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 1.109 by jsr166, Fri Jul 1 18:20:11 2011 UTC vs.
Revision 1.121 by jsr166, Tue Jan 31 01:33:21 2012 UTC

#	Line 20 | Line 20 | import java.util.concurrent.RejectedExec
20		import java.util.concurrent.RunnableFuture;
21		import java.util.concurrent.TimeUnit;
22		import java.util.concurrent.atomic.AtomicInteger;
23	<	import java.util.concurrent.locks.LockSupport;
24	<	import java.util.concurrent.locks.ReentrantLock;
23	>	import java.util.concurrent.atomic.AtomicLong;
24	>	import java.util.concurrent.locks.AbstractQueuedSynchronizer;
25		import java.util.concurrent.locks.Condition;
26
27		/**
#	Line 33 | 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 58 | Line 60 | import java.util.concurrent.locks.Condit
60		* convenient form for informal monitoring.
61		*
62		* <p> As is the case with other ExecutorServices, there are three
63	<	* main task execution methods summarized in the following
64	<	* table. These are designed to be used by clients not already engaged
65	<	* in fork/join computations in the current pool.  The main forms of
66	<	* these methods accept instances of {@code ForkJoinTask}, but
67	<	* overloaded forms also allow mixed execution of plain {@code
63	>	* main task execution methods summarized in the following table.
64	>	* These are designed to be used primarily by clients not already
65	>	* engaged in fork/join computations in the current pool.  The main
66	>	* forms of these methods accept instances of {@code ForkJoinTask},
67	>	* but overloaded forms also allow mixed execution of plain {@code
68		* Runnable}- or {@code Callable}- based activities as well.  However,
69	<	* tasks that are already executing in a pool should normally
70	<	* <em>NOT</em> use these pool execution methods, but instead use the
71	<	* within-computation forms listed in the table.
69	>	* tasks that are already executing in a pool should normally instead
70	>	* use the within-computation forms listed in the table unless using
71	>	* async event-style tasks that are not usually joined, in which case
72	>	* there is little difference among choice of methods.
73		*
74		* <table BORDER CELLPADDING=3 CELLSPACING=1>
75		*  <tr>
#	Line 125 | Line 128 | public class ForkJoinPool extends Abstra
128		/*
129		* Implementation Overview
130		*
131	<	* This class provides the central bookkeeping and control for a
132	<	* set of worker threads: Submissions from non-FJ threads enter
133	<	* into a submission queue. Workers take these tasks and typically
134	<	* split them into subtasks that may be stolen by other workers.
135	<	* Preference rules give first priority to processing tasks from
136	<	* their own queues (LIFO or FIFO, depending on mode), then to
137	<	* randomized FIFO steals of tasks in other worker queues, and
138	<	* lastly to new submissions.
131	>	* This class and its nested classes provide the main
132	>	* functionality and control for a set of worker threads:
133	>	* Submissions from non-FJ threads enter into submission queues.
134	>	* Workers take these tasks and typically split them into subtasks
135	>	* that may be stolen by other workers.  Preference rules give
136	>	* first priority to processing tasks from their own queues (LIFO
137	>	* or FIFO, depending on mode), then to randomized FIFO steals of
138	>	* tasks in other queues.
139	>	*
140	>	* WorkQueues
141	>	* ==========
142	>	*
143	>	* Most operations occur within work-stealing queues (in nested
144	>	* class WorkQueue).  These are special forms of Deques that
145	>	* support only three of the four possible end-operations -- push,
146	>	* pop, and poll (aka steal), under the further constraints that
147	>	* push and pop are called only from the owning thread (or, as
148	>	* extended here, under a lock), while poll may be called from
149	>	* other threads.  (If you are unfamiliar with them, you probably
150	>	* want to read Herlihy and Shavit's book "The Art of
151	>	* Multiprocessor programming", chapter 16 describing these in
152	>	* more detail before proceeding.)  The main work-stealing queue
153	>	* design is roughly similar to those in the papers "Dynamic
154	>	* Circular Work-Stealing Deque" by Chase and Lev, SPAA 2005
155	>	* (http://research.sun.com/scalable/pubs/index.html) and
156	>	* "Idempotent work stealing" by Michael, Saraswat, and Vechev,
157	>	* PPoPP 2009 (http://portal.acm.org/citation.cfm?id=1504186).
158	>	* The main differences ultimately stem from GC requirements that
159	>	* we null out taken slots as soon as we can, to maintain as small
160	>	* a footprint as possible even in programs generating huge
161	>	* numbers of tasks. To accomplish this, we shift the CAS
162	>	* arbitrating pop vs poll (steal) from being on the indices
163	>	* ("base" and "top") to the slots themselves.  So, both a
164	>	* successful pop and poll mainly entail a CAS of a slot from
165	>	* non-null to null.  Because we rely on CASes of references, we
166	>	* do not need tag bits on base or top.  They are simple ints as
167	>	* used in any circular array-based queue (see for example
168	>	* ArrayDeque).  Updates to the indices must still be ordered in a
169	>	* way that guarantees that top == base means the queue is empty,
170	>	* but otherwise may err on the side of possibly making the queue
171	>	* appear nonempty when a push, pop, or poll have not fully
172	>	* committed. Note that this means that the poll operation,
173	>	* considered individually, is not wait-free. One thief cannot
174	>	* successfully continue until another in-progress one (or, if
175	>	* previously empty, a push) completes.  However, in the
176	>	* aggregate, we ensure at least probabilistic non-blockingness.
177	>	* If an attempted steal fails, a thief always chooses a different
178	>	* random victim target to try next. So, in order for one thief to
179	>	* progress, it suffices for any in-progress poll or new push on
180	>	* any empty queue to complete.
181	>	*
182	>	* This approach also enables support of a user mode in which local
183	>	* task processing is in FIFO, not LIFO order, simply by using
184	>	* poll rather than pop.  This can be useful in message-passing
185	>	* frameworks in which tasks are never joined.  However neither
186	>	* mode considers affinities, loads, cache localities, etc, so
187	>	* rarely provide the best possible performance on a given
188	>	* machine, but portably provide good throughput by averaging over
189	>	* these factors.  (Further, even if we did try to use such
190	>	* information, we do not usually have a basis for exploiting it.
191	>	* For example, some sets of tasks profit from cache affinities,
192	>	* but others are harmed by cache pollution effects.)
193	>	*
194	>	* WorkQueues are also used in a similar way for tasks submitted
195	>	* to the pool. We cannot mix these tasks in the same queues used
196	>	* for work-stealing (this would contaminate lifo/fifo
197	>	* processing). Instead, we loosely associate submission queues
198	>	* with submitting threads, using a form of hashing.  The
199	>	* ThreadLocal Submitter class contains a value initially used as
200	>	* a hash code for choosing existing queues, but may be randomly
201	>	* repositioned upon contention with other submitters.  In
202	>	* essence, submitters act like workers except that they never
203	>	* take tasks, and they are multiplexed on to a finite number of
204	>	* shared work queues. However, classes are set up so that future
205	>	* extensions could allow submitters to optionally help perform
206	>	* tasks as well. Insertion of tasks in shared mode requires a
207	>	* lock (mainly to protect in the case of resizing) but we use
208	>	* only a simple spinlock (using bits in field runState), because
209	>	* submitters encountering a busy queue move on to try or create
210	>	* other queues, so never block.
211	>	*
212	>	* Management
213	>	* ==========
214		*
215		* The main throughput advantages of work-stealing stem from
216		* decentralized control -- workers mostly take tasks from
217		* themselves or each other. We cannot negate this in the
218		* implementation of other management responsibilities. The main
219		* tactic for avoiding bottlenecks is packing nearly all
220	<	* essentially atomic control state into a single 64bit volatile
221	<	* variable ("ctl"). This variable is read on the order of 10-100
222	<	* times as often as it is modified (always via CAS). (There is
223	<	* some additional control state, for example variable "shutdown"
224	<	* for which we can cope with uncoordinated updates.)  This
225	<	* streamlines synchronization and control at the expense of messy
226	<	* constructions needed to repack status bits upon updates.
227	<	* Updates tend not to contend with each other except during
228	<	* bursts while submitted tasks begin or end.  In some cases when
229	<	* they do contend, threads can instead do something else
230	<	* (usually, scan for tasks) until contention subsides.
231	<	*
232	<	* To enable packing, we restrict maximum parallelism to (1<<15)-1
233	<	* (which is far in excess of normal operating range) to allow
234	<	* ids, counts, and their negations (used for thresholding) to fit
235	<	* into 16bit fields.
236	<	*
237	<	* Recording Workers.  Workers are recorded in the "workers" array
238	<	* that is created upon pool construction and expanded if (rarely)
239	<	* necessary.  This is an array as opposed to some other data
240	<	* structure to support index-based random steals by workers.
241	<	* Updates to the array recording new workers and unrecording
242	<	* terminated ones are protected from each other by a seqLock
243	<	* (scanGuard) but the array is otherwise concurrently readable,
244	<	* and accessed directly by workers. To simplify index-based
220	>	* essentially atomic control state into two volatile variables
221	>	* that are by far most often read (not written) as status and
222	>	* consistency checks.
223	>	*
224	>	* Field "ctl" contains 64 bits holding all the information needed
225	>	* to atomically decide to add, inactivate, enqueue (on an event
226	>	* queue), dequeue, and/or re-activate workers.  To enable this
227	>	* packing, we restrict maximum parallelism to (1<<15)-1 (which is
228	>	* far in excess of normal operating range) to allow ids, counts,
229	>	* and their negations (used for thresholding) to fit into 16bit
230	>	* fields.
231	>	*
232	>	* Field "runState" contains 32 bits needed to register and
233	>	* deregister WorkQueues, as well as to enable shutdown. It is
234	>	* only modified under a lock (normally briefly held, but
235	>	* occasionally protecting allocations and resizings) but even
236	>	* when locked remains available to check consistency. An
237	>	* auxiliary field "growHints", also only modified under lock,
238	>	* contains a candidate index for the next WorkQueue and
239	>	* a mask for submission queue indices.
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 currently
256	>	* used indices.
257	>	*
258	>	* All worker thread creation is on-demand, triggered by task
259	>	* submissions, replacement of terminated workers, and/or
260	>	* compensation for blocked workers. However, all other support
261	>	* code is set up to work with other policies.  To ensure that we
262	>	* do not hold on to worker references that would prevent GC, ALL
263	>	* accesses to workQueues are via indices into the workQueues
264	>	* array (which is one source of some of the messy code
265	>	* constructions here). In essence, the workQueues array serves as
266	>	* a weak reference mechanism. Thus for example the wait queue
267	>	* field of ctl stores indices, not references.  Access to the
268	>	* workQueues in associated methods (for example signalWork) must
269	>	* both index-check and null-check the IDs. All such accesses
270	>	* ignore bad IDs by returning out early from what they are doing,
271	>	* since this can only be associated with termination, in which
272	>	* case it is OK to give up.  All uses of the workQueues array
273	>	* also check that it is non-null (even if previously
274	>	* non-null). This allows nulling during termination, which is
275	>	* currently not necessary, but remains an option for
276	>	* resource-revocation-based shutdown schemes. It also helps
277	>	* reduce JIT issuance of uncommon-trap code, which tends to
278	>	* unnecessarily complicate control flow in some methods.
279		*
280	<	* Wait Queuing. Unlike HPC work-stealing frameworks, we cannot
280	>	* Event Queuing. Unlike HPC work-stealing frameworks, we cannot
281		* let workers spin indefinitely scanning for tasks when none can
282		* be found immediately, and we cannot start/resume workers unless
283		* there appear to be tasks available.  On the other hand, we must
284		* quickly prod them into action when new tasks are submitted or
285	<	* generated.  We park/unpark workers after placing in an event
286	<	* wait queue when they cannot find work. This "queue" is actually
287	<	* a simple Treiber stack, headed by the "id" field of ctl, plus a
288	<	* 15bit counter value to both wake up waiters (by advancing their
289	<	* count) and avoid ABA effects. Successors are held in worker
290	<	* field "nextWait".  Queuing deals with several intrinsic races,
291	<	* mainly that a task-producing thread can miss seeing (and
285	>	* generated. In many usages, ramp-up time to activate workers is
286	>	* the main limiting factor in overall performance (this is
287	>	* compounded at program start-up by JIT compilation and
288	>	* allocation). So we try to streamline this as much as possible.
289	>	* We park/unpark workers after placing in an event wait queue
290	>	* when they cannot find work. This "queue" is actually a simple
291	>	* Treiber stack, headed by the "id" field of ctl, plus a 15bit
292	>	* counter value (that reflects the number of times a worker has
293	>	* been inactivated) to avoid ABA effects (we need only as many
294	>	* version numbers as worker threads). Successors are held in
295	>	* field WorkQueue.nextWait.  Queuing deals with several intrinsic
296	>	* races, mainly that a task-producing thread can miss seeing (and
297		* signalling) another thread that gave up looking for work but
298		* has not yet entered the wait queue. We solve this by requiring
299	<	* a full sweep of all workers both before (in scan()) and after
300	<	* (in tryAwaitWork()) a newly waiting worker is added to the wait
301	<	* queue. During a rescan, the worker might release some other
302	<	* queued worker rather than itself, which has the same net
303	<	* effect. Because enqueued workers may actually be rescanning
304	<	* rather than waiting, we set and clear the "parked" field of
305	<	* ForkJoinWorkerThread to reduce unnecessary calls to unpark.
306	<	* (Use of the parked field requires a secondary recheck to avoid
307	<	* missed signals.)
299	>	* a full sweep of all workers (via repeated calls to method
300	>	* scan()) both before and after a newly waiting worker is added
301	>	* to the wait queue. During a rescan, the worker might release
302	>	* some other queued worker rather than itself, which has the same
303	>	* net effect. Because enqueued workers may actually be rescanning
304	>	* rather than waiting, we set and clear the "parker" field of
305	>	* WorkQueues to reduce unnecessary calls to unpark.  (This
306	>	* requires a secondary recheck to avoid missed signals.)  Note
307	>	* the unusual conventions about Thread.interrupts surrounding
308	>	* parking and other blocking: Because interrupts are used solely
309	>	* to alert threads to check termination, which is checked anyway
310	>	* upon blocking, we clear status (using Thread.interrupted)
311	>	* before any call to park, so that park does not immediately
312	>	* return due to status being set via some other unrelated call to
313	>	* interrupt in user code.
314		*
315		* Signalling.  We create or wake up workers only when there
316		* appears to be at least one task they might be able to find and
317		* execute.  When a submission is added or another worker adds a
318	<	* task to a queue that previously had two or fewer tasks, they
318	>	* task to a queue that previously had fewer than two tasks, they
319		* signal waiting workers (or trigger creation of new ones if
320		* fewer than the given parallelism level -- see signalWork).
321	<	* These primary signals are buttressed by signals during rescans
322	<	* as well as those performed when a worker steals a task and
323	<	* notices that there are more tasks too; together these cover the
324	<	* signals needed in cases when more than two tasks are pushed
229	<	* but untaken.
321	>	* These primary signals are buttressed by signals during rescans;
322	>	* together these cover the signals needed in cases when more
323	>	* tasks are pushed but untaken, and improve performance compared
324	>	* to having one thread wake up all workers.
325		*
326		* Trimming workers. To release resources after periods of lack of
327		* use, a worker starting to wait when the pool is quiescent will
#	Line 234 | Line 329 | public class ForkJoinPool extends Abstra
329		* SHRINK_RATE nanosecs. This will slowly propagate, eventually
330		* terminating all workers after long periods of non-use.
331		*
332	<	* Submissions. External submissions are maintained in an
333	<	* array-based queue that is structured identically to
334	<	* ForkJoinWorkerThread queues except for the use of
335	<	* submissionLock in method addSubmission. Unlike the case for
336	<	* worker queues, multiple external threads can add new
337	<	* submissions, so adding requires a lock.
338	<	*
339	<	* Compensation. Beyond work-stealing support and lifecycle
340	<	* control, the main responsibility of this framework is to take
341	<	* actions when one worker is waiting to join a task stolen (or
342	<	* always held by) another.  Because we are multiplexing many
343	<	* tasks on to a pool of workers, we can't just let them block (as
344	<	* in Thread.join).  We also cannot just reassign the joiner's
345	<	* run-time stack with another and replace it later, which would
346	<	* be a form of "continuation", that even if possible is not
347	<	* necessarily a good idea since we sometimes need both an
348	<	* unblocked task and its continuation to progress. Instead we
349	<	* combine two tactics:
332	>	* Shutdown and Termination. A call to shutdownNow atomically sets
333	>	* a runState bit and then (non-atomically) sets each worker's
334	>	* runState status, cancels all unprocessed tasks, and wakes up
335	>	* all waiting workers.  Detecting whether termination should
336	>	* commence after a non-abrupt shutdown() call requires more work
337	>	* and bookkeeping. We need consensus about quiescence (i.e., that
338	>	* there is no more work). The active count provides a primary
339	>	* indication but non-abrupt shutdown still requires a rechecking
340	>	* scan for any workers that are inactive but not queued.
341	>	*
342	>	* Joining Tasks
343	>	* =============
344	>	*
345	>	* Any of several actions may be taken when one worker is waiting
346	>	* to join a task stolen (or always held) by another.  Because we
347	>	* are multiplexing many tasks on to a pool of workers, we can't
348	>	* just let them block (as in Thread.join).  We also cannot just
349	>	* reassign the joiner's run-time stack with another and replace
350	>	* it later, which would be a form of "continuation", that even if
351	>	* possible is not necessarily a good idea since we sometimes need
352	>	* both an unblocked task and its continuation to progress.
353	>	* Instead we combine two tactics:
354		*
355		*   Helping: Arranging for the joiner to execute some task that it
356	<	*      would be running if the steal had not occurred.  Method
258	<	*      ForkJoinWorkerThread.joinTask tracks joining->stealing
259	<	*      links to try to find such a task.
356	>	*      would be running if the steal had not occurred.
357		*
358		*   Compensating: Unless there are already enough live threads,
359	<	*      method tryPreBlock() may create or re-activate a spare
360	<	*      thread to compensate for blocked joiners until they
361	<	*      unblock.
359	>	*      method tryCompensate() may create or re-activate a spare
360	>	*      thread to compensate for blocked joiners until they unblock.
361	>	*
362	>	* A third form (implemented in tryRemoveAndExec and
363	>	* tryPollForAndExec) amounts to helping a hypothetical
364	>	* compensator: If we can readily tell that a possible action of a
365	>	* compensator is to steal and execute the task being joined, the
366	>	* joining thread can do so directly, without the need for a
367	>	* compensation thread (although at the expense of larger run-time
368	>	* stacks, but the tradeoff is typically worthwhile).
369		*
370		* The ManagedBlocker extension API can't use helping so relies
371		* only on compensation in method awaitBlocker.
372		*
373	+	* The algorithm in tryHelpStealer entails a form of "linear"
374	+	* helping: Each worker records (in field currentSteal) the most
375	+	* recent task it stole from some other worker. Plus, it records
376	+	* (in field currentJoin) the task it is currently actively
377	+	* joining. Method tryHelpStealer uses these markers to try to
378	+	* find a worker to help (i.e., steal back a task from and execute
379	+	* it) that could hasten completion of the actively joined task.
380	+	* In essence, the joiner executes a task that would be on its own
381	+	* local deque had the to-be-joined task not been stolen. This may
382	+	* be seen as a conservative variant of the approach in Wagner &
383	+	* Calder "Leapfrogging: a portable technique for implementing
384	+	* efficient futures" SIGPLAN Notices, 1993
385	+	* (http://portal.acm.org/citation.cfm?id=155354). It differs in
386	+	* that: (1) We only maintain dependency links across workers upon
387	+	* steals, rather than use per-task bookkeeping.  This sometimes
388	+	* requires a linear scan of workQueues array to locate stealers, but
389	+	* often doesn't because stealers leave hints (that may become
390	+	* stale/wrong) of where to locate them.  A stealHint is only a
391	+	* hint because a worker might have had multiple steals and the
392	+	* hint records only one of them (usually the most current).
393	+	* Hinting isolates cost to when it is needed, rather than adding
394	+	* to per-task overhead.  (2) It is "shallow", ignoring nesting
395	+	* and potentially cyclic mutual steals.  (3) It is intentionally
396	+	* racy: field currentJoin is updated only while actively joining,
397	+	* which means that we miss links in the chain during long-lived
398	+	* tasks, GC stalls etc (which is OK since blocking in such cases
399	+	* is usually a good idea).  (4) We bound the number of attempts
400	+	* to find work (see MAX_HELP_DEPTH) and fall back to suspending
401	+	* the worker and if necessary replacing it with another.
402	+	*
403		* It is impossible to keep exactly the target parallelism number
404		* of threads running at any given time.  Determining the
405		* existence of conservatively safe helping targets, the
406		* availability of already-created spares, and the apparent need
407	<	* to create new spares are all racy and require heuristic
408	<	* guidance, so we rely on multiple retries of each.  Currently,
409	<	* in keeping with on-demand signalling policy, we compensate only
410	<	* if blocking would leave less than one active (non-waiting,
411	<	* non-blocked) worker. Additionally, to avoid some false alarms
412	<	* due to GC, lagging counters, system activity, etc, compensated
413	<	* blocking for joins is only attempted after rechecks stabilize
414	<	* (retries are interspersed with Thread.yield, for good
415	<	* citizenship).  The variable blockedCount, incremented before
282	<	* blocking and decremented after, is sometimes needed to
283	<	* distinguish cases of waiting for work vs blocking on joins or
284	<	* other managed sync. Both cases are equivalent for most pool
285	<	* control, so we can update non-atomically. (Additionally,
286	<	* contention on blockedCount alleviates some contention on ctl).
287	<	*
288	<	* Shutdown and Termination. A call to shutdownNow atomically sets
289	<	* the ctl stop bit and then (non-atomically) sets each workers
290	<	* "terminate" status, cancels all unprocessed tasks, and wakes up
291	<	* all waiting workers.  Detecting whether termination should
292	<	* commence after a non-abrupt shutdown() call requires more work
293	<	* and bookkeeping. We need consensus about quiescence (i.e., that
294	<	* there is no more work) which is reflected in active counts so
295	<	* long as there are no current blockers, as well as possible
296	<	* re-evaluations during independent changes in blocking or
297	<	* quiescing workers.
407	>	* to create new spares are all racy, so we rely on multiple
408	>	* retries of each.  Currently, in keeping with on-demand
409	>	* signalling policy, we compensate only if blocking would leave
410	>	* less than one active (non-waiting, non-blocked) worker.
411	>	* Additionally, to avoid some false alarms due to GC, lagging
412	>	* counters, system activity, etc, compensated blocking for joins
413	>	* is only attempted after rechecks stabilize in
414	>	* ForkJoinTask.awaitJoin. (Retries are interspersed with
415	>	* Thread.yield, for good citizenship.)
416		*
417		* Style notes: There is a lot of representation-level coupling
418		* among classes ForkJoinPool, ForkJoinWorkerThread, and
419	<	* ForkJoinTask.  Most fields of ForkJoinWorkerThread maintain
420	<	* data structures managed by ForkJoinPool, so are directly
421	<	* accessed.  Conversely we allow access to "workers" array by
422	<	* workers, and direct access to ForkJoinTask.status by both
423	<	* ForkJoinPool and ForkJoinWorkerThread.  There is little point
424	<	* trying to reduce this, since any associated future changes in
425	<	* representations will need to be accompanied by algorithmic
426	<	* changes anyway. All together, these low-level implementation
427	<	* choices produce as much as a factor of 4 performance
428	<	* improvement compared to naive implementations, and enable the
429	<	* processing of billions of tasks per second, at the expense of
430	<	* some ugliness.
431	<	*
432	<	* Methods signalWork() and scan() are the main bottlenecks so are
433	<	* especially heavily micro-optimized/mangled.  There are lots of
434	<	* inline assignments (of form "while ((local = field) != 0)")
435	<	* which are usually the simplest way to ensure the required read
436	<	* orderings (which are sometimes critical). This leads to a
437	<	* "C"-like style of listing declarations of these locals at the
438	<	* heads of methods or blocks.  There are several occurrences of
439	<	* the unusual "do {} while (!cas...)"  which is the simplest way
440	<	* to force an update of a CAS'ed variable. There are also other
441	<	* coding oddities that help some methods perform reasonably even
442	<	* when interpreted (not compiled).
443	<	*
444	<	* The order of declarations in this file is: (1) declarations of
445	<	* statics (2) fields (along with constants used when unpacking
446	<	* some of them), listed in an order that tends to reduce
447	<	* contention among them a bit under most JVMs.  (3) internal
448	<	* control methods (4) callbacks and other support for
449	<	* ForkJoinTask and ForkJoinWorkerThread classes, (5) exported
450	<	* methods (plus a few little helpers). (6) static block
451	<	* initializing all statics in a minimally dependent order.
419	>	* ForkJoinTask.  The fields of WorkQueue maintain data structures
420	>	* managed by ForkJoinPool, so are directly accessed.  There is
421	>	* little point trying to reduce this, since any associated future
422	>	* changes in representations will need to be accompanied by
423	>	* algorithmic changes anyway. Several methods intrinsically
424	>	* sprawl because they must accumulate sets of consistent reads of
425	>	* volatiles held in local variables.  Methods signalWork() and
426	>	* scan() are the main bottlenecks, so are especially heavily
427	>	* micro-optimized/mangled.  There are lots of inline assignments
428	>	* (of form "while ((local = field) != 0)") which are usually the
429	>	* simplest way to ensure the required read orderings (which are
430	>	* sometimes critical). This leads to a "C"-like style of listing
431	>	* declarations of these locals at the heads of methods or blocks.
432	>	* There are several occurrences of the unusual "do {} while
433	>	* (!cas...)"  which is the simplest way to force an update of a
434	>	* CAS'ed variable. There are also other coding oddities that help
435	>	* some methods perform reasonably even when interpreted (not
436	>	* compiled).
437	>	*
438	>	* The order of declarations in this file is:
439	>	* (1) Static utility functions
440	>	* (2) Nested (static) classes
441	>	* (3) Static fields
442	>	* (4) Fields, along with constants used when unpacking some of them
443	>	* (5) Internal control methods
444	>	* (6) Callbacks and other support for ForkJoinTask methods
445	>	* (7) Exported methods
446	>	* (8) Static block initializing statics in minimally dependent order
447	>	*/
448	>
449	>	// Static utilities
450	>
451	>	/**
452	>	* Computes an initial hash code (also serving as a non-zero
453	>	* random seed) for a thread id. This method is expected to
454	>	* provide higher-quality hash codes than using method hashCode().
455	>	*/
456	>	static final int hashId(long id) {
457	>	int h = (int)id ^ (int)(id >>> 32); // Use MurmurHash of thread id
458	>	h ^= h >>> 16; h *= 0x85ebca6b;
459	>	h ^= h >>> 13; h *= 0xc2b2ae35;
460	>	h ^= h >>> 16;
461	>	return (h == 0) ? 1 : h; // ensure nonzero
462	>	}
463	>
464	>	/**
465	>	* If there is a security manager, makes sure caller has
466	>	* permission to modify threads.
467		*/
468	+	private static void checkPermission() {
469	+	SecurityManager security = System.getSecurityManager();
470	+	if (security != null)
471	+	security.checkPermission(modifyThreadPermission);
472	+	}
473	+
474	+	// Nested classes
475
476		/**
477		* Factory for creating new {@link ForkJoinWorkerThread}s.
#	Line 361 | Line 501 | public class ForkJoinPool extends Abstra
501		}
502
503		/**
504	<	* Creates a new ForkJoinWorkerThread. This factory is used unless
505	<	* overridden in ForkJoinPool constructors.
504	>	* A simple non-reentrant lock used for exclusion when managing
505	>	* queues and workers. We use a custom lock so that we can readily
506	>	* probe lock state in constructions that check among alternative
507	>	* actions. The lock is normally only very briefly held, and
508	>	* sometimes treated as a spinlock, but other usages block to
509	>	* reduce overall contention in those cases where locked code
510	>	* bodies perform allocation/resizing.
511	>	*/
512	>	static final class Mutex extends AbstractQueuedSynchronizer {
513	>	public final boolean tryAcquire(int ignore) {
514	>	return compareAndSetState(0, 1);
515	>	}
516	>	public final boolean tryRelease(int ignore) {
517	>	setState(0);
518	>	return true;
519	>	}
520	>	public final void lock() { acquire(0); }
521	>	public final void unlock() { release(0); }
522	>	public final boolean isHeldExclusively() { return getState() == 1; }
523	>	public final Condition newCondition() { return new ConditionObject(); }
524	>	}
525	>
526	>	/**
527	>	* Class for artificial tasks that are used to replace the target
528	>	* of local joins if they are removed from an interior queue slot
529	>	* in WorkQueue.tryRemoveAndExec. We don't need the proxy to
530	>	* actually do anything beyond having a unique identity.
531	>	*/
532	>	static final class EmptyTask extends ForkJoinTask<Void> {
533	>	EmptyTask() { status = ForkJoinTask.NORMAL; } // force done
534	>	public final Void getRawResult() { return null; }
535	>	public final void setRawResult(Void x) {}
536	>	public final boolean exec() { return true; }
537	>	}
538	>
539	>	/**
540	>	* Queues supporting work-stealing as well as external task
541	>	* submission. See above for main rationale and algorithms.
542	>	* Implementation relies heavily on "Unsafe" intrinsics
543	>	* and selective use of "volatile":
544	>	*
545	>	* Field "base" is the index (mod array.length) of the least valid
546	>	* queue slot, which is always the next position to steal (poll)
547	>	* from if nonempty. Reads and writes require volatile orderings
548	>	* but not CAS, because updates are only performed after slot
549	>	* CASes.
550	>	*
551	>	* Field "top" is the index (mod array.length) of the next queue
552	>	* slot to push to or pop from. It is written only by owner thread
553	>	* for push, or under lock for trySharedPush, and accessed by
554	>	* other threads only after reading (volatile) base.  Both top and
555	>	* base are allowed to wrap around on overflow, but (top - base)
556	>	* (or more commonly -(base - top) to force volatile read of base
557	>	* before top) still estimates size.
558	>	*
559	>	* The array slots are read and written using the emulation of
560	>	* volatiles/atomics provided by Unsafe. Insertions must in
561	>	* general use putOrderedObject as a form of releasing store to
562	>	* ensure that all writes to the task object are ordered before
563	>	* its publication in the queue. (Although we can avoid one case
564	>	* of this when locked in trySharedPush.) All removals entail a
565	>	* CAS to null.  The array is always a power of two. To ensure
566	>	* safety of Unsafe array operations, all accesses perform
567	>	* explicit null checks and implicit bounds checks via
568	>	* power-of-two masking.
569	>	*
570	>	* In addition to basic queuing support, this class contains
571	>	* fields described elsewhere to control execution. It turns out
572	>	* to work better memory-layout-wise to include them in this
573	>	* class rather than a separate class.
574	>	*
575	>	* Performance on most platforms is very sensitive to placement of
576	>	* instances of both WorkQueues and their arrays -- we absolutely
577	>	* do not want multiple WorkQueue instances or multiple queue
578	>	* arrays sharing cache lines. (It would be best for queue objects
579	>	* and their arrays to share, but there is nothing available to
580	>	* help arrange that).  Unfortunately, because they are recorded
581	>	* in a common array, WorkQueue instances are often moved to be
582	>	* adjacent by garbage collectors. To reduce impact, we use field
583	>	* padding that works OK on common platforms; this effectively
584	>	* trades off slightly slower average field access for the sake of
585	>	* avoiding really bad worst-case access. (Until better JVM
586	>	* support is in place, this padding is dependent on transient
587	>	* properties of JVM field layout rules.)  We also take care in
588	>	* allocating, sizing and resizing the array. Non-shared queue
589	>	* arrays are initialized (via method growArray) by workers before
590	>	* use. Others are allocated on first use.
591		*/
592	<	public static final ForkJoinWorkerThreadFactory
593	<	defaultForkJoinWorkerThreadFactory;
592	>	static final class WorkQueue {
593	>	/**
594	>	* Capacity of work-stealing queue array upon initialization.
595	>	* Must be a power of two; at least 4, but set larger to
596	>	* reduce cacheline sharing among queues.
597	>	*/
598	>	static final int INITIAL_QUEUE_CAPACITY = 1 << 8;
599
600	<	/**
601	<	* Permission required for callers of methods that may start or
602	<	* kill threads.
603	<	*/
604	<	private static final RuntimePermission modifyThreadPermission;
600	>	/**
601	>	* Maximum size for queue arrays. Must be a power of two less
602	>	* than or equal to 1 << (31 - width of array entry) to ensure
603	>	* lack of wraparound of index calculations, but defined to a
604	>	* value a bit less than this to help users trap runaway
605	>	* programs before saturating systems.
606	>	*/
607	>	static final int MAXIMUM_QUEUE_CAPACITY = 1 << 26; // 64M
608
609	<	/**
610	<	* If there is a security manager, makes sure caller has
611	<	* permission to modify threads.
612	<	*/
613	<	private static void checkPermission() {
614	<	SecurityManager security = System.getSecurityManager();
615	<	if (security != null)
616	<	security.checkPermission(modifyThreadPermission);
617	<	}
609	>	volatile long totalSteals; // cumulative number of steals
610	>	int seed;                  // for random scanning; initialize nonzero
611	>	volatile int eventCount;   // encoded inactivation count; < 0 if inactive
612	>	int nextWait;              // encoded record of next event waiter
613	>	int rescans;               // remaining scans until block
614	>	int nsteals;               // top-level task executions since last idle
615	>	final int mode;            // lifo, fifo, or shared
616	>	int poolIndex;             // index of this queue in pool (or 0)
617	>	int stealHint;             // index of most recent known stealer
618	>	volatile int runState;     // 1: locked, -1: terminate; else 0
619	>	volatile int base;         // index of next slot for poll
620	>	int top;                   // index of next slot for push
621	>	ForkJoinTask<?>[] array;   // the elements (initially unallocated)
622	>	final ForkJoinWorkerThread owner; // owning thread or null if shared
623	>	volatile Thread parker;    // == owner during call to park; else null
624	>	ForkJoinTask<?> currentJoin;  // task being joined in awaitJoin
625	>	ForkJoinTask<?> currentSteal; // current non-local task being executed
626	>	// Heuristic padding to ameliorate unfortunate memory placements
627	>	Object p00, p01, p02, p03, p04, p05, p06, p07, p08, p09, p0a;
628	>
629	>	WorkQueue(ForkJoinWorkerThread owner, int mode) {
630	>	this.owner = owner;
631	>	this.mode = mode;
632	>	// Place indices in the center of array (that is not yet allocated)
633	>	base = top = INITIAL_QUEUE_CAPACITY >>> 1;
634	>	}
635
636	<	/**
637	<	* Generator for assigning sequence numbers as pool names.
638	<	*/
639	<	private static final AtomicInteger poolNumberGenerator;
636	>	/**
637	>	* Returns number of tasks in the queue.
638	>	*/
639	>	final int queueSize() {
640	>	int n = base - top; // non-owner callers must read base first
641	>	return (n >= 0) ? 0 : -n;
642	>	}
643
644	<	/**
645	<	* Generator for initial random seeds for worker victim
646	<	* selection. This is used only to create initial seeds. Random
647	<	* steals use a cheaper xorshift generator per steal attempt. We
648	<	* don't expect much contention on seedGenerator, so just use a
649	<	* plain Random.
650	<	*/
651	<	static final Random workerSeedGenerator;
644	>	/**
645	>	* Pushes a task. Call only by owner in unshared queues.
646	>	*
647	>	* @param task the task. Caller must ensure non-null.
648	>	* @param p if non-null, pool to signal if necessary
649	>	* @throw RejectedExecutionException if array cannot be resized
650	>	*/
651	>	final void push(ForkJoinTask<?> task, ForkJoinPool p) {
652	>	ForkJoinTask<?>[] a;
653	>	int s = top, m, n;
654	>	if ((a = array) != null) {    // ignore if queue removed
655	>	U.putOrderedObject
656	>	(a, (((m = a.length - 1) & s) << ASHIFT) + ABASE, task);
657	>	if ((n = (top = s + 1) - base) <= 2) {
658	>	if (p != null)
659	>	p.signalWork();
660	>	}
661	>	else if (n >= m)
662	>	growArray(true);
663	>	}
664	>	}
665
666	<	/**
667	<	* Array holding all worker threads in the pool.  Initialized upon
668	<	* construction. Array size must be a power of two.  Updates and
669	<	* replacements are protected by scanGuard, but the array is
670	<	* always kept in a consistent enough state to be randomly
671	<	* accessed without locking by workers performing work-stealing,
672	<	* as well as other traversal-based methods in this class, so long
673	<	* as reads memory-acquire by first reading ctl. All readers must
674	<	* tolerate that some array slots may be null.
675	<	*/
676	<	ForkJoinWorkerThread[] workers;
666	>	/**
667	>	* Pushes a task if lock is free and array is either big
668	>	* enough or can be resized to be big enough.
669	>	*
670	>	* @param task the task. Caller must ensure non-null.
671	>	* @return true if submitted
672	>	*/
673	>	final boolean trySharedPush(ForkJoinTask<?> task) {
674	>	boolean submitted = false;
675	>	if (runState == 0 && U.compareAndSwapInt(this, RUNSTATE, 0, 1)) {
676	>	ForkJoinTask<?>[] a = array;
677	>	int s = top;
678	>	try {
679	>	if ((a != null && a.length > s + 1 - base) ||
680	>	(a = growArray(false)) != null) { // must presize
681	>	int j = (((a.length - 1) & s) << ASHIFT) + ABASE;
682	>	U.putObject(a, (long)j, task);    // don't need "ordered"
683	>	top = s + 1;
684	>	submitted = true;
685	>	}
686	>	} finally {
687	>	runState = 0;                         // unlock
688	>	}
689	>	}
690	>	return submitted;
691	>	}
692
693	<	/**
694	<	* Initial size for submission queue array. Must be a power of
695	<	* two.  In many applications, these always stay small so we use a
696	<	* small initial cap.
697	<	*/
698	<	private static final int INITIAL_QUEUE_CAPACITY = 8;
693	>	/**
694	>	* Takes next task, if one exists, in FIFO order.
695	>	*/
696	>	final ForkJoinTask<?> poll() {
697	>	ForkJoinTask<?>[] a; int b; ForkJoinTask<?> t;
698	>	while ((b = base) - top < 0 && (a = array) != null) {
699	>	int j = (((a.length - 1) & b) << ASHIFT) + ABASE;
700	>	if ((t = (ForkJoinTask<?>)U.getObjectVolatile(a, j)) != null &&
701	>	base == b &&
702	>	U.compareAndSwapObject(a, j, t, null)) {
703	>	base = b + 1;
704	>	return t;
705	>	}
706	>	}
707	>	return null;
708	>	}
709	>
710	>	/**
711	>	* Takes next task, if one exists, in LIFO order.  Call only
712	>	* by owner in unshared queues. (We do not have a shared
713	>	* version of this method because it is never needed.)
714	>	*/
715	>	final ForkJoinTask<?> pop() {
716	>	ForkJoinTask<?> t; int m;
717	>	ForkJoinTask<?>[] a = array;
718	>	if (a != null && (m = a.length - 1) >= 0) {
719	>	for (int s; (s = top - 1) - base >= 0;) {
720	>	int j = ((m & s) << ASHIFT) + ABASE;
721	>	if ((t = (ForkJoinTask<?>)U.getObjectVolatile(a, j)) == null)
722	>	break;
723	>	if (U.compareAndSwapObject(a, j, t, null)) {
724	>	top = s;
725	>	return t;
726	>	}
727	>	}
728	>	}
729	>	return null;
730	>	}
731	>
732	>	/**
733	>	* Takes next task, if one exists, in order specified by mode.
734	>	*/
735	>	final ForkJoinTask<?> nextLocalTask() {
736	>	return mode == 0 ? pop() : poll();
737	>	}
738	>
739	>	/**
740	>	* Returns next task, if one exists, in order specified by mode.
741	>	*/
742	>	final ForkJoinTask<?> peek() {
743	>	ForkJoinTask<?>[] a = array; int m;
744	>	if (a == null || (m = a.length - 1) < 0)
745	>	return null;
746	>	int i = mode == 0 ? top - 1 : base;
747	>	int j = ((i & m) << ASHIFT) + ABASE;
748	>	return (ForkJoinTask<?>)U.getObjectVolatile(a, j);
749	>	}
750	>
751	>	/**
752	>	* Returns task at index b if b is current base of queue.
753	>	*/
754	>	final ForkJoinTask<?> pollAt(int b) {
755	>	ForkJoinTask<?> t; ForkJoinTask<?>[] a;
756	>	if ((a = array) != null) {
757	>	int j = (((a.length - 1) & b) << ASHIFT) + ABASE;
758	>	if ((t = (ForkJoinTask<?>)U.getObjectVolatile(a, j)) != null &&
759	>	base == b &&
760	>	U.compareAndSwapObject(a, j, t, null)) {
761	>	base = b + 1;
762	>	return t;
763	>	}
764	>	}
765	>	return null;
766	>	}
767	>
768	>	/**
769	>	* Pops the given task only if it is at the current top.
770	>	*/
771	>	final boolean tryUnpush(ForkJoinTask<?> t) {
772	>	ForkJoinTask<?>[] a; int s;
773	>	if ((a = array) != null && (s = top) != base &&
774	>	U.compareAndSwapObject
775	>	(a, (((a.length - 1) & --s) << ASHIFT) + ABASE, t, null)) {
776	>	top = s;
777	>	return true;
778	>	}
779	>	return false;
780	>	}
781	>
782	>	/**
783	>	* Polls the given task only if it is at the current base.
784	>	*/
785	>	final boolean pollFor(ForkJoinTask<?> task) {
786	>	ForkJoinTask<?>[] a; int b;
787	>	if ((b = base) - top < 0 && (a = array) != null) {
788	>	int j = (((a.length - 1) & b) << ASHIFT) + ABASE;
789	>	if (U.getObjectVolatile(a, j) == task && base == b &&
790	>	U.compareAndSwapObject(a, j, task, null)) {
791	>	base = b + 1;
792	>	return true;
793	>	}
794	>	}
795	>	return false;
796	>	}
797	>
798	>	/**
799	>	* If present, removes from queue and executes the given task, or
800	>	* any other cancelled task. Returns (true) immediately on any CAS
801	>	* or consistency check failure so caller can retry.
802	>	*
803	>	* @return false if no progress can be made
804	>	*/
805	>	final boolean tryRemoveAndExec(ForkJoinTask<?> task) {
806	>	boolean removed = false, empty = true, progress = true;
807	>	ForkJoinTask<?>[] a; int m, s, b, n;
808	>	if ((a = array) != null && (m = a.length - 1) >= 0 &&
809	>	(n = (s = top) - (b = base)) > 0) {
810	>	for (ForkJoinTask<?> t;;) {           // traverse from s to b
811	>	int j = ((--s & m) << ASHIFT) + ABASE;
812	>	t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
813	>	if (t == null)                    // inconsistent length
814	>	break;
815	>	else if (t == task) {
816	>	if (s + 1 == top) {           // pop
817	>	if (!U.compareAndSwapObject(a, j, task, null))
818	>	break;
819	>	top = s;
820	>	removed = true;
821	>	}
822	>	else if (base == b)           // replace with proxy
823	>	removed = U.compareAndSwapObject(a, j, task,
824	>	new EmptyTask());
825	>	break;
826	>	}
827	>	else if (t.status >= 0)
828	>	empty = false;
829	>	else if (s + 1 == top) {          // pop and throw away
830	>	if (U.compareAndSwapObject(a, j, t, null))
831	>	top = s;
832	>	break;
833	>	}
834	>	if (--n == 0) {
835	>	if (!empty && base == b)
836	>	progress = false;
837	>	break;
838	>	}
839	>	}
840	>	}
841	>	if (removed)
842	>	task.doExec();
843	>	return progress;
844	>	}
845	>
846	>	/**
847	>	* Initializes or doubles the capacity of array. Call either
848	>	* by owner or with lock held -- it is OK for base, but not
849	>	* top, to move while resizings are in progress.
850	>	*
851	>	* @param rejectOnFailure if true, throw exception if capacity
852	>	* exceeded (relayed ultimately to user); else return null.
853	>	*/
854	>	final ForkJoinTask<?>[] growArray(boolean rejectOnFailure) {
855	>	ForkJoinTask<?>[] oldA = array;
856	>	int size = oldA != null ? oldA.length << 1 : INITIAL_QUEUE_CAPACITY;
857	>	if (size <= MAXIMUM_QUEUE_CAPACITY) {
858	>	int oldMask, t, b;
859	>	ForkJoinTask<?>[] a = array = new ForkJoinTask<?>[size];
860	>	if (oldA != null && (oldMask = oldA.length - 1) >= 0 &&
861	>	(t = top) - (b = base) > 0) {
862	>	int mask = size - 1;
863	>	do {
864	>	ForkJoinTask<?> x;
865	>	int oldj = ((b & oldMask) << ASHIFT) + ABASE;
866	>	int j    = ((b &    mask) << ASHIFT) + ABASE;
867	>	x = (ForkJoinTask<?>)U.getObjectVolatile(oldA, oldj);
868	>	if (x != null &&
869	>	U.compareAndSwapObject(oldA, oldj, x, null))
870	>	U.putObjectVolatile(a, j, x);
871	>	} while (++b != t);
872	>	}
873	>	return a;
874	>	}
875	>	else if (!rejectOnFailure)
876	>	return null;
877	>	else
878	>	throw new RejectedExecutionException("Queue capacity exceeded");
879	>	}
880	>
881	>	/**
882	>	* Removes and cancels all known tasks, ignoring any exceptions.
883	>	*/
884	>	final void cancelAll() {
885	>	ForkJoinTask.cancelIgnoringExceptions(currentJoin);
886	>	ForkJoinTask.cancelIgnoringExceptions(currentSteal);
887	>	for (ForkJoinTask<?> t; (t = poll()) != null; )
888	>	ForkJoinTask.cancelIgnoringExceptions(t);
889	>	}
890	>
891	>	/**
892	>	* Computes next value for random probes.  Scans don't require
893	>	* a very high quality generator, but also not a crummy one.
894	>	* Marsaglia xor-shift is cheap and works well enough.  Note:
895	>	* This is manually inlined in several usages in ForkJoinPool
896	>	* to avoid writes inside busy scan loops.
897	>	*/
898	>	final int nextSeed() {
899	>	int r = seed;
900	>	r ^= r << 13;
901	>	r ^= r >>> 17;
902	>	return seed = r ^= r << 5;
903	>	}
904	>
905	>	// Execution methods
906	>
907	>	/**
908	>	* Removes and runs tasks until empty, using local mode
909	>	* ordering.
910	>	*/
911	>	final void runLocalTasks() {
912	>	if (base - top < 0) {
913	>	for (ForkJoinTask<?> t; (t = nextLocalTask()) != null; )
914	>	t.doExec();
915	>	}
916	>	}
917	>
918	>	/**
919	>	* Executes a top-level task and any local tasks remaining
920	>	* after execution.
921	>	*
922	>	* @return true unless terminating
923	>	*/
924	>	final boolean runTask(ForkJoinTask<?> t) {
925	>	boolean alive = true;
926	>	if (t != null) {
927	>	currentSteal = t;
928	>	t.doExec();
929	>	runLocalTasks();
930	>	++nsteals;
931	>	currentSteal = null;
932	>	}
933	>	else if (runState < 0)            // terminating
934	>	alive = false;
935	>	return alive;
936	>	}
937	>
938	>	/**
939	>	* Executes a non-top-level (stolen) task.
940	>	*/
941	>	final void runSubtask(ForkJoinTask<?> t) {
942	>	if (t != null) {
943	>	ForkJoinTask<?> ps = currentSteal;
944	>	currentSteal = t;
945	>	t.doExec();
946	>	currentSteal = ps;
947	>	}
948	>	}
949	>
950	>	/**
951	>	* Returns true if owned and not known to be blocked.
952	>	*/
953	>	final boolean isApparentlyUnblocked() {
954	>	Thread wt; Thread.State s;
955	>	return (eventCount >= 0 &&
956	>	(wt = owner) != null &&
957	>	(s = wt.getState()) != Thread.State.BLOCKED &&
958	>	s != Thread.State.WAITING &&
959	>	s != Thread.State.TIMED_WAITING);
960	>	}
961	>
962	>	/**
963	>	* If this owned and is not already interrupted, try to
964	>	* interrupt and/or unpark, ignoring exceptions.
965	>	*/
966	>	final void interruptOwner() {
967	>	Thread wt, p;
968	>	if ((wt = owner) != null && !wt.isInterrupted()) {
969	>	try {
970	>	wt.interrupt();
971	>	} catch (SecurityException ignore) {
972	>	}
973	>	}
974	>	if ((p = parker) != null)
975	>	U.unpark(p);
976	>	}
977	>
978	>	// Unsafe mechanics
979	>	private static final sun.misc.Unsafe U;
980	>	private static final long RUNSTATE;
981	>	private static final int ABASE;
982	>	private static final int ASHIFT;
983	>	static {
984	>	int s;
985	>	try {
986	>	U = getUnsafe();
987	>	Class<?> k = WorkQueue.class;
988	>	Class<?> ak = ForkJoinTask[].class;
989	>	RUNSTATE = U.objectFieldOffset
990	>	(k.getDeclaredField("runState"));
991	>	ABASE = U.arrayBaseOffset(ak);
992	>	s = U.arrayIndexScale(ak);
993	>	} catch (Exception e) {
994	>	throw new Error(e);
995	>	}
996	>	if ((s & (s-1)) != 0)
997	>	throw new Error("data type scale not a power of two");
998	>	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
999	>	}
1000	>	}
1001
1002		/**
1003	<	* Maximum size for submission queue array. Must be a power of two
1004	<	* less than or equal to 1 << (31 - width of array entry) to
1005	<	* ensure lack of index wraparound, but is capped at a lower
1006	<	* value to help users trap runaway computations.
1003	>	* Per-thread records for threads that submit to pools. Currently
1004	>	* holds only pseudo-random seed / index that is used to choose
1005	>	* submission queues in method doSubmit. In the future, this may
1006	>	* also incorporate a means to implement different task rejection
1007	>	* and resubmission policies.
1008		*/
1009	<	private static final int MAXIMUM_QUEUE_CAPACITY = 1 << 24; // 16M
1009	>	static final class Submitter {
1010	>	int seed;
1011	>	Submitter() { seed = hashId(Thread.currentThread().getId()); }
1012	>	}
1013	>
1014	>	/** ThreadLocal class for Submitters */
1015	>	static final class ThreadSubmitter extends ThreadLocal<Submitter> {
1016	>	public Submitter initialValue() { return new Submitter(); }
1017	>	}
1018	>
1019	>	// static fields (initialized in static initializer below)
1020
1021		/**
1022	<	* Array serving as submission queue. Initialized upon construction.
1022	>	* Creates a new ForkJoinWorkerThread. This factory is used unless
1023	>	* overridden in ForkJoinPool constructors.
1024		*/
1025	<	private ForkJoinTask<?>[] submissionQueue;
1025	>	public static final ForkJoinWorkerThreadFactory
1026	>	defaultForkJoinWorkerThreadFactory;
1027
1028		/**
1029	<	* Lock protecting submissions array for addSubmission
1029	>	* Generator for assigning sequence numbers as pool names.
1030		*/
1031	<	private final ReentrantLock submissionLock;
1031	>	private static final AtomicInteger poolNumberGenerator;
1032
1033		/**
1034	<	* Condition for awaitTermination, using submissionLock for
1035	<	* convenience.
1034	>	* Permission required for callers of methods that may start or
1035	>	* kill threads.
1036		*/
1037	<	private final Condition termination;
1037	>	private static final RuntimePermission modifyThreadPermission;
1038
1039		/**
1040	<	* Creation factory for worker threads.
1040	>	* Per-thread submission bookeeping. Shared across all pools
1041	>	* to reduce ThreadLocal pollution and because random motion
1042	>	* to avoid contention in one pool is likely to hold for others.
1043		*/
1044	<	private final ForkJoinWorkerThreadFactory factory;
1044	>	private static final ThreadSubmitter submitters;
1045	>
1046	>	// static constants
1047
1048		/**
1049	<	* The uncaught exception handler used when any worker abruptly
1050	<	* terminates.
1049	>	* The wakeup interval (in nanoseconds) for a worker waiting for a
1050	>	* task when the pool is quiescent to instead try to shrink the
1051	>	* number of workers.  The exact value does not matter too
1052	>	* much. It must be short enough to release resources during
1053	>	* sustained periods of idleness, but not so short that threads
1054	>	* are continually re-created.
1055		*/
1056	<	final Thread.UncaughtExceptionHandler ueh;
1056	>	private static final long SHRINK_RATE =
1057	>	4L * 1000L * 1000L * 1000L; // 4 seconds
1058
1059		/**
1060	<	* Prefix for assigning names to worker threads
1060	>	* The timeout value for attempted shrinkage, includes
1061	>	* some slop to cope with system timer imprecision.
1062		*/
1063	<	private final String workerNamePrefix;
1063	>	private static final long SHRINK_TIMEOUT = SHRINK_RATE - (SHRINK_RATE / 10);
1064
1065		/**
1066	<	* Sum of per-thread steal counts, updated only when threads are
1067	<	* idle or terminating.
1066	>	* The maximum stolen->joining link depth allowed in tryHelpStealer.
1067	>	* Depths for legitimate chains are unbounded, but we use a fixed
1068	>	* constant to avoid (otherwise unchecked) cycles and to bound
1069	>	* staleness of traversal parameters at the expense of sometimes
1070	>	* blocking when we could be helping.
1071		*/
1072	<	private volatile long stealCount;
1072	>	private static final int MAX_HELP_DEPTH = 16;
1073
1074		/**
1075	<	* Main pool control -- a long packed with:
1075	>	* Bits and masks for control variables
1076	>	*
1077	>	* Field ctl is a long packed with:
1078		* AC: Number of active running workers minus target parallelism (16 bits)
1079		* TC: Number of total workers minus target parallelism (16 bits)
1080		* ST: true if pool is terminating (1 bit)
1081		* EC: the wait count of top waiting thread (15 bits)
1082	<	* ID: ~poolIndex of top of Treiber stack of waiting threads (16 bits)
1082	>	* ID: poolIndex of top of Treiber stack of waiters (16 bits)
1083		*
1084		* When convenient, we can extract the upper 32 bits of counts and
1085		* the lower 32 bits of queue state, u = (int)(ctl >>> 32) and e =
#	Line 477 | Line 1088 | public class ForkJoinPool extends Abstra
1088		* parallelism and the positionings of fields makes it possible to
1089		* perform the most common checks via sign tests of fields: When
1090		* ac is negative, there are not enough active workers, when tc is
1091	<	* negative, there are not enough total workers, when id is
481	<	* negative, there is at least one waiting worker, and when e is
1091	>	* negative, there are not enough total workers, and when e is
1092		* negative, the pool is terminating.  To deal with these possibly
1093		* negative fields, we use casts in and out of "short" and/or
1094		* signed shifts to maintain signedness.
1095	+	*
1096	+	* When a thread is queued (inactivated), its eventCount field is
1097	+	* set negative, which is the only way to tell if a worker is
1098	+	* prevented from executing tasks, even though it must continue to
1099	+	* scan for them to avoid queuing races. Note however that
1100	+	* eventCount updates lag releases so usage requires care.
1101	+	*
1102	+	* Field runState is an int packed with:
1103	+	* SHUTDOWN: true if shutdown is enabled (1 bit)
1104	+	* SEQ:  a sequence number updated upon (de)registering workers (15 bits)
1105	+	* MASK: mask (power of 2 - 1) covering all registered poolIndexes (16 bits)
1106	+	*
1107	+	* The combination of mask and sequence number enables simple
1108	+	* consistency checks: Staleness of read-only operations on the
1109	+	* workQueues array can be checked by comparing runState before vs
1110	+	* after the reads. The low 16 bits (i.e, anding with SMASK) hold
1111	+	* the smallest power of two covering all indices, minus
1112	+	* one.
1113		*/
486	–	volatile long ctl;
1114
1115		// bit positions/shifts for fields
1116		private static final int  AC_SHIFT   = 48;
#	Line 492 | Line 1119 | public class ForkJoinPool extends Abstra
1119		private static final int  EC_SHIFT   = 16;
1120
1121		// bounds
1122	<	private static final int  MAX_ID     = 0x7fff;  // max poolIndex
1123	<	private static final int  SMASK      = 0xffff;  // mask short bits
1122	>	private static final int  POOL_MAX   = 0x7fff;  // max #workers - 1
1123	>	private static final int  SMASK      = 0xffff;  // short bits
1124	>	private static final int  SQMASK     = 0xfffe;  // even short bits
1125		private static final int  SHORT_SIGN = 1 << 15;
1126		private static final int  INT_SIGN   = 1 << 31;
1127
#	Line 515 | Line 1143 | public class ForkJoinPool extends Abstra
1143		private static final int  UTC_UNIT   = 1 << UTC_SHIFT;
1144
1145		// masks and units for dealing with e = (int)ctl
1146	<	private static final int  E_MASK     = 0x7fffffff; // no STOP_BIT
1147	<	private static final int  EC_UNIT    = 1 << EC_SHIFT;
520	<
521	<	/**
522	<	* The target parallelism level.
523	<	*/
524	<	final int parallelism;
525	<
526	<	/**
527	<	* Index (mod submission queue length) of next element to take
528	<	* from submission queue. Usage is identical to that for
529	<	* per-worker queues -- see ForkJoinWorkerThread internal
530	<	* documentation.
531	<	*/
532	<	volatile int queueBase;
533	<
534	<	/**
535	<	* Index (mod submission queue length) of next element to add
536	<	* in submission queue. Usage is identical to that for
537	<	* per-worker queues -- see ForkJoinWorkerThread internal
538	<	* documentation.
539	<	*/
540	<	int queueTop;
541	<
542	<	/**
543	<	* True when shutdown() has been called.
544	<	*/
545	<	volatile boolean shutdown;
546	<
547	<	/**
548	<	* True if use local fifo, not default lifo, for local polling.
549	<	* Read by, and replicated by ForkJoinWorkerThreads.
550	<	*/
551	<	final boolean locallyFifo;
552	<
553	<	/**
554	<	* The number of threads in ForkJoinWorkerThreads.helpQuiescePool.
555	<	* When non-zero, suppresses automatic shutdown when active
556	<	* counts become zero.
557	<	*/
558	<	volatile int quiescerCount;
1146	>	private static final int E_MASK      = 0x7fffffff; // no STOP_BIT
1147	>	private static final int E_SEQ       = 1 << EC_SHIFT;
1148
1149	<	/**
1150	<	* The number of threads blocked in join.
1151	<	*/
1152	<	volatile int blockedCount;
1149	>	// runState bits
1150	>	private static final int SHUTDOWN    = 1 << 31;
1151	>	private static final int RS_SEQ      = 1 << 16;
1152	>	private static final int RS_SEQ_MASK = 0x7fff0000;
1153	>
1154	>	// access mode for WorkQueue
1155	>	static final int LIFO_QUEUE          =  0;
1156	>	static final int FIFO_QUEUE          =  1;
1157	>	static final int SHARED_QUEUE        = -1;
1158
1159	<	/**
566	<	* Counter for worker Thread names (unrelated to their poolIndex)
567	<	*/
568	<	private volatile int nextWorkerNumber;
569	<
570	<	/**
571	<	* The index for the next created worker. Accessed under scanGuard.
572	<	*/
573	<	private int nextWorkerIndex;
1159	>	// Instance fields
1160
1161	<	/**
1162	<	* SeqLock and index masking for updates to workers array.  Locked
1163	<	* when SG_UNIT is set. Unlocking clears bit by adding
1164	<	* SG_UNIT. Staleness of read-only operations can be checked by
1165	<	* comparing scanGuard to value before the reads. The low 16 bits
1166	<	* (i.e, anding with SMASK) hold (the smallest power of two
1167	<	* covering all worker indices, minus one, and is used to avoid
1168	<	* dealing with large numbers of null slots when the workers array
1169	<	* is overallocated.
1170	<	*/
1171	<	volatile int scanGuard;
1172	<
1173	<	private static final int SG_UNIT = 1 << 16;
1174	<
1175	<	/**
1176	<	* The wakeup interval (in nanoseconds) for a worker waiting for a
1177	<	* task when the pool is quiescent to instead try to shrink the
1178	<	* number of workers.  The exact value does not matter too
1179	<	* much. It must be short enough to release resources during
1180	<	* sustained periods of idleness, but not so short that threads
595	<	* are continually re-created.
596	<	*/
597	<	private static final long SHRINK_RATE =
598	<	4L * 1000L * 1000L * 1000L; // 4 seconds
599	<
600	<	/**
601	<	* Top-level loop for worker threads: On each step: if the
602	<	* previous step swept through all queues and found no tasks, or
603	<	* there are excess threads, then possibly blocks. Otherwise,
604	<	* scans for and, if found, executes a task. Returns when pool
605	<	* and/or worker terminate.
606	<	*
607	<	* @param w the worker
608	<	*/
609	<	final void work(ForkJoinWorkerThread w) {
610	<	boolean swept = false;                // true on empty scans
611	<	long c;
612	<	while (!w.terminate && (int)(c = ctl) >= 0) {
613	<	int a;                            // active count
614	<	if (!swept && (a = (int)(c >> AC_SHIFT)) <= 0)
615	<	swept = scan(w, a);
616	<	else if (tryAwaitWork(w, c))
617	<	swept = false;
618	<	}
619	<	}
1161	>	/*
1162	>	* Field layout order in this class tends to matter more than one
1163	>	* would like. Runtime layout order is only loosely related to
1164	>	* declaration order and may differ across JVMs, but the following
1165	>	* empirically works OK on current JVMs.
1166	>	*/
1167	>
1168	>	volatile long ctl;                         // main pool control
1169	>	final int parallelism;                     // parallelism level
1170	>	final int localMode;                       // per-worker scheduling mode
1171	>	int growHints;                             // for expanding indices/ranges
1172	>	volatile int runState;                     // shutdown status, seq, and mask
1173	>	WorkQueue[] workQueues;                    // main registry
1174	>	final Mutex lock;                          // for registration
1175	>	final Condition termination;               // for awaitTermination
1176	>	final ForkJoinWorkerThreadFactory factory; // factory for new workers
1177	>	final Thread.UncaughtExceptionHandler ueh; // per-worker UEH
1178	>	final AtomicLong stealCount;               // collect counts when terminated
1179	>	final AtomicInteger nextWorkerNumber;      // to create worker name string
1180	>	final String workerNamePrefix;             // to create worker name string
1181
1182	<	// Signalling
1182	>	//  Creating, registering, deregistering and running workers
1183
1184		/**
1185	<	* Wakes up or creates a worker.
1185	>	* Tries to create and start a worker
1186		*/
1187	<	final void signalWork() {
1188	<	/*
1189	<	* The while condition is true if: (there is are too few total
1190	<	* workers OR there is at least one waiter) AND (there are too
1191	<	* few active workers OR the pool is terminating).  The value
1192	<	* of e distinguishes the remaining cases: zero (no waiters)
1193	<	* for create, negative if terminating (in which case do
633	<	* nothing), else release a waiter. The secondary checks for
634	<	* release (non-null array etc) can fail if the pool begins
635	<	* terminating after the test, and don't impose any added cost
636	<	* because JVMs must perform null and bounds checks anyway.
637	<	*/
638	<	long c; int e, u;
639	<	while ((((e = (int)(c = ctl)) | (u = (int)(c >>> 32))) &
640	<	(INT_SIGN|SHORT_SIGN)) == (INT_SIGN|SHORT_SIGN) && e >= 0) {
641	<	if (e > 0) {                         // release a waiting worker
642	<	int i; ForkJoinWorkerThread w; ForkJoinWorkerThread[] ws;
643	<	if ((ws = workers) == null ||
644	<	(i = ~e & SMASK) >= ws.length ||
645	<	(w = ws[i]) == null)
646	<	break;
647	<	long nc = (((long)(w.nextWait & E_MASK)) |
648	<	((long)(u + UAC_UNIT) << 32));
649	<	if (w.eventCount == e &&
650	<	UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
651	<	w.eventCount = (e + EC_UNIT) & E_MASK;
652	<	if (w.parked)
653	<	UNSAFE.unpark(w);
654	<	break;
655	<	}
656	<	}
657	<	else if (UNSAFE.compareAndSwapLong
658	<	(this, ctlOffset, c,
659	<	(long)(((u + UTC_UNIT) & UTC_MASK) |
660	<	((u + UAC_UNIT) & UAC_MASK)) << 32)) {
661	<	addWorker();
662	<	break;
1187	>	private void addWorker() {
1188	>	Throwable ex = null;
1189	>	ForkJoinWorkerThread wt = null;
1190	>	try {
1191	>	if ((wt = factory.newThread(this)) != null) {
1192	>	wt.start();
1193	>	return;
1194		}
1195	+	} catch (Throwable e) {
1196	+	ex = e;
1197		}
1198	+	deregisterWorker(wt, ex); // adjust counts etc on failure
1199		}
1200
1201		/**
1202	<	* Variant of signalWork to help release waiters on rescans.
1203	<	* Tries once to release a waiter if active count < 0.
1204	<	*
1205	<	* @return false if failed due to contention, else true
1202	>	* Callback from ForkJoinWorkerThread constructor to assign a
1203	>	* public name. This must be separate from registerWorker because
1204	>	* it is called during the "super" constructor call in
1205	>	* ForkJoinWorkerThread.
1206		*/
1207	<	private boolean tryReleaseWaiter() {
1208	<	long c; int e, i; ForkJoinWorkerThread w; ForkJoinWorkerThread[] ws;
1209	<	if ((e = (int)(c = ctl)) > 0 &&
676	<	(int)(c >> AC_SHIFT) < 0 &&
677	<	(ws = workers) != null &&
678	<	(i = ~e & SMASK) < ws.length &&
679	<	(w = ws[i]) != null) {
680	<	long nc = ((long)(w.nextWait & E_MASK) |
681	<	((c + AC_UNIT) & (AC_MASK|TC_MASK)));
682	<	if (w.eventCount != e ||
683	<	!UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc))
684	<	return false;
685	<	w.eventCount = (e + EC_UNIT) & E_MASK;
686	<	if (w.parked)
687	<	UNSAFE.unpark(w);
688	<	}
689	<	return true;
690	<	}
691	<
692	<	// Scanning for tasks
693	<
694	<	/**
695	<	* Scans for and, if found, executes one task. Scans start at a
696	<	* random index of workers array, and randomly select the first
697	<	* (2*#workers)-1 probes, and then, if all empty, resort to 2
698	<	* circular sweeps, which is necessary to check quiescence. and
699	<	* taking a submission only if no stealable tasks were found.  The
700	<	* steal code inside the loop is a specialized form of
701	<	* ForkJoinWorkerThread.deqTask, followed bookkeeping to support
702	<	* helpJoinTask and signal propagation. The code for submission
703	<	* queues is almost identical. On each steal, the worker completes
704	<	* not only the task, but also all local tasks that this task may
705	<	* have generated. On detecting staleness or contention when
706	<	* trying to take a task, this method returns without finishing
707	<	* sweep, which allows global state rechecks before retry.
708	<	*
709	<	* @param w the worker
710	<	* @param a the number of active workers
711	<	* @return true if swept all queues without finding a task
712	<	*/
713	<	private boolean scan(ForkJoinWorkerThread w, int a) {
714	<	int g = scanGuard; // mask 0 avoids useless scans if only one active
715	<	int m = (parallelism == 1 - a && blockedCount == 0) ? 0 : g & SMASK;
716	<	ForkJoinWorkerThread[] ws = workers;
717	<	if (ws == null || ws.length <= m)         // staleness check
718	<	return false;
719	<	for (int r = w.seed, k = r, j = -(m + m); j <= m + m; ++j) {
720	<	ForkJoinTask<?> t; ForkJoinTask<?>[] q; int b, i;
721	<	ForkJoinWorkerThread v = ws[k & m];
722	<	if (v != null && (b = v.queueBase) != v.queueTop &&
723	<	(q = v.queue) != null && (i = (q.length - 1) & b) >= 0) {
724	<	long u = (i << ASHIFT) + ABASE;
725	<	if ((t = q[i]) != null && v.queueBase == b &&
726	<	UNSAFE.compareAndSwapObject(q, u, t, null)) {
727	<	int d = (v.queueBase = b + 1) - v.queueTop;
728	<	v.stealHint = w.poolIndex;
729	<	if (d != 0)
730	<	signalWork();             // propagate if nonempty
731	<	w.execTask(t);
732	<	}
733	<	r ^= r << 13; r ^= r >>> 17; w.seed = r ^ (r << 5);
734	<	return false;                     // store next seed
735	<	}
736	<	else if (j < 0) {                     // xorshift
737	<	r ^= r << 13; r ^= r >>> 17; k = r ^= r << 5;
738	<	}
739	<	else
740	<	++k;
741	<	}
742	<	if (scanGuard != g)                       // staleness check
743	<	return false;
744	<	else {                                    // try to take submission
745	<	ForkJoinTask<?> t; ForkJoinTask<?>[] q; int b, i;
746	<	if ((b = queueBase) != queueTop &&
747	<	(q = submissionQueue) != null &&
748	<	(i = (q.length - 1) & b) >= 0) {
749	<	long u = (i << ASHIFT) + ABASE;
750	<	if ((t = q[i]) != null && queueBase == b &&
751	<	UNSAFE.compareAndSwapObject(q, u, t, null)) {
752	<	queueBase = b + 1;
753	<	w.execTask(t);
754	<	}
755	<	return false;
756	<	}
757	<	return true;                         // all queues empty
758	<	}
1207	>	final String nextWorkerName() {
1208	>	return workerNamePrefix.concat
1209	>	(Integer.toString(nextWorkerNumber.addAndGet(1)));
1210		}
1211
1212		/**
1213	<	* Tries to enqueue worker w in wait queue and await change in
1214	<	* worker's eventCount.  If the pool is quiescent and there is
764	<	* more than one worker, possibly terminates worker upon exit.
765	<	* Otherwise, before blocking, rescans queues to avoid missed
766	<	* signals.  Upon finding work, releases at least one worker
767	<	* (which may be the current worker). Rescans restart upon
768	<	* detected staleness or failure to release due to
769	<	* contention. Note the unusual conventions about Thread.interrupt
770	<	* here and elsewhere: Because interrupts are used solely to alert
771	<	* threads to check termination, which is checked here anyway, we
772	<	* clear status (using Thread.interrupted) before any call to
773	<	* park, so that park does not immediately return due to status
774	<	* being set via some other unrelated call to interrupt in user
775	<	* code.
1213	>	* Callback from ForkJoinWorkerThread constructor to establish and
1214	>	* record its WorkQueue.
1215		*
1216	<	* @param w the calling worker
778	<	* @param c the ctl value on entry
779	<	* @return true if waited or another thread was released upon enq
1216	>	* @param wt the worker thread
1217		*/
1218	<	private boolean tryAwaitWork(ForkJoinWorkerThread w, long c) {
1219	<	int v = w.eventCount;
1220	<	w.nextWait = (int)c;                      // w's successor record
1221	<	long nc = (long)(v & E_MASK) | ((c - AC_UNIT) & (AC_MASK|TC_MASK));
1222	<	if (ctl != c || !UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
1223	<	long d = ctl; // return true if lost to a deq, to force scan
1224	<	return (int)d != (int)c && (d & AC_MASK) >= (c & AC_MASK);
1225	<	}
1226	<	for (int sc = w.stealCount; sc != 0;) {   // accumulate stealCount
1227	<	long s = stealCount;
1228	<	if (UNSAFE.compareAndSwapLong(this, stealCountOffset, s, s + sc))
1229	<	sc = w.stealCount = 0;
1230	<	else if (w.eventCount != v)
1231	<	return true;                      // update next time
795	<	}
796	<	if ((!shutdown || !tryTerminate(false)) &&
797	<	(int)c != 0 && 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;
1218	>	final void registerWorker(ForkJoinWorkerThread wt) {
1219	>	WorkQueue w = wt.workQueue;
1220	>	Mutex lock = this.lock;
1221	>	lock.lock();
1222	>	try {
1223	>	int g = growHints, k = g & SMASK;
1224	>	WorkQueue[] ws = workQueues;
1225	>	if (ws != null) {                       // ignore on shutdown
1226	>	int n = ws.length;
1227	>	if ((k & 1) == 0 || k >= n || ws[k] != null) {
1228	>	for (k = 1; k < n && ws[k] != null; k += 2)
1229	>	;                           // workers are at odd indices
1230	>	if (k >= n)                     // resize
1231	>	workQueues = ws = Arrays.copyOf(ws, n << 1);
1232		}
1233	<	LockSupport.park(this);
1234	<	rescanned = w.parked = false;
1233	>	w.eventCount = w.poolIndex = k;     // establish before recording
1234	>	ws[k] = w;
1235	>	growHints = (g & ~SMASK) | ((k + 2) & SMASK);
1236	>	int rs = runState;
1237	>	int m = rs & SMASK;                 // recalculate runState mask
1238	>	if (k > m)
1239	>	m = (m << 1) + 1;
1240	>	runState = (rs & SHUTDOWN) | ((rs + RS_SEQ) & RS_SEQ_MASK) | m;
1241		}
1242	+	} finally {
1243	+	lock.unlock();
1244		}
1245		}
1246
1247		/**
1248	<	* If inactivating worker w has caused pool to become
1249	<	* quiescent, check for pool termination, and wait for event
1250	<	* for up to SHRINK_RATE nanosecs (rescans are unnecessary in
1251	<	* this case because quiescence reflects consensus about lack
844	<	* of work). On timeout, if ctl has not changed, terminate the
845	<	* worker. Upon its termination (see deregisterWorker), it may
846	<	* wake up another worker to possibly repeat this process.
1248	>	* Final callback from terminating worker, as well as upon failure
1249	>	* to construct or start a worker in addWorker.  Removes record of
1250	>	* worker from array, and adjusts counts. If pool is shutting
1251	>	* down, tries to complete termination.
1252		*
1253	<	* @param w the calling worker
1254	<	* @param currentCtl the ctl value after enqueuing w
1255	<	* @param prevCtl the ctl value if w terminated
1256	<	* @param v the eventCount w awaits change
1257	<	*/
1258	<	private void idleAwaitWork(ForkJoinWorkerThread w, long currentCtl,
1259	<	long prevCtl, int v) {
1260	<	if (w.eventCount == v) {
1261	<	if (shutdown)
1262	<	tryTerminate(false);
1263	<	ForkJoinTask.helpExpungeStaleExceptions(); // help clean weak refs
1264	<	while (ctl == currentCtl) {
1265	<	long startTime = System.nanoTime();
1266	<	w.parked = true;
1267	<	if (w.eventCount == v)             // must recheck
1268	<	LockSupport.parkNanos(this, SHRINK_RATE);
864	<	w.parked = false;
865	<	if (w.eventCount != v)
866	<	break;
867	<	else if (System.nanoTime() - startTime <
868	<	SHRINK_RATE - (SHRINK_RATE / 10)) // timing slop
869	<	Thread.interrupted();          // spurious wakeup
870	<	else if (UNSAFE.compareAndSwapLong(this, ctlOffset,
871	<	currentCtl, prevCtl)) {
872	<	w.terminate = true;            // restore previous
873	<	w.eventCount = ((int)currentCtl + EC_UNIT) & E_MASK;
874	<	break;
1253	>	* @param wt the worker thread or null if addWorker failed
1254	>	* @param ex the exception causing failure, or null if none
1255	>	*/
1256	>	final void deregisterWorker(ForkJoinWorkerThread wt, Throwable ex) {
1257	>	WorkQueue w = null;
1258	>	if (wt != null && (w = wt.workQueue) != null) {
1259	>	w.runState = -1;                // ensure runState is set
1260	>	stealCount.getAndAdd(w.totalSteals + w.nsteals);
1261	>	int idx = w.poolIndex;
1262	>	Mutex lock = this.lock;
1263	>	lock.lock();
1264	>	try {                           // remove record from array
1265	>	WorkQueue[] ws = workQueues;
1266	>	if (ws != null && idx >= 0 && idx < ws.length && ws[idx] == w) {
1267	>	ws[idx] = null;
1268	>	growHints = (growHints & ~SMASK) | idx;
1269		}
1270	+	} finally {
1271	+	lock.unlock();
1272		}
1273		}
878	–	}
1274
1275	<	// Submissions
1275	>	long c;                             // adjust ctl counts
1276	>	do {} while (!U.compareAndSwapLong
1277	>	(this, CTL, c = ctl, (((c - AC_UNIT) & AC_MASK) |
1278	>	((c - TC_UNIT) & TC_MASK) |
1279	>	(c & ~(AC_MASK|TC_MASK)))));
1280
1281	<	/**
1282	<	* Enqueues the given task in the submissionQueue.  Same idea as
1283	<	* ForkJoinWorkerThread.pushTask except for use of submissionLock.
1284	<	*
1285	<	* @param t the task
1286	<	*/
888	<	private void addSubmission(ForkJoinTask<?> t) {
889	<	final ReentrantLock lock = this.submissionLock;
890	<	lock.lock();
891	<	try {
892	<	ForkJoinTask<?>[] q; int s, m;
893	<	if ((q = submissionQueue) != null) {    // ignore if queue removed
894	<	long u = (((s = queueTop) & (m = q.length-1)) << ASHIFT)+ABASE;
895	<	UNSAFE.putOrderedObject(q, u, t);
896	<	queueTop = s + 1;
897	<	if (s - queueBase == m)
898	<	growSubmissionQueue();
899	<	}
900	<	} finally {
901	<	lock.unlock();
1281	>	if (!tryTerminate(false, false) && w != null) {
1282	>	w.cancelAll();                  // cancel remaining tasks
1283	>	if (w.array != null)            // suppress signal if never ran
1284	>	signalWork();               // wake up or create replacement
1285	>	if (ex == null)                 // help clean refs on way out
1286	>	ForkJoinTask.helpExpungeStaleExceptions();
1287		}
903	–	signalWork();
904	–	}
1288
1289	<	//  (pollSubmission is defined below with exported methods)
1289	>	if (ex != null)                     // rethrow
1290	>	U.throwException(ex);
1291	>	}
1292
1293		/**
1294	<	* Creates or doubles submissionQueue array.
910	<	* Basically identical to ForkJoinWorkerThread version.
1294	>	* Top-level runloop for workers, called by ForkJoinWorkerThread.run.
1295		*/
1296	<	private void growSubmissionQueue() {
1297	<	ForkJoinTask<?>[] oldQ = submissionQueue;
1298	<	int size = oldQ != null ? oldQ.length << 1 : INITIAL_QUEUE_CAPACITY;
1299	<	if (size > MAXIMUM_QUEUE_CAPACITY)
1300	<	throw new RejectedExecutionException("Queue capacity exceeded");
1301	<	if (size < INITIAL_QUEUE_CAPACITY)
1302	<	size = INITIAL_QUEUE_CAPACITY;
919	<	ForkJoinTask<?>[] q = submissionQueue = new ForkJoinTask<?>[size];
920	<	int mask = size - 1;
921	<	int top = queueTop;
922	<	int oldMask;
923	<	if (oldQ != null && (oldMask = oldQ.length - 1) >= 0) {
924	<	for (int b = queueBase; b != top; ++b) {
925	<	long u = ((b & oldMask) << ASHIFT) + ABASE;
926	<	Object x = UNSAFE.getObjectVolatile(oldQ, u);
927	<	if (x != null && UNSAFE.compareAndSwapObject(oldQ, u, x, null))
928	<	UNSAFE.putObjectVolatile
929	<	(q, ((b & mask) << ASHIFT) + ABASE, x);
930	<	}
931	<	}
1296	>	final void runWorker(ForkJoinWorkerThread wt) {
1297	>	// Initialize queue array and seed in this thread
1298	>	WorkQueue w = wt.workQueue;
1299	>	w.growArray(false);
1300	>	w.seed = hashId(Thread.currentThread().getId());
1301	>
1302	>	do {} while (w.runTask(scan(w)));
1303		}
1304
1305	<	// Blocking support
1305	>	// Submissions
1306
1307		/**
1308	<	* Tries to increment blockedCount, decrement active count
1309	<	* (sometimes implicitly) and possibly release or create a
1310	<	* compensating worker in preparation for blocking. Fails
1311	<	* on contention or termination.
1312	<	*
1313	<	* @return true if the caller can block, else should recheck and retry
1308	>	* Unless shutting down, adds the given task to a submission queue
1309	>	* at submitter's current queue index (modulo submission
1310	>	* range). If no queue exists at the index, one is created unless
1311	>	* pool lock is busy.  If the queue and/or lock are busy, another
1312	>	* index is randomly chosen. The mask in growHints controls the
1313	>	* effective index range of queues considered. The mask is
1314	>	* expanded, up to the current workerQueue mask, upon any detected
1315	>	* contention but otherwise remains small to avoid needlessly
1316	>	* creating queues when there is no contention.
1317		*/
1318	<	private boolean tryPreBlock() {
1319	<	int b = blockedCount;
1320	<	if (UNSAFE.compareAndSwapInt(this, blockedCountOffset, b, b + 1)) {
1321	<	int pc = parallelism;
1322	<	do {
1323	<	ForkJoinWorkerThread[] ws; ForkJoinWorkerThread w;
1324	<	int e, ac, tc, i;
1325	<	long c = ctl;
1326	<	int u = (int)(c >>> 32);
1327	<	if ((e = (int)c) < 0) {
1328	<	// skip -- terminating
1329	<	}
1330	<	else if ((ac = (u >> UAC_SHIFT)) <= 0 && e != 0 &&
1331	<	(ws = workers) != null &&
1332	<	(i = ~e & SMASK) < ws.length &&
1333	<	(w = ws[i]) != null) {
960	<	long nc = ((long)(w.nextWait & E_MASK) |
961	<	(c & (AC_MASK|TC_MASK)));
962	<	if (w.eventCount == e &&
963	<	UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
964	<	w.eventCount = (e + EC_UNIT) & E_MASK;
965	<	if (w.parked)
966	<	UNSAFE.unpark(w);
967	<	return true;             // release an idle worker
1318	>	private void doSubmit(ForkJoinTask<?> task) {
1319	>	if (task == null)
1320	>	throw new NullPointerException();
1321	>	Submitter s = submitters.get();
1322	>	for (int r = s.seed, m = growHints >>> 16;;) {
1323	>	WorkQueue[] ws; WorkQueue q; Mutex lk;
1324	>	int k = r & m & SQMASK;          // use only even indices
1325	>	if (runState < 0 || (ws = workQueues) == null || ws.length <= k)
1326	>	throw new RejectedExecutionException(); // shutting down
1327	>	if ((q = ws[k]) == null && (lk = lock).tryAcquire(0)) {
1328	>	try {                        // try to create new queue
1329	>	if (ws == workQueues && (q = ws[k]) == null) {
1330	>	int rs;              // update runState seq
1331	>	ws[k] = q = new WorkQueue(null, SHARED_QUEUE);
1332	>	runState = (((rs = runState) & SHUTDOWN) |
1333	>	((rs + RS_SEQ) & ~SHUTDOWN));
1334		}
1335	+	} finally {
1336	+	lk.unlock();
1337		}
1338	<	else if ((tc = (short)(u >>> UTC_SHIFT)) >= 0 && ac + pc > 1) {
1339	<	long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK);
1340	<	if (UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc))
1341	<	return true;             // no compensation needed
1342	<	}
975	<	else if (tc + pc < MAX_ID) {
976	<	long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK);
977	<	if (UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
978	<	addWorker();
979	<	return true;            // create a replacement
980	<	}
1338	>	}
1339	>	if (q != null) {
1340	>	if (q.trySharedPush(task)) {
1341	>	signalWork();
1342	>	return;
1343		}
1344	<	// try to back out on any failure and let caller retry
1345	<	} while (!UNSAFE.compareAndSwapInt(this, blockedCountOffset,
1346	<	b = blockedCount, b - 1));
1344	>	else if (m < parallelism - 1 && m < (runState & SMASK)) {
1345	>	Mutex lock = this.lock;
1346	>	lock.lock();             // block until lock free
1347	>	int g = growHints;
1348	>	if (g >>> 16 == m)       // expand range
1349	>	growHints = (((m << 1) + 1) << 16) | (g & SMASK);
1350	>	lock.unlock();           // no need for try/finally
1351	>	}
1352	>	else if ((r & m) == 0)
1353	>	Thread.yield();          // occasionally yield if busy
1354	>	}
1355	>	if (m == (m = growHints >>> 16)) {
1356	>	r ^= r << 13;                // update seed unless new range
1357	>	r ^= r >>> 17;               // same xorshift as WorkQueues
1358	>	s.seed = r ^= r << 5;
1359	>	}
1360		}
986	–	return false;
1361		}
1362
1363	+	// Maintaining ctl counts
1364	+
1365		/**
1366	<	* Decrements blockedCount and increments active count.
1366	>	* Increments active count; mainly called upon return from blocking.
1367		*/
1368	<	private void postBlock() {
1368	>	final void incrementActiveCount() {
1369		long c;
1370	<	do {} while (!UNSAFE.compareAndSwapLong(this, ctlOffset,  // no mask
995	<	c = ctl, c + AC_UNIT));
996	<	int b;
997	<	do {} while (!UNSAFE.compareAndSwapInt(this, blockedCountOffset,
998	<	b = blockedCount, b - 1));
1370	>	do {} while (!U.compareAndSwapLong(this, CTL, c = ctl, c + AC_UNIT));
1371		}
1372
1373		/**
1374	<	* Possibly blocks waiting for the given task to complete, or
1003	<	* cancels the task if terminating.  Fails to wait if contended.
1004	<	*
1005	<	* @param joinMe the task
1374	>	* Tries to activate or create a worker if too few are active.
1375		*/
1376	<	final void tryAwaitJoin(ForkJoinTask<?> joinMe) {
1377	<	Thread.interrupted(); // clear interrupts before checking termination
1378	<	if (joinMe.status >= 0) {
1379	<	if (tryPreBlock()) {
1380	<	joinMe.tryAwaitDone(0L);
1381	<	postBlock();
1376	>	final void signalWork() {
1377	>	long c; int u;
1378	>	while ((u = (int)((c = ctl) >>> 32)) < 0) {     // too few active
1379	>	WorkQueue[] ws = workQueues; int e, i; WorkQueue w; Thread p;
1380	>	if ((e = (int)c) > 0) {                     // at least one waiting
1381	>	if (ws != null && (i = e & SMASK) < ws.length &&
1382	>	(w = ws[i]) != null && w.eventCount == (e | INT_SIGN)) {
1383	>	long nc = (((long)(w.nextWait & E_MASK)) |
1384	>	((long)(u + UAC_UNIT) << 32));
1385	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1386	>	w.eventCount = (e + E_SEQ) & E_MASK;
1387	>	if ((p = w.parker) != null)
1388	>	U.unpark(p);                // activate and release
1389	>	break;
1390	>	}
1391	>	}
1392	>	else
1393	>	break;
1394		}
1395	<	else if ((ctl & STOP_BIT) != 0L)
1396	<	joinMe.cancelIgnoringExceptions();
1395	>	else if (e == 0 && (u & SHORT_SIGN) != 0) { // too few total
1396	>	long nc = (long)(((u + UTC_UNIT) & UTC_MASK) |
1397	>	((u + UAC_UNIT) & UAC_MASK)) << 32;
1398	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1399	>	addWorker();
1400	>	break;
1401	>	}
1402	>	}
1403	>	else
1404	>	break;
1405		}
1406		}
1407
1408		/**
1409	<	* Possibly blocks the given worker waiting for joinMe to
1410	<	* complete or timeout.
1409	>	* Tries to decrement active count (sometimes implicitly) and
1410	>	* possibly release or create a compensating worker in preparation
1411	>	* for blocking. Fails on contention or termination.
1412		*
1413	<	* @param joinMe the task
1024	<	* @param millis the wait time for underlying Object.wait
1413	>	* @return true if the caller can block, else should recheck and retry
1414		*/
1415	<	final void timedAwaitJoin(ForkJoinTask<?> joinMe, long nanos) {
1416	<	while (joinMe.status >= 0) {
1417	<	Thread.interrupted();
1418	<	if ((ctl & STOP_BIT) != 0L) {
1419	<	joinMe.cancelIgnoringExceptions();
1420	<	break;
1415	>	final boolean tryCompensate() {
1416	>	WorkQueue w; Thread p;
1417	>	int pc = parallelism, e, u, ac, tc, i;
1418	>	long c = ctl;
1419	>	WorkQueue[] ws = workQueues;
1420	>	if ((e = (int)c) >= 0) {
1421	>	if ((ac = ((u = (int)(c >>> 32)) >> UAC_SHIFT)) <= 0 &&
1422	>	e != 0 && ws != null && (i = e & SMASK) < ws.length &&
1423	>	(w = ws[i]) != null) {
1424	>	long nc = (long)(w.nextWait & E_MASK) | (c & (AC_MASK|TC_MASK));
1425	>	if (w.eventCount == (e | INT_SIGN) &&
1426	>	U.compareAndSwapLong(this, CTL, c, nc)) {
1427	>	w.eventCount = (e + E_SEQ) & E_MASK;
1428	>	if ((p = w.parker) != null)
1429	>	U.unpark(p);
1430	>	return true;             // release an idle worker
1431	>	}
1432		}
1433	<	if (tryPreBlock()) {
1434	<	long last = System.nanoTime();
1435	<	while (joinMe.status >= 0) {
1436	<	long millis = TimeUnit.NANOSECONDS.toMillis(nanos);
1437	<	if (millis <= 0)
1438	<	break;
1439	<	joinMe.tryAwaitDone(millis);
1440	<	if (joinMe.status < 0)
1441	<	break;
1442	<	if ((ctl & STOP_BIT) != 0L) {
1043	<	joinMe.cancelIgnoringExceptions();
1044	<	break;
1045	<	}
1046	<	long now = System.nanoTime();
1047	<	nanos -= now - last;
1048	<	last = now;
1433	>	else if ((tc = (short)(u >>> UTC_SHIFT)) >= 0 && ac + pc > 1) {
1434	>	long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK);
1435	>	if (U.compareAndSwapLong(this, CTL, c, nc))
1436	>	return true;             // no compensation needed
1437	>	}
1438	>	else if (tc + pc < POOL_MAX) {
1439	>	long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK);
1440	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1441	>	addWorker();
1442	>	return true;             // create replacement
1443		}
1050	–	postBlock();
1051	–	break;
1444		}
1445		}
1446	+	return false;
1447		}
1448
1449	+	// Scanning for tasks
1450	+
1451		/**
1452	<	* If necessary, compensates for blocker, and blocks.
1453	<	*/
1454	<	private void awaitBlocker(ManagedBlocker blocker)
1455	<	throws InterruptedException {
1456	<	while (!blocker.isReleasable()) {
1457	<	if (tryPreBlock()) {
1458	<	try {
1459	<	do {} while (!blocker.isReleasable() && !blocker.block());
1460	<	} finally {
1461	<	postBlock();
1452	>	* Scans for and, if found, returns one task, else possibly
1453	>	* inactivates the worker. This method operates on single reads of
1454	>	* volatile state and is designed to be re-invoked continuously in
1455	>	* part because it returns upon detecting inconsistencies,
1456	>	* contention, or state changes that indicate possible success on
1457	>	* re-invocation.
1458	>	*
1459	>	* The scan searches for tasks across queues, randomly selecting
1460	>	* the first #queues probes, favoring steals over submissions
1461	>	* (by exploiting even/odd indexing), and then performing a
1462	>	* circular sweep of all queues.  The scan terminates upon either
1463	>	* finding a non-empty queue, or completing a full sweep. If the
1464	>	* worker is not inactivated, it takes and returns a task from
1465	>	* this queue.  On failure to find a task, we take one of the
1466	>	* following actions, after which the caller will retry calling
1467	>	* this method unless terminated.
1468	>	*
1469	>	* * If pool is terminating, terminate the worker.
1470	>	*
1471	>	* * If not a complete sweep, try to release a waiting worker.  If
1472	>	* the scan terminated because the worker is inactivated, then the
1473	>	* released worker will often be the calling worker, and it can
1474	>	* succeed obtaining a task on the next call. Or maybe it is
1475	>	* another worker, but with same net effect. Releasing in other
1476	>	* cases as well ensures that we have enough workers running.
1477	>	*
1478	>	* * If the caller has run a task since the last empty scan,
1479	>	* return (to allow rescan) if other workers are not also yet
1480	>	* enqueued.  Field WorkQueue.rescans counts down on each scan to
1481	>	* ensure eventual inactivation and blocking.
1482	>	*
1483	>	* * If not already enqueued, try to inactivate and enqueue the
1484	>	* worker on wait queue.
1485	>	*
1486	>	* * If already enqueued and none of the above apply, either park
1487	>	* awaiting signal, or if this is the most recent waiter and pool
1488	>	* is quiescent, relay to idleAwaitWork to check for termination
1489	>	* and possibly shrink pool.
1490	>	*
1491	>	* @param w the worker (via its WorkQueue)
1492	>	* @return a task or null of none found
1493	>	*/
1494	>	private final ForkJoinTask<?> scan(WorkQueue w) {
1495	>	boolean swept = false;               // true after full empty scan
1496	>	WorkQueue[] ws;                      // volatile read order matters
1497	>	int r = w.seed, ec = w.eventCount;   // ec is negative if inactive
1498	>	int rs = runState, m = rs & SMASK;
1499	>	if ((ws = workQueues) != null && ws.length > m) { // consistency check
1500	>	for (int k = 0, j = -1 - m; ; ++j) {
1501	>	WorkQueue q; int b;
1502	>	if (j < 0) {                 // random probes while j negative
1503	>	r ^= r << 13; r ^= r >>> 17; k = (r ^= r << 5) | (j & 1);
1504	>	}                            // worker (not submit) for odd j
1505	>	else                         // cyclic scan when j >= 0
1506	>	k += 7;                  // step 7 reduces array packing bias
1507	>	if ((q = ws[k & m]) != null && (b = q.base) - q.top < 0) {
1508	>	ForkJoinTask<?> t = (ec >= 0) ? q.pollAt(b) : null;
1509	>	w.seed = r;              // save seed for next scan
1510	>	if (t != null)
1511	>	return t;
1512	>	break;
1513	>	}
1514	>	else if (j - m > m) {
1515	>	if (rs == runState)      // staleness check
1516	>	swept = true;
1517	>	break;
1518		}
1068	–	break;
1519		}
1070	–	}
1071	–	}
1072	–
1073	–	// Creating, registering and deregistring workers
1520
1521	<	/**
1522	<	* Tries to create and start a worker; minimally rolls back counts
1523	<	* on failure.
1524	<	*/
1525	<	private void addWorker() {
1526	<	Throwable ex = null;
1527	<	ForkJoinWorkerThread t = null;
1528	<	try {
1529	<	t = factory.newThread(this);
1530	<	} catch (Throwable e) {
1531	<	ex = e;
1532	<	}
1533	<	if (t == null) {  // null or exceptional factory return
1534	<	long c;       // adjust counts
1535	<	do {} while (!UNSAFE.compareAndSwapLong
1536	<	(this, ctlOffset, c = ctl,
1537	<	(((c - AC_UNIT) & AC_MASK) |
1538	<	((c - TC_UNIT) & TC_MASK) |
1539	<	(c & ~(AC_MASK|TC_MASK)))));
1540	<	// Propagate exception if originating from an external caller
1541	<	if (!tryTerminate(false) && ex != null &&
1542	<	!(Thread.currentThread() instanceof ForkJoinWorkerThread))
1543	<	UNSAFE.throwException(ex);
1521	>	// Decode ctl on empty scan
1522	>	long c = ctl; int e = (int)c, a = (int)(c >> AC_SHIFT), nr, ns;
1523	>	if (e < 0)                       // pool is terminating
1524	>	w.runState = -1;
1525	>	else if (!swept) {               // try to release a waiter
1526	>	WorkQueue v; Thread p;
1527	>	if (e > 0 && a < 0 && (v = ws[e & m]) != null &&
1528	>	v.eventCount == (e | INT_SIGN)) {
1529	>	long nc = ((long)(v.nextWait & E_MASK) |
1530	>	((c + AC_UNIT) & (AC_MASK|TC_MASK)));
1531	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1532	>	v.eventCount = (e + E_SEQ) & E_MASK;
1533	>	if ((p = v.parker) != null)
1534	>	U.unpark(p);
1535	>	}
1536	>	}
1537	>	}
1538	>	else if ((nr = w.rescans) > 0) { // continue rescanning
1539	>	int ac = a + parallelism;
1540	>	if (((w.rescans = (ac < nr) ? ac : nr - 1) & 3) == 0 &&
1541	>	w.eventCount == ec)
1542	>	Thread.yield();          // occasionally yield
1543	>	}
1544	>	else if (ec >= 0) {              // try to enqueue
1545	>	long nc = (long)ec | ((c - AC_UNIT) & (AC_MASK|TC_MASK));
1546	>	w.nextWait = e;
1547	>	w.eventCount = ec | INT_SIGN;// mark as inactive
1548	>	if (!U.compareAndSwapLong(this, CTL, c, nc))
1549	>	w.eventCount = ec;       // unmark on CAS failure
1550	>	else if ((ns = w.nsteals) != 0) {
1551	>	w.nsteals = 0;           // set rescans if ran task
1552	>	w.rescans = a + parallelism;
1553	>	w.totalSteals += ns;
1554	>	}
1555	>	}
1556	>	else{                            // already queued
1557	>	if (parallelism == -a)
1558	>	idleAwaitWork(w);        // quiescent
1559	>	if (w.eventCount == ec) {
1560	>	Thread.interrupted();    // clear status
1561	>	ForkJoinWorkerThread wt = w.owner;
1562	>	U.putObject(wt, PARKBLOCKER, this);
1563	>	w.parker = wt;           // emulate LockSupport.park
1564	>	if (w.eventCount == ec)  // recheck
1565	>	U.park(false, 0L);   // block
1566	>	w.parker = null;
1567	>	U.putObject(wt, PARKBLOCKER, null);
1568	>	}
1569	>	}
1570		}
1571	<	else
1100	<	t.start();
1571	>	return null;
1572		}
1573
1574		/**
1575	<	* Callback from ForkJoinWorkerThread constructor to assign a
1576	<	* public name
1575	>	* If inactivating worker w has caused pool to become quiescent,
1576	>	* checks for pool termination, and, so long as this is not the
1577	>	* only worker, waits for event for up to SHRINK_RATE nanosecs.
1578	>	* On timeout, if ctl has not changed, terminates the worker,
1579	>	* which will in turn wake up another worker to possibly repeat
1580	>	* this process.
1581	>	*
1582	>	* @param w the calling worker
1583		*/
1584	<	final String nextWorkerName() {
1585	<	for (int n;;) {
1586	<	if (UNSAFE.compareAndSwapInt(this, nextWorkerNumberOffset,
1587	<	n = nextWorkerNumber, ++n))
1588	<	return workerNamePrefix + n;
1584	>	private void idleAwaitWork(WorkQueue w) {
1585	>	long c; int nw, ec;
1586	>	if (!tryTerminate(false, false) &&
1587	>	(int)((c = ctl) >> AC_SHIFT) + parallelism == 0 &&
1588	>	(ec = w.eventCount) == ((int)c | INT_SIGN) &&
1589	>	(nw = w.nextWait) != 0) {
1590	>	long nc = ((long)(nw & E_MASK) | // ctl to restore on timeout
1591	>	((c + AC_UNIT) & AC_MASK) | (c & TC_MASK));
1592	>	ForkJoinWorkerThread wt = w.owner;
1593	>	while (ctl == c) {
1594	>	long startTime = System.nanoTime();
1595	>	Thread.interrupted();  // timed variant of version in scan()
1596	>	U.putObject(wt, PARKBLOCKER, this);
1597	>	w.parker = wt;
1598	>	if (ctl == c)
1599	>	U.park(false, SHRINK_RATE);
1600	>	w.parker = null;
1601	>	U.putObject(wt, PARKBLOCKER, null);
1602	>	if (ctl != c)
1603	>	break;
1604	>	if (System.nanoTime() - startTime >= SHRINK_TIMEOUT &&
1605	>	U.compareAndSwapLong(this, CTL, c, nc)) {
1606	>	w.eventCount = (ec + E_SEQ) | E_MASK;
1607	>	w.runState = -1;          // shrink
1608	>	break;
1609	>	}
1610	>	}
1611		}
1612		}
1613
1614		/**
1615	<	* Callback from ForkJoinWorkerThread constructor to
1616	<	* determine its poolIndex and record in workers array.
1617	<	*
1618	<	* @param w the worker
1619	<	* @return the worker's pool index
1620	<	*/
1621	<	final int registerWorker(ForkJoinWorkerThread w) {
1622	<	/*
1623	<	* In the typical case, a new worker acquires the lock, uses
1624	<	* next available index and returns quickly.  Since we should
1625	<	* not block callers (ultimately from signalWork or
1626	<	* tryPreBlock) waiting for the lock needed to do this, we
1627	<	* instead help release other workers while waiting for the
1628	<	* lock.
1629	<	*/
1630	<	for (int g;;) {
1631	<	ForkJoinWorkerThread[] ws;
1632	<	if (((g = scanGuard) & SG_UNIT) == 0 &&
1633	<	UNSAFE.compareAndSwapInt(this, scanGuardOffset,
1634	<	g, g | SG_UNIT)) {
1635	<	int k = nextWorkerIndex;
1636	<	try {
1637	<	if ((ws = workers) != null) { // ignore on shutdown
1638	<	int n = ws.length;
1639	<	if (k < 0 || k >= n || ws[k] != null) {
1640	<	for (k = 0; k < n && ws[k] != null; ++k)
1641	<	;
1642	<	if (k == n)
1643	<	ws = workers = Arrays.copyOf(ws, n << 1);
1615	>	* Tries to locate and execute tasks for a stealer of the given
1616	>	* task, or in turn one of its stealers, Traces currentSteal ->
1617	>	* currentJoin links looking for a thread working on a descendant
1618	>	* of the given task and with a non-empty queue to steal back and
1619	>	* execute tasks from. The first call to this method upon a
1620	>	* waiting join will often entail scanning/search, (which is OK
1621	>	* because the joiner has nothing better to do), but this method
1622	>	* leaves hints in workers to speed up subsequent calls. The
1623	>	* implementation is very branchy to cope with potential
1624	>	* inconsistencies or loops encountering chains that are stale,
1625	>	* unknown, or of length greater than MAX_HELP_DEPTH links.  All
1626	>	* of these cases are dealt with by just retrying by caller.
1627	>	*
1628	>	* @param joiner the joining worker
1629	>	* @param task the task to join
1630	>	* @return true if found or ran a task (and so is immediately retryable)
1631	>	*/
1632	>	final boolean tryHelpStealer(WorkQueue joiner, ForkJoinTask<?> task) {
1633	>	ForkJoinTask<?> subtask;    // current target
1634	>	boolean progress = false;
1635	>	int depth = 0;              // current chain depth
1636	>	int m = runState & SMASK;
1637	>	WorkQueue[] ws = workQueues;
1638	>
1639	>	if (ws != null && ws.length > m && (subtask = task).status >= 0) {
1640	>	outer:for (WorkQueue j = joiner;;) {
1641	>	// Try to find the stealer of subtask, by first using hint
1642	>	WorkQueue stealer = null;
1643	>	WorkQueue v = ws[j.stealHint & m];
1644	>	if (v != null && v.currentSteal == subtask)
1645	>	stealer = v;
1646	>	else {
1647	>	for (int i = 1; i <= m; i += 2) {
1648	>	if ((v = ws[i]) != null && v.currentSteal == subtask) {
1649	>	stealer = v;
1650	>	j.stealHint = i; // save hint
1651	>	break;
1652		}
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);
1653		}
1654	<	} finally {
1655	<	scanGuard = g;
1654	>	if (stealer == null)
1655	>	break;
1656		}
1657	<	return k;
1658	<	}
1659	<	else if ((ws = workers) != null) { // help release others
1660	<	for (ForkJoinWorkerThread u : ws) {
1661	<	if (u != null && u.queueBase != u.queueTop) {
1662	<	if (tryReleaseWaiter())
1663	<	break;
1657	>
1658	>	for (WorkQueue q = stealer;;) { // Try to help stealer
1659	>	ForkJoinTask<?> t; int b;
1660	>	if (task.status < 0)
1661	>	break outer;
1662	>	if ((b = q.base) - q.top < 0) {
1663	>	progress = true;
1664	>	if (subtask.status < 0)
1665	>	break outer;               // stale
1666	>	if ((t = q.pollAt(b)) != null) {
1667	>	stealer.stealHint = joiner.poolIndex;
1668	>	joiner.runSubtask(t);
1669	>	}
1670	>	}
1671	>	else { // empty - try to descend to find stealer's stealer
1672	>	ForkJoinTask<?> next = stealer.currentJoin;
1673	>	if (++depth == MAX_HELP_DEPTH || subtask.status < 0 ||
1674	>	next == null || next == subtask)
1675	>	break outer;  // max depth, stale, dead-end, cyclic
1676	>	subtask = next;
1677	>	j = stealer;
1678	>	break;
1679		}
1680		}
1681		}
1682		}
1683	+	return progress;
1684		}
1685
1686		/**
1687	<	* 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.
1687	>	* If task is at base of some steal queue, steals and executes it.
1688		*
1689	<	* @param w the worker
1690	<	*/
1174	<	final void deregisterWorker(ForkJoinWorkerThread w, Throwable ex) {
1175	<	int idx = w.poolIndex;
1176	<	int sc = w.stealCount;
1177	<	int steps = 0;
1178	<	// Remove from array, adjust worker counts and collect steal count.
1179	<	// 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();
1210	<	}
1211	<	}
1212	<
1213	<	// Shutdown and termination
1214	<
1215	<	/**
1216	<	* Possibly initiates and/or completes termination.
1217	<	*
1218	<	* @param now if true, unconditionally terminate, else only
1219	<	* if shutdown and empty queue and no active workers
1220	<	* @return true if now terminating or terminated
1689	>	* @param joiner the joining worker
1690	>	* @param task the task
1691		*/
1692	<	private boolean tryTerminate(boolean now) {
1693	<	long c;
1694	<	while (((c = ctl) & STOP_BIT) == 0) {
1695	<	if (!now) {
1696	<	if ((int)(c >> AC_SHIFT) != -parallelism)
1697	<	return false;
1698	<	if (!shutdown || blockedCount != 0 || quiescerCount != 0 ||
1699	<	queueBase != queueTop) {
1700	<	if (ctl == c) // staleness check
1231	<	return false;
1232	<	continue;
1692	>	final void tryPollForAndExec(WorkQueue joiner, ForkJoinTask<?> task) {
1693	>	WorkQueue[] ws;
1694	>	int m = runState & SMASK;
1695	>	if ((ws = workQueues) != null && ws.length > m) {
1696	>	for (int j = 1; j <= m && task.status >= 0; j += 2) {
1697	>	WorkQueue q = ws[j];
1698	>	if (q != null && q.pollFor(task)) {
1699	>	joiner.runSubtask(task);
1700	>	break;
1701		}
1702		}
1235	–	if (UNSAFE.compareAndSwapLong(this, ctlOffset, c, c | STOP_BIT))
1236	–	startTerminating();
1703		}
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;
1704		}
1705
1706		/**
1707	<	* Runs up to three passes through workers: (0) Setting
1708	<	* termination status for each worker, followed by wakeups up to
1709	<	* queued workers; (1) helping cancel tasks; (2) interrupting
1710	<	* lagging threads (likely in external tasks, but possibly also
1711	<	* blocked in joins).  Each pass repeats previous steps because of
1712	<	* potential lagging thread creation.
1713	<	*/
1714	<	private void startTerminating() {
1715	<	cancelSubmissions();
1716	<	for (int pass = 0; pass < 3; ++pass) {
1717	<	ForkJoinWorkerThread[] ws = workers;
1718	<	if (ws != null) {
1719	<	for (ForkJoinWorkerThread w : ws) {
1720	<	if (w != null) {
1721	<	w.terminate = true;
1266	<	if (pass > 0) {
1267	<	w.cancelTasks();
1268	<	if (pass > 1 && !w.isInterrupted()) {
1269	<	try {
1270	<	w.interrupt();
1271	<	} catch (SecurityException ignore) {
1272	<	}
1273	<	}
1274	<	}
1707	>	* Returns a non-empty steal queue, if one is found during a random,
1708	>	* then cyclic scan, else null.  This method must be retried by
1709	>	* caller if, by the time it tries to use the queue, it is empty.
1710	>	*/
1711	>	private WorkQueue findNonEmptyStealQueue(WorkQueue w) {
1712	>	int r = w.seed;    // Same idea as scan(), but ignoring submissions
1713	>	for (WorkQueue[] ws;;) {
1714	>	int m = runState & SMASK;
1715	>	if ((ws = workQueues) == null)
1716	>	return null;
1717	>	if (ws.length > m) {
1718	>	WorkQueue q;
1719	>	for (int k = 0, j = -1 - m;; ++j) {
1720	>	if (j < 0) {
1721	>	r ^= r << 13; r ^= r >>> 17; k = r ^= r << 5;
1722		}
1723	+	else
1724	+	k += 7;
1725	+	if ((q = ws[(k | 1) & m]) != null && q.base - q.top < 0) {
1726	+	w.seed = r;
1727	+	return q;
1728	+	}
1729	+	else if (j - m > m)
1730	+	return null;
1731		}
1277	–	terminateWaiters();
1732		}
1733		}
1734		}
1735
1736		/**
1737	<	* Polls and cancels all submissions. Called only during termination.
1738	<	*/
1739	<	private void cancelSubmissions() {
1740	<	while (queueBase != queueTop) {
1741	<	ForkJoinTask<?> task = pollSubmission();
1742	<	if (task != null) {
1743	<	try {
1744	<	task.cancel(false);
1745	<	} catch (Throwable ignore) {
1737	>	* Runs tasks until {@code isQuiescent()}. We piggyback on
1738	>	* active count ctl maintenance, but rather than blocking
1739	>	* when tasks cannot be found, we rescan until all others cannot
1740	>	* find tasks either.
1741	>	*/
1742	>	final void helpQuiescePool(WorkQueue w) {
1743	>	for (boolean active = true;;) {
1744	>	w.runLocalTasks();      // exhaust local queue
1745	>	WorkQueue q = findNonEmptyStealQueue(w);
1746	>	if (q != null) {
1747	>	ForkJoinTask<?> t;
1748	>	if (!active) {      // re-establish active count
1749	>	long c;
1750	>	active = true;
1751	>	do {} while (!U.compareAndSwapLong
1752	>	(this, CTL, c = ctl, c + AC_UNIT));
1753		}
1754	+	if ((t = q.poll()) != null)
1755	+	w.runSubtask(t);
1756		}
1757	<	}
1758	<	}
1759	<
1760	<	/**
1761	<	* Tries to set the termination status of waiting workers, and
1762	<	* then wakes them up (after which they will terminate).
1763	<	*/
1764	<	private void terminateWaiters() {
1765	<	ForkJoinWorkerThread[] ws = workers;
1766	<	if (ws != null) {
1767	<	ForkJoinWorkerThread w; long c; int i, e;
1768	<	int n = ws.length;
1769	<	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);
1757	>	else {
1758	>	long c;
1759	>	if (active) {       // decrement active count without queuing
1760	>	active = false;
1761	>	do {} while (!U.compareAndSwapLong
1762	>	(this, CTL, c = ctl, c -= AC_UNIT));
1763	>	}
1764	>	else
1765	>	c = ctl;        // re-increment on exit
1766	>	if ((int)(c >> AC_SHIFT) + parallelism == 0) {
1767	>	do {} while (!U.compareAndSwapLong
1768	>	(this, CTL, c = ctl, c + AC_UNIT));
1769	>	break;
1770		}
1771		}
1772		}
1773		}
1774
1321	–	// misc ForkJoinWorkerThread support
1322	–
1775		/**
1776	<	* Increments or decrements quiescerCount. Needed only to prevent
1325	<	* triggering shutdown if a worker is transiently inactive while
1326	<	* checking quiescence.
1776	>	* Gets and removes a local or stolen task for the given worker.
1777		*
1778	<	* @param delta 1 for increment, -1 for decrement
1778	>	* @return a task, if available
1779		*/
1780	<	final void addQuiescerCount(int delta) {
1781	<	int c;
1782	<	do {} while (!UNSAFE.compareAndSwapInt(this, quiescerCountOffset,
1783	<	c = quiescerCount, c + delta));
1784	<	}
1785	<
1786	<	/**
1787	<	* Directly increments or decrements active count without queuing.
1788	<	* This method is used to transiently assert inactivation while
1789	<	* checking quiescence.
1340	<	*
1341	<	* @param delta 1 for increment, -1 for decrement
1342	<	*/
1343	<	final void addActiveCount(int delta) {
1344	<	long d = (long)delta << AC_SHIFT;
1345	<	long c;
1346	<	do {} while (!UNSAFE.compareAndSwapLong(this, ctlOffset,
1347	<	c = ctl, c + d));
1780	>	final ForkJoinTask<?> nextTaskFor(WorkQueue w) {
1781	>	for (ForkJoinTask<?> t;;) {
1782	>	WorkQueue q;
1783	>	if ((t = w.nextLocalTask()) != null)
1784	>	return t;
1785	>	if ((q = findNonEmptyStealQueue(w)) == null)
1786	>	return null;
1787	>	if ((t = q.poll()) != null)
1788	>	return t;
1789	>	}
1790		}
1791
1792		/**
1793		* Returns the approximate (non-atomic) number of idle threads per
1794	<	* active thread.
1794	>	* active thread to offset steal queue size for method
1795	>	* ForkJoinTask.getSurplusQueuedTaskCount().
1796		*/
1797		final int idlePerActive() {
1798		// Approximate at powers of two for small values, saturate past 4
#	Line 1362 | Line 1805 | public class ForkJoinPool extends Abstra
1805		8);
1806		}
1807
1808	+	//  Termination
1809	+
1810	+	/**
1811	+	* Possibly initiates and/or completes termination.  The caller
1812	+	* triggering termination runs three passes through workQueues:
1813	+	* (0) Setting termination status, followed by wakeups of queued
1814	+	* workers; (1) cancelling all tasks; (2) interrupting lagging
1815	+	* threads (likely in external tasks, but possibly also blocked in
1816	+	* joins).  Each pass repeats previous steps because of potential
1817	+	* lagging thread creation.
1818	+	*
1819	+	* @param now if true, unconditionally terminate, else only
1820	+	* if no work and no active workers
1821	+	* @param enable if true, enable shutdown when next possible
1822	+	* @return true if now terminating or terminated
1823	+	*/
1824	+	private boolean tryTerminate(boolean now, boolean enable) {
1825	+	Mutex lock = this.lock;
1826	+	for (long c;;) {
1827	+	if (((c = ctl) & STOP_BIT) != 0) {      // already terminating
1828	+	if ((short)(c >>> TC_SHIFT) == -parallelism) {
1829	+	lock.lock();                    // don't need try/finally
1830	+	termination.signalAll();        // signal when 0 workers
1831	+	lock.unlock();
1832	+	}
1833	+	return true;
1834	+	}
1835	+	if (runState >= 0) {                    // not yet enabled
1836	+	if (!enable)
1837	+	return false;
1838	+	lock.lock();
1839	+	runState |= SHUTDOWN;
1840	+	lock.unlock();
1841	+	}
1842	+	if (!now) {                             // check if idle & no tasks
1843	+	if ((int)(c >> AC_SHIFT) != -parallelism ||
1844	+	hasQueuedSubmissions())
1845	+	return false;
1846	+	// Check for unqueued inactive workers. One pass suffices.
1847	+	WorkQueue[] ws = workQueues; WorkQueue w;
1848	+	if (ws != null) {
1849	+	for (int i = 1; i < ws.length; i += 2) {
1850	+	if ((w = ws[i]) != null && w.eventCount >= 0)
1851	+	return false;
1852	+	}
1853	+	}
1854	+	}
1855	+	if (U.compareAndSwapLong(this, CTL, c, c | STOP_BIT)) {
1856	+	for (int pass = 0; pass < 3; ++pass) {
1857	+	WorkQueue[] ws = workQueues;
1858	+	if (ws != null) {
1859	+	WorkQueue w;
1860	+	int n = ws.length;
1861	+	for (int i = 0; i < n; ++i) {
1862	+	if ((w = ws[i]) != null) {
1863	+	w.runState = -1;
1864	+	if (pass > 0) {
1865	+	w.cancelAll();
1866	+	if (pass > 1)
1867	+	w.interruptOwner();
1868	+	}
1869	+	}
1870	+	}
1871	+	// Wake up workers parked on event queue
1872	+	int i, e; long cc; Thread p;
1873	+	while ((e = (int)(cc = ctl) & E_MASK) != 0 &&
1874	+	(i = e & SMASK) < n &&
1875	+	(w = ws[i]) != null) {
1876	+	long nc = ((long)(w.nextWait & E_MASK) |
1877	+	((cc + AC_UNIT) & AC_MASK) |
1878	+	(cc & (TC_MASK|STOP_BIT)));
1879	+	if (w.eventCount == (e | INT_SIGN) &&
1880	+	U.compareAndSwapLong(this, CTL, cc, nc)) {
1881	+	w.eventCount = (e + E_SEQ) & E_MASK;
1882	+	w.runState = -1;
1883	+	if ((p = w.parker) != null)
1884	+	U.unpark(p);
1885	+	}
1886	+	}
1887	+	}
1888	+	}
1889	+	}
1890	+	}
1891	+	}
1892	+
1893		// Exported methods
1894
1895		// Constructors
#	Line 1431 | Line 1959 | public class ForkJoinPool extends Abstra
1959		checkPermission();
1960		if (factory == null)
1961		throw new NullPointerException();
1962	<	if (parallelism <= 0 || parallelism > MAX_ID)
1962	>	if (parallelism <= 0 || parallelism > POOL_MAX)
1963		throw new IllegalArgumentException();
1964		this.parallelism = parallelism;
1965		this.factory = factory;
1966		this.ueh = handler;
1967	<	this.locallyFifo = asyncMode;
1967	>	this.localMode = asyncMode ? FIFO_QUEUE : LIFO_QUEUE;
1968	>	this.growHints = 1;
1969		long np = (long)(-parallelism); // offset ctl counts
1970		this.ctl = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK);
1971	<	this.submissionQueue = new ForkJoinTask<?>[INITIAL_QUEUE_CAPACITY];
1443	<	// initialize workers array with room for 2*parallelism if possible
1971	>	// initialize workQueues array with room for 2*parallelism if possible
1972		int n = parallelism << 1;
1973	<	if (n >= MAX_ID)
1974	<	n = MAX_ID;
1973	>	if (n >= POOL_MAX)
1974	>	n = POOL_MAX;
1975		else { // See Hackers Delight, sec 3.2, where n < (1 << 16)
1976		n |= n >>> 1; n |= n >>> 2; n |= n >>> 4; n |= n >>> 8;
1977		}
1978	<	workers = new ForkJoinWorkerThread[n + 1];
1979	<	this.submissionLock = new ReentrantLock();
1980	<	this.termination = submissionLock.newCondition();
1978	>	this.workQueues = new WorkQueue[(n + 1) << 1]; // #slots = 2 * #workers
1979	>	this.termination = (this.lock = new Mutex()).newCondition();
1980	>	this.stealCount = new AtomicLong();
1981	>	this.nextWorkerNumber = new AtomicInteger();
1982		StringBuilder sb = new StringBuilder("ForkJoinPool-");
1983		sb.append(poolNumberGenerator.incrementAndGet());
1984		sb.append("-worker-");
#	Line 1475 | Line 2004 | public class ForkJoinPool extends Abstra
2004		*         scheduled for execution
2005		*/
2006		public <T> T invoke(ForkJoinTask<T> task) {
2007	<	Thread t = Thread.currentThread();
2008	<	if (task == null)
1480	<	throw new NullPointerException();
1481	<	if (shutdown)
1482	<	throw new RejectedExecutionException();
1483	<	if ((t instanceof ForkJoinWorkerThread) &&
1484	<	((ForkJoinWorkerThread)t).pool == this)
1485	<	return task.invoke();  // bypass submit if in same pool
1486	<	else {
1487	<	addSubmission(task);
1488	<	return task.join();
1489	<	}
1490	<	}
1491	<
1492	<	/**
1493	<	* Unless terminating, forks task if within an ongoing FJ
1494	<	* computation in the current pool, else submits as external task.
1495	<	*/
1496	<	private <T> void forkOrSubmit(ForkJoinTask<T> task) {
1497	<	ForkJoinWorkerThread w;
1498	<	Thread t = Thread.currentThread();
1499	<	if (shutdown)
1500	<	throw new RejectedExecutionException();
1501	<	if ((t instanceof ForkJoinWorkerThread) &&
1502	<	(w = (ForkJoinWorkerThread)t).pool == this)
1503	<	w.pushTask(task);
1504	<	else
1505	<	addSubmission(task);
2007	>	doSubmit(task);
2008	>	return task.join();
2009		}
2010
2011		/**
#	Line 1514 | Line 2017 | public class ForkJoinPool extends Abstra
2017		*         scheduled for execution
2018		*/
2019		public void execute(ForkJoinTask<?> task) {
2020	<	if (task == null)
1518	<	throw new NullPointerException();
1519	<	forkOrSubmit(task);
2020	>	doSubmit(task);
2021		}
2022
2023		// AbstractExecutorService methods
#	Line 1534 | Line 2035 | public class ForkJoinPool extends Abstra
2035		job = (ForkJoinTask<?>) task;
2036		else
2037		job = ForkJoinTask.adapt(task, null);
2038	<	forkOrSubmit(job);
2038	>	doSubmit(job);
2039		}
2040
2041		/**
#	Line 1547 | Line 2048 | public class ForkJoinPool extends Abstra
2048		*         scheduled for execution
2049		*/
2050		public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) {
2051	<	if (task == null)
1551	<	throw new NullPointerException();
1552	<	forkOrSubmit(task);
2051	>	doSubmit(task);
2052		return task;
2053		}
2054
#	Line 1562 | Line 2061 | public class ForkJoinPool extends Abstra
2061		if (task == null)
2062		throw new NullPointerException();
2063		ForkJoinTask<T> job = ForkJoinTask.adapt(task);
2064	<	forkOrSubmit(job);
2064	>	doSubmit(job);
2065		return job;
2066		}
2067
#	Line 1575 | Line 2074 | public class ForkJoinPool extends Abstra
2074		if (task == null)
2075		throw new NullPointerException();
2076		ForkJoinTask<T> job = ForkJoinTask.adapt(task, result);
2077	<	forkOrSubmit(job);
2077	>	doSubmit(job);
2078		return job;
2079		}
2080
#	Line 1592 | Line 2091 | public class ForkJoinPool extends Abstra
2091		job = (ForkJoinTask<?>) task;
2092		else
2093		job = ForkJoinTask.adapt(task, null);
2094	<	forkOrSubmit(job);
2094	>	doSubmit(job);
2095		return job;
2096		}
2097
#	Line 1601 | Line 2100 | public class ForkJoinPool extends Abstra
2100		* @throws RejectedExecutionException {@inheritDoc}
2101		*/
2102		public <T> List<Future<T>> invokeAll(Collection<? extends Callable<T>> tasks) {
2103	<	ArrayList<ForkJoinTask<T>> forkJoinTasks =
2104	<	new ArrayList<ForkJoinTask<T>>(tasks.size());
2105	<	for (Callable<T> task : tasks)
2106	<	forkJoinTasks.add(ForkJoinTask.adapt(task));
2107	<	invoke(new InvokeAll<T>(forkJoinTasks));
2108	<
2103	>	// In previous versions of this class, this method constructed
2104	>	// a task to run ForkJoinTask.invokeAll, but now external
2105	>	// invocation of multiple tasks is at least as efficient.
2106	>	List<ForkJoinTask<T>> fs = new ArrayList<ForkJoinTask<T>>(tasks.size());
2107	>	// Workaround needed because method wasn't declared with
2108	>	// wildcards in return type but should have been.
2109		@SuppressWarnings({"unchecked", "rawtypes"})
2110	<	List<Future<T>> futures = (List<Future<T>>) (List) forkJoinTasks;
1612	<	return futures;
1613	<	}
2110	>	List<Future<T>> futures = (List<Future<T>>) (List) fs;
2111
2112	<	static final class InvokeAll<T> extends RecursiveAction {
2113	<	final ArrayList<ForkJoinTask<T>> tasks;
2114	<	InvokeAll(ArrayList<ForkJoinTask<T>> tasks) { this.tasks = tasks; }
2115	<	public void compute() {
2116	<	try { invokeAll(tasks); }
2117	<	catch (Exception ignore) {}
2112	>	boolean done = false;
2113	>	try {
2114	>	for (Callable<T> t : tasks) {
2115	>	ForkJoinTask<T> f = ForkJoinTask.adapt(t);
2116	>	doSubmit(f);
2117	>	fs.add(f);
2118	>	}
2119	>	for (ForkJoinTask<T> f : fs)
2120	>	f.quietlyJoin();
2121	>	done = true;
2122	>	return futures;
2123	>	} finally {
2124	>	if (!done)
2125	>	for (ForkJoinTask<T> f : fs)
2126	>	f.cancel(false);
2127		}
1622	–	private static final long serialVersionUID = -7914297376763021607L;
2128		}
2129
2130		/**
#	Line 1669 | Line 2174 | public class ForkJoinPool extends Abstra
2174		* @return {@code true} if this pool uses async mode
2175		*/
2176		public boolean getAsyncMode() {
2177	<	return locallyFifo;
2177	>	return localMode != 0;
2178		}
2179
2180		/**
#	Line 1681 | Line 2186 | public class ForkJoinPool extends Abstra
2186		* @return the number of worker threads
2187		*/
2188		public int getRunningThreadCount() {
2189	<	int r = parallelism + (int)(ctl >> AC_SHIFT);
2190	<	return (r <= 0) ? 0 : r; // suppress momentarily negative values
2189	>	int rc = 0;
2190	>	WorkQueue[] ws; WorkQueue w;
2191	>	if ((ws = workQueues) != null) {
2192	>	for (int i = 1; i < ws.length; i += 2) {
2193	>	if ((w = ws[i]) != null && w.isApparentlyUnblocked())
2194	>	++rc;
2195	>	}
2196	>	}
2197	>	return rc;
2198		}
2199
2200		/**
#	Line 1693 | Line 2205 | public class ForkJoinPool extends Abstra
2205		* @return the number of active threads
2206		*/
2207		public int getActiveThreadCount() {
2208	<	int r = parallelism + (int)(ctl >> AC_SHIFT) + blockedCount;
2208	>	int r = parallelism + (int)(ctl >> AC_SHIFT);
2209		return (r <= 0) ? 0 : r; // suppress momentarily negative values
2210		}
2211
#	Line 1709 | Line 2221 | public class ForkJoinPool extends Abstra
2221		* @return {@code true} if all threads are currently idle
2222		*/
2223		public boolean isQuiescent() {
2224	<	return parallelism + (int)(ctl >> AC_SHIFT) + blockedCount == 0;
2224	>	return (int)(ctl >> AC_SHIFT) + parallelism == 0;
2225		}
2226
2227		/**
#	Line 1724 | Line 2236 | public class ForkJoinPool extends Abstra
2236		* @return the number of steals
2237		*/
2238		public long getStealCount() {
2239	<	return stealCount;
2239	>	long count = stealCount.get();
2240	>	WorkQueue[] ws; WorkQueue w;
2241	>	if ((ws = workQueues) != null) {
2242	>	for (int i = 1; i < ws.length; i += 2) {
2243	>	if ((w = ws[i]) != null)
2244	>	count += w.totalSteals;
2245	>	}
2246	>	}
2247	>	return count;
2248		}
2249
2250		/**
#	Line 1739 | Line 2259 | public class ForkJoinPool extends Abstra
2259		*/
2260		public long getQueuedTaskCount() {
2261		long count = 0;
2262	<	ForkJoinWorkerThread[] ws;
2263	<	if ((short)(ctl >>> TC_SHIFT) > -parallelism &&
2264	<	(ws = workers) != null) {
2265	<	for (ForkJoinWorkerThread w : ws)
2266	<	if (w != null)
2267	<	count -= w.queueBase - w.queueTop; // must read base first
2262	>	WorkQueue[] ws; WorkQueue w;
2263	>	if ((ws = workQueues) != null) {
2264	>	for (int i = 1; i < ws.length; i += 2) {
2265	>	if ((w = ws[i]) != null)
2266	>	count += w.queueSize();
2267	>	}
2268		}
2269		return count;
2270		}
#	Line 1757 | Line 2277 | public class ForkJoinPool extends Abstra
2277		* @return the number of queued submissions
2278		*/
2279		public int getQueuedSubmissionCount() {
2280	<	return -queueBase + queueTop;
2280	>	int count = 0;
2281	>	WorkQueue[] ws; WorkQueue w;
2282	>	if ((ws = workQueues) != null) {
2283	>	for (int i = 0; i < ws.length; i += 2) {
2284	>	if ((w = ws[i]) != null)
2285	>	count += w.queueSize();
2286	>	}
2287	>	}
2288	>	return count;
2289		}
2290
2291		/**
#	Line 1767 | Line 2295 | public class ForkJoinPool extends Abstra
2295		* @return {@code true} if there are any queued submissions
2296		*/
2297		public boolean hasQueuedSubmissions() {
2298	<	return queueBase != queueTop;
2298	>	WorkQueue[] ws; WorkQueue w;
2299	>	if ((ws = workQueues) != null) {
2300	>	for (int i = 0; i < ws.length; i += 2) {
2301	>	if ((w = ws[i]) != null && w.queueSize() != 0)
2302	>	return true;
2303	>	}
2304	>	}
2305	>	return false;
2306		}
2307
2308		/**
#	Line 1778 | Line 2313 | public class ForkJoinPool extends Abstra
2313		* @return the next submission, or {@code null} if none
2314		*/
2315		protected ForkJoinTask<?> pollSubmission() {
2316	<	ForkJoinTask<?> t; ForkJoinTask<?>[] q; int b, i;
2317	<	while ((b = queueBase) != queueTop &&
2318	<	(q = submissionQueue) != null &&
2319	<	(i = (q.length - 1) & b) >= 0) {
2320	<	long u = (i << ASHIFT) + ABASE;
1786	<	if ((t = q[i]) != null &&
1787	<	queueBase == b &&
1788	<	UNSAFE.compareAndSwapObject(q, u, t, null)) {
1789	<	queueBase = b + 1;
1790	<	return t;
2316	>	WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2317	>	if ((ws = workQueues) != null) {
2318	>	for (int i = 0; i < ws.length; i += 2) {
2319	>	if ((w = ws[i]) != null && (t = w.poll()) != null)
2320	>	return t;
2321		}
2322		}
2323		return null;
#	Line 1812 | Line 2342 | public class ForkJoinPool extends Abstra
2342		*/
2343		protected int drainTasksTo(Collection<? super ForkJoinTask<?>> c) {
2344		int count = 0;
2345	<	while (queueBase != queueTop) {
2346	<	ForkJoinTask<?> t = pollSubmission();
2347	<	if (t != null) {
2348	<	c.add(t);
2349	<	++count;
2345	>	WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2346	>	if ((ws = workQueues) != null) {
2347	>	for (int i = 0; i < ws.length; ++i) {
2348	>	if ((w = ws[i]) != null) {
2349	>	while ((t = w.poll()) != null) {
2350	>	c.add(t);
2351	>	++count;
2352	>	}
2353	>	}
2354		}
2355		}
1822	–	ForkJoinWorkerThread[] ws;
1823	–	if ((short)(ctl >>> TC_SHIFT) > -parallelism &&
1824	–	(ws = workers) != null) {
1825	–	for (ForkJoinWorkerThread w : ws)
1826	–	if (w != null)
1827	–	count += w.drainTasksTo(c);
1828	–	}
2356		return count;
2357		}
2358
#	Line 1837 | Line 2364 | public class ForkJoinPool extends Abstra
2364		* @return a string identifying this pool, as well as its state
2365		*/
2366		public String toString() {
2367	<	long st = getStealCount();
2368	<	long qt = getQueuedTaskCount();
2369	<	long qs = getQueuedSubmissionCount();
1843	<	int pc = parallelism;
2367	>	// Use a single pass through workQueues to collect counts
2368	>	long qt = 0L, qs = 0L; int rc = 0;
2369	>	long st = stealCount.get();
2370		long c = ctl;
2371	+	WorkQueue[] ws; WorkQueue w;
2372	+	if ((ws = workQueues) != null) {
2373	+	for (int i = 0; i < ws.length; ++i) {
2374	+	if ((w = ws[i]) != null) {
2375	+	int size = w.queueSize();
2376	+	if ((i & 1) == 0)
2377	+	qs += size;
2378	+	else {
2379	+	qt += size;
2380	+	st += w.totalSteals;
2381	+	if (w.isApparentlyUnblocked())
2382	+	++rc;
2383	+	}
2384	+	}
2385	+	}
2386	+	}
2387	+	int pc = parallelism;
2388		int tc = pc + (short)(c >>> TC_SHIFT);
2389	<	int rc = pc + (int)(c >> AC_SHIFT);
2390	<	if (rc < 0) // ignore transient negative
2391	<	rc = 0;
1849	<	int ac = rc + blockedCount;
2389	>	int ac = pc + (int)(c >> AC_SHIFT);
2390	>	if (ac < 0) // ignore transient negative
2391	>	ac = 0;
2392		String level;
2393		if ((c & STOP_BIT) != 0)
2394		level = (tc == 0) ? "Terminated" : "Terminating";
2395		else
2396	<	level = shutdown ? "Shutting down" : "Running";
2396	>	level = runState < 0 ? "Shutting down" : "Running";
2397		return super.toString() +
2398		"[" + level +
2399		", parallelism = " + pc +
#	Line 1878 | Line 2420 | public class ForkJoinPool extends Abstra
2420		*/
2421		public void shutdown() {
2422		checkPermission();
2423	<	shutdown = true;
1882	<	tryTerminate(false);
2423	>	tryTerminate(false, true);
2424		}
2425
2426		/**
#	Line 1900 | Line 2441 | public class ForkJoinPool extends Abstra
2441		*/
2442		public List<Runnable> shutdownNow() {
2443		checkPermission();
2444	<	shutdown = true;
1904	<	tryTerminate(true);
2444	>	tryTerminate(true, true);
2445		return Collections.emptyList();
2446		}
2447
#	Line 1936 | Line 2476 | public class ForkJoinPool extends Abstra
2476		}
2477
2478		/**
1939	–	* Returns true if terminating or terminated. Used by ForkJoinWorkerThread.
1940	–	*/
1941	–	final boolean isAtLeastTerminating() {
1942	–	return (ctl & STOP_BIT) != 0L;
1943	–	}
1944	–
1945	–	/**
2479		* Returns {@code true} if this pool has been shut down.
2480		*
2481		* @return {@code true} if this pool has been shut down
2482		*/
2483		public boolean isShutdown() {
2484	<	return shutdown;
2484	>	return runState < 0;
2485		}
2486
2487		/**
#	Line 1965 | Line 2498 | public class ForkJoinPool extends Abstra
2498		public boolean awaitTermination(long timeout, TimeUnit unit)
2499		throws InterruptedException {
2500		long nanos = unit.toNanos(timeout);
2501	<	final ReentrantLock lock = this.submissionLock;
2501	>	final Mutex lock = this.lock;
2502		lock.lock();
2503		try {
2504		for (;;) {
#	Line 2076 | Line 2609 | public class ForkJoinPool extends Abstra
2609		public static void managedBlock(ManagedBlocker blocker)
2610		throws InterruptedException {
2611		Thread t = Thread.currentThread();
2612	<	if (t instanceof ForkJoinWorkerThread) {
2613	<	ForkJoinWorkerThread w = (ForkJoinWorkerThread) t;
2614	<	w.pool.awaitBlocker(blocker);
2615	<	}
2616	<	else {
2617	<	do {} while (!blocker.isReleasable() && !blocker.block());
2612	>	ForkJoinPool p = ((t instanceof ForkJoinWorkerThread) ?
2613	>	((ForkJoinWorkerThread)t).pool : null);
2614	>	while (!blocker.isReleasable()) {
2615	>	if (p == null || p.tryCompensate()) {
2616	>	try {
2617	>	do {} while (!blocker.isReleasable() && !blocker.block());
2618	>	} finally {
2619	>	if (p != null)
2620	>	p.incrementActiveCount();
2621	>	}
2622	>	break;
2623	>	}
2624		}
2625		}
2626
#	Line 2098 | Line 2637 | public class ForkJoinPool extends Abstra
2637		}
2638
2639		// Unsafe mechanics
2640	<	private static final sun.misc.Unsafe UNSAFE;
2641	<	private static final long ctlOffset;
2642	<	private static final long stealCountOffset;
2104	<	private static final long blockedCountOffset;
2105	<	private static final long quiescerCountOffset;
2106	<	private static final long scanGuardOffset;
2107	<	private static final long nextWorkerNumberOffset;
2108	<	private static final long ABASE;
2109	<	private static final int ASHIFT;
2640	>	private static final sun.misc.Unsafe U;
2641	>	private static final long CTL;
2642	>	private static final long PARKBLOCKER;
2643
2644		static {
2645		poolNumberGenerator = new AtomicInteger();
2113	–	workerSeedGenerator = new Random();
2646		modifyThreadPermission = new RuntimePermission("modifyThread");
2647		defaultForkJoinWorkerThreadFactory =
2648		new DefaultForkJoinWorkerThreadFactory();
2649	+	submitters = new ThreadSubmitter();
2650		try {
2651	<	UNSAFE = getUnsafe();
2651	>	U = getUnsafe();
2652		Class<?> k = ForkJoinPool.class;
2653	<	ctlOffset = UNSAFE.objectFieldOffset
2653	>	CTL = U.objectFieldOffset
2654		(k.getDeclaredField("ctl"));
2655	<	stealCountOffset = UNSAFE.objectFieldOffset
2656	<	(k.getDeclaredField("stealCount"));
2657	<	blockedCountOffset = UNSAFE.objectFieldOffset
2125	<	(k.getDeclaredField("blockedCount"));
2126	<	quiescerCountOffset = UNSAFE.objectFieldOffset
2127	<	(k.getDeclaredField("quiescerCount"));
2128	<	scanGuardOffset = UNSAFE.objectFieldOffset
2129	<	(k.getDeclaredField("scanGuard"));
2130	<	nextWorkerNumberOffset = UNSAFE.objectFieldOffset
2131	<	(k.getDeclaredField("nextWorkerNumber"));
2655	>	Class<?> tk = Thread.class;
2656	>	PARKBLOCKER = U.objectFieldOffset
2657	>	(tk.getDeclaredField("parkBlocker"));
2658		} catch (Exception e) {
2659		throw new Error(e);
2660		}
2135	–	Class<?> a = ForkJoinTask[].class;
2136	–	ABASE = UNSAFE.arrayBaseOffset(a);
2137	–	int s = UNSAFE.arrayIndexScale(a);
2138	–	if ((s & (s-1)) != 0)
2139	–	throw new Error("data type scale not a power of two");
2140	–	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
2661		}
2662
2663		/**
#	Line 2167 | Line 2687 | public class ForkJoinPool extends Abstra
2687		}
2688		}
2689		}
2690	+
2691		}

Diff Legend

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