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 1.106 by jsr166, Fri Jul 1 01:15:06 2011 UTC vs.
Revision 1.130 by jsr166, Mon Aug 13 18:25:53 2012 UTC

#	Line 5 | Line 5
5		*/
6
7		package jsr166y;
8	–
8		import java.util.ArrayList;
9		import java.util.Arrays;
10		import java.util.Collection;
#	Line 20 | Line 19 | import java.util.concurrent.RejectedExec
19		import java.util.concurrent.RunnableFuture;
20		import java.util.concurrent.TimeUnit;
21		import java.util.concurrent.atomic.AtomicInteger;
22	<	import java.util.concurrent.locks.LockSupport;
23	<	import java.util.concurrent.locks.ReentrantLock;
22	>	import java.util.concurrent.atomic.AtomicLong;
23	>	import java.util.concurrent.locks.AbstractQueuedSynchronizer;
24		import java.util.concurrent.locks.Condition;
25
26		/**
#	Line 33 | Line 32 | import java.util.concurrent.locks.Condit
32		* <p>A {@code ForkJoinPool} differs from other kinds of {@link
33		* ExecutorService} mainly by virtue of employing
34		* <em>work-stealing</em>: all threads in the pool attempt to find and
35	<	* execute subtasks created by other active tasks (eventually blocking
36	<	* waiting for work if none exist). This enables efficient processing
37	<	* when most tasks spawn other subtasks (as do most {@code
38	<	* ForkJoinTask}s). When setting <em>asyncMode</em> to true in
39	<	* constructors, {@code ForkJoinPool}s may also be appropriate for use
40	<	* with event-style tasks that are never joined.
35	>	* execute tasks submitted to the pool and/or created by other active
36	>	* tasks (eventually blocking waiting for work if none exist). This
37	>	* enables efficient processing when most tasks spawn other subtasks
38	>	* (as do most {@code ForkJoinTask}s), as well as when many small
39	>	* tasks are submitted to the pool from external clients.  Especially
40	>	* when setting <em>asyncMode</em> to true in constructors, {@code
41	>	* ForkJoinPool}s may also be appropriate for use with event-style
42	>	* tasks that are never joined.
43		*
44		* <p>A {@code ForkJoinPool} is constructed with a given target
45		* parallelism level; by default, equal to the number of available
#	Line 58 | Line 59 | import java.util.concurrent.locks.Condit
59		* convenient form for informal monitoring.
60		*
61		* <p> As is the case with other ExecutorServices, there are three
62	<	* main task execution methods summarized in the following
63	<	* table. These are designed to be used by clients not already engaged
64	<	* in fork/join computations in the current pool.  The main forms of
65	<	* these methods accept instances of {@code ForkJoinTask}, but
66	<	* overloaded forms also allow mixed execution of plain {@code
62	>	* main task execution methods summarized in the following table.
63	>	* These are designed to be used primarily by clients not already
64	>	* engaged in fork/join computations in the current pool.  The main
65	>	* forms of these methods accept instances of {@code ForkJoinTask},
66	>	* but overloaded forms also allow mixed execution of plain {@code
67		* Runnable}- or {@code Callable}- based activities as well.  However,
68	<	* tasks that are already executing in a pool should normally
69	<	* <em>NOT</em> use these pool execution methods, but instead use the
70	<	* within-computation forms listed in the table.
68	>	* tasks that are already executing in a pool should normally instead
69	>	* use the within-computation forms listed in the table unless using
70	>	* async event-style tasks that are not usually joined, in which case
71	>	* there is little difference among choice of methods.
72		*
73		* <table BORDER CELLPADDING=3 CELLSPACING=1>
74		*  <tr>
#	Line 125 | Line 127 | public class ForkJoinPool extends Abstra
127		/*
128		* Implementation Overview
129		*
130	<	* This class provides the central bookkeeping and control for a
131	<	* set of worker threads: Submissions from non-FJ threads enter
132	<	* into a submission queue. Workers take these tasks and typically
133	<	* split them into subtasks that may be stolen by other workers.
134	<	* Preference rules give first priority to processing tasks from
135	<	* their own queues (LIFO or FIFO, depending on mode), then to
136	<	* randomized FIFO steals of tasks in other worker queues, and
137	<	* lastly to new submissions.
130	>	* This class and its nested classes provide the main
131	>	* functionality and control for a set of worker threads:
132	>	* Submissions from non-FJ threads enter into submission queues.
133	>	* Workers take these tasks and typically split them into subtasks
134	>	* that may be stolen by other workers.  Preference rules give
135	>	* first priority to processing tasks from their own queues (LIFO
136	>	* or FIFO, depending on mode), then to randomized FIFO steals of
137	>	* tasks in other queues.
138	>	*
139	>	* WorkQueues
140	>	* ==========
141	>	*
142	>	* Most operations occur within work-stealing queues (in nested
143	>	* class WorkQueue).  These are special forms of Deques that
144	>	* support only three of the four possible end-operations -- push,
145	>	* pop, and poll (aka steal), under the further constraints that
146	>	* push and pop are called only from the owning thread (or, as
147	>	* extended here, under a lock), while poll may be called from
148	>	* other threads.  (If you are unfamiliar with them, you probably
149	>	* want to read Herlihy and Shavit's book "The Art of
150	>	* Multiprocessor programming", chapter 16 describing these in
151	>	* more detail before proceeding.)  The main work-stealing queue
152	>	* design is roughly similar to those in the papers "Dynamic
153	>	* Circular Work-Stealing Deque" by Chase and Lev, SPAA 2005
154	>	* (http://research.sun.com/scalable/pubs/index.html) and
155	>	* "Idempotent work stealing" by Michael, Saraswat, and Vechev,
156	>	* PPoPP 2009 (http://portal.acm.org/citation.cfm?id=1504186).
157	>	* The main differences ultimately stem from GC requirements that
158	>	* we null out taken slots as soon as we can, to maintain as small
159	>	* a footprint as possible even in programs generating huge
160	>	* numbers of tasks. To accomplish this, we shift the CAS
161	>	* arbitrating pop vs poll (steal) from being on the indices
162	>	* ("base" and "top") to the slots themselves.  So, both a
163	>	* successful pop and poll mainly entail a CAS of a slot from
164	>	* non-null to null.  Because we rely on CASes of references, we
165	>	* do not need tag bits on base or top.  They are simple ints as
166	>	* used in any circular array-based queue (see for example
167	>	* ArrayDeque).  Updates to the indices must still be ordered in a
168	>	* way that guarantees that top == base means the queue is empty,
169	>	* but otherwise may err on the side of possibly making the queue
170	>	* appear nonempty when a push, pop, or poll have not fully
171	>	* committed. Note that this means that the poll operation,
172	>	* considered individually, is not wait-free. One thief cannot
173	>	* successfully continue until another in-progress one (or, if
174	>	* previously empty, a push) completes.  However, in the
175	>	* aggregate, we ensure at least probabilistic non-blockingness.
176	>	* If an attempted steal fails, a thief always chooses a different
177	>	* random victim target to try next. So, in order for one thief to
178	>	* progress, it suffices for any in-progress poll or new push on
179	>	* any empty queue to complete. (This is why we normally use
180	>	* method pollAt and its variants that try once at the apparent
181	>	* base index, else consider alternative actions, rather than
182	>	* method poll.)
183	>	*
184	>	* This approach also enables support of a user mode in which local
185	>	* task processing is in FIFO, not LIFO order, simply by using
186	>	* poll rather than pop.  This can be useful in message-passing
187	>	* frameworks in which tasks are never joined.  However neither
188	>	* mode considers affinities, loads, cache localities, etc, so
189	>	* rarely provide the best possible performance on a given
190	>	* machine, but portably provide good throughput by averaging over
191	>	* these factors.  (Further, even if we did try to use such
192	>	* information, we do not usually have a basis for exploiting it.
193	>	* For example, some sets of tasks profit from cache affinities,
194	>	* but others are harmed by cache pollution effects.)
195	>	*
196	>	* WorkQueues are also used in a similar way for tasks submitted
197	>	* to the pool. We cannot mix these tasks in the same queues used
198	>	* for work-stealing (this would contaminate lifo/fifo
199	>	* processing). Instead, we loosely associate submission queues
200	>	* with submitting threads, using a form of hashing.  The
201	>	* ThreadLocal Submitter class contains a value initially used as
202	>	* a hash code for choosing existing queues, but may be randomly
203	>	* repositioned upon contention with other submitters.  In
204	>	* essence, submitters act like workers except that they never
205	>	* take tasks, and they are multiplexed on to a finite number of
206	>	* shared work queues. However, classes are set up so that future
207	>	* extensions could allow submitters to optionally help perform
208	>	* tasks as well. Insertion of tasks in shared mode requires a
209	>	* lock (mainly to protect in the case of resizing) but we use
210	>	* only a simple spinlock (using bits in field runState), because
211	>	* submitters encountering a busy queue move on to try or create
212	>	* other queues -- they block only when creating and registering
213	>	* new queues.
214	>	*
215	>	* Management
216	>	* ==========
217		*
218		* The main throughput advantages of work-stealing stem from
219		* decentralized control -- workers mostly take tasks from
220		* themselves or each other. We cannot negate this in the
221		* implementation of other management responsibilities. The main
222		* tactic for avoiding bottlenecks is packing nearly all
223	<	* essentially atomic control state into a single 64bit volatile
224	<	* variable ("ctl"). This variable is read on the order of 10-100
225	<	* times as often as it is modified (always via CAS). (There is
226	<	* some additional control state, for example variable "shutdown"
227	<	* for which we can cope with uncoordinated updates.)  This
228	<	* streamlines synchronization and control at the expense of messy
229	<	* constructions needed to repack status bits upon updates.
230	<	* Updates tend not to contend with each other except during
231	<	* bursts while submitted tasks begin or end.  In some cases when
232	<	* they do contend, threads can instead do something else
233	<	* (usually, scan for tasks) until contention subsides.
234	<	*
235	<	* To enable packing, we restrict maximum parallelism to (1<<15)-1
236	<	* (which is far in excess of normal operating range) to allow
237	<	* ids, counts, and their negations (used for thresholding) to fit
238	<	* into 16bit fields.
239	<	*
240	<	* Recording Workers.  Workers are recorded in the "workers" array
241	<	* that is created upon pool construction and expanded if (rarely)
242	<	* necessary.  This is an array as opposed to some other data
243	<	* structure to support index-based random steals by workers.
244	<	* Updates to the array recording new workers and unrecording
245	<	* terminated ones are protected from each other by a seqLock
246	<	* (scanGuard) but the array is otherwise concurrently readable,
166	<	* and accessed directly by workers. To simplify index-based
223	>	* essentially atomic control state into two volatile variables
224	>	* that are by far most often read (not written) as status and
225	>	* consistency checks.
226	>	*
227	>	* Field "ctl" contains 64 bits holding all the information needed
228	>	* to atomically decide to add, inactivate, enqueue (on an event
229	>	* queue), dequeue, and/or re-activate workers.  To enable this
230	>	* packing, we restrict maximum parallelism to (1<<15)-1 (which is
231	>	* far in excess of normal operating range) to allow ids, counts,
232	>	* and their negations (used for thresholding) to fit into 16bit
233	>	* fields.
234	>	*
235	>	* Field "runState" contains 32 bits needed to register and
236	>	* deregister WorkQueues, as well as to enable shutdown. It is
237	>	* only modified under a lock (normally briefly held, but
238	>	* occasionally protecting allocations and resizings) but even
239	>	* when locked remains available to check consistency.
240	>	*
241	>	* Recording WorkQueues.  WorkQueues are recorded in the
242	>	* "workQueues" array that is created upon pool construction and
243	>	* expanded if necessary.  Updates to the array while recording
244	>	* new workers and unrecording terminated ones are protected from
245	>	* each other by a lock but the array is otherwise concurrently
246	>	* readable, and accessed directly.  To simplify index-based
247		* operations, the array size is always a power of two, and all
248	<	* readers must tolerate null slots. To avoid flailing during
249	<	* start-up, the array is presized to hold twice #parallelism
250	<	* workers (which is unlikely to need further resizing during
251	<	* execution). But to avoid dealing with so many null slots,
252	<	* variable scanGuard includes a mask for the nearest power of two
253	<	* that contains all current workers.  All worker thread creation
254	<	* is on-demand, triggered by task submissions, replacement of
255	<	* terminated workers, and/or compensation for blocked
256	<	* workers. However, all other support code is set up to work with
257	<	* other policies.  To ensure that we do not hold on to worker
258	<	* references that would prevent GC, ALL accesses to workers are
259	<	* via indices into the workers array (which is one source of some
260	<	* of the messy code constructions here). In essence, the workers
261	<	* array serves as a weak reference mechanism. Thus for example
262	<	* the wait queue field of ctl stores worker indices, not worker
263	<	* references.  Access to the workers in associated methods (for
264	<	* example signalWork) must both index-check and null-check the
265	<	* IDs. All such accesses ignore bad IDs by returning out early
266	<	* from what they are doing, since this can only be associated
267	<	* with termination, in which case it is OK to give up.
268	<	*
269	<	* All uses of the workers array, as well as queue arrays, check
270	<	* that the array is non-null (even if previously non-null). This
271	<	* allows nulling during termination, which is currently not
272	<	* necessary, but remains an option for resource-revocation-based
273	<	* shutdown schemes.
248	>	* readers must tolerate null slots. Shared (submission) queues
249	>	* are at even indices, worker queues at odd indices. Grouping
250	>	* them together in this way simplifies and speeds up task
251	>	* scanning.
252	>	*
253	>	* All worker thread creation is on-demand, triggered by task
254	>	* submissions, replacement of terminated workers, and/or
255	>	* compensation for blocked workers. However, all other support
256	>	* code is set up to work with other policies.  To ensure that we
257	>	* do not hold on to worker references that would prevent GC, ALL
258	>	* accesses to workQueues are via indices into the workQueues
259	>	* array (which is one source of some of the messy code
260	>	* constructions here). In essence, the workQueues array serves as
261	>	* a weak reference mechanism. Thus for example the wait queue
262	>	* field of ctl stores indices, not references.  Access to the
263	>	* workQueues in associated methods (for example signalWork) must
264	>	* both index-check and null-check the IDs. All such accesses
265	>	* ignore bad IDs by returning out early from what they are doing,
266	>	* since this can only be associated with termination, in which
267	>	* case it is OK to give up.  All uses of the workQueues array
268	>	* also check that it is non-null (even if previously
269	>	* non-null). This allows nulling during termination, which is
270	>	* currently not necessary, but remains an option for
271	>	* resource-revocation-based shutdown schemes. It also helps
272	>	* reduce JIT issuance of uncommon-trap code, which tends to
273	>	* unnecessarily complicate control flow in some methods.
274		*
275	<	* Wait Queuing. Unlike HPC work-stealing frameworks, we cannot
275	>	* Event Queuing. Unlike HPC work-stealing frameworks, we cannot
276		* let workers spin indefinitely scanning for tasks when none can
277		* be found immediately, and we cannot start/resume workers unless
278		* there appear to be tasks available.  On the other hand, we must
279		* quickly prod them into action when new tasks are submitted or
280	<	* generated.  We park/unpark workers after placing in an event
281	<	* wait queue when they cannot find work. This "queue" is actually
282	<	* a simple Treiber stack, headed by the "id" field of ctl, plus a
283	<	* 15bit counter value to both wake up waiters (by advancing their
284	<	* count) and avoid ABA effects. Successors are held in worker
285	<	* field "nextWait".  Queuing deals with several intrinsic races,
286	<	* mainly that a task-producing thread can miss seeing (and
280	>	* generated. In many usages, ramp-up time to activate workers is
281	>	* the main limiting factor in overall performance (this is
282	>	* compounded at program start-up by JIT compilation and
283	>	* allocation). So we try to streamline this as much as possible.
284	>	* We park/unpark workers after placing in an event wait queue
285	>	* when they cannot find work. This "queue" is actually a simple
286	>	* Treiber stack, headed by the "id" field of ctl, plus a 15bit
287	>	* counter value (that reflects the number of times a worker has
288	>	* been inactivated) to avoid ABA effects (we need only as many
289	>	* version numbers as worker threads). Successors are held in
290	>	* field WorkQueue.nextWait.  Queuing deals with several intrinsic
291	>	* races, mainly that a task-producing thread can miss seeing (and
292		* signalling) another thread that gave up looking for work but
293		* has not yet entered the wait queue. We solve this by requiring
294	<	* a full sweep of all workers both before (in scan()) and after
295	<	* (in tryAwaitWork()) a newly waiting worker is added to the wait
296	<	* queue. During a rescan, the worker might release some other
297	<	* queued worker rather than itself, which has the same net
298	<	* effect. Because enqueued workers may actually be rescanning
299	<	* rather than waiting, we set and clear the "parked" field of
300	<	* ForkJoinWorkerThread to reduce unnecessary calls to unpark.
301	<	* (Use of the parked field requires a secondary recheck to avoid
302	<	* missed signals.)
294	>	* a full sweep of all workers (via repeated calls to method
295	>	* scan()) both before and after a newly waiting worker is added
296	>	* to the wait queue. During a rescan, the worker might release
297	>	* some other queued worker rather than itself, which has the same
298	>	* net effect. Because enqueued workers may actually be rescanning
299	>	* rather than waiting, we set and clear the "parker" field of
300	>	* WorkQueues to reduce unnecessary calls to unpark.  (This
301	>	* requires a secondary recheck to avoid missed signals.)  Note
302	>	* the unusual conventions about Thread.interrupts surrounding
303	>	* parking and other blocking: Because interrupts are used solely
304	>	* to alert threads to check termination, which is checked anyway
305	>	* upon blocking, we clear status (using Thread.interrupted)
306	>	* before any call to park, so that park does not immediately
307	>	* return due to status being set via some other unrelated call to
308	>	* interrupt in user code.
309		*
310		* Signalling.  We create or wake up workers only when there
311		* appears to be at least one task they might be able to find and
312		* execute.  When a submission is added or another worker adds a
313	<	* task to a queue that previously had two or fewer tasks, they
313	>	* task to a queue that previously had fewer than two tasks, they
314		* signal waiting workers (or trigger creation of new ones if
315		* fewer than the given parallelism level -- see signalWork).
316	<	* These primary signals are buttressed by signals during rescans
317	<	* as well as those performed when a worker steals a task and
318	<	* notices that there are more tasks too; together these cover the
319	<	* signals needed in cases when more than two tasks are pushed
229	<	* but untaken.
316	>	* These primary signals are buttressed by signals during rescans;
317	>	* together these cover the signals needed in cases when more
318	>	* tasks are pushed but untaken, and improve performance compared
319	>	* to having one thread wake up all workers.
320		*
321		* Trimming workers. To release resources after periods of lack of
322		* use, a worker starting to wait when the pool is quiescent will
#	Line 234 | Line 324 | public class ForkJoinPool extends Abstra
324		* SHRINK_RATE nanosecs. This will slowly propagate, eventually
325		* terminating all workers after long periods of non-use.
326		*
327	<	* Submissions. External submissions are maintained in an
328	<	* array-based queue that is structured identically to
329	<	* ForkJoinWorkerThread queues except for the use of
330	<	* submissionLock in method addSubmission. Unlike the case for
331	<	* worker queues, multiple external threads can add new
332	<	* submissions, so adding requires a lock.
333	<	*
334	<	* Compensation. Beyond work-stealing support and lifecycle
335	<	* control, the main responsibility of this framework is to take
336	<	* actions when one worker is waiting to join a task stolen (or
337	<	* always held by) another.  Because we are multiplexing many
338	<	* tasks on to a pool of workers, we can't just let them block (as
339	<	* in Thread.join).  We also cannot just reassign the joiner's
340	<	* run-time stack with another and replace it later, which would
341	<	* be a form of "continuation", that even if possible is not
342	<	* necessarily a good idea since we sometimes need both an
343	<	* unblocked task and its continuation to progress. Instead we
344	<	* combine two tactics:
327	>	* Shutdown and Termination. A call to shutdownNow atomically sets
328	>	* a runState bit and then (non-atomically) sets each worker's
329	>	* runState status, cancels all unprocessed tasks, and wakes up
330	>	* all waiting workers.  Detecting whether termination should
331	>	* commence after a non-abrupt shutdown() call requires more work
332	>	* and bookkeeping. We need consensus about quiescence (i.e., that
333	>	* there is no more work). The active count provides a primary
334	>	* indication but non-abrupt shutdown still requires a rechecking
335	>	* scan for any workers that are inactive but not queued.
336	>	*
337	>	* Joining Tasks
338	>	* =============
339	>	*
340	>	* Any of several actions may be taken when one worker is waiting
341	>	* to join a task stolen (or always held) by another.  Because we
342	>	* are multiplexing many tasks on to a pool of workers, we can't
343	>	* just let them block (as in Thread.join).  We also cannot just
344	>	* reassign the joiner's run-time stack with another and replace
345	>	* it later, which would be a form of "continuation", that even if
346	>	* possible is not necessarily a good idea since we sometimes need
347	>	* both an unblocked task and its continuation to progress.
348	>	* Instead we combine two tactics:
349		*
350		*   Helping: Arranging for the joiner to execute some task that it
351	<	*      would be running if the steal had not occurred.  Method
258	<	*      ForkJoinWorkerThread.joinTask tracks joining->stealing
259	<	*      links to try to find such a task.
351	>	*      would be running if the steal had not occurred.
352		*
353		*   Compensating: Unless there are already enough live threads,
354	<	*      method tryPreBlock() may create or re-activate a spare
355	<	*      thread to compensate for blocked joiners until they
356	<	*      unblock.
354	>	*      method tryCompensate() may create or re-activate a spare
355	>	*      thread to compensate for blocked joiners until they unblock.
356	>	*
357	>	* A third form (implemented in tryRemoveAndExec and
358	>	* tryPollForAndExec) amounts to helping a hypothetical
359	>	* compensator: If we can readily tell that a possible action of a
360	>	* compensator is to steal and execute the task being joined, the
361	>	* joining thread can do so directly, without the need for a
362	>	* compensation thread (although at the expense of larger run-time
363	>	* stacks, but the tradeoff is typically worthwhile).
364		*
365		* The ManagedBlocker extension API can't use helping so relies
366		* only on compensation in method awaitBlocker.
367		*
368	+	* The algorithm in tryHelpStealer entails a form of "linear"
369	+	* helping: Each worker records (in field currentSteal) the most
370	+	* recent task it stole from some other worker. Plus, it records
371	+	* (in field currentJoin) the task it is currently actively
372	+	* joining. Method tryHelpStealer uses these markers to try to
373	+	* find a worker to help (i.e., steal back a task from and execute
374	+	* it) that could hasten completion of the actively joined task.
375	+	* In essence, the joiner executes a task that would be on its own
376	+	* local deque had the to-be-joined task not been stolen. This may
377	+	* be seen as a conservative variant of the approach in Wagner &
378	+	* Calder "Leapfrogging: a portable technique for implementing
379	+	* efficient futures" SIGPLAN Notices, 1993
380	+	* (http://portal.acm.org/citation.cfm?id=155354). It differs in
381	+	* that: (1) We only maintain dependency links across workers upon
382	+	* steals, rather than use per-task bookkeeping.  This sometimes
383	+	* requires a linear scan of workQueues array to locate stealers,
384	+	* but often doesn't because stealers leave hints (that may become
385	+	* stale/wrong) of where to locate them.  A stealHint is only a
386	+	* hint because a worker might have had multiple steals and the
387	+	* hint records only one of them (usually the most current).
388	+	* Hinting isolates cost to when it is needed, rather than adding
389	+	* to per-task overhead.  (2) It is "shallow", ignoring nesting
390	+	* and potentially cyclic mutual steals.  (3) It is intentionally
391	+	* racy: field currentJoin is updated only while actively joining,
392	+	* which means that we miss links in the chain during long-lived
393	+	* tasks, GC stalls etc (which is OK since blocking in such cases
394	+	* is usually a good idea).  (4) We bound the number of attempts
395	+	* to find work (see MAX_HELP) and fall back to suspending the
396	+	* worker and if necessary replacing it with another.
397	+	*
398		* It is impossible to keep exactly the target parallelism number
399		* of threads running at any given time.  Determining the
400		* existence of conservatively safe helping targets, the
401		* availability of already-created spares, and the apparent need
402	<	* to create new spares are all racy and require heuristic
403	<	* guidance, so we rely on multiple retries of each.  Currently,
404	<	* in keeping with on-demand signalling policy, we compensate only
405	<	* if blocking would leave less than one active (non-waiting,
406	<	* non-blocked) worker. Additionally, to avoid some false alarms
407	<	* due to GC, lagging counters, system activity, etc, compensated
408	<	* blocking for joins is only attempted after rechecks stabilize
409	<	* (retries are interspersed with Thread.yield, for good
410	<	* citizenship).  The variable blockedCount, incremented before
411	<	* blocking and decremented after, is sometimes needed to
412	<	* distinguish cases of waiting for work vs blocking on joins or
413	<	* other managed sync. Both cases are equivalent for most pool
414	<	* control, so we can update non-atomically. (Additionally,
415	<	* contention on blockedCount alleviates some contention on ctl).
416	<	*
417	<	* Shutdown and Termination. A call to shutdownNow atomically sets
418	<	* the ctl stop bit and then (non-atomically) sets each workers
419	<	* "terminate" status, cancels all unprocessed tasks, and wakes up
420	<	* all waiting workers.  Detecting whether termination should
421	<	* commence after a non-abrupt shutdown() call requires more work
422	<	* and bookkeeping. We need consensus about quiescence (i.e., that
423	<	* there is no more work) which is reflected in active counts so
424	<	* long as there are no current blockers, as well as possible
425	<	* re-evaluations during independent changes in blocking or
426	<	* quiescing workers.
402	>	* to create new spares are all racy, so we rely on multiple
403	>	* retries of each.  Compensation in the apparent absence of
404	>	* helping opportunities is challenging to control on JVMs, where
405	>	* GC and other activities can stall progress of tasks that in
406	>	* turn stall out many other dependent tasks, without us being
407	>	* able to determine whether they will ever require compensation.
408	>	* Even though work-stealing otherwise encounters little
409	>	* degradation in the presence of more threads than cores,
410	>	* aggressively adding new threads in such cases entails risk of
411	>	* unwanted positive feedback control loops in which more threads
412	>	* cause more dependent stalls (as well as delayed progress of
413	>	* unblocked threads to the point that we know they are available)
414	>	* leading to more situations requiring more threads, and so
415	>	* on. This aspect of control can be seen as an (analytically
416	>	* intractable) game with an opponent that may choose the worst
417	>	* (for us) active thread to stall at any time.  We take several
418	>	* precautions to bound losses (and thus bound gains), mainly in
419	>	* methods tryCompensate and awaitJoin: (1) We only try
420	>	* compensation after attempting enough helping steps (measured
421	>	* via counting and timing) that we have already consumed the
422	>	* estimated cost of creating and activating a new thread.  (2) We
423	>	* allow up to 50% of threads to be blocked before initially
424	>	* adding any others, and unless completely saturated, check that
425	>	* some work is available for a new worker before adding. Also, we
426	>	* create up to only 50% more threads until entering a mode that
427	>	* only adds a thread if all others are possibly blocked.  All
428	>	* together, this means that we might be half as fast to react,
429	>	* and create half as many threads as possible in the ideal case,
430	>	* but present vastly fewer anomalies in all other cases compared
431	>	* to both more aggressive and more conservative alternatives.
432		*
433		* Style notes: There is a lot of representation-level coupling
434		* among classes ForkJoinPool, ForkJoinWorkerThread, and
435	<	* ForkJoinTask.  Most fields of ForkJoinWorkerThread maintain
436	<	* data structures managed by ForkJoinPool, so are directly
437	<	* accessed.  Conversely we allow access to "workers" array by
438	<	* workers, and direct access to ForkJoinTask.status by both
439	<	* ForkJoinPool and ForkJoinWorkerThread.  There is little point
440	<	* trying to reduce this, since any associated future changes in
441	<	* representations will need to be accompanied by algorithmic
442	<	* changes anyway. All together, these low-level implementation
443	<	* choices produce as much as a factor of 4 performance
444	<	* improvement compared to naive implementations, and enable the
445	<	* processing of billions of tasks per second, at the expense of
446	<	* some ugliness.
447	<	*
448	<	* Methods signalWork() and scan() are the main bottlenecks so are
449	<	* especially heavily micro-optimized/mangled.  There are lots of
450	<	* inline assignments (of form "while ((local = field) != 0)")
451	<	* which are usually the simplest way to ensure the required read
452	<	* orderings (which are sometimes critical). This leads to a
453	<	* "C"-like style of listing declarations of these locals at the
454	<	* heads of methods or blocks.  There are several occurrences of
455	<	* the unusual "do {} while (!cas...)"  which is the simplest way
456	<	* to force an update of a CAS'ed variable. There are also other
457	<	* coding oddities that help some methods perform reasonably even
458	<	* when interpreted (not compiled).
459	<	*
460	<	* The order of declarations in this file is: (1) declarations of
461	<	* statics (2) fields (along with constants used when unpacking
462	<	* some of them), listed in an order that tends to reduce
329	<	* contention among them a bit under most JVMs.  (3) internal
330	<	* control methods (4) callbacks and other support for
331	<	* ForkJoinTask and ForkJoinWorkerThread classes, (5) exported
332	<	* methods (plus a few little helpers). (6) static block
333	<	* initializing all statics in a minimally dependent order.
435	>	* ForkJoinTask.  The fields of WorkQueue maintain data structures
436	>	* managed by ForkJoinPool, so are directly accessed.  There is
437	>	* little point trying to reduce this, since any associated future
438	>	* changes in representations will need to be accompanied by
439	>	* algorithmic changes anyway. Several methods intrinsically
440	>	* sprawl because they must accumulate sets of consistent reads of
441	>	* volatiles held in local variables.  Methods signalWork() and
442	>	* scan() are the main bottlenecks, so are especially heavily
443	>	* micro-optimized/mangled.  There are lots of inline assignments
444	>	* (of form "while ((local = field) != 0)") which are usually the
445	>	* simplest way to ensure the required read orderings (which are
446	>	* sometimes critical). This leads to a "C"-like style of listing
447	>	* declarations of these locals at the heads of methods or blocks.
448	>	* There are several occurrences of the unusual "do {} while
449	>	* (!cas...)"  which is the simplest way to force an update of a
450	>	* CAS'ed variable. There are also other coding oddities that help
451	>	* some methods perform reasonably even when interpreted (not
452	>	* compiled).
453	>	*
454	>	* The order of declarations in this file is:
455	>	* (1) Static utility functions
456	>	* (2) Nested (static) classes
457	>	* (3) Static fields
458	>	* (4) Fields, along with constants used when unpacking some of them
459	>	* (5) Internal control methods
460	>	* (6) Callbacks and other support for ForkJoinTask methods
461	>	* (7) Exported methods
462	>	* (8) Static block initializing statics in minimally dependent order
463		*/
464
465	+	// Static utilities
466	+
467	+	/**
468	+	* If there is a security manager, makes sure caller has
469	+	* permission to modify threads.
470	+	*/
471	+	private static void checkPermission() {
472	+	SecurityManager security = System.getSecurityManager();
473	+	if (security != null)
474	+	security.checkPermission(modifyThreadPermission);
475	+	}
476	+
477	+	// Nested classes
478	+
479		/**
480		* Factory for creating new {@link ForkJoinWorkerThread}s.
481		* A {@code ForkJoinWorkerThreadFactory} must be defined and used
#	Line 361 | Line 504 | public class ForkJoinPool extends Abstra
504		}
505
506		/**
507	<	* Creates a new ForkJoinWorkerThread. This factory is used unless
508	<	* overridden in ForkJoinPool constructors.
507	>	* A simple non-reentrant lock used for exclusion when managing
508	>	* queues and workers. We use a custom lock so that we can readily
509	>	* probe lock state in constructions that check among alternative
510	>	* actions. The lock is normally only very briefly held, and
511	>	* sometimes treated as a spinlock, but other usages block to
512	>	* reduce overall contention in those cases where locked code
513	>	* bodies perform allocation/resizing.
514	>	*/
515	>	static final class Mutex extends AbstractQueuedSynchronizer {
516	>	public final boolean tryAcquire(int ignore) {
517	>	return compareAndSetState(0, 1);
518	>	}
519	>	public final boolean tryRelease(int ignore) {
520	>	setState(0);
521	>	return true;
522	>	}
523	>	public final void lock() { acquire(0); }
524	>	public final void unlock() { release(0); }
525	>	public final boolean isHeldExclusively() { return getState() == 1; }
526	>	public final Condition newCondition() { return new ConditionObject(); }
527	>	}
528	>
529	>	/**
530	>	* Class for artificial tasks that are used to replace the target
531	>	* of local joins if they are removed from an interior queue slot
532	>	* in WorkQueue.tryRemoveAndExec. We don't need the proxy to
533	>	* actually do anything beyond having a unique identity.
534	>	*/
535	>	static final class EmptyTask extends ForkJoinTask<Void> {
536	>	EmptyTask() { status = ForkJoinTask.NORMAL; } // force done
537	>	public final Void getRawResult() { return null; }
538	>	public final void setRawResult(Void x) {}
539	>	public final boolean exec() { return true; }
540	>	}
541	>
542	>	/**
543	>	* Queues supporting work-stealing as well as external task
544	>	* submission. See above for main rationale and algorithms.
545	>	* Implementation relies heavily on "Unsafe" intrinsics
546	>	* and selective use of "volatile":
547	>	*
548	>	* Field "base" is the index (mod array.length) of the least valid
549	>	* queue slot, which is always the next position to steal (poll)
550	>	* from if nonempty. Reads and writes require volatile orderings
551	>	* but not CAS, because updates are only performed after slot
552	>	* CASes.
553	>	*
554	>	* Field "top" is the index (mod array.length) of the next queue
555	>	* slot to push to or pop from. It is written only by owner thread
556	>	* for push, or under lock for trySharedPush, and accessed by
557	>	* other threads only after reading (volatile) base.  Both top and
558	>	* base are allowed to wrap around on overflow, but (top - base)
559	>	* (or more commonly -(base - top) to force volatile read of base
560	>	* before top) still estimates size.
561	>	*
562	>	* The array slots are read and written using the emulation of
563	>	* volatiles/atomics provided by Unsafe. Insertions must in
564	>	* general use putOrderedObject as a form of releasing store to
565	>	* ensure that all writes to the task object are ordered before
566	>	* its publication in the queue. (Although we can avoid one case
567	>	* of this when locked in trySharedPush.) All removals entail a
568	>	* CAS to null.  The array is always a power of two. To ensure
569	>	* safety of Unsafe array operations, all accesses perform
570	>	* explicit null checks and implicit bounds checks via
571	>	* power-of-two masking.
572	>	*
573	>	* In addition to basic queuing support, this class contains
574	>	* fields described elsewhere to control execution. It turns out
575	>	* to work better memory-layout-wise to include them in this
576	>	* class rather than a separate class.
577	>	*
578	>	* Performance on most platforms is very sensitive to placement of
579	>	* instances of both WorkQueues and their arrays -- we absolutely
580	>	* do not want multiple WorkQueue instances or multiple queue
581	>	* arrays sharing cache lines. (It would be best for queue objects
582	>	* and their arrays to share, but there is nothing available to
583	>	* help arrange that).  Unfortunately, because they are recorded
584	>	* in a common array, WorkQueue instances are often moved to be
585	>	* adjacent by garbage collectors. To reduce impact, we use field
586	>	* padding that works OK on common platforms; this effectively
587	>	* trades off slightly slower average field access for the sake of
588	>	* avoiding really bad worst-case access. (Until better JVM
589	>	* support is in place, this padding is dependent on transient
590	>	* properties of JVM field layout rules.)  We also take care in
591	>	* allocating, sizing and resizing the array. Non-shared queue
592	>	* arrays are initialized (via method growArray) by workers before
593	>	* use. Others are allocated on first use.
594		*/
595	<	public static final ForkJoinWorkerThreadFactory
596	<	defaultForkJoinWorkerThreadFactory;
595	>	static final class WorkQueue {
596	>	/**
597	>	* Capacity of work-stealing queue array upon initialization.
598	>	* Must be a power of two; at least 4, but should be larger to
599	>	* reduce or eliminate cacheline sharing among queues.
600	>	* Currently, it is much larger, as a partial workaround for
601	>	* the fact that JVMs often place arrays in locations that
602	>	* share GC bookkeeping (especially cardmarks) such that
603	>	* per-write accesses encounter serious memory contention.
604	>	*/
605	>	static final int INITIAL_QUEUE_CAPACITY = 1 << 13;
606
607	<	/**
608	<	* Permission required for callers of methods that may start or
609	<	* kill threads.
610	<	*/
611	<	private static final RuntimePermission modifyThreadPermission;
607	>	/**
608	>	* Maximum size for queue arrays. Must be a power of two less
609	>	* than or equal to 1 << (31 - width of array entry) to ensure
610	>	* lack of wraparound of index calculations, but defined to a
611	>	* value a bit less than this to help users trap runaway
612	>	* programs before saturating systems.
613	>	*/
614	>	static final int MAXIMUM_QUEUE_CAPACITY = 1 << 26; // 64M
615	>
616	>	volatile long totalSteals; // cumulative number of steals
617	>	int seed;                  // for random scanning; initialize nonzero
618	>	volatile int eventCount;   // encoded inactivation count; < 0 if inactive
619	>	int nextWait;              // encoded record of next event waiter
620	>	int rescans;               // remaining scans until block
621	>	int nsteals;               // top-level task executions since last idle
622	>	final int mode;            // lifo, fifo, or shared
623	>	int poolIndex;             // index of this queue in pool (or 0)
624	>	int stealHint;             // index of most recent known stealer
625	>	volatile int runState;     // 1: locked, -1: terminate; else 0
626	>	volatile int base;         // index of next slot for poll
627	>	int top;                   // index of next slot for push
628	>	ForkJoinTask<?>[] array;   // the elements (initially unallocated)
629	>	final ForkJoinPool pool;   // the containing pool (may be null)
630	>	final ForkJoinWorkerThread owner; // owning thread or null if shared
631	>	volatile Thread parker;    // == owner during call to park; else null
632	>	volatile ForkJoinTask<?> currentJoin;  // task being joined in awaitJoin
633	>	ForkJoinTask<?> currentSteal; // current non-local task being executed
634	>	// Heuristic padding to ameliorate unfortunate memory placements
635	>	Object p00, p01, p02, p03, p04, p05, p06, p07;
636	>	Object p08, p09, p0a, p0b, p0c, p0d, p0e;
637	>
638	>	WorkQueue(ForkJoinPool pool, ForkJoinWorkerThread owner, int mode) {
639	>	this.mode = mode;
640	>	this.pool = pool;
641	>	this.owner = owner;
642	>	// Place indices in the center of array (that is not yet allocated)
643	>	base = top = INITIAL_QUEUE_CAPACITY >>> 1;
644	>	}
645	>
646	>	/**
647	>	* Returns the approximate number of tasks in the queue.
648	>	*/
649	>	final int queueSize() {
650	>	int n = base - top;       // non-owner callers must read base first
651	>	return (n >= 0) ? 0 : -n; // ignore transient negative
652	>	}
653	>
654	>	/**
655	>	* Provides a more accurate estimate of whether this queue has
656	>	* any tasks than does queueSize, by checking whether a
657	>	* near-empty queue has at least one unclaimed task.
658	>	*/
659	>	final boolean isEmpty() {
660	>	ForkJoinTask<?>[] a; int m, s;
661	>	int n = base - (s = top);
662	>	return (n >= 0 ||
663	>	(n == -1 &&
664	>	((a = array) == null ||
665	>	(m = a.length - 1) < 0 ||
666	>	U.getObjectVolatile
667	>	(a, ((m & (s - 1)) << ASHIFT) + ABASE) == null)));
668	>	}
669	>
670	>	/**
671	>	* Pushes a task. Call only by owner in unshared queues.
672	>	*
673	>	* @param task the task. Caller must ensure non-null.
674	>	* @throw RejectedExecutionException if array cannot be resized
675	>	*/
676	>	final void push(ForkJoinTask<?> task) {
677	>	ForkJoinTask<?>[] a; ForkJoinPool p;
678	>	int s = top, m, n;
679	>	if ((a = array) != null) {    // ignore if queue removed
680	>	U.putOrderedObject
681	>	(a, (((m = a.length - 1) & s) << ASHIFT) + ABASE, task);
682	>	if ((n = (top = s + 1) - base) <= 2) {
683	>	if ((p = pool) != null)
684	>	p.signalWork();
685	>	}
686	>	else if (n >= m)
687	>	growArray(true);
688	>	}
689	>	}
690	>
691	>	/**
692	>	* Pushes a task if lock is free and array is either big
693	>	* enough or can be resized to be big enough.
694	>	*
695	>	* @param task the task. Caller must ensure non-null.
696	>	* @return true if submitted
697	>	*/
698	>	final boolean trySharedPush(ForkJoinTask<?> task) {
699	>	boolean submitted = false;
700	>	if (runState == 0 && U.compareAndSwapInt(this, RUNSTATE, 0, 1)) {
701	>	ForkJoinTask<?>[] a = array;
702	>	int s = top;
703	>	try {
704	>	if ((a != null && a.length > s + 1 - base) ||
705	>	(a = growArray(false)) != null) { // must presize
706	>	int j = (((a.length - 1) & s) << ASHIFT) + ABASE;
707	>	U.putObject(a, (long)j, task);    // don't need "ordered"
708	>	top = s + 1;
709	>	submitted = true;
710	>	}
711	>	} finally {
712	>	runState = 0;                         // unlock
713	>	}
714	>	}
715	>	return submitted;
716	>	}
717	>
718	>	/**
719	>	* Takes next task, if one exists, in LIFO order.  Call only
720	>	* by owner in unshared queues. (We do not have a shared
721	>	* version of this method because it is never needed.)
722	>	*/
723	>	final ForkJoinTask<?> pop() {
724	>	ForkJoinTask<?>[] a; ForkJoinTask<?> t; int m;
725	>	if ((a = array) != null && (m = a.length - 1) >= 0) {
726	>	for (int s; (s = top - 1) - base >= 0;) {
727	>	long j = ((m & s) << ASHIFT) + ABASE;
728	>	if ((t = (ForkJoinTask<?>)U.getObject(a, j)) == null)
729	>	break;
730	>	if (U.compareAndSwapObject(a, j, t, null)) {
731	>	top = s;
732	>	return t;
733	>	}
734	>	}
735	>	}
736	>	return null;
737	>	}
738	>
739	>	/**
740	>	* Takes a task in FIFO order if b is base of queue and a task
741	>	* can be claimed without contention. Specialized versions
742	>	* appear in ForkJoinPool methods scan and tryHelpStealer.
743	>	*/
744	>	final ForkJoinTask<?> pollAt(int b) {
745	>	ForkJoinTask<?> t; ForkJoinTask<?>[] a;
746	>	if ((a = array) != null) {
747	>	int j = (((a.length - 1) & b) << ASHIFT) + ABASE;
748	>	if ((t = (ForkJoinTask<?>)U.getObjectVolatile(a, j)) != null &&
749	>	base == b &&
750	>	U.compareAndSwapObject(a, j, t, null)) {
751	>	base = b + 1;
752	>	return t;
753	>	}
754	>	}
755	>	return null;
756	>	}
757	>
758	>	/**
759	>	* Takes next task, if one exists, in FIFO order.
760	>	*/
761	>	final ForkJoinTask<?> poll() {
762	>	ForkJoinTask<?>[] a; int b; ForkJoinTask<?> t;
763	>	while ((b = base) - top < 0 && (a = array) != null) {
764	>	int j = (((a.length - 1) & b) << ASHIFT) + ABASE;
765	>	t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
766	>	if (t != null) {
767	>	if (base == b &&
768	>	U.compareAndSwapObject(a, j, t, null)) {
769	>	base = b + 1;
770	>	return t;
771	>	}
772	>	}
773	>	else if (base == b) {
774	>	if (b + 1 == top)
775	>	break;
776	>	Thread.yield(); // wait for lagging update
777	>	}
778	>	}
779	>	return null;
780	>	}
781	>
782	>	/**
783	>	* Takes next task, if one exists, in order specified by mode.
784	>	*/
785	>	final ForkJoinTask<?> nextLocalTask() {
786	>	return mode == 0 ? pop() : poll();
787	>	}
788	>
789	>	/**
790	>	* Returns next task, if one exists, in order specified by mode.
791	>	*/
792	>	final ForkJoinTask<?> peek() {
793	>	ForkJoinTask<?>[] a = array; int m;
794	>	if (a == null || (m = a.length - 1) < 0)
795	>	return null;
796	>	int i = mode == 0 ? top - 1 : base;
797	>	int j = ((i & m) << ASHIFT) + ABASE;
798	>	return (ForkJoinTask<?>)U.getObjectVolatile(a, j);
799	>	}
800	>
801	>	/**
802	>	* Pops the given task only if it is at the current top.
803	>	*/
804	>	final boolean tryUnpush(ForkJoinTask<?> t) {
805	>	ForkJoinTask<?>[] a; int s;
806	>	if ((a = array) != null && (s = top) != base &&
807	>	U.compareAndSwapObject
808	>	(a, (((a.length - 1) & --s) << ASHIFT) + ABASE, t, null)) {
809	>	top = s;
810	>	return true;
811	>	}
812	>	return false;
813	>	}
814	>
815	>	/**
816	>	* Polls the given task only if it is at the current base.
817	>	*/
818	>	final boolean pollFor(ForkJoinTask<?> task) {
819	>	ForkJoinTask<?>[] a; int b;
820	>	if ((b = base) - top < 0 && (a = array) != null) {
821	>	int j = (((a.length - 1) & b) << ASHIFT) + ABASE;
822	>	if (U.getObjectVolatile(a, j) == task && base == b &&
823	>	U.compareAndSwapObject(a, j, task, null)) {
824	>	base = b + 1;
825	>	return true;
826	>	}
827	>	}
828	>	return false;
829	>	}
830	>
831	>	/**
832	>	* Initializes or doubles the capacity of array. Call either
833	>	* by owner or with lock held -- it is OK for base, but not
834	>	* top, to move while resizings are in progress.
835	>	*
836	>	* @param rejectOnFailure if true, throw exception if capacity
837	>	* exceeded (relayed ultimately to user); else return null.
838	>	*/
839	>	final ForkJoinTask<?>[] growArray(boolean rejectOnFailure) {
840	>	ForkJoinTask<?>[] oldA = array;
841	>	int size = oldA != null ? oldA.length << 1 : INITIAL_QUEUE_CAPACITY;
842	>	if (size <= MAXIMUM_QUEUE_CAPACITY) {
843	>	int oldMask, t, b;
844	>	ForkJoinTask<?>[] a = array = new ForkJoinTask<?>[size];
845	>	if (oldA != null && (oldMask = oldA.length - 1) >= 0 &&
846	>	(t = top) - (b = base) > 0) {
847	>	int mask = size - 1;
848	>	do {
849	>	ForkJoinTask<?> x;
850	>	int oldj = ((b & oldMask) << ASHIFT) + ABASE;
851	>	int j    = ((b &    mask) << ASHIFT) + ABASE;
852	>	x = (ForkJoinTask<?>)U.getObjectVolatile(oldA, oldj);
853	>	if (x != null &&
854	>	U.compareAndSwapObject(oldA, oldj, x, null))
855	>	U.putObjectVolatile(a, j, x);
856	>	} while (++b != t);
857	>	}
858	>	return a;
859	>	}
860	>	else if (!rejectOnFailure)
861	>	return null;
862	>	else
863	>	throw new RejectedExecutionException("Queue capacity exceeded");
864	>	}
865	>
866	>	/**
867	>	* Removes and cancels all known tasks, ignoring any exceptions.
868	>	*/
869	>	final void cancelAll() {
870	>	ForkJoinTask.cancelIgnoringExceptions(currentJoin);
871	>	ForkJoinTask.cancelIgnoringExceptions(currentSteal);
872	>	for (ForkJoinTask<?> t; (t = poll()) != null; )
873	>	ForkJoinTask.cancelIgnoringExceptions(t);
874	>	}
875
876	+	/**
877	+	* Computes next value for random probes.  Scans don't require
878	+	* a very high quality generator, but also not a crummy one.
879	+	* Marsaglia xor-shift is cheap and works well enough.  Note:
880	+	* This is manually inlined in its usages in ForkJoinPool to
881	+	* avoid writes inside busy scan loops.
882	+	*/
883	+	final int nextSeed() {
884	+	int r = seed;
885	+	r ^= r << 13;
886	+	r ^= r >>> 17;
887	+	return seed = r ^= r << 5;
888	+	}
889	+
890	+	// Execution methods
891	+
892	+	/**
893	+	* Pops and runs tasks until empty.
894	+	*/
895	+	private void popAndExecAll() {
896	+	// A bit faster than repeated pop calls
897	+	ForkJoinTask<?>[] a; int m, s; long j; ForkJoinTask<?> t;
898	+	while ((a = array) != null && (m = a.length - 1) >= 0 &&
899	+	(s = top - 1) - base >= 0 &&
900	+	(t = ((ForkJoinTask<?>)
901	+	U.getObject(a, j = ((m & s) << ASHIFT) + ABASE)))
902	+	!= null) {
903	+	if (U.compareAndSwapObject(a, j, t, null)) {
904	+	top = s;
905	+	t.doExec();
906	+	}
907	+	}
908	+	}
909	+
910	+	/**
911	+	* Polls and runs tasks until empty.
912	+	*/
913	+	private void pollAndExecAll() {
914	+	for (ForkJoinTask<?> t; (t = poll()) != null;)
915	+	t.doExec();
916	+	}
917	+
918	+	/**
919	+	* If present, removes from queue and executes the given task, or
920	+	* any other cancelled task. Returns (true) immediately on any CAS
921	+	* or consistency check failure so caller can retry.
922	+	*
923	+	* @return 0 if no progress can be made, else positive
924	+	* (this unusual convention simplifies use with tryHelpStealer.)
925	+	*/
926	+	final int tryRemoveAndExec(ForkJoinTask<?> task) {
927	+	int stat = 1;
928	+	boolean removed = false, empty = true;
929	+	ForkJoinTask<?>[] a; int m, s, b, n;
930	+	if ((a = array) != null && (m = a.length - 1) >= 0 &&
931	+	(n = (s = top) - (b = base)) > 0) {
932	+	for (ForkJoinTask<?> t;;) {           // traverse from s to b
933	+	int j = ((--s & m) << ASHIFT) + ABASE;
934	+	t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
935	+	if (t == null)                    // inconsistent length
936	+	break;
937	+	else if (t == task) {
938	+	if (s + 1 == top) {           // pop
939	+	if (!U.compareAndSwapObject(a, j, task, null))
940	+	break;
941	+	top = s;
942	+	removed = true;
943	+	}
944	+	else if (base == b)           // replace with proxy
945	+	removed = U.compareAndSwapObject(a, j, task,
946	+	new EmptyTask());
947	+	break;
948	+	}
949	+	else if (t.status >= 0)
950	+	empty = false;
951	+	else if (s + 1 == top) {          // pop and throw away
952	+	if (U.compareAndSwapObject(a, j, t, null))
953	+	top = s;
954	+	break;
955	+	}
956	+	if (--n == 0) {
957	+	if (!empty && base == b)
958	+	stat = 0;
959	+	break;
960	+	}
961	+	}
962	+	}
963	+	if (removed)
964	+	task.doExec();
965	+	return stat;
966	+	}
967	+
968	+	/**
969	+	* Executes a top-level task and any local tasks remaining
970	+	* after execution.
971	+	*/
972	+	final void runTask(ForkJoinTask<?> t) {
973	+	if (t != null) {
974	+	currentSteal = t;
975	+	t.doExec();
976	+	if (top != base) {       // process remaining local tasks
977	+	if (mode == 0)
978	+	popAndExecAll();
979	+	else
980	+	pollAndExecAll();
981	+	}
982	+	++nsteals;
983	+	currentSteal = null;
984	+	}
985	+	}
986	+
987	+	/**
988	+	* Executes a non-top-level (stolen) task.
989	+	*/
990	+	final void runSubtask(ForkJoinTask<?> t) {
991	+	if (t != null) {
992	+	ForkJoinTask<?> ps = currentSteal;
993	+	currentSteal = t;
994	+	t.doExec();
995	+	currentSteal = ps;
996	+	}
997	+	}
998	+
999	+	/**
1000	+	* Returns true if owned and not known to be blocked.
1001	+	*/
1002	+	final boolean isApparentlyUnblocked() {
1003	+	Thread wt; Thread.State s;
1004	+	return (eventCount >= 0 &&
1005	+	(wt = owner) != null &&
1006	+	(s = wt.getState()) != Thread.State.BLOCKED &&
1007	+	s != Thread.State.WAITING &&
1008	+	s != Thread.State.TIMED_WAITING);
1009	+	}
1010	+
1011	+	/**
1012	+	* If this owned and is not already interrupted, try to
1013	+	* interrupt and/or unpark, ignoring exceptions.
1014	+	*/
1015	+	final void interruptOwner() {
1016	+	Thread wt, p;
1017	+	if ((wt = owner) != null && !wt.isInterrupted()) {
1018	+	try {
1019	+	wt.interrupt();
1020	+	} catch (SecurityException ignore) {
1021	+	}
1022	+	}
1023	+	if ((p = parker) != null)
1024	+	U.unpark(p);
1025	+	}
1026	+
1027	+	// Unsafe mechanics
1028	+	private static final sun.misc.Unsafe U;
1029	+	private static final long RUNSTATE;
1030	+	private static final int ABASE;
1031	+	private static final int ASHIFT;
1032	+	static {
1033	+	int s;
1034	+	try {
1035	+	U = getUnsafe();
1036	+	Class<?> k = WorkQueue.class;
1037	+	Class<?> ak = ForkJoinTask[].class;
1038	+	RUNSTATE = U.objectFieldOffset
1039	+	(k.getDeclaredField("runState"));
1040	+	ABASE = U.arrayBaseOffset(ak);
1041	+	s = U.arrayIndexScale(ak);
1042	+	} catch (Exception e) {
1043	+	throw new Error(e);
1044	+	}
1045	+	if ((s & (s-1)) != 0)
1046	+	throw new Error("data type scale not a power of two");
1047	+	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
1048	+	}
1049	+	}
1050		/**
1051	<	* If there is a security manager, makes sure caller has
1052	<	* permission to modify threads.
1053	<	*/
1054	<	private static void checkPermission() {
1055	<	SecurityManager security = System.getSecurityManager();
1056	<	if (security != null)
1057	<	security.checkPermission(modifyThreadPermission);
1051	>	* Per-thread records for threads that submit to pools. Currently
1052	>	* holds only pseudo-random seed / index that is used to choose
1053	>	* submission queues in method doSubmit. In the future, this may
1054	>	* also incorporate a means to implement different task rejection
1055	>	* and resubmission policies.
1056	>	*
1057	>	* Seeds for submitters and workers/workQueues work in basically
1058	>	* the same way but are initialized and updated using slightly
1059	>	* different mechanics. Both are initialized using the same
1060	>	* approach as in class ThreadLocal, where successive values are
1061	>	* unlikely to collide with previous values. This is done during
1062	>	* registration for workers, but requires a separate AtomicInteger
1063	>	* for submitters. Seeds are then randomly modified upon
1064	>	* collisions using xorshifts, which requires a non-zero seed.
1065	>	*/
1066	>	static final class Submitter {
1067	>	int seed;
1068	>	Submitter() {
1069	>	int s = nextSubmitterSeed.getAndAdd(SEED_INCREMENT);
1070	>	seed = (s == 0) ? 1 : s; // ensure non-zero
1071	>	}
1072		}
1073
1074	<	/**
1075	<	* Generator for assigning sequence numbers as pool names.
1076	<	*/
1077	<	private static final AtomicInteger poolNumberGenerator;
1074	>	/** ThreadLocal class for Submitters */
1075	>	static final class ThreadSubmitter extends ThreadLocal<Submitter> {
1076	>	public Submitter initialValue() { return new Submitter(); }
1077	>	}
1078
1079	<	/**
392	<	* Generator for initial random seeds for worker victim
393	<	* selection. This is used only to create initial seeds. Random
394	<	* steals use a cheaper xorshift generator per steal attempt. We
395	<	* don't expect much contention on seedGenerator, so just use a
396	<	* plain Random.
397	<	*/
398	<	static final Random workerSeedGenerator;
1079	>	// static fields (initialized in static initializer below)
1080
1081		/**
1082	<	* Array holding all worker threads in the pool.  Initialized upon
1083	<	* construction. Array size must be a power of two.  Updates and
403	<	* replacements are protected by scanGuard, but the array is
404	<	* always kept in a consistent enough state to be randomly
405	<	* accessed without locking by workers performing work-stealing,
406	<	* as well as other traversal-based methods in this class, so long
407	<	* as reads memory-acquire by first reading ctl. All readers must
408	<	* tolerate that some array slots may be null.
1082	>	* Creates a new ForkJoinWorkerThread. This factory is used unless
1083	>	* overridden in ForkJoinPool constructors.
1084		*/
1085	<	ForkJoinWorkerThread[] workers;
1085	>	public static final ForkJoinWorkerThreadFactory
1086	>	defaultForkJoinWorkerThreadFactory;
1087
1088		/**
1089	<	* Initial size for submission queue array. Must be a power of
414	<	* two.  In many applications, these always stay small so we use a
415	<	* small initial cap.
1089	>	* Generator for assigning sequence numbers as pool names.
1090		*/
1091	<	private static final int INITIAL_QUEUE_CAPACITY = 8;
1091	>	private static final AtomicInteger poolNumberGenerator;
1092
1093		/**
1094	<	* Maximum size for submission queue array. Must be a power of two
1095	<	* less than or equal to 1 << (31 - width of array entry) to
422	<	* ensure lack of index wraparound, but is capped at a lower
423	<	* value to help users trap runaway computations.
1094	>	* Generator for initial hashes/seeds for submitters. Accessed by
1095	>	* Submitter class constructor.
1096		*/
1097	<	private static final int MAXIMUM_QUEUE_CAPACITY = 1 << 24; // 16M
1097	>	static final AtomicInteger nextSubmitterSeed;
1098
1099		/**
1100	<	* Array serving as submission queue. Initialized upon construction.
1100	>	* Permission required for callers of methods that may start or
1101	>	* kill threads.
1102		*/
1103	<	private ForkJoinTask<?>[] submissionQueue;
1103	>	private static final RuntimePermission modifyThreadPermission;
1104
1105		/**
1106	<	* Lock protecting submissions array for addSubmission
1106	>	* Per-thread submission bookkeeping. Shared across all pools
1107	>	* to reduce ThreadLocal pollution and because random motion
1108	>	* to avoid contention in one pool is likely to hold for others.
1109		*/
1110	<	private final ReentrantLock submissionLock;
1110	>	private static final ThreadSubmitter submitters;
1111	>
1112	>	// static constants
1113
1114		/**
1115	<	* Condition for awaitTermination, using submissionLock for
1116	<	* convenience.
1115	>	* The wakeup interval (in nanoseconds) for a worker waiting for a
1116	>	* task when the pool is quiescent to instead try to shrink the
1117	>	* number of workers.  The exact value does not matter too
1118	>	* much. It must be short enough to release resources during
1119	>	* sustained periods of idleness, but not so short that threads
1120	>	* are continually re-created.
1121		*/
1122	<	private final Condition termination;
1122	>	private static final long SHRINK_RATE =
1123	>	4L * 1000L * 1000L * 1000L; // 4 seconds
1124
1125		/**
1126	<	* Creation factory for worker threads.
1126	>	* The timeout value for attempted shrinkage, includes
1127	>	* some slop to cope with system timer imprecision.
1128		*/
1129	<	private final ForkJoinWorkerThreadFactory factory;
1129	>	private static final long SHRINK_TIMEOUT = SHRINK_RATE - (SHRINK_RATE / 10);
1130
1131		/**
1132	<	* The uncaught exception handler used when any worker abruptly
1133	<	* terminates.
1132	>	* The maximum stolen->joining link depth allowed in method
1133	>	* tryHelpStealer.  Must be a power of two. This value also
1134	>	* controls the maximum number of times to try to help join a task
1135	>	* without any apparent progress or change in pool state before
1136	>	* giving up and blocking (see awaitJoin).  Depths for legitimate
1137	>	* chains are unbounded, but we use a fixed constant to avoid
1138	>	* (otherwise unchecked) cycles and to bound staleness of
1139	>	* traversal parameters at the expense of sometimes blocking when
1140	>	* we could be helping.
1141		*/
1142	<	final Thread.UncaughtExceptionHandler ueh;
1142	>	private static final int MAX_HELP = 64;
1143
1144		/**
1145	<	* Prefix for assigning names to worker threads
1145	>	* Secondary time-based bound (in nanosecs) for helping attempts
1146	>	* before trying compensated blocking in awaitJoin. Used in
1147	>	* conjunction with MAX_HELP to reduce variance due to different
1148	>	* polling rates associated with different helping options. The
1149	>	* value should roughly approximate the time required to create
1150	>	* and/or activate a worker thread.
1151		*/
1152	<	private final String workerNamePrefix;
1152	>	private static final long COMPENSATION_DELAY = 1L << 18; // ~0.25 millisec
1153
1154		/**
1155	<	* Sum of per-thread steal counts, updated only when threads are
1156	<	* idle or terminating.
1155	>	* Increment for seed generators. See class ThreadLocal for
1156	>	* explanation.
1157		*/
1158	<	private volatile long stealCount;
1158	>	private static final int SEED_INCREMENT = 0x61c88647;
1159
1160		/**
1161	<	* Main pool control -- a long packed with:
1161	>	* Bits and masks for control variables
1162	>	*
1163	>	* Field ctl is a long packed with:
1164		* AC: Number of active running workers minus target parallelism (16 bits)
1165		* TC: Number of total workers minus target parallelism (16 bits)
1166		* ST: true if pool is terminating (1 bit)
1167		* EC: the wait count of top waiting thread (15 bits)
1168	<	* ID: ~poolIndex of top of Treiber stack of waiting threads (16 bits)
1168	>	* ID: poolIndex of top of Treiber stack of waiters (16 bits)
1169		*
1170		* When convenient, we can extract the upper 32 bits of counts and
1171		* the lower 32 bits of queue state, u = (int)(ctl >>> 32) and e =
#	Line 477 | Line 1174 | public class ForkJoinPool extends Abstra
1174		* parallelism and the positionings of fields makes it possible to
1175		* perform the most common checks via sign tests of fields: When
1176		* ac is negative, there are not enough active workers, when tc is
1177	<	* negative, there are not enough total workers, when id is
481	<	* negative, there is at least one waiting worker, and when e is
1177	>	* negative, there are not enough total workers, and when e is
1178		* negative, the pool is terminating.  To deal with these possibly
1179		* negative fields, we use casts in and out of "short" and/or
1180		* signed shifts to maintain signedness.
1181	+	*
1182	+	* When a thread is queued (inactivated), its eventCount field is
1183	+	* set negative, which is the only way to tell if a worker is
1184	+	* prevented from executing tasks, even though it must continue to
1185	+	* scan for them to avoid queuing races. Note however that
1186	+	* eventCount updates lag releases so usage requires care.
1187	+	*
1188	+	* Field runState is an int packed with:
1189	+	* SHUTDOWN: true if shutdown is enabled (1 bit)
1190	+	* SEQ:  a sequence number updated upon (de)registering workers (30 bits)
1191	+	* INIT: set true after workQueues array construction (1 bit)
1192	+	*
1193	+	* The sequence number enables simple consistency checks:
1194	+	* Staleness of read-only operations on the workQueues array can
1195	+	* be checked by comparing runState before vs after the reads.
1196		*/
486	–	volatile long ctl;
1197
1198		// bit positions/shifts for fields
1199		private static final int  AC_SHIFT   = 48;
#	Line 492 | Line 1202 | public class ForkJoinPool extends Abstra
1202		private static final int  EC_SHIFT   = 16;
1203
1204		// bounds
1205	<	private static final int  MAX_ID     = 0x7fff;  // max poolIndex
1206	<	private static final int  SMASK      = 0xffff;  // mask short bits
1205	>	private static final int  SMASK      = 0xffff;  // short bits
1206	>	private static final int  MAX_CAP    = 0x7fff;  // max #workers - 1
1207	>	private static final int  SQMASK     = 0xfffe;  // even short bits
1208		private static final int  SHORT_SIGN = 1 << 15;
1209		private static final int  INT_SIGN   = 1 << 31;
1210
#	Line 515 | Line 1226 | public class ForkJoinPool extends Abstra
1226		private static final int  UTC_UNIT   = 1 << UTC_SHIFT;
1227
1228		// masks and units for dealing with e = (int)ctl
1229	<	private static final int  E_MASK     = 0x7fffffff; // no STOP_BIT
1230	<	private static final int  EC_UNIT    = 1 << EC_SHIFT;
1229	>	private static final int E_MASK      = 0x7fffffff; // no STOP_BIT
1230	>	private static final int E_SEQ       = 1 << EC_SHIFT;
1231
1232	<	/**
1233	<	* The target parallelism level.
523	<	*/
524	<	final int parallelism;
1232	>	// runState bits
1233	>	private static final int SHUTDOWN    = 1 << 31;
1234
1235	<	/**
1236	<	* Index (mod submission queue length) of next element to take
1237	<	* from submission queue. Usage is identical to that for
1238	<	* per-worker queues -- see ForkJoinWorkerThread internal
530	<	* documentation.
531	<	*/
532	<	volatile int queueBase;
1235	>	// access mode for WorkQueue
1236	>	static final int LIFO_QUEUE          =  0;
1237	>	static final int FIFO_QUEUE          =  1;
1238	>	static final int SHARED_QUEUE        = -1;
1239
1240	<	/**
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;
1240	>	// Instance fields
1241
1242	<	/**
1243	<	* True when shutdown() has been called.
1244	<	*/
1245	<	volatile boolean shutdown;
1242	>	/*
1243	>	* Field layout order in this class tends to matter more than one
1244	>	* would like. Runtime layout order is only loosely related to
1245	>	* declaration order and may differ across JVMs, but the following
1246	>	* empirically works OK on current JVMs.
1247	>	*/
1248	>
1249	>	volatile long ctl;                         // main pool control
1250	>	final int parallelism;                     // parallelism level
1251	>	final int localMode;                       // per-worker scheduling mode
1252	>	final int submitMask;                      // submit queue index bound
1253	>	int nextSeed;                              // for initializing worker seeds
1254	>	volatile int runState;                     // shutdown status and seq
1255	>	WorkQueue[] workQueues;                    // main registry
1256	>	final Mutex lock;                          // for registration
1257	>	final Condition termination;               // for awaitTermination
1258	>	final ForkJoinWorkerThreadFactory factory; // factory for new workers
1259	>	final Thread.UncaughtExceptionHandler ueh; // per-worker UEH
1260	>	final AtomicLong stealCount;               // collect counts when terminated
1261	>	final AtomicInteger nextWorkerNumber;      // to create worker name string
1262	>	final String workerNamePrefix;             // to create worker name string
1263
1264	<	/**
548	<	* True if use local fifo, not default lifo, for local polling
549	<	* Read by, and replicated by ForkJoinWorkerThreads
550	<	*/
551	<	final boolean locallyFifo;
1264	>	//  Creating, registering, and deregistering workers
1265
1266		/**
1267	<	* The number of threads in ForkJoinWorkerThreads.helpQuiescePool.
555	<	* When non-zero, suppresses automatic shutdown when active
556	<	* counts become zero.
1267	>	* Tries to create and start a worker
1268		*/
1269	<	volatile int quiescerCount;
1269	>	private void addWorker() {
1270	>	Throwable ex = null;
1271	>	ForkJoinWorkerThread wt = null;
1272	>	try {
1273	>	if ((wt = factory.newThread(this)) != null) {
1274	>	wt.start();
1275	>	return;
1276	>	}
1277	>	} catch (Throwable e) {
1278	>	ex = e;
1279	>	}
1280	>	deregisterWorker(wt, ex); // adjust counts etc on failure
1281	>	}
1282
1283		/**
1284	<	* The number of threads blocked in join.
1284	>	* Callback from ForkJoinWorkerThread constructor to assign a
1285	>	* public name. This must be separate from registerWorker because
1286	>	* it is called during the "super" constructor call in
1287	>	* ForkJoinWorkerThread.
1288		*/
1289	<	volatile int blockedCount;
1289	>	final String nextWorkerName() {
1290	>	return workerNamePrefix.concat
1291	>	(Integer.toString(nextWorkerNumber.addAndGet(1)));
1292	>	}
1293
1294		/**
1295	<	* Counter for worker Thread names (unrelated to their poolIndex)
1295	>	* Callback from ForkJoinWorkerThread constructor to establish its
1296	>	* poolIndex and record its WorkQueue. To avoid scanning bias due
1297	>	* to packing entries in front of the workQueues array, we treat
1298	>	* the array as a simple power-of-two hash table using per-thread
1299	>	* seed as hash, expanding as needed.
1300	>	*
1301	>	* @param w the worker's queue
1302		*/
568	–	private volatile int nextWorkerNumber;
1303
1304	<	/**
1305	<	* The index for the next created worker. Accessed under scanGuard.
1306	<	*/
1307	<	private int nextWorkerIndex;
1304	>	final void registerWorker(WorkQueue w) {
1305	>	Mutex lock = this.lock;
1306	>	lock.lock();
1307	>	try {
1308	>	WorkQueue[] ws = workQueues;
1309	>	if (w != null && ws != null) {          // skip on shutdown/failure
1310	>	int rs, n =  ws.length, m = n - 1;
1311	>	int s = nextSeed += SEED_INCREMENT; // rarely-colliding sequence
1312	>	w.seed = (s == 0) ? 1 : s;          // ensure non-zero seed
1313	>	int r = (s << 1) | 1;               // use odd-numbered indices
1314	>	if (ws[r &= m] != null) {           // collision
1315	>	int probes = 0;                 // step by approx half size
1316	>	int step = (n <= 4) ? 2 : ((n >>> 1) & SQMASK) + 2;
1317	>	while (ws[r = (r + step) & m] != null) {
1318	>	if (++probes >= n) {
1319	>	workQueues = ws = Arrays.copyOf(ws, n <<= 1);
1320	>	m = n - 1;
1321	>	probes = 0;
1322	>	}
1323	>	}
1324	>	}
1325	>	w.eventCount = w.poolIndex = r;     // establish before recording
1326	>	ws[r] = w;                          // also update seq
1327	>	runState = ((rs = runState) & SHUTDOWN) | ((rs + 2) & ~SHUTDOWN);
1328	>	}
1329	>	} finally {
1330	>	lock.unlock();
1331	>	}
1332	>	}
1333
1334		/**
1335	<	* SeqLock and index masking for updates to workers array.  Locked
1336	<	* when SG_UNIT is set. Unlocking clears bit by adding
1337	<	* SG_UNIT. Staleness of read-only operations can be checked by
1338	<	* comparing scanGuard to value before the reads. The low 16 bits
1339	<	* (i.e, anding with SMASK) hold (the smallest power of two
1340	<	* covering all worker indices, minus one, and is used to avoid
1341	<	* dealing with large numbers of null slots when the workers array
583	<	* is overallocated.
1335	>	* Final callback from terminating worker, as well as upon failure
1336	>	* to construct or start a worker in addWorker.  Removes record of
1337	>	* worker from array, and adjusts counts. If pool is shutting
1338	>	* down, tries to complete termination.
1339	>	*
1340	>	* @param wt the worker thread or null if addWorker failed
1341	>	* @param ex the exception causing failure, or null if none
1342		*/
1343	<	volatile int scanGuard;
1344	<
1345	<	private static final int SG_UNIT = 1 << 16;
1343	>	final void deregisterWorker(ForkJoinWorkerThread wt, Throwable ex) {
1344	>	Mutex lock = this.lock;
1345	>	WorkQueue w = null;
1346	>	if (wt != null && (w = wt.workQueue) != null) {
1347	>	w.runState = -1;                // ensure runState is set
1348	>	stealCount.getAndAdd(w.totalSteals + w.nsteals);
1349	>	int idx = w.poolIndex;
1350	>	lock.lock();
1351	>	try {                           // remove record from array
1352	>	WorkQueue[] ws = workQueues;
1353	>	if (ws != null && idx >= 0 && idx < ws.length && ws[idx] == w)
1354	>	ws[idx] = null;
1355	>	} finally {
1356	>	lock.unlock();
1357	>	}
1358	>	}
1359
1360	<	/**
1361	<	* The wakeup interval (in nanoseconds) for a worker waiting for a
1362	<	* task when the pool is quiescent to instead try to shrink the
1363	<	* number of workers.  The exact value does not matter too
1364	<	* much. It must be short enough to release resources during
594	<	* 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
1360	>	long c;                             // adjust ctl counts
1361	>	do {} while (!U.compareAndSwapLong
1362	>	(this, CTL, c = ctl, (((c - AC_UNIT) & AC_MASK) |
1363	>	((c - TC_UNIT) & TC_MASK) |
1364	>	(c & ~(AC_MASK|TC_MASK)))));
1365
1366	<	/**
1367	<	* Top-level loop for worker threads: On each step: if the
1368	<	* previous step swept through all queues and found no tasks, or
1369	<	* there are excess threads, then possibly blocks. Otherwise,
1370	<	* scans for and, if found, executes a task. Returns when pool
1371	<	* 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;
1366	>	if (!tryTerminate(false, false) && w != null) {
1367	>	w.cancelAll();                  // cancel remaining tasks
1368	>	if (w.array != null)            // suppress signal if never ran
1369	>	signalWork();               // wake up or create replacement
1370	>	if (ex == null)                 // help clean refs on way out
1371	>	ForkJoinTask.helpExpungeStaleExceptions();
1372		}
1373	+
1374	+	if (ex != null)                     // rethrow
1375	+	U.throwException(ex);
1376		}
1377
1378	<	// Signalling
1378	>
1379	>	// Submissions
1380
1381		/**
1382	<	* Wakes up or creates a worker.
1383	<	*/
1384	<	final void signalWork() {
1385	<	/*
1386	<	* The while condition is true if: (there is are too few total
1387	<	* workers OR there is at least one waiter) AND (there are too
1388	<	* few active workers OR the pool is terminating).  The value
1389	<	* of e distinguishes the remaining cases: zero (no waiters)
1390	<	* for create, negative if terminating (in which case do
1391	<	* nothing), else release a waiter. The secondary checks for
1392	<	* release (non-null array etc) can fail if the pool begins
1393	<	* terminating after the test, and don't impose any added cost
1394	<	* because JVMs must perform null and bounds checks anyway.
1395	<	*/
1396	<	long c; int e, u;
1397	<	while ((((e = (int)(c = ctl)) | (u = (int)(c >>> 32))) &
1398	<	(INT_SIGN|SHORT_SIGN)) == (INT_SIGN|SHORT_SIGN) && e >= 0) {
1399	<	if (e > 0) {                         // release a waiting worker
1400	<	int i; ForkJoinWorkerThread w; ForkJoinWorkerThread[] ws;
1401	<	if ((ws = workers) == null ||
1402	<	(i = ~e & SMASK) >= ws.length ||
1403	<	(w = ws[i]) == null)
1404	<	break;
1405	<	long nc = (((long)(w.nextWait & E_MASK)) |
1406	<	((long)(u + UAC_UNIT) << 32));
1407	<	if (w.eventCount == e &&
1408	<	UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
1409	<	w.eventCount = (e + EC_UNIT) & E_MASK;
652	<	if (w.parked)
653	<	UNSAFE.unpark(w);
654	<	break;
1382	>	* Unless shutting down, adds the given task to a submission queue
1383	>	* at submitter's current queue index (modulo submission
1384	>	* range). If no queue exists at the index, one is created.  If
1385	>	* the queue is busy, another index is randomly chosen. The
1386	>	* submitMask bounds the effective number of queues to the
1387	>	* (nearest power of two for) parallelism level.
1388	>	*
1389	>	* @param task the task. Caller must ensure non-null.
1390	>	*/
1391	>	private void doSubmit(ForkJoinTask<?> task) {
1392	>	Submitter s = submitters.get();
1393	>	for (int r = s.seed, m = submitMask;;) {
1394	>	WorkQueue[] ws; WorkQueue q;
1395	>	int k = r & m & SQMASK;          // use only even indices
1396	>	if (runState < 0 || (ws = workQueues) == null || ws.length <= k)
1397	>	throw new RejectedExecutionException(); // shutting down
1398	>	else if ((q = ws[k]) == null) {  // create new queue
1399	>	WorkQueue nq = new WorkQueue(this, null, SHARED_QUEUE);
1400	>	Mutex lock = this.lock;      // construct outside lock
1401	>	lock.lock();
1402	>	try {                        // recheck under lock
1403	>	int rs = runState;       // to update seq
1404	>	if (ws == workQueues && ws[k] == null) {
1405	>	ws[k] = nq;
1406	>	runState = ((rs & SHUTDOWN) | ((rs + 2) & ~SHUTDOWN));
1407	>	}
1408	>	} finally {
1409	>	lock.unlock();
1410		}
1411		}
1412	<	else if (UNSAFE.compareAndSwapLong
1413	<	(this, ctlOffset, c,
1414	<	(long)(((u + UTC_UNIT) & UTC_MASK) |
1415	<	((u + UAC_UNIT) & UAC_MASK)) << 32)) {
1416	<	addWorker();
1417	<	break;
1412	>	else if (q.trySharedPush(task)) {
1413	>	signalWork();
1414	>	return;
1415	>	}
1416	>	else if (m > 1) {                // move to a different index
1417	>	r ^= r << 13;                // same xorshift as WorkQueues
1418	>	r ^= r >>> 17;
1419	>	s.seed = r ^= r << 5;
1420		}
1421	+	else
1422	+	Thread.yield();              // yield if no alternatives
1423		}
1424		}
1425
1426	+	// Maintaining ctl counts
1427	+
1428		/**
1429	<	* Variant of signalWork to help release waiters on rescans.
669	<	* Tries once to release a waiter if active count < 0.
670	<	*
671	<	* @return false if failed due to contention, else true
1429	>	* Increments active count; mainly called upon return from blocking.
1430		*/
1431	<	private boolean tryReleaseWaiter() {
1432	<	long c; int e, i; ForkJoinWorkerThread w; ForkJoinWorkerThread[] ws;
1433	<	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;
1431	>	final void incrementActiveCount() {
1432	>	long c;
1433	>	do {} while (!U.compareAndSwapLong(this, CTL, c = ctl, c + AC_UNIT));
1434		}
1435
692	–	// Scanning for tasks
693	–
1436		/**
1437	<	* Scans for and, if found, executes one task. Scans start at a
1438	<	* random index of workers array, and randomly select the first
1439	<	* (2*#workers)-1 probes, and then, if all empty, resort to 2
1440	<	* circular sweeps, which is necessary to check quiescence. and
1441	<	* taking a submission only if no stealable tasks were found.  The
1442	<	* steal code inside the loop is a specialized form of
1443	<	* ForkJoinWorkerThread.deqTask, followed bookkeeping to support
1444	<	* helpJoinTask and signal propagation. The code for submission
1445	<	* queues is almost identical. On each steal, the worker completes
1446	<	* not only the task, but also all local tasks that this task may
1447	<	* have generated. On detecting staleness or contention when
1448	<	* trying to take a task, this method returns without finishing
1449	<	* sweep, which allows global state rechecks before retry.
1450	<	*
1451	<	* @param w the worker
1452	<	* @param a the number of active workers
1453	<	* @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);
1437	>	* Tries to activate or create a worker if too few are active.
1438	>	*/
1439	>	final void signalWork() {
1440	>	long c; int u;
1441	>	while ((u = (int)((c = ctl) >>> 32)) < 0) {     // too few active
1442	>	WorkQueue[] ws = workQueues; int e, i; WorkQueue w; Thread p;
1443	>	if ((e = (int)c) > 0) {                     // at least one waiting
1444	>	if (ws != null && (i = e & SMASK) < ws.length &&
1445	>	(w = ws[i]) != null && w.eventCount == (e | INT_SIGN)) {
1446	>	long nc = (((long)(w.nextWait & E_MASK)) |
1447	>	((long)(u + UAC_UNIT) << 32));
1448	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1449	>	w.eventCount = (e + E_SEQ) & E_MASK;
1450	>	if ((p = w.parker) != null)
1451	>	U.unpark(p);                // activate and release
1452	>	break;
1453	>	}
1454		}
1455	<	r ^= r << 13; r ^= r >>> 17; w.seed = r ^ (r << 5);
1456	<	return false;                     // store next seed
1455	>	else
1456	>	break;
1457		}
1458	<	else if (j < 0) {                     // xorshift
1459	<	r ^= r << 13; r ^= r >>> 17; k = r ^= r << 5;
1458	>	else if (e == 0 && (u & SHORT_SIGN) != 0) { // too few total
1459	>	long nc = (long)(((u + UTC_UNIT) & UTC_MASK) |
1460	>	((u + UAC_UNIT) & UAC_MASK)) << 32;
1461	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1462	>	addWorker();
1463	>	break;
1464	>	}
1465		}
1466		else
1467	<	++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
1467	>	break;
1468		}
1469		}
1470
1471	+	// Scanning for tasks
1472	+
1473		/**
1474	<	* Tries to enqueue worker w in wait queue and await change in
763	<	* 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.
776	<	*
777	<	* @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
1474	>	* Top-level runloop for workers, called by ForkJoinWorkerThread.run.
1475		*/
1476	<	private boolean tryAwaitWork(ForkJoinWorkerThread w, long c) {
1477	<	int v = w.eventCount;
1478	<	w.nextWait = (int)c;                      // w's successor record
1479	<	long nc = (long)(v & E_MASK) | ((c - AC_UNIT) & (AC_MASK|TC_MASK));
1480	<	if (ctl != c || !UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
1481	<	long d = ctl; // return true if lost to a deq, to force scan
1482	<	return (int)d != (int)c && ((d - c) & AC_MASK) >= 0L;
1483	<	}
1484	<	for (int sc = w.stealCount; sc != 0;) {   // accumulate stealCount
1485	<	long s = stealCount;
1486	<	if (UNSAFE.compareAndSwapLong(this, stealCountOffset, s, s + sc))
1487	<	sc = w.stealCount = 0;
1488	<	else if (w.eventCount != v)
1489	<	return true;                      // update next time
1490	<	}
1491	<	if ((!shutdown || !tryTerminate(false)) &&
1492	<	(int)c != 0 && parallelism + (int)(nc >> AC_SHIFT) == 0 &&
1493	<	blockedCount == 0 && quiescerCount == 0)
1494	<	idleAwaitWork(w, nc, c, v);           // quiescent
1495	<	for (boolean rescanned = false;;) {
1496	<	if (w.eventCount != v)
1497	<	return true;
1498	<	if (!rescanned) {
1499	<	int g = scanGuard, m = g & SMASK;
1500	<	ForkJoinWorkerThread[] ws = workers;
1501	<	if (ws != null && m < ws.length) {
1502	<	rescanned = true;
1503	<	for (int i = 0; i <= m; ++i) {
1504	<	ForkJoinWorkerThread u = ws[i];
1505	<	if (u != null) {
1506	<	if (u.queueBase != u.queueTop &&
1507	<	!tryReleaseWaiter())
1508	<	rescanned = false; // contended
1509	<	if (w.eventCount != v)
1510	<	return true;
1511	<	}
1476	>	final void runWorker(WorkQueue w) {
1477	>	w.growArray(false);         // initialize queue array in this thread
1478	>	do { w.runTask(scan(w)); } while (w.runState >= 0);
1479	>	}
1480	>
1481	>	/**
1482	>	* Scans for and, if found, returns one task, else possibly
1483	>	* inactivates the worker. This method operates on single reads of
1484	>	* volatile state and is designed to be re-invoked continuously,
1485	>	* in part because it returns upon detecting inconsistencies,
1486	>	* contention, or state changes that indicate possible success on
1487	>	* re-invocation.
1488	>	*
1489	>	* The scan searches for tasks across a random permutation of
1490	>	* queues (starting at a random index and stepping by a random
1491	>	* relative prime, checking each at least once).  The scan
1492	>	* terminates upon either finding a non-empty queue, or completing
1493	>	* the sweep. If the worker is not inactivated, it takes and
1494	>	* returns a task from this queue.  On failure to find a task, we
1495	>	* take one of the following actions, after which the caller will
1496	>	* retry calling this method unless terminated.
1497	>	*
1498	>	* * If pool is terminating, terminate the worker.
1499	>	*
1500	>	* * If not a complete sweep, try to release a waiting worker.  If
1501	>	* the scan terminated because the worker is inactivated, then the
1502	>	* released worker will often be the calling worker, and it can
1503	>	* succeed obtaining a task on the next call. Or maybe it is
1504	>	* another worker, but with same net effect. Releasing in other
1505	>	* cases as well ensures that we have enough workers running.
1506	>	*
1507	>	* * If not already enqueued, try to inactivate and enqueue the
1508	>	* worker on wait queue. Or, if inactivating has caused the pool
1509	>	* to be quiescent, relay to idleAwaitWork to check for
1510	>	* termination and possibly shrink pool.
1511	>	*
1512	>	* * If already inactive, and the caller has run a task since the
1513	>	* last empty scan, return (to allow rescan) unless others are
1514	>	* also inactivated.  Field WorkQueue.rescans counts down on each
1515	>	* scan to ensure eventual inactivation and blocking.
1516	>	*
1517	>	* * If already enqueued and none of the above apply, park
1518	>	* awaiting signal,
1519	>	*
1520	>	* @param w the worker (via its WorkQueue)
1521	>	* @return a task or null of none found
1522	>	*/
1523	>	private final ForkJoinTask<?> scan(WorkQueue w) {
1524	>	WorkQueue[] ws;                       // first update random seed
1525	>	int r = w.seed; r ^= r << 13; r ^= r >>> 17; w.seed = r ^= r << 5;
1526	>	int rs = runState, m;                 // volatile read order matters
1527	>	if ((ws = workQueues) != null && (m = ws.length - 1) > 0) {
1528	>	int ec = w.eventCount;            // ec is negative if inactive
1529	>	int step = (r >>> 16) | 1;        // relative prime
1530	>	for (int j = (m + 1) << 2; ; r += step) {
1531	>	WorkQueue q; ForkJoinTask<?> t; ForkJoinTask<?>[] a; int b;
1532	>	if ((q = ws[r & m]) != null && (b = q.base) - q.top < 0 &&
1533	>	(a = q.array) != null) {  // probably nonempty
1534	>	int i = (((a.length - 1) & b) << ASHIFT) + ABASE;
1535	>	t = (ForkJoinTask<?>)U.getObjectVolatile(a, i);
1536	>	if (q.base == b && ec >= 0 && t != null &&
1537	>	U.compareAndSwapObject(a, i, t, null)) {
1538	>	if (q.top - (q.base = b + 1) > 1)
1539	>	signalWork();    // help pushes signal
1540	>	return t;
1541	>	}
1542	>	else if (ec < 0 || j <= m) {
1543	>	rs = 0;               // mark scan as imcomplete
1544	>	break;                // caller can retry after release
1545		}
1546		}
1547	<	if (scanGuard != g ||              // stale
1548	<	(queueBase != queueTop && !tryReleaseWaiter()))
821	<	rescanned = false;
822	<	if (!rescanned)
823	<	Thread.yield();                // reduce contention
824	<	else
825	<	Thread.interrupted();          // clear before park
1547	>	if (--j < 0)
1548	>	break;
1549		}
1550	<	else {
1551	<	w.parked = true;                   // must recheck
1552	<	if (w.eventCount != v) {
1553	<	w.parked = false;
1554	<	return true;
1550	>
1551	>	long c = ctl; int e = (int)c, a = (int)(c >> AC_SHIFT), nr, ns;
1552	>	if (e < 0)                        // decode ctl on empty scan
1553	>	w.runState = -1;              // pool is terminating
1554	>	else if (rs == 0 || rs != runState) { // incomplete scan
1555	>	WorkQueue v; Thread p;        // try to release a waiter
1556	>	if (e > 0 && a < 0 && w.eventCount == ec &&
1557	>	(v = ws[e & m]) != null && v.eventCount == (e | INT_SIGN)) {
1558	>	long nc = ((long)(v.nextWait & E_MASK) |
1559	>	((c + AC_UNIT) & (AC_MASK|TC_MASK)));
1560	>	if (ctl == c && U.compareAndSwapLong(this, CTL, c, nc)) {
1561	>	v.eventCount = (e + E_SEQ) & E_MASK;
1562	>	if ((p = v.parker) != null)
1563	>	U.unpark(p);
1564	>	}
1565	>	}
1566	>	}
1567	>	else if (ec >= 0) {               // try to enqueue/inactivate
1568	>	long nc = (long)ec | ((c - AC_UNIT) & (AC_MASK|TC_MASK));
1569	>	w.nextWait = e;
1570	>	w.eventCount = ec | INT_SIGN; // mark as inactive
1571	>	if (ctl != c || !U.compareAndSwapLong(this, CTL, c, nc))
1572	>	w.eventCount = ec;        // unmark on CAS failure
1573	>	else {
1574	>	if ((ns = w.nsteals) != 0) {
1575	>	w.nsteals = 0;        // set rescans if ran task
1576	>	w.rescans = (a > 0) ? 0 : a + parallelism;
1577	>	w.totalSteals += ns;
1578	>	}
1579	>	if (a == 1 - parallelism) // quiescent
1580	>	idleAwaitWork(w, nc, c);
1581	>	}
1582	>	}
1583	>	else if (w.eventCount < 0) {      // already queued
1584	>	if ((nr = w.rescans) > 0) {   // continue rescanning
1585	>	int ac = a + parallelism;
1586	>	if (((w.rescans = (ac < nr) ? ac : nr - 1) & 3) == 0)
1587	>	Thread.yield();       // yield before block
1588	>	}
1589	>	else {
1590	>	Thread.interrupted();     // clear status
1591	>	Thread wt = Thread.currentThread();
1592	>	U.putObject(wt, PARKBLOCKER, this);
1593	>	w.parker = wt;            // emulate LockSupport.park
1594	>	if (w.eventCount < 0)     // recheck
1595	>	U.park(false, 0L);
1596	>	w.parker = null;
1597	>	U.putObject(wt, PARKBLOCKER, null);
1598		}
833	–	LockSupport.park(this);
834	–	rescanned = w.parked = false;
1599		}
1600		}
1601	+	return null;
1602		}
1603
1604		/**
1605	<	* If inactivating worker w has caused pool to become
1606	<	* quiescent, check for pool termination, and wait for event
1607	<	* for up to SHRINK_RATE nanosecs (rescans are unnecessary in
1608	<	* this case because quiescence reflects consensus about lack
1609	<	* of work). On timeout, if ctl has not changed, terminate the
1610	<	* worker. Upon its termination (see deregisterWorker), it may
846	<	* wake up another worker to possibly repeat this process.
1605	>	* If inactivating worker w has caused the pool to become
1606	>	* quiescent, checks for pool termination, and, so long as this is
1607	>	* not the only worker, waits for event for up to SHRINK_RATE
1608	>	* nanosecs.  On timeout, if ctl has not changed, terminates the
1609	>	* worker, which will in turn wake up another worker to possibly
1610	>	* repeat this process.
1611		*
1612		* @param w the calling worker
1613	<	* @param currentCtl the ctl value after enqueuing w
1614	<	* @param prevCtl the ctl value if w terminated
1615	<	* @param v the eventCount w awaits change
1616	<	*/
1617	<	private void idleAwaitWork(ForkJoinWorkerThread w, long currentCtl,
1618	<	long prevCtl, int v) {
1619	<	if (w.eventCount == v) {
1620	<	if (shutdown)
857	<	tryTerminate(false);
858	<	ForkJoinTask.helpExpungeStaleExceptions(); // help clean weak refs
1613	>	* @param currentCtl the ctl value triggering possible quiescence
1614	>	* @param prevCtl the ctl value to restore if thread is terminated
1615	>	*/
1616	>	private void idleAwaitWork(WorkQueue w, long currentCtl, long prevCtl) {
1617	>	if (w.eventCount < 0 && !tryTerminate(false, false) &&
1618	>	(int)prevCtl != 0 && !hasQueuedSubmissions() && ctl == currentCtl) {
1619	>	Thread wt = Thread.currentThread();
1620	>	Thread.yield();            // yield before block
1621		while (ctl == currentCtl) {
1622		long startTime = System.nanoTime();
1623	<	w.parked = true;
1624	<	if (w.eventCount == v)             // must recheck
1625	<	LockSupport.parkNanos(this, SHRINK_RATE);
1626	<	w.parked = false;
1627	<	if (w.eventCount != v)
1623	>	Thread.interrupted();  // timed variant of version in scan()
1624	>	U.putObject(wt, PARKBLOCKER, this);
1625	>	w.parker = wt;
1626	>	if (ctl == currentCtl)
1627	>	U.park(false, SHRINK_RATE);
1628	>	w.parker = null;
1629	>	U.putObject(wt, PARKBLOCKER, null);
1630	>	if (ctl != currentCtl)
1631		break;
1632	<	else if (System.nanoTime() - startTime <
1633	<	SHRINK_RATE - (SHRINK_RATE / 10)) // timing slop
1634	<	Thread.interrupted();          // spurious wakeup
1635	<	else if (UNSAFE.compareAndSwapLong(this, ctlOffset,
871	<	currentCtl, prevCtl)) {
872	<	w.terminate = true;            // restore previous
873	<	w.eventCount = ((int)currentCtl + EC_UNIT) & E_MASK;
1632	>	if (System.nanoTime() - startTime >= SHRINK_TIMEOUT &&
1633	>	U.compareAndSwapLong(this, CTL, currentCtl, prevCtl)) {
1634	>	w.eventCount = (w.eventCount + E_SEQ) | E_MASK;
1635	>	w.runState = -1;   // shrink
1636		break;
1637		}
1638		}
1639		}
1640		}
1641
880	–	// Submissions
881	–
1642		/**
1643	<	* Enqueues the given task in the submissionQueue.  Same idea as
1644	<	* ForkJoinWorkerThread.pushTask except for use of submissionLock.
1645	<	*
1646	<	* @param t the task
1647	<	*/
1648	<	private void addSubmission(ForkJoinTask<?> t) {
1649	<	final ReentrantLock lock = this.submissionLock;
1650	<	lock.lock();
1651	<	try {
1652	<	ForkJoinTask<?>[] q; int s, m;
1653	<	if ((q = submissionQueue) != null) {    // ignore if queue removed
1654	<	long u = (((s = queueTop) & (m = q.length-1)) << ASHIFT)+ABASE;
1655	<	UNSAFE.putOrderedObject(q, u, t);
1656	<	queueTop = s + 1;
1657	<	if (s - queueBase == m)
1658	<	growSubmissionQueue();
1643	>	* Tries to locate and execute tasks for a stealer of the given
1644	>	* task, or in turn one of its stealers, Traces currentSteal ->
1645	>	* currentJoin links looking for a thread working on a descendant
1646	>	* of the given task and with a non-empty queue to steal back and
1647	>	* execute tasks from. The first call to this method upon a
1648	>	* waiting join will often entail scanning/search, (which is OK
1649	>	* because the joiner has nothing better to do), but this method
1650	>	* leaves hints in workers to speed up subsequent calls. The
1651	>	* implementation is very branchy to cope with potential
1652	>	* inconsistencies or loops encountering chains that are stale,
1653	>	* unknown, or so long that they are likely cyclic.
1654	>	*
1655	>	* @param joiner the joining worker
1656	>	* @param task the task to join
1657	>	* @return 0 if no progress can be made, negative if task
1658	>	* known complete, else positive
1659	>	*/
1660	>	private int tryHelpStealer(WorkQueue joiner, ForkJoinTask<?> task) {
1661	>	int stat = 0, steps = 0;                    // bound to avoid cycles
1662	>	if (joiner != null && task != null) {       // hoist null checks
1663	>	restart: for (;;) {
1664	>	ForkJoinTask<?> subtask = task;     // current target
1665	>	for (WorkQueue j = joiner, v;;) {   // v is stealer of subtask
1666	>	WorkQueue[] ws; int m, s, h;
1667	>	if ((s = task.status) < 0) {
1668	>	stat = s;
1669	>	break restart;
1670	>	}
1671	>	if ((ws = workQueues) == null || (m = ws.length - 1) <= 0)
1672	>	break restart;              // shutting down
1673	>	if ((v = ws[h = (j.stealHint | 1) & m]) == null ||
1674	>	v.currentSteal != subtask) {
1675	>	for (int origin = h;;) {    // find stealer
1676	>	if (((h = (h + 2) & m) & 15) == 1 &&
1677	>	(subtask.status < 0 || j.currentJoin != subtask))
1678	>	continue restart;   // occasional staleness check
1679	>	if ((v = ws[h]) != null &&
1680	>	v.currentSteal == subtask) {
1681	>	j.stealHint = h;    // save hint
1682	>	break;
1683	>	}
1684	>	if (h == origin)
1685	>	break restart;      // cannot find stealer
1686	>	}
1687	>	}
1688	>	for (;;) { // help stealer or descend to its stealer
1689	>	ForkJoinTask[] a;  int b;
1690	>	if (subtask.status < 0)     // surround probes with
1691	>	continue restart;       //   consistency checks
1692	>	if ((b = v.base) - v.top < 0 && (a = v.array) != null) {
1693	>	int i = (((a.length - 1) & b) << ASHIFT) + ABASE;
1694	>	ForkJoinTask<?> t =
1695	>	(ForkJoinTask<?>)U.getObjectVolatile(a, i);
1696	>	if (subtask.status < 0 || j.currentJoin != subtask ||
1697	>	v.currentSteal != subtask)
1698	>	continue restart;   // stale
1699	>	stat = 1;               // apparent progress
1700	>	if (t != null && v.base == b &&
1701	>	U.compareAndSwapObject(a, i, t, null)) {
1702	>	v.base = b + 1;     // help stealer
1703	>	joiner.runSubtask(t);
1704	>	}
1705	>	else if (v.base == b && ++steps == MAX_HELP)
1706	>	break restart;      // v apparently stalled
1707	>	}
1708	>	else {                      // empty -- try to descend
1709	>	ForkJoinTask<?> next = v.currentJoin;
1710	>	if (subtask.status < 0 || j.currentJoin != subtask ||
1711	>	v.currentSteal != subtask)
1712	>	continue restart;   // stale
1713	>	else if (next == null || ++steps == MAX_HELP)
1714	>	break restart;      // dead-end or maybe cyclic
1715	>	else {
1716	>	subtask = next;
1717	>	j = v;
1718	>	break;
1719	>	}
1720	>	}
1721	>	}
1722	>	}
1723		}
900	–	} finally {
901	–	lock.unlock();
1724		}
1725	<	signalWork();
1725	>	return stat;
1726		}
1727
906	–	//  (pollSubmission is defined below with exported methods)
907	–
1728		/**
1729	<	* Creates or doubles submissionQueue array.
1730	<	* Basically identical to ForkJoinWorkerThread version.
1729	>	* If task is at base of some steal queue, steals and executes it.
1730	>	*
1731	>	* @param joiner the joining worker
1732	>	* @param task the task
1733		*/
1734	<	private void growSubmissionQueue() {
1735	<	ForkJoinTask<?>[] oldQ = submissionQueue;
1736	<	int size = oldQ != null ? oldQ.length << 1 : INITIAL_QUEUE_CAPACITY;
1737	<	if (size > MAXIMUM_QUEUE_CAPACITY)
1738	<	throw new RejectedExecutionException("Queue capacity exceeded");
1739	<	if (size < INITIAL_QUEUE_CAPACITY)
1740	<	size = INITIAL_QUEUE_CAPACITY;
1741	<	ForkJoinTask<?>[] q = submissionQueue = new ForkJoinTask<?>[size];
1742	<	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);
1734	>	private void tryPollForAndExec(WorkQueue joiner, ForkJoinTask<?> task) {
1735	>	WorkQueue[] ws;
1736	>	if ((ws = workQueues) != null) {
1737	>	for (int j = 1; j < ws.length && task.status >= 0; j += 2) {
1738	>	WorkQueue q = ws[j];
1739	>	if (q != null && q.pollFor(task)) {
1740	>	joiner.runSubtask(task);
1741	>	break;
1742	>	}
1743		}
1744		}
1745		}
1746
934	–	// Blocking support
935	–
1747		/**
1748	<	* Tries to increment blockedCount, decrement active count
1749	<	* (sometimes implicitly) and possibly release or create a
1750	<	* compensating worker in preparation for blocking. Fails
1751	<	* on contention or termination.
1748	>	* Tries to decrement active count (sometimes implicitly) and
1749	>	* possibly release or create a compensating worker in preparation
1750	>	* for blocking. Fails on contention or termination. Otherwise,
1751	>	* adds a new thread if no idle workers are available and either
1752	>	* pool would become completely starved or: (at least half
1753	>	* starved, and fewer than 50% spares exist, and there is at least
1754	>	* one task apparently available). Even though the availability
1755	>	* check requires a full scan, it is worthwhile in reducing false
1756	>	* alarms.
1757		*
1758	+	* @param task if non-null, a task being waited for
1759	+	* @param blocker if non-null, a blocker being waited for
1760		* @return true if the caller can block, else should recheck and retry
1761		*/
1762	<	private boolean tryPreBlock() {
1763	<	int b = blockedCount;
1764	<	if (UNSAFE.compareAndSwapInt(this, blockedCountOffset, b, b + 1)) {
1765	<	int pc = parallelism;
1766	<	do {
1767	<	ForkJoinWorkerThread[] ws; ForkJoinWorkerThread w;
1768	<	int e, ac, tc, i;
1769	<	long c = ctl;
1770	<	int u = (int)(c >>> 32);
1771	<	if ((e = (int)c) < 0) {
1772	<	// skip -- terminating
1773	<	}
1774	<	else if ((ac = (u >> UAC_SHIFT)) <= 0 && e != 0 &&
1775	<	(ws = workers) != null &&
1776	<	(i = ~e & SMASK) < ws.length &&
1777	<	(w = ws[i]) != null) {
1778	<	long nc = ((long)(w.nextWait & E_MASK) |
1779	<	(c & (AC_MASK|TC_MASK)));
1780	<	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
1762	>	final boolean tryCompensate(ForkJoinTask<?> task, ManagedBlocker blocker) {
1763	>	int pc = parallelism, e;
1764	>	long c = ctl;
1765	>	WorkQueue[] ws = workQueues;
1766	>	if ((e = (int)c) >= 0 && ws != null) {
1767	>	int u, a, ac, hc;
1768	>	int tc = (short)((u = (int)(c >>> 32)) >>> UTC_SHIFT) + pc;
1769	>	boolean replace = false;
1770	>	if ((a = u >> UAC_SHIFT) <= 0) {
1771	>	if ((ac = a + pc) <= 1)
1772	>	replace = true;
1773	>	else if ((e > 0 || (task != null &&
1774	>	ac <= (hc = pc >>> 1) && tc < pc + hc))) {
1775	>	WorkQueue w;
1776	>	for (int j = 0; j < ws.length; ++j) {
1777	>	if ((w = ws[j]) != null && !w.isEmpty()) {
1778	>	replace = true;
1779	>	break;   // in compensation range and tasks available
1780	>	}
1781		}
1782		}
1783	<	else if ((tc = (short)(u >>> UTC_SHIFT)) >= 0 && ac + pc > 1) {
1783	>	}
1784	>	if ((task == null || task.status >= 0) && // recheck need to block
1785	>	(blocker == null || !blocker.isReleasable()) && ctl == c) {
1786	>	if (!replace) {          // no compensation
1787		long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK);
1788	<	if (UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc))
1789	<	return true;             // no compensation needed
1788	>	if (U.compareAndSwapLong(this, CTL, c, nc))
1789	>	return true;
1790		}
1791	<	else if (tc + pc < MAX_ID) {
1791	>	else if (e != 0) {       // release an idle worker
1792	>	WorkQueue w; Thread p; int i;
1793	>	if ((i = e & SMASK) < ws.length && (w = ws[i]) != null) {
1794	>	long nc = ((long)(w.nextWait & E_MASK) |
1795	>	(c & (AC_MASK|TC_MASK)));
1796	>	if (w.eventCount == (e | INT_SIGN) &&
1797	>	U.compareAndSwapLong(this, CTL, c, nc)) {
1798	>	w.eventCount = (e + E_SEQ) & E_MASK;
1799	>	if ((p = w.parker) != null)
1800	>	U.unpark(p);
1801	>	return true;
1802	>	}
1803	>	}
1804	>	}
1805	>	else if (tc < MAX_CAP) { // create replacement
1806		long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK);
1807	<	if (UNSAFE.compareAndSwapLong(this, ctlOffset, c, nc)) {
1807	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1808		addWorker();
1809	<	return true;            // create a replacement
1809	>	return true;
1810		}
1811		}
1812	<	// try to back out on any failure and let caller retry
983	<	} while (!UNSAFE.compareAndSwapInt(this, blockedCountOffset,
984	<	b = blockedCount, b - 1));
1812	>	}
1813		}
1814		return false;
1815		}
1816
1817		/**
1818	<	* Decrements blockedCount and increments active count.
991	<	*/
992	<	private void postBlock() {
993	<	long c;
994	<	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));
999	<	}
1000	<
1001	<	/**
1002	<	* Possibly blocks waiting for the given task to complete, or
1003	<	* cancels the task if terminating.  Fails to wait if contended.
1818	>	* Helps and/or blocks until the given task is done.
1819		*
1820	<	* @param joinMe the task
1820	>	* @param joiner the joining worker
1821	>	* @param task the task
1822	>	* @return task status on exit
1823		*/
1824	<	final void tryAwaitJoin(ForkJoinTask<?> joinMe) {
1825	<	Thread.interrupted(); // clear interrupts before checking termination
1826	<	if (joinMe.status >= 0) {
1827	<	if (tryPreBlock()) {
1828	<	joinMe.tryAwaitDone(0L);
1829	<	postBlock();
1824	>	final int awaitJoin(WorkQueue joiner, ForkJoinTask<?> task) {
1825	>	int s;
1826	>	if ((s = task.status) >= 0) {
1827	>	ForkJoinTask<?> prevJoin = joiner.currentJoin;
1828	>	joiner.currentJoin = task;
1829	>	long startTime = 0L;
1830	>	for (int k = 0;;) {
1831	>	if ((s = (joiner.isEmpty() ?           // try to help
1832	>	tryHelpStealer(joiner, task) :
1833	>	joiner.tryRemoveAndExec(task))) == 0 &&
1834	>	(s = task.status) >= 0) {
1835	>	if (k == 0) {
1836	>	startTime = System.nanoTime();
1837	>	tryPollForAndExec(joiner, task); // check uncommon case
1838	>	}
1839	>	else if ((k & (MAX_HELP - 1)) == 0 &&
1840	>	System.nanoTime() - startTime >=
1841	>	COMPENSATION_DELAY &&
1842	>	tryCompensate(task, null)) {
1843	>	if (task.trySetSignal()) {
1844	>	synchronized (task) {
1845	>	if (task.status >= 0) {
1846	>	try {                // see ForkJoinTask
1847	>	task.wait();     //  for explanation
1848	>	} catch (InterruptedException ie) {
1849	>	}
1850	>	}
1851	>	else
1852	>	task.notifyAll();
1853	>	}
1854	>	}
1855	>	long c;                          // re-activate
1856	>	do {} while (!U.compareAndSwapLong
1857	>	(this, CTL, c = ctl, c + AC_UNIT));
1858	>	}
1859	>	}
1860	>	if (s < 0 || (s = task.status) < 0) {
1861	>	joiner.currentJoin = prevJoin;
1862	>	break;
1863	>	}
1864	>	else if ((k++ & (MAX_HELP - 1)) == MAX_HELP >>> 1)
1865	>	Thread.yield();                     // for politeness
1866		}
1014	–	else if ((ctl & STOP_BIT) != 0L)
1015	–	joinMe.cancelIgnoringExceptions();
1867		}
1868	+	return s;
1869		}
1870
1871		/**
1872	<	* Possibly blocks the given worker waiting for joinMe to
1873	<	* complete or timeout.
1872	>	* Stripped-down variant of awaitJoin used by timed joins. Tries
1873	>	* to help join only while there is continuous progress. (Caller
1874	>	* will then enter a timed wait.)
1875		*
1876	<	* @param joinMe the task
1877	<	* @param millis the wait time for underlying Object.wait
1876	>	* @param joiner the joining worker
1877	>	* @param task the task
1878	>	* @return task status on exit
1879		*/
1880	<	final void timedAwaitJoin(ForkJoinTask<?> joinMe, long nanos) {
1881	<	while (joinMe.status >= 0) {
1882	<	Thread.interrupted();
1883	<	if ((ctl & STOP_BIT) != 0L) {
1884	<	joinMe.cancelIgnoringExceptions();
1885	<	break;
1886	<	}
1887	<	if (tryPreBlock()) {
1034	<	long last = System.nanoTime();
1035	<	while (joinMe.status >= 0) {
1036	<	long millis = TimeUnit.NANOSECONDS.toMillis(nanos);
1037	<	if (millis <= 0)
1038	<	break;
1039	<	joinMe.tryAwaitDone(millis);
1040	<	if (joinMe.status < 0)
1041	<	break;
1042	<	if ((ctl & STOP_BIT) != 0L) {
1043	<	joinMe.cancelIgnoringExceptions();
1044	<	break;
1045	<	}
1046	<	long now = System.nanoTime();
1047	<	nanos -= now - last;
1048	<	last = now;
1049	<	}
1050	<	postBlock();
1051	<	break;
1052	<	}
1053	<	}
1880	>	final int helpJoinOnce(WorkQueue joiner, ForkJoinTask<?> task) {
1881	>	int s;
1882	>	while ((s = task.status) >= 0 &&
1883	>	(joiner.isEmpty() ?
1884	>	tryHelpStealer(joiner, task) :
1885	>	joiner.tryRemoveAndExec(task)) != 0)
1886	>	;
1887	>	return s;
1888		}
1889
1890		/**
1891	<	* If necessary, compensates for blocker, and blocks.
1892	<	*/
1893	<	private void awaitBlocker(ManagedBlocker blocker)
1894	<	throws InterruptedException {
1895	<	while (!blocker.isReleasable()) {
1896	<	if (tryPreBlock()) {
1897	<	try {
1898	<	do {} while (!blocker.isReleasable() && !blocker.block());
1899	<	} finally {
1900	<	postBlock();
1891	>	* Returns a (probably) non-empty steal queue, if one is found
1892	>	* during a random, then cyclic scan, else null.  This method must
1893	>	* be retried by caller if, by the time it tries to use the queue,
1894	>	* it is empty.
1895	>	*/
1896	>	private WorkQueue findNonEmptyStealQueue(WorkQueue w) {
1897	>	// Similar to loop in scan(), but ignoring submissions
1898	>	int r = w.seed; r ^= r << 13; r ^= r >>> 17; w.seed = r ^= r << 5;
1899	>	int step = (r >>> 16) | 1;
1900	>	for (WorkQueue[] ws;;) {
1901	>	int rs = runState, m;
1902	>	if ((ws = workQueues) == null || (m = ws.length - 1) < 1)
1903	>	return null;
1904	>	for (int j = (m + 1) << 2; ; r += step) {
1905	>	WorkQueue q = ws[((r << 1) | 1) & m];
1906	>	if (q != null && !q.isEmpty())
1907	>	return q;
1908	>	else if (--j < 0) {
1909	>	if (runState == rs)
1910	>	return null;
1911	>	break;
1912		}
1068	–	break;
1913		}
1914		}
1915		}
1916
1073	–	// Creating, registering and deregistring workers
1074	–
1075	–	/**
1076	–	* Tries to create and start a worker; minimally rolls back counts
1077	–	* on failure.
1078	–	*/
1079	–	private void addWorker() {
1080	–	Throwable ex = null;
1081	–	ForkJoinWorkerThread t = null;
1082	–	try {
1083	–	t = factory.newThread(this);
1084	–	} catch (Throwable e) {
1085	–	ex = e;
1086	–	}
1087	–	if (t == null) {  // null or exceptional factory return
1088	–	long c;       // adjust counts
1089	–	do {} while (!UNSAFE.compareAndSwapLong
1090	–	(this, ctlOffset, c = ctl,
1091	–	(((c - AC_UNIT) & AC_MASK) |
1092	–	((c - TC_UNIT) & TC_MASK) |
1093	–	(c & ~(AC_MASK|TC_MASK)))));
1094	–	// Propagate exception if originating from an external caller
1095	–	if (!tryTerminate(false) && ex != null &&
1096	–	!(Thread.currentThread() instanceof ForkJoinWorkerThread))
1097	–	UNSAFE.throwException(ex);
1098	–	}
1099	–	else
1100	–	t.start();
1101	–	}
1917
1918		/**
1919	<	* Callback from ForkJoinWorkerThread constructor to assign a
1920	<	* public name
1921	<	*/
1922	<	final String nextWorkerName() {
1923	<	for (int n;;) {
1924	<	if (UNSAFE.compareAndSwapInt(this, nextWorkerNumberOffset,
1925	<	n = nextWorkerNumber, ++n))
1926	<	return workerNamePrefix + n;
1927	<	}
1928	<	}
1929	<
1930	<	/**
1931	<	* Callback from ForkJoinWorkerThread constructor to
1932	<	* determine its poolIndex and record in workers array.
1933	<	*
1934	<	* @param w the worker
1935	<	* @return the worker's pool index
1936	<	*/
1122	<	final int registerWorker(ForkJoinWorkerThread w) {
1123	<	/*
1124	<	* In the typical case, a new worker acquires the lock, uses
1125	<	* next available index and returns quickly.  Since we should
1126	<	* not block callers (ultimately from signalWork or
1127	<	* tryPreBlock) waiting for the lock needed to do this, we
1128	<	* instead help release other workers while waiting for the
1129	<	* lock.
1130	<	*/
1131	<	for (int g;;) {
1132	<	ForkJoinWorkerThread[] ws;
1133	<	if (((g = scanGuard) & SG_UNIT) == 0 &&
1134	<	UNSAFE.compareAndSwapInt(this, scanGuardOffset,
1135	<	g, g | SG_UNIT)) {
1136	<	int k = nextWorkerIndex;
1137	<	try {
1138	<	if ((ws = workers) != null) { // ignore on shutdown
1139	<	int n = ws.length;
1140	<	if (k < 0 || k >= n || ws[k] != null) {
1141	<	for (k = 0; k < n && ws[k] != null; ++k)
1142	<	;
1143	<	if (k == n)
1144	<	ws = workers = Arrays.copyOf(ws, n << 1);
1145	<	}
1146	<	ws[k] = w;
1147	<	nextWorkerIndex = k + 1;
1148	<	int m = g & SMASK;
1149	<	g = (k > m) ? ((m << 1) + 1) & SMASK : g + (SG_UNIT<<1);
1150	<	}
1151	<	} finally {
1152	<	scanGuard = g;
1919	>	* Runs tasks until {@code isQuiescent()}. We piggyback on
1920	>	* active count ctl maintenance, but rather than blocking
1921	>	* when tasks cannot be found, we rescan until all others cannot
1922	>	* find tasks either.
1923	>	*/
1924	>	final void helpQuiescePool(WorkQueue w) {
1925	>	for (boolean active = true;;) {
1926	>	ForkJoinTask<?> localTask; // exhaust local queue
1927	>	while ((localTask = w.nextLocalTask()) != null)
1928	>	localTask.doExec();
1929	>	WorkQueue q = findNonEmptyStealQueue(w);
1930	>	if (q != null) {
1931	>	ForkJoinTask<?> t; int b;
1932	>	if (!active) {      // re-establish active count
1933	>	long c;
1934	>	active = true;
1935	>	do {} while (!U.compareAndSwapLong
1936	>	(this, CTL, c = ctl, c + AC_UNIT));
1937		}
1938	<	return k;
1938	>	if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null)
1939	>	w.runSubtask(t);
1940		}
1941	<	else if ((ws = workers) != null) { // help release others
1942	<	for (ForkJoinWorkerThread u : ws) {
1943	<	if (u != null && u.queueBase != u.queueTop) {
1944	<	if (tryReleaseWaiter())
1945	<	break;
1946	<	}
1941	>	else {
1942	>	long c;
1943	>	if (active) {       // decrement active count without queuing
1944	>	active = false;
1945	>	do {} while (!U.compareAndSwapLong
1946	>	(this, CTL, c = ctl, c -= AC_UNIT));
1947	>	}
1948	>	else
1949	>	c = ctl;        // re-increment on exit
1950	>	if ((int)(c >> AC_SHIFT) + parallelism == 0) {
1951	>	do {} while (!U.compareAndSwapLong
1952	>	(this, CTL, c = ctl, c + AC_UNIT));
1953	>	break;
1954		}
1955		}
1956		}
1957		}
1958
1959		/**
1960	<	* 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.
1960	>	* Gets and removes a local or stolen task for the given worker.
1961		*
1962	<	* @param w the worker
1962	>	* @return a task, if available
1963		*/
1964	<	final void deregisterWorker(ForkJoinWorkerThread w, Throwable ex) {
1965	<	int idx = w.poolIndex;
1966	<	int sc = w.stealCount;
1967	<	int steps = 0;
1968	<	// Remove from array, adjust worker counts and collect steal count.
1969	<	// We can intermix failed removes or adjusts with steal updates
1970	<	do {
1971	<	long s, c;
1972	<	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();
1964	>	final ForkJoinTask<?> nextTaskFor(WorkQueue w) {
1965	>	for (ForkJoinTask<?> t;;) {
1966	>	WorkQueue q; int b;
1967	>	if ((t = w.nextLocalTask()) != null)
1968	>	return t;
1969	>	if ((q = findNonEmptyStealQueue(w)) == null)
1970	>	return null;
1971	>	if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null)
1972	>	return t;
1973		}
1974		}
1975
1976	<	// Shutdown and termination
1976	>	/**
1977	>	* Returns the approximate (non-atomic) number of idle threads per
1978	>	* active thread to offset steal queue size for method
1979	>	* ForkJoinTask.getSurplusQueuedTaskCount().
1980	>	*/
1981	>	final int idlePerActive() {
1982	>	// Approximate at powers of two for small values, saturate past 4
1983	>	int p = parallelism;
1984	>	int a = p + (int)(ctl >> AC_SHIFT);
1985	>	return (a > (p >>>= 1) ? 0 :
1986	>	a > (p >>>= 1) ? 1 :
1987	>	a > (p >>>= 1) ? 2 :
1988	>	a > (p >>>= 1) ? 4 :
1989	>	8);
1990	>	}
1991	>
1992	>	//  Termination
1993
1994		/**
1995	<	* Possibly initiates and/or completes termination.
1995	>	* Possibly initiates and/or completes termination.  The caller
1996	>	* triggering termination runs three passes through workQueues:
1997	>	* (0) Setting termination status, followed by wakeups of queued
1998	>	* workers; (1) cancelling all tasks; (2) interrupting lagging
1999	>	* threads (likely in external tasks, but possibly also blocked in
2000	>	* joins).  Each pass repeats previous steps because of potential
2001	>	* lagging thread creation.
2002		*
2003		* @param now if true, unconditionally terminate, else only
2004	<	* if shutdown and empty queue and no active workers
2004	>	* if no work and no active workers
2005	>	* @param enable if true, enable shutdown when next possible
2006		* @return true if now terminating or terminated
2007		*/
2008	<	private boolean tryTerminate(boolean now) {
2009	<	long c;
2010	<	while (((c = ctl) & STOP_BIT) == 0) {
2011	<	if (!now) {
2012	<	if ((int)(c >> AC_SHIFT) != -parallelism)
2013	<	return false;
2014	<	if (!shutdown || blockedCount != 0 || quiescerCount != 0 ||
2015	<	queueBase != queueTop) {
1230	<	if (ctl == c) // staleness check
1231	<	return false;
1232	<	continue;
2008	>	private boolean tryTerminate(boolean now, boolean enable) {
2009	>	Mutex lock = this.lock;
2010	>	for (long c;;) {
2011	>	if (((c = ctl) & STOP_BIT) != 0) {      // already terminating
2012	>	if ((short)(c >>> TC_SHIFT) == -parallelism) {
2013	>	lock.lock();                    // don't need try/finally
2014	>	termination.signalAll();        // signal when 0 workers
2015	>	lock.unlock();
2016		}
2017	+	return true;
2018		}
2019	<	if (UNSAFE.compareAndSwapLong(this, ctlOffset, c, c | STOP_BIT))
2020	<	startTerminating();
2021	<	}
2022	<	if ((short)(c >>> TC_SHIFT) == -parallelism) { // signal when 0 workers
2023	<	final ReentrantLock lock = this.submissionLock;
1240	<	lock.lock();
1241	<	try {
1242	<	termination.signalAll();
1243	<	} finally {
2019	>	if (runState >= 0) {                    // not yet enabled
2020	>	if (!enable)
2021	>	return false;
2022	>	lock.lock();
2023	>	runState |= SHUTDOWN;
2024		lock.unlock();
2025		}
2026	<	}
2027	<	return true;
2028	<	}
2029	<
2030	<	/**
2031	<	* Runs up to three passes through workers: (0) Setting
2032	<	* termination status for each worker, followed by wakeups up to
2033	<	* queued workers; (1) helping cancel tasks; (2) interrupting
2034	<	* lagging threads (likely in external tasks, but possibly also
2035	<	* blocked in joins).  Each pass repeats previous steps because of
2036	<	* potential lagging thread creation.
2037	<	*/
2038	<	private void startTerminating() {
2039	<	cancelSubmissions();
2040	<	for (int pass = 0; pass < 3; ++pass) {
2041	<	ForkJoinWorkerThread[] ws = workers;
2042	<	if (ws != null) {
2043	<	for (ForkJoinWorkerThread w : ws) {
2044	<	if (w != null) {
2045	<	w.terminate = true;
2046	<	if (pass > 0) {
2047	<	w.cancelTasks();
2048	<	if (pass > 1 && !w.isInterrupted()) {
2049	<	try {
2050	<	w.interrupt();
2051	<	} catch (SecurityException ignore) {
2026	>	if (!now) {                             // check if idle & no tasks
2027	>	if ((int)(c >> AC_SHIFT) != -parallelism ||
2028	>	hasQueuedSubmissions())
2029	>	return false;
2030	>	// Check for unqueued inactive workers. One pass suffices.
2031	>	WorkQueue[] ws = workQueues; WorkQueue w;
2032	>	if (ws != null) {
2033	>	for (int i = 1; i < ws.length; i += 2) {
2034	>	if ((w = ws[i]) != null && w.eventCount >= 0)
2035	>	return false;
2036	>	}
2037	>	}
2038	>	}
2039	>	if (U.compareAndSwapLong(this, CTL, c, c | STOP_BIT)) {
2040	>	for (int pass = 0; pass < 3; ++pass) {
2041	>	WorkQueue[] ws = workQueues;
2042	>	if (ws != null) {
2043	>	WorkQueue w;
2044	>	int n = ws.length;
2045	>	for (int i = 0; i < n; ++i) {
2046	>	if ((w = ws[i]) != null) {
2047	>	w.runState = -1;
2048	>	if (pass > 0) {
2049	>	w.cancelAll();
2050	>	if (pass > 1)
2051	>	w.interruptOwner();
2052		}
2053		}
2054		}
2055	+	// Wake up workers parked on event queue
2056	+	int i, e; long cc; Thread p;
2057	+	while ((e = (int)(cc = ctl) & E_MASK) != 0 &&
2058	+	(i = e & SMASK) < n &&
2059	+	(w = ws[i]) != null) {
2060	+	long nc = ((long)(w.nextWait & E_MASK) |
2061	+	((cc + AC_UNIT) & AC_MASK) |
2062	+	(cc & (TC_MASK|STOP_BIT)));
2063	+	if (w.eventCount == (e | INT_SIGN) &&
2064	+	U.compareAndSwapLong(this, CTL, cc, nc)) {
2065	+	w.eventCount = (e + E_SEQ) & E_MASK;
2066	+	w.runState = -1;
2067	+	if ((p = w.parker) != null)
2068	+	U.unpark(p);
2069	+	}
2070	+	}
2071		}
2072		}
1277	–	terminateWaiters();
1278	–	}
1279	–	}
1280	–	}
1281	–
1282	–	/**
1283	–	* Polls and cancels all submissions. Called only during termination.
1284	–	*/
1285	–	private void cancelSubmissions() {
1286	–	while (queueBase != queueTop) {
1287	–	ForkJoinTask<?> task = pollSubmission();
1288	–	if (task != null) {
1289	–	try {
1290	–	task.cancel(false);
1291	–	} catch (Throwable ignore) {
1292	–	}
2073		}
2074		}
2075		}
2076
1297	–	/**
1298	–	* Tries to set the termination status of waiting workers, and
1299	–	* then wakes them up (after which they will terminate).
1300	–	*/
1301	–	private void terminateWaiters() {
1302	–	ForkJoinWorkerThread[] ws = workers;
1303	–	if (ws != null) {
1304	–	ForkJoinWorkerThread w; long c; int i, e;
1305	–	int n = ws.length;
1306	–	while ((i = ~(e = (int)(c = ctl)) & SMASK) < n &&
1307	–	(w = ws[i]) != null && w.eventCount == (e & E_MASK)) {
1308	–	if (UNSAFE.compareAndSwapLong(this, ctlOffset, c,
1309	–	(long)(w.nextWait & E_MASK) |
1310	–	((c + AC_UNIT) & AC_MASK) |
1311	–	(c & (TC_MASK|STOP_BIT)))) {
1312	–	w.terminate = true;
1313	–	w.eventCount = e + EC_UNIT;
1314	–	if (w.parked)
1315	–	UNSAFE.unpark(w);
1316	–	}
1317	–	}
1318	–	}
1319	–	}
1320	–
1321	–	// misc ForkJoinWorkerThread support
1322	–
1323	–	/**
1324	–	* Increment or decrement quiescerCount. Needed only to prevent
1325	–	* triggering shutdown if a worker is transiently inactive while
1326	–	* checking quiescence.
1327	–	*
1328	–	* @param delta 1 for increment, -1 for decrement
1329	–	*/
1330	–	final void addQuiescerCount(int delta) {
1331	–	int c;
1332	–	do {} while (!UNSAFE.compareAndSwapInt(this, quiescerCountOffset,
1333	–	c = quiescerCount, c + delta));
1334	–	}
1335	–
1336	–	/**
1337	–	* Directly increment or decrement active count without
1338	–	* queuing. This method is used to transiently assert inactivation
1339	–	* while checking quiescence.
1340	–	*
1341	–	* @param delta 1 for increment, -1 for decrement
1342	–	*/
1343	–	final void addActiveCount(int delta) {
1344	–	long d = delta < 0 ? -AC_UNIT : AC_UNIT;
1345	–	long c;
1346	–	do {} while (!UNSAFE.compareAndSwapLong(this, ctlOffset, c = ctl,
1347	–	((c + d) & AC_MASK) |
1348	–	(c & ~AC_MASK)));
1349	–	}
1350	–
1351	–	/**
1352	–	* Returns the approximate (non-atomic) number of idle threads per
1353	–	* active thread.
1354	–	*/
1355	–	final int idlePerActive() {
1356	–	// Approximate at powers of two for small values, saturate past 4
1357	–	int p = parallelism;
1358	–	int a = p + (int)(ctl >> AC_SHIFT);
1359	–	return (a > (p >>>= 1) ? 0 :
1360	–	a > (p >>>= 1) ? 1 :
1361	–	a > (p >>>= 1) ? 2 :
1362	–	a > (p >>>= 1) ? 4 :
1363	–	8);
1364	–	}
1365	–
2077		// Exported methods
2078
2079		// Constructors
#	Line 1432 | Line 2143 | public class ForkJoinPool extends Abstra
2143		checkPermission();
2144		if (factory == null)
2145		throw new NullPointerException();
2146	<	if (parallelism <= 0 || parallelism > MAX_ID)
2146	>	if (parallelism <= 0 || parallelism > MAX_CAP)
2147		throw new IllegalArgumentException();
2148		this.parallelism = parallelism;
2149		this.factory = factory;
2150		this.ueh = handler;
2151	<	this.locallyFifo = asyncMode;
2151	>	this.localMode = asyncMode ? FIFO_QUEUE : LIFO_QUEUE;
2152		long np = (long)(-parallelism); // offset ctl counts
2153		this.ctl = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK);
2154	<	this.submissionQueue = new ForkJoinTask<?>[INITIAL_QUEUE_CAPACITY];
2155	<	// initialize workers array with room for 2*parallelism if possible
2156	<	int n = parallelism << 1;
2157	<	if (n >= MAX_ID)
2158	<	n = MAX_ID;
2159	<	else { // See Hackers Delight, sec 3.2, where n < (1 << 16)
2160	<	n |= n >>> 1; n |= n >>> 2; n |= n >>> 4; n |= n >>> 8;
2161	<	}
2162	<	workers = new ForkJoinWorkerThread[n + 1];
2163	<	this.submissionLock = new ReentrantLock();
1453	<	this.termination = submissionLock.newCondition();
2154	>	// Use nearest power 2 for workQueues size. See Hackers Delight sec 3.2.
2155	>	int n = parallelism - 1;
2156	>	n |= n >>> 1; n |= n >>> 2; n |= n >>> 4; n |= n >>> 8; n |= n >>> 16;
2157	>	int size = (n + 1) << 1;        // #slots = 2*#workers
2158	>	this.submitMask = size - 1;     // room for max # of submit queues
2159	>	this.workQueues = new WorkQueue[size];
2160	>	this.termination = (this.lock = new Mutex()).newCondition();
2161	>	this.stealCount = new AtomicLong();
2162	>	this.nextWorkerNumber = new AtomicInteger();
2163	>	int pn = poolNumberGenerator.incrementAndGet();
2164		StringBuilder sb = new StringBuilder("ForkJoinPool-");
2165	<	sb.append(poolNumberGenerator.incrementAndGet());
2165	>	sb.append(Integer.toString(pn));
2166		sb.append("-worker-");
2167		this.workerNamePrefix = sb.toString();
2168	+	lock.lock();
2169	+	this.runState = 1;              // set init flag
2170	+	lock.unlock();
2171		}
2172
2173		// Execution methods
#	Line 1476 | Line 2189 | public class ForkJoinPool extends Abstra
2189		*         scheduled for execution
2190		*/
2191		public <T> T invoke(ForkJoinTask<T> task) {
1479	–	Thread t = Thread.currentThread();
2192		if (task == null)
2193		throw new NullPointerException();
2194	<	if (shutdown)
2195	<	throw new RejectedExecutionException();
1484	<	if ((t instanceof ForkJoinWorkerThread) &&
1485	<	((ForkJoinWorkerThread)t).pool == this)
1486	<	return task.invoke();  // bypass submit if in same pool
1487	<	else {
1488	<	addSubmission(task);
1489	<	return task.join();
1490	<	}
1491	<	}
1492	<
1493	<	/**
1494	<	* Unless terminating, forks task if within an ongoing FJ
1495	<	* computation in the current pool, else submits as external task.
1496	<	*/
1497	<	private <T> void forkOrSubmit(ForkJoinTask<T> task) {
1498	<	ForkJoinWorkerThread w;
1499	<	Thread t = Thread.currentThread();
1500	<	if (shutdown)
1501	<	throw new RejectedExecutionException();
1502	<	if ((t instanceof ForkJoinWorkerThread) &&
1503	<	(w = (ForkJoinWorkerThread)t).pool == this)
1504	<	w.pushTask(task);
1505	<	else
1506	<	addSubmission(task);
2194	>	doSubmit(task);
2195	>	return task.join();
2196		}
2197
2198		/**
#	Line 1517 | Line 2206 | public class ForkJoinPool extends Abstra
2206		public void execute(ForkJoinTask<?> task) {
2207		if (task == null)
2208		throw new NullPointerException();
2209	<	forkOrSubmit(task);
2209	>	doSubmit(task);
2210		}
2211
2212		// AbstractExecutorService methods
#	Line 1534 | Line 2223 | public class ForkJoinPool extends Abstra
2223		if (task instanceof ForkJoinTask<?>) // avoid re-wrap
2224		job = (ForkJoinTask<?>) task;
2225		else
2226	<	job = ForkJoinTask.adapt(task, null);
2227	<	forkOrSubmit(job);
2226	>	job = new ForkJoinTask.AdaptedRunnableAction(task);
2227	>	doSubmit(job);
2228		}
2229
2230		/**
#	Line 1550 | Line 2239 | public class ForkJoinPool extends Abstra
2239		public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) {
2240		if (task == null)
2241		throw new NullPointerException();
2242	<	forkOrSubmit(task);
2242	>	doSubmit(task);
2243		return task;
2244		}
2245
#	Line 1560 | Line 2249 | public class ForkJoinPool extends Abstra
2249		*         scheduled for execution
2250		*/
2251		public <T> ForkJoinTask<T> submit(Callable<T> task) {
2252	<	if (task == null)
2253	<	throw new NullPointerException();
1565	<	ForkJoinTask<T> job = ForkJoinTask.adapt(task);
1566	<	forkOrSubmit(job);
2252	>	ForkJoinTask<T> job = new ForkJoinTask.AdaptedCallable<T>(task);
2253	>	doSubmit(job);
2254		return job;
2255		}
2256
#	Line 1573 | Line 2260 | public class ForkJoinPool extends Abstra
2260		*         scheduled for execution
2261		*/
2262		public <T> ForkJoinTask<T> submit(Runnable task, T result) {
2263	<	if (task == null)
2264	<	throw new NullPointerException();
1578	<	ForkJoinTask<T> job = ForkJoinTask.adapt(task, result);
1579	<	forkOrSubmit(job);
2263	>	ForkJoinTask<T> job = new ForkJoinTask.AdaptedRunnable<T>(task, result);
2264	>	doSubmit(job);
2265		return job;
2266		}
2267
#	Line 1592 | Line 2277 | public class ForkJoinPool extends Abstra
2277		if (task instanceof ForkJoinTask<?>) // avoid re-wrap
2278		job = (ForkJoinTask<?>) task;
2279		else
2280	<	job = ForkJoinTask.adapt(task, null);
2281	<	forkOrSubmit(job);
2280	>	job = new ForkJoinTask.AdaptedRunnableAction(task);
2281	>	doSubmit(job);
2282		return job;
2283		}
2284
#	Line 1602 | Line 2287 | public class ForkJoinPool extends Abstra
2287		* @throws RejectedExecutionException {@inheritDoc}
2288		*/
2289		public <T> List<Future<T>> invokeAll(Collection<? extends Callable<T>> tasks) {
2290	<	ArrayList<ForkJoinTask<T>> forkJoinTasks =
2291	<	new ArrayList<ForkJoinTask<T>>(tasks.size());
2292	<	for (Callable<T> task : tasks)
2293	<	forkJoinTasks.add(ForkJoinTask.adapt(task));
2294	<	invoke(new InvokeAll<T>(forkJoinTasks));
2295	<
2290	>	// In previous versions of this class, this method constructed
2291	>	// a task to run ForkJoinTask.invokeAll, but now external
2292	>	// invocation of multiple tasks is at least as efficient.
2293	>	List<ForkJoinTask<T>> fs = new ArrayList<ForkJoinTask<T>>(tasks.size());
2294	>	// Workaround needed because method wasn't declared with
2295	>	// wildcards in return type but should have been.
2296		@SuppressWarnings({"unchecked", "rawtypes"})
2297	<	List<Future<T>> futures = (List<Future<T>>) (List) forkJoinTasks;
1613	<	return futures;
1614	<	}
2297	>	List<Future<T>> futures = (List<Future<T>>) (List) fs;
2298
2299	<	static final class InvokeAll<T> extends RecursiveAction {
2300	<	final ArrayList<ForkJoinTask<T>> tasks;
2301	<	InvokeAll(ArrayList<ForkJoinTask<T>> tasks) { this.tasks = tasks; }
2302	<	public void compute() {
2303	<	try { invokeAll(tasks); }
2304	<	catch (Exception ignore) {}
2299	>	boolean done = false;
2300	>	try {
2301	>	for (Callable<T> t : tasks) {
2302	>	ForkJoinTask<T> f = new ForkJoinTask.AdaptedCallable<T>(t);
2303	>	doSubmit(f);
2304	>	fs.add(f);
2305	>	}
2306	>	for (ForkJoinTask<T> f : fs)
2307	>	f.quietlyJoin();
2308	>	done = true;
2309	>	return futures;
2310	>	} finally {
2311	>	if (!done)
2312	>	for (ForkJoinTask<T> f : fs)
2313	>	f.cancel(false);
2314		}
1623	–	private static final long serialVersionUID = -7914297376763021607L;
2315		}
2316
2317		/**
#	Line 1670 | Line 2361 | public class ForkJoinPool extends Abstra
2361		* @return {@code true} if this pool uses async mode
2362		*/
2363		public boolean getAsyncMode() {
2364	<	return locallyFifo;
2364	>	return localMode != 0;
2365		}
2366
2367		/**
#	Line 1682 | Line 2373 | public class ForkJoinPool extends Abstra
2373		* @return the number of worker threads
2374		*/
2375		public int getRunningThreadCount() {
2376	<	int r = parallelism + (int)(ctl >> AC_SHIFT);
2377	<	return (r <= 0) ? 0 : r; // suppress momentarily negative values
2376	>	int rc = 0;
2377	>	WorkQueue[] ws; WorkQueue w;
2378	>	if ((ws = workQueues) != null) {
2379	>	for (int i = 1; i < ws.length; i += 2) {
2380	>	if ((w = ws[i]) != null && w.isApparentlyUnblocked())
2381	>	++rc;
2382	>	}
2383	>	}
2384	>	return rc;
2385		}
2386
2387		/**
#	Line 1694 | Line 2392 | public class ForkJoinPool extends Abstra
2392		* @return the number of active threads
2393		*/
2394		public int getActiveThreadCount() {
2395	<	int r = parallelism + (int)(ctl >> AC_SHIFT) + blockedCount;
2395	>	int r = parallelism + (int)(ctl >> AC_SHIFT);
2396		return (r <= 0) ? 0 : r; // suppress momentarily negative values
2397		}
2398
#	Line 1710 | Line 2408 | public class ForkJoinPool extends Abstra
2408		* @return {@code true} if all threads are currently idle
2409		*/
2410		public boolean isQuiescent() {
2411	<	return parallelism + (int)(ctl >> AC_SHIFT) + blockedCount == 0;
2411	>	return (int)(ctl >> AC_SHIFT) + parallelism == 0;
2412		}
2413
2414		/**
#	Line 1725 | Line 2423 | public class ForkJoinPool extends Abstra
2423		* @return the number of steals
2424		*/
2425		public long getStealCount() {
2426	<	return stealCount;
2426	>	long count = stealCount.get();
2427	>	WorkQueue[] ws; WorkQueue w;
2428	>	if ((ws = workQueues) != null) {
2429	>	for (int i = 1; i < ws.length; i += 2) {
2430	>	if ((w = ws[i]) != null)
2431	>	count += w.totalSteals;
2432	>	}
2433	>	}
2434	>	return count;
2435		}
2436
2437		/**
#	Line 1740 | Line 2446 | public class ForkJoinPool extends Abstra
2446		*/
2447		public long getQueuedTaskCount() {
2448		long count = 0;
2449	<	ForkJoinWorkerThread[] ws;
2450	<	if ((short)(ctl >>> TC_SHIFT) > -parallelism &&
2451	<	(ws = workers) != null) {
2452	<	for (ForkJoinWorkerThread w : ws)
2453	<	if (w != null)
2454	<	count -= w.queueBase - w.queueTop; // must read base first
2449	>	WorkQueue[] ws; WorkQueue w;
2450	>	if ((ws = workQueues) != null) {
2451	>	for (int i = 1; i < ws.length; i += 2) {
2452	>	if ((w = ws[i]) != null)
2453	>	count += w.queueSize();
2454	>	}
2455		}
2456		return count;
2457		}
#	Line 1758 | Line 2464 | public class ForkJoinPool extends Abstra
2464		* @return the number of queued submissions
2465		*/
2466		public int getQueuedSubmissionCount() {
2467	<	return -queueBase + queueTop;
2467	>	int count = 0;
2468	>	WorkQueue[] ws; WorkQueue w;
2469	>	if ((ws = workQueues) != null) {
2470	>	for (int i = 0; i < ws.length; i += 2) {
2471	>	if ((w = ws[i]) != null)
2472	>	count += w.queueSize();
2473	>	}
2474	>	}
2475	>	return count;
2476		}
2477
2478		/**
#	Line 1768 | Line 2482 | public class ForkJoinPool extends Abstra
2482		* @return {@code true} if there are any queued submissions
2483		*/
2484		public boolean hasQueuedSubmissions() {
2485	<	return queueBase != queueTop;
2485	>	WorkQueue[] ws; WorkQueue w;
2486	>	if ((ws = workQueues) != null) {
2487	>	for (int i = 0; i < ws.length; i += 2) {
2488	>	if ((w = ws[i]) != null && !w.isEmpty())
2489	>	return true;
2490	>	}
2491	>	}
2492	>	return false;
2493		}
2494
2495		/**
#	Line 1779 | Line 2500 | public class ForkJoinPool extends Abstra
2500		* @return the next submission, or {@code null} if none
2501		*/
2502		protected ForkJoinTask<?> pollSubmission() {
2503	<	ForkJoinTask<?> t; ForkJoinTask<?>[] q; int b, i;
2504	<	while ((b = queueBase) != queueTop &&
2505	<	(q = submissionQueue) != null &&
2506	<	(i = (q.length - 1) & b) >= 0) {
2507	<	long u = (i << ASHIFT) + ABASE;
1787	<	if ((t = q[i]) != null &&
1788	<	queueBase == b &&
1789	<	UNSAFE.compareAndSwapObject(q, u, t, null)) {
1790	<	queueBase = b + 1;
1791	<	return t;
2503	>	WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2504	>	if ((ws = workQueues) != null) {
2505	>	for (int i = 0; i < ws.length; i += 2) {
2506	>	if ((w = ws[i]) != null && (t = w.poll()) != null)
2507	>	return t;
2508		}
2509		}
2510		return null;
#	Line 1813 | Line 2529 | public class ForkJoinPool extends Abstra
2529		*/
2530		protected int drainTasksTo(Collection<? super ForkJoinTask<?>> c) {
2531		int count = 0;
2532	<	while (queueBase != queueTop) {
2533	<	ForkJoinTask<?> t = pollSubmission();
2534	<	if (t != null) {
2535	<	c.add(t);
2536	<	++count;
2532	>	WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2533	>	if ((ws = workQueues) != null) {
2534	>	for (int i = 0; i < ws.length; ++i) {
2535	>	if ((w = ws[i]) != null) {
2536	>	while ((t = w.poll()) != null) {
2537	>	c.add(t);
2538	>	++count;
2539	>	}
2540	>	}
2541		}
2542		}
1823	–	ForkJoinWorkerThread[] ws;
1824	–	if ((short)(ctl >>> TC_SHIFT) > -parallelism &&
1825	–	(ws = workers) != null) {
1826	–	for (ForkJoinWorkerThread w : ws)
1827	–	if (w != null)
1828	–	count += w.drainTasksTo(c);
1829	–	}
2543		return count;
2544		}
2545
#	Line 1838 | Line 2551 | public class ForkJoinPool extends Abstra
2551		* @return a string identifying this pool, as well as its state
2552		*/
2553		public String toString() {
2554	<	long st = getStealCount();
2555	<	long qt = getQueuedTaskCount();
2556	<	long qs = getQueuedSubmissionCount();
1844	<	int pc = parallelism;
2554	>	// Use a single pass through workQueues to collect counts
2555	>	long qt = 0L, qs = 0L; int rc = 0;
2556	>	long st = stealCount.get();
2557		long c = ctl;
2558	+	WorkQueue[] ws; WorkQueue w;
2559	+	if ((ws = workQueues) != null) {
2560	+	for (int i = 0; i < ws.length; ++i) {
2561	+	if ((w = ws[i]) != null) {
2562	+	int size = w.queueSize();
2563	+	if ((i & 1) == 0)
2564	+	qs += size;
2565	+	else {
2566	+	qt += size;
2567	+	st += w.totalSteals;
2568	+	if (w.isApparentlyUnblocked())
2569	+	++rc;
2570	+	}
2571	+	}
2572	+	}
2573	+	}
2574	+	int pc = parallelism;
2575		int tc = pc + (short)(c >>> TC_SHIFT);
2576	<	int rc = pc + (int)(c >> AC_SHIFT);
2577	<	if (rc < 0) // ignore transient negative
2578	<	rc = 0;
1850	<	int ac = rc + blockedCount;
2576	>	int ac = pc + (int)(c >> AC_SHIFT);
2577	>	if (ac < 0) // ignore transient negative
2578	>	ac = 0;
2579		String level;
2580		if ((c & STOP_BIT) != 0)
2581		level = (tc == 0) ? "Terminated" : "Terminating";
2582		else
2583	<	level = shutdown ? "Shutting down" : "Running";
2583	>	level = runState < 0 ? "Shutting down" : "Running";
2584		return super.toString() +
2585		"[" + level +
2586		", parallelism = " + pc +
#	Line 1879 | Line 2607 | public class ForkJoinPool extends Abstra
2607		*/
2608		public void shutdown() {
2609		checkPermission();
2610	<	shutdown = true;
1883	<	tryTerminate(false);
2610	>	tryTerminate(false, true);
2611		}
2612
2613		/**
#	Line 1901 | Line 2628 | public class ForkJoinPool extends Abstra
2628		*/
2629		public List<Runnable> shutdownNow() {
2630		checkPermission();
2631	<	shutdown = true;
1905	<	tryTerminate(true);
2631	>	tryTerminate(true, true);
2632		return Collections.emptyList();
2633		}
2634
#	Line 1937 | Line 2663 | public class ForkJoinPool extends Abstra
2663		}
2664
2665		/**
1940	–	* Returns true if terminating or terminated. Used by ForkJoinWorkerThread.
1941	–	*/
1942	–	final boolean isAtLeastTerminating() {
1943	–	return (ctl & STOP_BIT) != 0L;
1944	–	}
1945	–
1946	–	/**
2666		* Returns {@code true} if this pool has been shut down.
2667		*
2668		* @return {@code true} if this pool has been shut down
2669		*/
2670		public boolean isShutdown() {
2671	<	return shutdown;
2671	>	return runState < 0;
2672		}
2673
2674		/**
#	Line 1966 | Line 2685 | public class ForkJoinPool extends Abstra
2685		public boolean awaitTermination(long timeout, TimeUnit unit)
2686		throws InterruptedException {
2687		long nanos = unit.toNanos(timeout);
2688	<	final ReentrantLock lock = this.submissionLock;
2688	>	final Mutex lock = this.lock;
2689		lock.lock();
2690		try {
2691		for (;;) {
#	Line 2077 | Line 2796 | public class ForkJoinPool extends Abstra
2796		public static void managedBlock(ManagedBlocker blocker)
2797		throws InterruptedException {
2798		Thread t = Thread.currentThread();
2799	<	if (t instanceof ForkJoinWorkerThread) {
2800	<	ForkJoinWorkerThread w = (ForkJoinWorkerThread) t;
2801	<	w.pool.awaitBlocker(blocker);
2802	<	}
2803	<	else {
2804	<	do {} while (!blocker.isReleasable() && !blocker.block());
2799	>	ForkJoinPool p = ((t instanceof ForkJoinWorkerThread) ?
2800	>	((ForkJoinWorkerThread)t).pool : null);
2801	>	while (!blocker.isReleasable()) {
2802	>	if (p == null || p.tryCompensate(null, blocker)) {
2803	>	try {
2804	>	do {} while (!blocker.isReleasable() && !blocker.block());
2805	>	} finally {
2806	>	if (p != null)
2807	>	p.incrementActiveCount();
2808	>	}
2809	>	break;
2810	>	}
2811		}
2812		}
2813
#	Line 2091 | Line 2816 | public class ForkJoinPool extends Abstra
2816		// implement RunnableFuture.
2817
2818		protected <T> RunnableFuture<T> newTaskFor(Runnable runnable, T value) {
2819	<	return (RunnableFuture<T>) ForkJoinTask.adapt(runnable, value);
2819	>	return new ForkJoinTask.AdaptedRunnable<T>(runnable, value);
2820		}
2821
2822		protected <T> RunnableFuture<T> newTaskFor(Callable<T> callable) {
2823	<	return (RunnableFuture<T>) ForkJoinTask.adapt(callable);
2823	>	return new ForkJoinTask.AdaptedCallable<T>(callable);
2824		}
2825
2826		// Unsafe mechanics
2827	<	private static final sun.misc.Unsafe UNSAFE;
2828	<	private static final long ctlOffset;
2829	<	private static final long stealCountOffset;
2830	<	private static final long blockedCountOffset;
2106	<	private static final long quiescerCountOffset;
2107	<	private static final long scanGuardOffset;
2108	<	private static final long nextWorkerNumberOffset;
2109	<	private static final long ABASE;
2827	>	private static final sun.misc.Unsafe U;
2828	>	private static final long CTL;
2829	>	private static final long PARKBLOCKER;
2830	>	private static final int ABASE;
2831		private static final int ASHIFT;
2832
2833		static {
2834		poolNumberGenerator = new AtomicInteger();
2835	<	workerSeedGenerator = new Random();
2835	>	nextSubmitterSeed = new AtomicInteger(0x55555555);
2836		modifyThreadPermission = new RuntimePermission("modifyThread");
2837		defaultForkJoinWorkerThreadFactory =
2838		new DefaultForkJoinWorkerThreadFactory();
2839	+	submitters = new ThreadSubmitter();
2840	+	int s;
2841		try {
2842	<	UNSAFE = getUnsafe();
2842	>	U = getUnsafe();
2843		Class<?> k = ForkJoinPool.class;
2844	<	ctlOffset = UNSAFE.objectFieldOffset
2844	>	Class<?> ak = ForkJoinTask[].class;
2845	>	CTL = U.objectFieldOffset
2846		(k.getDeclaredField("ctl"));
2847	<	stealCountOffset = UNSAFE.objectFieldOffset
2848	<	(k.getDeclaredField("stealCount"));
2849	<	blockedCountOffset = UNSAFE.objectFieldOffset
2850	<	(k.getDeclaredField("blockedCount"));
2851	<	quiescerCountOffset = UNSAFE.objectFieldOffset
2128	<	(k.getDeclaredField("quiescerCount"));
2129	<	scanGuardOffset = UNSAFE.objectFieldOffset
2130	<	(k.getDeclaredField("scanGuard"));
2131	<	nextWorkerNumberOffset = UNSAFE.objectFieldOffset
2132	<	(k.getDeclaredField("nextWorkerNumber"));
2847	>	Class<?> tk = Thread.class;
2848	>	PARKBLOCKER = U.objectFieldOffset
2849	>	(tk.getDeclaredField("parkBlocker"));
2850	>	ABASE = U.arrayBaseOffset(ak);
2851	>	s = U.arrayIndexScale(ak);
2852		} catch (Exception e) {
2853		throw new Error(e);
2854		}
2136	–	Class<?> a = ForkJoinTask[].class;
2137	–	ABASE = UNSAFE.arrayBaseOffset(a);
2138	–	int s = UNSAFE.arrayIndexScale(a);
2855		if ((s & (s-1)) != 0)
2856		throw new Error("data type scale not a power of two");
2857		ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
#	Line 2168 | Line 2884 | public class ForkJoinPool extends Abstra
2884		}
2885		}
2886		}
2887	+
2888		}

Diff Legend

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