s = stealCount, s + sc))
1203	<	sc = 0;
1204	<	} while (steps != 2 || sc != 0);
1205	<	if (!tryTerminate(false)) {
1206	<	if (ex != null)   // possibly replace if died abnormally
1207	<	signalWork();
1208	<	else
1209	<	tryReleaseWaiter();
1786	>	private void enableShutdown() {
1787	>	ReentrantLock lock = this.lock;
1788	>	if (runState >= 0) {
1789	>	lock.lock();                       // don't need try/finally
1790	>	runState |= SHUTDOWN;
1791	>	lock.unlock();
1792		}
1793		}
1794
1213	–	// Shutdown and termination
1214	–
1795		/**
1796	<	* Possibly initiates and/or completes termination.
1796	>	* Possibly initiates and/or completes termination.  Upon
1797	>	* termination, cancels all queued tasks and then
1798		*
1799		* @param now if true, unconditionally terminate, else only
1800	<	* if shutdown and empty queue and no active workers
1800	>	* if no work and no active workers
1801		* @return true if now terminating or terminated
1802		*/
1803		private boolean tryTerminate(boolean now) {
1804	<	long c;
1805	<	while (((c = ctl) & STOP_BIT) == 0) {
1804	>	for (long c;;) {
1805	>	if (((c = ctl) & STOP_BIT) != 0) {      // already terminating
1806	>	if ((short)(c >>> TC_SHIFT) == -parallelism) {
1807	>	ReentrantLock lock = this.lock; // signal when no workers
1808	>	lock.lock();                    // don't need try/finally
1809	>	termination.signalAll();        // signal when 0 workers
1810	>	lock.unlock();
1811	>	}
1812	>	return true;
1813	>	}
1814		if (!now) {
1815	<	if ((int)(c >> AC_SHIFT) != -parallelism)
1815	>	if ((int)(c >> AC_SHIFT) != -parallelism || runState >= 0 ||
1816	>	hasQueuedSubmissions())
1817		return false;
1818	<	if (!shutdown || blockedCount != 0 || quiescerCount != 0 ||
1819	<	queueBase != queueTop) {
1820	<	if (ctl == c) // staleness check
1821	<	return false;
1822	<	continue;
1818	>	// Check for unqueued inactive workers. One pass suffices.
1819	>	WorkQueue[] ws = workQueues; WorkQueue w;
1820	>	if (ws != null) {
1821	>	int n = ws.length;
1822	>	for (int i = 1; i < n; i += 2) {
1823	>	if ((w = ws[i]) != null && w.eventCount >= 0)
1824	>	return false;
1825	>	}
1826		}
1827		}
1828	<	if (UNSAFE.compareAndSwapLong(this, ctlOffset, c, c | STOP_BIT))
1828	>	if (U.compareAndSwapLong(this, CTL, c, c | STOP_BIT))
1829		startTerminating();
1830		}
1238	–	if ((short)(c >>> TC_SHIFT) == -parallelism) { // signal when 0 workers
1239	–	final ReentrantLock lock = this.submissionLock;
1240	–	lock.lock();
1241	–	try {
1242	–	termination.signalAll();
1243	–	} finally {
1244	–	lock.unlock();
1245	–	}
1246	–	}
1247	–	return true;
1831		}
1832
1833		/**
1834	<	* Runs up to three passes through workers: (0) Setting
1835	<	* termination status for each worker, followed by wakeups up to
1836	<	* queued workers; (1) helping cancel tasks; (2) interrupting
1837	<	* lagging threads (likely in external tasks, but possibly also
1838	<	* blocked in joins).  Each pass repeats previous steps because of
1839	<	* potential lagging thread creation.
1834	>	* Initiates termination: Runs three passes through workQueues:
1835	>	* (0) Setting termination status, followed by wakeups of queued
1836	>	* workers; (1) cancelling all tasks; (2) interrupting lagging
1837	>	* threads (likely in external tasks, but possibly also blocked in
1838	>	* joins).  Each pass repeats previous steps because of potential
1839	>	* lagging thread creation.
1840		*/
1841		private void startTerminating() {
1259	–	cancelSubmissions();
1842		for (int pass = 0; pass < 3; ++pass) {
1843	<	ForkJoinWorkerThread[] ws = workers;
1843	>	WorkQueue[] ws = workQueues;
1844		if (ws != null) {
1845	<	for (ForkJoinWorkerThread w : ws) {
1846	<	if (w != null) {
1847	<	w.terminate = true;
1845	>	WorkQueue w; Thread wt;
1846	>	int n = ws.length;
1847	>	for (int j = 0; j < n; ++j) {
1848	>	if ((w = ws[j]) != null) {
1849	>	w.runState = -1;
1850		if (pass > 0) {
1851	<	w.cancelTasks();
1852	<	if (pass > 1 && !w.isInterrupted()) {
1851	>	w.cancelAll();
1852	>	if (pass > 1 && (wt = w.owner) != null &&
1853	>	!wt.isInterrupted()) {
1854		try {
1855	<	w.interrupt();
1855	>	wt.interrupt();
1856		} catch (SecurityException ignore) {
1857		}
1858		}
1859		}
1860		}
1861		}
1862	<	terminateWaiters();
1863	<	}
1864	<	}
1865	<	}
1866	<
1867	<	/**
1868	<	* Polls and cancels all submissions. Called only during termination.
1869	<	*/
1870	<	private void cancelSubmissions() {
1871	<	while (queueBase != queueTop) {
1872	<	ForkJoinTask<?> task = pollSubmission();
1873	<	if (task != null) {
1874	<	try {
1290	<	task.cancel(false);
1291	<	} catch (Throwable ignore) {
1292	<	}
1293	<	}
1294	<	}
1295	<	}
1296	<
1297	<	/**
1298	<	* Tries to set the termination status of waiting workers, and
1299	<	* then wakes them up (after which they will terminate).
1300	<	*/
1301	<	private void terminateWaiters() {
1302	<	ForkJoinWorkerThread[] ws = workers;
1303	<	if (ws != null) {
1304	<	ForkJoinWorkerThread w; long c; int i, e;
1305	<	int n = ws.length;
1306	<	while ((i = ~(e = (int)(c = ctl)) & SMASK) < n &&
1307	<	(w = ws[i]) != null && w.eventCount == (e & E_MASK)) {
1308	<	if (UNSAFE.compareAndSwapLong(this, ctlOffset, c,
1309	<	(long)(w.nextWait & E_MASK) |
1310	<	((c + AC_UNIT) & AC_MASK) |
1311	<	(c & (TC_MASK|STOP_BIT)))) {
1312	<	w.terminate = true;
1313	<	w.eventCount = e + EC_UNIT;
1314	<	if (w.parked)
1315	<	UNSAFE.unpark(w);
1862	>	// Wake up workers parked on event queue
1863	>	int i, e; long c; Thread p;
1864	>	while ((i = ((~(e = (int)(c = ctl)) << 1) | 1) & SMASK) < n &&
1865	>	(w = ws[i]) != null &&
1866	>	w.eventCount == (e | INT_SIGN)) {
1867	>	long nc = ((long)(w.nextWait & E_MASK) |
1868	>	((c + AC_UNIT) & AC_MASK) |
1869	>	(c & (TC_MASK|STOP_BIT)));
1870	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1871	>	w.eventCount = (e + E_SEQ) & E_MASK;
1872	>	if ((p = w.parker) != null)
1873	>	U.unpark(p);
1874	>	}
1875		}
1876		}
1877		}
1878		}
1879
1321	–	// misc ForkJoinWorkerThread support
1322	–
1323	–	/**
1324	–	* Increment or decrement quiescerCount. Needed only to prevent
1325	–	* triggering shutdown if a worker is transiently inactive while
1326	–	* checking quiescence.
1327	–	*
1328	–	* @param delta 1 for increment, -1 for decrement
1329	–	*/
1330	–	final void addQuiescerCount(int delta) {
1331	–	int c;
1332	–	do {} while(!UNSAFE.compareAndSwapInt(this, quiescerCountOffset,
1333	–	c = quiescerCount, c + delta));
1334	–	}
1335	–
1336	–	/**
1337	–	* Directly increment or decrement active count without
1338	–	* queuing. This method is used to transiently assert inactivation
1339	–	* while checking quiescence.
1340	–	*
1341	–	* @param delta 1 for increment, -1 for decrement
1342	–	*/
1343	–	final void addActiveCount(int delta) {
1344	–	long d = delta < 0 ? -AC_UNIT : AC_UNIT;
1345	–	long c;
1346	–	do {} while (!UNSAFE.compareAndSwapLong(this, ctlOffset, c = ctl,
1347	–	((c + d) & AC_MASK) |
1348	–	(c & ~AC_MASK)));
1349	–	}
1350	–
1351	–	/**
1352	–	* Returns the approximate (non-atomic) number of idle threads per
1353	–	* active thread.
1354	–	*/
1355	–	final int idlePerActive() {
1356	–	// Approximate at powers of two for small values, saturate past 4
1357	–	int p = parallelism;
1358	–	int a = p + (int)(ctl >> AC_SHIFT);
1359	–	return (a > (p >>>= 1) ? 0 :
1360	–	a > (p >>>= 1) ? 1 :
1361	–	a > (p >>>= 1) ? 2 :
1362	–	a > (p >>>= 1) ? 4 :
1363	–	8);
1364	–	}
1365	–
1880		// Exported methods
1881
1882		// Constructors
#	Line 1437 | Line 1951 | public class ForkJoinPool extends Abstra
1951		this.parallelism = parallelism;
1952		this.factory = factory;
1953		this.ueh = handler;
1954	<	this.locallyFifo = asyncMode;
1954	>	this.localMode = asyncMode ? FIFO_QUEUE : LIFO_QUEUE;
1955	>	this.nextPoolIndex = 1;
1956		long np = (long)(-parallelism); // offset ctl counts
1957		this.ctl = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK);
1958	<	this.submissionQueue = new ForkJoinTask<?>[INITIAL_QUEUE_CAPACITY];
1444	<	// initialize workers array with room for 2*parallelism if possible
1958	>	// initialize workQueues array with room for 2*parallelism if possible
1959		int n = parallelism << 1;
1960		if (n >= MAX_ID)
1961		n = MAX_ID;
1962		else { // See Hackers Delight, sec 3.2, where n < (1 << 16)
1963		n |= n >>> 1; n |= n >>> 2; n |= n >>> 4; n |= n >>> 8;
1964		}
1965	<	workers = new ForkJoinWorkerThread[n + 1];
1966	<	this.submissionLock = new ReentrantLock();
1967	<	this.termination = submissionLock.newCondition();
1965	>	this.workQueues = new WorkQueue[(n + 1) << 1];
1966	>	ReentrantLock lck = this.lock = new ReentrantLock();
1967	>	this.termination = lck.newCondition();
1968	>	this.stealCount = new AtomicLong();
1969	>	this.nextWorkerNumber = new AtomicInteger();
1970		StringBuilder sb = new StringBuilder("ForkJoinPool-");
1971		sb.append(poolNumberGenerator.incrementAndGet());
1972		sb.append("-worker-");
1973		this.workerNamePrefix = sb.toString();
1974	+	// Create initial submission queue
1975	+	WorkQueue sq = tryAddSharedQueue(0);
1976	+	if (sq != null)
1977	+	sq.growArray(false);
1978		}
1979
1980		// Execution methods
#	Line 1476 | Line 1996 | public class ForkJoinPool extends Abstra
1996		*         scheduled for execution
1997		*/
1998		public <T> T invoke(ForkJoinTask<T> task) {
1999	<	Thread t = Thread.currentThread();
2000	<	if (task == null)
1481	<	throw new NullPointerException();
1482	<	if (shutdown)
1483	<	throw new RejectedExecutionException();
1484	<	if ((t instanceof ForkJoinWorkerThread) &&
1485	<	((ForkJoinWorkerThread)t).pool == this)
1486	<	return task.invoke();  // bypass submit if in same pool
1487	<	else {
1488	<	addSubmission(task);
1489	<	return task.join();
1490	<	}
1491	<	}
1492	<
1493	<	/**
1494	<	* Unless terminating, forks task if within an ongoing FJ
1495	<	* computation in the current pool, else submits as external task.
1496	<	*/
1497	<	private <T> void forkOrSubmit(ForkJoinTask<T> task) {
1498	<	ForkJoinWorkerThread w;
1499	<	Thread t = Thread.currentThread();
1500	<	if (shutdown)
1501	<	throw new RejectedExecutionException();
1502	<	if ((t instanceof ForkJoinWorkerThread) &&
1503	<	(w = (ForkJoinWorkerThread)t).pool == this)
1504	<	w.pushTask(task);
1505	<	else
1506	<	addSubmission(task);
1999	>	doSubmit(task);
2000	>	return task.join();
2001		}
2002
2003		/**
#	Line 1515 | Line 2009 | public class ForkJoinPool extends Abstra
2009		*         scheduled for execution
2010		*/
2011		public void execute(ForkJoinTask<?> task) {
2012	<	if (task == null)
1519	<	throw new NullPointerException();
1520	<	forkOrSubmit(task);
2012	>	doSubmit(task);
2013		}
2014
2015		// AbstractExecutorService methods
#	Line 1535 | Line 2027 | public class ForkJoinPool extends Abstra
2027		job = (ForkJoinTask<?>) task;
2028		else
2029		job = ForkJoinTask.adapt(task, null);
2030	<	forkOrSubmit(job);
2030	>	doSubmit(job);
2031		}
2032
2033		/**
#	Line 1548 | Line 2040 | public class ForkJoinPool extends Abstra
2040		*         scheduled for execution
2041		*/
2042		public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) {
2043	<	if (task == null)
1552	<	throw new NullPointerException();
1553	<	forkOrSubmit(task);
2043	>	doSubmit(task);
2044		return task;
2045		}
2046
#	Line 1563 | Line 2053 | public class ForkJoinPool extends Abstra
2053		if (task == null)
2054		throw new NullPointerException();
2055		ForkJoinTask<T> job = ForkJoinTask.adapt(task);
2056	<	forkOrSubmit(job);
2056	>	doSubmit(job);
2057		return job;
2058		}
2059
#	Line 1576 | Line 2066 | public class ForkJoinPool extends Abstra
2066		if (task == null)
2067		throw new NullPointerException();
2068		ForkJoinTask<T> job = ForkJoinTask.adapt(task, result);
2069	<	forkOrSubmit(job);
2069	>	doSubmit(job);
2070		return job;
2071		}
2072
#	Line 1593 | Line 2083 | public class ForkJoinPool extends Abstra
2083		job = (ForkJoinTask<?>) task;
2084		else
2085		job = ForkJoinTask.adapt(task, null);
2086	<	forkOrSubmit(job);
2086	>	doSubmit(job);
2087		return job;
2088		}
2089
#	Line 1670 | Line 2160 | public class ForkJoinPool extends Abstra
2160		* @return {@code true} if this pool uses async mode
2161		*/
2162		public boolean getAsyncMode() {
2163	<	return locallyFifo;
2163	>	return localMode != 0;
2164		}
2165
2166		/**
#	Line 1682 | Line 2172 | public class ForkJoinPool extends Abstra
2172		* @return the number of worker threads
2173		*/
2174		public int getRunningThreadCount() {
2175	<	int r = parallelism + (int)(ctl >> AC_SHIFT);
2176	<	return r <= 0? 0 : r; // suppress momentarily negative values
2175	>	int rc = 0;
2176	>	WorkQueue[] ws; WorkQueue w;
2177	>	if ((ws = workQueues) != null) {
2178	>	int n = ws.length;
2179	>	for (int i = 1; i < n; i += 2) {
2180	>	Thread.State s; ForkJoinWorkerThread wt;
2181	>	if ((w = ws[i]) != null && (wt = w.owner) != null &&
2182	>	w.eventCount >= 0 &&
2183	>	(s = wt.getState()) != Thread.State.BLOCKED &&
2184	>	s != Thread.State.WAITING &&
2185	>	s != Thread.State.TIMED_WAITING)
2186	>	++rc;
2187	>	}
2188	>	}
2189	>	return rc;
2190		}
2191
2192		/**
#	Line 1694 | Line 2197 | public class ForkJoinPool extends Abstra
2197		* @return the number of active threads
2198		*/
2199		public int getActiveThreadCount() {
2200	<	int r = parallelism + (int)(ctl >> AC_SHIFT) + blockedCount;
2201	<	return r <= 0? 0 : r; // suppress momentarily negative values
2200	>	int r = parallelism + (int)(ctl >> AC_SHIFT);
2201	>	return (r <= 0) ? 0 : r; // suppress momentarily negative values
2202		}
2203
2204		/**
#	Line 1710 | Line 2213 | public class ForkJoinPool extends Abstra
2213		* @return {@code true} if all threads are currently idle
2214		*/
2215		public boolean isQuiescent() {
2216	<	return parallelism + (int)(ctl >> AC_SHIFT) + blockedCount == 0;
2216	>	return (int)(ctl >> AC_SHIFT) + parallelism == 0;
2217		}
2218
2219		/**
#	Line 1725 | Line 2228 | public class ForkJoinPool extends Abstra
2228		* @return the number of steals
2229		*/
2230		public long getStealCount() {
2231	<	return stealCount;
2231	>	long count = stealCount.get();
2232	>	WorkQueue[] ws; WorkQueue w;
2233	>	if ((ws = workQueues) != null) {
2234	>	int n = ws.length;
2235	>	for (int i = 1; i < n; i += 2) {
2236	>	if ((w = ws[i]) != null)
2237	>	count += w.totalSteals;
2238	>	}
2239	>	}
2240	>	return count;
2241		}
2242
2243		/**
#	Line 1740 | Line 2252 | public class ForkJoinPool extends Abstra
2252		*/
2253		public long getQueuedTaskCount() {
2254		long count = 0;
2255	<	ForkJoinWorkerThread[] ws;
2256	<	if ((short)(ctl >>> TC_SHIFT) > -parallelism &&
2257	<	(ws = workers) != null) {
2258	<	for (ForkJoinWorkerThread w : ws)
2259	<	if (w != null)
2260	<	count -= w.queueBase - w.queueTop; // must read base first
2255	>	WorkQueue[] ws; WorkQueue w;
2256	>	if ((ws = workQueues) != null) {
2257	>	int n = ws.length;
2258	>	for (int i = 1; i < n; i += 2) {
2259	>	if ((w = ws[i]) != null)
2260	>	count += w.queueSize();
2261	>	}
2262		}
2263		return count;
2264		}
#	Line 1758 | Line 2271 | public class ForkJoinPool extends Abstra
2271		* @return the number of queued submissions
2272		*/
2273		public int getQueuedSubmissionCount() {
2274	<	return -queueBase + queueTop;
2274	>	int count = 0;
2275	>	WorkQueue[] ws; WorkQueue w;
2276	>	if ((ws = workQueues) != null) {
2277	>	int n = ws.length;
2278	>	for (int i = 0; i < n; i += 2) {
2279	>	if ((w = ws[i]) != null)
2280	>	count += w.queueSize();
2281	>	}
2282	>	}
2283	>	return count;
2284		}
2285
2286		/**
#	Line 1768 | Line 2290 | public class ForkJoinPool extends Abstra
2290		* @return {@code true} if there are any queued submissions
2291		*/
2292		public boolean hasQueuedSubmissions() {
2293	<	return queueBase != queueTop;
2293	>	WorkQueue[] ws; WorkQueue w;
2294	>	if ((ws = workQueues) != null) {
2295	>	int n = ws.length;
2296	>	for (int i = 0; i < n; i += 2) {
2297	>	if ((w = ws[i]) != null && w.queueSize() != 0)
2298	>	return true;
2299	>	}
2300	>	}
2301	>	return false;
2302		}
2303
2304		/**
#	Line 1779 | Line 2309 | public class ForkJoinPool extends Abstra
2309		* @return the next submission, or {@code null} if none
2310		*/
2311		protected ForkJoinTask<?> pollSubmission() {
2312	<	ForkJoinTask<?> t; ForkJoinTask<?>[] q; int b, i;
2313	<	while ((b = queueBase) != queueTop &&
2314	<	(q = submissionQueue) != null &&
2315	<	(i = (q.length - 1) & b) >= 0) {
2316	<	long u = (i << ASHIFT) + ABASE;
2317	<	if ((t = q[i]) != null &&
1788	<	queueBase == b &&
1789	<	UNSAFE.compareAndSwapObject(q, u, t, null)) {
1790	<	queueBase = b + 1;
1791	<	return t;
2312	>	WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2313	>	if ((ws = workQueues) != null) {
2314	>	int n = ws.length;
2315	>	for (int i = 0; i < n; i += 2) {
2316	>	if ((w = ws[i]) != null && (t = w.poll()) != null)
2317	>	return t;
2318		}
2319		}
2320		return null;
#	Line 1813 | Line 2339 | public class ForkJoinPool extends Abstra
2339		*/
2340		protected int drainTasksTo(Collection<? super ForkJoinTask<?>> c) {
2341		int count = 0;
2342	<	while (queueBase != queueTop) {
2343	<	ForkJoinTask<?> t = pollSubmission();
2344	<	if (t != null) {
2345	<	c.add(t);
2346	<	++count;
2342	>	WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2343	>	if ((ws = workQueues) != null) {
2344	>	int n = ws.length;
2345	>	for (int i = 0; i < n; ++i) {
2346	>	if ((w = ws[i]) != null) {
2347	>	while ((t = w.poll()) != null) {
2348	>	c.add(t);
2349	>	++count;
2350	>	}
2351	>	}
2352		}
2353		}
1823	–	ForkJoinWorkerThread[] ws;
1824	–	if ((short)(ctl >>> TC_SHIFT) > -parallelism &&
1825	–	(ws = workers) != null) {
1826	–	for (ForkJoinWorkerThread w : ws)
1827	–	if (w != null)
1828	–	count += w.drainTasksTo(c);
1829	–	}
2354		return count;
2355		}
2356
#	Line 1841 | Line 2365 | public class ForkJoinPool extends Abstra
2365		long st = getStealCount();
2366		long qt = getQueuedTaskCount();
2367		long qs = getQueuedSubmissionCount();
2368	+	int rc = getRunningThreadCount();
2369		int pc = parallelism;
2370		long c = ctl;
2371		int tc = pc + (short)(c >>> TC_SHIFT);
2372	<	int rc = pc + (int)(c >> AC_SHIFT);
2373	<	if (rc < 0) // ignore transient negative
2374	<	rc = 0;
1850	<	int ac = rc + blockedCount;
2372	>	int ac = pc + (int)(c >> AC_SHIFT);
2373	>	if (ac < 0) // ignore transient negative
2374	>	ac = 0;
2375		String level;
2376		if ((c & STOP_BIT) != 0)
2377	<	level = (tc == 0)? "Terminated" : "Terminating";
2377	>	level = (tc == 0) ? "Terminated" : "Terminating";
2378		else
2379	<	level = shutdown? "Shutting down" : "Running";
2379	>	level = runState < 0 ? "Shutting down" : "Running";
2380		return super.toString() +
2381		"[" + level +
2382		", parallelism = " + pc +
#	Line 1879 | Line 2403 | public class ForkJoinPool extends Abstra
2403		*/
2404		public void shutdown() {
2405		checkPermission();
2406	<	shutdown = true;
2406	>	enableShutdown();
2407		tryTerminate(false);
2408		}
2409
#	Line 1901 | Line 2425 | public class ForkJoinPool extends Abstra
2425		*/
2426		public List<Runnable> shutdownNow() {
2427		checkPermission();
2428	<	shutdown = true;
2428	>	enableShutdown();
2429		tryTerminate(true);
2430		return Collections.emptyList();
2431		}
#	Line 1937 | Line 2461 | public class ForkJoinPool extends Abstra
2461		}
2462
2463		/**
1940	–	* Returns true if terminating or terminated. Used by ForkJoinWorkerThread.
1941	–	*/
1942	–	final boolean isAtLeastTerminating() {
1943	–	return (ctl & STOP_BIT) != 0L;
1944	–	}
1945	–
1946	–	/**
2464		* Returns {@code true} if this pool has been shut down.
2465		*
2466		* @return {@code true} if this pool has been shut down
2467		*/
2468		public boolean isShutdown() {
2469	<	return shutdown;
2469	>	return runState < 0;
2470		}
2471
2472		/**
#	Line 1966 | Line 2483 | public class ForkJoinPool extends Abstra
2483		public boolean awaitTermination(long timeout, TimeUnit unit)
2484		throws InterruptedException {
2485		long nanos = unit.toNanos(timeout);
2486	<	final ReentrantLock lock = this.submissionLock;
2486	>	final ReentrantLock lock = this.lock;
2487		lock.lock();
2488		try {
2489		for (;;) {
#	Line 2077 | Line 2594 | public class ForkJoinPool extends Abstra
2594		public static void managedBlock(ManagedBlocker blocker)
2595		throws InterruptedException {
2596		Thread t = Thread.currentThread();
2597	<	if (t instanceof ForkJoinWorkerThread) {
2598	<	ForkJoinWorkerThread w = (ForkJoinWorkerThread) t;
2599	<	w.pool.awaitBlocker(blocker);
2600	<	}
2601	<	else {
2602	<	do {} while (!blocker.isReleasable() && !blocker.block());
2597	>	ForkJoinPool p = ((t instanceof ForkJoinWorkerThread) ?
2598	>	((ForkJoinWorkerThread)t).pool : null);
2599	>	while (!blocker.isReleasable()) {
2600	>	if (p == null || p.tryCompensate()) {
2601	>	try {
2602	>	do {} while (!blocker.isReleasable() && !blocker.block());
2603	>	} finally {
2604	>	if (p != null)
2605	>	p.incrementActiveCount();
2606	>	}
2607	>	break;
2608	>	}
2609		}
2610		}
2611
#	Line 2099 | Line 2622 | public class ForkJoinPool extends Abstra
2622		}
2623
2624		// Unsafe mechanics
2625	<	private static final sun.misc.Unsafe UNSAFE;
2626	<	private static final long ctlOffset;
2627	<	private static final long stealCountOffset;
2628	<	private static final long blockedCountOffset;
2106	<	private static final long quiescerCountOffset;
2107	<	private static final long scanGuardOffset;
2108	<	private static final long nextWorkerNumberOffset;
2109	<	private static final long ABASE;
2110	<	private static final int ASHIFT;
2625	>	private static final sun.misc.Unsafe U;
2626	>	private static final long CTL;
2627	>	private static final long RUNSTATE;
2628	>	private static final long PARKBLOCKER;
2629
2630		static {
2631		poolNumberGenerator = new AtomicInteger();
2114	–	workerSeedGenerator = new Random();
2632		modifyThreadPermission = new RuntimePermission("modifyThread");
2633		defaultForkJoinWorkerThreadFactory =
2634		new DefaultForkJoinWorkerThreadFactory();
2635		int s;
2636		try {
2637	<	UNSAFE = getUnsafe();
2638	<	Class k = ForkJoinPool.class;
2639	<	ctlOffset = UNSAFE.objectFieldOffset
2637	>	U = getUnsafe();
2638	>	Class<?> k = ForkJoinPool.class;
2639	>	Class<?> tk = Thread.class;
2640	>	CTL = U.objectFieldOffset
2641		(k.getDeclaredField("ctl"));
2642	<	stealCountOffset = UNSAFE.objectFieldOffset
2643	<	(k.getDeclaredField("stealCount"));
2644	<	blockedCountOffset = UNSAFE.objectFieldOffset
2645	<	(k.getDeclaredField("blockedCount"));
2128	<	quiescerCountOffset = UNSAFE.objectFieldOffset
2129	<	(k.getDeclaredField("quiescerCount"));
2130	<	scanGuardOffset = UNSAFE.objectFieldOffset
2131	<	(k.getDeclaredField("scanGuard"));
2132	<	nextWorkerNumberOffset = UNSAFE.objectFieldOffset
2133	<	(k.getDeclaredField("nextWorkerNumber"));
2134	<	Class a = ForkJoinTask[].class;
2135	<	ABASE = UNSAFE.arrayBaseOffset(a);
2136	<	s = UNSAFE.arrayIndexScale(a);
2642	>	RUNSTATE = U.objectFieldOffset
2643	>	(k.getDeclaredField("runState"));
2644	>	PARKBLOCKER = U.objectFieldOffset
2645	>	(tk.getDeclaredField("parkBlocker"));
2646		} catch (Exception e) {
2647		throw new Error(e);
2648		}
2140	–	if ((s & (s-1)) != 0)
2141	–	throw new Error("data type scale not a power of two");
2142	–	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
2649		}
2650
2651		/**
#	Line 2169 | Line 2675 | public class ForkJoinPool extends Abstra
2675		}
2676		}
2677		}
2678	+
2679		}

Diff Legend

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