ViewVC Help

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

root/jsr166/jsr166/src/main/java/util/concurrent/ForkJoinPool.java

Comparing jsr166/src/main/java/util/concurrent/ForkJoinPool.java (file contents):
Revision 1.18 by dl, Wed Jul 7 20:41:24 2010 UTC vs.
Revision 1.19 by dl, Wed Aug 11 18:45:45 2010 UTC

#	Line 50 | Line 50 | import java.util.concurrent.CountDownLat
50		* convenient form for informal monitoring.
51		*
52		* <p> As is the case with other ExecutorServices, there are three
53	<	* main task execution methods summarized in the follwoing
53	>	* main task execution methods summarized in the following
54		* table. These are designed to be used by clients not already engaged
55		* in fork/join computations in the current pool.  The main forms of
56		* these methods accept instances of {@code ForkJoinTask}, but
#	Line 58 | Line 58 | import java.util.concurrent.CountDownLat
58		* Runnable}- or {@code Callable}- based activities as well.  However,
59		* tasks that are already executing in a pool should normally
60		* <em>NOT</em> use these pool execution methods, but instead use the
61	<	* within-computation forms listed in the table. To avoid inadvertant
62	<	* cyclic task dependencies and to improve performance, task
63	<	* submissions to the current pool by an ongoing fork/join
64	<	* computations may be implicitly translated to the corresponding
65	<	* ForkJoinTask forms.
61	>	* within-computation forms listed in the table.
62		*
63		* <table BORDER CELLPADDING=3 CELLSPACING=1>
64		*  <tr>
#	Line 86 | Line 82 | import java.util.concurrent.CountDownLat
82		*    <td> {@link ForkJoinTask#fork} (ForkJoinTasks <em>are</em> Futures)</td>
83		*  </tr>
84		* </table>
85	<	*
85	>	*
86		* <p><b>Sample Usage.</b> Normally a single {@code ForkJoinPool} is
87		* used for all parallel task execution in a program or subsystem.
88		* Otherwise, use would not usually outweigh the construction and
#	Line 111 | Line 107 | import java.util.concurrent.CountDownLat
107		* {@code IllegalArgumentException}.
108		*
109		* <p>This implementation rejects submitted tasks (that is, by throwing
110	<	* {@link RejectedExecutionException}) only when the pool is shut down.
110	>	* {@link RejectedExecutionException}) only when the pool is shut down
111	>	* or internal resources have been exhuasted.
112		*
113		* @since 1.7
114		* @author Doug Lea
#	Line 138 | Line 135 | public class ForkJoinPool extends Abstra
135		* of tasks profit from cache affinities, but others are harmed by
136		* cache pollution effects.)
137		*
138	+	* Beyond work-stealing support and essential bookkeeping, the
139	+	* main responsibility of this framework is to take actions when
140	+	* one worker is waiting to join a task stolen (or always held by)
141	+	* another.  Becauae we are multiplexing many tasks on to a pool
142	+	* of workers, we can't just let them block (as in Thread.join).
143	+	* We also cannot just reassign the joiner's run-time stack with
144	+	* another and replace it later, which would be a form of
145	+	* "continuation", that even if possible is not necessarily a good
146	+	* idea. Given that the creation costs of most threads on most
147	+	* systems mainly surrounds setting up runtime stacks, thread
148	+	* creation and switching is usually not much more expensive than
149	+	* stack creation and switching, and is more flexible). Instead we
150	+	* combine two tactics:
151	+	*
152	+	*   Helping: Arranging for the joiner to execute some task that it
153	+	*      would be running if the steal had not occurred.  Method
154	+	*      ForkJoinWorkerThread.helpJoinTask tracks joining->stealing
155	+	*      links to try to find such a task.
156	+	*
157	+	*   Compensating: Unless there are already enough live threads,
158	+	*      method helpMaintainParallelism() may create or or
159	+	*      re-activate a spare thread to compensate for blocked
160	+	*      joiners until they unblock.
161	+	*
162	+	* Because the determining existence of conservatively safe
163	+	* helping targets, the availability of already-created spares,
164	+	* and the apparent need to create new spares are all racy and
165	+	* require heuristic guidance, we rely on multiple retries of
166	+	* each. Further, because it is impossible to keep exactly the
167	+	* target (parallelism) number of threads running at any given
168	+	* time, we allow compensation during joins to fail, and enlist
169	+	* all other threads to help out whenever they are not otherwise
170	+	* occupied (i.e., mainly in method preStep).
171	+	*
172	+	* The ManagedBlocker extension API can't use helping so relies
173	+	* only on compensation in method awaitBlocker.
174	+	*
175		* The main throughput advantages of work-stealing stem from
176		* decentralized control -- workers mostly steal tasks from each
177		* other. We do not want to negate this by creating bottlenecks
178	<	* implementing the management responsibilities of this class. So
179	<	* we use a collection of techniques that avoid, reduce, or cope
180	<	* well with contention. These entail several instances of
181	<	* bit-packing into CASable fields to maintain only the minimally
182	<	* required atomicity. To enable such packing, we restrict maximum
183	<	* parallelism to (1<<15)-1 (enabling twice this to fit into a 16
184	<	* bit field), which is far in excess of normal operating range.
185	<	* Even though updates to some of these bookkeeping fields do
186	<	* sometimes contend with each other, they don't normally
187	<	* cache-contend with updates to others enough to warrant memory
188	<	* padding or isolation. So they are all held as fields of
189	<	* ForkJoinPool objects.  The main capabilities are as follows:
178	>	* implementing other management responsibilities. So we use a
179	>	* collection of techniques that avoid, reduce, or cope well with
180	>	* contention. These entail several instances of bit-packing into
181	>	* CASable fields to maintain only the minimally required
182	>	* atomicity. To enable such packing, we restrict maximum
183	>	* parallelism to (1<<15)-1 (enabling twice this (to accommodate
184	>	* unbalanced increments and decrements) to fit into a 16 bit
185	>	* field, which is far in excess of normal operating range.  Even
186	>	* though updates to some of these bookkeeping fields do sometimes
187	>	* contend with each other, they don't normally cache-contend with
188	>	* updates to others enough to warrant memory padding or
189	>	* isolation. So they are all held as fields of ForkJoinPool
190	>	* objects.  The main capabilities are as follows:
191		*
192		* 1. Creating and removing workers. Workers are recorded in the
193		* "workers" array. This is an array as opposed to some other data
#	Line 168 | Line 203 | public class ForkJoinPool extends Abstra
203		* blocked workers. However, all other support code is set up to
204		* work with other policies.
205		*
206	+	* To ensure that we do not hold on to worker references that
207	+	* would prevent GC, ALL accesses to workers are via indices into
208	+	* the workers array (which is one source of some of the unusual
209	+	* code constructions here). In essence, the workers array serves
210	+	* as a WeakReference mechanism. Thus for example the event queue
211	+	* stores worker indices, not worker references. Access to the
212	+	* workers in associated methods (for example releaseEventWaiters)
213	+	* must both index-check and null-check the IDs. All such accesses
214	+	* ignore bad IDs by returning out early from what they are doing,
215	+	* since this can only be associated with shutdown, in which case
216	+	* it is OK to give up. On termination, we just clobber these
217	+	* data structures without trying to use them.
218	+	*
219		* 2. Bookkeeping for dynamically adding and removing workers. We
220		* aim to approximately maintain the given level of parallelism.
221		* When some workers are known to be blocked (on joins or via
#	Line 177 | Line 225 | public class ForkJoinPool extends Abstra
225		* that are neither blocked nor artifically suspended) as well as
226		* the total number.  These two values are packed into one field,
227		* "workerCounts" because we need accurate snapshots when deciding
228	<	* to create, resume or suspend.  To support these decisions,
229	<	* updates to spare counts must be prospective (not
230	<	* retrospective).  For example, the running count is decremented
231	<	* before blocking by a thread about to block as a spare, but
184	<	* incremented by the thread about to unblock it. Updates upon
185	<	* resumption ofr threads blocking in awaitJoin or awaitBlocker
186	<	* cannot usually be prospective, so the running count is in
187	<	* general an upper bound of the number of productively running
188	<	* threads Updates to the workerCounts field sometimes transiently
189	<	* encounter a fair amount of contention when join dependencies
190	<	* are such that many threads block or unblock at about the same
191	<	* time. We alleviate this by sometimes performing an alternative
192	<	* action on contention like releasing waiters or locating spares.
228	>	* to create, resume or suspend.  Note however that the
229	>	* correspondance of these counts to reality is not guaranteed. In
230	>	* particular updates for unblocked threads may lag until they
231	>	* actually wake up.
232		*
233		* 3. Maintaining global run state. The run state of the pool
234		* consists of a runLevel (SHUTDOWN, TERMINATING, etc) similar to
#	Line 218 | Line 257 | public class ForkJoinPool extends Abstra
257		* workers that previously could not find a task to now find one:
258		* Submission of a new task to the pool, or another worker pushing
259		* a task onto a previously empty queue.  (We also use this
260	<	* mechanism for termination and reconfiguration actions that
261	<	* require wakeups of idle workers).  Each worker maintains its
262	<	* last known event count, and blocks when a scan for work did not
263	<	* find a task AND its lastEventCount matches the current
264	<	* eventCount. Waiting idle workers are recorded in a variant of
265	<	* Treiber stack headed by field eventWaiters which, when nonzero,
266	<	* encodes the thread index and count awaited for by the worker
267	<	* thread most recently calling eventSync. This thread in turn has
268	<	* a record (field nextEventWaiter) for the next waiting worker.
269	<	* In addition to allowing simpler decisions about need for
270	<	* wakeup, the event count bits in eventWaiters serve the role of
271	<	* tags to avoid ABA errors in Treiber stacks.  To reduce delays
272	<	* in task diffusion, workers not otherwise occupied may invoke
273	<	* method releaseWaiters, that removes and signals (unparks)
274	<	* workers not waiting on current count. To minimize task
275	<	* production stalls associate with signalling, any worker pushing
276	<	* a task on an empty queue invokes the weaker method signalWork,
277	<	* that only releases idle workers until it detects interference
278	<	* by other threads trying to release, and lets them take
279	<	* over. The net effect is a tree-like diffusion of signals, where
280	<	* released threads (and possibly others) help with unparks.  To
281	<	* further reduce contention effects a bit, failed CASes to
260	>	* mechanism for termination actions that require wakeups of idle
261	>	* workers).  Each worker maintains its last known event count,
262	>	* and blocks when a scan for work did not find a task AND its
263	>	* lastEventCount matches the current eventCount. Waiting idle
264	>	* workers are recorded in a variant of Treiber stack headed by
265	>	* field eventWaiters which, when nonzero, encodes the thread
266	>	* index and count awaited for by the worker thread most recently
267	>	* calling eventSync. This thread in turn has a record (field
268	>	* nextEventWaiter) for the next waiting worker.  In addition to
269	>	* allowing simpler decisions about need for wakeup, the event
270	>	* count bits in eventWaiters serve the role of tags to avoid ABA
271	>	* errors in Treiber stacks.  To reduce delays in task diffusion,
272	>	* workers not otherwise occupied may invoke method
273	>	* releaseEventWaiters, that removes and signals (unparks) workers
274	>	* not waiting on current count. To reduce stalls, To minimize
275	>	* task production stalls associate with signalling, any worker
276	>	* pushing a task on an empty queue invokes the weaker method
277	>	* signalWork, that only releases idle workers until it detects
278	>	* interference by other threads trying to release, and lets them
279	>	* take over.  The net effect is a tree-like diffusion of signals,
280	>	* where released threads (and possibly others) help with unparks.
281	>	* To further reduce contention effects a bit, failed CASes to
282		* increment field eventCount are tolerated without retries.
283		* Conceptually they are merged into the same event, which is OK
284		* when their only purpose is to enable workers to scan for work.
#	Line 247 | Line 286 | public class ForkJoinPool extends Abstra
286		* 5. Managing suspension of extra workers. When a worker is about
287		* to block waiting for a join (or via ManagedBlockers), we may
288		* create a new thread to maintain parallelism level, or at least
289	<	* avoid starvation (see below). Usually, extra threads are needed
290	<	* for only very short periods, yet join dependencies are such
291	<	* that we sometimes need them in bursts. Rather than create new
292	<	* threads each time this happens, we suspend no-longer-needed
293	<	* extra ones as "spares". For most purposes, we don't distinguish
294	<	* "extra" spare threads from normal "core" threads: On each call
295	<	* to preStep (the only point at which we can do this) a worker
289	>	* avoid starvation. Usually, extra threads are needed for only
290	>	* very short periods, yet join dependencies are such that we
291	>	* sometimes need them in bursts. Rather than create new threads
292	>	* each time this happens, we suspend no-longer-needed extra ones
293	>	* as "spares". For most purposes, we don't distinguish "extra"
294	>	* spare threads from normal "core" threads: On each call to
295	>	* preStep (the only point at which we can do this) a worker
296		* checks to see if there are now too many running workers, and if
297	<	* so, suspends itself.  Methods awaitJoin and awaitBlocker look
298	<	* for suspended threads to resume before considering creating a
299	<	* new replacement. We don't need a special data structure to
300	<	* maintain spares; simply scanning the workers array looking for
301	<	* worker.isSuspended() is fine because the calling thread is
263	<	* otherwise not doing anything useful anyway; we are at least as
264	<	* happy if after locating a spare, the caller doesn't actually
265	<	* block because the join is ready before we try to adjust and
266	<	* compensate.  Note that this is intrinsically racy.  One thread
297	>	* so, suspends itself.  Method helpMaintainParallelism looks for
298	>	* suspended threads to resume before considering creating a new
299	>	* replacement. The spares themselves are encoded on another
300	>	* variant of a Treiber Stack, headed at field "spareWaiters".
301	>	* Note that the use of spares is intrinsically racy.  One thread
302		* may become a spare at about the same time as another is
303		* needlessly being created. We counteract this and related slop
304		* in part by requiring resumed spares to immediately recheck (in
305	<	* preStep) to see whether they they should re-suspend. The only
306	<	* effective difference between "extra" and "core" threads is that
307	<	* we allow the "extra" ones to time out and die if they are not
308	<	* resumed within a keep-alive interval of a few seconds. This is
309	<	* implemented mainly within ForkJoinWorkerThread, but requires
310	<	* some coordination (isTrimmed() -- meaning killed while
311	<	* suspended) to correctly maintain pool counts.
305	>	* preStep) to see whether they they should re-suspend.  To avoid
306	>	* long-term build-up of spares, the oldest spare (see
307	>	* ForkJoinWorkerThread.suspendAsSpare) occasionally wakes up if
308	>	* not signalled and calls tryTrimSpare, which uses two different
309	>	* thresholds: Always killing if the number of spares is greater
310	>	* that 25% of total, and killing others only at a slower rate
311	>	* (UNUSED_SPARE_TRIM_RATE_NANOS).
312		*
313		* 6. Deciding when to create new workers. The main dynamic
314	<	* control in this class is deciding when to create extra threads,
315	<	* in methods awaitJoin and awaitBlocker. We always need to create
316	<	* one when the number of running threads becomes zero. But
317	<	* because blocked joins are typically dependent, we don't
318	<	* necessarily need or want one-to-one replacement. Instead, we
319	<	* use a combination of heuristics that adds threads only when the
320	<	* pool appears to be approaching starvation.  These effectively
321	<	* reduce churn at the price of systematically undershooting
322	<	* target parallelism when many threads are blocked.  However,
323	<	* biasing toward undeshooting partially compensates for the above
324	<	* mechanics to suspend extra threads, that normally lead to
325	<	* overshoot because we can only suspend workers in-between
326	<	* top-level actions. It also better copes with the fact that some
327	<	* of the methods in this class tend to never become compiled (but
328	<	* are interpreted), so some components of the entire set of
329	<	* controls might execute many times faster than others. And
330	<	* similarly for cases where the apparent lack of work is just due
331	<	* to GC stalls and other transient system activity.
314	>	* control in this class is deciding when to create extra threads
315	>	* in method helpMaintainParallelism. We would like to keep
316	>	* exactly #parallelism threads running, which is an impossble
317	>	* task. We always need to create one when the number of running
318	>	* threads would become zero and all workers are busy. Beyond
319	>	* this, we must rely on heuristics that work well in the the
320	>	* presence of transients phenomena such as GC stalls, dynamic
321	>	* compilation, and wake-up lags. These transients are extremely
322	>	* common -- we are normally trying to fully saturate the CPUs on
323	>	* a machine, so almost any activity other than running tasks
324	>	* impedes accuracy. Our main defense is to allow some slack in
325	>	* creation thresholds, using rules that reflect the fact that the
326	>	* more threads we have running, the more likely that we are
327	>	* underestimating the number running threads. The rules also
328	>	* better cope with the fact that some of the methods in this
329	>	* class tend to never become compiled (but are interpreted), so
330	>	* some components of the entire set of controls might execute 100
331	>	* times faster than others. And similarly for cases where the
332	>	* apparent lack of work is just due to GC stalls and other
333	>	* transient system activity.
334		*
335		* Beware that there is a lot of representation-level coupling
336		* among classes ForkJoinPool, ForkJoinWorkerThread, and
#	Line 306 | Line 343 | public class ForkJoinPool extends Abstra
343		*
344		* Style notes: There are lots of inline assignments (of form
345		* "while ((local = field) != 0)") which are usually the simplest
346	<	* way to ensure read orderings. Also several occurrences of the
347	<	* unusual "do {} while(!cas...)" which is the simplest way to
348	<	* force an update of a CAS'ed variable. There are also a few
349	<	* other coding oddities that help some methods perform reasonably
350	<	* even when interpreted (not compiled).
346	>	* way to ensure the required read orderings (which are sometimes
347	>	* critical). Also several occurrences of the unusual "do {}
348	>	* while(!cas...)" which is the simplest way to force an update of
349	>	* a CAS'ed variable. There are also other coding oddities that
350	>	* help some methods perform reasonably even when interpreted (not
351	>	* compiled), at the expense of some messy constructions that
352	>	* reduce byte code counts.
353		*
354		* The order of declarations in this file is: (1) statics (2)
355		* fields (along with constants used when unpacking some of them)
#	Line 378 | Line 417 | public class ForkJoinPool extends Abstra
417		new AtomicInteger();
418
419		/**
420	<	* Absolute bound for parallelism level. Twice this number must
421	<	* fit into a 16bit field to enable word-packing for some counts.
420	>	* Absolute bound for parallelism level. Twice this number plus
421	>	* one (i.e., 0xfff) must fit into a 16bit field to enable
422	>	* word-packing for some counts and indices.
423		*/
424	<	private static final int MAX_THREADS = 0x7fff;
424	>	private static final int MAX_WORKERS   = 0x7fff;
425
426		/**
427		* Array holding all worker threads in the pool.  Array size must
#	Line 421 | Line 461 | public class ForkJoinPool extends Abstra
461		private volatile long stealCount;
462
463		/**
464	+	* The last nanoTime that a spare thread was trimmed
465	+	*/
466	+	private volatile long trimTime;
467	+
468	+	/**
469	+	* The rate at which to trim unused spares
470	+	*/
471	+	static final long UNUSED_SPARE_TRIM_RATE_NANOS =
472	+	1000L * 1000L * 1000L; // 1 sec
473	+
474	+	/**
475		* Encoded record of top of treiber stack of threads waiting for
476		* events. The top 32 bits contain the count being waited for. The
477	<	* bottom word contains one plus the pool index of waiting worker
478	<	* thread.
477	>	* bottom 16 bits contains one plus the pool index of waiting
478	>	* worker thread. (Bits 16-31 are unused.)
479		*/
480		private volatile long eventWaiters;
481
482		private static final int  EVENT_COUNT_SHIFT = 32;
483	<	private static final long WAITER_INDEX_MASK = (1L << EVENT_COUNT_SHIFT)-1L;
483	>	private static final long WAITER_ID_MASK    = (1L << 16) - 1L;
484
485		/**
486		* A counter for events that may wake up worker threads:
487		*   - Submission of a new task to the pool
488		*   - A worker pushing a task on an empty queue
489	<	*   - termination and reconfiguration
489	>	*   - termination
490		*/
491		private volatile int eventCount;
492
493		/**
494	+	* Encoded record of top of treiber stack of spare threads waiting
495	+	* for resumption. The top 16 bits contain an arbitrary count to
496	+	* avoid ABA effects. The bottom 16bits contains one plus the pool
497	+	* index of waiting worker thread.
498	+	*/
499	+	private volatile int spareWaiters;
500	+
501	+	private static final int SPARE_COUNT_SHIFT = 16;
502	+	private static final int SPARE_ID_MASK     = (1 << 16) - 1;
503	+
504	+	/**
505		* Lifecycle control. The low word contains the number of workers
506		* that are (probably) executing tasks. This value is atomically
507		* incremented before a worker gets a task to run, and decremented
#	Line 468 | Line 530 | public class ForkJoinPool extends Abstra
530		* making decisions about creating and suspending spare
531		* threads. Updated only by CAS. Note that adding a new worker
532		* requires incrementing both counts, since workers start off in
533	<	* running state.  This field is also used for memory-fencing
472	<	* configuration parameters.
533	>	* running state.
534		*/
535		private volatile int workerCounts;
536
#	Line 501 | Line 562 | public class ForkJoinPool extends Abstra
562		*/
563		private final int poolNumber;
564
565	<	// utilities for updating fields
565	>
566	>	// Utilities for CASing fields. Note that several of these
567	>	// are manually inlined by callers
568
569		/**
570	<	* Increments running count.  Also used by ForkJoinTask.
570	>	* Increments running count part of workerCounts
571		*/
572		final void incrementRunningCount() {
573		int c;
574		do {} while (!UNSAFE.compareAndSwapInt(this, workerCountsOffset,
575	<	c = workerCounts,
575	>	c = workerCounts,
576		c + ONE_RUNNING));
577		}
578	<
578	>
579		/**
580		* Tries to decrement running count unless already zero
581		*/
#	Line 525 | Line 588 | public class ForkJoinPool extends Abstra
588		}
589
590		/**
591	+	* Forces decrement of encoded workerCounts, awaiting nonzero if
592	+	* (rarely) necessary when other count updates lag.
593	+	*
594	+	* @param dr -- either zero or ONE_RUNNING
595	+	* @param dt == either zero or ONE_TOTAL
596	+	*/
597	+	private void decrementWorkerCounts(int dr, int dt) {
598	+	for (;;) {
599	+	int wc = workerCounts;
600	+	if (wc == 0 && (runState & TERMINATED) != 0)
601	+	return; // lagging termination on a backout
602	+	if ((wc & RUNNING_COUNT_MASK)  - dr < 0 ||
603	+	(wc >>> TOTAL_COUNT_SHIFT) - dt < 0)
604	+	Thread.yield();
605	+	if (UNSAFE.compareAndSwapInt(this, workerCountsOffset,
606	+	wc, wc - (dr + dt)))
607	+	return;
608	+	}
609	+	}
610	+
611	+	/**
612	+	* Increments event count
613	+	*/
614	+	private void advanceEventCount() {
615	+	int c;
616	+	do {} while(!UNSAFE.compareAndSwapInt(this, eventCountOffset,
617	+	c = eventCount, c+1));
618	+	}
619	+
620	+	/**
621		* Tries incrementing active count; fails on contention.
622		* Called by workers before executing tasks.
623		*
#	Line 572 | Line 665 | public class ForkJoinPool extends Abstra
665		lock.lock();
666		try {
667		ForkJoinWorkerThread[] ws = workers;
668	<	int nws = ws.length;
669	<	if (k < 0 || k >= nws || ws[k] != null) {
670	<	for (k = 0; k < nws && ws[k] != null; ++k)
668	>	int n = ws.length;
669	>	if (k < 0 || k >= n || ws[k] != null) {
670	>	for (k = 0; k < n && ws[k] != null; ++k)
671		;
672	<	if (k == nws)
673	<	ws = Arrays.copyOf(ws, nws << 1);
672	>	if (k == n)
673	>	ws = Arrays.copyOf(ws, n << 1);
674		}
675		ws[k] = w;
676		workers = ws; // volatile array write ensures slot visibility
#	Line 610 | Line 703 | public class ForkJoinPool extends Abstra
703		* Tries to create and add new worker. Assumes that worker counts
704		* are already updated to accommodate the worker, so adjusts on
705		* failure.
613	–	*
614	–	* @return new worker or null if creation failed
706		*/
707	<	private ForkJoinWorkerThread addWorker() {
707	>	private void addWorker() {
708		ForkJoinWorkerThread w = null;
709		try {
710		w = factory.newThread(this);
711		} finally { // Adjust on either null or exceptional factory return
712		if (w == null) {
713	<	onWorkerCreationFailure();
714	<	return null;
713	>	decrementWorkerCounts(ONE_RUNNING, ONE_TOTAL);
714	>	tryTerminate(false); // in case of failure during shutdown
715		}
716		}
717	<	w.start(recordWorker(w), ueh);
718	<	return w;
628	<	}
629	<
630	<	/**
631	<	* Adjusts counts upon failure to create worker
632	<	*/
633	<	private void onWorkerCreationFailure() {
634	<	for (;;) {
635	<	int wc = workerCounts;
636	<	if ((wc >>> TOTAL_COUNT_SHIFT) > 0 &&
637	<	UNSAFE.compareAndSwapInt(this, workerCountsOffset,
638	<	wc, wc - (ONE_RUNNING|ONE_TOTAL)))
639	<	break;
640	<	}
641	<	tryTerminate(false); // in case of failure during shutdown
642	<	}
643	<
644	<	/**
645	<	* Create enough total workers to establish target parallelism,
646	<	* giving up if terminating or addWorker fails
647	<	*/
648	<	private void ensureEnoughTotalWorkers() {
649	<	int wc;
650	<	while (((wc = workerCounts) >>> TOTAL_COUNT_SHIFT) < parallelism &&
651	<	runState < TERMINATING) {
652	<	if ((UNSAFE.compareAndSwapInt(this, workerCountsOffset,
653	<	wc, wc + (ONE_RUNNING|ONE_TOTAL)) &&
654	<	addWorker() == null))
655	<	break;
656	<	}
717	>	if (w != null)
718	>	w.start(recordWorker(w), ueh);
719		}
720
721		/**
722		* Final callback from terminating worker.  Removes record of
723		* worker from array, and adjusts counts. If pool is shutting
724	<	* down, tries to complete terminatation, else possibly replaces
663	<	* the worker.
724	>	* down, tries to complete terminatation.
725		*
726		* @param w the worker
727		*/
728		final void workerTerminated(ForkJoinWorkerThread w) {
668	–	if (w.active) { // force inactive
669	–	w.active = false;
670	–	do {} while (!tryDecrementActiveCount());
671	–	}
729		forgetWorker(w);
730	<
731	<	// Decrement total count, and if was running, running count
732	<	// Spin (waiting for other updates) if either would be negative
733	<	int nr = w.isTrimmed() ? 0 : ONE_RUNNING;
677	<	int unit = ONE_TOTAL + nr;
678	<	for (;;) {
679	<	int wc = workerCounts;
680	<	int rc = wc & RUNNING_COUNT_MASK;
681	<	if (rc - nr < 0 || (wc >>> TOTAL_COUNT_SHIFT) == 0)
682	<	Thread.yield(); // back off if waiting for other updates
683	<	else if (UNSAFE.compareAndSwapInt(this, workerCountsOffset,
684	<	wc, wc - unit))
685	<	break;
686	<	}
687	<
688	<	accumulateStealCount(w); // collect final count
689	<	if (!tryTerminate(false))
690	<	ensureEnoughTotalWorkers();
730	>	decrementWorkerCounts(w.isTrimmed()? 0 : ONE_RUNNING, ONE_TOTAL);
731	>	while (w.stealCount != 0) // collect final count
732	>	tryAccumulateStealCount(w);
733	>	tryTerminate(false);
734		}
735
736		// Waiting for and signalling events
737
738		/**
739		* Releases workers blocked on a count not equal to current count.
740	+	* Normally called after precheck that eventWaiters isn't zero to
741	+	* avoid wasted array checks.
742	+	*
743	+	* @param signalling true if caller is a signalling worker so can
744	+	* exit upon (conservatively) detected contention by other threads
745	+	* who will continue to release
746		*/
747	<	private void releaseWaiters() {
748	<	long top;
749	<	int id;
750	<	while ((id = (int)((top = eventWaiters) & WAITER_INDEX_MASK)) > 0 &&
751	<	(int)(top >>> EVENT_COUNT_SHIFT) != eventCount) {
752	<	ForkJoinWorkerThread[] ws = workers;
753	<	ForkJoinWorkerThread w;
754	<	if (ws.length >= id && (w = ws[id - 1]) != null &&
755	<	UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
756	<	top, w.nextWaiter))
747	>	private void releaseEventWaiters(boolean signalling) {
748	>	ForkJoinWorkerThread[] ws = workers;
749	>	int n = ws.length;
750	>	long h; // head of stack
751	>	ForkJoinWorkerThread w; int id, ec;
752	>	while ((id = ((int)((h = eventWaiters) & WAITER_ID_MASK)) - 1) >= 0 &&
753	>	(int)(h >>> EVENT_COUNT_SHIFT) != (ec = eventCount) &&
754	>	id < n && (w = ws[id]) != null) {
755	>	if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
756	>	h, h = w.nextWaiter))
757		LockSupport.unpark(w);
758	+	if (signalling && (eventCount != ec || eventWaiters != h))
759	+	break;
760		}
761		}
762
763		/**
764	<	* Ensures eventCount on exit is different (mod 2^32) than on
765	<	* entry and wakes up all waiters
764	>	* Tries to advance eventCount and releases waiters. Called only
765	>	* from workers.
766		*/
767	<	private void signalEvent() {
768	<	int c;
769	<	do {} while (!UNSAFE.compareAndSwapInt(this, eventCountOffset,
770	<	c = eventCount, c+1));
771	<	releaseWaiters();
767	>	final void signalWork() {
768	>	int c; // try to increment event count -- CAS failure OK
769	>	UNSAFE.compareAndSwapInt(this, eventCountOffset, c = eventCount, c+1);
770	>	if (eventWaiters != 0L)
771	>	releaseEventWaiters(true);
772		}
773
774		/**
775	<	* Advances eventCount and releases waiters until interference by
776	<	* other releasing threads is detected.
775	>	* Blocks worker until terminating or event count
776	>	* advances from last value held by worker
777	>	*
778	>	* @param w the calling worker thread
779		*/
780	<	final void signalWork() {
781	<	// EventCount CAS failures are OK -- any change in count suffices.
782	<	int ec;
783	<	UNSAFE.compareAndSwapInt(this, eventCountOffset, ec=eventCount, ec+1);
784	<	outer:for (;;) {
785	<	long top = eventWaiters;
786	<	ec = eventCount;
787	<	for (;;) {
788	<	ForkJoinWorkerThread[] ws; ForkJoinWorkerThread w;
789	<	int id = (int)(top & WAITER_INDEX_MASK);
790	<	if (id <= 0 || (int)(top >>> EVENT_COUNT_SHIFT) == ec)
791	<	return;
792	<	if ((ws = workers).length < id || (w = ws[id - 1]) == null ||
793	<	!UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
794	<	top, top = w.nextWaiter))
795	<	continue outer;      // possibly stale; reread
796	<	LockSupport.unpark(w);
797	<	if (top != eventWaiters) // let someone else take over
798	<	return;
780	>	private void eventSync(ForkJoinWorkerThread w) {
781	>	int wec = w.lastEventCount;
782	>	long nh = (((long)wec) << EVENT_COUNT_SHIFT) | ((long)(w.poolIndex+1));
783	>	long h;
784	>	while ((runState < SHUTDOWN || !tryTerminate(false)) &&
785	>	((h = eventWaiters) == 0L ||
786	>	(int)(h >>> EVENT_COUNT_SHIFT) == wec) &&
787	>	eventCount == wec) {
788	>	if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
789	>	w.nextWaiter = h, nh)) {
790	>	while (runState < TERMINATING && eventCount == wec) {
791	>	if (!tryAccumulateStealCount(w))  // transfer while idle
792	>	continue;
793	>	Thread.interrupted();             // clear/ignore interrupt
794	>	if (eventCount != wec)
795	>	break;
796	>	LockSupport.park(w);
797	>	}
798	>	break;
799		}
800		}
801	+	w.lastEventCount = eventCount;
802		}
803
804	+	// Maintaining spares
805	+
806		/**
807	<	* If worker is inactive, blocks until terminating or event count
752	<	* advances from last value held by worker; in any case helps
753	<	* release others.
754	<	*
755	<	* @param w the calling worker thread
807	>	* Pushes worker onto the spare stack
808		*/
809	<	private void eventSync(ForkJoinWorkerThread w) {
810	<	if (!w.active) {
811	<	int prev = w.lastEventCount;
812	<	long nextTop = (((long)prev << EVENT_COUNT_SHIFT) |
813	<	((long)(w.poolIndex + 1)));
814	<	long top;
815	<	while ((runState < SHUTDOWN || !tryTerminate(false)) &&
816	<	(((int)(top = eventWaiters) & WAITER_INDEX_MASK) == 0 ||
817	<	(int)(top >>> EVENT_COUNT_SHIFT) == prev) &&
818	<	eventCount == prev) {
819	<	if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
820	<	w.nextWaiter = top, nextTop)) {
821	<	accumulateStealCount(w); // transfer steals while idle
822	<	Thread.interrupted();    // clear/ignore interrupt
823	<	while (eventCount == prev)
824	<	w.doPark();
825	<	break;
826	<	}
809	>	final void pushSpare(ForkJoinWorkerThread w) {
810	>	int ns = (++w.spareCount << SPARE_COUNT_SHIFT) | (w.poolIndex+1);
811	>	do {} while (!UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
812	>	w.nextSpare = spareWaiters,ns));
813	>	}
814	>
815	>	/**
816	>	* Tries (once) to resume a spare if running count is less than
817	>	* target parallelism. Fails on contention or stale workers.
818	>	*/
819	>	private void tryResumeSpare() {
820	>	int sw, id;
821	>	ForkJoinWorkerThread w;
822	>	ForkJoinWorkerThread[] ws;
823	>	if ((id = ((sw = spareWaiters) & SPARE_ID_MASK) - 1) >= 0 &&
824	>	id < (ws = workers).length && (w = ws[id]) != null &&
825	>	(workerCounts & RUNNING_COUNT_MASK) < parallelism &&
826	>	eventWaiters == 0L &&
827	>	spareWaiters == sw &&
828	>	UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
829	>	sw, w.nextSpare) &&
830	>	w.tryUnsuspend()) {
831	>	int c; // try increment; if contended, finish after unpark
832	>	boolean inc = UNSAFE.compareAndSwapInt(this, workerCountsOffset,
833	>	c = workerCounts,
834	>	c + ONE_RUNNING);
835	>	LockSupport.unpark(w);
836	>	if (!inc) {
837	>	do {} while(!UNSAFE.compareAndSwapInt(this, workerCountsOffset,
838	>	c = workerCounts,
839	>	c + ONE_RUNNING));
840		}
776	–	w.lastEventCount = eventCount;
841		}
842	<	releaseWaiters();
842	>	}
843	>
844	>	/**
845	>	* Callback from oldest spare occasionally waking up.  Tries
846	>	* (once) to shutdown a spare if more than 25% spare overage, or
847	>	* if UNUSED_SPARE_TRIM_RATE_NANOS have elapsed and there are at
848	>	* least #parallelism running threads. Note that we don't need CAS
849	>	* or locks here because the method is called only from the oldest
850	>	* suspended spare occasionally waking (and even misfires are OK).
851	>	*
852	>	* @param now the wake up nanoTime of caller
853	>	*/
854	>	final void tryTrimSpare(long now) {
855	>	long lastTrim = trimTime;
856	>	trimTime = now;
857	>	helpMaintainParallelism(); // first, help wake up any needed spares
858	>	int sw, id;
859	>	ForkJoinWorkerThread w;
860	>	ForkJoinWorkerThread[] ws;
861	>	int pc = parallelism;
862	>	int wc = workerCounts;
863	>	if ((wc & RUNNING_COUNT_MASK) >= pc &&
864	>	(((wc >>> TOTAL_COUNT_SHIFT) - pc) > (pc >>> 2) + 1 ||// approx 25%
865	>	now - lastTrim >= UNUSED_SPARE_TRIM_RATE_NANOS) &&
866	>	(id = ((sw = spareWaiters) & SPARE_ID_MASK) - 1) >= 0 &&
867	>	id < (ws = workers).length && (w = ws[id]) != null &&
868	>	UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
869	>	sw, w.nextSpare))
870	>	w.shutdown(false);
871	>	}
872	>
873	>	/**
874	>	* Does at most one of:
875	>	*
876	>	* 1. Help wake up existing workers waiting for work via
877	>	*    releaseEventWaiters. (If any exist, then it probably doesn't
878	>	*    matter right now if under target parallelism level.)
879	>	*
880	>	* 2. If below parallelism level and a spare exists, try (once)
881	>	*    to resume it via tryResumeSpare.
882	>	*
883	>	* 3. If neither of the above, tries (once) to add a new
884	>	*    worker if either there are not enough total, or if all
885	>	*    existing workers are busy, there are either no running
886	>	*    workers or the deficit is at least twice the surplus.
887	>	*/
888	>	private void helpMaintainParallelism() {
889	>	// uglified to work better when not compiled
890	>	int pc, wc, rc, tc, rs; long h;
891	>	if ((h = eventWaiters) != 0L) {
892	>	if ((int)(h >>> EVENT_COUNT_SHIFT) != eventCount)
893	>	releaseEventWaiters(false); // avoid useless call
894	>	}
895	>	else if ((pc = parallelism) >
896	>	(rc = ((wc = workerCounts) & RUNNING_COUNT_MASK))) {
897	>	if (spareWaiters != 0)
898	>	tryResumeSpare();
899	>	else if ((rs = runState) < TERMINATING &&
900	>	((tc = wc >>> TOTAL_COUNT_SHIFT) < pc ||
901	>	(tc == (rs & ACTIVE_COUNT_MASK) && // all busy
902	>	(rc == 0 ||                       // must add
903	>	rc < pc - ((tc - pc) << 1)) &&   // within slack
904	>	tc < MAX_WORKERS && runState == rs)) && // recheck busy
905	>	workerCounts == wc &&
906	>	UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
907	>	wc + (ONE_RUNNING|ONE_TOTAL)))
908	>	addWorker();
909	>	}
910		}
911
912		/**
913		* Callback from workers invoked upon each top-level action (i.e.,
914		* stealing a task or taking a submission and running
915	<	* it). Performs one or both of the following:
915	>	* it). Performs one or more of the following:
916	>	*
917	>	* 1. If the worker cannot find work (misses > 0), updates its
918	>	*    active status to inactive and updates activeCount unless
919	>	*    this is the first miss and there is contention, in which
920	>	*    case it may try again (either in this or a subsequent
921	>	*    call).
922	>	*
923	>	* 2. If there are at least 2 misses, awaits the next task event
924	>	*    via eventSync
925	>	*
926	>	* 3. If there are too many running threads, suspends this worker
927	>	*    (first forcing inactivation if necessary).  If it is not
928	>	*    needed, it may be killed while suspended via
929	>	*    tryTrimSpare. Otherwise, upon resume it rechecks to make
930	>	*    sure that it is still needed.
931		*
932	<	* * If the worker cannot find work, updates its active status to
933	<	* inactive and updates activeCount unless there is contention, in
788	<	* which case it may try again (either in this or a subsequent
789	<	* call).  Additionally, awaits the next task event and/or helps
790	<	* wake up other releasable waiters.
791	<	*
792	<	* * If there are too many running threads, suspends this worker
793	<	* (first forcing inactivation if necessary).  If it is not
794	<	* resumed before a keepAlive elapses, the worker may be "trimmed"
795	<	* -- killed while suspended within suspendAsSpare. Otherwise,
796	<	* upon resume it rechecks to make sure that it is still needed.
932	>	* 4. Helps release and/or reactivate other workers via
933	>	*    helpMaintainParallelism
934		*
935		* @param w the worker
936	<	* @param worked false if the worker scanned for work but didn't
937	<	* find any (in which case it may block waiting for work).
936	>	* @param misses the number of scans by caller failing to find work
937	>	* (saturating at 2 just to avoid wraparound)
938		*/
939	<	final void preStep(ForkJoinWorkerThread w, boolean worked) {
939	>	final void preStep(ForkJoinWorkerThread w, int misses) {
940		boolean active = w.active;
941	<	boolean inactivate = !worked & active;
941	>	int pc = parallelism;
942		for (;;) {
943	<	if (inactivate) {
944	<	int rs = runState;
943	>	int wc = workerCounts;
944	>	int rc = wc & RUNNING_COUNT_MASK;
945	>	if (active && (misses > 0 || rc > pc)) {
946	>	int rs;                      // try inactivate
947		if (UNSAFE.compareAndSwapInt(this, runStateOffset,
948	<	rs, rs - ONE_ACTIVE))
949	<	inactivate = active = w.active = false;
948	>	rs = runState, rs - ONE_ACTIVE))
949	>	active = w.active = false;
950	>	else if (misses > 1 || rc > pc ||
951	>	(rs & ACTIVE_COUNT_MASK) >= pc)
952	>	continue;                // force inactivate
953		}
954	<	int wc = workerCounts;
955	<	if ((wc & RUNNING_COUNT_MASK) <= parallelism) {
956	<	if (!worked)
957	<	eventSync(w);
958	<	return;
954	>	if (misses > 1) {
955	>	misses = 0;                  // don't re-sync
956	>	eventSync(w);                // continue loop to recheck rc
957	>	}
958	>	else if (rc > pc) {
959	>	if (workerCounts == wc &&   // try to suspend as spare
960	>	UNSAFE.compareAndSwapInt(this, workerCountsOffset,
961	>	wc, wc - ONE_RUNNING) &&
962	>	!w.suspendAsSpare())    // false if killed
963	>	break;
964	>	}
965	>	else {
966	>	if (rc < pc || eventWaiters != 0L)
967	>	helpMaintainParallelism();
968	>	break;
969		}
818	–	if (!(inactivate |= active) &&  // must inactivate to suspend
819	–	UNSAFE.compareAndSwapInt(this, workerCountsOffset,
820	–	wc, wc - ONE_RUNNING) &&
821	–	!w.suspendAsSpare())        // false if trimmed
822	–	return;
970		}
971		}
972
973		/**
974	<	* Tries to decrement running count, and if so, possibly creates
975	<	* or resumes compensating threads before blocking on task joinMe.
976	<	* This code is sprawled out with manual inlining to evade some
977	<	* JIT oddities.
974	>	* Helps and/or blocks awaiting join of the given task.
975	>	* Alternates between helpJoinTask() and helpMaintainParallelism()
976	>	* as many times as there is a deficit in running count (or longer
977	>	* if running count would become zero), then blocks if task still
978	>	* not done.
979		*
980		* @param joinMe the task to join
833	–	* @return task status on exit
981		*/
982	<	final int tryAwaitJoin(ForkJoinTask<?> joinMe) {
983	<	int cw = workerCounts; // read now to spoil CAS if counts change as ...
984	<	releaseWaiters();      // ... a byproduct of releaseWaiters
985	<	int stat = joinMe.status;
986	<	if (stat >= 0 && // inline variant of tryDecrementRunningCount
987	<	(cw & RUNNING_COUNT_MASK) > 0 &&
988	<	UNSAFE.compareAndSwapInt(this, workerCountsOffset,
989	<	cw, cw - ONE_RUNNING)) {
990	<	int pc = parallelism;
991	<	int scans = 0;  // to require confirming passes to add threads
992	<	outer: while ((workerCounts & RUNNING_COUNT_MASK) < pc) {
993	<	if ((stat = joinMe.status) < 0)
994	<	break;
995	<	ForkJoinWorkerThread spare = null;
996	<	ForkJoinWorkerThread[] ws = workers;
997	<	int nws = ws.length;
998	<	for (int i = 0; i < nws; ++i) {
999	<	ForkJoinWorkerThread w = ws[i];
1000	<	if (w != null && w.isSuspended()) {
1001	<	spare = w;
1002	<	break;
1003	<	}
1004	<	}
1005	<	if ((stat = joinMe.status) < 0) // recheck to narrow race
1006	<	break;
860	<	int wc = workerCounts;
861	<	int rc = wc & RUNNING_COUNT_MASK;
862	<	if (rc >= pc)
863	<	break;
864	<	if (spare != null) {
865	<	if (spare.tryUnsuspend()) {
866	<	int c; // inline incrementRunningCount
867	<	do {} while (!UNSAFE.compareAndSwapInt
868	<	(this, workerCountsOffset,
869	<	c = workerCounts, c + ONE_RUNNING));
870	<	LockSupport.unpark(spare);
871	<	break;
872	<	}
873	<	continue;
874	<	}
875	<	int tc = wc >>> TOTAL_COUNT_SHIFT;
876	<	int sc = tc - pc;
877	<	if (rc > 0) {
878	<	int p = pc;
879	<	int s = sc;
880	<	while (s-- >= 0) { // try keeping 3/4 live
881	<	if (rc > (p -= (p >>> 2) + 1))
882	<	break outer;
883	<	}
884	<	}
885	<	if (scans++ > sc && tc < MAX_THREADS &&
886	<	UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
887	<	wc + (ONE_RUNNING|ONE_TOTAL))) {
888	<	addWorker();
889	<	break;
890	<	}
982	>	final void awaitJoin(ForkJoinTask<?> joinMe, ForkJoinWorkerThread worker) {
983	>	int threshold = parallelism;         // descend blocking thresholds
984	>	while (joinMe.status >= 0) {
985	>	boolean block; int wc;
986	>	worker.helpJoinTask(joinMe);
987	>	if (joinMe.status < 0)
988	>	break;
989	>	if (((wc = workerCounts) & RUNNING_COUNT_MASK) <= threshold) {
990	>	if (threshold > 0)
991	>	--threshold;
992	>	else
993	>	advanceEventCount(); // force release
994	>	block = false;
995	>	}
996	>	else
997	>	block = UNSAFE.compareAndSwapInt(this, workerCountsOffset,
998	>	wc, wc - ONE_RUNNING);
999	>	helpMaintainParallelism();
1000	>	if (block) {
1001	>	int c;
1002	>	joinMe.internalAwaitDone();
1003	>	do {} while (!UNSAFE.compareAndSwapInt
1004	>	(this, workerCountsOffset,
1005	>	c = workerCounts, c + ONE_RUNNING));
1006	>	break;
1007		}
892	–	if (stat >= 0)
893	–	stat = joinMe.internalAwaitDone();
894	–	int c; // inline incrementRunningCount
895	–	do {} while (!UNSAFE.compareAndSwapInt
896	–	(this, workerCountsOffset,
897	–	c = workerCounts, c + ONE_RUNNING));
1008		}
899	–	return stat;
1009		}
1010
1011		/**
1012	<	* Same idea as (and mostly pasted from) tryAwaitJoin, but
904	<	* self-contained
1012	>	* Same idea as awaitJoin, but no helping
1013		*/
1014		final void awaitBlocker(ManagedBlocker blocker)
1015		throws InterruptedException {
1016	<	for (;;) {
1017	<	if (blocker.isReleasable())
1018	<	return;
1019	<	int cw = workerCounts;
1020	<	releaseWaiters();
1021	<	if ((cw & RUNNING_COUNT_MASK) > 0 &&
1022	<	UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1023	<	cw, cw - ONE_RUNNING))
1024	<	break;
917	<	}
918	<	boolean done = false;
919	<	int pc = parallelism;
920	<	int scans = 0;
921	<	outer: while ((workerCounts & RUNNING_COUNT_MASK) < pc) {
922	<	if (done = blocker.isReleasable())
923	<	break;
924	<	ForkJoinWorkerThread spare = null;
925	<	ForkJoinWorkerThread[] ws = workers;
926	<	int nws = ws.length;
927	<	for (int i = 0; i < nws; ++i) {
928	<	ForkJoinWorkerThread w = ws[i];
929	<	if (w != null && w.isSuspended()) {
930	<	spare = w;
931	<	break;
932	<	}
1016	>	int threshold = parallelism;
1017	>	while (!blocker.isReleasable()) {
1018	>	boolean block; int wc;
1019	>	if (((wc = workerCounts) & RUNNING_COUNT_MASK) <= threshold) {
1020	>	if (threshold > 0)
1021	>	--threshold;
1022	>	else
1023	>	advanceEventCount();
1024	>	block = false;
1025		}
1026	<	if (done = blocker.isReleasable())
1027	<	break;
1028	<	int wc = workerCounts;
1029	<	int rc = wc & RUNNING_COUNT_MASK;
1030	<	if (rc >= pc)
1031	<	break;
1032	<	if (spare != null) {
1033	<	if (spare.tryUnsuspend()) {
1026	>	else
1027	>	block = UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1028	>	wc, wc - ONE_RUNNING);
1029	>	helpMaintainParallelism();
1030	>	if (block) {
1031	>	try {
1032	>	do {} while (!blocker.isReleasable() && !blocker.block());
1033	>	} finally {
1034		int c;
1035		do {} while (!UNSAFE.compareAndSwapInt
1036		(this, workerCountsOffset,
1037		c = workerCounts, c + ONE_RUNNING));
946	–	LockSupport.unpark(spare);
947	–	break;
948	–	}
949	–	continue;
950	–	}
951	–	int tc = wc >>> TOTAL_COUNT_SHIFT;
952	–	int sc = tc - pc;
953	–	if (rc > 0) {
954	–	int p = pc;
955	–	int s = sc;
956	–	while (s-- >= 0) {
957	–	if (rc > (p -= (p >>> 2) + 1))
958	–	break outer;
1038		}
960	–	}
961	–	if (scans++ > sc && tc < MAX_THREADS &&
962	–	UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
963	–	wc + (ONE_RUNNING|ONE_TOTAL))) {
964	–	addWorker();
1039		break;
1040		}
1041		}
1042	<	try {
969	<	if (!done)
970	<	do {} while (!blocker.isReleasable() &&
971	<	!blocker.block());
972	<	} finally {
973	<	int c;
974	<	do {} while (!UNSAFE.compareAndSwapInt
975	<	(this, workerCountsOffset,
976	<	c = workerCounts, c + ONE_RUNNING));
977	<	}
978	<	}
1042	>	}
1043
1044		/**
1045		* Possibly initiates and/or completes termination.
#	Line 1005 | Line 1069 | public class ForkJoinPool extends Abstra
1069
1070		/**
1071		* Actions on transition to TERMINATING
1072	+	*
1073	+	* Runs up to four passes through workers: (0) shutting down each
1074	+	* quietly (without waking up if parked) to quickly spread
1075	+	* notifications without unnecessary bouncing around event queues
1076	+	* etc (1) wake up and help cancel tasks (2) interrupt (3) mop up
1077	+	* races with interrupted workers
1078		*/
1079		private void startTerminating() {
1080	<	for (int i = 0; i < 2; ++i) { // twice to mop up newly created workers
1081	<	cancelSubmissions();
1082	<	shutdownWorkers();
1083	<	cancelWorkerTasks();
1084	<	signalEvent();
1085	<	interruptWorkers();
1080	>	cancelSubmissions();
1081	>	for (int passes = 0; passes < 4 && workerCounts != 0; ++passes) {
1082	>	advanceEventCount();
1083	>	eventWaiters = 0L; // clobber lists
1084	>	spareWaiters = 0;
1085	>	ForkJoinWorkerThread[] ws = workers;
1086	>	int n = ws.length;
1087	>	for (int i = 0; i < n; ++i) {
1088	>	ForkJoinWorkerThread w = ws[i];
1089	>	if (w != null) {
1090	>	w.shutdown(true);
1091	>	if (passes > 0 && !w.isTerminated()) {
1092	>	w.cancelTasks();
1093	>	LockSupport.unpark(w);
1094	>	if (passes > 1) {
1095	>	try {
1096	>	w.interrupt();
1097	>	} catch (SecurityException ignore) {
1098	>	}
1099	>	}
1100	>	}
1101	>	}
1102	>	}
1103		}
1104		}
1105
#	Line 1029 | Line 1116 | public class ForkJoinPool extends Abstra
1116		}
1117		}
1118
1032	–	/**
1033	–	* Sets all worker run states to at least shutdown,
1034	–	* also resuming suspended workers
1035	–	*/
1036	–	private void shutdownWorkers() {
1037	–	ForkJoinWorkerThread[] ws = workers;
1038	–	int nws = ws.length;
1039	–	for (int i = 0; i < nws; ++i) {
1040	–	ForkJoinWorkerThread w = ws[i];
1041	–	if (w != null)
1042	–	w.shutdown();
1043	–	}
1044	–	}
1045	–
1046	–	/**
1047	–	* Clears out and cancels all locally queued tasks
1048	–	*/
1049	–	private void cancelWorkerTasks() {
1050	–	ForkJoinWorkerThread[] ws = workers;
1051	–	int nws = ws.length;
1052	–	for (int i = 0; i < nws; ++i) {
1053	–	ForkJoinWorkerThread w = ws[i];
1054	–	if (w != null)
1055	–	w.cancelTasks();
1056	–	}
1057	–	}
1058	–
1059	–	/**
1060	–	* Unsticks all workers blocked on joins etc
1061	–	*/
1062	–	private void interruptWorkers() {
1063	–	ForkJoinWorkerThread[] ws = workers;
1064	–	int nws = ws.length;
1065	–	for (int i = 0; i < nws; ++i) {
1066	–	ForkJoinWorkerThread w = ws[i];
1067	–	if (w != null && !w.isTerminated()) {
1068	–	try {
1069	–	w.interrupt();
1070	–	} catch (SecurityException ignore) {
1071	–	}
1072	–	}
1073	–	}
1074	–	}
1075	–
1119		// misc support for ForkJoinWorkerThread
1120
1121		/**
#	Line 1083 | Line 1126 | public class ForkJoinPool extends Abstra
1126		}
1127
1128		/**
1129	<	* Accumulates steal count from a worker, clearing
1130	<	* the worker's value
1129	>	* Tries to accumulates steal count from a worker, clearing
1130	>	* the worker's value.
1131	>	*
1132	>	* @return true if worker steal count now zero
1133		*/
1134	<	final void accumulateStealCount(ForkJoinWorkerThread w) {
1134	>	final boolean tryAccumulateStealCount(ForkJoinWorkerThread w) {
1135		int sc = w.stealCount;
1136	<	if (sc != 0) {
1137	<	long c;
1138	<	w.stealCount = 0;
1139	<	do {} while (!UNSAFE.compareAndSwapLong(this, stealCountOffset,
1140	<	c = stealCount, c + sc));
1136	>	long c = stealCount;
1137	>	// CAS even if zero, for fence effects
1138	>	if (UNSAFE.compareAndSwapLong(this, stealCountOffset, c, c + sc)) {
1139	>	if (sc != 0)
1140	>	w.stealCount = 0;
1141	>	return true;
1142		}
1143	+	return sc == 0;
1144		}
1145
1146		/**
#	Line 1101 | Line 1148 | public class ForkJoinPool extends Abstra
1148		* active thread.
1149		*/
1150		final int idlePerActive() {
1151	<	int pc = parallelism; // use targeted parallelism, not rc
1151	>	int pc = parallelism; // use parallelism, not rc
1152		int ac = runState;    // no mask -- artifically boosts during shutdown
1153		// Use exact results for small values, saturate past 4
1154		return pc <= ac? 0 : pc >>> 1 <= ac? 1 : pc >>> 2 <= ac? 3 : pc >>> 3;
#	Line 1152 | Line 1199 | public class ForkJoinPool extends Abstra
1199		* use {@link java.lang.Runtime#availableProcessors}.
1200		* @param factory the factory for creating new threads. For default value,
1201		* use {@link #defaultForkJoinWorkerThreadFactory}.
1202	<	* @param handler the handler for internal worker threads that
1203	<	* terminate due to unrecoverable errors encountered while executing
1202	>	* @param handler the handler for internal worker threads that
1203	>	* terminate due to unrecoverable errors encountered while executing
1204		* tasks. For default value, use <code>null</code>.
1205	<	* @param asyncMode if true,
1205	>	* @param asyncMode if true,
1206		* establishes local first-in-first-out scheduling mode for forked
1207		* tasks that are never joined. This mode may be more appropriate
1208		* than default locally stack-based mode in applications in which
#	Line 1169 | Line 1216 | public class ForkJoinPool extends Abstra
1216		*         because it does not hold {@link
1217		*         java.lang.RuntimePermission}{@code ("modifyThread")}
1218		*/
1219	<	public ForkJoinPool(int parallelism,
1219	>	public ForkJoinPool(int parallelism,
1220		ForkJoinWorkerThreadFactory factory,
1221		Thread.UncaughtExceptionHandler handler,
1222		boolean asyncMode) {
1223		checkPermission();
1224		if (factory == null)
1225		throw new NullPointerException();
1226	<	if (parallelism <= 0 || parallelism > MAX_THREADS)
1226	>	if (parallelism <= 0 || parallelism > MAX_WORKERS)
1227		throw new IllegalArgumentException();
1228		this.parallelism = parallelism;
1229		this.factory = factory;
#	Line 1188 | Line 1235 | public class ForkJoinPool extends Abstra
1235		this.workerLock = new ReentrantLock();
1236		this.termination = new Phaser(1);
1237		this.poolNumber = poolNumberGenerator.incrementAndGet();
1238	+	this.trimTime = System.nanoTime();
1239		}
1240
1241		/**
#	Line 1195 | Line 1243 | public class ForkJoinPool extends Abstra
1243		* @param pc the initial parallelism level
1244		*/
1245		private static int initialArraySizeFor(int pc) {
1246	<	// See Hackers Delight, sec 3.2. We know MAX_THREADS < (1 >>> 16)
1247	<	int size = pc < MAX_THREADS ? pc + 1 : MAX_THREADS;
1246	>	// See Hackers Delight, sec 3.2. We know MAX_WORKERS < (1 >>> 16)
1247	>	int size = pc < MAX_WORKERS ? pc + 1 : MAX_WORKERS;
1248		size |= size >>> 1;
1249		size |= size >>> 2;
1250		size |= size >>> 4;
#	Line 1214 | Line 1262 | public class ForkJoinPool extends Abstra
1262		throw new NullPointerException();
1263		if (runState >= SHUTDOWN)
1264		throw new RejectedExecutionException();
1265	<	// Convert submissions to current pool into forks
1266	<	Thread t = Thread.currentThread();
1267	<	ForkJoinWorkerThread w;
1220	<	if ((t instanceof ForkJoinWorkerThread) &&
1221	<	(w = (ForkJoinWorkerThread) t).pool == this)
1222	<	w.pushTask(task);
1223	<	else {
1224	<	submissionQueue.offer(task);
1225	<	signalEvent();
1226	<	ensureEnoughTotalWorkers();
1227	<	}
1265	>	submissionQueue.offer(task);
1266	>	advanceEventCount();
1267	>	helpMaintainParallelism();         // start or wake up workers
1268		}
1269
1270		/**
1271		* Performs the given task, returning its result upon completion.
1272		* If the caller is already engaged in a fork/join computation in
1273	<	* the current pool, this method is equivalent in effect to
1273	>	* the current pool, this method is equivalent in effect to
1274		* {@link ForkJoinTask#invoke}.
1275		*
1276		* @param task the task
#	Line 1247 | Line 1287 | public class ForkJoinPool extends Abstra
1287		/**
1288		* Arranges for (asynchronous) execution of the given task.
1289		* If the caller is already engaged in a fork/join computation in
1290	<	* the current pool, this method is equivalent in effect to
1290	>	* the current pool, this method is equivalent in effect to
1291		* {@link ForkJoinTask#fork}.
1292		*
1293		* @param task the task
#	Line 1278 | Line 1318 | public class ForkJoinPool extends Abstra
1318		/**
1319		* Submits a ForkJoinTask for execution.
1320		* If the caller is already engaged in a fork/join computation in
1321	<	* the current pool, this method is equivalent in effect to
1321	>	* the current pool, this method is equivalent in effect to
1322		* {@link ForkJoinTask#fork}.
1323		*
1324		* @param task the task to submit
#	Line 1471 | Line 1511 | public class ForkJoinPool extends Abstra
1511		public long getQueuedTaskCount() {
1512		long count = 0;
1513		ForkJoinWorkerThread[] ws = workers;
1514	<	int nws = ws.length;
1515	<	for (int i = 0; i < nws; ++i) {
1514	>	int n = ws.length;
1515	>	for (int i = 0; i < n; ++i) {
1516		ForkJoinWorkerThread w = ws[i];
1517		if (w != null)
1518		count += w.getQueueSize();
#	Line 1530 | Line 1570 | public class ForkJoinPool extends Abstra
1570		* @return the number of elements transferred
1571		*/
1572		protected int drainTasksTo(Collection<? super ForkJoinTask<?>> c) {
1573	<	int n = submissionQueue.drainTo(c);
1534	<	ForkJoinWorkerThread[] ws = workers;
1535	<	int nws = ws.length;
1536	<	for (int i = 0; i < nws; ++i) {
1537	<	ForkJoinWorkerThread w = ws[i];
1538	<	if (w != null)
1539	<	n += w.drainTasksTo(c);
1540	<	}
1541	<	return n;
1542	<	}
1543	<
1544	<	/**
1545	<	* Returns count of total parks by existing workers.
1546	<	* Used during development only since not meaningful to users.
1547	<	*/
1548	<	private int collectParkCount() {
1549	<	int count = 0;
1573	>	int count = submissionQueue.drainTo(c);
1574		ForkJoinWorkerThread[] ws = workers;
1575	<	int nws = ws.length;
1576	<	for (int i = 0; i < nws; ++i) {
1575	>	int n = ws.length;
1576	>	for (int i = 0; i < n; ++i) {
1577		ForkJoinWorkerThread w = ws[i];
1578		if (w != null)
1579	<	count += w.parkCount;
1579	>	count += w.drainTasksTo(c);
1580		}
1581		return count;
1582		}
#	Line 1574 | Line 1598 | public class ForkJoinPool extends Abstra
1598		int pc = parallelism;
1599		int rs = runState;
1600		int ac = rs & ACTIVE_COUNT_MASK;
1577	–	//        int pk = collectParkCount();
1601		return super.toString() +
1602		"[" + runLevelToString(rs) +
1603		", parallelism = " + pc +
#	Line 1584 | Line 1607 | public class ForkJoinPool extends Abstra
1607		", steals = " + st +
1608		", tasks = " + qt +
1609		", submissions = " + qs +
1587	–	//            ", parks = " + pk +
1610		"]";
1611		}
1612
#	Line 1691 | Line 1713 | public class ForkJoinPool extends Abstra
1713		* Interface for extending managed parallelism for tasks running
1714		* in {@link ForkJoinPool}s.
1715		*
1716	<	* <p>A {@code ManagedBlocker} provides two methods.
1717	<	* Method {@code isReleasable} must return {@code true} if
1718	<	* blocking is not necessary. Method {@code block} blocks the
1719	<	* current thread if necessary (perhaps internally invoking
1720	<	* {@code isReleasable} before actually blocking).
1716	>	* <p>A {@code ManagedBlocker} provides two methods.  Method
1717	>	* {@code isReleasable} must return {@code true} if blocking is
1718	>	* not necessary. Method {@code block} blocks the current thread
1719	>	* if necessary (perhaps internally invoking {@code isReleasable}
1720	>	* before actually blocking). The unusual methods in this API
1721	>	* accommodate synchronizers that may, but don't usually, block
1722	>	* for long periods. Similarly, they allow more efficient internal
1723	>	* handling of cases in which additional workers may be, but
1724	>	* usually are not, needed to ensure sufficient parallelism.
1725	>	* Toward this end, implementations of method {@code isReleasable}
1726	>	* must be amenable to repeated invocation.
1727		*
1728		* <p>For example, here is a ManagedBlocker based on a
1729		* ReentrantLock:
#	Line 1713 | Line 1741 | public class ForkJoinPool extends Abstra
1741		*     return hasLock || (hasLock = lock.tryLock());
1742		*   }
1743		* }}</pre>
1744	+	*
1745	+	* <p>Here is a class that possibly blocks waiting for an
1746	+	* item on a given queue:
1747	+	*  <pre> {@code
1748	+	* class QueueTaker<E> implements ManagedBlocker {
1749	+	*   final BlockingQueue<E> queue;
1750	+	*   volatile E item = null;
1751	+	*   QueueTaker(BlockingQueue<E> q) { this.queue = q; }
1752	+	*   public boolean block() throws InterruptedException {
1753	+	*     if (item == null)
1754	+	*       item = queue.take
1755	+	*     return true;
1756	+	*   }
1757	+	*   public boolean isReleasable() {
1758	+	*     return item != null || (item = queue.poll) != null;
1759	+	*   }
1760	+	*   public E getItem() { // call after pool.managedBlock completes
1761	+	*     return item;
1762	+	*   }
1763	+	* }}</pre>
1764		*/
1765		public static interface ManagedBlocker {
1766		/**
#	Line 1755 | Line 1803 | public class ForkJoinPool extends Abstra
1803		public static void managedBlock(ManagedBlocker blocker)
1804		throws InterruptedException {
1805		Thread t = Thread.currentThread();
1806	<	if (t instanceof ForkJoinWorkerThread)
1807	<	((ForkJoinWorkerThread) t).pool.awaitBlocker(blocker);
1806	>	if (t instanceof ForkJoinWorkerThread) {
1807	>	ForkJoinWorkerThread w = (ForkJoinWorkerThread) t;
1808	>	w.pool.awaitBlocker(blocker);
1809	>	}
1810		else {
1811		do {} while (!blocker.isReleasable() && !blocker.block());
1812		}
#	Line 1787 | Line 1837 | public class ForkJoinPool extends Abstra
1837		objectFieldOffset("eventWaiters",ForkJoinPool.class);
1838		private static final long stealCountOffset =
1839		objectFieldOffset("stealCount",ForkJoinPool.class);
1840	<
1840	>	private static final long spareWaitersOffset =
1841	>	objectFieldOffset("spareWaiters",ForkJoinPool.class);
1842
1843		private static long objectFieldOffset(String field, Class<?> klazz) {
1844		try {

Diff Legend

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