[ViewVC] Diff of: jsr166/jsr166/src/jsr166y/ForkJoinPool.java

Comparing jsr166/src/jsr166y/ForkJoinPool.java (file contents):
Revision 1.125 by jsr166, Tue Feb 21 00:19:23 2012 UTC vs.
Revision 1.146 by dl, Mon Nov 19 18:12:42 2012 UTC

#	Line 5 \| Line 5
5		*/
6
7		package jsr166y;
8	+
9		import java.util.ArrayList;
10		import java.util.Arrays;
11		import java.util.Collection;
12		import java.util.Collections;
13		import java.util.List;
13	–	import java.util.Random;
14		import java.util.concurrent.AbstractExecutorService;
15		import java.util.concurrent.Callable;
16		import java.util.concurrent.ExecutorService;
#	Line 18 \| Line 18 \| import java.util.concurrent.Future;
18		import java.util.concurrent.RejectedExecutionException;
19		import java.util.concurrent.RunnableFuture;
20		import java.util.concurrent.TimeUnit;
21	–	import java.util.concurrent.atomic.AtomicInteger;
22	–	import java.util.concurrent.atomic.AtomicLong;
23	–	import java.util.concurrent.locks.AbstractQueuedSynchronizer;
24	–	import java.util.concurrent.locks.Condition;
21
22		/**
23		* An {@link ExecutorService} for running {@link ForkJoinTask}s.
#	Line 41 \| 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 58 \| Line 62 \| import java.util.concurrent.locks.Condit
62		* {@link #toString} returns indications of pool state in a
63		* convenient form for informal monitoring.
64		*
65	<	* <p> As is the case with other ExecutorServices, there are three
65	>	* <p>As is the case with other ExecutorServices, there are three
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
#	Line 93 \| 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.
109	<	*
110	<	* <pre> {@code
111	<	* static final ForkJoinPool mainPool = new ForkJoinPool();
112	<	* ...
113	<	* public void sort(long[] array) {
110	<	* mainPool.invoke(new SortTask(array, 0, array.length));
111	<	* }}</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	>	<<<<<<< ForkJoinPool.java
107	>	* exceptionHandler} -- the class name of a {@code
108	>	=======
109	>	* exceptionHandler} -- the class name of a {@link
110	>	* java.lang.Thread.UncaughtExceptionHandler
111	>	>>>>>>> 1.111
112	>	* Thread.UncaughtExceptionHandler}. Upon any error in establishing
113	>	* these settings, default parameters are used.
114		*
115		* <p><b>Implementation notes</b>: This implementation restricts the
116		* maximum number of running threads to 32767. Attempts to create
#	Line 196 \| Line 198 \| public class ForkJoinPool extends Abstra
198		* WorkQueues are also used in a similar way for tasks submitted
199		* to the pool. We cannot mix these tasks in the same queues used
200		* for work-stealing (this would contaminate lifo/fifo
201	<	* processing). Instead, we loosely associate submission queues
201	>	* processing). Instead, we randomly associate submission queues
202		* with submitting threads, using a form of hashing. The
203		* ThreadLocal Submitter class contains a value initially used as
204		* a hash code for choosing existing queues, but may be randomly
205		* repositioned upon contention with other submitters. In
206	<	* essence, submitters act like workers except that they never
207	<	* take tasks, and they are multiplexed on to a finite number of
208	<	* shared work queues. However, classes are set up so that future
209	<	* extensions could allow submitters to optionally help perform
210	<	* tasks as well. Insertion of tasks in shared mode requires a
211	<	* lock (mainly to protect in the case of resizing) but we use
212	<	* only a simple spinlock (using bits in field runState), because
213	<	* submitters encountering a busy queue move on to try or create
214	<	* other queues -- they block only when creating and registering
215	<	* new queues.
206	>	* essence, submitters act like workers except that they are
207	>	* restricted to executing local tasks that they submitted (or in
208	>	* the case of CountedCompleters, others with the same root task).
209	>	* However, because most shared/external queue operations are more
210	>	* expensive than internal, and because, at steady state, external
211	>	* submitters will compete for CPU with workers, ForkJoinTask.join
212	>	* and related methods disable them from repeatedly helping to
213	>	* process tasks if all workers are active. Insertion of tasks in
214	>	* shared mode requires a lock (mainly to protect in the case of
215	>	* resizing) but we use only a simple spinlock (using bits in
216	>	* field qlock), because submitters encountering a busy queue move
217	>	* on to try or create other queues -- they block only when
218	>	* creating and registering new queues.
219		*
220		* Management
221		* ==========
#	Line 232 \| Line 237 \| public class ForkJoinPool extends Abstra
237		* and their negations (used for thresholding) to fit into 16bit
238		* fields.
239		*
240	<	* Field "runState" contains 32 bits needed to register and
241	<	* deregister WorkQueues, as well as to enable shutdown. It is
242	<	* only modified under a lock (normally briefly held, but
243	<	* occasionally protecting allocations and resizings) but even
244	<	* when locked remains available to check consistency.
240	>	* Field "plock" is a form of sequence lock with a saturating
241	>	* shutdown bit (similarly for per-queue "qlocks"), mainly
242	>	* protecting updates to the workQueues array, as well as to
243	>	* enable shutdown. When used as a lock, it is normally only very
244	>	* briefly held, so is nearly always available after at most a
245	>	* brief spin, but we use a monitor-based backup strategy to
246	>	* block when needed.
247		*
248		* Recording WorkQueues. WorkQueues are recorded in the
249	<	* "workQueues" array that is created upon pool construction and
250	<	* expanded if necessary. Updates to the array while recording
251	<	* new workers and unrecording terminated ones are protected from
252	<	* each other by a lock but the array is otherwise concurrently
253	<	* readable, and accessed directly. To simplify index-based
254	<	* operations, the array size is always a power of two, and all
255	<	* readers must tolerate null slots. Shared (submission) queues
256	<	* are at even indices, worker queues at odd indices. Grouping
257	<	* them together in this way simplifies and speeds up task
258	<	* scanning.
249	>	* "workQueues" array that is created upon first use and expanded
250	>	* if necessary. Updates to the array while recording new workers
251	>	* and unrecording terminated ones are protected from each other
252	>	* by a lock but the array is otherwise concurrently readable, and
253	>	* accessed directly. To simplify index-based operations, the
254	>	* array size is always a power of two, and all readers must
255	>	* tolerate null slots. Worker queues are at odd indices. Shared
256	>	* (submission) queues are at even indices, up to a maximum of 64
257	>	* slots, to limit growth even if array needs to expand to add
258	>	* more workers. Grouping them together in this way simplifies and
259	>	* speeds up task scanning.
260		*
261		* All worker thread creation is on-demand, triggered by task
262		* submissions, replacement of terminated workers, and/or
#	Line 309 \| Line 317 \| public class ForkJoinPool extends Abstra
317		*
318		* Signalling. We create or wake up workers only when there
319		* appears to be at least one task they might be able to find and
320	<	* execute. When a submission is added or another worker adds a
321	<	* task to a queue that previously had fewer than two tasks, they
322	<	* signal waiting workers (or trigger creation of new ones if
323	<	* fewer than the given parallelism level -- see signalWork).
324	<	* These primary signals are buttressed by signals during rescans;
325	<	* together these cover the signals needed in cases when more
326	<	* tasks are pushed but untaken, and improve performance compared
327	<	* to having one thread wake up all workers.
320	>	* execute. However, many other threads may notice the same task
321	>	* and each signal to wake up a thread that might take it. So in
322	>	* general, pools will be over-signalled. When a submission is
323	>	* added or another worker adds a task to a queue that is
324	>	* apparently empty, they signal waiting workers (or trigger
325	>	* creation of new ones if fewer than the given parallelism
326	>	* level). These primary signals are buttressed by signals
327	>	* whenever other threads scan for work or do not have a task to
328	>	* process (including the case of leaving a hint to unparked
329	>	* threads to help signal others upon wakeup). On most platforms,
330	>	* signalling (unpark) overhead time is noticeably long, and the
331	>	* time between signalling a thread and it actually making
332	>	* progress can be very noticeably long, so it is worth offloading
333	>	* these delays from critical paths as much as possible.
334		*
335		* Trimming workers. To release resources after periods of lack of
336		* use, a worker starting to wait when the pool is quiescent will
337	<	* time out and terminate if the pool has remained quiescent for
338	<	* SHRINK_RATE nanosecs. This will slowly propagate, eventually
339	<	* terminating all workers after long periods of non-use.
337	>	* time out and terminate if the pool has remained quiescent for a
338	>	* given period -- a short period if there are more threads than
339	>	* parallelism, longer as the number of threads decreases. This
340	>	* will slowly propagate, eventually terminating all workers after
341	>	* periods of non-use.
342		*
343		* Shutdown and Termination. A call to shutdownNow atomically sets
344	<	* a runState bit and then (non-atomically) sets each worker's
345	<	* runState status, cancels all unprocessed tasks, and wakes up
344	>	* a plock bit and then (non-atomically) sets each worker's
345	>	* qlock status, cancels all unprocessed tasks, and wakes up
346		* all waiting workers. Detecting whether termination should
347		* commence after a non-abrupt shutdown() call requires more work
348		* and bookkeeping. We need consensus about quiescence (i.e., that
#	Line 354 \| Line 370 \| public class ForkJoinPool extends Abstra
370		* method tryCompensate() may create or re-activate a spare
371		* thread to compensate for blocked joiners until they unblock.
372		*
373	<	* A third form (implemented in tryRemoveAndExec and
374	<	* tryPollForAndExec) amounts to helping a hypothetical
375	<	* compensator: If we can readily tell that a possible action of a
376	<	* compensator is to steal and execute the task being joined, the
377	<	* joining thread can do so directly, without the need for a
378	<	* compensation thread (although at the expense of larger run-time
379	<	* stacks, but the tradeoff is typically worthwhile).
373	>	* A third form (implemented in tryRemoveAndExec) amounts to
374	>	* helping a hypothetical compensator: If we can readily tell that
375	>	* a possible action of a compensator is to steal and execute the
376	>	* task being joined, the joining thread can do so directly,
377	>	* without the need for a compensation thread (although at the
378	>	* expense of larger run-time stacks, but the tradeoff is
379	>	* typically worthwhile).
380		*
381		* The ManagedBlocker extension API can't use helping so relies
382		* only on compensation in method awaitBlocker.
#	Line 382 \| Line 398 \| public class ForkJoinPool extends Abstra
398		* steals, rather than use per-task bookkeeping. This sometimes
399		* requires a linear scan of workQueues array to locate stealers,
400		* but often doesn't because stealers leave hints (that may become
401	<	* stale/wrong) of where to locate them. A stealHint is only a
402	<	* hint because a worker might have had multiple steals and the
403	<	* hint records only one of them (usually the most current).
404	<	* Hinting isolates cost to when it is needed, rather than adding
405	<	* to per-task overhead. (2) It is "shallow", ignoring nesting
406	<	* and potentially cyclic mutual steals. (3) It is intentionally
401	>	* stale/wrong) of where to locate them. It is only a hint
402	>	* because a worker might have had multiple steals and the hint
403	>	* records only one of them (usually the most current). Hinting
404	>	* isolates cost to when it is needed, rather than adding to
405	>	* per-task overhead. (2) It is "shallow", ignoring nesting and
406	>	* potentially cyclic mutual steals. (3) It is intentionally
407		* racy: field currentJoin is updated only while actively joining,
408		* which means that we miss links in the chain during long-lived
409		* tasks, GC stalls etc (which is OK since blocking in such cases
#	Line 395 \| Line 411 \| public class ForkJoinPool extends Abstra
411		* to find work (see MAX_HELP) and fall back to suspending the
412		* worker and if necessary replacing it with another.
413		*
414	+	* Helping actions for CountedCompleters are much simpler: Method
415	+	* helpComplete can take and execute any task with the same root
416	+	* as the task being waited on. However, this still entails some
417	+	* traversal of completer chains, so is less efficient than using
418	+	* CountedCompleters without explicit joins.
419	+	*
420		* It is impossible to keep exactly the target parallelism number
421		* of threads running at any given time. Determining the
422		* existence of conservatively safe helping targets, the
#	Line 416 \| Line 438 \| public class ForkJoinPool extends Abstra
438		* intractable) game with an opponent that may choose the worst
439		* (for us) active thread to stall at any time. We take several
440		* precautions to bound losses (and thus bound gains), mainly in
441	<	* methods tryCompensate and awaitJoin: (1) We only try
442	<	* compensation after attempting enough helping steps (measured
443	<	* via counting and timing) that we have already consumed the
444	<	* estimated cost of creating and activating a new thread. (2) We
445	<	* allow up to 50% of threads to be blocked before initially
446	<	* adding any others, and unless completely saturated, check that
447	<	* some work is available for a new worker before adding. Also, we
448	<	* create up to only 50% more threads until entering a mode that
449	<	* only adds a thread if all others are possibly blocked. All
450	<	* together, this means that we might be half as fast to react,
451	<	* and create half as many threads as possible in the ideal case,
452	<	* but present vastly fewer anomalies in all other cases compared
453	<	* to both more aggressive and more conservative alternatives.
454	<	*
455	<	* Style notes: There is a lot of representation-level coupling
456	<	* among classes ForkJoinPool, ForkJoinWorkerThread, and
457	<	* ForkJoinTask. The fields of WorkQueue maintain data structures
458	<	* managed by ForkJoinPool, so are directly accessed. There is
459	<	* little point trying to reduce this, since any associated future
460	<	* changes in representations will need to be accompanied by
461	<	* algorithmic changes anyway. Several methods intrinsically
462	<	* sprawl because they must accumulate sets of consistent reads of
463	<	* volatiles held in local variables. Methods signalWork() and
464	<	* scan() are the main bottlenecks, so are especially heavily
441	>	* methods tryCompensate and awaitJoin.
442	>	*
443	>	* Common Pool
444	>	* ===========
445	>	*
446	>	* The static commonPool always exists after static
447	>	* initialization. Since it (or any other created pool) need
448	>	* never be used, we minimize initial construction overhead and
449	>	* footprint to the setup of about a dozen fields, with no nested
450	>	* allocation. Most bootstrapping occurs within method
451	>	* fullExternalPush during the first submission to the pool.
452	>	*
453	>	* When external threads submit to the common pool, they can
454	>	* perform some subtask processing (see externalHelpJoin and
455	>	* related methods). We do not need to record whether these
456	>	* submissions are to the common pool -- if not, externalHelpJoin
457	>	* returns quickly (at the most helping to signal some common pool
458	>	* workers). These submitters would otherwise be blocked waiting
459	>	* for completion, so the extra effort (with liberally sprinkled
460	>	* task status checks) in inapplicable cases amounts to an odd
461	>	* form of limited spin-wait before blocking in ForkJoinTask.join.
462	>	*
463	>	* Style notes
464	>	* ===========
465	>	*
466	>	* There is a lot of representation-level coupling among classes
467	>	* ForkJoinPool, ForkJoinWorkerThread, and ForkJoinTask. The
468	>	* fields of WorkQueue maintain data structures managed by
469	>	* ForkJoinPool, so are directly accessed. There is little point
470	>	* trying to reduce this, since any associated future changes in
471	>	* representations will need to be accompanied by algorithmic
472	>	* changes anyway. Several methods intrinsically sprawl because
473	>	* they must accumulate sets of consistent reads of volatiles held
474	>	* in local variables. Methods signalWork() and scan() are the
475	>	* main bottlenecks, so are especially heavily
476		* micro-optimized/mangled. There are lots of inline assignments
477		* (of form "while ((local = field) != 0)") which are usually the
478		* simplest way to ensure the required read orderings (which are
#	Line 447 \| Line 480 \| public class ForkJoinPool extends Abstra
480		* declarations of these locals at the heads of methods or blocks.
481		* There are several occurrences of the unusual "do {} while
482		* (!cas...)" which is the simplest way to force an update of a
483	<	* CAS'ed variable. There are also other coding oddities that help
483	>	* CAS'ed variable. There are also other coding oddities (including
484	>	* several unnecessary-looking hoisted null checks) that help
485		* some methods perform reasonably even when interpreted (not
486		* compiled).
487		*
#	Line 496 \| Line 530 \| public class ForkJoinPool extends Abstra
530		* Default ForkJoinWorkerThreadFactory implementation; creates a
531		* new ForkJoinWorkerThread.
532		*/
533	<	static class DefaultForkJoinWorkerThreadFactory
533	>	static final class DefaultForkJoinWorkerThreadFactory
534		implements ForkJoinWorkerThreadFactory {
535	<	public ForkJoinWorkerThread newThread(ForkJoinPool pool) {
535	>	public final ForkJoinWorkerThread newThread(ForkJoinPool pool) {
536		return new ForkJoinWorkerThread(pool);
537		}
538		}
539
540		/**
507	–	* A simple non-reentrant lock used for exclusion when managing
508	–	* queues and workers. We use a custom lock so that we can readily
509	–	* probe lock state in constructions that check among alternative
510	–	* actions. The lock is normally only very briefly held, and
511	–	* sometimes treated as a spinlock, but other usages block to
512	–	* reduce overall contention in those cases where locked code
513	–	* bodies perform allocation/resizing.
514	–	*/
515	–	static final class Mutex extends AbstractQueuedSynchronizer {
516	–	public final boolean tryAcquire(int ignore) {
517	–	return compareAndSetState(0, 1);
518	–	}
519	–	public final boolean tryRelease(int ignore) {
520	–	setState(0);
521	–	return true;
522	–	}
523	–	public final void lock() { acquire(0); }
524	–	public final void unlock() { release(0); }
525	–	public final boolean isHeldExclusively() { return getState() == 1; }
526	–	public final Condition newCondition() { return new ConditionObject(); }
527	–	}
528	–
529	–	/**
541		* Class for artificial tasks that are used to replace the target
542		* of local joins if they are removed from an interior queue slot
543		* in WorkQueue.tryRemoveAndExec. We don't need the proxy to
544		* actually do anything beyond having a unique identity.
545		*/
546		static final class EmptyTask extends ForkJoinTask<Void> {
547	+	private static final long serialVersionUID = -7721805057305804111L;
548		EmptyTask() { status = ForkJoinTask.NORMAL; } // force done
549		public final Void getRawResult() { return null; }
550		public final void setRawResult(Void x) {}
#	Line 553 \| Line 565 \| public class ForkJoinPool extends Abstra
565		*
566		* Field "top" is the index (mod array.length) of the next queue
567		* slot to push to or pop from. It is written only by owner thread
568	<	* for push, or under lock for trySharedPush, and accessed by
569	<	* other threads only after reading (volatile) base. Both top and
570	<	* base are allowed to wrap around on overflow, but (top - base)
571	<	* (or more commonly -(base - top) to force volatile read of base
572	<	* before top) still estimates size.
568	>	* for push, or under lock for external/shared push, and accessed
569	>	* by other threads only after reading (volatile) base. Both top
570	>	* and base are allowed to wrap around on overflow, but (top -
571	>	* base) (or more commonly -(base - top) to force volatile read of
572	>	* base before top) still estimates size. The lock ("qlock") is
573	>	* forced to -1 on termination, causing all further lock attempts
574	>	* to fail. (Note: we don't need CAS for termination state because
575	>	* upon pool shutdown, all shared-queues will stop being used
576	>	* anyway.) Nearly all lock bodies are set up so that exceptions
577	>	* within lock bodies are "impossible" (modulo JVM errors that
578	>	* would cause failure anyway.)
579		*
580		* The array slots are read and written using the emulation of
581		* volatiles/atomics provided by Unsafe. Insertions must in
582		* general use putOrderedObject as a form of releasing store to
583		* ensure that all writes to the task object are ordered before
584	<	* its publication in the queue. (Although we can avoid one case
585	<	* of this when locked in trySharedPush.) All removals entail a
586	<	* CAS to null. The array is always a power of two. To ensure
587	<	* safety of Unsafe array operations, all accesses perform
570	<	* explicit null checks and implicit bounds checks via
571	<	* power-of-two masking.
584	>	* its publication in the queue. All removals entail a CAS to
585	>	* null. The array is always a power of two. To ensure safety of
586	>	* Unsafe array operations, all accesses perform explicit null
587	>	* checks and implicit bounds checks via power-of-two masking.
588		*
589		* In addition to basic queuing support, this class contains
590		* fields described elsewhere to control execution. It turns out
591	<	* to work better memory-layout-wise to include them in this
592	<	* class rather than a separate class.
591	>	* to work better memory-layout-wise to include them in this class
592	>	* rather than a separate class.
593		*
594		* Performance on most platforms is very sensitive to placement of
595		* instances of both WorkQueues and their arrays -- we absolutely
#	Line 587 \| Line 603 \| public class ForkJoinPool extends Abstra
603		* trades off slightly slower average field access for the sake of
604		* avoiding really bad worst-case access. (Until better JVM
605		* support is in place, this padding is dependent on transient
606	<	* properties of JVM field layout rules.) We also take care in
591	<	* allocating, sizing and resizing the array. Non-shared queue
592	<	* arrays are initialized (via method growArray) by workers before
593	<	* use. Others are allocated on first use.
606	>	* properties of JVM field layout rules.)
607		*/
608		static final class WorkQueue {
609		/**
#	Line 613 \| Line 626 \| public class ForkJoinPool extends Abstra
626		*/
627		static final int MAXIMUM_QUEUE_CAPACITY = 1 << 26; // 64M
628
616	–	volatile long totalSteals; // cumulative number of steals
629		int seed; // for random scanning; initialize nonzero
630		volatile int eventCount; // encoded inactivation count; < 0 if inactive
631		int nextWait; // encoded record of next event waiter
632	<	int rescans; // remaining scans until block
621	<	int nsteals; // top-level task executions since last idle
622	<	final int mode; // lifo, fifo, or shared
632	>	int hint; // steal or signal hint (index)
633		int poolIndex; // index of this queue in pool (or 0)
634	<	int stealHint; // index of most recent known stealer
635	<	volatile int runState; // 1: locked, -1: terminate; else 0
634	>	final int mode; // 0: lifo, > 0: fifo, < 0: shared
635	>	int nsteals; // number of steals
636	>	volatile int qlock; // 1: locked, -1: terminate; else 0
637		volatile int base; // index of next slot for poll
638		int top; // index of next slot for push
639		ForkJoinTask<?>[] array; // the elements (initially unallocated)
640		final ForkJoinPool pool; // the containing pool (may be null)
641		final ForkJoinWorkerThread owner; // owning thread or null if shared
642		volatile Thread parker; // == owner during call to park; else null
643	<	ForkJoinTask<?> currentJoin; // task being joined in awaitJoin
643	>	volatile ForkJoinTask<?> currentJoin; // task being joined in awaitJoin
644		ForkJoinTask<?> currentSteal; // current non-local task being executed
645	+
646		// Heuristic padding to ameliorate unfortunate memory placements
647		Object p00, p01, p02, p03, p04, p05, p06, p07;
648	<	Object p08, p09, p0a, p0b, p0c, p0d, p0e;
648	>	Object p08, p09, p0a, p0b, p0c;
649
650	<	WorkQueue(ForkJoinPool pool, ForkJoinWorkerThread owner, int mode) {
651	<	this.mode = mode;
650	>	WorkQueue(ForkJoinPool pool, ForkJoinWorkerThread owner, int mode,
651	>	int seed) {
652	>	this.array = new ForkJoinTask<?>[WorkQueue.INITIAL_QUEUE_CAPACITY];
653		this.pool = pool;
654		this.owner = owner;
655	<	// Place indices in the center of array (that is not yet allocated)
655	>	this.mode = mode;
656	>	this.seed = seed;
657	>	// Place indices in the center of array
658		base = top = INITIAL_QUEUE_CAPACITY >>> 1;
659		}
660
661		/**
647	–	* Returns the approximate number of tasks in the queue.
648	–	*/
649	–	final int queueSize() {
650	–	int n = base - top; // non-owner callers must read base first
651	–	return (n >= 0) ? 0 : -n; // ignore transient negative
652	–	}
653	–
654	–	/**
655	–	* Provides a more accurate estimate of whether this queue has
656	–	* any tasks than does queueSize, by checking whether a
657	–	* near-empty queue has at least one unclaimed task.
658	–	*/
659	–	final boolean isEmpty() {
660	–	ForkJoinTask<?>[] a; int m, s;
661	–	int n = base - (s = top);
662	–	return (n >= 0 \|\|
663	–	(n == -1 &&
664	–	((a = array) == null \|\|
665	–	(m = a.length - 1) < 0 \|\|
666	–	U.getObjectVolatile
667	–	(a, ((m & (s - 1)) << ASHIFT) + ABASE) == null)));
668	–	}
669	–
670	–	/**
662		* Pushes a task. Call only by owner in unshared queues.
663	+	* Cases needing resizing or rejection are relayed to fullPush
664	+	* (that also handles shared queues).
665		*
666		* @param task the task. Caller must ensure non-null.
667		* @throw RejectedExecutionException if array cannot be resized
#	Line 679 \| Line 672 \| public class ForkJoinPool extends Abstra
672		if ((a = array) != null) { // ignore if queue removed
673		U.putOrderedObject
674		(a, (((m = a.length - 1) & s) << ASHIFT) + ABASE, task);
675	<	if ((n = (top = s + 1) - base) <= 2) {
675	>	if ((n = (top = s + 1) - base) <= 1) {
676		if ((p = pool) != null)
677	<	p.signalWork();
677	>	p.signalWork(this, 0);
678		}
679		else if (n >= m)
680	<	growArray(true);
680	>	growArray();
681		}
682		}
683
#	Line 697 \| Line 690 \| public class ForkJoinPool extends Abstra
690		*/
691		final boolean trySharedPush(ForkJoinTask<?> task) {
692		boolean submitted = false;
693	<	if (runState == 0 && U.compareAndSwapInt(this, RUNSTATE, 0, 1)) {
694	<	ForkJoinTask<?>[] a = array;
693	>	if (qlock == 0 && U.compareAndSwapInt(this, QLOCK, 0, 1)) {
694	>	ForkJoinTask<?>[] a = array; ForkJoinPool p;
695		int s = top;
696		try {
697		if ((a != null && a.length > s + 1 - base) \|\|
698	<	(a = growArray(false)) != null) { // must presize
698	>	(a = growArray()) != null) { // must presize
699		int j = (((a.length - 1) & s) << ASHIFT) + ABASE;
700	<	U.putObject(a, (long)j, task); // don't need "ordered"
700	>	U.putOrderedObject(a, j, task);
701		top = s + 1;
702		submitted = true;
703		}
704		} finally {
705	<	runState = 0; // unlock
705	>	qlock = 0; // unlock
706		}
707	+	if (submitted && (p = pool) != null)
708	+	p.signalWork(this, 0);
709		}
710		return submitted;
711		}
712
713	+	/**
714	+	* Initializes or doubles the capacity of array. Call either
715	+	* by owner or with lock held -- it is OK for base, but not
716	+	* top, to move while resizings are in progress.
717	+	*/
718	+	final ForkJoinTask<?>[] growArray() {
719	+	ForkJoinTask<?>[] oldA = array;
720	+	int size = oldA != null ? oldA.length << 1 : INITIAL_QUEUE_CAPACITY;
721	+	if (size > MAXIMUM_QUEUE_CAPACITY)
722	+	throw new RejectedExecutionException("Queue capacity exceeded");
723	+	int oldMask, t, b;
724	+	ForkJoinTask<?>[] a = array = new ForkJoinTask<?>[size];
725	+	if (oldA != null && (oldMask = oldA.length - 1) >= 0 &&
726	+	(t = top) - (b = base) > 0) {
727	+	int mask = size - 1;
728	+	do {
729	+	ForkJoinTask<?> x;
730	+	int oldj = ((b & oldMask) << ASHIFT) + ABASE;
731	+	int j = ((b & mask) << ASHIFT) + ABASE;
732	+	x = (ForkJoinTask<?>)U.getObjectVolatile(oldA, oldj);
733	+	if (x != null &&
734	+	U.compareAndSwapObject(oldA, oldj, x, null))
735	+	U.putObjectVolatile(a, j, x);
736	+	} while (++b != t);
737	+	}
738	+	return a;
739	+	}
740	+
741		/**
742		* Takes next task, if one exists, in LIFO order. Call only
743	<	* by owner in unshared queues. (We do not have a shared
721	<	* version of this method because it is never needed.)
743	>	* by owner in unshared queues.
744		*/
745		final ForkJoinTask<?> pop() {
746	<	ForkJoinTask<?> t; int m;
747	<	ForkJoinTask<?>[] a = array;
726	<	if (a != null && (m = a.length - 1) >= 0) {
746	>	ForkJoinTask<?>[] a; ForkJoinTask<?> t; int m;
747	>	if ((a = array) != null && (m = a.length - 1) >= 0) {
748		for (int s; (s = top - 1) - base >= 0;) {
749	<	int j = ((m & s) << ASHIFT) + ABASE;
750	<	if ((t = (ForkJoinTask<?>)U.getObjectVolatile(a, j)) == null)
749	>	long j = ((m & s) << ASHIFT) + ABASE;
750	>	if ((t = (ForkJoinTask<?>)U.getObject(a, j)) == null)
751		break;
752		if (U.compareAndSwapObject(a, j, t, null)) {
753		top = s;
#	Line 774 \| Line 795 \| public class ForkJoinPool extends Abstra
795		else if (base == b) {
796		if (b + 1 == top)
797		break;
798	<	Thread.yield(); // wait for lagging update
798	>	Thread.yield(); // wait for lagging update (very rare)
799		}
800		}
801		return null;
#	Line 801 \| Line 822 \| public class ForkJoinPool extends Abstra
822
823		/**
824		* Pops the given task only if it is at the current top.
825	+	* (A shared version is available only via FJP.tryExternalUnpush)
826		*/
827		final boolean tryUnpush(ForkJoinTask<?> t) {
828		ForkJoinTask<?>[] a; int s;
#	Line 814 \| Line 836 \| public class ForkJoinPool extends Abstra
836		}
837
838		/**
839	<	* Polls the given task only if it is at the current base.
839	>	* Removes and cancels all known tasks, ignoring any exceptions.
840		*/
841	<	final boolean pollFor(ForkJoinTask<?> task) {
842	<	ForkJoinTask<?>[] a; int b;
843	<	if ((b = base) - top < 0 && (a = array) != null) {
844	<	int j = (((a.length - 1) & b) << ASHIFT) + ABASE;
845	<	if (U.getObjectVolatile(a, j) == task && base == b &&
846	<	U.compareAndSwapObject(a, j, task, null)) {
847	<	base = b + 1;
848	<	return true;
841	>	final void cancelAll() {
842	>	ForkJoinTask.cancelIgnoringExceptions(currentJoin);
843	>	ForkJoinTask.cancelIgnoringExceptions(currentSteal);
844	>	for (ForkJoinTask<?> t; (t = poll()) != null; )
845	>	ForkJoinTask.cancelIgnoringExceptions(t);
846	>	}
847	>
848	>	/**
849	>	* Computes next value for random probes. Scans don't require
850	>	* a very high quality generator, but also not a crummy one.
851	>	* Marsaglia xor-shift is cheap and works well enough. Note:
852	>	* This is manually inlined in its usages in ForkJoinPool to
853	>	* avoid writes inside busy scan loops.
854	>	*/
855	>	final int nextSeed() {
856	>	int r = seed;
857	>	r ^= r << 13;
858	>	r ^= r >>> 17;
859	>	return seed = r ^= r << 5;
860	>	}
861	>
862	>	/**
863	>	* Provides a more accurate estimate of size than (top - base)
864	>	* by ordering reads and checking whether a near-empty queue
865	>	* has at least one unclaimed task.
866	>	*/
867	>	final int queueSize() {
868	>	ForkJoinTask<?>[] a; int k, s, n;
869	>	return ((n = base - (s = top)) < 0 &&
870	>	(n != -1 \|\|
871	>	((a = array) != null && (k = a.length) > 0 &&
872	>	U.getObject
873	>	(a, (long)((((k - 1) & (s - 1)) << ASHIFT) + ABASE)) != null))) ?
874	>	-n : 0;
875	>	}
876	>
877	>	// Specialized execution methods
878	>
879	>	/**
880	>	* Pops and runs tasks until empty.
881	>	*/
882	>	private void popAndExecAll() {
883	>	// A bit faster than repeated pop calls
884	>	ForkJoinTask<?>[] a; int m, s; long j; ForkJoinTask<?> t;
885	>	while ((a = array) != null && (m = a.length - 1) >= 0 &&
886	>	(s = top - 1) - base >= 0 &&
887	>	(t = ((ForkJoinTask<?>)
888	>	U.getObject(a, j = ((m & s) << ASHIFT) + ABASE)))
889	>	!= null) {
890	>	if (U.compareAndSwapObject(a, j, t, null)) {
891	>	top = s;
892	>	t.doExec();
893		}
894		}
829	–	return false;
895		}
896
897		/**
898	<	* If present, removes from queue and executes the given task, or
899	<	* any other cancelled task. Returns (true) immediately on any CAS
898	>	* Polls and runs tasks until empty.
899	>	*/
900	>	private void pollAndExecAll() {
901	>	for (ForkJoinTask<?> t; (t = poll()) != null;)
902	>	t.doExec();
903	>	}
904	>
905	>	/**
906	>	* If present, removes from queue and executes the given task,
907	>	* or any other cancelled task. Returns (true) on any CAS
908		* or consistency check failure so caller can retry.
909		*
910	<	* @return false if no progress can be made
910	>	* @return false if no progress can be made, else true;
911		*/
912		final boolean tryRemoveAndExec(ForkJoinTask<?> task) {
913	<	boolean removed = false, empty = true, progress = true;
913	>	boolean stat = true, removed = false, empty = true;
914		ForkJoinTask<?>[] a; int m, s, b, n;
915		if ((a = array) != null && (m = a.length - 1) >= 0 &&
916		(n = (s = top) - (b = base)) > 0) {
#	Line 867 \| Line 940 \| public class ForkJoinPool extends Abstra
940		}
941		if (--n == 0) {
942		if (!empty && base == b)
943	<	progress = false;
943	>	stat = false;
944		break;
945		}
946		}
947		}
948		if (removed)
949		task.doExec();
950	<	return progress;
950	>	return stat;
951		}
952
953		/**
954	<	* Initializes or doubles the capacity of array. Call either
955	<	* by owner or with lock held -- it is OK for base, but not
883	<	* top, to move while resizings are in progress.
884	<	*
885	<	* @param rejectOnFailure if true, throw exception if capacity
886	<	* exceeded (relayed ultimately to user); else return null.
954	>	* Polls for and executes the given task or any other task in
955	>	* its CountedCompleter computation
956		*/
957	<	final ForkJoinTask<?>[] growArray(boolean rejectOnFailure) {
958	<	ForkJoinTask<?>[] oldA = array;
959	<	int size = oldA != null ? oldA.length << 1 : INITIAL_QUEUE_CAPACITY;
960	<	if (size <= MAXIMUM_QUEUE_CAPACITY) {
961	<	int oldMask, t, b;
962	<	ForkJoinTask<?>[] a = array = new ForkJoinTask<?>[size];
963	<	if (oldA != null && (oldMask = oldA.length - 1) >= 0 &&
964	<	(t = top) - (b = base) > 0) {
965	<	int mask = size - 1;
966	<	do {
967	<	ForkJoinTask<?> x;
968	<	int oldj = ((b & oldMask) << ASHIFT) + ABASE;
969	<	int j = ((b & mask) << ASHIFT) + ABASE;
970	<	x = (ForkJoinTask<?>)U.getObjectVolatile(oldA, oldj);
902	<	if (x != null &&
903	<	U.compareAndSwapObject(oldA, oldj, x, null))
904	<	U.putObjectVolatile(a, j, x);
905	<	} while (++b != t);
906	<	}
907	<	return a;
908	<	}
909	<	else if (!rejectOnFailure)
910	<	return null;
911	<	else
912	<	throw new RejectedExecutionException("Queue capacity exceeded");
913	<	}
914	<
915	<	/**
916	<	* Removes and cancels all known tasks, ignoring any exceptions.
917	<	*/
918	<	final void cancelAll() {
919	<	ForkJoinTask.cancelIgnoringExceptions(currentJoin);
920	<	ForkJoinTask.cancelIgnoringExceptions(currentSteal);
921	<	for (ForkJoinTask<?> t; (t = poll()) != null; )
922	<	ForkJoinTask.cancelIgnoringExceptions(t);
923	<	}
924	<
925	<	/**
926	<	* Computes next value for random probes. Scans don't require
927	<	* a very high quality generator, but also not a crummy one.
928	<	* Marsaglia xor-shift is cheap and works well enough. Note:
929	<	* This is manually inlined in its usages in ForkJoinPool to
930	<	* avoid writes inside busy scan loops.
931	<	*/
932	<	final int nextSeed() {
933	<	int r = seed;
934	<	r ^= r << 13;
935	<	r ^= r >>> 17;
936	<	return seed = r ^= r << 5;
937	<	}
938	<
939	<	// Execution methods
940	<
941	<	/**
942	<	* Removes and runs tasks until empty, using local mode
943	<	* ordering. Normally called only after checking for apparent
944	<	* non-emptiness.
945	<	*/
946	<	final void runLocalTasks() {
947	<	// hoist checks from repeated pop/poll
948	<	ForkJoinTask<?>[] a; int m;
949	<	if ((a = array) != null && (m = a.length - 1) >= 0) {
950	<	if (mode == 0) {
951	<	for (int s; (s = top - 1) - base >= 0;) {
952	<	int j = ((m & s) << ASHIFT) + ABASE;
953	<	ForkJoinTask<?> t =
954	<	(ForkJoinTask<?>)U.getObjectVolatile(a, j);
955	<	if (t != null) {
956	<	if (U.compareAndSwapObject(a, j, t, null)) {
957	<	top = s;
958	<	t.doExec();
959	<	}
957	>	final boolean pollAndExecCC(ForkJoinTask<?> root) {
958	>	ForkJoinTask<?>[] a; int b; Object o;
959	>	outer: while ((b = base) - top < 0 && (a = array) != null) {
960	>	long j = (((a.length - 1) & b) << ASHIFT) + ABASE;
961	>	if ((o = U.getObject(a, j)) == null \|\|
962	>	!(o instanceof CountedCompleter))
963	>	break;
964	>	for (CountedCompleter<?> t = (CountedCompleter<?>)o, r = t;;) {
965	>	if (r == root) {
966	>	if (base == b &&
967	>	U.compareAndSwapObject(a, j, t, null)) {
968	>	base = b + 1;
969	>	t.doExec();
970	>	return true;
971		}
972		else
973	<	break;
963	<	}
964	<	}
965	<	else {
966	<	for (int b; (b = base) - top < 0;) {
967	<	int j = ((m & b) << ASHIFT) + ABASE;
968	<	ForkJoinTask<?> t =
969	<	(ForkJoinTask<?>)U.getObjectVolatile(a, j);
970	<	if (t != null) {
971	<	if (base == b &&
972	<	U.compareAndSwapObject(a, j, t, null)) {
973	<	base = b + 1;
974	<	t.doExec();
975	<	}
976	<	} else if (base == b) {
977	<	if (b + 1 == top)
978	<	break;
979	<	Thread.yield(); // wait for lagging update
980	<	}
973	>	break; // restart
974		}
975	+	if ((r = r.completer) == null)
976	+	break outer; // not part of root computation
977		}
978		}
979	+	return false;
980		}
981
982		/**
983		* Executes a top-level task and any local tasks remaining
984		* after execution.
989	–	*
990	–	* @return true unless terminating
985		*/
986	<	final boolean runTask(ForkJoinTask<?> t) {
993	<	boolean alive = true;
986	>	final void runTask(ForkJoinTask<?> t) {
987		if (t != null) {
988	<	currentSteal = t;
996	<	t.doExec();
997	<	if (top != base) // conservative guard
998	<	runLocalTasks();
999	<	++nsteals;
988	>	(currentSteal = t).doExec();
989		currentSteal = null;
990	+	++nsteals;
991	+	if (top != base) { // process remaining local tasks
992	+	if (mode == 0)
993	+	popAndExecAll();
994	+	else
995	+	pollAndExecAll();
996	+	}
997		}
1002	–	else if (runState < 0) // terminating
1003	–	alive = false;
1004	–	return alive;
998		}
999
1000		/**
#	Line 1010 \| Line 1003 \| public class ForkJoinPool extends Abstra
1003		final void runSubtask(ForkJoinTask<?> t) {
1004		if (t != null) {
1005		ForkJoinTask<?> ps = currentSteal;
1006	<	currentSteal = t;
1014	<	t.doExec();
1006	>	(currentSteal = t).doExec();
1007		currentSteal = ps;
1008		}
1009		}
#	Line 1046 \| Line 1038 \| public class ForkJoinPool extends Abstra
1038
1039		// Unsafe mechanics
1040		private static final sun.misc.Unsafe U;
1041	<	private static final long RUNSTATE;
1041	>	private static final long QLOCK;
1042		private static final int ABASE;
1043		private static final int ASHIFT;
1044		static {
#	Line 1055 \| Line 1047 \| public class ForkJoinPool extends Abstra
1047		U = getUnsafe();
1048		Class<?> k = WorkQueue.class;
1049		Class<?> ak = ForkJoinTask[].class;
1050	<	RUNSTATE = U.objectFieldOffset
1051	<	(k.getDeclaredField("runState"));
1050	>	QLOCK = U.objectFieldOffset
1051	>	(k.getDeclaredField("qlock"));
1052		ABASE = U.arrayBaseOffset(ak);
1053		s = U.arrayIndexScale(ak);
1054		} catch (Exception e) {
#	Line 1068 \| Line 1060 \| public class ForkJoinPool extends Abstra
1060		}
1061		}
1062
1063	+	// static fields (initialized in static initializer below)
1064	+
1065	+	/**
1066	+	* Creates a new ForkJoinWorkerThread. This factory is used unless
1067	+	* overridden in ForkJoinPool constructors.
1068	+	*/
1069	+	public static final ForkJoinWorkerThreadFactory
1070	+	defaultForkJoinWorkerThreadFactory;
1071	+
1072		/**
1073		* Per-thread records for threads that submit to pools. Currently
1074		* holds only pseudo-random seed / index that is used to choose
1075	<	* submission queues in method doSubmit. In the future, this may
1075	>	* submission queues in method externalPush. In the future, this may
1076		* also incorporate a means to implement different task rejection
1077		* and resubmission policies.
1078		*
#	Line 1079 \| Line 1080 \| public class ForkJoinPool extends Abstra
1080		* the same way but are initialized and updated using slightly
1081		* different mechanics. Both are initialized using the same
1082		* approach as in class ThreadLocal, where successive values are
1083	<	* unlikely to collide with previous values. This is done during
1084	<	* registration for workers, but requires a separate AtomicInteger
1085	<	* for submitters. Seeds are then randomly modified upon
1085	<	* collisions using xorshifts, which requires a non-zero seed.
1083	>	* unlikely to collide with previous values. Seeds are then
1084	>	* randomly modified upon collisions using xorshifts, which
1085	>	* requires a non-zero seed.
1086		*/
1087		static final class Submitter {
1088		int seed;
1089	<	Submitter() {
1090	<	int s = nextSubmitterSeed.getAndAdd(SEED_INCREMENT);
1091	<	seed = (s == 0) ? 1 : s; // ensure non-zero
1092	<	}
1093	<	}
1094	<
1095	<	/** ThreadLocal class for Submitters */
1096	<	static final class ThreadSubmitter extends ThreadLocal<Submitter> {
1097	<	public Submitter initialValue() { return new Submitter(); }
1089	>	Submitter(int s) { seed = s; }
1090		}
1091
1092	<	// static fields (initialized in static initializer below)
1092	>	/**
1093	>	* Per-thread submission bookkeeping. Shared across all pools
1094	>	* to reduce ThreadLocal pollution and because random motion
1095	>	* to avoid contention in one pool is likely to hold for others.
1096	>	* Lazily initialized on first submission (but null-checked
1097	>	* in other contexts to avoid unnecessary initialization).
1098	>	*/
1099	>	static final ThreadLocal<Submitter> submitters;
1100
1101		/**
1102	<	* Creates a new ForkJoinWorkerThread. This factory is used unless
1103	<	* overridden in ForkJoinPool constructors.
1102	>	* Common (static) pool. Non-null for public use unless a static
1103	>	* construction exception, but internal usages null-check on use
1104	>	* to paranoically avoid potential initialization circularities
1105	>	* as well as to simplify generated code.
1106		*/
1107	<	public static final ForkJoinWorkerThreadFactory
1107	<	defaultForkJoinWorkerThreadFactory;
1107	>	static final ForkJoinPool commonPool;
1108
1109		/**
1110	<	* Generator for assigning sequence numbers as pool names.
1110	>	* Permission required for callers of methods that may start or
1111	>	* kill threads.
1112		*/
1113	<	private static final AtomicInteger poolNumberGenerator;
1113	>	private static final RuntimePermission modifyThreadPermission;
1114
1115		/**
1116	<	* Generator for initial hashes/seeds for submitters. Accessed by
1116	<	* Submitter class constructor.
1116	>	* Common pool parallelism. Must equal commonPool.parallelism.
1117		*/
1118	<	static final AtomicInteger nextSubmitterSeed;
1118	>	static final int commonPoolParallelism;
1119
1120		/**
1121	<	* Permission required for callers of methods that may start or
1122	<	* kill threads.
1121	>	* Sequence number for creating workerNamePrefix.
1122		*/
1123	<	private static final RuntimePermission modifyThreadPermission;
1123	>	private static int poolNumberSequence;
1124
1125		/**
1126	<	* Per-thread submission bookeeping. Shared across all pools
1127	<	* to reduce ThreadLocal pollution and because random motion
1129	<	* to avoid contention in one pool is likely to hold for others.
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 ThreadSubmitter submitters;
1129	>	private static final synchronized int nextPoolId() {
1130	>	return ++poolNumberSequence;
1131	>	}
1132
1133		// static constants
1134
1135		/**
1136	<	* The wakeup interval (in nanoseconds) for a worker waiting for a
1137	<	* task when the pool is quiescent to instead try to shrink the
1138	<	* number of workers. The exact value does not matter too
1139	<	* much. It must be short enough to release resources during
1140	<	* sustained periods of idleness, but not so short that threads
1141	<	* are continually re-created.
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 static final long SHRINK_RATE =
1144	<	4L * 1000L * 1000L * 1000L; // 4 seconds
1143	>	private static final long IDLE_TIMEOUT = 2000L * 1000L * 1000L; // 2sec
1144
1145		/**
1146	<	* The timeout value for attempted shrinkage, includes
1148	<	* some slop to cope with system timer imprecision.
1146	>	* Timeout value when there are more threads than parallelism level
1147		*/
1148	<	private static final long SHRINK_TIMEOUT = SHRINK_RATE - (SHRINK_RATE / 10);
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. This value also
1155	<	* controls the maximum number of times to try to help join a task
1156	<	* without any apparent progress or change in pool state before
1157	<	* giving up and blocking (see awaitJoin). Depths for legitimate
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 = 32;
1164	<
1165	<	/**
1166	<	* Secondary time-based bound (in nanosecs) for helping attempts
1167	<	* before trying compensated blocking in awaitJoin. Used in
1168	<	* conjunction with MAX_HELP to reduce variance due to different
1169	<	* polling rates associated with different helping options. The
1170	<	* value should roughly approximate the time required to create
1171	<	* and/or activate a worker thread.
1172	<	*/
1173	<	private static final long COMPENSATION_DELAY = 100L * 1000L; // 0.1 millisec
1158	>	private static final int MAX_HELP = 64;
1159
1160		/**
1161		* Increment for seed generators. See class ThreadLocal for
#	Line 1206 \| Line 1191 \| public class ForkJoinPool extends Abstra
1191		* scan for them to avoid queuing races. Note however that
1192		* eventCount updates lag releases so usage requires care.
1193		*
1194	<	* Field runState is an int packed with:
1194	>	* Field plock is an int packed with:
1195		* SHUTDOWN: true if shutdown is enabled (1 bit)
1196	<	* SEQ: a sequence number updated upon (de)registering workers (30 bits)
1197	<	* INIT: set true after workQueues array construction (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 runState before vs after the reads.
1201	>	* be checked by comparing plock before vs after the reads.
1202		*/
1203
1204		// bit positions/shifts for fields
#	Line 1225 \| Line 1210 \| public class ForkJoinPool extends Abstra
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 SQMASK = 0xfffe; // even short bits
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
#	Line 1250 \| Line 1236 \| public class ForkJoinPool extends Abstra
1236		private static final int E_MASK = 0x7fffffff; // no STOP_BIT
1237		private static final int E_SEQ = 1 << EC_SHIFT;
1238
1239	<	// runState bits
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	+	// bounds for #steps in scan loop -- must be power 2 minus 1
1251	+	private static final int MIN_SCAN = 0x1ff; // cover estimation slop
1252	+	private static final int MAX_SCAN = 0x1ffff; // 4 * max workers
1253	+
1254		// Instance fields
1255
1256		/*
1257	<	* Field layout order in this class tends to matter more than one
1258	<	* would like. Runtime layout order is only loosely related to
1257	>	* Field layout of this class tends to matter more than one would
1258	>	* like. Runtime layout order is only loosely related to
1259		* declaration order and may differ across JVMs, but the following
1260		* empirically works OK on current JVMs.
1261		*/
1262	<
1262	>	volatile long stealCount; // collects worker counts
1263		volatile long ctl; // main pool control
1264	<	final int parallelism; // parallelism level
1265	<	final int localMode; // per-worker scheduling mode
1266	<	final int submitMask; // submit queue index bound
1274	<	int nextSeed; // for initializing worker seeds
1275	<	volatile int runState; // shutdown status and seq
1264	>	volatile int plock; // shutdown status and seqLock
1265	>	volatile int indexSeed; // worker/submitter index seed
1266	>	final int config; // mode and parallelism level
1267		WorkQueue[] workQueues; // main registry
1268	<	final Mutex lock; // for registration
1278	<	final Condition termination; // for awaitTermination
1279	<	final ForkJoinWorkerThreadFactory factory; // factory for new workers
1268	>	final ForkJoinWorkerThreadFactory factory;
1269		final Thread.UncaughtExceptionHandler ueh; // per-worker UEH
1281	–	final AtomicLong stealCount; // collect counts when terminated
1282	–	final AtomicInteger nextWorkerNumber; // to create worker name string
1270		final String workerNamePrefix; // to create worker name string
1271
1272	<	// Creating, registering, and deregistering workers
1272	>	/*
1273	>	* Acquires the plock lock to protect worker array and related
1274	>	* updates. This method is called only if an initial CAS on plock
1275	>	* fails. This acts as a spinLock for normal cases, but falls back
1276	>	* to builtin monitor to block when (rarely) needed. This would be
1277	>	* a terrible idea for a highly contended lock, but works fine as
1278	>	* a more conservative alternative to a pure spinlock. See
1279	>	* internal ConcurrentHashMap documentation for further
1280	>	* explanation of nearly the same construction.
1281	>	*/
1282	>	private int acquirePlock() {
1283	>	int spins = PL_SPINS, r = 0, ps, nps;
1284	>	for (;;) {
1285	>	if (((ps = plock) & PL_LOCK) == 0 &&
1286	>	U.compareAndSwapInt(this, PLOCK, ps, nps = ps + PL_LOCK))
1287	>	return nps;
1288	>	else if (r == 0) { // randomize spins if possible
1289	>	Thread t = Thread.currentThread(); WorkQueue w; Submitter z;
1290	>	if ((t instanceof ForkJoinWorkerThread) &&
1291	>	(w = ((ForkJoinWorkerThread)t).workQueue) != null)
1292	>	r = w.seed;
1293	>	else if ((z = submitters.get()) != null)
1294	>	r = z.seed;
1295	>	else
1296	>	r = 1;
1297	>	}
1298	>	else if (spins >= 0) {
1299	>	r ^= r << 1; r ^= r >>> 3; r ^= r << 10; // xorshift
1300	>	if (r >= 0)
1301	>	--spins;
1302	>	}
1303	>	else if (U.compareAndSwapInt(this, PLOCK, ps, ps \| PL_SIGNAL)) {
1304	>	synchronized (this) {
1305	>	if ((plock & PL_SIGNAL) != 0) {
1306	>	try {
1307	>	wait();
1308	>	} catch (InterruptedException ie) {
1309	>	try {
1310	>	Thread.currentThread().interrupt();
1311	>	} catch (SecurityException ignore) {
1312	>	}
1313	>	}
1314	>	}
1315	>	else
1316	>	notifyAll();
1317	>	}
1318	>	}
1319	>	}
1320	>	}
1321	>
1322	>	/**
1323	>	* Unlocks and signals any thread waiting for plock. Called only
1324	>	* when CAS of seq value for unlock fails.
1325	>	*/
1326	>	private void releasePlock(int ps) {
1327	>	plock = ps;
1328	>	synchronized (this) { notifyAll(); }
1329	>	}
1330
1331		/**
1332	<	* Tries to create and start a worker
1332	>	* Tries to create and start a worker; adjusts counts etc on failure
1333		*/
1334		private void addWorker() {
1291	–	Throwable ex = null;
1335		ForkJoinWorkerThread wt = null;
1336		try {
1337	<	if ((wt = factory.newThread(this)) != null) {
1338	<	wt.start();
1339	<	return;
1297	<	}
1298	<	} catch (Throwable e) {
1299	<	ex = e;
1337	>	(wt = factory.newThread(this)).start();
1338	>	} catch (Throwable ex) {
1339	>	deregisterWorker(wt, ex); // adjust on failure
1340		}
1301	–	deregisterWorker(wt, ex); // adjust counts etc on failure
1341		}
1342
1343		/**
1344	<	* Callback from ForkJoinWorkerThread constructor to assign a
1345	<	* public name. This must be separate from registerWorker because
1346	<	* it is called during the "super" constructor call in
1347	<	* ForkJoinWorkerThread.
1348	<	*/
1349	<	final String nextWorkerName() {
1350	<	return workerNamePrefix.concat
1351	<	(Integer.toString(nextWorkerNumber.addAndGet(1)));
1344	>	* Performs secondary initialization, called when plock is zero.
1345	>	* Creates workQueue array and sets plock to a valid value. The
1346	>	* lock body must be exception-free (so no try/finally) so we
1347	>	* optimistically allocate new array outside the lock and throw
1348	>	* away if (very rarely) not needed. (A similar tactic is used in
1349	>	* fullExternalPush.) Because the plock seq value can eventually
1350	>	* wrap around zero, this method harmlessly fails to reinitialize
1351	>	* if workQueues exists, while still advancing plock.
1352	>	*/
1353	>	private void initWorkQueuesArray() {
1354	>	WorkQueue[] ws; int ps;
1355	>	int p = config & SMASK; // find power of two table size
1356	>	int n = (p > 1) ? p - 1 : 1; // ensure at least 2 slots
1357	>	n \|= n >>> 1; n \|= n >>> 2; n \|= n >>> 4; n \|= n >>> 8; n \|= n >>> 16;
1358	>	WorkQueue[] nws = new WorkQueue[(n + 1) << 1];
1359	>	if (((ps = plock) & PL_LOCK) != 0 \|\|
1360	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1361	>	ps = acquirePlock();
1362	>	if ((ws = workQueues) == null \|\| ws.length == 0)
1363	>	workQueues = nws;
1364	>	int nps = (ps & SHUTDOWN) \| ((ps + PL_LOCK) & ~SHUTDOWN);
1365	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1366	>	releasePlock(nps);
1367	>	long c; int u;
1368	>	if ((u = (int)((c = ctl) >>> 32)) < 0 && (int)c == 0) {
1369	>	long nc = (long)(((u + UTC_UNIT) & UTC_MASK) \|
1370	>	((u + UAC_UNIT) & UAC_MASK)) << 32;
1371	>	if (U.compareAndSwapLong(this, CTL, c, nc))
1372	>	addWorker();
1373	>	}
1374	>
1375		}
1376
1377	+	// Registering and deregistering workers
1378	+
1379		/**
1380	<	* Callback from ForkJoinWorkerThread constructor to establish its
1381	<	* poolIndex and record its WorkQueue. To avoid scanning bias due
1382	<	* to packing entries in front of the workQueues array, we treat
1383	<	* the array as a simple power-of-two hash table using per-thread
1384	<	* seed as hash, expanding as needed.
1380	>	* Callback from ForkJoinWorkerThread to establish and record its
1381	>	* WorkQueue. To avoid scanning bias due to packing entries in
1382	>	* front of the workQueues array, we treat the array as a simple
1383	>	* power-of-two hash table using per-thread seed as hash,
1384	>	* expanding as needed.
1385		*
1386	<	* @param w the worker's queue
1386	>	* @param wt the worker thread
1387		*/
1388	<	final void registerWorker(WorkQueue w) {
1389	<	Mutex lock = this.lock;
1390	<	lock.lock();
1391	<	try {
1392	<	WorkQueue[] ws = workQueues;
1393	<	if (w != null && ws != null) { // skip on shutdown/failure
1394	<	int rs, n;
1395	<	while ((n = ws.length) < // ensure can hold total
1396	<	(parallelism + (short)(ctl >>> TC_SHIFT) << 1))
1397	<	workQueues = ws = Arrays.copyOf(ws, n << 1);
1398	<	int m = n - 1;
1399	<	int s = nextSeed += SEED_INCREMENT; // rarely-colliding sequence
1400	<	w.seed = (s == 0) ? 1 : s; // ensure non-zero seed
1401	<	int r = (s << 1) \| 1; // use odd-numbered indices
1402	<	while (ws[r &= m] != null) // step by approx half size
1403	<	r += ((n >>> 1) & SQMASK) + 2;
1404	<	w.eventCount = w.poolIndex = r; // establish before recording
1405	<	ws[r] = w; // also update seq
1406	<	runState = ((rs = runState) & SHUTDOWN) \| ((rs + 2) & ~SHUTDOWN);
1388	>	final void registerWorker(ForkJoinWorkerThread wt) {
1389	>	if (wt != null && wt.workQueue == null) {
1390	>	int s, ps; // generate a rarely colliding candidate index seed
1391	>	do {} while (!U.compareAndSwapInt(this, INDEXSEED, s = indexSeed,
1392	>	s += SEED_INCREMENT) \|\|
1393	>	s == 0); // skip 0
1394	>	WorkQueue w = new WorkQueue(this, wt, config >>> 16, s);
1395	>	if (((ps = plock) & PL_LOCK) != 0 \|\|
1396	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1397	>	ps = acquirePlock();
1398	>	int nps = (ps & SHUTDOWN) \| ((ps + PL_LOCK) & ~SHUTDOWN);
1399	>	try {
1400	>	WorkQueue[] ws;
1401	>	if ((ws = workQueues) != null && wt.workQueue == null) {
1402	>	int n = ws.length, m = n - 1;
1403	>	int r = (s << 1) \| 1; // use odd-numbered indices
1404	>	if (ws[r &= m] != null) { // collision
1405	>	int probes = 0; // step by approx half size
1406	>	int step = (n <= 4) ? 2 : ((n >>> 1) & EVENMASK) + 2;
1407	>	while (ws[r = (r + step) & m] != null) {
1408	>	if (++probes >= n) {
1409	>	workQueues = ws = Arrays.copyOf(ws, n <<= 1);
1410	>	m = n - 1;
1411	>	probes = 0;
1412	>	}
1413	>	}
1414	>	}
1415	>	w.eventCount = w.poolIndex = r; // volatile write orders
1416	>	wt.workQueue = ws[r] = w;
1417	>	}
1418	>	} finally {
1419	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1420	>	releasePlock(nps);
1421		}
1344	–	} finally {
1345	–	lock.unlock();
1422		}
1423		}
1424
1425		/**
1426		* Final callback from terminating worker, as well as upon failure
1427	<	* to construct or start a worker in addWorker. Removes record of
1428	<	* worker from array, and adjusts counts. If pool is shutting
1429	<	* down, tries to complete termination.
1427	>	* to construct or start a worker. Removes record of worker from
1428	>	* array, and adjusts counts. If pool is shutting down, tries to
1429	>	* complete termination.
1430		*
1431	<	* @param wt the worker thread or null if addWorker failed
1431	>	* @param wt the worker thread or null if construction failed
1432		* @param ex the exception causing failure, or null if none
1433		*/
1434		final void deregisterWorker(ForkJoinWorkerThread wt, Throwable ex) {
1359	–	Mutex lock = this.lock;
1435		WorkQueue w = null;
1436		if (wt != null && (w = wt.workQueue) != null) {
1437	<	w.runState = -1; // ensure runState is set
1438	<	stealCount.getAndAdd(w.totalSteals + w.nsteals);
1439	<	int idx = w.poolIndex;
1440	<	lock.lock();
1441	<	try { // remove record from array
1437	>	int ps;
1438	>	w.qlock = -1; // ensure set
1439	>	long ns = w.nsteals, sc; // collect steal count
1440	>	do {} while (!U.compareAndSwapLong(this, STEALCOUNT,
1441	>	sc = stealCount, sc + ns));
1442	>	if (((ps = plock) & PL_LOCK) != 0 \|\|
1443	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1444	>	ps = acquirePlock();
1445	>	int nps = (ps & SHUTDOWN) \| ((ps + PL_LOCK) & ~SHUTDOWN);
1446	>	try {
1447	>	int idx = w.poolIndex;
1448		WorkQueue[] ws = workQueues;
1449		if (ws != null && idx >= 0 && idx < ws.length && ws[idx] == w)
1450		ws[idx] = null;
1451		} finally {
1452	<	lock.unlock();
1452	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1453	>	releasePlock(nps);
1454		}
1455		}
1456
#	Line 1381 \| Line 1463 \| public class ForkJoinPool extends Abstra
1463		if (!tryTerminate(false, false) && w != null) {
1464		w.cancelAll(); // cancel remaining tasks
1465		if (w.array != null) // suppress signal if never ran
1466	<	signalWork(); // wake up or create replacement
1466	>	helpSignal(null, 0); // wake up or create replacement
1467		if (ex == null) // help clean refs on way out
1468		ForkJoinTask.helpExpungeStaleExceptions();
1469		}
1470
1471		if (ex != null) // rethrow
1472	<	U.throwException(ex);
1472	>	ForkJoinTask.rethrow(ex);
1473		}
1474
1393	–
1475		// Submissions
1476
1477		/**
1478		* Unless shutting down, adds the given task to a submission queue
1479		* at submitter's current queue index (modulo submission
1480	<	* range). If no queue exists at the index, one is created. If
1481	<	* the queue is busy, another index is randomly chosen. The
1401	<	* submitMask bounds the effective number of queues to the
1402	<	* (nearest power of two for) parallelism level.
1480	>	* range). Only the most common path is directly handled in this
1481	>	* method. All others are relayed to fullExternalPush.
1482		*
1483		* @param task the task. Caller must ensure non-null.
1484		*/
1485	<	private void doSubmit(ForkJoinTask<?> task) {
1486	<	Submitter s = submitters.get();
1487	<	for (int r = s.seed, m = submitMask;;) {
1488	<	WorkQueue[] ws; WorkQueue q;
1489	<	int k = r & m & SQMASK; // use only even indices
1490	<	if (runState < 0 \|\| (ws = workQueues) == null \|\| ws.length <= k)
1491	<	throw new RejectedExecutionException(); // shutting down
1492	<	else if ((q = ws[k]) == null) { // create new queue
1493	<	WorkQueue nq = new WorkQueue(this, null, SHARED_QUEUE);
1494	<	Mutex lock = this.lock; // construct outside lock
1495	<	lock.lock();
1496	<	try { // recheck under lock
1497	<	int rs = runState; // to update seq
1419	<	if (ws == workQueues && ws[k] == null) {
1420	<	ws[k] = nq;
1421	<	runState = ((rs & SHUTDOWN) \| ((rs + 2) & ~SHUTDOWN));
1422	<	}
1423	<	} finally {
1424	<	lock.unlock();
1425	<	}
1426	<	}
1427	<	else if (q.trySharedPush(task)) {
1428	<	signalWork();
1485	>	final void externalPush(ForkJoinTask<?> task) {
1486	>	WorkQueue[] ws; WorkQueue q; Submitter z; int m; ForkJoinTask<?>[] a;
1487	>	if ((z = submitters.get()) != null && plock > 0 &&
1488	>	(ws = workQueues) != null && (m = (ws.length - 1)) >= 0 &&
1489	>	(q = ws[m & z.seed & SQMASK]) != null &&
1490	>	U.compareAndSwapInt(q, QLOCK, 0, 1)) { // lock
1491	>	int b = q.base, s = q.top, n, an;
1492	>	if ((a = q.array) != null && (an = a.length) > (n = s + 1 - b)) {
1493	>	U.putObject(a, (long)(((an - 1) & s) << ASHIFT) + ABASE, task);
1494	>	q.top = s + 1; // push on to deque
1495	>	q.qlock = 0;
1496	>	if (n <= 2)
1497	>	signalWork(q, 0);
1498		return;
1499		}
1500	<	else if (m > 1) { // move to a different index
1500	>	q.qlock = 0;
1501	>	}
1502	>	fullExternalPush(task);
1503	>	}
1504	>
1505	>	/**
1506	>	* Full version of externalPush. This method is called, among
1507	>	* other times, upon the first submission of the first task to the
1508	>	* pool, so must perform secondary initialization (via
1509	>	* initWorkQueuesArray). It also detects first submission by an
1510	>	* external thread by looking up its ThreadLocal, and creates a
1511	>	* new shared queue if the one at index if empty or contended. The
1512	>	* lock body must be exception-free (so no try/finally) so we
1513	>	* optimistically allocate new queues outside the lock and throw
1514	>	* them away if (very rarely) not needed.
1515	>	*/
1516	>	private void fullExternalPush(ForkJoinTask<?> task) {
1517	>	int r = 0;
1518	>	for (Submitter z = submitters.get();;) {
1519	>	WorkQueue[] ws; WorkQueue q; int ps, m, k;
1520	>	if (z == null) {
1521	>	if (U.compareAndSwapInt(this, INDEXSEED, r = indexSeed,
1522	>	r += SEED_INCREMENT) && r != 0)
1523	>	submitters.set(z = new Submitter(r));
1524	>	}
1525	>	else if (r == 0) { // move to a different index
1526	>	r = z.seed;
1527		r ^= r << 13; // same xorshift as WorkQueues
1528		r ^= r >>> 17;
1529	<	s.seed = r ^= r << 5;
1529	>	z.seed = r ^ (r << 5);
1530	>	}
1531	>	else if ((ps = plock) < 0)
1532	>	throw new RejectedExecutionException();
1533	>	else if (ps == 0 \|\| (ws = workQueues) == null \|\|
1534	>	(m = ws.length - 1) < 0)
1535	>	initWorkQueuesArray();
1536	>	else if ((q = ws[k = r & m & SQMASK]) != null) {
1537	>	if (q.trySharedPush(task))
1538	>	return;
1539	>	else
1540	>	r = 0; // move on contention
1541	>	}
1542	>	else if (((ps = plock) & PL_LOCK) == 0) { // create new queue
1543	>	q = new WorkQueue(this, null, SHARED_QUEUE, r);
1544	>	if (((ps = plock) & PL_LOCK) != 0 \|\|
1545	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1546	>	ps = acquirePlock();
1547	>	if ((ws = workQueues) != null && k < ws.length && ws[k] == null)
1548	>	ws[k] = q;
1549	>	int nps = (ps & SHUTDOWN) \| ((ps + PL_LOCK) & ~SHUTDOWN);
1550	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1551	>	releasePlock(nps);
1552		}
1553		else
1554	<	Thread.yield(); // yield if no alternatives
1554	>	r = 0; // try elsewhere while lock held
1555		}
1556		}
1557
#	Line 1449 \| Line 1566 \| public class ForkJoinPool extends Abstra
1566		}
1567
1568		/**
1569	<	* Tries to activate or create a worker if too few are active.
1570	<	*/
1571	<	final void signalWork() {
1572	<	long c; int u;
1573	<	while ((u = (int)((c = ctl) >>> 32)) < 0) { // too few active
1574	<	WorkQueue[] ws = workQueues; int e, i; WorkQueue w; Thread p;
1575	<	if ((e = (int)c) > 0) { // at least one waiting
1576	<	if (ws != null && (i = e & SMASK) < ws.length &&
1569	>	* Tries to create (at most one) or activate (possibly several)
1570	>	* workers if too few are active. On contention failure, continues
1571	>	* until at least one worker is signalled or the given queue is
1572	>	* empty or all workers are active.
1573	>	*
1574	>	* @param q if non-null, the queue holding tasks to be signalled
1575	>	* @param signals the target number of signals (at least one --
1576	>	* if argument is zero also sets signallee hint if parked).
1577	>	*/
1578	>	final void signalWork(WorkQueue q, int signals) {
1579	>	long c; int e, u, i, s; WorkQueue[] ws; WorkQueue w; Thread p;
1580	>	while ((u = (int)((c = ctl) >>> 32)) < 0) {
1581	>	if ((e = (int)c) > 0) {
1582	>	if ((ws = workQueues) != null && ws.length > (i = e & SMASK) &&
1583		(w = ws[i]) != null && w.eventCount == (e \| INT_SIGN)) {
1584		long nc = (((long)(w.nextWait & E_MASK)) \|
1585		((long)(u + UAC_UNIT) << 32));
1586		if (U.compareAndSwapLong(this, CTL, c, nc)) {
1587		w.eventCount = (e + E_SEQ) & E_MASK;
1588	<	if ((p = w.parker) != null)
1589	<	U.unpark(p); // activate and release
1590	<	break;
1588	>	if ((p = w.parker) != null) {
1589	>	if (q != null && signals == 0)
1590	>	w.hint = q.poolIndex;
1591	>	U.unpark(p);
1592	>	}
1593	>	if (--signals <= 0)
1594	>	break;
1595		}
1596	+	if (q != null && (s = q.queueSize()) <= signals &&
1597	+	(signals = s) <= 0)
1598	+	break;
1599		}
1600		else
1601		break;
1602		}
1603	<	else if (e == 0 && (u & SHORT_SIGN) != 0) { // too few total
1603	>	else if (e == 0 && (u & SHORT_SIGN) != 0) {
1604		long nc = (long)(((u + UTC_UNIT) & UTC_MASK) \|
1605		((u + UAC_UNIT) & UAC_MASK)) << 32;
1606		if (U.compareAndSwapLong(this, CTL, c, nc)) {
#	Line 1483 \| Line 1613 \| public class ForkJoinPool extends Abstra
1613		}
1614		}
1615
1486	–
1616		// Scanning for tasks
1617
1618		/**
1619		* Top-level runloop for workers, called by ForkJoinWorkerThread.run.
1620		*/
1621		final void runWorker(WorkQueue w) {
1622	<	w.growArray(false); // initialize queue array in this thread
1623	<	do {} while (w.runTask(scan(w)));
1622	>	if (w != null) // skip on initialization failure
1623	>	do { w.runTask(scan(w)); } while (w.qlock >= 0);
1624		}
1625
1626		/**
#	Line 1502 \| Line 1631 \| public class ForkJoinPool extends Abstra
1631		* contention, or state changes that indicate possible success on
1632		* re-invocation.
1633		*
1634	<	* The scan searches for tasks across a random permutation of
1635	<	* queues (starting at a random index and stepping by a random
1636	<	* relative prime, checking each at least once). The scan
1637	<	* terminates upon either finding a non-empty queue, or completing
1638	<	* the sweep. If the worker is not inactivated, it takes and
1639	<	* returns a task from this queue. On failure to find a task, we
1640	<	* take one of the following actions, after which the caller will
1641	<	* retry calling this method unless terminated.
1634	>	* The scan searches for tasks across queues (starting at a random
1635	>	* index, and relying on registerWorker to irregularly scatter
1636	>	* them within array to avoid bias), checking each at least twice.
1637	>	* The scan terminates upon either finding a non-empty queue, or
1638	>	* completing the sweep. If the worker is not inactivated, it
1639	>	* takes and returns a task from this queue. Otherwise, if not
1640	>	* activated, it signals workers (that may include itself) and
1641	>	* returns so caller can retry. Also returns for true if the
1642	>	* worker array may have changed during an empty scan. On failure
1643	>	* to find a task, we take one of the following actions, after
1644	>	* which the caller will retry calling this method unless
1645	>	* terminated.
1646		*
1647		* * If pool is terminating, terminate the worker.
1648		*
1516	–	* * If not a complete sweep, try to release a waiting worker. If
1517	–	* the scan terminated because the worker is inactivated, then the
1518	–	* released worker will often be the calling worker, and it can
1519	–	* succeed obtaining a task on the next call. Or maybe it is
1520	–	* another worker, but with same net effect. Releasing in other
1521	–	* cases as well ensures that we have enough workers running.
1522	–	*
1649		* * If not already enqueued, try to inactivate and enqueue the
1650		* worker on wait queue. Or, if inactivating has caused the pool
1651		* to be quiescent, relay to idleAwaitWork to check for
1652		* termination and possibly shrink pool.
1653		*
1654	<	* * If already inactive, and the caller has run a task since the
1655	<	* last empty scan, return (to allow rescan) unless others are
1656	<	* also inactivated. Field WorkQueue.rescans counts down on each
1531	<	* scan to ensure eventual inactivation and blocking.
1532	<	*
1533	<	* * If already enqueued and none of the above apply, park
1534	<	* awaiting signal,
1654	>	* * If already enqueued and none of the above apply, possibly
1655	>	* (with 1/2 probability) park awaiting signal, else lingering to
1656	>	* help scan and signal.
1657		*
1658		* @param w the worker (via its WorkQueue)
1659	<	* @return a task or null of none found
1659	>	* @return a task or null if none found
1660		*/
1661		private final ForkJoinTask<?> scan(WorkQueue w) {
1662	<	WorkQueue[] ws; // first update random seed
1663	<	int r = w.seed; r ^= r << 13; r ^= r >>> 17; w.seed = r ^= r << 5;
1664	<	int rs = runState, m; // volatile read order matters
1665	<	if ((ws = workQueues) != null && (m = ws.length - 1) > 0) {
1666	<	int ec = w.eventCount; // ec is negative if inactive
1667	<	int step = (r >>> 16) \| 1; // relative prime
1668	<	for (int j = (m + 1) << 2; ; r += step) {
1669	<	WorkQueue q; ForkJoinTask<?> t; ForkJoinTask<?>[] a; int b;
1670	<	if ((q = ws[r & m]) != null && (b = q.base) - q.top < 0 &&
1549	<	(a = q.array) != null) { // probably nonempty
1662	>	WorkQueue[] ws; int m, hint;
1663	>	int ps = plock; // read plock before ws
1664	>	if (w != null && (ws = workQueues) != null && (m = ws.length - 1) >= 0) {
1665	>	int ec = w.eventCount; // ec is negative if inactive
1666	>	int r = w.seed; r ^= r << 13; r ^= r >>> 17; w.seed = r ^= r << 5;
1667	>	for (int j = ((m + m + 1) \| MIN_SCAN) & MAX_SCAN; ; --j) {
1668	>	WorkQueue q; ForkJoinTask<?>[] a; int b;
1669	>	if ((q = ws[(r + j) & m]) != null && (b = q.base) - q.top < 0 &&
1670	>	(a = q.array) != null) { // probably nonempty
1671		int i = (((a.length - 1) & b) << ASHIFT) + ABASE;
1672	<	t = (ForkJoinTask<?>)U.getObjectVolatile(a, i);
1672	>	ForkJoinTask<?> t = (ForkJoinTask<?>)
1673	>	U.getObjectVolatile(a, i);
1674		if (q.base == b && ec >= 0 && t != null &&
1675		U.compareAndSwapObject(a, i, t, null)) {
1676	<	q.base = b + 1; // specialization of pollAt
1677	<	return t;
1676	>	if ((q.base = b + 1) - q.top < 0)
1677	>	signalWork(q, 0);
1678	>	return t; // taken
1679		}
1680	<	else if ((t != null \|\| b + 1 != q.top) &&
1681	<	(ec < 0 \|\| j <= m)) {
1682	<	rs = 0; // mark scan as imcomplete
1560	<	break; // caller can retry after release
1680	>	else if (ec < 0 \|\| j < m) { // cannot take or cannot rescan
1681	>	w.hint = q.poolIndex; // use hint below
1682	>	break; // let caller retry after signal
1683		}
1684		}
1685	<	if (--j < 0)
1686	<	break;
1687	<	}
1688	<	long c = ctl; int e = (int)c, a = (int)(c >> AC_SHIFT), nr, ns;
1689	<	if (e < 0) // decode ctl on empty scan
1690	<	w.runState = -1; // pool is terminating
1569	<	else if (rs == 0 \|\| rs != runState) { // incomplete scan
1570	<	WorkQueue v; Thread p; // try to release a waiter
1571	<	if (e > 0 && a < 0 && w.eventCount == ec &&
1572	<	(v = ws[e & m]) != null && v.eventCount == (e \| INT_SIGN)) {
1573	<	long nc = ((long)(v.nextWait & E_MASK) \|
1574	<	((c + AC_UNIT) & (AC_MASK\|TC_MASK)));
1575	<	if (ctl == c && U.compareAndSwapLong(this, CTL, c, nc)) {
1576	<	v.eventCount = (e + E_SEQ) & E_MASK;
1577	<	if ((p = v.parker) != null)
1578	<	U.unpark(p);
1685	>	else if (j < 0) { // end of scan; in loop to simplify code
1686	>	long c, sc; int e, ns;
1687	>	if ((ns = w.nsteals) != 0) {
1688	>	if (U.compareAndSwapLong(this, STEALCOUNT,
1689	>	sc = stealCount, sc + ns))
1690	>	w.nsteals = 0; // collect steals
1691		}
1692	+	else if (plock != ps) // ws may have changed
1693	+	break;
1694	+	else if ((e = (int)(c = ctl)) < 0)
1695	+	w.qlock = -1; // pool is terminating
1696	+	else if (ec >= 0) { // try to enqueue/inactivate
1697	+	long nc = ((long)ec \|
1698	+	((c - AC_UNIT) & (AC_MASK\|TC_MASK)));
1699	+	w.nextWait = e; // link and mark inactive
1700	+	w.hint = -1; // use hint if set while parked
1701	+	w.eventCount = ec \| INT_SIGN;
1702	+	if (ctl != c \|\|
1703	+	!U.compareAndSwapLong(this, CTL, c, nc))
1704	+	w.eventCount = ec; // unmark on CAS failure
1705	+	else if ((int)(c >> AC_SHIFT) == 1 - (config & SMASK))
1706	+	idleAwaitWork(w, nc, c);
1707	+	}
1708	+	else if (w.eventCount < 0) { // block
1709	+	Thread wt = Thread.currentThread();
1710	+	Thread.interrupted(); // clear status
1711	+	U.putObject(wt, PARKBLOCKER, this);
1712	+	w.parker = wt; // emulate LockSupport.park
1713	+	if (w.eventCount < 0) // recheck
1714	+	U.park(false, 0L);
1715	+	w.parker = null;
1716	+	U.putObject(wt, PARKBLOCKER, null);
1717	+	}
1718	+	break;
1719		}
1720		}
1721	<	else if (ec >= 0) { // try to enqueue/inactivate
1722	<	long nc = (long)ec \| ((c - AC_UNIT) & (AC_MASK\|TC_MASK));
1723	<	w.nextWait = e;
1724	<	w.eventCount = ec \| INT_SIGN; // mark as inactive
1725	<	if (ctl != c \|\| !U.compareAndSwapLong(this, CTL, c, nc))
1726	<	w.eventCount = ec; // unmark on CAS failure
1588	<	else {
1589	<	if ((ns = w.nsteals) != 0) {
1590	<	w.nsteals = 0; // set rescans if ran task
1591	<	w.rescans = (a > 0) ? 0 : a + parallelism;
1592	<	w.totalSteals += ns;
1593	<	}
1594	<	if (a == 1 - parallelism) // quiescent
1595	<	idleAwaitWork(w, nc, c);
1596	<	}
1597	<	}
1598	<	else if (w.eventCount < 0) { // already queued
1599	<	if ((nr = w.rescans) > 0) { // continue rescanning
1600	<	int ac = a + parallelism;
1601	<	if (((w.rescans = (ac < nr) ? ac : nr - 1) & 3) == 0)
1602	<	Thread.yield(); // yield before block
1603	<	}
1604	<	else {
1605	<	Thread.interrupted(); // clear status
1606	<	Thread wt = Thread.currentThread();
1607	<	U.putObject(wt, PARKBLOCKER, this);
1608	<	w.parker = wt; // emulate LockSupport.park
1609	<	if (w.eventCount < 0) // recheck
1610	<	U.park(false, 0L);
1611	<	w.parker = null;
1612	<	U.putObject(wt, PARKBLOCKER, null);
1613	<	}
1721	>	if ((hint = w.hint) >= 0) { // help signal
1722	>	WorkQueue[] vs; WorkQueue v; int k;
1723	>	w.hint = -1; // suppress resignal
1724	>	if ((vs = workQueues) != null && hint < vs.length &&
1725	>	(v = vs[hint]) != null && (k = v.base - v.top) < -1)
1726	>	signalWork(v, 1 - k);
1727		}
1728		}
1729		return null;
#	Line 1619 \| Line 1732 \| public class ForkJoinPool extends Abstra
1732		/**
1733		* If inactivating worker w has caused the pool to become
1734		* quiescent, checks for pool termination, and, so long as this is
1735	<	* not the only worker, waits for event for up to SHRINK_RATE
1736	<	* nanosecs. On timeout, if ctl has not changed, terminates the
1735	>	* not the only worker, waits for event for up to a given
1736	>	* duration. On timeout, if ctl has not changed, terminates the
1737		* worker, which will in turn wake up another worker to possibly
1738		* repeat this process.
1739		*
#	Line 1629 \| Line 1742 \| public class ForkJoinPool extends Abstra
1742		* @param prevCtl the ctl value to restore if thread is terminated
1743		*/
1744		private void idleAwaitWork(WorkQueue w, long currentCtl, long prevCtl) {
1745	<	if (w.eventCount < 0 && !tryTerminate(false, false) &&
1746	<	(int)prevCtl != 0 && ctl == currentCtl) {
1745	>	if (w != null && w.eventCount < 0 &&
1746	>	!tryTerminate(false, false) && (int)prevCtl != 0) {
1747	>	int dc = -(short)(currentCtl >>> TC_SHIFT);
1748	>	long parkTime = dc < 0 ? FAST_IDLE_TIMEOUT: (dc + 1) * IDLE_TIMEOUT;
1749	>	long deadline = System.nanoTime() + parkTime - 100000L; // 1ms slop
1750		Thread wt = Thread.currentThread();
1635	–	Thread.yield(); // yield before block
1751		while (ctl == currentCtl) {
1637	–	long startTime = System.nanoTime();
1752		Thread.interrupted(); // timed variant of version in scan()
1753		U.putObject(wt, PARKBLOCKER, this);
1754		w.parker = wt;
1755		if (ctl == currentCtl)
1756	<	U.park(false, SHRINK_RATE);
1756	>	U.park(false, parkTime);
1757		w.parker = null;
1758		U.putObject(wt, PARKBLOCKER, null);
1759		if (ctl != currentCtl)
1760		break;
1761	<	if (System.nanoTime() - startTime >= SHRINK_TIMEOUT &&
1761	>	if (deadline - System.nanoTime() <= 0L &&
1762		U.compareAndSwapLong(this, CTL, currentCtl, prevCtl)) {
1763		w.eventCount = (w.eventCount + E_SEQ) \| E_MASK;
1764	<	w.runState = -1; // shrink
1764	>	w.qlock = -1; // shrink
1765	>	w.hint = -1; // suppress helping
1766		break;
1767		}
1768		}
#	Line 1655 \| Line 1770 \| public class ForkJoinPool extends Abstra
1770		}
1771
1772		/**
1773	+	* Scans through queues looking for work (optionally, while
1774	+	* joining a task); if any are present, signals. May return early
1775	+	* if more signalling is detectably unneeded.
1776	+	*
1777	+	* @param task if non-null, return early if done
1778	+	* @param origin an index to start scan
1779	+	*/
1780	+	final int helpSignal(ForkJoinTask<?> task, int origin) {
1781	+	WorkQueue[] ws; WorkQueue q; int m, n, s, u;
1782	+	if ((ws = workQueues) != null && (m = ws.length - 1) >= 0) {
1783	+	for (int i = 0; i <= m; ++i) {
1784	+	if (task != null && (s = task.status) < 0)
1785	+	return s;
1786	+	if ((q = ws[(i + origin) & m]) != null &&
1787	+	(n = q.queueSize()) > 0) {
1788	+	signalWork(q, n);
1789	+	if ((u = (int)(ctl >>> 32)) >= 0 \|\| (u >> UAC_SHIFT) >= 0)
1790	+	break;
1791	+	}
1792	+	}
1793	+	}
1794	+	return 0;
1795	+	}
1796	+
1797	+	/**
1798		* Tries to locate and execute tasks for a stealer of the given
1799		* task, or in turn one of its stealers, Traces currentSteal ->
1800		* currentJoin links looking for a thread working on a descendant
#	Line 1665 \| Line 1805 \| public class ForkJoinPool extends Abstra
1805		* leaves hints in workers to speed up subsequent calls. The
1806		* implementation is very branchy to cope with potential
1807		* inconsistencies or loops encountering chains that are stale,
1808	<	* unknown, or so long that they are likely cyclic. All of these
1669	<	* cases are dealt with by just retrying by caller.
1808	>	* unknown, or so long that they are likely cyclic.
1809		*
1810		* @param joiner the joining worker
1811		* @param task the task to join
1812	<	* @return true if found or ran a task (and so is immediately retryable)
1812	>	* @return 0 if no progress can be made, negative if task
1813	>	* known complete, else positive
1814		*/
1815	<	private boolean tryHelpStealer(WorkQueue joiner, ForkJoinTask<?> task) {
1816	<	WorkQueue[] ws;
1817	<	int m, depth = MAX_HELP; // remaining chain depth
1818	<	boolean progress = false;
1819	<	if ((ws = workQueues) != null && (m = ws.length - 1) > 0 &&
1820	<	task.status >= 0) {
1821	<	ForkJoinTask<?> subtask = task; // current target
1822	<	outer: for (WorkQueue j = joiner;;) {
1823	<	WorkQueue stealer = null; // find stealer of subtask
1824	<	WorkQueue v = ws[j.stealHint & m]; // try hint
1825	<	if (v != null && v.currentSteal == subtask)
1826	<	stealer = v;
1827	<	else { // scan
1828	<	for (int i = 1; i <= m; i += 2) {
1829	<	if ((v = ws[i]) != null && v.currentSteal == subtask &&
1830	<	v != joiner) {
1831	<	stealer = v;
1832	<	j.stealHint = i; // save hint
1833	<	break;
1815	>	private int tryHelpStealer(WorkQueue joiner, ForkJoinTask<?> task) {
1816	>	int stat = 0, steps = 0; // bound to avoid cycles
1817	>	if (joiner != null && task != null) { // hoist null checks
1818	>	restart: for (;;) {
1819	>	ForkJoinTask<?> subtask = task; // current target
1820	>	for (WorkQueue j = joiner, v;;) { // v is stealer of subtask
1821	>	WorkQueue[] ws; int m, s, h;
1822	>	if ((s = task.status) < 0) {
1823	>	stat = s;
1824	>	break restart;
1825	>	}
1826	>	if ((ws = workQueues) == null \|\| (m = ws.length - 1) <= 0)
1827	>	break restart; // shutting down
1828	>	if ((v = ws[h = (j.hint \| 1) & m]) == null \|\|
1829	>	v.currentSteal != subtask) {
1830	>	for (int origin = h;;) { // find stealer
1831	>	if (((h = (h + 2) & m) & 15) == 1 &&
1832	>	(subtask.status < 0 \|\| j.currentJoin != subtask))
1833	>	continue restart; // occasional staleness check
1834	>	if ((v = ws[h]) != null &&
1835	>	v.currentSteal == subtask) {
1836	>	j.hint = h; // save hint
1837	>	break;
1838	>	}
1839	>	if (h == origin)
1840	>	break restart; // cannot find stealer
1841		}
1842		}
1843	<	if (stealer == null)
1844	<	break;
1845	<	}
1846	<
1847	<	for (WorkQueue q = stealer;;) { // try to help stealer
1848	<	ForkJoinTask[] a; ForkJoinTask<?> t; int b;
1849	<	if (task.status < 0)
1850	<	break outer;
1851	<	if ((b = q.base) - q.top < 0 && (a = q.array) != null) {
1852	<	progress = true;
1853	<	int i = (((a.length - 1) & b) << ASHIFT) + ABASE;
1854	<	t = (ForkJoinTask<?>)U.getObjectVolatile(a, i);
1855	<	if (subtask.status < 0) // must recheck before taking
1856	<	break outer;
1857	<	if (t != null &&
1858	<	q.base == b &&
1859	<	U.compareAndSwapObject(a, i, t, null)) {
1860	<	q.base = b + 1;
1861	<	joiner.runSubtask(t);
1843	>	for (;;) { // help stealer or descend to its stealer
1844	>	ForkJoinTask[] a; int b;
1845	>	if (subtask.status < 0) // surround probes with
1846	>	continue restart; // consistency checks
1847	>	if ((b = v.base) - v.top < 0 && (a = v.array) != null) {
1848	>	int i = (((a.length - 1) & b) << ASHIFT) + ABASE;
1849	>	ForkJoinTask<?> t =
1850	>	(ForkJoinTask<?>)U.getObjectVolatile(a, i);
1851	>	if (subtask.status < 0 \|\| j.currentJoin != subtask \|\|
1852	>	v.currentSteal != subtask)
1853	>	continue restart; // stale
1854	>	stat = 1; // apparent progress
1855	>	if (t != null && v.base == b &&
1856	>	U.compareAndSwapObject(a, i, t, null)) {
1857	>	v.base = b + 1; // help stealer
1858	>	joiner.runSubtask(t);
1859	>	}
1860	>	else if (v.base == b && ++steps == MAX_HELP)
1861	>	break restart; // v apparently stalled
1862	>	}
1863	>	else { // empty -- try to descend
1864	>	ForkJoinTask<?> next = v.currentJoin;
1865	>	if (subtask.status < 0 \|\| j.currentJoin != subtask \|\|
1866	>	v.currentSteal != subtask)
1867	>	continue restart; // stale
1868	>	else if (next == null \|\| ++steps == MAX_HELP)
1869	>	break restart; // dead-end or maybe cyclic
1870	>	else {
1871	>	subtask = next;
1872	>	j = v;
1873	>	break;
1874	>	}
1875		}
1716	–	else if (q.base == b)
1717	–	break outer; // possibly stalled
1718	–	}
1719	–	else { // descend
1720	–	ForkJoinTask<?> next = stealer.currentJoin;
1721	–	if (--depth <= 0 \|\| subtask.status < 0 \|\|
1722	–	next == null \|\| next == subtask)
1723	–	break outer; // stale, dead-end, or cyclic
1724	–	subtask = next;
1725	–	j = stealer;
1726	–	break;
1876		}
1877		}
1878		}
1879		}
1880	<	return progress;
1880	>	return stat;
1881		}
1882
1883		/**
1884	<	* If task is at base of some steal queue, steals and executes it.
1884	>	* Analog of tryHelpStealer for CountedCompleters. Tries to steal
1885	>	* and run tasks within the target's computation.
1886	>	*
1887	>	* @param task the task to join
1888	>	* @param mode if shared, exit upon completing any task
1889	>	* if all workers are active
1890		*
1737	–	* @param joiner the joining worker
1738	–	* @param task the task
1891		*/
1892	<	private void tryPollForAndExec(WorkQueue joiner, ForkJoinTask<?> task) {
1893	<	WorkQueue[] ws;
1894	<	if ((ws = workQueues) != null) {
1895	<	for (int j = 1; j < ws.length && task.status >= 0; j += 2) {
1896	<	WorkQueue q = ws[j];
1897	<	if (q != null && q.pollFor(task)) {
1898	<	joiner.runSubtask(task);
1899	<	break;
1892	>	private int helpComplete(ForkJoinTask<?> task, int mode) {
1893	>	WorkQueue[] ws; WorkQueue q; int m, n, s, u;
1894	>	if (task != null && (ws = workQueues) != null &&
1895	>	(m = ws.length - 1) >= 0) {
1896	>	for (int j = 1, origin = j;;) {
1897	>	if ((s = task.status) < 0)
1898	>	return s;
1899	>	if ((q = ws[j & m]) != null && q.pollAndExecCC(task)) {
1900	>	origin = j;
1901	>	if (mode == SHARED_QUEUE &&
1902	>	((u = (int)(ctl >>> 32)) >= 0 \|\| (u >> UAC_SHIFT) >= 0))
1903	>	break;
1904		}
1905	+	else if ((j = (j + 2) & m) == origin)
1906	+	break;
1907		}
1908		}
1909	+	return 0;
1910		}
1911
1912		/**
1913		* Tries to decrement active count (sometimes implicitly) and
1914		* possibly release or create a compensating worker in preparation
1915		* for blocking. Fails on contention or termination. Otherwise,
1916	<	* adds a new thread if no idle workers are available and either
1917	<	* pool would become completely starved or: (at least half
1759	<	* starved, and fewer than 50% spares exist, and there is at least
1760	<	* one task apparently available). Even though the availability
1761	<	* check requires a full scan, it is worthwhile in reducing false
1762	<	* alarms.
1763	<	*
1764	<	* @param task if nonnull, a task being waited for
1765	<	* @param blocker if nonnull, a blocker being waited for
1766	<	* @return true if the caller can block, else should recheck and retry
1916	>	* adds a new thread if no idle workers are available and pool
1917	>	* may become starved.
1918		*/
1919	<	final boolean tryCompensate(ForkJoinTask<?> task, ManagedBlocker blocker) {
1920	<	int pc = parallelism, e;
1921	<	long c = ctl;
1922	<	WorkQueue[] ws = workQueues;
1923	<	if ((e = (int)c) >= 0 && ws != null) {
1924	<	int u, a, ac, hc;
1925	<	int tc = (short)((u = (int)(c >>> 32)) >>> UTC_SHIFT) + pc;
1926	<	boolean replace = false;
1927	<	if ((a = u >> UAC_SHIFT) <= 0) {
1928	<	if ((ac = a + pc) <= 1)
1929	<	replace = true;
1930	<	else if ((e > 0 \|\| (task != null &&
1931	<	ac <= (hc = pc >>> 1) && tc < pc + hc))) {
1781	<	WorkQueue w;
1782	<	for (int j = 0; j < ws.length; ++j) {
1783	<	if ((w = ws[j]) != null && !w.isEmpty()) {
1784	<	replace = true;
1785	<	break; // in compensation range and tasks available
1786	<	}
1787	<	}
1919	>	final boolean tryCompensate() {
1920	>	int pc = config & SMASK, e, i, tc; long c;
1921	>	WorkQueue[] ws; WorkQueue w; Thread p;
1922	>	if ((ws = workQueues) != null && (e = (int)(c = ctl)) >= 0) {
1923	>	if (e != 0 && (i = e & SMASK) < ws.length &&
1924	>	(w = ws[i]) != null && w.eventCount == (e \| INT_SIGN)) {
1925	>	long nc = ((long)(w.nextWait & E_MASK) \|
1926	>	(c & (AC_MASK\|TC_MASK)));
1927	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1928	>	w.eventCount = (e + E_SEQ) & E_MASK;
1929	>	if ((p = w.parker) != null)
1930	>	U.unpark(p);
1931	>	return true; // replace with idle worker
1932		}
1933		}
1934	<	if ((task == null \|\| task.status >= 0) && // recheck need to block
1935	<	(blocker == null \|\| !blocker.isReleasable()) && ctl == c) {
1936	<	if (!replace) { // no compensation
1937	<	long nc = ((c - AC_UNIT) & AC_MASK) \| (c & ~AC_MASK);
1938	<	if (U.compareAndSwapLong(this, CTL, c, nc))
1939	<	return true;
1940	<	}
1941	<	else if (e != 0) { // release an idle worker
1942	<	WorkQueue w; Thread p; int i;
1943	<	if ((i = e & SMASK) < ws.length && (w = ws[i]) != null) {
1944	<	long nc = ((long)(w.nextWait & E_MASK) \|
1801	<	(c & (AC_MASK\|TC_MASK)));
1802	<	if (w.eventCount == (e \| INT_SIGN) &&
1803	<	U.compareAndSwapLong(this, CTL, c, nc)) {
1804	<	w.eventCount = (e + E_SEQ) & E_MASK;
1805	<	if ((p = w.parker) != null)
1806	<	U.unpark(p);
1807	<	return true;
1808	<	}
1809	<	}
1810	<	}
1811	<	else if (tc < MAX_CAP) { // create replacement
1812	<	long nc = ((c + TC_UNIT) & TC_MASK) \| (c & ~TC_MASK);
1813	<	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1814	<	addWorker();
1815	<	return true;
1816	<	}
1934	>	else if ((tc = (short)(c >>> TC_SHIFT)) >= 0 &&
1935	>	(int)(c >> AC_SHIFT) + pc > 1) {
1936	>	long nc = ((c - AC_UNIT) & AC_MASK) \| (c & ~AC_MASK);
1937	>	if (U.compareAndSwapLong(this, CTL, c, nc))
1938	>	return true; // no compensation
1939	>	}
1940	>	else if (tc + pc < MAX_CAP) {
1941	>	long nc = ((c + TC_UNIT) & TC_MASK) \| (c & ~TC_MASK);
1942	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1943	>	addWorker();
1944	>	return true;
1945		}
1946		}
1947		}
#	Line 1821 \| Line 1949 \| public class ForkJoinPool extends Abstra
1949		}
1950
1951		/**
1952	<	* Helps and/or blocks until the given task is done
1952	>	* Helps and/or blocks until the given task is done.
1953		*
1954		* @param joiner the joining worker
1955		* @param task the task
1956		* @return task status on exit
1957		*/
1958		final int awaitJoin(WorkQueue joiner, ForkJoinTask<?> task) {
1959	<	ForkJoinTask<?> prevJoin = joiner.currentJoin;
1960	<	joiner.currentJoin = task;
1961	<	long startTime = 0L;
1962	<	for (int k = 0, s; ; ++k) {
1963	<	if ((joiner.isEmpty() ? // try to help
1964	<	!tryHelpStealer(joiner, task) :
1965	<	!joiner.tryRemoveAndExec(task))) {
1966	<	if (k == 0) {
1967	<	startTime = System.nanoTime();
1968	<	tryPollForAndExec(joiner, task); // check uncommon case
1969	<	}
1970	<	else if ((k & (MAX_HELP - 1)) == 0 &&
1971	<	System.nanoTime() - startTime >= COMPENSATION_DELAY &&
1972	<	tryCompensate(task, null)) {
1973	<	if (task.trySetSignal() && task.status >= 0) {
1974	<	synchronized (task) {
1975	<	if (task.status >= 0) {
1976	<	try { // see ForkJoinTask
1977	<	task.wait(); // for explanation
1978	<	} catch (InterruptedException ie) {
1959	>	int s = 0;
1960	>	if (joiner != null && task != null && (s = task.status) >= 0) {
1961	>	ForkJoinTask<?> prevJoin = joiner.currentJoin;
1962	>	joiner.currentJoin = task;
1963	>	do {} while ((s = task.status) >= 0 &&
1964	>	joiner.queueSize() > 0 &&
1965	>	joiner.tryRemoveAndExec(task)); // process local tasks
1966	>	if (s >= 0 && (s = task.status) >= 0 &&
1967	>	(s = helpSignal(task, joiner.poolIndex)) >= 0 &&
1968	>	(task instanceof CountedCompleter))
1969	>	s = helpComplete(task, LIFO_QUEUE);
1970	>	int k = 0; // to perform pre-block yield for politeness
1971	>	while (s >= 0 && (s = task.status) >= 0) {
1972	>	if ((joiner.queueSize() > 0 \|\| // try helping
1973	>	(s = tryHelpStealer(joiner, task)) == 0) &&
1974	>	(s = task.status) >= 0) {
1975	>	if (k < 3) {
1976	>	if (++k < 3)
1977	>	s = helpSignal(task, joiner.poolIndex);
1978	>	else
1979	>	Thread.yield();
1980	>	}
1981	>	else if (!tryCompensate())
1982	>	k = 0;
1983	>	else {
1984	>	if (task.trySetSignal() && (s = task.status) >= 0) {
1985	>	synchronized (task) {
1986	>	if (task.status >= 0) {
1987	>	try { // see ForkJoinTask
1988	>	task.wait(); // for explanation
1989	>	} catch (InterruptedException ie) {
1990	>	}
1991		}
1992	+	else
1993	+	task.notifyAll();
1994		}
1853	–	else
1854	–	task.notifyAll();
1995		}
1996	+	long c; // re-activate
1997	+	do {} while (!U.compareAndSwapLong
1998	+	(this, CTL, c = ctl, c + AC_UNIT));
1999		}
1857	–	long c; // re-activate
1858	–	do {} while (!U.compareAndSwapLong
1859	–	(this, CTL, c = ctl, c + AC_UNIT));
2000		}
2001		}
2002	<	if ((s = task.status) < 0) {
1863	<	joiner.currentJoin = prevJoin;
1864	<	return s;
1865	<	}
1866	<	else if ((k & (MAX_HELP - 1)) == MAX_HELP >>> 1)
1867	<	Thread.yield(); // for politeness
2002	>	joiner.currentJoin = prevJoin;
2003		}
2004	+	return s;
2005		}
2006
2007		/**
#	Line 1875 \| Line 2011 \| public class ForkJoinPool extends Abstra
2011		*
2012		* @param joiner the joining worker
2013		* @param task the task
1878	–	* @return task status on exit
2014		*/
2015	<	final int helpJoinOnce(WorkQueue joiner, ForkJoinTask<?> task) {
2015	>	final void helpJoinOnce(WorkQueue joiner, ForkJoinTask<?> task) {
2016		int s;
2017	<	while ((s = task.status) >= 0 &&
2018	<	(joiner.isEmpty() ?
2019	<	tryHelpStealer(joiner, task) :
2020	<	joiner.tryRemoveAndExec(task)))
2021	<	;
2022	<	return s;
2017	>	if (joiner != null && task != null && (s = task.status) >= 0) {
2018	>	ForkJoinTask<?> prevJoin = joiner.currentJoin;
2019	>	joiner.currentJoin = task;
2020	>	do {} while ((s = task.status) >= 0 &&
2021	>	joiner.queueSize() > 0 &&
2022	>	joiner.tryRemoveAndExec(task));
2023	>	if (s >= 0 && (s = task.status) >= 0 &&
2024	>	(s = helpSignal(task, joiner.poolIndex)) >= 0 &&
2025	>	(task instanceof CountedCompleter))
2026	>	s = helpComplete(task, LIFO_QUEUE);
2027	>	if (s >= 0 && joiner.queueSize() == 0) {
2028	>	do {} while (task.status >= 0 &&
2029	>	tryHelpStealer(joiner, task) > 0);
2030	>	}
2031	>	joiner.currentJoin = prevJoin;
2032	>	}
2033		}
2034
2035		/**
#	Line 1892 \| Line 2037 \| public class ForkJoinPool extends Abstra
2037		* during a random, then cyclic scan, else null. This method must
2038		* be retried by caller if, by the time it tries to use the queue,
2039		* it is empty.
2040	+	* @param r a (random) seed for scanning
2041		*/
2042	<	private WorkQueue findNonEmptyStealQueue(WorkQueue w) {
1897	<	// Similar to loop in scan(), but ignoring submissions
1898	<	int r = w.seed; r ^= r << 13; r ^= r >>> 17; w.seed = r ^= r << 5;
1899	<	int step = (r >>> 16) \| 1;
2042	>	private WorkQueue findNonEmptyStealQueue(int r) {
2043		for (WorkQueue[] ws;;) {
2044	<	int rs = runState, m;
2044	>	int ps = plock, m, n;
2045		if ((ws = workQueues) == null \|\| (m = ws.length - 1) < 1)
2046		return null;
2047	<	for (int j = (m + 1) << 2; ; r += step) {
2048	<	WorkQueue q = ws[((r << 1) \| 1) & m];
2049	<	if (q != null && !q.isEmpty())
2047	>	for (int j = (m + 1) << 2; ;) {
2048	>	WorkQueue q = ws[(((r + j) << 1) \| 1) & m];
2049	>	if (q != null && (n = q.queueSize()) > 0) {
2050	>	if (n > 1)
2051	>	signalWork(q, 0);
2052		return q;
2053	+	}
2054		else if (--j < 0) {
2055	<	if (runState == rs)
2055	>	if (plock == ps)
2056		return null;
2057		break;
2058		}
#	Line 1922 \| Line 2068 \| public class ForkJoinPool extends Abstra
2068		*/
2069		final void helpQuiescePool(WorkQueue w) {
2070		for (boolean active = true;;) {
2071	<	if (w.base - w.top < 0)
2072	<	w.runLocalTasks(); // exhaust local queue
2073	<	WorkQueue q = findNonEmptyStealQueue(w);
2071	>	ForkJoinTask<?> localTask; // exhaust local queue
2072	>	while ((localTask = w.nextLocalTask()) != null)
2073	>	localTask.doExec();
2074	>	// Similar to loop in scan(), but ignoring submissions
2075	>	WorkQueue q = findNonEmptyStealQueue(w.nextSeed());
2076		if (q != null) {
2077		ForkJoinTask<?> t; int b;
2078		if (!active) { // re-establish active count
#	Line 1945 \| Line 2093 \| public class ForkJoinPool extends Abstra
2093		}
2094		else
2095		c = ctl; // re-increment on exit
2096	<	if ((int)(c >> AC_SHIFT) + parallelism == 0) {
2096	>	if ((int)(c >> AC_SHIFT) + (config & SMASK) == 0) {
2097		do {} while (!U.compareAndSwapLong
2098		(this, CTL, c = ctl, c + AC_UNIT));
2099		break;
#	Line 1964 \| Line 2112 \| public class ForkJoinPool extends Abstra
2112		WorkQueue q; int b;
2113		if ((t = w.nextLocalTask()) != null)
2114		return t;
2115	<	if ((q = findNonEmptyStealQueue(w)) == null)
2115	>	if ((q = findNonEmptyStealQueue(w.nextSeed())) == null)
2116		return null;
2117		if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null)
2118		return t;
#	Line 1972 \| Line 2120 \| public class ForkJoinPool extends Abstra
2120		}
2121
2122		/**
2123	<	* Returns the approximate (non-atomic) number of idle threads per
2124	<	* active thread to offset steal queue size for method
2125	<	* ForkJoinTask.getSurplusQueuedTaskCount().
2126	<	*/
2127	<	final int idlePerActive() {
2128	<	// Approximate at powers of two for small values, saturate past 4
2129	<	int p = parallelism;
2130	<	int a = p + (int)(ctl >> AC_SHIFT);
2131	<	return (a > (p >>>= 1) ? 0 :
2132	<	a > (p >>>= 1) ? 1 :
2133	<	a > (p >>>= 1) ? 2 :
2134	<	a > (p >>>= 1) ? 4 :
2135	<	8);
2123	>	* Returns a cheap heuristic guide for task partitioning when
2124	>	* programmers, frameworks, tools, or languages have little or no
2125	>	* idea about task granularity. In essence by offering this
2126	>	* method, we ask users only about tradeoffs in overhead vs
2127	>	* expected throughput and its variance, rather than how finely to
2128	>	* partition tasks.
2129	>	*
2130	>	* In a steady state strict (tree-structured) computation, each
2131	>	* thread makes available for stealing enough tasks for other
2132	>	* threads to remain active. Inductively, if all threads play by
2133	>	* the same rules, each thread should make available only a
2134	>	* constant number of tasks.
2135	>	*
2136	>	* The minimum useful constant is just 1. But using a value of 1
2137	>	* would require immediate replenishment upon each steal to
2138	>	* maintain enough tasks, which is infeasible. Further,
2139	>	* partitionings/granularities of offered tasks should minimize
2140	>	* steal rates, which in general means that threads nearer the top
2141	>	* of computation tree should generate more than those nearer the
2142	>	* bottom. In perfect steady state, each thread is at
2143	>	* approximately the same level of computation tree. However,
2144	>	* producing extra tasks amortizes the uncertainty of progress and
2145	>	* diffusion assumptions.
2146	>	*
2147	>	* So, users will want to use values larger, but not much larger
2148	>	* than 1 to both smooth over transient shortages and hedge
2149	>	* against uneven progress; as traded off against the cost of
2150	>	* extra task overhead. We leave the user to pick a threshold
2151	>	* value to compare with the results of this call to guide
2152	>	* decisions, but recommend values such as 3.
2153	>	*
2154	>	* When all threads are active, it is on average OK to estimate
2155	>	* surplus strictly locally. In steady-state, if one thread is
2156	>	* maintaining say 2 surplus tasks, then so are others. So we can
2157	>	* just use estimated queue length. However, this strategy alone
2158	>	* leads to serious mis-estimates in some non-steady-state
2159	>	* conditions (ramp-up, ramp-down, other stalls). We can detect
2160	>	* many of these by further considering the number of "idle"
2161	>	* threads, that are known to have zero queued tasks, so
2162	>	* compensate by a factor of (#idle/#active) threads.
2163	>	*
2164	>	* Note: The approximation of #busy workers as #active workers is
2165	>	* not very good under current signalling scheme, and should be
2166	>	* improved.
2167	>	*/
2168	>	static int getSurplusQueuedTaskCount() {
2169	>	Thread t; ForkJoinWorkerThread wt; ForkJoinPool pool; WorkQueue q;
2170	>	if (((t = Thread.currentThread()) instanceof ForkJoinWorkerThread)) {
2171	>	int p = (pool = (wt = (ForkJoinWorkerThread)t).pool).config & SMASK;
2172	>	int n = (q = wt.workQueue).top - q.base;
2173	>	int a = (int)(pool.ctl >> AC_SHIFT) + p;
2174	>	return n - (a > (p >>>= 1) ? 0 :
2175	>	a > (p >>>= 1) ? 1 :
2176	>	a > (p >>>= 1) ? 2 :
2177	>	a > (p >>>= 1) ? 4 :
2178	>	8);
2179	>	}
2180	>	return 0;
2181		}
2182
2183		// Termination
#	Line 2004 \| Line 2197 \| public class ForkJoinPool extends Abstra
2197		* @return true if now terminating or terminated
2198		*/
2199		private boolean tryTerminate(boolean now, boolean enable) {
2200	<	Mutex lock = this.lock;
2200	>	if (this == commonPool) // cannot shut down
2201	>	return false;
2202		for (long c;;) {
2203		if (((c = ctl) & STOP_BIT) != 0) { // already terminating
2204	<	if ((short)(c >>> TC_SHIFT) == -parallelism) {
2205	<	lock.lock(); // don't need try/finally
2206	<	termination.signalAll(); // signal when 0 workers
2207	<	lock.unlock();
2204	>	if ((short)(c >>> TC_SHIFT) == -(config & SMASK)) {
2205	>	synchronized (this) {
2206	>	notifyAll(); // signal when 0 workers
2207	>	}
2208		}
2209		return true;
2210		}
2211	<	if (runState >= 0) { // not yet enabled
2211	>	if (plock >= 0) { // not yet enabled
2212	>	int ps;
2213		if (!enable)
2214		return false;
2215	<	lock.lock();
2216	<	runState \|= SHUTDOWN;
2217	<	lock.unlock();
2215	>	if (((ps = plock) & PL_LOCK) != 0 \|\|
2216	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
2217	>	ps = acquirePlock();
2218	>	int nps = SHUTDOWN;
2219	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
2220	>	releasePlock(nps);
2221		}
2222		if (!now) { // check if idle & no tasks
2223	<	if ((int)(c >> AC_SHIFT) != -parallelism \|\|
2223	>	if ((int)(c >> AC_SHIFT) != -(config & SMASK) \|\|
2224		hasQueuedSubmissions())
2225		return false;
2226		// Check for unqueued inactive workers. One pass suffices.
#	Line 2042 \| Line 2240 \| public class ForkJoinPool extends Abstra
2240		int n = ws.length;
2241		for (int i = 0; i < n; ++i) {
2242		if ((w = ws[i]) != null) {
2243	<	w.runState = -1;
2243	>	w.qlock = -1;
2244		if (pass > 0) {
2245		w.cancelAll();
2246		if (pass > 1)
#	Line 2061 \| Line 2259 \| public class ForkJoinPool extends Abstra
2259		if (w.eventCount == (e \| INT_SIGN) &&
2260		U.compareAndSwapLong(this, CTL, cc, nc)) {
2261		w.eventCount = (e + E_SEQ) & E_MASK;
2262	<	w.runState = -1;
2262	>	w.qlock = -1;
2263		if ((p = w.parker) != null)
2264		U.unpark(p);
2265		}
#	Line 2072 \| Line 2270 \| public class ForkJoinPool extends Abstra
2270		}
2271		}
2272
2273	+	// external operations on common pool
2274	+
2275	+	/**
2276	+	* Returns common pool queue for a thread that has submitted at
2277	+	* least one task.
2278	+	*/
2279	+	static WorkQueue commonSubmitterQueue() {
2280	+	ForkJoinPool p; WorkQueue[] ws; int m; Submitter z;
2281	+	return ((z = submitters.get()) != null &&
2282	+	(p = commonPool) != null &&
2283	+	(ws = p.workQueues) != null &&
2284	+	(m = ws.length - 1) >= 0) ?
2285	+	ws[m & z.seed & SQMASK] : null;
2286	+	}
2287	+
2288	+	/**
2289	+	* Tries to pop the given task from submitter's queue in common pool.
2290	+	*/
2291	+	static boolean tryExternalUnpush(ForkJoinTask<?> t) {
2292	+	ForkJoinPool p; WorkQueue[] ws; WorkQueue q; Submitter z;
2293	+	ForkJoinTask<?>[] a; int m, s; long j;
2294	+	if ((z = submitters.get()) != null &&
2295	+	(p = commonPool) != null &&
2296	+	(ws = p.workQueues) != null &&
2297	+	(m = ws.length - 1) >= 0 &&
2298	+	(q = ws[m & z.seed & SQMASK]) != null &&
2299	+	(s = q.top) != q.base &&
2300	+	(a = q.array) != null &&
2301	+	U.getObjectVolatile
2302	+	(a, j = (((a.length - 1) & (s - 1)) << ASHIFT) + ABASE) == t &&
2303	+	U.compareAndSwapInt(q, QLOCK, 0, 1)) {
2304	+	if (q.array == a && q.top == s && // recheck
2305	+	U.compareAndSwapObject(a, j, t, null)) {
2306	+	q.top = s - 1;
2307	+	q.qlock = 0;
2308	+	return true;
2309	+	}
2310	+	q.qlock = 0;
2311	+	}
2312	+	return false;
2313	+	}
2314	+
2315	+	/**
2316	+	* Tries to pop and run local tasks within the same computation
2317	+	* as the given root. On failure, tries to help complete from
2318	+	* other queues via helpComplete.
2319	+	*/
2320	+	private void externalHelpComplete(WorkQueue q, ForkJoinTask<?> root) {
2321	+	ForkJoinTask<?>[] a; int m;
2322	+	if (q != null && (a = q.array) != null && (m = (a.length - 1)) >= 0 &&
2323	+	root != null && root.status >= 0) {
2324	+	for (;;) {
2325	+	int s, u; Object o; CountedCompleter<?> task = null;
2326	+	if ((s = q.top) - q.base > 0) {
2327	+	long j = ((m & (s - 1)) << ASHIFT) + ABASE;
2328	+	if ((o = U.getObject(a, j)) != null &&
2329	+	(o instanceof CountedCompleter)) {
2330	+	CountedCompleter<?> t = (CountedCompleter<?>)o, r = t;
2331	+	do {
2332	+	if (r == root) {
2333	+	if (U.compareAndSwapInt(q, QLOCK, 0, 1)) {
2334	+	if (q.array == a && q.top == s &&
2335	+	U.compareAndSwapObject(a, j, t, null)) {
2336	+	q.top = s - 1;
2337	+	task = t;
2338	+	}
2339	+	q.qlock = 0;
2340	+	}
2341	+	break;
2342	+	}
2343	+	} while ((r = r.completer) != null);
2344	+	}
2345	+	}
2346	+	if (task != null)
2347	+	task.doExec();
2348	+	if (root.status < 0 \|\|
2349	+	(u = (int)(ctl >>> 32)) >= 0 \|\| (u >> UAC_SHIFT) >= 0)
2350	+	break;
2351	+	if (task == null) {
2352	+	if (helpSignal(root, q.poolIndex) >= 0)
2353	+	helpComplete(root, SHARED_QUEUE);
2354	+	break;
2355	+	}
2356	+	}
2357	+	}
2358	+	}
2359	+
2360	+	/**
2361	+	* Tries to help execute or signal availability of the given task
2362	+	* from submitter's queue in common pool.
2363	+	*/
2364	+	static void externalHelpJoin(ForkJoinTask<?> t) {
2365	+	// Some hard-to-avoid overlap with tryExternalUnpush
2366	+	ForkJoinPool p; WorkQueue[] ws; WorkQueue q, w; Submitter z;
2367	+	ForkJoinTask<?>[] a; int m, s, n; long j;
2368	+	if (t != null &&
2369	+	(z = submitters.get()) != null &&
2370	+	(p = commonPool) != null &&
2371	+	(ws = p.workQueues) != null &&
2372	+	(m = ws.length - 1) >= 0 &&
2373	+	(q = ws[m & z.seed & SQMASK]) != null &&
2374	+	(a = q.array) != null &&
2375	+	t.status >= 0) {
2376	+	if ((s = q.top) != q.base &&
2377	+	U.getObjectVolatile
2378	+	(a, j = (((a.length - 1) & (s - 1)) << ASHIFT) + ABASE) == t &&
2379	+	U.compareAndSwapInt(q, QLOCK, 0, 1)) {
2380	+	if (q.array == a && q.top == s &&
2381	+	U.compareAndSwapObject(a, j, t, null)) {
2382	+	q.top = s - 1;
2383	+	q.qlock = 0;
2384	+	t.doExec();
2385	+	}
2386	+	else
2387	+	q.qlock = 0;
2388	+	}
2389	+	if (t.status >= 0) {
2390	+	if (t instanceof CountedCompleter)
2391	+	p.externalHelpComplete(q, t);
2392	+	else
2393	+	p.helpSignal(t, q.poolIndex);
2394	+	}
2395	+	}
2396	+	}
2397	+
2398	+	/**
2399	+	* Restricted version of helpQuiescePool for external callers
2400	+	*/
2401	+	static void externalHelpQuiescePool() {
2402	+	ForkJoinPool p; ForkJoinTask<?> t; WorkQueue q; int b;
2403	+	if ((p = commonPool) != null &&
2404	+	(q = p.findNonEmptyStealQueue(1)) != null &&
2405	+	(b = q.base) - q.top < 0 &&
2406	+	(t = q.pollAt(b)) != null)
2407	+	t.doExec();
2408	+	}
2409	+
2410		// Exported methods
2411
2412		// Constructors
#	Line 2143 \| Line 2478 \| public class ForkJoinPool extends Abstra
2478		throw new NullPointerException();
2479		if (parallelism <= 0 \|\| parallelism > MAX_CAP)
2480		throw new IllegalArgumentException();
2146	–	this.parallelism = parallelism;
2481		this.factory = factory;
2482		this.ueh = handler;
2483	<	this.localMode = asyncMode ? FIFO_QUEUE : LIFO_QUEUE;
2483	>	this.config = parallelism \| (asyncMode? (FIFO_QUEUE << 16) : 0);
2484		long np = (long)(-parallelism); // offset ctl counts
2485		this.ctl = ((np << AC_SHIFT) & AC_MASK) \| ((np << TC_SHIFT) & TC_MASK);
2486	<	// Use nearest power 2 for workQueues size. See Hackers Delight sec 3.2.
2153	<	int n = parallelism - 1;
2154	<	n \|= n >>> 1; n \|= n >>> 2; n \|= n >>> 4; n \|= n >>> 8; n \|= n >>> 16;
2155	<	int size = (n + 1) << 1; // #slots = 2*#workers
2156	<	this.submitMask = size - 1; // room for max # of submit queues
2157	<	this.workQueues = new WorkQueue[size];
2158	<	this.termination = (this.lock = new Mutex()).newCondition();
2159	<	this.stealCount = new AtomicLong();
2160	<	this.nextWorkerNumber = new AtomicInteger();
2161	<	int pn = poolNumberGenerator.incrementAndGet();
2486	>	int pn = nextPoolId();
2487		StringBuilder sb = new StringBuilder("ForkJoinPool-");
2488		sb.append(Integer.toString(pn));
2489		sb.append("-worker-");
2490		this.workerNamePrefix = sb.toString();
2491	<	lock.lock();
2492	<	this.runState = 1; // set init flag
2493	<	lock.unlock();
2491	>	}
2492	>
2493	>	/**
2494	>	* Constructor for common pool, suitable only for static initialization.
2495	>	* Basically the same as above, but uses smallest possible initial footprint.
2496	>	*/
2497	>	ForkJoinPool(int parallelism, long ctl,
2498	>	ForkJoinWorkerThreadFactory factory,
2499	>	Thread.UncaughtExceptionHandler handler) {
2500	>	this.config = parallelism;
2501	>	this.ctl = ctl;
2502	>	this.factory = factory;
2503	>	this.ueh = handler;
2504	>	this.workerNamePrefix = "ForkJoinPool.commonPool-worker-";
2505	>	}
2506	>
2507	>	/**
2508	>	* Returns the common pool instance.
2509	>	*
2510	>	* @return the common pool instance
2511	>	*/
2512	>	public static ForkJoinPool commonPool() {
2513	>	// assert commonPool != null : "static init error";
2514	>	return commonPool;
2515		}
2516
2517		// Execution methods
#	Line 2189 \| Line 2535 \| public class ForkJoinPool extends Abstra
2535		public <T> T invoke(ForkJoinTask<T> task) {
2536		if (task == null)
2537		throw new NullPointerException();
2538	<	doSubmit(task);
2538	>	externalPush(task);
2539		return task.join();
2540		}
2541
#	Line 2204 \| Line 2550 \| public class ForkJoinPool extends Abstra
2550		public void execute(ForkJoinTask<?> task) {
2551		if (task == null)
2552		throw new NullPointerException();
2553	<	doSubmit(task);
2553	>	externalPush(task);
2554		}
2555
2556		// AbstractExecutorService methods
#	Line 2222 \| Line 2568 \| public class ForkJoinPool extends Abstra
2568		job = (ForkJoinTask<?>) task;
2569		else
2570		job = new ForkJoinTask.AdaptedRunnableAction(task);
2571	<	doSubmit(job);
2571	>	externalPush(job);
2572		}
2573
2574		/**
#	Line 2237 \| Line 2583 \| public class ForkJoinPool extends Abstra
2583		public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) {
2584		if (task == null)
2585		throw new NullPointerException();
2586	<	doSubmit(task);
2586	>	externalPush(task);
2587		return task;
2588		}
2589
#	Line 2248 \| Line 2594 \| public class ForkJoinPool extends Abstra
2594		*/
2595		public <T> ForkJoinTask<T> submit(Callable<T> task) {
2596		ForkJoinTask<T> job = new ForkJoinTask.AdaptedCallable<T>(task);
2597	<	doSubmit(job);
2597	>	externalPush(job);
2598		return job;
2599		}
2600
#	Line 2259 \| Line 2605 \| public class ForkJoinPool extends Abstra
2605		*/
2606		public <T> ForkJoinTask<T> submit(Runnable task, T result) {
2607		ForkJoinTask<T> job = new ForkJoinTask.AdaptedRunnable<T>(task, result);
2608	<	doSubmit(job);
2608	>	externalPush(job);
2609		return job;
2610		}
2611
#	Line 2276 \| Line 2622 \| public class ForkJoinPool extends Abstra
2622		job = (ForkJoinTask<?>) task;
2623		else
2624		job = new ForkJoinTask.AdaptedRunnableAction(task);
2625	<	doSubmit(job);
2625	>	externalPush(job);
2626		return job;
2627		}
2628
#	Line 2298 \| Line 2644 \| public class ForkJoinPool extends Abstra
2644		try {
2645		for (Callable<T> t : tasks) {
2646		ForkJoinTask<T> f = new ForkJoinTask.AdaptedCallable<T>(t);
2647	<	doSubmit(f);
2647	>	externalPush(f);
2648		fs.add(f);
2649		}
2650		for (ForkJoinTask<T> f : fs)
#	Line 2337 \| Line 2683 \| public class ForkJoinPool extends Abstra
2683		* @return the targeted parallelism level of this pool
2684		*/
2685		public int getParallelism() {
2686	<	return parallelism;
2686	>	return config & SMASK;
2687	>	}
2688	>
2689	>	/**
2690	>	* Returns the targeted parallelism level of the common pool.
2691	>	*
2692	>	* @return the targeted parallelism level of the common pool
2693	>	*/
2694	>	public static int getCommonPoolParallelism() {
2695	>	return commonPoolParallelism;
2696		}
2697
2698		/**
#	Line 2349 \| Line 2704 \| public class ForkJoinPool extends Abstra
2704		* @return the number of worker threads
2705		*/
2706		public int getPoolSize() {
2707	<	return parallelism + (short)(ctl >>> TC_SHIFT);
2707	>	return (config & SMASK) + (short)(ctl >>> TC_SHIFT);
2708		}
2709
2710		/**
#	Line 2359 \| Line 2714 \| public class ForkJoinPool extends Abstra
2714		* @return {@code true} if this pool uses async mode
2715		*/
2716		public boolean getAsyncMode() {
2717	<	return localMode != 0;
2717	>	return (config >>> 16) == FIFO_QUEUE;
2718		}
2719
2720		/**
#	Line 2390 \| Line 2745 \| public class ForkJoinPool extends Abstra
2745		* @return the number of active threads
2746		*/
2747		public int getActiveThreadCount() {
2748	<	int r = parallelism + (int)(ctl >> AC_SHIFT);
2748	>	int r = (config & SMASK) + (int)(ctl >> AC_SHIFT);
2749		return (r <= 0) ? 0 : r; // suppress momentarily negative values
2750		}
2751
#	Line 2406 \| Line 2761 \| public class ForkJoinPool extends Abstra
2761		* @return {@code true} if all threads are currently idle
2762		*/
2763		public boolean isQuiescent() {
2764	<	return (int)(ctl >> AC_SHIFT) + parallelism == 0;
2764	>	return (int)(ctl >> AC_SHIFT) + (config & SMASK) == 0;
2765		}
2766
2767		/**
#	Line 2421 \| Line 2776 \| public class ForkJoinPool extends Abstra
2776		* @return the number of steals
2777		*/
2778		public long getStealCount() {
2779	<	long count = stealCount.get();
2779	>	long count = stealCount;
2780		WorkQueue[] ws; WorkQueue w;
2781		if ((ws = workQueues) != null) {
2782		for (int i = 1; i < ws.length; i += 2) {
2783		if ((w = ws[i]) != null)
2784	<	count += w.totalSteals;
2784	>	count += w.nsteals;
2785		}
2786		}
2787		return count;
#	Line 2483 \| Line 2838 \| public class ForkJoinPool extends Abstra
2838		WorkQueue[] ws; WorkQueue w;
2839		if ((ws = workQueues) != null) {
2840		for (int i = 0; i < ws.length; i += 2) {
2841	<	if ((w = ws[i]) != null && !w.isEmpty())
2841	>	if ((w = ws[i]) != null && w.queueSize() != 0)
2842		return true;
2843		}
2844		}
#	Line 2551 \| Line 2906 \| public class ForkJoinPool extends Abstra
2906		public String toString() {
2907		// Use a single pass through workQueues to collect counts
2908		long qt = 0L, qs = 0L; int rc = 0;
2909	<	long st = stealCount.get();
2909	>	long st = stealCount;
2910		long c = ctl;
2911		WorkQueue[] ws; WorkQueue w;
2912		if ((ws = workQueues) != null) {
#	Line 2562 \| Line 2917 \| public class ForkJoinPool extends Abstra
2917		qs += size;
2918		else {
2919		qt += size;
2920	<	st += w.totalSteals;
2920	>	st += w.nsteals;
2921		if (w.isApparentlyUnblocked())
2922		++rc;
2923		}
2924		}
2925		}
2926		}
2927	<	int pc = parallelism;
2927	>	int pc = (config & SMASK);
2928		int tc = pc + (short)(c >>> TC_SHIFT);
2929		int ac = pc + (int)(c >> AC_SHIFT);
2930		if (ac < 0) // ignore transient negative
#	Line 2578 \| Line 2933 \| public class ForkJoinPool extends Abstra
2933		if ((c & STOP_BIT) != 0)
2934		level = (tc == 0) ? "Terminated" : "Terminating";
2935		else
2936	<	level = runState < 0 ? "Shutting down" : "Running";
2936	>	level = plock < 0 ? "Shutting down" : "Running";
2937		return super.toString() +
2938		"[" + level +
2939		", parallelism = " + pc +
#	Line 2592 \| Line 2947 \| public class ForkJoinPool extends Abstra
2947		}
2948
2949		/**
2950	<	* Initiates an orderly shutdown in which previously submitted
2951	<	* tasks are executed, but no new tasks will be accepted.
2952	<	* Invocation has no additional effect if already shut down.
2953	<	* Tasks that are in the process of being submitted concurrently
2954	<	* during the course of this method may or may not be rejected.
2950	>	* Possibly initiates an orderly shutdown in which previously
2951	>	* submitted tasks are executed, but no new tasks will be
2952	>	* accepted. Invocation has no effect on execution state if this
2953	>	* is the {@link #commonPool}, and no additional effect if
2954	>	* already shut down. Tasks that are in the process of being
2955	>	* submitted concurrently during the course of this method may or
2956	>	* may not be rejected.
2957		*
2958		* @throws SecurityException if a security manager exists and
2959		* the caller is not permitted to modify threads
#	Line 2609 \| Line 2966 \| public class ForkJoinPool extends Abstra
2966		}
2967
2968		/**
2969	<	* Attempts to cancel and/or stop all tasks, and reject all
2970	<	* subsequently submitted tasks. Tasks that are in the process of
2971	<	* being submitted or executed concurrently during the course of
2972	<	* this method may or may not be rejected. This method cancels
2973	<	* both existing and unexecuted tasks, in order to permit
2974	<	* termination in the presence of task dependencies. So the method
2975	<	* always returns an empty list (unlike the case for some other
2976	<	* Executors).
2969	>	* Possibly attempts to cancel and/or stop all tasks, and reject
2970	>	* all subsequently submitted tasks. Invocation has no effect on
2971	>	* execution state if this is the {@link #commonPool}, and no
2972	>	* additional effect if already shut down. Otherwise, tasks that
2973	>	* are in the process of being submitted or executed concurrently
2974	>	* during the course of this method may or may not be
2975	>	* rejected. This method cancels both existing and unexecuted
2976	>	* tasks, in order to permit termination in the presence of task
2977	>	* dependencies. So the method always returns an empty list
2978	>	* (unlike the case for some other Executors).
2979		*
2980		* @return an empty list
2981		* @throws SecurityException if a security manager exists and
#	Line 2638 \| Line 2997 \| public class ForkJoinPool extends Abstra
2997		public boolean isTerminated() {
2998		long c = ctl;
2999		return ((c & STOP_BIT) != 0L &&
3000	<	(short)(c >>> TC_SHIFT) == -parallelism);
3000	>	(short)(c >>> TC_SHIFT) == -(config & SMASK));
3001		}
3002
3003		/**
#	Line 2657 \| Line 3016 \| public class ForkJoinPool extends Abstra
3016		public boolean isTerminating() {
3017		long c = ctl;
3018		return ((c & STOP_BIT) != 0L &&
3019	<	(short)(c >>> TC_SHIFT) != -parallelism);
3019	>	(short)(c >>> TC_SHIFT) != -(config & SMASK));
3020		}
3021
3022		/**
#	Line 2666 \| Line 3025 \| public class ForkJoinPool extends Abstra
3025		* @return {@code true} if this pool has been shut down
3026		*/
3027		public boolean isShutdown() {
3028	<	return runState < 0;
3028	>	return plock < 0;
3029		}
3030
3031		/**
3032	<	* Blocks until all tasks have completed execution after a shutdown
3033	<	* request, or the timeout occurs, or the current thread is
3034	<	* interrupted, whichever happens first.
3032	>	* Blocks until all tasks have completed execution after a
3033	>	* shutdown request, or the timeout occurs, or the current thread
3034	>	* is interrupted, whichever happens first. Note that the {@link
3035	>	* #commonPool()} never terminates until program shutdown so
3036	>	* this method will always time out.
3037		*
3038		* @param timeout the maximum time to wait
3039		* @param unit the time unit of the timeout argument
#	Line 2683 \| Line 3044 \| public class ForkJoinPool extends Abstra
3044		public boolean awaitTermination(long timeout, TimeUnit unit)
3045		throws InterruptedException {
3046		long nanos = unit.toNanos(timeout);
3047	<	final Mutex lock = this.lock;
3048	<	lock.lock();
3049	<	try {
3050	<	for (;;) {
3051	<	if (isTerminated())
3052	<	return true;
3053	<	if (nanos <= 0)
3054	<	return false;
3055	<	nanos = termination.awaitNanos(nanos);
3047	>	if (isTerminated())
3048	>	return true;
3049	>	long startTime = System.nanoTime();
3050	>	boolean terminated = false;
3051	>	synchronized (this) {
3052	>	for (long waitTime = nanos, millis = 0L;;) {
3053	>	if (terminated = isTerminated() \|\|
3054	>	waitTime <= 0L \|\|
3055	>	(millis = unit.toMillis(waitTime)) <= 0L)
3056	>	break;
3057	>	wait(millis);
3058	>	waitTime = nanos - (System.nanoTime() - startTime);
3059		}
2696	–	} finally {
2697	–	lock.unlock();
3060		}
3061	+	return terminated;
3062		}
3063
3064		/**
#	Line 2794 \| Line 3157 \| public class ForkJoinPool extends Abstra
3157		public static void managedBlock(ManagedBlocker blocker)
3158		throws InterruptedException {
3159		Thread t = Thread.currentThread();
3160	<	ForkJoinPool p = ((t instanceof ForkJoinWorkerThread) ?
3161	<	((ForkJoinWorkerThread)t).pool : null);
3162	<	while (!blocker.isReleasable()) {
3163	<	if (p == null \|\| p.tryCompensate(null, blocker)) {
3164	<	try {
3165	<	do {} while (!blocker.isReleasable() && !blocker.block());
3166	<	} finally {
3167	<	if (p != null)
3160	>	if (t instanceof ForkJoinWorkerThread) {
3161	>	ForkJoinPool p = ((ForkJoinWorkerThread)t).pool;
3162	>	while (!blocker.isReleasable()) { // variant of helpSignal
3163	>	WorkQueue[] ws; WorkQueue q; int m, n, u;
3164	>	if ((ws = p.workQueues) != null && (m = ws.length - 1) >= 0) {
3165	>	for (int i = 0; i <= m; ++i) {
3166	>	if (blocker.isReleasable())
3167	>	return;
3168	>	if ((q = ws[i]) != null && (n = q.queueSize()) > 0) {
3169	>	p.signalWork(q, n);
3170	>	if ((u = (int)(p.ctl >>> 32)) >= 0 \|\|
3171	>	(u >> UAC_SHIFT) >= 0)
3172	>	break;
3173	>	}
3174	>	}
3175	>	}
3176	>	if (p.tryCompensate()) {
3177	>	try {
3178	>	do {} while (!blocker.isReleasable() &&
3179	>	!blocker.block());
3180	>	} finally {
3181		p.incrementActiveCount();
3182	+	}
3183	+	break;
3184		}
2807	–	break;
3185		}
3186		}
3187	+	else {
3188	+	do {} while (!blocker.isReleasable() &&
3189	+	!blocker.block());
3190	+	}
3191		}
3192
3193		// AbstractExecutorService overrides. These rely on undocumented
#	Line 2827 \| Line 3208 \| public class ForkJoinPool extends Abstra
3208		private static final long PARKBLOCKER;
3209		private static final int ABASE;
3210		private static final int ASHIFT;
3211	+	private static final long STEALCOUNT;
3212	+	private static final long PLOCK;
3213	+	private static final long INDEXSEED;
3214	+	private static final long QLOCK;
3215
3216		static {
3217	<	poolNumberGenerator = new AtomicInteger();
2833	<	nextSubmitterSeed = new AtomicInteger(0x55555555);
2834	<	modifyThreadPermission = new RuntimePermission("modifyThread");
2835	<	defaultForkJoinWorkerThreadFactory =
2836	<	new DefaultForkJoinWorkerThreadFactory();
2837	<	submitters = new ThreadSubmitter();
2838	<	int s;
3217	>	int s; // initialize field offsets for CAS etc
3218		try {
3219		U = getUnsafe();
3220		Class<?> k = ForkJoinPool.class;
2842	–	Class<?> ak = ForkJoinTask[].class;
3221		CTL = U.objectFieldOffset
3222		(k.getDeclaredField("ctl"));
3223	+	STEALCOUNT = U.objectFieldOffset
3224	+	(k.getDeclaredField("stealCount"));
3225	+	PLOCK = U.objectFieldOffset
3226	+	(k.getDeclaredField("plock"));
3227	+	INDEXSEED = U.objectFieldOffset
3228	+	(k.getDeclaredField("indexSeed"));
3229		Class<?> tk = Thread.class;
3230		PARKBLOCKER = U.objectFieldOffset
3231		(tk.getDeclaredField("parkBlocker"));
3232	+	Class<?> wk = WorkQueue.class;
3233	+	QLOCK = U.objectFieldOffset
3234	+	(wk.getDeclaredField("qlock"));
3235	+	Class<?> ak = ForkJoinTask[].class;
3236		ABASE = U.arrayBaseOffset(ak);
3237		s = U.arrayIndexScale(ak);
3238	+	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
3239		} catch (Exception e) {
3240		throw new Error(e);
3241		}
3242		if ((s & (s-1)) != 0)
3243		throw new Error("data type scale not a power of two");
3244	<	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
3244	>
3245	>	submitters = new ThreadLocal<Submitter>();
3246	>	ForkJoinWorkerThreadFactory fac = defaultForkJoinWorkerThreadFactory =
3247	>	new DefaultForkJoinWorkerThreadFactory();
3248	>	/*
3249	>	* Establish common pool parameters. For extra caution,
3250	>	* computations to set up common pool state are here; the
3251	>	* constructor just assigns these values to fields.
3252	>	*/
3253	>
3254	>	int par = 0;
3255	>	Thread.UncaughtExceptionHandler handler = null;
3256	>	try { // TBD: limit or report ignored exceptions?
3257	>	String pp = System.getProperty
3258	>	("java.util.concurrent.ForkJoinPool.common.parallelism");
3259	>	String hp = System.getProperty
3260	>	("java.util.concurrent.ForkJoinPool.common.exceptionHandler");
3261	>	String fp = System.getProperty
3262	>	("java.util.concurrent.ForkJoinPool.common.threadFactory");
3263	>	if (fp != null)
3264	>	fac = ((ForkJoinWorkerThreadFactory)ClassLoader.
3265	>	getSystemClassLoader().loadClass(fp).newInstance());
3266	>	if (hp != null)
3267	>	handler = ((Thread.UncaughtExceptionHandler)ClassLoader.
3268	>	getSystemClassLoader().loadClass(hp).newInstance());
3269	>	if (pp != null)
3270	>	par = Integer.parseInt(pp);
3271	>	} catch (Exception ignore) {
3272	>	}
3273	>
3274	>	if (par <= 0)
3275	>	par = Runtime.getRuntime().availableProcessors();
3276	>	if (par > MAX_CAP)
3277	>	par = MAX_CAP;
3278	>	commonPoolParallelism = par;
3279	>	long np = (long)(-par); // precompute initial ctl value
3280	>	long ct = ((np << AC_SHIFT) & AC_MASK) \| ((np << TC_SHIFT) & TC_MASK);
3281	>
3282	>	commonPool = new ForkJoinPool(par, ct, fac, handler);
3283	>	modifyThreadPermission = new RuntimePermission("modifyThread");
3284		}
3285
3286		/**

Diff Legend

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

Comparing jsr166/src/jsr166y/ForkJoinPool.java (file contents): Revision 1.125 by jsr166, Tue Feb 21 00:19:23 2012 UTC vs. Revision 1.146 by dl, Mon Nov 19 18:12:42 2012 UTC

Diff Legend

Comparing jsr166/src/jsr166y/ForkJoinPool.java (file contents):
Revision 1.125 by jsr166, Tue Feb 21 00:19:23 2012 UTC vs.
Revision 1.146 by dl, Mon Nov 19 18:12:42 2012 UTC