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

Comparing jsr166/src/jsr166y/ForkJoinPool.java (file contents):
Revision 1.112 by dl, Thu Jan 26 18:15:12 2012 UTC vs.
Revision 1.135 by dl, Sun Oct 28 22:36:01 2012 UTC

#	Line 21 \| Line 21 \| import java.util.concurrent.RunnableFutu
21		import java.util.concurrent.TimeUnit;
22		import java.util.concurrent.atomic.AtomicInteger;
23		import java.util.concurrent.atomic.AtomicLong;
24	<	import java.util.concurrent.locks.ReentrantLock;
24	>	import java.util.concurrent.locks.AbstractQueuedSynchronizer;
25		import java.util.concurrent.locks.Condition;
26
27		/**
#	Line 42 \| Line 42 \| import java.util.concurrent.locks.Condit
42		* ForkJoinPool}s may also be appropriate for use with event-style
43		* tasks that are never joined.
44		*
45	<	* <p>A {@code ForkJoinPool} is constructed with a given target
46	<	* parallelism level; by default, equal to the number of available
47	<	* processors. The pool attempts to maintain enough active (or
48	<	* available) threads by dynamically adding, suspending, or resuming
49	<	* internal worker threads, even if some tasks are stalled waiting to
50	<	* join others. However, no such adjustments are guaranteed in the
51	<	* face of blocked IO or other unmanaged synchronization. The nested
52	<	* {@link ManagedBlocker} interface enables extension of the kinds of
45	>	* <p>A static {@link #commonPool} is available and appropriate for
46	>	* most applications. The common pool is constructed upon first
47	>	* access, or upon usage by any ForkJoinTask that is not explictly
48	>	* submitted to a specified pool. Using the common pool normally
49	>	* reduces resource usage (its threads are slowly reclaimed during
50	>	* periods of non-use, and reinstated upon subsequent use). The
51	>	* common pool is by default constructed with default parameters, but
52	>	* these may be controlled by setting any or all of the three
53	>	* properties {@code
54	>	* java.util.concurrent.ForkJoinPool.common.{parallelism,
55	>	* threadFactory, exceptionHandler}}.
56	>	*
57	>	* <p>For applications that require separate or custom pools, a {@code
58	>	* ForkJoinPool} may be constructed with a given target parallelism
59	>	* level; by default, equal to the number of available processors. The
60	>	* pool attempts to maintain enough active (or available) threads by
61	>	* dynamically adding, suspending, or resuming internal worker
62	>	* threads, even if some tasks are stalled waiting to join
63	>	* others. However, no such adjustments are guaranteed in the face of
64	>	* blocked IO or other unmanaged synchronization. The nested {@link
65	>	* ManagedBlocker} interface enables extension of the kinds of
66		* synchronization accommodated.
67		*
68		* <p>In addition to execution and lifecycle control methods, this
#	Line 60 \| Line 73 \| import java.util.concurrent.locks.Condit
73		* convenient form for informal monitoring.
74		*
75		* <p> As is the case with other ExecutorServices, there are three
76	<	* main task execution methods summarized in the following
77	<	* table. These are designed to be used primarily by clients not
78	<	* already engaged in fork/join computations in the current pool. The
79	<	* main forms of these methods accept instances of {@code
80	<	* ForkJoinTask}, but overloaded forms also allow mixed execution of
81	<	* plain {@code Runnable}- or {@code Callable}- based activities as
82	<	* well. However, tasks that are already executing in a pool should
83	<	* normally instead use the within-computation forms listed in the
84	<	* table unless using async event-style tasks that are not usually
85	<	* joined, in which case there is little difference among choice of
73	<	* methods.
76	>	* main task execution methods summarized in the following table.
77	>	* These are designed to be used primarily by clients not already
78	>	* engaged in fork/join computations in the current pool. The main
79	>	* forms of these methods accept instances of {@code ForkJoinTask},
80	>	* but overloaded forms also allow mixed execution of plain {@code
81	>	* Runnable}- or {@code Callable}- based activities as well. However,
82	>	* tasks that are already executing in a pool should normally instead
83	>	* use the within-computation forms listed in the table unless using
84	>	* async event-style tasks that are not usually joined, in which case
85	>	* there is little difference among choice of methods.
86		*
87		* <table BORDER CELLPADDING=3 CELLSPACING=1>
88		* <tr>
#	Line 95 \| Line 107 \| import java.util.concurrent.locks.Condit
107		* </tr>
108		* </table>
109		*
98	–	* <p><b>Sample Usage.</b> Normally a single {@code ForkJoinPool} is
99	–	* used for all parallel task execution in a program or subsystem.
100	–	* Otherwise, use would not usually outweigh the construction and
101	–	* bookkeeping overhead of creating a large set of threads. For
102	–	* example, a common pool could be used for the {@code SortTasks}
103	–	* illustrated in {@link RecursiveAction}. Because {@code
104	–	* ForkJoinPool} uses threads in {@linkplain java.lang.Thread#isDaemon
105	–	* daemon} mode, there is typically no need to explicitly {@link
106	–	* #shutdown} such a pool upon program exit.
107	–	*
108	–	* <pre> {@code
109	–	* static final ForkJoinPool mainPool = new ForkJoinPool();
110	–	* ...
111	–	* public void sort(long[] array) {
112	–	* mainPool.invoke(new SortTask(array, 0, array.length));
113	–	* }}</pre>
114	–	*
110		* <p><b>Implementation notes</b>: This implementation restricts the
111		* maximum number of running threads to 32767. Attempts to create
112		* pools with greater than the maximum number result in
#	Line 131 \| Line 126 \| public class ForkJoinPool extends Abstra
126		*
127		* This class and its nested classes provide the main
128		* functionality and control for a set of worker threads:
129	<	* Submissions from non-FJ threads enter into submission
130	<	* queues. Workers take these tasks and typically split them into
131	<	* subtasks that may be stolen by other workers. Preference rules
132	<	* give first priority to processing tasks from their own queues
133	<	* (LIFO or FIFO, depending on mode), then to randomized FIFO
134	<	* steals of tasks in other queues.
129	>	* Submissions from non-FJ threads enter into submission queues.
130	>	* Workers take these tasks and typically split them into subtasks
131	>	* that may be stolen by other workers. Preference rules give
132	>	* first priority to processing tasks from their own queues (LIFO
133	>	* or FIFO, depending on mode), then to randomized FIFO steals of
134	>	* tasks in other queues.
135		*
136	<	* WorkQueues.
136	>	* WorkQueues
137		* ==========
138		*
139		* Most operations occur within work-stealing queues (in nested
#	Line 156 \| Line 151 \| public class ForkJoinPool extends Abstra
151		* (http://research.sun.com/scalable/pubs/index.html) and
152		* "Idempotent work stealing" by Michael, Saraswat, and Vechev,
153		* PPoPP 2009 (http://portal.acm.org/citation.cfm?id=1504186).
154	<	* The main differences ultimately stem from gc requirements that
154	>	* The main differences ultimately stem from GC requirements that
155		* we null out taken slots as soon as we can, to maintain as small
156		* a footprint as possible even in programs generating huge
157		* numbers of tasks. To accomplish this, we shift the CAS
#	Line 178 \| Line 173 \| public class ForkJoinPool extends Abstra
173		* If an attempted steal fails, a thief always chooses a different
174		* random victim target to try next. So, in order for one thief to
175		* progress, it suffices for any in-progress poll or new push on
176	<	* any empty queue to complete.
176	>	* any empty queue to complete. (This is why we normally use
177	>	* method pollAt and its variants that try once at the apparent
178	>	* base index, else consider alternative actions, rather than
179	>	* method poll.)
180		*
181		* This approach also enables support of a user mode in which local
182		* task processing is in FIFO, not LIFO order, simply by using
#	Line 188 \| Line 186 \| public class ForkJoinPool extends Abstra
186		* rarely provide the best possible performance on a given
187		* machine, but portably provide good throughput by averaging over
188		* these factors. (Further, even if we did try to use such
189	<	* information, we do not usually have a basis for exploiting
190	<	* it. For example, some sets of tasks profit from cache
191	<	* affinities, but others are harmed by cache pollution effects.)
189	>	* information, we do not usually have a basis for exploiting it.
190	>	* For example, some sets of tasks profit from cache affinities,
191	>	* but others are harmed by cache pollution effects.)
192		*
193		* WorkQueues are also used in a similar way for tasks submitted
194		* to the pool. We cannot mix these tasks in the same queues used
195		* for work-stealing (this would contaminate lifo/fifo
196	<	* processing). Instead, we loosely associate (via hashing)
197	<	* submission queues with submitting threads, and randomly scan
198	<	* these queues as well when looking for work. In essence,
199	<	* submitters act like workers except that they never take tasks,
200	<	* and they are multiplexed on to a finite number of shared work
201	<	* queues. However, classes are set up so that future extensions
202	<	* could allow submitters to optionally help perform tasks as
203	<	* well. Pool submissions from internal workers are also allowed,
204	<	* but use randomized rather than thread-hashed queue indices to
205	<	* avoid imbalance. Insertion of tasks in shared mode requires a
196	>	* processing). Instead, we loosely associate submission queues
197	>	* with submitting threads, using a form of hashing. The
198	>	* ThreadLocal Submitter class contains a value initially used as
199	>	* a hash code for choosing existing queues, but may be randomly
200	>	* repositioned upon contention with other submitters. In
201	>	* essence, submitters act like workers except that they never
202	>	* take tasks, and they are multiplexed on to a finite number of
203	>	* shared work queues. However, classes are set up so that future
204	>	* extensions could allow submitters to optionally help perform
205	>	* tasks as well. Insertion of tasks in shared mode requires a
206		* lock (mainly to protect in the case of resizing) but we use
207		* only a simple spinlock (using bits in field runState), because
208	<	* submitters encountering a busy queue try or create others so
209	<	* never block.
208	>	* submitters encountering a busy queue move on to try or create
209	>	* other queues -- they block only when creating and registering
210	>	* new queues.
211		*
212	<	* Management.
212	>	* Management
213		* ==========
214		*
215		* The main throughput advantages of work-stealing stem from
#	Line 220 \| Line 219 \| public class ForkJoinPool extends Abstra
219		* tactic for avoiding bottlenecks is packing nearly all
220		* essentially atomic control state into two volatile variables
221		* that are by far most often read (not written) as status and
222	<	* consistency checks
222	>	* consistency checks.
223		*
224		* Field "ctl" contains 64 bits holding all the information needed
225		* to atomically decide to add, inactivate, enqueue (on an event
#	Line 246 \| Line 245 \| public class ForkJoinPool extends Abstra
245		* readers must tolerate null slots. Shared (submission) queues
246		* are at even indices, worker queues at odd indices. Grouping
247		* them together in this way simplifies and speeds up task
248	<	* scanning. To avoid flailing during start-up, the array is
249	<	* presized to hold twice #parallelism workers (which is unlikely
250	<	* to need further resizing during execution). But to avoid
251	<	* dealing with so many null slots, variable runState includes a
253	<	* mask for the nearest power of two that contains all current
254	<	* workers. All worker thread creation is on-demand, triggered by
255	<	* task submissions, replacement of terminated workers, and/or
248	>	* scanning.
249	>	*
250	>	* All worker thread creation is on-demand, triggered by task
251	>	* submissions, replacement of terminated workers, and/or
252		* compensation for blocked workers. However, all other support
253		* code is set up to work with other policies. To ensure that we
254		* do not hold on to worker references that would prevent GC, ALL
#	Line 265 \| Line 261 \| public class ForkJoinPool extends Abstra
261		* both index-check and null-check the IDs. All such accesses
262		* ignore bad IDs by returning out early from what they are doing,
263		* since this can only be associated with termination, in which
264	<	* case it is OK to give up.
265	<	*
266	<	* All uses of the workQueues array check that it is non-null
267	<	* (even if previously non-null). This allows nulling during
268	<	* termination, which is currently not necessary, but remains an
269	<	* option for resource-revocation-based shutdown schemes. It also
274	<	* helps reduce JIT issuance of uncommon-trap code, which tends to
264	>	* case it is OK to give up. All uses of the workQueues array
265	>	* also check that it is non-null (even if previously
266	>	* non-null). This allows nulling during termination, which is
267	>	* currently not necessary, but remains an option for
268	>	* resource-revocation-based shutdown schemes. It also helps
269	>	* reduce JIT issuance of uncommon-trap code, which tends to
270		* unnecessarily complicate control flow in some methods.
271		*
272		* Event Queuing. Unlike HPC work-stealing frameworks, we cannot
#	Line 299 \| Line 294 \| public class ForkJoinPool extends Abstra
294		* some other queued worker rather than itself, which has the same
295		* net effect. Because enqueued workers may actually be rescanning
296		* rather than waiting, we set and clear the "parker" field of
297	<	* Workqueues to reduce unnecessary calls to unpark. (This
297	>	* WorkQueues to reduce unnecessary calls to unpark. (This
298		* requires a secondary recheck to avoid missed signals.) Note
299		* the unusual conventions about Thread.interrupts surrounding
300		* parking and other blocking: Because interrupts are used solely
#	Line 322 \| Line 317 \| public class ForkJoinPool extends Abstra
317		*
318		* Trimming workers. To release resources after periods of lack of
319		* use, a worker starting to wait when the pool is quiescent will
320	<	* time out and terminate if the pool has remained quiescent for
321	<	* SHRINK_RATE nanosecs. This will slowly propagate, eventually
322	<	* terminating all workers after long periods of non-use.
320	>	* time out and terminate if the pool has remained quiescent for a
321	>	* given period -- a short period if there are more threads than
322	>	* parallelism, longer as the number of threads decreases. This
323	>	* will slowly propagate, eventually terminating all workers after
324	>	* periods of non-use.
325		*
326		* Shutdown and Termination. A call to shutdownNow atomically sets
327	<	* a runState bit and then (non-atomically) sets each workers
327	>	* a runState bit and then (non-atomically) sets each worker's
328		* runState status, cancels all unprocessed tasks, and wakes up
329		* all waiting workers. Detecting whether termination should
330		* commence after a non-abrupt shutdown() call requires more work
#	Line 336 \| Line 333 \| public class ForkJoinPool extends Abstra
333		* indication but non-abrupt shutdown still requires a rechecking
334		* scan for any workers that are inactive but not queued.
335		*
336	<	* Joining Tasks.
337	<	* ==============
336	>	* Joining Tasks
337	>	* =============
338		*
339		* Any of several actions may be taken when one worker is waiting
340	<	* to join a task stolen (or always held by) another. Because we
340	>	* to join a task stolen (or always held) by another. Because we
341		* are multiplexing many tasks on to a pool of workers, we can't
342		* just let them block (as in Thread.join). We also cannot just
343		* reassign the joiner's run-time stack with another and replace
344		* it later, which would be a form of "continuation", that even if
345		* possible is not necessarily a good idea since we sometimes need
346	<	* both an unblocked task and its continuation to
347	<	* progress. Instead we combine two tactics:
346	>	* both an unblocked task and its continuation to progress.
347	>	* Instead we combine two tactics:
348		*
349		* Helping: Arranging for the joiner to execute some task that it
350		* would be running if the steal had not occurred.
#	Line 382 \| Line 379 \| public class ForkJoinPool extends Abstra
379		* (http://portal.acm.org/citation.cfm?id=155354). It differs in
380		* that: (1) We only maintain dependency links across workers upon
381		* steals, rather than use per-task bookkeeping. This sometimes
382	<	* requires a linear scan of workers array to locate stealers, but
383	<	* often doesn't because stealers leave hints (that may become
382	>	* requires a linear scan of workQueues array to locate stealers,
383	>	* but often doesn't because stealers leave hints (that may become
384		* stale/wrong) of where to locate them. A stealHint is only a
385		* hint because a worker might have had multiple steals and the
386		* hint records only one of them (usually the most current).
#	Line 394 \| Line 391 \| public class ForkJoinPool extends Abstra
391		* which means that we miss links in the chain during long-lived
392		* tasks, GC stalls etc (which is OK since blocking in such cases
393		* is usually a good idea). (4) We bound the number of attempts
394	<	* to find work (see MAX_HELP_DEPTH) and fall back to suspending
395	<	* the worker and if necessary replacing it with another.
394	>	* to find work (see MAX_HELP) and fall back to suspending the
395	>	* worker and if necessary replacing it with another.
396		*
397		* It is impossible to keep exactly the target parallelism number
398		* of threads running at any given time. Determining the
399		* existence of conservatively safe helping targets, the
400		* availability of already-created spares, and the apparent need
401		* to create new spares are all racy, so we rely on multiple
402	<	* retries of each. Currently, in keeping with on-demand
403	<	* signalling policy, we compensate only if blocking would leave
404	<	* less than one active (non-waiting, non-blocked) worker.
405	<	* Additionally, to avoid some false alarms due to GC, lagging
406	<	* counters, system activity, etc, compensated blocking for joins
407	<	* is only attempted after rechecks stabilize in
408	<	* ForkJoinTask.awaitJoin. (Retries are interspersed with
409	<	* Thread.yield, for good citizenship.)
402	>	* retries of each. Compensation in the apparent absence of
403	>	* helping opportunities is challenging to control on JVMs, where
404	>	* GC and other activities can stall progress of tasks that in
405	>	* turn stall out many other dependent tasks, without us being
406	>	* able to determine whether they will ever require compensation.
407	>	* Even though work-stealing otherwise encounters little
408	>	* degradation in the presence of more threads than cores,
409	>	* aggressively adding new threads in such cases entails risk of
410	>	* unwanted positive feedback control loops in which more threads
411	>	* cause more dependent stalls (as well as delayed progress of
412	>	* unblocked threads to the point that we know they are available)
413	>	* leading to more situations requiring more threads, and so
414	>	* on. This aspect of control can be seen as an (analytically
415	>	* intractable) game with an opponent that may choose the worst
416	>	* (for us) active thread to stall at any time. We take several
417	>	* precautions to bound losses (and thus bound gains), mainly in
418	>	* methods tryCompensate and awaitJoin: (1) We only try
419	>	* compensation after attempting enough helping steps (measured
420	>	* via counting and timing) that we have already consumed the
421	>	* estimated cost of creating and activating a new thread. (2) We
422	>	* allow up to 50% of threads to be blocked before initially
423	>	* adding any others, and unless completely saturated, check that
424	>	* some work is available for a new worker before adding. Also, we
425	>	* create up to only 50% more threads until entering a mode that
426	>	* only adds a thread if all others are possibly blocked. All
427	>	* together, this means that we might be half as fast to react,
428	>	* and create half as many threads as possible in the ideal case,
429	>	* but present vastly fewer anomalies in all other cases compared
430	>	* to both more aggressive and more conservative alternatives.
431		*
432		* Style notes: There is a lot of representation-level coupling
433		* among classes ForkJoinPool, ForkJoinWorkerThread, and
#	Line 417 \| Line 435 \| public class ForkJoinPool extends Abstra
435		* managed by ForkJoinPool, so are directly accessed. There is
436		* little point trying to reduce this, since any associated future
437		* changes in representations will need to be accompanied by
438	<	* algorithmic changes anyway. All together, these low-level
439	<	* implementation choices produce as much as a factor of 4
440	<	* performance improvement compared to naive implementations, and
441	<	* enable the processing of billions of tasks per second, at the
442	<	* expense of some ugliness.
443	<	*
444	<	* Methods signalWork() and scan() are the main bottlenecks so are
445	<	* especially heavily micro-optimized/mangled. There are lots of
446	<	* inline assignments (of form "while ((local = field) != 0)")
447	<	* which are usually the simplest way to ensure the required read
448	<	* orderings (which are sometimes critical). This leads to a
449	<	* "C"-like style of listing declarations of these locals at the
450	<	* heads of methods or blocks. There are several occurrences of
451	<	* the unusual "do {} while (!cas...)" which is the simplest way
452	<	* to force an update of a CAS'ed variable. There are also other
453	<	* coding oddities that help some methods perform reasonably even
454	<	* when interpreted (not compiled).
455	<	*
456	<	* The order of declarations in this file is: (1) declarations of
457	<	* statics (2) fields (along with constants used when unpacking
458	<	* some of them), listed in an order that tends to reduce
459	<	* contention among them a bit under most JVMs; (3) nested
460	<	* classes; (4) internal control methods; (5) callbacks and other
461	<	* support for ForkJoinTask methods; (6) exported methods (plus a
462	<	* few little helpers); (7) static block initializing all statics
463	<	* in a minimally dependent order.
438	>	* algorithmic changes anyway. Several methods intrinsically
439	>	* sprawl because they must accumulate sets of consistent reads of
440	>	* volatiles held in local variables. Methods signalWork() and
441	>	* scan() are the main bottlenecks, so are especially heavily
442	>	* micro-optimized/mangled. There are lots of inline assignments
443	>	* (of form "while ((local = field) != 0)") which are usually the
444	>	* simplest way to ensure the required read orderings (which are
445	>	* sometimes critical). This leads to a "C"-like style of listing
446	>	* declarations of these locals at the heads of methods or blocks.
447	>	* There are several occurrences of the unusual "do {} while
448	>	* (!cas...)" which is the simplest way to force an update of a
449	>	* CAS'ed variable. There are also other coding oddities that help
450	>	* some methods perform reasonably even when interpreted (not
451	>	* compiled).
452	>	*
453	>	* The order of declarations in this file is:
454	>	* (1) Static utility functions
455	>	* (2) Nested (static) classes
456	>	* (3) Static fields
457	>	* (4) Fields, along with constants used when unpacking some of them
458	>	* (5) Internal control methods
459	>	* (6) Callbacks and other support for ForkJoinTask methods
460	>	* (7) Exported methods
461	>	* (8) Static block initializing statics in minimally dependent order
462	>	*/
463	>
464	>	// Static utilities
465	>
466	>	/**
467	>	* If there is a security manager, makes sure caller has
468	>	* permission to modify threads.
469		*/
470	+	private static void checkPermission() {
471	+	SecurityManager security = System.getSecurityManager();
472	+	if (security != null)
473	+	security.checkPermission(modifyThreadPermission);
474	+	}
475	+
476	+	// Nested classes
477
478		/**
479		* Factory for creating new {@link ForkJoinWorkerThread}s.
#	Line 473 \| Line 503 \| public class ForkJoinPool extends Abstra
503		}
504
505		/**
506	<	* Creates a new ForkJoinWorkerThread. This factory is used unless
507	<	* overridden in ForkJoinPool constructors.
508	<	*/
509	<	public static final ForkJoinWorkerThreadFactory
510	<	defaultForkJoinWorkerThreadFactory;
511	<
512	<	/**
513	<	* Permission required for callers of methods that may start or
514	<	* kill threads.
515	<	*/
516	<	private static final RuntimePermission modifyThreadPermission;
517	<
518	<	/**
519	<	* If there is a security manager, makes sure caller has
520	<	* permission to modify threads.
521	<	*/
522	<	private static void checkPermission() {
523	<	SecurityManager security = System.getSecurityManager();
524	<	if (security != null)
525	<	security.checkPermission(modifyThreadPermission);
506	>	* A simple non-reentrant lock used for exclusion when managing
507	>	* queues and workers. We use a custom lock so that we can readily
508	>	* probe lock state in constructions that check among alternative
509	>	* actions. The lock is normally only very briefly held, and
510	>	* sometimes treated as a spinlock, but other usages block to
511	>	* reduce overall contention in those cases where locked code
512	>	* bodies perform allocation/resizing.
513	>	*/
514	>	static final class Mutex extends AbstractQueuedSynchronizer {
515	>	public final boolean tryAcquire(int ignore) {
516	>	return compareAndSetState(0, 1);
517	>	}
518	>	public final boolean tryRelease(int ignore) {
519	>	setState(0);
520	>	return true;
521	>	}
522	>	public final void lock() { acquire(0); }
523	>	public final void unlock() { release(0); }
524	>	public final boolean isHeldExclusively() { return getState() == 1; }
525	>	public final Condition newCondition() { return new ConditionObject(); }
526		}
527
528		/**
529	<	* Generator for assigning sequence numbers as pool names.
530	<	*/
531	<	private static final AtomicInteger poolNumberGenerator;
532	<
503	<	/**
504	<	* Bits and masks for control variables
505	<	*
506	<	* Field ctl is a long packed with:
507	<	* AC: Number of active running workers minus target parallelism (16 bits)
508	<	* TC: Number of total workers minus target parallelism (16 bits)
509	<	* ST: true if pool is terminating (1 bit)
510	<	* EC: the wait count of top waiting thread (15 bits)
511	<	* ID: ~(poolIndex >>> 1) of top of Treiber stack of waiters (16 bits)
512	<	*
513	<	* When convenient, we can extract the upper 32 bits of counts and
514	<	* the lower 32 bits of queue state, u = (int)(ctl >>> 32) and e =
515	<	* (int)ctl. The ec field is never accessed alone, but always
516	<	* together with id and st. The offsets of counts by the target
517	<	* parallelism and the positionings of fields makes it possible to
518	<	* perform the most common checks via sign tests of fields: When
519	<	* ac is negative, there are not enough active workers, when tc is
520	<	* negative, there are not enough total workers, when id is
521	<	* negative, there is at least one waiting worker, and when e is
522	<	* negative, the pool is terminating. To deal with these possibly
523	<	* negative fields, we use casts in and out of "short" and/or
524	<	* signed shifts to maintain signedness.
525	<	*
526	<	* When a thread is queued (inactivated), its eventCount field is
527	<	* negative, which is the only way to tell if a worker is
528	<	* prevented from executing tasks, even though it must continue to
529	<	* scan for them to avoid queuing races.
530	<	*
531	<	* Field runState is an int packed with:
532	<	* SHUTDOWN: true if shutdown is enabled (1 bit)
533	<	* SEQ: a sequence number updated upon (de)registering workers (15 bits)
534	<	* MASK: mask (power of 2 - 1) covering all registered poolIndexes (16 bits)
535	<	*
536	<	* The combination of mask and sequence number enables simple
537	<	* consistency checks: Staleness of read-only operations on the
538	<	* workers and queues arrays can be checked by comparing runState
539	<	* before vs after the reads. The low 16 bits (i.e, anding with
540	<	* SMASK) hold (the smallest power of two covering all worker
541	<	* indices, minus one. The mask for queues (vs workers) is twice
542	<	* this value plus 1.
543	<	*/
544	<
545	<	// bit positions/shifts for fields
546	<	private static final int AC_SHIFT = 48;
547	<	private static final int TC_SHIFT = 32;
548	<	private static final int ST_SHIFT = 31;
549	<	private static final int EC_SHIFT = 16;
550	<
551	<	// bounds
552	<	private static final int MAX_ID = 0x7fff; // max poolIndex
553	<	private static final int SMASK = 0xffff; // mask short bits
554	<	private static final int SHORT_SIGN = 1 << 15;
555	<	private static final int INT_SIGN = 1 << 31;
556	<
557	<	// masks
558	<	private static final long STOP_BIT = 0x0001L << ST_SHIFT;
559	<	private static final long AC_MASK = ((long)SMASK) << AC_SHIFT;
560	<	private static final long TC_MASK = ((long)SMASK) << TC_SHIFT;
561	<
562	<	// units for incrementing and decrementing
563	<	private static final long TC_UNIT = 1L << TC_SHIFT;
564	<	private static final long AC_UNIT = 1L << AC_SHIFT;
565	<
566	<	// masks and units for dealing with u = (int)(ctl >>> 32)
567	<	private static final int UAC_SHIFT = AC_SHIFT - 32;
568	<	private static final int UTC_SHIFT = TC_SHIFT - 32;
569	<	private static final int UAC_MASK = SMASK << UAC_SHIFT;
570	<	private static final int UTC_MASK = SMASK << UTC_SHIFT;
571	<	private static final int UAC_UNIT = 1 << UAC_SHIFT;
572	<	private static final int UTC_UNIT = 1 << UTC_SHIFT;
573	<
574	<	// masks and units for dealing with e = (int)ctl
575	<	private static final int E_MASK = 0x7fffffff; // no STOP_BIT
576	<	private static final int E_SEQ = 1 << EC_SHIFT;
577	<
578	<	// runState bits
579	<	private static final int SHUTDOWN = 1 << 31;
580	<	private static final int RS_SEQ = 1 << 16;
581	<	private static final int RS_SEQ_MASK = 0x7fff0000;
582	<
583	<	// access mode for WorkQueue
584	<	static final int LIFO_QUEUE = 0;
585	<	static final int FIFO_QUEUE = 1;
586	<	static final int SHARED_QUEUE = -1;
587	<
588	<	/**
589	<	* The wakeup interval (in nanoseconds) for a worker waiting for a
590	<	* task when the pool is quiescent to instead try to shrink the
591	<	* number of workers. The exact value does not matter too
592	<	* much. It must be short enough to release resources during
593	<	* sustained periods of idleness, but not so short that threads
594	<	* are continually re-created.
595	<	*/
596	<	private static final long SHRINK_RATE =
597	<	4L * 1000L * 1000L * 1000L; // 4 seconds
598	<
599	<	/**
600	<	* The timeout value for attempted shrinkage, includes
601	<	* some slop to cope with system timer imprecision.
602	<	*/
603	<	private static final long SHRINK_TIMEOUT = SHRINK_RATE - (SHRINK_RATE / 10);
604	<
605	<	/**
606	<	* The maximum stolen->joining link depth allowed in tryHelpStealer.
607	<	* Depths for legitimate chains are unbounded, but we use a fixed
608	<	* constant to avoid (otherwise unchecked) cycles and to bound
609	<	* staleness of traversal parameters at the expense of sometimes
610	<	* blocking when we could be helping.
611	<	*/
612	<	private static final int MAX_HELP_DEPTH = 16;
613	<
614	<	/*
615	<	* Field layout order in this class tends to matter more than one
616	<	* would like. Runtime layout order is only loosely related to
617	<	* declaration order and may differ across JVMs, but the following
618	<	* empirically works OK on current JVMs.
529	>	* Class for artificial tasks that are used to replace the target
530	>	* of local joins if they are removed from an interior queue slot
531	>	* in WorkQueue.tryRemoveAndExec. We don't need the proxy to
532	>	* actually do anything beyond having a unique identity.
533		*/
534	<
535	<	volatile long ctl; // main pool control
536	<	final int parallelism; // parallelism level
537	<	final int localMode; // per-worker scheduling mode
538	<	int nextPoolIndex; // hint used in registerWorker
539	<	volatile int runState; // shutdown status, seq, and mask
626	<	WorkQueue[] workQueues; // main registry
627	<	final ReentrantLock lock; // for registration
628	<	final Condition termination; // for awaitTermination
629	<	final ForkJoinWorkerThreadFactory factory; // factory for new workers
630	<	final Thread.UncaughtExceptionHandler ueh; // per-worker UEH
631	<	final AtomicLong stealCount; // collect counts when terminated
632	<	final AtomicInteger nextWorkerNumber; // to create worker name string
633	<	final String workerNamePrefix; // Prefix for assigning worker names
534	>	static final class EmptyTask extends ForkJoinTask<Void> {
535	>	EmptyTask() { status = ForkJoinTask.NORMAL; } // force done
536	>	public final Void getRawResult() { return null; }
537	>	public final void setRawResult(Void x) {}
538	>	public final boolean exec() { return true; }
539	>	}
540
541		/**
542		* Queues supporting work-stealing as well as external task
#	Line 649 \| Line 555 \| public class ForkJoinPool extends Abstra
555		* for push, or under lock for trySharedPush, and accessed by
556		* other threads only after reading (volatile) base. Both top and
557		* base are allowed to wrap around on overflow, but (top - base)
558	<	* (or more comonly -(base - top) to force volatile read of base
558	>	* (or more commonly -(base - top) to force volatile read of base
559		* before top) still estimates size.
560		*
561		* The array slots are read and written using the emulation of
#	Line 681 \| Line 587 \| public class ForkJoinPool extends Abstra
587		* avoiding really bad worst-case access. (Until better JVM
588		* support is in place, this padding is dependent on transient
589		* properties of JVM field layout rules.) We also take care in
590	<	* allocating and sizing and resizing the array. Non-shared queue
590	>	* allocating, sizing and resizing the array. Non-shared queue
591		* arrays are initialized (via method growArray) by workers before
592		* use. Others are allocated on first use.
593		*/
594		static final class WorkQueue {
595		/**
596		* Capacity of work-stealing queue array upon initialization.
597	<	* Must be a power of two; at least 4, but set larger to
598	<	* reduce cacheline sharing among queues.
597	>	* Must be a power of two; at least 4, but should be larger to
598	>	* reduce or eliminate cacheline sharing among queues.
599	>	* Currently, it is much larger, as a partial workaround for
600	>	* the fact that JVMs often place arrays in locations that
601	>	* share GC bookkeeping (especially cardmarks) such that
602	>	* per-write accesses encounter serious memory contention.
603		*/
604	<	static final int INITIAL_QUEUE_CAPACITY = 1 << 8;
604	>	static final int INITIAL_QUEUE_CAPACITY = 1 << 13;
605
606		/**
607		* Maximum size for queue arrays. Must be a power of two less
#	Line 715 \| Line 625 \| public class ForkJoinPool extends Abstra
625		volatile int base; // index of next slot for poll
626		int top; // index of next slot for push
627		ForkJoinTask<?>[] array; // the elements (initially unallocated)
628	+	final ForkJoinPool pool; // the containing pool (may be null)
629		final ForkJoinWorkerThread owner; // owning thread or null if shared
630		volatile Thread parker; // == owner during call to park; else null
631	<	ForkJoinTask<?> currentJoin; // task being joined in awaitJoin
631	>	volatile ForkJoinTask<?> currentJoin; // task being joined in awaitJoin
632		ForkJoinTask<?> currentSteal; // current non-local task being executed
633		// Heuristic padding to ameliorate unfortunate memory placements
634	<	Object p00, p01, p02, p03, p04, p05, p06, p07, p08, p09, p0a;
634	>	Object p00, p01, p02, p03, p04, p05, p06, p07;
635	>	Object p08, p09, p0a, p0b, p0c, p0d, p0e;
636
637	<	WorkQueue(ForkJoinWorkerThread owner, int mode) {
726	<	this.owner = owner;
637	>	WorkQueue(ForkJoinPool pool, ForkJoinWorkerThread owner, int mode) {
638		this.mode = mode;
639	+	this.pool = pool;
640	+	this.owner = owner;
641		// Place indices in the center of array (that is not yet allocated)
642		base = top = INITIAL_QUEUE_CAPACITY >>> 1;
643		}
644
645		/**
646	<	* Returns number of tasks in the queue
646	>	* Returns the approximate number of tasks in the queue.
647		*/
648		final int queueSize() {
649	<	int n = base - top; // non-owner callers must read base first
650	<	return (n >= 0) ? 0 : -n;
649	>	int n = base - top; // non-owner callers must read base first
650	>	return (n >= 0) ? 0 : -n; // ignore transient negative
651	>	}
652	>
653	>	/**
654	>	* Provides a more accurate estimate of whether this queue has
655	>	* any tasks than does queueSize, by checking whether a
656	>	* near-empty queue has at least one unclaimed task.
657	>	*/
658	>	final boolean isEmpty() {
659	>	ForkJoinTask<?>[] a; int m, s;
660	>	int n = base - (s = top);
661	>	return (n >= 0 \|\|
662	>	(n == -1 &&
663	>	((a = array) == null \|\|
664	>	(m = a.length - 1) < 0 \|\|
665	>	U.getObjectVolatile
666	>	(a, ((m & (s - 1)) << ASHIFT) + ABASE) == null)));
667		}
668
669		/**
670		* Pushes a task. Call only by owner in unshared queues.
671		*
672		* @param task the task. Caller must ensure non-null.
673	<	* @param p, if non-null, pool to signal if necessary
745	<	* @throw RejectedExecutionException if array cannot
746	<	* be resized
673	>	* @throw RejectedExecutionException if array cannot be resized
674		*/
675	<	final void push(ForkJoinTask<?> task, ForkJoinPool p) {
676	<	ForkJoinTask<?>[] a;
675	>	final void push(ForkJoinTask<?> task) {
676	>	ForkJoinTask<?>[] a; ForkJoinPool p;
677		int s = top, m, n;
678		if ((a = array) != null) { // ignore if queue removed
679		U.putOrderedObject
680		(a, (((m = a.length - 1) & s) << ASHIFT) + ABASE, task);
681		if ((n = (top = s + 1) - base) <= 2) {
682	<	if (p != null)
682	>	if ((p = pool) != null)
683		p.signalWork();
684		}
685		else if (n >= m)
#	Line 771 \| Line 698 \| public class ForkJoinPool extends Abstra
698		boolean submitted = false;
699		if (runState == 0 && U.compareAndSwapInt(this, RUNSTATE, 0, 1)) {
700		ForkJoinTask<?>[] a = array;
701	<	int s = top, n = s - base;
701	>	int s = top;
702		try {
703	<	if ((a != null && n < a.length - 1) \|\|
703	>	if ((a != null && a.length > s + 1 - base) \|\|
704		(a = growArray(false)) != null) { // must presize
705		int j = (((a.length - 1) & s) << ASHIFT) + ABASE;
706		U.putObject(a, (long)j, task); // don't need "ordered"
#	Line 788 \| Line 715 \| public class ForkJoinPool extends Abstra
715		}
716
717		/**
718	<	* Takes next task, if one exists, in FIFO order.
718	>	* Takes next task, if one exists, in LIFO order. Call only
719	>	* by owner in unshared queues. (We do not have a shared
720	>	* version of this method because it is never needed.)
721		*/
722	<	final ForkJoinTask<?> poll() {
723	<	ForkJoinTask<?>[] a; int b, i;
724	<	while ((b = base) - top < 0 && (a = array) != null &&
725	<	(i = (a.length - 1) & b) >= 0) {
726	<	int j = (i << ASHIFT) + ABASE;
727	<	ForkJoinTask<?> t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
728	<	if (t != null && base == b &&
722	>	final ForkJoinTask<?> pop() {
723	>	ForkJoinTask<?>[] a; ForkJoinTask<?> t; int m;
724	>	if ((a = array) != null && (m = a.length - 1) >= 0) {
725	>	for (int s; (s = top - 1) - base >= 0;) {
726	>	long j = ((m & s) << ASHIFT) + ABASE;
727	>	if ((t = (ForkJoinTask<?>)U.getObject(a, j)) == null)
728	>	break;
729	>	if (U.compareAndSwapObject(a, j, t, null)) {
730	>	top = s;
731	>	return t;
732	>	}
733	>	}
734	>	}
735	>	return null;
736	>	}
737	>
738	>	/**
739	>	* Takes a task in FIFO order if b is base of queue and a task
740	>	* can be claimed without contention. Specialized versions
741	>	* appear in ForkJoinPool methods scan and tryHelpStealer.
742	>	*/
743	>	final ForkJoinTask<?> pollAt(int b) {
744	>	ForkJoinTask<?> t; ForkJoinTask<?>[] a;
745	>	if ((a = array) != null) {
746	>	int j = (((a.length - 1) & b) << ASHIFT) + ABASE;
747	>	if ((t = (ForkJoinTask<?>)U.getObjectVolatile(a, j)) != null &&
748	>	base == b &&
749		U.compareAndSwapObject(a, j, t, null)) {
750		base = b + 1;
751		return t;
#	Line 806 \| Line 755 \| public class ForkJoinPool extends Abstra
755		}
756
757		/**
758	<	* Takes next task, if one exists, in LIFO order.
810	<	* Call only by owner in unshared queues.
758	>	* Takes next task, if one exists, in FIFO order.
759		*/
760	<	final ForkJoinTask<?> pop() {
761	<	ForkJoinTask<?> t; int m;
762	<	ForkJoinTask<?>[] a = array;
763	<	if (a != null && (m = a.length - 1) >= 0) {
764	<	for (int s; (s = top - 1) - base >= 0;) {
765	<	int j = ((m & s) << ASHIFT) + ABASE;
766	<	if ((t = (ForkJoinTask<?>)U.getObjectVolatile(a, j)) == null)
767	<	break;
768	<	if (U.compareAndSwapObject(a, j, t, null)) {
821	<	top = s;
760	>	final ForkJoinTask<?> poll() {
761	>	ForkJoinTask<?>[] a; int b; ForkJoinTask<?> t;
762	>	while ((b = base) - top < 0 && (a = array) != null) {
763	>	int j = (((a.length - 1) & b) << ASHIFT) + ABASE;
764	>	t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
765	>	if (t != null) {
766	>	if (base == b &&
767	>	U.compareAndSwapObject(a, j, t, null)) {
768	>	base = b + 1;
769		return t;
770		}
771		}
772	+	else if (base == b) {
773	+	if (b + 1 == top)
774	+	break;
775	+	Thread.yield(); // wait for lagging update
776	+	}
777		}
778		return null;
779		}
#	Line 846 \| Line 798 \| public class ForkJoinPool extends Abstra
798		}
799
800		/**
849	–	* Returns task at index b if b is current base of queue.
850	–	*/
851	–	final ForkJoinTask<?> pollAt(int b) {
852	–	ForkJoinTask<?>[] a; int i;
853	–	ForkJoinTask<?> task = null;
854	–	if ((a = array) != null && (i = ((a.length - 1) & b)) >= 0) {
855	–	int j = (i << ASHIFT) + ABASE;
856	–	ForkJoinTask<?> t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
857	–	if (t != null && base == b &&
858	–	U.compareAndSwapObject(a, j, t, null)) {
859	–	base = b + 1;
860	–	task = t;
861	–	}
862	–	}
863	–	return task;
864	–	}
865	–
866	–	/**
801		* Pops the given task only if it is at the current top.
802		*/
803		final boolean tryUnpush(ForkJoinTask<?> t) {
#	Line 878 \| Line 812 \| public class ForkJoinPool extends Abstra
812		}
813
814		/**
815	+	* Version of tryUnpush for shared queues; called by non-FJ
816	+	* submitters. Conservatively fails to unpush if all workers
817	+	* are active unless there are multiple tasks in queue.
818	+	*/
819	+	final boolean trySharedUnpush(ForkJoinTask<?> task, ForkJoinPool p) {
820	+	boolean success = false;
821	+	if (task != null && top != base && runState == 0 &&
822	+	U.compareAndSwapInt(this, RUNSTATE, 0, 1)) {
823	+	try {
824	+	ForkJoinTask<?>[] a; int n, s;
825	+	if ((a = array) != null && (n = (s = top) - base) > 0 &&
826	+	(n > 1 \|\| p == null \|\| (int)(p.ctl >> AC_SHIFT) < 0)) {
827	+	int j = (((a.length - 1) & --s) << ASHIFT) + ABASE;
828	+	if (U.getObjectVolatile(a, j) == task &&
829	+	U.compareAndSwapObject(a, j, task, null)) {
830	+	top = s;
831	+	success = true;
832	+	}
833	+	}
834	+	} finally {
835	+	runState = 0; // unlock
836	+	}
837	+	}
838	+	return success;
839	+	}
840	+
841	+	/**
842		* Polls the given task only if it is at the current base.
843		*/
844		final boolean pollFor(ForkJoinTask<?> task) {
845	<	ForkJoinTask<?>[] a; int b, i;
846	<	if ((b = base) - top < 0 && (a = array) != null &&
847	<	(i = (a.length - 1) & b) >= 0) {
887	<	int j = (i << ASHIFT) + ABASE;
845	>	ForkJoinTask<?>[] a; int b;
846	>	if ((b = base) - top < 0 && (a = array) != null) {
847	>	int j = (((a.length - 1) & b) << ASHIFT) + ABASE;
848		if (U.getObjectVolatile(a, j) == task && base == b &&
849		U.compareAndSwapObject(a, j, task, null)) {
850		base = b + 1;
#	Line 895 \| Line 855 \| public class ForkJoinPool extends Abstra
855		}
856
857		/**
898	–	* If present, removes from queue and executes the given task, or
899	–	* any other cancelled task. Returns (true) immediately on any CAS
900	–	* or consistency check failure so caller can retry.
901	–	*
902	–	* @return false if no progress can be made
903	–	*/
904	–	final boolean tryRemoveAndExec(ForkJoinTask<?> task) {
905	–	boolean removed = false, empty = true, progress = true;
906	–	ForkJoinTask<?>[] a; int m, s, b, n;
907	–	if ((a = array) != null && (m = a.length - 1) >= 0 &&
908	–	(n = (s = top) - (b = base)) > 0) {
909	–	for (ForkJoinTask<?> t;;) { // traverse from s to b
910	–	int j = ((--s & m) << ASHIFT) + ABASE;
911	–	t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
912	–	if (t == null) // inconsistent length
913	–	break;
914	–	else if (t == task) {
915	–	if (s + 1 == top) { // pop
916	–	if (!U.compareAndSwapObject(a, j, task, null))
917	–	break;
918	–	top = s;
919	–	removed = true;
920	–	}
921	–	else if (base == b) // replace with proxy
922	–	removed = U.compareAndSwapObject(a, j, task,
923	–	new EmptyTask());
924	–	break;
925	–	}
926	–	else if (t.status >= 0)
927	–	empty = false;
928	–	else if (s + 1 == top) { // pop and throw away
929	–	if (U.compareAndSwapObject(a, j, t, null))
930	–	top = s;
931	–	break;
932	–	}
933	–	if (--n == 0) {
934	–	if (!empty && base == b)
935	–	progress = false;
936	–	break;
937	–	}
938	–	}
939	–	}
940	–	if (removed)
941	–	task.doExec();
942	–	return progress;
943	–	}
944	–
945	–	/**
858		* Initializes or doubles the capacity of array. Call either
859		* by owner or with lock held -- it is OK for base, but not
860		* top, to move while resizings are in progress.
#	Line 978 \| Line 890 \| public class ForkJoinPool extends Abstra
890		}
891
892		/**
893	<	* Removes and cancels all known tasks, ignoring any exceptions
893	>	* Removes and cancels all known tasks, ignoring any exceptions.
894		*/
895		final void cancelAll() {
896		ForkJoinTask.cancelIgnoringExceptions(currentJoin);
#	Line 987 \| Line 899 \| public class ForkJoinPool extends Abstra
899		ForkJoinTask.cancelIgnoringExceptions(t);
900		}
901
902	+	/**
903	+	* Computes next value for random probes. Scans don't require
904	+	* a very high quality generator, but also not a crummy one.
905	+	* Marsaglia xor-shift is cheap and works well enough. Note:
906	+	* This is manually inlined in its usages in ForkJoinPool to
907	+	* avoid writes inside busy scan loops.
908	+	*/
909	+	final int nextSeed() {
910	+	int r = seed;
911	+	r ^= r << 13;
912	+	r ^= r >>> 17;
913	+	return seed = r ^= r << 5;
914	+	}
915	+
916		// Execution methods
917
918		/**
919	<	* Removes and runs tasks until empty, using local mode
994	<	* ordering.
919	>	* Pops and runs tasks until empty.
920		*/
921	<	final void runLocalTasks() {
922	<	if (base - top < 0) {
923	<	for (ForkJoinTask<?> t; (t = nextLocalTask()) != null; )
921	>	private void popAndExecAll() {
922	>	// A bit faster than repeated pop calls
923	>	ForkJoinTask<?>[] a; int m, s; long j; ForkJoinTask<?> t;
924	>	while ((a = array) != null && (m = a.length - 1) >= 0 &&
925	>	(s = top - 1) - base >= 0 &&
926	>	(t = ((ForkJoinTask<?>)
927	>	U.getObject(a, j = ((m & s) << ASHIFT) + ABASE)))
928	>	!= null) {
929	>	if (U.compareAndSwapObject(a, j, t, null)) {
930	>	top = s;
931		t.doExec();
932	+	}
933		}
934		}
935
936		/**
937	+	* Polls and runs tasks until empty.
938	+	*/
939	+	private void pollAndExecAll() {
940	+	for (ForkJoinTask<?> t; (t = poll()) != null;)
941	+	t.doExec();
942	+	}
943	+
944	+	/**
945	+	* If present, removes from queue and executes the given task, or
946	+	* any other cancelled task. Returns (true) immediately on any CAS
947	+	* or consistency check failure so caller can retry.
948	+	*
949	+	* @return 0 if no progress can be made, else positive
950	+	* (this unusual convention simplifies use with tryHelpStealer.)
951	+	*/
952	+	final int tryRemoveAndExec(ForkJoinTask<?> task) {
953	+	int stat = 1;
954	+	boolean removed = false, empty = true;
955	+	ForkJoinTask<?>[] a; int m, s, b, n;
956	+	if ((a = array) != null && (m = a.length - 1) >= 0 &&
957	+	(n = (s = top) - (b = base)) > 0) {
958	+	for (ForkJoinTask<?> t;;) { // traverse from s to b
959	+	int j = ((--s & m) << ASHIFT) + ABASE;
960	+	t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
961	+	if (t == null) // inconsistent length
962	+	break;
963	+	else if (t == task) {
964	+	if (s + 1 == top) { // pop
965	+	if (!U.compareAndSwapObject(a, j, task, null))
966	+	break;
967	+	top = s;
968	+	removed = true;
969	+	}
970	+	else if (base == b) // replace with proxy
971	+	removed = U.compareAndSwapObject(a, j, task,
972	+	new EmptyTask());
973	+	break;
974	+	}
975	+	else if (t.status >= 0)
976	+	empty = false;
977	+	else if (s + 1 == top) { // pop and throw away
978	+	if (U.compareAndSwapObject(a, j, t, null))
979	+	top = s;
980	+	break;
981	+	}
982	+	if (--n == 0) {
983	+	if (!empty && base == b)
984	+	stat = 0;
985	+	break;
986	+	}
987	+	}
988	+	}
989	+	if (removed)
990	+	task.doExec();
991	+	return stat;
992	+	}
993	+
994	+	/**
995		* Executes a top-level task and any local tasks remaining
996		* after execution.
1006	–	*
1007	–	* @return true unless terminating
997		*/
998	<	final boolean runTask(ForkJoinTask<?> t) {
1010	<	boolean alive = true;
998	>	final void runTask(ForkJoinTask<?> t) {
999		if (t != null) {
1000		currentSteal = t;
1001		t.doExec();
1002	<	runLocalTasks();
1002	>	if (top != base) { // process remaining local tasks
1003	>	if (mode == 0)
1004	>	popAndExecAll();
1005	>	else
1006	>	pollAndExecAll();
1007	>	}
1008		++nsteals;
1009		currentSteal = null;
1010		}
1018	–	else if (runState < 0) // terminating
1019	–	alive = false;
1020	–	return alive;
1011		}
1012
1013		/**
1014	<	* Executes a non-top-level (stolen) task
1014	>	* Executes a non-top-level (stolen) task.
1015		*/
1016		final void runSubtask(ForkJoinTask<?> t) {
1017		if (t != null) {
#	Line 1033 \| Line 1023 \| public class ForkJoinPool extends Abstra
1023		}
1024
1025		/**
1026	<	* Computes next value for random probes. Scans don't require
1037	<	* a very high quality generator, but also not a crummy one.
1038	<	* Marsaglia xor-shift is cheap and works well enough. Note:
1039	<	* This is manually inlined in several usages in ForkJoinPool
1040	<	* to avoid writes inside busy scan loops.
1026	>	* Returns true if owned and not known to be blocked.
1027		*/
1028	<	final int nextSeed() {
1029	<	int r = seed;
1030	<	r ^= r << 13;
1031	<	r ^= r >>> 17;
1032	<	r ^= r << 5;
1033	<	return seed = r;
1028	>	final boolean isApparentlyUnblocked() {
1029	>	Thread wt; Thread.State s;
1030	>	return (eventCount >= 0 &&
1031	>	(wt = owner) != null &&
1032	>	(s = wt.getState()) != Thread.State.BLOCKED &&
1033	>	s != Thread.State.WAITING &&
1034	>	s != Thread.State.TIMED_WAITING);
1035	>	}
1036	>
1037	>	/**
1038	>	* If this owned and is not already interrupted, try to
1039	>	* interrupt and/or unpark, ignoring exceptions.
1040	>	*/
1041	>	final void interruptOwner() {
1042	>	Thread wt, p;
1043	>	if ((wt = owner) != null && !wt.isInterrupted()) {
1044	>	try {
1045	>	wt.interrupt();
1046	>	} catch (SecurityException ignore) {
1047	>	}
1048	>	}
1049	>	if ((p = parker) != null)
1050	>	U.unpark(p);
1051		}
1052
1053		// Unsafe mechanics
#	Line 1072 \| Line 1075 \| public class ForkJoinPool extends Abstra
1075		}
1076
1077		/**
1078	<	* Class for artificial tasks that are used to replace the target
1079	<	* of local joins if they are removed from an interior queue slot
1080	<	* in WorkQueue.tryRemoveAndExec. We don't need the proxy to
1081	<	* actually do anything beyond having a unique identity.
1082	<	*/
1083	<	static final class EmptyTask extends ForkJoinTask<Void> {
1084	<	EmptyTask() { status = ForkJoinTask.NORMAL; } // force done
1085	<	public Void getRawResult() { return null; }
1086	<	public void setRawResult(Void x) {}
1087	<	public boolean exec() { return true; }
1078	>	* Per-thread records for threads that submit to pools. Currently
1079	>	* holds only pseudo-random seed / index that is used to choose
1080	>	* submission queues in method doSubmit. In the future, this may
1081	>	* also incorporate a means to implement different task rejection
1082	>	* and resubmission policies.
1083	>	*
1084	>	* Seeds for submitters and workers/workQueues work in basically
1085	>	* the same way but are initialized and updated using slightly
1086	>	* different mechanics. Both are initialized using the same
1087	>	* approach as in class ThreadLocal, where successive values are
1088	>	* unlikely to collide with previous values. This is done during
1089	>	* registration for workers, but requires a separate AtomicInteger
1090	>	* for submitters. Seeds are then randomly modified upon
1091	>	* collisions using xorshifts, which requires a non-zero seed.
1092	>	*/
1093	>	static final class Submitter {
1094	>	int seed;
1095	>	Submitter() {
1096	>	int s = nextSubmitterSeed.getAndAdd(SEED_INCREMENT);
1097	>	seed = (s == 0) ? 1 : s; // ensure non-zero
1098	>	}
1099	>	}
1100	>
1101	>	/** ThreadLocal class for Submitters */
1102	>	static final class ThreadSubmitter extends ThreadLocal<Submitter> {
1103	>	public Submitter initialValue() { return new Submitter(); }
1104		}
1105
1106	+	// static fields (initialized in static initializer below)
1107	+
1108		/**
1109	<	* Computes a hash code for the given thread. This method is
1110	<	* expected to provide higher-quality hash codes than those using
1090	<	* method hashCode().
1109	>	* Creates a new ForkJoinWorkerThread. This factory is used unless
1110	>	* overridden in ForkJoinPool constructors.
1111		*/
1112	<	static final int hashThread(Thread t) {
1113	<	long id = (t == null)? 0L : t.getId(); // Use MurmurHash of thread id
1094	<	int h = (int)id ^ (int)(id >>> 32);
1095	<	h ^= h >>> 16;
1096	<	h *= 0x85ebca6b;
1097	<	h ^= h >>> 13;
1098	<	h *= 0xc2b2ae35;
1099	<	return h ^ (h >>> 16);
1100	<	}
1112	>	public static final ForkJoinWorkerThreadFactory
1113	>	defaultForkJoinWorkerThreadFactory;
1114
1115		/**
1116	<	* Top-level runloop for workers
1116	>	* Generator for assigning sequence numbers as pool names.
1117		*/
1118	<	final void runWorker(ForkJoinWorkerThread wt) {
1106	<	WorkQueue w = wt.workQueue;
1107	<	w.growArray(false); // Initialize queue array and seed in this thread
1108	<	w.seed = hashThread(Thread.currentThread()) \| (1 << 31); // force < 0
1118	>	private static final AtomicInteger poolNumberGenerator;
1119
1120	<	do {} while (w.runTask(scan(w)));
1121	<	}
1120	>	/**
1121	>	* Generator for initial hashes/seeds for submitters. Accessed by
1122	>	* Submitter class constructor.
1123	>	*/
1124	>	static final AtomicInteger nextSubmitterSeed;
1125
1126	<	// Creating, registering and deregistering workers
1126	>	/**
1127	>	* Permission required for callers of methods that may start or
1128	>	* kill threads.
1129	>	*/
1130	>	private static final RuntimePermission modifyThreadPermission;
1131	>
1132	>	/**
1133	>	* Per-thread submission bookkeeping. Shared across all pools
1134	>	* to reduce ThreadLocal pollution and because random motion
1135	>	* to avoid contention in one pool is likely to hold for others.
1136	>	*/
1137	>	private static final ThreadSubmitter submitters;
1138	>
1139	>	/** Common default pool */
1140	>	static volatile ForkJoinPool commonPool;
1141	>
1142	>	// commonPool construction parameters
1143	>	private static final String propPrefix =
1144	>	"java.util.concurrent.ForkJoinPool.common.";
1145	>	private static final Thread.UncaughtExceptionHandler commonPoolUEH;
1146	>	private static final ForkJoinWorkerThreadFactory commonPoolFactory;
1147	>	static final int commonPoolParallelism;
1148	>
1149	>	/** Static initialization lock */
1150	>	private static final Mutex initializationLock;
1151	>
1152	>	// static constants
1153	>
1154	>	/**
1155	>	* Initial timeout value (in nanoseconds) for the tread triggering
1156	>	* quiescence to park waiting for new work. On timeout, the thread
1157	>	* will instead try to shrink the number of workers.
1158	>	*/
1159	>	private static final long IDLE_TIMEOUT = 1000L * 1000L * 1000L; // 1sec
1160	>
1161	>	/**
1162	>	* Timeout value when there are more threads than parallelism level
1163	>	*/
1164	>	private static final long FAST_IDLE_TIMEOUT = 100L * 1000L * 1000L;
1165	>
1166	>	/**
1167	>	* The maximum stolen->joining link depth allowed in method
1168	>	* tryHelpStealer. Must be a power of two. This value also
1169	>	* controls the maximum number of times to try to help join a task
1170	>	* without any apparent progress or change in pool state before
1171	>	* giving up and blocking (see awaitJoin). Depths for legitimate
1172	>	* chains are unbounded, but we use a fixed constant to avoid
1173	>	* (otherwise unchecked) cycles and to bound staleness of
1174	>	* traversal parameters at the expense of sometimes blocking when
1175	>	* we could be helping.
1176	>	*/
1177	>	private static final int MAX_HELP = 64;
1178	>
1179	>	/**
1180	>	* Secondary time-based bound (in nanosecs) for helping attempts
1181	>	* before trying compensated blocking in awaitJoin. Used in
1182	>	* conjunction with MAX_HELP to reduce variance due to different
1183	>	* polling rates associated with different helping options. The
1184	>	* value should roughly approximate the time required to create
1185	>	* and/or activate a worker thread.
1186	>	*/
1187	>	private static final long COMPENSATION_DELAY = 1L << 18; // ~0.25 millisec
1188	>
1189	>	/**
1190	>	* Increment for seed generators. See class ThreadLocal for
1191	>	* explanation.
1192	>	*/
1193	>	private static final int SEED_INCREMENT = 0x61c88647;
1194	>
1195	>	/**
1196	>	* Bits and masks for control variables
1197	>	*
1198	>	* Field ctl is a long packed with:
1199	>	* AC: Number of active running workers minus target parallelism (16 bits)
1200	>	* TC: Number of total workers minus target parallelism (16 bits)
1201	>	* ST: true if pool is terminating (1 bit)
1202	>	* EC: the wait count of top waiting thread (15 bits)
1203	>	* ID: poolIndex of top of Treiber stack of waiters (16 bits)
1204	>	*
1205	>	* When convenient, we can extract the upper 32 bits of counts and
1206	>	* the lower 32 bits of queue state, u = (int)(ctl >>> 32) and e =
1207	>	* (int)ctl. The ec field is never accessed alone, but always
1208	>	* together with id and st. The offsets of counts by the target
1209	>	* parallelism and the positionings of fields makes it possible to
1210	>	* perform the most common checks via sign tests of fields: When
1211	>	* ac is negative, there are not enough active workers, when tc is
1212	>	* negative, there are not enough total workers, and when e is
1213	>	* negative, the pool is terminating. To deal with these possibly
1214	>	* negative fields, we use casts in and out of "short" and/or
1215	>	* signed shifts to maintain signedness.
1216	>	*
1217	>	* When a thread is queued (inactivated), its eventCount field is
1218	>	* set negative, which is the only way to tell if a worker is
1219	>	* prevented from executing tasks, even though it must continue to
1220	>	* scan for them to avoid queuing races. Note however that
1221	>	* eventCount updates lag releases so usage requires care.
1222	>	*
1223	>	* Field runState is an int packed with:
1224	>	* SHUTDOWN: true if shutdown is enabled (1 bit)
1225	>	* SEQ: a sequence number updated upon (de)registering workers (30 bits)
1226	>	* INIT: set true after workQueues array construction (1 bit)
1227	>	*
1228	>	* The sequence number enables simple consistency checks:
1229	>	* Staleness of read-only operations on the workQueues array can
1230	>	* be checked by comparing runState before vs after the reads.
1231	>	*/
1232	>
1233	>	// bit positions/shifts for fields
1234	>	private static final int AC_SHIFT = 48;
1235	>	private static final int TC_SHIFT = 32;
1236	>	private static final int ST_SHIFT = 31;
1237	>	private static final int EC_SHIFT = 16;
1238	>
1239	>	// bounds
1240	>	private static final int SMASK = 0xffff; // short bits
1241	>	private static final int MAX_CAP = 0x7fff; // max #workers - 1
1242	>	private static final int SQMASK = 0xfffe; // even short bits
1243	>	private static final int SHORT_SIGN = 1 << 15;
1244	>	private static final int INT_SIGN = 1 << 31;
1245	>
1246	>	// masks
1247	>	private static final long STOP_BIT = 0x0001L << ST_SHIFT;
1248	>	private static final long AC_MASK = ((long)SMASK) << AC_SHIFT;
1249	>	private static final long TC_MASK = ((long)SMASK) << TC_SHIFT;
1250	>
1251	>	// units for incrementing and decrementing
1252	>	private static final long TC_UNIT = 1L << TC_SHIFT;
1253	>	private static final long AC_UNIT = 1L << AC_SHIFT;
1254	>
1255	>	// masks and units for dealing with u = (int)(ctl >>> 32)
1256	>	private static final int UAC_SHIFT = AC_SHIFT - 32;
1257	>	private static final int UTC_SHIFT = TC_SHIFT - 32;
1258	>	private static final int UAC_MASK = SMASK << UAC_SHIFT;
1259	>	private static final int UTC_MASK = SMASK << UTC_SHIFT;
1260	>	private static final int UAC_UNIT = 1 << UAC_SHIFT;
1261	>	private static final int UTC_UNIT = 1 << UTC_SHIFT;
1262	>
1263	>	// masks and units for dealing with e = (int)ctl
1264	>	private static final int E_MASK = 0x7fffffff; // no STOP_BIT
1265	>	private static final int E_SEQ = 1 << EC_SHIFT;
1266	>
1267	>	// runState bits
1268	>	private static final int SHUTDOWN = 1 << 31;
1269	>
1270	>	// access mode for WorkQueue
1271	>	static final int LIFO_QUEUE = 0;
1272	>	static final int FIFO_QUEUE = 1;
1273	>	static final int SHARED_QUEUE = -1;
1274	>
1275	>	// Instance fields
1276	>
1277	>	/*
1278	>	* Field layout order in this class tends to matter more than one
1279	>	* would like. Runtime layout order is only loosely related to
1280	>	* declaration order and may differ across JVMs, but the following
1281	>	* empirically works OK on current JVMs.
1282	>	*/
1283	>
1284	>	volatile long ctl; // main pool control
1285	>	final int parallelism; // parallelism level
1286	>	final int localMode; // per-worker scheduling mode
1287	>	final int submitMask; // submit queue index bound
1288	>	int nextSeed; // for initializing worker seeds
1289	>	volatile int runState; // shutdown status and seq
1290	>	WorkQueue[] workQueues; // main registry
1291	>	final Mutex lock; // for registration
1292	>	final Condition termination; // for awaitTermination
1293	>	final ForkJoinWorkerThreadFactory factory; // factory for new workers
1294	>	final Thread.UncaughtExceptionHandler ueh; // per-worker UEH
1295	>	final AtomicLong stealCount; // collect counts when terminated
1296	>	final AtomicInteger nextWorkerNumber; // to create worker name string
1297	>	String workerNamePrefix; // to create worker name string
1298	>
1299	>	// Creating, registering, and deregistering workers
1300
1301		/**
1302		* Tries to create and start a worker
1303		*/
1304		private void addWorker() {
1305		Throwable ex = null;
1306	<	ForkJoinWorkerThread w = null;
1306	>	ForkJoinWorkerThread wt = null;
1307		try {
1308	<	if ((w = factory.newThread(this)) != null) {
1309	<	w.start();
1308	>	if ((wt = factory.newThread(this)) != null) {
1309	>	wt.start();
1310		return;
1311		}
1312		} catch (Throwable e) {
1313		ex = e;
1314		}
1315	<	deregisterWorker(w, ex);
1315	>	deregisterWorker(wt, ex); // adjust counts etc on failure
1316		}
1317
1318		/**
#	Line 1141 \| Line 1327 \| public class ForkJoinPool extends Abstra
1327		}
1328
1329		/**
1330	<	* Callback from ForkJoinWorkerThread constructor to establish and
1331	<	* record its WorkQueue
1330	>	* Callback from ForkJoinWorkerThread constructor to establish its
1331	>	* poolIndex and record its WorkQueue. To avoid scanning bias due
1332	>	* to packing entries in front of the workQueues array, we treat
1333	>	* the array as a simple power-of-two hash table using per-thread
1334	>	* seed as hash, expanding as needed.
1335		*
1336	<	* @param wt the worker thread
1336	>	* @param w the worker's queue
1337		*/
1338	<	final void registerWorker(ForkJoinWorkerThread wt) {
1339	<	WorkQueue w = wt.workQueue;
1151	<	ReentrantLock lock = this.lock;
1338	>	final void registerWorker(WorkQueue w) {
1339	>	Mutex lock = this.lock;
1340		lock.lock();
1341		try {
1154	–	int k = nextPoolIndex;
1342		WorkQueue[] ws = workQueues;
1343	<	if (ws != null) { // ignore on shutdown
1344	<	int n = ws.length;
1345	<	if (k < 0 \|\| (k & 1) == 0 \|\| k >= n \|\| ws[k] != null) {
1346	<	for (k = 1; k < n && ws[k] != null; k += 2)
1347	<	; // workers are at odd indices
1348	<	if (k >= n) // resize
1349	<	workQueues = ws = Arrays.copyOf(ws, n << 1);
1350	<	}
1351	<	w.poolIndex = k;
1352	<	w.eventCount = ~(k >>> 1) & SMASK; // Set up wait count
1353	<	ws[k] = w; // record worker
1354	<	nextPoolIndex = k + 2;
1355	<	int rs = runState;
1356	<	int m = rs & SMASK; // recalculate runState mask
1357	<	if (k > m)
1358	<	m = (m << 1) + 1;
1359	<	runState = (rs & SHUTDOWN) \| ((rs + RS_SEQ) & RS_SEQ_MASK) \| m;
1343	>	if (w != null && ws != null) { // skip on shutdown/failure
1344	>	int rs, n = ws.length, m = n - 1;
1345	>	int s = nextSeed += SEED_INCREMENT; // rarely-colliding sequence
1346	>	w.seed = (s == 0) ? 1 : s; // ensure non-zero seed
1347	>	int r = (s << 1) \| 1; // use odd-numbered indices
1348	>	if (ws[r &= m] != null) { // collision
1349	>	int probes = 0; // step by approx half size
1350	>	int step = (n <= 4) ? 2 : ((n >>> 1) & SQMASK) + 2;
1351	>	while (ws[r = (r + step) & m] != null) {
1352	>	if (++probes >= n) {
1353	>	workQueues = ws = Arrays.copyOf(ws, n <<= 1);
1354	>	m = n - 1;
1355	>	probes = 0;
1356	>	}
1357	>	}
1358	>	}
1359	>	w.eventCount = w.poolIndex = r; // establish before recording
1360	>	ws[r] = w; // also update seq
1361	>	runState = ((rs = runState) & SHUTDOWN) \| ((rs + 2) & ~SHUTDOWN);
1362		}
1363		} finally {
1364		lock.unlock();
#	Line 1177 \| Line 1366 \| public class ForkJoinPool extends Abstra
1366		}
1367
1368		/**
1369	<	* Final callback from terminating worker, as well as failure to
1370	<	* construct or start a worker in addWorker. Removes record of
1369	>	* Final callback from terminating worker, as well as upon failure
1370	>	* to construct or start a worker in addWorker. Removes record of
1371		* worker from array, and adjusts counts. If pool is shutting
1372		* down, tries to complete termination.
1373		*
#	Line 1186 \| Line 1375 \| public class ForkJoinPool extends Abstra
1375		* @param ex the exception causing failure, or null if none
1376		*/
1377		final void deregisterWorker(ForkJoinWorkerThread wt, Throwable ex) {
1378	+	Mutex lock = this.lock;
1379		WorkQueue w = null;
1380		if (wt != null && (w = wt.workQueue) != null) {
1381		w.runState = -1; // ensure runState is set
1382		stealCount.getAndAdd(w.totalSteals + w.nsteals);
1383		int idx = w.poolIndex;
1194	–	ReentrantLock lock = this.lock;
1384		lock.lock();
1385		try { // remove record from array
1386		WorkQueue[] ws = workQueues;
1387		if (ws != null && idx >= 0 && idx < ws.length && ws[idx] == w)
1388	<	ws[nextPoolIndex = idx] = null;
1388	>	ws[idx] = null;
1389		} finally {
1390		lock.unlock();
1391		}
#	Line 1208 \| Line 1397 \| public class ForkJoinPool extends Abstra
1397		((c - TC_UNIT) & TC_MASK) \|
1398		(c & ~(AC_MASK\|TC_MASK)))));
1399
1400	<	if (!tryTerminate(false) && w != null) {
1400	>	if (!tryTerminate(false, false) && w != null) {
1401		w.cancelAll(); // cancel remaining tasks
1402		if (w.array != null) // suppress signal if never ran
1403		signalWork(); // wake up or create replacement
1404	+	if (ex == null) // help clean refs on way out
1405	+	ForkJoinTask.helpExpungeStaleExceptions();
1406		}
1407
1408		if (ex != null) // rethrow
1409		U.throwException(ex);
1410		}
1411
1412	<
1222	<	// Maintaining ctl counts
1223	<
1224	<	/**
1225	<	* Increments active count; mainly called upon return from blocking
1226	<	*/
1227	<	final void incrementActiveCount() {
1228	<	long c;
1229	<	do {} while (!U.compareAndSwapLong(this, CTL, c = ctl, c + AC_UNIT));
1230	<	}
1412	>	// Submissions
1413
1414		/**
1415	<	* Activates or creates a worker
1415	>	* Unless shutting down, adds the given task to a submission queue
1416	>	* at submitter's current queue index (modulo submission
1417	>	* range). If no queue exists at the index, one is created. If
1418	>	* the queue is busy, another index is randomly chosen. The
1419	>	* submitMask bounds the effective number of queues to the
1420	>	* (nearest power of two for) parallelism level.
1421	>	*
1422	>	* @param task the task. Caller must ensure non-null.
1423		*/
1424	<	final void signalWork() {
1425	<	/*
1426	<	* The while condition is true if: (there is are too few total
1427	<	* workers OR there is at least one waiter) AND (there are too
1428	<	* few active workers OR the pool is terminating). The value
1429	<	* of e distinguishes the remaining cases: zero (no waiters)
1430	<	* for create, negative if terminating (in which case do
1431	<	* nothing), else release a waiter. The secondary checks for
1432	<	* release (non-null array etc) can fail if the pool begins
1433	<	* terminating after the test, and don't impose any added cost
1434	<	* because JVMs must perform null and bounds checks anyway.
1435	<	*/
1436	<	long c; int e, u;
1437	<	while ((((e = (int)(c = ctl)) \| (u = (int)(c >>> 32))) &
1438	<	(INT_SIGN\|SHORT_SIGN)) == (INT_SIGN\|SHORT_SIGN)) {
1439	<	WorkQueue[] ws = workQueues; int i; WorkQueue w; Thread p;
1440	<	if (e == 0) { // add a new worker
1441	<	if (U.compareAndSwapLong
1442	<	(this, CTL, c, (long)(((u + UTC_UNIT) & UTC_MASK) \|
1254	<	((u + UAC_UNIT) & UAC_MASK)) << 32)) {
1255	<	addWorker();
1256	<	break;
1424	>	private void doSubmit(ForkJoinTask<?> task) {
1425	>	Submitter s = submitters.get();
1426	>	for (int r = s.seed, m = submitMask;;) {
1427	>	WorkQueue[] ws; WorkQueue q;
1428	>	int k = r & m & SQMASK; // use only even indices
1429	>	if (runState < 0 \|\| (ws = workQueues) == null \|\| ws.length <= k)
1430	>	throw new RejectedExecutionException(); // shutting down
1431	>	else if ((q = ws[k]) == null) { // create new queue
1432	>	WorkQueue nq = new WorkQueue(this, null, SHARED_QUEUE);
1433	>	Mutex lock = this.lock; // construct outside lock
1434	>	lock.lock();
1435	>	try { // recheck under lock
1436	>	int rs = runState; // to update seq
1437	>	if (ws == workQueues && ws[k] == null) {
1438	>	ws[k] = nq;
1439	>	runState = ((rs & SHUTDOWN) \| ((rs + 2) & ~SHUTDOWN));
1440	>	}
1441	>	} finally {
1442	>	lock.unlock();
1443		}
1444		}
1445	<	else if (e > 0 && ws != null &&
1446	<	(i = ((~e << 1) \| 1) & SMASK) < ws.length &&
1447	<	(w = ws[i]) != null &&
1448	<	w.eventCount == (e \| INT_SIGN)) {
1449	<	if (U.compareAndSwapLong
1450	<	(this, CTL, c, (((long)(w.nextWait & E_MASK)) \|
1451	<	((long)(u + UAC_UNIT) << 32)))) {
1452	<	w.eventCount = (e + E_SEQ) & E_MASK;
1267	<	if ((p = w.parker) != null)
1268	<	U.unpark(p); // release a waiting worker
1269	<	break;
1270	<	}
1445	>	else if (q.trySharedPush(task)) {
1446	>	signalWork();
1447	>	return;
1448	>	}
1449	>	else if (m > 1) { // move to a different index
1450	>	r ^= r << 13; // same xorshift as WorkQueues
1451	>	r ^= r >>> 17;
1452	>	s.seed = r ^= r << 5;
1453		}
1454		else
1455	<	break;
1455	>	Thread.yield(); // yield if no alternatives
1456		}
1457		}
1458
1459		/**
1460	<	* Tries to decrement active count (sometimes implicitly) and
1279	<	* possibly release or create a compensating worker in preparation
1280	<	* for blocking. Fails on contention or termination.
1281	<	*
1282	<	* @return true if the caller can block, else should recheck and retry
1460	>	* Submits the given (non-null) task to the common pool, if possible.
1461		*/
1462	<	final boolean tryCompensate() {
1463	<	WorkQueue[] ws; WorkQueue w; Thread p;
1464	<	int pc = parallelism, e, u, ac, tc, i;
1465	<	long c = ctl;
1462	>	static void submitToCommonPool(ForkJoinTask<?> task) {
1463	>	ForkJoinPool p;
1464	>	if ((p = commonPool) == null)
1465	>	p = ensureCommonPool();
1466	>	p.doSubmit(task);
1467	>	}
1468
1469	<	if ((e = (int)c) >= 0) {
1470	<	if ((ac = ((u = (int)(c >>> 32)) >> UAC_SHIFT)) <= 0 &&
1471	<	e != 0 && (ws = workQueues) != null &&
1472	<	(i = ((~e << 1) \| 1) & SMASK) < ws.length &&
1473	<	(w = ws[i]) != null) {
1474	<	if (w.eventCount == (e \| INT_SIGN) &&
1475	<	U.compareAndSwapLong
1476	<	(this, CTL, c, ((long)(w.nextWait & E_MASK) \|
1477	<	(c & (AC_MASK\|TC_MASK))))) {
1478	<	w.eventCount = (e + E_SEQ) & E_MASK;
1479	<	if ((p = w.parker) != null)
1480	<	U.unpark(p);
1481	<	return true; // release an idle worker
1482	<	}
1483	<	}
1484	<	else if ((tc = (short)(u >>> UTC_SHIFT)) >= 0 && ac + pc > 1) {
1485	<	long nc = ((c - AC_UNIT) & AC_MASK) \| (c & ~AC_MASK);
1306	<	if (U.compareAndSwapLong(this, CTL, c, nc))
1307	<	return true; // no compensation needed
1308	<	}
1309	<	else if (tc + pc < MAX_ID) {
1310	<	long nc = ((c + TC_UNIT) & TC_MASK) \| (c & ~TC_MASK);
1311	<	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1312	<	addWorker();
1313	<	return true; // create replacement
1314	<	}
1315	<	}
1316	<	}
1317	<	return false;
1469	>	/**
1470	>	* Returns true if the given task was submitted to common pool
1471	>	* and has not yet commenced execution, and is available for
1472	>	* removal according to execution policies; if so removing the
1473	>	* submission from the pool.
1474	>	*
1475	>	* @param task the task
1476	>	* @return true if successful
1477	>	*/
1478	>	static boolean tryUnsubmitFromCommonPool(ForkJoinTask<?> task) {
1479	>	ForkJoinPool p; WorkQueue[] ws; WorkQueue q;
1480	>	int k = submitters.get().seed & SQMASK;
1481	>	return ((p = commonPool) != null &&
1482	>	(ws = p.workQueues) != null &&
1483	>	ws.length > (k &= p.submitMask) &&
1484	>	(q = ws[k]) != null &&
1485	>	q.trySharedUnpush(task, p));
1486		}
1487
1488	<	// Submissions
1488	>	// Maintaining ctl counts
1489
1490		/**
1491	<	* Unless shutting down, adds the given task to some submission
1324	<	* queue; using a randomly chosen queue index if the caller is a
1325	<	* ForkJoinWorkerThread, else one based on caller thread's hash
1326	<	* code. If no queue exists at the index, one is created. If the
1327	<	* queue is busy, another is chosen by sweeping through the queues
1328	<	* array.
1491	>	* Increments active count; mainly called upon return from blocking.
1492		*/
1493	<	private void doSubmit(ForkJoinTask<?> task) {
1494	<	if (task == null)
1495	<	throw new NullPointerException();
1333	<	Thread t = Thread.currentThread();
1334	<	int r = ((t instanceof ForkJoinWorkerThread) ?
1335	<	((ForkJoinWorkerThread)t).workQueue.nextSeed() : hashThread(t));
1336	<	for (;;) {
1337	<	int rs = runState, m = rs & SMASK;
1338	<	int j = r &= (m & ~1); // even numbered queues
1339	<	WorkQueue[] ws = workQueues;
1340	<	if (rs < 0 \|\| ws == null)
1341	<	throw new RejectedExecutionException(); // shutting down
1342	<	if (ws.length > m) { // consistency check
1343	<	for (WorkQueue q;;) { // circular sweep
1344	<	if (((q = ws[j]) != null \|\|
1345	<	(q = tryAddSharedQueue(j)) != null) &&
1346	<	q.trySharedPush(task)) {
1347	<	signalWork();
1348	<	return;
1349	<	}
1350	<	if ((j = (j + 2) & m) == r) {
1351	<	Thread.yield(); // all queues busy
1352	<	break;
1353	<	}
1354	<	}
1355	<	}
1356	<	}
1493	>	final void incrementActiveCount() {
1494	>	long c;
1495	>	do {} while (!U.compareAndSwapLong(this, CTL, c = ctl, c + AC_UNIT));
1496		}
1497
1498		/**
1499	<	* Tries to add and register a new queue at the given index.
1500	<	*
1501	<	* @param idx the workQueues array index to register the queue
1502	<	* @return the queue, or null if could not add because could
1503	<	* not acquire lock or idx is unusable
1504	<	*/
1505	<	private WorkQueue tryAddSharedQueue(int idx) {
1506	<	WorkQueue q = null;
1507	<	ReentrantLock lock = this.lock;
1508	<	if (idx >= 0 && (idx & 1) == 0 && !lock.isLocked()) {
1509	<	// create queue outside of lock but only if apparently free
1510	<	WorkQueue nq = new WorkQueue(null, SHARED_QUEUE);
1511	<	if (lock.tryLock()) {
1512	<	try {
1513	<	WorkQueue[] ws = workQueues;
1514	<	if (ws != null && idx < ws.length) {
1376	<	if ((q = ws[idx]) == null) {
1377	<	int rs; // update runState seq
1378	<	ws[idx] = q = nq;
1379	<	runState = (((rs = runState) & SHUTDOWN) \|
1380	<	((rs + RS_SEQ) & ~SHUTDOWN));
1381	<	}
1499	>	* Tries to create one or activate one or more workers if too few are active.
1500	>	*/
1501	>	final void signalWork() {
1502	>	long c; int u;
1503	>	while ((u = (int)((c = ctl) >>> 32)) < 0) { // too few active
1504	>	WorkQueue[] ws = workQueues; int e, i; WorkQueue w; Thread p;
1505	>	if ((e = (int)c) > 0) { // at least one waiting
1506	>	if (ws != null && (i = e & SMASK) < ws.length &&
1507	>	(w = ws[i]) != null && w.eventCount == (e \| INT_SIGN)) {
1508	>	long nc = (((long)(w.nextWait & E_MASK)) \|
1509	>	((long)(u + UAC_UNIT) << 32));
1510	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1511	>	w.eventCount = (e + E_SEQ) & E_MASK;
1512	>	if ((p = w.parker) != null)
1513	>	U.unpark(p); // activate and release
1514	>	break;
1515		}
1383	–	} finally {
1384	–	lock.unlock();
1516		}
1517	+	else
1518	+	break;
1519		}
1520	+	else if (e == 0 && (u & SHORT_SIGN) != 0) { // too few total
1521	+	long nc = (long)(((u + UTC_UNIT) & UTC_MASK) \|
1522	+	((u + UAC_UNIT) & UAC_MASK)) << 32;
1523	+	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1524	+	addWorker();
1525	+	break;
1526	+	}
1527	+	}
1528	+	else
1529	+	break;
1530		}
1388	–	return q;
1531		}
1532
1533		// Scanning for tasks
1534
1535		/**
1536	+	* Top-level runloop for workers, called by ForkJoinWorkerThread.run.
1537	+	*/
1538	+	final void runWorker(WorkQueue w) {
1539	+	w.growArray(false); // initialize queue array in this thread
1540	+	do { w.runTask(scan(w)); } while (w.runState >= 0);
1541	+	}
1542	+
1543	+	/**
1544		* Scans for and, if found, returns one task, else possibly
1545		* inactivates the worker. This method operates on single reads of
1546	<	* volatile state and is designed to be re-invoked continuously in
1547	<	* part because it returns upon detecting inconsistencies,
1546	>	* volatile state and is designed to be re-invoked continuously,
1547	>	* in part because it returns upon detecting inconsistencies,
1548		* contention, or state changes that indicate possible success on
1549		* re-invocation.
1550		*
1551	<	* The scan searches for tasks across queues, randomly selecting
1552	<	* the first #queues probes, favoring steals 2:1 over submissions
1553	<	* (by exploiting even/odd indexing), and then performing a
1554	<	* circular sweep of all queues. The scan terminates upon either
1555	<	* finding a non-empty queue, or completing a full sweep. If the
1556	<	* worker is not inactivated, it takes and returns a task from
1557	<	* this queue. On failure to find a task, we take one of the
1558	<	* following actions, after which the caller will retry calling
1559	<	* this method unless terminated.
1551	>	* The scan searches for tasks across a random permutation of
1552	>	* queues (starting at a random index and stepping by a random
1553	>	* relative prime, checking each at least once). The scan
1554	>	* terminates upon either finding a non-empty queue, or completing
1555	>	* the sweep. If the worker is not inactivated, it takes and
1556	>	* returns a task from this queue. On failure to find a task, we
1557	>	* take one of the following actions, after which the caller will
1558	>	* retry calling this method unless terminated.
1559	>	*
1560	>	* * If pool is terminating, terminate the worker.
1561		*
1562		* * If not a complete sweep, try to release a waiting worker. If
1563		* the scan terminated because the worker is inactivated, then the
#	Line 1415 \| Line 1566 \| public class ForkJoinPool extends Abstra
1566		* another worker, but with same net effect. Releasing in other
1567		* cases as well ensures that we have enough workers running.
1568		*
1418	–	* * If the caller has run a task since the the last empty scan,
1419	–	* return (to allow rescan) if other workers are not also yet
1420	–	* enqueued. Field WorkQueue.rescans counts down on each scan to
1421	–	* ensure eventual inactivation, and occasional calls to
1422	–	* Thread.yield to help avoid interference with more useful
1423	–	* activities on the system.
1424	–	*
1425	–	* * If pool is terminating, terminate the worker
1426	–	*
1569		* * If not already enqueued, try to inactivate and enqueue the
1570	<	* worker on wait queue.
1570	>	* worker on wait queue. Or, if inactivating has caused the pool
1571	>	* to be quiescent, relay to idleAwaitWork to check for
1572	>	* termination and possibly shrink pool.
1573	>	*
1574	>	* * If already inactive, and the caller has run a task since the
1575	>	* last empty scan, return (to allow rescan) unless others are
1576	>	* also inactivated. Field WorkQueue.rescans counts down on each
1577	>	* scan to ensure eventual inactivation and blocking.
1578		*
1579	<	* * If already enqueued and none of the above apply, either park
1580	<	* awaiting signal, or if this is the most recent waiter and pool
1432	<	* is quiescent, relay to idleAwaitWork to check for termination
1433	<	* and possibly shrink pool.
1579	>	* * If already enqueued and none of the above apply, park
1580	>	* awaiting signal,
1581		*
1582		* @param w the worker (via its WorkQueue)
1583	<	* @return a task or null of none found
1583	>	* @return a task or null if none found
1584		*/
1585		private final ForkJoinTask<?> scan(WorkQueue w) {
1586	<	boolean swept = false; // true after full empty scan
1587	<	WorkQueue[] ws; // volatile read order matters
1588	<	int r = w.seed, ec = w.eventCount; // ec is negative if inactive
1589	<	int rs = runState, m = rs & SMASK;
1590	<	if ((ws = workQueues) != null && ws.length > m) {
1591	<	ForkJoinTask<?> task = null;
1592	<	for (int k = 0, j = -2 - m; ; ++j) {
1593	<	WorkQueue q; int b;
1594	<	if (j < 0) { // random probes while j negative
1595	<	r ^= r << 13; r ^= r >>> 17; k = (r ^= r << 5) \| (j & 1);
1596	<	} // worker (not submit) for odd j
1597	<	else // cyclic scan when j >= 0
1598	<	k += (m >>> 1) \| 1; // step by half to reduce bias
1599	<
1600	<	if ((q = ws[k & m]) != null && (b = q.base) - q.top < 0) {
1601	<	if (ec >= 0)
1602	<	task = q.pollAt(b); // steal
1603	<	break;
1586	>	WorkQueue[] ws; // first update random seed
1587	>	int r = w.seed; r ^= r << 13; r ^= r >>> 17; w.seed = r ^= r << 5;
1588	>	int rs = runState, m; // volatile read order matters
1589	>	if ((ws = workQueues) != null && (m = ws.length - 1) > 0) {
1590	>	int ec = w.eventCount; // ec is negative if inactive
1591	>	int step = (r >>> 16) \| 1; // relative prime
1592	>	for (int j = (m + 1) << 2; ; r += step) {
1593	>	WorkQueue q; ForkJoinTask<?> t; ForkJoinTask<?>[] a; int b;
1594	>	if ((q = ws[r & m]) != null && (b = q.base) - q.top < 0 &&
1595	>	(a = q.array) != null) { // probably nonempty
1596	>	int i = (((a.length - 1) & b) << ASHIFT) + ABASE;
1597	>	t = (ForkJoinTask<?>)U.getObjectVolatile(a, i);
1598	>	if (q.base == b && ec >= 0 && t != null &&
1599	>	U.compareAndSwapObject(a, i, t, null)) {
1600	>	if (q.top - (q.base = b + 1) > 0)
1601	>	signalWork(); // help pushes signal
1602	>	return t;
1603	>	}
1604	>	else if (ec < 0 \|\| j <= m) {
1605	>	rs = 0; // mark scan as imcomplete
1606	>	break; // caller can retry after release
1607	>	}
1608		}
1609	<	else if (j > m) {
1459	<	if (rs == runState) // staleness check
1460	<	swept = true;
1609	>	if (--j < 0)
1610		break;
1611	+	}
1612	+
1613	+	long c = ctl; int e = (int)c, a = (int)(c >> AC_SHIFT), nr, ns;
1614	+	if (e < 0) // decode ctl on empty scan
1615	+	w.runState = -1; // pool is terminating
1616	+	else if (rs == 0 \|\| rs != runState) { // incomplete scan
1617	+	WorkQueue v; Thread p; // try to release a waiter
1618	+	if (e > 0 && a < 0 && w.eventCount == ec &&
1619	+	(v = ws[e & m]) != null && v.eventCount == (e \| INT_SIGN)) {
1620	+	long nc = ((long)(v.nextWait & E_MASK) \|
1621	+	((c + AC_UNIT) & (AC_MASK\|TC_MASK)));
1622	+	if (ctl == c && U.compareAndSwapLong(this, CTL, c, nc)) {
1623	+	v.eventCount = (e + E_SEQ) & E_MASK;
1624	+	if ((p = v.parker) != null)
1625	+	U.unpark(p);
1626	+	}
1627		}
1628		}
1629	<	w.seed = r; // save seed for next scan
1630	<	if (task != null)
1631	<	return task;
1632	<	}
1633	<
1634	<	// Decode ctl on empty scan
1635	<	long c = ctl; int e = (int)c, a = (int)(c >> AC_SHIFT), nr, ns;
1636	<	if (!swept) { // try to release a waiter
1637	<	WorkQueue v; Thread p;
1638	<	if (e > 0 && a < 0 && ws != null &&
1639	<	(v = ws[((~e << 1) \| 1) & m]) != null &&
1640	<	v.eventCount == (e \| INT_SIGN) && U.compareAndSwapLong
1641	<	(this, CTL, c, ((long)(v.nextWait & E_MASK) \|
1642	<	((c + AC_UNIT) & (AC_MASK\|TC_MASK))))) {
1643	<	v.eventCount = (e + E_SEQ) & E_MASK;
1644	<	if ((p = v.parker) != null)
1645	<	U.unpark(p);
1646	<	}
1647	<	}
1648	<	else if ((nr = w.rescans) > 0) { // continue rescanning
1649	<	int ac = a + parallelism;
1650	<	if ((w.rescans = (ac < nr) ? ac : nr - 1) > 0 && w.seed < 0 &&
1651	<	w.eventCount == ec)
1652	<	Thread.yield(); // 1 bit randomness for yield call
1653	<	}
1654	<	else if (e < 0) // pool is terminating
1655	<	w.runState = -1;
1656	<	else if (ec >= 0) { // try to enqueue
1657	<	long nc = (long)ec \| ((c - AC_UNIT) & (AC_MASK\|TC_MASK));
1658	<	w.nextWait = e;
1494	<	w.eventCount = ec \| INT_SIGN; // mark as inactive
1495	<	if (!U.compareAndSwapLong(this, CTL, c, nc))
1496	<	w.eventCount = ec; // back out on CAS failure
1497	<	else if ((ns = w.nsteals) != 0) { // set rescans if ran task
1498	<	if (a <= 0) // ... unless too many active
1499	<	w.rescans = a + parallelism;
1500	<	w.nsteals = 0;
1501	<	w.totalSteals += ns;
1502	<	}
1503	<	}
1504	<	else{ // already queued
1505	<	if (parallelism == -a)
1506	<	idleAwaitWork(w); // quiescent
1507	<	if (w.eventCount == ec) {
1508	<	Thread.interrupted(); // clear status
1509	<	ForkJoinWorkerThread wt = w.owner;
1510	<	U.putObject(wt, PARKBLOCKER, this);
1511	<	w.parker = wt; // emulate LockSupport.park
1512	<	if (w.eventCount == ec) // recheck
1513	<	U.park(false, 0L); // block
1514	<	w.parker = null;
1515	<	U.putObject(wt, PARKBLOCKER, null);
1629	>	else if (ec >= 0) { // try to enqueue/inactivate
1630	>	long nc = (long)ec \| ((c - AC_UNIT) & (AC_MASK\|TC_MASK));
1631	>	w.nextWait = e;
1632	>	w.eventCount = ec \| INT_SIGN; // mark as inactive
1633	>	if (ctl != c \|\| !U.compareAndSwapLong(this, CTL, c, nc))
1634	>	w.eventCount = ec; // unmark on CAS failure
1635	>	else {
1636	>	if ((ns = w.nsteals) != 0) {
1637	>	w.nsteals = 0; // set rescans if ran task
1638	>	w.rescans = (a > 0) ? 0 : a + parallelism;
1639	>	w.totalSteals += ns;
1640	>	}
1641	>	if (a == 1 - parallelism) // quiescent
1642	>	idleAwaitWork(w, nc, c);
1643	>	}
1644	>	}
1645	>	else if (w.eventCount < 0) { // already queued
1646	>	int ac = a + parallelism;
1647	>	if ((nr = w.rescans) > 0) // continue rescanning
1648	>	w.rescans = (ac < nr) ? ac : nr - 1;
1649	>	else if (((w.seed >>> 16) & ac) == 0) { // randomize park
1650	>	Thread.interrupted(); // clear status
1651	>	Thread wt = Thread.currentThread();
1652	>	U.putObject(wt, PARKBLOCKER, this);
1653	>	w.parker = wt; // emulate LockSupport.park
1654	>	if (w.eventCount < 0) // recheck
1655	>	U.park(false, 0L);
1656	>	w.parker = null;
1657	>	U.putObject(wt, PARKBLOCKER, null);
1658	>	}
1659		}
1660		}
1661		return null;
1662		}
1663
1664		/**
1665	<	* If inactivating worker w has caused pool to become quiescent,
1666	<	* check for pool termination, and, so long as this is not the
1667	<	* only worker, wait for event for up to SHRINK_RATE nanosecs On
1668	<	* timeout, if ctl has not changed, terminate the worker, which
1669	<	* will in turn wake up another worker to possibly repeat this
1670	<	* process.
1665	>	* If inactivating worker w has caused the pool to become
1666	>	* quiescent, checks for pool termination, and, so long as this is
1667	>	* not the only worker, waits for event for up to a given
1668	>	* duration. On timeout, if ctl has not changed, terminates the
1669	>	* worker, which will in turn wake up another worker to possibly
1670	>	* repeat this process.
1671		*
1672		* @param w the calling worker
1673	+	* @param currentCtl the ctl value triggering possible quiescence
1674	+	* @param prevCtl the ctl value to restore if thread is terminated
1675		*/
1676	<	private void idleAwaitWork(WorkQueue w) {
1677	<	long c; int nw, ec;
1678	<	if (!tryTerminate(false) &&
1679	<	(int)((c = ctl) >> AC_SHIFT) + parallelism == 0 &&
1680	<	(ec = w.eventCount) == ((int)c \| INT_SIGN) &&
1681	<	(nw = w.nextWait) != 0) {
1682	<	long nc = ((long)(nw & E_MASK) \| // ctl to restore on timeout
1683	<	((c + AC_UNIT) & AC_MASK) \| (c & TC_MASK));
1539	<	ForkJoinTask.helpExpungeStaleExceptions(); // help clean
1540	<	ForkJoinWorkerThread wt = w.owner;
1541	<	while (ctl == c) {
1542	<	long startTime = System.nanoTime();
1676	>	private void idleAwaitWork(WorkQueue w, long currentCtl, long prevCtl) {
1677	>	if (w.eventCount < 0 && !tryTerminate(false, false) &&
1678	>	(int)prevCtl != 0 && !hasQueuedSubmissions() && ctl == currentCtl) {
1679	>	int dc = -(short)(currentCtl >>> TC_SHIFT);
1680	>	long parkTime = dc < 0 ? FAST_IDLE_TIMEOUT: (dc + 1) * IDLE_TIMEOUT;
1681	>	long deadline = System.nanoTime() + parkTime - 100000L; // 1ms slop
1682	>	Thread wt = Thread.currentThread();
1683	>	while (ctl == currentCtl) {
1684		Thread.interrupted(); // timed variant of version in scan()
1685		U.putObject(wt, PARKBLOCKER, this);
1686		w.parker = wt;
1687	<	if (ctl == c)
1688	<	U.park(false, SHRINK_RATE);
1687	>	if (ctl == currentCtl)
1688	>	U.park(false, parkTime);
1689		w.parker = null;
1690		U.putObject(wt, PARKBLOCKER, null);
1691	<	if (ctl != c)
1691	>	if (ctl != currentCtl)
1692		break;
1693	<	if (System.nanoTime() - startTime >= SHRINK_TIMEOUT &&
1694	<	U.compareAndSwapLong(this, CTL, c, nc)) {
1695	<	w.runState = -1; // shrink
1696	<	w.eventCount = (ec + E_SEQ) \| E_MASK;
1693	>	if (deadline - System.nanoTime() <= 0L &&
1694	>	U.compareAndSwapLong(this, CTL, currentCtl, prevCtl)) {
1695	>	w.eventCount = (w.eventCount + E_SEQ) \| E_MASK;
1696	>	w.runState = -1; // shrink
1697		break;
1698		}
1699		}
#	Line 1570 \| Line 1711 \| public class ForkJoinPool extends Abstra
1711		* leaves hints in workers to speed up subsequent calls. The
1712		* implementation is very branchy to cope with potential
1713		* inconsistencies or loops encountering chains that are stale,
1714	<	* unknown, or of length greater than MAX_HELP_DEPTH links. All
1574	<	* of these cases are dealt with by just retrying by caller.
1714	>	* unknown, or so long that they are likely cyclic.
1715		*
1716		* @param joiner the joining worker
1717		* @param task the task to join
1718	<	* @return true if found or ran a task (and so is immediately retryable)
1718	>	* @return 0 if no progress can be made, negative if task
1719	>	* known complete, else positive
1720		*/
1721	<	final boolean tryHelpStealer(WorkQueue joiner, ForkJoinTask<?> task) {
1722	<	ForkJoinTask<?> subtask; // current target
1723	<	boolean progress = false;
1724	<	int depth = 0; // current chain depth
1725	<	int m = runState & SMASK;
1726	<	WorkQueue[] ws = workQueues;
1727	<
1728	<	if (ws != null && ws.length > m && (subtask = task).status >= 0) {
1729	<	outer:for (WorkQueue j = joiner;;) {
1730	<	// Try to find the stealer of subtask, by first using hint
1590	<	WorkQueue stealer = null;
1591	<	WorkQueue v = ws[j.stealHint & m];
1592	<	if (v != null && v.currentSteal == subtask)
1593	<	stealer = v;
1594	<	else {
1595	<	for (int i = 1; i <= m; i += 2) {
1596	<	if ((v = ws[i]) != null && v.currentSteal == subtask) {
1597	<	stealer = v;
1598	<	j.stealHint = i; // save hint
1599	<	break;
1600	<	}
1721	>	private int tryHelpStealer(WorkQueue joiner, ForkJoinTask<?> task) {
1722	>	int stat = 0, steps = 0; // bound to avoid cycles
1723	>	if (joiner != null && task != null) { // hoist null checks
1724	>	restart: for (;;) {
1725	>	ForkJoinTask<?> subtask = task; // current target
1726	>	for (WorkQueue j = joiner, v;;) { // v is stealer of subtask
1727	>	WorkQueue[] ws; int m, s, h;
1728	>	if ((s = task.status) < 0) {
1729	>	stat = s;
1730	>	break restart;
1731		}
1732	<	if (stealer == null)
1733	<	break;
1734	<	}
1735	<
1736	<	for (WorkQueue q = stealer;;) { // Try to help stealer
1737	<	ForkJoinTask<?> t; int b;
1738	<	if (task.status < 0)
1739	<	break outer;
1740	<	if ((b = q.base) - q.top < 0) {
1741	<	progress = true;
1742	<	if (subtask.status < 0)
1743	<	break outer; // stale
1744	<	if ((t = q.pollAt(b)) != null) {
1745	<	stealer.stealHint = joiner.poolIndex;
1746	<	joiner.runSubtask(t);
1732	>	if ((ws = workQueues) == null \|\| (m = ws.length - 1) <= 0)
1733	>	break restart; // shutting down
1734	>	if ((v = ws[h = (j.stealHint \| 1) & m]) == null \|\|
1735	>	v.currentSteal != subtask) {
1736	>	for (int origin = h;;) { // find stealer
1737	>	if (((h = (h + 2) & m) & 15) == 1 &&
1738	>	(subtask.status < 0 \|\| j.currentJoin != subtask))
1739	>	continue restart; // occasional staleness check
1740	>	if ((v = ws[h]) != null &&
1741	>	v.currentSteal == subtask) {
1742	>	j.stealHint = h; // save hint
1743	>	break;
1744	>	}
1745	>	if (h == origin)
1746	>	break restart; // cannot find stealer
1747		}
1748		}
1749	<	else { // empty - try to descend to find stealer's stealer
1750	<	ForkJoinTask<?> next = stealer.currentJoin;
1751	<	if (++depth == MAX_HELP_DEPTH \|\| subtask.status < 0 \|\|
1752	<	next == null \|\| next == subtask)
1753	<	break outer; // max depth, stale, dead-end, cyclic
1754	<	subtask = next;
1755	<	j = stealer;
1756	<	break;
1749	>	for (;;) { // help stealer or descend to its stealer
1750	>	ForkJoinTask[] a; int b;
1751	>	if (subtask.status < 0) // surround probes with
1752	>	continue restart; // consistency checks
1753	>	if ((b = v.base) - v.top < 0 && (a = v.array) != null) {
1754	>	int i = (((a.length - 1) & b) << ASHIFT) + ABASE;
1755	>	ForkJoinTask<?> t =
1756	>	(ForkJoinTask<?>)U.getObjectVolatile(a, i);
1757	>	if (subtask.status < 0 \|\| j.currentJoin != subtask \|\|
1758	>	v.currentSteal != subtask)
1759	>	continue restart; // stale
1760	>	stat = 1; // apparent progress
1761	>	if (t != null && v.base == b &&
1762	>	U.compareAndSwapObject(a, i, t, null)) {
1763	>	v.base = b + 1; // help stealer
1764	>	joiner.runSubtask(t);
1765	>	}
1766	>	else if (v.base == b && ++steps == MAX_HELP)
1767	>	break restart; // v apparently stalled
1768	>	}
1769	>	else { // empty -- try to descend
1770	>	ForkJoinTask<?> next = v.currentJoin;
1771	>	if (subtask.status < 0 \|\| j.currentJoin != subtask \|\|
1772	>	v.currentSteal != subtask)
1773	>	continue restart; // stale
1774	>	else if (next == null \|\| ++steps == MAX_HELP)
1775	>	break restart; // dead-end or maybe cyclic
1776	>	else {
1777	>	subtask = next;
1778	>	j = v;
1779	>	break;
1780	>	}
1781	>	}
1782		}
1783		}
1784		}
1785		}
1786	<	return progress;
1786	>	return stat;
1787		}
1788
1789		/**
#	Line 1637 \| Line 1792 \| public class ForkJoinPool extends Abstra
1792		* @param joiner the joining worker
1793		* @param task the task
1794		*/
1795	<	final void tryPollForAndExec(WorkQueue joiner, ForkJoinTask<?> task) {
1795	>	private void tryPollForAndExec(WorkQueue joiner, ForkJoinTask<?> task) {
1796		WorkQueue[] ws;
1797	<	int m = runState & SMASK;
1798	<	if ((ws = workQueues) != null && ws.length > m) {
1644	<	for (int j = 1; j <= m && task.status >= 0; j += 2) {
1797	>	if ((ws = workQueues) != null) {
1798	>	for (int j = 1; j < ws.length && task.status >= 0; j += 2) {
1799		WorkQueue q = ws[j];
1800		if (q != null && q.pollFor(task)) {
1801		joiner.runSubtask(task);
#	Line 1652 \| Line 1806 \| public class ForkJoinPool extends Abstra
1806		}
1807
1808		/**
1809	<	* Returns a non-empty steal queue, if one is found during a random,
1810	<	* then cyclic scan, else null. This method must be retried by
1811	<	* caller if, by the time it tries to use the queue, it is empty.
1809	>	* Tries to decrement active count (sometimes implicitly) and
1810	>	* possibly release or create a compensating worker in preparation
1811	>	* for blocking. Fails on contention or termination. Otherwise,
1812	>	* adds a new thread if no idle workers are available and either
1813	>	* pool would become completely starved or: (at least half
1814	>	* starved, and fewer than 50% spares exist, and there is at least
1815	>	* one task apparently available). Even though the availability
1816	>	* check requires a full scan, it is worthwhile in reducing false
1817	>	* alarms.
1818	>	*
1819	>	* @param task if non-null, a task being waited for
1820	>	* @param blocker if non-null, a blocker being waited for
1821	>	* @return true if the caller can block, else should recheck and retry
1822	>	*/
1823	>	final boolean tryCompensate(ForkJoinTask<?> task, ManagedBlocker blocker) {
1824	>	int pc = parallelism, e;
1825	>	long c = ctl;
1826	>	WorkQueue[] ws = workQueues;
1827	>	if ((e = (int)c) >= 0 && ws != null) {
1828	>	int u, a, ac, hc;
1829	>	int tc = (short)((u = (int)(c >>> 32)) >>> UTC_SHIFT) + pc;
1830	>	boolean replace = false;
1831	>	if ((a = u >> UAC_SHIFT) <= 0) {
1832	>	if ((ac = a + pc) <= 1)
1833	>	replace = true;
1834	>	else if ((e > 0 \|\| (task != null &&
1835	>	ac <= (hc = pc >>> 1) && tc < pc + hc))) {
1836	>	WorkQueue w;
1837	>	for (int j = 0; j < ws.length; ++j) {
1838	>	if ((w = ws[j]) != null && !w.isEmpty()) {
1839	>	replace = true;
1840	>	break; // in compensation range and tasks available
1841	>	}
1842	>	}
1843	>	}
1844	>	}
1845	>	if ((task == null \|\| task.status >= 0) && // recheck need to block
1846	>	(blocker == null \|\| !blocker.isReleasable()) && ctl == c) {
1847	>	if (!replace) { // no compensation
1848	>	long nc = ((c - AC_UNIT) & AC_MASK) \| (c & ~AC_MASK);
1849	>	if (U.compareAndSwapLong(this, CTL, c, nc))
1850	>	return true;
1851	>	}
1852	>	else if (e != 0) { // release an idle worker
1853	>	WorkQueue w; Thread p; int i;
1854	>	if ((i = e & SMASK) < ws.length && (w = ws[i]) != null) {
1855	>	long nc = ((long)(w.nextWait & E_MASK) \|
1856	>	(c & (AC_MASK\|TC_MASK)));
1857	>	if (w.eventCount == (e \| INT_SIGN) &&
1858	>	U.compareAndSwapLong(this, CTL, c, nc)) {
1859	>	w.eventCount = (e + E_SEQ) & E_MASK;
1860	>	if ((p = w.parker) != null)
1861	>	U.unpark(p);
1862	>	return true;
1863	>	}
1864	>	}
1865	>	}
1866	>	else if (tc < MAX_CAP) { // create replacement
1867	>	long nc = ((c + TC_UNIT) & TC_MASK) \| (c & ~TC_MASK);
1868	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1869	>	addWorker();
1870	>	return true;
1871	>	}
1872	>	}
1873	>	}
1874	>	}
1875	>	return false;
1876	>	}
1877	>
1878	>	/**
1879	>	* Helps and/or blocks until the given task is done.
1880	>	*
1881	>	* @param joiner the joining worker
1882	>	* @param task the task
1883	>	* @return task status on exit
1884	>	*/
1885	>	final int awaitJoin(WorkQueue joiner, ForkJoinTask<?> task) {
1886	>	int s;
1887	>	if ((s = task.status) >= 0) {
1888	>	ForkJoinTask<?> prevJoin = joiner.currentJoin;
1889	>	joiner.currentJoin = task;
1890	>	long startTime = 0L;
1891	>	for (int k = 0;;) {
1892	>	if ((s = (joiner.isEmpty() ? // try to help
1893	>	tryHelpStealer(joiner, task) :
1894	>	joiner.tryRemoveAndExec(task))) == 0 &&
1895	>	(s = task.status) >= 0) {
1896	>	if (k == 0) {
1897	>	startTime = System.nanoTime();
1898	>	tryPollForAndExec(joiner, task); // check uncommon case
1899	>	}
1900	>	else if ((k & (MAX_HELP - 1)) == 0 &&
1901	>	System.nanoTime() - startTime >=
1902	>	COMPENSATION_DELAY &&
1903	>	tryCompensate(task, null)) {
1904	>	if (task.trySetSignal()) {
1905	>	synchronized (task) {
1906	>	if (task.status >= 0) {
1907	>	try { // see ForkJoinTask
1908	>	task.wait(); // for explanation
1909	>	} catch (InterruptedException ie) {
1910	>	}
1911	>	}
1912	>	else
1913	>	task.notifyAll();
1914	>	}
1915	>	}
1916	>	long c; // re-activate
1917	>	do {} while (!U.compareAndSwapLong
1918	>	(this, CTL, c = ctl, c + AC_UNIT));
1919	>	}
1920	>	}
1921	>	if (s < 0 \|\| (s = task.status) < 0) {
1922	>	joiner.currentJoin = prevJoin;
1923	>	break;
1924	>	}
1925	>	else if ((k++ & (MAX_HELP - 1)) == MAX_HELP >>> 1)
1926	>	Thread.yield(); // for politeness
1927	>	}
1928	>	}
1929	>	return s;
1930	>	}
1931	>
1932	>	/**
1933	>	* Stripped-down variant of awaitJoin used by timed joins. Tries
1934	>	* to help join only while there is continuous progress. (Caller
1935	>	* will then enter a timed wait.)
1936	>	*
1937	>	* @param joiner the joining worker
1938	>	* @param task the task
1939	>	* @return task status on exit
1940	>	*/
1941	>	final int helpJoinOnce(WorkQueue joiner, ForkJoinTask<?> task) {
1942	>	int s;
1943	>	while ((s = task.status) >= 0 &&
1944	>	(joiner.isEmpty() ?
1945	>	tryHelpStealer(joiner, task) :
1946	>	joiner.tryRemoveAndExec(task)) != 0)
1947	>	;
1948	>	return s;
1949	>	}
1950	>
1951	>	/**
1952	>	* Returns a (probably) non-empty steal queue, if one is found
1953	>	* during a random, then cyclic scan, else null. This method must
1954	>	* be retried by caller if, by the time it tries to use the queue,
1955	>	* it is empty.
1956		*/
1957		private WorkQueue findNonEmptyStealQueue(WorkQueue w) {
1958	<	int r = w.seed; // Same idea as scan(), but ignoring submissions
1958	>	// Similar to loop in scan(), but ignoring submissions
1959	>	int r;
1960	>	if (w == null) // allow external callers
1961	>	r = ThreadLocalRandom.current().nextInt();
1962	>	else {
1963	>	r = w.seed; r ^= r << 13; r ^= r >>> 17; w.seed = r ^= r << 5;
1964	>	}
1965	>	int step = (r >>> 16) \| 1;
1966		for (WorkQueue[] ws;;) {
1967	<	int m = runState & SMASK;
1968	<	if ((ws = workQueues) == null)
1967	>	int rs = runState, m;
1968	>	if ((ws = workQueues) == null \|\| (m = ws.length - 1) < 1)
1969		return null;
1970	<	if (ws.length > m) {
1971	<	WorkQueue q;
1972	<	for (int n = m << 2, k = r, j = -n;;) {
1973	<	r ^= r << 13; r ^= r >>> 17; r ^= r << 5;
1974	<	if ((q = ws[(k \| 1) & m]) != null && q.base - q.top < 0) {
1975	<	w.seed = r;
1671	<	return q;
1672	<	}
1673	<	else if (j > n)
1970	>	for (int j = (m + 1) << 2; ; r += step) {
1971	>	WorkQueue q = ws[((r << 1) \| 1) & m];
1972	>	if (q != null && !q.isEmpty())
1973	>	return q;
1974	>	else if (--j < 0) {
1975	>	if (runState == rs)
1976		return null;
1977	<	else
1676	<	k = (j++ < 0) ? r : k + ((m >>> 1) \| 1);
1677	<
1977	>	break;
1978		}
1979		}
1980		}
#	Line 1688 \| Line 1988 \| public class ForkJoinPool extends Abstra
1988		*/
1989		final void helpQuiescePool(WorkQueue w) {
1990		for (boolean active = true;;) {
1991	<	w.runLocalTasks(); // exhaust local queue
1991	>	ForkJoinTask<?> localTask; // exhaust local queue
1992	>	while ((localTask = w.nextLocalTask()) != null)
1993	>	localTask.doExec();
1994		WorkQueue q = findNonEmptyStealQueue(w);
1995		if (q != null) {
1996	<	ForkJoinTask<?> t;
1996	>	ForkJoinTask<?> t; int b;
1997		if (!active) { // re-establish active count
1998		long c;
1999		active = true;
2000		do {} while (!U.compareAndSwapLong
2001		(this, CTL, c = ctl, c + AC_UNIT));
2002		}
2003	<	if ((t = q.poll()) != null)
2003	>	if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null)
2004		w.runSubtask(t);
2005		}
2006		else {
#	Line 1720 \| Line 2022 \| public class ForkJoinPool extends Abstra
2022		}
2023
2024		/**
2025	<	* Gets and removes a local or stolen task for the given worker
2025	>	* Restricted version of helpQuiescePool for non-FJ callers
2026	>	*/
2027	>	static void externalHelpQuiescePool() {
2028	>	ForkJoinPool p; WorkQueue[] ws; WorkQueue w, q;
2029	>	ForkJoinTask<?> t; int b;
2030	>	int k = submitters.get().seed & SQMASK;
2031	>	if ((p = commonPool) != null &&
2032	>	(ws = p.workQueues) != null &&
2033	>	ws.length > (k &= p.submitMask) &&
2034	>	(w = ws[k]) != null &&
2035	>	(q = p.findNonEmptyStealQueue(w)) != null &&
2036	>	(b = q.base) - q.top < 0 &&
2037	>	(t = q.pollAt(b)) != null)
2038	>	t.doExec();
2039	>	}
2040	>
2041	>	/**
2042	>	* Gets and removes a local or stolen task for the given worker.
2043		*
2044		* @return a task, if available
2045		*/
2046		final ForkJoinTask<?> nextTaskFor(WorkQueue w) {
2047		for (ForkJoinTask<?> t;;) {
2048	<	WorkQueue q;
2048	>	WorkQueue q; int b;
2049		if ((t = w.nextLocalTask()) != null)
2050		return t;
2051		if ((q = findNonEmptyStealQueue(w)) == null)
2052		return null;
2053	<	if ((t = q.poll()) != null)
2053	>	if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null)
2054		return t;
2055		}
2056		}
#	Line 1752 \| Line 2071 \| public class ForkJoinPool extends Abstra
2071		8);
2072		}
2073
1755	–	// Termination
1756	–
2074		/**
2075	<	* Sets SHUTDOWN bit of runState under lock
2075	>	* Returns approximate submission queue length for the given caller
2076		*/
2077	<	private void enableShutdown() {
2078	<	ReentrantLock lock = this.lock;
2079	<	if (runState >= 0) {
2080	<	lock.lock(); // don't need try/finally
2081	<	runState \|= SHUTDOWN;
2082	<	lock.unlock();
2083	<	}
2077	>	static int getEstimatedSubmitterQueueLength() {
2078	>	ForkJoinPool p; WorkQueue[] ws; WorkQueue q;
2079	>	int k = submitters.get().seed & SQMASK;
2080	>	return ((p = commonPool) != null &&
2081	>	p.runState >= 0 &&
2082	>	(ws = p.workQueues) != null &&
2083	>	ws.length > (k &= p.submitMask) &&
2084	>	(q = ws[k]) != null) ?
2085	>	q.queueSize() : 0;
2086		}
2087
2088	+	// Termination
2089	+
2090		/**
2091	<	* Possibly initiates and/or completes termination. Upon
2092	<	* termination, cancels all queued tasks and then
2091	>	* Possibly initiates and/or completes termination. The caller
2092	>	* triggering termination runs three passes through workQueues:
2093	>	* (0) Setting termination status, followed by wakeups of queued
2094	>	* workers; (1) cancelling all tasks; (2) interrupting lagging
2095	>	* threads (likely in external tasks, but possibly also blocked in
2096	>	* joins). Each pass repeats previous steps because of potential
2097	>	* lagging thread creation.
2098		*
2099		* @param now if true, unconditionally terminate, else only
2100		* if no work and no active workers
2101	+	* @param enable if true, enable shutdown when next possible
2102		* @return true if now terminating or terminated
2103		*/
2104	<	private boolean tryTerminate(boolean now) {
2104	>	private boolean tryTerminate(boolean now, boolean enable) {
2105	>	Mutex lock = this.lock;
2106		for (long c;;) {
2107		if (((c = ctl) & STOP_BIT) != 0) { // already terminating
2108		if ((short)(c >>> TC_SHIFT) == -parallelism) {
1781	–	ReentrantLock lock = this.lock; // signal when no workers
2109		lock.lock(); // don't need try/finally
2110		termination.signalAll(); // signal when 0 workers
2111		lock.unlock();
2112		}
2113		return true;
2114		}
2115	<	if (!now) {
2116	<	if ((int)(c >> AC_SHIFT) != -parallelism \|\| runState >= 0 \|\|
2115	>	if (runState >= 0) { // not yet enabled
2116	>	if (!enable)
2117	>	return false;
2118	>	lock.lock();
2119	>	runState \|= SHUTDOWN;
2120	>	lock.unlock();
2121	>	}
2122	>	if (!now) { // check if idle & no tasks
2123	>	if ((int)(c >> AC_SHIFT) != -parallelism \|\|
2124		hasQueuedSubmissions())
2125		return false;
2126		// Check for unqueued inactive workers. One pass suffices.
2127		WorkQueue[] ws = workQueues; WorkQueue w;
2128		if (ws != null) {
2129	<	int n = ws.length;
1796	<	for (int i = 1; i < n; i += 2) {
2129	>	for (int i = 1; i < ws.length; i += 2) {
2130		if ((w = ws[i]) != null && w.eventCount >= 0)
2131		return false;
2132		}
2133		}
2134		}
2135	<	if (U.compareAndSwapLong(this, CTL, c, c \| STOP_BIT))
2136	<	startTerminating();
2137	<	}
2138	<	}
2139	<
2140	<	/**
2141	<	* Initiates termination: Runs three passes through workQueues:
2142	<	* (0) Setting termination status, followed by wakeups of queued
2143	<	* workers; (1) cancelling all tasks; (2) interrupting lagging
2144	<	* threads (likely in external tasks, but possibly also blocked in
2145	<	* joins). Each pass repeats previous steps because of potential
2146	<	* lagging thread creation.
2147	<	*/
1815	<	private void startTerminating() {
1816	<	for (int pass = 0; pass < 3; ++pass) {
1817	<	WorkQueue[] ws = workQueues;
1818	<	if (ws != null) {
1819	<	WorkQueue w; Thread wt;
1820	<	int n = ws.length;
1821	<	for (int j = 0; j < n; ++j) {
1822	<	if ((w = ws[j]) != null) {
1823	<	w.runState = -1;
1824	<	if (pass > 0) {
1825	<	w.cancelAll();
1826	<	if (pass > 1 && (wt = w.owner) != null &&
1827	<	!wt.isInterrupted()) {
1828	<	try {
1829	<	wt.interrupt();
1830	<	} catch (SecurityException ignore) {
2135	>	if (U.compareAndSwapLong(this, CTL, c, c \| STOP_BIT)) {
2136	>	for (int pass = 0; pass < 3; ++pass) {
2137	>	WorkQueue[] ws = workQueues;
2138	>	if (ws != null) {
2139	>	WorkQueue w;
2140	>	int n = ws.length;
2141	>	for (int i = 0; i < n; ++i) {
2142	>	if ((w = ws[i]) != null) {
2143	>	w.runState = -1;
2144	>	if (pass > 0) {
2145	>	w.cancelAll();
2146	>	if (pass > 1)
2147	>	w.interruptOwner();
2148		}
2149		}
2150		}
2151	<	}
2152	<	}
2153	<	// Wake up workers parked on event queue
2154	<	int i, e; long c; Thread p;
2155	<	while ((i = ((~(e = (int)(c = ctl)) << 1) \| 1) & SMASK) < n &&
2156	<	(w = ws[i]) != null &&
2157	<	w.eventCount == (e \| INT_SIGN)) {
2158	<	long nc = ((long)(w.nextWait & E_MASK) \|
2159	<	((c + AC_UNIT) & AC_MASK) \|
2160	<	(c & (TC_MASK\|STOP_BIT)));
2161	<	if (U.compareAndSwapLong(this, CTL, c, nc)) {
2162	<	w.eventCount = (e + E_SEQ) & E_MASK;
2163	<	if ((p = w.parker) != null)
2164	<	U.unpark(p);
2151	>	// Wake up workers parked on event queue
2152	>	int i, e; long cc; Thread p;
2153	>	while ((e = (int)(cc = ctl) & E_MASK) != 0 &&
2154	>	(i = e & SMASK) < n &&
2155	>	(w = ws[i]) != null) {
2156	>	long nc = ((long)(w.nextWait & E_MASK) \|
2157	>	((cc + AC_UNIT) & AC_MASK) \|
2158	>	(cc & (TC_MASK\|STOP_BIT)));
2159	>	if (w.eventCount == (e \| INT_SIGN) &&
2160	>	U.compareAndSwapLong(this, CTL, cc, nc)) {
2161	>	w.eventCount = (e + E_SEQ) & E_MASK;
2162	>	w.runState = -1;
2163	>	if ((p = w.parker) != null)
2164	>	U.unpark(p);
2165	>	}
2166	>	}
2167		}
2168		}
2169		}
#	Line 1920 \| Line 2239 \| public class ForkJoinPool extends Abstra
2239		checkPermission();
2240		if (factory == null)
2241		throw new NullPointerException();
2242	<	if (parallelism <= 0 \|\| parallelism > MAX_ID)
2242	>	if (parallelism <= 0 \|\| parallelism > MAX_CAP)
2243		throw new IllegalArgumentException();
2244		this.parallelism = parallelism;
2245		this.factory = factory;
2246		this.ueh = handler;
2247		this.localMode = asyncMode ? FIFO_QUEUE : LIFO_QUEUE;
1929	–	this.nextPoolIndex = 1;
2248		long np = (long)(-parallelism); // offset ctl counts
2249		this.ctl = ((np << AC_SHIFT) & AC_MASK) \| ((np << TC_SHIFT) & TC_MASK);
2250	<	// initialize workQueues array with room for 2*parallelism if possible
2251	<	int n = parallelism << 1;
2252	<	if (n >= MAX_ID)
2253	<	n = MAX_ID;
2254	<	else { // See Hackers Delight, sec 3.2, where n < (1 << 16)
2255	<	n \|= n >>> 1; n \|= n >>> 2; n \|= n >>> 4; n \|= n >>> 8;
2256	<	}
1939	<	this.workQueues = new WorkQueue[(n + 1) << 1];
1940	<	ReentrantLock lck = this.lock = new ReentrantLock();
1941	<	this.termination = lck.newCondition();
2250	>	// Use nearest power 2 for workQueues size. See Hackers Delight sec 3.2.
2251	>	int n = parallelism - 1;
2252	>	n \|= n >>> 1; n \|= n >>> 2; n \|= n >>> 4; n \|= n >>> 8; n \|= n >>> 16;
2253	>	int size = (n + 1) << 1; // #slots = 2*#workers
2254	>	this.submitMask = size - 1; // room for max # of submit queues
2255	>	this.workQueues = new WorkQueue[size];
2256	>	this.termination = (this.lock = new Mutex()).newCondition();
2257		this.stealCount = new AtomicLong();
2258		this.nextWorkerNumber = new AtomicInteger();
2259	+	int pn = poolNumberGenerator.incrementAndGet();
2260		StringBuilder sb = new StringBuilder("ForkJoinPool-");
2261	<	sb.append(poolNumberGenerator.incrementAndGet());
2261	>	sb.append(Integer.toString(pn));
2262		sb.append("-worker-");
2263		this.workerNamePrefix = sb.toString();
2264	<	// Create initial submission queue
2265	<	WorkQueue sq = tryAddSharedQueue(0);
2266	<	if (sq != null)
2267	<	sq.growArray(false);
2264	>	lock.lock();
2265	>	this.runState = 1; // set init flag
2266	>	lock.unlock();
2267	>	}
2268	>
2269	>	/**
2270	>	* Returns the common pool instance
2271	>	*
2272	>	* @return the common pool instance
2273	>	*/
2274	>	public static ForkJoinPool commonPool() {
2275	>	ForkJoinPool p;
2276	>	return (p = commonPool) != null? p : ensureCommonPool();
2277	>	}
2278	>
2279	>	private static ForkJoinPool ensureCommonPool() {
2280	>	ForkJoinPool p;
2281	>	if ((p = commonPool) == null) {
2282	>	final Mutex lock = initializationLock;
2283	>	lock.lock();
2284	>	try {
2285	>	if ((p = commonPool) == null) {
2286	>	p = commonPool = new ForkJoinPool(commonPoolParallelism,
2287	>	commonPoolFactory,
2288	>	commonPoolUEH, false);
2289	>	// use a more informative name string for workers
2290	>	p.workerNamePrefix = "ForkJoinPool.commonPool-worker-";
2291	>	}
2292	>	} finally {
2293	>	lock.unlock();
2294	>	}
2295	>	}
2296	>	return p;
2297		}
2298
2299		// Execution methods
#	Line 1970 \| Line 2315 \| public class ForkJoinPool extends Abstra
2315		* scheduled for execution
2316		*/
2317		public <T> T invoke(ForkJoinTask<T> task) {
2318	+	if (task == null)
2319	+	throw new NullPointerException();
2320		doSubmit(task);
2321		return task.join();
2322		}
#	Line 1983 \| Line 2330 \| public class ForkJoinPool extends Abstra
2330		* scheduled for execution
2331		*/
2332		public void execute(ForkJoinTask<?> task) {
2333	+	if (task == null)
2334	+	throw new NullPointerException();
2335		doSubmit(task);
2336		}
2337
#	Line 2000 \| Line 2349 \| public class ForkJoinPool extends Abstra
2349		if (task instanceof ForkJoinTask<?>) // avoid re-wrap
2350		job = (ForkJoinTask<?>) task;
2351		else
2352	<	job = ForkJoinTask.adapt(task, null);
2352	>	job = new ForkJoinTask.AdaptedRunnableAction(task);
2353		doSubmit(job);
2354		}
2355
#	Line 2014 \| Line 2363 \| public class ForkJoinPool extends Abstra
2363		* scheduled for execution
2364		*/
2365		public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) {
2366	+	if (task == null)
2367	+	throw new NullPointerException();
2368		doSubmit(task);
2369		return task;
2370		}
#	Line 2024 \| Line 2375 \| public class ForkJoinPool extends Abstra
2375		* scheduled for execution
2376		*/
2377		public <T> ForkJoinTask<T> submit(Callable<T> task) {
2378	<	if (task == null)
2028	<	throw new NullPointerException();
2029	<	ForkJoinTask<T> job = ForkJoinTask.adapt(task);
2378	>	ForkJoinTask<T> job = new ForkJoinTask.AdaptedCallable<T>(task);
2379		doSubmit(job);
2380		return job;
2381		}
#	Line 2037 \| Line 2386 \| public class ForkJoinPool extends Abstra
2386		* scheduled for execution
2387		*/
2388		public <T> ForkJoinTask<T> submit(Runnable task, T result) {
2389	<	if (task == null)
2041	<	throw new NullPointerException();
2042	<	ForkJoinTask<T> job = ForkJoinTask.adapt(task, result);
2389	>	ForkJoinTask<T> job = new ForkJoinTask.AdaptedRunnable<T>(task, result);
2390		doSubmit(job);
2391		return job;
2392		}
#	Line 2056 \| Line 2403 \| public class ForkJoinPool extends Abstra
2403		if (task instanceof ForkJoinTask<?>) // avoid re-wrap
2404		job = (ForkJoinTask<?>) task;
2405		else
2406	<	job = ForkJoinTask.adapt(task, null);
2406	>	job = new ForkJoinTask.AdaptedRunnableAction(task);
2407		doSubmit(job);
2408		return job;
2409		}
#	Line 2066 \| Line 2413 \| public class ForkJoinPool extends Abstra
2413		* @throws RejectedExecutionException {@inheritDoc}
2414		*/
2415		public <T> List<Future<T>> invokeAll(Collection<? extends Callable<T>> tasks) {
2416	<	ArrayList<ForkJoinTask<T>> forkJoinTasks =
2417	<	new ArrayList<ForkJoinTask<T>>(tasks.size());
2418	<	for (Callable<T> task : tasks)
2419	<	forkJoinTasks.add(ForkJoinTask.adapt(task));
2420	<	invoke(new InvokeAll<T>(forkJoinTasks));
2421	<
2416	>	// In previous versions of this class, this method constructed
2417	>	// a task to run ForkJoinTask.invokeAll, but now external
2418	>	// invocation of multiple tasks is at least as efficient.
2419	>	List<ForkJoinTask<T>> fs = new ArrayList<ForkJoinTask<T>>(tasks.size());
2420	>	// Workaround needed because method wasn't declared with
2421	>	// wildcards in return type but should have been.
2422		@SuppressWarnings({"unchecked", "rawtypes"})
2423	<	List<Future<T>> futures = (List<Future<T>>) (List) forkJoinTasks;
2077	<	return futures;
2078	<	}
2423	>	List<Future<T>> futures = (List<Future<T>>) (List) fs;
2424
2425	<	static final class InvokeAll<T> extends RecursiveAction {
2426	<	final ArrayList<ForkJoinTask<T>> tasks;
2427	<	InvokeAll(ArrayList<ForkJoinTask<T>> tasks) { this.tasks = tasks; }
2428	<	public void compute() {
2429	<	try { invokeAll(tasks); }
2430	<	catch (Exception ignore) {}
2425	>	boolean done = false;
2426	>	try {
2427	>	for (Callable<T> t : tasks) {
2428	>	ForkJoinTask<T> f = new ForkJoinTask.AdaptedCallable<T>(t);
2429	>	doSubmit(f);
2430	>	fs.add(f);
2431	>	}
2432	>	for (ForkJoinTask<T> f : fs)
2433	>	f.quietlyJoin();
2434	>	done = true;
2435	>	return futures;
2436	>	} finally {
2437	>	if (!done)
2438	>	for (ForkJoinTask<T> f : fs)
2439	>	f.cancel(false);
2440		}
2087	–	private static final long serialVersionUID = -7914297376763021607L;
2441		}
2442
2443		/**
#	Line 2116 \| Line 2469 \| public class ForkJoinPool extends Abstra
2469		}
2470
2471		/**
2472	+	* Returns the targeted parallelism level of the common pool.
2473	+	*
2474	+	* @return the targeted parallelism level of the common pool
2475	+	*/
2476	+	public static int getCommonPoolParallelism() {
2477	+	return commonPoolParallelism;
2478	+	}
2479	+
2480	+	/**
2481		* Returns the number of worker threads that have started but not
2482		* yet terminated. The result returned by this method may differ
2483		* from {@link #getParallelism} when threads are created to
#	Line 2149 \| Line 2511 \| public class ForkJoinPool extends Abstra
2511		int rc = 0;
2512		WorkQueue[] ws; WorkQueue w;
2513		if ((ws = workQueues) != null) {
2514	<	int n = ws.length;
2515	<	for (int i = 1; i < n; i += 2) {
2154	<	Thread.State s; ForkJoinWorkerThread wt;
2155	<	if ((w = ws[i]) != null && (wt = w.owner) != null &&
2156	<	w.eventCount >= 0 &&
2157	<	(s = wt.getState()) != Thread.State.BLOCKED &&
2158	<	s != Thread.State.WAITING &&
2159	<	s != Thread.State.TIMED_WAITING)
2514	>	for (int i = 1; i < ws.length; i += 2) {
2515	>	if ((w = ws[i]) != null && w.isApparentlyUnblocked())
2516		++rc;
2517		}
2518		}
#	Line 2205 \| Line 2561 \| public class ForkJoinPool extends Abstra
2561		long count = stealCount.get();
2562		WorkQueue[] ws; WorkQueue w;
2563		if ((ws = workQueues) != null) {
2564	<	int n = ws.length;
2209	<	for (int i = 1; i < n; i += 2) {
2564	>	for (int i = 1; i < ws.length; i += 2) {
2565		if ((w = ws[i]) != null)
2566		count += w.totalSteals;
2567		}
#	Line 2228 \| Line 2583 \| public class ForkJoinPool extends Abstra
2583		long count = 0;
2584		WorkQueue[] ws; WorkQueue w;
2585		if ((ws = workQueues) != null) {
2586	<	int n = ws.length;
2232	<	for (int i = 1; i < n; i += 2) {
2586	>	for (int i = 1; i < ws.length; i += 2) {
2587		if ((w = ws[i]) != null)
2588		count += w.queueSize();
2589		}
#	Line 2248 \| Line 2602 \| public class ForkJoinPool extends Abstra
2602		int count = 0;
2603		WorkQueue[] ws; WorkQueue w;
2604		if ((ws = workQueues) != null) {
2605	<	int n = ws.length;
2252	<	for (int i = 0; i < n; i += 2) {
2605	>	for (int i = 0; i < ws.length; i += 2) {
2606		if ((w = ws[i]) != null)
2607		count += w.queueSize();
2608		}
#	Line 2266 \| Line 2619 \| public class ForkJoinPool extends Abstra
2619		public boolean hasQueuedSubmissions() {
2620		WorkQueue[] ws; WorkQueue w;
2621		if ((ws = workQueues) != null) {
2622	<	int n = ws.length;
2623	<	for (int i = 0; i < n; i += 2) {
2271	<	if ((w = ws[i]) != null && w.queueSize() != 0)
2622	>	for (int i = 0; i < ws.length; i += 2) {
2623	>	if ((w = ws[i]) != null && !w.isEmpty())
2624		return true;
2625		}
2626		}
#	Line 2285 \| Line 2637 \| public class ForkJoinPool extends Abstra
2637		protected ForkJoinTask<?> pollSubmission() {
2638		WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2639		if ((ws = workQueues) != null) {
2640	<	int n = ws.length;
2289	<	for (int i = 0; i < n; i += 2) {
2640	>	for (int i = 0; i < ws.length; i += 2) {
2641		if ((w = ws[i]) != null && (t = w.poll()) != null)
2642		return t;
2643		}
#	Line 2315 \| Line 2666 \| public class ForkJoinPool extends Abstra
2666		int count = 0;
2667		WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2668		if ((ws = workQueues) != null) {
2669	<	int n = ws.length;
2319	<	for (int i = 0; i < n; ++i) {
2669	>	for (int i = 0; i < ws.length; ++i) {
2670		if ((w = ws[i]) != null) {
2671		while ((t = w.poll()) != null) {
2672		c.add(t);
#	Line 2336 \| Line 2686 \| public class ForkJoinPool extends Abstra
2686		* @return a string identifying this pool, as well as its state
2687		*/
2688		public String toString() {
2689	<	long st = getStealCount();
2690	<	long qt = getQueuedTaskCount();
2691	<	long qs = getQueuedSubmissionCount();
2342	<	int rc = getRunningThreadCount();
2343	<	int pc = parallelism;
2689	>	// Use a single pass through workQueues to collect counts
2690	>	long qt = 0L, qs = 0L; int rc = 0;
2691	>	long st = stealCount.get();
2692		long c = ctl;
2693	+	WorkQueue[] ws; WorkQueue w;
2694	+	if ((ws = workQueues) != null) {
2695	+	for (int i = 0; i < ws.length; ++i) {
2696	+	if ((w = ws[i]) != null) {
2697	+	int size = w.queueSize();
2698	+	if ((i & 1) == 0)
2699	+	qs += size;
2700	+	else {
2701	+	qt += size;
2702	+	st += w.totalSteals;
2703	+	if (w.isApparentlyUnblocked())
2704	+	++rc;
2705	+	}
2706	+	}
2707	+	}
2708	+	}
2709	+	int pc = parallelism;
2710		int tc = pc + (short)(c >>> TC_SHIFT);
2711		int ac = pc + (int)(c >> AC_SHIFT);
2712		if (ac < 0) // ignore transient negative
#	Line 2364 \| Line 2729 \| public class ForkJoinPool extends Abstra
2729		}
2730
2731		/**
2732	<	* Initiates an orderly shutdown in which previously submitted
2733	<	* tasks are executed, but no new tasks will be accepted.
2734	<	* Invocation has no additional effect if already shut down.
2735	<	* Tasks that are in the process of being submitted concurrently
2736	<	* during the course of this method may or may not be rejected.
2732	>	* Possibly initiates an orderly shutdown in which previously
2733	>	* submitted tasks are executed, but no new tasks will be
2734	>	* accepted. Invocation has no effect on execution state if this
2735	>	* is the {@link #commonPool}, and no additional effect if
2736	>	* already shut down. Tasks that are in the process of being
2737	>	* submitted concurrently during the course of this method may or
2738	>	* may not be rejected.
2739		*
2740		* @throws SecurityException if a security manager exists and
2741		* the caller is not permitted to modify threads
#	Line 2377 \| Line 2744 \| public class ForkJoinPool extends Abstra
2744		*/
2745		public void shutdown() {
2746		checkPermission();
2747	<	enableShutdown();
2748	<	tryTerminate(false);
2747	>	if (this != commonPool)
2748	>	tryTerminate(false, true);
2749		}
2750
2751		/**
2752	<	* Attempts to cancel and/or stop all tasks, and reject all
2753	<	* subsequently submitted tasks. Tasks that are in the process of
2754	<	* being submitted or executed concurrently during the course of
2755	<	* this method may or may not be rejected. This method cancels
2756	<	* both existing and unexecuted tasks, in order to permit
2757	<	* termination in the presence of task dependencies. So the method
2758	<	* always returns an empty list (unlike the case for some other
2759	<	* Executors).
2752	>	* Possibly attempts to cancel and/or stop all tasks, and reject
2753	>	* all subsequently submitted tasks. Invocation has no effect on
2754	>	* execution state if this is the {@link #commonPool}, and no
2755	>	* additional effect if already shut down. Otherwise, tasks that
2756	>	* are in the process of being submitted or executed concurrently
2757	>	* during the course of this method may or may not be
2758	>	* rejected. This method cancels both existing and unexecuted
2759	>	* tasks, in order to permit termination in the presence of task
2760	>	* dependencies. So the method always returns an empty list
2761	>	* (unlike the case for some other Executors).
2762		*
2763		* @return an empty list
2764		* @throws SecurityException if a security manager exists and
#	Line 2399 \| Line 2768 \| public class ForkJoinPool extends Abstra
2768		*/
2769		public List<Runnable> shutdownNow() {
2770		checkPermission();
2771	<	enableShutdown();
2772	<	tryTerminate(true);
2771	>	if (this != commonPool)
2772	>	tryTerminate(true, true);
2773		return Collections.emptyList();
2774		}
2775
#	Line 2457 \| Line 2826 \| public class ForkJoinPool extends Abstra
2826		public boolean awaitTermination(long timeout, TimeUnit unit)
2827		throws InterruptedException {
2828		long nanos = unit.toNanos(timeout);
2829	<	final ReentrantLock lock = this.lock;
2829	>	final Mutex lock = this.lock;
2830		lock.lock();
2831		try {
2832		for (;;) {
#	Line 2553 \| Line 2922 \| public class ForkJoinPool extends Abstra
2922		*
2923		* <p>If the caller is not a {@link ForkJoinTask}, this method is
2924		* behaviorally equivalent to
2925	<	a * <pre> {@code
2925	>	* <pre> {@code
2926		* while (!blocker.isReleasable())
2927		* if (blocker.block())
2928		* return;
#	Line 2571 \| Line 2940 \| *a <pre> {@code**
2940		ForkJoinPool p = ((t instanceof ForkJoinWorkerThread) ?
2941		((ForkJoinWorkerThread)t).pool : null);
2942		while (!blocker.isReleasable()) {
2943	<	if (p == null \|\| p.tryCompensate()) {
2943	>	if (p == null \|\| p.tryCompensate(null, blocker)) {
2944		try {
2945		do {} while (!blocker.isReleasable() && !blocker.block());
2946		} finally {
#	Line 2588 \| Line 2957 \| *a <pre> {@code**
2957		// implement RunnableFuture.
2958
2959		protected <T> RunnableFuture<T> newTaskFor(Runnable runnable, T value) {
2960	<	return (RunnableFuture<T>) ForkJoinTask.adapt(runnable, value);
2960	>	return new ForkJoinTask.AdaptedRunnable<T>(runnable, value);
2961		}
2962
2963		protected <T> RunnableFuture<T> newTaskFor(Callable<T> callable) {
2964	<	return (RunnableFuture<T>) ForkJoinTask.adapt(callable);
2964	>	return new ForkJoinTask.AdaptedCallable<T>(callable);
2965		}
2966
2967		// Unsafe mechanics
2968		private static final sun.misc.Unsafe U;
2969		private static final long CTL;
2601	–	private static final long RUNSTATE;
2970		private static final long PARKBLOCKER;
2971	+	private static final int ABASE;
2972	+	private static final int ASHIFT;
2973
2974		static {
2975		poolNumberGenerator = new AtomicInteger();
2976	+	nextSubmitterSeed = new AtomicInteger(0x55555555);
2977		modifyThreadPermission = new RuntimePermission("modifyThread");
2978		defaultForkJoinWorkerThreadFactory =
2979		new DefaultForkJoinWorkerThreadFactory();
2980	+	submitters = new ThreadSubmitter();
2981	+	initializationLock = new Mutex();
2982		int s;
2983		try {
2984		U = getUnsafe();
2985		Class<?> k = ForkJoinPool.class;
2986	<	Class<?> tk = Thread.class;
2986	>	Class<?> ak = ForkJoinTask[].class;
2987		CTL = U.objectFieldOffset
2988		(k.getDeclaredField("ctl"));
2989	<	RUNSTATE = U.objectFieldOffset
2617	<	(k.getDeclaredField("runState"));
2989	>	Class<?> tk = Thread.class;
2990		PARKBLOCKER = U.objectFieldOffset
2991		(tk.getDeclaredField("parkBlocker"));
2992	+	ABASE = U.arrayBaseOffset(ak);
2993	+	s = U.arrayIndexScale(ak);
2994	+	} catch (Exception e) {
2995	+	throw new Error(e);
2996	+	}
2997	+	if ((s & (s-1)) != 0)
2998	+	throw new Error("data type scale not a power of two");
2999	+	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
3000	+
3001	+	// Establish configuration for default pool
3002	+	try {
3003	+	String pp = System.getProperty(propPrefix + "parallelism");
3004	+	String fp = System.getProperty(propPrefix + "threadFactory");
3005	+	String up = System.getProperty(propPrefix + "exceptionHandler");
3006	+	int par;
3007	+	if ((pp == null \|\| (par = Integer.parseInt(pp)) <= 0))
3008	+	par = Runtime.getRuntime().availableProcessors();
3009	+	commonPoolParallelism = par;
3010	+	if (fp != null)
3011	+	commonPoolFactory = (ForkJoinWorkerThreadFactory)
3012	+	ClassLoader.getSystemClassLoader().loadClass(fp).newInstance();
3013	+	else
3014	+	commonPoolFactory = defaultForkJoinWorkerThreadFactory;
3015	+	if (up != null)
3016	+	commonPoolUEH = (Thread.UncaughtExceptionHandler)
3017	+	ClassLoader.getSystemClassLoader().loadClass(up).newInstance();
3018	+	else
3019	+	commonPoolUEH = null;
3020		} catch (Exception e) {
3021		throw new Error(e);
3022		}

Diff Legend

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

Comparing jsr166/src/jsr166y/ForkJoinPool.java (file contents): Revision 1.112 by dl, Thu Jan 26 18:15:12 2012 UTC vs. Revision 1.135 by dl, Sun Oct 28 22:36:01 2012 UTC

Diff Legend

Comparing jsr166/src/jsr166y/ForkJoinPool.java (file contents):
Revision 1.112 by dl, Thu Jan 26 18:15:12 2012 UTC vs.
Revision 1.135 by dl, Sun Oct 28 22:36:01 2012 UTC