ViewVC Help

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

root/jsr166/jsr166/src/jsr166y/ForkJoinPool.java

Comparing jsr166/src/jsr166y/ForkJoinPool.java (file contents):
Revision 1.115 by jsr166, Thu Jan 26 19:10:27 2012 UTC vs.
Revision 1.141 by jsr166, Wed Nov 14 18:45:53 2012 UTC

#	Line 11 | Line 11 | import java.util.Arrays;
11		import java.util.Collection;
12		import java.util.Collections;
13		import java.util.List;
14	–	import java.util.Random;
14		import java.util.concurrent.AbstractExecutorService;
15		import java.util.concurrent.Callable;
16		import java.util.concurrent.ExecutorService;
#	Line 19 | Line 18 | import java.util.concurrent.Future;
18		import java.util.concurrent.RejectedExecutionException;
19		import java.util.concurrent.RunnableFuture;
20		import java.util.concurrent.TimeUnit;
22	–	import java.util.concurrent.atomic.AtomicInteger;
23	–	import java.util.concurrent.atomic.AtomicLong;
24	–	import java.util.concurrent.locks.ReentrantLock;
25	–	import java.util.concurrent.locks.Condition;
21
22		/**
23		* An {@link ExecutorService} for running {@link ForkJoinTask}s.
#	Line 42 | Line 37 | import java.util.concurrent.locks.Condit
37		* ForkJoinPool}s may also be appropriate for use with event-style
38		* tasks that are never joined.
39		*
40	<	* <p>A {@code ForkJoinPool} is constructed with a given target
41	<	* parallelism level; by default, equal to the number of available
42	<	* processors. The pool attempts to maintain enough active (or
43	<	* available) threads by dynamically adding, suspending, or resuming
44	<	* internal worker threads, even if some tasks are stalled waiting to
45	<	* join others. However, no such adjustments are guaranteed in the
46	<	* face of blocked IO or other unmanaged synchronization. The nested
47	<	* {@link ManagedBlocker} interface enables extension of the kinds of
40	>	* <p>A static {@link #commonPool} is available and appropriate for
41	>	* most applications. The common pool is used by any ForkJoinTask that
42	>	* is not explicitly submitted to a specified pool. Using the common
43	>	* pool normally reduces resource usage (its threads are slowly
44	>	* reclaimed during periods of non-use, and reinstated upon subsequent
45	>	* use).
46	>	*
47	>	* <p>For applications that require separate or custom pools, a {@code
48	>	* ForkJoinPool} may be constructed with a given target parallelism
49	>	* level; by default, equal to the number of available processors. The
50	>	* pool attempts to maintain enough active (or available) threads by
51	>	* dynamically adding, suspending, or resuming internal worker
52	>	* threads, even if some tasks are stalled waiting to join
53	>	* others. However, no such adjustments are guaranteed in the face of
54	>	* blocked IO or other unmanaged synchronization. The nested {@link
55	>	* ManagedBlocker} interface enables extension of the kinds of
56		* synchronization accommodated.
57		*
58		* <p>In addition to execution and lifecycle control methods, this
#	Line 60 | Line 63 | import java.util.concurrent.locks.Condit
63		* convenient form for informal monitoring.
64		*
65		* <p> As is the case with other ExecutorServices, there are three
66	<	* main task execution methods summarized in the following
67	<	* table. These are designed to be used primarily by clients not
68	<	* already engaged in fork/join computations in the current pool.  The
69	<	* main forms of these methods accept instances of {@code
70	<	* ForkJoinTask}, but overloaded forms also allow mixed execution of
71	<	* plain {@code Runnable}- or {@code Callable}- based activities as
72	<	* well.  However, tasks that are already executing in a pool should
73	<	* normally instead use the within-computation forms listed in the
74	<	* table unless using async event-style tasks that are not usually
75	<	* joined, in which case there is little difference among choice of
73	<	* methods.
66	>	* main task execution methods summarized in the following table.
67	>	* These are designed to be used primarily by clients not already
68	>	* engaged in fork/join computations in the current pool.  The main
69	>	* forms of these methods accept instances of {@code ForkJoinTask},
70	>	* but overloaded forms also allow mixed execution of plain {@code
71	>	* Runnable}- or {@code Callable}- based activities as well.  However,
72	>	* tasks that are already executing in a pool should normally instead
73	>	* use the within-computation forms listed in the table unless using
74	>	* async event-style tasks that are not usually joined, in which case
75	>	* there is little difference among choice of methods.
76		*
77		* <table BORDER CELLPADDING=3 CELLSPACING=1>
78		*  <tr>
#	Line 95 | Line 97 | import java.util.concurrent.locks.Condit
97		*  </tr>
98		* </table>
99		*
100	<	* <p><b>Sample Usage.</b> Normally a single {@code ForkJoinPool} is
101	<	* used for all parallel task execution in a program or subsystem.
102	<	* Otherwise, use would not usually outweigh the construction and
103	<	* bookkeeping overhead of creating a large set of threads. For
104	<	* example, a common pool could be used for the {@code SortTasks}
105	<	* illustrated in {@link RecursiveAction}. Because {@code
106	<	* ForkJoinPool} uses threads in {@linkplain java.lang.Thread#isDaemon
107	<	* daemon} mode, there is typically no need to explicitly {@link
108	<	* #shutdown} such a pool upon program exit.
107	<	*
108	<	*  <pre> {@code
109	<	* static final ForkJoinPool mainPool = new ForkJoinPool();
110	<	* ...
111	<	* public void sort(long[] array) {
112	<	*   mainPool.invoke(new SortTask(array, 0, array.length));
113	<	* }}</pre>
100	>	* <p>The common pool is by default constructed with default
101	>	* parameters, but these may be controlled by setting three {@link
102	>	* System#getProperty properties} with prefix {@code
103	>	* java.util.concurrent.ForkJoinPool.common}: {@code parallelism} --
104	>	* an integer greater than zero, {@code threadFactory} -- the class
105	>	* name of a {@link ForkJoinWorkerThreadFactory}, and {@code
106	>	* exceptionHandler} -- the class name of a {@link
107	>	* Thread.UncaughtExceptionHandler}. Upon any error in establishing
108	>	* these settings, default parameters are used.
109		*
110		* <p><b>Implementation notes</b>: This implementation restricts the
111		* maximum number of running threads to 32767. Attempts to create
#	Line 131 | Line 126 | public class ForkJoinPool extends Abstra
126		*
127		* This class and its nested classes provide the main
128		* functionality and control for a set of worker threads:
129	<	* Submissions from non-FJ threads enter into submission
130	<	* queues. Workers take these tasks and typically split them into
131	<	* subtasks that may be stolen by other workers.  Preference rules
132	<	* give first priority to processing tasks from their own queues
133	<	* (LIFO or FIFO, depending on mode), then to randomized FIFO
134	<	* steals of tasks in other queues.
129	>	* Submissions from non-FJ threads enter into submission queues.
130	>	* Workers take these tasks and typically split them into subtasks
131	>	* that may be stolen by other workers.  Preference rules give
132	>	* first priority to processing tasks from their own queues (LIFO
133	>	* or FIFO, depending on mode), then to randomized FIFO steals of
134	>	* tasks in other queues.
135		*
136	<	* WorkQueues.
136	>	* WorkQueues
137		* ==========
138		*
139		* Most operations occur within work-stealing queues (in nested
#	Line 156 | Line 151 | public class ForkJoinPool extends Abstra
151		* (http://research.sun.com/scalable/pubs/index.html) and
152		* "Idempotent work stealing" by Michael, Saraswat, and Vechev,
153		* PPoPP 2009 (http://portal.acm.org/citation.cfm?id=1504186).
154	<	* The main differences ultimately stem from gc requirements that
154	>	* The main differences ultimately stem from GC requirements that
155		* we null out taken slots as soon as we can, to maintain as small
156		* a footprint as possible even in programs generating huge
157		* numbers of tasks. To accomplish this, we shift the CAS
#	Line 178 | Line 173 | public class ForkJoinPool extends Abstra
173		* If an attempted steal fails, a thief always chooses a different
174		* random victim target to try next. So, in order for one thief to
175		* progress, it suffices for any in-progress poll or new push on
176	<	* any empty queue to complete.
176	>	* any empty queue to complete. (This is why we normally use
177	>	* method pollAt and its variants that try once at the apparent
178	>	* base index, else consider alternative actions, rather than
179	>	* method poll.)
180		*
181		* This approach also enables support of a user mode in which local
182		* task processing is in FIFO, not LIFO order, simply by using
#	Line 188 | Line 186 | public class ForkJoinPool extends Abstra
186		* rarely provide the best possible performance on a given
187		* machine, but portably provide good throughput by averaging over
188		* these factors.  (Further, even if we did try to use such
189	<	* information, we do not usually have a basis for exploiting
190	<	* it. For example, some sets of tasks profit from cache
191	<	* affinities, but others are harmed by cache pollution effects.)
189	>	* information, we do not usually have a basis for exploiting it.
190	>	* For example, some sets of tasks profit from cache affinities,
191	>	* but others are harmed by cache pollution effects.)
192		*
193		* WorkQueues are also used in a similar way for tasks submitted
194		* to the pool. We cannot mix these tasks in the same queues used
195		* for work-stealing (this would contaminate lifo/fifo
196	<	* processing). Instead, we loosely associate (via hashing)
197	<	* submission queues with submitting threads, and randomly scan
198	<	* these queues as well when looking for work. In essence,
199	<	* submitters act like workers except that they never take tasks,
200	<	* and they are multiplexed on to a finite number of shared work
201	<	* queues. However, classes are set up so that future extensions
202	<	* could allow submitters to optionally help perform tasks as
203	<	* well. Pool submissions from internal workers are also allowed,
204	<	* but use randomized rather than thread-hashed queue indices to
205	<	* avoid imbalance.  Insertion of tasks in shared mode requires a
206	<	* lock (mainly to protect in the case of resizing) but we use
207	<	* only a simple spinlock (using bits in field runState), because
208	<	* submitters encountering a busy queue try or create others so
209	<	* never block.
196	>	* processing). Instead, we randomly associate submission queues
197	>	* with submitting threads, using a form of hashing.  The
198	>	* ThreadLocal Submitter class contains a value initially used as
199	>	* a hash code for choosing existing queues, but may be randomly
200	>	* repositioned upon contention with other submitters.  In
201	>	* essence, submitters act like workers except that they are
202	>	* restricted to executing local tasks that they submitted (or in
203	>	* the case of CountedCompleters, others with the same root task).
204	>	* However, because most shared/external queue operations are more
205	>	* expensive than internal, and because, at steady state, external
206	>	* submitters will compete for CPU with workers, ForkJoinTask.join
207	>	* and related methods disable them from repeatedly helping to
208	>	* process tasks if all workers are active.  Insertion of tasks in
209	>	* shared mode requires a lock (mainly to protect in the case of
210	>	* resizing) but we use only a simple spinlock (using bits in
211	>	* field qlock), because submitters encountering a busy queue move
212	>	* on to try or create other queues -- they block only when
213	>	* creating and registering new queues.
214		*
215	<	* Management.
215	>	* Management
216		* ==========
217		*
218		* The main throughput advantages of work-stealing stem from
#	Line 220 | Line 222 | public class ForkJoinPool extends Abstra
222		* tactic for avoiding bottlenecks is packing nearly all
223		* essentially atomic control state into two volatile variables
224		* that are by far most often read (not written) as status and
225	<	* consistency checks
225	>	* consistency checks.
226		*
227		* Field "ctl" contains 64 bits holding all the information needed
228		* to atomically decide to add, inactivate, enqueue (on an event
#	Line 230 | Line 232 | public class ForkJoinPool extends Abstra
232		* and their negations (used for thresholding) to fit into 16bit
233		* fields.
234		*
235	<	* Field "runState" contains 32 bits needed to register and
236	<	* deregister WorkQueues, as well as to enable shutdown. It is
237	<	* only modified under a lock (normally briefly held, but
238	<	* occasionally protecting allocations and resizings) but even
239	<	* when locked remains available to check consistency.
235	>	* Field "plock" is a form of sequence lock with a saturating
236	>	* shutdown bit (similarly for per-queue "qlocks"), mainly
237	>	* protecting updates to the workQueues array, as well as to
238	>	* enable shutdown.  When used as a lock, it is normally only very
239	>	* briefly held, so is nearly always available after at most a
240	>	* brief spin, but we use a monitor-based backup strategy to
241	>	* blocking when needed.
242		*
243		* Recording WorkQueues.  WorkQueues are recorded in the
244	<	* "workQueues" array that is created upon pool construction and
245	<	* expanded if necessary.  Updates to the array while recording
246	<	* new workers and unrecording terminated ones are protected from
247	<	* each other by a lock but the array is otherwise concurrently
248	<	* readable, and accessed directly.  To simplify index-based
249	<	* operations, the array size is always a power of two, and all
250	<	* readers must tolerate null slots. Shared (submission) queues
251	<	* are at even indices, worker queues at odd indices. Grouping
252	<	* them together in this way simplifies and speeds up task
253	<	* scanning. To avoid flailing during start-up, the array is
254	<	* presized to hold twice #parallelism workers (which is unlikely
255	<	* to need further resizing during execution). But to avoid
256	<	* dealing with so many null slots, variable runState includes a
257	<	* mask for the nearest power of two that contains all current
254	<	* workers.  All worker thread creation is on-demand, triggered by
255	<	* task submissions, replacement of terminated workers, and/or
244	>	* "workQueues" array that is created upon first use and expanded
245	>	* if necessary.  Updates to the array while recording new workers
246	>	* and unrecording terminated ones are protected from each other
247	>	* by a lock but the array is otherwise concurrently readable, and
248	>	* accessed directly.  To simplify index-based operations, the
249	>	* array size is always a power of two, and all readers must
250	>	* tolerate null slots. Worker queues are at odd indices Shared
251	>	* (submission) queues are at even indices, up to a maximum of 64
252	>	* slots, to limit growth even if array needs to expand to add
253	>	* more workers. Grouping them together in this way simplifies and
254	>	* speeds up task scanning.
255	>	*
256	>	* All worker thread creation is on-demand, triggered by task
257	>	* submissions, replacement of terminated workers, and/or
258		* compensation for blocked workers. However, all other support
259		* code is set up to work with other policies.  To ensure that we
260		* do not hold on to worker references that would prevent GC, ALL
#	Line 265 | Line 267 | public class ForkJoinPool extends Abstra
267		* both index-check and null-check the IDs. All such accesses
268		* ignore bad IDs by returning out early from what they are doing,
269		* since this can only be associated with termination, in which
270	<	* case it is OK to give up.
271	<	*
272	<	* All uses of the workQueues array check that it is non-null
273	<	* (even if previously non-null). This allows nulling during
274	<	* termination, which is currently not necessary, but remains an
275	<	* option for resource-revocation-based shutdown schemes. It also
274	<	* helps reduce JIT issuance of uncommon-trap code, which tends to
270	>	* case it is OK to give up.  All uses of the workQueues array
271	>	* also check that it is non-null (even if previously
272	>	* non-null). This allows nulling during termination, which is
273	>	* currently not necessary, but remains an option for
274	>	* resource-revocation-based shutdown schemes. It also helps
275	>	* reduce JIT issuance of uncommon-trap code, which tends to
276		* unnecessarily complicate control flow in some methods.
277		*
278		* Event Queuing. Unlike HPC work-stealing frameworks, we cannot
#	Line 299 | Line 300 | public class ForkJoinPool extends Abstra
300		* some other queued worker rather than itself, which has the same
301		* net effect. Because enqueued workers may actually be rescanning
302		* rather than waiting, we set and clear the "parker" field of
303	<	* Workqueues to reduce unnecessary calls to unpark.  (This
303	>	* WorkQueues to reduce unnecessary calls to unpark.  (This
304		* requires a secondary recheck to avoid missed signals.)  Note
305		* the unusual conventions about Thread.interrupts surrounding
306		* parking and other blocking: Because interrupts are used solely
#	Line 311 | Line 312 | public class ForkJoinPool extends Abstra
312		*
313		* Signalling.  We create or wake up workers only when there
314		* appears to be at least one task they might be able to find and
315	<	* execute.  When a submission is added or another worker adds a
316	<	* task to a queue that previously had fewer than two tasks, they
317	<	* signal waiting workers (or trigger creation of new ones if
318	<	* fewer than the given parallelism level -- see signalWork).
319	<	* These primary signals are buttressed by signals during rescans;
320	<	* together these cover the signals needed in cases when more
321	<	* tasks are pushed but untaken, and improve performance compared
322	<	* to having one thread wake up all workers.
315	>	* execute. However, many other threads may notice the same task
316	>	* and each signal to wake up a thread that might take it. So in
317	>	* general, pools will be over-signalled.  When a submission is
318	>	* added or another worker adds a task to a queue that is
319	>	* apparently empty, they signal waiting workers (or trigger
320	>	* creation of new ones if fewer than the given parallelism level
321	>	* -- see signalWork).  These primary signals are buttressed by
322	>	* signals whenever other threads scan for work or do not have a
323	>	* task to process. On most platforms, signalling (unpark)
324	>	* overhead time is noticeably long, and the time between
325	>	* signalling a thread and it actually making progress can be very
326	>	* noticeably long, so it is worth offloading these delays from
327	>	* critical paths as much as possible.
328		*
329		* Trimming workers. To release resources after periods of lack of
330		* use, a worker starting to wait when the pool is quiescent will
331	<	* time out and terminate if the pool has remained quiescent for
332	<	* SHRINK_RATE nanosecs. This will slowly propagate, eventually
333	<	* terminating all workers after long periods of non-use.
331	>	* time out and terminate if the pool has remained quiescent for a
332	>	* given period -- a short period if there are more threads than
333	>	* parallelism, longer as the number of threads decreases. This
334	>	* will slowly propagate, eventually terminating all workers after
335	>	* periods of non-use.
336		*
337		* Shutdown and Termination. A call to shutdownNow atomically sets
338	<	* a runState bit and then (non-atomically) sets each workers
339	<	* runState status, cancels all unprocessed tasks, and wakes up
338	>	* a plock bit and then (non-atomically) sets each worker's
339	>	* qlock status, cancels all unprocessed tasks, and wakes up
340		* all waiting workers.  Detecting whether termination should
341		* commence after a non-abrupt shutdown() call requires more work
342		* and bookkeeping. We need consensus about quiescence (i.e., that
#	Line 336 | Line 344 | public class ForkJoinPool extends Abstra
344		* indication but non-abrupt shutdown still requires a rechecking
345		* scan for any workers that are inactive but not queued.
346		*
347	<	* Joining Tasks.
348	<	* ==============
347	>	* Joining Tasks
348	>	* =============
349		*
350		* Any of several actions may be taken when one worker is waiting
351	<	* to join a task stolen (or always held by) another.  Because we
351	>	* to join a task stolen (or always held) by another.  Because we
352		* are multiplexing many tasks on to a pool of workers, we can't
353		* just let them block (as in Thread.join).  We also cannot just
354		* reassign the joiner's run-time stack with another and replace
355		* it later, which would be a form of "continuation", that even if
356		* possible is not necessarily a good idea since we sometimes need
357	<	* both an unblocked task and its continuation to
358	<	* progress. Instead we combine two tactics:
357	>	* both an unblocked task and its continuation to progress.
358	>	* Instead we combine two tactics:
359		*
360		*   Helping: Arranging for the joiner to execute some task that it
361		*      would be running if the steal had not occurred.
#	Line 356 | Line 364 | public class ForkJoinPool extends Abstra
364		*      method tryCompensate() may create or re-activate a spare
365		*      thread to compensate for blocked joiners until they unblock.
366		*
367	<	* A third form (implemented in tryRemoveAndExec and
368	<	* tryPollForAndExec) amounts to helping a hypothetical
369	<	* compensator: If we can readily tell that a possible action of a
370	<	* compensator is to steal and execute the task being joined, the
371	<	* joining thread can do so directly, without the need for a
372	<	* compensation thread (although at the expense of larger run-time
373	<	* stacks, but the tradeoff is typically worthwhile).
367	>	* A third form (implemented in tryRemoveAndExec) amounts to
368	>	* helping a hypothetical compensator: If we can readily tell that
369	>	* a possible action of a compensator is to steal and execute the
370	>	* task being joined, the joining thread can do so directly,
371	>	* without the need for a compensation thread (although at the
372	>	* expense of larger run-time stacks, but the tradeoff is
373	>	* typically worthwhile).
374		*
375		* The ManagedBlocker extension API can't use helping so relies
376		* only on compensation in method awaitBlocker.
#	Line 382 | Line 390 | public class ForkJoinPool extends Abstra
390		* (http://portal.acm.org/citation.cfm?id=155354). It differs in
391		* that: (1) We only maintain dependency links across workers upon
392		* steals, rather than use per-task bookkeeping.  This sometimes
393	<	* requires a linear scan of workers array to locate stealers, but
394	<	* often doesn't because stealers leave hints (that may become
393	>	* requires a linear scan of workQueues array to locate stealers,
394	>	* but often doesn't because stealers leave hints (that may become
395		* stale/wrong) of where to locate them.  A stealHint is only a
396		* hint because a worker might have had multiple steals and the
397		* hint records only one of them (usually the most current).
#	Line 394 | Line 402 | public class ForkJoinPool extends Abstra
402		* which means that we miss links in the chain during long-lived
403		* tasks, GC stalls etc (which is OK since blocking in such cases
404		* is usually a good idea).  (4) We bound the number of attempts
405	<	* to find work (see MAX_HELP_DEPTH) and fall back to suspending
406	<	* the worker and if necessary replacing it with another.
405	>	* to find work (see MAX_HELP) and fall back to suspending the
406	>	* worker and if necessary replacing it with another.
407	>	*
408	>	* Helping actions for CountedCompleters are much simpler: Method
409	>	* helpComplete can take and execute any task with the same root
410	>	* as the task being waited on. However, this still entails some
411	>	* traversal of completer chains, so is less efficient than using
412	>	* CountedCompleters without explicit joins.
413		*
414		* It is impossible to keep exactly the target parallelism number
415		* of threads running at any given time.  Determining the
416		* existence of conservatively safe helping targets, the
417		* availability of already-created spares, and the apparent need
418		* to create new spares are all racy, so we rely on multiple
419	<	* retries of each.  Currently, in keeping with on-demand
420	<	* signalling policy, we compensate only if blocking would leave
421	<	* less than one active (non-waiting, non-blocked) worker.
422	<	* Additionally, to avoid some false alarms due to GC, lagging
423	<	* counters, system activity, etc, compensated blocking for joins
424	<	* is only attempted after rechecks stabilize in
425	<	* ForkJoinTask.awaitJoin. (Retries are interspersed with
426	<	* Thread.yield, for good citizenship.)
427	<	*
428	<	* Style notes: There is a lot of representation-level coupling
429	<	* among classes ForkJoinPool, ForkJoinWorkerThread, and
430	<	* ForkJoinTask.  The fields of WorkQueue maintain data structures
431	<	* managed by ForkJoinPool, so are directly accessed.  There is
432	<	* little point trying to reduce this, since any associated future
433	<	* changes in representations will need to be accompanied by
434	<	* algorithmic changes anyway. All together, these low-level
435	<	* implementation choices produce as much as a factor of 4
436	<	* performance improvement compared to naive implementations, and
437	<	* enable the processing of billions of tasks per second, at the
438	<	* expense of some ugliness.
439	<	*
440	<	* Methods signalWork() and scan() are the main bottlenecks so are
441	<	* especially heavily micro-optimized/mangled.  There are lots of
442	<	* inline assignments (of form "while ((local = field) != 0)")
443	<	* which are usually the simplest way to ensure the required read
444	<	* orderings (which are sometimes critical). This leads to a
445	<	* "C"-like style of listing declarations of these locals at the
446	<	* heads of methods or blocks.  There are several occurrences of
447	<	* the unusual "do {} while (!cas...)"  which is the simplest way
448	<	* to force an update of a CAS'ed variable. There are also other
449	<	* coding oddities that help some methods perform reasonably even
450	<	* when interpreted (not compiled).
451	<	*
452	<	* The order of declarations in this file is: (1) declarations of
453	<	* statics (2) fields (along with constants used when unpacking
454	<	* some of them), listed in an order that tends to reduce
455	<	* contention among them a bit under most JVMs; (3) nested
456	<	* classes; (4) internal control methods; (5) callbacks and other
457	<	* support for ForkJoinTask methods; (6) exported methods (plus a
458	<	* few little helpers); (7) static block initializing all statics
459	<	* in a minimally dependent order.
419	>	* retries of each.  Compensation in the apparent absence of
420	>	* helping opportunities is challenging to control on JVMs, where
421	>	* GC and other activities can stall progress of tasks that in
422	>	* turn stall out many other dependent tasks, without us being
423	>	* able to determine whether they will ever require compensation.
424	>	* Even though work-stealing otherwise encounters little
425	>	* degradation in the presence of more threads than cores,
426	>	* aggressively adding new threads in such cases entails risk of
427	>	* unwanted positive feedback control loops in which more threads
428	>	* cause more dependent stalls (as well as delayed progress of
429	>	* unblocked threads to the point that we know they are available)
430	>	* leading to more situations requiring more threads, and so
431	>	* on. This aspect of control can be seen as an (analytically
432	>	* intractable) game with an opponent that may choose the worst
433	>	* (for us) active thread to stall at any time.  We take several
434	>	* precautions to bound losses (and thus bound gains), mainly in
435	>	* methods tryCompensate and awaitJoin.
436	>	*
437	>	* Common Pool
438	>	* ===========
439	>	*
440	>	* The static commonPool always exists after static
441	>	* initialization.  Since it (or any other created pool) need
442	>	* never be used, we minimize initial construction overhead and
443	>	* footprint to the setup of about a dozen fields, with no nested
444	>	* allocation. Most bootstrapping occurs within method
445	>	* fullExternalPush during the first submission to the pool.
446	>	*
447	>	* When external threads submit to the common pool, they can
448	>	* perform some subtask processing (see externalHelpJoin and
449	>	* related methods).  We do not need to record whether these
450	>	* submissions are to the common pool -- if not, externalHelpJoin
451	>	* returns quicky (at the most helping to signal some common pool
452	>	* workers). These submitters would otherwise be blocked waiting
453	>	* for completion, so the extra effort (with liberally sprinkled
454	>	* task status checks) in inapplicable cases amounts to an odd
455	>	* form of limited spin-wait before blocking in ForkJoinTask.join.
456	>	*
457	>	* Style notes
458	>	* ===========
459	>	*
460	>	* There is a lot of representation-level coupling among classes
461	>	* ForkJoinPool, ForkJoinWorkerThread, and ForkJoinTask.  The
462	>	* fields of WorkQueue maintain data structures managed by
463	>	* ForkJoinPool, so are directly accessed.  There is little point
464	>	* trying to reduce this, since any associated future changes in
465	>	* representations will need to be accompanied by algorithmic
466	>	* changes anyway. Several methods intrinsically sprawl because
467	>	* they must accumulate sets of consistent reads of volatiles held
468	>	* in local variables.  Methods signalWork() and scan() are the
469	>	* main bottlenecks, so are especially heavily
470	>	* micro-optimized/mangled.  There are lots of inline assignments
471	>	* (of form "while ((local = field) != 0)") which are usually the
472	>	* simplest way to ensure the required read orderings (which are
473	>	* sometimes critical). This leads to a "C"-like style of listing
474	>	* declarations of these locals at the heads of methods or blocks.
475	>	* There are several occurrences of the unusual "do {} while
476	>	* (!cas...)"  which is the simplest way to force an update of a
477	>	* CAS'ed variable. There are also other coding oddities (including
478	>	* several unnecessary-looking hoisted null checks) that help
479	>	* some methods perform reasonably even when interpreted (not
480	>	* compiled).
481	>	*
482	>	* The order of declarations in this file is:
483	>	* (1) Static utility functions
484	>	* (2) Nested (static) classes
485	>	* (3) Static fields
486	>	* (4) Fields, along with constants used when unpacking some of them
487	>	* (5) Internal control methods
488	>	* (6) Callbacks and other support for ForkJoinTask methods
489	>	* (7) Exported methods
490	>	* (8) Static block initializing statics in minimally dependent order
491	>	*/
492	>
493	>	// Static utilities
494	>
495	>	/**
496	>	* If there is a security manager, makes sure caller has
497	>	* permission to modify threads.
498		*/
499	+	private static void checkPermission() {
500	+	SecurityManager security = System.getSecurityManager();
501	+	if (security != null)
502	+	security.checkPermission(modifyThreadPermission);
503	+	}
504	+
505	+	// Nested classes
506
507		/**
508		* Factory for creating new {@link ForkJoinWorkerThread}s.
#	Line 473 | Line 532 | public class ForkJoinPool extends Abstra
532		}
533
534		/**
535	<	* Creates a new ForkJoinWorkerThread. This factory is used unless
536	<	* overridden in ForkJoinPool constructors.
537	<	*/
538	<	public static final ForkJoinWorkerThreadFactory
480	<	defaultForkJoinWorkerThreadFactory;
481	<
482	<	/**
483	<	* Permission required for callers of methods that may start or
484	<	* kill threads.
485	<	*/
486	<	private static final RuntimePermission modifyThreadPermission;
487	<
488	<	/**
489	<	* If there is a security manager, makes sure caller has
490	<	* permission to modify threads.
535	>	* Class for artificial tasks that are used to replace the target
536	>	* of local joins if they are removed from an interior queue slot
537	>	* in WorkQueue.tryRemoveAndExec. We don't need the proxy to
538	>	* actually do anything beyond having a unique identity.
539		*/
540	<	private static void checkPermission() {
541	<	SecurityManager security = System.getSecurityManager();
542	<	if (security != null)
543	<	security.checkPermission(modifyThreadPermission);
540	>	static final class EmptyTask extends ForkJoinTask<Void> {
541	>	private static final long serialVersionUID = -7721805057305804111L;
542	>	EmptyTask() { status = ForkJoinTask.NORMAL; } // force done
543	>	public final Void getRawResult() { return null; }
544	>	public final void setRawResult(Void x) {}
545	>	public final boolean exec() { return true; }
546		}
547
548		/**
499	–	* Generator for assigning sequence numbers as pool names.
500	–	*/
501	–	private static final AtomicInteger poolNumberGenerator;
502	–
503	–	/**
504	–	* Bits and masks for control variables
505	–	*
506	–	* Field ctl is a long packed with:
507	–	* AC: Number of active running workers minus target parallelism (16 bits)
508	–	* TC: Number of total workers minus target parallelism (16 bits)
509	–	* ST: true if pool is terminating (1 bit)
510	–	* EC: the wait count of top waiting thread (15 bits)
511	–	* ID: ~(poolIndex >>> 1) of top of Treiber stack of waiters (16 bits)
512	–	*
513	–	* When convenient, we can extract the upper 32 bits of counts and
514	–	* the lower 32 bits of queue state, u = (int)(ctl >>> 32) and e =
515	–	* (int)ctl.  The ec field is never accessed alone, but always
516	–	* together with id and st. The offsets of counts by the target
517	–	* parallelism and the positionings of fields makes it possible to
518	–	* perform the most common checks via sign tests of fields: When
519	–	* ac is negative, there are not enough active workers, when tc is
520	–	* negative, there are not enough total workers, when id is
521	–	* negative, there is at least one waiting worker, and when e is
522	–	* negative, the pool is terminating.  To deal with these possibly
523	–	* negative fields, we use casts in and out of "short" and/or
524	–	* signed shifts to maintain signedness.
525	–	*
526	–	* When a thread is queued (inactivated), its eventCount field is
527	–	* negative, which is the only way to tell if a worker is
528	–	* prevented from executing tasks, even though it must continue to
529	–	* scan for them to avoid queuing races.
530	–	*
531	–	* Field runState is an int packed with:
532	–	* SHUTDOWN: true if shutdown is enabled (1 bit)
533	–	* SEQ:  a sequence number updated upon (de)registering workers (15 bits)
534	–	* MASK: mask (power of 2 - 1) covering all registered poolIndexes (16 bits)
535	–	*
536	–	* The combination of mask and sequence number enables simple
537	–	* consistency checks: Staleness of read-only operations on the
538	–	* workers and queues arrays can be checked by comparing runState
539	–	* before vs after the reads. The low 16 bits (i.e, anding with
540	–	* SMASK) hold (the smallest power of two covering all worker
541	–	* indices, minus one.  The mask for queues (vs workers) is twice
542	–	* this value plus 1.
543	–	*/
544	–
545	–	// bit positions/shifts for fields
546	–	private static final int  AC_SHIFT   = 48;
547	–	private static final int  TC_SHIFT   = 32;
548	–	private static final int  ST_SHIFT   = 31;
549	–	private static final int  EC_SHIFT   = 16;
550	–
551	–	// bounds
552	–	private static final int  MAX_ID     = 0x7fff;  // max poolIndex
553	–	private static final int  SMASK      = 0xffff;  // mask short bits
554	–	private static final int  SHORT_SIGN = 1 << 15;
555	–	private static final int  INT_SIGN   = 1 << 31;
556	–
557	–	// masks
558	–	private static final long STOP_BIT   = 0x0001L << ST_SHIFT;
559	–	private static final long AC_MASK    = ((long)SMASK) << AC_SHIFT;
560	–	private static final long TC_MASK    = ((long)SMASK) << TC_SHIFT;
561	–
562	–	// units for incrementing and decrementing
563	–	private static final long TC_UNIT    = 1L << TC_SHIFT;
564	–	private static final long AC_UNIT    = 1L << AC_SHIFT;
565	–
566	–	// masks and units for dealing with u = (int)(ctl >>> 32)
567	–	private static final int  UAC_SHIFT  = AC_SHIFT - 32;
568	–	private static final int  UTC_SHIFT  = TC_SHIFT - 32;
569	–	private static final int  UAC_MASK   = SMASK << UAC_SHIFT;
570	–	private static final int  UTC_MASK   = SMASK << UTC_SHIFT;
571	–	private static final int  UAC_UNIT   = 1 << UAC_SHIFT;
572	–	private static final int  UTC_UNIT   = 1 << UTC_SHIFT;
573	–
574	–	// masks and units for dealing with e = (int)ctl
575	–	private static final int E_MASK      = 0x7fffffff; // no STOP_BIT
576	–	private static final int E_SEQ       = 1 << EC_SHIFT;
577	–
578	–	// runState bits
579	–	private static final int SHUTDOWN    = 1 << 31;
580	–	private static final int RS_SEQ      = 1 << 16;
581	–	private static final int RS_SEQ_MASK = 0x7fff0000;
582	–
583	–	// access mode for WorkQueue
584	–	static final int LIFO_QUEUE          =  0;
585	–	static final int FIFO_QUEUE          =  1;
586	–	static final int SHARED_QUEUE        = -1;
587	–
588	–	/**
589	–	* The wakeup interval (in nanoseconds) for a worker waiting for a
590	–	* task when the pool is quiescent to instead try to shrink the
591	–	* number of workers.  The exact value does not matter too
592	–	* much. It must be short enough to release resources during
593	–	* sustained periods of idleness, but not so short that threads
594	–	* are continually re-created.
595	–	*/
596	–	private static final long SHRINK_RATE =
597	–	4L * 1000L * 1000L * 1000L; // 4 seconds
598	–
599	–	/**
600	–	* The timeout value for attempted shrinkage, includes
601	–	* some slop to cope with system timer imprecision.
602	–	*/
603	–	private static final long SHRINK_TIMEOUT = SHRINK_RATE - (SHRINK_RATE / 10);
604	–
605	–	/**
606	–	* The maximum stolen->joining link depth allowed in tryHelpStealer.
607	–	* Depths for legitimate chains are unbounded, but we use a fixed
608	–	* constant to avoid (otherwise unchecked) cycles and to bound
609	–	* staleness of traversal parameters at the expense of sometimes
610	–	* blocking when we could be helping.
611	–	*/
612	–	private static final int MAX_HELP_DEPTH = 16;
613	–
614	–	/*
615	–	* Field layout order in this class tends to matter more than one
616	–	* would like. Runtime layout order is only loosely related to
617	–	* declaration order and may differ across JVMs, but the following
618	–	* empirically works OK on current JVMs.
619	–	*/
620	–
621	–	volatile long ctl;                       // main pool control
622	–	final int parallelism;                   // parallelism level
623	–	final int localMode;                     // per-worker scheduling mode
624	–	int nextPoolIndex;                       // hint used in registerWorker
625	–	volatile int runState;                   // shutdown status, seq, and mask
626	–	WorkQueue[] workQueues;                  // main registry
627	–	final ReentrantLock lock;                // for registration
628	–	final Condition termination;             // for awaitTermination
629	–	final ForkJoinWorkerThreadFactory factory; // factory for new workers
630	–	final Thread.UncaughtExceptionHandler ueh; // per-worker UEH
631	–	final AtomicLong stealCount;             // collect counts when terminated
632	–	final AtomicInteger nextWorkerNumber;    // to create worker name string
633	–	final String workerNamePrefix;           // Prefix for assigning worker names
634	–
635	–	/**
549		* Queues supporting work-stealing as well as external task
550		* submission. See above for main rationale and algorithms.
551		* Implementation relies heavily on "Unsafe" intrinsics
#	Line 646 | Line 559 | public class ForkJoinPool extends Abstra
559		*
560		* Field "top" is the index (mod array.length) of the next queue
561		* slot to push to or pop from. It is written only by owner thread
562	<	* for push, or under lock for trySharedPush, and accessed by
563	<	* other threads only after reading (volatile) base.  Both top and
564	<	* base are allowed to wrap around on overflow, but (top - base)
565	<	* (or more commonly -(base - top) to force volatile read of base
566	<	* before top) still estimates size.
562	>	* for push, or under lock for external/shared push, and accessed
563	>	* by other threads only after reading (volatile) base.  Both top
564	>	* and base are allowed to wrap around on overflow, but (top -
565	>	* base) (or more commonly -(base - top) to force volatile read of
566	>	* base before top) still estimates size. The lock ("qlock") is
567	>	* forced to -1 on termination, causing all further lock attempts
568	>	* to fail. (Note: we don't need CAS for termination state because
569	>	* upon pool shutdown, all shared-queues will stop being used
570	>	* anyway.)  Nearly all lock bodies are set up so that exceptions
571	>	* within lock bodies are "impossible" (modulo JVM errors that
572	>	* would cause failure anyway.)
573		*
574		* The array slots are read and written using the emulation of
575		* volatiles/atomics provided by Unsafe. Insertions must in
576		* general use putOrderedObject as a form of releasing store to
577		* ensure that all writes to the task object are ordered before
578	<	* its publication in the queue. (Although we can avoid one case
579	<	* of this when locked in trySharedPush.) All removals entail a
580	<	* CAS to null.  The array is always a power of two. To ensure
581	<	* safety of Unsafe array operations, all accesses perform
663	<	* explicit null checks and implicit bounds checks via
664	<	* power-of-two masking.
578	>	* its publication in the queue.  All removals entail a CAS to
579	>	* null.  The array is always a power of two. To ensure safety of
580	>	* Unsafe array operations, all accesses perform explicit null
581	>	* checks and implicit bounds checks via power-of-two masking.
582		*
583		* In addition to basic queuing support, this class contains
584		* fields described elsewhere to control execution. It turns out
585	<	* to work better memory-layout-wise to include them in this
586	<	* class rather than a separate class.
585	>	* to work better memory-layout-wise to include them in this class
586	>	* rather than a separate class.
587		*
588		* Performance on most platforms is very sensitive to placement of
589		* instances of both WorkQueues and their arrays -- we absolutely
#	Line 681 | Line 598 | public class ForkJoinPool extends Abstra
598		* avoiding really bad worst-case access. (Until better JVM
599		* support is in place, this padding is dependent on transient
600		* properties of JVM field layout rules.)  We also take care in
601	<	* allocating and sizing and resizing the array. Non-shared queue
602	<	* arrays are initialized (via method growArray) by workers before
603	<	* use. Others are allocated on first use.
601	>	* allocating, sizing and resizing the array. Non-shared queue
602	>	* arrays are initialized by workers before use. Others are
603	>	* allocated on first use.
604		*/
605		static final class WorkQueue {
606		/**
607		* Capacity of work-stealing queue array upon initialization.
608	<	* Must be a power of two; at least 4, but set larger to
609	<	* reduce cacheline sharing among queues.
608	>	* Must be a power of two; at least 4, but should be larger to
609	>	* reduce or eliminate cacheline sharing among queues.
610	>	* Currently, it is much larger, as a partial workaround for
611	>	* the fact that JVMs often place arrays in locations that
612	>	* share GC bookkeeping (especially cardmarks) such that
613	>	* per-write accesses encounter serious memory contention.
614		*/
615	<	static final int INITIAL_QUEUE_CAPACITY = 1 << 8;
615	>	static final int INITIAL_QUEUE_CAPACITY = 1 << 13;
616
617		/**
618		* Maximum size for queue arrays. Must be a power of two less
#	Line 702 | Line 623 | public class ForkJoinPool extends Abstra
623		*/
624		static final int MAXIMUM_QUEUE_CAPACITY = 1 << 26; // 64M
625
705	–	volatile long totalSteals; // cumulative number of steals
626		int seed;                  // for random scanning; initialize nonzero
627		volatile int eventCount;   // encoded inactivation count; < 0 if inactive
628		int nextWait;              // encoded record of next event waiter
709	–	int rescans;               // remaining scans until block
710	–	int nsteals;               // top-level task executions since last idle
629		final int mode;            // lifo, fifo, or shared
630	+	int nsteals;               // cumulative number of steals
631		int poolIndex;             // index of this queue in pool (or 0)
632		int stealHint;             // index of most recent known stealer
633	<	volatile int runState;     // 1: locked, -1: terminate; else 0
633	>	volatile int qlock;        // 1: locked, -1: terminate; else 0
634		volatile int base;         // index of next slot for poll
635		int top;                   // index of next slot for push
636		ForkJoinTask<?>[] array;   // the elements (initially unallocated)
637	+	final ForkJoinPool pool;   // the containing pool (may be null)
638		final ForkJoinWorkerThread owner; // owning thread or null if shared
639		volatile Thread parker;    // == owner during call to park; else null
640	<	ForkJoinTask<?> currentJoin;  // task being joined in awaitJoin
640	>	volatile ForkJoinTask<?> currentJoin;  // task being joined in awaitJoin
641		ForkJoinTask<?> currentSteal; // current non-local task being executed
642		// Heuristic padding to ameliorate unfortunate memory placements
643	<	Object p00, p01, p02, p03, p04, p05, p06, p07, p08, p09, p0a;
643	>	Object p00, p01, p02, p03, p04, p05, p06, p07;
644	>	Object p08, p09, p0a, p0b, p0c, p0d, p0e;
645
646	<	WorkQueue(ForkJoinWorkerThread owner, int mode) {
726	<	this.owner = owner;
646	>	WorkQueue(ForkJoinPool pool, ForkJoinWorkerThread owner, int mode) {
647		this.mode = mode;
648	+	this.pool = pool;
649	+	this.owner = owner;
650		// Place indices in the center of array (that is not yet allocated)
651		base = top = INITIAL_QUEUE_CAPACITY >>> 1;
652		}
653
654		/**
733	–	* Returns number of tasks in the queue
734	–	*/
735	–	final int queueSize() {
736	–	int n = base - top; // non-owner callers must read base first
737	–	return (n >= 0) ? 0 : -n;
738	–	}
739	–
740	–	/**
655		* Pushes a task. Call only by owner in unshared queues.
656	+	* Cases needing resizing or rejection are relyaed to fullPush
657	+	* (that also handles shared queues).
658		*
659		* @param task the task. Caller must ensure non-null.
660	<	* @param p, if non-null, pool to signal if necessary
661	<	* @throw RejectedExecutionException if array cannot
662	<	* be resized
663	<	*/
664	<	final void push(ForkJoinTask<?> task, ForkJoinPool p) {
665	<	ForkJoinTask<?>[] a;
750	<	int s = top, m, n;
751	<	if ((a = array) != null) {    // ignore if queue removed
660	>	* @throw RejectedExecutionException if array cannot be resized
661	>	*/
662	>	final void push(ForkJoinTask<?> task) {
663	>	ForkJoinPool p; ForkJoinTask<?>[] a;
664	>	int s = top, n;
665	>	if ((a = array) != null && a.length > (n = s + 1 - base)) {
666		U.putOrderedObject
667	<	(a, (((m = a.length - 1) & s) << ASHIFT) + ABASE, task);
668	<	if ((n = (top = s + 1) - base) <= 2) {
669	<	if (p != null)
670	<	p.signalWork();
757	<	}
758	<	else if (n >= m)
759	<	growArray(true);
667	>	(a, (((a.length - 1) & s) << ASHIFT) + ABASE, task);
668	>	top = s + 1;
669	>	if (n <= 1 && (p = pool) != null)
670	>	p.signalWork(this, 1);
671		}
672	+	else
673	+	fullPush(task, true);
674		}
675
676		/**
677		* Pushes a task if lock is free and array is either big
678	<	* enough or can be resized to be big enough.
678	>	* enough or can be resized to be big enough. Note: a
679	>	* specialization of a common fast path of this method is in
680	>	* ForkJoinPool.externalPush. When called from a FJWT queue,
681	>	* this can fail only if the pool has been shut down or
682	>	* an out of memory error.
683		*
684		* @param task the task. Caller must ensure non-null.
685	<	* @return true if submitted
685	>	* @param owned if true, throw RJE on failure
686		*/
687	<	final boolean trySharedPush(ForkJoinTask<?> task) {
688	<	boolean submitted = false;
689	<	if (runState == 0 && U.compareAndSwapInt(this, RUNSTATE, 0, 1)) {
690	<	ForkJoinTask<?>[] a = array;
691	<	int s = top, n = s - base;
692	<	try {
693	<	if ((a != null && n < a.length - 1) ||
694	<	(a = growArray(false)) != null) { // must presize
695	<	int j = (((a.length - 1) & s) << ASHIFT) + ABASE;
696	<	U.putObject(a, (long)j, task);    // don't need "ordered"
697	<	top = s + 1;
698	<	submitted = true;
687	>	final boolean fullPush(ForkJoinTask<?> task, boolean owned) {
688	>	ForkJoinPool p; ForkJoinTask<?>[] a;
689	>	if (owned) {
690	>	if (qlock < 0) // must be shutting down
691	>	throw new RejectedExecutionException();
692	>	}
693	>	else if (!U.compareAndSwapInt(this, QLOCK, 0, 1))
694	>	return false;
695	>	try {
696	>	int s = top, oldLen, len;
697	>	if ((a = array) == null)
698	>	a = array = new ForkJoinTask<?>[len=INITIAL_QUEUE_CAPACITY];
699	>	else if ((oldLen = a.length) > s + 1 - base)
700	>	len = oldLen;
701	>	else if ((len = oldLen << 1) > MAXIMUM_QUEUE_CAPACITY)
702	>	throw new RejectedExecutionException("Capacity exceeded");
703	>	else {
704	>	int oldMask, b;
705	>	ForkJoinTask<?>[] oldA = a;
706	>	a = array = new ForkJoinTask<?>[len];
707	>	if ((oldMask = oldLen - 1) >= 0 && s - (b = base) > 0) {
708	>	int mask = len - 1;
709	>	do {
710	>	ForkJoinTask<?> x;
711	>	int oldj = ((b & oldMask) << ASHIFT) + ABASE;
712	>	int j    = ((b &    mask) << ASHIFT) + ABASE;
713	>	x = (ForkJoinTask<?>)
714	>	U.getObjectVolatile(oldA, oldj);
715	>	if (x != null &&
716	>	U.compareAndSwapObject(oldA, oldj, x, null))
717	>	U.putObjectVolatile(a, j, x);
718	>	} while (++b != s);
719		}
783	–	} finally {
784	–	runState = 0;                         // unlock
720		}
721	+	U.putOrderedObject
722	+	(a, (((len - 1) & s) << ASHIFT) + ABASE, task);
723	+	top = s + 1;
724	+	} finally {
725	+	if (!owned)
726	+	qlock = 0;
727		}
728	<	return submitted;
728	>	if ((p = pool) != null)
729	>	p.signalWork(this, 1);
730	>	return true;
731		}
732
733		/**
734	<	* Takes next task, if one exists, in FIFO order.
734	>	* Takes next task, if one exists, in LIFO order.  Call only
735	>	* by owner in unshared queues.
736		*/
737	<	final ForkJoinTask<?> poll() {
738	<	ForkJoinTask<?>[] a; int b, i;
739	<	while ((b = base) - top < 0 && (a = array) != null &&
740	<	(i = (a.length - 1) & b) >= 0) {
741	<	int j = (i << ASHIFT) + ABASE;
742	<	ForkJoinTask<?> t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
743	<	if (t != null && base == b &&
737	>	final ForkJoinTask<?> pop() {
738	>	ForkJoinTask<?>[] a; ForkJoinTask<?> t; int m;
739	>	if ((a = array) != null && (m = a.length - 1) >= 0) {
740	>	for (int s; (s = top - 1) - base >= 0;) {
741	>	long j = ((m & s) << ASHIFT) + ABASE;
742	>	if ((t = (ForkJoinTask<?>)U.getObject(a, j)) == null)
743	>	break;
744	>	if (U.compareAndSwapObject(a, j, t, null)) {
745	>	top = s;
746	>	return t;
747	>	}
748	>	}
749	>	}
750	>	return null;
751	>	}
752	>
753	>	/**
754	>	* Takes a task in FIFO order if b is base of queue and a task
755	>	* can be claimed without contention. Specialized versions
756	>	* appear in ForkJoinPool methods scan and tryHelpStealer.
757	>	*/
758	>	final ForkJoinTask<?> pollAt(int b) {
759	>	ForkJoinTask<?> t; ForkJoinTask<?>[] a;
760	>	if ((a = array) != null) {
761	>	int j = (((a.length - 1) & b) << ASHIFT) + ABASE;
762	>	if ((t = (ForkJoinTask<?>)U.getObjectVolatile(a, j)) != null &&
763	>	base == b &&
764		U.compareAndSwapObject(a, j, t, null)) {
765		base = b + 1;
766		return t;
#	Line 806 | Line 770 | public class ForkJoinPool extends Abstra
770		}
771
772		/**
773	<	* Takes next task, if one exists, in LIFO order.
810	<	* Call only by owner in unshared queues.
773	>	* Takes next task, if one exists, in FIFO order.
774		*/
775	<	final ForkJoinTask<?> pop() {
776	<	ForkJoinTask<?> t; int m;
777	<	ForkJoinTask<?>[] a = array;
778	<	if (a != null && (m = a.length - 1) >= 0) {
779	<	for (int s; (s = top - 1) - base >= 0;) {
780	<	int j = ((m & s) << ASHIFT) + ABASE;
781	<	if ((t = (ForkJoinTask<?>)U.getObjectVolatile(a, j)) == null)
782	<	break;
783	<	if (U.compareAndSwapObject(a, j, t, null)) {
821	<	top = s;
775	>	final ForkJoinTask<?> poll() {
776	>	ForkJoinTask<?>[] a; int b; ForkJoinTask<?> t;
777	>	while ((b = base) - top < 0 && (a = array) != null) {
778	>	int j = (((a.length - 1) & b) << ASHIFT) + ABASE;
779	>	t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
780	>	if (t != null) {
781	>	if (base == b &&
782	>	U.compareAndSwapObject(a, j, t, null)) {
783	>	base = b + 1;
784		return t;
785		}
786		}
787	+	else if (base == b) {
788	+	if (b + 1 == top)
789	+	break;
790	+	Thread.yield(); // wait for lagging update (very rare)
791	+	}
792		}
793		return null;
794		}
#	Line 846 | Line 813 | public class ForkJoinPool extends Abstra
813		}
814
815		/**
849	–	* Returns task at index b if b is current base of queue.
850	–	*/
851	–	final ForkJoinTask<?> pollAt(int b) {
852	–	ForkJoinTask<?>[] a; int i;
853	–	ForkJoinTask<?> task = null;
854	–	if ((a = array) != null && (i = ((a.length - 1) & b)) >= 0) {
855	–	int j = (i << ASHIFT) + ABASE;
856	–	ForkJoinTask<?> t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
857	–	if (t != null && base == b &&
858	–	U.compareAndSwapObject(a, j, t, null)) {
859	–	base = b + 1;
860	–	task = t;
861	–	}
862	–	}
863	–	return task;
864	–	}
865	–
866	–	/**
816		* Pops the given task only if it is at the current top.
817	+	* (A shared version is available only via FJP.tryExternalUnpush)
818		*/
819		final boolean tryUnpush(ForkJoinTask<?> t) {
820		ForkJoinTask<?>[] a; int s;
#	Line 878 | Line 828 | public class ForkJoinPool extends Abstra
828		}
829
830		/**
831	<	* Polls the given task only if it is at the current base.
831	>	* Removes and cancels all known tasks, ignoring any exceptions.
832		*/
833	<	final boolean pollFor(ForkJoinTask<?> task) {
834	<	ForkJoinTask<?>[] a; int b, i;
835	<	if ((b = base) - top < 0 && (a = array) != null &&
836	<	(i = (a.length - 1) & b) >= 0) {
837	<	int j = (i << ASHIFT) + ABASE;
838	<	if (U.getObjectVolatile(a, j) == task && base == b &&
839	<	U.compareAndSwapObject(a, j, task, null)) {
840	<	base = b + 1;
841	<	return true;
833	>	final void cancelAll() {
834	>	ForkJoinTask.cancelIgnoringExceptions(currentJoin);
835	>	ForkJoinTask.cancelIgnoringExceptions(currentSteal);
836	>	for (ForkJoinTask<?> t; (t = poll()) != null; )
837	>	ForkJoinTask.cancelIgnoringExceptions(t);
838	>	}
839	>
840	>	/**
841	>	* Computes next value for random probes.  Scans don't require
842	>	* a very high quality generator, but also not a crummy one.
843	>	* Marsaglia xor-shift is cheap and works well enough.  Note:
844	>	* This is manually inlined in its usages in ForkJoinPool to
845	>	* avoid writes inside busy scan loops.
846	>	*/
847	>	final int nextSeed() {
848	>	int r = seed;
849	>	r ^= r << 13;
850	>	r ^= r >>> 17;
851	>	return seed = r ^= r << 5;
852	>	}
853	>
854	>	/**
855	>	* Provides a more accurate estimate of size than (top - base)
856	>	* by ordering reads and checking whether a near-empty queue
857	>	* has at least one unclaimed task.
858	>	*/
859	>	final int queueSize() {
860	>	ForkJoinTask<?>[] a; int k, s, n;
861	>	return ((n = base - (s = top)) < 0 &&
862	>	(n != -1 ||
863	>	((a = array) != null && (k = a.length) > 0 &&
864	>	U.getObject
865	>	(a, (long)((((k - 1) & (s - 1)) << ASHIFT) + ABASE)) != null))) ?
866	>	-n : 0;
867	>	}
868	>
869	>	// Specialized execution methods
870	>
871	>	/**
872	>	* Pops and runs tasks until empty.
873	>	*/
874	>	private void popAndExecAll() {
875	>	// A bit faster than repeated pop calls
876	>	ForkJoinTask<?>[] a; int m, s; long j; ForkJoinTask<?> t;
877	>	while ((a = array) != null && (m = a.length - 1) >= 0 &&
878	>	(s = top - 1) - base >= 0 &&
879	>	(t = ((ForkJoinTask<?>)
880	>	U.getObject(a, j = ((m & s) << ASHIFT) + ABASE)))
881	>	!= null) {
882	>	if (U.compareAndSwapObject(a, j, t, null)) {
883	>	top = s;
884	>	t.doExec();
885		}
886		}
894	–	return false;
887		}
888
889		/**
890	<	* If present, removes from queue and executes the given task, or
891	<	* any other cancelled task. Returns (true) immediately on any CAS
890	>	* Polls and runs tasks until empty.
891	>	*/
892	>	private void pollAndExecAll() {
893	>	for (ForkJoinTask<?> t; (t = poll()) != null;)
894	>	t.doExec();
895	>	}
896	>
897	>	/**
898	>	* If present, removes from queue and executes the given task,
899	>	* or any other cancelled task. Returns (true) on any CAS
900		* or consistency check failure so caller can retry.
901		*
902	<	* @return false if no progress can be made
902	>	* @return false if no progress can be made, else true;
903		*/
904		final boolean tryRemoveAndExec(ForkJoinTask<?> task) {
905	<	boolean removed = false, empty = true, progress = true;
905	>	boolean stat = true, removed = false, empty = true;
906		ForkJoinTask<?>[] a; int m, s, b, n;
907		if ((a = array) != null && (m = a.length - 1) >= 0 &&
908		(n = (s = top) - (b = base)) > 0) {
#	Line 932 | Line 932 | public class ForkJoinPool extends Abstra
932		}
933		if (--n == 0) {
934		if (!empty && base == b)
935	<	progress = false;
935	>	stat = false;
936		break;
937		}
938		}
939		}
940		if (removed)
941		task.doExec();
942	<	return progress;
942	>	return stat;
943		}
944
945		/**
946	<	* Initializes or doubles the capacity of array. Call either
947	<	* by owner or with lock held -- it is OK for base, but not
948	<	* top, to move while resizings are in progress.
949	<	*
950	<	* @param rejectOnFailure if true, throw exception if capacity
951	<	* exceeded (relayed ultimately to user); else return null.
946	>	* Polls for and executes the given task or any other task in
947	>	* its CountedCompleter computation
948		*/
949	<	final ForkJoinTask<?>[] growArray(boolean rejectOnFailure) {
950	<	ForkJoinTask<?>[] oldA = array;
951	<	int size = oldA != null ? oldA.length << 1 : INITIAL_QUEUE_CAPACITY;
952	<	if (size <= MAXIMUM_QUEUE_CAPACITY) {
953	<	int oldMask, t, b;
954	<	ForkJoinTask<?>[] a = array = new ForkJoinTask<?>[size];
955	<	if (oldA != null && (oldMask = oldA.length - 1) >= 0 &&
956	<	(t = top) - (b = base) > 0) {
957	<	int mask = size - 1;
958	<	do {
959	<	ForkJoinTask<?> x;
960	<	int oldj = ((b & oldMask) << ASHIFT) + ABASE;
961	<	int j    = ((b &    mask) << ASHIFT) + ABASE;
962	<	x = (ForkJoinTask<?>)U.getObjectVolatile(oldA, oldj);
963	<	if (x != null &&
964	<	U.compareAndSwapObject(oldA, oldj, x, null))
965	<	U.putObjectVolatile(a, j, x);
966	<	} while (++b != t);
949	>	final boolean pollAndExecCC(ForkJoinTask<?> root) {
950	>	ForkJoinTask<?>[] a; int b; Object o;
951	>	outer: while ((b = base) - top < 0 && (a = array) != null) {
952	>	long j = (((a.length - 1) & b) << ASHIFT) + ABASE;
953	>	if ((o = U.getObject(a, j)) == null ||
954	>	!(o instanceof CountedCompleter))
955	>	break;
956	>	for (CountedCompleter<?> t = (CountedCompleter<?>)o, r = t;;) {
957	>	if (r == root) {
958	>	if (base == b &&
959	>	U.compareAndSwapObject(a, j, t, null)) {
960	>	base = b + 1;
961	>	t.doExec();
962	>	return true;
963	>	}
964	>	else
965	>	break; // restart
966	>	}
967	>	if ((r = r.completer) == null)
968	>	break outer; // not part of root computation
969		}
972	–	return a;
973	–	}
974	–	else if (!rejectOnFailure)
975	–	return null;
976	–	else
977	–	throw new RejectedExecutionException("Queue capacity exceeded");
978	–	}
979	–
980	–	/**
981	–	* Removes and cancels all known tasks, ignoring any exceptions
982	–	*/
983	–	final void cancelAll() {
984	–	ForkJoinTask.cancelIgnoringExceptions(currentJoin);
985	–	ForkJoinTask.cancelIgnoringExceptions(currentSteal);
986	–	for (ForkJoinTask<?> t; (t = poll()) != null; )
987	–	ForkJoinTask.cancelIgnoringExceptions(t);
988	–	}
989	–
990	–	// Execution methods
991	–
992	–	/**
993	–	* Removes and runs tasks until empty, using local mode
994	–	* ordering.
995	–	*/
996	–	final void runLocalTasks() {
997	–	if (base - top < 0) {
998	–	for (ForkJoinTask<?> t; (t = nextLocalTask()) != null; )
999	–	t.doExec();
970		}
971	+	return false;
972		}
973
974		/**
975		* Executes a top-level task and any local tasks remaining
976		* after execution.
1006	–	*
1007	–	* @return true unless terminating
977		*/
978	<	final boolean runTask(ForkJoinTask<?> t) {
1010	<	boolean alive = true;
978	>	final void runTask(ForkJoinTask<?> t) {
979		if (t != null) {
980	<	currentSteal = t;
1013	<	t.doExec();
1014	<	runLocalTasks();
1015	<	++nsteals;
980	>	(currentSteal = t).doExec();
981		currentSteal = null;
982	+	if (++nsteals < 0) {     // spill on overflow
983	+	ForkJoinPool p;
984	+	if ((p = pool) != null)
985	+	p.collectStealCount(this);
986	+	}
987	+	if (top != base) {       // process remaining local tasks
988	+	if (mode == 0)
989	+	popAndExecAll();
990	+	else
991	+	pollAndExecAll();
992	+	}
993		}
1018	–	else if (runState < 0)            // terminating
1019	–	alive = false;
1020	–	return alive;
994		}
995
996		/**
997	<	* Executes a non-top-level (stolen) task
997	>	* Executes a non-top-level (stolen) task.
998		*/
999		final void runSubtask(ForkJoinTask<?> t) {
1000		if (t != null) {
1001		ForkJoinTask<?> ps = currentSteal;
1002	<	currentSteal = t;
1030	<	t.doExec();
1002	>	(currentSteal = t).doExec();
1003		currentSteal = ps;
1004		}
1005		}
1006
1007		/**
1008	<	* Computes next value for random probes.  Scans don't require
1037	<	* a very high quality generator, but also not a crummy one.
1038	<	* Marsaglia xor-shift is cheap and works well enough.  Note:
1039	<	* This is manually inlined in several usages in ForkJoinPool
1040	<	* to avoid writes inside busy scan loops.
1008	>	* Returns true if owned and not known to be blocked.
1009		*/
1010	<	final int nextSeed() {
1011	<	int r = seed;
1012	<	r ^= r << 13;
1013	<	r ^= r >>> 17;
1014	<	r ^= r << 5;
1015	<	return seed = r;
1010	>	final boolean isApparentlyUnblocked() {
1011	>	Thread wt; Thread.State s;
1012	>	return (eventCount >= 0 &&
1013	>	(wt = owner) != null &&
1014	>	(s = wt.getState()) != Thread.State.BLOCKED &&
1015	>	s != Thread.State.WAITING &&
1016	>	s != Thread.State.TIMED_WAITING);
1017	>	}
1018	>
1019	>	/**
1020	>	* If this owned and is not already interrupted, try to
1021	>	* interrupt and/or unpark, ignoring exceptions.
1022	>	*/
1023	>	final void interruptOwner() {
1024	>	Thread wt, p;
1025	>	if ((wt = owner) != null && !wt.isInterrupted()) {
1026	>	try {
1027	>	wt.interrupt();
1028	>	} catch (SecurityException ignore) {
1029	>	}
1030	>	}
1031	>	if ((p = parker) != null)
1032	>	U.unpark(p);
1033		}
1034
1035		// Unsafe mechanics
1036		private static final sun.misc.Unsafe U;
1037	<	private static final long RUNSTATE;
1037	>	private static final long QLOCK;
1038		private static final int ABASE;
1039		private static final int ASHIFT;
1040		static {
#	Line 1058 | Line 1043 | public class ForkJoinPool extends Abstra
1043		U = getUnsafe();
1044		Class<?> k = WorkQueue.class;
1045		Class<?> ak = ForkJoinTask[].class;
1046	<	RUNSTATE = U.objectFieldOffset
1047	<	(k.getDeclaredField("runState"));
1046	>	QLOCK = U.objectFieldOffset
1047	>	(k.getDeclaredField("qlock"));
1048		ABASE = U.arrayBaseOffset(ak);
1049		s = U.arrayIndexScale(ak);
1050		} catch (Exception e) {
#	Line 1072 | Line 1057 | public class ForkJoinPool extends Abstra
1057		}
1058
1059		/**
1060	<	* Class for artificial tasks that are used to replace the target
1061	<	* of local joins if they are removed from an interior queue slot
1062	<	* in WorkQueue.tryRemoveAndExec. We don't need the proxy to
1063	<	* actually do anything beyond having a unique identity.
1060	>	* Per-thread records for threads that submit to pools. Currently
1061	>	* holds only pseudo-random seed / index that is used to choose
1062	>	* submission queues in method externalPush. In the future, this may
1063	>	* also incorporate a means to implement different task rejection
1064	>	* and resubmission policies.
1065	>	*
1066	>	* Seeds for submitters and workers/workQueues work in basically
1067	>	* the same way but are initialized and updated using slightly
1068	>	* different mechanics. Both are initialized using the same
1069	>	* approach as in class ThreadLocal, where successive values are
1070	>	* unlikely to collide with previous values. Seeds are then
1071	>	* randomly modified upon collisions using xorshifts, which
1072	>	* requires a non-zero seed.
1073	>	*/
1074	>	static final class Submitter {
1075	>	int seed;
1076	>	Submitter(int s) { seed = s; }
1077	>	}
1078	>
1079	>	/** Property prefix for constructing common pool */
1080	>	private static final String propPrefix =
1081	>	"java.util.concurrent.ForkJoinPool.common.";
1082	>
1083	>	// static fields (initialized in static initializer below)
1084	>
1085	>	/**
1086	>	* Creates a new ForkJoinWorkerThread. This factory is used unless
1087	>	* overridden in ForkJoinPool constructors.
1088		*/
1089	<	static final class EmptyTask extends ForkJoinTask<Void> {
1090	<	EmptyTask() { status = ForkJoinTask.NORMAL; } // force done
1082	<	public Void getRawResult() { return null; }
1083	<	public void setRawResult(Void x) {}
1084	<	public boolean exec() { return true; }
1085	<	}
1089	>	public static final ForkJoinWorkerThreadFactory
1090	>	defaultForkJoinWorkerThreadFactory;
1091
1092		/**
1093	<	* Computes a hash code for the given thread. This method is
1094	<	* expected to provide higher-quality hash codes than those using
1095	<	* method hashCode().
1093	>	* Common (static) pool. Non-null for public use unless a static
1094	>	* construction exception, but internal usages null-check on use
1095	>	* to paranoically avoid potential initialization circularities
1096	>	* as well as to simplify generated code.
1097		*/
1098	<	static final int hashThread(Thread t) {
1093	<	long id = (t == null) ? 0L : t.getId(); // Use MurmurHash of thread id
1094	<	int h = (int)id ^ (int)(id >>> 32);
1095	<	h ^= h >>> 16;
1096	<	h *= 0x85ebca6b;
1097	<	h ^= h >>> 13;
1098	<	h *= 0xc2b2ae35;
1099	<	return h ^ (h >>> 16);
1100	<	}
1098	>	static final ForkJoinPool commonPool;
1099
1100		/**
1101	<	* Top-level runloop for workers
1101	>	* Permission required for callers of methods that may start or
1102	>	* kill threads.
1103	>	*/
1104	>	private static final RuntimePermission modifyThreadPermission;
1105	>
1106	>	/**
1107	>	* 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.
1110	>	* Lazily initialized on first submission (but null-checked
1111	>	* in other contexts to avoid unnecessary initialization).
1112		*/
1113	<	final void runWorker(ForkJoinWorkerThread wt) {
1106	<	WorkQueue w = wt.workQueue;
1107	<	w.growArray(false);     // Initialize queue array and seed in this thread
1108	<	w.seed = hashThread(Thread.currentThread()) | (1 << 31); // force < 0
1113	>	static final ThreadLocal<Submitter> submitters;
1114
1115	<	do {} while (w.runTask(scan(w)));
1115	>	/**
1116	>	* Common pool parallelism. Must equal commonPool.parallelism.
1117	>	*/
1118	>	static final int commonPoolParallelism;
1119	>
1120	>	/**
1121	>	* Sequence number for creating workerNamePrefix.
1122	>	*/
1123	>	private static int poolNumberSequence;
1124	>
1125	>	/**
1126	>	* Return the next sequence number. We don't expect this to
1127	>	* ever contend so use simple builtin sync.
1128	>	*/
1129	>	private static final synchronized int nextPoolId() {
1130	>	return ++poolNumberSequence;
1131		}
1132
1133	<	// Creating, registering and deregistering workers
1133	>	// static constants
1134
1135		/**
1136	<	* Tries to create and start a worker
1136	>	* Initial timeout value (in nanoseconds) for the thread
1137	>	* triggering quiescence to park waiting for new work. On timeout,
1138	>	* the thread will instead try to shrink the number of
1139	>	* workers. The value should be large enough to avoid overly
1140	>	* aggressive shrinkage during most transient stalls (long GCs
1141	>	* etc).
1142		*/
1143	<	private void addWorker() {
1144	<	Throwable ex = null;
1145	<	ForkJoinWorkerThread w = null;
1146	<	try {
1147	<	if ((w = factory.newThread(this)) != null) {
1148	<	w.start();
1149	<	return;
1143	>	private static final long IDLE_TIMEOUT      = 2000L * 1000L * 1000L; // 2sec
1144	>
1145	>	/**
1146	>	* Timeout value when there are more threads than parallelism level
1147	>	*/
1148	>	private static final long FAST_IDLE_TIMEOUT =  200L * 1000L * 1000L;
1149	>
1150	>	/**
1151	>	* The maximum stolen->joining link depth allowed in method
1152	>	* tryHelpStealer.  Must be a power of two.  Depths for legitimate
1153	>	* chains are unbounded, but we use a fixed constant to avoid
1154	>	* (otherwise unchecked) cycles and to bound staleness of
1155	>	* traversal parameters at the expense of sometimes blocking when
1156	>	* we could be helping.
1157	>	*/
1158	>	private static final int MAX_HELP = 64;
1159	>
1160	>	/**
1161	>	* Increment for seed generators. See class ThreadLocal for
1162	>	* explanation.
1163	>	*/
1164	>	private static final int SEED_INCREMENT = 0x61c88647;
1165	>
1166	>	/**
1167	>	* Bits and masks for control variables
1168	>	*
1169	>	* Field ctl is a long packed with:
1170	>	* AC: Number of active running workers minus target parallelism (16 bits)
1171	>	* TC: Number of total workers minus target parallelism (16 bits)
1172	>	* ST: true if pool is terminating (1 bit)
1173	>	* EC: the wait count of top waiting thread (15 bits)
1174	>	* ID: poolIndex of top of Treiber stack of waiters (16 bits)
1175	>	*
1176	>	* When convenient, we can extract the upper 32 bits of counts and
1177	>	* the lower 32 bits of queue state, u = (int)(ctl >>> 32) and e =
1178	>	* (int)ctl.  The ec field is never accessed alone, but always
1179	>	* together with id and st. The offsets of counts by the target
1180	>	* parallelism and the positionings of fields makes it possible to
1181	>	* perform the most common checks via sign tests of fields: When
1182	>	* ac is negative, there are not enough active workers, when tc is
1183	>	* negative, there are not enough total workers, and when e is
1184	>	* negative, the pool is terminating.  To deal with these possibly
1185	>	* negative fields, we use casts in and out of "short" and/or
1186	>	* signed shifts to maintain signedness.
1187	>	*
1188	>	* When a thread is queued (inactivated), its eventCount field is
1189	>	* set negative, which is the only way to tell if a worker is
1190	>	* prevented from executing tasks, even though it must continue to
1191	>	* scan for them to avoid queuing races. Note however that
1192	>	* eventCount updates lag releases so usage requires care.
1193	>	*
1194	>	* Field plock is an int packed with:
1195	>	* SHUTDOWN: true if shutdown is enabled (1 bit)
1196	>	* SEQ:  a sequence lock, with PL_LOCK bit set if locked (30 bits)
1197	>	* SIGNAL: set when threads may be waiting on the lock (1 bit)
1198	>	*
1199	>	* The sequence number enables simple consistency checks:
1200	>	* Staleness of read-only operations on the workQueues array can
1201	>	* be checked by comparing plock before vs after the reads.
1202	>	*/
1203	>
1204	>	// bit positions/shifts for fields
1205	>	private static final int  AC_SHIFT   = 48;
1206	>	private static final int  TC_SHIFT   = 32;
1207	>	private static final int  ST_SHIFT   = 31;
1208	>	private static final int  EC_SHIFT   = 16;
1209	>
1210	>	// bounds
1211	>	private static final int  SMASK      = 0xffff;  // short bits
1212	>	private static final int  MAX_CAP    = 0x7fff;  // max #workers - 1
1213	>	private static final int  EVENMASK   = 0xfffe;  // even short bits
1214	>	private static final int  SQMASK     = 0x007e;  // max 64 (even) slots
1215	>	private static final int  SHORT_SIGN = 1 << 15;
1216	>	private static final int  INT_SIGN   = 1 << 31;
1217	>
1218	>	// masks
1219	>	private static final long STOP_BIT   = 0x0001L << ST_SHIFT;
1220	>	private static final long AC_MASK    = ((long)SMASK) << AC_SHIFT;
1221	>	private static final long TC_MASK    = ((long)SMASK) << TC_SHIFT;
1222	>
1223	>	// units for incrementing and decrementing
1224	>	private static final long TC_UNIT    = 1L << TC_SHIFT;
1225	>	private static final long AC_UNIT    = 1L << AC_SHIFT;
1226	>
1227	>	// masks and units for dealing with u = (int)(ctl >>> 32)
1228	>	private static final int  UAC_SHIFT  = AC_SHIFT - 32;
1229	>	private static final int  UTC_SHIFT  = TC_SHIFT - 32;
1230	>	private static final int  UAC_MASK   = SMASK << UAC_SHIFT;
1231	>	private static final int  UTC_MASK   = SMASK << UTC_SHIFT;
1232	>	private static final int  UAC_UNIT   = 1 << UAC_SHIFT;
1233	>	private static final int  UTC_UNIT   = 1 << UTC_SHIFT;
1234	>
1235	>	// masks and units for dealing with e = (int)ctl
1236	>	private static final int E_MASK      = 0x7fffffff; // no STOP_BIT
1237	>	private static final int E_SEQ       = 1 << EC_SHIFT;
1238	>
1239	>	// plock bits
1240	>	private static final int SHUTDOWN    = 1 << 31;
1241	>	private static final int PL_LOCK     = 2;
1242	>	private static final int PL_SIGNAL   = 1;
1243	>	private static final int PL_SPINS    = 1 << 8;
1244	>
1245	>	// access mode for WorkQueue
1246	>	static final int LIFO_QUEUE          =  0;
1247	>	static final int FIFO_QUEUE          =  1;
1248	>	static final int SHARED_QUEUE        = -1;
1249	>
1250	>	// Instance fields
1251	>
1252	>	/*
1253	>	* Field layout order in this class tends to matter more than one
1254	>	* would like. Runtime layout order is only loosely related to
1255	>	* declaration order and may differ across JVMs, but the following
1256	>	* empirically works OK on current JVMs.
1257	>	*/
1258	>	volatile long stealCount;                  // collects worker counts
1259	>	volatile long ctl;                         // main pool control
1260	>	final int parallelism;                     // parallelism level
1261	>	final int localMode;                       // per-worker scheduling mode
1262	>	volatile int indexSeed;                    // worker/submitter index seed
1263	>	volatile int plock;                        // shutdown status and seqLock
1264	>	WorkQueue[] workQueues;                    // main registry
1265	>	final ForkJoinWorkerThreadFactory factory; // factory for new workers
1266	>	final Thread.UncaughtExceptionHandler ueh; // per-worker UEH
1267	>	final String workerNamePrefix;             // to create worker name string
1268	>
1269	>	/*
1270	>	* Acquires the plock lock to protect worker array and related
1271	>	* updates. This method is called only if an initial CAS on plock
1272	>	* fails. This acts as a spinLock for normal cases, but falls back
1273	>	* to builtin monitor to block when (rarely) needed. This would be
1274	>	* a terrible idea for a highly contended lock, but works fine as
1275	>	* a more conservative alternative to a pure spinlock.  See
1276	>	* internal ConcurrentHashMap documentation for further
1277	>	* explanation of nearly the same construction.
1278	>	*/
1279	>	private int acquirePlock() {
1280	>	int spins = PL_SPINS, r = 0, ps, nps;
1281	>	for (;;) {
1282	>	if (((ps = plock) & PL_LOCK) == 0 &&
1283	>	U.compareAndSwapInt(this, PLOCK, ps, nps = ps + PL_LOCK))
1284	>	return nps;
1285	>	else if (r == 0)
1286	>	r = ThreadLocalRandom.current().nextInt(); // randomize spins
1287	>	else if (spins >= 0) {
1288	>	r ^= r << 1; r ^= r >>> 3; r ^= r << 10; // xorshift
1289	>	if (r >= 0)
1290	>	--spins;
1291	>	}
1292	>	else if (U.compareAndSwapInt(this, PLOCK, ps, ps | PL_SIGNAL)) {
1293	>	synchronized (this) {
1294	>	if ((plock & PL_SIGNAL) != 0) {
1295	>	try {
1296	>	wait();
1297	>	} catch (InterruptedException ie) {
1298	>	try {
1299	>	Thread.currentThread().interrupt();
1300	>	} catch (SecurityException ignore) {
1301	>	}
1302	>	}
1303	>	}
1304	>	else
1305	>	notifyAll();
1306	>	}
1307		}
1126	–	} catch (Throwable e) {
1127	–	ex = e;
1308		}
1129	–	deregisterWorker(w, ex);
1309		}
1310
1311		/**
1312	<	* Callback from ForkJoinWorkerThread constructor to assign a
1313	<	* public name. This must be separate from registerWorker because
1135	<	* it is called during the "super" constructor call in
1136	<	* ForkJoinWorkerThread.
1312	>	* Unlocks and signals any thread waiting for plock. Called only
1313	>	* when CAS of seq value for unlock fails.
1314		*/
1315	<	final String nextWorkerName() {
1316	<	return workerNamePrefix.concat
1317	<	(Integer.toString(nextWorkerNumber.addAndGet(1)));
1315	>	private void releasePlock(int ps) {
1316	>	plock = ps;
1317	>	synchronized (this) { notifyAll(); }
1318		}
1319
1320	+	//  Registering and deregistering workers
1321	+
1322		/**
1323	<	* Callback from ForkJoinWorkerThread constructor to establish and
1324	<	* record its WorkQueue
1323	>	* Callback from ForkJoinWorkerThread constructor to establish its
1324	>	* poolIndex and record its WorkQueue. To avoid scanning bias due
1325	>	* to packing entries in front of the workQueues array, we treat
1326	>	* the array as a simple power-of-two hash table using per-thread
1327	>	* seed as hash, expanding as needed.
1328		*
1329	<	* @param wt the worker thread
1329	>	* @param w the worker's queue
1330		*/
1331	<	final void registerWorker(ForkJoinWorkerThread wt) {
1332	<	WorkQueue w = wt.workQueue;
1333	<	ReentrantLock lock = this.lock;
1334	<	lock.lock();
1331	>	final void registerWorker(WorkQueue w) {
1332	>	int s, ps; // generate a rarely colliding candidate index seed
1333	>	do {} while (!U.compareAndSwapInt(this, INDEXSEED,
1334	>	s = indexSeed, s += SEED_INCREMENT) ||
1335	>	s == 0); // skip 0
1336	>	if (((ps = plock) & PL_LOCK) != 0 ||
1337	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1338	>	ps = acquirePlock();
1339	>	int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1340		try {
1341	<	int k = nextPoolIndex;
1342	<	WorkQueue[] ws = workQueues;
1343	<	if (ws != null) {                       // ignore on shutdown
1344	<	int n = ws.length;
1345	<	if (k < 0 || (k & 1) == 0 || k >= n || ws[k] != null) {
1346	<	for (k = 1; k < n && ws[k] != null; k += 2)
1347	<	;                           // workers are at odd indices
1348	<	if (k >= n)                     // resize
1349	<	workQueues = ws = Arrays.copyOf(ws, n << 1);
1350	<	}
1351	<	w.poolIndex = k;
1352	<	w.eventCount = ~(k >>> 1) & SMASK;  // Set up wait count
1353	<	ws[k] = w;                          // record worker
1354	<	nextPoolIndex = k + 2;
1355	<	int rs = runState;
1356	<	int m = rs & SMASK;                 // recalculate runState mask
1357	<	if (k > m)
1358	<	m = (m << 1) + 1;
1172	<	runState = (rs & SHUTDOWN) | ((rs + RS_SEQ) & RS_SEQ_MASK) | m;
1341	>	WorkQueue[] ws;
1342	>	if (w != null && (ws = workQueues) != null) {
1343	>	w.seed = s;
1344	>	int n = ws.length, m = n - 1;
1345	>	int r = (s << 1) | 1;               // use odd-numbered indices
1346	>	if (ws[r &= m] != null) {           // collision
1347	>	int probes = 0;                 // step by approx half size
1348	>	int step = (n <= 4) ? 2 : ((n >>> 1) & EVENMASK) + 2;
1349	>	while (ws[r = (r + step) & m] != null) {
1350	>	if (++probes >= n) {
1351	>	workQueues = ws = Arrays.copyOf(ws, n <<= 1);
1352	>	m = n - 1;
1353	>	probes = 0;
1354	>	}
1355	>	}
1356	>	}
1357	>	w.eventCount = w.poolIndex = r;     // establish before recording
1358	>	ws[r] = w;
1359		}
1360		} finally {
1361	<	lock.unlock();
1361	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1362	>	releasePlock(nps);
1363		}
1364		}
1365
1366		/**
1367	<	* Final callback from terminating worker, as well as failure to
1368	<	* construct or start a worker in addWorker.  Removes record of
1369	<	* worker from array, and adjusts counts. If pool is shutting
1370	<	* down, tries to complete termination.
1367	>	* Final callback from terminating worker, as well as upon failure
1368	>	* to construct or start a worker.  Removes record of worker from
1369	>	* array, and adjusts counts. If pool is shutting down, tries to
1370	>	* complete termination.
1371		*
1372	<	* @param wt the worker thread or null if addWorker failed
1372	>	* @param wt the worker thread or null if construction failed
1373		* @param ex the exception causing failure, or null if none
1374		*/
1375		final void deregisterWorker(ForkJoinWorkerThread wt, Throwable ex) {
1376		WorkQueue w = null;
1377		if (wt != null && (w = wt.workQueue) != null) {
1378	<	w.runState = -1;                // ensure runState is set
1379	<	stealCount.getAndAdd(w.totalSteals + w.nsteals);
1380	<	int idx = w.poolIndex;
1381	<	ReentrantLock lock = this.lock;
1382	<	lock.lock();
1383	<	try {                           // remove record from array
1378	>	int ps;
1379	>	collectStealCount(w);
1380	>	w.qlock = -1;                // ensure set
1381	>	if (((ps = plock) & PL_LOCK) != 0 ||
1382	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1383	>	ps = acquirePlock();
1384	>	int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1385	>	try {
1386	>	int idx = w.poolIndex;
1387		WorkQueue[] ws = workQueues;
1388		if (ws != null && idx >= 0 && idx < ws.length && ws[idx] == w)
1389	<	ws[nextPoolIndex = idx] = null;
1389	>	ws[idx] = null;
1390		} finally {
1391	<	lock.unlock();
1391	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1392	>	releasePlock(nps);
1393		}
1394		}
1395
#	Line 1208 | Line 1399 | public class ForkJoinPool extends Abstra
1399		((c - TC_UNIT) & TC_MASK) |
1400		(c & ~(AC_MASK|TC_MASK)))));
1401
1402	<	if (!tryTerminate(false) && w != null) {
1402	>	if (!tryTerminate(false, false) && w != null) {
1403		w.cancelAll();                  // cancel remaining tasks
1404		if (w.array != null)            // suppress signal if never ran
1405	<	signalWork();               // wake up or create replacement
1405	>	signalWork(null, 1);        // wake up or create replacement
1406	>	if (ex == null)                 // help clean refs on way out
1407	>	ForkJoinTask.helpExpungeStaleExceptions();
1408		}
1409
1410		if (ex != null)                     // rethrow
1411	<	U.throwException(ex);
1411	>	ForkJoinTask.rethrow(ex);
1412		}
1413
1221	–
1222	–	// Maintaining ctl counts
1223	–
1414		/**
1415	<	* Increments active count; mainly called upon return from blocking
1416	<	*/
1417	<	final void incrementActiveCount() {
1418	<	long c;
1419	<	do {} while (!U.compareAndSwapLong(this, CTL, c = ctl, c + AC_UNIT));
1415	>	* Collect worker steal count into total. Called on termination
1416	>	* and upon int overflow of local count. (There is a possible race
1417	>	* in the latter case vs any caller of getStealCount, which can
1418	>	* make its results less accurate than usual.)
1419	>	*/
1420	>	final void collectStealCount(WorkQueue w) {
1421	>	if (w != null) {
1422	>	long sc;
1423	>	int ns = w.nsteals;
1424	>	w.nsteals = 0; // handle overflow
1425	>	long steals = (ns >= 0) ? ns : 1L + (long)(Integer.MAX_VALUE);
1426	>	do {} while (!U.compareAndSwapLong(this, STEALCOUNT,
1427	>	sc = stealCount, sc + steals));
1428	>	}
1429		}
1430
1431	+	// Submissions
1432	+
1433		/**
1434	<	* Activates or creates a worker
1435	<	*/
1436	<	final void signalWork() {
1437	<	/*
1438	<	* The while condition is true if: (there is are too few total
1439	<	* workers OR there is at least one waiter) AND (there are too
1440	<	* few active workers OR the pool is terminating).  The value
1441	<	* of e distinguishes the remaining cases: zero (no waiters)
1442	<	* for create, negative if terminating (in which case do
1443	<	* nothing), else release a waiter. The secondary checks for
1444	<	* release (non-null array etc) can fail if the pool begins
1445	<	* terminating after the test, and don't impose any added cost
1446	<	* because JVMs must perform null and bounds checks anyway.
1447	<	*/
1448	<	long c; int e, u;
1449	<	while ((((e = (int)(c = ctl)) | (u = (int)(c >>> 32))) &
1450	<	(INT_SIGN|SHORT_SIGN)) == (INT_SIGN|SHORT_SIGN)) {
1451	<	WorkQueue[] ws = workQueues; int i; WorkQueue w; Thread p;
1452	<	if (e == 0) {                    // add a new worker
1453	<	if (U.compareAndSwapLong
1454	<	(this, CTL, c, (long)(((u + UTC_UNIT) & UTC_MASK) |
1455	<	((u + UAC_UNIT) & UAC_MASK)) << 32)) {
1255	<	addWorker();
1256	<	break;
1257	<	}
1258	<	}
1259	<	else if (e > 0 && ws != null &&
1260	<	(i = ((~e << 1) | 1) & SMASK) < ws.length &&
1261	<	(w = ws[i]) != null &&
1262	<	w.eventCount == (e | INT_SIGN)) {
1263	<	if (U.compareAndSwapLong
1264	<	(this, CTL, c, (((long)(w.nextWait & E_MASK)) |
1265	<	((long)(u + UAC_UNIT) << 32)))) {
1266	<	w.eventCount = (e + E_SEQ) & E_MASK;
1267	<	if ((p = w.parker) != null)
1268	<	U.unpark(p);         // release a waiting worker
1269	<	break;
1270	<	}
1434	>	* Unless shutting down, adds the given task to a submission queue
1435	>	* at submitter's current queue index (modulo submission
1436	>	* range). Only the most common path is directly handled in this
1437	>	* method. All others are relayed to fullExternalPush.
1438	>	*
1439	>	* @param task the task. Caller must ensure non-null.
1440	>	*/
1441	>	final void externalPush(ForkJoinTask<?> task) {
1442	>	WorkQueue[] ws; WorkQueue q; Submitter z; int m; ForkJoinTask<?>[] a;
1443	>	if ((z = submitters.get()) != null && plock > 0 &&
1444	>	(ws = workQueues) != null && (m = (ws.length - 1)) >= 0 &&
1445	>	(q = ws[m & z.seed & SQMASK]) != null &&
1446	>	U.compareAndSwapInt(q, QLOCK, 0, 1)) { // lock
1447	>	int s = q.top, n;
1448	>	if ((a = q.array) != null && a.length > (n = s + 1 - q.base)) {
1449	>	U.putObject(a, (long)(((a.length - 1) & s) << ASHIFT) + ABASE,
1450	>	task);
1451	>	q.top = s + 1;                     // push on to deque
1452	>	q.qlock = 0;
1453	>	if (n <= 1)
1454	>	signalWork(q, 1);
1455	>	return;
1456		}
1457	<	else
1273	<	break;
1457	>	q.qlock = 0;
1458		}
1459	+	fullExternalPush(task);
1460		}
1461
1462		/**
1463	<	* Tries to decrement active count (sometimes implicitly) and
1464	<	* possibly release or create a compensating worker in preparation
1465	<	* for blocking. Fails on contention or termination.
1466	<	*
1467	<	* @return true if the caller can block, else should recheck and retry
1468	<	*/
1469	<	final boolean tryCompensate() {
1470	<	WorkQueue[] ws; WorkQueue w; Thread p;
1471	<	int pc = parallelism, e, u, ac, tc, i;
1472	<	long c = ctl;
1473	<
1474	<	if ((e = (int)c) >= 0) {
1475	<	if ((ac = ((u = (int)(c >>> 32)) >> UAC_SHIFT)) <= 0 &&
1476	<	e != 0 && (ws = workQueues) != null &&
1477	<	(i = ((~e << 1) | 1) & SMASK) < ws.length &&
1478	<	(w = ws[i]) != null) {
1479	<	if (w.eventCount == (e | INT_SIGN) &&
1480	<	U.compareAndSwapLong
1481	<	(this, CTL, c, ((long)(w.nextWait & E_MASK) |
1482	<	(c & (AC_MASK|TC_MASK))))) {
1483	<	w.eventCount = (e + E_SEQ) & E_MASK;
1484	<	if ((p = w.parker) != null)
1485	<	U.unpark(p);
1486	<	return true;             // release an idle worker
1487	<	}
1488	<	}
1489	<	else if ((tc = (short)(u >>> UTC_SHIFT)) >= 0 && ac + pc > 1) {
1490	<	long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK);
1491	<	if (U.compareAndSwapLong(this, CTL, c, nc))
1492	<	return true;             // no compensation needed
1463	>	* Full version of externalPush. This method is called, among
1464	>	* other times, upon the first submission of the first task to the
1465	>	* pool, so must perform secondary initialization: creating
1466	>	* workQueue array and setting plock to a valid value. It also
1467	>	* detects first submission by an external thread by looking up
1468	>	* its ThreadLocal, and creates a new shared queue if the one at
1469	>	* index if empty or contended. The lock bodies must be
1470	>	* exception-free (so no try/finally) so we optimistically
1471	>	* allocate new queues/arrays outside the locks and throw them
1472	>	* away if (very rarely) not needed. Note that the plock seq value
1473	>	* can eventually wrap around zero, but if so harmlessly fails to
1474	>	* reinitialize.
1475	>	*/
1476	>	private void fullExternalPush(ForkJoinTask<?> task) {
1477	>	for (Submitter z = null;;) {
1478	>	WorkQueue[] ws; WorkQueue q; int ps, m, r, s;
1479	>	if ((ps = plock) < 0)
1480	>	throw new RejectedExecutionException();
1481	>	else if ((ws = workQueues) == null || (m = ws.length - 1) < 0) {
1482	>	int n = parallelism - 1; n |= n >>> 1; n |= n >>> 2;
1483	>	n |= n >>> 4; n |= n >>> 8; n |= n >>> 16;
1484	>	WorkQueue[] nws = new WorkQueue[(n + 1) << 1]; // power of two
1485	>	if ((ps & PL_LOCK) != 0 ||
1486	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1487	>	ps = acquirePlock();
1488	>	if ((ws = workQueues) == null)
1489	>	workQueues = nws;
1490	>	int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1491	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1492	>	releasePlock(nps);
1493	>	}
1494	>	else if (z == null && (z = submitters.get()) == null) {
1495	>	if (U.compareAndSwapInt(this, INDEXSEED,
1496	>	s = indexSeed, s += SEED_INCREMENT) &&
1497	>	s != 0) // skip 0
1498	>	submitters.set(z = new Submitter(s));
1499		}
1500	<	else if (tc + pc < MAX_ID) {
1501	<	long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK);
1502	<	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1503	<	addWorker();
1504	<	return true;             // create replacement
1505	<	}
1500	>	else {
1501	>	int k = (r = z.seed) & m & SQMASK;
1502	>	if ((q = ws[k]) == null && (ps & PL_LOCK) == 0) {
1503	>	(q = new WorkQueue(this, null, SHARED_QUEUE)).poolIndex = k;
1504	>	if (((ps = plock) & PL_LOCK) != 0 ||
1505	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1506	>	ps = acquirePlock();
1507	>	WorkQueue w = null;
1508	>	if ((ws = workQueues) != null && k < ws.length &&
1509	>	(w = ws[k]) == null)
1510	>	ws[k] = q;
1511	>	else
1512	>	q = w;
1513	>	int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1514	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1515	>	releasePlock(nps);
1516	>	}
1517	>	if (q != null && q.qlock == 0 && q.fullPush(task, false))
1518	>	return;
1519	>	r ^= r << 13;                // same xorshift as WorkQueues
1520	>	r ^= r >>> 17;
1521	>	z.seed = r ^= r << 5;        // move to a different index
1522		}
1523		}
1317	–	return false;
1524		}
1525
1526	<	// Submissions
1526	>	// Maintaining ctl counts
1527
1528		/**
1529	<	* Unless shutting down, adds the given task to some submission
1324	<	* queue; using a randomly chosen queue index if the caller is a
1325	<	* ForkJoinWorkerThread, else one based on caller thread's hash
1326	<	* code. If no queue exists at the index, one is created.  If the
1327	<	* queue is busy, another is chosen by sweeping through the queues
1328	<	* array.
1529	>	* Increments active count; mainly called upon return from blocking.
1530		*/
1531	<	private void doSubmit(ForkJoinTask<?> task) {
1532	<	if (task == null)
1533	<	throw new NullPointerException();
1534	<	Thread t = Thread.currentThread();
1535	<	int r = ((t instanceof ForkJoinWorkerThread) ?
1536	<	((ForkJoinWorkerThread)t).workQueue.nextSeed() : hashThread(t));
1537	<	for (;;) {
1538	<	int rs = runState, m = rs & SMASK;
1539	<	int j = r &= (m & ~1);                      // even numbered queues
1540	<	WorkQueue[] ws = workQueues;
1541	<	if (rs < 0 || ws == null)
1542	<	throw new RejectedExecutionException(); // shutting down
1543	<	if (ws.length > m) {                        // consistency check
1544	<	for (WorkQueue q;;) {                   // circular sweep
1545	<	if (((q = ws[j]) != null ||
1546	<	(q = tryAddSharedQueue(j)) != null) &&
1547	<	q.trySharedPush(task)) {
1548	<	signalWork();
1549	<	return;
1531	>	final void incrementActiveCount() {
1532	>	long c;
1533	>	do {} while (!U.compareAndSwapLong(this, CTL, c = ctl, c + AC_UNIT));
1534	>	}
1535	>
1536	>	/**
1537	>	* Tries to create (at most one) or activate (possibly several)
1538	>	* workers if too few are active. On contention failure, continues
1539	>	* until at least one worker is signalled or the given queue is
1540	>	* empty or all workers are active.
1541	>	*
1542	>	* @param q if non-null, the queue holding tasks to be signalled
1543	>	* @param signals the target number of signals.
1544	>	*/
1545	>	final void signalWork(WorkQueue q, int signals) {
1546	>	long c; int e, u, i; WorkQueue[] ws; WorkQueue w; Thread p;
1547	>	while ((u = (int)((c = ctl) >>> 32)) < 0) {
1548	>	if ((e = (int)c) > 0) {
1549	>	if ((ws = workQueues) != null && ws.length > (i = e & SMASK) &&
1550	>	(w = ws[i]) != null && w.eventCount == (e | INT_SIGN)) {
1551	>	long nc = (((long)(w.nextWait & E_MASK)) |
1552	>	((long)(u + UAC_UNIT) << 32));
1553	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1554	>	w.eventCount = (e + E_SEQ) & E_MASK;
1555	>	if ((p = w.parker) != null)
1556	>	U.unpark(p);
1557	>	if (--signals <= 0)
1558	>	break;
1559		}
1560	<	if ((j = (j + 2) & m) == r) {
1561	<	Thread.yield();                 // all queues busy
1560	>	else
1561	>	signals = 1;
1562	>	if ((q != null && q.queueSize() == 0))
1563		break;
1353	–	}
1564		}
1565	+	else
1566	+	break;
1567		}
1568	<	}
1569	<	}
1570	<
1571	<	/**
1572	<	* Tries to add and register a new queue at the given index.
1573	<	*
1574	<	* @param idx the workQueues array index to register the queue
1575	<	* @return the queue, or null if could not add because could
1576	<	* not acquire lock or idx is unusable
1577	<	*/
1578	<	private WorkQueue tryAddSharedQueue(int idx) {
1579	<	WorkQueue q = null;
1580	<	ReentrantLock lock = this.lock;
1369	<	if (idx >= 0 && (idx & 1) == 0 && !lock.isLocked()) {
1370	<	// create queue outside of lock but only if apparently free
1371	<	WorkQueue nq = new WorkQueue(null, SHARED_QUEUE);
1372	<	if (lock.tryLock()) {
1373	<	try {
1374	<	WorkQueue[] ws = workQueues;
1375	<	if (ws != null && idx < ws.length) {
1376	<	if ((q = ws[idx]) == null) {
1377	<	int rs;         // update runState seq
1378	<	ws[idx] = q = nq;
1379	<	runState = (((rs = runState) & SHUTDOWN) |
1380	<	((rs + RS_SEQ) & ~SHUTDOWN));
1568	>	else if (e == 0 && (u & SHORT_SIGN) != 0) {
1569	>	long nc = (long)(((u + UTC_UNIT) & UTC_MASK) |
1570	>	((u + UAC_UNIT) & UAC_MASK)) << 32;
1571	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1572	>	ForkJoinWorkerThread wt = null;
1573	>	Throwable ex = null;
1574	>	boolean started = false;
1575	>	try {
1576	>	ForkJoinWorkerThreadFactory fac;
1577	>	if ((fac = factory) != null &&
1578	>	(wt = fac.newThread(this)) != null) {
1579	>	wt.start();
1580	>	started = true;
1581		}
1582	+	} catch (Throwable rex) {
1583	+	ex = rex;
1584		}
1585	<	} finally {
1586	<	lock.unlock();
1585	>	if (!started)
1586	>	deregisterWorker(wt, ex); // adjust counts on failure
1587	>	break;
1588		}
1589		}
1590	+	else
1591	+	break;
1592		}
1388	–	return q;
1593		}
1594
1595		// Scanning for tasks
1596
1597		/**
1598	+	* Top-level runloop for workers, called by ForkJoinWorkerThread.run.
1599	+	*/
1600	+	final void runWorker(WorkQueue w) {
1601	+	// initialize queue array in this thread
1602	+	w.array = new ForkJoinTask<?>[WorkQueue.INITIAL_QUEUE_CAPACITY];
1603	+	do { w.runTask(scan(w)); } while (w.qlock >= 0);
1604	+	}
1605	+
1606	+	/**
1607		* Scans for and, if found, returns one task, else possibly
1608		* inactivates the worker. This method operates on single reads of
1609	<	* volatile state and is designed to be re-invoked continuously in
1610	<	* part because it returns upon detecting inconsistencies,
1609	>	* volatile state and is designed to be re-invoked continuously,
1610	>	* in part because it returns upon detecting inconsistencies,
1611		* contention, or state changes that indicate possible success on
1612		* re-invocation.
1613		*
1614	<	* The scan searches for tasks across queues, randomly selecting
1615	<	* the first #queues probes, favoring steals 2:1 over submissions
1616	<	* (by exploiting even/odd indexing), and then performing a
1617	<	* circular sweep of all queues.  The scan terminates upon either
1618	<	* finding a non-empty queue, or completing a full sweep. If the
1619	<	* worker is not inactivated, it takes and returns a task from
1620	<	* this queue.  On failure to find a task, we take one of the
1621	<	* following actions, after which the caller will retry calling
1622	<	* this method unless terminated.
1623	<	*
1624	<	* * If not a complete sweep, try to release a waiting worker.  If
1412	<	* the scan terminated because the worker is inactivated, then the
1413	<	* released worker will often be the calling worker, and it can
1414	<	* succeed obtaining a task on the next call. Or maybe it is
1415	<	* another worker, but with same net effect. Releasing in other
1416	<	* cases as well ensures that we have enough workers running.
1417	<	*
1418	<	* * If the caller has run a task since the the last empty scan,
1419	<	* return (to allow rescan) if other workers are not also yet
1420	<	* enqueued.  Field WorkQueue.rescans counts down on each scan to
1421	<	* ensure eventual inactivation, and occasional calls to
1422	<	* Thread.yield to help avoid interference with more useful
1423	<	* activities on the system.
1614	>	* The scan searches for tasks across a random permutation of
1615	>	* queues (starting at a random index and stepping by a random
1616	>	* relative prime, checking each at least once).  The scan
1617	>	* terminates upon either finding a non-empty queue, or completing
1618	>	* the sweep. If the worker is not inactivated, it takes and
1619	>	* returns a task from this queue. Otherwise, if not activated, it
1620	>	* signals workers (that may include itself) and returns so caller
1621	>	* can retry. Also returns for trtry if the worker array may have
1622	>	* changed during an empty scan.  On failure to find a task, we
1623	>	* take one of the following actions, after which the caller will
1624	>	* retry calling this method unless terminated.
1625		*
1626	<	* * If pool is terminating, terminate the worker
1626	>	* * If pool is terminating, terminate the worker.
1627		*
1628		* * If not already enqueued, try to inactivate and enqueue the
1629	<	* worker on wait queue.
1630	<	*
1631	<	* * If already enqueued and none of the above apply, either park
1632	<	* awaiting signal, or if this is the most recent waiter and pool
1633	<	* is quiescent, relay to idleAwaitWork to check for termination
1634	<	* and possibly shrink pool.
1629	>	* worker on wait queue. Or, if inactivating has caused the pool
1630	>	* to be quiescent, relay to idleAwaitWork to check for
1631	>	* termination and possibly shrink pool.
1632	>	*
1633	>	* * If already enqueued and none of the above apply, possibly
1634	>	* (with 1/2 probablility) park awaiting signal, else lingering to
1635	>	* help scan and signal.
1636		*
1637		* @param w the worker (via its WorkQueue)
1638	<	* @return a task or null of none found
1638	>	* @return a task or null if none found
1639		*/
1640		private final ForkJoinTask<?> scan(WorkQueue w) {
1641	<	boolean swept = false;                 // true after full empty scan
1642	<	WorkQueue[] ws;                        // volatile read order matters
1643	<	int r = w.seed, ec = w.eventCount;     // ec is negative if inactive
1644	<	int rs = runState, m = rs & SMASK;
1645	<	if ((ws = workQueues) != null && ws.length > m) {
1646	<	ForkJoinTask<?> task = null;
1647	<	for (int k = 0, j = -2 - m; ; ++j) {
1648	<	WorkQueue q; int b;
1649	<	if (j < 0) {                    // random probes while j negative
1650	<	r ^= r << 13; r ^= r >>> 17; k = (r ^= r << 5) | (j & 1);
1651	<	}                               // worker (not submit) for odd j
1652	<	else                            // cyclic scan when j >= 0
1653	<	k += (m >>> 1) | 1;         // step by half to reduce bias
1654	<
1655	<	if ((q = ws[k & m]) != null && (b = q.base) - q.top < 0) {
1656	<	if (ec >= 0)
1657	<	task = q.pollAt(b);     // steal
1658	<	break;
1641	>	WorkQueue[] ws; WorkQueue q;           // first update random seed
1642	>	int r = w.seed; r ^= r << 13; r ^= r >>> 17; w.seed = r ^= r << 5;
1643	>	int ps = plock, m;                     // volatile read order matters
1644	>	if ((ws = workQueues) != null && (m = ws.length - 1) > 0) {
1645	>	int ec = w.eventCount;             // ec is negative if inactive
1646	>	int step = (r >>> 16) | 1;         // relatively prime
1647	>	for (int j = (m + 1) << 2;  ; --j, r += step) {
1648	>	ForkJoinTask<?> t; ForkJoinTask<?>[] a; int b, n;
1649	>	if ((q = ws[r & m]) != null && (b = q.base) - q.top < 0 &&
1650	>	(a = q.array) != null) {   // probably nonempty
1651	>	int i = (((a.length - 1) & b) << ASHIFT) + ABASE;
1652	>	t = (ForkJoinTask<?>)U.getObjectVolatile(a, i);
1653	>	if (q.base == b && ec >= 0 && t != null &&
1654	>	U.compareAndSwapObject(a, i, t, null)) {
1655	>	if ((n = q.top - (q.base = b + 1)) > 0)
1656	>	signalWork(q, n);
1657	>	return t;              // taken
1658	>	}
1659	>	if (j < m || (ec < 0 && (ec = w.eventCount) < 0)) {
1660	>	if ((n = q.queueSize() - 1) > 0)
1661	>	signalWork(q, n);
1662	>	break;                 // let caller retry after signal
1663	>	}
1664		}
1665	<	else if (j > m) {
1666	<	if (rs == runState)        // staleness check
1667	<	swept = true;
1665	>	else if (j < 0) {              // end of scan
1666	>	long c = ctl; int e;
1667	>	if (plock != ps)           // incomplete sweep
1668	>	break;
1669	>	if ((e = (int)c) < 0)      // pool is terminating
1670	>	w.qlock = -1;
1671	>	else if (ec >= 0) {        // try to enqueue/inactivate
1672	>	long nc = ((long)ec |
1673	>	((c - AC_UNIT) & (AC_MASK|TC_MASK)));
1674	>	w.nextWait = e;
1675	>	w.eventCount = ec | INT_SIGN; // mark as inactive
1676	>	if (ctl != c ||
1677	>	!U.compareAndSwapLong(this, CTL, c, nc))
1678	>	w.eventCount = ec; // unmark on CAS failure
1679	>	else if ((int)(c >> AC_SHIFT) == 1 - parallelism)
1680	>	idleAwaitWork(w, nc, c);  // quiescent
1681	>	}
1682	>	else if (w.seed >= 0 && w.eventCount < 0) {
1683	>	Thread wt = Thread.currentThread();
1684	>	Thread.interrupted();  // clear status
1685	>	U.putObject(wt, PARKBLOCKER, this);
1686	>	w.parker = wt;         // emulate LockSupport.park
1687	>	if (w.eventCount < 0)  // recheck
1688	>	U.park(false, 0L);
1689	>	w.parker = null;
1690	>	U.putObject(wt, PARKBLOCKER, null);
1691	>	}
1692		break;
1693		}
1694		}
1464	–	w.seed = r;                        // save seed for next scan
1465	–	if (task != null)
1466	–	return task;
1467	–	}
1468	–
1469	–	// Decode ctl on empty scan
1470	–	long c = ctl; int e = (int)c, a = (int)(c >> AC_SHIFT), nr, ns;
1471	–	if (!swept) {                          // try to release a waiter
1472	–	WorkQueue v; Thread p;
1473	–	if (e > 0 && a < 0 && ws != null &&
1474	–	(v = ws[((~e << 1) | 1) & m]) != null &&
1475	–	v.eventCount == (e | INT_SIGN) && U.compareAndSwapLong
1476	–	(this, CTL, c, ((long)(v.nextWait & E_MASK) |
1477	–	((c + AC_UNIT) & (AC_MASK|TC_MASK))))) {
1478	–	v.eventCount = (e + E_SEQ) & E_MASK;
1479	–	if ((p = v.parker) != null)
1480	–	U.unpark(p);
1481	–	}
1482	–	}
1483	–	else if ((nr = w.rescans) > 0) {       // continue rescanning
1484	–	int ac = a + parallelism;
1485	–	if ((w.rescans = (ac < nr) ? ac : nr - 1) > 0 && w.seed < 0 &&
1486	–	w.eventCount == ec)
1487	–	Thread.yield();                // 1 bit randomness for yield call
1488	–	}
1489	–	else if (e < 0)                        // pool is terminating
1490	–	w.runState = -1;
1491	–	else if (ec >= 0) {                    // try to enqueue
1492	–	long nc = (long)ec | ((c - AC_UNIT) & (AC_MASK|TC_MASK));
1493	–	w.nextWait = e;
1494	–	w.eventCount = ec | INT_SIGN;      // mark as inactive
1495	–	if (!U.compareAndSwapLong(this, CTL, c, nc))
1496	–	w.eventCount = ec;             // back out on CAS failure
1497	–	else if ((ns = w.nsteals) != 0) {  // set rescans if ran task
1498	–	if (a <= 0)                    // ... unless too many active
1499	–	w.rescans = a + parallelism;
1500	–	w.nsteals = 0;
1501	–	w.totalSteals += ns;
1502	–	}
1503	–	}
1504	–	else{                                  // already queued
1505	–	if (parallelism == -a)
1506	–	idleAwaitWork(w);              // quiescent
1507	–	if (w.eventCount == ec) {
1508	–	Thread.interrupted();          // clear status
1509	–	ForkJoinWorkerThread wt = w.owner;
1510	–	U.putObject(wt, PARKBLOCKER, this);
1511	–	w.parker = wt;                 // emulate LockSupport.park
1512	–	if (w.eventCount == ec)        // recheck
1513	–	U.park(false, 0L);         // block
1514	–	w.parker = null;
1515	–	U.putObject(wt, PARKBLOCKER, null);
1516	–	}
1695		}
1696		return null;
1697		}
1698
1699		/**
1700	<	* If inactivating worker w has caused pool to become quiescent,
1701	<	* check for pool termination, and, so long as this is not the
1702	<	* only worker, wait for event for up to SHRINK_RATE nanosecs On
1703	<	* timeout, if ctl has not changed, terminate the worker, which
1704	<	* will in turn wake up another worker to possibly repeat this
1705	<	* process.
1700	>	* If inactivating worker w has caused the pool to become
1701	>	* quiescent, checks for pool termination, and, so long as this is
1702	>	* not the only worker, waits for event for up to a given
1703	>	* duration.  On timeout, if ctl has not changed, terminates the
1704	>	* worker, which will in turn wake up another worker to possibly
1705	>	* repeat this process.
1706		*
1707		* @param w the calling worker
1708	+	* @param currentCtl the ctl value triggering possible quiescence
1709	+	* @param prevCtl the ctl value to restore if thread is terminated
1710		*/
1711	<	private void idleAwaitWork(WorkQueue w) {
1712	<	long c; int nw, ec;
1713	<	if (!tryTerminate(false) &&
1714	<	(int)((c = ctl) >> AC_SHIFT) + parallelism == 0 &&
1715	<	(ec = w.eventCount) == ((int)c | INT_SIGN) &&
1716	<	(nw = w.nextWait) != 0) {
1717	<	long nc = ((long)(nw & E_MASK) | // ctl to restore on timeout
1718	<	((c + AC_UNIT) & AC_MASK) | (c & TC_MASK));
1719	<	ForkJoinTask.helpExpungeStaleExceptions(); // help clean
1540	<	ForkJoinWorkerThread wt = w.owner;
1541	<	while (ctl == c) {
1542	<	long startTime = System.nanoTime();
1711	>	private void idleAwaitWork(WorkQueue w, long currentCtl, long prevCtl) {
1712	>	if (w.eventCount < 0 &&
1713	>	(this == commonPool || !tryTerminate(false, false)) &&
1714	>	(int)prevCtl != 0) {
1715	>	int dc = -(short)(currentCtl >>> TC_SHIFT);
1716	>	long parkTime = dc < 0 ? FAST_IDLE_TIMEOUT: (dc + 1) * IDLE_TIMEOUT;
1717	>	long deadline = System.nanoTime() + parkTime - 100000L; // 1ms slop
1718	>	Thread wt = Thread.currentThread();
1719	>	while (ctl == currentCtl) {
1720		Thread.interrupted();  // timed variant of version in scan()
1721		U.putObject(wt, PARKBLOCKER, this);
1722		w.parker = wt;
1723	<	if (ctl == c)
1724	<	U.park(false, SHRINK_RATE);
1723	>	if (ctl == currentCtl)
1724	>	U.park(false, parkTime);
1725		w.parker = null;
1726		U.putObject(wt, PARKBLOCKER, null);
1727	<	if (ctl != c)
1727	>	if (ctl != currentCtl)
1728		break;
1729	<	if (System.nanoTime() - startTime >= SHRINK_TIMEOUT &&
1730	<	U.compareAndSwapLong(this, CTL, c, nc)) {
1731	<	w.runState = -1;          // shrink
1732	<	w.eventCount = (ec + E_SEQ) | E_MASK;
1729	>	if (deadline - System.nanoTime() <= 0L &&
1730	>	U.compareAndSwapLong(this, CTL, currentCtl, prevCtl)) {
1731	>	w.eventCount = (w.eventCount + E_SEQ) | E_MASK;
1732	>	w.qlock = -1;   // shrink
1733		break;
1734		}
1735		}
#	Line 1560 | Line 1737 | public class ForkJoinPool extends Abstra
1737		}
1738
1739		/**
1740	+	* Scans through queues looking for work while joining a task;
1741	+	* if any are present, signals.
1742	+	*
1743	+	* @param task to return early if done
1744	+	* @param origin an index to start scan
1745	+	*/
1746	+	final int helpSignal(ForkJoinTask<?> task, int origin) {
1747	+	WorkQueue[] ws; WorkQueue q; int m, n, s;
1748	+	if (task != null && (ws = workQueues) != null &&
1749	+	(m = ws.length - 1) >= 0) {
1750	+	for (int i = 0; i <= m; ++i) {
1751	+	if ((s = task.status) < 0)
1752	+	return s;
1753	+	if ((q = ws[(i + origin) & m]) != null &&
1754	+	(n = q.queueSize()) > 0) {
1755	+	signalWork(q, n);
1756	+	if ((int)(ctl >> AC_SHIFT) >= 0)
1757	+	break;
1758	+	}
1759	+	}
1760	+	}
1761	+	return 0;
1762	+	}
1763	+
1764	+	/**
1765		* Tries to locate and execute tasks for a stealer of the given
1766		* task, or in turn one of its stealers, Traces currentSteal ->
1767		* currentJoin links looking for a thread working on a descendant
#	Line 1570 | Line 1772 | public class ForkJoinPool extends Abstra
1772		* leaves hints in workers to speed up subsequent calls. The
1773		* implementation is very branchy to cope with potential
1774		* inconsistencies or loops encountering chains that are stale,
1775	<	* unknown, or of length greater than MAX_HELP_DEPTH links.  All
1574	<	* of these cases are dealt with by just retrying by caller.
1775	>	* unknown, or so long that they are likely cyclic.
1776		*
1777		* @param joiner the joining worker
1778		* @param task the task to join
1779	<	* @return true if found or ran a task (and so is immediately retryable)
1779	>	* @return 0 if no progress can be made, negative if task
1780	>	* known complete, else positive
1781		*/
1782	<	final boolean tryHelpStealer(WorkQueue joiner, ForkJoinTask<?> task) {
1783	<	ForkJoinTask<?> subtask;    // current target
1784	<	boolean progress = false;
1785	<	int depth = 0;              // current chain depth
1786	<	int m = runState & SMASK;
1787	<	WorkQueue[] ws = workQueues;
1788	<
1789	<	if (ws != null && ws.length > m && (subtask = task).status >= 0) {
1790	<	outer:for (WorkQueue j = joiner;;) {
1791	<	// Try to find the stealer of subtask, by first using hint
1792	<	WorkQueue stealer = null;
1793	<	WorkQueue v = ws[j.stealHint & m];
1794	<	if (v != null && v.currentSteal == subtask)
1795	<	stealer = v;
1796	<	else {
1797	<	for (int i = 1; i <= m; i += 2) {
1798	<	if ((v = ws[i]) != null && v.currentSteal == subtask) {
1799	<	stealer = v;
1800	<	j.stealHint = i; // save hint
1801	<	break;
1782	>	private int tryHelpStealer(WorkQueue joiner, ForkJoinTask<?> task) {
1783	>	int stat = 0, steps = 0;                    // bound to avoid cycles
1784	>	if (joiner != null && task != null) {       // hoist null checks
1785	>	restart: for (;;) {
1786	>	ForkJoinTask<?> subtask = task;     // current target
1787	>	for (WorkQueue j = joiner, v;;) {   // v is stealer of subtask
1788	>	WorkQueue[] ws; int m, s, h;
1789	>	if ((s = task.status) < 0) {
1790	>	stat = s;
1791	>	break restart;
1792	>	}
1793	>	if ((ws = workQueues) == null || (m = ws.length - 1) <= 0)
1794	>	break restart;              // shutting down
1795	>	if ((v = ws[h = (j.stealHint | 1) & m]) == null ||
1796	>	v.currentSteal != subtask) {
1797	>	for (int origin = h;;) {    // find stealer
1798	>	if (((h = (h + 2) & m) & 15) == 1 &&
1799	>	(subtask.status < 0 || j.currentJoin != subtask))
1800	>	continue restart;   // occasional staleness check
1801	>	if ((v = ws[h]) != null &&
1802	>	v.currentSteal == subtask) {
1803	>	j.stealHint = h;    // save hint
1804	>	break;
1805	>	}
1806	>	if (h == origin)
1807	>	break restart;      // cannot find stealer
1808		}
1809		}
1810	<	if (stealer == null)
1810	>	for (;;) { // help stealer or descend to its stealer
1811	>	ForkJoinTask[] a;  int b;
1812	>	if (subtask.status < 0)     // surround probes with
1813	>	continue restart;       //   consistency checks
1814	>	if ((b = v.base) - v.top < 0 && (a = v.array) != null) {
1815	>	int i = (((a.length - 1) & b) << ASHIFT) + ABASE;
1816	>	ForkJoinTask<?> t =
1817	>	(ForkJoinTask<?>)U.getObjectVolatile(a, i);
1818	>	if (subtask.status < 0 || j.currentJoin != subtask ||
1819	>	v.currentSteal != subtask)
1820	>	continue restart;   // stale
1821	>	stat = 1;               // apparent progress
1822	>	if (t != null && v.base == b &&
1823	>	U.compareAndSwapObject(a, i, t, null)) {
1824	>	v.base = b + 1;     // help stealer
1825	>	joiner.runSubtask(t);
1826	>	}
1827	>	else if (v.base == b && ++steps == MAX_HELP)
1828	>	break restart;      // v apparently stalled
1829	>	}
1830	>	else {                      // empty -- try to descend
1831	>	ForkJoinTask<?> next = v.currentJoin;
1832	>	if (subtask.status < 0 || j.currentJoin != subtask ||
1833	>	v.currentSteal != subtask)
1834	>	continue restart;   // stale
1835	>	else if (next == null || ++steps == MAX_HELP)
1836	>	break restart;      // dead-end or maybe cyclic
1837	>	else {
1838	>	subtask = next;
1839	>	j = v;
1840	>	break;
1841	>	}
1842	>	}
1843	>	}
1844	>	}
1845	>	}
1846	>	}
1847	>	return stat;
1848	>	}
1849	>
1850	>	/**
1851	>	* Analog of tryHelpStealer for CountedCompleters. Tries to steal
1852	>	* and run tasks within the target's computation
1853	>	*
1854	>	* @param task the task to join
1855	>	* @param mode if shared, exit upon completing any task
1856	>	* if all workers are active
1857	>	*
1858	>	*/
1859	>	private int helpComplete(ForkJoinTask<?> task, int mode) {
1860	>	WorkQueue[] ws; WorkQueue q; int m, n, s;
1861	>	if (task != null && (ws = workQueues) != null &&
1862	>	(m = ws.length - 1) >= 0) {
1863	>	for (int j = 1, origin = j;;) {
1864	>	if ((s = task.status) < 0)
1865	>	return s;
1866	>	if ((q = ws[j & m]) != null && q.pollAndExecCC(task)) {
1867	>	origin = j;
1868	>	if (mode == SHARED_QUEUE && (int)(ctl >> AC_SHIFT) >= 0)
1869		break;
1870		}
1871	+	else if ((j = (j + 2) & m) == origin)
1872	+	break;
1873	+	}
1874	+	}
1875	+	return 0;
1876	+	}
1877
1878	<	for (WorkQueue q = stealer;;) { // Try to help stealer
1879	<	ForkJoinTask<?> t; int b;
1880	<	if (task.status < 0)
1881	<	break outer;
1882	<	if ((b = q.base) - q.top < 0) {
1883	<	progress = true;
1884	<	if (subtask.status < 0)
1885	<	break outer;               // stale
1886	<	if ((t = q.pollAt(b)) != null) {
1887	<	stealer.stealHint = joiner.poolIndex;
1888	<	joiner.runSubtask(t);
1878	>	/**
1879	>	* Tries to decrement active count (sometimes implicitly) and
1880	>	* possibly release or create a compensating worker in preparation
1881	>	* for blocking. Fails on contention or termination. Otherwise,
1882	>	* adds a new thread if no idle workers are available and pool
1883	>	* may become starved.
1884	>	*/
1885	>	final boolean tryCompensate() {
1886	>	int pc = parallelism, e, u, i, tc; long c;
1887	>	WorkQueue[] ws; WorkQueue w; Thread p;
1888	>	if ((e = (int)(c = ctl)) >= 0 && (ws = workQueues) != null) {
1889	>	if (e != 0 && (i = e & SMASK) < ws.length &&
1890	>	(w = ws[i]) != null && w.eventCount == (e | INT_SIGN)) {
1891	>	long nc = ((long)(w.nextWait & E_MASK) |
1892	>	(c & (AC_MASK|TC_MASK)));
1893	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1894	>	w.eventCount = (e + E_SEQ) & E_MASK;
1895	>	if ((p = w.parker) != null)
1896	>	U.unpark(p);
1897	>	return true;   // replace with idle worker
1898	>	}
1899	>	}
1900	>	else if ((short)((u = (int)(c >>> 32)) >>> UTC_SHIFT) >= 0 &&
1901	>	(u >> UAC_SHIFT) + pc > 1) {
1902	>	long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK);
1903	>	if (U.compareAndSwapLong(this, CTL, c, nc))
1904	>	return true;    // no compensation
1905	>	}
1906	>	else if ((tc = u + pc) < MAX_CAP) {
1907	>	long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK);
1908	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1909	>	Throwable ex = null;
1910	>	ForkJoinWorkerThread wt = null;
1911	>	try {
1912	>	ForkJoinWorkerThreadFactory fac;
1913	>	if ((fac = factory) != null &&
1914	>	(wt = fac.newThread(this)) != null) {
1915	>	wt.start();
1916	>	return true;
1917		}
1918	+	} catch (Throwable rex) {
1919	+	ex = rex;
1920		}
1921	<	else { // empty - try to descend to find stealer's stealer
1620	<	ForkJoinTask<?> next = stealer.currentJoin;
1621	<	if (++depth == MAX_HELP_DEPTH || subtask.status < 0 ||
1622	<	next == null || next == subtask)
1623	<	break outer;  // max depth, stale, dead-end, cyclic
1624	<	subtask = next;
1625	<	j = stealer;
1626	<	break;
1627	<	}
1921	>	deregisterWorker(wt, ex); // adjust counts etc
1922		}
1923		}
1924		}
1925	<	return progress;
1925	>	return false;
1926		}
1927
1928		/**
1929	<	* If task is at base of some steal queue, steals and executes it.
1929	>	* Helps and/or blocks until the given task is done.
1930		*
1931		* @param joiner the joining worker
1932		* @param task the task
1933	+	* @return task status on exit
1934		*/
1935	<	final void tryPollForAndExec(WorkQueue joiner, ForkJoinTask<?> task) {
1936	<	WorkQueue[] ws;
1937	<	int m = runState & SMASK;
1938	<	if ((ws = workQueues) != null && ws.length > m) {
1939	<	for (int j = 1; j <= m && task.status >= 0; j += 2) {
1940	<	WorkQueue q = ws[j];
1941	<	if (q != null && q.pollFor(task)) {
1942	<	joiner.runSubtask(task);
1943	<	break;
1935	>	final int awaitJoin(WorkQueue joiner, ForkJoinTask<?> task) {
1936	>	int s = 0;
1937	>	if (joiner != null && task != null && (s = task.status) >= 0) {
1938	>	ForkJoinTask<?> prevJoin = joiner.currentJoin;
1939	>	joiner.currentJoin = task;
1940	>	do {} while ((s = task.status) >= 0 &&
1941	>	joiner.queueSize() > 0 &&
1942	>	joiner.tryRemoveAndExec(task)); // process local tasks
1943	>	if (s >= 0 && (s = task.status) >= 0 &&
1944	>	(s = helpSignal(task, joiner.poolIndex)) >= 0 &&
1945	>	(task instanceof CountedCompleter))
1946	>	s = helpComplete(task, LIFO_QUEUE);
1947	>	while (s >= 0 && (s = task.status) >= 0) {
1948	>	if ((joiner.queueSize() > 0 ||           // try helping
1949	>	(s = tryHelpStealer(joiner, task)) == 0) &&
1950	>	(s = task.status) >= 0 && tryCompensate()) {
1951	>	if (task.trySetSignal() && (s = task.status) >= 0) {
1952	>	synchronized (task) {
1953	>	if (task.status >= 0) {
1954	>	try {                // see ForkJoinTask
1955	>	task.wait();     //  for explanation
1956	>	} catch (InterruptedException ie) {
1957	>	}
1958	>	}
1959	>	else
1960	>	task.notifyAll();
1961	>	}
1962	>	}
1963	>	long c;                          // re-activate
1964	>	do {} while (!U.compareAndSwapLong
1965	>	(this, CTL, c = ctl, c + AC_UNIT));
1966		}
1967		}
1968	+	joiner.currentJoin = prevJoin;
1969		}
1970	+	return s;
1971		}
1972
1973		/**
1974	<	* Returns a non-empty steal queue, if one is found during a random,
1975	<	* then cyclic scan, else null.  This method must be retried by
1976	<	* caller if, by the time it tries to use the queue, it is empty.
1974	>	* Stripped-down variant of awaitJoin used by timed joins. Tries
1975	>	* to help join only while there is continuous progress. (Caller
1976	>	* will then enter a timed wait.)
1977	>	*
1978	>	* @param joiner the joining worker
1979	>	* @param task the task
1980		*/
1981	<	private WorkQueue findNonEmptyStealQueue(WorkQueue w) {
1982	<	int r = w.seed;    // Same idea as scan(), but ignoring submissions
1981	>	final void helpJoinOnce(WorkQueue joiner, ForkJoinTask<?> task) {
1982	>	int s;
1983	>	if (joiner != null && task != null && (s = task.status) >= 0) {
1984	>	ForkJoinTask<?> prevJoin = joiner.currentJoin;
1985	>	joiner.currentJoin = task;
1986	>	do {} while ((s = task.status) >= 0 &&
1987	>	joiner.queueSize() > 0 &&
1988	>	joiner.tryRemoveAndExec(task));
1989	>	if (s >= 0 && (s = task.status) >= 0 &&
1990	>	(s = helpSignal(task, joiner.poolIndex)) >= 0 &&
1991	>	(task instanceof CountedCompleter))
1992	>	s = helpComplete(task, LIFO_QUEUE);
1993	>	if (s >= 0 && joiner.queueSize() == 0) {
1994	>	do {} while (task.status >= 0 &&
1995	>	tryHelpStealer(joiner, task) > 0);
1996	>	}
1997	>	joiner.currentJoin = prevJoin;
1998	>	}
1999	>	}
2000	>
2001	>	/**
2002	>	* Returns a (probably) non-empty steal queue, if one is found
2003	>	* during a random, then cyclic scan, else null.  This method must
2004	>	* be retried by caller if, by the time it tries to use the queue,
2005	>	* it is empty.
2006	>	* @param r a (random) seed for scanning
2007	>	*/
2008	>	private WorkQueue findNonEmptyStealQueue(int r) {
2009	>	int step = (r >>> 16) | 1;
2010		for (WorkQueue[] ws;;) {
2011	<	int m = runState & SMASK;
2012	<	if ((ws = workQueues) == null)
2011	>	int ps = plock, m;
2012	>	if ((ws = workQueues) == null || (m = ws.length - 1) < 1)
2013		return null;
2014	<	if (ws.length > m) {
2015	<	WorkQueue q;
2016	<	for (int n = m << 2, k = r, j = -n;;) {
2017	<	r ^= r << 13; r ^= r >>> 17; r ^= r << 5;
2018	<	if ((q = ws[(k | 1) & m]) != null && q.base - q.top < 0) {
2019	<	w.seed = r;
1671	<	return q;
1672	<	}
1673	<	else if (j > n)
2014	>	for (int j = (m + 1) << 2; ; r += step) {
2015	>	WorkQueue q = ws[((r << 1) | 1) & m];
2016	>	if (q != null && q.queueSize() > 0)
2017	>	return q;
2018	>	else if (--j < 0) {
2019	>	if (plock == ps)
2020		return null;
2021	<	else
1676	<	k = (j++ < 0) ? r : k + ((m >>> 1) | 1);
1677	<
2021	>	break;
2022		}
2023		}
2024		}
#	Line 1688 | Line 2032 | public class ForkJoinPool extends Abstra
2032		*/
2033		final void helpQuiescePool(WorkQueue w) {
2034		for (boolean active = true;;) {
2035	<	w.runLocalTasks();      // exhaust local queue
2036	<	WorkQueue q = findNonEmptyStealQueue(w);
2035	>	ForkJoinTask<?> localTask; // exhaust local queue
2036	>	while ((localTask = w.nextLocalTask()) != null)
2037	>	localTask.doExec();
2038	>	// Similar to loop in scan(), but ignoring submissions
2039	>	WorkQueue q = findNonEmptyStealQueue(w.nextSeed());
2040		if (q != null) {
2041	<	ForkJoinTask<?> t;
2041	>	ForkJoinTask<?> t; int b;
2042		if (!active) {      // re-establish active count
2043		long c;
2044		active = true;
2045		do {} while (!U.compareAndSwapLong
2046		(this, CTL, c = ctl, c + AC_UNIT));
2047		}
2048	<	if ((t = q.poll()) != null)
2048	>	if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null)
2049		w.runSubtask(t);
2050		}
2051		else {
#	Line 1720 | Line 2067 | public class ForkJoinPool extends Abstra
2067		}
2068
2069		/**
2070	<	* Gets and removes a local or stolen task for the given worker
2070	>	* Gets and removes a local or stolen task for the given worker.
2071		*
2072		* @return a task, if available
2073		*/
2074		final ForkJoinTask<?> nextTaskFor(WorkQueue w) {
2075		for (ForkJoinTask<?> t;;) {
2076	<	WorkQueue q;
2076	>	WorkQueue q; int b;
2077		if ((t = w.nextLocalTask()) != null)
2078		return t;
2079	<	if ((q = findNonEmptyStealQueue(w)) == null)
2079	>	if ((q = findNonEmptyStealQueue(w.nextSeed())) == null)
2080		return null;
2081	<	if ((t = q.poll()) != null)
2081	>	if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null)
2082		return t;
2083		}
2084		}
2085
2086		/**
2087	<	* Returns the approximate (non-atomic) number of idle threads per
2088	<	* active thread to offset steal queue size for method
2089	<	* ForkJoinTask.getSurplusQueuedTaskCount().
2090	<	*/
2091	<	final int idlePerActive() {
2092	<	// Approximate at powers of two for small values, saturate past 4
2093	<	int p = parallelism;
2094	<	int a = p + (int)(ctl >> AC_SHIFT);
2095	<	return (a > (p >>>= 1) ? 0 :
2096	<	a > (p >>>= 1) ? 1 :
2097	<	a > (p >>>= 1) ? 2 :
2098	<	a > (p >>>= 1) ? 4 :
2099	<	8);
2100	<	}
2101	<
2102	<	// Termination
2103	<
2104	<	/**
2105	<	* Sets SHUTDOWN bit of runState under lock
2106	<	*/
2107	<	private void enableShutdown() {
2108	<	ReentrantLock lock = this.lock;
2109	<	if (runState >= 0) {
2110	<	lock.lock();                       // don't need try/finally
2111	<	runState |= SHUTDOWN;
2112	<	lock.unlock();
2087	>	* Returns a cheap heuristic guide for task partitioning when
2088	>	* programmers, frameworks, tools, or languages have little or no
2089	>	* idea about task granularity.  In essence by offering this
2090	>	* method, we ask users only about tradeoffs in overhead vs
2091	>	* expected throughput and its variance, rather than how finely to
2092	>	* partition tasks.
2093	>	*
2094	>	* In a steady state strict (tree-structured) computation, each
2095	>	* thread makes available for stealing enough tasks for other
2096	>	* threads to remain active. Inductively, if all threads play by
2097	>	* the same rules, each thread should make available only a
2098	>	* constant number of tasks.
2099	>	*
2100	>	* The minimum useful constant is just 1. But using a value of 1
2101	>	* would require immediate replenishment upon each steal to
2102	>	* maintain enough tasks, which is infeasible.  Further,
2103	>	* partitionings/granularities of offered tasks should minimize
2104	>	* steal rates, which in general means that threads nearer the top
2105	>	* of computation tree should generate more than those nearer the
2106	>	* bottom. In perfect steady state, each thread is at
2107	>	* approximately the same level of computation tree. However,
2108	>	* producing extra tasks amortizes the uncertainty of progress and
2109	>	* diffusion assumptions.
2110	>	*
2111	>	* So, users will want to use values larger, but not much larger
2112	>	* than 1 to both smooth over transient shortages and hedge
2113	>	* against uneven progress; as traded off against the cost of
2114	>	* extra task overhead. We leave the user to pick a threshold
2115	>	* value to compare with the results of this call to guide
2116	>	* decisions, but recommend values such as 3.
2117	>	*
2118	>	* When all threads are active, it is on average OK to estimate
2119	>	* surplus strictly locally. In steady-state, if one thread is
2120	>	* maintaining say 2 surplus tasks, then so are others. So we can
2121	>	* just use estimated queue length.  However, this strategy alone
2122	>	* leads to serious mis-estimates in some non-steady-state
2123	>	* conditions (ramp-up, ramp-down, other stalls). We can detect
2124	>	* many of these by further considering the number of "idle"
2125	>	* threads, that are known to have zero queued tasks, so
2126	>	* compensate by a factor of (#idle/#active) threads.
2127	>	*
2128	>	* Note: The approximation of #busy workers as #active workers is
2129	>	* not very good under current signalling scheme, and should be
2130	>	* improved.
2131	>	*/
2132	>	static int getSurplusQueuedTaskCount() {
2133	>	Thread t; ForkJoinWorkerThread wt; ForkJoinPool pool; WorkQueue q;
2134	>	if (((t = Thread.currentThread()) instanceof ForkJoinWorkerThread)) {
2135	>	int b = (q = (wt = (ForkJoinWorkerThread)t).workQueue).base;
2136	>	int p = (pool = wt.pool).parallelism;
2137	>	int a = (int)(pool.ctl >> AC_SHIFT) + p;
2138	>	return q.top - b - (a > (p >>>= 1) ? 0 :
2139	>	a > (p >>>= 1) ? 1 :
2140	>	a > (p >>>= 1) ? 2 :
2141	>	a > (p >>>= 1) ? 4 :
2142	>	8);
2143		}
2144	+	return 0;
2145		}
2146
2147	+	//  Termination
2148	+
2149		/**
2150	<	* Possibly initiates and/or completes termination.  Upon
2151	<	* termination, cancels all queued tasks and then
2150	>	* Possibly initiates and/or completes termination.  The caller
2151	>	* triggering termination runs three passes through workQueues:
2152	>	* (0) Setting termination status, followed by wakeups of queued
2153	>	* workers; (1) cancelling all tasks; (2) interrupting lagging
2154	>	* threads (likely in external tasks, but possibly also blocked in
2155	>	* joins).  Each pass repeats previous steps because of potential
2156	>	* lagging thread creation.
2157		*
2158		* @param now if true, unconditionally terminate, else only
2159		* if no work and no active workers
2160	+	* @param enable if true, enable shutdown when next possible
2161		* @return true if now terminating or terminated
2162		*/
2163	<	private boolean tryTerminate(boolean now) {
2163	>	private boolean tryTerminate(boolean now, boolean enable) {
2164	>	if (this == commonPool)                     // cannot shut down
2165	>	return false;
2166		for (long c;;) {
2167		if (((c = ctl) & STOP_BIT) != 0) {      // already terminating
2168		if ((short)(c >>> TC_SHIFT) == -parallelism) {
2169	<	ReentrantLock lock = this.lock; // signal when no workers
2170	<	lock.lock();                    // don't need try/finally
2171	<	termination.signalAll();        // signal when 0 workers
1784	<	lock.unlock();
2169	>	synchronized (this) {
2170	>	notifyAll();                // signal when 0 workers
2171	>	}
2172		}
2173		return true;
2174		}
2175	<	if (!now) {
2176	<	if ((int)(c >> AC_SHIFT) != -parallelism || runState >= 0 ||
2175	>	if (plock >= 0) {                       // not yet enabled
2176	>	int ps;
2177	>	if (!enable)
2178	>	return false;
2179	>	if (((ps = plock) & PL_LOCK) != 0 ||
2180	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
2181	>	ps = acquirePlock();
2182	>	int nps = SHUTDOWN;
2183	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
2184	>	releasePlock(nps);
2185	>	}
2186	>	if (!now) {                             // check if idle & no tasks
2187	>	if ((int)(c >> AC_SHIFT) != -parallelism ||
2188		hasQueuedSubmissions())
2189		return false;
2190		// Check for unqueued inactive workers. One pass suffices.
2191		WorkQueue[] ws = workQueues; WorkQueue w;
2192		if (ws != null) {
2193	<	int n = ws.length;
1796	<	for (int i = 1; i < n; i += 2) {
2193	>	for (int i = 1; i < ws.length; i += 2) {
2194		if ((w = ws[i]) != null && w.eventCount >= 0)
2195		return false;
2196		}
2197		}
2198		}
2199	<	if (U.compareAndSwapLong(this, CTL, c, c | STOP_BIT))
2200	<	startTerminating();
2199	>	if (U.compareAndSwapLong(this, CTL, c, c | STOP_BIT)) {
2200	>	for (int pass = 0; pass < 3; ++pass) {
2201	>	WorkQueue[] ws = workQueues;
2202	>	if (ws != null) {
2203	>	WorkQueue w;
2204	>	int n = ws.length;
2205	>	for (int i = 0; i < n; ++i) {
2206	>	if ((w = ws[i]) != null) {
2207	>	w.qlock = -1;
2208	>	if (pass > 0) {
2209	>	w.cancelAll();
2210	>	if (pass > 1)
2211	>	w.interruptOwner();
2212	>	}
2213	>	}
2214	>	}
2215	>	// Wake up workers parked on event queue
2216	>	int i, e; long cc; Thread p;
2217	>	while ((e = (int)(cc = ctl) & E_MASK) != 0 &&
2218	>	(i = e & SMASK) < n &&
2219	>	(w = ws[i]) != null) {
2220	>	long nc = ((long)(w.nextWait & E_MASK) |
2221	>	((cc + AC_UNIT) & AC_MASK) |
2222	>	(cc & (TC_MASK|STOP_BIT)));
2223	>	if (w.eventCount == (e | INT_SIGN) &&
2224	>	U.compareAndSwapLong(this, CTL, cc, nc)) {
2225	>	w.eventCount = (e + E_SEQ) & E_MASK;
2226	>	w.qlock = -1;
2227	>	if ((p = w.parker) != null)
2228	>	U.unpark(p);
2229	>	}
2230	>	}
2231	>	}
2232	>	}
2233	>	}
2234		}
2235		}
2236
2237	+	// external operations on common pool
2238	+
2239		/**
2240	<	* Initiates termination: Runs three passes through workQueues:
2241	<	* (0) Setting termination status, followed by wakeups of queued
1810	<	* workers; (1) cancelling all tasks; (2) interrupting lagging
1811	<	* threads (likely in external tasks, but possibly also blocked in
1812	<	* joins).  Each pass repeats previous steps because of potential
1813	<	* lagging thread creation.
2240	>	* Returns common pool queue for a thread that has submitted at
2241	>	* least one task.
2242		*/
2243	<	private void startTerminating() {
2244	<	for (int pass = 0; pass < 3; ++pass) {
2245	<	WorkQueue[] ws = workQueues;
2246	<	if (ws != null) {
2247	<	WorkQueue w; Thread wt;
2248	<	int n = ws.length;
2249	<	for (int j = 0; j < n; ++j) {
2250	<	if ((w = ws[j]) != null) {
2251	<	w.runState = -1;
2252	<	if (pass > 0) {
2253	<	w.cancelAll();
2254	<	if (pass > 1 && (wt = w.owner) != null &&
2255	<	!wt.isInterrupted()) {
2256	<	try {
2257	<	wt.interrupt();
2258	<	} catch (SecurityException ignore) {
2243	>	static WorkQueue commonSubmitterQueue() {
2244	>	ForkJoinPool p; WorkQueue[] ws; int m; Submitter z;
2245	>	return ((z = submitters.get()) != null &&
2246	>	(p = commonPool) != null &&
2247	>	(ws = p.workQueues) != null &&
2248	>	(m = ws.length - 1) >= 0) ?
2249	>	ws[m & z.seed & SQMASK] : null;
2250	>	}
2251	>
2252	>	/**
2253	>	* Tries to pop the given task from submitter's queue in common pool.
2254	>	*/
2255	>	static boolean tryExternalUnpush(ForkJoinTask<?> t) {
2256	>	ForkJoinPool p; WorkQueue[] ws; WorkQueue q; Submitter z;
2257	>	ForkJoinTask<?>[] a;  int m, s; long j;
2258	>	if ((z = submitters.get()) != null &&
2259	>	(p = commonPool) != null &&
2260	>	(ws = p.workQueues) != null &&
2261	>	(m = ws.length - 1) >= 0 &&
2262	>	(q = ws[m & z.seed & SQMASK]) != null &&
2263	>	(s = q.top) != q.base &&
2264	>	(a = q.array) != null &&
2265	>	U.getObjectVolatile
2266	>	(a, j = (((a.length - 1) & (s - 1)) << ASHIFT) + ABASE) == t &&
2267	>	U.compareAndSwapInt(q, QLOCK, 0, 1)) {
2268	>	if (q.array == a && q.top == s && // recheck
2269	>	U.compareAndSwapObject(a, j, t, null)) {
2270	>	q.top = s - 1;
2271	>	q.qlock = 0;
2272	>	return true;
2273	>	}
2274	>	q.qlock = 0;
2275	>	}
2276	>	return false;
2277	>	}
2278	>
2279	>	/**
2280	>	* Tries to pop and run local tasks within the same computation
2281	>	* as the given root. On failure, tries to help complete from
2282	>	* other queues via helpComplete.
2283	>	*/
2284	>	private void externalHelpComplete(WorkQueue q, ForkJoinTask<?> root) {
2285	>	ForkJoinTask<?>[] a; int m;
2286	>	if (q != null && (a = q.array) != null && (m = (a.length - 1)) >= 0 &&
2287	>	root != null && root.status >= 0) {
2288	>	for (;;) {
2289	>	int s; Object o; CountedCompleter<?> task = null;
2290	>	if ((s = q.top) - q.base > 0) {
2291	>	long j = ((m & (s - 1)) << ASHIFT) + ABASE;
2292	>	if ((o = U.getObject(a, j)) != null &&
2293	>	(o instanceof CountedCompleter)) {
2294	>	CountedCompleter<?> t = (CountedCompleter<?>)o, r = t;
2295	>	do {
2296	>	if (r == root) {
2297	>	if (U.compareAndSwapInt(q, QLOCK, 0, 1)) {
2298	>	if (q.array == a && q.top == s &&
2299	>	U.compareAndSwapObject(a, j, t, null)) {
2300	>	q.top = s - 1;
2301	>	task = t;
2302	>	}
2303	>	q.qlock = 0;
2304		}
2305	+	break;
2306		}
2307	<	}
2307	>	} while ((r = r.completer) != null);
2308		}
2309		}
2310	<	// Wake up workers parked on event queue
2311	<	int i, e; long c; Thread p;
2312	<	while ((i = ((~(e = (int)(c = ctl)) << 1) | 1) & SMASK) < n &&
2313	<	(w = ws[i]) != null &&
2314	<	w.eventCount == (e | INT_SIGN)) {
2315	<	long nc = ((long)(w.nextWait & E_MASK) |
2316	<	((c + AC_UNIT) & AC_MASK) |
2317	<	(c & (TC_MASK|STOP_BIT)));
1844	<	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1845	<	w.eventCount = (e + E_SEQ) & E_MASK;
1846	<	if ((p = w.parker) != null)
1847	<	U.unpark(p);
1848	<	}
2310	>	if (task != null)
2311	>	task.doExec();
2312	>	if (root.status < 0 || (int)(ctl >> AC_SHIFT) >= 0)
2313	>	break;
2314	>	if (task == null) {
2315	>	if (helpSignal(root, q.poolIndex) >= 0)
2316	>	helpComplete(root, SHARED_QUEUE);
2317	>	break;
2318		}
2319		}
2320		}
2321		}
2322
2323	+	/**
2324	+	* Tries to help execute or signal availability of the given task
2325	+	* from submitter's queue in common pool.
2326	+	*/
2327	+	static void externalHelpJoin(ForkJoinTask<?> t) {
2328	+	// Some hard-to-avoid overlap with tryExternalUnpush
2329	+	ForkJoinPool p; WorkQueue[] ws; WorkQueue q, w; Submitter z;
2330	+	ForkJoinTask<?>[] a;  int m, s, n; long j;
2331	+	if (t != null && t.status >= 0 &&
2332	+	(z = submitters.get()) != null &&
2333	+	(p = commonPool) != null &&
2334	+	(ws = p.workQueues) != null &&
2335	+	(m = ws.length - 1) >= 0 &&
2336	+	(q = ws[m & z.seed & SQMASK]) != null &&
2337	+	(a = q.array) != null) {
2338	+	if ((s = q.top) != q.base &&
2339	+	U.getObjectVolatile
2340	+	(a, j = (((a.length - 1) & (s - 1)) << ASHIFT) + ABASE) == t &&
2341	+	U.compareAndSwapInt(q, QLOCK, 0, 1)) {
2342	+	if (q.array == a && q.top == s &&
2343	+	U.compareAndSwapObject(a, j, t, null)) {
2344	+	q.top = s - 1;
2345	+	q.qlock = 0;
2346	+	t.doExec();
2347	+	}
2348	+	else
2349	+	q.qlock = 0;
2350	+	}
2351	+	if (t.status >= 0) {
2352	+	if (t instanceof CountedCompleter)
2353	+	p.externalHelpComplete(q, t);
2354	+	else
2355	+	p.helpSignal(t, q.poolIndex);
2356	+	}
2357	+	}
2358	+	}
2359	+
2360	+	/**
2361	+	* Restricted version of helpQuiescePool for external callers
2362	+	*/
2363	+	static void externalHelpQuiescePool() {
2364	+	ForkJoinPool p; ForkJoinTask<?> t; WorkQueue q; int b;
2365	+	int r = ThreadLocalRandom.current().nextInt();
2366	+	if ((p = commonPool) != null &&
2367	+	(q = p.findNonEmptyStealQueue(r)) != null &&
2368	+	(b = q.base) - q.top < 0 &&
2369	+	(t = q.pollAt(b)) != null)
2370	+	t.doExec();
2371	+	}
2372	+
2373		// Exported methods
2374
2375		// Constructors
#	Line 1920 | Line 2439 | public class ForkJoinPool extends Abstra
2439		checkPermission();
2440		if (factory == null)
2441		throw new NullPointerException();
2442	<	if (parallelism <= 0 || parallelism > MAX_ID)
2442	>	if (parallelism <= 0 || parallelism > MAX_CAP)
2443		throw new IllegalArgumentException();
2444		this.parallelism = parallelism;
2445		this.factory = factory;
2446		this.ueh = handler;
2447		this.localMode = asyncMode ? FIFO_QUEUE : LIFO_QUEUE;
1929	–	this.nextPoolIndex = 1;
2448		long np = (long)(-parallelism); // offset ctl counts
2449		this.ctl = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK);
2450	<	// initialize workQueues array with room for 2*parallelism if possible
1933	<	int n = parallelism << 1;
1934	<	if (n >= MAX_ID)
1935	<	n = MAX_ID;
1936	<	else { // See Hackers Delight, sec 3.2, where n < (1 << 16)
1937	<	n |= n >>> 1; n |= n >>> 2; n |= n >>> 4; n |= n >>> 8;
1938	<	}
1939	<	this.workQueues = new WorkQueue[(n + 1) << 1];
1940	<	ReentrantLock lck = this.lock = new ReentrantLock();
1941	<	this.termination = lck.newCondition();
1942	<	this.stealCount = new AtomicLong();
1943	<	this.nextWorkerNumber = new AtomicInteger();
2450	>	int pn = nextPoolId();
2451		StringBuilder sb = new StringBuilder("ForkJoinPool-");
2452	<	sb.append(poolNumberGenerator.incrementAndGet());
2452	>	sb.append(Integer.toString(pn));
2453		sb.append("-worker-");
2454		this.workerNamePrefix = sb.toString();
2455	<	// Create initial submission queue
2456	<	WorkQueue sq = tryAddSharedQueue(0);
2457	<	if (sq != null)
2458	<	sq.growArray(false);
2455	>	}
2456	>
2457	>	/**
2458	>	* Constructor for common pool, suitable only for static initialization.
2459	>	* Basically the same as above, but uses smallest possible initial footprint.
2460	>	*/
2461	>	ForkJoinPool(int parallelism, long ctl,
2462	>	ForkJoinWorkerThreadFactory factory,
2463	>	Thread.UncaughtExceptionHandler handler) {
2464	>	this.parallelism = parallelism;
2465	>	this.ctl = ctl;
2466	>	this.factory = factory;
2467	>	this.ueh = handler;
2468	>	this.localMode = LIFO_QUEUE;
2469	>	this.workerNamePrefix = "ForkJoinPool.commonPool-worker-";
2470	>	}
2471	>
2472	>	/**
2473	>	* Returns the common pool instance.
2474	>	*
2475	>	* @return the common pool instance
2476	>	*/
2477	>	public static ForkJoinPool commonPool() {
2478	>	return commonPool; // cannot be null (if so, a static init error)
2479		}
2480
2481		// Execution methods
#	Line 1970 | Line 2497 | public class ForkJoinPool extends Abstra
2497		*         scheduled for execution
2498		*/
2499		public <T> T invoke(ForkJoinTask<T> task) {
2500	<	doSubmit(task);
2500	>	if (task == null)
2501	>	throw new NullPointerException();
2502	>	externalPush(task);
2503		return task.join();
2504		}
2505
#	Line 1983 | Line 2512 | public class ForkJoinPool extends Abstra
2512		*         scheduled for execution
2513		*/
2514		public void execute(ForkJoinTask<?> task) {
2515	<	doSubmit(task);
2515	>	if (task == null)
2516	>	throw new NullPointerException();
2517	>	externalPush(task);
2518		}
2519
2520		// AbstractExecutorService methods
#	Line 2000 | Line 2531 | public class ForkJoinPool extends Abstra
2531		if (task instanceof ForkJoinTask<?>) // avoid re-wrap
2532		job = (ForkJoinTask<?>) task;
2533		else
2534	<	job = ForkJoinTask.adapt(task, null);
2535	<	doSubmit(job);
2534	>	job = new ForkJoinTask.AdaptedRunnableAction(task);
2535	>	externalPush(job);
2536		}
2537
2538		/**
#	Line 2014 | Line 2545 | public class ForkJoinPool extends Abstra
2545		*         scheduled for execution
2546		*/
2547		public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) {
2548	<	doSubmit(task);
2548	>	if (task == null)
2549	>	throw new NullPointerException();
2550	>	externalPush(task);
2551		return task;
2552		}
2553
#	Line 2024 | Line 2557 | public class ForkJoinPool extends Abstra
2557		*         scheduled for execution
2558		*/
2559		public <T> ForkJoinTask<T> submit(Callable<T> task) {
2560	<	if (task == null)
2561	<	throw new NullPointerException();
2029	<	ForkJoinTask<T> job = ForkJoinTask.adapt(task);
2030	<	doSubmit(job);
2560	>	ForkJoinTask<T> job = new ForkJoinTask.AdaptedCallable<T>(task);
2561	>	externalPush(job);
2562		return job;
2563		}
2564
#	Line 2037 | Line 2568 | public class ForkJoinPool extends Abstra
2568		*         scheduled for execution
2569		*/
2570		public <T> ForkJoinTask<T> submit(Runnable task, T result) {
2571	<	if (task == null)
2572	<	throw new NullPointerException();
2042	<	ForkJoinTask<T> job = ForkJoinTask.adapt(task, result);
2043	<	doSubmit(job);
2571	>	ForkJoinTask<T> job = new ForkJoinTask.AdaptedRunnable<T>(task, result);
2572	>	externalPush(job);
2573		return job;
2574		}
2575
#	Line 2056 | Line 2585 | public class ForkJoinPool extends Abstra
2585		if (task instanceof ForkJoinTask<?>) // avoid re-wrap
2586		job = (ForkJoinTask<?>) task;
2587		else
2588	<	job = ForkJoinTask.adapt(task, null);
2589	<	doSubmit(job);
2588	>	job = new ForkJoinTask.AdaptedRunnableAction(task);
2589	>	externalPush(job);
2590		return job;
2591		}
2592
#	Line 2066 | Line 2595 | public class ForkJoinPool extends Abstra
2595		* @throws RejectedExecutionException {@inheritDoc}
2596		*/
2597		public <T> List<Future<T>> invokeAll(Collection<? extends Callable<T>> tasks) {
2598	<	ArrayList<ForkJoinTask<T>> forkJoinTasks =
2599	<	new ArrayList<ForkJoinTask<T>>(tasks.size());
2600	<	for (Callable<T> task : tasks)
2601	<	forkJoinTasks.add(ForkJoinTask.adapt(task));
2602	<	invoke(new InvokeAll<T>(forkJoinTasks));
2603	<
2598	>	// In previous versions of this class, this method constructed
2599	>	// a task to run ForkJoinTask.invokeAll, but now external
2600	>	// invocation of multiple tasks is at least as efficient.
2601	>	List<ForkJoinTask<T>> fs = new ArrayList<ForkJoinTask<T>>(tasks.size());
2602	>	// Workaround needed because method wasn't declared with
2603	>	// wildcards in return type but should have been.
2604		@SuppressWarnings({"unchecked", "rawtypes"})
2605	<	List<Future<T>> futures = (List<Future<T>>) (List) forkJoinTasks;
2077	<	return futures;
2078	<	}
2605	>	List<Future<T>> futures = (List<Future<T>>) (List) fs;
2606
2607	<	static final class InvokeAll<T> extends RecursiveAction {
2608	<	final ArrayList<ForkJoinTask<T>> tasks;
2609	<	InvokeAll(ArrayList<ForkJoinTask<T>> tasks) { this.tasks = tasks; }
2610	<	public void compute() {
2611	<	try { invokeAll(tasks); }
2612	<	catch (Exception ignore) {}
2607	>	boolean done = false;
2608	>	try {
2609	>	for (Callable<T> t : tasks) {
2610	>	ForkJoinTask<T> f = new ForkJoinTask.AdaptedCallable<T>(t);
2611	>	externalPush(f);
2612	>	fs.add(f);
2613	>	}
2614	>	for (ForkJoinTask<T> f : fs)
2615	>	f.quietlyJoin();
2616	>	done = true;
2617	>	return futures;
2618	>	} finally {
2619	>	if (!done)
2620	>	for (ForkJoinTask<T> f : fs)
2621	>	f.cancel(false);
2622		}
2087	–	private static final long serialVersionUID = -7914297376763021607L;
2623		}
2624
2625		/**
#	Line 2116 | Line 2651 | public class ForkJoinPool extends Abstra
2651		}
2652
2653		/**
2654	+	* Returns the targeted parallelism level of the common pool.
2655	+	*
2656	+	* @return the targeted parallelism level of the common pool
2657	+	*/
2658	+	public static int getCommonPoolParallelism() {
2659	+	return commonPoolParallelism;
2660	+	}
2661	+
2662	+	/**
2663		* Returns the number of worker threads that have started but not
2664		* yet terminated.  The result returned by this method may differ
2665		* from {@link #getParallelism} when threads are created to
#	Line 2149 | Line 2693 | public class ForkJoinPool extends Abstra
2693		int rc = 0;
2694		WorkQueue[] ws; WorkQueue w;
2695		if ((ws = workQueues) != null) {
2696	<	int n = ws.length;
2697	<	for (int i = 1; i < n; i += 2) {
2154	<	Thread.State s; ForkJoinWorkerThread wt;
2155	<	if ((w = ws[i]) != null && (wt = w.owner) != null &&
2156	<	w.eventCount >= 0 &&
2157	<	(s = wt.getState()) != Thread.State.BLOCKED &&
2158	<	s != Thread.State.WAITING &&
2159	<	s != Thread.State.TIMED_WAITING)
2696	>	for (int i = 1; i < ws.length; i += 2) {
2697	>	if ((w = ws[i]) != null && w.isApparentlyUnblocked())
2698		++rc;
2699		}
2700		}
#	Line 2202 | Line 2740 | public class ForkJoinPool extends Abstra
2740		* @return the number of steals
2741		*/
2742		public long getStealCount() {
2743	<	long count = stealCount.get();
2743	>	long count = stealCount;
2744		WorkQueue[] ws; WorkQueue w;
2745		if ((ws = workQueues) != null) {
2746	<	int n = ws.length;
2209	<	for (int i = 1; i < n; i += 2) {
2746	>	for (int i = 1; i < ws.length; i += 2) {
2747		if ((w = ws[i]) != null)
2748	<	count += w.totalSteals;
2748	>	count += w.nsteals;
2749		}
2750		}
2751		return count;
#	Line 2228 | Line 2765 | public class ForkJoinPool extends Abstra
2765		long count = 0;
2766		WorkQueue[] ws; WorkQueue w;
2767		if ((ws = workQueues) != null) {
2768	<	int n = ws.length;
2232	<	for (int i = 1; i < n; i += 2) {
2768	>	for (int i = 1; i < ws.length; i += 2) {
2769		if ((w = ws[i]) != null)
2770		count += w.queueSize();
2771		}
#	Line 2248 | Line 2784 | public class ForkJoinPool extends Abstra
2784		int count = 0;
2785		WorkQueue[] ws; WorkQueue w;
2786		if ((ws = workQueues) != null) {
2787	<	int n = ws.length;
2252	<	for (int i = 0; i < n; i += 2) {
2787	>	for (int i = 0; i < ws.length; i += 2) {
2788		if ((w = ws[i]) != null)
2789		count += w.queueSize();
2790		}
#	Line 2266 | Line 2801 | public class ForkJoinPool extends Abstra
2801		public boolean hasQueuedSubmissions() {
2802		WorkQueue[] ws; WorkQueue w;
2803		if ((ws = workQueues) != null) {
2804	<	int n = ws.length;
2270	<	for (int i = 0; i < n; i += 2) {
2804	>	for (int i = 0; i < ws.length; i += 2) {
2805		if ((w = ws[i]) != null && w.queueSize() != 0)
2806		return true;
2807		}
#	Line 2285 | Line 2819 | public class ForkJoinPool extends Abstra
2819		protected ForkJoinTask<?> pollSubmission() {
2820		WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2821		if ((ws = workQueues) != null) {
2822	<	int n = ws.length;
2289	<	for (int i = 0; i < n; i += 2) {
2822	>	for (int i = 0; i < ws.length; i += 2) {
2823		if ((w = ws[i]) != null && (t = w.poll()) != null)
2824		return t;
2825		}
#	Line 2315 | Line 2848 | public class ForkJoinPool extends Abstra
2848		int count = 0;
2849		WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2850		if ((ws = workQueues) != null) {
2851	<	int n = ws.length;
2319	<	for (int i = 0; i < n; ++i) {
2851	>	for (int i = 0; i < ws.length; ++i) {
2852		if ((w = ws[i]) != null) {
2853		while ((t = w.poll()) != null) {
2854		c.add(t);
#	Line 2336 | Line 2868 | public class ForkJoinPool extends Abstra
2868		* @return a string identifying this pool, as well as its state
2869		*/
2870		public String toString() {
2871	<	long st = getStealCount();
2872	<	long qt = getQueuedTaskCount();
2873	<	long qs = getQueuedSubmissionCount();
2342	<	int rc = getRunningThreadCount();
2343	<	int pc = parallelism;
2871	>	// Use a single pass through workQueues to collect counts
2872	>	long qt = 0L, qs = 0L; int rc = 0;
2873	>	long st = stealCount;
2874		long c = ctl;
2875	+	WorkQueue[] ws; WorkQueue w;
2876	+	if ((ws = workQueues) != null) {
2877	+	for (int i = 0; i < ws.length; ++i) {
2878	+	if ((w = ws[i]) != null) {
2879	+	int size = w.queueSize();
2880	+	if ((i & 1) == 0)
2881	+	qs += size;
2882	+	else {
2883	+	qt += size;
2884	+	st += w.nsteals;
2885	+	if (w.isApparentlyUnblocked())
2886	+	++rc;
2887	+	}
2888	+	}
2889	+	}
2890	+	}
2891	+	int pc = parallelism;
2892		int tc = pc + (short)(c >>> TC_SHIFT);
2893		int ac = pc + (int)(c >> AC_SHIFT);
2894		if (ac < 0) // ignore transient negative
#	Line 2350 | Line 2897 | public class ForkJoinPool extends Abstra
2897		if ((c & STOP_BIT) != 0)
2898		level = (tc == 0) ? "Terminated" : "Terminating";
2899		else
2900	<	level = runState < 0 ? "Shutting down" : "Running";
2900	>	level = plock < 0 ? "Shutting down" : "Running";
2901		return super.toString() +
2902		"[" + level +
2903		", parallelism = " + pc +
#	Line 2364 | Line 2911 | public class ForkJoinPool extends Abstra
2911		}
2912
2913		/**
2914	<	* Initiates an orderly shutdown in which previously submitted
2915	<	* tasks are executed, but no new tasks will be accepted.
2916	<	* Invocation has no additional effect if already shut down.
2917	<	* Tasks that are in the process of being submitted concurrently
2918	<	* during the course of this method may or may not be rejected.
2914	>	* Possibly initiates an orderly shutdown in which previously
2915	>	* submitted tasks are executed, but no new tasks will be
2916	>	* accepted. Invocation has no effect on execution state if this
2917	>	* is the {@link #commonPool}, and no additional effect if
2918	>	* already shut down.  Tasks that are in the process of being
2919	>	* submitted concurrently during the course of this method may or
2920	>	* may not be rejected.
2921		*
2922		* @throws SecurityException if a security manager exists and
2923		*         the caller is not permitted to modify threads
#	Line 2377 | Line 2926 | public class ForkJoinPool extends Abstra
2926		*/
2927		public void shutdown() {
2928		checkPermission();
2929	<	enableShutdown();
2381	<	tryTerminate(false);
2929	>	tryTerminate(false, true);
2930		}
2931
2932		/**
2933	<	* Attempts to cancel and/or stop all tasks, and reject all
2934	<	* subsequently submitted tasks.  Tasks that are in the process of
2935	<	* being submitted or executed concurrently during the course of
2936	<	* this method may or may not be rejected. This method cancels
2937	<	* both existing and unexecuted tasks, in order to permit
2938	<	* termination in the presence of task dependencies. So the method
2939	<	* always returns an empty list (unlike the case for some other
2940	<	* Executors).
2933	>	* Possibly attempts to cancel and/or stop all tasks, and reject
2934	>	* all subsequently submitted tasks.  Invocation has no effect on
2935	>	* execution state if this is the {@link #commonPool}, and no
2936	>	* additional effect if already shut down. Otherwise, tasks that
2937	>	* are in the process of being submitted or executed concurrently
2938	>	* during the course of this method may or may not be
2939	>	* rejected. This method cancels both existing and unexecuted
2940	>	* tasks, in order to permit termination in the presence of task
2941	>	* dependencies. So the method always returns an empty list
2942	>	* (unlike the case for some other Executors).
2943		*
2944		* @return an empty list
2945		* @throws SecurityException if a security manager exists and
#	Line 2399 | Line 2949 | public class ForkJoinPool extends Abstra
2949		*/
2950		public List<Runnable> shutdownNow() {
2951		checkPermission();
2952	<	enableShutdown();
2403	<	tryTerminate(true);
2952	>	tryTerminate(true, true);
2953		return Collections.emptyList();
2954		}
2955
#	Line 2440 | Line 2989 | public class ForkJoinPool extends Abstra
2989		* @return {@code true} if this pool has been shut down
2990		*/
2991		public boolean isShutdown() {
2992	<	return runState < 0;
2992	>	return plock < 0;
2993		}
2994
2995		/**
2996	<	* Blocks until all tasks have completed execution after a shutdown
2997	<	* request, or the timeout occurs, or the current thread is
2998	<	* interrupted, whichever happens first.
2996	>	* Blocks until all tasks have completed execution after a
2997	>	* shutdown request, or the timeout occurs, or the current thread
2998	>	* is interrupted, whichever happens first. Note that the {@link
2999	>	* #commonPool()} never terminates until program shutdown so
3000	>	* this method will always time out.
3001		*
3002		* @param timeout the maximum time to wait
3003		* @param unit the time unit of the timeout argument
#	Line 2457 | Line 3008 | public class ForkJoinPool extends Abstra
3008		public boolean awaitTermination(long timeout, TimeUnit unit)
3009		throws InterruptedException {
3010		long nanos = unit.toNanos(timeout);
3011	<	final ReentrantLock lock = this.lock;
3012	<	lock.lock();
3013	<	try {
3014	<	for (;;) {
3015	<	if (isTerminated())
3016	<	return true;
3017	<	if (nanos <= 0)
3018	<	return false;
3019	<	nanos = termination.awaitNanos(nanos);
3011	>	if (isTerminated())
3012	>	return true;
3013	>	long startTime = System.nanoTime();
3014	>	boolean terminated = false;
3015	>	synchronized (this) {
3016	>	for (long waitTime = nanos, millis = 0L;;) {
3017	>	if (terminated = isTerminated() ||
3018	>	waitTime <= 0L ||
3019	>	(millis = unit.toMillis(waitTime)) <= 0L)
3020	>	break;
3021	>	wait(millis);
3022	>	waitTime = nanos - (System.nanoTime() - startTime);
3023		}
2470	–	} finally {
2471	–	lock.unlock();
3024		}
3025	+	return terminated;
3026		}
3027
3028		/**
#	Line 2568 | Line 3121 | public class ForkJoinPool extends Abstra
3121		public static void managedBlock(ManagedBlocker blocker)
3122		throws InterruptedException {
3123		Thread t = Thread.currentThread();
3124	<	ForkJoinPool p = ((t instanceof ForkJoinWorkerThread) ?
3125	<	((ForkJoinWorkerThread)t).pool : null);
3126	<	while (!blocker.isReleasable()) {
3127	<	if (p == null || p.tryCompensate()) {
3128	<	try {
3129	<	do {} while (!blocker.isReleasable() && !blocker.block());
3130	<	} finally {
3131	<	if (p != null)
3124	>	if (t instanceof ForkJoinWorkerThread) {
3125	>	ForkJoinPool p = ((ForkJoinWorkerThread)t).pool;
3126	>	while (!blocker.isReleasable()) { // variant of helpSignal
3127	>	WorkQueue[] ws; WorkQueue q; int m, n;
3128	>	if ((ws = p.workQueues) != null && (m = ws.length - 1) >= 0) {
3129	>	for (int i = 0; i <= m; ++i) {
3130	>	if (blocker.isReleasable())
3131	>	return;
3132	>	if ((q = ws[i]) != null && (n = q.queueSize()) > 0) {
3133	>	p.signalWork(q, n);
3134	>	if ((int)(p.ctl >> AC_SHIFT) >= 0)
3135	>	break;
3136	>	}
3137	>	}
3138	>	}
3139	>	if (p.tryCompensate()) {
3140	>	try {
3141	>	do {} while (!blocker.isReleasable() &&
3142	>	!blocker.block());
3143	>	} finally {
3144		p.incrementActiveCount();
3145	+	}
3146	+	break;
3147		}
2581	–	break;
3148		}
3149		}
3150	+	else {
3151	+	do {} while (!blocker.isReleasable() &&
3152	+	!blocker.block());
3153	+	}
3154		}
3155
3156		// AbstractExecutorService overrides.  These rely on undocumented
#	Line 2588 | Line 3158 | public class ForkJoinPool extends Abstra
3158		// implement RunnableFuture.
3159
3160		protected <T> RunnableFuture<T> newTaskFor(Runnable runnable, T value) {
3161	<	return (RunnableFuture<T>) ForkJoinTask.adapt(runnable, value);
3161	>	return new ForkJoinTask.AdaptedRunnable<T>(runnable, value);
3162		}
3163
3164		protected <T> RunnableFuture<T> newTaskFor(Callable<T> callable) {
3165	<	return (RunnableFuture<T>) ForkJoinTask.adapt(callable);
3165	>	return new ForkJoinTask.AdaptedCallable<T>(callable);
3166		}
3167
3168		// Unsafe mechanics
3169		private static final sun.misc.Unsafe U;
3170		private static final long CTL;
2601	–	private static final long RUNSTATE;
3171		private static final long PARKBLOCKER;
3172	+	private static final int ABASE;
3173	+	private static final int ASHIFT;
3174	+	private static final long STEALCOUNT;
3175	+	private static final long PLOCK;
3176	+	private static final long INDEXSEED;
3177	+	private static final long QLOCK;
3178
3179		static {
3180	<	poolNumberGenerator = new AtomicInteger();
3181	<	modifyThreadPermission = new RuntimePermission("modifyThread");
3182	<	defaultForkJoinWorkerThreadFactory =
3183	<	new DefaultForkJoinWorkerThreadFactory();
3184	<	int s;
3180	>	// Establish common pool parameters
3181	>	// TBD: limit or report ignored exceptions?
3182	>
3183	>	int par = 0;
3184	>	ForkJoinWorkerThreadFactory fac = null;
3185	>	Thread.UncaughtExceptionHandler handler = null;
3186	>	try {
3187	>	String pp = System.getProperty(propPrefix + "parallelism");
3188	>	String hp = System.getProperty(propPrefix + "exceptionHandler");
3189	>	String fp = System.getProperty(propPrefix + "threadFactory");
3190	>	if (fp != null)
3191	>	fac = ((ForkJoinWorkerThreadFactory)ClassLoader.
3192	>	getSystemClassLoader().loadClass(fp).newInstance());
3193	>	if (hp != null)
3194	>	handler = ((Thread.UncaughtExceptionHandler)ClassLoader.
3195	>	getSystemClassLoader().loadClass(hp).newInstance());
3196	>	if (pp != null)
3197	>	par = Integer.parseInt(pp);
3198	>	} catch (Exception ignore) {
3199	>	}
3200	>
3201	>	int s; // initialize field offsets for CAS etc
3202		try {
3203		U = getUnsafe();
3204		Class<?> k = ForkJoinPool.class;
2613	–	Class<?> tk = Thread.class;
3205		CTL = U.objectFieldOffset
3206		(k.getDeclaredField("ctl"));
3207	<	RUNSTATE = U.objectFieldOffset
3208	<	(k.getDeclaredField("runState"));
3207	>	STEALCOUNT = U.objectFieldOffset
3208	>	(k.getDeclaredField("stealCount"));
3209	>	PLOCK = U.objectFieldOffset
3210	>	(k.getDeclaredField("plock"));
3211	>	INDEXSEED = U.objectFieldOffset
3212	>	(k.getDeclaredField("indexSeed"));
3213	>	Class<?> tk = Thread.class;
3214		PARKBLOCKER = U.objectFieldOffset
3215		(tk.getDeclaredField("parkBlocker"));
3216	+	Class<?> wk = WorkQueue.class;
3217	+	QLOCK = U.objectFieldOffset
3218	+	(wk.getDeclaredField("qlock"));
3219	+	Class<?> ak = ForkJoinTask[].class;
3220	+	ABASE = U.arrayBaseOffset(ak);
3221	+	s = U.arrayIndexScale(ak);
3222	+	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
3223		} catch (Exception e) {
3224		throw new Error(e);
3225		}
3226	+	if ((s & (s-1)) != 0)
3227	+	throw new Error("data type scale not a power of two");
3228	+
3229	+	/*
3230	+	* For extra caution, computations to set up pool state are
3231	+	* here; the constructor just assigns these values to fields.
3232	+	*/
3233	+	ForkJoinWorkerThreadFactory defaultFac =
3234	+	defaultForkJoinWorkerThreadFactory =
3235	+	new DefaultForkJoinWorkerThreadFactory();
3236	+	if (fac == null)
3237	+	fac = defaultFac;
3238	+	if (par <= 0)
3239	+	par = Runtime.getRuntime().availableProcessors();
3240	+	if (par > MAX_CAP)
3241	+	par = MAX_CAP;
3242	+	long np = (long)(-par); // precompute initial ctl value
3243	+	long ct = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK);
3244	+
3245	+	commonPoolParallelism = par;
3246	+	commonPool = new ForkJoinPool(par, ct, fac, handler);
3247	+	modifyThreadPermission = new RuntimePermission("modifyThread");
3248	+	submitters = new ThreadLocal<Submitter>();
3249		}
3250
3251		/**

Diff Legend

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