ViewVC Help

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

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

Comparing jsr166/src/jsr166y/ForkJoinPool.java (file contents):
Revision 1.65 by dl, Wed Aug 18 14:05:27 2010 UTC vs.
Revision 1.71 by jsr166, Mon Sep 6 21:36:43 2010 UTC

#	Line 7 | Line 7
7		package jsr166y;
8
9		import java.util.concurrent.*;
10	–
10		import java.util.ArrayList;
11		import java.util.Arrays;
12		import java.util.Collection;
#	Line 69 | Line 68 | import java.util.concurrent.CountDownLat
68		*    <td ALIGN=CENTER> <b>Call from within fork/join computations</b></td>
69		*  </tr>
70		*  <tr>
71	<	*    <td> <b>Arange async execution</td>
71	>	*    <td> <b>Arrange async execution</td>
72		*    <td> {@link #execute(ForkJoinTask)}</td>
73		*    <td> {@link ForkJoinTask#fork}</td>
74		*  </tr>
#	Line 140 | Line 139 | public class ForkJoinPool extends Abstra
139		* Beyond work-stealing support and essential bookkeeping, the
140		* main responsibility of this framework is to take actions when
141		* one worker is waiting to join a task stolen (or always held by)
142	<	* another.  Becauae we are multiplexing many tasks on to a pool
142	>	* another.  Because we are multiplexing many tasks on to a pool
143		* of workers, we can't just let them block (as in Thread.join).
144		* We also cannot just reassign the joiner's run-time stack with
145		* another and replace it later, which would be a form of
#	Line 157 | Line 156 | public class ForkJoinPool extends Abstra
156		*      links to try to find such a task.
157		*
158		*   Compensating: Unless there are already enough live threads,
159	<	*      method helpMaintainParallelism() may create or or
159	>	*      method helpMaintainParallelism() may create or
160		*      re-activate a spare thread to compensate for blocked
161		*      joiners until they unblock.
162		*
163	<	* Because the determining existence of conservatively safe
164	<	* helping targets, the availability of already-created spares,
165	<	* and the apparent need to create new spares are all racy and
166	<	* require heuristic guidance, we rely on multiple retries of
167	<	* each. Further, because it is impossible to keep exactly the
168	<	* target (parallelism) number of threads running at any given
169	<	* time, we allow compensation during joins to fail, and enlist
170	<	* all other threads to help out whenever they are not otherwise
171	<	* occupied (i.e., mainly in method preStep).
163	>	* It is impossible to keep exactly the target (parallelism)
164	>	* number of threads running at any given time.  Determining
165	>	* existence of conservatively safe helping targets, the
166	>	* availability of already-created spares, and the apparent need
167	>	* to create new spares are all racy and require heuristic
168	>	* guidance, so we rely on multiple retries of each.  Compensation
169	>	* occurs in slow-motion. It is triggered only upon timeouts of
170	>	* Object.wait used for joins. This reduces poor decisions that
171	>	* would otherwise be made when threads are waiting for others
172	>	* that are stalled because of unrelated activities such as
173	>	* garbage collection.
174		*
175		* The ManagedBlocker extension API can't use helping so relies
176		* only on compensation in method awaitBlocker.
#	Line 224 | Line 225 | public class ForkJoinPool extends Abstra
225		* ManagedBlocker), we may create or resume others to take their
226		* place until they unblock (see below). Implementing this
227		* requires counts of the number of "running" threads (i.e., those
228	<	* that are neither blocked nor artifically suspended) as well as
228	>	* that are neither blocked nor artificially suspended) as well as
229		* the total number.  These two values are packed into one field,
230		* "workerCounts" because we need accurate snapshots when deciding
231		* to create, resume or suspend.  Note however that the
232	<	* correspondance of these counts to reality is not guaranteed. In
232	>	* correspondence of these counts to reality is not guaranteed. In
233		* particular updates for unblocked threads may lag until they
234		* actually wake up.
235		*
#	Line 271 | Line 272 | public class ForkJoinPool extends Abstra
272		* In addition to allowing simpler decisions about need for
273		* wakeup, the event count bits in eventWaiters serve the role of
274		* tags to avoid ABA errors in Treiber stacks. Upon any wakeup,
275	<	* released threads also try to release others (but give up upon
276	<	* contention to reduce useless flailing).  The net effect is a
277	<	* tree-like diffusion of signals, where released threads (and
278	<	* possibly others) help with unparks.  To further reduce
279	<	* contention effects a bit, failed CASes to increment field
279	<	* eventCount are tolerated without retries in signalWork.
275	>	* released threads also try to release at most two others.  The
276	>	* net effect is a tree-like diffusion of signals, where released
277	>	* threads (and possibly others) help with unparks.  To further
278	>	* reduce contention effects a bit, failed CASes to increment
279	>	* field eventCount are tolerated without retries in signalWork.
280		* Conceptually they are merged into the same event, which is OK
281		* when their only purpose is to enable workers to scan for work.
282		*
283	<	* 5. Managing suspension of extra workers. When a worker is about
284	<	* to block waiting for a join (or via ManagedBlockers), we may
285	<	* create a new thread to maintain parallelism level, or at least
286	<	* avoid starvation. Usually, extra threads are needed for only
287	<	* very short periods, yet join dependencies are such that we
288	<	* sometimes need them in bursts. Rather than create new threads
289	<	* each time this happens, we suspend no-longer-needed extra ones
290	<	* as "spares". For most purposes, we don't distinguish "extra"
291	<	* spare threads from normal "core" threads: On each call to
292	<	* preStep (the only point at which we can do this) a worker
293	<	* checks to see if there are now too many running workers, and if
294	<	* so, suspends itself.  Method helpMaintainParallelism looks for
295	<	* suspended threads to resume before considering creating a new
296	<	* replacement. The spares themselves are encoded on another
297	<	* variant of a Treiber Stack, headed at field "spareWaiters".
298	<	* Note that the use of spares is intrinsically racy.  One thread
299	<	* may become a spare at about the same time as another is
300	<	* needlessly being created. We counteract this and related slop
301	<	* in part by requiring resumed spares to immediately recheck (in
302	<	* preStep) to see whether they they should re-suspend.
303	<	*
304	<	* 6. Killing off unneeded workers. The Spare and Event queues use
305	<	* similar mechanisms to shed unused workers: The oldest (first)
306	<	* waiter uses a timed rather than hard wait. When this wait times
307	<	* out without a normal wakeup, it tries to shutdown any one (for
308	<	* convenience the newest) other waiter via tryShutdownSpare or
309	<	* tryShutdownWaiter, respectively. The wakeup rates for spares
310	<	* are much shorter than for waiters. Together, they will
311	<	* eventually reduce the number of worker threads to a minimum of
312	<	* one after a long enough period without use.
283	>	* 5. Managing suspension of extra workers. When a worker notices
284	>	* (usually upon timeout of a wait()) that there are too few
285	>	* running threads, we may create a new thread to maintain
286	>	* parallelism level, or at least avoid starvation. Usually, extra
287	>	* threads are needed for only very short periods, yet join
288	>	* dependencies are such that we sometimes need them in
289	>	* bursts. Rather than create new threads each time this happens,
290	>	* we suspend no-longer-needed extra ones as "spares". For most
291	>	* purposes, we don't distinguish "extra" spare threads from
292	>	* normal "core" threads: On each call to preStep (the only point
293	>	* at which we can do this) a worker checks to see if there are
294	>	* now too many running workers, and if so, suspends itself.
295	>	* Method helpMaintainParallelism looks for suspended threads to
296	>	* resume before considering creating a new replacement. The
297	>	* spares themselves are encoded on another variant of a Treiber
298	>	* Stack, headed at field "spareWaiters".  Note that the use of
299	>	* spares is intrinsically racy.  One thread may become a spare at
300	>	* about the same time as another is needlessly being created. We
301	>	* counteract this and related slop in part by requiring resumed
302	>	* spares to immediately recheck (in preStep) to see whether they
303	>	* they should re-suspend.
304	>	*
305	>	* 6. Killing off unneeded workers. A timeout mechanism is used to
306	>	* shed unused workers: The oldest (first) event queue waiter uses
307	>	* a timed rather than hard wait. When this wait times out without
308	>	* a normal wakeup, it tries to shutdown any one (for convenience
309	>	* the newest) other spare or event waiter via
310	>	* tryShutdownUnusedWorker. This eventually reduces the number of
311	>	* worker threads to a minimum of one after a long enough period
312	>	* without use.
313		*
314		* 7. Deciding when to create new workers. The main dynamic
315		* control in this class is deciding when to create extra threads
316		* in method helpMaintainParallelism. We would like to keep
317	<	* exactly #parallelism threads running, which is an impossble
317	>	* exactly #parallelism threads running, which is an impossible
318		* task. We always need to create one when the number of running
319		* threads would become zero and all workers are busy. Beyond
320	<	* this, we must rely on heuristics that work well in the the
321	<	* presence of transients phenomena such as GC stalls, dynamic
320	>	* this, we must rely on heuristics that work well in the
321	>	* presence of transient phenomena such as GC stalls, dynamic
322		* compilation, and wake-up lags. These transients are extremely
323		* common -- we are normally trying to fully saturate the CPUs on
324		* a machine, so almost any activity other than running tasks
325	<	* impedes accuracy. Our main defense is to allow some slack in
326	<	* creation thresholds, using rules that reflect the fact that the
327	<	* more threads we have running, the more likely that we are
328	<	* underestimating the number running threads. (We also include
329	<	* some heuristic use of Thread.yield when all workers appear to
330	<	* be busy, to improve likelihood of counts settling.) The rules
331	<	* also better cope with the fact that some of the methods in this
332	<	* class tend to never become compiled (but are interpreted), so
333	<	* some components of the entire set of controls might execute 100
334	<	* times faster than others. And similarly for cases where the
335	<	* apparent lack of work is just due to GC stalls and other
336	<	* transient system activity.
325	>	* impedes accuracy. Our main defense is to allow parallelism to
326	>	* lapse for a while during joins, and use a timeout to see if,
327	>	* after the resulting settling, there is still a need for
328	>	* additional workers.  This also better copes with the fact that
329	>	* some of the methods in this class tend to never become compiled
330	>	* (but are interpreted), so some components of the entire set of
331	>	* controls might execute 100 times faster than others. And
332	>	* similarly for cases where the apparent lack of work is just due
333	>	* to GC stalls and other transient system activity.
334		*
335		* Beware that there is a lot of representation-level coupling
336		* among classes ForkJoinPool, ForkJoinWorkerThread, and
#	Line 348 | Line 345 | public class ForkJoinPool extends Abstra
345		* "while ((local = field) != 0)") which are usually the simplest
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
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
#	Line 420 | Line 417 | public class ForkJoinPool extends Abstra
417		new AtomicInteger();
418
419		/**
420	+	* The time to block in a join (see awaitJoin) before checking if
421	+	* a new worker should be (re)started to maintain parallelism
422	+	* level. The value should be short enough to maintain global
423	+	* responsiveness and progress but long enough to avoid
424	+	* counterproductive firings during GC stalls or unrelated system
425	+	* activity, and to not bog down systems with continual re-firings
426	+	* on GCs or legitimately long waits.
427	+	*/
428	+	private static final long JOIN_TIMEOUT_MILLIS = 250L; // 4 per second
429	+
430	+	/**
431		* The wakeup interval (in nanoseconds) for the oldest worker
432	<	* worker waiting for an event invokes tryShutdownWaiter to shrink
432	>	* worker waiting for an event invokes tryShutdownUnusedWorker to shrink
433		* the number of workers.  The exact value does not matter too
434		* much, but should be long enough to slowly release resources
435		* during long periods without use without disrupting normal use.
436		*/
437		private static final long SHRINK_RATE_NANOS =
438	<	60L * 1000L * 1000L * 1000L; // one minute
438	>	30L * 1000L * 1000L * 1000L; // 2 per minute
439
440		/**
441		* Absolute bound for parallelism level. Twice this number plus
#	Line 474 | Line 482 | public class ForkJoinPool extends Abstra
482		private volatile long stealCount;
483
484		/**
485	<	* Encoded record of top of treiber stack of threads waiting for
485	>	* Encoded record of top of Treiber stack of threads waiting for
486		* events. The top 32 bits contain the count being waited for. The
487		* bottom 16 bits contains one plus the pool index of waiting
488		* worker thread. (Bits 16-31 are unused.)
#	Line 493 | Line 501 | public class ForkJoinPool extends Abstra
501		private volatile int eventCount;
502
503		/**
504	<	* Encoded record of top of treiber stack of spare threads waiting
504	>	* Encoded record of top of Treiber stack of spare threads waiting
505		* for resumption. The top 16 bits contain an arbitrary count to
506		* avoid ABA effects. The bottom 16bits contains one plus the pool
507		* index of waiting worker thread.
#	Line 567 | Line 575 | public class ForkJoinPool extends Abstra
575		*/
576		private final int poolNumber;
577
570	–
578		// Utilities for CASing fields. Note that most of these
579		// are usually manually inlined by callers
580
#	Line 615 | Line 622 | public class ForkJoinPool extends Abstra
622		}
623
624		/**
618	–	* Increments event count
619	–	*/
620	–	private void advanceEventCount() {
621	–	int c;
622	–	do {} while(!UNSAFE.compareAndSwapInt(this, eventCountOffset,
623	–	c = eventCount, c+1));
624	–	}
625	–
626	–	/**
627	–	* Tries incrementing active count; fails on contention.
628	–	* Called by workers before executing tasks.
629	–	*
630	–	* @return true on success
631	–	*/
632	–	final boolean tryIncrementActiveCount() {
633	–	int c;
634	–	return UNSAFE.compareAndSwapInt(this, runStateOffset,
635	–	c = runState, c + 1);
636	–	}
637	–
638	–	/**
625		* Tries decrementing active count; fails on contention.
626		* Called when workers cannot find tasks to run.
627		*/
#	Line 687 | Line 673 | public class ForkJoinPool extends Abstra
673		}
674
675		/**
676	<	* Nulls out record of worker in workers array
676	>	* Nulls out record of worker in workers array.
677		*/
678		private void forgetWorker(ForkJoinWorkerThread w) {
679		int idx = w.poolIndex;
680	<	// Locking helps method recordWorker avoid unecessary expansion
680	>	// Locking helps method recordWorker avoid unnecessary expansion
681		final ReentrantLock lock = this.workerLock;
682		lock.lock();
683		try {
#	Line 703 | Line 689 | public class ForkJoinPool extends Abstra
689		}
690		}
691
706	–	// adding and removing workers
707	–
708	–	/**
709	–	* Tries to create and add new worker. Assumes that worker counts
710	–	* are already updated to accommodate the worker, so adjusts on
711	–	* failure.
712	–	*
713	–	* @return the worker, or null on failure
714	–	*/
715	–	private ForkJoinWorkerThread addWorker() {
716	–	ForkJoinWorkerThread w = null;
717	–	try {
718	–	w = factory.newThread(this);
719	–	} finally { // Adjust on either null or exceptional factory return
720	–	if (w == null) {
721	–	decrementWorkerCounts(ONE_RUNNING, ONE_TOTAL);
722	–	tryTerminate(false); // in case of failure during shutdown
723	–	}
724	–	}
725	–	if (w != null) {
726	–	w.start(recordWorker(w), ueh);
727	–	advanceEventCount();
728	–	}
729	–	return w;
730	–	}
731	–
692		/**
693		* Final callback from terminating worker.  Removes record of
694		* worker from array, and adjusts counts. If pool is shutting
695	<	* down, tries to complete terminatation.
695	>	* down, tries to complete termination.
696		*
697		* @param w the worker
698		*/
#	Line 750 | Line 710 | public class ForkJoinPool extends Abstra
710		* Releases workers blocked on a count not equal to current count.
711		* Normally called after precheck that eventWaiters isn't zero to
712		* avoid wasted array checks. Gives up upon a change in count or
713	<	* contention, letting other workers take over.
713	>	* upon releasing two workers, letting others take over.
714		*/
715		private void releaseEventWaiters() {
716		ForkJoinWorkerThread[] ws = workers;
717		int n = ws.length;
718		long h = eventWaiters;
719		int ec = eventCount;
720	+	boolean releasedOne = false;
721		ForkJoinWorkerThread w; int id;
722	<	while ((int)(h >>> EVENT_COUNT_SHIFT) != ec &&
723	<	(id = ((int)(h & WAITER_ID_MASK)) - 1) >= 0 &&
724	<	id < n && (w = ws[id]) != null &&
725	<	UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
726	<	h,  h = w.nextWaiter)) {
727	<	LockSupport.unpark(w);
728	<	if (eventWaiters != h || eventCount != ec)
722	>	while ((id = ((int)(h & WAITER_ID_MASK)) - 1) >= 0 &&
723	>	(int)(h >>> EVENT_COUNT_SHIFT) != ec &&
724	>	id < n && (w = ws[id]) != null) {
725	>	if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
726	>	h,  w.nextWaiter)) {
727	>	LockSupport.unpark(w);
728	>	if (releasedOne) // exit on second release
729	>	break;
730	>	releasedOne = true;
731	>	}
732	>	if (eventCount != ec)
733		break;
734	+	h = eventWaiters;
735		}
736		}
737
#	Line 782 | Line 748 | public class ForkJoinPool extends Abstra
748
749		/**
750		* Adds the given worker to event queue and blocks until
751	<	* terminating or event count advances from the workers
786	<	* lastEventCount value
751	>	* terminating or event count advances from the given value
752		*
753		* @param w the calling worker thread
754	+	* @param ec the count
755		*/
756	<	private void eventSync(ForkJoinWorkerThread w) {
791	<	int ec = w.lastEventCount;
756	>	private void eventSync(ForkJoinWorkerThread w, int ec) {
757		long nh = (((long)ec) << EVENT_COUNT_SHIFT) | ((long)(w.poolIndex+1));
758		long h;
759		while ((runState < SHUTDOWN || !tryTerminate(false)) &&
#	Line 808 | Line 773 | public class ForkJoinPool extends Abstra
773		* event waiter) until terminating or event count advances from
774		* the given value. The oldest (first) waiter uses a timed wait to
775		* occasionally one-by-one shrink the number of workers (to a
776	<	* minumum of one) if the pool has not been used for extended
776	>	* minimum of one) if the pool has not been used for extended
777		* periods.
778		*
779		* @param w the calling worker thread
#	Line 821 | Line 786 | public class ForkJoinPool extends Abstra
786		(workerCounts & RUNNING_COUNT_MASK) <= 1);
787		long startTime = untimed? 0 : System.nanoTime();
788		Thread.interrupted();         // clear/ignore interrupt
789	<	if (eventCount != ec || !w.isRunning() ||
789	>	if (eventCount != ec || w.runState != 0 ||
790		runState >= TERMINATING)  // recheck after clear
791		break;
792		if (untimed)
793		LockSupport.park(w);
794		else {
795		LockSupport.parkNanos(w, SHRINK_RATE_NANOS);
796	<	if (eventCount != ec || !w.isRunning() ||
796	>	if (eventCount != ec || w.runState != 0 ||
797		runState >= TERMINATING)
798		break;
799		if (System.nanoTime() - startTime >= SHRINK_RATE_NANOS)
800	<	tryShutdownWaiter(ec);
800	>	tryShutdownUnusedWorker(ec);
801		}
802		}
803		}
804		}
805
806	<	/**
842	<	* Callback from the oldest waiter in awaitEvent waking up after a
843	<	* period of non-use. Tries (once) to shutdown an event waiter (or
844	<	* a spare, if one exists). Note that we don't need CAS or locks
845	<	* here because the method is called only from one thread
846	<	* occasionally waking (and even misfires are OK). Note that
847	<	* until the shutdown worker fully terminates, workerCounts
848	<	* will overestimate total count, which is tolerable.
849	<	*
850	<	* @param ec the event count waited on by caller (to abort
851	<	* attempt if count has since changed).
852	<	*/
853	<	private void tryShutdownWaiter(int ec) {
854	<	if (spareWaiters != 0) { // prefer killing spares
855	<	tryShutdownSpare();
856	<	return;
857	<	}
858	<	ForkJoinWorkerThread[] ws = workers;
859	<	int n = ws.length;
860	<	long h = eventWaiters;
861	<	ForkJoinWorkerThread w; int id; long nh;
862	<	if (runState == 0 &&
863	<	submissionQueue.isEmpty() &&
864	<	eventCount == ec &&
865	<	(id = ((int)(h & WAITER_ID_MASK)) - 1) >= 0 &&
866	<	id < n && (w = ws[id]) != null &&
867	<	(nh = w.nextWaiter) != 0L && // keep at least one worker
868	<	UNSAFE.compareAndSwapLong(this, eventWaitersOffset, h, nh)) {
869	<	w.shutdown();
870	<	LockSupport.unpark(w);
871	<	}
872	<	releaseEventWaiters();
873	<	}
874	<
875	<	// Maintaining spares
806	>	// Maintaining parallelism
807
808		/**
809		* Pushes worker onto the spare stack
#	Line 884 | Line 815 | public class ForkJoinPool extends Abstra
815		}
816
817		/**
818	<	* Callback from oldest spare occasionally waking up.  Tries
819	<	* (once) to shutdown a spare. Same idea as tryShutdownWaiter.
818	>	* Tries (once) to resume a spare if the number of running
819	>	* threads is less than target.
820		*/
821	<	final void tryShutdownSpare() {
821	>	private void tryResumeSpare() {
822		int sw, id;
892	–	ForkJoinWorkerThread w;
893	–	ForkJoinWorkerThread[] ws;
894	–	if ((id = ((sw = spareWaiters) & SPARE_ID_MASK) - 1) >= 0 &&
895	–	id < (ws = workers).length && (w = ws[id]) != null &&
896	–	(workerCounts & RUNNING_COUNT_MASK) >= parallelism &&
897	–	UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
898	–	sw, w.nextSpare)) {
899	–	w.shutdown();
900	–	LockSupport.unpark(w);
901	–	advanceEventCount();
902	–	}
903	–	}
904	–
905	–	/**
906	–	* Tries (once) to resume a spare if worker counts match
907	–	* the given count.
908	–	*
909	–	* @param wc workerCounts value on invocation of this method
910	–	*/
911	–	private void tryResumeSpare(int wc) {
823		ForkJoinWorkerThread[] ws = workers;
824		int n = ws.length;
825	<	int sw, id, rs;  ForkJoinWorkerThread w;
826	<	if ((id = ((sw = spareWaiters) & SPARE_ID_MASK) - 1) >= 0 &&
825	>	ForkJoinWorkerThread w;
826	>	if ((sw = spareWaiters) != 0 &&
827	>	(id = (sw & SPARE_ID_MASK) - 1) >= 0 &&
828		id < n && (w = ws[id]) != null &&
829	<	(rs = runState) < TERMINATING &&
830	<	eventWaiters == 0L && workerCounts == wc) {
831	<	// In case all workers busy, heuristically back off to let settle
832	<	Thread.yield();
833	<	if (eventWaiters == 0L && runState == rs && // recheck
834	<	workerCounts == wc && spareWaiters == sw &&
835	<	UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
836	<	sw, w.nextSpare)) {
837	<	int c;              // increment running count before resume
838	<	do {} while(!UNSAFE.compareAndSwapInt
839	<	(this, workerCountsOffset,
840	<	c = workerCounts, c + ONE_RUNNING));
929	<	if (w.tryUnsuspend())
930	<	LockSupport.unpark(w);
931	<	else               // back out if w was shutdown
932	<	decrementWorkerCounts(ONE_RUNNING, 0);
933	<	}
829	>	(workerCounts & RUNNING_COUNT_MASK) < parallelism &&
830	>	spareWaiters == sw &&
831	>	UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
832	>	sw, w.nextSpare)) {
833	>	int c; // increment running count before resume
834	>	do {} while (!UNSAFE.compareAndSwapInt
835	>	(this, workerCountsOffset,
836	>	c = workerCounts, c + ONE_RUNNING));
837	>	if (w.tryUnsuspend())
838	>	LockSupport.unpark(w);
839	>	else   // back out if w was shutdown
840	>	decrementWorkerCounts(ONE_RUNNING, 0);
841		}
842		}
843
937	–	// adding workers on demand
938	–
844		/**
845	<	* Adds one or more workers if needed to establish target parallelism.
846	<	* Retries upon contention.
845	>	* Tries to increase the number of running workers if below target
846	>	* parallelism: If a spare exists tries to resume it via
847	>	* tryResumeSpare.  Otherwise, if not enough total workers or all
848	>	* existing workers are busy, adds a new worker. In all cases also
849	>	* helps wake up releasable workers waiting for work.
850		*/
851	<	private void addWorkerIfBelowTarget() {
851	>	private void helpMaintainParallelism() {
852		int pc = parallelism;
853	<	int wc;
854	<	while (((wc = workerCounts) >>> TOTAL_COUNT_SHIFT) < pc &&
855	<	runState < TERMINATING) {
856	<	if (UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
857	<	wc + (ONE_RUNNING|ONE_TOTAL))) {
858	<	if (addWorker() == null)
853	>	int wc, rs, tc;
854	>	while (((wc = workerCounts) & RUNNING_COUNT_MASK) < pc &&
855	>	(rs = runState) < TERMINATING) {
856	>	if (spareWaiters != 0)
857	>	tryResumeSpare();
858	>	else if ((tc = wc >>> TOTAL_COUNT_SHIFT) >= MAX_WORKERS ||
859	>	(tc >= pc && (rs & ACTIVE_COUNT_MASK) != tc))
860	>	break;   // enough total
861	>	else if (runState == rs && workerCounts == wc &&
862	>	UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
863	>	wc + (ONE_RUNNING|ONE_TOTAL))) {
864	>	ForkJoinWorkerThread w = null;
865	>	try {
866	>	w = factory.newThread(this);
867	>	} finally { // adjust on null or exceptional factory return
868	>	if (w == null) {
869	>	decrementWorkerCounts(ONE_RUNNING, ONE_TOTAL);
870	>	tryTerminate(false); // handle failure during shutdown
871	>	}
872	>	}
873	>	if (w == null)
874		break;
875	+	w.start(recordWorker(w), ueh);
876	+	if ((workerCounts >>> TOTAL_COUNT_SHIFT) >= pc) {
877	+	int c; // advance event count
878	+	UNSAFE.compareAndSwapInt(this, eventCountOffset,
879	+	c = eventCount, c+1);
880	+	break; // add at most one unless total below target
881	+	}
882		}
883		}
884	+	if (eventWaiters != 0L)
885	+	releaseEventWaiters();
886		}
887
888		/**
889	<	* Tries (once) to add a new worker if all existing workers are
890	<	* busy, and there are either no running workers or the deficit is
891	<	* at least twice the surplus.
892	<	*
893	<	* @param wc workerCounts value on invocation of this method
894	<	*/
895	<	private void tryAddWorkerIfBusy(int wc) {
964	<	int tc, rc, rs;
965	<	int pc = parallelism;
966	<	if ((tc = wc >>> TOTAL_COUNT_SHIFT) < MAX_WORKERS &&
967	<	((rc = wc & RUNNING_COUNT_MASK) == 0 ||
968	<	rc < pc - ((tc - pc) << 1)) &&
969	<	(rs = runState) < TERMINATING &&
970	<	(rs & ACTIVE_COUNT_MASK) == tc) {
971	<	// Since all workers busy, heuristically back off to let settle
972	<	Thread.yield();
973	<	if (eventWaiters == 0L && spareWaiters == 0 && // recheck
974	<	runState == rs && workerCounts == wc &&
975	<	UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
976	<	wc + (ONE_RUNNING|ONE_TOTAL)))
977	<	addWorker();
978	<	}
979	<	}
980	<
981	<	/**
982	<	* Does at most one of:
983	<	*
984	<	* 1. Help wake up existing workers waiting for work via
985	<	*    releaseEventWaiters. (If any exist, then it doesn't
986	<	*    matter right now if under target parallelism level.)
987	<	*
988	<	* 2. If a spare exists, try (once) to resume it via tryResumeSpare.
989	<	*
990	<	* 3. If there are not enough total workers, add some
991	<	*    via addWorkerIfBelowTarget;
889	>	* Callback from the oldest waiter in awaitEvent waking up after a
890	>	* period of non-use. If all workers are idle, tries (once) to
891	>	* shutdown an event waiter or a spare, if one exists. Note that
892	>	* we don't need CAS or locks here because the method is called
893	>	* only from one thread occasionally waking (and even misfires are
894	>	* OK). Note that until the shutdown worker fully terminates,
895	>	* workerCounts will overestimate total count, which is tolerable.
896		*
897	<	* 4. Try (once) to add a new worker if all existing workers
898	<	*     are busy, via tryAddWorkerIfBusy
897	>	* @param ec the event count waited on by caller (to abort
898	>	* attempt if count has since changed).
899		*/
900	<	private void helpMaintainParallelism() {
901	<	long h; int pc, wc;
902	<	if (((int)((h = eventWaiters) & WAITER_ID_MASK)) != 0) {
903	<	if ((int)(h >>> EVENT_COUNT_SHIFT) != eventCount)
904	<	releaseEventWaiters(); // avoid useless call
905	<	}
906	<	else if ((pc = parallelism) >
907	<	((wc = workerCounts) & RUNNING_COUNT_MASK)) {
908	<	if (spareWaiters != 0)
909	<	tryResumeSpare(wc);
910	<	else if ((wc >>> TOTAL_COUNT_SHIFT) < pc)
911	<	addWorkerIfBelowTarget();
912	<	else
913	<	tryAddWorkerIfBusy(wc);
900	>	private void tryShutdownUnusedWorker(int ec) {
901	>	if (runState == 0 && eventCount == ec) { // only trigger if all idle
902	>	ForkJoinWorkerThread[] ws = workers;
903	>	int n = ws.length;
904	>	ForkJoinWorkerThread w = null;
905	>	boolean shutdown = false;
906	>	int sw;
907	>	long h;
908	>	if ((sw = spareWaiters) != 0) { // prefer killing spares
909	>	int id = (sw & SPARE_ID_MASK) - 1;
910	>	if (id >= 0 && id < n && (w = ws[id]) != null &&
911	>	UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
912	>	sw, w.nextSpare))
913	>	shutdown = true;
914	>	}
915	>	else if ((h = eventWaiters) != 0L) {
916	>	long nh;
917	>	int id = ((int)(h & WAITER_ID_MASK)) - 1;
918	>	if (id >= 0 && id < n && (w = ws[id]) != null &&
919	>	(nh = w.nextWaiter) != 0L && // keep at least one worker
920	>	UNSAFE.compareAndSwapLong(this, eventWaitersOffset, h, nh))
921	>	shutdown = true;
922	>	}
923	>	if (w != null && shutdown) {
924	>	w.shutdown();
925	>	LockSupport.unpark(w);
926	>	}
927		}
928	+	releaseEventWaiters(); // in case of interference
929		}
930
931		/**
#	Line 1015 | Line 933 | public class ForkJoinPool extends Abstra
933		* stealing a task or taking a submission and running it).
934		* Performs one or more of the following:
935		*
936	<	* 1. If the worker is active, try to set its active status to
937	<	*    inactive and update activeCount. On contention, we may try
938	<	*    again on this or subsequent call.
939	<	*
940	<	* 2. Release any existing event waiters that are now relesable
941	<	*
942	<	* 3. If there are too many running threads, suspend this worker
943	<	*    (first forcing inactive if necessary).  If it is not
944	<	*    needed, it may be killed while suspended via
945	<	*    tryShutdownSpare. Otherwise, upon resume it rechecks to make
946	<	*    sure that it is still needed.
947	<	*
948	<	* 4. If more than 1 miss, await the next task event via
949	<	*    eventSync (first forcing inactivation if necessary), upon
950	<	*    which worker may also be killed, via tryShutdownWaiter.
951	<	*
952	<	* 5. Help reactivate other workers via helpMaintainParallelism
936	>	* 1. If the worker is active and either did not run a task
937	>	*    or there are too many workers, try to set its active status
938	>	*    to inactive and update activeCount. On contention, we may
939	>	*    try again in this or a subsequent call.
940	>	*
941	>	* 2. If not enough total workers, help create some.
942	>	*
943	>	* 3. If there are too many running workers, suspend this worker
944	>	*    (first forcing inactive if necessary).  If it is not needed,
945	>	*    it may be shutdown while suspended (via
946	>	*    tryShutdownUnusedWorker).  Otherwise, upon resume it
947	>	*    rechecks running thread count and need for event sync.
948	>	*
949	>	* 4. If worker did not run a task, await the next task event via
950	>	*    eventSync if necessary (first forcing inactivation), upon
951	>	*    which the worker may be shutdown via
952	>	*    tryShutdownUnusedWorker.  Otherwise, help release any
953	>	*    existing event waiters that are now releasable,
954		*
955		* @param w the worker
956	<	* @param misses the number of scans by caller failing to find work
1038	<	* (saturating at 2 to avoid wraparound)
956	>	* @param ran true if worker ran a task since last call to this method
957		*/
958	<	final void preStep(ForkJoinWorkerThread w, int misses) {
958	>	final void preStep(ForkJoinWorkerThread w, boolean ran) {
959	>	int wec = w.lastEventCount;
960		boolean active = w.active;
961	+	boolean inactivate = false;
962		int pc = parallelism;
963	<	for (;;) {
964	<	int rs, wc, rc, ec; long h;
965	<	if (active && UNSAFE.compareAndSwapInt(this, runStateOffset,
966	<	rs = runState, rs - 1))
967	<	active = w.active = false;
968	<	if (((int)((h = eventWaiters) & WAITER_ID_MASK)) != 0 &&
969	<	(int)(h >>> EVENT_COUNT_SHIFT) != eventCount) {
970	<	releaseEventWaiters();
1051	<	if (misses > 1)
1052	<	continue;                  // clear before sync below
1053	<	}
1054	<	if ((rc = ((wc = workerCounts) & RUNNING_COUNT_MASK)) > pc) {
1055	<	if (!active &&                 // must inactivate to suspend
963	>	int rs;
964	>	while (w.runState == 0 && (rs = runState) < TERMINATING) {
965	>	if ((inactivate || (active && (rs & ACTIVE_COUNT_MASK) >= pc)) &&
966	>	UNSAFE.compareAndSwapInt(this, runStateOffset, rs, rs - 1))
967	>	inactivate = active = w.active = false;
968	>	int wc = workerCounts;
969	>	if ((wc & RUNNING_COUNT_MASK) > pc) {
970	>	if (!(inactivate |= active) && // must inactivate to suspend
971		workerCounts == wc &&      // try to suspend as spare
972		UNSAFE.compareAndSwapInt(this, workerCountsOffset,
973	<	wc, wc - ONE_RUNNING)) {
973	>	wc, wc - ONE_RUNNING))
974		w.suspendAsSpare();
1060	–	if (!w.isRunning())
1061	–	break;                 // was killed while spare
1062	–	}
1063	–	continue;
975		}
976	<	if (misses > 0) {
977	<	if ((ec = eventCount) == w.lastEventCount && misses > 1) {
978	<	if (!active) {             // must inactivate to sync
979	<	eventSync(w);
980	<	if (w.isRunning())
981	<	misses = 1;        // don't re-sync
982	<	else
983	<	break;             // was killed while waiting
984	<	}
985	<	continue;
976	>	else if ((wc >>> TOTAL_COUNT_SHIFT) < pc)
977	>	helpMaintainParallelism();     // not enough workers
978	>	else if (!ran) {
979	>	long h = eventWaiters;
980	>	int ec = eventCount;
981	>	if (h != 0L && (int)(h >>> EVENT_COUNT_SHIFT) != ec)
982	>	releaseEventWaiters();     // release others before waiting
983	>	else if (ec != wec) {
984	>	w.lastEventCount = ec;     // no need to wait
985	>	break;
986		}
987	<	w.lastEventCount = ec;
987	>	else if (!(inactivate |= active))
988	>	eventSync(w, wec);         // must inactivate before sync
989		}
990	<	if (rc < pc)
991	<	helpMaintainParallelism();
1080	<	break;
990	>	else
991	>	break;
992		}
993		}
994
995		/**
996		* Helps and/or blocks awaiting join of the given task.
997	<	* Alternates between helpJoinTask() and helpMaintainParallelism()
1087	<	* as many times as there is a deficit in running count (or longer
1088	<	* if running count would become zero), then blocks if task still
1089	<	* not done.
997	>	* See above for explanation.
998		*
999		* @param joinMe the task to join
1000	+	* @param worker the current worker thread
1001		*/
1002		final void awaitJoin(ForkJoinTask<?> joinMe, ForkJoinWorkerThread worker) {
1003	<	int threshold = parallelism;         // descend blocking thresholds
1003	>	int retries = 2 + (parallelism >> 2); // #helpJoins before blocking
1004		while (joinMe.status >= 0) {
1005	<	boolean block; int wc;
1005	>	int wc;
1006		worker.helpJoinTask(joinMe);
1007		if (joinMe.status < 0)
1008		break;
1009	<	if (((wc = workerCounts) & RUNNING_COUNT_MASK) <= threshold) {
1010	<	if (threshold > 0)
1011	<	--threshold;
1012	<	else
1013	<	advanceEventCount(); // force release
1014	<	block = false;
1015	<	}
1016	<	else
1017	<	block = UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1018	<	wc, wc - ONE_RUNNING);
1019	<	helpMaintainParallelism();
1020	<	if (block) {
1021	<	int c;
1022	<	joinMe.internalAwaitDone();
1009	>	else if (retries > 0)
1010	>	--retries;
1011	>	else if (((wc = workerCounts) & RUNNING_COUNT_MASK) != 0 &&
1012	>	UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1013	>	wc, wc - ONE_RUNNING)) {
1014	>	int stat, c; long h;
1015	>	while ((stat = joinMe.status) >= 0 &&
1016	>	(h = eventWaiters) != 0L && // help release others
1017	>	(int)(h >>> EVENT_COUNT_SHIFT) != eventCount)
1018	>	releaseEventWaiters();
1019	>	if (stat >= 0 &&
1020	>	((workerCounts & RUNNING_COUNT_MASK) == 0 ||
1021	>	(stat =
1022	>	joinMe.internalAwaitDone(JOIN_TIMEOUT_MILLIS)) >= 0))
1023	>	helpMaintainParallelism(); // timeout or no running workers
1024		do {} while (!UNSAFE.compareAndSwapInt
1025		(this, workerCountsOffset,
1026		c = workerCounts, c + ONE_RUNNING));
1027	<	break;
1027	>	if (stat < 0)
1028	>	break;   // else restart
1029		}
1030		}
1031		}
1032
1033		/**
1034	<	* Same idea as awaitJoin, but no helping
1034	>	* Same idea as awaitJoin, but no helping, retries, or timeouts.
1035		*/
1036		final void awaitBlocker(ManagedBlocker blocker)
1037		throws InterruptedException {
1127	–	int threshold = parallelism;
1038		while (!blocker.isReleasable()) {
1039	<	boolean block; int wc;
1040	<	if (((wc = workerCounts) & RUNNING_COUNT_MASK) <= threshold) {
1041	<	if (threshold > 0)
1042	<	--threshold;
1133	<	else
1134	<	advanceEventCount();
1135	<	block = false;
1136	<	}
1137	<	else
1138	<	block = UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1139	<	wc, wc - ONE_RUNNING);
1140	<	helpMaintainParallelism();
1141	<	if (block) {
1039	>	int wc = workerCounts;
1040	>	if ((wc & RUNNING_COUNT_MASK) != 0 &&
1041	>	UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1042	>	wc, wc - ONE_RUNNING)) {
1043		try {
1044	<	do {} while (!blocker.isReleasable() && !blocker.block());
1044	>	while (!blocker.isReleasable()) {
1045	>	long h = eventWaiters;
1046	>	if (h != 0L &&
1047	>	(int)(h >>> EVENT_COUNT_SHIFT) != eventCount)
1048	>	releaseEventWaiters();
1049	>	else if ((workerCounts & RUNNING_COUNT_MASK) == 0 &&
1050	>	runState < TERMINATING)
1051	>	helpMaintainParallelism();
1052	>	else if (blocker.block())
1053	>	break;
1054	>	}
1055		} finally {
1056		int c;
1057		do {} while (!UNSAFE.compareAndSwapInt
#	Line 1190 | Line 1101 | public class ForkJoinPool extends Abstra
1101		private void startTerminating() {
1102		cancelSubmissions();
1103		for (int passes = 0; passes < 4 && workerCounts != 0; ++passes) {
1104	<	advanceEventCount();
1104	>	int c; // advance event count
1105	>	UNSAFE.compareAndSwapInt(this, eventCountOffset,
1106	>	c = eventCount, c+1);
1107		eventWaiters = 0L; // clobber lists
1108		spareWaiters = 0;
1109	<	ForkJoinWorkerThread[] ws = workers;
1197	<	int n = ws.length;
1198	<	for (int i = 0; i < n; ++i) {
1199	<	ForkJoinWorkerThread w = ws[i];
1109	>	for (ForkJoinWorkerThread w : workers) {
1110		if (w != null) {
1111		w.shutdown();
1112		if (passes > 0 && !w.isTerminated()) {
#	Line 1260 | Line 1170 | public class ForkJoinPool extends Abstra
1170		*/
1171		final int idlePerActive() {
1172		int pc = parallelism; // use parallelism, not rc
1173	<	int ac = runState;    // no mask -- artifically boosts during shutdown
1173	>	int ac = runState;    // no mask -- artificially boosts during shutdown
1174		// Use exact results for small values, saturate past 4
1175		return pc <= ac? 0 : pc >>> 1 <= ac? 1 : pc >>> 2 <= ac? 3 : pc >>> 3;
1176		}
#	Line 1353 | Line 1263 | public class ForkJoinPool extends Abstra
1263		* @param pc the initial parallelism level
1264		*/
1265		private static int initialArraySizeFor(int pc) {
1266	<	// See Hackers Delight, sec 3.2. We know MAX_WORKERS < (1 >>> 16)
1266	>	// If possible, initially allocate enough space for one spare
1267		int size = pc < MAX_WORKERS ? pc + 1 : MAX_WORKERS;
1268	+	// See Hackers Delight, sec 3.2. We know MAX_WORKERS < (1 >>> 16)
1269		size |= size >>> 1;
1270		size |= size >>> 2;
1271		size |= size >>> 4;
#	Line 1373 | Line 1284 | public class ForkJoinPool extends Abstra
1284		if (runState >= SHUTDOWN)
1285		throw new RejectedExecutionException();
1286		submissionQueue.offer(task);
1287	<	advanceEventCount();
1288	<	if (eventWaiters != 0L)
1289	<	releaseEventWaiters();
1379	<	if ((workerCounts >>> TOTAL_COUNT_SHIFT) < parallelism)
1380	<	addWorkerIfBelowTarget();
1287	>	int c; // try to increment event count -- CAS failure OK
1288	>	UNSAFE.compareAndSwapInt(this, eventCountOffset, c = eventCount, c+1);
1289	>	helpMaintainParallelism(); // create, start, or resume some workers
1290		}
1291
1292		/**
#	Line 1614 | Line 1523 | public class ForkJoinPool extends Abstra
1523		*/
1524		public long getQueuedTaskCount() {
1525		long count = 0;
1526	<	ForkJoinWorkerThread[] ws = workers;
1618	<	int n = ws.length;
1619	<	for (int i = 0; i < n; ++i) {
1620	<	ForkJoinWorkerThread w = ws[i];
1526	>	for (ForkJoinWorkerThread w : workers)
1527		if (w != null)
1528		count += w.getQueueSize();
1623	–	}
1529		return count;
1530		}
1531
#	Line 1675 | Line 1580 | public class ForkJoinPool extends Abstra
1580		*/
1581		protected int drainTasksTo(Collection<? super ForkJoinTask<?>> c) {
1582		int count = submissionQueue.drainTo(c);
1583	<	ForkJoinWorkerThread[] ws = workers;
1679	<	int n = ws.length;
1680	<	for (int i = 0; i < n; ++i) {
1681	<	ForkJoinWorkerThread w = ws[i];
1583	>	for (ForkJoinWorkerThread w : workers)
1584		if (w != null)
1585		count += w.drainTasksTo(c);
1684	–	}
1586		return count;
1587		}
1588
#	Line 1808 | Line 1709 | public class ForkJoinPool extends Abstra
1709		throws InterruptedException {
1710		try {
1711		return termination.awaitAdvanceInterruptibly(0, timeout, unit) > 0;
1712	<	} catch(TimeoutException ex) {
1712	>	} catch (TimeoutException ex) {
1713		return false;
1714		}
1715		}

Diff Legend

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