ViewVC Help

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

root/jsr166/jsr166/src/jsr166y/ForkJoinWorkerThread.java

Comparing jsr166/src/jsr166y/ForkJoinWorkerThread.java (file contents):
Revision 1.34 by dl, Fri Jun 4 14:37:54 2010 UTC vs.
Revision 1.69 by dl, Mon Feb 20 18:20:06 2012 UTC

#	Line 1 | Line 1
1		/*
2		* Written by Doug Lea with assistance from members of JCP JSR-166
3		* Expert Group and released to the public domain, as explained at
4	<	* http://creativecommons.org/licenses/publicdomain
4	>	* http://creativecommons.org/publicdomain/zero/1.0/
5		*/
6
7		package jsr166y;
8
9	–	import java.util.concurrent.*;
10	–
11	–	import java.util.Random;
12	–	import java.util.Collection;
13	–	import java.util.concurrent.locks.LockSupport;
14	–
9		/**
10	<	* A thread managed by a {@link ForkJoinPool}.  This class is
11	<	* subclassable solely for the sake of adding functionality -- there
12	<	* are no overridable methods dealing with scheduling or execution.
13	<	* However, you can override initialization and termination methods
14	<	* surrounding the main task processing loop.  If you do create such a
15	<	* subclass, you will also need to supply a custom {@link
16	<	* ForkJoinPool.ForkJoinWorkerThreadFactory} to use it in a {@code
17	<	* ForkJoinPool}.
10	>	* A thread managed by a {@link ForkJoinPool}, which executes
11	>	* {@link ForkJoinTask}s.
12	>	* This class is subclassable solely for the sake of adding
13	>	* functionality -- there are no overridable methods dealing with
14	>	* scheduling or execution.  However, you can override initialization
15	>	* and termination methods surrounding the main task processing loop.
16	>	* If you do create such a subclass, you will also need to supply a
17	>	* custom {@link ForkJoinPool.ForkJoinWorkerThreadFactory} to use it
18	>	* in a {@code ForkJoinPool}.
19		*
20		* @since 1.7
21		* @author Doug Lea
22		*/
23		public class ForkJoinWorkerThread extends Thread {
24		/*
30	–	* Overview:
31	–	*
25		* ForkJoinWorkerThreads are managed by ForkJoinPools and perform
26	<	* ForkJoinTasks. This class includes bookkeeping in support of
27	<	* worker activation, suspension, and lifecycle control described
35	<	* in more detail in the internal documentation of class
36	<	* ForkJoinPool. And as described further below, this class also
37	<	* includes special-cased support for some ForkJoinTask
38	<	* methods. But the main mechanics involve work-stealing:
39	<	*
40	<	* Work-stealing queues are special forms of Deques that support
41	<	* only three of the four possible end-operations -- push, pop,
42	<	* and deq (aka steal), under the further constraints that push
43	<	* and pop are called only from the owning thread, while deq may
44	<	* be called from other threads.  (If you are unfamiliar with
45	<	* them, you probably want to read Herlihy and Shavit's book "The
46	<	* Art of Multiprocessor programming", chapter 16 describing these
47	<	* in more detail before proceeding.)  The main work-stealing
48	<	* queue design is roughly similar to those in the papers "Dynamic
49	<	* Circular Work-Stealing Deque" by Chase and Lev, SPAA 2005
50	<	* (http://research.sun.com/scalable/pubs/index.html) and
51	<	* "Idempotent work stealing" by Michael, Saraswat, and Vechev,
52	<	* PPoPP 2009 (http://portal.acm.org/citation.cfm?id=1504186).
53	<	* The main differences ultimately stem from gc requirements that
54	<	* we null out taken slots as soon as we can, to maintain as small
55	<	* a footprint as possible even in programs generating huge
56	<	* numbers of tasks. To accomplish this, we shift the CAS
57	<	* arbitrating pop vs deq (steal) from being on the indices
58	<	* ("base" and "sp") to the slots themselves (mainly via method
59	<	* "casSlotNull()"). So, both a successful pop and deq mainly
60	<	* entail a CAS of a slot from non-null to null.  Because we rely
61	<	* on CASes of references, we do not need tag bits on base or sp.
62	<	* They are simple ints as used in any circular array-based queue
63	<	* (see for example ArrayDeque).  Updates to the indices must
64	<	* still be ordered in a way that guarantees that sp == base means
65	<	* the queue is empty, but otherwise may err on the side of
66	<	* possibly making the queue appear nonempty when a push, pop, or
67	<	* deq have not fully committed. Note that this means that the deq
68	<	* operation, considered individually, is not wait-free. One thief
69	<	* cannot successfully continue until another in-progress one (or,
70	<	* if previously empty, a push) completes.  However, in the
71	<	* aggregate, we ensure at least probabilistic non-blockingness.
72	<	* If an attempted steal fails, a thief always chooses a different
73	<	* random victim target to try next. So, in order for one thief to
74	<	* progress, it suffices for any in-progress deq or new push on
75	<	* any empty queue to complete. One reason this works well here is
76	<	* that apparently-nonempty often means soon-to-be-stealable,
77	<	* which gives threads a chance to set activation status if
78	<	* necessary before stealing.
79	<	*
80	<	* This approach also enables support for "async mode" where local
81	<	* task processing is in FIFO, not LIFO order; simply by using a
82	<	* version of deq rather than pop when locallyFifo is true (as set
83	<	* by the ForkJoinPool).  This allows use in message-passing
84	<	* frameworks in which tasks are never joined.
85	<	*
86	<	* Efficient implementation of this approach currently relies on
87	<	* an uncomfortable amount of "Unsafe" mechanics. To maintain
88	<	* correct orderings, reads and writes of variable base require
89	<	* volatile ordering.  Variable sp does not require volatile
90	<	* writes but still needs store-ordering, which we accomplish by
91	<	* pre-incrementing sp before filling the slot with an ordered
92	<	* store.  (Pre-incrementing also enables backouts used in
93	<	* scanWhileJoining.)  Because they are protected by volatile base
94	<	* reads, reads of the queue array and its slots by other threads
95	<	* do not need volatile load semantics, but writes (in push)
96	<	* require store order and CASes (in pop and deq) require
97	<	* (volatile) CAS semantics.  (Michael, Saraswat, and Vechev's
98	<	* algorithm has similar properties, but without support for
99	<	* nulling slots.)  Since these combinations aren't supported
100	<	* using ordinary volatiles, the only way to accomplish these
101	<	* efficiently is to use direct Unsafe calls. (Using external
102	<	* AtomicIntegers and AtomicReferenceArrays for the indices and
103	<	* array is significantly slower because of memory locality and
104	<	* indirection effects.)
105	<	*
106	<	* Further, performance on most platforms is very sensitive to
107	<	* placement and sizing of the (resizable) queue array.  Even
108	<	* though these queues don't usually become all that big, the
109	<	* initial size must be large enough to counteract cache
110	<	* contention effects across multiple queues (especially in the
111	<	* presence of GC cardmarking). Also, to improve thread-locality,
112	<	* queues are initialized after starting.  All together, these
113	<	* low-level implementation choices produce as much as a factor of
114	<	* 4 performance improvement compared to naive implementations,
115	<	* and enable the processing of billions of tasks per second,
116	<	* sometimes at the expense of ugliness.
117	<	*/
118	<
119	<	/**
120	<	* Generator for initial random seeds for random victim
121	<	* selection. This is used only to create initial seeds. Random
122	<	* steals use a cheaper xorshift generator per steal attempt. We
123	<	* expect only rare contention on seedGenerator, so just use a
124	<	* plain Random.
26	>	* ForkJoinTasks. For explanation, see the internal documentation
27	>	* of class ForkJoinPool.
28		*/
126	–	private static final Random seedGenerator = new Random();
29
30	<	/**
31	<	* The timeout value for suspending spares. Spare workers that
130	<	* remain unsignalled for more than this time may be trimmed
131	<	* (killed and removed from pool).  Since our goal is to avoid
132	<	* long-term thread buildup, the exact value of timeout does not
133	<	* matter too much so long as it avoids most false-alarm timeouts
134	<	* under GC stalls or momentarily high system load.
135	<	*/
136	<	private static final long SPARE_KEEPALIVE_NANOS =
137	<	5L * 1000L * 1000L * 1000L; // 5 secs
138	<
139	<	/**
140	<	* Capacity of work-stealing queue array upon initialization.
141	<	* Must be a power of two. Initial size must be at least 4, but is
142	<	* padded to minimize cache effects.
143	<	*/
144	<	private static final int INITIAL_QUEUE_CAPACITY = 1 << 13;
145	<
146	<	/**
147	<	* Maximum work-stealing queue array size.  Must be less than or
148	<	* equal to 1 << 28 to ensure lack of index wraparound. (This
149	<	* is less than usual bounds, because we need leftshift by 3
150	<	* to be in int range).
151	<	*/
152	<	private static final int MAXIMUM_QUEUE_CAPACITY = 1 << 28;
153	<
154	<	/**
155	<	* The pool this thread works in. Accessed directly by ForkJoinTask.
156	<	*/
157	<	final ForkJoinPool pool;
158	<
159	<	/**
160	<	* The work-stealing queue array. Size must be a power of two.
161	<	* Initialized in onStart, to improve memory locality.
162	<	*/
163	<	private ForkJoinTask<?>[] queue;
164	<
165	<	/**
166	<	* Index (mod queue.length) of least valid queue slot, which is
167	<	* always the next position to steal from if nonempty.
168	<	*/
169	<	private volatile int base;
170	<
171	<	/**
172	<	* Index (mod queue.length) of next queue slot to push to or pop
173	<	* from. It is written only by owner thread, and accessed by other
174	<	* threads only after reading (volatile) base.  Both sp and base
175	<	* are allowed to wrap around on overflow, but (sp - base) still
176	<	* estimates size.
177	<	*/
178	<	private int sp;
179	<
180	<	/**
181	<	* Run state of this worker. In addition to the usual run levels,
182	<	* tracks if this worker is suspended as a spare, and if it was
183	<	* killed (trimmed) while suspended. However, "active" status is
184	<	* maintained separately.
185	<	*/
186	<	private volatile int runState;
187	<
188	<	private static final int TERMINATING = 0x01;
189	<	private static final int TERMINATED  = 0x02;
190	<	private static final int SUSPENDED   = 0x04; // inactive spare
191	<	private static final int TRIMMED     = 0x08; // killed while suspended
192	<
193	<	/**
194	<	* Number of LockSupport.park calls to block this thread for
195	<	* suspension or event waits. Used for internal instrumention;
196	<	* currently not exported but included because volatile write upon
197	<	* park also provides a workaround for a JVM bug.
198	<	*/
199	<	private volatile int parkCount;
200	<
201	<	/**
202	<	* Number of steals, transferred and reset in pool callbacks pool
203	<	* when idle Accessed directly by pool.
204	<	*/
205	<	int stealCount;
206	<
207	<	/**
208	<	* Seed for random number generator for choosing steal victims.
209	<	* Uses Marsaglia xorshift. Must be initialized as nonzero.
210	<	*/
211	<	private int seed;
212	<
213	<	/**
214	<	* Activity status. When true, this worker is considered active.
215	<	* Accessed directly by pool.  Must be false upon construction.
216	<	*/
217	<	boolean active;
218	<
219	<	/**
220	<	* True if use local fifo, not default lifo, for local polling.
221	<	* Shadows value from ForkJoinPool, which resets it if changed
222	<	* pool-wide.
223	<	*/
224	<	private boolean locallyFifo;
225	<
226	<	/**
227	<	* Index of this worker in pool array. Set once by pool before
228	<	* running, and accessed directly by pool to locate this worker in
229	<	* its workers array.
230	<	*/
231	<	int poolIndex;
232	<
233	<	/**
234	<	* The last pool event waited for. Accessed only by pool in
235	<	* callback methods invoked within this thread.
236	<	*/
237	<	int lastEventCount;
238	<
239	<	/**
240	<	* Encoded index and event count of next event waiter. Used only
241	<	* by ForkJoinPool for managing event waiters.
242	<	*/
243	<	volatile long nextWaiter;
30	>	final ForkJoinPool.WorkQueue workQueue; // Work-stealing mechanics
31	>	final ForkJoinPool pool;                // the pool this thread works in
32
33		/**
34		* Creates a ForkJoinWorkerThread operating in the given pool.
#	Line 249 | Line 37 | public class ForkJoinWorkerThread extend
37		* @throws NullPointerException if pool is null
38		*/
39		protected ForkJoinWorkerThread(ForkJoinPool pool) {
40	<	if (pool == null) throw new NullPointerException();
41	<	this.pool = pool;
42	<	// To avoid exposing construction details to subclasses,
255	<	// remaining initialization is in start() and onStart()
256	<	}
257	<
258	<	/**
259	<	* Performs additional initialization and starts this thread
260	<	*/
261	<	final void start(int poolIndex, boolean locallyFifo,
262	<	UncaughtExceptionHandler ueh) {
263	<	this.poolIndex = poolIndex;
264	<	this.locallyFifo = locallyFifo;
40	>	super(pool.nextWorkerName());
41	>	setDaemon(true);
42	>	Thread.UncaughtExceptionHandler ueh = pool.ueh;
43		if (ueh != null)
44		setUncaughtExceptionHandler(ueh);
45	<	setDaemon(true);
46	<	start();
45	>	this.pool = pool;
46	>	pool.registerWorker(this.workQueue = new ForkJoinPool.WorkQueue
47	>	(pool, this, pool.localMode));
48		}
49
271	–	// Public/protected methods
272	–
50		/**
51		* Returns the pool hosting this thread.
52		*
#	Line 289 | Line 66 | public class ForkJoinWorkerThread extend
66		* @return the index number
67		*/
68		public int getPoolIndex() {
69	<	return poolIndex;
69	>	return workQueue.poolIndex;
70		}
71
72		/**
73		* Initializes internal state after construction but before
74		* processing any tasks. If you override this method, you must
75	<	* invoke super.onStart() at the beginning of the method.
75	>	* invoke {@code super.onStart()} at the beginning of the method.
76		* Initialization requires care: Most fields must have legal
77		* default values, to ensure that attempted accesses from other
78		* threads work correctly even before this thread starts
79		* processing tasks.
80		*/
81		protected void onStart() {
305	–	int rs = seedGenerator.nextInt();
306	–	seed = rs == 0? 1 : rs; // seed must be nonzero
307	–
308	–	// Allocate name string and queue array in this thread
309	–	String pid = Integer.toString(pool.getPoolNumber());
310	–	String wid = Integer.toString(poolIndex);
311	–	setName("ForkJoinPool-" + pid + "-worker-" + wid);
312	–
313	–	queue = new ForkJoinTask<?>[INITIAL_QUEUE_CAPACITY];
82		}
83
84		/**
#	Line 322 | Line 90 | public class ForkJoinWorkerThread extend
90		* to an unrecoverable error, or {@code null} if completed normally
91		*/
92		protected void onTermination(Throwable exception) {
325	–	try {
326	–	cancelTasks();
327	–	setTerminated();
328	–	pool.workerTerminated(this);
329	–	} catch (Throwable ex) {        // Shouldn't ever happen
330	–	if (exception == null)      // but if so, at least rethrown
331	–	exception = ex;
332	–	} finally {
333	–	if (exception != null)
334	–	UNSAFE.throwException(exception);
335	–	}
93		}
94
95		/**
96		* This method is required to be public, but should never be
97		* called explicitly. It performs the main run loop to execute
98	<	* ForkJoinTasks.
98	>	* {@link ForkJoinTask}s.
99		*/
100		public void run() {
101		Throwable exception = null;
102		try {
103		onStart();
104	<	mainLoop();
104	>	pool.runWorker(workQueue);
105		} catch (Throwable ex) {
106		exception = ex;
107		} finally {
351	–	onTermination(exception);
352	–	}
353	–	}
354	–
355	–	// helpers for run()
356	–
357	–	/**
358	–	* Find and execute tasks and check status while running
359	–	*/
360	–	private void mainLoop() {
361	–	boolean ran = false;      // true if ran task in last loop iter
362	–	boolean prevRan = false;  // true if ran on last or previous step
363	–	ForkJoinPool p = pool;
364	–	for (;;) {
365	–	p.preStep(this, prevRan);
366	–	if (runState != 0)
367	–	return;
368	–	ForkJoinTask<?> t; // try to get and run stolen or submitted task
369	–	if ((t = scan()) != null || (t = pollSubmission()) != null) {
370	–	t.tryExec();
371	–	if (base != sp)
372	–	runLocalTasks();
373	–	prevRan = ran = true;
374	–	}
375	–	else {
376	–	prevRan = ran;
377	–	ran = false;
378	–	}
379	–	}
380	–	}
381	–
382	–	/**
383	–	* Runs local tasks until queue is empty or shut down.  Call only
384	–	* while active.
385	–	*/
386	–	private void runLocalTasks() {
387	–	while (runState == 0) {
388	–	ForkJoinTask<?> t = locallyFifo? locallyDeqTask() : popTask();
389	–	if (t != null)
390	–	t.tryExec();
391	–	else if (base == sp)
392	–	break;
393	–	}
394	–	}
395	–
396	–	/**
397	–	* If a submission exists, try to activate and take it
398	–	*
399	–	* @return a task, if available
400	–	*/
401	–	private ForkJoinTask<?> pollSubmission() {
402	–	ForkJoinPool p = pool;
403	–	while (p.hasQueuedSubmissions()) {
404	–	if (active || (active = p.tryIncrementActiveCount())) {
405	–	ForkJoinTask<?> t = p.pollSubmission();
406	–	return t != null ? t : scan(); // if missed, rescan
407	–	}
408	–	}
409	–	return null;
410	–	}
411	–
412	–	/*
413	–	* Intrinsics-based atomic writes for queue slots. These are
414	–	* basically the same as methods in AtomicObjectArray, but
415	–	* specialized for (1) ForkJoinTask elements (2) requirement that
416	–	* nullness and bounds checks have already been performed by
417	–	* callers and (3) effective offsets are known not to overflow
418	–	* from int to long (because of MAXIMUM_QUEUE_CAPACITY). We don't
419	–	* need corresponding version for reads: plain array reads are OK
420	–	* because they protected by other volatile reads and are
421	–	* confirmed by CASes.
422	–	*
423	–	* Most uses don't actually call these methods, but instead contain
424	–	* inlined forms that enable more predictable optimization.  We
425	–	* don't define the version of write used in pushTask at all, but
426	–	* instead inline there a store-fenced array slot write.
427	–	*/
428	–
429	–	/**
430	–	* CASes slot i of array q from t to null. Caller must ensure q is
431	–	* non-null and index is in range.
432	–	*/
433	–	private static final boolean casSlotNull(ForkJoinTask<?>[] q, int i,
434	–	ForkJoinTask<?> t) {
435	–	return UNSAFE.compareAndSwapObject(q, (i << qShift) + qBase, t, null);
436	–	}
437	–
438	–	/**
439	–	* Performs a volatile write of the given task at given slot of
440	–	* array q.  Caller must ensure q is non-null and index is in
441	–	* range. This method is used only during resets and backouts.
442	–	*/
443	–	private static final void writeSlot(ForkJoinTask<?>[] q, int i,
444	–	ForkJoinTask<?> t) {
445	–	UNSAFE.putObjectVolatile(q, (i << qShift) + qBase, t);
446	–	}
447	–
448	–	// queue methods
449	–
450	–	/**
451	–	* Pushes a task. Call only from this thread.
452	–	*
453	–	* @param t the task. Caller must ensure non-null.
454	–	*/
455	–	final void pushTask(ForkJoinTask<?> t) {
456	–	ForkJoinTask<?>[] q = queue;
457	–	int mask = q.length - 1; // implicit assert q != null
458	–	int s = sp++;            // ok to increment sp before slot write
459	–	UNSAFE.putOrderedObject(q, ((s & mask) << qShift) + qBase, t);
460	–	if ((s -= base) == 0)
461	–	pool.signalWork();   // was empty
462	–	else if (s == mask)
463	–	growQueue();         // is full
464	–	}
465	–
466	–	/**
467	–	* Tries to take a task from the base of the queue, failing if
468	–	* empty or contended. Note: Specializations of this code appear
469	–	* in scan and scanWhileJoining.
470	–	*
471	–	* @return a task, or null if none or contended
472	–	*/
473	–	final ForkJoinTask<?> deqTask() {
474	–	ForkJoinTask<?> t;
475	–	ForkJoinTask<?>[] q;
476	–	int b, i;
477	–	if ((b = base) != sp &&
478	–	(q = queue) != null && // must read q after b
479	–	(t = q[i = (q.length - 1) & b]) != null &&
480	–	UNSAFE.compareAndSwapObject(q, (i << qShift) + qBase, t, null)) {
481	–	base = b + 1;
482	–	return t;
483	–	}
484	–	return null;
485	–	}
486	–
487	–	/**
488	–	* Tries to take a task from the base of own queue. Assumes active
489	–	* status.  Called only by current thread.
490	–	*
491	–	* @return a task, or null if none
492	–	*/
493	–	final ForkJoinTask<?> locallyDeqTask() {
494	–	ForkJoinTask<?>[] q = queue;
495	–	if (q != null) {
496	–	ForkJoinTask<?> t;
497	–	int b, i;
498	–	while (sp != (b = base)) {
499	–	if ((t = q[i = (q.length - 1) & b]) != null &&
500	–	UNSAFE.compareAndSwapObject(q, (i << qShift) + qBase,
501	–	t, null)) {
502	–	base = b + 1;
503	–	return t;
504	–	}
505	–	}
506	–	}
507	–	return null;
508	–	}
509	–
510	–	/**
511	–	* Returns a popped task, or null if empty. Assumes active status.
512	–	* Called only by current thread. (Note: a specialization of this
513	–	* code appears in popWhileJoining.)
514	–	*/
515	–	final ForkJoinTask<?> popTask() {
516	–	int s;
517	–	ForkJoinTask<?>[] q;
518	–	if (base != (s = sp) && (q = queue) != null) {
519	–	int i = (q.length - 1) & --s;
520	–	ForkJoinTask<?> t = q[i];
521	–	if (t != null && UNSAFE.compareAndSwapObject
522	–	(q, (i << qShift) + qBase, t, null)) {
523	–	sp = s;
524	–	return t;
525	–	}
526	–	}
527	–	return null;
528	–	}
529	–
530	–	/**
531	–	* Specialized version of popTask to pop only if topmost element
532	–	* is the given task. Called only by current thread while
533	–	* active.
534	–	*
535	–	* @param t the task. Caller must ensure non-null.
536	–	*/
537	–	final boolean unpushTask(ForkJoinTask<?> t) {
538	–	int s;
539	–	ForkJoinTask<?>[] q;
540	–	if (base != (s = sp) && (q = queue) != null &&
541	–	UNSAFE.compareAndSwapObject
542	–	(q, (((q.length - 1) & --s) << qShift) + qBase, t, null)) {
543	–	sp = s;
544	–	return true;
545	–	}
546	–	return false;
547	–	}
548	–
549	–	/**
550	–	* Returns next task or null if empty or contended
551	–	*/
552	–	final ForkJoinTask<?> peekTask() {
553	–	ForkJoinTask<?>[] q = queue;
554	–	if (q == null)
555	–	return null;
556	–	int mask = q.length - 1;
557	–	int i = locallyFifo ? base : (sp - 1);
558	–	return q[i & mask];
559	–	}
560	–
561	–	/**
562	–	* Doubles queue array size. Transfers elements by emulating
563	–	* steals (deqs) from old array and placing, oldest first, into
564	–	* new array.
565	–	*/
566	–	private void growQueue() {
567	–	ForkJoinTask<?>[] oldQ = queue;
568	–	int oldSize = oldQ.length;
569	–	int newSize = oldSize << 1;
570	–	if (newSize > MAXIMUM_QUEUE_CAPACITY)
571	–	throw new RejectedExecutionException("Queue capacity exceeded");
572	–	ForkJoinTask<?>[] newQ = queue = new ForkJoinTask<?>[newSize];
573	–
574	–	int b = base;
575	–	int bf = b + oldSize;
576	–	int oldMask = oldSize - 1;
577	–	int newMask = newSize - 1;
578	–	do {
579	–	int oldIndex = b & oldMask;
580	–	ForkJoinTask<?> t = oldQ[oldIndex];
581	–	if (t != null && !casSlotNull(oldQ, oldIndex, t))
582	–	t = null;
583	–	writeSlot(newQ, b & newMask, t);
584	–	} while (++b != bf);
585	–	pool.signalWork();
586	–	}
587	–
588	–	/**
589	–	* Computes next value for random victim probe in scan().  Scans
590	–	* don't require a very high quality generator, but also not a
591	–	* crummy one.  Marsaglia xor-shift is cheap and works well enough.
592	–	* Note: This is manually inlined in scan()
593	–	*/
594	–	private static final int xorShift(int r) {
595	–	r ^= r << 13;
596	–	r ^= r >>> 17;
597	–	return r ^ (r << 5);
598	–	}
599	–
600	–	/**
601	–	* Tries to steal a task from another worker. Starts at a random
602	–	* index of workers array, and probes workers until finding one
603	–	* with non-empty queue or finding that all are empty.  It
604	–	* randomly selects the first n probes. If these are empty, it
605	–	* resorts to a circular sweep, which is necessary to accurately
606	–	* set active status. (The circular sweep uses steps of
607	–	* approximately half the array size plus 1, to avoid bias
608	–	* stemming from leftmost packing of the array in ForkJoinPool.)
609	–	*
610	–	* This method must be both fast and quiet -- usually avoiding
611	–	* memory accesses that could disrupt cache sharing etc other than
612	–	* those needed to check for and take tasks (or to activate if not
613	–	* already active). This accounts for, among other things,
614	–	* updating random seed in place without storing it until exit.
615	–	*
616	–	* @return a task, or null if none found
617	–	*/
618	–	private ForkJoinTask<?> scan() {
619	–	ForkJoinPool p = pool;
620	–	ForkJoinWorkerThread[] ws;        // worker array
621	–	int n;                            // upper bound of #workers
622	–	if ((ws = p.workers) != null && (n = ws.length) > 1) {
623	–	boolean canSteal = active;    // shadow active status
624	–	int r = seed;                 // extract seed once
625	–	int mask = n - 1;
626	–	int j = -n;                   // loop counter
627	–	int k = r;                    // worker index, random if j < 0
628	–	for (;;) {
629	–	ForkJoinWorkerThread v = ws[k & mask];
630	–	r ^= r << 13; r ^= r >>> 17; r ^= r << 5; // inline xorshift
631	–	if (v != null && v.base != v.sp) {
632	–	int b, i;             // inline specialized deqTask
633	–	ForkJoinTask<?>[] q;
634	–	ForkJoinTask<?> t;
635	–	if ((canSteal ||      // ensure active status
636	–	(canSteal = active = p.tryIncrementActiveCount())) &&
637	–	(q = v.queue) != null &&
638	–	(t = q[i = (q.length - 1) & (b = v.base)]) != null &&
639	–	UNSAFE.compareAndSwapObject
640	–	(q, (i << qShift) + qBase, t, null)) {
641	–	v.base = b + 1;
642	–	seed = r;
643	–	++stealCount;
644	–	return t;
645	–	}
646	–	j = -n;
647	–	k = r;                // restart on contention
648	–	}
649	–	else if (++j <= 0)
650	–	k = r;
651	–	else if (j <= n)
652	–	k += (n >>> 1) | 1;
653	–	else
654	–	break;
655	–	}
656	–	}
657	–	return null;
658	–	}
659	–
660	–	// Run State management
661	–
662	–	// status check methods used mainly by ForkJoinPool
663	–	final boolean isTerminating() { return (runState & TERMINATING) != 0; }
664	–	final boolean isTerminated()  { return (runState & TERMINATED) != 0; }
665	–	final boolean isSuspended()   { return (runState & SUSPENDED) != 0; }
666	–	final boolean isTrimmed()     { return (runState & TRIMMED) != 0; }
667	–
668	–	/**
669	–	* Sets state to TERMINATING, also resuming if suspended.
670	–	*/
671	–	final void shutdown() {
672	–	for (;;) {
673	–	int s = runState;
674	–	if ((s & SUSPENDED) != 0) { // kill and wakeup if suspended
675	–	if (UNSAFE.compareAndSwapInt(this, runStateOffset, s,
676	–	(s & ~SUSPENDED) |
677	–	(TRIMMED|TERMINATING))) {
678	–	LockSupport.unpark(this);
679	–	break;
680	–	}
681	–	}
682	–	else if (UNSAFE.compareAndSwapInt(this, runStateOffset, s,
683	–	s | TERMINATING))
684	–	break;
685	–	}
686	–	}
687	–
688	–	/**
689	–	* Sets state to TERMINATED. Called only by this thread.
690	–	*/
691	–	private void setTerminated() {
692	–	int s;
693	–	do {} while (!UNSAFE.compareAndSwapInt(this, runStateOffset,
694	–	s = runState,
695	–	s | (TERMINATING|TERMINATED)));
696	–	}
697	–
698	–	/**
699	–	* Instrumented version of park. Also used by ForkJoinPool.awaitEvent
700	–	*/
701	–	final void doPark() {
702	–	++parkCount;
703	–	LockSupport.park(this);
704	–	}
705	–
706	–	/**
707	–	* If suspended, tries to set status to unsuspended.
708	–	* Caller must unpark to actually resume
709	–	*
710	–	* @return true if successful
711	–	*/
712	–	final boolean tryUnsuspend() {
713	–	int s;
714	–	return (((s = runState) & SUSPENDED) != 0 &&
715	–	UNSAFE.compareAndSwapInt(this, runStateOffset, s,
716	–	s & ~SUSPENDED));
717	–	}
718	–
719	–	/**
720	–	* Sets suspended status and blocks as spare until resumed,
721	–	* shutdown, or timed out.
722	–	*
723	–	* @return false if trimmed
724	–	*/
725	–	final boolean suspendAsSpare() {
726	–	for (;;) {               // set suspended unless terminating
727	–	int s = runState;
728	–	if ((s & TERMINATING) != 0) { // must kill
729	–	if (UNSAFE.compareAndSwapInt(this, runStateOffset, s,
730	–	s | (TRIMMED | TERMINATING)))
731	–	return false;
732	–	}
733	–	else if (UNSAFE.compareAndSwapInt(this, runStateOffset, s,
734	–	s | SUSPENDED))
735	–	break;
736	–	}
737	–	lastEventCount = 0;      // reset upon resume
738	–	ForkJoinPool p = pool;
739	–	p.releaseWaiters();      // help others progress
740	–	p.accumulateStealCount(this);
741	–	interrupted();           // clear/ignore interrupts
742	–	if (poolIndex < p.getParallelism()) { // untimed wait
743	–	while ((runState & SUSPENDED) != 0)
744	–	doPark();
745	–	return true;
746	–	}
747	–	return timedSuspend();   // timed wait if apparently non-core
748	–	}
749	–
750	–	/**
751	–	* Blocks as spare until resumed or timed out
752	–	* @return false if trimmed
753	–	*/
754	–	private boolean timedSuspend() {
755	–	long nanos = SPARE_KEEPALIVE_NANOS;
756	–	long startTime = System.nanoTime();
757	–	while ((runState & SUSPENDED) != 0) {
758	–	++parkCount;
759	–	if ((nanos -= (System.nanoTime() - startTime)) > 0)
760	–	LockSupport.parkNanos(this, nanos);
761	–	else { // try to trim on timeout
762	–	int s = runState;
763	–	if (UNSAFE.compareAndSwapInt(this, runStateOffset, s,
764	–	(s & ~SUSPENDED) |
765	–	(TRIMMED|TERMINATING)))
766	–	return false;
767	–	}
768	–	}
769	–	return true;
770	–	}
771	–
772	–	// Misc support methods for ForkJoinPool
773	–
774	–	/**
775	–	* Returns an estimate of the number of tasks in the queue.  Also
776	–	* used by ForkJoinTask.
777	–	*/
778	–	final int getQueueSize() {
779	–	return -base + sp;
780	–	}
781	–
782	–	/**
783	–	* Set locallyFifo mode. Called only by ForkJoinPool
784	–	*/
785	–	final void setAsyncMode(boolean async) {
786	–	locallyFifo = async;
787	–	}
788	–
789	–	/**
790	–	* Removes and cancels all tasks in queue.  Can be called from any
791	–	* thread.
792	–	*/
793	–	final void cancelTasks() {
794	–	while (base != sp) {
795	–	ForkJoinTask<?> t = deqTask();
796	–	if (t != null)
797	–	t.cancelIgnoringExceptions();
798	–	}
799	–	}
800	–
801	–	/**
802	–	* Drains tasks to given collection c.
803	–	*
804	–	* @return the number of tasks drained
805	–	*/
806	–	final int drainTasksTo(Collection<? super ForkJoinTask<?>> c) {
807	–	int n = 0;
808	–	while (base != sp) {
809	–	ForkJoinTask<?> t = deqTask();
810	–	if (t != null) {
811	–	c.add(t);
812	–	++n;
813	–	}
814	–	}
815	–	return n;
816	–	}
817	–
818	–	// Support methods for ForkJoinTask
819	–
820	–	/**
821	–	* Returns an estimate of the number of tasks, offset by a
822	–	* function of number of idle workers.
823	–	*
824	–	* This method provides a cheap heuristic guide for task
825	–	* partitioning when programmers, frameworks, tools, or languages
826	–	* have little or no idea about task granularity.  In essence by
827	–	* offering this method, we ask users only about tradeoffs in
828	–	* overhead vs expected throughput and its variance, rather than
829	–	* how finely to partition tasks.
830	–	*
831	–	* In a steady state strict (tree-structured) computation, each
832	–	* thread makes available for stealing enough tasks for other
833	–	* threads to remain active. Inductively, if all threads play by
834	–	* the same rules, each thread should make available only a
835	–	* constant number of tasks.
836	–	*
837	–	* The minimum useful constant is just 1. But using a value of 1
838	–	* would require immediate replenishment upon each steal to
839	–	* maintain enough tasks, which is infeasible.  Further,
840	–	* partitionings/granularities of offered tasks should minimize
841	–	* steal rates, which in general means that threads nearer the top
842	–	* of computation tree should generate more than those nearer the
843	–	* bottom. In perfect steady state, each thread is at
844	–	* approximately the same level of computation tree. However,
845	–	* producing extra tasks amortizes the uncertainty of progress and
846	–	* diffusion assumptions.
847	–	*
848	–	* So, users will want to use values larger, but not much larger
849	–	* than 1 to both smooth over transient shortages and hedge
850	–	* against uneven progress; as traded off against the cost of
851	–	* extra task overhead. We leave the user to pick a threshold
852	–	* value to compare with the results of this call to guide
853	–	* decisions, but recommend values such as 3.
854	–	*
855	–	* When all threads are active, it is on average OK to estimate
856	–	* surplus strictly locally. In steady-state, if one thread is
857	–	* maintaining say 2 surplus tasks, then so are others. So we can
858	–	* just use estimated queue length (although note that (sp - base)
859	–	* can be an overestimate because of stealers lagging increments
860	–	* of base).  However, this strategy alone leads to serious
861	–	* mis-estimates in some non-steady-state conditions (ramp-up,
862	–	* ramp-down, other stalls). We can detect many of these by
863	–	* further considering the number of "idle" threads, that are
864	–	* known to have zero queued tasks, so compensate by a factor of
865	–	* (#idle/#active) threads.
866	–	*/
867	–	final int getEstimatedSurplusTaskCount() {
868	–	return sp - base - pool.idlePerActive();
869	–	}
870	–
871	–	/**
872	–	* Gets and removes a local task.
873	–	*
874	–	* @return a task, if available
875	–	*/
876	–	final ForkJoinTask<?> pollLocalTask() {
877	–	while (base != sp) {
878	–	if (active || (active = pool.tryIncrementActiveCount()))
879	–	return locallyFifo? locallyDeqTask() : popTask();
880	–	}
881	–	return null;
882	–	}
883	–
884	–	/**
885	–	* Gets and removes a local or stolen task.
886	–	*
887	–	* @return a task, if available
888	–	*/
889	–	final ForkJoinTask<?> pollTask() {
890	–	ForkJoinTask<?> t;
891	–	return (t = pollLocalTask()) != null ? t : scan();
892	–	}
893	–
894	–	/**
895	–	* Executes or processes other tasks awaiting the given task
896	–	* @return task completion status
897	–	*/
898	–	final int execWhileJoining(ForkJoinTask<?> joinMe) {
899	–	int s;
900	–	while ((s = joinMe.status) >= 0) {
901	–	ForkJoinTask<?> t = base != sp?
902	–	popWhileJoining(joinMe) :
903	–	scanWhileJoining(joinMe);
904	–	if (t != null)
905	–	t.tryExec();
906	–	}
907	–	return s;
908	–	}
909	–
910	–	/**
911	–	* Returns or stolen task, if available, unless joinMe is done
912	–	*
913	–	* This method is intrinsically nonmodular. To maintain the
914	–	* property that tasks are never stolen if the awaited task is
915	–	* ready, we must interleave mechanics of scan with status
916	–	* checks. We rely here on the commit points of deq that allow us
917	–	* to cancel a steal even after CASing slot to null, but before
918	–	* adjusting base index: If, after the CAS, we see that joinMe is
919	–	* ready, we can back out by placing the task back into the slot,
920	–	* without adjusting index. The loop is otherwise a variant of the
921	–	* one in scan().
922	–	*
923	–	*/
924	–	private ForkJoinTask<?> scanWhileJoining(ForkJoinTask<?> joinMe) {
925	–	int r = seed;
926	–	ForkJoinPool p = pool;
927	–	ForkJoinWorkerThread[] ws;
928	–	int n;
929	–	outer:while ((ws = p.workers) != null && (n = ws.length) > 1) {
930	–	int mask = n - 1;
931	–	int k = r;
932	–	boolean contended = false; // to retry loop if deq contends
933	–	for (int j = -n; j <= n; ++j) {
934	–	if (joinMe.status < 0)
935	–	break outer;
936	–	int b;
937	–	ForkJoinTask<?>[] q;
938	–	ForkJoinWorkerThread v = ws[k & mask];
939	–	r ^= r << 13; r ^= r >>> 17; r ^= r << 5; // xorshift
940	–	if (v != null && (b=v.base) != v.sp && (q=v.queue) != null) {
941	–	int i = (q.length - 1) & b;
942	–	ForkJoinTask<?> t = q[i];
943	–	if (t != null && UNSAFE.compareAndSwapObject
944	–	(q, (i << qShift) + qBase, t, null)) {
945	–	if (joinMe.status >= 0) {
946	–	v.base = b + 1;
947	–	seed = r;
948	–	++stealCount;
949	–	return t;
950	–	}
951	–	UNSAFE.putObjectVolatile(q, (i<<qShift)+qBase, t);
952	–	break outer; // back out
953	–	}
954	–	contended = true;
955	–	}
956	–	k = j < 0 ? r : (k + ((n >>> 1) | 1));
957	–	}
958	–	if (!contended && p.tryAwaitBusyJoin(joinMe))
959	–	break;
960	–	}
961	–	return null;
962	–	}
963	–
964	–	/**
965	–	* Version of popTask with join checks surrounding extraction.
966	–	* Uses the same backout strategy as helpJoinTask. Note that
967	–	* we ignore locallyFifo flag for local tasks here since helping
968	–	* joins only make sense in LIFO mode.
969	–	*
970	–	* @return a popped task, if available, unless joinMe is done
971	–	*/
972	–	private ForkJoinTask<?> popWhileJoining(ForkJoinTask<?> joinMe) {
973	–	int s;
974	–	ForkJoinTask<?>[] q;
975	–	while ((s = sp) != base && (q = queue) != null && joinMe.status >= 0) {
976	–	int i = (q.length - 1) & --s;
977	–	ForkJoinTask<?> t = q[i];
978	–	if (t != null && UNSAFE.compareAndSwapObject
979	–	(q, (i << qShift) + qBase, t, null)) {
980	–	if (joinMe.status >= 0) {
981	–	sp = s;
982	–	return t;
983	–	}
984	–	UNSAFE.putObjectVolatile(q, (i << qShift) + qBase, t);
985	–	break;  // back out
986	–	}
987	–	}
988	–	return null;
989	–	}
990	–
991	–	/**
992	–	* Runs tasks until {@code pool.isQuiescent()}.
993	–	*/
994	–	final void helpQuiescePool() {
995	–	for (;;) {
996	–	ForkJoinTask<?> t = pollLocalTask();
997	–	if (t != null || (t = scan()) != null)
998	–	t.tryExec();
999	–	else {
1000	–	ForkJoinPool p = pool;
1001	–	if (active) {
1002	–	active = false; // inactivate
1003	–	do {} while (!p.tryDecrementActiveCount());
1004	–	}
1005	–	if (p.isQuiescent()) {
1006	–	active = true; // re-activate
1007	–	do {} while (!p.tryIncrementActiveCount());
1008	–	return;
1009	–	}
1010	–	}
1011	–	}
1012	–	}
1013	–
1014	–	// Unsafe mechanics
1015	–
1016	–	private static final sun.misc.Unsafe UNSAFE = getUnsafe();
1017	–	private static final long runStateOffset =
1018	–	objectFieldOffset("runState", ForkJoinWorkerThread.class);
1019	–	private static final long qBase =
1020	–	UNSAFE.arrayBaseOffset(ForkJoinTask[].class);
1021	–	private static final int qShift;
1022	–
1023	–	static {
1024	–	int s = UNSAFE.arrayIndexScale(ForkJoinTask[].class);
1025	–	if ((s & (s-1)) != 0)
1026	–	throw new Error("data type scale not a power of two");
1027	–	qShift = 31 - Integer.numberOfLeadingZeros(s);
1028	–	}
1029	–
1030	–	private static long objectFieldOffset(String field, Class<?> klazz) {
1031	–	try {
1032	–	return UNSAFE.objectFieldOffset(klazz.getDeclaredField(field));
1033	–	} catch (NoSuchFieldException e) {
1034	–	// Convert Exception to corresponding Error
1035	–	NoSuchFieldError error = new NoSuchFieldError(field);
1036	–	error.initCause(e);
1037	–	throw error;
1038	–	}
1039	–	}
1040	–
1041	–	/**
1042	–	* Returns a sun.misc.Unsafe.  Suitable for use in a 3rd party package.
1043	–	* Replace with a simple call to Unsafe.getUnsafe when integrating
1044	–	* into a jdk.
1045	–	*
1046	–	* @return a sun.misc.Unsafe
1047	–	*/
1048	–	private static sun.misc.Unsafe getUnsafe() {
1049	–	try {
1050	–	return sun.misc.Unsafe.getUnsafe();
1051	–	} catch (SecurityException se) {
108		try {
109	<	return java.security.AccessController.doPrivileged
110	<	(new java.security
111	<	.PrivilegedExceptionAction<sun.misc.Unsafe>() {
112	<	public sun.misc.Unsafe run() throws Exception {
113	<	java.lang.reflect.Field f = sun.misc
114	<	.Unsafe.class.getDeclaredField("theUnsafe");
1059	<	f.setAccessible(true);
1060	<	return (sun.misc.Unsafe) f.get(null);
1061	<	}});
1062	<	} catch (java.security.PrivilegedActionException e) {
1063	<	throw new RuntimeException("Could not initialize intrinsics",
1064	<	e.getCause());
109	>	onTermination(exception);
110	>	} catch (Throwable ex) {
111	>	if (exception == null)
112	>	exception = ex;
113	>	} finally {
114	>	pool.deregisterWorker(this, exception);
115		}
116		}
117		}
118		}
119	+

Diff Legend

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