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 1.85 by dl, Sun Nov 21 13:55:04 2010 UTC vs.
Revision 1.127 by dl, Sun Mar 4 15:52:45 2012 UTC

#	Line 1 | Line 1
1		/*
2		* Written by Doug Lea with assistance from members of JCP JSR-166
3		* Expert Group and released to the public domain, as explained at
4	<	* http://creativecommons.org/licenses/publicdomain
4	>	* http://creativecommons.org/publicdomain/zero/1.0/
5		*/
6
7		package jsr166y;
8	–
8		import java.util.ArrayList;
9		import java.util.Arrays;
10		import java.util.Collection;
11		import java.util.Collections;
12		import java.util.List;
13	+	import java.util.Random;
14		import java.util.concurrent.AbstractExecutorService;
15		import java.util.concurrent.Callable;
16		import java.util.concurrent.ExecutorService;
#	Line 18 | Line 18 | import java.util.concurrent.Future;
18		import java.util.concurrent.RejectedExecutionException;
19		import java.util.concurrent.RunnableFuture;
20		import java.util.concurrent.TimeUnit;
21	–	import java.util.concurrent.TimeoutException;
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		/**
27		* An {@link ExecutorService} for running {@link ForkJoinTask}s.
#	Line 32 | Line 32 | import java.util.concurrent.locks.Reentr
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 57 | Line 59 | import java.util.concurrent.locks.Reentr
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 100 | Line 103 | import java.util.concurrent.locks.Reentr
103		* daemon} mode, there is typically no need to explicitly {@link
104		* #shutdown} such a pool upon program exit.
105		*
106	<	* <pre>
106	>	*  <pre> {@code
107		* static final ForkJoinPool mainPool = new ForkJoinPool();
108		* ...
109		* public void sort(long[] array) {
110		*   mainPool.invoke(new SortTask(array, 0, array.length));
111	<	* }
109	<	* </pre>
111	>	* }}</pre>
112		*
113		* <p><b>Implementation notes</b>: This implementation restricts the
114		* maximum number of running threads to 32767. Attempts to create
#	Line 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	<	* The main work-stealing mechanics implemented in class
135	<	* ForkJoinWorkerThread give first priority to processing tasks
136	<	* from their own queues (LIFO or FIFO, depending on mode), then
137	<	* to randomized FIFO steals of tasks in other worker queues, and
138	<	* lastly to new submissions. These mechanics do not consider
139	<	* affinities, loads, cache localities, etc, so rarely provide the
140	<	* best possible performance on a given machine, but portably
141	<	* provide good throughput by averaging over these factors.
142	<	* (Further, even if we did try to use such information, we do not
143	<	* usually have a basis for exploiting it. For example, some sets
144	<	* of tasks profit from cache affinities, but others are harmed by
145	<	* cache pollution effects.)
146	<	*
147	<	* Beyond work-stealing support and essential bookkeeping, the
148	<	* main responsibility of this framework is to take actions when
149	<	* one worker is waiting to join a task stolen (or always held by)
150	<	* another.  Because we are multiplexing many tasks on to a pool
151	<	* of workers, we can't just let them block (as in Thread.join).
152	<	* We also cannot just reassign the joiner's run-time stack with
153	<	* another and replace it later, which would be a form of
154	<	* "continuation", that even if possible is not necessarily a good
155	<	* idea. Given that the creation costs of most threads on most
156	<	* systems mainly surrounds setting up runtime stacks, thread
157	<	* creation and switching is usually not much more expensive than
158	<	* stack creation and switching, and is more flexible). Instead we
159	<	* combine two tactics:
130	>	* This class and its nested classes provide the main
131	>	* functionality and control for a set of worker threads:
132	>	* Submissions from non-FJ threads enter into submission queues.
133	>	* Workers take these tasks and typically split them into subtasks
134	>	* that may be stolen by other workers.  Preference rules give
135	>	* first priority to processing tasks from their own queues (LIFO
136	>	* or FIFO, depending on mode), then to randomized FIFO steals of
137	>	* tasks in other queues.
138	>	*
139	>	* WorkQueues
140	>	* ==========
141	>	*
142	>	* Most operations occur within work-stealing queues (in nested
143	>	* class WorkQueue).  These are special forms of Deques that
144	>	* support only three of the four possible end-operations -- push,
145	>	* pop, and poll (aka steal), under the further constraints that
146	>	* push and pop are called only from the owning thread (or, as
147	>	* extended here, under a lock), while poll may be called from
148	>	* other threads.  (If you are unfamiliar with them, you probably
149	>	* want to read Herlihy and Shavit's book "The Art of
150	>	* Multiprocessor programming", chapter 16 describing these in
151	>	* more detail before proceeding.)  The main work-stealing queue
152	>	* design is roughly similar to those in the papers "Dynamic
153	>	* Circular Work-Stealing Deque" by Chase and Lev, SPAA 2005
154	>	* (http://research.sun.com/scalable/pubs/index.html) and
155	>	* "Idempotent work stealing" by Michael, Saraswat, and Vechev,
156	>	* PPoPP 2009 (http://portal.acm.org/citation.cfm?id=1504186).
157	>	* The main differences ultimately stem from GC requirements that
158	>	* we null out taken slots as soon as we can, to maintain as small
159	>	* a footprint as possible even in programs generating huge
160	>	* numbers of tasks. To accomplish this, we shift the CAS
161	>	* arbitrating pop vs poll (steal) from being on the indices
162	>	* ("base" and "top") to the slots themselves.  So, both a
163	>	* successful pop and poll mainly entail a CAS of a slot from
164	>	* non-null to null.  Because we rely on CASes of references, we
165	>	* do not need tag bits on base or top.  They are simple ints as
166	>	* used in any circular array-based queue (see for example
167	>	* ArrayDeque).  Updates to the indices must still be ordered in a
168	>	* way that guarantees that top == base means the queue is empty,
169	>	* but otherwise may err on the side of possibly making the queue
170	>	* appear nonempty when a push, pop, or poll have not fully
171	>	* committed. Note that this means that the poll operation,
172	>	* considered individually, is not wait-free. One thief cannot
173	>	* successfully continue until another in-progress one (or, if
174	>	* previously empty, a push) completes.  However, in the
175	>	* aggregate, we ensure at least probabilistic non-blockingness.
176	>	* If an attempted steal fails, a thief always chooses a different
177	>	* random victim target to try next. So, in order for one thief to
178	>	* progress, it suffices for any in-progress poll or new push on
179	>	* any empty queue to complete. (This is why we normally use
180	>	* method pollAt and its variants that try once at the apparent
181	>	* base index, else consider alternative actions, rather than
182	>	* method poll.)
183	>	*
184	>	* This approach also enables support of a user mode in which local
185	>	* task processing is in FIFO, not LIFO order, simply by using
186	>	* poll rather than pop.  This can be useful in message-passing
187	>	* frameworks in which tasks are never joined.  However neither
188	>	* mode considers affinities, loads, cache localities, etc, so
189	>	* rarely provide the best possible performance on a given
190	>	* machine, but portably provide good throughput by averaging over
191	>	* these factors.  (Further, even if we did try to use such
192	>	* information, we do not usually have a basis for exploiting it.
193	>	* For example, some sets of tasks profit from cache affinities,
194	>	* but others are harmed by cache pollution effects.)
195	>	*
196	>	* WorkQueues are also used in a similar way for tasks submitted
197	>	* to the pool. We cannot mix these tasks in the same queues used
198	>	* for work-stealing (this would contaminate lifo/fifo
199	>	* processing). Instead, we loosely associate submission queues
200	>	* with submitting threads, using a form of hashing.  The
201	>	* ThreadLocal Submitter class contains a value initially used as
202	>	* a hash code for choosing existing queues, but may be randomly
203	>	* repositioned upon contention with other submitters.  In
204	>	* essence, submitters act like workers except that they never
205	>	* take tasks, and they are multiplexed on to a finite number of
206	>	* shared work queues. However, classes are set up so that future
207	>	* extensions could allow submitters to optionally help perform
208	>	* tasks as well. Insertion of tasks in shared mode requires a
209	>	* lock (mainly to protect in the case of resizing) but we use
210	>	* only a simple spinlock (using bits in field runState), because
211	>	* submitters encountering a busy queue move on to try or create
212	>	* other queues -- they block only when creating and registering
213	>	* new queues.
214	>	*
215	>	* Management
216	>	* ==========
217	>	*
218	>	* The main throughput advantages of work-stealing stem from
219	>	* decentralized control -- workers mostly take tasks from
220	>	* themselves or each other. We cannot negate this in the
221	>	* implementation of other management responsibilities. The main
222	>	* tactic for avoiding bottlenecks is packing nearly all
223	>	* essentially atomic control state into two volatile variables
224	>	* that are by far most often read (not written) as status and
225	>	* consistency checks.
226	>	*
227	>	* Field "ctl" contains 64 bits holding all the information needed
228	>	* to atomically decide to add, inactivate, enqueue (on an event
229	>	* queue), dequeue, and/or re-activate workers.  To enable this
230	>	* packing, we restrict maximum parallelism to (1<<15)-1 (which is
231	>	* far in excess of normal operating range) to allow ids, counts,
232	>	* and their negations (used for thresholding) to fit into 16bit
233	>	* fields.
234	>	*
235	>	* Field "runState" contains 32 bits needed to register and
236	>	* deregister WorkQueues, as well as to enable shutdown. It is
237	>	* only modified under a lock (normally briefly held, but
238	>	* occasionally protecting allocations and resizings) but even
239	>	* when locked remains available to check consistency.
240	>	*
241	>	* Recording WorkQueues.  WorkQueues are recorded in the
242	>	* "workQueues" array that is created upon pool construction and
243	>	* expanded if necessary.  Updates to the array while recording
244	>	* new workers and unrecording terminated ones are protected from
245	>	* each other by a lock but the array is otherwise concurrently
246	>	* readable, and accessed directly.  To simplify index-based
247	>	* operations, the array size is always a power of two, and all
248	>	* readers must tolerate null slots. Shared (submission) queues
249	>	* are at even indices, worker queues at odd indices. Grouping
250	>	* them together in this way simplifies and speeds up task
251	>	* scanning.
252	>	*
253	>	* All worker thread creation is on-demand, triggered by task
254	>	* submissions, replacement of terminated workers, and/or
255	>	* compensation for blocked workers. However, all other support
256	>	* code is set up to work with other policies.  To ensure that we
257	>	* do not hold on to worker references that would prevent GC, ALL
258	>	* accesses to workQueues are via indices into the workQueues
259	>	* array (which is one source of some of the messy code
260	>	* constructions here). In essence, the workQueues array serves as
261	>	* a weak reference mechanism. Thus for example the wait queue
262	>	* field of ctl stores indices, not references.  Access to the
263	>	* workQueues in associated methods (for example signalWork) must
264	>	* both index-check and null-check the IDs. All such accesses
265	>	* ignore bad IDs by returning out early from what they are doing,
266	>	* since this can only be associated with termination, in which
267	>	* case it is OK to give up.  All uses of the workQueues array
268	>	* also check that it is non-null (even if previously
269	>	* non-null). This allows nulling during termination, which is
270	>	* currently not necessary, but remains an option for
271	>	* resource-revocation-based shutdown schemes. It also helps
272	>	* reduce JIT issuance of uncommon-trap code, which tends to
273	>	* unnecessarily complicate control flow in some methods.
274	>	*
275	>	* Event Queuing. Unlike HPC work-stealing frameworks, we cannot
276	>	* let workers spin indefinitely scanning for tasks when none can
277	>	* be found immediately, and we cannot start/resume workers unless
278	>	* there appear to be tasks available.  On the other hand, we must
279	>	* quickly prod them into action when new tasks are submitted or
280	>	* generated. In many usages, ramp-up time to activate workers is
281	>	* the main limiting factor in overall performance (this is
282	>	* compounded at program start-up by JIT compilation and
283	>	* allocation). So we try to streamline this as much as possible.
284	>	* We park/unpark workers after placing in an event wait queue
285	>	* when they cannot find work. This "queue" is actually a simple
286	>	* Treiber stack, headed by the "id" field of ctl, plus a 15bit
287	>	* counter value (that reflects the number of times a worker has
288	>	* been inactivated) to avoid ABA effects (we need only as many
289	>	* version numbers as worker threads). Successors are held in
290	>	* field WorkQueue.nextWait.  Queuing deals with several intrinsic
291	>	* races, mainly that a task-producing thread can miss seeing (and
292	>	* signalling) another thread that gave up looking for work but
293	>	* has not yet entered the wait queue. We solve this by requiring
294	>	* a full sweep of all workers (via repeated calls to method
295	>	* scan()) both before and after a newly waiting worker is added
296	>	* to the wait queue. During a rescan, the worker might release
297	>	* some other queued worker rather than itself, which has the same
298	>	* net effect. Because enqueued workers may actually be rescanning
299	>	* rather than waiting, we set and clear the "parker" field of
300	>	* WorkQueues to reduce unnecessary calls to unpark.  (This
301	>	* requires a secondary recheck to avoid missed signals.)  Note
302	>	* the unusual conventions about Thread.interrupts surrounding
303	>	* parking and other blocking: Because interrupts are used solely
304	>	* to alert threads to check termination, which is checked anyway
305	>	* upon blocking, we clear status (using Thread.interrupted)
306	>	* before any call to park, so that park does not immediately
307	>	* return due to status being set via some other unrelated call to
308	>	* interrupt in user code.
309	>	*
310	>	* Signalling.  We create or wake up workers only when there
311	>	* appears to be at least one task they might be able to find and
312	>	* execute.  When a submission is added or another worker adds a
313	>	* task to a queue that previously had fewer than two tasks, they
314	>	* signal waiting workers (or trigger creation of new ones if
315	>	* fewer than the given parallelism level -- see signalWork).
316	>	* These primary signals are buttressed by signals during rescans;
317	>	* together these cover the signals needed in cases when more
318	>	* tasks are pushed but untaken, and improve performance compared
319	>	* to having one thread wake up all workers.
320	>	*
321	>	* Trimming workers. To release resources after periods of lack of
322	>	* use, a worker starting to wait when the pool is quiescent will
323	>	* time out and terminate if the pool has remained quiescent for
324	>	* SHRINK_RATE nanosecs. This will slowly propagate, eventually
325	>	* terminating all workers after long periods of non-use.
326	>	*
327	>	* Shutdown and Termination. A call to shutdownNow atomically sets
328	>	* a runState bit and then (non-atomically) sets each worker's
329	>	* runState status, cancels all unprocessed tasks, and wakes up
330	>	* all waiting workers.  Detecting whether termination should
331	>	* commence after a non-abrupt shutdown() call requires more work
332	>	* and bookkeeping. We need consensus about quiescence (i.e., that
333	>	* there is no more work). The active count provides a primary
334	>	* indication but non-abrupt shutdown still requires a rechecking
335	>	* scan for any workers that are inactive but not queued.
336	>	*
337	>	* Joining Tasks
338	>	* =============
339	>	*
340	>	* Any of several actions may be taken when one worker is waiting
341	>	* to join a task stolen (or always held) by another.  Because we
342	>	* are multiplexing many tasks on to a pool of workers, we can't
343	>	* just let them block (as in Thread.join).  We also cannot just
344	>	* reassign the joiner's run-time stack with another and replace
345	>	* it later, which would be a form of "continuation", that even if
346	>	* possible is not necessarily a good idea since we sometimes need
347	>	* both an unblocked task and its continuation to progress.
348	>	* Instead we combine two tactics:
349		*
350		*   Helping: Arranging for the joiner to execute some task that it
351	<	*      would be running if the steal had not occurred.  Method
161	<	*      ForkJoinWorkerThread.helpJoinTask tracks joining->stealing
162	<	*      links to try to find such a task.
351	>	*      would be running if the steal had not occurred.
352		*
353		*   Compensating: Unless there are already enough live threads,
354	<	*      method helpMaintainParallelism() may create or
355	<	*      re-activate a spare thread to compensate for blocked
167	<	*      joiners until they unblock.
354	>	*      method tryCompensate() may create or re-activate a spare
355	>	*      thread to compensate for blocked joiners until they unblock.
356		*
357	<	* It is impossible to keep exactly the target (parallelism)
358	<	* number of threads running at any given time.  Determining
359	<	* existence of conservatively safe helping targets, the
360	<	* availability of already-created spares, and the apparent need
361	<	* to create new spares are all racy and require heuristic
362	<	* guidance, so we rely on multiple retries of each.  Compensation
363	<	* occurs in slow-motion. It is triggered only upon timeouts of
176	<	* Object.wait used for joins. This reduces poor decisions that
177	<	* would otherwise be made when threads are waiting for others
178	<	* that are stalled because of unrelated activities such as
179	<	* garbage collection.
357	>	* A third form (implemented in tryRemoveAndExec and
358	>	* tryPollForAndExec) amounts to helping a hypothetical
359	>	* compensator: If we can readily tell that a possible action of a
360	>	* compensator is to steal and execute the task being joined, the
361	>	* joining thread can do so directly, without the need for a
362	>	* compensation thread (although at the expense of larger run-time
363	>	* stacks, but the tradeoff is typically worthwhile).
364		*
365		* The ManagedBlocker extension API can't use helping so relies
366		* only on compensation in method awaitBlocker.
367		*
368	<	* The main throughput advantages of work-stealing stem from
369	<	* decentralized control -- workers mostly steal tasks from each
370	<	* other. We do not want to negate this by creating bottlenecks
371	<	* implementing other management responsibilities. So we use a
372	<	* collection of techniques that avoid, reduce, or cope well with
373	<	* contention. These entail several instances of bit-packing into
374	<	* CASable fields to maintain only the minimally required
375	<	* atomicity. To enable such packing, we restrict maximum
376	<	* parallelism to (1<<15)-1 (enabling twice this (to accommodate
377	<	* unbalanced increments and decrements) to fit into a 16 bit
378	<	* field, which is far in excess of normal operating range.  Even
379	<	* though updates to some of these bookkeeping fields do sometimes
380	<	* contend with each other, they don't normally cache-contend with
381	<	* updates to others enough to warrant memory padding or
382	<	* isolation. So they are all held as fields of ForkJoinPool
383	<	* objects.  The main capabilities are as follows:
384	<	*
385	<	* 1. Creating and removing workers. Workers are recorded in the
386	<	* "workers" array. This is an array as opposed to some other data
387	<	* structure to support index-based random steals by workers.
388	<	* Updates to the array recording new workers and unrecording
389	<	* terminated ones are protected from each other by a lock
390	<	* (workerLock) but the array is otherwise concurrently readable,
391	<	* and accessed directly by workers. To simplify index-based
392	<	* operations, the array size is always a power of two, and all
393	<	* readers must tolerate null slots. Currently, all worker thread
394	<	* creation is on-demand, triggered by task submissions,
395	<	* replacement of terminated workers, and/or compensation for
396	<	* blocked workers. However, all other support code is set up to
213	<	* work with other policies.
214	<	*
215	<	* To ensure that we do not hold on to worker references that
216	<	* would prevent GC, ALL accesses to workers are via indices into
217	<	* the workers array (which is one source of some of the unusual
218	<	* code constructions here). In essence, the workers array serves
219	<	* as a WeakReference mechanism. Thus for example the event queue
220	<	* stores worker indices, not worker references. Access to the
221	<	* workers in associated methods (for example releaseEventWaiters)
222	<	* must both index-check and null-check the IDs. All such accesses
223	<	* ignore bad IDs by returning out early from what they are doing,
224	<	* since this can only be associated with shutdown, in which case
225	<	* it is OK to give up. On termination, we just clobber these
226	<	* data structures without trying to use them.
227	<	*
228	<	* 2. Bookkeeping for dynamically adding and removing workers. We
229	<	* aim to approximately maintain the given level of parallelism.
230	<	* When some workers are known to be blocked (on joins or via
231	<	* ManagedBlocker), we may create or resume others to take their
232	<	* place until they unblock (see below). Implementing this
233	<	* requires counts of the number of "running" threads (i.e., those
234	<	* that are neither blocked nor artificially suspended) as well as
235	<	* the total number.  These two values are packed into one field,
236	<	* "workerCounts" because we need accurate snapshots when deciding
237	<	* to create, resume or suspend.  Note however that the
238	<	* correspondence of these counts to reality is not guaranteed. In
239	<	* particular updates for unblocked threads may lag until they
240	<	* actually wake up.
241	<	*
242	<	* 3. Maintaining global run state. The run state of the pool
243	<	* consists of a runLevel (SHUTDOWN, TERMINATING, etc) similar to
244	<	* those in other Executor implementations, as well as a count of
245	<	* "active" workers -- those that are, or soon will be, or
246	<	* recently were executing tasks. The runLevel and active count
247	<	* are packed together in order to correctly trigger shutdown and
248	<	* termination. Without care, active counts can be subject to very
249	<	* high contention.  We substantially reduce this contention by
250	<	* relaxing update rules.  A worker must claim active status
251	<	* prospectively, by activating if it sees that a submitted or
252	<	* stealable task exists (it may find after activating that the
253	<	* task no longer exists). It stays active while processing this
254	<	* task (if it exists) and any other local subtasks it produces,
255	<	* until it cannot find any other tasks. It then tries
256	<	* inactivating (see method preStep), but upon update contention
257	<	* instead scans for more tasks, later retrying inactivation if it
258	<	* doesn't find any.
259	<	*
260	<	* 4. Managing idle workers waiting for tasks. We cannot let
261	<	* workers spin indefinitely scanning for tasks when none are
262	<	* available. On the other hand, we must quickly prod them into
263	<	* action when new tasks are submitted or generated.  We
264	<	* park/unpark these idle workers using an event-count scheme.
265	<	* Field eventCount is incremented upon events that may enable
266	<	* workers that previously could not find a task to now find one:
267	<	* Submission of a new task to the pool, or another worker pushing
268	<	* a task onto a previously empty queue.  (We also use this
269	<	* mechanism for configuration and termination actions that
270	<	* require wakeups of idle workers).  Each worker maintains its
271	<	* last known event count, and blocks when a scan for work did not
272	<	* find a task AND its lastEventCount matches the current
273	<	* eventCount. Waiting idle workers are recorded in a variant of
274	<	* Treiber stack headed by field eventWaiters which, when nonzero,
275	<	* encodes the thread index and count awaited for by the worker
276	<	* thread most recently calling eventSync. This thread in turn has
277	<	* a record (field nextEventWaiter) for the next waiting worker.
278	<	* In addition to allowing simpler decisions about need for
279	<	* wakeup, the event count bits in eventWaiters serve the role of
280	<	* tags to avoid ABA errors in Treiber stacks. Upon any wakeup,
281	<	* released threads also try to release at most two others.  The
282	<	* net effect is a tree-like diffusion of signals, where released
283	<	* threads (and possibly others) help with unparks.  To further
284	<	* reduce contention effects a bit, failed CASes to increment
285	<	* field eventCount are tolerated without retries in signalWork.
286	<	* Conceptually they are merged into the same event, which is OK
287	<	* when their only purpose is to enable workers to scan for work.
288	<	*
289	<	* 5. Managing suspension of extra workers. When a worker notices
290	<	* (usually upon timeout of a wait()) that there are too few
291	<	* running threads, we may create a new thread to maintain
292	<	* parallelism level, or at least avoid starvation. Usually, extra
293	<	* threads are needed for only very short periods, yet join
294	<	* dependencies are such that we sometimes need them in
295	<	* bursts. Rather than create new threads each time this happens,
296	<	* we suspend no-longer-needed extra ones as "spares". For most
297	<	* purposes, we don't distinguish "extra" spare threads from
298	<	* normal "core" threads: On each call to preStep (the only point
299	<	* at which we can do this) a worker checks to see if there are
300	<	* now too many running workers, and if so, suspends itself.
301	<	* Method helpMaintainParallelism looks for suspended threads to
302	<	* resume before considering creating a new replacement. The
303	<	* spares themselves are encoded on another variant of a Treiber
304	<	* Stack, headed at field "spareWaiters".  Note that the use of
305	<	* spares is intrinsically racy.  One thread may become a spare at
306	<	* about the same time as another is needlessly being created. We
307	<	* counteract this and related slop in part by requiring resumed
308	<	* spares to immediately recheck (in preStep) to see whether they
309	<	* should re-suspend.
310	<	*
311	<	* 6. Killing off unneeded workers. A timeout mechanism is used to
312	<	* shed unused workers: The oldest (first) event queue waiter uses
313	<	* a timed rather than hard wait. When this wait times out without
314	<	* a normal wakeup, it tries to shutdown any one (for convenience
315	<	* the newest) other spare or event waiter via
316	<	* tryShutdownUnusedWorker. This eventually reduces the number of
317	<	* worker threads to a minimum of one after a long enough period
318	<	* without use.
319	<	*
320	<	* 7. Deciding when to create new workers. The main dynamic
321	<	* control in this class is deciding when to create extra threads
322	<	* in method helpMaintainParallelism. We would like to keep
323	<	* exactly #parallelism threads running, which is an impossible
324	<	* task. We always need to create one when the number of running
325	<	* threads would become zero and all workers are busy. Beyond
326	<	* this, we must rely on heuristics that work well in the
327	<	* presence of transient phenomena such as GC stalls, dynamic
328	<	* compilation, and wake-up lags. These transients are extremely
329	<	* common -- we are normally trying to fully saturate the CPUs on
330	<	* a machine, so almost any activity other than running tasks
331	<	* impedes accuracy. Our main defense is to allow parallelism to
332	<	* lapse for a while during joins, and use a timeout to see if,
333	<	* after the resulting settling, there is still a need for
334	<	* additional workers.  This also better copes with the fact that
335	<	* some of the methods in this class tend to never become compiled
336	<	* (but are interpreted), so some components of the entire set of
337	<	* controls might execute 100 times faster than others. And
338	<	* similarly for cases where the apparent lack of work is just due
339	<	* to GC stalls and other transient system activity.
368	>	* The algorithm in tryHelpStealer entails a form of "linear"
369	>	* helping: Each worker records (in field currentSteal) the most
370	>	* recent task it stole from some other worker. Plus, it records
371	>	* (in field currentJoin) the task it is currently actively
372	>	* joining. Method tryHelpStealer uses these markers to try to
373	>	* find a worker to help (i.e., steal back a task from and execute
374	>	* it) that could hasten completion of the actively joined task.
375	>	* In essence, the joiner executes a task that would be on its own
376	>	* local deque had the to-be-joined task not been stolen. This may
377	>	* be seen as a conservative variant of the approach in Wagner &
378	>	* Calder "Leapfrogging: a portable technique for implementing
379	>	* efficient futures" SIGPLAN Notices, 1993
380	>	* (http://portal.acm.org/citation.cfm?id=155354). It differs in
381	>	* that: (1) We only maintain dependency links across workers upon
382	>	* steals, rather than use per-task bookkeeping.  This sometimes
383	>	* requires a linear scan of workQueues array to locate stealers,
384	>	* but often doesn't because stealers leave hints (that may become
385	>	* stale/wrong) of where to locate them.  A stealHint is only a
386	>	* hint because a worker might have had multiple steals and the
387	>	* hint records only one of them (usually the most current).
388	>	* Hinting isolates cost to when it is needed, rather than adding
389	>	* to per-task overhead.  (2) It is "shallow", ignoring nesting
390	>	* and potentially cyclic mutual steals.  (3) It is intentionally
391	>	* racy: field currentJoin is updated only while actively joining,
392	>	* which means that we miss links in the chain during long-lived
393	>	* tasks, GC stalls etc (which is OK since blocking in such cases
394	>	* is usually a good idea).  (4) We bound the number of attempts
395	>	* to find work (see MAX_HELP) and fall back to suspending the
396	>	* worker and if necessary replacing it with another.
397		*
398	<	* Beware that there is a lot of representation-level coupling
398	>	* It is impossible to keep exactly the target parallelism number
399	>	* of threads running at any given time.  Determining the
400	>	* existence of conservatively safe helping targets, the
401	>	* availability of already-created spares, and the apparent need
402	>	* to create new spares are all racy, so we rely on multiple
403	>	* retries of each.  Compensation in the apparent absence of
404	>	* helping opportunities is challenging to control on JVMs, where
405	>	* GC and other activities can stall progress of tasks that in
406	>	* turn stall out many other dependent tasks, without us being
407	>	* able to determine whether they will ever require compensation.
408	>	* Even though work-stealing otherwise encounters little
409	>	* degradation in the presence of more threads than cores,
410	>	* aggressively adding new threads in such cases entails risk of
411	>	* unwanted positive feedback control loops in which more threads
412	>	* cause more dependent stalls (as well as delayed progress of
413	>	* unblocked threads to the point that we know they are available)
414	>	* leading to more situations requiring more threads, and so
415	>	* on. This aspect of control can be seen as an (analytically
416	>	* 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.  For example, direct access to "workers" array by
436	<	* workers, and direct access to ForkJoinTask.status by both
437	<	* ForkJoinPool and ForkJoinWorkerThread.  There is little point
438	<	* trying to reduce this, since any associated future changes in
439	<	* representations will need to be accompanied by algorithmic
440	<	* changes anyway.
441	<	*
442	<	* Style notes: There are lots of inline assignments (of form
443	<	* "while ((local = field) != 0)") which are usually the simplest
444	<	* way to ensure the required read orderings (which are sometimes
445	<	* critical). Also several occurrences of the unusual "do {}
446	<	* while (!cas...)" which is the simplest way to force an update of
447	<	* a CAS'ed variable. There are also other coding oddities that
448	<	* help some methods perform reasonably even when interpreted (not
449	<	* compiled), at the expense of some messy constructions that
450	<	* reduce byte code counts.
451	<	*
452	<	* The order of declarations in this file is: (1) statics (2)
453	<	* fields (along with constants used when unpacking some of them)
454	<	* (3) internal control methods (4) callbacks and other support
455	<	* for ForkJoinTask and ForkJoinWorkerThread classes, (5) exported
456	<	* methods (plus a few little helpers).
435	>	* ForkJoinTask.  The fields of WorkQueue maintain data structures
436	>	* managed by ForkJoinPool, so are directly accessed.  There is
437	>	* little point trying to reduce this, since any associated future
438	>	* changes in representations will need to be accompanied by
439	>	* algorithmic changes anyway. Several methods intrinsically
440	>	* sprawl because they must accumulate sets of consistent reads of
441	>	* volatiles held in local variables.  Methods signalWork() and
442	>	* scan() are the main bottlenecks, so are especially heavily
443	>	* micro-optimized/mangled.  There are lots of inline assignments
444	>	* (of form "while ((local = field) != 0)") which are usually the
445	>	* simplest way to ensure the required read orderings (which are
446	>	* sometimes critical). This leads to a "C"-like style of listing
447	>	* declarations of these locals at the heads of methods or blocks.
448	>	* There are several occurrences of the unusual "do {} while
449	>	* (!cas...)"  which is the simplest way to force an update of a
450	>	* CAS'ed variable. There are also other coding oddities that help
451	>	* some methods perform reasonably even when interpreted (not
452	>	* compiled).
453	>	*
454	>	* The order of declarations in this file is:
455	>	* (1) Static utility functions
456	>	* (2) Nested (static) classes
457	>	* (3) Static fields
458	>	* (4) Fields, along with constants used when unpacking some of them
459	>	* (5) Internal control methods
460	>	* (6) Callbacks and other support for ForkJoinTask methods
461	>	* (7) Exported methods
462	>	* (8) Static block initializing statics in minimally dependent order
463		*/
464
465	+	// Static utilities
466	+
467	+	/**
468	+	* If there is a security manager, makes sure caller has
469	+	* permission to modify threads.
470	+	*/
471	+	private static void checkPermission() {
472	+	SecurityManager security = System.getSecurityManager();
473	+	if (security != null)
474	+	security.checkPermission(modifyThreadPermission);
475	+	}
476	+
477	+	// Nested classes
478	+
479		/**
480		* Factory for creating new {@link ForkJoinWorkerThread}s.
481		* A {@code ForkJoinWorkerThreadFactory} must be defined and used
#	Line 392 | Line 504 | public class ForkJoinPool extends Abstra
504		}
505
506		/**
507	<	* Creates a new ForkJoinWorkerThread. This factory is used unless
508	<	* overridden in ForkJoinPool constructors.
507	>	* A simple non-reentrant lock used for exclusion when managing
508	>	* queues and workers. We use a custom lock so that we can readily
509	>	* probe lock state in constructions that check among alternative
510	>	* actions. The lock is normally only very briefly held, and
511	>	* sometimes treated as a spinlock, but other usages block to
512	>	* reduce overall contention in those cases where locked code
513	>	* bodies perform allocation/resizing.
514	>	*/
515	>	static final class Mutex extends AbstractQueuedSynchronizer {
516	>	public final boolean tryAcquire(int ignore) {
517	>	return compareAndSetState(0, 1);
518	>	}
519	>	public final boolean tryRelease(int ignore) {
520	>	setState(0);
521	>	return true;
522	>	}
523	>	public final void lock() { acquire(0); }
524	>	public final void unlock() { release(0); }
525	>	public final boolean isHeldExclusively() { return getState() == 1; }
526	>	public final Condition newCondition() { return new ConditionObject(); }
527	>	}
528	>
529	>	/**
530	>	* Class for artificial tasks that are used to replace the target
531	>	* of local joins if they are removed from an interior queue slot
532	>	* in WorkQueue.tryRemoveAndExec. We don't need the proxy to
533	>	* actually do anything beyond having a unique identity.
534	>	*/
535	>	static final class EmptyTask extends ForkJoinTask<Void> {
536	>	EmptyTask() { status = ForkJoinTask.NORMAL; } // force done
537	>	public final Void getRawResult() { return null; }
538	>	public final void setRawResult(Void x) {}
539	>	public final boolean exec() { return true; }
540	>	}
541	>
542	>	/**
543	>	* Queues supporting work-stealing as well as external task
544	>	* submission. See above for main rationale and algorithms.
545	>	* Implementation relies heavily on "Unsafe" intrinsics
546	>	* and selective use of "volatile":
547	>	*
548	>	* Field "base" is the index (mod array.length) of the least valid
549	>	* queue slot, which is always the next position to steal (poll)
550	>	* from if nonempty. Reads and writes require volatile orderings
551	>	* but not CAS, because updates are only performed after slot
552	>	* CASes.
553	>	*
554	>	* Field "top" is the index (mod array.length) of the next queue
555	>	* slot to push to or pop from. It is written only by owner thread
556	>	* for push, or under lock for trySharedPush, and accessed by
557	>	* other threads only after reading (volatile) base.  Both top and
558	>	* base are allowed to wrap around on overflow, but (top - base)
559	>	* (or more commonly -(base - top) to force volatile read of base
560	>	* before top) still estimates size.
561	>	*
562	>	* The array slots are read and written using the emulation of
563	>	* volatiles/atomics provided by Unsafe. Insertions must in
564	>	* general use putOrderedObject as a form of releasing store to
565	>	* ensure that all writes to the task object are ordered before
566	>	* its publication in the queue. (Although we can avoid one case
567	>	* of this when locked in trySharedPush.) All removals entail a
568	>	* CAS to null.  The array is always a power of two. To ensure
569	>	* safety of Unsafe array operations, all accesses perform
570	>	* explicit null checks and implicit bounds checks via
571	>	* power-of-two masking.
572	>	*
573	>	* In addition to basic queuing support, this class contains
574	>	* fields described elsewhere to control execution. It turns out
575	>	* to work better memory-layout-wise to include them in this
576	>	* class rather than a separate class.
577	>	*
578	>	* Performance on most platforms is very sensitive to placement of
579	>	* instances of both WorkQueues and their arrays -- we absolutely
580	>	* do not want multiple WorkQueue instances or multiple queue
581	>	* arrays sharing cache lines. (It would be best for queue objects
582	>	* and their arrays to share, but there is nothing available to
583	>	* help arrange that).  Unfortunately, because they are recorded
584	>	* in a common array, WorkQueue instances are often moved to be
585	>	* adjacent by garbage collectors. To reduce impact, we use field
586	>	* padding that works OK on common platforms; this effectively
587	>	* trades off slightly slower average field access for the sake of
588	>	* avoiding really bad worst-case access. (Until better JVM
589	>	* support is in place, this padding is dependent on transient
590	>	* properties of JVM field layout rules.)  We also take care in
591	>	* allocating, sizing and resizing the array. Non-shared queue
592	>	* arrays are initialized (via method growArray) by workers before
593	>	* use. Others are allocated on first use.
594		*/
595	<	public static final ForkJoinWorkerThreadFactory
596	<	defaultForkJoinWorkerThreadFactory =
597	<	new DefaultForkJoinWorkerThreadFactory();
595	>	static final class WorkQueue {
596	>	/**
597	>	* Capacity of work-stealing queue array upon initialization.
598	>	* Must be a power of two; at least 4, but should be larger to
599	>	* reduce or eliminate cacheline sharing among queues.
600	>	* Currently, it is much larger, as a partial workaround for
601	>	* the fact that JVMs often place arrays in locations that
602	>	* share GC bookkeeping (especially cardmarks) such that
603	>	* per-write accesses encounter serious memory contention.
604	>	*/
605	>	static final int INITIAL_QUEUE_CAPACITY = 1 << 13;
606
607	<	/**
608	<	* Permission required for callers of methods that may start or
609	<	* kill threads.
610	<	*/
611	<	private static final RuntimePermission modifyThreadPermission =
612	<	new RuntimePermission("modifyThread");
607	>	/**
608	>	* Maximum size for queue arrays. Must be a power of two less
609	>	* than or equal to 1 << (31 - width of array entry) to ensure
610	>	* lack of wraparound of index calculations, but defined to a
611	>	* value a bit less than this to help users trap runaway
612	>	* programs before saturating systems.
613	>	*/
614	>	static final int MAXIMUM_QUEUE_CAPACITY = 1 << 26; // 64M
615
616	<	/**
617	<	* If there is a security manager, makes sure caller has
618	<	* permission to modify threads.
619	<	*/
620	<	private static void checkPermission() {
621	<	SecurityManager security = System.getSecurityManager();
622	<	if (security != null)
623	<	security.checkPermission(modifyThreadPermission);
624	<	}
616	>	volatile long totalSteals; // cumulative number of steals
617	>	int seed;                  // for random scanning; initialize nonzero
618	>	volatile int eventCount;   // encoded inactivation count; < 0 if inactive
619	>	int nextWait;              // encoded record of next event waiter
620	>	int rescans;               // remaining scans until block
621	>	int nsteals;               // top-level task executions since last idle
622	>	final int mode;            // lifo, fifo, or shared
623	>	int poolIndex;             // index of this queue in pool (or 0)
624	>	int stealHint;             // index of most recent known stealer
625	>	volatile int runState;     // 1: locked, -1: terminate; else 0
626	>	volatile int base;         // index of next slot for poll
627	>	int top;                   // index of next slot for push
628	>	ForkJoinTask<?>[] array;   // the elements (initially unallocated)
629	>	final ForkJoinPool pool;   // the containing pool (may be null)
630	>	final ForkJoinWorkerThread owner; // owning thread or null if shared
631	>	volatile Thread parker;    // == owner during call to park; else null
632	>	ForkJoinTask<?> currentJoin;  // task being joined in awaitJoin
633	>	ForkJoinTask<?> currentSteal; // current non-local task being executed
634	>	// Heuristic padding to ameliorate unfortunate memory placements
635	>	Object p00, p01, p02, p03, p04, p05, p06, p07;
636	>	Object p08, p09, p0a, p0b, p0c, p0d, p0e;
637	>
638	>	WorkQueue(ForkJoinPool pool, ForkJoinWorkerThread owner, int mode) {
639	>	this.mode = mode;
640	>	this.pool = pool;
641	>	this.owner = owner;
642	>	// Place indices in the center of array (that is not yet allocated)
643	>	base = top = INITIAL_QUEUE_CAPACITY >>> 1;
644	>	}
645
646	<	/**
647	<	* Generator for assigning sequence numbers as pool names.
648	<	*/
649	<	private static final AtomicInteger poolNumberGenerator =
650	<	new AtomicInteger();
646	>	/**
647	>	* Returns the approximate number of tasks in the queue.
648	>	*/
649	>	final int queueSize() {
650	>	int n = base - top;       // non-owner callers must read base first
651	>	return (n >= 0) ? 0 : -n; // ignore transient negative
652	>	}
653
654	<	/**
655	<	* The time to block in a join (see awaitJoin) before checking if
656	<	* a new worker should be (re)started to maintain parallelism
657	<	* level. The value should be short enough to maintain global
658	<	* responsiveness and progress but long enough to avoid
659	<	* counterproductive firings during GC stalls or unrelated system
660	<	* activity, and to not bog down systems with continual re-firings
661	<	* on GCs or legitimately long waits.
662	<	*/
663	<	private static final long JOIN_TIMEOUT_MILLIS = 250L; // 4 per second
654	>	/**
655	>	* Provides a more accurate estimate of whether this queue has
656	>	* any tasks than does queueSize, by checking whether a
657	>	* near-empty queue has at least one unclaimed task.
658	>	*/
659	>	final boolean isEmpty() {
660	>	ForkJoinTask<?>[] a; int m, s;
661	>	int n = base - (s = top);
662	>	return (n >= 0 ||
663	>	(n == -1 &&
664	>	((a = array) == null ||
665	>	(m = a.length - 1) < 0 ||
666	>	U.getObjectVolatile
667	>	(a, ((m & (s - 1)) << ASHIFT) + ABASE) == null)));
668	>	}
669
670	<	/**
671	<	* The wakeup interval (in nanoseconds) for the oldest worker
672	<	* waiting for an event to invoke tryShutdownUnusedWorker to
673	<	* shrink the number of workers.  The exact value does not matter
674	<	* too much. It must be short enough to release resources during
675	<	* sustained periods of idleness, but not so short that threads
676	<	* are continually re-created.
677	<	*/
678	<	private static final long SHRINK_RATE_NANOS =
679	<	30L * 1000L * 1000L * 1000L; // 2 per minute
670	>	/**
671	>	* Pushes a task. Call only by owner in unshared queues.
672	>	*
673	>	* @param task the task. Caller must ensure non-null.
674	>	* @throw RejectedExecutionException if array cannot be resized
675	>	*/
676	>	final void push(ForkJoinTask<?> task) {
677	>	ForkJoinTask<?>[] a; ForkJoinPool p;
678	>	int s = top, m, n;
679	>	if ((a = array) != null) {    // ignore if queue removed
680	>	U.putOrderedObject
681	>	(a, (((m = a.length - 1) & s) << ASHIFT) + ABASE, task);
682	>	if ((n = (top = s + 1) - base) <= 2) {
683	>	if ((p = pool) != null)
684	>	p.signalWork();
685	>	}
686	>	else if (n >= m)
687	>	growArray(true);
688	>	}
689	>	}
690
691	<	/**
692	<	* Absolute bound for parallelism level. Twice this number plus
693	<	* one (i.e., 0xfff) must fit into a 16bit field to enable
694	<	* word-packing for some counts and indices.
695	<	*/
696	<	private static final int MAX_WORKERS   = 0x7fff;
691	>	/**
692	>	* Pushes a task if lock is free and array is either big
693	>	* enough or can be resized to be big enough.
694	>	*
695	>	* @param task the task. Caller must ensure non-null.
696	>	* @return true if submitted
697	>	*/
698	>	final boolean trySharedPush(ForkJoinTask<?> task) {
699	>	boolean submitted = false;
700	>	if (runState == 0 && U.compareAndSwapInt(this, RUNSTATE, 0, 1)) {
701	>	ForkJoinTask<?>[] a = array;
702	>	int s = top;
703	>	try {
704	>	if ((a != null && a.length > s + 1 - base) ||
705	>	(a = growArray(false)) != null) { // must presize
706	>	int j = (((a.length - 1) & s) << ASHIFT) + ABASE;
707	>	U.putObject(a, (long)j, task);    // don't need "ordered"
708	>	top = s + 1;
709	>	submitted = true;
710	>	}
711	>	} finally {
712	>	runState = 0;                         // unlock
713	>	}
714	>	}
715	>	return submitted;
716	>	}
717
718	<	/**
719	<	* Array holding all worker threads in the pool.  Array size must
720	<	* be a power of two.  Updates and replacements are protected by
721	<	* workerLock, but the array is always kept in a consistent enough
722	<	* state to be randomly accessed without locking by workers
723	<	* performing work-stealing, as well as other traversal-based
724	<	* methods in this class. All readers must tolerate that some
725	<	* array slots may be null.
726	<	*/
727	<	volatile ForkJoinWorkerThread[] workers;
718	>	/**
719	>	* Takes next task, if one exists, in LIFO order.  Call only
720	>	* by owner in unshared queues. (We do not have a shared
721	>	* version of this method because it is never needed.)
722	>	*/
723	>	final ForkJoinTask<?> pop() {
724	>	ForkJoinTask<?>[] 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	<	* Queue for external submissions.
741	<	*/
742	<	private final LinkedTransferQueue<ForkJoinTask<?>> submissionQueue;
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	<	* Lock protecting updates to workers array.
760	<	*/
761	<	private final ReentrantLock workerLock;
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	<	* Latch released upon termination.
784	<	*/
785	<	private final Phaser termination;
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	<	* Creation factory for worker threads.
791	<	*/
792	<	private final ForkJoinWorkerThreadFactory factory;
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	<	* Sum of per-thread steal counts, updated only when threads are
803	<	* idle or terminating.
804	<	*/
805	<	private volatile long stealCount;
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	<	* Encoded record of top of Treiber stack of threads waiting for
817	<	* events. The top 32 bits contain the count being waited for. The
818	<	* bottom 16 bits contains one plus the pool index of waiting
819	<	* worker thread. (Bits 16-31 are unused.)
820	<	*/
821	<	private volatile long eventWaiters;
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	<	private static final int  EVENT_COUNT_SHIFT = 32;
832	<	private static final long WAITER_ID_MASK    = (1L << 16) - 1L;
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	<	* A counter for events that may wake up worker threads:
868	<	*   - Submission of a new task to the pool
869	<	*   - A worker pushing a task on an empty queue
870	<	*   - termination
871	<	*/
872	<	private volatile int eventCount;
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	<	* Encoded record of top of Treiber stack of spare threads waiting
878	<	* for resumption. The top 16 bits contain an arbitrary count to
879	<	* avoid ABA effects. The bottom 16bits contains one plus the pool
880	<	* index of waiting worker thread.
881	<	*/
882	<	private volatile int spareWaiters;
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	<	private static final int SPARE_COUNT_SHIFT = 16;
891	<	private static final int SPARE_ID_MASK     = (1 << 16) - 1;
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 false if no progress can be made
924	>	*/
925	>	final boolean tryRemoveAndExec(ForkJoinTask<?> task) {
926	>	boolean removed = false, empty = true, progress = true;
927	>	ForkJoinTask<?>[] a; int m, s, b, n;
928	>	if ((a = array) != null && (m = a.length - 1) >= 0 &&
929	>	(n = (s = top) - (b = base)) > 0) {
930	>	for (ForkJoinTask<?> t;;) {           // traverse from s to b
931	>	int j = ((--s & m) << ASHIFT) + ABASE;
932	>	t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
933	>	if (t == null)                    // inconsistent length
934	>	break;
935	>	else if (t == task) {
936	>	if (s + 1 == top) {           // pop
937	>	if (!U.compareAndSwapObject(a, j, task, null))
938	>	break;
939	>	top = s;
940	>	removed = true;
941	>	}
942	>	else if (base == b)           // replace with proxy
943	>	removed = U.compareAndSwapObject(a, j, task,
944	>	new EmptyTask());
945	>	break;
946	>	}
947	>	else if (t.status >= 0)
948	>	empty = false;
949	>	else if (s + 1 == top) {          // pop and throw away
950	>	if (U.compareAndSwapObject(a, j, t, null))
951	>	top = s;
952	>	break;
953	>	}
954	>	if (--n == 0) {
955	>	if (!empty && base == b)
956	>	progress = false;
957	>	break;
958	>	}
959	>	}
960	>	}
961	>	if (removed)
962	>	task.doExec();
963	>	return progress;
964	>	}
965	>
966	>	/**
967	>	* Executes a top-level task and any local tasks remaining
968	>	* after execution.
969	>	*/
970	>	final void runTask(ForkJoinTask<?> t) {
971	>	if (t != null) {
972	>	currentSteal = t;
973	>	t.doExec();
974	>	if (top != base) {       // process remaining local tasks
975	>	if (mode == 0)
976	>	popAndExecAll();
977	>	else
978	>	pollAndExecAll();
979	>	}
980	>	++nsteals;
981	>	currentSteal = null;
982	>	}
983	>	}
984	>
985	>	/**
986	>	* Executes a non-top-level (stolen) task.
987	>	*/
988	>	final void runSubtask(ForkJoinTask<?> t) {
989	>	if (t != null) {
990	>	ForkJoinTask<?> ps = currentSteal;
991	>	currentSteal = t;
992	>	t.doExec();
993	>	currentSteal = ps;
994	>	}
995	>	}
996	>
997	>	/**
998	>	* Returns true if owned and not known to be blocked.
999	>	*/
1000	>	final boolean isApparentlyUnblocked() {
1001	>	Thread wt; Thread.State s;
1002	>	return (eventCount >= 0 &&
1003	>	(wt = owner) != null &&
1004	>	(s = wt.getState()) != Thread.State.BLOCKED &&
1005	>	s != Thread.State.WAITING &&
1006	>	s != Thread.State.TIMED_WAITING);
1007	>	}
1008	>
1009	>	/**
1010	>	* If this owned and is not already interrupted, try to
1011	>	* interrupt and/or unpark, ignoring exceptions.
1012	>	*/
1013	>	final void interruptOwner() {
1014	>	Thread wt, p;
1015	>	if ((wt = owner) != null && !wt.isInterrupted()) {
1016	>	try {
1017	>	wt.interrupt();
1018	>	} catch (SecurityException ignore) {
1019	>	}
1020	>	}
1021	>	if ((p = parker) != null)
1022	>	U.unpark(p);
1023	>	}
1024	>
1025	>	// Unsafe mechanics
1026	>	private static final sun.misc.Unsafe U;
1027	>	private static final long RUNSTATE;
1028	>	private static final int ABASE;
1029	>	private static final int ASHIFT;
1030	>	static {
1031	>	int s;
1032	>	try {
1033	>	U = getUnsafe();
1034	>	Class<?> k = WorkQueue.class;
1035	>	Class<?> ak = ForkJoinTask[].class;
1036	>	RUNSTATE = U.objectFieldOffset
1037	>	(k.getDeclaredField("runState"));
1038	>	ABASE = U.arrayBaseOffset(ak);
1039	>	s = U.arrayIndexScale(ak);
1040	>	} catch (Exception e) {
1041	>	throw new Error(e);
1042	>	}
1043	>	if ((s & (s-1)) != 0)
1044	>	throw new Error("data type scale not a power of two");
1045	>	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
1046	>	}
1047	>	}
1048
1049		/**
1050	<	* Lifecycle control. The low word contains the number of workers
1051	<	* that are (probably) executing tasks. This value is atomically
1052	<	* incremented before a worker gets a task to run, and decremented
1053	<	* when a worker has no tasks and cannot find any.  Bits 16-18
1054	<	* contain runLevel value. When all are zero, the pool is
527	<	* running. Level transitions are monotonic (running -> shutdown
528	<	* -> terminating -> terminated) so each transition adds a bit.
529	<	* These are bundled together to ensure consistent read for
530	<	* termination checks (i.e., that runLevel is at least SHUTDOWN
531	<	* and active threads is zero).
1050	>	* Per-thread records for threads that submit to pools. Currently
1051	>	* holds only pseudo-random seed / index that is used to choose
1052	>	* submission queues in method doSubmit. In the future, this may
1053	>	* also incorporate a means to implement different task rejection
1054	>	* and resubmission policies.
1055		*
1056	<	* Notes: Most direct CASes are dependent on these bitfield
1057	<	* positions.  Also, this field is non-private to enable direct
1058	<	* performance-sensitive CASes in ForkJoinWorkerThread.
1056	>	* Seeds for submitters and workers/workQueues work in basically
1057	>	* the same way but are initialized and updated using slightly
1058	>	* different mechanics. Both are initialized using the same
1059	>	* approach as in class ThreadLocal, where successive values are
1060	>	* unlikely to collide with previous values. This is done during
1061	>	* registration for workers, but requires a separate AtomicInteger
1062	>	* for submitters. Seeds are then randomly modified upon
1063	>	* collisions using xorshifts, which requires a non-zero seed.
1064		*/
1065	<	volatile int runState;
1065	>	static final class Submitter {
1066	>	int seed;
1067	>	Submitter() {
1068	>	int s = nextSubmitterSeed.getAndAdd(SEED_INCREMENT);
1069	>	seed = (s == 0) ? 1 : s; // ensure non-zero
1070	>	}
1071	>	}
1072
1073	<	// Note: The order among run level values matters.
1074	<	private static final int RUNLEVEL_SHIFT     = 16;
1075	<	private static final int SHUTDOWN           = 1 << RUNLEVEL_SHIFT;
1076	<	private static final int TERMINATING        = 1 << (RUNLEVEL_SHIFT + 1);
1077	<	private static final int TERMINATED         = 1 << (RUNLEVEL_SHIFT + 2);
1078	<	private static final int ACTIVE_COUNT_MASK  = (1 << RUNLEVEL_SHIFT) - 1;
1073	>	/** ThreadLocal class for Submitters */
1074	>	static final class ThreadSubmitter extends ThreadLocal<Submitter> {
1075	>	public Submitter initialValue() { return new Submitter(); }
1076	>	}
1077	>
1078	>	// static fields (initialized in static initializer below)
1079
1080		/**
1081	<	* Holds number of total (i.e., created and not yet terminated)
1082	<	* and running (i.e., not blocked on joins or other managed sync)
549	<	* threads, packed together to ensure consistent snapshot when
550	<	* making decisions about creating and suspending spare
551	<	* threads. Updated only by CAS. Note that adding a new worker
552	<	* requires incrementing both counts, since workers start off in
553	<	* running state.
1081	>	* Creates a new ForkJoinWorkerThread. This factory is used unless
1082	>	* overridden in ForkJoinPool constructors.
1083		*/
1084	<	private volatile int workerCounts;
1085	<
557	<	private static final int TOTAL_COUNT_SHIFT  = 16;
558	<	private static final int RUNNING_COUNT_MASK = (1 << TOTAL_COUNT_SHIFT) - 1;
559	<	private static final int ONE_RUNNING        = 1;
560	<	private static final int ONE_TOTAL          = 1 << TOTAL_COUNT_SHIFT;
1084	>	public static final ForkJoinWorkerThreadFactory
1085	>	defaultForkJoinWorkerThreadFactory;
1086
1087		/**
1088	<	* The target parallelism level.
564	<	* Accessed directly by ForkJoinWorkerThreads.
1088	>	* Generator for assigning sequence numbers as pool names.
1089		*/
1090	<	final int parallelism;
1090	>	private static final AtomicInteger poolNumberGenerator;
1091
1092		/**
1093	<	* True if use local fifo, not default lifo, for local polling
1094	<	* Read by, and replicated by ForkJoinWorkerThreads
1093	>	* Generator for initial hashes/seeds for submitters. Accessed by
1094	>	* Submitter class constructor.
1095		*/
1096	<	final boolean locallyFifo;
1096	>	static final AtomicInteger nextSubmitterSeed;
1097
1098		/**
1099	<	* The uncaught exception handler used when any worker abruptly
1100	<	* terminates.
1099	>	* Permission required for callers of methods that may start or
1100	>	* kill threads.
1101		*/
1102	<	private final Thread.UncaughtExceptionHandler ueh;
1102	>	private static final RuntimePermission modifyThreadPermission;
1103
1104		/**
1105	<	* Pool number, just for assigning useful names to worker threads
1105	>	* Per-thread submission bookeeping. Shared across all pools
1106	>	* to reduce ThreadLocal pollution and because random motion
1107	>	* to avoid contention in one pool is likely to hold for others.
1108		*/
1109	<	private final int poolNumber;
1109	>	private static final ThreadSubmitter submitters;
1110
1111	<	// Utilities for CASing fields. Note that most of these
586	<	// are usually manually inlined by callers
1111	>	// static constants
1112
1113		/**
1114	<	* Increments running count part of workerCounts
1114	>	* The wakeup interval (in nanoseconds) for a worker waiting for a
1115	>	* task when the pool is quiescent to instead try to shrink the
1116	>	* number of workers.  The exact value does not matter too
1117	>	* much. It must be short enough to release resources during
1118	>	* sustained periods of idleness, but not so short that threads
1119	>	* are continually re-created.
1120		*/
1121	<	final void incrementRunningCount() {
1122	<	int c;
593	<	do {} while (!UNSAFE.compareAndSwapInt(this, workerCountsOffset,
594	<	c = workerCounts,
595	<	c + ONE_RUNNING));
596	<	}
1121	>	private static final long SHRINK_RATE =
1122	>	4L * 1000L * 1000L * 1000L; // 4 seconds
1123
1124		/**
1125	<	* Tries to increment running count part of workerCounts
1125	>	* The timeout value for attempted shrinkage, includes
1126	>	* some slop to cope with system timer imprecision.
1127		*/
1128	<	final boolean tryIncrementRunningCount() {
602	<	int c;
603	<	return UNSAFE.compareAndSwapInt(this, workerCountsOffset,
604	<	c = workerCounts,
605	<	c + ONE_RUNNING);
606	<	}
1128	>	private static final long SHRINK_TIMEOUT = SHRINK_RATE - (SHRINK_RATE / 10);
1129
1130		/**
1131	<	* Tries to decrement running count unless already zero
1132	<	*/
1133	<	final boolean tryDecrementRunningCount() {
1134	<	int wc = workerCounts;
1135	<	if ((wc & RUNNING_COUNT_MASK) == 0)
1136	<	return false;
1137	<	return UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1138	<	wc, wc - ONE_RUNNING);
1139	<	}
1131	>	* The maximum stolen->joining link depth allowed in method
1132	>	* tryHelpStealer.  Must be a power of two. This value also
1133	>	* controls the maximum number of times to try to help join a task
1134	>	* without any apparent progress or change in pool state before
1135	>	* giving up and blocking (see awaitJoin).  Depths for legitimate
1136	>	* chains are unbounded, but we use a fixed constant to avoid
1137	>	* (otherwise unchecked) cycles and to bound staleness of
1138	>	* traversal parameters at the expense of sometimes blocking when
1139	>	* we could be helping.
1140	>	*/
1141	>	private static final int MAX_HELP = 32;
1142	>
1143	>	/**
1144	>	* Secondary time-based bound (in nanosecs) for helping attempts
1145	>	* before trying compensated blocking in awaitJoin. Used in
1146	>	* conjunction with MAX_HELP to reduce variance due to different
1147	>	* polling rates associated with different helping options. The
1148	>	* value should roughly approximate the time required to create
1149	>	* and/or activate a worker thread.
1150	>	*/
1151	>	private static final long COMPENSATION_DELAY = 100L * 1000L; // 0.1 millisec
1152	>
1153	>	/**
1154	>	* Increment for seed generators. See class ThreadLocal for
1155	>	* explanation.
1156	>	*/
1157	>	private static final int SEED_INCREMENT = 0x61c88647;
1158	>
1159	>	/**
1160	>	* Bits and masks for control variables
1161	>	*
1162	>	* Field ctl is a long packed with:
1163	>	* AC: Number of active running workers minus target parallelism (16 bits)
1164	>	* TC: Number of total workers minus target parallelism (16 bits)
1165	>	* ST: true if pool is terminating (1 bit)
1166	>	* EC: the wait count of top waiting thread (15 bits)
1167	>	* ID: poolIndex of top of Treiber stack of waiters (16 bits)
1168	>	*
1169	>	* When convenient, we can extract the upper 32 bits of counts and
1170	>	* the lower 32 bits of queue state, u = (int)(ctl >>> 32) and e =
1171	>	* (int)ctl.  The ec field is never accessed alone, but always
1172	>	* together with id and st. The offsets of counts by the target
1173	>	* parallelism and the positionings of fields makes it possible to
1174	>	* perform the most common checks via sign tests of fields: When
1175	>	* ac is negative, there are not enough active workers, when tc is
1176	>	* negative, there are not enough total workers, and when e is
1177	>	* negative, the pool is terminating.  To deal with these possibly
1178	>	* negative fields, we use casts in and out of "short" and/or
1179	>	* signed shifts to maintain signedness.
1180	>	*
1181	>	* When a thread is queued (inactivated), its eventCount field is
1182	>	* set negative, which is the only way to tell if a worker is
1183	>	* prevented from executing tasks, even though it must continue to
1184	>	* scan for them to avoid queuing races. Note however that
1185	>	* eventCount updates lag releases so usage requires care.
1186	>	*
1187	>	* Field runState is an int packed with:
1188	>	* SHUTDOWN: true if shutdown is enabled (1 bit)
1189	>	* SEQ:  a sequence number updated upon (de)registering workers (30 bits)
1190	>	* INIT: set true after workQueues array construction (1 bit)
1191	>	*
1192	>	* The sequence number enables simple consistency checks:
1193	>	* Staleness of read-only operations on the workQueues array can
1194	>	* be checked by comparing runState before vs after the reads.
1195	>	*/
1196	>
1197	>	// bit positions/shifts for fields
1198	>	private static final int  AC_SHIFT   = 48;
1199	>	private static final int  TC_SHIFT   = 32;
1200	>	private static final int  ST_SHIFT   = 31;
1201	>	private static final int  EC_SHIFT   = 16;
1202	>
1203	>	// bounds
1204	>	private static final int  SMASK      = 0xffff;  // short bits
1205	>	private static final int  MAX_CAP    = 0x7fff;  // max #workers - 1
1206	>	private static final int  SQMASK     = 0xfffe;  // even short bits
1207	>	private static final int  SHORT_SIGN = 1 << 15;
1208	>	private static final int  INT_SIGN   = 1 << 31;
1209	>
1210	>	// masks
1211	>	private static final long STOP_BIT   = 0x0001L << ST_SHIFT;
1212	>	private static final long AC_MASK    = ((long)SMASK) << AC_SHIFT;
1213	>	private static final long TC_MASK    = ((long)SMASK) << TC_SHIFT;
1214	>
1215	>	// units for incrementing and decrementing
1216	>	private static final long TC_UNIT    = 1L << TC_SHIFT;
1217	>	private static final long AC_UNIT    = 1L << AC_SHIFT;
1218	>
1219	>	// masks and units for dealing with u = (int)(ctl >>> 32)
1220	>	private static final int  UAC_SHIFT  = AC_SHIFT - 32;
1221	>	private static final int  UTC_SHIFT  = TC_SHIFT - 32;
1222	>	private static final int  UAC_MASK   = SMASK << UAC_SHIFT;
1223	>	private static final int  UTC_MASK   = SMASK << UTC_SHIFT;
1224	>	private static final int  UAC_UNIT   = 1 << UAC_SHIFT;
1225	>	private static final int  UTC_UNIT   = 1 << UTC_SHIFT;
1226	>
1227	>	// masks and units for dealing with e = (int)ctl
1228	>	private static final int E_MASK      = 0x7fffffff; // no STOP_BIT
1229	>	private static final int E_SEQ       = 1 << EC_SHIFT;
1230	>
1231	>	// runState bits
1232	>	private static final int SHUTDOWN    = 1 << 31;
1233	>
1234	>	// access mode for WorkQueue
1235	>	static final int LIFO_QUEUE          =  0;
1236	>	static final int FIFO_QUEUE          =  1;
1237	>	static final int SHARED_QUEUE        = -1;
1238
1239	<	/**
1240	<	* Forces decrement of encoded workerCounts, awaiting nonzero if
1241	<	* (rarely) necessary when other count updates lag.
1242	<	*
1243	<	* @param dr -- either zero or ONE_RUNNING
1244	<	* @param dt -- either zero or ONE_TOTAL
1245	<	*/
1246	<	private void decrementWorkerCounts(int dr, int dt) {
1247	<	for (;;) {
1248	<	int wc = workerCounts;
1249	<	if ((wc & RUNNING_COUNT_MASK)  - dr < 0 ||
1250	<	(wc >>> TOTAL_COUNT_SHIFT) - dt < 0) {
1251	<	if ((runState & TERMINATED) != 0)
1252	<	return; // lagging termination on a backout
1253	<	Thread.yield();
1254	<	}
1255	<	if (UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1256	<	wc, wc - (dr + dt)))
1239	>	// Instance fields
1240	>
1241	>	/*
1242	>	* Field layout order in this class tends to matter more than one
1243	>	* would like. Runtime layout order is only loosely related to
1244	>	* declaration order and may differ across JVMs, but the following
1245	>	* empirically works OK on current JVMs.
1246	>	*/
1247	>
1248	>	volatile long ctl;                         // main pool control
1249	>	final int parallelism;                     // parallelism level
1250	>	final int localMode;                       // per-worker scheduling mode
1251	>	final int submitMask;                      // submit queue index bound
1252	>	int nextSeed;                              // for initializing worker seeds
1253	>	volatile int runState;                     // shutdown status and seq
1254	>	WorkQueue[] workQueues;                    // main registry
1255	>	final Mutex lock;                          // for registration
1256	>	final Condition termination;               // for awaitTermination
1257	>	final ForkJoinWorkerThreadFactory factory; // factory for new workers
1258	>	final Thread.UncaughtExceptionHandler ueh; // per-worker UEH
1259	>	final AtomicLong stealCount;               // collect counts when terminated
1260	>	final AtomicInteger nextWorkerNumber;      // to create worker name string
1261	>	final String workerNamePrefix;             // to create worker name string
1262	>
1263	>	//  Creating, registering, and deregistering workers
1264	>
1265	>	/**
1266	>	* Tries to create and start a worker
1267	>	*/
1268	>	private void addWorker() {
1269	>	Throwable ex = null;
1270	>	ForkJoinWorkerThread wt = null;
1271	>	try {
1272	>	if ((wt = factory.newThread(this)) != null) {
1273	>	wt.start();
1274		return;
1275	+	}
1276	+	} catch (Throwable e) {
1277	+	ex = e;
1278		}
1279	+	deregisterWorker(wt, ex); // adjust counts etc on failure
1280		}
1281
1282		/**
1283	<	* Tries decrementing active count; fails on contention.
1284	<	* Called when workers cannot find tasks to run.
1285	<	*/
1286	<	final boolean tryDecrementActiveCount() {
646	<	int c;
647	<	return UNSAFE.compareAndSwapInt(this, runStateOffset,
648	<	c = runState, c - 1);
649	<	}
650	<
651	<	/**
652	<	* Advances to at least the given level. Returns true if not
653	<	* already in at least the given level.
1283	>	* Callback from ForkJoinWorkerThread constructor to assign a
1284	>	* public name. This must be separate from registerWorker because
1285	>	* it is called during the "super" constructor call in
1286	>	* ForkJoinWorkerThread.
1287		*/
1288	<	private boolean advanceRunLevel(int level) {
1289	<	for (;;) {
1290	<	int s = runState;
658	<	if ((s & level) != 0)
659	<	return false;
660	<	if (UNSAFE.compareAndSwapInt(this, runStateOffset, s, s | level))
661	<	return true;
662	<	}
1288	>	final String nextWorkerName() {
1289	>	return workerNamePrefix.concat
1290	>	(Integer.toString(nextWorkerNumber.addAndGet(1)));
1291		}
1292
665	–	// workers array maintenance
666	–
1293		/**
1294	<	* Records and returns a workers array index for new worker.
1294	>	* Callback from ForkJoinWorkerThread constructor to establish its
1295	>	* poolIndex and record its WorkQueue. To avoid scanning bias due
1296	>	* to packing entries in front of the workQueues array, we treat
1297	>	* the array as a simple power-of-two hash table using per-thread
1298	>	* seed as hash, expanding as needed.
1299	>	*
1300	>	* @param w the worker's queue
1301		*/
670	–	private int recordWorker(ForkJoinWorkerThread w) {
671	–	// Try using slot totalCount-1. If not available, scan and/or resize
672	–	int k = (workerCounts >>> TOTAL_COUNT_SHIFT) - 1;
673	–	final ReentrantLock lock = this.workerLock;
674	–	lock.lock();
675	–	try {
676	–	ForkJoinWorkerThread[] ws = workers;
677	–	int n = ws.length;
678	–	if (k < 0 || k >= n || ws[k] != null) {
679	–	for (k = 0; k < n && ws[k] != null; ++k)
680	–	;
681	–	if (k == n)
682	–	ws = workers = Arrays.copyOf(ws, n << 1);
683	–	}
684	–	ws[k] = w;
685	–	int c = eventCount; // advance event count to ensure visibility
686	–	UNSAFE.compareAndSwapInt(this, eventCountOffset, c, c+1);
687	–	} finally {
688	–	lock.unlock();
689	–	}
690	–	return k;
691	–	}
1302
1303	<	/**
1304	<	* Nulls out record of worker in workers array.
695	<	*/
696	<	private void forgetWorker(ForkJoinWorkerThread w) {
697	<	int idx = w.poolIndex;
698	<	// Locking helps method recordWorker avoid unnecessary expansion
699	<	final ReentrantLock lock = this.workerLock;
1303	>	final void registerWorker(WorkQueue w) {
1304	>	Mutex lock = this.lock;
1305		lock.lock();
1306		try {
1307	<	ForkJoinWorkerThread[] ws = workers;
1308	<	if (idx >= 0 && idx < ws.length && ws[idx] == w) // verify
1309	<	ws[idx] = null;
1307	>	WorkQueue[] ws = workQueues;
1308	>	if (w != null && ws != null) {          // skip on shutdown/failure
1309	>	int rs, n =  ws.length, m = n - 1;
1310	>	int s = nextSeed += SEED_INCREMENT; // rarely-colliding sequence
1311	>	w.seed = (s == 0) ? 1 : s;          // ensure non-zero seed
1312	>	int r = (s << 1) | 1;               // use odd-numbered indices
1313	>	if (ws[r &= m] != null) {           // collision
1314	>	int probes = 0;                 // step by approx half size
1315	>	int step = (n <= 4) ? 2 : ((n >>> 1) & SQMASK) + 2;
1316	>	while (ws[r = (r + step) & m] != null) {
1317	>	if (++probes >= n) {
1318	>	workQueues = ws = Arrays.copyOf(ws, n <<= 1);
1319	>	m = n - 1;
1320	>	probes = 0;
1321	>	}
1322	>	}
1323	>	}
1324	>	w.eventCount = w.poolIndex = r;     // establish before recording
1325	>	ws[r] = w;                          // also update seq
1326	>	runState = ((rs = runState) & SHUTDOWN) | ((rs + 2) & ~SHUTDOWN);
1327	>	}
1328		} finally {
1329		lock.unlock();
1330		}
1331		}
1332
1333		/**
1334	<	* Final callback from terminating worker.  Removes record of
1334	>	* Final callback from terminating worker, as well as upon failure
1335	>	* to construct or start a worker in addWorker.  Removes record of
1336		* worker from array, and adjusts counts. If pool is shutting
1337		* down, tries to complete termination.
1338		*
1339	<	* @param w the worker
1339	>	* @param wt the worker thread or null if addWorker failed
1340	>	* @param ex the exception causing failure, or null if none
1341		*/
1342	<	final void workerTerminated(ForkJoinWorkerThread w) {
1343	<	forgetWorker(w);
1344	<	decrementWorkerCounts(w.isTrimmed() ? 0 : ONE_RUNNING, ONE_TOTAL);
1345	<	while (w.stealCount != 0) // collect final count
1346	<	tryAccumulateStealCount(w);
1347	<	tryTerminate(false);
1348	<	}
1349	<
1350	<	// Waiting for and signalling events
1351	<
1352	<	/**
1353	<	* Releases workers blocked on a count not equal to current count.
1354	<	* Normally called after precheck that eventWaiters isn't zero to
1355	<	* avoid wasted array checks. Gives up upon a change in count or
731	<	* upon releasing two workers, letting others take over.
732	<	*/
733	<	private void releaseEventWaiters() {
734	<	ForkJoinWorkerThread[] ws = workers;
735	<	int n = ws.length;
736	<	long h = eventWaiters;
737	<	int ec = eventCount;
738	<	boolean releasedOne = false;
739	<	ForkJoinWorkerThread w; int id;
740	<	while ((id = ((int)(h & WAITER_ID_MASK)) - 1) >= 0 &&
741	<	(int)(h >>> EVENT_COUNT_SHIFT) != ec &&
742	<	id < n && (w = ws[id]) != null) {
743	<	if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
744	<	h,  w.nextWaiter)) {
745	<	LockSupport.unpark(w);
746	<	if (releasedOne) // exit on second release
747	<	break;
748	<	releasedOne = true;
1342	>	final void deregisterWorker(ForkJoinWorkerThread wt, Throwable ex) {
1343	>	Mutex lock = this.lock;
1344	>	WorkQueue w = null;
1345	>	if (wt != null && (w = wt.workQueue) != null) {
1346	>	w.runState = -1;                // ensure runState is set
1347	>	stealCount.getAndAdd(w.totalSteals + w.nsteals);
1348	>	int idx = w.poolIndex;
1349	>	lock.lock();
1350	>	try {                           // remove record from array
1351	>	WorkQueue[] ws = workQueues;
1352	>	if (ws != null && idx >= 0 && idx < ws.length && ws[idx] == w)
1353	>	ws[idx] = null;
1354	>	} finally {
1355	>	lock.unlock();
1356		}
750	–	if (eventCount != ec)
751	–	break;
752	–	h = eventWaiters;
1357		}
754	–	}
1358
1359	<	/**
1360	<	* Tries to advance eventCount and releases waiters. Called only
1361	<	* from workers.
1362	<	*/
1363	<	final void signalWork() {
1364	<	int c; // try to increment event count -- CAS failure OK
1365	<	UNSAFE.compareAndSwapInt(this, eventCountOffset, c = eventCount, c+1);
1366	<	if (eventWaiters != 0L)
1367	<	releaseEventWaiters();
1359	>	long c;                             // adjust ctl counts
1360	>	do {} while (!U.compareAndSwapLong
1361	>	(this, CTL, c = ctl, (((c - AC_UNIT) & AC_MASK) |
1362	>	((c - TC_UNIT) & TC_MASK) |
1363	>	(c & ~(AC_MASK|TC_MASK)))));
1364	>
1365	>	if (!tryTerminate(false, false) && w != null) {
1366	>	w.cancelAll();                  // cancel remaining tasks
1367	>	if (w.array != null)            // suppress signal if never ran
1368	>	signalWork();               // wake up or create replacement
1369	>	if (ex == null)                 // help clean refs on way out
1370	>	ForkJoinTask.helpExpungeStaleExceptions();
1371	>	}
1372	>
1373	>	if (ex != null)                     // rethrow
1374	>	U.throwException(ex);
1375		}
1376
767	–	/**
768	–	* Adds the given worker to event queue and blocks until
769	–	* terminating or event count advances from the given value
770	–	*
771	–	* @param w the calling worker thread
772	–	* @param ec the count
773	–	*/
774	–	private void eventSync(ForkJoinWorkerThread w, int ec) {
775	–	long nh = (((long)ec) << EVENT_COUNT_SHIFT) | ((long)(w.poolIndex+1));
776	–	long h;
777	–	while ((runState < SHUTDOWN || !tryTerminate(false)) &&
778	–	(((int)((h = eventWaiters) & WAITER_ID_MASK)) == 0 ||
779	–	(int)(h >>> EVENT_COUNT_SHIFT) == ec) &&
780	–	eventCount == ec) {
781	–	if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
782	–	w.nextWaiter = h, nh)) {
783	–	awaitEvent(w, ec);
784	–	break;
785	–	}
786	–	}
787	–	}
1377
1378	<	/**
1379	<	* Blocks the given worker (that has already been entered as an
1380	<	* event waiter) until terminating or event count advances from
1381	<	* the given value. The oldest (first) waiter uses a timed wait to
1382	<	* occasionally one-by-one shrink the number of workers (to a
1383	<	* minimum of one) if the pool has not been used for extended
1384	<	* periods.
1385	<	*
1386	<	* @param w the calling worker thread
1387	<	* @param ec the count
1388	<	*/
1389	<	private void awaitEvent(ForkJoinWorkerThread w, int ec) {
1390	<	while (eventCount == ec) {
1391	<	if (tryAccumulateStealCount(w)) { // transfer while idle
1392	<	boolean untimed = (w.nextWaiter != 0L ||
1393	<	(workerCounts & RUNNING_COUNT_MASK) <= 1);
1394	<	long startTime = untimed ? 0 : System.nanoTime();
1395	<	Thread.interrupted();         // clear/ignore interrupt
1396	<	if (w.isTerminating() || eventCount != ec)
1397	<	break;                    // recheck after clear
1398	<	if (untimed)
1399	<	LockSupport.park(w);
1400	<	else {
1401	<	LockSupport.parkNanos(w, SHRINK_RATE_NANOS);
1402	<	if (eventCount != ec || w.isTerminating())
1403	<	break;
1404	<	if (System.nanoTime() - startTime >= SHRINK_RATE_NANOS)
1405	<	tryShutdownUnusedWorker(ec);
1378	>	// Submissions
1379	>
1380	>	/**
1381	>	* Unless shutting down, adds the given task to a submission queue
1382	>	* at submitter's current queue index (modulo submission
1383	>	* range). If no queue exists at the index, one is created.  If
1384	>	* the queue is busy, another index is randomly chosen. The
1385	>	* submitMask bounds the effective number of queues to the
1386	>	* (nearest power of two for) parallelism level.
1387	>	*
1388	>	* @param task the task. Caller must ensure non-null.
1389	>	*/
1390	>	private void doSubmit(ForkJoinTask<?> task) {
1391	>	Submitter s = submitters.get();
1392	>	for (int r = s.seed, m = submitMask;;) {
1393	>	WorkQueue[] ws; WorkQueue q;
1394	>	int k = r & m & SQMASK;          // use only even indices
1395	>	if (runState < 0 || (ws = workQueues) == null || ws.length <= k)
1396	>	throw new RejectedExecutionException(); // shutting down
1397	>	else if ((q = ws[k]) == null) {  // create new queue
1398	>	WorkQueue nq = new WorkQueue(this, null, SHARED_QUEUE);
1399	>	Mutex lock = this.lock;      // construct outside lock
1400	>	lock.lock();
1401	>	try {                        // recheck under lock
1402	>	int rs = runState;       // to update seq
1403	>	if (ws == workQueues && ws[k] == null) {
1404	>	ws[k] = nq;
1405	>	runState = ((rs & SHUTDOWN) | ((rs + 2) & ~SHUTDOWN));
1406	>	}
1407	>	} finally {
1408	>	lock.unlock();
1409		}
1410		}
1411	+	else if (q.trySharedPush(task)) {
1412	+	signalWork();
1413	+	return;
1414	+	}
1415	+	else if (m > 1) {                // move to a different index
1416	+	r ^= r << 13;                // same xorshift as WorkQueues
1417	+	r ^= r >>> 17;
1418	+	s.seed = r ^= r << 5;
1419	+	}
1420	+	else
1421	+	Thread.yield();              // yield if no alternatives
1422		}
1423		}
1424
1425	<	// Maintaining parallelism
1425	>	// Maintaining ctl counts
1426
1427		/**
1428	<	* Pushes worker onto the spare stack.
1428	>	* Increments active count; mainly called upon return from blocking.
1429		*/
1430	<	final void pushSpare(ForkJoinWorkerThread w) {
1431	<	int ns = (++w.spareCount << SPARE_COUNT_SHIFT) | (w.poolIndex + 1);
1432	<	do {} while (!UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
830	<	w.nextSpare = spareWaiters,ns));
1430	>	final void incrementActiveCount() {
1431	>	long c;
1432	>	do {} while (!U.compareAndSwapLong(this, CTL, c = ctl, c + AC_UNIT));
1433		}
1434
1435		/**
1436	<	* Tries (once) to resume a spare if the number of running
835	<	* threads is less than target.
1436	>	* Tries to activate or create a worker if too few are active.
1437		*/
1438	<	private void tryResumeSpare() {
1439	<	int sw, id;
1440	<	ForkJoinWorkerThread[] ws = workers;
1441	<	int n = ws.length;
1442	<	ForkJoinWorkerThread w;
1443	<	if ((sw = spareWaiters) != 0 &&
1444	<	(id = (sw & SPARE_ID_MASK) - 1) >= 0 &&
1445	<	id < n && (w = ws[id]) != null &&
1446	<	(runState >= TERMINATING ||
1447	<	(workerCounts & RUNNING_COUNT_MASK) < parallelism) &&
1448	<	spareWaiters == sw &&
1449	<	UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
1450	<	sw, w.nextSpare)) {
1451	<	int c; // increment running count before resume
1452	<	do {} while (!UNSAFE.compareAndSwapInt
1453	<	(this, workerCountsOffset,
1454	<	c = workerCounts, c + ONE_RUNNING));
1455	<	if (w.tryUnsuspend())
1456	<	LockSupport.unpark(w);
1457	<	else   // back out if w was shutdown
1458	<	decrementWorkerCounts(ONE_RUNNING, 0);
1438	>	final void signalWork() {
1439	>	long c; int u;
1440	>	while ((u = (int)((c = ctl) >>> 32)) < 0) {     // too few active
1441	>	WorkQueue[] ws = workQueues; int e, i; WorkQueue w; Thread p;
1442	>	if ((e = (int)c) > 0) {                     // at least one waiting
1443	>	if (ws != null && (i = e & SMASK) < ws.length &&
1444	>	(w = ws[i]) != null && w.eventCount == (e | INT_SIGN)) {
1445	>	long nc = (((long)(w.nextWait & E_MASK)) |
1446	>	((long)(u + UAC_UNIT) << 32));
1447	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1448	>	w.eventCount = (e + E_SEQ) & E_MASK;
1449	>	if ((p = w.parker) != null)
1450	>	U.unpark(p);                // activate and release
1451	>	break;
1452	>	}
1453	>	}
1454	>	else
1455	>	break;
1456	>	}
1457	>	else if (e == 0 && (u & SHORT_SIGN) != 0) { // too few total
1458	>	long nc = (long)(((u + UTC_UNIT) & UTC_MASK) |
1459	>	((u + UAC_UNIT) & UAC_MASK)) << 32;
1460	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1461	>	addWorker();
1462	>	break;
1463	>	}
1464	>	}
1465	>	else
1466	>	break;
1467		}
1468		}
1469
1470	+	// Scanning for tasks
1471	+
1472		/**
1473	<	* Tries to increase the number of running workers if below target
863	<	* parallelism: If a spare exists tries to resume it via
864	<	* tryResumeSpare.  Otherwise, if not enough total workers or all
865	<	* existing workers are busy, adds a new worker. In all cases also
866	<	* helps wake up releasable workers waiting for work.
1473	>	* Top-level runloop for workers, called by ForkJoinWorkerThread.run.
1474		*/
1475	<	private void helpMaintainParallelism() {
1476	<	int pc = parallelism;
1477	<	int wc, rs, tc;
1478	<	while (((wc = workerCounts) & RUNNING_COUNT_MASK) < pc &&
1479	<	(rs = runState) < TERMINATING) {
1480	<	if (spareWaiters != 0)
1481	<	tryResumeSpare();
1482	<	else if ((tc = wc >>> TOTAL_COUNT_SHIFT) >= MAX_WORKERS ||
1483	<	(tc >= pc && (rs & ACTIVE_COUNT_MASK) != tc))
1484	<	break;   // enough total
1485	<	else if (runState == rs && workerCounts == wc &&
1486	<	UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
1487	<	wc + (ONE_RUNNING|ONE_TOTAL))) {
1488	<	ForkJoinWorkerThread w = null;
1489	<	Throwable fail = null;
1490	<	try {
1491	<	w = factory.newThread(this);
1492	<	} catch (Throwable ex) {
1493	<	fail = ex;
1494	<	}
1495	<	if (w == null) { // null or exceptional factory return
1496	<	decrementWorkerCounts(ONE_RUNNING, ONE_TOTAL);
1497	<	tryTerminate(false); // handle failure during shutdown
1498	<	// If originating from an external caller,
1499	<	// propagate exception, else ignore
1500	<	if (fail != null && runState < TERMINATING &&
1501	<	!(Thread.currentThread() instanceof
1502	<	ForkJoinWorkerThread))
1503	<	UNSAFE.throwException(fail);
1475	>	final void runWorker(WorkQueue w) {
1476	>	w.growArray(false);         // initialize queue array in this thread
1477	>	do { w.runTask(scan(w)); } while (w.runState >= 0);
1478	>	}
1479	>
1480	>	/**
1481	>	* Scans for and, if found, returns one task, else possibly
1482	>	* inactivates the worker. This method operates on single reads of
1483	>	* volatile state and is designed to be re-invoked continuously,
1484	>	* in part because it returns upon detecting inconsistencies,
1485	>	* contention, or state changes that indicate possible success on
1486	>	* re-invocation.
1487	>	*
1488	>	* The scan searches for tasks across a random permutation of
1489	>	* queues (starting at a random index and stepping by a random
1490	>	* relative prime, checking each at least once).  The scan
1491	>	* terminates upon either finding a non-empty queue, or completing
1492	>	* the sweep. If the worker is not inactivated, it takes and
1493	>	* returns a task from this queue.  On failure to find a task, we
1494	>	* take one of the following actions, after which the caller will
1495	>	* retry calling this method unless terminated.
1496	>	*
1497	>	* * If pool is terminating, terminate the worker.
1498	>	*
1499	>	* * If not a complete sweep, try to release a waiting worker.  If
1500	>	* the scan terminated because the worker is inactivated, then the
1501	>	* released worker will often be the calling worker, and it can
1502	>	* succeed obtaining a task on the next call. Or maybe it is
1503	>	* another worker, but with same net effect. Releasing in other
1504	>	* cases as well ensures that we have enough workers running.
1505	>	*
1506	>	* * If not already enqueued, try to inactivate and enqueue the
1507	>	* worker on wait queue. Or, if inactivating has caused the pool
1508	>	* to be quiescent, relay to idleAwaitWork to check for
1509	>	* termination and possibly shrink pool.
1510	>	*
1511	>	* * If already inactive, and the caller has run a task since the
1512	>	* last empty scan, return (to allow rescan) unless others are
1513	>	* also inactivated.  Field WorkQueue.rescans counts down on each
1514	>	* scan to ensure eventual inactivation and blocking.
1515	>	*
1516	>	* * If already enqueued and none of the above apply, park
1517	>	* awaiting signal,
1518	>	*
1519	>	* @param w the worker (via its WorkQueue)
1520	>	* @return a task or null of none found
1521	>	*/
1522	>	private final ForkJoinTask<?> scan(WorkQueue w) {
1523	>	WorkQueue[] ws;                       // first update random seed
1524	>	int r = w.seed; r ^= r << 13; r ^= r >>> 17; w.seed = r ^= r << 5;
1525	>	int rs = runState, m;                 // volatile read order matters
1526	>	if ((ws = workQueues) != null && (m = ws.length - 1) > 0) {
1527	>	int ec = w.eventCount;            // ec is negative if inactive
1528	>	int step = (r >>> 16) | 1;        // relative prime
1529	>	for (int j = (m + 1) << 2; ; r += step) {
1530	>	WorkQueue q; ForkJoinTask<?> t; ForkJoinTask<?>[] a; int b;
1531	>	if ((q = ws[r & m]) != null && (b = q.base) - q.top < 0 &&
1532	>	(a = q.array) != null) {  // probably nonempty
1533	>	int i = (((a.length - 1) & b) << ASHIFT) + ABASE;
1534	>	t = (ForkJoinTask<?>)U.getObjectVolatile(a, i);
1535	>	if (q.base == b && ec >= 0 && t != null &&
1536	>	U.compareAndSwapObject(a, i, t, null)) {
1537	>	q.base = b + 1;       // specialization of pollAt
1538	>	return t;
1539	>	}
1540	>	else if (ec < 0 || j <= m) {
1541	>	rs = 0;               // mark scan as imcomplete
1542	>	break;                // caller can retry after release
1543	>	}
1544	>	}
1545	>	if (--j < 0)
1546		break;
1547	+	}
1548	+	long c = ctl; int e = (int)c, a = (int)(c >> AC_SHIFT), nr, ns;
1549	+	if (e < 0)                        // decode ctl on empty scan
1550	+	w.runState = -1;              // pool is terminating
1551	+	else if (rs == 0 || rs != runState) { // incomplete scan
1552	+	WorkQueue v; Thread p;        // try to release a waiter
1553	+	if (e > 0 && a < 0 && w.eventCount == ec &&
1554	+	(v = ws[e & m]) != null && v.eventCount == (e | INT_SIGN)) {
1555	+	long nc = ((long)(v.nextWait & E_MASK) |
1556	+	((c + AC_UNIT) & (AC_MASK|TC_MASK)));
1557	+	if (ctl == c && U.compareAndSwapLong(this, CTL, c, nc)) {
1558	+	v.eventCount = (e + E_SEQ) & E_MASK;
1559	+	if ((p = v.parker) != null)
1560	+	U.unpark(p);
1561	+	}
1562		}
899	–	w.start(recordWorker(w), ueh);
900	–	if ((workerCounts >>> TOTAL_COUNT_SHIFT) >= pc)
901	–	break; // add at most one unless total below target
902	–	}
903	–	}
904	–	if (eventWaiters != 0L)
905	–	releaseEventWaiters();
906	–	}
907	–
908	–	/**
909	–	* Callback from the oldest waiter in awaitEvent waking up after a
910	–	* period of non-use. If all workers are idle, tries (once) to
911	–	* shutdown an event waiter or a spare, if one exists. Note that
912	–	* we don't need CAS or locks here because the method is called
913	–	* only from one thread occasionally waking (and even misfires are
914	–	* OK). Note that until the shutdown worker fully terminates,
915	–	* workerCounts will overestimate total count, which is tolerable.
916	–	*
917	–	* @param ec the event count waited on by caller (to abort
918	–	* attempt if count has since changed).
919	–	*/
920	–	private void tryShutdownUnusedWorker(int ec) {
921	–	if (runState == 0 && eventCount == ec) { // only trigger if all idle
922	–	ForkJoinWorkerThread[] ws = workers;
923	–	int n = ws.length;
924	–	ForkJoinWorkerThread w = null;
925	–	boolean shutdown = false;
926	–	int sw;
927	–	long h;
928	–	if ((sw = spareWaiters) != 0) { // prefer killing spares
929	–	int id = (sw & SPARE_ID_MASK) - 1;
930	–	if (id >= 0 && id < n && (w = ws[id]) != null &&
931	–	UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
932	–	sw, w.nextSpare))
933	–	shutdown = true;
934	–	}
935	–	else if ((h = eventWaiters) != 0L) {
936	–	long nh;
937	–	int id = ((int)(h & WAITER_ID_MASK)) - 1;
938	–	if (id >= 0 && id < n && (w = ws[id]) != null &&
939	–	(nh = w.nextWaiter) != 0L && // keep at least one worker
940	–	UNSAFE.compareAndSwapLong(this, eventWaitersOffset, h, nh))
941	–	shutdown = true;
942	–	}
943	–	if (w != null && shutdown) {
944	–	w.shutdown();
945	–	LockSupport.unpark(w);
946	–	}
947	–	}
948	–	releaseEventWaiters(); // in case of interference
949	–	}
950	–
951	–	/**
952	–	* Callback from workers invoked upon each top-level action (i.e.,
953	–	* stealing a task or taking a submission and running it).
954	–	* Performs one or more of the following:
955	–	*
956	–	* 1. If the worker is active and either did not run a task
957	–	*    or there are too many workers, try to set its active status
958	–	*    to inactive and update activeCount. On contention, we may
959	–	*    try again in this or a subsequent call.
960	–	*
961	–	* 2. If not enough total workers, help create some.
962	–	*
963	–	* 3. If there are too many running workers, suspend this worker
964	–	*    (first forcing inactive if necessary).  If it is not needed,
965	–	*    it may be shutdown while suspended (via
966	–	*    tryShutdownUnusedWorker).  Otherwise, upon resume it
967	–	*    rechecks running thread count and need for event sync.
968	–	*
969	–	* 4. If worker did not run a task, await the next task event via
970	–	*    eventSync if necessary (first forcing inactivation), upon
971	–	*    which the worker may be shutdown via
972	–	*    tryShutdownUnusedWorker.  Otherwise, help release any
973	–	*    existing event waiters that are now releasable,
974	–	*
975	–	* @param w the worker
976	–	* @param ran true if worker ran a task since last call to this method
977	–	*/
978	–	final void preStep(ForkJoinWorkerThread w, boolean ran) {
979	–	int wec = w.lastEventCount;
980	–	boolean active = w.active;
981	–	boolean inactivate = false;
982	–	int pc = parallelism;
983	–	while (w.runState == 0) {
984	–	int rs = runState;
985	–	if (rs >= TERMINATING) {           // propagate shutdown
986	–	w.shutdown();
987	–	break;
1563		}
1564	<	if ((inactivate || (active && (rs & ACTIVE_COUNT_MASK) >= pc)) &&
1565	<	UNSAFE.compareAndSwapInt(this, runStateOffset, rs, --rs)) {
1566	<	inactivate = active = w.active = false;
1567	<	if (rs == SHUTDOWN) {          // all inactive and shut down
1568	<	tryTerminate(false);
1569	<	continue;
1570	<	}
1571	<	}
1572	<	int wc = workerCounts;             // try to suspend as spare
1573	<	if ((wc & RUNNING_COUNT_MASK) > pc) {
1574	<	if (!(inactivate |= active) && // must inactivate to suspend
1575	<	workerCounts == wc &&
1576	<	UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1577	<	wc, wc - ONE_RUNNING))
1578	<	w.suspendAsSpare();
1579	<	}
1580	<	else if ((wc >>> TOTAL_COUNT_SHIFT) < pc)
1581	<	helpMaintainParallelism();     // not enough workers
1582	<	else if (ran)
1583	<	break;
1584	<	else {
1585	<	long h = eventWaiters;
1586	<	int ec = eventCount;
1587	<	if (h != 0L && (int)(h >>> EVENT_COUNT_SHIFT) != ec)
1588	<	releaseEventWaiters();     // release others before waiting
1589	<	else if (ec != wec) {
1590	<	w.lastEventCount = ec;     // no need to wait
1591	<	break;
1564	>	else if (ec >= 0) {               // try to enqueue/inactivate
1565	>	long nc = (long)ec | ((c - AC_UNIT) & (AC_MASK|TC_MASK));
1566	>	w.nextWait = e;
1567	>	w.eventCount = ec | INT_SIGN; // mark as inactive
1568	>	if (ctl != c || !U.compareAndSwapLong(this, CTL, c, nc))
1569	>	w.eventCount = ec;        // unmark on CAS failure
1570	>	else {
1571	>	if ((ns = w.nsteals) != 0) {
1572	>	w.nsteals = 0;        // set rescans if ran task
1573	>	w.rescans = (a > 0) ? 0 : a + parallelism;
1574	>	w.totalSteals += ns;
1575	>	}
1576	>	if (a == 1 - parallelism) // quiescent
1577	>	idleAwaitWork(w, nc, c);
1578	>	}
1579	>	}
1580	>	else if (w.eventCount < 0) {      // already queued
1581	>	if ((nr = w.rescans) > 0) {   // continue rescanning
1582	>	int ac = a + parallelism;
1583	>	if (((w.rescans = (ac < nr) ? ac : nr - 1) & 3) == 0)
1584	>	Thread.yield();       // yield before block
1585	>	}
1586	>	else {
1587	>	Thread.interrupted();     // clear status
1588	>	Thread wt = Thread.currentThread();
1589	>	U.putObject(wt, PARKBLOCKER, this);
1590	>	w.parker = wt;            // emulate LockSupport.park
1591	>	if (w.eventCount < 0)     // recheck
1592	>	U.park(false, 0L);
1593	>	w.parker = null;
1594	>	U.putObject(wt, PARKBLOCKER, null);
1595		}
1018	–	else if (!(inactivate |= active))
1019	–	eventSync(w, wec);         // must inactivate before sync
1596		}
1597		}
1598	+	return null;
1599		}
1600
1601		/**
1602	<	* Helps and/or blocks awaiting join of the given task.
1603	<	* See above for explanation.
1604	<	*
1605	<	* @param joinMe the task to join
1606	<	* @param worker the current worker thread
1607	<	* @param timed true if wait should time out
1608	<	* @param nanos timeout value if timed
1609	<	*/
1610	<	final void awaitJoin(ForkJoinTask<?> joinMe, ForkJoinWorkerThread worker,
1611	<	boolean timed, long nanos) {
1612	<	long startTime = timed? System.nanoTime() : 0L;
1613	<	int retries = 2 + (parallelism >> 2); // #helpJoins before blocking
1614	<	boolean running = true;               // false when count decremented
1615	<	while (joinMe.status >= 0) {
1616	<	if (runState >= TERMINATING) {
1617	<	joinMe.cancelIgnoringExceptions();
1618	<	break;
1619	<	}
1620	<	running = worker.helpJoinTask(joinMe, running);
1621	<	if (joinMe.status < 0)
1622	<	break;
1623	<	if (retries > 0) {
1624	<	--retries;
1625	<	continue;
1626	<	}
1627	<	int wc = workerCounts;
1628	<	if ((wc & RUNNING_COUNT_MASK) != 0) {
1629	<	if (running) {
1630	<	if (!UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1631	<	wc, wc - ONE_RUNNING))
1632	<	continue;
1056	<	running = false;
1057	<	}
1058	<	long h = eventWaiters;
1059	<	if (h != 0L && (int)(h >>> EVENT_COUNT_SHIFT) != eventCount)
1060	<	releaseEventWaiters();
1061	<	if (joinMe.status < 0)
1602	>	* If inactivating worker w has caused the pool to become
1603	>	* quiescent, checks for pool termination, and, so long as this is
1604	>	* not the only worker, waits for event for up to SHRINK_RATE
1605	>	* nanosecs.  On timeout, if ctl has not changed, terminates the
1606	>	* worker, which will in turn wake up another worker to possibly
1607	>	* repeat this process.
1608	>	*
1609	>	* @param w the calling worker
1610	>	* @param currentCtl the ctl value triggering possible quiescence
1611	>	* @param prevCtl the ctl value to restore if thread is terminated
1612	>	*/
1613	>	private void idleAwaitWork(WorkQueue w, long currentCtl, long prevCtl) {
1614	>	if (w.eventCount < 0 && !tryTerminate(false, false) &&
1615	>	(int)prevCtl != 0 && !hasQueuedSubmissions() && ctl == currentCtl) {
1616	>	Thread wt = Thread.currentThread();
1617	>	Thread.yield();            // yield before block
1618	>	while (ctl == currentCtl) {
1619	>	long startTime = System.nanoTime();
1620	>	Thread.interrupted();  // timed variant of version in scan()
1621	>	U.putObject(wt, PARKBLOCKER, this);
1622	>	w.parker = wt;
1623	>	if (ctl == currentCtl)
1624	>	U.park(false, SHRINK_RATE);
1625	>	w.parker = null;
1626	>	U.putObject(wt, PARKBLOCKER, null);
1627	>	if (ctl != currentCtl)
1628	>	break;
1629	>	if (System.nanoTime() - startTime >= SHRINK_TIMEOUT &&
1630	>	U.compareAndSwapLong(this, CTL, currentCtl, prevCtl)) {
1631	>	w.eventCount = (w.eventCount + E_SEQ) | E_MASK;
1632	>	w.runState = -1;   // shrink
1633		break;
1063	–	if ((workerCounts & RUNNING_COUNT_MASK) != 0) {
1064	–	long ms; int ns;
1065	–	if (!timed) {
1066	–	ms = JOIN_TIMEOUT_MILLIS;
1067	–	ns = 0;
1068	–	}
1069	–	else { // at most JOIN_TIMEOUT_MILLIS per wait
1070	–	long nt = nanos - (System.nanoTime() - startTime);
1071	–	if (nt <= 0L)
1072	–	break;
1073	–	ms = nt / 1000000;
1074	–	if (ms > JOIN_TIMEOUT_MILLIS) {
1075	–	ms = JOIN_TIMEOUT_MILLIS;
1076	–	ns = 0;
1077	–	}
1078	–	else
1079	–	ns = (int) (nt % 1000000);
1080	–	}
1081	–	if (joinMe.internalAwaitDone(ms, ns) < 0)
1082	–	break;
1634		}
1635		}
1085	–	helpMaintainParallelism();
1086	–	}
1087	–	if (!running) {
1088	–	int c;
1089	–	do {} while (!UNSAFE.compareAndSwapInt
1090	–	(this, workerCountsOffset,
1091	–	c = workerCounts, c + ONE_RUNNING));
1636		}
1637		}
1638
1639		/**
1640	<	* Same idea as awaitJoin, but no helping, retries, or timeouts.
1641	<	*/
1642	<	final void awaitBlocker(ManagedBlocker blocker)
1643	<	throws InterruptedException {
1644	<	while (!blocker.isReleasable()) {
1645	<	int wc = workerCounts;
1646	<	if ((wc & RUNNING_COUNT_MASK) == 0)
1647	<	helpMaintainParallelism();
1648	<	else if (UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1649	<	wc, wc - ONE_RUNNING)) {
1650	<	try {
1651	<	while (!blocker.isReleasable()) {
1652	<	long h = eventWaiters;
1653	<	if (h != 0L &&
1654	<	(int)(h >>> EVENT_COUNT_SHIFT) != eventCount)
1655	<	releaseEventWaiters();
1656	<	else if ((workerCounts & RUNNING_COUNT_MASK) == 0 &&
1657	<	runState < TERMINATING)
1658	<	helpMaintainParallelism();
1659	<	else if (blocker.block())
1640	>	* Tries to locate and execute tasks for a stealer of the given
1641	>	* task, or in turn one of its stealers, Traces currentSteal ->
1642	>	* currentJoin links looking for a thread working on a descendant
1643	>	* of the given task and with a non-empty queue to steal back and
1644	>	* execute tasks from. The first call to this method upon a
1645	>	* waiting join will often entail scanning/search, (which is OK
1646	>	* because the joiner has nothing better to do), but this method
1647	>	* leaves hints in workers to speed up subsequent calls. The
1648	>	* implementation is very branchy to cope with potential
1649	>	* inconsistencies or loops encountering chains that are stale,
1650	>	* unknown, or so long that they are likely cyclic.  All of these
1651	>	* cases are dealt with by just retrying by caller.
1652	>	*
1653	>	* @param joiner the joining worker
1654	>	* @param task the task to join
1655	>	* @return true if found or ran a task (and so is immediately retryable)
1656	>	*/
1657	>	private boolean tryHelpStealer(WorkQueue joiner, ForkJoinTask<?> task) {
1658	>	WorkQueue[] ws;
1659	>	int m, depth = MAX_HELP;                // remaining chain depth
1660	>	boolean progress = false;
1661	>	if ((ws = workQueues) != null && (m = ws.length - 1) > 0 &&
1662	>	task.status >= 0) {
1663	>	ForkJoinTask<?> subtask = task;     // current target
1664	>	outer: for (WorkQueue j = joiner;;) {
1665	>	WorkQueue stealer = null;       // find stealer of subtask
1666	>	WorkQueue v = ws[j.stealHint & m]; // try hint
1667	>	if (v != null && v.currentSteal == subtask)
1668	>	stealer = v;
1669	>	else {                          // scan
1670	>	for (int i = 1; i <= m; i += 2) {
1671	>	if ((v = ws[i]) != null && v.currentSteal == subtask &&
1672	>	v != joiner) {
1673	>	stealer = v;
1674	>	j.stealHint = i;    // save hint
1675		break;
1676	+	}
1677	+	}
1678	+	if (stealer == null)
1679	+	break;
1680	+	}
1681	+
1682	+	for (WorkQueue q = stealer;;) { // try to help stealer
1683	+	ForkJoinTask[] a; ForkJoinTask<?> t; int b;
1684	+	if (task.status < 0)
1685	+	break outer;
1686	+	if ((b = q.base) - q.top < 0 && (a = q.array) != null) {
1687	+	progress = true;
1688	+	int i = (((a.length - 1) & b) << ASHIFT) + ABASE;
1689	+	t = (ForkJoinTask<?>)U.getObjectVolatile(a, i);
1690	+	if (subtask.status < 0) // must recheck before taking
1691	+	break outer;
1692	+	if (t != null &&
1693	+	q.base == b &&
1694	+	U.compareAndSwapObject(a, i, t, null)) {
1695	+	q.base = b + 1;
1696	+	joiner.runSubtask(t);
1697	+	}
1698	+	else if (q.base == b)
1699	+	break outer;        // possibly stalled
1700	+	}
1701	+	else {                      // descend
1702	+	ForkJoinTask<?> next = stealer.currentJoin;
1703	+	if (--depth <= 0 || subtask.status < 0 ||
1704	+	next == null || next == subtask)
1705	+	break outer;        // stale, dead-end, or cyclic
1706	+	subtask = next;
1707	+	j = stealer;
1708	+	break;
1709		}
1118	–	} finally {
1119	–	int c;
1120	–	do {} while (!UNSAFE.compareAndSwapInt
1121	–	(this, workerCountsOffset,
1122	–	c = workerCounts, c + ONE_RUNNING));
1710		}
1124	–	break;
1711		}
1712		}
1713	+	return progress;
1714		}
1715
1716		/**
1717	<	* Possibly initiates and/or completes termination.
1717	>	* If task is at base of some steal queue, steals and executes it.
1718		*
1719	<	* @param now if true, unconditionally terminate, else only
1720	<	* if shutdown and empty queue and no active workers
1134	<	* @return true if now terminating or terminated
1719	>	* @param joiner the joining worker
1720	>	* @param task the task
1721		*/
1722	<	private boolean tryTerminate(boolean now) {
1723	<	if (now)
1724	<	advanceRunLevel(SHUTDOWN); // ensure at least SHUTDOWN
1725	<	else if (runState < SHUTDOWN ||
1726	<	!submissionQueue.isEmpty() ||
1727	<	(runState & ACTIVE_COUNT_MASK) != 0)
1728	<	return false;
1729	<
1730	<	if (advanceRunLevel(TERMINATING))
1731	<	startTerminating();
1722	>	private void tryPollForAndExec(WorkQueue joiner, ForkJoinTask<?> task) {
1723	>	WorkQueue[] ws;
1724	>	if ((ws = workQueues) != null) {
1725	>	for (int j = 1; j < ws.length && task.status >= 0; j += 2) {
1726	>	WorkQueue q = ws[j];
1727	>	if (q != null && q.pollFor(task)) {
1728	>	joiner.runSubtask(task);
1729	>	break;
1730	>	}
1731	>	}
1732	>	}
1733	>	}
1734
1735	<	// Finish now if all threads terminated; else in some subsequent call
1736	<	if ((workerCounts >>> TOTAL_COUNT_SHIFT) == 0) {
1737	<	advanceRunLevel(TERMINATED);
1738	<	termination.forceTermination();
1739	<	}
1740	<	return true;
1735	>	/**
1736	>	* Tries to decrement active count (sometimes implicitly) and
1737	>	* possibly release or create a compensating worker in preparation
1738	>	* for blocking. Fails on contention or termination. Otherwise,
1739	>	* adds a new thread if no idle workers are available and either
1740	>	* pool would become completely starved or: (at least half
1741	>	* starved, and fewer than 50% spares exist, and there is at least
1742	>	* one task apparently available). Even though the availability
1743	>	* check requires a full scan, it is worthwhile in reducing false
1744	>	* alarms.
1745	>	*
1746	>	* @param task if non-null, a task being waited for
1747	>	* @param blocker if non-null, a blocker being waited for
1748	>	* @return true if the caller can block, else should recheck and retry
1749	>	*/
1750	>	final boolean tryCompensate(ForkJoinTask<?> task, ManagedBlocker blocker) {
1751	>	int pc = parallelism, e;
1752	>	long c = ctl;
1753	>	WorkQueue[] ws = workQueues;
1754	>	if ((e = (int)c) >= 0 && ws != null) {
1755	>	int u, a, ac, hc;
1756	>	int tc = (short)((u = (int)(c >>> 32)) >>> UTC_SHIFT) + pc;
1757	>	boolean replace = false;
1758	>	if ((a = u >> UAC_SHIFT) <= 0) {
1759	>	if ((ac = a + pc) <= 1)
1760	>	replace = true;
1761	>	else if ((e > 0 || (task != null &&
1762	>	ac <= (hc = pc >>> 1) && tc < pc + hc))) {
1763	>	WorkQueue w;
1764	>	for (int j = 0; j < ws.length; ++j) {
1765	>	if ((w = ws[j]) != null && !w.isEmpty()) {
1766	>	replace = true;
1767	>	break;   // in compensation range and tasks available
1768	>	}
1769	>	}
1770	>	}
1771	>	}
1772	>	if ((task == null || task.status >= 0) && // recheck need to block
1773	>	(blocker == null || !blocker.isReleasable()) && ctl == c) {
1774	>	if (!replace) {          // no compensation
1775	>	long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK);
1776	>	if (U.compareAndSwapLong(this, CTL, c, nc))
1777	>	return true;
1778	>	}
1779	>	else if (e != 0) {       // release an idle worker
1780	>	WorkQueue w; Thread p; int i;
1781	>	if ((i = e & SMASK) < ws.length && (w = ws[i]) != null) {
1782	>	long nc = ((long)(w.nextWait & E_MASK) |
1783	>	(c & (AC_MASK|TC_MASK)));
1784	>	if (w.eventCount == (e | INT_SIGN) &&
1785	>	U.compareAndSwapLong(this, CTL, c, nc)) {
1786	>	w.eventCount = (e + E_SEQ) & E_MASK;
1787	>	if ((p = w.parker) != null)
1788	>	U.unpark(p);
1789	>	return true;
1790	>	}
1791	>	}
1792	>	}
1793	>	else if (tc < MAX_CAP) { // create replacement
1794	>	long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK);
1795	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1796	>	addWorker();
1797	>	return true;
1798	>	}
1799	>	}
1800	>	}
1801	>	}
1802	>	return false;
1803		}
1804
1805		/**
1806	<	* Actions on transition to TERMINATING
1807	<	*
1808	<	* Runs up to four passes through workers: (0) shutting down each
1809	<	* (without waking up if parked) to quickly spread notifications
1810	<	* without unnecessary bouncing around event queues etc (1) wake
1811	<	* up and help cancel tasks (2) interrupt (3) mop up races with
1812	<	* interrupted workers
1813	<	*/
1814	<	private void startTerminating() {
1815	<	cancelSubmissions();
1816	<	for (int passes = 0; passes < 4 && workerCounts != 0; ++passes) {
1817	<	int c; // advance event count
1818	<	UNSAFE.compareAndSwapInt(this, eventCountOffset,
1819	<	c = eventCount, c+1);
1820	<	eventWaiters = 0L; // clobber lists
1821	<	spareWaiters = 0;
1822	<	for (ForkJoinWorkerThread w : workers) {
1823	<	if (w != null) {
1824	<	w.shutdown();
1825	<	if (passes > 0 && !w.isTerminated()) {
1826	<	w.cancelTasks();
1827	<	LockSupport.unpark(w);
1828	<	if (passes > 1 && !w.isInterrupted()) {
1829	<	try {
1830	<	w.interrupt();
1831	<	} catch (SecurityException ignore) {
1806	>	* Helps and/or blocks until the given task is done.
1807	>	*
1808	>	* @param joiner the joining worker
1809	>	* @param task the task
1810	>	* @return task status on exit
1811	>	*/
1812	>	final int awaitJoin(WorkQueue joiner, ForkJoinTask<?> task) {
1813	>	int s;
1814	>	ForkJoinTask<?> prevJoin = joiner.currentJoin;
1815	>	if ((s = task.status) >= 0) {
1816	>	joiner.currentJoin = task;
1817	>	long startTime = 0L;
1818	>	for (int k = 0;;) {
1819	>	if ((joiner.isEmpty() ?                  // try to help
1820	>	!tryHelpStealer(joiner, task) :
1821	>	!joiner.tryRemoveAndExec(task))) {
1822	>	if (k == 0) {
1823	>	startTime = System.nanoTime();
1824	>	tryPollForAndExec(joiner, task); // check uncommon case
1825	>	}
1826	>	else if ((k & (MAX_HELP - 1)) == 0 &&
1827	>	System.nanoTime() - startTime >=
1828	>	COMPENSATION_DELAY &&
1829	>	tryCompensate(task, null)) {
1830	>	if (task.trySetSignal() && task.status >= 0) {
1831	>	synchronized (task) {
1832	>	if (task.status >= 0) {
1833	>	try {                // see ForkJoinTask
1834	>	task.wait();     //  for explanation
1835	>	} catch (InterruptedException ie) {
1836	>	}
1837	>	}
1838	>	else
1839	>	task.notifyAll();
1840		}
1841		}
1842	+	long c;                          // re-activate
1843	+	do {} while (!U.compareAndSwapLong
1844	+	(this, CTL, c = ctl, c + AC_UNIT));
1845		}
1846		}
1847	+	if ((s = task.status) < 0) {
1848	+	joiner.currentJoin = prevJoin;
1849	+	break;
1850	+	}
1851	+	else if ((k++ & (MAX_HELP - 1)) == MAX_HELP >>> 1)
1852	+	Thread.yield();                     // for politeness
1853		}
1854		}
1855	+	return s;
1856		}
1857
1858		/**
1859	<	* Clears out and cancels submissions, ignoring exceptions.
1859	>	* Stripped-down variant of awaitJoin used by timed joins. Tries
1860	>	* to help join only while there is continuous progress. (Caller
1861	>	* will then enter a timed wait.)
1862	>	*
1863	>	* @param joiner the joining worker
1864	>	* @param task the task
1865	>	* @return task status on exit
1866		*/
1867	<	private void cancelSubmissions() {
1868	<	ForkJoinTask<?> task;
1869	<	while ((task = submissionQueue.poll()) != null) {
1870	<	try {
1871	<	task.cancel(false);
1872	<	} catch (Throwable ignore) {
1867	>	final int helpJoinOnce(WorkQueue joiner, ForkJoinTask<?> task) {
1868	>	int s;
1869	>	while ((s = task.status) >= 0 &&
1870	>	(joiner.isEmpty() ?
1871	>	tryHelpStealer(joiner, task) :
1872	>	joiner.tryRemoveAndExec(task)))
1873	>	;
1874	>	return s;
1875	>	}
1876	>
1877	>	/**
1878	>	* Returns a (probably) non-empty steal queue, if one is found
1879	>	* during a random, then cyclic scan, else null.  This method must
1880	>	* be retried by caller if, by the time it tries to use the queue,
1881	>	* it is empty.
1882	>	*/
1883	>	private WorkQueue findNonEmptyStealQueue(WorkQueue w) {
1884	>	// Similar to loop in scan(), but ignoring submissions
1885	>	int r = w.seed; r ^= r << 13; r ^= r >>> 17; w.seed = r ^= r << 5;
1886	>	int step = (r >>> 16) | 1;
1887	>	for (WorkQueue[] ws;;) {
1888	>	int rs = runState, m;
1889	>	if ((ws = workQueues) == null || (m = ws.length - 1) < 1)
1890	>	return null;
1891	>	for (int j = (m + 1) << 2; ; r += step) {
1892	>	WorkQueue q = ws[((r << 1) | 1) & m];
1893	>	if (q != null && !q.isEmpty())
1894	>	return q;
1895	>	else if (--j < 0) {
1896	>	if (runState == rs)
1897	>	return null;
1898	>	break;
1899	>	}
1900		}
1901		}
1902		}
1903
1203	–	// misc support for ForkJoinWorkerThread
1204	–
1904		/**
1905	<	* Returns pool number.
1906	<	*/
1907	<	final int getPoolNumber() {
1908	<	return poolNumber;
1905	>	* Runs tasks until {@code isQuiescent()}. We piggyback on
1906	>	* active count ctl maintenance, but rather than blocking
1907	>	* when tasks cannot be found, we rescan until all others cannot
1908	>	* find tasks either.
1909	>	*/
1910	>	final void helpQuiescePool(WorkQueue w) {
1911	>	for (boolean active = true;;) {
1912	>	ForkJoinTask<?> localTask; // exhaust local queue
1913	>	while ((localTask = w.nextLocalTask()) != null)
1914	>	localTask.doExec();
1915	>	WorkQueue q = findNonEmptyStealQueue(w);
1916	>	if (q != null) {
1917	>	ForkJoinTask<?> t; int b;
1918	>	if (!active) {      // re-establish active count
1919	>	long c;
1920	>	active = true;
1921	>	do {} while (!U.compareAndSwapLong
1922	>	(this, CTL, c = ctl, c + AC_UNIT));
1923	>	}
1924	>	if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null)
1925	>	w.runSubtask(t);
1926	>	}
1927	>	else {
1928	>	long c;
1929	>	if (active) {       // decrement active count without queuing
1930	>	active = false;
1931	>	do {} while (!U.compareAndSwapLong
1932	>	(this, CTL, c = ctl, c -= AC_UNIT));
1933	>	}
1934	>	else
1935	>	c = ctl;        // re-increment on exit
1936	>	if ((int)(c >> AC_SHIFT) + parallelism == 0) {
1937	>	do {} while (!U.compareAndSwapLong
1938	>	(this, CTL, c = ctl, c + AC_UNIT));
1939	>	break;
1940	>	}
1941	>	}
1942	>	}
1943		}
1944
1945		/**
1946	<	* Tries to accumulate steal count from a worker, clearing
1214	<	* the worker's value if successful.
1946	>	* Gets and removes a local or stolen task for the given worker.
1947		*
1948	<	* @return true if worker steal count now zero
1948	>	* @return a task, if available
1949		*/
1950	<	final boolean tryAccumulateStealCount(ForkJoinWorkerThread w) {
1951	<	int sc = w.stealCount;
1952	<	long c = stealCount;
1953	<	// CAS even if zero, for fence effects
1954	<	if (UNSAFE.compareAndSwapLong(this, stealCountOffset, c, c + sc)) {
1955	<	if (sc != 0)
1956	<	w.stealCount = 0;
1957	<	return true;
1950	>	final ForkJoinTask<?> nextTaskFor(WorkQueue w) {
1951	>	for (ForkJoinTask<?> t;;) {
1952	>	WorkQueue q; int b;
1953	>	if ((t = w.nextLocalTask()) != null)
1954	>	return t;
1955	>	if ((q = findNonEmptyStealQueue(w)) == null)
1956	>	return null;
1957	>	if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null)
1958	>	return t;
1959		}
1227	–	return sc == 0;
1960		}
1961
1962		/**
1963		* Returns the approximate (non-atomic) number of idle threads per
1964	<	* active thread.
1964	>	* active thread to offset steal queue size for method
1965	>	* ForkJoinTask.getSurplusQueuedTaskCount().
1966		*/
1967		final int idlePerActive() {
1968	<	int pc = parallelism; // use parallelism, not rc
1969	<	int ac = runState;    // no mask -- artificially boosts during shutdown
1970	<	// Use exact results for small values, saturate past 4
1971	<	return ((pc <= ac) ? 0 :
1972	<	(pc >>> 1 <= ac) ? 1 :
1973	<	(pc >>> 2 <= ac) ? 3 :
1974	<	pc >>> 3);
1968	>	// Approximate at powers of two for small values, saturate past 4
1969	>	int p = parallelism;
1970	>	int a = p + (int)(ctl >> AC_SHIFT);
1971	>	return (a > (p >>>= 1) ? 0 :
1972	>	a > (p >>>= 1) ? 1 :
1973	>	a > (p >>>= 1) ? 2 :
1974	>	a > (p >>>= 1) ? 4 :
1975	>	8);
1976	>	}
1977	>
1978	>	//  Termination
1979	>
1980	>	/**
1981	>	* Possibly initiates and/or completes termination.  The caller
1982	>	* triggering termination runs three passes through workQueues:
1983	>	* (0) Setting termination status, followed by wakeups of queued
1984	>	* workers; (1) cancelling all tasks; (2) interrupting lagging
1985	>	* threads (likely in external tasks, but possibly also blocked in
1986	>	* joins).  Each pass repeats previous steps because of potential
1987	>	* lagging thread creation.
1988	>	*
1989	>	* @param now if true, unconditionally terminate, else only
1990	>	* if no work and no active workers
1991	>	* @param enable if true, enable shutdown when next possible
1992	>	* @return true if now terminating or terminated
1993	>	*/
1994	>	private boolean tryTerminate(boolean now, boolean enable) {
1995	>	Mutex lock = this.lock;
1996	>	for (long c;;) {
1997	>	if (((c = ctl) & STOP_BIT) != 0) {      // already terminating
1998	>	if ((short)(c >>> TC_SHIFT) == -parallelism) {
1999	>	lock.lock();                    // don't need try/finally
2000	>	termination.signalAll();        // signal when 0 workers
2001	>	lock.unlock();
2002	>	}
2003	>	return true;
2004	>	}
2005	>	if (runState >= 0) {                    // not yet enabled
2006	>	if (!enable)
2007	>	return false;
2008	>	lock.lock();
2009	>	runState |= SHUTDOWN;
2010	>	lock.unlock();
2011	>	}
2012	>	if (!now) {                             // check if idle & no tasks
2013	>	if ((int)(c >> AC_SHIFT) != -parallelism ||
2014	>	hasQueuedSubmissions())
2015	>	return false;
2016	>	// Check for unqueued inactive workers. One pass suffices.
2017	>	WorkQueue[] ws = workQueues; WorkQueue w;
2018	>	if (ws != null) {
2019	>	for (int i = 1; i < ws.length; i += 2) {
2020	>	if ((w = ws[i]) != null && w.eventCount >= 0)
2021	>	return false;
2022	>	}
2023	>	}
2024	>	}
2025	>	if (U.compareAndSwapLong(this, CTL, c, c | STOP_BIT)) {
2026	>	for (int pass = 0; pass < 3; ++pass) {
2027	>	WorkQueue[] ws = workQueues;
2028	>	if (ws != null) {
2029	>	WorkQueue w;
2030	>	int n = ws.length;
2031	>	for (int i = 0; i < n; ++i) {
2032	>	if ((w = ws[i]) != null) {
2033	>	w.runState = -1;
2034	>	if (pass > 0) {
2035	>	w.cancelAll();
2036	>	if (pass > 1)
2037	>	w.interruptOwner();
2038	>	}
2039	>	}
2040	>	}
2041	>	// Wake up workers parked on event queue
2042	>	int i, e; long cc; Thread p;
2043	>	while ((e = (int)(cc = ctl) & E_MASK) != 0 &&
2044	>	(i = e & SMASK) < n &&
2045	>	(w = ws[i]) != null) {
2046	>	long nc = ((long)(w.nextWait & E_MASK) |
2047	>	((cc + AC_UNIT) & AC_MASK) |
2048	>	(cc & (TC_MASK|STOP_BIT)));
2049	>	if (w.eventCount == (e | INT_SIGN) &&
2050	>	U.compareAndSwapLong(this, CTL, cc, nc)) {
2051	>	w.eventCount = (e + E_SEQ) & E_MASK;
2052	>	w.runState = -1;
2053	>	if ((p = w.parker) != null)
2054	>	U.unpark(p);
2055	>	}
2056	>	}
2057	>	}
2058	>	}
2059	>	}
2060	>	}
2061		}
2062
2063	<	// Public and protected methods
2063	>	// Exported methods
2064
2065		// Constructors
2066
#	Line 1310 | Line 2129 | public class ForkJoinPool extends Abstra
2129		checkPermission();
2130		if (factory == null)
2131		throw new NullPointerException();
2132	<	if (parallelism <= 0 || parallelism > MAX_WORKERS)
2132	>	if (parallelism <= 0 || parallelism > MAX_CAP)
2133		throw new IllegalArgumentException();
2134		this.parallelism = parallelism;
2135		this.factory = factory;
2136		this.ueh = handler;
2137	<	this.locallyFifo = asyncMode;
2138	<	int arraySize = initialArraySizeFor(parallelism);
2139	<	this.workers = new ForkJoinWorkerThread[arraySize];
2140	<	this.submissionQueue = new LinkedTransferQueue<ForkJoinTask<?>>();
2141	<	this.workerLock = new ReentrantLock();
2142	<	this.termination = new Phaser(1);
2143	<	this.poolNumber = poolNumberGenerator.incrementAndGet();
2144	<	}
2145	<
2146	<	/**
2147	<	* Returns initial power of two size for workers array.
2148	<	* @param pc the initial parallelism level
2149	<	*/
2150	<	private static int initialArraySizeFor(int pc) {
2151	<	// If possible, initially allocate enough space for one spare
2152	<	int size = pc < MAX_WORKERS ? pc + 1 : MAX_WORKERS;
2153	<	// See Hackers Delight, sec 3.2. We know MAX_WORKERS < (1 >>> 16)
2154	<	size |= size >>> 1;
2155	<	size |= size >>> 2;
2156	<	size |= size >>> 4;
1338	<	size |= size >>> 8;
1339	<	return size + 1;
2137	>	this.localMode = asyncMode ? FIFO_QUEUE : LIFO_QUEUE;
2138	>	long np = (long)(-parallelism); // offset ctl counts
2139	>	this.ctl = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK);
2140	>	// Use nearest power 2 for workQueues size. See Hackers Delight sec 3.2.
2141	>	int n = parallelism - 1;
2142	>	n |= n >>> 1; n |= n >>> 2; n |= n >>> 4; n |= n >>> 8; n |= n >>> 16;
2143	>	int size = (n + 1) << 1;        // #slots = 2*#workers
2144	>	this.submitMask = size - 1;     // room for max # of submit queues
2145	>	this.workQueues = new WorkQueue[size];
2146	>	this.termination = (this.lock = new Mutex()).newCondition();
2147	>	this.stealCount = new AtomicLong();
2148	>	this.nextWorkerNumber = new AtomicInteger();
2149	>	int pn = poolNumberGenerator.incrementAndGet();
2150	>	StringBuilder sb = new StringBuilder("ForkJoinPool-");
2151	>	sb.append(Integer.toString(pn));
2152	>	sb.append("-worker-");
2153	>	this.workerNamePrefix = sb.toString();
2154	>	lock.lock();
2155	>	this.runState = 1;              // set init flag
2156	>	lock.unlock();
2157		}
2158
2159		// Execution methods
2160
2161		/**
1345	–	* Submits task and creates, starts, or resumes some workers if necessary
1346	–	*/
1347	–	private <T> void doSubmit(ForkJoinTask<T> task) {
1348	–	submissionQueue.offer(task);
1349	–	int c; // try to increment event count -- CAS failure OK
1350	–	UNSAFE.compareAndSwapInt(this, eventCountOffset, c = eventCount, c+1);
1351	–	helpMaintainParallelism();
1352	–	}
1353	–
1354	–	/**
2162		* Performs the given task, returning its result upon completion.
2163	+	* If the computation encounters an unchecked Exception or Error,
2164	+	* it is rethrown as the outcome of this invocation.  Rethrown
2165	+	* exceptions behave in the same way as regular exceptions, but,
2166	+	* when possible, contain stack traces (as displayed for example
2167	+	* using {@code ex.printStackTrace()}) of both the current thread
2168	+	* as well as the thread actually encountering the exception;
2169	+	* minimally only the latter.
2170		*
2171		* @param task the task
2172		* @return the task's result
#	Line 1363 | Line 2177 | public class ForkJoinPool extends Abstra
2177		public <T> T invoke(ForkJoinTask<T> task) {
2178		if (task == null)
2179		throw new NullPointerException();
2180	<	if (runState >= SHUTDOWN)
2181	<	throw new RejectedExecutionException();
1368	<	Thread t = Thread.currentThread();
1369	<	if ((t instanceof ForkJoinWorkerThread) &&
1370	<	((ForkJoinWorkerThread)t).pool == this)
1371	<	return task.invoke();  // bypass submit if in same pool
1372	<	else {
1373	<	doSubmit(task);
1374	<	return task.join();
1375	<	}
1376	<	}
1377	<
1378	<	/**
1379	<	* Unless terminating, forks task if within an ongoing FJ
1380	<	* computation in the current pool, else submits as external task.
1381	<	*/
1382	<	private <T> void forkOrSubmit(ForkJoinTask<T> task) {
1383	<	if (runState >= SHUTDOWN)
1384	<	throw new RejectedExecutionException();
1385	<	Thread t = Thread.currentThread();
1386	<	if ((t instanceof ForkJoinWorkerThread) &&
1387	<	((ForkJoinWorkerThread)t).pool == this)
1388	<	task.fork();
1389	<	else
1390	<	doSubmit(task);
2180	>	doSubmit(task);
2181	>	return task.join();
2182		}
2183
2184		/**
#	Line 1401 | Line 2192 | public class ForkJoinPool extends Abstra
2192		public void execute(ForkJoinTask<?> task) {
2193		if (task == null)
2194		throw new NullPointerException();
2195	<	forkOrSubmit(task);
2195	>	doSubmit(task);
2196		}
2197
2198		// AbstractExecutorService methods
#	Line 1418 | Line 2209 | public class ForkJoinPool extends Abstra
2209		if (task instanceof ForkJoinTask<?>) // avoid re-wrap
2210		job = (ForkJoinTask<?>) task;
2211		else
2212	<	job = ForkJoinTask.adapt(task, null);
2213	<	forkOrSubmit(job);
2212	>	job = new ForkJoinTask.AdaptedRunnableAction(task);
2213	>	doSubmit(job);
2214		}
2215
2216		/**
#	Line 1434 | Line 2225 | public class ForkJoinPool extends Abstra
2225		public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) {
2226		if (task == null)
2227		throw new NullPointerException();
2228	<	forkOrSubmit(task);
2228	>	doSubmit(task);
2229		return task;
2230		}
2231
#	Line 1444 | Line 2235 | public class ForkJoinPool extends Abstra
2235		*         scheduled for execution
2236		*/
2237		public <T> ForkJoinTask<T> submit(Callable<T> task) {
2238	<	if (task == null)
2239	<	throw new NullPointerException();
1449	<	ForkJoinTask<T> job = ForkJoinTask.adapt(task);
1450	<	forkOrSubmit(job);
2238	>	ForkJoinTask<T> job = new ForkJoinTask.AdaptedCallable<T>(task);
2239	>	doSubmit(job);
2240		return job;
2241		}
2242
#	Line 1457 | Line 2246 | public class ForkJoinPool extends Abstra
2246		*         scheduled for execution
2247		*/
2248		public <T> ForkJoinTask<T> submit(Runnable task, T result) {
2249	<	if (task == null)
2250	<	throw new NullPointerException();
1462	<	ForkJoinTask<T> job = ForkJoinTask.adapt(task, result);
1463	<	forkOrSubmit(job);
2249	>	ForkJoinTask<T> job = new ForkJoinTask.AdaptedRunnable<T>(task, result);
2250	>	doSubmit(job);
2251		return job;
2252		}
2253
#	Line 1476 | Line 2263 | public class ForkJoinPool extends Abstra
2263		if (task instanceof ForkJoinTask<?>) // avoid re-wrap
2264		job = (ForkJoinTask<?>) task;
2265		else
2266	<	job = ForkJoinTask.adapt(task, null);
2267	<	forkOrSubmit(job);
2266	>	job = new ForkJoinTask.AdaptedRunnableAction(task);
2267	>	doSubmit(job);
2268		return job;
2269		}
2270
#	Line 1486 | Line 2273 | public class ForkJoinPool extends Abstra
2273		* @throws RejectedExecutionException {@inheritDoc}
2274		*/
2275		public <T> List<Future<T>> invokeAll(Collection<? extends Callable<T>> tasks) {
2276	<	ArrayList<ForkJoinTask<T>> forkJoinTasks =
2277	<	new ArrayList<ForkJoinTask<T>>(tasks.size());
2278	<	for (Callable<T> task : tasks)
2279	<	forkJoinTasks.add(ForkJoinTask.adapt(task));
2280	<	invoke(new InvokeAll<T>(forkJoinTasks));
2281	<
2276	>	// In previous versions of this class, this method constructed
2277	>	// a task to run ForkJoinTask.invokeAll, but now external
2278	>	// invocation of multiple tasks is at least as efficient.
2279	>	List<ForkJoinTask<T>> fs = new ArrayList<ForkJoinTask<T>>(tasks.size());
2280	>	// Workaround needed because method wasn't declared with
2281	>	// wildcards in return type but should have been.
2282		@SuppressWarnings({"unchecked", "rawtypes"})
2283	<	List<Future<T>> futures = (List<Future<T>>) (List) forkJoinTasks;
1497	<	return futures;
1498	<	}
2283	>	List<Future<T>> futures = (List<Future<T>>) (List) fs;
2284
2285	<	static final class InvokeAll<T> extends RecursiveAction {
2286	<	final ArrayList<ForkJoinTask<T>> tasks;
2287	<	InvokeAll(ArrayList<ForkJoinTask<T>> tasks) { this.tasks = tasks; }
2288	<	public void compute() {
2289	<	try { invokeAll(tasks); }
2290	<	catch (Exception ignore) {}
2285	>	boolean done = false;
2286	>	try {
2287	>	for (Callable<T> t : tasks) {
2288	>	ForkJoinTask<T> f = new ForkJoinTask.AdaptedCallable<T>(t);
2289	>	doSubmit(f);
2290	>	fs.add(f);
2291	>	}
2292	>	for (ForkJoinTask<T> f : fs)
2293	>	f.quietlyJoin();
2294	>	done = true;
2295	>	return futures;
2296	>	} finally {
2297	>	if (!done)
2298	>	for (ForkJoinTask<T> f : fs)
2299	>	f.cancel(false);
2300		}
1507	–	private static final long serialVersionUID = -7914297376763021607L;
2301		}
2302
2303		/**
#	Line 1544 | Line 2337 | public class ForkJoinPool extends Abstra
2337		* @return the number of worker threads
2338		*/
2339		public int getPoolSize() {
2340	<	return workerCounts >>> TOTAL_COUNT_SHIFT;
2340	>	return parallelism + (short)(ctl >>> TC_SHIFT);
2341		}
2342
2343		/**
#	Line 1554 | Line 2347 | public class ForkJoinPool extends Abstra
2347		* @return {@code true} if this pool uses async mode
2348		*/
2349		public boolean getAsyncMode() {
2350	<	return locallyFifo;
2350	>	return localMode != 0;
2351		}
2352
2353		/**
#	Line 1566 | Line 2359 | public class ForkJoinPool extends Abstra
2359		* @return the number of worker threads
2360		*/
2361		public int getRunningThreadCount() {
2362	<	return workerCounts & RUNNING_COUNT_MASK;
2362	>	int rc = 0;
2363	>	WorkQueue[] ws; WorkQueue w;
2364	>	if ((ws = workQueues) != null) {
2365	>	for (int i = 1; i < ws.length; i += 2) {
2366	>	if ((w = ws[i]) != null && w.isApparentlyUnblocked())
2367	>	++rc;
2368	>	}
2369	>	}
2370	>	return rc;
2371		}
2372
2373		/**
#	Line 1577 | Line 2378 | public class ForkJoinPool extends Abstra
2378		* @return the number of active threads
2379		*/
2380		public int getActiveThreadCount() {
2381	<	return runState & ACTIVE_COUNT_MASK;
2381	>	int r = parallelism + (int)(ctl >> AC_SHIFT);
2382	>	return (r <= 0) ? 0 : r; // suppress momentarily negative values
2383		}
2384
2385		/**
#	Line 1592 | Line 2394 | public class ForkJoinPool extends Abstra
2394		* @return {@code true} if all threads are currently idle
2395		*/
2396		public boolean isQuiescent() {
2397	<	return (runState & ACTIVE_COUNT_MASK) == 0;
2397	>	return (int)(ctl >> AC_SHIFT) + parallelism == 0;
2398		}
2399
2400		/**
#	Line 1607 | Line 2409 | public class ForkJoinPool extends Abstra
2409		* @return the number of steals
2410		*/
2411		public long getStealCount() {
2412	<	return stealCount;
2412	>	long count = stealCount.get();
2413	>	WorkQueue[] ws; WorkQueue w;
2414	>	if ((ws = workQueues) != null) {
2415	>	for (int i = 1; i < ws.length; i += 2) {
2416	>	if ((w = ws[i]) != null)
2417	>	count += w.totalSteals;
2418	>	}
2419	>	}
2420	>	return count;
2421		}
2422
2423		/**
#	Line 1622 | Line 2432 | public class ForkJoinPool extends Abstra
2432		*/
2433		public long getQueuedTaskCount() {
2434		long count = 0;
2435	<	for (ForkJoinWorkerThread w : workers)
2436	<	if (w != null)
2437	<	count += w.getQueueSize();
2435	>	WorkQueue[] ws; WorkQueue w;
2436	>	if ((ws = workQueues) != null) {
2437	>	for (int i = 1; i < ws.length; i += 2) {
2438	>	if ((w = ws[i]) != null)
2439	>	count += w.queueSize();
2440	>	}
2441	>	}
2442		return count;
2443		}
2444
2445		/**
2446		* Returns an estimate of the number of tasks submitted to this
2447	<	* pool that have not yet begun executing.  This method takes time
2448	<	* proportional to the number of submissions.
2447	>	* pool that have not yet begun executing.  This method may take
2448	>	* time proportional to the number of submissions.
2449		*
2450		* @return the number of queued submissions
2451		*/
2452		public int getQueuedSubmissionCount() {
2453	<	return submissionQueue.size();
2453	>	int count = 0;
2454	>	WorkQueue[] ws; WorkQueue w;
2455	>	if ((ws = workQueues) != null) {
2456	>	for (int i = 0; i < ws.length; i += 2) {
2457	>	if ((w = ws[i]) != null)
2458	>	count += w.queueSize();
2459	>	}
2460	>	}
2461	>	return count;
2462		}
2463
2464		/**
#	Line 1646 | Line 2468 | public class ForkJoinPool extends Abstra
2468		* @return {@code true} if there are any queued submissions
2469		*/
2470		public boolean hasQueuedSubmissions() {
2471	<	return !submissionQueue.isEmpty();
2471	>	WorkQueue[] ws; WorkQueue w;
2472	>	if ((ws = workQueues) != null) {
2473	>	for (int i = 0; i < ws.length; i += 2) {
2474	>	if ((w = ws[i]) != null && !w.isEmpty())
2475	>	return true;
2476	>	}
2477	>	}
2478	>	return false;
2479		}
2480
2481		/**
#	Line 1657 | Line 2486 | public class ForkJoinPool extends Abstra
2486		* @return the next submission, or {@code null} if none
2487		*/
2488		protected ForkJoinTask<?> pollSubmission() {
2489	<	return submissionQueue.poll();
2489	>	WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2490	>	if ((ws = workQueues) != null) {
2491	>	for (int i = 0; i < ws.length; i += 2) {
2492	>	if ((w = ws[i]) != null && (t = w.poll()) != null)
2493	>	return t;
2494	>	}
2495	>	}
2496	>	return null;
2497		}
2498
2499		/**
#	Line 1678 | Line 2514 | public class ForkJoinPool extends Abstra
2514		* @return the number of elements transferred
2515		*/
2516		protected int drainTasksTo(Collection<? super ForkJoinTask<?>> c) {
2517	<	int count = submissionQueue.drainTo(c);
2518	<	for (ForkJoinWorkerThread w : workers)
2519	<	if (w != null)
2520	<	count += w.drainTasksTo(c);
2517	>	int count = 0;
2518	>	WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2519	>	if ((ws = workQueues) != null) {
2520	>	for (int i = 0; i < ws.length; ++i) {
2521	>	if ((w = ws[i]) != null) {
2522	>	while ((t = w.poll()) != null) {
2523	>	c.add(t);
2524	>	++count;
2525	>	}
2526	>	}
2527	>	}
2528	>	}
2529		return count;
2530		}
2531
#	Line 1693 | Line 2537 | public class ForkJoinPool extends Abstra
2537		* @return a string identifying this pool, as well as its state
2538		*/
2539		public String toString() {
2540	<	long st = getStealCount();
2541	<	long qt = getQueuedTaskCount();
2542	<	long qs = getQueuedSubmissionCount();
2543	<	int wc = workerCounts;
2544	<	int tc = wc >>> TOTAL_COUNT_SHIFT;
2545	<	int rc = wc & RUNNING_COUNT_MASK;
2540	>	// Use a single pass through workQueues to collect counts
2541	>	long qt = 0L, qs = 0L; int rc = 0;
2542	>	long st = stealCount.get();
2543	>	long c = ctl;
2544	>	WorkQueue[] ws; WorkQueue w;
2545	>	if ((ws = workQueues) != null) {
2546	>	for (int i = 0; i < ws.length; ++i) {
2547	>	if ((w = ws[i]) != null) {
2548	>	int size = w.queueSize();
2549	>	if ((i & 1) == 0)
2550	>	qs += size;
2551	>	else {
2552	>	qt += size;
2553	>	st += w.totalSteals;
2554	>	if (w.isApparentlyUnblocked())
2555	>	++rc;
2556	>	}
2557	>	}
2558	>	}
2559	>	}
2560		int pc = parallelism;
2561	<	int rs = runState;
2562	<	int ac = rs & ACTIVE_COUNT_MASK;
2561	>	int tc = pc + (short)(c >>> TC_SHIFT);
2562	>	int ac = pc + (int)(c >> AC_SHIFT);
2563	>	if (ac < 0) // ignore transient negative
2564	>	ac = 0;
2565	>	String level;
2566	>	if ((c & STOP_BIT) != 0)
2567	>	level = (tc == 0) ? "Terminated" : "Terminating";
2568	>	else
2569	>	level = runState < 0 ? "Shutting down" : "Running";
2570		return super.toString() +
2571	<	"[" + runLevelToString(rs) +
2571	>	"[" + level +
2572		", parallelism = " + pc +
2573		", size = " + tc +
2574		", active = " + ac +
#	Line 1714 | Line 2579 | public class ForkJoinPool extends Abstra
2579		"]";
2580		}
2581
1717	–	private static String runLevelToString(int s) {
1718	–	return ((s & TERMINATED) != 0 ? "Terminated" :
1719	–	((s & TERMINATING) != 0 ? "Terminating" :
1720	–	((s & SHUTDOWN) != 0 ? "Shutting down" :
1721	–	"Running")));
1722	–	}
1723	–
2582		/**
2583		* Initiates an orderly shutdown in which previously submitted
2584		* tasks are executed, but no new tasks will be accepted.
#	Line 1735 | Line 2593 | public class ForkJoinPool extends Abstra
2593		*/
2594		public void shutdown() {
2595		checkPermission();
2596	<	advanceRunLevel(SHUTDOWN);
1739	<	tryTerminate(false);
2596	>	tryTerminate(false, true);
2597		}
2598
2599		/**
#	Line 1757 | Line 2614 | public class ForkJoinPool extends Abstra
2614		*/
2615		public List<Runnable> shutdownNow() {
2616		checkPermission();
2617	<	tryTerminate(true);
2617	>	tryTerminate(true, true);
2618		return Collections.emptyList();
2619		}
2620
#	Line 1767 | Line 2624 | public class ForkJoinPool extends Abstra
2624		* @return {@code true} if all tasks have completed following shut down
2625		*/
2626		public boolean isTerminated() {
2627	<	return runState >= TERMINATED;
2627	>	long c = ctl;
2628	>	return ((c & STOP_BIT) != 0L &&
2629	>	(short)(c >>> TC_SHIFT) == -parallelism);
2630		}
2631
2632		/**
#	Line 1775 | Line 2634 | public class ForkJoinPool extends Abstra
2634		* commenced but not yet completed.  This method may be useful for
2635		* debugging. A return of {@code true} reported a sufficient
2636		* period after shutdown may indicate that submitted tasks have
2637	<	* ignored or suppressed interruption, causing this executor not
2638	<	* to properly terminate.
2637	>	* ignored or suppressed interruption, or are waiting for IO,
2638	>	* causing this executor not to properly terminate. (See the
2639	>	* advisory notes for class {@link ForkJoinTask} stating that
2640	>	* tasks should not normally entail blocking operations.  But if
2641	>	* they do, they must abort them on interrupt.)
2642		*
2643		* @return {@code true} if terminating but not yet terminated
2644		*/
2645		public boolean isTerminating() {
2646	<	return (runState & (TERMINATING|TERMINATED)) == TERMINATING;
2647	<	}
2648	<
1787	<	/**
1788	<	* Returns true if terminating or terminated. Used by ForkJoinWorkerThread.
1789	<	*/
1790	<	final boolean isAtLeastTerminating() {
1791	<	return runState >= TERMINATING;
2646	>	long c = ctl;
2647	>	return ((c & STOP_BIT) != 0L &&
2648	>	(short)(c >>> TC_SHIFT) != -parallelism);
2649		}
2650
2651		/**
#	Line 1797 | Line 2654 | public class ForkJoinPool extends Abstra
2654		* @return {@code true} if this pool has been shut down
2655		*/
2656		public boolean isShutdown() {
2657	<	return runState >= SHUTDOWN;
2657	>	return runState < 0;
2658		}
2659
2660		/**
#	Line 1813 | Line 2670 | public class ForkJoinPool extends Abstra
2670		*/
2671		public boolean awaitTermination(long timeout, TimeUnit unit)
2672		throws InterruptedException {
2673	+	long nanos = unit.toNanos(timeout);
2674	+	final Mutex lock = this.lock;
2675	+	lock.lock();
2676		try {
2677	<	termination.awaitAdvanceInterruptibly(0, timeout, unit);
2678	<	} catch (TimeoutException ex) {
2679	<	return false;
2677	>	for (;;) {
2678	>	if (isTerminated())
2679	>	return true;
2680	>	if (nanos <= 0)
2681	>	return false;
2682	>	nanos = termination.awaitNanos(nanos);
2683	>	}
2684	>	} finally {
2685	>	lock.unlock();
2686		}
1821	–	return true;
2687		}
2688
2689		/**
#	Line 1829 | Line 2694 | public class ForkJoinPool extends Abstra
2694		* {@code isReleasable} must return {@code true} if blocking is
2695		* not necessary. Method {@code block} blocks the current thread
2696		* if necessary (perhaps internally invoking {@code isReleasable}
2697	<	* before actually blocking). The unusual methods in this API
2698	<	* accommodate synchronizers that may, but don't usually, block
2699	<	* for long periods. Similarly, they allow more efficient internal
2700	<	* handling of cases in which additional workers may be, but
2701	<	* usually are not, needed to ensure sufficient parallelism.
2702	<	* Toward this end, implementations of method {@code isReleasable}
2703	<	* must be amenable to repeated invocation.
2697	>	* before actually blocking). These actions are performed by any
2698	>	* thread invoking {@link ForkJoinPool#managedBlock}.  The
2699	>	* unusual methods in this API accommodate synchronizers that may,
2700	>	* but don't usually, block for long periods. Similarly, they
2701	>	* allow more efficient internal handling of cases in which
2702	>	* additional workers may be, but usually are not, needed to
2703	>	* ensure sufficient parallelism.  Toward this end,
2704	>	* implementations of method {@code isReleasable} must be amenable
2705	>	* to repeated invocation.
2706		*
2707		* <p>For example, here is a ManagedBlocker based on a
2708		* ReentrantLock:
#	Line 1915 | Line 2782 | public class ForkJoinPool extends Abstra
2782		public static void managedBlock(ManagedBlocker blocker)
2783		throws InterruptedException {
2784		Thread t = Thread.currentThread();
2785	<	if (t instanceof ForkJoinWorkerThread) {
2786	<	ForkJoinWorkerThread w = (ForkJoinWorkerThread) t;
2787	<	w.pool.awaitBlocker(blocker);
2788	<	}
2789	<	else {
2790	<	do {} while (!blocker.isReleasable() && !blocker.block());
2785	>	ForkJoinPool p = ((t instanceof ForkJoinWorkerThread) ?
2786	>	((ForkJoinWorkerThread)t).pool : null);
2787	>	while (!blocker.isReleasable()) {
2788	>	if (p == null || p.tryCompensate(null, blocker)) {
2789	>	try {
2790	>	do {} while (!blocker.isReleasable() && !blocker.block());
2791	>	} finally {
2792	>	if (p != null)
2793	>	p.incrementActiveCount();
2794	>	}
2795	>	break;
2796	>	}
2797		}
2798		}
2799
#	Line 1929 | Line 2802 | public class ForkJoinPool extends Abstra
2802		// implement RunnableFuture.
2803
2804		protected <T> RunnableFuture<T> newTaskFor(Runnable runnable, T value) {
2805	<	return (RunnableFuture<T>) ForkJoinTask.adapt(runnable, value);
2805	>	return new ForkJoinTask.AdaptedRunnable<T>(runnable, value);
2806		}
2807
2808		protected <T> RunnableFuture<T> newTaskFor(Callable<T> callable) {
2809	<	return (RunnableFuture<T>) ForkJoinTask.adapt(callable);
2809	>	return new ForkJoinTask.AdaptedCallable<T>(callable);
2810		}
2811
2812		// Unsafe mechanics
2813	<
2814	<	private static final sun.misc.Unsafe UNSAFE = getUnsafe();
2815	<	private static final long workerCountsOffset =
2816	<	objectFieldOffset("workerCounts", ForkJoinPool.class);
2817	<	private static final long runStateOffset =
2818	<	objectFieldOffset("runState", ForkJoinPool.class);
2819	<	private static final long eventCountOffset =
2820	<	objectFieldOffset("eventCount", ForkJoinPool.class);
2821	<	private static final long eventWaitersOffset =
2822	<	objectFieldOffset("eventWaiters", ForkJoinPool.class);
2823	<	private static final long stealCountOffset =
2824	<	objectFieldOffset("stealCount", ForkJoinPool.class);
2825	<	private static final long spareWaitersOffset =
2826	<	objectFieldOffset("spareWaiters", ForkJoinPool.class);
1954	<
1955	<	private static long objectFieldOffset(String field, Class<?> klazz) {
2813	>	private static final sun.misc.Unsafe U;
2814	>	private static final long CTL;
2815	>	private static final long PARKBLOCKER;
2816	>	private static final int ABASE;
2817	>	private static final int ASHIFT;
2818	>
2819	>	static {
2820	>	poolNumberGenerator = new AtomicInteger();
2821	>	nextSubmitterSeed = new AtomicInteger(0x55555555);
2822	>	modifyThreadPermission = new RuntimePermission("modifyThread");
2823	>	defaultForkJoinWorkerThreadFactory =
2824	>	new DefaultForkJoinWorkerThreadFactory();
2825	>	submitters = new ThreadSubmitter();
2826	>	int s;
2827		try {
2828	<	return UNSAFE.objectFieldOffset(klazz.getDeclaredField(field));
2829	<	} catch (NoSuchFieldException e) {
2830	<	// Convert Exception to corresponding Error
2831	<	NoSuchFieldError error = new NoSuchFieldError(field);
2832	<	error.initCause(e);
2833	<	throw error;
2834	<	}
2828	>	U = getUnsafe();
2829	>	Class<?> k = ForkJoinPool.class;
2830	>	Class<?> ak = ForkJoinTask[].class;
2831	>	CTL = U.objectFieldOffset
2832	>	(k.getDeclaredField("ctl"));
2833	>	Class<?> tk = Thread.class;
2834	>	PARKBLOCKER = U.objectFieldOffset
2835	>	(tk.getDeclaredField("parkBlocker"));
2836	>	ABASE = U.arrayBaseOffset(ak);
2837	>	s = U.arrayIndexScale(ak);
2838	>	} catch (Exception e) {
2839	>	throw new Error(e);
2840	>	}
2841	>	if ((s & (s-1)) != 0)
2842	>	throw new Error("data type scale not a power of two");
2843	>	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
2844		}
2845
2846		/**
#	Line 1990 | Line 2870 | public class ForkJoinPool extends Abstra
2870		}
2871		}
2872		}
2873	+
2874		}

Diff Legend

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