ViewVC Help

View File | Revision Log | Show Annotations | Download File | Root Listing

root/jsr166/jsr166/src/jsr166e/ForkJoinPool.java

Comparing jsr166/src/jsr166e/ForkJoinPool.java (file contents):
Revision 1.5 by jsr166, Sun Oct 21 04:14:31 2012 UTC vs.
Revision 1.65 by jsr166, Sat Sep 12 19:16:45 2015 UTC

#	Line 5 | Line 5
5		*/
6
7		package jsr166e;
8	+
9	+	import java.lang.Thread.UncaughtExceptionHandler;
10		import java.util.ArrayList;
11		import java.util.Arrays;
12		import java.util.Collection;
13		import java.util.Collections;
14		import java.util.List;
13	–	import java.util.Random;
15		import java.util.concurrent.AbstractExecutorService;
16		import java.util.concurrent.Callable;
17		import java.util.concurrent.ExecutorService;
#	Line 18 | Line 19 | import java.util.concurrent.Future;
19		import java.util.concurrent.RejectedExecutionException;
20		import java.util.concurrent.RunnableFuture;
21		import java.util.concurrent.TimeUnit;
21	–	import java.util.concurrent.atomic.AtomicInteger;
22	–	import java.util.concurrent.atomic.AtomicLong;
23	–	import java.util.concurrent.locks.AbstractQueuedSynchronizer;
24	–	import java.util.concurrent.locks.Condition;
22
23		/**
24		* An {@link ExecutorService} for running {@link ForkJoinTask}s.
#	Line 41 | Line 38 | import java.util.concurrent.locks.Condit
38		* ForkJoinPool}s may also be appropriate for use with event-style
39		* tasks that are never joined.
40		*
41	<	* <p>A {@code ForkJoinPool} is constructed with a given target
42	<	* parallelism level; by default, equal to the number of available
43	<	* processors. The pool attempts to maintain enough active (or
44	<	* available) threads by dynamically adding, suspending, or resuming
45	<	* internal worker threads, even if some tasks are stalled waiting to
46	<	* join others. However, no such adjustments are guaranteed in the
47	<	* face of blocked IO or other unmanaged synchronization. The nested
48	<	* {@link ManagedBlocker} interface enables extension of the kinds of
41	>	* <p>A static {@link #commonPool()} is available and appropriate for
42	>	* most applications. The common pool is used by any ForkJoinTask that
43	>	* is not explicitly submitted to a specified pool. Using the common
44	>	* pool normally reduces resource usage (its threads are slowly
45	>	* reclaimed during periods of non-use, and reinstated upon subsequent
46	>	* use).
47	>	*
48	>	* <p>For applications that require separate or custom pools, a {@code
49	>	* ForkJoinPool} may be constructed with a given target parallelism
50	>	* level; by default, equal to the number of available processors. The
51	>	* pool attempts to maintain enough active (or available) threads by
52	>	* dynamically adding, suspending, or resuming internal worker
53	>	* threads, even if some tasks are stalled waiting to join others.
54	>	* However, no such adjustments are guaranteed in the face of blocked
55	>	* I/O or other unmanaged synchronization. The nested {@link
56	>	* ManagedBlocker} interface enables extension of the kinds of
57		* synchronization accommodated.
58		*
59		* <p>In addition to execution and lifecycle control methods, this
#	Line 58 | Line 63 | import java.util.concurrent.locks.Condit
63		* {@link #toString} returns indications of pool state in a
64		* convenient form for informal monitoring.
65		*
66	<	* <p> As is the case with other ExecutorServices, there are three
66	>	* <p>As is the case with other ExecutorServices, there are three
67		* main task execution methods summarized in the following table.
68		* These are designed to be used primarily by clients not already
69		* engaged in fork/join computations in the current pool.  The main
#	Line 71 | Line 76 | import java.util.concurrent.locks.Condit
76		* there is little difference among choice of methods.
77		*
78		* <table BORDER CELLPADDING=3 CELLSPACING=1>
79	+	* <caption>Summary of task execution methods</caption>
80		*  <tr>
81		*    <td></td>
82		*    <td ALIGN=CENTER> <b>Call from non-fork/join clients</b></td>
83		*    <td ALIGN=CENTER> <b>Call from within fork/join computations</b></td>
84		*  </tr>
85		*  <tr>
86	<	*    <td> <b>Arrange async execution</td>
86	>	*    <td> <b>Arrange async execution</b></td>
87		*    <td> {@link #execute(ForkJoinTask)}</td>
88		*    <td> {@link ForkJoinTask#fork}</td>
89		*  </tr>
90		*  <tr>
91	<	*    <td> <b>Await and obtain result</td>
91	>	*    <td> <b>Await and obtain result</b></td>
92		*    <td> {@link #invoke(ForkJoinTask)}</td>
93		*    <td> {@link ForkJoinTask#invoke}</td>
94		*  </tr>
95		*  <tr>
96	<	*    <td> <b>Arrange exec and obtain Future</td>
96	>	*    <td> <b>Arrange exec and obtain Future</b></td>
97		*    <td> {@link #submit(ForkJoinTask)}</td>
98		*    <td> {@link ForkJoinTask#fork} (ForkJoinTasks <em>are</em> Futures)</td>
99		*  </tr>
100		* </table>
101		*
102	<	* <p><b>Sample Usage.</b> Normally a single {@code ForkJoinPool} is
103	<	* used for all parallel task execution in a program or subsystem.
104	<	* Otherwise, use would not usually outweigh the construction and
105	<	* bookkeeping overhead of creating a large set of threads. For
106	<	* example, a common pool could be used for the {@code SortTasks}
107	<	* illustrated in {@link RecursiveAction}. Because {@code
108	<	* ForkJoinPool} uses threads in {@linkplain java.lang.Thread#isDaemon
109	<	* daemon} mode, there is typically no need to explicitly {@link
110	<	* #shutdown} such a pool upon program exit.
111	<	*
112	<	*  <pre> {@code
113	<	* static final ForkJoinPool mainPool = new ForkJoinPool();
114	<	* ...
115	<	* public void sort(long[] array) {
116	<	*   mainPool.invoke(new SortTask(array, 0, array.length));
117	<	* }}</pre>
102	>	* <p>The common pool is by default constructed with default
103	>	* parameters, but these may be controlled by setting three
104	>	* {@linkplain System#getProperty system properties}:
105	>	* <ul>
106	>	* <li>{@code java.util.concurrent.ForkJoinPool.common.parallelism}
107	>	* - the parallelism level, a non-negative integer
108	>	* <li>{@code java.util.concurrent.ForkJoinPool.common.threadFactory}
109	>	* - the class name of a {@link ForkJoinWorkerThreadFactory}
110	>	* <li>{@code java.util.concurrent.ForkJoinPool.common.exceptionHandler}
111	>	* - the class name of a {@link UncaughtExceptionHandler}
112	>	* </ul>
113	>	* The system class loader is used to load these classes.
114	>	* Upon any error in establishing these settings, default parameters
115	>	* are used. It is possible to disable or limit the use of threads in
116	>	* the common pool by setting the parallelism property to zero, and/or
117	>	* using a factory that may return {@code null}.
118		*
119		* <p><b>Implementation notes</b>: This implementation restricts the
120		* maximum number of running threads to 32767. Attempts to create
#	Line 154 | Line 160 | public class ForkJoinPool extends Abstra
160		* (http://research.sun.com/scalable/pubs/index.html) and
161		* "Idempotent work stealing" by Michael, Saraswat, and Vechev,
162		* PPoPP 2009 (http://portal.acm.org/citation.cfm?id=1504186).
163	<	* The main differences ultimately stem from GC requirements that
164	<	* we null out taken slots as soon as we can, to maintain as small
165	<	* a footprint as possible even in programs generating huge
166	<	* numbers of tasks. To accomplish this, we shift the CAS
167	<	* arbitrating pop vs poll (steal) from being on the indices
168	<	* ("base" and "top") to the slots themselves.  So, both a
169	<	* successful pop and poll mainly entail a CAS of a slot from
170	<	* non-null to null.  Because we rely on CASes of references, we
171	<	* do not need tag bits on base or top.  They are simple ints as
172	<	* used in any circular array-based queue (see for example
173	<	* ArrayDeque).  Updates to the indices must still be ordered in a
174	<	* way that guarantees that top == base means the queue is empty,
175	<	* but otherwise may err on the side of possibly making the queue
176	<	* appear nonempty when a push, pop, or poll have not fully
177	<	* committed. Note that this means that the poll operation,
178	<	* considered individually, is not wait-free. One thief cannot
179	<	* successfully continue until another in-progress one (or, if
180	<	* previously empty, a push) completes.  However, in the
181	<	* aggregate, we ensure at least probabilistic non-blockingness.
182	<	* If an attempted steal fails, a thief always chooses a different
183	<	* random victim target to try next. So, in order for one thief to
184	<	* progress, it suffices for any in-progress poll or new push on
185	<	* any empty queue to complete. (This is why we normally use
186	<	* method pollAt and its variants that try once at the apparent
187	<	* base index, else consider alternative actions, rather than
188	<	* method poll.)
163	>	* See also "Correct and Efficient Work-Stealing for Weak Memory
164	>	* Models" by Le, Pop, Cohen, and Nardelli, PPoPP 2013
165	>	* (http://www.di.ens.fr/~zappa/readings/ppopp13.pdf) for an
166	>	* analysis of memory ordering (atomic, volatile etc) issues.  The
167	>	* main differences ultimately stem from GC requirements that we
168	>	* null out taken slots as soon as we can, to maintain as small a
169	>	* footprint as possible even in programs generating huge numbers
170	>	* of tasks. To accomplish this, we shift the CAS arbitrating pop
171	>	* vs poll (steal) from being on the indices ("base" and "top") to
172	>	* the slots themselves.  So, both a successful pop and poll
173	>	* mainly entail a CAS of a slot from non-null to null.  Because
174	>	* we rely on CASes of references, we do not need tag bits on base
175	>	* or top.  They are simple ints as used in any circular
176	>	* array-based queue (see for example ArrayDeque).  Updates to the
177	>	* indices must still be ordered in a way that guarantees that top
178	>	* == base means the queue is empty, but otherwise may err on the
179	>	* side of possibly making the queue appear nonempty when a push,
180	>	* pop, or poll have not fully committed. Note that this means
181	>	* that the poll operation, considered individually, is not
182	>	* wait-free. One thief cannot successfully continue until another
183	>	* in-progress one (or, if previously empty, a push) completes.
184	>	* However, in the aggregate, we ensure at least probabilistic
185	>	* non-blockingness.  If an attempted steal fails, a thief always
186	>	* chooses a different random victim target to try next. So, in
187	>	* order for one thief to progress, it suffices for any
188	>	* in-progress poll or new push on any empty queue to
189	>	* complete. (This is why we normally use method pollAt and its
190	>	* variants that try once at the apparent base index, else
191	>	* consider alternative actions, rather than method poll.)
192		*
193		* This approach also enables support of a user mode in which local
194		* task processing is in FIFO, not LIFO order, simply by using
#	Line 196 | Line 205 | public class ForkJoinPool extends Abstra
205		* WorkQueues are also used in a similar way for tasks submitted
206		* to the pool. We cannot mix these tasks in the same queues used
207		* for work-stealing (this would contaminate lifo/fifo
208	<	* processing). Instead, we loosely associate submission queues
208	>	* processing). Instead, we randomly associate submission queues
209		* with submitting threads, using a form of hashing.  The
210	<	* ThreadLocal Submitter class contains a value initially used as
211	<	* a hash code for choosing existing queues, but may be randomly
212	<	* repositioned upon contention with other submitters.  In
213	<	* essence, submitters act like workers except that they never
214	<	* take tasks, and they are multiplexed on to a finite number of
215	<	* shared work queues. However, classes are set up so that future
216	<	* extensions could allow submitters to optionally help perform
217	<	* tasks as well. Insertion of tasks in shared mode requires a
218	<	* lock (mainly to protect in the case of resizing) but we use
219	<	* only a simple spinlock (using bits in field runState), because
220	<	* submitters encountering a busy queue move on to try or create
221	<	* other queues -- they block only when creating and registering
222	<	* new queues.
210	>	* Submitter probe value serves as a hash code for
211	>	* choosing existing queues, and may be randomly repositioned upon
212	>	* contention with other submitters.  In essence, submitters act
213	>	* like workers except that they are restricted to executing local
214	>	* tasks that they submitted (or in the case of CountedCompleters,
215	>	* others with the same root task).  However, because most
216	>	* shared/external queue operations are more expensive than
217	>	* internal, and because, at steady state, external submitters
218	>	* will compete for CPU with workers, ForkJoinTask.join and
219	>	* related methods disable them from repeatedly helping to process
220	>	* tasks if all workers are active.  Insertion of tasks in shared
221	>	* mode requires a lock (mainly to protect in the case of
222	>	* resizing) but we use only a simple spinlock (using bits in
223	>	* field qlock), because submitters encountering a busy queue move
224	>	* on to try or create other queues -- they block only when
225	>	* creating and registering new queues.
226		*
227		* Management
228		* ==========
#	Line 232 | Line 244 | public class ForkJoinPool extends Abstra
244		* and their negations (used for thresholding) to fit into 16bit
245		* fields.
246		*
247	<	* Field "runState" contains 32 bits needed to register and
248	<	* deregister WorkQueues, as well as to enable shutdown. It is
249	<	* only modified under a lock (normally briefly held, but
250	<	* occasionally protecting allocations and resizings) but even
251	<	* when locked remains available to check consistency.
247	>	* Field "plock" is a form of sequence lock with a saturating
248	>	* shutdown bit (similarly for per-queue "qlocks"), mainly
249	>	* protecting updates to the workQueues array, as well as to
250	>	* enable shutdown.  When used as a lock, it is normally only very
251	>	* briefly held, so is nearly always available after at most a
252	>	* brief spin, but we use a monitor-based backup strategy to
253	>	* block when needed.
254		*
255		* Recording WorkQueues.  WorkQueues are recorded in the
256	<	* "workQueues" array that is created upon pool construction and
257	<	* expanded if necessary.  Updates to the array while recording
258	<	* new workers and unrecording terminated ones are protected from
259	<	* each other by a lock but the array is otherwise concurrently
260	<	* readable, and accessed directly.  To simplify index-based
261	<	* operations, the array size is always a power of two, and all
262	<	* readers must tolerate null slots. Shared (submission) queues
263	<	* are at even indices, worker queues at odd indices. Grouping
264	<	* them together in this way simplifies and speeds up task
265	<	* scanning.
256	>	* "workQueues" array that is created upon first use and expanded
257	>	* if necessary.  Updates to the array while recording new workers
258	>	* and unrecording terminated ones are protected from each other
259	>	* by a lock but the array is otherwise concurrently readable, and
260	>	* accessed directly.  To simplify index-based operations, the
261	>	* array size is always a power of two, and all readers must
262	>	* tolerate null slots. Worker queues are at odd indices. Shared
263	>	* (submission) queues are at even indices, up to a maximum of 64
264	>	* slots, to limit growth even if array needs to expand to add
265	>	* more workers. Grouping them together in this way simplifies and
266	>	* speeds up task scanning.
267		*
268		* All worker thread creation is on-demand, triggered by task
269		* submissions, replacement of terminated workers, and/or
#	Line 293 | Line 308 | public class ForkJoinPool extends Abstra
308		* has not yet entered the wait queue. We solve this by requiring
309		* a full sweep of all workers (via repeated calls to method
310		* scan()) both before and after a newly waiting worker is added
311	<	* to the wait queue. During a rescan, the worker might release
312	<	* some other queued worker rather than itself, which has the same
313	<	* net effect. Because enqueued workers may actually be rescanning
314	<	* rather than waiting, we set and clear the "parker" field of
315	<	* WorkQueues to reduce unnecessary calls to unpark.  (This
316	<	* requires a secondary recheck to avoid missed signals.)  Note
317	<	* the unusual conventions about Thread.interrupts surrounding
318	<	* parking and other blocking: Because interrupts are used solely
319	<	* to alert threads to check termination, which is checked anyway
320	<	* upon blocking, we clear status (using Thread.interrupted)
321	<	* 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.
311	>	* to the wait queue.  Because enqueued workers may actually be
312	>	* rescanning rather than waiting, we set and clear the "parker"
313	>	* field of WorkQueues to reduce unnecessary calls to unpark.
314	>	* (This requires a secondary recheck to avoid missed signals.)
315	>	* Note the unusual conventions about Thread.interrupts
316	>	* surrounding parking and other blocking: Because interrupts are
317	>	* used solely to alert threads to check termination, which is
318	>	* checked anyway upon blocking, we clear status (using
319	>	* Thread.interrupted) before any call to park, so that park does
320	>	* not immediately return due to status being set via some other
321	>	* unrelated call to interrupt in user code.
322		*
323		* Signalling.  We create or wake up workers only when there
324		* appears to be at least one task they might be able to find and
325		* execute.  When a submission is added or another worker adds a
326	<	* task to a queue that previously had fewer than two tasks, they
327	<	* signal waiting workers (or trigger creation of new ones if
328	<	* fewer than the given parallelism level -- see signalWork).
329	<	* These primary signals are buttressed by signals during rescans;
330	<	* together these cover the signals needed in cases when more
331	<	* tasks are pushed but untaken, and improve performance compared
332	<	* to having one thread wake up all workers.
326	>	* task to a queue that has fewer than two tasks, they signal
327	>	* waiting workers (or trigger creation of new ones if fewer than
328	>	* the given parallelism level -- signalWork).  These primary
329	>	* signals are buttressed by others whenever other threads remove
330	>	* a task from a queue and notice that there are other tasks there
331	>	* as well.  So in general, pools will be over-signalled. On most
332	>	* platforms, signalling (unpark) overhead time is noticeably
333	>	* long, and the time between signalling a thread and it actually
334	>	* making progress can be very noticeably long, so it is worth
335	>	* offloading these delays from critical paths as much as
336	>	* possible. Additionally, workers spin-down gradually, by staying
337	>	* alive so long as they see the ctl state changing.  Similar
338	>	* stability-sensing techniques are also used before blocking in
339	>	* awaitJoin and helpComplete.
340		*
341		* Trimming workers. To release resources after periods of lack of
342		* use, a worker starting to wait when the pool is quiescent will
343	<	* time out and terminate if the pool has remained quiescent for
344	<	* SHRINK_RATE nanosecs. This will slowly propagate, eventually
345	<	* terminating all workers after long periods of non-use.
343	>	* time out and terminate if the pool has remained quiescent for a
344	>	* given period -- a short period if there are more threads than
345	>	* parallelism, longer as the number of threads decreases. This
346	>	* will slowly propagate, eventually terminating all workers after
347	>	* periods of non-use.
348		*
349		* Shutdown and Termination. A call to shutdownNow atomically sets
350	<	* a runState bit and then (non-atomically) sets each worker's
351	<	* runState status, cancels all unprocessed tasks, and wakes up
350	>	* a plock bit and then (non-atomically) sets each worker's
351	>	* qlock status, cancels all unprocessed tasks, and wakes up
352		* all waiting workers.  Detecting whether termination should
353		* commence after a non-abrupt shutdown() call requires more work
354		* and bookkeeping. We need consensus about quiescence (i.e., that
#	Line 354 | Line 376 | public class ForkJoinPool extends Abstra
376		*      method tryCompensate() may create or re-activate a spare
377		*      thread to compensate for blocked joiners until they unblock.
378		*
379	<	* A third form (implemented in tryRemoveAndExec and
380	<	* tryPollForAndExec) amounts to helping a hypothetical
381	<	* compensator: If we can readily tell that a possible action of a
382	<	* compensator is to steal and execute the task being joined, the
383	<	* joining thread can do so directly, without the need for a
384	<	* compensation thread (although at the expense of larger run-time
385	<	* stacks, but the tradeoff is typically worthwhile).
379	>	* A third form (implemented in tryRemoveAndExec) amounts to
380	>	* helping a hypothetical compensator: If we can readily tell that
381	>	* a possible action of a compensator is to steal and execute the
382	>	* task being joined, the joining thread can do so directly,
383	>	* without the need for a compensation thread (although at the
384	>	* expense of larger run-time stacks, but the tradeoff is
385	>	* typically worthwhile).
386		*
387		* The ManagedBlocker extension API can't use helping so relies
388		* only on compensation in method awaitBlocker.
#	Line 382 | Line 404 | public class ForkJoinPool extends Abstra
404		* steals, rather than use per-task bookkeeping.  This sometimes
405		* requires a linear scan of workQueues array to locate stealers,
406		* but often doesn't because stealers leave hints (that may become
407	<	* stale/wrong) of where to locate them.  A stealHint is only a
408	<	* hint because a worker might have had multiple steals and the
409	<	* hint records only one of them (usually the most current).
410	<	* Hinting isolates cost to when it is needed, rather than adding
411	<	* to per-task overhead.  (2) It is "shallow", ignoring nesting
412	<	* and potentially cyclic mutual steals.  (3) It is intentionally
407	>	* stale/wrong) of where to locate them.  It is only a hint
408	>	* because a worker might have had multiple steals and the hint
409	>	* records only one of them (usually the most current).  Hinting
410	>	* isolates cost to when it is needed, rather than adding to
411	>	* per-task overhead.  (2) It is "shallow", ignoring nesting and
412	>	* potentially cyclic mutual steals.  (3) It is intentionally
413		* racy: field currentJoin is updated only while actively joining,
414		* which means that we miss links in the chain during long-lived
415		* tasks, GC stalls etc (which is OK since blocking in such cases
#	Line 395 | Line 417 | public class ForkJoinPool extends Abstra
417		* to find work (see MAX_HELP) and fall back to suspending the
418		* worker and if necessary replacing it with another.
419		*
420	+	* Helping actions for CountedCompleters are much simpler: Method
421	+	* helpComplete can take and execute any task with the same root
422	+	* as the task being waited on. However, this still entails some
423	+	* traversal of completer chains, so is less efficient than using
424	+	* CountedCompleters without explicit joins.
425	+	*
426		* It is impossible to keep exactly the target parallelism number
427		* of threads running at any given time.  Determining the
428		* existence of conservatively safe helping targets, the
#	Line 416 | Line 444 | public class ForkJoinPool extends Abstra
444		* intractable) game with an opponent that may choose the worst
445		* (for us) active thread to stall at any time.  We take several
446		* precautions to bound losses (and thus bound gains), mainly in
447	<	* methods tryCompensate and awaitJoin: (1) We only try
448	<	* compensation after attempting enough helping steps (measured
449	<	* via counting and timing) that we have already consumed the
450	<	* estimated cost of creating and activating a new thread.  (2) We
451	<	* allow up to 50% of threads to be blocked before initially
452	<	* adding any others, and unless completely saturated, check that
453	<	* some work is available for a new worker before adding. Also, we
454	<	* create up to only 50% more threads until entering a mode that
455	<	* only adds a thread if all others are possibly blocked.  All
456	<	* together, this means that we might be half as fast to react,
457	<	* and create half as many threads as possible in the ideal case,
458	<	* but present vastly fewer anomalies in all other cases compared
459	<	* to both more aggressive and more conservative alternatives.
460	<	*
461	<	* Style notes: There is a lot of representation-level coupling
462	<	* among classes ForkJoinPool, ForkJoinWorkerThread, and
463	<	* ForkJoinTask.  The fields of WorkQueue maintain data structures
464	<	* managed by ForkJoinPool, so are directly accessed.  There is
465	<	* little point trying to reduce this, since any associated future
466	<	* changes in representations will need to be accompanied by
467	<	* algorithmic changes anyway. Several methods intrinsically
468	<	* sprawl because they must accumulate sets of consistent reads of
469	<	* volatiles held in local variables.  Methods signalWork() and
470	<	* scan() are the main bottlenecks, so are especially heavily
447	>	* methods tryCompensate and awaitJoin.
448	>	*
449	>	* Common Pool
450	>	* ===========
451	>	*
452	>	* The static common pool always exists after static
453	>	* initialization.  Since it (or any other created pool) need
454	>	* never be used, we minimize initial construction overhead and
455	>	* footprint to the setup of about a dozen fields, with no nested
456	>	* allocation. Most bootstrapping occurs within method
457	>	* fullExternalPush during the first submission to the pool.
458	>	*
459	>	* When external threads submit to the common pool, they can
460	>	* perform subtask processing (see externalHelpJoin and related
461	>	* methods).  This caller-helps policy makes it sensible to set
462	>	* common pool parallelism level to one (or more) less than the
463	>	* total number of available cores, or even zero for pure
464	>	* caller-runs.  We do not need to record whether external
465	>	* submissions are to the common pool -- if not, externalHelpJoin
466	>	* returns quickly (at the most helping to signal some common pool
467	>	* workers). These submitters would otherwise be blocked waiting
468	>	* for completion, so the extra effort (with liberally sprinkled
469	>	* task status checks) in inapplicable cases amounts to an odd
470	>	* form of limited spin-wait before blocking in ForkJoinTask.join.
471	>	*
472	>	* Style notes
473	>	* ===========
474	>	*
475	>	* There is a lot of representation-level coupling among classes
476	>	* ForkJoinPool, ForkJoinWorkerThread, and ForkJoinTask.  The
477	>	* fields of WorkQueue maintain data structures managed by
478	>	* ForkJoinPool, so are directly accessed.  There is little point
479	>	* trying to reduce this, since any associated future changes in
480	>	* representations will need to be accompanied by algorithmic
481	>	* changes anyway. Several methods intrinsically sprawl because
482	>	* they must accumulate sets of consistent reads of volatiles held
483	>	* in local variables.  Methods signalWork() and scan() are the
484	>	* main bottlenecks, so are especially heavily
485		* micro-optimized/mangled.  There are lots of inline assignments
486		* (of form "while ((local = field) != 0)") which are usually the
487		* simplest way to ensure the required read orderings (which are
#	Line 447 | Line 489 | public class ForkJoinPool extends Abstra
489		* declarations of these locals at the heads of methods or blocks.
490		* There are several occurrences of the unusual "do {} while
491		* (!cas...)"  which is the simplest way to force an update of a
492	<	* CAS'ed variable. There are also other coding oddities that help
492	>	* CAS'ed variable. There are also other coding oddities (including
493	>	* several unnecessary-looking hoisted null checks) that help
494		* some methods perform reasonably even when interpreted (not
495		* compiled).
496		*
#	Line 487 | Line 530 | public class ForkJoinPool extends Abstra
530		* Returns a new worker thread operating in the given pool.
531		*
532		* @param pool the pool this thread works in
533	+	* @return the new worker thread
534		* @throws NullPointerException if the pool is null
535		*/
536		public ForkJoinWorkerThread newThread(ForkJoinPool pool);
#	Line 496 | Line 540 | public class ForkJoinPool extends Abstra
540		* Default ForkJoinWorkerThreadFactory implementation; creates a
541		* new ForkJoinWorkerThread.
542		*/
543	<	static class DefaultForkJoinWorkerThreadFactory
543	>	static final class DefaultForkJoinWorkerThreadFactory
544		implements ForkJoinWorkerThreadFactory {
545	<	public ForkJoinWorkerThread newThread(ForkJoinPool pool) {
545	>	public final ForkJoinWorkerThread newThread(ForkJoinPool pool) {
546		return new ForkJoinWorkerThread(pool);
547		}
548		}
549
550		/**
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	–	/**
551		* Class for artificial tasks that are used to replace the target
552		* of local joins if they are removed from an interior queue slot
553		* in WorkQueue.tryRemoveAndExec. We don't need the proxy to
554		* actually do anything beyond having a unique identity.
555		*/
556		static final class EmptyTask extends ForkJoinTask<Void> {
557	+	private static final long serialVersionUID = -7721805057305804111L;
558		EmptyTask() { status = ForkJoinTask.NORMAL; } // force done
559		public final Void getRawResult() { return null; }
560		public final void setRawResult(Void x) {}
#	Line 553 | Line 575 | public class ForkJoinPool extends Abstra
575		*
576		* Field "top" is the index (mod array.length) of the next queue
577		* slot to push to or pop from. It is written only by owner thread
578	<	* for push, or under lock for trySharedPush, and accessed by
579	<	* other threads only after reading (volatile) base.  Both top and
580	<	* base are allowed to wrap around on overflow, but (top - base)
581	<	* (or more commonly -(base - top) to force volatile read of base
582	<	* before top) still estimates size.
578	>	* for push, or under lock for external/shared push, and accessed
579	>	* by other threads only after reading (volatile) base.  Both top
580	>	* and base are allowed to wrap around on overflow, but (top -
581	>	* base) (or more commonly -(base - top) to force volatile read of
582	>	* base before top) still estimates size. The lock ("qlock") is
583	>	* forced to -1 on termination, causing all further lock attempts
584	>	* to fail. (Note: we don't need CAS for termination state because
585	>	* upon pool shutdown, all shared-queues will stop being used
586	>	* anyway.)  Nearly all lock bodies are set up so that exceptions
587	>	* within lock bodies are "impossible" (modulo JVM errors that
588	>	* would cause failure anyway.)
589		*
590		* The array slots are read and written using the emulation of
591		* volatiles/atomics provided by Unsafe. Insertions must in
592		* general use putOrderedObject as a form of releasing store to
593		* ensure that all writes to the task object are ordered before
594	<	* its publication in the queue. (Although we can avoid one case
595	<	* of this when locked in trySharedPush.) All removals entail a
596	<	* CAS to null.  The array is always a power of two. To ensure
597	<	* safety of Unsafe array operations, all accesses perform
570	<	* explicit null checks and implicit bounds checks via
571	<	* power-of-two masking.
594	>	* its publication in the queue.  All removals entail a CAS to
595	>	* null.  The array is always a power of two. To ensure safety of
596	>	* Unsafe array operations, all accesses perform explicit null
597	>	* checks and implicit bounds checks via power-of-two masking.
598		*
599		* In addition to basic queuing support, this class contains
600		* fields described elsewhere to control execution. It turns out
601	<	* to work better memory-layout-wise to include them in this
602	<	* class rather than a separate class.
601	>	* to work better memory-layout-wise to include them in this class
602	>	* rather than a separate class.
603		*
604		* Performance on most platforms is very sensitive to placement of
605		* instances of both WorkQueues and their arrays -- we absolutely
606		* do not want multiple WorkQueue instances or multiple queue
607		* arrays sharing cache lines. (It would be best for queue objects
608		* and their arrays to share, but there is nothing available to
609	<	* help arrange that).  Unfortunately, because they are recorded
610	<	* 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.
609	>	* help arrange that). The @Contended annotation alerts JVMs to
610	>	* try to keep instances apart.
611		*/
612		static final class WorkQueue {
613		/**
#	Line 613 | Line 630 | public class ForkJoinPool extends Abstra
630		*/
631		static final int MAXIMUM_QUEUE_CAPACITY = 1 << 26; // 64M
632
633	<	volatile long totalSteals; // cumulative number of steals
634	<	int seed;                  // for random scanning; initialize nonzero
633	>	// Heuristic padding to ameliorate unfortunate memory placements
634	>	volatile long pad00, pad01, pad02, pad03, pad04, pad05, pad06;
635	>
636		volatile int eventCount;   // encoded inactivation count; < 0 if inactive
637		int nextWait;              // encoded record of next event waiter
638	<	int rescans;               // remaining scans until block
639	<	int nsteals;               // top-level task executions since last idle
640	<	final int mode;            // lifo, fifo, or shared
641	<	int poolIndex;             // index of this queue in pool (or 0)
642	<	int stealHint;             // index of most recent known stealer
625	<	volatile int runState;     // 1: locked, -1: terminate; else 0
638	>	int nsteals;               // number of steals
639	>	int hint;                  // steal index hint
640	>	short poolIndex;           // index of this queue in pool
641	>	final short mode;          // 0: lifo, > 0: fifo, < 0: shared
642	>	volatile int qlock;        // 1: locked, -1: terminate; else 0
643		volatile int base;         // index of next slot for poll
644		int top;                   // index of next slot for push
645		ForkJoinTask<?>[] array;   // the elements (initially unallocated)
#	Line 631 | Line 648 | public class ForkJoinPool extends Abstra
648		volatile Thread parker;    // == owner during call to park; else null
649		volatile ForkJoinTask<?> currentJoin;  // task being joined in awaitJoin
650		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;
651
652	<	WorkQueue(ForkJoinPool pool, ForkJoinWorkerThread owner, int mode) {
653	<	this.mode = mode;
652	>	volatile Object pad10, pad11, pad12, pad13, pad14, pad15, pad16, pad17;
653	>	volatile Object pad18, pad19, pad1a, pad1b, pad1c, pad1d;
654	>
655	>	WorkQueue(ForkJoinPool pool, ForkJoinWorkerThread owner, int mode,
656	>	int seed) {
657		this.pool = pool;
658		this.owner = owner;
659	+	this.mode = (short)mode;
660	+	this.hint = seed; // store initial seed for runWorker
661		// Place indices in the center of array (that is not yet allocated)
662		base = top = INITIAL_QUEUE_CAPACITY >>> 1;
663		}
#	Line 663 | Line 682 | public class ForkJoinPool extends Abstra
682		(n == -1 &&
683		((a = array) == null ||
684		(m = a.length - 1) < 0 ||
685	<	U.getObjectVolatile
686	<	(a, ((m & (s - 1)) << ASHIFT) + ABASE) == null)));
685	>	U.getObject
686	>	(a, (long)((m & (s - 1)) << ASHIFT) + ABASE) == null)));
687		}
688
689		/**
690	<	* Pushes a task. Call only by owner in unshared queues.
690	>	* Pushes a task. Call only by owner in unshared queues.  (The
691	>	* shared-queue version is embedded in method externalPush.)
692		*
693		* @param task the task. Caller must ensure non-null.
694	<	* @throw RejectedExecutionException if array cannot be resized
694	>	* @throws RejectedExecutionException if array cannot be resized
695		*/
696		final void push(ForkJoinTask<?> task) {
697		ForkJoinTask<?>[] a; ForkJoinPool p;
698	<	int s = top, m, n;
698	>	int s = top, n;
699		if ((a = array) != null) {    // ignore if queue removed
700	<	U.putOrderedObject
701	<	(a, (((m = a.length - 1) & s) << ASHIFT) + ABASE, task);
702	<	if ((n = (top = s + 1) - base) <= 2) {
703	<	if ((p = pool) != null)
684	<	p.signalWork();
685	<	}
700	>	int m = a.length - 1;
701	>	U.putOrderedObject(a, ((m & s) << ASHIFT) + ABASE, task);
702	>	if ((n = (top = s + 1) - base) <= 2)
703	>	(p = pool).signalWork(p.workQueues, this);
704		else if (n >= m)
705	<	growArray(true);
705	>	growArray();
706		}
707		}
708
709		/**
710	<	* Pushes a task if lock is free and array is either big
711	<	* enough or can be resized to be big enough.
712	<	*
695	<	* @param task the task. Caller must ensure non-null.
696	<	* @return true if submitted
710	>	* Initializes or doubles the capacity of array. Call either
711	>	* by owner or with lock held -- it is OK for base, but not
712	>	* top, to move while resizings are in progress.
713		*/
714	<	final boolean trySharedPush(ForkJoinTask<?> task) {
715	<	boolean submitted = false;
716	<	if (runState == 0 && U.compareAndSwapInt(this, RUNSTATE, 0, 1)) {
717	<	ForkJoinTask<?>[] a = array;
718	<	int s = top;
719	<	try {
720	<	if ((a != null && a.length > s + 1 - base) ||
721	<	(a = growArray(false)) != null) { // must presize
722	<	int j = (((a.length - 1) & s) << ASHIFT) + ABASE;
723	<	U.putObject(a, (long)j, task);    // don't need "ordered"
724	<	top = s + 1;
725	<	submitted = true;
726	<	}
727	<	} finally {
728	<	runState = 0;                         // unlock
729	<	}
714	>	final ForkJoinTask<?>[] growArray() {
715	>	ForkJoinTask<?>[] oldA = array;
716	>	int size = oldA != null ? oldA.length << 1 : INITIAL_QUEUE_CAPACITY;
717	>	if (size > MAXIMUM_QUEUE_CAPACITY)
718	>	throw new RejectedExecutionException("Queue capacity exceeded");
719	>	int oldMask, t, b;
720	>	ForkJoinTask<?>[] a = array = new ForkJoinTask<?>[size];
721	>	if (oldA != null && (oldMask = oldA.length - 1) >= 0 &&
722	>	(t = top) - (b = base) > 0) {
723	>	int mask = size - 1;
724	>	do {
725	>	ForkJoinTask<?> x;
726	>	int oldj = ((b & oldMask) << ASHIFT) + ABASE;
727	>	int j    = ((b &    mask) << ASHIFT) + ABASE;
728	>	x = (ForkJoinTask<?>)U.getObjectVolatile(oldA, oldj);
729	>	if (x != null &&
730	>	U.compareAndSwapObject(oldA, oldj, x, null))
731	>	U.putObjectVolatile(a, j, x);
732	>	} while (++b != t);
733		}
734	<	return submitted;
734	>	return a;
735		}
736
737		/**
738		* Takes next task, if one exists, in LIFO order.  Call only
739	<	* by owner in unshared queues. (We do not have a shared
721	<	* version of this method because it is never needed.)
739	>	* by owner in unshared queues.
740		*/
741		final ForkJoinTask<?> pop() {
742		ForkJoinTask<?>[] a; ForkJoinTask<?> t; int m;
#	Line 746 | Line 764 | public class ForkJoinPool extends Abstra
764		if ((a = array) != null) {
765		int j = (((a.length - 1) & b) << ASHIFT) + ABASE;
766		if ((t = (ForkJoinTask<?>)U.getObjectVolatile(a, j)) != null &&
767	<	base == b &&
768	<	U.compareAndSwapObject(a, j, t, null)) {
751	<	base = b + 1;
767	>	base == b && U.compareAndSwapObject(a, j, t, null)) {
768	>	U.putOrderedInt(this, QBASE, b + 1);
769		return t;
770		}
771		}
#	Line 764 | Line 781 | public class ForkJoinPool extends Abstra
781		int j = (((a.length - 1) & b) << ASHIFT) + ABASE;
782		t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
783		if (t != null) {
784	<	if (base == b &&
785	<	U.compareAndSwapObject(a, j, t, null)) {
769	<	base = b + 1;
784	>	if (U.compareAndSwapObject(a, j, t, null)) {
785	>	U.putOrderedInt(this, QBASE, b + 1);
786		return t;
787		}
788		}
789		else if (base == b) {
790		if (b + 1 == top)
791		break;
792	<	Thread.yield(); // wait for lagging update
792	>	Thread.yield(); // wait for lagging update (very rare)
793		}
794		}
795		return null;
#	Line 800 | Line 816 | public class ForkJoinPool extends Abstra
816
817		/**
818		* Pops the given task only if it is at the current top.
819	+	* (A shared version is available only via FJP.tryExternalUnpush)
820		*/
821		final boolean tryUnpush(ForkJoinTask<?> t) {
822		ForkJoinTask<?>[] a; int s;
#	Line 813 | Line 830 | public class ForkJoinPool extends Abstra
830		}
831
832		/**
816	–	* Polls the given task only if it is at the current base.
817	–	*/
818	–	final boolean pollFor(ForkJoinTask<?> task) {
819	–	ForkJoinTask<?>[] a; int b;
820	–	if ((b = base) - top < 0 && (a = array) != null) {
821	–	int j = (((a.length - 1) & b) << ASHIFT) + ABASE;
822	–	if (U.getObjectVolatile(a, j) == task && base == b &&
823	–	U.compareAndSwapObject(a, j, task, null)) {
824	–	base = b + 1;
825	–	return true;
826	–	}
827	–	}
828	–	return false;
829	–	}
830	–
831	–	/**
832	–	* Initializes or doubles the capacity of array. Call either
833	–	* by owner or with lock held -- it is OK for base, but not
834	–	* top, to move while resizings are in progress.
835	–	*
836	–	* @param rejectOnFailure if true, throw exception if capacity
837	–	* exceeded (relayed ultimately to user); else return null.
838	–	*/
839	–	final ForkJoinTask<?>[] growArray(boolean rejectOnFailure) {
840	–	ForkJoinTask<?>[] oldA = array;
841	–	int size = oldA != null ? oldA.length << 1 : INITIAL_QUEUE_CAPACITY;
842	–	if (size <= MAXIMUM_QUEUE_CAPACITY) {
843	–	int oldMask, t, b;
844	–	ForkJoinTask<?>[] a = array = new ForkJoinTask<?>[size];
845	–	if (oldA != null && (oldMask = oldA.length - 1) >= 0 &&
846	–	(t = top) - (b = base) > 0) {
847	–	int mask = size - 1;
848	–	do {
849	–	ForkJoinTask<?> x;
850	–	int oldj = ((b & oldMask) << ASHIFT) + ABASE;
851	–	int j    = ((b &    mask) << ASHIFT) + ABASE;
852	–	x = (ForkJoinTask<?>)U.getObjectVolatile(oldA, oldj);
853	–	if (x != null &&
854	–	U.compareAndSwapObject(oldA, oldj, x, null))
855	–	U.putObjectVolatile(a, j, x);
856	–	} while (++b != t);
857	–	}
858	–	return a;
859	–	}
860	–	else if (!rejectOnFailure)
861	–	return null;
862	–	else
863	–	throw new RejectedExecutionException("Queue capacity exceeded");
864	–	}
865	–
866	–	/**
833		* Removes and cancels all known tasks, ignoring any exceptions.
834		*/
835		final void cancelAll() {
#	Line 873 | Line 839 | public class ForkJoinPool extends Abstra
839		ForkJoinTask.cancelIgnoringExceptions(t);
840		}
841
842	<	/**
877	<	* Computes next value for random probes.  Scans don't require
878	<	* a very high quality generator, but also not a crummy one.
879	<	* Marsaglia xor-shift is cheap and works well enough.  Note:
880	<	* This is manually inlined in its usages in ForkJoinPool to
881	<	* avoid writes inside busy scan loops.
882	<	*/
883	<	final int nextSeed() {
884	<	int r = seed;
885	<	r ^= r << 13;
886	<	r ^= r >>> 17;
887	<	return seed = r ^= r << 5;
888	<	}
889	<
890	<	// Execution methods
891	<
892	<	/**
893	<	* Pops and runs tasks until empty.
894	<	*/
895	<	private void popAndExecAll() {
896	<	// A bit faster than repeated pop calls
897	<	ForkJoinTask<?>[] a; int m, s; long j; ForkJoinTask<?> t;
898	<	while ((a = array) != null && (m = a.length - 1) >= 0 &&
899	<	(s = top - 1) - base >= 0 &&
900	<	(t = ((ForkJoinTask<?>)
901	<	U.getObject(a, j = ((m & s) << ASHIFT) + ABASE)))
902	<	!= null) {
903	<	if (U.compareAndSwapObject(a, j, t, null)) {
904	<	top = s;
905	<	t.doExec();
906	<	}
907	<	}
908	<	}
842	>	// Specialized execution methods
843
844		/**
845		* Polls and runs tasks until empty.
846		*/
847	<	private void pollAndExecAll() {
847	>	final void pollAndExecAll() {
848		for (ForkJoinTask<?> t; (t = poll()) != null;)
849		t.doExec();
850		}
851
852		/**
853	<	* If present, removes from queue and executes the given task, or
854	<	* any other cancelled task. Returns (true) immediately on any CAS
853	>	* Executes a top-level task and any local tasks remaining
854	>	* after execution.
855	>	*/
856	>	final void runTask(ForkJoinTask<?> task) {
857	>	if ((currentSteal = task) != null) {
858	>	task.doExec();
859	>	ForkJoinTask<?>[] a = array;
860	>	int md = mode;
861	>	++nsteals;
862	>	currentSteal = null;
863	>	if (md != 0)
864	>	pollAndExecAll();
865	>	else if (a != null) {
866	>	int s, m = a.length - 1;
867	>	while ((s = top - 1) - base >= 0) {
868	>	long i = ((m & s) << ASHIFT) + ABASE;
869	>	ForkJoinTask<?> t = (ForkJoinTask<?>)U.getObject(a, i);
870	>	if (t == null)
871	>	break;
872	>	if (U.compareAndSwapObject(a, i, t, null)) {
873	>	top = s;
874	>	t.doExec();
875	>	}
876	>	}
877	>	}
878	>	}
879	>	}
880	>
881	>	/**
882	>	* If present, removes from queue and executes the given task,
883	>	* or any other cancelled task. Returns (true) on any CAS
884		* or consistency check failure so caller can retry.
885		*
886	<	* @return 0 if no progress can be made, else positive
924	<	* (this unusual convention simplifies use with tryHelpStealer.)
886	>	* @return false if no progress can be made, else true
887		*/
888	<	final int tryRemoveAndExec(ForkJoinTask<?> task) {
889	<	int stat = 1;
928	<	boolean removed = false, empty = true;
888	>	final boolean tryRemoveAndExec(ForkJoinTask<?> task) {
889	>	boolean stat;
890		ForkJoinTask<?>[] a; int m, s, b, n;
891	<	if ((a = array) != null && (m = a.length - 1) >= 0 &&
891	>	if (task != null && (a = array) != null && (m = a.length - 1) >= 0 &&
892		(n = (s = top) - (b = base)) > 0) {
893	+	boolean removed = false, empty = true;
894	+	stat = true;
895		for (ForkJoinTask<?> t;;) {           // traverse from s to b
896	<	int j = ((--s & m) << ASHIFT) + ABASE;
897	<	t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
896	>	long j = ((--s & m) << ASHIFT) + ABASE;
897	>	t = (ForkJoinTask<?>)U.getObject(a, j);
898		if (t == null)                    // inconsistent length
899		break;
900		else if (t == task) {
#	Line 955 | Line 918 | public class ForkJoinPool extends Abstra
918		}
919		if (--n == 0) {
920		if (!empty && base == b)
921	<	stat = 0;
921	>	stat = false;
922		break;
923		}
924		}
925	+	if (removed)
926	+	task.doExec();
927		}
928	<	if (removed)
929	<	task.doExec();
928	>	else
929	>	stat = false;
930		return stat;
931		}
932
933		/**
934	<	* Executes a top-level task and any local tasks remaining
935	<	* after execution.
934	>	* Tries to poll for and execute the given task or any other
935	>	* task in its CountedCompleter computation.
936		*/
937	<	final void runTask(ForkJoinTask<?> t) {
938	<	if (t != null) {
939	<	currentSteal = t;
940	<	t.doExec();
941	<	if (top != base) {       // process remaining local tasks
942	<	if (mode == 0)
943	<	popAndExecAll();
944	<	else
945	<	pollAndExecAll();
937	>	final boolean pollAndExecCC(CountedCompleter<?> root) {
938	>	ForkJoinTask<?>[] a; int b; Object o; CountedCompleter<?> t, r;
939	>	if ((b = base) - top < 0 && (a = array) != null) {
940	>	long j = (((a.length - 1) & b) << ASHIFT) + ABASE;
941	>	if ((o = U.getObjectVolatile(a, j)) == null)
942	>	return true; // retry
943	>	if (o instanceof CountedCompleter) {
944	>	for (t = (CountedCompleter<?>)o, r = t;;) {
945	>	if (r == root) {
946	>	if (base == b &&
947	>	U.compareAndSwapObject(a, j, t, null)) {
948	>	U.putOrderedInt(this, QBASE, b + 1);
949	>	t.doExec();
950	>	}
951	>	return true;
952	>	}
953	>	else if ((r = r.completer) == null)
954	>	break; // not part of root computation
955	>	}
956		}
982	–	++nsteals;
983	–	currentSteal = null;
957		}
958	+	return false;
959		}
960
961		/**
962	<	* Executes a non-top-level (stolen) task.
962	>	* Tries to pop and execute the given task or any other task
963	>	* in its CountedCompleter computation.
964		*/
965	<	final void runSubtask(ForkJoinTask<?> t) {
966	<	if (t != null) {
967	<	ForkJoinTask<?> ps = currentSteal;
968	<	currentSteal = t;
969	<	t.doExec();
970	<	currentSteal = ps;
965	>	final boolean externalPopAndExecCC(CountedCompleter<?> root) {
966	>	ForkJoinTask<?>[] a; int s; Object o; CountedCompleter<?> t, r;
967	>	if (base - (s = top) < 0 && (a = array) != null) {
968	>	long j = (((a.length - 1) & (s - 1)) << ASHIFT) + ABASE;
969	>	if ((o = U.getObject(a, j)) instanceof CountedCompleter) {
970	>	for (t = (CountedCompleter<?>)o, r = t;;) {
971	>	if (r == root) {
972	>	if (U.compareAndSwapInt(this, QLOCK, 0, 1)) {
973	>	if (top == s && array == a &&
974	>	U.compareAndSwapObject(a, j, t, null)) {
975	>	top = s - 1;
976	>	qlock = 0;
977	>	t.doExec();
978	>	}
979	>	else
980	>	qlock = 0;
981	>	}
982	>	return true;
983	>	}
984	>	else if ((r = r.completer) == null)
985	>	break;
986	>	}
987	>	}
988		}
989	+	return false;
990	+	}
991	+
992	+	/**
993	+	* Internal version
994	+	*/
995	+	final boolean internalPopAndExecCC(CountedCompleter<?> root) {
996	+	ForkJoinTask<?>[] a; int s; Object o; CountedCompleter<?> t, r;
997	+	if (base - (s = top) < 0 && (a = array) != null) {
998	+	long j = (((a.length - 1) & (s - 1)) << ASHIFT) + ABASE;
999	+	if ((o = U.getObject(a, j)) instanceof CountedCompleter) {
1000	+	for (t = (CountedCompleter<?>)o, r = t;;) {
1001	+	if (r == root) {
1002	+	if (U.compareAndSwapObject(a, j, t, null)) {
1003	+	top = s - 1;
1004	+	t.doExec();
1005	+	}
1006	+	return true;
1007	+	}
1008	+	else if ((r = r.completer) == null)
1009	+	break;
1010	+	}
1011	+	}
1012	+	}
1013	+	return false;
1014		}
1015
1016		/**
#	Line 1008 | Line 1025 | public class ForkJoinPool extends Abstra
1025		s != Thread.State.TIMED_WAITING);
1026		}
1027
1011	–	/**
1012	–	* If this owned and is not already interrupted, try to
1013	–	* interrupt and/or unpark, ignoring exceptions.
1014	–	*/
1015	–	final void interruptOwner() {
1016	–	Thread wt, p;
1017	–	if ((wt = owner) != null && !wt.isInterrupted()) {
1018	–	try {
1019	–	wt.interrupt();
1020	–	} catch (SecurityException ignore) {
1021	–	}
1022	–	}
1023	–	if ((p = parker) != null)
1024	–	U.unpark(p);
1025	–	}
1026	–
1028		// Unsafe mechanics
1029		private static final sun.misc.Unsafe U;
1030	<	private static final long RUNSTATE;
1030	>	private static final long QBASE;
1031	>	private static final long QLOCK;
1032		private static final int ABASE;
1033		private static final int ASHIFT;
1034		static {
1033	–	int s;
1035		try {
1036		U = getUnsafe();
1037		Class<?> k = WorkQueue.class;
1038		Class<?> ak = ForkJoinTask[].class;
1039	<	RUNSTATE = U.objectFieldOffset
1040	<	(k.getDeclaredField("runState"));
1039	>	QBASE = U.objectFieldOffset
1040	>	(k.getDeclaredField("base"));
1041	>	QLOCK = U.objectFieldOffset
1042	>	(k.getDeclaredField("qlock"));
1043		ABASE = U.arrayBaseOffset(ak);
1044	<	s = U.arrayIndexScale(ak);
1044	>	int scale = U.arrayIndexScale(ak);
1045	>	if ((scale & (scale - 1)) != 0)
1046	>	throw new Error("data type scale not a power of two");
1047	>	ASHIFT = 31 - Integer.numberOfLeadingZeros(scale);
1048		} catch (Exception e) {
1049		throw new Error(e);
1050		}
1045	–	if ((s & (s-1)) != 0)
1046	–	throw new Error("data type scale not a power of two");
1047	–	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
1051		}
1052		}
1053
1054	+	// static fields (initialized in static initializer below)
1055	+
1056		/**
1057	<	* Per-thread records for threads that submit to pools. Currently
1058	<	* holds only pseudo-random seed / index that is used to choose
1059	<	* submission queues in method doSubmit. In the future, this may
1060	<	* also incorporate a means to implement different task rejection
1061	<	* and resubmission policies.
1057	<	*
1058	<	* Seeds for submitters and workers/workQueues work in basically
1059	<	* the same way but are initialized and updated using slightly
1060	<	* different mechanics. Both are initialized using the same
1061	<	* approach as in class ThreadLocal, where successive values are
1062	<	* unlikely to collide with previous values. This is done during
1063	<	* registration for workers, but requires a separate AtomicInteger
1064	<	* for submitters. Seeds are then randomly modified upon
1065	<	* collisions using xorshifts, which requires a non-zero seed.
1057	>	* Per-thread submission bookkeeping. Shared across all pools
1058	>	* to reduce ThreadLocal pollution and because random motion
1059	>	* to avoid contention in one pool is likely to hold for others.
1060	>	* Lazily initialized on first submission (but null-checked
1061	>	* in other contexts to avoid unnecessary initialization).
1062		*/
1063	<	static final class Submitter {
1068	<	int seed;
1069	<	Submitter() {
1070	<	int s = nextSubmitterSeed.getAndAdd(SEED_INCREMENT);
1071	<	seed = (s == 0) ? 1 : s; // ensure non-zero
1072	<	}
1073	<	}
1074	<
1075	<	/** ThreadLocal class for Submitters */
1076	<	static final class ThreadSubmitter extends ThreadLocal<Submitter> {
1077	<	public Submitter initialValue() { return new Submitter(); }
1078	<	}
1079	<
1080	<	// static fields (initialized in static initializer below)
1063	>	static final ThreadLocal<Submitter> submitters;
1064
1065		/**
1066		* Creates a new ForkJoinWorkerThread. This factory is used unless
#	Line 1087 | Line 1070 | public class ForkJoinPool extends Abstra
1070		defaultForkJoinWorkerThreadFactory;
1071
1072		/**
1073	<	* Generator for assigning sequence numbers as pool names.
1073	>	* Permission required for callers of methods that may start or
1074	>	* kill threads.
1075		*/
1076	<	private static final AtomicInteger poolNumberGenerator;
1076	>	private static final RuntimePermission modifyThreadPermission;
1077
1078		/**
1079	<	* Generator for initial hashes/seeds for submitters. Accessed by
1080	<	* Submitter class constructor.
1079	>	* Common (static) pool. Non-null for public use unless a static
1080	>	* construction exception, but internal usages null-check on use
1081	>	* to paranoically avoid potential initialization circularities
1082	>	* as well as to simplify generated code.
1083		*/
1084	<	static final AtomicInteger nextSubmitterSeed;
1084	>	static final ForkJoinPool common;
1085
1086		/**
1087	<	* Permission required for callers of methods that may start or
1088	<	* kill threads.
1087	>	* Common pool parallelism. To allow simpler use and management
1088	>	* when common pool threads are disabled, we allow the underlying
1089	>	* common.parallelism field to be zero, but in that case still report
1090	>	* parallelism as 1 to reflect resulting caller-runs mechanics.
1091		*/
1092	<	private static final RuntimePermission modifyThreadPermission;
1092	>	static final int commonParallelism;
1093
1094		/**
1095	<	* Per-thread submission bookkeeping. Shared across all pools
1108	<	* to reduce ThreadLocal pollution and because random motion
1109	<	* to avoid contention in one pool is likely to hold for others.
1095	>	* Sequence number for creating workerNamePrefix.
1096		*/
1097	<	private static final ThreadSubmitter submitters;
1097	>	private static int poolNumberSequence;
1098	>
1099	>	/**
1100	>	* Returns the next sequence number. We don't expect this to
1101	>	* ever contend, so use simple builtin sync.
1102	>	*/
1103	>	private static final synchronized int nextPoolId() {
1104	>	return ++poolNumberSequence;
1105	>	}
1106
1107		// static constants
1108
1109		/**
1110	<	* The wakeup interval (in nanoseconds) for a worker waiting for a
1111	<	* task when the pool is quiescent to instead try to shrink the
1112	<	* number of workers.  The exact value does not matter too
1113	<	* much. It must be short enough to release resources during
1114	<	* sustained periods of idleness, but not so short that threads
1115	<	* are continually re-created.
1110	>	* Initial timeout value (in nanoseconds) for the thread
1111	>	* triggering quiescence to park waiting for new work. On timeout,
1112	>	* the thread will instead try to shrink the number of
1113	>	* workers. The value should be large enough to avoid overly
1114	>	* aggressive shrinkage during most transient stalls (long GCs
1115	>	* etc).
1116		*/
1117	<	private static final long SHRINK_RATE =
1124	<	4L * 1000L * 1000L * 1000L; // 4 seconds
1117	>	private static final long IDLE_TIMEOUT      = 2000L * 1000L * 1000L; // 2sec
1118
1119		/**
1120	<	* The timeout value for attempted shrinkage, includes
1128	<	* some slop to cope with system timer imprecision.
1120	>	* Timeout value when there are more threads than parallelism level
1121		*/
1122	<	private static final long SHRINK_TIMEOUT = SHRINK_RATE - (SHRINK_RATE / 10);
1122	>	private static final long FAST_IDLE_TIMEOUT =  200L * 1000L * 1000L;
1123	>
1124	>	/**
1125	>	* Tolerance for idle timeouts, to cope with timer undershoots
1126	>	*/
1127	>	private static final long TIMEOUT_SLOP = 2000000L;
1128
1129		/**
1130		* The maximum stolen->joining link depth allowed in method
1131	<	* tryHelpStealer.  Must be a power of two. This value also
1135	<	* controls the maximum number of times to try to help join a task
1136	<	* without any apparent progress or change in pool state before
1137	<	* giving up and blocking (see awaitJoin).  Depths for legitimate
1131	>	* tryHelpStealer.  Must be a power of two.  Depths for legitimate
1132		* chains are unbounded, but we use a fixed constant to avoid
1133		* (otherwise unchecked) cycles and to bound staleness of
1134		* traversal parameters at the expense of sometimes blocking when
#	Line 1143 | Line 1137 | public class ForkJoinPool extends Abstra
1137		private static final int MAX_HELP = 64;
1138
1139		/**
1146	–	* Secondary time-based bound (in nanosecs) for helping attempts
1147	–	* before trying compensated blocking in awaitJoin. Used in
1148	–	* conjunction with MAX_HELP to reduce variance due to different
1149	–	* polling rates associated with different helping options. The
1150	–	* value should roughly approximate the time required to create
1151	–	* and/or activate a worker thread.
1152	–	*/
1153	–	private static final long COMPENSATION_DELAY = 1L << 18; // ~0.25 millisec
1154	–
1155	–	/**
1140		* Increment for seed generators. See class ThreadLocal for
1141		* explanation.
1142		*/
1143		private static final int SEED_INCREMENT = 0x61c88647;
1144
1145	<	/**
1145	>	/*
1146		* Bits and masks for control variables
1147		*
1148		* Field ctl is a long packed with:
#	Line 1186 | Line 1170 | public class ForkJoinPool extends Abstra
1170		* scan for them to avoid queuing races. Note however that
1171		* eventCount updates lag releases so usage requires care.
1172		*
1173	<	* Field runState is an int packed with:
1173	>	* Field plock is an int packed with:
1174		* SHUTDOWN: true if shutdown is enabled (1 bit)
1175	<	* SEQ:  a sequence number updated upon (de)registering workers (30 bits)
1176	<	* INIT: set true after workQueues array construction (1 bit)
1175	>	* SEQ:  a sequence lock, with PL_LOCK bit set if locked (30 bits)
1176	>	* SIGNAL: set when threads may be waiting on the lock (1 bit)
1177		*
1178		* The sequence number enables simple consistency checks:
1179		* Staleness of read-only operations on the workQueues array can
1180	<	* be checked by comparing runState before vs after the reads.
1180	>	* be checked by comparing plock before vs after the reads.
1181		*/
1182
1183		// bit positions/shifts for fields
#	Line 1205 | Line 1189 | public class ForkJoinPool extends Abstra
1189		// bounds
1190		private static final int  SMASK      = 0xffff;  // short bits
1191		private static final int  MAX_CAP    = 0x7fff;  // max #workers - 1
1192	<	private static final int  SQMASK     = 0xfffe;  // even short bits
1192	>	private static final int  EVENMASK   = 0xfffe;  // even short bits
1193	>	private static final int  SQMASK     = 0x007e;  // max 64 (even) slots
1194		private static final int  SHORT_SIGN = 1 << 15;
1195		private static final int  INT_SIGN   = 1 << 31;
1196
#	Line 1230 | Line 1215 | public class ForkJoinPool extends Abstra
1215		private static final int E_MASK      = 0x7fffffff; // no STOP_BIT
1216		private static final int E_SEQ       = 1 << EC_SHIFT;
1217
1218	<	// runState bits
1218	>	// plock bits
1219		private static final int SHUTDOWN    = 1 << 31;
1220	+	private static final int PL_LOCK     = 2;
1221	+	private static final int PL_SIGNAL   = 1;
1222	+	private static final int PL_SPINS    = 1 << 8;
1223
1224		// access mode for WorkQueue
1225		static final int LIFO_QUEUE          =  0;
1226		static final int FIFO_QUEUE          =  1;
1227		static final int SHARED_QUEUE        = -1;
1228
1229	<	// Instance fields
1230	<
1243	<	/*
1244	<	* Field layout order in this class tends to matter more than one
1245	<	* would like. Runtime layout order is only loosely related to
1246	<	* declaration order and may differ across JVMs, but the following
1247	<	* empirically works OK on current JVMs.
1248	<	*/
1229	>	// Heuristic padding to ameliorate unfortunate memory placements
1230	>	volatile long pad00, pad01, pad02, pad03, pad04, pad05, pad06;
1231
1232	+	// Instance fields
1233	+	volatile long stealCount;                  // collects worker counts
1234		volatile long ctl;                         // main pool control
1235	<	final int parallelism;                     // parallelism level
1236	<	final int localMode;                       // per-worker scheduling mode
1237	<	final int submitMask;                      // submit queue index bound
1238	<	int nextSeed;                              // for initializing worker seeds
1255	<	volatile int runState;                     // shutdown status and seq
1235	>	volatile int plock;                        // shutdown status and seqLock
1236	>	volatile int indexSeed;                    // worker/submitter index seed
1237	>	final short parallelism;                   // parallelism level
1238	>	final short mode;                          // LIFO/FIFO
1239		WorkQueue[] workQueues;                    // main registry
1240	<	final Mutex lock;                          // for registration
1241	<	final Condition termination;               // for awaitTermination
1259	<	final ForkJoinWorkerThreadFactory factory; // factory for new workers
1260	<	final Thread.UncaughtExceptionHandler ueh; // per-worker UEH
1261	<	final AtomicLong stealCount;               // collect counts when terminated
1262	<	final AtomicInteger nextWorkerNumber;      // to create worker name string
1240	>	final ForkJoinWorkerThreadFactory factory;
1241	>	final UncaughtExceptionHandler ueh;        // per-worker UEH
1242		final String workerNamePrefix;             // to create worker name string
1243
1244	<	//  Creating, registering, and deregistering workers
1244	>	volatile Object pad10, pad11, pad12, pad13, pad14, pad15, pad16, pad17;
1245	>	volatile Object pad18, pad19, pad1a, pad1b;
1246
1247		/**
1248	<	* Tries to create and start a worker
1249	<	*/
1250	<	private void addWorker() {
1251	<	Throwable ex = null;
1252	<	ForkJoinWorkerThread wt = null;
1253	<	try {
1254	<	if ((wt = factory.newThread(this)) != null) {
1255	<	wt.start();
1256	<	return;
1248	>	* Acquires the plock lock to protect worker array and related
1249	>	* updates. This method is called only if an initial CAS on plock
1250	>	* fails. This acts as a spinlock for normal cases, but falls back
1251	>	* to builtin monitor to block when (rarely) needed. This would be
1252	>	* a terrible idea for a highly contended lock, but works fine as
1253	>	* a more conservative alternative to a pure spinlock.
1254	>	*/
1255	>	private int acquirePlock() {
1256	>	int spins = PL_SPINS, ps, nps;
1257	>	for (;;) {
1258	>	if (((ps = plock) & PL_LOCK) == 0 &&
1259	>	U.compareAndSwapInt(this, PLOCK, ps, nps = ps + PL_LOCK))
1260	>	return nps;
1261	>	else if (spins >= 0) {
1262	>	if (ThreadLocalRandom.current().nextInt() >= 0)
1263	>	--spins;
1264	>	}
1265	>	else if (U.compareAndSwapInt(this, PLOCK, ps, ps | PL_SIGNAL)) {
1266	>	synchronized (this) {
1267	>	if ((plock & PL_SIGNAL) != 0) {
1268	>	try {
1269	>	wait();
1270	>	} catch (InterruptedException ie) {
1271	>	try {
1272	>	Thread.currentThread().interrupt();
1273	>	} catch (SecurityException ignore) {
1274	>	}
1275	>	}
1276	>	}
1277	>	else
1278	>	notifyAll();
1279	>	}
1280		}
1278	–	} catch (Throwable e) {
1279	–	ex = e;
1281		}
1281	–	deregisterWorker(wt, ex); // adjust counts etc on failure
1282		}
1283
1284		/**
1285	<	* Callback from ForkJoinWorkerThread constructor to assign a
1286	<	* public name. This must be separate from registerWorker because
1287	<	* it is called during the "super" constructor call in
1288	<	* ForkJoinWorkerThread.
1285	>	* Unlocks and signals any thread waiting for plock. Called only
1286	>	* when CAS of seq value for unlock fails.
1287		*/
1288	<	final String nextWorkerName() {
1289	<	return workerNamePrefix.concat
1290	<	(Integer.toString(nextWorkerNumber.addAndGet(1)));
1288	>	private void releasePlock(int ps) {
1289	>	plock = ps;
1290	>	synchronized (this) { notifyAll(); }
1291		}
1292
1293		/**
1294	<	* Callback from ForkJoinWorkerThread constructor to establish its
1295	<	* poolIndex and record its WorkQueue. To avoid scanning bias due
1298	<	* to packing entries in front of the workQueues array, we treat
1299	<	* the array as a simple power-of-two hash table using per-thread
1300	<	* seed as hash, expanding as needed.
1301	<	*
1302	<	* @param w the worker's queue
1294	>	* Tries to create and start one worker if fewer than target
1295	>	* parallelism level exist. Adjusts counts etc on failure.
1296		*/
1297	+	private void tryAddWorker() {
1298	+	long c; int u, e;
1299	+	while ((u = (int)((c = ctl) >>> 32)) < 0 &&
1300	+	(u & SHORT_SIGN) != 0 && (e = (int)c) >= 0) {
1301	+	long nc = ((long)(((u + UTC_UNIT) & UTC_MASK) |
1302	+	((u + UAC_UNIT) & UAC_MASK)) << 32) | (long)e;
1303	+	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1304	+	ForkJoinWorkerThreadFactory fac;
1305	+	Throwable ex = null;
1306	+	ForkJoinWorkerThread wt = null;
1307	+	try {
1308	+	if ((fac = factory) != null &&
1309	+	(wt = fac.newThread(this)) != null) {
1310	+	wt.start();
1311	+	break;
1312	+	}
1313	+	} catch (Throwable rex) {
1314	+	ex = rex;
1315	+	}
1316	+	deregisterWorker(wt, ex);
1317	+	break;
1318	+	}
1319	+	}
1320	+	}
1321	+
1322	+	//  Registering and deregistering workers
1323
1324	<	final void registerWorker(WorkQueue w) {
1325	<	Mutex lock = this.lock;
1326	<	lock.lock();
1324	>	/**
1325	>	* Callback from ForkJoinWorkerThread to establish and record its
1326	>	* WorkQueue. To avoid scanning bias due to packing entries in
1327	>	* front of the workQueues array, we treat the array as a simple
1328	>	* power-of-two hash table using per-thread seed as hash,
1329	>	* expanding as needed.
1330	>	*
1331	>	* @param wt the worker thread
1332	>	* @return the worker's queue
1333	>	*/
1334	>	final WorkQueue registerWorker(ForkJoinWorkerThread wt) {
1335	>	UncaughtExceptionHandler handler; WorkQueue[] ws; int s, ps;
1336	>	wt.setDaemon(true);
1337	>	if ((handler = ueh) != null)
1338	>	wt.setUncaughtExceptionHandler(handler);
1339	>	do {} while (!U.compareAndSwapInt(this, INDEXSEED, s = indexSeed,
1340	>	s += SEED_INCREMENT) ||
1341	>	s == 0); // skip 0
1342	>	WorkQueue w = new WorkQueue(this, wt, mode, s);
1343	>	if (((ps = plock) & PL_LOCK) != 0 ||
1344	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1345	>	ps = acquirePlock();
1346	>	int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1347		try {
1348	<	WorkQueue[] ws = workQueues;
1349	<	if (w != null && ws != null) {          // skip on shutdown/failure
1350	<	int rs, n = ws.length, m = n - 1;
1351	<	int s = nextSeed += SEED_INCREMENT; // rarely-colliding sequence
1352	<	w.seed = (s == 0) ? 1 : s;          // ensure non-zero seed
1353	<	int r = (s << 1) | 1;               // use odd-numbered indices
1315	<	if (ws[r &= m] != null) {           // collision
1316	<	int probes = 0;                 // step by approx half size
1317	<	int step = (n <= 4) ? 2 : ((n >>> 1) & SQMASK) + 2;
1348	>	if ((ws = workQueues) != null) {    // skip if shutting down
1349	>	int n = ws.length, m = n - 1;
1350	>	int r = (s << 1) | 1;           // use odd-numbered indices
1351	>	if (ws[r &= m] != null) {       // collision
1352	>	int probes = 0;             // step by approx half size
1353	>	int step = (n <= 4) ? 2 : ((n >>> 1) & EVENMASK) + 2;
1354		while (ws[r = (r + step) & m] != null) {
1355		if (++probes >= n) {
1356		workQueues = ws = Arrays.copyOf(ws, n <<= 1);
#	Line 1323 | Line 1359 | public class ForkJoinPool extends Abstra
1359		}
1360		}
1361		}
1362	<	w.eventCount = w.poolIndex = r;     // establish before recording
1363	<	ws[r] = w;                          // also update seq
1364	<	runState = ((rs = runState) & SHUTDOWN) | ((rs + 2) & ~SHUTDOWN);
1362	>	w.poolIndex = (short)r;
1363	>	w.eventCount = r; // volatile write orders
1364	>	ws[r] = w;
1365		}
1366		} finally {
1367	<	lock.unlock();
1367	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1368	>	releasePlock(nps);
1369		}
1370	+	wt.setName(workerNamePrefix.concat(Integer.toString(w.poolIndex >>> 1)));
1371	+	return w;
1372		}
1373
1374		/**
1375		* Final callback from terminating worker, as well as upon failure
1376	<	* to construct or start a worker in addWorker.  Removes record of
1377	<	* worker from array, and adjusts counts. If pool is shutting
1378	<	* down, tries to complete termination.
1376	>	* to construct or start a worker.  Removes record of worker from
1377	>	* array, and adjusts counts. If pool is shutting down, tries to
1378	>	* complete termination.
1379		*
1380	<	* @param wt the worker thread or null if addWorker failed
1380	>	* @param wt the worker thread, or null if construction failed
1381		* @param ex the exception causing failure, or null if none
1382		*/
1383		final void deregisterWorker(ForkJoinWorkerThread wt, Throwable ex) {
1345	–	Mutex lock = this.lock;
1384		WorkQueue w = null;
1385		if (wt != null && (w = wt.workQueue) != null) {
1386	<	w.runState = -1;                // ensure runState is set
1387	<	stealCount.getAndAdd(w.totalSteals + w.nsteals);
1388	<	int idx = w.poolIndex;
1389	<	lock.lock();
1390	<	try {                           // remove record from array
1386	>	int ps; long sc;
1387	>	w.qlock = -1;                // ensure set
1388	>	do {} while (!U.compareAndSwapLong(this, STEALCOUNT,
1389	>	sc = stealCount,
1390	>	sc + w.nsteals));
1391	>	if (((ps = plock) & PL_LOCK) != 0 ||
1392	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1393	>	ps = acquirePlock();
1394	>	int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1395	>	try {
1396	>	int idx = w.poolIndex;
1397		WorkQueue[] ws = workQueues;
1398		if (ws != null && idx >= 0 && idx < ws.length && ws[idx] == w)
1399		ws[idx] = null;
1400		} finally {
1401	<	lock.unlock();
1401	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1402	>	releasePlock(nps);
1403		}
1404		}
1405
1406	<	long c;                             // adjust ctl counts
1406	>	long c;                          // adjust ctl counts
1407		do {} while (!U.compareAndSwapLong
1408		(this, CTL, c = ctl, (((c - AC_UNIT) & AC_MASK) |
1409		((c - TC_UNIT) & TC_MASK) |
1410		(c & ~(AC_MASK|TC_MASK)))));
1411
1412	<	if (!tryTerminate(false, false) && w != null) {
1413	<	w.cancelAll();                  // cancel remaining tasks
1414	<	if (w.array != null)            // suppress signal if never ran
1415	<	signalWork();               // wake up or create replacement
1416	<	if (ex == null)                 // help clean refs on way out
1417	<	ForkJoinTask.helpExpungeStaleExceptions();
1412	>	if (!tryTerminate(false, false) && w != null && w.array != null) {
1413	>	w.cancelAll();               // cancel remaining tasks
1414	>	WorkQueue[] ws; WorkQueue v; Thread p; int u, i, e;
1415	>	while ((u = (int)((c = ctl) >>> 32)) < 0 && (e = (int)c) >= 0) {
1416	>	if (e > 0) {             // activate or create replacement
1417	>	if ((ws = workQueues) == null ||
1418	>	(i = e & SMASK) >= ws.length ||
1419	>	(v = ws[i]) == null)
1420	>	break;
1421	>	long nc = (((long)(v.nextWait & E_MASK)) |
1422	>	((long)(u + UAC_UNIT) << 32));
1423	>	if (v.eventCount != (e | INT_SIGN))
1424	>	break;
1425	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1426	>	v.eventCount = (e + E_SEQ) & E_MASK;
1427	>	if ((p = v.parker) != null)
1428	>	U.unpark(p);
1429	>	break;
1430	>	}
1431	>	}
1432	>	else {
1433	>	if ((short)u < 0)
1434	>	tryAddWorker();
1435	>	break;
1436	>	}
1437	>	}
1438		}
1439	<
1440	<	if (ex != null)                     // rethrow
1441	<	U.throwException(ex);
1439	>	if (ex == null)                     // help clean refs on way out
1440	>	ForkJoinTask.helpExpungeStaleExceptions();
1441	>	else                                // rethrow
1442	>	ForkJoinTask.rethrow(ex);
1443		}
1444
1379	–
1445		// Submissions
1446
1447		/**
1448	+	* Per-thread records for threads that submit to pools. Currently
1449	+	* holds only pseudo-random seed / index that is used to choose
1450	+	* submission queues in method externalPush. In the future, this may
1451	+	* also incorporate a means to implement different task rejection
1452	+	* and resubmission policies.
1453	+	*
1454	+	* Seeds for submitters and workers/workQueues work in basically
1455	+	* the same way but are initialized and updated using slightly
1456	+	* different mechanics. Both are initialized using the same
1457	+	* approach as in class ThreadLocal, where successive values are
1458	+	* unlikely to collide with previous values. Seeds are then
1459	+	* randomly modified upon collisions using xorshifts, which
1460	+	* requires a non-zero seed.
1461	+	*/
1462	+	static final class Submitter {
1463	+	int seed;
1464	+	Submitter(int s) { seed = s; }
1465	+	}
1466	+
1467	+	/**
1468		* Unless shutting down, adds the given task to a submission queue
1469		* at submitter's current queue index (modulo submission
1470	<	* range). If no queue exists at the index, one is created.  If
1471	<	* the queue is busy, another index is randomly chosen. The
1387	<	* submitMask bounds the effective number of queues to the
1388	<	* (nearest power of two for) parallelism level.
1470	>	* range). Only the most common path is directly handled in this
1471	>	* method. All others are relayed to fullExternalPush.
1472		*
1473		* @param task the task. Caller must ensure non-null.
1474		*/
1475	<	private void doSubmit(ForkJoinTask<?> task) {
1476	<	Submitter s = submitters.get();
1477	<	for (int r = s.seed, m = submitMask;;) {
1478	<	WorkQueue[] ws; WorkQueue q;
1479	<	int k = r & m & SQMASK;          // use only even indices
1480	<	if (runState < 0 || (ws = workQueues) == null || ws.length <= k)
1481	<	throw new RejectedExecutionException(); // shutting down
1482	<	else if ((q = ws[k]) == null) {  // create new queue
1483	<	WorkQueue nq = new WorkQueue(this, null, SHARED_QUEUE);
1484	<	Mutex lock = this.lock;      // construct outside lock
1485	<	lock.lock();
1486	<	try {                        // recheck under lock
1487	<	int rs = runState;       // to update seq
1488	<	if (ws == workQueues && ws[k] == null) {
1489	<	ws[k] = nq;
1490	<	runState = ((rs & SHUTDOWN) | ((rs + 2) & ~SHUTDOWN));
1408	<	}
1409	<	} finally {
1410	<	lock.unlock();
1411	<	}
1412	<	}
1413	<	else if (q.trySharedPush(task)) {
1414	<	signalWork();
1475	>	final void externalPush(ForkJoinTask<?> task) {
1476	>	Submitter z = submitters.get();
1477	>	WorkQueue q; int r, m, s, n, am; ForkJoinTask<?>[] a;
1478	>	int ps = plock;
1479	>	WorkQueue[] ws = workQueues;
1480	>	if (z != null && ps > 0 && ws != null && (m = (ws.length - 1)) >= 0 &&
1481	>	(q = ws[m & (r = z.seed) & SQMASK]) != null && r != 0 &&
1482	>	U.compareAndSwapInt(q, QLOCK, 0, 1)) { // lock
1483	>	if ((a = q.array) != null &&
1484	>	(am = a.length - 1) > (n = (s = q.top) - q.base)) {
1485	>	int j = ((am & s) << ASHIFT) + ABASE;
1486	>	U.putOrderedObject(a, j, task);
1487	>	q.top = s + 1;                     // push on to deque
1488	>	q.qlock = 0;
1489	>	if (n <= 1)
1490	>	signalWork(ws, q);
1491		return;
1492		}
1493	<	else if (m > 1) {                // move to a different index
1494	<	r ^= r << 13;                // same xorshift as WorkQueues
1493	>	q.qlock = 0;
1494	>	}
1495	>	fullExternalPush(task);
1496	>	}
1497	>
1498	>	/**
1499	>	* Full version of externalPush. This method is called, among
1500	>	* other times, upon the first submission of the first task to the
1501	>	* pool, so must perform secondary initialization.  It also
1502	>	* detects first submission by an external thread by looking up
1503	>	* its ThreadLocal, and creates a new shared queue if the one at
1504	>	* index if empty or contended. The plock lock body must be
1505	>	* exception-free (so no try/finally) so we optimistically
1506	>	* allocate new queues outside the lock and throw them away if
1507	>	* (very rarely) not needed.
1508	>	*
1509	>	* Secondary initialization occurs when plock is zero, to create
1510	>	* workQueue array and set plock to a valid value.  This lock body
1511	>	* must also be exception-free. Because the plock seq value can
1512	>	* eventually wrap around zero, this method harmlessly fails to
1513	>	* reinitialize if workQueues exists, while still advancing plock.
1514	>	*/
1515	>	private void fullExternalPush(ForkJoinTask<?> task) {
1516	>	int r = 0; // random index seed
1517	>	for (Submitter z = submitters.get();;) {
1518	>	WorkQueue[] ws; WorkQueue q; int ps, m, k;
1519	>	if (z == null) {
1520	>	if (U.compareAndSwapInt(this, INDEXSEED, r = indexSeed,
1521	>	r += SEED_INCREMENT) && r != 0)
1522	>	submitters.set(z = new Submitter(r));
1523	>	}
1524	>	else if (r == 0) {                  // move to a different index
1525	>	r = z.seed;
1526	>	r ^= r << 13;                   // same xorshift as WorkQueues
1527		r ^= r >>> 17;
1528	<	s.seed = r ^= r << 5;
1528	>	z.seed = r ^= (r << 5);
1529	>	}
1530	>	if ((ps = plock) < 0)
1531	>	throw new RejectedExecutionException();
1532	>	else if (ps == 0 || (ws = workQueues) == null ||
1533	>	(m = ws.length - 1) < 0) { // initialize workQueues
1534	>	int p = parallelism;            // find power of two table size
1535	>	int n = (p > 1) ? p - 1 : 1;    // ensure at least 2 slots
1536	>	n |= n >>> 1;
1537	>	n |= n >>> 2;
1538	>	n |= n >>> 4;
1539	>	n |= n >>> 8;
1540	>	n |= n >>> 16;
1541	>	n = (n + 1) << 1;
1542	>	WorkQueue[] nws = ((ws = workQueues) == null || ws.length == 0 ?
1543	>	new WorkQueue[n] : null);
1544	>	if (((ps = plock) & PL_LOCK) != 0 ||
1545	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1546	>	ps = acquirePlock();
1547	>	if (((ws = workQueues) == null || ws.length == 0) && nws != null)
1548	>	workQueues = nws;
1549	>	int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1550	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1551	>	releasePlock(nps);
1552	>	}
1553	>	else if ((q = ws[k = r & m & SQMASK]) != null) {
1554	>	if (q.qlock == 0 && U.compareAndSwapInt(q, QLOCK, 0, 1)) {
1555	>	ForkJoinTask<?>[] a = q.array;
1556	>	int s = q.top;
1557	>	boolean submitted = false;
1558	>	try {                      // locked version of push
1559	>	if ((a != null && a.length > s + 1 - q.base) ||
1560	>	(a = q.growArray()) != null) {   // must presize
1561	>	int j = (((a.length - 1) & s) << ASHIFT) + ABASE;
1562	>	U.putOrderedObject(a, j, task);
1563	>	q.top = s + 1;
1564	>	submitted = true;
1565	>	}
1566	>	} finally {
1567	>	q.qlock = 0;  // unlock
1568	>	}
1569	>	if (submitted) {
1570	>	signalWork(ws, q);
1571	>	return;
1572	>	}
1573	>	}
1574	>	r = 0; // move on failure
1575	>	}
1576	>	else if (((ps = plock) & PL_LOCK) == 0) { // create new queue
1577	>	q = new WorkQueue(this, null, SHARED_QUEUE, r);
1578	>	q.poolIndex = (short)k;
1579	>	if (((ps = plock) & PL_LOCK) != 0 ||
1580	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1581	>	ps = acquirePlock();
1582	>	if ((ws = workQueues) != null && k < ws.length && ws[k] == null)
1583	>	ws[k] = q;
1584	>	int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1585	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1586	>	releasePlock(nps);
1587		}
1588		else
1589	<	Thread.yield();              // yield if no alternatives
1589	>	r = 0;
1590		}
1591		}
1592
#	Line 1431 | Line 1597 | public class ForkJoinPool extends Abstra
1597		*/
1598		final void incrementActiveCount() {
1599		long c;
1600	<	do {} while (!U.compareAndSwapLong(this, CTL, c = ctl, c + AC_UNIT));
1600	>	do {} while (!U.compareAndSwapLong
1601	>	(this, CTL, c = ctl, ((c & ~AC_MASK) |
1602	>	((c & AC_MASK) + AC_UNIT))));
1603		}
1604
1605		/**
1606	<	* Tries to activate or create a worker if too few are active.
1606	>	* Tries to create or activate a worker if too few are active.
1607	>	*
1608	>	* @param ws the worker array to use to find signallees
1609	>	* @param q if non-null, the queue holding tasks to be processed
1610		*/
1611	<	final void signalWork() {
1612	<	long c; int u;
1613	<	while ((u = (int)((c = ctl) >>> 32)) < 0) {     // too few active
1614	<	WorkQueue[] ws = workQueues; int e, i; WorkQueue w; Thread p;
1615	<	if ((e = (int)c) > 0) {                     // at least one waiting
1616	<	if (ws != null && (i = e & SMASK) < ws.length &&
1617	<	(w = ws[i]) != null && w.eventCount == (e | INT_SIGN)) {
1618	<	long nc = (((long)(w.nextWait & E_MASK)) |
1619	<	((long)(u + UAC_UNIT) << 32));
1449	<	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1450	<	w.eventCount = (e + E_SEQ) & E_MASK;
1451	<	if ((p = w.parker) != null)
1452	<	U.unpark(p);                // activate and release
1453	<	break;
1454	<	}
1455	<	}
1456	<	else
1457	<	break;
1611	>	final void signalWork(WorkQueue[] ws, WorkQueue q) {
1612	>	for (;;) {
1613	>	long c; int e, u, i; WorkQueue w; Thread p;
1614	>	if ((u = (int)((c = ctl) >>> 32)) >= 0)
1615	>	break;
1616	>	if ((e = (int)c) <= 0) {
1617	>	if ((short)u < 0)
1618	>	tryAddWorker();
1619	>	break;
1620		}
1621	<	else if (e == 0 && (u & SHORT_SIGN) != 0) { // too few total
1622	<	long nc = (long)(((u + UTC_UNIT) & UTC_MASK) |
1623	<	((u + UAC_UNIT) & UAC_MASK)) << 32;
1624	<	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1625	<	addWorker();
1626	<	break;
1627	<	}
1621	>	if (ws == null || ws.length <= (i = e & SMASK) ||
1622	>	(w = ws[i]) == null)
1623	>	break;
1624	>	long nc = (((long)(w.nextWait & E_MASK)) |
1625	>	((long)(u + UAC_UNIT)) << 32);
1626	>	int ne = (e + E_SEQ) & E_MASK;
1627	>	if (w.eventCount == (e | INT_SIGN) &&
1628	>	U.compareAndSwapLong(this, CTL, c, nc)) {
1629	>	w.eventCount = ne;
1630	>	if ((p = w.parker) != null)
1631	>	U.unpark(p);
1632	>	break;
1633		}
1634	<	else
1634	>	if (q != null && q.base >= q.top)
1635		break;
1636		}
1637		}
#	Line 1475 | Line 1642 | public class ForkJoinPool extends Abstra
1642		* Top-level runloop for workers, called by ForkJoinWorkerThread.run.
1643		*/
1644		final void runWorker(WorkQueue w) {
1645	<	w.growArray(false);         // initialize queue array in this thread
1646	<	do { w.runTask(scan(w)); } while (w.runState >= 0);
1645	>	w.growArray(); // allocate queue
1646	>	for (int r = w.hint; scan(w, r) == 0; ) {
1647	>	r ^= r << 13; r ^= r >>> 17; r ^= r << 5; // xorshift
1648	>	}
1649		}
1650
1651		/**
1652	<	* Scans for and, if found, returns one task, else possibly
1652	>	* Scans for and, if found, runs one task, else possibly
1653		* inactivates the worker. This method operates on single reads of
1654		* volatile state and is designed to be re-invoked continuously,
1655		* in part because it returns upon detecting inconsistencies,
1656		* contention, or state changes that indicate possible success on
1657		* re-invocation.
1658		*
1659	<	* The scan searches for tasks across a random permutation of
1660	<	* queues (starting at a random index and stepping by a random
1661	<	* relative prime, checking each at least once).  The scan
1662	<	* terminates upon either finding a non-empty queue, or completing
1663	<	* the sweep. If the worker is not inactivated, it takes and
1664	<	* returns a task from this queue.  On failure to find a task, we
1665	<	* take one of the following actions, after which the caller will
1666	<	* retry calling this method unless terminated.
1667	<	*
1499	<	* * If pool is terminating, terminate the worker.
1500	<	*
1501	<	* * If not a complete sweep, try to release a waiting worker.  If
1502	<	* the scan terminated because the worker is inactivated, then the
1503	<	* released worker will often be the calling worker, and it can
1504	<	* succeed obtaining a task on the next call. Or maybe it is
1505	<	* another worker, but with same net effect. Releasing in other
1506	<	* cases as well ensures that we have enough workers running.
1507	<	*
1508	<	* * If not already enqueued, try to inactivate and enqueue the
1509	<	* worker on wait queue. Or, if inactivating has caused the pool
1510	<	* to be quiescent, relay to idleAwaitWork to check for
1511	<	* termination and possibly shrink pool.
1512	<	*
1513	<	* * If already inactive, and the caller has run a task since the
1514	<	* last empty scan, return (to allow rescan) unless others are
1515	<	* also inactivated.  Field WorkQueue.rescans counts down on each
1516	<	* scan to ensure eventual inactivation and blocking.
1517	<	*
1518	<	* * If already enqueued and none of the above apply, park
1519	<	* awaiting signal,
1659	>	* The scan searches for tasks across queues starting at a random
1660	>	* index, checking each at least twice.  The scan terminates upon
1661	>	* either finding a non-empty queue, or completing the sweep. If
1662	>	* the worker is not inactivated, it takes and runs a task from
1663	>	* this queue. Otherwise, if not activated, it tries to activate
1664	>	* itself or some other worker by signalling. On failure to find a
1665	>	* task, returns (for retry) if pool state may have changed during
1666	>	* an empty scan, or tries to inactivate if active, else possibly
1667	>	* blocks or terminates via method awaitWork.
1668		*
1669		* @param w the worker (via its WorkQueue)
1670	<	* @return a task or null if none found
1670	>	* @param r a random seed
1671	>	* @return worker qlock status if would have waited, else 0
1672		*/
1673	<	private final ForkJoinTask<?> scan(WorkQueue w) {
1674	<	WorkQueue[] ws;                       // first update random seed
1675	<	int r = w.seed; r ^= r << 13; r ^= r >>> 17; w.seed = r ^= r << 5;
1676	<	int rs = runState, m;                 // volatile read order matters
1677	<	if ((ws = workQueues) != null && (m = ws.length - 1) > 0) {
1678	<	int ec = w.eventCount;            // ec is negative if inactive
1679	<	int step = (r >>> 16) | 1;        // relative prime
1680	<	for (int j = (m + 1) << 2; ; r += step) {
1681	<	WorkQueue q; ForkJoinTask<?> t; ForkJoinTask<?>[] a; int b;
1682	<	if ((q = ws[r & m]) != null && (b = q.base) - q.top < 0 &&
1683	<	(a = q.array) != null) {  // probably nonempty
1684	<	int i = (((a.length - 1) & b) << ASHIFT) + ABASE;
1685	<	t = (ForkJoinTask<?>)U.getObjectVolatile(a, i);
1686	<	if (q.base == b && ec >= 0 && t != null &&
1687	<	U.compareAndSwapObject(a, i, t, null)) {
1688	<	if (q.top - (q.base = b + 1) > 1)
1689	<	signalWork();    // help pushes signal
1690	<	return t;
1691	<	}
1692	<	else if (ec < 0 || j <= m) {
1544	<	rs = 0;               // mark scan as imcomplete
1545	<	break;                // caller can retry after release
1673	>	private final int scan(WorkQueue w, int r) {
1674	>	WorkQueue[] ws; int m;
1675	>	long c = ctl;                            // for consistency check
1676	>	if ((ws = workQueues) != null && (m = ws.length - 1) >= 0 && w != null) {
1677	>	for (int j = m + m + 1, ec = w.eventCount;;) {
1678	>	WorkQueue q; int b, e; ForkJoinTask<?>[] a; ForkJoinTask<?> t;
1679	>	if ((q = ws[(r - j) & m]) != null &&
1680	>	(b = q.base) - q.top < 0 && (a = q.array) != null) {
1681	>	long i = (((a.length - 1) & b) << ASHIFT) + ABASE;
1682	>	if ((t = ((ForkJoinTask<?>)
1683	>	U.getObjectVolatile(a, i))) != null) {
1684	>	if (ec < 0)
1685	>	helpRelease(c, ws, w, q, b);
1686	>	else if (q.base == b &&
1687	>	U.compareAndSwapObject(a, i, t, null)) {
1688	>	U.putOrderedInt(q, QBASE, b + 1);
1689	>	if ((b + 1) - q.top < 0)
1690	>	signalWork(ws, q);
1691	>	w.runTask(t);
1692	>	}
1693		}
1547	–	}
1548	–	if (--j < 0)
1694		break;
1695	<	}
1696	<
1697	<	long c = ctl; int e = (int)c, a = (int)(c >> AC_SHIFT), nr, ns;
1698	<	if (e < 0)                        // decode ctl on empty scan
1699	<	w.runState = -1;              // pool is terminating
1700	<	else if (rs == 0 || rs != runState) { // incomplete scan
1701	<	WorkQueue v; Thread p;        // try to release a waiter
1702	<	if (e > 0 && a < 0 && w.eventCount == ec &&
1703	<	(v = ws[e & m]) != null && v.eventCount == (e | INT_SIGN)) {
1704	<	long nc = ((long)(v.nextWait & E_MASK) |
1560	<	((c + AC_UNIT) & (AC_MASK|TC_MASK)));
1561	<	if (ctl == c && U.compareAndSwapLong(this, CTL, c, nc)) {
1562	<	v.eventCount = (e + E_SEQ) & E_MASK;
1563	<	if ((p = v.parker) != null)
1564	<	U.unpark(p);
1695	>	}
1696	>	else if (--j < 0) {
1697	>	if ((ec | (e = (int)c)) < 0) // inactive or terminating
1698	>	return awaitWork(w, c, ec);
1699	>	else if (ctl == c) {         // try to inactivate and enqueue
1700	>	long nc = (long)ec | ((c - AC_UNIT) & (AC_MASK|TC_MASK));
1701	>	w.nextWait = e;
1702	>	w.eventCount = ec | INT_SIGN;
1703	>	if (!U.compareAndSwapLong(this, CTL, c, nc))
1704	>	w.eventCount = ec;   // back out
1705		}
1706	+	break;
1707		}
1708		}
1709	<	else if (ec >= 0) {               // try to enqueue/inactivate
1710	<	long nc = (long)ec | ((c - AC_UNIT) & (AC_MASK|TC_MASK));
1711	<	w.nextWait = e;
1712	<	w.eventCount = ec | INT_SIGN; // mark as inactive
1713	<	if (ctl != c || !U.compareAndSwapLong(this, CTL, c, nc))
1714	<	w.eventCount = ec;        // unmark on CAS failure
1715	<	else {
1716	<	if ((ns = w.nsteals) != 0) {
1717	<	w.nsteals = 0;        // set rescans if ran task
1718	<	w.rescans = (a > 0) ? 0 : a + parallelism;
1719	<	w.totalSteals += ns;
1720	<	}
1721	<	if (a == 1 - parallelism) // quiescent
1722	<	idleAwaitWork(w, nc, c);
1723	<	}
1709	>	}
1710	>	return 0;
1711	>	}
1712	>
1713	>	/**
1714	>	* A continuation of scan(), possibly blocking or terminating
1715	>	* worker w. Returns without blocking if pool state has apparently
1716	>	* changed since last invocation.  Also, if inactivating w has
1717	>	* caused the pool to become quiescent, checks for pool
1718	>	* termination, and, so long as this is not the only worker, waits
1719	>	* for event for up to a given duration.  On timeout, if ctl has
1720	>	* not changed, terminates the worker, which will in turn wake up
1721	>	* another worker to possibly repeat this process.
1722	>	*
1723	>	* @param w the calling worker
1724	>	* @param c the ctl value on entry to scan
1725	>	* @param ec the worker's eventCount on entry to scan
1726	>	*/
1727	>	private final int awaitWork(WorkQueue w, long c, int ec) {
1728	>	int stat, ns; long parkTime, deadline;
1729	>	if ((stat = w.qlock) >= 0 && w.eventCount == ec && ctl == c &&
1730	>	!Thread.interrupted()) {
1731	>	int e = (int)c;
1732	>	int u = (int)(c >>> 32);
1733	>	int d = (u >> UAC_SHIFT) + parallelism; // active count
1734	>
1735	>	if (e < 0 || (d <= 0 && tryTerminate(false, false)))
1736	>	stat = w.qlock = -1;          // pool is terminating
1737	>	else if ((ns = w.nsteals) != 0) { // collect steals and retry
1738	>	long sc;
1739	>	w.nsteals = 0;
1740	>	do {} while (!U.compareAndSwapLong(this, STEALCOUNT,
1741	>	sc = stealCount, sc + ns));
1742		}
1743	<	else if (w.eventCount < 0) {      // already queued
1744	<	if ((nr = w.rescans) > 0) {   // continue rescanning
1745	<	int ac = a + parallelism;
1746	<	if (((w.rescans = (ac < nr) ? ac : nr - 1) & 3) == 0)
1747	<	Thread.yield();       // yield before block
1743	>	else {
1744	>	long pc = ((d > 0 || ec != (e | INT_SIGN)) ? 0L :
1745	>	((long)(w.nextWait & E_MASK)) | // ctl to restore
1746	>	((long)(u + UAC_UNIT)) << 32);
1747	>	if (pc != 0L) {               // timed wait if last waiter
1748	>	int dc = -(short)(c >>> TC_SHIFT);
1749	>	parkTime = (dc < 0 ? FAST_IDLE_TIMEOUT:
1750	>	(dc + 1) * IDLE_TIMEOUT);
1751	>	deadline = System.nanoTime() + parkTime - TIMEOUT_SLOP;
1752		}
1753	<	else {
1754	<	Thread.interrupted();     // clear status
1753	>	else
1754	>	parkTime = deadline = 0L;
1755	>	if (w.eventCount == ec && ctl == c) {
1756		Thread wt = Thread.currentThread();
1757		U.putObject(wt, PARKBLOCKER, this);
1758		w.parker = wt;            // emulate LockSupport.park
1759	<	if (w.eventCount < 0)     // recheck
1760	<	U.park(false, 0L);
1759	>	if (w.eventCount == ec && ctl == c)
1760	>	U.park(false, parkTime);  // must recheck before park
1761		w.parker = null;
1762		U.putObject(wt, PARKBLOCKER, null);
1763	+	if (parkTime != 0L && ctl == c &&
1764	+	deadline - System.nanoTime() <= 0L &&
1765	+	U.compareAndSwapLong(this, CTL, c, pc))
1766	+	stat = w.qlock = -1;  // shrink pool
1767		}
1768		}
1769		}
1770	<	return null;
1770	>	return stat;
1771		}
1772
1773		/**
1774	<	* If inactivating worker w has caused the pool to become
1775	<	* quiescent, checks for pool termination, and, so long as this is
1776	<	* not the only worker, waits for event for up to SHRINK_RATE
1777	<	* nanosecs.  On timeout, if ctl has not changed, terminates the
1778	<	* worker, which will in turn wake up another worker to possibly
1779	<	* repeat this process.
1780	<	*
1781	<	* @param w the calling worker
1782	<	* @param currentCtl the ctl value triggering possible quiescence
1783	<	* @param prevCtl the ctl value to restore if thread is terminated
1784	<	*/
1785	<	private void idleAwaitWork(WorkQueue w, long currentCtl, long prevCtl) {
1786	<	if (w.eventCount < 0 && !tryTerminate(false, false) &&
1787	<	(int)prevCtl != 0 && !hasQueuedSubmissions() && ctl == currentCtl) {
1788	<	Thread wt = Thread.currentThread();
1789	<	Thread.yield();            // yield before block
1790	<	while (ctl == currentCtl) {
1791	<	long startTime = System.nanoTime();
1792	<	Thread.interrupted();  // timed variant of version in scan()
1793	<	U.putObject(wt, PARKBLOCKER, this);
1626	<	w.parker = wt;
1627	<	if (ctl == currentCtl)
1628	<	U.park(false, SHRINK_RATE);
1629	<	w.parker = null;
1630	<	U.putObject(wt, PARKBLOCKER, null);
1631	<	if (ctl != currentCtl)
1632	<	break;
1633	<	if (System.nanoTime() - startTime >= SHRINK_TIMEOUT &&
1634	<	U.compareAndSwapLong(this, CTL, currentCtl, prevCtl)) {
1635	<	w.eventCount = (w.eventCount + E_SEQ) | E_MASK;
1636	<	w.runState = -1;   // shrink
1637	<	break;
1638	<	}
1774	>	* Possibly releases (signals) a worker. Called only from scan()
1775	>	* when a worker with apparently inactive status finds a non-empty
1776	>	* queue. This requires revalidating all of the associated state
1777	>	* from caller.
1778	>	*/
1779	>	private final void helpRelease(long c, WorkQueue[] ws, WorkQueue w,
1780	>	WorkQueue q, int b) {
1781	>	WorkQueue v; int e, i; Thread p;
1782	>	if (w != null && w.eventCount < 0 && (e = (int)c) > 0 &&
1783	>	ws != null && ws.length > (i = e & SMASK) &&
1784	>	(v = ws[i]) != null && ctl == c) {
1785	>	long nc = (((long)(v.nextWait & E_MASK)) |
1786	>	((long)((int)(c >>> 32) + UAC_UNIT)) << 32);
1787	>	int ne = (e + E_SEQ) & E_MASK;
1788	>	if (q != null && q.base == b && w.eventCount < 0 &&
1789	>	v.eventCount == (e | INT_SIGN) &&
1790	>	U.compareAndSwapLong(this, CTL, c, nc)) {
1791	>	v.eventCount = ne;
1792	>	if ((p = v.parker) != null)
1793	>	U.unpark(p);
1794		}
1795		}
1796		}
#	Line 1660 | Line 1815 | public class ForkJoinPool extends Abstra
1815		*/
1816		private int tryHelpStealer(WorkQueue joiner, ForkJoinTask<?> task) {
1817		int stat = 0, steps = 0;                    // bound to avoid cycles
1818	<	if (joiner != null && task != null) {       // hoist null checks
1818	>	if (task != null && joiner != null &&
1819	>	joiner.base - joiner.top >= 0) {        // hoist checks
1820		restart: for (;;) {
1821		ForkJoinTask<?> subtask = task;     // current target
1822		for (WorkQueue j = joiner, v;;) {   // v is stealer of subtask
#	Line 1671 | Line 1827 | public class ForkJoinPool extends Abstra
1827		}
1828		if ((ws = workQueues) == null || (m = ws.length - 1) <= 0)
1829		break restart;              // shutting down
1830	<	if ((v = ws[h = (j.stealHint | 1) & m]) == null ||
1830	>	if ((v = ws[h = (j.hint | 1) & m]) == null ||
1831		v.currentSteal != subtask) {
1832		for (int origin = h;;) {    // find stealer
1833		if (((h = (h + 2) & m) & 15) == 1 &&
#	Line 1679 | Line 1835 | public class ForkJoinPool extends Abstra
1835		continue restart;   // occasional staleness check
1836		if ((v = ws[h]) != null &&
1837		v.currentSteal == subtask) {
1838	<	j.stealHint = h;    // save hint
1838	>	j.hint = h;        // save hint
1839		break;
1840		}
1841		if (h == origin)
#	Line 1687 | Line 1843 | public class ForkJoinPool extends Abstra
1843		}
1844		}
1845		for (;;) { // help stealer or descend to its stealer
1846	<	ForkJoinTask[] a;  int b;
1846	>	ForkJoinTask[] a; int b;
1847		if (subtask.status < 0)     // surround probes with
1848		continue restart;       //   consistency checks
1849		if ((b = v.base) - v.top < 0 && (a = v.array) != null) {
#	Line 1698 | Line 1854 | public class ForkJoinPool extends Abstra
1854		v.currentSteal != subtask)
1855		continue restart;   // stale
1856		stat = 1;               // apparent progress
1857	<	if (t != null && v.base == b &&
1858	<	U.compareAndSwapObject(a, i, t, null)) {
1859	<	v.base = b + 1;     // help stealer
1860	<	joiner.runSubtask(t);
1857	>	if (v.base == b) {
1858	>	if (t == null)
1859	>	break restart;
1860	>	if (U.compareAndSwapObject(a, i, t, null)) {
1861	>	U.putOrderedInt(v, QBASE, b + 1);
1862	>	ForkJoinTask<?> ps = joiner.currentSteal;
1863	>	int jt = joiner.top;
1864	>	do {
1865	>	joiner.currentSteal = t;
1866	>	t.doExec(); // clear local tasks too
1867	>	} while (task.status >= 0 &&
1868	>	joiner.top != jt &&
1869	>	(t = joiner.pop()) != null);
1870	>	joiner.currentSteal = ps;
1871	>	break restart;
1872	>	}
1873		}
1706	–	else if (v.base == b && ++steps == MAX_HELP)
1707	–	break restart;      // v apparently stalled
1874		}
1875		else {                      // empty -- try to descend
1876		ForkJoinTask<?> next = v.currentJoin;
#	Line 1727 | Line 1893 | public class ForkJoinPool extends Abstra
1893		}
1894
1895		/**
1896	<	* If task is at base of some steal queue, steals and executes it.
1896	>	* Analog of tryHelpStealer for CountedCompleters. Tries to steal
1897	>	* and run tasks within the target's computation.
1898		*
1899	<	* @param joiner the joining worker
1733	<	* @param task the task
1899	>	* @param task the task to join
1900		*/
1901	<	private void tryPollForAndExec(WorkQueue joiner, ForkJoinTask<?> task) {
1902	<	WorkQueue[] ws;
1903	<	if ((ws = workQueues) != null) {
1904	<	for (int j = 1; j < ws.length && task.status >= 0; j += 2) {
1905	<	WorkQueue q = ws[j];
1906	<	if (q != null && q.pollFor(task)) {
1907	<	joiner.runSubtask(task);
1901	>	private int helpComplete(WorkQueue joiner, CountedCompleter<?> task) {
1902	>	WorkQueue[] ws; int m;
1903	>	int s = 0;
1904	>	if ((ws = workQueues) != null && (m = ws.length - 1) >= 0 &&
1905	>	joiner != null && task != null) {
1906	>	int j = joiner.poolIndex;
1907	>	int scans = m + m + 1;
1908	>	long c = 0L;              // for stability check
1909	>	for (int k = scans; ; j += 2) {
1910	>	WorkQueue q;
1911	>	if ((s = task.status) < 0)
1912	>	break;
1913	>	else if (joiner.internalPopAndExecCC(task))
1914	>	k = scans;
1915	>	else if ((s = task.status) < 0)
1916		break;
1917	+	else if ((q = ws[j & m]) != null && q.pollAndExecCC(task))
1918	+	k = scans;
1919	+	else if (--k < 0) {
1920	+	if (c == (c = ctl))
1921	+	break;
1922	+	k = scans;
1923		}
1924		}
1925		}
1926	+	return s;
1927		}
1928
1929		/**
1930		* Tries to decrement active count (sometimes implicitly) and
1931		* possibly release or create a compensating worker in preparation
1932		* for blocking. Fails on contention or termination. Otherwise,
1933	<	* adds a new thread if no idle workers are available and either
1934	<	* pool would become completely starved or: (at least half
1935	<	* starved, and fewer than 50% spares exist, and there is at least
1936	<	* one task apparently available). Even though the availability
1756	<	* check requires a full scan, it is worthwhile in reducing false
1757	<	* alarms.
1758	<	*
1759	<	* @param task if non-null, a task being waited for
1760	<	* @param blocker if non-null, a blocker being waited for
1761	<	* @return true if the caller can block, else should recheck and retry
1933	>	* adds a new thread if no idle workers are available and pool
1934	>	* may become starved.
1935	>	*
1936	>	* @param c the assumed ctl value
1937		*/
1938	<	final boolean tryCompensate(ForkJoinTask<?> task, ManagedBlocker blocker) {
1764	<	int pc = parallelism, e;
1765	<	long c = ctl;
1938	>	final boolean tryCompensate(long c) {
1939		WorkQueue[] ws = workQueues;
1940	<	if ((e = (int)c) >= 0 && ws != null) {
1941	<	int u, a, ac, hc;
1942	<	int tc = (short)((u = (int)(c >>> 32)) >>> UTC_SHIFT) + pc;
1943	<	boolean replace = false;
1944	<	if ((a = u >> UAC_SHIFT) <= 0) {
1945	<	if ((ac = a + pc) <= 1)
1946	<	replace = true;
1947	<	else if ((e > 0 || (task != null &&
1948	<	ac <= (hc = pc >>> 1) && tc < pc + hc))) {
1949	<	WorkQueue w;
1950	<	for (int j = 0; j < ws.length; ++j) {
1951	<	if ((w = ws[j]) != null && !w.isEmpty()) {
1952	<	replace = true;
1953	<	break;   // in compensation range and tasks available
1954	<	}
1955	<	}
1956	<	}
1940	>	int pc = parallelism, e = (int)c, m, tc;
1941	>	if (ws != null && (m = ws.length - 1) >= 0 && e >= 0 && ctl == c) {
1942	>	WorkQueue w = ws[e & m];
1943	>	if (e != 0 && w != null) {
1944	>	Thread p;
1945	>	long nc = ((long)(w.nextWait & E_MASK) |
1946	>	(c & (AC_MASK|TC_MASK)));
1947	>	int ne = (e + E_SEQ) & E_MASK;
1948	>	if (w.eventCount == (e | INT_SIGN) &&
1949	>	U.compareAndSwapLong(this, CTL, c, nc)) {
1950	>	w.eventCount = ne;
1951	>	if ((p = w.parker) != null)
1952	>	U.unpark(p);
1953	>	return true;   // replace with idle worker
1954	>	}
1955	>	}
1956	>	else if ((tc = (short)(c >>> TC_SHIFT)) >= 0 &&
1957	>	(int)(c >> AC_SHIFT) + pc > 1) {
1958	>	long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK);
1959	>	if (U.compareAndSwapLong(this, CTL, c, nc))
1960	>	return true;   // no compensation
1961		}
1962	<	if ((task == null || task.status >= 0) && // recheck need to block
1963	<	(blocker == null || !blocker.isReleasable()) && ctl == c) {
1964	<	if (!replace) {          // no compensation
1965	<	long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK);
1966	<	if (U.compareAndSwapLong(this, CTL, c, nc))
1967	<	return true;
1968	<	}
1969	<	else if (e != 0) {       // release an idle worker
1970	<	WorkQueue w; Thread p; int i;
1971	<	if ((i = e & SMASK) < ws.length && (w = ws[i]) != null) {
1795	<	long nc = ((long)(w.nextWait & E_MASK) |
1796	<	(c & (AC_MASK|TC_MASK)));
1797	<	if (w.eventCount == (e | INT_SIGN) &&
1798	<	U.compareAndSwapLong(this, CTL, c, nc)) {
1799	<	w.eventCount = (e + E_SEQ) & E_MASK;
1800	<	if ((p = w.parker) != null)
1801	<	U.unpark(p);
1962	>	else if (tc + pc < MAX_CAP) {
1963	>	long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK);
1964	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1965	>	ForkJoinWorkerThreadFactory fac;
1966	>	Throwable ex = null;
1967	>	ForkJoinWorkerThread wt = null;
1968	>	try {
1969	>	if ((fac = factory) != null &&
1970	>	(wt = fac.newThread(this)) != null) {
1971	>	wt.start();
1972		return true;
1973		}
1974	+	} catch (Throwable rex) {
1975	+	ex = rex;
1976		}
1977	<	}
1806	<	else if (tc < MAX_CAP) { // create replacement
1807	<	long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK);
1808	<	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1809	<	addWorker();
1810	<	return true;
1811	<	}
1977	>	deregisterWorker(wt, ex); // clean up and return false
1978		}
1979		}
1980		}
#	Line 1823 | Line 1989 | public class ForkJoinPool extends Abstra
1989		* @return task status on exit
1990		*/
1991		final int awaitJoin(WorkQueue joiner, ForkJoinTask<?> task) {
1992	<	int s;
1993	<	if ((s = task.status) >= 0) {
1992	>	int s = 0;
1993	>	if (task != null && (s = task.status) >= 0 && joiner != null) {
1994		ForkJoinTask<?> prevJoin = joiner.currentJoin;
1995		joiner.currentJoin = task;
1996	<	long startTime = 0L;
1997	<	for (int k = 0;;) {
1998	<	if ((s = (joiner.isEmpty() ?           // try to help
1999	<	tryHelpStealer(joiner, task) :
2000	<	joiner.tryRemoveAndExec(task))) == 0 &&
1996	>	do {} while (joiner.tryRemoveAndExec(task) && // process local tasks
1997	>	(s = task.status) >= 0);
1998	>	if (s >= 0 && (task instanceof CountedCompleter))
1999	>	s = helpComplete(joiner, (CountedCompleter<?>)task);
2000	>	long cc = 0;        // for stability checks
2001	>	while (s >= 0 && (s = task.status) >= 0) {
2002	>	if ((s = tryHelpStealer(joiner, task)) == 0 &&
2003		(s = task.status) >= 0) {
2004	<	if (k == 0) {
2005	<	startTime = System.nanoTime();
2006	<	tryPollForAndExec(joiner, task); // check uncommon case
2007	<	}
1840	<	else if ((k & (MAX_HELP - 1)) == 0 &&
1841	<	System.nanoTime() - startTime >=
1842	<	COMPENSATION_DELAY &&
1843	<	tryCompensate(task, null)) {
1844	<	if (task.trySetSignal()) {
2004	>	if (!tryCompensate(cc))
2005	>	cc = ctl;
2006	>	else {
2007	>	if (task.trySetSignal() && (s = task.status) >= 0) {
2008		synchronized (task) {
2009		if (task.status >= 0) {
2010		try {                // see ForkJoinTask
#	Line 1853 | Line 2016 | public class ForkJoinPool extends Abstra
2016		task.notifyAll();
2017		}
2018		}
2019	<	long c;                          // re-activate
2019	>	long c; // reactivate
2020		do {} while (!U.compareAndSwapLong
2021	<	(this, CTL, c = ctl, c + AC_UNIT));
2021	>	(this, CTL, c = ctl,
2022	>	((c & ~AC_MASK) |
2023	>	((c & AC_MASK) + AC_UNIT))));
2024		}
2025		}
1861	–	if (s < 0 || (s = task.status) < 0) {
1862	–	joiner.currentJoin = prevJoin;
1863	–	break;
1864	–	}
1865	–	else if ((k++ & (MAX_HELP - 1)) == MAX_HELP >>> 1)
1866	–	Thread.yield();                     // for politeness
2026		}
2027	+	joiner.currentJoin = prevJoin;
2028		}
2029		return s;
2030		}
#	Line 1876 | Line 2036 | public class ForkJoinPool extends Abstra
2036		*
2037		* @param joiner the joining worker
2038		* @param task the task
1879	–	* @return task status on exit
2039		*/
2040	<	final int helpJoinOnce(WorkQueue joiner, ForkJoinTask<?> task) {
2040	>	final void helpJoinOnce(WorkQueue joiner, ForkJoinTask<?> task) {
2041		int s;
2042	<	while ((s = task.status) >= 0 &&
2043	<	(joiner.isEmpty() ?
2044	<	tryHelpStealer(joiner, task) :
2045	<	joiner.tryRemoveAndExec(task)) != 0)
2046	<	;
2047	<	return s;
2042	>	if (joiner != null && task != null && (s = task.status) >= 0) {
2043	>	ForkJoinTask<?> prevJoin = joiner.currentJoin;
2044	>	joiner.currentJoin = task;
2045	>	do {} while (joiner.tryRemoveAndExec(task) && // process local tasks
2046	>	(s = task.status) >= 0);
2047	>	if (s >= 0) {
2048	>	if (task instanceof CountedCompleter)
2049	>	helpComplete(joiner, (CountedCompleter<?>)task);
2050	>	do {} while (task.status >= 0 &&
2051	>	tryHelpStealer(joiner, task) > 0);
2052	>	}
2053	>	joiner.currentJoin = prevJoin;
2054	>	}
2055		}
2056
2057		/**
2058		* Returns a (probably) non-empty steal queue, if one is found
2059	<	* during a random, then cyclic scan, else null.  This method must
2060	<	* be retried by caller if, by the time it tries to use the queue,
2061	<	* it is empty.
2062	<	*/
2063	<	private WorkQueue findNonEmptyStealQueue(WorkQueue w) {
2064	<	// Similar to loop in scan(), but ignoring submissions
2065	<	int r = w.seed; r ^= r << 13; r ^= r >>> 17; w.seed = r ^= r << 5;
2066	<	int step = (r >>> 16) | 1;
2067	<	for (WorkQueue[] ws;;) {
2068	<	int rs = runState, m;
2069	<	if ((ws = workQueues) == null || (m = ws.length - 1) < 1)
2070	<	return null;
1905	<	for (int j = (m + 1) << 2; ; r += step) {
1906	<	WorkQueue q = ws[((r << 1) | 1) & m];
1907	<	if (q != null && !q.isEmpty())
1908	<	return q;
1909	<	else if (--j < 0) {
1910	<	if (runState == rs)
1911	<	return null;
1912	<	break;
2059	>	* during a scan, else null.  This method must be retried by
2060	>	* caller if, by the time it tries to use the queue, it is empty.
2061	>	*/
2062	>	private WorkQueue findNonEmptyStealQueue() {
2063	>	int r = ThreadLocalRandom.current().nextInt();
2064	>	for (;;) {
2065	>	int ps = plock, m; WorkQueue[] ws; WorkQueue q;
2066	>	if ((ws = workQueues) != null && (m = ws.length - 1) >= 0) {
2067	>	for (int j = (m + 1) << 2; j >= 0; --j) {
2068	>	if ((q = ws[(((r - j) << 1) | 1) & m]) != null &&
2069	>	q.base - q.top < 0)
2070	>	return q;
2071		}
2072		}
2073	+	if (plock == ps)
2074	+	return null;
2075		}
2076		}
2077
1918	–
2078		/**
2079		* Runs tasks until {@code isQuiescent()}. We piggyback on
2080		* active count ctl maintenance, but rather than blocking
#	Line 1923 | Line 2082 | public class ForkJoinPool extends Abstra
2082		* find tasks either.
2083		*/
2084		final void helpQuiescePool(WorkQueue w) {
2085	+	ForkJoinTask<?> ps = w.currentSteal;
2086		for (boolean active = true;;) {
2087	<	ForkJoinTask<?> localTask; // exhaust local queue
2088	<	while ((localTask = w.nextLocalTask()) != null)
2089	<	localTask.doExec();
2090	<	WorkQueue q = findNonEmptyStealQueue(w);
1931	<	if (q != null) {
1932	<	ForkJoinTask<?> t; int b;
2087	>	long c; WorkQueue q; ForkJoinTask<?> t; int b;
2088	>	while ((t = w.nextLocalTask()) != null)
2089	>	t.doExec();
2090	>	if ((q = findNonEmptyStealQueue()) != null) {
2091		if (!active) {      // re-establish active count
1934	–	long c;
2092		active = true;
2093		do {} while (!U.compareAndSwapLong
2094	<	(this, CTL, c = ctl, c + AC_UNIT));
2094	>	(this, CTL, c = ctl,
2095	>	((c & ~AC_MASK) |
2096	>	((c & AC_MASK) + AC_UNIT))));
2097	>	}
2098	>	if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null) {
2099	>	(w.currentSteal = t).doExec();
2100	>	w.currentSteal = ps;
2101		}
1939	–	if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null)
1940	–	w.runSubtask(t);
2102		}
2103	<	else {
2104	<	long c;
2105	<	if (active) {       // decrement active count without queuing
2103	>	else if (active) {      // decrement active count without queuing
2104	>	long nc = ((c = ctl) & ~AC_MASK) | ((c & AC_MASK) - AC_UNIT);
2105	>	if ((int)(nc >> AC_SHIFT) + parallelism == 0)
2106	>	break;          // bypass decrement-then-increment
2107	>	if (U.compareAndSwapLong(this, CTL, c, nc))
2108		active = false;
1946	–	do {} while (!U.compareAndSwapLong
1947	–	(this, CTL, c = ctl, c -= AC_UNIT));
1948	–	}
1949	–	else
1950	–	c = ctl;        // re-increment on exit
1951	–	if ((int)(c >> AC_SHIFT) + parallelism == 0) {
1952	–	do {} while (!U.compareAndSwapLong
1953	–	(this, CTL, c = ctl, c + AC_UNIT));
1954	–	break;
1955	–	}
2109		}
2110	+	else if ((int)((c = ctl) >> AC_SHIFT) + parallelism <= 0 &&
2111	+	U.compareAndSwapLong
2112	+	(this, CTL, c, ((c & ~AC_MASK) |
2113	+	((c & AC_MASK) + AC_UNIT))))
2114	+	break;
2115		}
2116		}
2117
#	Line 1967 | Line 2125 | public class ForkJoinPool extends Abstra
2125		WorkQueue q; int b;
2126		if ((t = w.nextLocalTask()) != null)
2127		return t;
2128	<	if ((q = findNonEmptyStealQueue(w)) == null)
2128	>	if ((q = findNonEmptyStealQueue()) == null)
2129		return null;
2130		if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null)
2131		return t;
#	Line 1975 | Line 2133 | public class ForkJoinPool extends Abstra
2133		}
2134
2135		/**
2136	<	* Returns the approximate (non-atomic) number of idle threads per
2137	<	* active thread to offset steal queue size for method
2138	<	* ForkJoinTask.getSurplusQueuedTaskCount().
2139	<	*/
2140	<	final int idlePerActive() {
2141	<	// Approximate at powers of two for small values, saturate past 4
2142	<	int p = parallelism;
2143	<	int a = p + (int)(ctl >> AC_SHIFT);
2144	<	return (a > (p >>>= 1) ? 0 :
2145	<	a > (p >>>= 1) ? 1 :
2146	<	a > (p >>>= 1) ? 2 :
2147	<	a > (p >>>= 1) ? 4 :
2148	<	8);
2136	>	* Returns a cheap heuristic guide for task partitioning when
2137	>	* programmers, frameworks, tools, or languages have little or no
2138	>	* idea about task granularity.  In essence by offering this
2139	>	* method, we ask users only about tradeoffs in overhead vs
2140	>	* expected throughput and its variance, rather than how finely to
2141	>	* partition tasks.
2142	>	*
2143	>	* In a steady state strict (tree-structured) computation, each
2144	>	* thread makes available for stealing enough tasks for other
2145	>	* threads to remain active. Inductively, if all threads play by
2146	>	* the same rules, each thread should make available only a
2147	>	* constant number of tasks.
2148	>	*
2149	>	* The minimum useful constant is just 1. But using a value of 1
2150	>	* would require immediate replenishment upon each steal to
2151	>	* maintain enough tasks, which is infeasible.  Further,
2152	>	* partitionings/granularities of offered tasks should minimize
2153	>	* steal rates, which in general means that threads nearer the top
2154	>	* of computation tree should generate more than those nearer the
2155	>	* bottom. In perfect steady state, each thread is at
2156	>	* approximately the same level of computation tree. However,
2157	>	* producing extra tasks amortizes the uncertainty of progress and
2158	>	* diffusion assumptions.
2159	>	*
2160	>	* So, users will want to use values larger (but not much larger)
2161	>	* than 1 to both smooth over transient shortages and hedge
2162	>	* against uneven progress; as traded off against the cost of
2163	>	* extra task overhead. We leave the user to pick a threshold
2164	>	* value to compare with the results of this call to guide
2165	>	* decisions, but recommend values such as 3.
2166	>	*
2167	>	* When all threads are active, it is on average OK to estimate
2168	>	* surplus strictly locally. In steady-state, if one thread is
2169	>	* maintaining say 2 surplus tasks, then so are others. So we can
2170	>	* just use estimated queue length.  However, this strategy alone
2171	>	* leads to serious mis-estimates in some non-steady-state
2172	>	* conditions (ramp-up, ramp-down, other stalls). We can detect
2173	>	* many of these by further considering the number of "idle"
2174	>	* threads, that are known to have zero queued tasks, so
2175	>	* compensate by a factor of (#idle/#active) threads.
2176	>	*
2177	>	* Note: The approximation of #busy workers as #active workers is
2178	>	* not very good under current signalling scheme, and should be
2179	>	* improved.
2180	>	*/
2181	>	static int getSurplusQueuedTaskCount() {
2182	>	Thread t; ForkJoinWorkerThread wt; ForkJoinPool pool; WorkQueue q;
2183	>	if (((t = Thread.currentThread()) instanceof ForkJoinWorkerThread)) {
2184	>	int p = (pool = (wt = (ForkJoinWorkerThread)t).pool).parallelism;
2185	>	int n = (q = wt.workQueue).top - q.base;
2186	>	int a = (int)(pool.ctl >> AC_SHIFT) + p;
2187	>	return n - (a > (p >>>= 1) ? 0 :
2188	>	a > (p >>>= 1) ? 1 :
2189	>	a > (p >>>= 1) ? 2 :
2190	>	a > (p >>>= 1) ? 4 :
2191	>	8);
2192	>	}
2193	>	return 0;
2194		}
2195
2196		//  Termination
#	Line 2007 | Line 2210 | public class ForkJoinPool extends Abstra
2210		* @return true if now terminating or terminated
2211		*/
2212		private boolean tryTerminate(boolean now, boolean enable) {
2213	<	Mutex lock = this.lock;
2213	>	int ps;
2214	>	if (this == common)                        // cannot shut down
2215	>	return false;
2216	>	if ((ps = plock) >= 0) {                   // enable by setting plock
2217	>	if (!enable)
2218	>	return false;
2219	>	if ((ps & PL_LOCK) != 0 ||
2220	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
2221	>	ps = acquirePlock();
2222	>	int nps = ((ps + PL_LOCK) & ~SHUTDOWN) | SHUTDOWN;
2223	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
2224	>	releasePlock(nps);
2225	>	}
2226		for (long c;;) {
2227	<	if (((c = ctl) & STOP_BIT) != 0) {      // already terminating
2228	<	if ((short)(c >>> TC_SHIFT) == -parallelism) {
2229	<	lock.lock();                    // don't need try/finally
2230	<	termination.signalAll();        // signal when 0 workers
2231	<	lock.unlock();
2227	>	if (((c = ctl) & STOP_BIT) != 0) {     // already terminating
2228	>	if ((short)(c >>> TC_SHIFT) + parallelism <= 0) {
2229	>	synchronized (this) {
2230	>	notifyAll();               // signal when 0 workers
2231	>	}
2232		}
2233		return true;
2234		}
2235	<	if (runState >= 0) {                    // not yet enabled
2236	<	if (!enable)
2237	<	return false;
2023	<	lock.lock();
2024	<	runState |= SHUTDOWN;
2025	<	lock.unlock();
2026	<	}
2027	<	if (!now) {                             // check if idle & no tasks
2028	<	if ((int)(c >> AC_SHIFT) != -parallelism ||
2029	<	hasQueuedSubmissions())
2235	>	if (!now) {                            // check if idle & no tasks
2236	>	WorkQueue[] ws; WorkQueue w;
2237	>	if ((int)(c >> AC_SHIFT) + parallelism > 0)
2238		return false;
2239	<	// Check for unqueued inactive workers. One pass suffices.
2240	<	WorkQueue[] ws = workQueues; WorkQueue w;
2241	<	if (ws != null) {
2242	<	for (int i = 1; i < ws.length; i += 2) {
2243	<	if ((w = ws[i]) != null && w.eventCount >= 0)
2239	>	if ((ws = workQueues) != null) {
2240	>	for (int i = 0; i < ws.length; ++i) {
2241	>	if ((w = ws[i]) != null &&
2242	>	(!w.isEmpty() ||
2243	>	((i & 1) != 0 && w.eventCount >= 0))) {
2244	>	signalWork(ws, w);
2245		return false;
2246	+	}
2247		}
2248		}
2249		}
2250		if (U.compareAndSwapLong(this, CTL, c, c | STOP_BIT)) {
2251		for (int pass = 0; pass < 3; ++pass) {
2252	<	WorkQueue[] ws = workQueues;
2253	<	if (ws != null) {
2044	<	WorkQueue w;
2252	>	WorkQueue[] ws; WorkQueue w; Thread wt;
2253	>	if ((ws = workQueues) != null) {
2254		int n = ws.length;
2255		for (int i = 0; i < n; ++i) {
2256		if ((w = ws[i]) != null) {
2257	<	w.runState = -1;
2257	>	w.qlock = -1;
2258		if (pass > 0) {
2259		w.cancelAll();
2260	<	if (pass > 1)
2261	<	w.interruptOwner();
2260	>	if (pass > 1 && (wt = w.owner) != null) {
2261	>	if (!wt.isInterrupted()) {
2262	>	try {
2263	>	wt.interrupt();
2264	>	} catch (Throwable ignore) {
2265	>	}
2266	>	}
2267	>	U.unpark(wt);
2268	>	}
2269		}
2270		}
2271		}
2272		// Wake up workers parked on event queue
2273		int i, e; long cc; Thread p;
2274		while ((e = (int)(cc = ctl) & E_MASK) != 0 &&
2275	<	(i = e & SMASK) < n &&
2275	>	(i = e & SMASK) < n && i >= 0 &&
2276		(w = ws[i]) != null) {
2277		long nc = ((long)(w.nextWait & E_MASK) |
2278		((cc + AC_UNIT) & AC_MASK) |
#	Line 2064 | Line 2280 | public class ForkJoinPool extends Abstra
2280		if (w.eventCount == (e | INT_SIGN) &&
2281		U.compareAndSwapLong(this, CTL, cc, nc)) {
2282		w.eventCount = (e + E_SEQ) & E_MASK;
2283	<	w.runState = -1;
2283	>	w.qlock = -1;
2284		if ((p = w.parker) != null)
2285		U.unpark(p);
2286		}
#	Line 2075 | Line 2291 | public class ForkJoinPool extends Abstra
2291		}
2292		}
2293
2294	+	// external operations on common pool
2295	+
2296	+	/**
2297	+	* Returns common pool queue for a thread that has submitted at
2298	+	* least one task.
2299	+	*/
2300	+	static WorkQueue commonSubmitterQueue() {
2301	+	Submitter z; ForkJoinPool p; WorkQueue[] ws; int m, r;
2302	+	return ((z = submitters.get()) != null &&
2303	+	(p = common) != null &&
2304	+	(ws = p.workQueues) != null &&
2305	+	(m = ws.length - 1) >= 0) ?
2306	+	ws[m & z.seed & SQMASK] : null;
2307	+	}
2308	+
2309	+	/**
2310	+	* Tries to pop the given task from submitter's queue in common pool.
2311	+	*/
2312	+	final boolean tryExternalUnpush(ForkJoinTask<?> task) {
2313	+	WorkQueue joiner; ForkJoinTask<?>[] a; int m, s;
2314	+	Submitter z = submitters.get();
2315	+	WorkQueue[] ws = workQueues;
2316	+	boolean popped = false;
2317	+	if (z != null && ws != null && (m = ws.length - 1) >= 0 &&
2318	+	(joiner = ws[z.seed & m & SQMASK]) != null &&
2319	+	joiner.base != (s = joiner.top) &&
2320	+	(a = joiner.array) != null) {
2321	+	long j = (((a.length - 1) & (s - 1)) << ASHIFT) + ABASE;
2322	+	if (U.getObject(a, j) == task &&
2323	+	U.compareAndSwapInt(joiner, QLOCK, 0, 1)) {
2324	+	if (joiner.top == s && joiner.array == a &&
2325	+	U.compareAndSwapObject(a, j, task, null)) {
2326	+	joiner.top = s - 1;
2327	+	popped = true;
2328	+	}
2329	+	joiner.qlock = 0;
2330	+	}
2331	+	}
2332	+	return popped;
2333	+	}
2334	+
2335	+	final int externalHelpComplete(CountedCompleter<?> task) {
2336	+	WorkQueue joiner; int m, j;
2337	+	Submitter z = submitters.get();
2338	+	WorkQueue[] ws = workQueues;
2339	+	int s = 0;
2340	+	if (z != null && ws != null && (m = ws.length - 1) >= 0 &&
2341	+	(joiner = ws[(j = z.seed) & m & SQMASK]) != null && task != null) {
2342	+	int scans = m + m + 1;
2343	+	long c = 0L;             // for stability check
2344	+	j |= 1;                  // poll odd queues
2345	+	for (int k = scans; ; j += 2) {
2346	+	WorkQueue q;
2347	+	if ((s = task.status) < 0)
2348	+	break;
2349	+	else if (joiner.externalPopAndExecCC(task))
2350	+	k = scans;
2351	+	else if ((s = task.status) < 0)
2352	+	break;
2353	+	else if ((q = ws[j & m]) != null && q.pollAndExecCC(task))
2354	+	k = scans;
2355	+	else if (--k < 0) {
2356	+	if (c == (c = ctl))
2357	+	break;
2358	+	k = scans;
2359	+	}
2360	+	}
2361	+	}
2362	+	return s;
2363	+	}
2364	+
2365		// Exported methods
2366
2367		// Constructors
#	Line 2091 | Line 2378 | public class ForkJoinPool extends Abstra
2378		*         java.lang.RuntimePermission}{@code ("modifyThread")}
2379		*/
2380		public ForkJoinPool() {
2381	<	this(Runtime.getRuntime().availableProcessors(),
2381	>	this(Math.min(MAX_CAP, Runtime.getRuntime().availableProcessors()),
2382		defaultForkJoinWorkerThreadFactory, null, false);
2383		}
2384
#	Line 2139 | Line 2426 | public class ForkJoinPool extends Abstra
2426		*/
2427		public ForkJoinPool(int parallelism,
2428		ForkJoinWorkerThreadFactory factory,
2429	<	Thread.UncaughtExceptionHandler handler,
2429	>	UncaughtExceptionHandler handler,
2430		boolean asyncMode) {
2431	+	this(checkParallelism(parallelism),
2432	+	checkFactory(factory),
2433	+	handler,
2434	+	(asyncMode ? FIFO_QUEUE : LIFO_QUEUE),
2435	+	"ForkJoinPool-" + nextPoolId() + "-worker-");
2436		checkPermission();
2437	<	if (factory == null)
2438	<	throw new NullPointerException();
2437	>	}
2438	>
2439	>	private static int checkParallelism(int parallelism) {
2440		if (parallelism <= 0 || parallelism > MAX_CAP)
2441		throw new IllegalArgumentException();
2442	<	this.parallelism = parallelism;
2442	>	return parallelism;
2443	>	}
2444	>
2445	>	private static ForkJoinWorkerThreadFactory checkFactory
2446	>	(ForkJoinWorkerThreadFactory factory) {
2447	>	if (factory == null)
2448	>	throw new NullPointerException();
2449	>	return factory;
2450	>	}
2451	>
2452	>	/**
2453	>	* Creates a {@code ForkJoinPool} with the given parameters, without
2454	>	* any security checks or parameter validation.  Invoked directly by
2455	>	* makeCommonPool.
2456	>	*/
2457	>	private ForkJoinPool(int parallelism,
2458	>	ForkJoinWorkerThreadFactory factory,
2459	>	UncaughtExceptionHandler handler,
2460	>	int mode,
2461	>	String workerNamePrefix) {
2462	>	this.workerNamePrefix = workerNamePrefix;
2463		this.factory = factory;
2464		this.ueh = handler;
2465	<	this.localMode = asyncMode ? FIFO_QUEUE : LIFO_QUEUE;
2465	>	this.mode = (short)mode;
2466	>	this.parallelism = (short)parallelism;
2467		long np = (long)(-parallelism); // offset ctl counts
2468		this.ctl = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK);
2469	<	// Use nearest power 2 for workQueues size. See Hackers Delight sec 3.2.
2470	<	int n = parallelism - 1;
2471	<	n |= n >>> 1; n |= n >>> 2; n |= n >>> 4; n |= n >>> 8; n |= n >>> 16;
2472	<	int size = (n + 1) << 1;        // #slots = 2*#workers
2473	<	this.submitMask = size - 1;     // room for max # of submit queues
2474	<	this.workQueues = new WorkQueue[size];
2475	<	this.termination = (this.lock = new Mutex()).newCondition();
2476	<	this.stealCount = new AtomicLong();
2477	<	this.nextWorkerNumber = new AtomicInteger();
2478	<	int pn = poolNumberGenerator.incrementAndGet();
2479	<	StringBuilder sb = new StringBuilder("ForkJoinPool-");
2480	<	sb.append(Integer.toString(pn));
2481	<	sb.append("-worker-");
2482	<	this.workerNamePrefix = sb.toString();
2483	<	lock.lock();
2484	<	this.runState = 1;              // set init flag
2485	<	lock.unlock();
2469	>	}
2470	>
2471	>	/**
2472	>	* Returns the common pool instance. This pool is statically
2473	>	* constructed; its run state is unaffected by attempts to {@link
2474	>	* #shutdown} or {@link #shutdownNow}. However this pool and any
2475	>	* ongoing processing are automatically terminated upon program
2476	>	* {@link System#exit}.  Any program that relies on asynchronous
2477	>	* task processing to complete before program termination should
2478	>	* invoke {@code commonPool().}{@link #awaitQuiescence awaitQuiescence},
2479	>	* before exit.
2480	>	*
2481	>	* @return the common pool instance
2482	>	* @since 1.8
2483	>	*/
2484	>	public static ForkJoinPool commonPool() {
2485	>	// assert common != null : "static init error";
2486	>	return common;
2487		}
2488
2489		// Execution methods
#	Line 2184 | Line 2499 | public class ForkJoinPool extends Abstra
2499		* minimally only the latter.
2500		*
2501		* @param task the task
2502	+	* @param <T> the type of the task's result
2503		* @return the task's result
2504		* @throws NullPointerException if the task is null
2505		* @throws RejectedExecutionException if the task cannot be
#	Line 2192 | Line 2508 | public class ForkJoinPool extends Abstra
2508		public <T> T invoke(ForkJoinTask<T> task) {
2509		if (task == null)
2510		throw new NullPointerException();
2511	<	doSubmit(task);
2511	>	externalPush(task);
2512		return task.join();
2513		}
2514
#	Line 2207 | Line 2523 | public class ForkJoinPool extends Abstra
2523		public void execute(ForkJoinTask<?> task) {
2524		if (task == null)
2525		throw new NullPointerException();
2526	<	doSubmit(task);
2526	>	externalPush(task);
2527		}
2528
2529		// AbstractExecutorService methods
#	Line 2224 | Line 2540 | public class ForkJoinPool extends Abstra
2540		if (task instanceof ForkJoinTask<?>) // avoid re-wrap
2541		job = (ForkJoinTask<?>) task;
2542		else
2543	<	job = new ForkJoinTask.AdaptedRunnableAction(task);
2544	<	doSubmit(job);
2543	>	job = new ForkJoinTask.RunnableExecuteAction(task);
2544	>	externalPush(job);
2545		}
2546
2547		/**
2548		* Submits a ForkJoinTask for execution.
2549		*
2550		* @param task the task to submit
2551	+	* @param <T> the type of the task's result
2552		* @return the task
2553		* @throws NullPointerException if the task is null
2554		* @throws RejectedExecutionException if the task cannot be
#	Line 2240 | Line 2557 | public class ForkJoinPool extends Abstra
2557		public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) {
2558		if (task == null)
2559		throw new NullPointerException();
2560	<	doSubmit(task);
2560	>	externalPush(task);
2561		return task;
2562		}
2563
#	Line 2251 | Line 2568 | public class ForkJoinPool extends Abstra
2568		*/
2569		public <T> ForkJoinTask<T> submit(Callable<T> task) {
2570		ForkJoinTask<T> job = new ForkJoinTask.AdaptedCallable<T>(task);
2571	<	doSubmit(job);
2571	>	externalPush(job);
2572		return job;
2573		}
2574
#	Line 2262 | Line 2579 | public class ForkJoinPool extends Abstra
2579		*/
2580		public <T> ForkJoinTask<T> submit(Runnable task, T result) {
2581		ForkJoinTask<T> job = new ForkJoinTask.AdaptedRunnable<T>(task, result);
2582	<	doSubmit(job);
2582	>	externalPush(job);
2583		return job;
2584		}
2585
#	Line 2279 | Line 2596 | public class ForkJoinPool extends Abstra
2596		job = (ForkJoinTask<?>) task;
2597		else
2598		job = new ForkJoinTask.AdaptedRunnableAction(task);
2599	<	doSubmit(job);
2599	>	externalPush(job);
2600		return job;
2601		}
2602
#	Line 2291 | Line 2608 | public class ForkJoinPool extends Abstra
2608		// In previous versions of this class, this method constructed
2609		// a task to run ForkJoinTask.invokeAll, but now external
2610		// invocation of multiple tasks is at least as efficient.
2611	<	List<ForkJoinTask<T>> fs = new ArrayList<ForkJoinTask<T>>(tasks.size());
2295	<	// Workaround needed because method wasn't declared with
2296	<	// wildcards in return type but should have been.
2297	<	@SuppressWarnings({"unchecked", "rawtypes"})
2298	<	List<Future<T>> futures = (List<Future<T>>) (List) fs;
2611	>	ArrayList<Future<T>> futures = new ArrayList<Future<T>>(tasks.size());
2612
2613		boolean done = false;
2614		try {
2615		for (Callable<T> t : tasks) {
2616		ForkJoinTask<T> f = new ForkJoinTask.AdaptedCallable<T>(t);
2617	<	doSubmit(f);
2618	<	fs.add(f);
2617	>	futures.add(f);
2618	>	externalPush(f);
2619		}
2620	<	for (ForkJoinTask<T> f : fs)
2621	<	f.quietlyJoin();
2620	>	for (int i = 0, size = futures.size(); i < size; i++)
2621	>	((ForkJoinTask<?>)futures.get(i)).quietlyJoin();
2622		done = true;
2623		return futures;
2624		} finally {
2625		if (!done)
2626	<	for (ForkJoinTask<T> f : fs)
2627	<	f.cancel(false);
2626	>	for (int i = 0, size = futures.size(); i < size; i++)
2627	>	futures.get(i).cancel(false);
2628		}
2629		}
2630
#	Line 2330 | Line 2643 | public class ForkJoinPool extends Abstra
2643		*
2644		* @return the handler, or {@code null} if none
2645		*/
2646	<	public Thread.UncaughtExceptionHandler getUncaughtExceptionHandler() {
2646	>	public UncaughtExceptionHandler getUncaughtExceptionHandler() {
2647		return ueh;
2648		}
2649
#	Line 2340 | Line 2653 | public class ForkJoinPool extends Abstra
2653		* @return the targeted parallelism level of this pool
2654		*/
2655		public int getParallelism() {
2656	<	return parallelism;
2656	>	int par;
2657	>	return ((par = parallelism) > 0) ? par : 1;
2658	>	}
2659	>
2660	>	/**
2661	>	* Returns the targeted parallelism level of the common pool.
2662	>	*
2663	>	* @return the targeted parallelism level of the common pool
2664	>	* @since 1.8
2665	>	*/
2666	>	public static int getCommonPoolParallelism() {
2667	>	return commonParallelism;
2668		}
2669
2670		/**
#	Line 2362 | Line 2686 | public class ForkJoinPool extends Abstra
2686		* @return {@code true} if this pool uses async mode
2687		*/
2688		public boolean getAsyncMode() {
2689	<	return localMode != 0;
2689	>	return mode == FIFO_QUEUE;
2690		}
2691
2692		/**
#	Line 2409 | Line 2733 | public class ForkJoinPool extends Abstra
2733		* @return {@code true} if all threads are currently idle
2734		*/
2735		public boolean isQuiescent() {
2736	<	return (int)(ctl >> AC_SHIFT) + parallelism == 0;
2736	>	return parallelism + (int)(ctl >> AC_SHIFT) <= 0;
2737		}
2738
2739		/**
#	Line 2424 | Line 2748 | public class ForkJoinPool extends Abstra
2748		* @return the number of steals
2749		*/
2750		public long getStealCount() {
2751	<	long count = stealCount.get();
2751	>	long count = stealCount;
2752		WorkQueue[] ws; WorkQueue w;
2753		if ((ws = workQueues) != null) {
2754		for (int i = 1; i < ws.length; i += 2) {
2755		if ((w = ws[i]) != null)
2756	<	count += w.totalSteals;
2756	>	count += w.nsteals;
2757		}
2758		}
2759		return count;
#	Line 2554 | Line 2878 | public class ForkJoinPool extends Abstra
2878		public String toString() {
2879		// Use a single pass through workQueues to collect counts
2880		long qt = 0L, qs = 0L; int rc = 0;
2881	<	long st = stealCount.get();
2881	>	long st = stealCount;
2882		long c = ctl;
2883		WorkQueue[] ws; WorkQueue w;
2884		if ((ws = workQueues) != null) {
#	Line 2565 | Line 2889 | public class ForkJoinPool extends Abstra
2889		qs += size;
2890		else {
2891		qt += size;
2892	<	st += w.totalSteals;
2892	>	st += w.nsteals;
2893		if (w.isApparentlyUnblocked())
2894		++rc;
2895		}
#	Line 2581 | Line 2905 | public class ForkJoinPool extends Abstra
2905		if ((c & STOP_BIT) != 0)
2906		level = (tc == 0) ? "Terminated" : "Terminating";
2907		else
2908	<	level = runState < 0 ? "Shutting down" : "Running";
2908	>	level = plock < 0 ? "Shutting down" : "Running";
2909		return super.toString() +
2910		"[" + level +
2911		", parallelism = " + pc +
#	Line 2595 | Line 2919 | public class ForkJoinPool extends Abstra
2919		}
2920
2921		/**
2922	<	* Initiates an orderly shutdown in which previously submitted
2923	<	* tasks are executed, but no new tasks will be accepted.
2924	<	* Invocation has no additional effect if already shut down.
2925	<	* Tasks that are in the process of being submitted concurrently
2926	<	* during the course of this method may or may not be rejected.
2922	>	* Possibly initiates an orderly shutdown in which previously
2923	>	* submitted tasks are executed, but no new tasks will be
2924	>	* accepted. Invocation has no effect on execution state if this
2925	>	* is the {@link #commonPool()}, and no additional effect if
2926	>	* already shut down.  Tasks that are in the process of being
2927	>	* submitted concurrently during the course of this method may or
2928	>	* may not be rejected.
2929		*
2930		* @throws SecurityException if a security manager exists and
2931		*         the caller is not permitted to modify threads
#	Line 2612 | Line 2938 | public class ForkJoinPool extends Abstra
2938		}
2939
2940		/**
2941	<	* Attempts to cancel and/or stop all tasks, and reject all
2942	<	* subsequently submitted tasks.  Tasks that are in the process of
2943	<	* being submitted or executed concurrently during the course of
2944	<	* this method may or may not be rejected. This method cancels
2945	<	* both existing and unexecuted tasks, in order to permit
2946	<	* termination in the presence of task dependencies. So the method
2947	<	* always returns an empty list (unlike the case for some other
2948	<	* Executors).
2941	>	* Possibly attempts to cancel and/or stop all tasks, and reject
2942	>	* all subsequently submitted tasks.  Invocation has no effect on
2943	>	* execution state if this is the {@link #commonPool()}, and no
2944	>	* additional effect if already shut down. Otherwise, tasks that
2945	>	* are in the process of being submitted or executed concurrently
2946	>	* during the course of this method may or may not be
2947	>	* rejected. This method cancels both existing and unexecuted
2948	>	* tasks, in order to permit termination in the presence of task
2949	>	* dependencies. So the method always returns an empty list
2950	>	* (unlike the case for some other Executors).
2951		*
2952		* @return an empty list
2953		* @throws SecurityException if a security manager exists and
#	Line 2641 | Line 2969 | public class ForkJoinPool extends Abstra
2969		public boolean isTerminated() {
2970		long c = ctl;
2971		return ((c & STOP_BIT) != 0L &&
2972	<	(short)(c >>> TC_SHIFT) == -parallelism);
2972	>	(short)(c >>> TC_SHIFT) + parallelism <= 0);
2973		}
2974
2975		/**
#	Line 2649 | Line 2977 | public class ForkJoinPool extends Abstra
2977		* commenced but not yet completed.  This method may be useful for
2978		* debugging. A return of {@code true} reported a sufficient
2979		* period after shutdown may indicate that submitted tasks have
2980	<	* ignored or suppressed interruption, or are waiting for IO,
2980	>	* ignored or suppressed interruption, or are waiting for I/O,
2981		* causing this executor not to properly terminate. (See the
2982		* advisory notes for class {@link ForkJoinTask} stating that
2983		* tasks should not normally entail blocking operations.  But if
#	Line 2660 | Line 2988 | public class ForkJoinPool extends Abstra
2988		public boolean isTerminating() {
2989		long c = ctl;
2990		return ((c & STOP_BIT) != 0L &&
2991	<	(short)(c >>> TC_SHIFT) != -parallelism);
2991	>	(short)(c >>> TC_SHIFT) + parallelism > 0);
2992		}
2993
2994		/**
#	Line 2669 | Line 2997 | public class ForkJoinPool extends Abstra
2997		* @return {@code true} if this pool has been shut down
2998		*/
2999		public boolean isShutdown() {
3000	<	return runState < 0;
3000	>	return plock < 0;
3001		}
3002
3003		/**
3004	<	* Blocks until all tasks have completed execution after a shutdown
3005	<	* request, or the timeout occurs, or the current thread is
3006	<	* interrupted, whichever happens first.
3004	>	* Blocks until all tasks have completed execution after a
3005	>	* shutdown request, or the timeout occurs, or the current thread
3006	>	* is interrupted, whichever happens first. Because the {@link
3007	>	* #commonPool()} never terminates until program shutdown, when
3008	>	* applied to the common pool, this method is equivalent to {@link
3009	>	* #awaitQuiescence(long, TimeUnit)} but always returns {@code false}.
3010		*
3011		* @param timeout the maximum time to wait
3012		* @param unit the time unit of the timeout argument
#	Line 2685 | Line 3016 | public class ForkJoinPool extends Abstra
3016		*/
3017		public boolean awaitTermination(long timeout, TimeUnit unit)
3018		throws InterruptedException {
3019	+	if (Thread.interrupted())
3020	+	throw new InterruptedException();
3021	+	if (this == common) {
3022	+	awaitQuiescence(timeout, unit);
3023	+	return false;
3024	+	}
3025		long nanos = unit.toNanos(timeout);
3026	<	final Mutex lock = this.lock;
3027	<	lock.lock();
3028	<	try {
3026	>	if (isTerminated())
3027	>	return true;
3028	>	if (nanos <= 0L)
3029	>	return false;
3030	>	long deadline = System.nanoTime() + nanos;
3031	>	synchronized (this) {
3032		for (;;) {
3033		if (isTerminated())
3034		return true;
3035	<	if (nanos <= 0)
3035	>	if (nanos <= 0L)
3036		return false;
3037	<	nanos = termination.awaitNanos(nanos);
3037	>	long millis = TimeUnit.NANOSECONDS.toMillis(nanos);
3038	>	wait(millis > 0L ? millis : 1L);
3039	>	nanos = deadline - System.nanoTime();
3040		}
2699	–	} finally {
2700	–	lock.unlock();
3041		}
3042		}
3043
3044		/**
3045	+	* If called by a ForkJoinTask operating in this pool, equivalent
3046	+	* in effect to {@link ForkJoinTask#helpQuiesce}. Otherwise,
3047	+	* waits and/or attempts to assist performing tasks until this
3048	+	* pool {@link #isQuiescent} or the indicated timeout elapses.
3049	+	*
3050	+	* @param timeout the maximum time to wait
3051	+	* @param unit the time unit of the timeout argument
3052	+	* @return {@code true} if quiescent; {@code false} if the
3053	+	* timeout elapsed.
3054	+	*/
3055	+	public boolean awaitQuiescence(long timeout, TimeUnit unit) {
3056	+	long nanos = unit.toNanos(timeout);
3057	+	ForkJoinWorkerThread wt;
3058	+	Thread thread = Thread.currentThread();
3059	+	if ((thread instanceof ForkJoinWorkerThread) &&
3060	+	(wt = (ForkJoinWorkerThread)thread).pool == this) {
3061	+	helpQuiescePool(wt.workQueue);
3062	+	return true;
3063	+	}
3064	+	long startTime = System.nanoTime();
3065	+	WorkQueue[] ws;
3066	+	int r = 0, m;
3067	+	boolean found = true;
3068	+	while (!isQuiescent() && (ws = workQueues) != null &&
3069	+	(m = ws.length - 1) >= 0) {
3070	+	if (!found) {
3071	+	if ((System.nanoTime() - startTime) > nanos)
3072	+	return false;
3073	+	Thread.yield(); // cannot block
3074	+	}
3075	+	found = false;
3076	+	for (int j = (m + 1) << 2; j >= 0; --j) {
3077	+	ForkJoinTask<?> t; WorkQueue q; int b;
3078	+	if ((q = ws[r++ & m]) != null && (b = q.base) - q.top < 0) {
3079	+	found = true;
3080	+	if ((t = q.pollAt(b)) != null)
3081	+	t.doExec();
3082	+	break;
3083	+	}
3084	+	}
3085	+	}
3086	+	return true;
3087	+	}
3088	+
3089	+	/**
3090	+	* Waits and/or attempts to assist performing tasks indefinitely
3091	+	* until the {@link #commonPool()} {@link #isQuiescent}.
3092	+	*/
3093	+	static void quiesceCommonPool() {
3094	+	common.awaitQuiescence(Long.MAX_VALUE, TimeUnit.NANOSECONDS);
3095	+	}
3096	+
3097	+	/**
3098		* Interface for extending managed parallelism for tasks running
3099		* in {@link ForkJoinPool}s.
3100		*
#	Line 2710 | Line 3103 | public class ForkJoinPool extends Abstra
3103		* not necessary. Method {@code block} blocks the current thread
3104		* if necessary (perhaps internally invoking {@code isReleasable}
3105		* before actually blocking). These actions are performed by any
3106	<	* thread invoking {@link ForkJoinPool#managedBlock}.  The
3107	<	* unusual methods in this API accommodate synchronizers that may,
3108	<	* but don't usually, block for long periods. Similarly, they
3106	>	* thread invoking {@link ForkJoinPool#managedBlock(ManagedBlocker)}.
3107	>	* The unusual methods in this API accommodate synchronizers that
3108	>	* may, but don't usually, block for long periods. Similarly, they
3109		* allow more efficient internal handling of cases in which
3110		* additional workers may be, but usually are not, needed to
3111		* ensure sufficient parallelism.  Toward this end,
#	Line 2770 | Line 3163 | public class ForkJoinPool extends Abstra
3163
3164		/**
3165		* Returns {@code true} if blocking is unnecessary.
3166	+	* @return {@code true} if blocking is unnecessary
3167		*/
3168		boolean isReleasable();
3169		}
#	Line 2797 | Line 3191 | public class ForkJoinPool extends Abstra
3191		public static void managedBlock(ManagedBlocker blocker)
3192		throws InterruptedException {
3193		Thread t = Thread.currentThread();
3194	<	ForkJoinPool p = ((t instanceof ForkJoinWorkerThread) ?
3195	<	((ForkJoinWorkerThread)t).pool : null);
3196	<	while (!blocker.isReleasable()) {
3197	<	if (p == null || p.tryCompensate(null, blocker)) {
3198	<	try {
3199	<	do {} while (!blocker.isReleasable() && !blocker.block());
3200	<	} finally {
3201	<	if (p != null)
3194	>	if (t instanceof ForkJoinWorkerThread) {
3195	>	ForkJoinPool p = ((ForkJoinWorkerThread)t).pool;
3196	>	while (!blocker.isReleasable()) {
3197	>	if (p.tryCompensate(p.ctl)) {
3198	>	try {
3199	>	do {} while (!blocker.isReleasable() &&
3200	>	!blocker.block());
3201	>	} finally {
3202		p.incrementActiveCount();
3203	+	}
3204	+	break;
3205		}
2810	–	break;
3206		}
3207		}
3208	+	else {
3209	+	do {} while (!blocker.isReleasable() &&
3210	+	!blocker.block());
3211	+	}
3212		}
3213
3214		// AbstractExecutorService overrides.  These rely on undocumented
#	Line 2830 | Line 3229 | public class ForkJoinPool extends Abstra
3229		private static final long PARKBLOCKER;
3230		private static final int ABASE;
3231		private static final int ASHIFT;
3232	+	private static final long STEALCOUNT;
3233	+	private static final long PLOCK;
3234	+	private static final long INDEXSEED;
3235	+	private static final long QBASE;
3236	+	private static final long QLOCK;
3237
3238		static {
3239	<	poolNumberGenerator = new AtomicInteger();
2836	<	nextSubmitterSeed = new AtomicInteger(0x55555555);
2837	<	modifyThreadPermission = new RuntimePermission("modifyThread");
2838	<	defaultForkJoinWorkerThreadFactory =
2839	<	new DefaultForkJoinWorkerThreadFactory();
2840	<	submitters = new ThreadSubmitter();
2841	<	int s;
3239	>	// initialize field offsets for CAS etc
3240		try {
3241		U = getUnsafe();
3242		Class<?> k = ForkJoinPool.class;
2845	–	Class<?> ak = ForkJoinTask[].class;
3243		CTL = U.objectFieldOffset
3244		(k.getDeclaredField("ctl"));
3245	+	STEALCOUNT = U.objectFieldOffset
3246	+	(k.getDeclaredField("stealCount"));
3247	+	PLOCK = U.objectFieldOffset
3248	+	(k.getDeclaredField("plock"));
3249	+	INDEXSEED = U.objectFieldOffset
3250	+	(k.getDeclaredField("indexSeed"));
3251		Class<?> tk = Thread.class;
3252		PARKBLOCKER = U.objectFieldOffset
3253		(tk.getDeclaredField("parkBlocker"));
3254	+	Class<?> wk = WorkQueue.class;
3255	+	QBASE = U.objectFieldOffset
3256	+	(wk.getDeclaredField("base"));
3257	+	QLOCK = U.objectFieldOffset
3258	+	(wk.getDeclaredField("qlock"));
3259	+	Class<?> ak = ForkJoinTask[].class;
3260		ABASE = U.arrayBaseOffset(ak);
3261	<	s = U.arrayIndexScale(ak);
3261	>	int scale = U.arrayIndexScale(ak);
3262	>	if ((scale & (scale - 1)) != 0)
3263	>	throw new Error("data type scale not a power of two");
3264	>	ASHIFT = 31 - Integer.numberOfLeadingZeros(scale);
3265		} catch (Exception e) {
3266		throw new Error(e);
3267		}
3268	<	if ((s & (s-1)) != 0)
3269	<	throw new Error("data type scale not a power of two");
3270	<	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
3268	>
3269	>	submitters = new ThreadLocal<Submitter>();
3270	>	defaultForkJoinWorkerThreadFactory =
3271	>	new DefaultForkJoinWorkerThreadFactory();
3272	>	modifyThreadPermission = new RuntimePermission("modifyThread");
3273	>
3274	>	common = java.security.AccessController.doPrivileged
3275	>	(new java.security.PrivilegedAction<ForkJoinPool>() {
3276	>	public ForkJoinPool run() { return makeCommonPool(); }});
3277	>	int par = common.parallelism; // report 1 even if threads disabled
3278	>	commonParallelism = par > 0 ? par : 1;
3279	>	}
3280	>
3281	>	/**
3282	>	* Creates and returns the common pool, respecting user settings
3283	>	* specified via system properties.
3284	>	*/
3285	>	private static ForkJoinPool makeCommonPool() {
3286	>	int parallelism = -1;
3287	>	ForkJoinWorkerThreadFactory factory
3288	>	= defaultForkJoinWorkerThreadFactory;
3289	>	UncaughtExceptionHandler handler = null;
3290	>	try {  // ignore exceptions in accessing/parsing properties
3291	>	String pp = System.getProperty
3292	>	("java.util.concurrent.ForkJoinPool.common.parallelism");
3293	>	String fp = System.getProperty
3294	>	("java.util.concurrent.ForkJoinPool.common.threadFactory");
3295	>	String hp = System.getProperty
3296	>	("java.util.concurrent.ForkJoinPool.common.exceptionHandler");
3297	>	if (pp != null)
3298	>	parallelism = Integer.parseInt(pp);
3299	>	if (fp != null)
3300	>	factory = ((ForkJoinWorkerThreadFactory)ClassLoader.
3301	>	getSystemClassLoader().loadClass(fp).newInstance());
3302	>	if (hp != null)
3303	>	handler = ((UncaughtExceptionHandler)ClassLoader.
3304	>	getSystemClassLoader().loadClass(hp).newInstance());
3305	>	} catch (Exception ignore) {
3306	>	}
3307	>
3308	>	if (parallelism < 0 && // default 1 less than #cores
3309	>	(parallelism = Runtime.getRuntime().availableProcessors() - 1) < 0)
3310	>	parallelism = 0;
3311	>	if (parallelism > MAX_CAP)
3312	>	parallelism = MAX_CAP;
3313	>	return new ForkJoinPool(parallelism, factory, handler, LIFO_QUEUE,
3314	>	"ForkJoinPool.commonPool-worker-");
3315		}
3316
3317		/**
#	Line 2868 | Line 3324 | public class ForkJoinPool extends Abstra
3324		private static sun.misc.Unsafe getUnsafe() {
3325		try {
3326		return sun.misc.Unsafe.getUnsafe();
3327	<	} catch (SecurityException se) {
3328	<	try {
3329	<	return java.security.AccessController.doPrivileged
3330	<	(new java.security
3331	<	.PrivilegedExceptionAction<sun.misc.Unsafe>() {
3332	<	public sun.misc.Unsafe run() throws Exception {
3333	<	java.lang.reflect.Field f = sun.misc
3334	<	.Unsafe.class.getDeclaredField("theUnsafe");
3335	<	f.setAccessible(true);
3336	<	return (sun.misc.Unsafe) f.get(null);
3337	<	}});
3338	<	} catch (java.security.PrivilegedActionException e) {
3339	<	throw new RuntimeException("Could not initialize intrinsics",
3340	<	e.getCause());
3341	<	}
3327	>	} catch (SecurityException tryReflectionInstead) {}
3328	>	try {
3329	>	return java.security.AccessController.doPrivileged
3330	>	(new java.security.PrivilegedExceptionAction<sun.misc.Unsafe>() {
3331	>	public sun.misc.Unsafe run() throws Exception {
3332	>	Class<sun.misc.Unsafe> k = sun.misc.Unsafe.class;
3333	>	for (java.lang.reflect.Field f : k.getDeclaredFields()) {
3334	>	f.setAccessible(true);
3335	>	Object x = f.get(null);
3336	>	if (k.isInstance(x))
3337	>	return k.cast(x);
3338	>	}
3339	>	throw new NoSuchFieldError("the Unsafe");
3340	>	}});
3341	>	} catch (java.security.PrivilegedActionException e) {
3342	>	throw new RuntimeException("Could not initialize intrinsics",
3343	>	e.getCause());
3344		}
3345		}
2888	–
3346		}

Diff Legend

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