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 1.23 by dl, Wed Aug 18 14:05:51 2010 UTC vs.
Revision 1.24 by dl, Sun Aug 29 23:35:07 2010 UTC

#	Line 159 | Line 159 | public class ForkJoinPool extends Abstra
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).
162	>	* It is impossible to keep exactly the target (parallelism)
163	>	* number of threads running at any given time.  Determining
164	>	* existence of conservatively safe helping targets, the
165	>	* availability of already-created spares, and the apparent need
166	>	* to create new spares are all racy and require heuristic
167	>	* guidance, so we rely on multiple retries of each.  Compensation
168	>	* occurs in slow-motion. It is triggered only upon timeouts of
169	>	* Object.wait used for joins. This reduces poor decisions that
170	>	* would otherwise be made when threads are waiting for others
171	>	* that are stalled because of unrelated activities such as
172	>	* garbage collection.
173		*
174		* The ManagedBlocker extension API can't use helping so relies
175		* only on compensation in method awaitBlocker.
#	Line 269 | Line 271 | public class ForkJoinPool extends Abstra
271		* In addition to allowing simpler decisions about need for
272		* wakeup, the event count bits in eventWaiters serve the role of
273		* tags to avoid ABA errors in Treiber stacks. Upon any wakeup,
274	<	* released threads also try to release others (but give up upon
275	<	* contention to reduce useless flailing).  The net effect is a
276	<	* tree-like diffusion of signals, where released threads (and
277	<	* possibly others) help with unparks.  To further reduce
278	<	* contention effects a bit, failed CASes to increment field
277	<	* eventCount are tolerated without retries in signalWork.
274	>	* released threads also try to release at most two others.  The
275	>	* net effect is a tree-like diffusion of signals, where released
276	>	* threads (and possibly others) help with unparks.  To further
277	>	* reduce contention effects a bit, failed CASes to increment
278	>	* field eventCount are tolerated without retries in signalWork.
279		* Conceptually they are merged into the same event, which is OK
280		* when their only purpose is to enable workers to scan for work.
281		*
282	<	* 5. Managing suspension of extra workers. When a worker is about
283	<	* to block waiting for a join (or via ManagedBlockers), we may
284	<	* create a new thread to maintain parallelism level, or at least
285	<	* avoid starvation. Usually, extra threads are needed for only
286	<	* very short periods, yet join dependencies are such that we
287	<	* sometimes need them in bursts. Rather than create new threads
288	<	* each time this happens, we suspend no-longer-needed extra ones
289	<	* as "spares". For most purposes, we don't distinguish "extra"
290	<	* spare threads from normal "core" threads: On each call to
291	<	* preStep (the only point at which we can do this) a worker
292	<	* checks to see if there are now too many running workers, and if
293	<	* so, suspends itself.  Method helpMaintainParallelism looks for
294	<	* suspended threads to resume before considering creating a new
295	<	* replacement. The spares themselves are encoded on another
296	<	* variant of a Treiber Stack, headed at field "spareWaiters".
297	<	* Note that the use of spares is intrinsically racy.  One thread
298	<	* may become a spare at about the same time as another is
299	<	* needlessly being created. We counteract this and related slop
300	<	* in part by requiring resumed spares to immediately recheck (in
301	<	* preStep) to see whether they they should re-suspend.
302	<	*
303	<	* 6. Killing off unneeded workers. The Spare and Event queues use
304	<	* similar mechanisms to shed unused workers: The oldest (first)
305	<	* waiter uses a timed rather than hard wait. When this wait times
306	<	* out without a normal wakeup, it tries to shutdown any one (for
307	<	* convenience the newest) other waiter via tryShutdownSpare or
308	<	* tryShutdownWaiter, respectively. The wakeup rates for spares
309	<	* are much shorter than for waiters. Together, they will
310	<	* eventually reduce the number of worker threads to a minimum of
311	<	* one after a long enough period without use.
282	>	* 5. Managing suspension of extra workers. When a worker notices
283	>	* (usually upon timeout of a wait()) that there are too few
284	>	* running threads, we may create a new thread to maintain
285	>	* parallelism level, or at least avoid starvation. Usually, extra
286	>	* threads are needed for only very short periods, yet join
287	>	* dependencies are such that we sometimes need them in
288	>	* bursts. Rather than create new threads each time this happens,
289	>	* we suspend no-longer-needed extra ones as "spares". For most
290	>	* purposes, we don't distinguish "extra" spare threads from
291	>	* normal "core" threads: On each call to preStep (the only point
292	>	* at which we can do this) a worker checks to see if there are
293	>	* now too many running workers, and if so, suspends itself.
294	>	* Method helpMaintainParallelism looks for suspended threads to
295	>	* resume before considering creating a new replacement. The
296	>	* spares themselves are encoded on another variant of a Treiber
297	>	* Stack, headed at field "spareWaiters".  Note that the use of
298	>	* spares is intrinsically racy.  One thread may become a spare at
299	>	* about the same time as another is needlessly being created. We
300	>	* counteract this and related slop in part by requiring resumed
301	>	* spares to immediately recheck (in preStep) to see whether they
302	>	* they should re-suspend.
303	>	*
304	>	* 6. Killing off unneeded workers. A timeout mechanism is used to
305	>	* shed unused workers: The oldest (first) event queue waiter uses
306	>	* a timed rather than hard wait. When this wait times out without
307	>	* a normal wakeup, it tries to shutdown any one (for convenience
308	>	* the newest) other spare or event waiter via
309	>	* tryShutdownUnusedWorker. This eventually reduces the number of
310	>	* worker threads to a minimum of one after a long enough period
311	>	* without use.
312		*
313		* 7. Deciding when to create new workers. The main dynamic
314		* control in this class is deciding when to create extra threads
#	Line 320 | Line 321 | public class ForkJoinPool extends Abstra
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. (We also include
328	<	* some heuristic use of Thread.yield when all workers appear to
329	<	* be busy, to improve likelihood of counts settling.) The rules
330	<	* also better cope with the fact that some of the methods in this
331	<	* class tend to never become compiled (but are interpreted), so
332	<	* some components of the entire set of controls might execute 100
332	<	* times faster than others. And similarly for cases where the
333	<	* apparent lack of work is just due to GC stalls and other
334	<	* transient system activity.
324	>	* impedes accuracy. Our main defense is to allow parallelism to
325	>	* lapse for a while during joins, and use a timeout to see if,
326	>	* after the resulting settling, there is still a need for
327	>	* additional workers.  This also better copes with the fact that
328	>	* some of the methods in this class tend to never become compiled
329	>	* (but are interpreted), so some components of the entire set of
330	>	* controls might execute 100 times faster than others. And
331	>	* similarly for cases where the apparent lack of work is just due
332	>	* to GC stalls and other transient system activity.
333		*
334		* Beware that there is a lot of representation-level coupling
335		* among classes ForkJoinPool, ForkJoinWorkerThread, and
#	Line 418 | Line 416 | public class ForkJoinPool extends Abstra
416		new AtomicInteger();
417
418		/**
419	+	* The time to block in a join (see awaitJoin) before checking if
420	+	* a new worker should be (re)started to maintain parallelism
421	+	* level. The value should be short enough to maintain gloabal
422	+	* responsiveness and progress but long enough to avoid
423	+	* counterproductive firings during GC stalls or unrelated system
424	+	* activity, and to not bog down systems with continual re-firings
425	+	* on GCs or legitimately long waits.
426	+	*/
427	+	private static final long JOIN_TIMEOUT_MILLIS = 250L; // 4 per second
428	+
429	+	/**
430		* The wakeup interval (in nanoseconds) for the oldest worker
431	<	* worker waiting for an event invokes tryShutdownWaiter to shrink
431	>	* worker waiting for an event invokes tryShutdownUnusedWorker to shrink
432		* the number of workers.  The exact value does not matter too
433		* much, but should be long enough to slowly release resources
434		* during long periods without use without disrupting normal use.
435		*/
436		private static final long SHRINK_RATE_NANOS =
437	<	60L * 1000L * 1000L * 1000L; // one minute
437	>	30L * 1000L * 1000L * 1000L; // 2 per minute
438
439		/**
440		* Absolute bound for parallelism level. Twice this number plus
#	Line 565 | Line 574 | public class ForkJoinPool extends Abstra
574		*/
575		private final int poolNumber;
576
568	–
577		// Utilities for CASing fields. Note that most of these
578		// are usually manually inlined by callers
579
#	Line 613 | Line 621 | public class ForkJoinPool extends Abstra
621		}
622
623		/**
616	–	* Increments event count
617	–	*/
618	–	private void advanceEventCount() {
619	–	int c;
620	–	do {} while(!UNSAFE.compareAndSwapInt(this, eventCountOffset,
621	–	c = eventCount, c+1));
622	–	}
623	–
624	–	/**
625	–	* Tries incrementing active count; fails on contention.
626	–	* Called by workers before executing tasks.
627	–	*
628	–	* @return true on success
629	–	*/
630	–	final boolean tryIncrementActiveCount() {
631	–	int c;
632	–	return UNSAFE.compareAndSwapInt(this, runStateOffset,
633	–	c = runState, c + 1);
634	–	}
635	–
636	–	/**
624		* Tries decrementing active count; fails on contention.
625		* Called when workers cannot find tasks to run.
626		*/
#	Line 701 | Line 688 | public class ForkJoinPool extends Abstra
688		}
689		}
690
704	–	// adding and removing workers
705	–
706	–	/**
707	–	* Tries to create and add new worker. Assumes that worker counts
708	–	* are already updated to accommodate the worker, so adjusts on
709	–	* failure.
710	–	*
711	–	* @return the worker, or null on failure
712	–	*/
713	–	private ForkJoinWorkerThread addWorker() {
714	–	ForkJoinWorkerThread w = null;
715	–	try {
716	–	w = factory.newThread(this);
717	–	} finally { // Adjust on either null or exceptional factory return
718	–	if (w == null) {
719	–	decrementWorkerCounts(ONE_RUNNING, ONE_TOTAL);
720	–	tryTerminate(false); // in case of failure during shutdown
721	–	}
722	–	}
723	–	if (w != null) {
724	–	w.start(recordWorker(w), ueh);
725	–	advanceEventCount();
726	–	}
727	–	return w;
728	–	}
729	–
691		/**
692		* Final callback from terminating worker.  Removes record of
693		* worker from array, and adjusts counts. If pool is shutting
#	Line 748 | Line 709 | public class ForkJoinPool extends Abstra
709		* Releases workers blocked on a count not equal to current count.
710		* Normally called after precheck that eventWaiters isn't zero to
711		* avoid wasted array checks. Gives up upon a change in count or
712	<	* contention, letting other workers take over.
712	>	* upon releasing two workers, letting others take over.
713		*/
714		private void releaseEventWaiters() {
715		ForkJoinWorkerThread[] ws = workers;
716		int n = ws.length;
717		long h = eventWaiters;
718		int ec = eventCount;
719	+	boolean releasedOne = false;
720		ForkJoinWorkerThread w; int id;
721	<	while ((int)(h >>> EVENT_COUNT_SHIFT) != ec &&
722	<	(id = ((int)(h & WAITER_ID_MASK)) - 1) >= 0 &&
723	<	id < n && (w = ws[id]) != null &&
724	<	UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
725	<	h,  h = w.nextWaiter)) {
726	<	LockSupport.unpark(w);
727	<	if (eventWaiters != h || eventCount != ec)
721	>	while ((id = ((int)(h & WAITER_ID_MASK)) - 1) >= 0 &&
722	>	(int)(h >>> EVENT_COUNT_SHIFT) != ec &&
723	>	id < n && (w = ws[id]) != null) {
724	>	if (UNSAFE.compareAndSwapLong(this, eventWaitersOffset,
725	>	h,  w.nextWaiter)) {
726	>	LockSupport.unpark(w);
727	>	if (releasedOne) // exit on second release
728	>	break;
729	>	releasedOne = true;
730	>	}
731	>	if (eventCount != ec)
732		break;
733	+	h = eventWaiters;
734		}
735		}
736
#	Line 780 | Line 747 | public class ForkJoinPool extends Abstra
747
748		/**
749		* Adds the given worker to event queue and blocks until
750	<	* terminating or event count advances from the workers
784	<	* lastEventCount value
750	>	* terminating or event count advances from the given value
751		*
752		* @param w the calling worker thread
753	+	* @param ec the count
754		*/
755	<	private void eventSync(ForkJoinWorkerThread w) {
789	<	int ec = w.lastEventCount;
755	>	private void eventSync(ForkJoinWorkerThread w, int ec) {
756		long nh = (((long)ec) << EVENT_COUNT_SHIFT) | ((long)(w.poolIndex+1));
757		long h;
758		while ((runState < SHUTDOWN || !tryTerminate(false)) &&
#	Line 806 | Line 772 | public class ForkJoinPool extends Abstra
772		* event waiter) until terminating or event count advances from
773		* the given value. The oldest (first) waiter uses a timed wait to
774		* occasionally one-by-one shrink the number of workers (to a
775	<	* minumum of one) if the pool has not been used for extended
775	>	* minimum of one) if the pool has not been used for extended
776		* periods.
777		*
778		* @param w the calling worker thread
#	Line 819 | Line 785 | public class ForkJoinPool extends Abstra
785		(workerCounts & RUNNING_COUNT_MASK) <= 1);
786		long startTime = untimed? 0 : System.nanoTime();
787		Thread.interrupted();         // clear/ignore interrupt
788	<	if (eventCount != ec || !w.isRunning() ||
788	>	if (eventCount != ec || w.runState != 0 ||
789		runState >= TERMINATING)  // recheck after clear
790		break;
791		if (untimed)
792		LockSupport.park(w);
793		else {
794		LockSupport.parkNanos(w, SHRINK_RATE_NANOS);
795	<	if (eventCount != ec || !w.isRunning() ||
795	>	if (eventCount != ec || w.runState != 0 ||
796		runState >= TERMINATING)
797		break;
798		if (System.nanoTime() - startTime >= SHRINK_RATE_NANOS)
799	<	tryShutdownWaiter(ec);
799	>	tryShutdownUnusedWorker(ec);
800		}
801		}
802		}
803		}
804
805	<	/**
840	<	* Callback from the oldest waiter in awaitEvent waking up after a
841	<	* period of non-use. Tries (once) to shutdown an event waiter (or
842	<	* a spare, if one exists). Note that we don't need CAS or locks
843	<	* here because the method is called only from one thread
844	<	* occasionally waking (and even misfires are OK). Note that
845	<	* until the shutdown worker fully terminates, workerCounts
846	<	* will overestimate total count, which is tolerable.
847	<	*
848	<	* @param ec the event count waited on by caller (to abort
849	<	* attempt if count has since changed).
850	<	*/
851	<	private void tryShutdownWaiter(int ec) {
852	<	if (spareWaiters != 0) { // prefer killing spares
853	<	tryShutdownSpare();
854	<	return;
855	<	}
856	<	ForkJoinWorkerThread[] ws = workers;
857	<	int n = ws.length;
858	<	long h = eventWaiters;
859	<	ForkJoinWorkerThread w; int id; long nh;
860	<	if (runState == 0 &&
861	<	submissionQueue.isEmpty() &&
862	<	eventCount == ec &&
863	<	(id = ((int)(h & WAITER_ID_MASK)) - 1) >= 0 &&
864	<	id < n && (w = ws[id]) != null &&
865	<	(nh = w.nextWaiter) != 0L && // keep at least one worker
866	<	UNSAFE.compareAndSwapLong(this, eventWaitersOffset, h, nh)) {
867	<	w.shutdown();
868	<	LockSupport.unpark(w);
869	<	}
870	<	releaseEventWaiters();
871	<	}
872	<
873	<	// Maintaining spares
805	>	// Maintaining parallelism
806
807		/**
808		* Pushes worker onto the spare stack
#	Line 882 | Line 814 | public class ForkJoinPool extends Abstra
814		}
815
816		/**
817	<	* Callback from oldest spare occasionally waking up.  Tries
818	<	* (once) to shutdown a spare. Same idea as tryShutdownWaiter.
817	>	* Tries (once) to resume a spare if the number of running
818	>	* threads is less than target.
819		*/
820	<	final void tryShutdownSpare() {
820	>	private void tryResumeSpare() {
821		int sw, id;
890	–	ForkJoinWorkerThread w;
891	–	ForkJoinWorkerThread[] ws;
892	–	if ((id = ((sw = spareWaiters) & SPARE_ID_MASK) - 1) >= 0 &&
893	–	id < (ws = workers).length && (w = ws[id]) != null &&
894	–	(workerCounts & RUNNING_COUNT_MASK) >= parallelism &&
895	–	UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
896	–	sw, w.nextSpare)) {
897	–	w.shutdown();
898	–	LockSupport.unpark(w);
899	–	advanceEventCount();
900	–	}
901	–	}
902	–
903	–	/**
904	–	* Tries (once) to resume a spare if worker counts match
905	–	* the given count.
906	–	*
907	–	* @param wc workerCounts value on invocation of this method
908	–	*/
909	–	private void tryResumeSpare(int wc) {
822		ForkJoinWorkerThread[] ws = workers;
823		int n = ws.length;
824	<	int sw, id, rs;  ForkJoinWorkerThread w;
825	<	if ((id = ((sw = spareWaiters) & SPARE_ID_MASK) - 1) >= 0 &&
824	>	ForkJoinWorkerThread w;
825	>	if ((sw = spareWaiters) != 0 &&
826	>	(id = (sw & SPARE_ID_MASK) - 1) >= 0 &&
827		id < n && (w = ws[id]) != null &&
828	<	(rs = runState) < TERMINATING &&
829	<	eventWaiters == 0L && workerCounts == wc) {
830	<	// In case all workers busy, heuristically back off to let settle
831	<	Thread.yield();
832	<	if (eventWaiters == 0L && runState == rs && // recheck
833	<	workerCounts == wc && spareWaiters == sw &&
834	<	UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
835	<	sw, w.nextSpare)) {
836	<	int c;              // increment running count before resume
837	<	do {} while(!UNSAFE.compareAndSwapInt
838	<	(this, workerCountsOffset,
839	<	c = workerCounts, c + ONE_RUNNING));
927	<	if (w.tryUnsuspend())
928	<	LockSupport.unpark(w);
929	<	else               // back out if w was shutdown
930	<	decrementWorkerCounts(ONE_RUNNING, 0);
931	<	}
828	>	(workerCounts & RUNNING_COUNT_MASK) < parallelism &&
829	>	spareWaiters == sw &&
830	>	UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
831	>	sw, w.nextSpare)) {
832	>	int c; // increment running count before resume
833	>	do {} while(!UNSAFE.compareAndSwapInt
834	>	(this, workerCountsOffset,
835	>	c = workerCounts, c + ONE_RUNNING));
836	>	if (w.tryUnsuspend())
837	>	LockSupport.unpark(w);
838	>	else   // back out if w was shutdown
839	>	decrementWorkerCounts(ONE_RUNNING, 0);
840		}
841		}
842
935	–	// adding workers on demand
936	–
843		/**
844	<	* Adds one or more workers if needed to establish target parallelism.
845	<	* Retries upon contention.
844	>	* Tries to increase the number of running workers if below target
845	>	* parallelism: If a spare exists tries to resume it via
846	>	* tryResumeSpare.  Otherwise, if not enough total workers or all
847	>	* existing workers are busy, adds a new worker. In all casses also
848	>	* helps wake up releasable workers waiting for work.
849		*/
850	<	private void addWorkerIfBelowTarget() {
850	>	private void helpMaintainParallelism() {
851		int pc = parallelism;
852	<	int wc;
853	<	while (((wc = workerCounts) >>> TOTAL_COUNT_SHIFT) < pc &&
854	<	runState < TERMINATING) {
855	<	if (UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
856	<	wc + (ONE_RUNNING|ONE_TOTAL))) {
857	<	if (addWorker() == null)
852	>	int wc, rs, tc;
853	>	while (((wc = workerCounts) & RUNNING_COUNT_MASK) < pc &&
854	>	(rs = runState) < TERMINATING) {
855	>	if (spareWaiters != 0)
856	>	tryResumeSpare();
857	>	else if ((tc = wc >>> TOTAL_COUNT_SHIFT) >= MAX_WORKERS ||
858	>	(tc >= pc && (rs & ACTIVE_COUNT_MASK) != tc))
859	>	break;   // enough total
860	>	else if (runState == rs && workerCounts == wc &&
861	>	UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
862	>	wc + (ONE_RUNNING|ONE_TOTAL))) {
863	>	ForkJoinWorkerThread w = null;
864	>	try {
865	>	w = factory.newThread(this);
866	>	} finally { // adjust on null or exceptional factory return
867	>	if (w == null) {
868	>	decrementWorkerCounts(ONE_RUNNING, ONE_TOTAL);
869	>	tryTerminate(false); // handle failure during shutdown
870	>	}
871	>	}
872	>	if (w == null)
873		break;
874	+	w.start(recordWorker(w), ueh);
875	+	if ((workerCounts >>> TOTAL_COUNT_SHIFT) >= pc) {
876	+	int c; // advance event count
877	+	UNSAFE.compareAndSwapInt(this, eventCountOffset,
878	+	c = eventCount, c+1);
879	+	break; // add at most one unless total below target
880	+	}
881		}
882		}
883	+	if (eventWaiters != 0L)
884	+	releaseEventWaiters();
885		}
886
887		/**
888	<	* Tries (once) to add a new worker if all existing workers are
889	<	* busy, and there are either no running workers or the deficit is
890	<	* at least twice the surplus.
891	<	*
892	<	* @param wc workerCounts value on invocation of this method
893	<	*/
894	<	private void tryAddWorkerIfBusy(int wc) {
962	<	int tc, rc, rs;
963	<	int pc = parallelism;
964	<	if ((tc = wc >>> TOTAL_COUNT_SHIFT) < MAX_WORKERS &&
965	<	((rc = wc & RUNNING_COUNT_MASK) == 0 ||
966	<	rc < pc - ((tc - pc) << 1)) &&
967	<	(rs = runState) < TERMINATING &&
968	<	(rs & ACTIVE_COUNT_MASK) == tc) {
969	<	// Since all workers busy, heuristically back off to let settle
970	<	Thread.yield();
971	<	if (eventWaiters == 0L && spareWaiters == 0 && // recheck
972	<	runState == rs && workerCounts == wc &&
973	<	UNSAFE.compareAndSwapInt(this, workerCountsOffset, wc,
974	<	wc + (ONE_RUNNING|ONE_TOTAL)))
975	<	addWorker();
976	<	}
977	<	}
978	<
979	<	/**
980	<	* Does at most one of:
981	<	*
982	<	* 1. Help wake up existing workers waiting for work via
983	<	*    releaseEventWaiters. (If any exist, then it doesn't
984	<	*    matter right now if under target parallelism level.)
985	<	*
986	<	* 2. If a spare exists, try (once) to resume it via tryResumeSpare.
987	<	*
988	<	* 3. If there are not enough total workers, add some
989	<	*    via addWorkerIfBelowTarget;
888	>	* Callback from the oldest waiter in awaitEvent waking up after a
889	>	* period of non-use. If all workers are idle, tries (once) to
890	>	* shutdown an event waiter or a spare, if one exists. Note that
891	>	* we don't need CAS or locks here because the method is called
892	>	* only from one thread occasionally waking (and even misfires are
893	>	* OK). Note that until the shutdown worker fully terminates,
894	>	* workerCounts will overestimate total count, which is tolerable.
895		*
896	<	* 4. Try (once) to add a new worker if all existing workers
897	<	*     are busy, via tryAddWorkerIfBusy
896	>	* @param ec the event count waited on by caller (to abort
897	>	* attempt if count has since changed).
898		*/
899	<	private void helpMaintainParallelism() {
900	<	long h; int pc, wc;
901	<	if (((int)((h = eventWaiters) & WAITER_ID_MASK)) != 0) {
902	<	if ((int)(h >>> EVENT_COUNT_SHIFT) != eventCount)
903	<	releaseEventWaiters(); // avoid useless call
904	<	}
905	<	else if ((pc = parallelism) >
906	<	((wc = workerCounts) & RUNNING_COUNT_MASK)) {
907	<	if (spareWaiters != 0)
908	<	tryResumeSpare(wc);
909	<	else if ((wc >>> TOTAL_COUNT_SHIFT) < pc)
910	<	addWorkerIfBelowTarget();
911	<	else
912	<	tryAddWorkerIfBusy(wc);
899	>	private void tryShutdownUnusedWorker(int ec) {
900	>	if (runState == 0 && eventCount == ec) { // only trigger if all idle
901	>	ForkJoinWorkerThread[] ws = workers;
902	>	int n = ws.length;
903	>	ForkJoinWorkerThread w = null;
904	>	boolean shutdown = false;
905	>	int sw;
906	>	long h;
907	>	if ((sw = spareWaiters) != 0) { // prefer killing spares
908	>	int id = (sw & SPARE_ID_MASK) - 1;
909	>	if (id >= 0 && id < n && (w = ws[id]) != null &&
910	>	UNSAFE.compareAndSwapInt(this, spareWaitersOffset,
911	>	sw, w.nextSpare))
912	>	shutdown = true;
913	>	}
914	>	else if ((h = eventWaiters) != 0L) {
915	>	long nh;
916	>	int id = ((int)(h & WAITER_ID_MASK)) - 1;
917	>	if (id >= 0 && id < n && (w = ws[id]) != null &&
918	>	(nh = w.nextWaiter) != 0L && // keep at least one worker
919	>	UNSAFE.compareAndSwapLong(this, eventWaitersOffset, h, nh))
920	>	shutdown = true;
921	>	}
922	>	if (w != null && shutdown) {
923	>	w.shutdown();
924	>	LockSupport.unpark(w);
925	>	}
926		}
927	+	releaseEventWaiters(); // in case of interference
928		}
929
930		/**
#	Line 1013 | Line 932 | public class ForkJoinPool extends Abstra
932		* stealing a task or taking a submission and running it).
933		* Performs one or more of the following:
934		*
935	<	* 1. If the worker is active, try to set its active status to
936	<	*    inactive and update activeCount. On contention, we may try
937	<	*    again on this or subsequent call.
938	<	*
939	<	* 2. Release any existing event waiters that are now relesable
940	<	*
941	<	* 3. If there are too many running threads, suspend this worker
942	<	*    (first forcing inactive if necessary).  If it is not
943	<	*    needed, it may be killed while suspended via
944	<	*    tryShutdownSpare. Otherwise, upon resume it rechecks to make
945	<	*    sure that it is still needed.
946	<	*
947	<	* 4. If more than 1 miss, await the next task event via
948	<	*    eventSync (first forcing inactivation if necessary), upon
949	<	*    which worker may also be killed, via tryShutdownWaiter.
950	<	*
951	<	* 5. Help reactivate other workers via helpMaintainParallelism
935	>	* 1. If the worker is active and either did not run a task
936	>	*    or there are too many workers, try to set its active status
937	>	*    to inactive and update activeCount. On contention, we may
938	>	*    try again in this or a subsequent call.
939	>	*
940	>	* 2. If not enough total workers, help create some.
941	>	*
942	>	* 3. If there are too many running workers, suspend this worker
943	>	*    (first forcing inactive if necessary).  If it is not needed,
944	>	*    it may be shutdown while suspended (via
945	>	*    tryShutdownUnusedWorker).  Otherwise, upon resume it
946	>	*    rechecks running thread count and need for event sync.
947	>	*
948	>	* 4. If worker did not run a task, await the next task event via
949	>	*    eventSync if necessary (first forcing inactivation), upon
950	>	*    which the worker may be shutdown via
951	>	*    tryShutdownUnusedWorker.  Otherwise, help release any
952	>	*    existing event waiters that are now releasable,
953		*
954		* @param w the worker
955	<	* @param misses the number of scans by caller failing to find work
1036	<	* (saturating at 2 to avoid wraparound)
955	>	* @param ran true if worker ran a task since last call to this method
956		*/
957	<	final void preStep(ForkJoinWorkerThread w, int misses) {
957	>	final void preStep(ForkJoinWorkerThread w, boolean ran) {
958	>	int wec = w.lastEventCount;
959		boolean active = w.active;
960	+	boolean inactivate = false;
961		int pc = parallelism;
962	<	for (;;) {
963	<	int rs, wc, rc, ec; long h;
964	<	if (active && UNSAFE.compareAndSwapInt(this, runStateOffset,
965	<	rs = runState, rs - 1))
966	<	active = w.active = false;
967	<	if (((int)((h = eventWaiters) & WAITER_ID_MASK)) != 0 &&
968	<	(int)(h >>> EVENT_COUNT_SHIFT) != eventCount) {
969	<	releaseEventWaiters();
1049	<	if (misses > 1)
1050	<	continue;                  // clear before sync below
1051	<	}
1052	<	if ((rc = ((wc = workerCounts) & RUNNING_COUNT_MASK)) > pc) {
1053	<	if (!active &&                 // must inactivate to suspend
962	>	int rs;
963	>	while (w.runState == 0 && (rs = runState) < TERMINATING) {
964	>	if ((inactivate || (active && (rs & ACTIVE_COUNT_MASK) >= pc)) &&
965	>	UNSAFE.compareAndSwapInt(this, runStateOffset, rs, rs - 1))
966	>	inactivate = active = w.active = false;
967	>	int wc = workerCounts;
968	>	if ((wc & RUNNING_COUNT_MASK) > pc) {
969	>	if (!(inactivate |= active) && // must inactivate to suspend
970		workerCounts == wc &&      // try to suspend as spare
971		UNSAFE.compareAndSwapInt(this, workerCountsOffset,
972	<	wc, wc - ONE_RUNNING)) {
972	>	wc, wc - ONE_RUNNING))
973		w.suspendAsSpare();
1058	–	if (!w.isRunning())
1059	–	break;                 // was killed while spare
1060	–	}
1061	–	continue;
974		}
975	<	if (misses > 0) {
976	<	if ((ec = eventCount) == w.lastEventCount && misses > 1) {
977	<	if (!active) {             // must inactivate to sync
978	<	eventSync(w);
979	<	if (w.isRunning())
980	<	misses = 1;        // don't re-sync
981	<	else
982	<	break;             // was killed while waiting
983	<	}
984	<	continue;
975	>	else if ((wc >>> TOTAL_COUNT_SHIFT) < pc)
976	>	helpMaintainParallelism();     // not enough workers
977	>	else if (!ran) {
978	>	long h = eventWaiters;
979	>	int ec = eventCount;
980	>	if (h != 0L && (int)(h >>> EVENT_COUNT_SHIFT) != ec)
981	>	releaseEventWaiters();     // release others before waiting
982	>	else if (ec != wec) {
983	>	w.lastEventCount = ec;     // no need to wait
984	>	break;
985		}
986	<	w.lastEventCount = ec;
986	>	else if (!(inactivate |= active))
987	>	eventSync(w, wec);         // must inactivate before sync
988		}
989	<	if (rc < pc)
990	<	helpMaintainParallelism();
1078	<	break;
989	>	else
990	>	break;
991		}
992		}
993
994		/**
995		* Helps and/or blocks awaiting join of the given task.
996	<	* Alternates between helpJoinTask() and helpMaintainParallelism()
1085	<	* as many times as there is a deficit in running count (or longer
1086	<	* if running count would become zero), then blocks if task still
1087	<	* not done.
996	>	* See above for explanation.
997		*
998		* @param joinMe the task to join
999	+	* @param worker the current worker thread
1000		*/
1001		final void awaitJoin(ForkJoinTask<?> joinMe, ForkJoinWorkerThread worker) {
1002	<	int threshold = parallelism;         // descend blocking thresholds
1002	>	int retries = 2 + (parallelism >> 2); // #helpJoins before blocking
1003		while (joinMe.status >= 0) {
1004	<	boolean block; int wc;
1004	>	int wc;
1005		worker.helpJoinTask(joinMe);
1006		if (joinMe.status < 0)
1007		break;
1008	<	if (((wc = workerCounts) & RUNNING_COUNT_MASK) <= threshold) {
1009	<	if (threshold > 0)
1010	<	--threshold;
1011	<	else
1012	<	advanceEventCount(); // force release
1013	<	block = false;
1014	<	}
1015	<	else
1016	<	block = UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1017	<	wc, wc - ONE_RUNNING);
1018	<	helpMaintainParallelism();
1019	<	if (block) {
1020	<	int c;
1021	<	joinMe.internalAwaitDone();
1008	>	else if (retries > 0)
1009	>	--retries;
1010	>	else if (((wc = workerCounts) & RUNNING_COUNT_MASK) != 0 &&
1011	>	UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1012	>	wc, wc - ONE_RUNNING)) {
1013	>	int stat, c; long h;
1014	>	while ((stat = joinMe.status) >= 0 &&
1015	>	(h = eventWaiters) != 0L && // help release others
1016	>	(int)(h >>> EVENT_COUNT_SHIFT) != eventCount)
1017	>	releaseEventWaiters();
1018	>	if (stat >= 0 &&
1019	>	((workerCounts & RUNNING_COUNT_MASK) == 0 ||
1020	>	(stat =
1021	>	joinMe.internalAwaitDone(JOIN_TIMEOUT_MILLIS)) >= 0))
1022	>	helpMaintainParallelism(); // timeout or no running workers
1023		do {} while (!UNSAFE.compareAndSwapInt
1024		(this, workerCountsOffset,
1025		c = workerCounts, c + ONE_RUNNING));
1026	<	break;
1026	>	if (stat < 0)
1027	>	break;   // else restart
1028		}
1029		}
1030		}
1031
1032		/**
1033	<	* Same idea as awaitJoin, but no helping
1033	>	* Same idea as awaitJoin, but no helping, retries, or timeouts.
1034		*/
1035		final void awaitBlocker(ManagedBlocker blocker)
1036		throws InterruptedException {
1125	–	int threshold = parallelism;
1037		while (!blocker.isReleasable()) {
1038	<	boolean block; int wc;
1039	<	if (((wc = workerCounts) & RUNNING_COUNT_MASK) <= threshold) {
1040	<	if (threshold > 0)
1041	<	--threshold;
1131	<	else
1132	<	advanceEventCount();
1133	<	block = false;
1134	<	}
1135	<	else
1136	<	block = UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1137	<	wc, wc - ONE_RUNNING);
1138	<	helpMaintainParallelism();
1139	<	if (block) {
1038	>	int wc = workerCounts;
1039	>	if ((wc & RUNNING_COUNT_MASK) != 0 &&
1040	>	UNSAFE.compareAndSwapInt(this, workerCountsOffset,
1041	>	wc, wc - ONE_RUNNING)) {
1042		try {
1043	<	do {} while (!blocker.isReleasable() && !blocker.block());
1043	>	while (!blocker.isReleasable()) {
1044	>	long h = eventWaiters;
1045	>	if (h != 0L &&
1046	>	(int)(h >>> EVENT_COUNT_SHIFT) != eventCount)
1047	>	releaseEventWaiters();
1048	>	else if ((workerCounts & RUNNING_COUNT_MASK) == 0 &&
1049	>	runState < TERMINATING)
1050	>	helpMaintainParallelism();
1051	>	else if (blocker.block())
1052	>	break;
1053	>	}
1054		} finally {
1055		int c;
1056		do {} while (!UNSAFE.compareAndSwapInt
#	Line 1188 | Line 1100 | public class ForkJoinPool extends Abstra
1100		private void startTerminating() {
1101		cancelSubmissions();
1102		for (int passes = 0; passes < 4 && workerCounts != 0; ++passes) {
1103	<	advanceEventCount();
1103	>	int c; // advance event count
1104	>	UNSAFE.compareAndSwapInt(this, eventCountOffset,
1105	>	c = eventCount, c+1);
1106		eventWaiters = 0L; // clobber lists
1107		spareWaiters = 0;
1108		ForkJoinWorkerThread[] ws = workers;
#	Line 1351 | Line 1265 | public class ForkJoinPool extends Abstra
1265		* @param pc the initial parallelism level
1266		*/
1267		private static int initialArraySizeFor(int pc) {
1268	<	// See Hackers Delight, sec 3.2. We know MAX_WORKERS < (1 >>> 16)
1268	>	// If possible, initially allocate enough space for one spare
1269		int size = pc < MAX_WORKERS ? pc + 1 : MAX_WORKERS;
1270	+	// See Hackers Delight, sec 3.2. We know MAX_WORKERS < (1 >>> 16)
1271		size |= size >>> 1;
1272		size |= size >>> 2;
1273		size |= size >>> 4;
#	Line 1371 | Line 1286 | public class ForkJoinPool extends Abstra
1286		if (runState >= SHUTDOWN)
1287		throw new RejectedExecutionException();
1288		submissionQueue.offer(task);
1289	<	advanceEventCount();
1290	<	if (eventWaiters != 0L)
1291	<	releaseEventWaiters();
1377	<	if ((workerCounts >>> TOTAL_COUNT_SHIFT) < parallelism)
1378	<	addWorkerIfBelowTarget();
1289	>	int c; // try to increment event count -- CAS failure OK
1290	>	UNSAFE.compareAndSwapInt(this, eventCountOffset, c = eventCount, c+1);
1291	>	helpMaintainParallelism(); // create, start, or resume some workers
1292		}
1293
1294		/**

Diff Legend

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