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 1.62 by jsr166, Mon Nov 29 20:58:06 2010 UTC vs.
Revision 1.63 by dl, Tue Feb 22 00:39:31 2011 UTC

#	Line 6 | Line 6
6
7		package jsr166y;
8
9	–	import java.util.Random;
9		import java.util.Collection;
11	–	import java.util.concurrent.locks.LockSupport;
10		import java.util.concurrent.RejectedExecutionException;
11
12		/**
#	Line 55 | Line 53 | public class ForkJoinWorkerThread extend
53		* a footprint as possible even in programs generating huge
54		* numbers of tasks. To accomplish this, we shift the CAS
55		* arbitrating pop vs deq (steal) from being on the indices
56	<	* ("base" and "sp") to the slots themselves (mainly via method
57	<	* "casSlotNull()"). So, both a successful pop and deq mainly
58	<	* entail a CAS of a slot from non-null to null.  Because we rely
59	<	* on CASes of references, we do not need tag bits on base or sp.
60	<	* They are simple ints as used in any circular array-based queue
61	<	* (see for example ArrayDeque).  Updates to the indices must
62	<	* still be ordered in a way that guarantees that sp == base means
63	<	* the queue is empty, but otherwise may err on the side of
64	<	* possibly making the queue appear nonempty when a push, pop, or
65	<	* deq have not fully committed. Note that this means that the deq
66	<	* operation, considered individually, is not wait-free. One thief
67	<	* cannot successfully continue until another in-progress one (or,
68	<	* if previously empty, a push) completes.  However, in the
56	>	* ("queueBase" and "queueTop") to the slots themselves (mainly
57	>	* via method "casSlotNull()"). So, both a successful pop and deq
58	>	* mainly entail a CAS of a slot from non-null to null.  Because
59	>	* we rely on CASes of references, we do not need tag bits on
60	>	* queueBase or queueTop.  They are simple ints as used in any
61	>	* circular array-based queue (see for example ArrayDeque).
62	>	* Updates to the indices must still be ordered in a way that
63	>	* guarantees that queueTop == queueBase means the queue is empty,
64	>	* but otherwise may err on the side of possibly making the queue
65	>	* appear nonempty when a push, pop, or deq have not fully
66	>	* committed. Note that this means that the deq operation,
67	>	* considered individually, is not wait-free. One thief cannot
68	>	* successfully continue until another in-progress one (or, if
69	>	* previously empty, a push) completes.  However, in the
70		* aggregate, we ensure at least probabilistic non-blockingness.
71		* If an attempted steal fails, a thief always chooses a different
72		* random victim target to try next. So, in order for one thief to
73		* progress, it suffices for any in-progress deq or new push on
74	<	* 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.
74	>	* any empty queue to complete.
75		*
76		* This approach also enables support for "async mode" where local
77		* task processing is in FIFO, not LIFO order; simply by using a
78		* version of deq rather than pop when locallyFifo is true (as set
79		* by the ForkJoinPool).  This allows use in message-passing
80	<	* frameworks in which tasks are never joined.
80	>	* frameworks in which tasks are never joined.  However neither
81	>	* mode considers affinities, loads, cache localities, etc, so
82	>	* rarely provide the best possible performance on a given
83	>	* machine, but portably provide good throughput by averaging over
84	>	* these factors.  (Further, even if we did try to use such
85	>	* information, we do not usually have a basis for exploiting
86	>	* it. For example, some sets of tasks profit from cache
87	>	* affinities, but others are harmed by cache pollution effects.)
88		*
89		* When a worker would otherwise be blocked waiting to join a
90		* task, it first tries a form of linear helping: Each worker
#	Line 108 | Line 111 | public class ForkJoinWorkerThread extend
111		* miss links in the chain during long-lived tasks, GC stalls etc
112		* (which is OK since blocking in such cases is usually a good
113		* idea).  (4) We bound the number of attempts to find work (see
114	<	* MAX_HELP_DEPTH) and fall back to suspending the worker and if
115	<	* necessary replacing it with a spare (see
113	<	* ForkJoinPool.awaitJoin).
114	>	* MAX_HELP) and fall back to suspending the worker and if
115	>	* necessary replacing it with another.
116		*
117		* Efficient implementation of these algorithms currently relies
118		* on an uncomfortable amount of "Unsafe" mechanics. To maintain
119	<	* correct orderings, reads and writes of variable base require
120	<	* volatile ordering.  Variable sp does not require volatile
121	<	* writes but still needs store-ordering, which we accomplish by
122	<	* pre-incrementing sp before filling the slot with an ordered
123	<	* store.  (Pre-incrementing also enables backouts used in
124	<	* joinTask.)  Because they are protected by volatile base reads,
125	<	* reads of the queue array and its slots by other threads do not
126	<	* need volatile load semantics, but writes (in push) require
127	<	* store order and CASes (in pop and deq) require (volatile) CAS
128	<	* semantics.  (Michael, Saraswat, and Vechev's algorithm has
129	<	* similar properties, but without support for nulling slots.)
130	<	* Since these combinations aren't supported using ordinary
131	<	* volatiles, the only way to accomplish these efficiently is to
132	<	* use direct Unsafe calls. (Using external AtomicIntegers and
133	<	* AtomicReferenceArrays for the indices and array is
132	<	* significantly slower because of memory locality and indirection
133	<	* effects.)
119	>	* correct orderings, reads and writes of variable queueBase
120	>	* require volatile ordering.  Variable queueTop need not be
121	>	* volatile because non-local reads always follow those of
122	>	* queueBase.  Similarly, because they are protected by volatile
123	>	* queueBase reads, reads of the queue array and its slots by
124	>	* other threads do not need volatile load semantics, but writes
125	>	* (in push) require store order and CASes (in pop and deq)
126	>	* require (volatile) CAS semantics.  (Michael, Saraswat, and
127	>	* Vechev's algorithm has similar properties, but without support
128	>	* for nulling slots.)  Since these combinations aren't supported
129	>	* using ordinary volatiles, the only way to accomplish these
130	>	* efficiently is to use direct Unsafe calls. (Using external
131	>	* AtomicIntegers and AtomicReferenceArrays for the indices and
132	>	* array is significantly slower because of memory locality and
133	>	* indirection effects.)
134		*
135		* Further, performance on most platforms is very sensitive to
136		* placement and sizing of the (resizable) queue array.  Even
#	Line 138 | Line 138 | public class ForkJoinWorkerThread extend
138		* initial size must be large enough to counteract cache
139		* contention effects across multiple queues (especially in the
140		* presence of GC cardmarking). Also, to improve thread-locality,
141	<	* queues are initialized after starting.  All together, these
142	<	* low-level implementation choices produce as much as a factor of
143	<	* 4 performance improvement compared to naive implementations,
144	<	* and enable the processing of billions of tasks per second,
145	<	* sometimes at the expense of ugliness.
141	>	* queues are initialized after starting.
142		*/
143
144		/**
145	<	* Generator for initial random seeds for random victim
150	<	* selection. This is used only to create initial seeds. Random
151	<	* steals use a cheaper xorshift generator per steal attempt. We
152	<	* expect only rare contention on seedGenerator, so just use a
153	<	* plain Random.
145	>	* Mask for pool indices encoded as shorts
146		*/
147	<	private static final Random seedGenerator = new Random();
156	<
157	<	/**
158	<	* The maximum stolen->joining link depth allowed in helpJoinTask.
159	<	* Depths for legitimate chains are unbounded, but we use a fixed
160	<	* constant to avoid (otherwise unchecked) cycles and bound
161	<	* staleness of traversal parameters at the expense of sometimes
162	<	* blocking when we could be helping.
163	<	*/
164	<	private static final int MAX_HELP_DEPTH = 8;
147	>	private static final int  SMASK  = 0xffff;
148
149		/**
150		* Capacity of work-stealing queue array upon initialization.
#	Line 171 | Line 154 | public class ForkJoinWorkerThread extend
154		private static final int INITIAL_QUEUE_CAPACITY = 1 << 13;
155
156		/**
157	<	* Maximum work-stealing queue array size.  Must be less than or
158	<	* equal to 1 << (31 - width of array entry) to ensure lack of
159	<	* index wraparound. The value is set in the static block
160	<	* at the end of this file after obtaining width.
157	>	* Maximum size for queue array. Must be a power of two
158	>	* less than or equal to 1 << (31 - width of array entry) to
159	>	* ensure lack of index wraparound, but is capped at a lower
160	>	* value to help users trap runaway computations.
161		*/
162	<	private static final int MAXIMUM_QUEUE_CAPACITY;
162	>	private static final int MAXIMUM_QUEUE_CAPACITY = 1 << 24; // 16M
163	>
164	>	/**
165	>	* The work-stealing queue array. Size must be a power of two.
166	>	* Initialized when started (as oposed to when constructed), to
167	>	* improve memory locality.
168	>	*/
169	>	ForkJoinTask<?>[] queue;
170
171		/**
172		* The pool this thread works in. Accessed directly by ForkJoinTask.
#	Line 184 | Line 174 | public class ForkJoinWorkerThread extend
174		final ForkJoinPool pool;
175
176		/**
177	<	* The work-stealing queue array. Size must be a power of two.
178	<	* Initialized in onStart, to improve memory locality.
177	>	* Index (mod queue.length) of next queue slot to push to or pop
178	>	* from. It is written only by owner thread, and accessed by other
179	>	* threads only after reading (volatile) queueBase.  Both queueTop
180	>	* and queueBase are allowed to wrap around on overflow, but
181	>	* (queueTop - queueBase) still estimates size.
182		*/
183	<	private ForkJoinTask<?>[] queue;
183	>	int queueTop;
184
185		/**
186		* Index (mod queue.length) of least valid queue slot, which is
187		* always the next position to steal from if nonempty.
188		*/
189	<	private volatile int base;
197	<
198	<	/**
199	<	* Index (mod queue.length) of next queue slot to push to or pop
200	<	* from. It is written only by owner thread, and accessed by other
201	<	* threads only after reading (volatile) base.  Both sp and base
202	<	* are allowed to wrap around on overflow, but (sp - base) still
203	<	* estimates size.
204	<	*/
205	<	private int sp;
189	>	volatile int queueBase;
190
191		/**
192		* The index of most recent stealer, used as a hint to avoid
#	Line 211 | Line 195 | public class ForkJoinWorkerThread extend
195		* of them (usually the most current). Declared non-volatile,
196		* relying on other prevailing sync to keep reasonably current.
197		*/
198	<	private int stealHint;
198	>	int stealHint;
199
200		/**
201	<	* Run state of this worker. In addition to the usual run levels,
202	<	* tracks if this worker is suspended as a spare, and if it was
203	<	* killed (trimmed) while suspended. However, "active" status is
220	<	* maintained separately and modified only in conjunction with
221	<	* CASes of the pool's runState (which are currently sadly
222	<	* manually inlined for performance.)  Accessed directly by pool
223	<	* to simplify checks for normal (zero) status.
224	<	*/
225	<	volatile int runState;
226	<
227	<	private static final int TERMINATING = 0x01;
228	<	private static final int TERMINATED  = 0x02;
229	<	private static final int SUSPENDED   = 0x04; // inactive spare
230	<	private static final int TRIMMED     = 0x08; // killed while suspended
231	<
232	<	/**
233	<	* Number of steals. Directly accessed (and reset) by
234	<	* pool.tryAccumulateStealCount when idle.
201	>	* Index of this worker in pool array. Set once by pool before
202	>	* running, and accessed directly by pool to locate this worker in
203	>	* its workers array.
204		*/
205	<	int stealCount;
205	>	final int poolIndex;
206
207		/**
208	<	* Seed for random number generator for choosing steal victims.
209	<	* Uses Marsaglia xorshift. Must be initialized as nonzero.
208	>	* Encoded record for pool task waits. Usages are always
209	>	* surrounded by volatile reads/writes
210		*/
211	<	private int seed;
211	>	int nextWait;
212
213		/**
214	<	* Activity status. When true, this worker is considered active.
215	<	* Accessed directly by pool.  Must be false upon construction.
214	>	* Complement of poolIndex, offset by count of entries of task
215	>	* waits. Accessed by ForkJoinPool to manage event waiters.
216		*/
217	<	boolean active;
217	>	volatile int eventCount;
218
219		/**
220	<	* True if use local fifo, not default lifo, for local polling.
221	<	* Shadows value from ForkJoinPool.
220	>	* Seed for random number generator for choosing steal victims.
221	>	* Uses Marsaglia xorshift. Must be initialized as nonzero.
222		*/
223	<	private final boolean locallyFifo;
223	>	int seed;
224
225		/**
226	<	* Index of this worker in pool array. Set once by pool before
227	<	* running, and accessed directly by pool to locate this worker in
259	<	* its workers array.
226	>	* Number of steals. Directly accessed (and reset) by pool when
227	>	* idle.
228		*/
229	<	int poolIndex;
229	>	int stealCount;
230
231		/**
232	<	* The last pool event waited for. Accessed only by pool in
265	<	* callback methods invoked within this thread.
232	>	* True if this worker should or did terminate
233		*/
234	<	int lastEventCount;
234	>	volatile boolean terminate;
235
236		/**
237	<	* Encoded index and event count of next event waiter. Accessed
271	<	* only by ForkJoinPool for managing event waiters.
237	>	* Set to true before LockSupport.park; false on return
238		*/
239	<	volatile long nextWaiter;
239	>	volatile boolean parked;
240
241		/**
242	<	* Number of times this thread suspended as spare. Accessed only
243	<	* by pool.
242	>	* True if use local fifo, not default lifo, for local polling.
243	>	* Shadows value from ForkJoinPool.
244		*/
245	<	int spareCount;
245	>	final boolean locallyFifo;
246
247		/**
248	<	* Encoded index and count of next spare waiter. Accessed only
249	<	* by ForkJoinPool for managing spares.
248	>	* The task most recently stolen from another worker (or
249	>	* submission queue).  All uses are surrounded by enough volatile
250	>	* reads/writes to maintain as non-volatile.
251		*/
252	<	volatile int nextSpare;
252	>	ForkJoinTask<?> currentSteal;
253
254		/**
255		* The task currently being joined, set only when actively trying
256	<	* to help other stealers in helpJoinTask. Written only by this
257	<	* thread, but read by others.
291	<	*/
292	<	private volatile ForkJoinTask<?> currentJoin;
293	<
294	<	/**
295	<	* The task most recently stolen from another worker (or
296	<	* submission queue).  Written only by this thread, but read by
297	<	* others.
256	>	* to help other stealers in helpJoinTask. All uses are surrounded
257	>	* by enough volatile reads/writes to maintain as non-volatile.
258		*/
259	<	private volatile ForkJoinTask<?> currentSteal;
259	>	ForkJoinTask<?> currentJoin;
260
261		/**
262		* Creates a ForkJoinWorkerThread operating in the given pool.
#	Line 305 | Line 265 | public class ForkJoinWorkerThread extend
265		* @throws NullPointerException if pool is null
266		*/
267		protected ForkJoinWorkerThread(ForkJoinPool pool) {
268	+	super(pool.nextWorkerName());
269		this.pool = pool;
270	<	this.locallyFifo = pool.locallyFifo;
271	<	setDaemon(true);
272	<	// To avoid exposing construction details to subclasses,
273	<	// remaining initialization is in start() and onStart()
274	<	}
314	<
315	<	/**
316	<	* Performs additional initialization and starts this thread.
317	<	*/
318	<	final void start(int poolIndex, UncaughtExceptionHandler ueh) {
319	<	this.poolIndex = poolIndex;
270	>	int k = pool.registerWorker(this);
271	>	poolIndex = k;
272	>	eventCount = ~k & SMASK; // clear wait count
273	>	locallyFifo = pool.locallyFifo;
274	>	Thread.UncaughtExceptionHandler ueh = pool.ueh;
275		if (ueh != null)
276		setUncaughtExceptionHandler(ueh);
277	<	start();
277	>	setDaemon(true);
278		}
279
280	<	// Public/protected methods
280	>	// Public methods
281
282		/**
283		* Returns the pool hosting this thread.
#	Line 346 | Line 301 | public class ForkJoinWorkerThread extend
301		return poolIndex;
302		}
303
304	+	// Randomization
305	+
306	+	/**
307	+	* Computes next value for random victim probes and backoffs.
308	+	* Scans don't require a very high quality generator, but also not
309	+	* a crummy one.  Marsaglia xor-shift is cheap and works well
310	+	* enough.  Note: This is manually inlined in FJP.scan() to avoid
311	+	* writes inside busy loops.
312	+	*/
313	+	private int nextSeed() {
314	+	int r = seed;
315	+	r ^= r << 13;
316	+	r ^= r >>> 17;
317	+	r ^= r << 5;
318	+	return seed = r;
319	+	}
320	+
321	+	// Run State management
322	+
323		/**
324		* Initializes internal state after construction but before
325		* processing any tasks. If you override this method, you must
#	Line 356 | Line 330 | public class ForkJoinWorkerThread extend
330		* processing tasks.
331		*/
332		protected void onStart() {
359	–	int rs = seedGenerator.nextInt();
360	–	seed = (rs == 0) ? 1 : rs; // seed must be nonzero
361	–
362	–	// Allocate name string and arrays in this thread
363	–	String pid = Integer.toString(pool.getPoolNumber());
364	–	String wid = Integer.toString(poolIndex);
365	–	setName("ForkJoinPool-" + pid + "-worker-" + wid);
366	–
333		queue = new ForkJoinTask<?>[INITIAL_QUEUE_CAPACITY];
334	+	int r = pool.workerSeedGenerator.nextInt();
335	+	seed = (r == 0)? 1 : r; //  must be nonzero
336		}
337
338		/**
#	Line 377 | Line 345 | public class ForkJoinWorkerThread extend
345		*/
346		protected void onTermination(Throwable exception) {
347		try {
348	<	ForkJoinPool p = pool;
381	<	if (active) {
382	<	int a; // inline p.tryDecrementActiveCount
383	<	active = false;
384	<	do {} while (!UNSAFE.compareAndSwapInt
385	<	(p, poolRunStateOffset, a = p.runState, a - 1));
386	<	}
348	>	terminate = true;
349		cancelTasks();
350	<	setTerminated();
389	<	p.workerTerminated(this);
350	>	pool.deregisterWorker(this, exception);
351		} catch (Throwable ex) {        // Shouldn't ever happen
352		if (exception == null)      // but if so, at least rethrown
353		exception = ex;
#	Line 405 | Line 366 | public class ForkJoinWorkerThread extend
366		Throwable exception = null;
367		try {
368		onStart();
369	<	mainLoop();
369	>	pool.work(this);
370		} catch (Throwable ex) {
371		exception = ex;
372		} finally {
#	Line 413 | Line 374 | public class ForkJoinWorkerThread extend
374		}
375		}
376
416	–	// helpers for run()
417	–
418	–	/**
419	–	* Finds and executes tasks, and checks status while running.
420	–	*/
421	–	private void mainLoop() {
422	–	boolean ran = false; // true if ran a task on last step
423	–	ForkJoinPool p = pool;
424	–	for (;;) {
425	–	p.preStep(this, ran);
426	–	if (runState != 0)
427	–	break;
428	–	ran = tryExecSteal() || tryExecSubmission();
429	–	}
430	–	}
431	–
432	–	/**
433	–	* Tries to steal a task and execute it.
434	–	*
435	–	* @return true if ran a task
436	–	*/
437	–	private boolean tryExecSteal() {
438	–	ForkJoinTask<?> t;
439	–	if ((t = scan()) != null) {
440	–	t.quietlyExec();
441	–	UNSAFE.putOrderedObject(this, currentStealOffset, null);
442	–	if (sp != base)
443	–	execLocalTasks();
444	–	return true;
445	–	}
446	–	return false;
447	–	}
448	–
449	–	/**
450	–	* If a submission exists, try to activate and run it.
451	–	*
452	–	* @return true if ran a task
453	–	*/
454	–	private boolean tryExecSubmission() {
455	–	ForkJoinPool p = pool;
456	–	// This loop is needed in case attempt to activate fails, in
457	–	// which case we only retry if there still appears to be a
458	–	// submission.
459	–	while (p.hasQueuedSubmissions()) {
460	–	ForkJoinTask<?> t; int a;
461	–	if (active || // inline p.tryIncrementActiveCount
462	–	(active = UNSAFE.compareAndSwapInt(p, poolRunStateOffset,
463	–	a = p.runState, a + 1))) {
464	–	if ((t = p.pollSubmission()) != null) {
465	–	UNSAFE.putOrderedObject(this, currentStealOffset, t);
466	–	t.quietlyExec();
467	–	UNSAFE.putOrderedObject(this, currentStealOffset, null);
468	–	if (sp != base)
469	–	execLocalTasks();
470	–	return true;
471	–	}
472	–	}
473	–	}
474	–	return false;
475	–	}
476	–
477	–	/**
478	–	* Runs local tasks until queue is empty or shut down.  Call only
479	–	* while active.
480	–	*/
481	–	private void execLocalTasks() {
482	–	while (runState == 0) {
483	–	ForkJoinTask<?> t = locallyFifo ? locallyDeqTask() : popTask();
484	–	if (t != null)
485	–	t.quietlyExec();
486	–	else if (sp == base)
487	–	break;
488	–	}
489	–	}
490	–
377		/*
378		* Intrinsics-based atomic writes for queue slots. These are
379		* basically the same as methods in AtomicReferenceArray, but
#	Line 499 | Line 385 | public class ForkJoinWorkerThread extend
385		* because they are protected by other volatile reads and are
386		* confirmed by CASes.
387		*
388	<	* Most uses don't actually call these methods, but instead contain
389	<	* inlined forms that enable more predictable optimization.  We
390	<	* don't define the version of write used in pushTask at all, but
391	<	* instead inline there a store-fenced array slot write.
388	>	* Most uses don't actually call these methods, but instead
389	>	* contain inlined forms that enable more predictable
390	>	* optimization.  We don't define the version of write used in
391	>	* pushTask at all, but instead inline there a store-fenced array
392	>	* slot write.
393	>	*
394	>	* Also in most methods, as a performance (not correctness) issue,
395	>	* we'd like to encourage compilers not to arbitrarily postpone
396	>	* setting queueTop after writing slot.  Currently there is no
397	>	* intrinsic for arranging this, but using Unsafe putOrderedInt
398	>	* may be a preferable strategy on some compilers even though its
399	>	* main effect is a pre-, not post- fence. To simplify possible
400	>	* changes, the option is left in comments next to the associated
401	>	* assignments.
402		*/
403
404		/**
#	Line 511 | Line 407 | public class ForkJoinWorkerThread extend
407		*/
408		private static final boolean casSlotNull(ForkJoinTask<?>[] q, int i,
409		ForkJoinTask<?> t) {
410	<	return UNSAFE.compareAndSwapObject(q, (i << qShift) + qBase, t, null);
410	>	return UNSAFE.compareAndSwapObject(q, (i << ASHIFT) + ABASE, t, null);
411		}
412
413		/**
#	Line 521 | Line 417 | public class ForkJoinWorkerThread extend
417		*/
418		private static final void writeSlot(ForkJoinTask<?>[] q, int i,
419		ForkJoinTask<?> t) {
420	<	UNSAFE.putObjectVolatile(q, (i << qShift) + qBase, t);
420	>	UNSAFE.putObjectVolatile(q, (i << ASHIFT) + ABASE, t);
421		}
422
423		// queue methods
#	Line 532 | Line 428 | public class ForkJoinWorkerThread extend
428		* @param t the task. Caller must ensure non-null.
429		*/
430		final void pushTask(ForkJoinTask<?> t) {
431	<	ForkJoinTask<?>[] q = queue;
432	<	int mask = q.length - 1; // implicit assert q != null
433	<	int s = sp++;            // ok to increment sp before slot write
434	<	UNSAFE.putOrderedObject(q, ((s & mask) << qShift) + qBase, t);
435	<	if ((s -= base) == 0)
436	<	pool.signalWork();   // was empty
437	<	else if (s == mask)
438	<	growQueue();         // is full
431	>	ForkJoinTask<?>[] q; int s, m;
432	>	if ((q = queue) != null) {    // ignore if queue removed
433	>	long u = (((s = queueTop) & (m = q.length - 1)) << ASHIFT) + ABASE;
434	>	UNSAFE.putOrderedObject(q, u, t);
435	>	queueTop = s + 1;         // or use putOrderedInt
436	>	if ((s -= queueBase) <= 2)
437	>	pool.signalWork();
438	>	else if (s == m)
439	>	growQueue();
440	>	}
441	>	}
442	>
443	>	/**
444	>	* Creates or doubles queue array.  Transfers elements by
445	>	* emulating steals (deqs) from old array and placing, oldest
446	>	* first, into new array.
447	>	*/
448	>	private void growQueue() {
449	>	ForkJoinTask<?>[] oldQ = queue;
450	>	int size = oldQ != null ? oldQ.length << 1 : INITIAL_QUEUE_CAPACITY;
451	>	if (size > MAXIMUM_QUEUE_CAPACITY)
452	>	throw new RejectedExecutionException("Queue capacity exceeded");
453	>	if (size < INITIAL_QUEUE_CAPACITY)
454	>	size = INITIAL_QUEUE_CAPACITY;
455	>	ForkJoinTask<?>[] q = queue = new ForkJoinTask<?>[size];
456	>	int mask = size - 1;
457	>	int top = queueTop;
458	>	int oldMask;
459	>	if (oldQ != null && (oldMask = oldQ.length - 1) >= 0) {
460	>	for (int b = queueBase; b != top; ++b) {
461	>	long u = ((b & oldMask) << ASHIFT) + ABASE;
462	>	Object x = UNSAFE.getObjectVolatile(oldQ, u);
463	>	if (x != null && UNSAFE.compareAndSwapObject(oldQ, u, x, null))
464	>	UNSAFE.putObjectVolatile
465	>	(q, ((b & mask) << ASHIFT) + ABASE, x);
466	>	}
467	>	}
468		}
469
470		/**
#	Line 550 | Line 475 | public class ForkJoinWorkerThread extend
475		* @return a task, or null if none or contended
476		*/
477		final ForkJoinTask<?> deqTask() {
478	<	ForkJoinTask<?> t;
479	<	ForkJoinTask<?>[] q;
555	<	int b, i;
556	<	if (sp != (b = base) &&
478	>	ForkJoinTask<?> t; ForkJoinTask<?>[] q; int b, i;
479	>	if (queueTop != (b = queueBase) &&
480		(q = queue) != null && // must read q after b
481	<	(t = q[i = (q.length - 1) & b]) != null && base == b &&
482	<	UNSAFE.compareAndSwapObject(q, (i << qShift) + qBase, t, null)) {
483	<	base = b + 1;
481	>	(i = (q.length - 1) & b) >= 0 &&
482	>	(t = q[i]) != null && queueBase == b &&
483	>	UNSAFE.compareAndSwapObject(q, (i << ASHIFT) + ABASE, t, null)) {
484	>	queueBase = b + 1;
485		return t;
486		}
487		return null;
488		}
489
490		/**
491	<	* Tries to take a task from the base of own queue. Assumes active
492	<	* status.  Called only by this thread.
491	>	* Tries to take a task from the base of own queue.  Called only
492	>	* by this thread.
493		*
494		* @return a task, or null if none
495		*/
496		final ForkJoinTask<?> locallyDeqTask() {
497	+	ForkJoinTask<?> t; int m, b, i;
498		ForkJoinTask<?>[] q = queue;
499	<	if (q != null) {
500	<	ForkJoinTask<?> t;
501	<	int b, i;
502	<	while (sp != (b = base)) {
503	<	if ((t = q[i = (q.length - 1) & b]) != null && base == b &&
579	<	UNSAFE.compareAndSwapObject(q, (i << qShift) + qBase,
499	>	if (q != null && (m = q.length - 1) >= 0) {
500	>	while (queueTop != (b = queueBase)) {
501	>	if ((t = q[i = m & b]) != null &&
502	>	queueBase == b &&
503	>	UNSAFE.compareAndSwapObject(q, (i << ASHIFT) + ABASE,
504		t, null)) {
505	<	base = b + 1;
505	>	queueBase = b + 1;
506		return t;
507		}
508		}
#	Line 587 | Line 511 | public class ForkJoinWorkerThread extend
511		}
512
513		/**
514	<	* Returns a popped task, or null if empty. Assumes active status.
514	>	* Returns a popped task, or null if empty.
515		* Called only by this thread.
516		*/
517		private ForkJoinTask<?> popTask() {
518	+	int m;
519		ForkJoinTask<?>[] q = queue;
520	<	if (q != null) {
521	<	int s;
522	<	while ((s = sp) != base) {
523	<	int i = (q.length - 1) & --s;
599	<	long u = (i << qShift) + qBase; // raw offset
520	>	if (q != null && (m = q.length - 1) >= 0) {
521	>	for (int s; (s = queueTop) != queueBase;) {
522	>	int i = m & --s;
523	>	long u = (i << ASHIFT) + ABASE; // raw offset
524		ForkJoinTask<?> t = q[i];
525		if (t == null)   // lost to stealer
526		break;
527		if (UNSAFE.compareAndSwapObject(q, u, t, null)) {
528	<	/*
605	<	* Note: here and in related methods, as a
606	<	* performance (not correctness) issue, we'd like
607	<	* to encourage compiler not to arbitrarily
608	<	* postpone setting sp after successful CAS.
609	<	* Currently there is no intrinsic for arranging
610	<	* this, but using Unsafe putOrderedInt may be a
611	<	* preferable strategy on some compilers even
612	<	* though its main effect is a pre-, not post-
613	<	* fence. To simplify possible changes, the option
614	<	* is left in comments next to the associated
615	<	* assignments.
616	<	*/
617	<	sp = s; // putOrderedInt may encourage more timely write
618	<	// UNSAFE.putOrderedInt(this, spOffset, s);
528	>	queueTop = s; // or putOrderedInt
529		return t;
530		}
531		}
#	Line 625 | Line 535 | public class ForkJoinWorkerThread extend
535
536		/**
537		* Specialized version of popTask to pop only if topmost element
538	<	* is the given task. Called only by this thread while active.
538	>	* is the given task. Called only by this thread.
539		*
540		* @param t the task. Caller must ensure non-null.
541		*/
542		final boolean unpushTask(ForkJoinTask<?> t) {
543	+	ForkJoinTask<?>[] q;
544		int s;
545	<	ForkJoinTask<?>[] q = queue;
635	<	if ((s = sp) != base && q != null &&
545	>	if ((q = queue) != null && (s = queueTop) != queueBase &&
546		UNSAFE.compareAndSwapObject
547	<	(q, (((q.length - 1) & --s) << qShift) + qBase, t, null)) {
548	<	sp = s; // putOrderedInt may encourage more timely write
639	<	// UNSAFE.putOrderedInt(this, spOffset, s);
547	>	(q, (((q.length - 1) & --s) << ASHIFT) + ABASE, t, null)) {
548	>	queueTop = s; // or putOrderedInt
549		return true;
550		}
551		return false;
#	Line 646 | Line 555 | public class ForkJoinWorkerThread extend
555		* Returns next task, or null if empty or contended.
556		*/
557		final ForkJoinTask<?> peekTask() {
558	+	int m;
559		ForkJoinTask<?>[] q = queue;
560	<	if (q == null)
560	>	if (q == null || (m = q.length - 1) < 0)
561		return null;
562	<	int mask = q.length - 1;
563	<	int i = locallyFifo ? base : (sp - 1);
654	<	return q[i & mask];
655	<	}
656	<
657	<	/**
658	<	* Doubles queue array size. Transfers elements by emulating
659	<	* steals (deqs) from old array and placing, oldest first, into
660	<	* new array.
661	<	*/
662	<	private void growQueue() {
663	<	ForkJoinTask<?>[] oldQ = queue;
664	<	int oldSize = oldQ.length;
665	<	int newSize = oldSize << 1;
666	<	if (newSize > MAXIMUM_QUEUE_CAPACITY)
667	<	throw new RejectedExecutionException("Queue capacity exceeded");
668	<	ForkJoinTask<?>[] newQ = queue = new ForkJoinTask<?>[newSize];
669	<
670	<	int b = base;
671	<	int bf = b + oldSize;
672	<	int oldMask = oldSize - 1;
673	<	int newMask = newSize - 1;
674	<	do {
675	<	int oldIndex = b & oldMask;
676	<	ForkJoinTask<?> t = oldQ[oldIndex];
677	<	if (t != null && !casSlotNull(oldQ, oldIndex, t))
678	<	t = null;
679	<	writeSlot(newQ, b & newMask, t);
680	<	} while (++b != bf);
681	<	pool.signalWork();
562	>	int i = locallyFifo ? queueBase : (queueTop - 1);
563	>	return q[i & m];
564		}
565
566	<	/**
685	<	* Computes next value for random victim probe in scan().  Scans
686	<	* don't require a very high quality generator, but also not a
687	<	* crummy one.  Marsaglia xor-shift is cheap and works well enough.
688	<	* Note: This is manually inlined in scan().
689	<	*/
690	<	private static final int xorShift(int r) {
691	<	r ^= r << 13;
692	<	r ^= r >>> 17;
693	<	return r ^ (r << 5);
694	<	}
566	>	// Support methods for ForkJoinPool
567
568		/**
569	<	* Tries to steal a task from another worker. Starts at a random
698	<	* index of workers array, and probes workers until finding one
699	<	* with non-empty queue or finding that all are empty.  It
700	<	* randomly selects the first n probes. If these are empty, it
701	<	* resorts to a circular sweep, which is necessary to accurately
702	<	* set active status. (The circular sweep uses steps of
703	<	* approximately half the array size plus 1, to avoid bias
704	<	* stemming from leftmost packing of the array in ForkJoinPool.)
705	<	*
706	<	* This method must be both fast and quiet -- usually avoiding
707	<	* memory accesses that could disrupt cache sharing etc other than
708	<	* those needed to check for and take tasks (or to activate if not
709	<	* already active). This accounts for, among other things,
710	<	* updating random seed in place without storing it until exit.
711	<	*
712	<	* @return a task, or null if none found
569	>	* Runs the given task, plus any local tasks until queue is empty
570		*/
571	<	private ForkJoinTask<?> scan() {
572	<	ForkJoinPool p = pool;
716	<	ForkJoinWorkerThread[] ws;        // worker array
717	<	int n;                            // upper bound of #workers
718	<	if ((ws = p.workers) != null && (n = ws.length) > 1) {
719	<	boolean canSteal = active;    // shadow active status
720	<	int r = seed;                 // extract seed once
721	<	int mask = n - 1;
722	<	int j = -n;                   // loop counter
723	<	int k = r;                    // worker index, random if j < 0
724	<	for (;;) {
725	<	ForkJoinWorkerThread v = ws[k & mask];
726	<	r ^= r << 13; r ^= r >>> 17; r ^= r << 5; // inline xorshift
727	<	ForkJoinTask<?>[] q; ForkJoinTask<?> t; int b, a;
728	<	if (v != null && (b = v.base) != v.sp &&
729	<	(q = v.queue) != null) {
730	<	int i = (q.length - 1) & b;
731	<	long u = (i << qShift) + qBase; // raw offset
732	<	int pid = poolIndex;
733	<	if ((t = q[i]) != null) {
734	<	if (!canSteal &&  // inline p.tryIncrementActiveCount
735	<	UNSAFE.compareAndSwapInt(p, poolRunStateOffset,
736	<	a = p.runState, a + 1))
737	<	canSteal = active = true;
738	<	if (canSteal && v.base == b++ &&
739	<	UNSAFE.compareAndSwapObject(q, u, t, null)) {
740	<	v.base = b;
741	<	v.stealHint = pid;
742	<	UNSAFE.putOrderedObject(this,
743	<	currentStealOffset, t);
744	<	seed = r;
745	<	++stealCount;
746	<	return t;
747	<	}
748	<	}
749	<	j = -n;
750	<	k = r;                // restart on contention
751	<	}
752	<	else if (++j <= 0)
753	<	k = r;
754	<	else if (j <= n)
755	<	k += (n >>> 1) | 1;
756	<	else
757	<	break;
758	<	}
759	<	}
760	<	return null;
761	<	}
762	<
763	<	// Run State management
764	<
765	<	// status check methods used mainly by ForkJoinPool
766	<	final boolean isRunning()    { return runState == 0; }
767	<	final boolean isTerminated() { return (runState & TERMINATED) != 0; }
768	<	final boolean isSuspended()  { return (runState & SUSPENDED) != 0; }
769	<	final boolean isTrimmed()    { return (runState & TRIMMED) != 0; }
770	<
771	<	final boolean isTerminating() {
772	<	if ((runState & TERMINATING) != 0)
773	<	return true;
774	<	if (pool.isAtLeastTerminating()) { // propagate pool state
775	<	shutdown();
776	<	return true;
777	<	}
778	<	return false;
779	<	}
780	<
781	<	/**
782	<	* Sets state to TERMINATING. Does NOT unpark or interrupt
783	<	* to wake up if currently blocked. Callers must do so if desired.
784	<	*/
785	<	final void shutdown() {
571	>	final void execTask(ForkJoinTask<?> t) {
572	>	currentSteal = t;
573		for (;;) {
574	<	int s = runState;
575	<	if ((s & (TERMINATING|TERMINATED)) != 0)
576	<	break;
790	<	if ((s & SUSPENDED) != 0) { // kill and wakeup if suspended
791	<	if (UNSAFE.compareAndSwapInt(this, runStateOffset, s,
792	<	(s & ~SUSPENDED) |
793	<	(TRIMMED|TERMINATING)))
794	<	break;
795	<	}
796	<	else if (UNSAFE.compareAndSwapInt(this, runStateOffset, s,
797	<	s | TERMINATING))
798	<	break;
799	<	}
800	<	}
801	<
802	<	/**
803	<	* Sets state to TERMINATED. Called only by onTermination().
804	<	*/
805	<	private void setTerminated() {
806	<	int s;
807	<	do {} while (!UNSAFE.compareAndSwapInt(this, runStateOffset,
808	<	s = runState,
809	<	s | (TERMINATING|TERMINATED)));
810	<	}
811	<
812	<	/**
813	<	* If suspended, tries to set status to unsuspended.
814	<	* Does NOT wake up if blocked.
815	<	*
816	<	* @return true if successful
817	<	*/
818	<	final boolean tryUnsuspend() {
819	<	int s;
820	<	while (((s = runState) & SUSPENDED) != 0) {
821	<	if (UNSAFE.compareAndSwapInt(this, runStateOffset, s,
822	<	s & ~SUSPENDED))
823	<	return true;
824	<	}
825	<	return false;
826	<	}
827	<
828	<	/**
829	<	* Sets suspended status and blocks as spare until resumed
830	<	* or shutdown.
831	<	*/
832	<	final void suspendAsSpare() {
833	<	for (;;) {                  // set suspended unless terminating
834	<	int s = runState;
835	<	if ((s & TERMINATING) != 0) { // must kill
836	<	if (UNSAFE.compareAndSwapInt(this, runStateOffset, s,
837	<	s | (TRIMMED | TERMINATING)))
838	<	return;
839	<	}
840	<	else if (UNSAFE.compareAndSwapInt(this, runStateOffset, s,
841	<	s | SUSPENDED))
574	>	if (t != null)
575	>	t.doExec();
576	>	if (queueTop == queueBase)
577		break;
578	+	t = locallyFifo ? locallyDeqTask() : popTask();
579		}
580	<	ForkJoinPool p = pool;
581	<	p.pushSpare(this);
846	<	while ((runState & SUSPENDED) != 0) {
847	<	if (p.tryAccumulateStealCount(this)) {
848	<	interrupted();          // clear/ignore interrupts
849	<	if ((runState & SUSPENDED) == 0)
850	<	break;
851	<	LockSupport.park(this);
852	<	}
853	<	}
854	<	}
855	<
856	<	// Misc support methods for ForkJoinPool
857	<
858	<	/**
859	<	* Returns an estimate of the number of tasks in the queue.  Also
860	<	* used by ForkJoinTask.
861	<	*/
862	<	final int getQueueSize() {
863	<	int n; // external calls must read base first
864	<	return (n = -base + sp) <= 0 ? 0 : n;
580	>	++stealCount;
581	>	currentSteal = null;
582		}
583
584		/**
#	Line 870 | Line 587 | public class ForkJoinWorkerThread extend
587		*/
588		final void cancelTasks() {
589		ForkJoinTask<?> cj = currentJoin; // try to cancel ongoing tasks
590	<	if (cj != null && cj.status >= 0) {
590	>	if (cj != null && cj.status >= 0)
591		cj.cancelIgnoringExceptions();
875	–	try {
876	–	this.interrupt(); // awaken wait
877	–	} catch (SecurityException ignore) {
878	–	}
879	–	}
592		ForkJoinTask<?> cs = currentSteal;
593		if (cs != null && cs.status >= 0)
594		cs.cancelIgnoringExceptions();
595	<	while (base != sp) {
595	>	while (queueBase != queueTop) {
596		ForkJoinTask<?> t = deqTask();
597		if (t != null)
598		t.cancelIgnoringExceptions();
#	Line 894 | Line 606 | public class ForkJoinWorkerThread extend
606		*/
607		final int drainTasksTo(Collection<? super ForkJoinTask<?>> c) {
608		int n = 0;
609	<	while (base != sp) {
609	>	while (queueBase != queueTop) {
610		ForkJoinTask<?> t = deqTask();
611		if (t != null) {
612		c.add(t);
#	Line 907 | Line 619 | public class ForkJoinWorkerThread extend
619		// Support methods for ForkJoinTask
620
621		/**
622	+	* Returns an estimate of the number of tasks in the queue.
623	+	*/
624	+	final int getQueueSize() {
625	+	return queueTop - queueBase;
626	+	}
627	+
628	+	/**
629		* Gets and removes a local task.
630		*
631		* @return a task, if available
632		*/
633		final ForkJoinTask<?> pollLocalTask() {
634	<	ForkJoinPool p = pool;
916	<	while (sp != base) {
917	<	int a; // inline p.tryIncrementActiveCount
918	<	if (active ||
919	<	(active = UNSAFE.compareAndSwapInt(p, poolRunStateOffset,
920	<	a = p.runState, a + 1)))
921	<	return locallyFifo ? locallyDeqTask() : popTask();
922	<	}
923	<	return null;
634	>	return locallyFifo ? locallyDeqTask() : popTask();
635		}
636
637		/**
#	Line 929 | Line 640 | public class ForkJoinWorkerThread extend
640		* @return a task, if available
641		*/
642		final ForkJoinTask<?> pollTask() {
643	+	ForkJoinWorkerThread[] ws;
644		ForkJoinTask<?> t = pollLocalTask();
645	<	if (t == null) {
646	<	t = scan();
647	<	// cannot retain/track/help steal
648	<	UNSAFE.putOrderedObject(this, currentStealOffset, null);
645	>	if (t != null || (ws = pool.workers) == null)
646	>	return t;
647	>	int n = ws.length; // cheap version of FJP.scan
648	>	int steps = n << 1;
649	>	int r = nextSeed();
650	>	int i = 0;
651	>	while (i < steps) {
652	>	ForkJoinWorkerThread w = ws[(i++ + r) & (n - 1)];
653	>	if (w != null && w.queueBase != w.queueTop && w.queue != null) {
654	>	if ((t = w.deqTask()) != null)
655	>	return t;
656	>	i = 0;
657	>	}
658		}
659	<	return t;
659	>	return null;
660		}
661
662		/**
663	<	* Possibly runs some tasks and/or blocks, until task is done.
664	<	*
665	<	* @param joinMe the task to join
666	<	* @param timed true if use timed wait
667	<	* @param nanos wait time if timed
663	>	* The maximum stolen->joining link depth allowed in helpJoinTask,
664	>	* as well as the maximum number of retries (allowing on average
665	>	* one staleness retry per level) per attempt to instead try
666	>	* compensation.  Depths for legitimate chains are unbounded, but
667	>	* we use a fixed constant to avoid (otherwise unchecked) cycles
668	>	* and bound staleness of traversal parameters at the expense of
669	>	* sometimes blocking when we could be helping.
670		*/
671	<	final void joinTask(ForkJoinTask<?> joinMe, boolean timed, long nanos) {
949	<	// currentJoin only written by this thread; only need ordered store
950	<	ForkJoinTask<?> prevJoin = currentJoin;
951	<	UNSAFE.putOrderedObject(this, currentJoinOffset, joinMe);
952	<	pool.awaitJoin(joinMe, this, timed, nanos);
953	<	UNSAFE.putOrderedObject(this, currentJoinOffset, prevJoin);
954	<	}
671	>	private static final int MAX_HELP = 16;
672
673		/**
674	<	* Tries to locate and help perform tasks for a stealer of the
958	<	* given task, or in turn one of its stealers.  Traces
959	<	* currentSteal->currentJoin links looking for a thread working on
960	<	* a descendant of the given task and with a non-empty queue to
961	<	* steal back and execute tasks from.
962	<	*
963	<	* The implementation is very branchy to cope with potential
964	<	* inconsistencies or loops encountering chains that are stale,
965	<	* unknown, or of length greater than MAX_HELP_DEPTH links.  All
966	<	* of these cases are dealt with by just returning back to the
967	<	* caller, who is expected to retry if other join mechanisms also
968	<	* don't work out.
674	>	* Possibly runs some tasks and/or blocks, until joinMe is done.
675		*
676		* @param joinMe the task to join
677	<	* @param running if false, then must update pool count upon
678	<	*  running a task
679	<	* @return value of running on exit
680	<	*/
681	<	final boolean helpJoinTask(ForkJoinTask<?> joinMe, boolean running) {
682	<	/*
683	<	* Initial checks to (1) abort if terminating; (2) clean out
684	<	* old cancelled tasks from local queue; (3) if joinMe is next
685	<	* task, run it; (4) omit scan if local queue nonempty (since
686	<	* it may contain non-descendents of joinMe).
687	<	*/
688	<	ForkJoinPool p = pool;
689	<	for (;;) {
690	<	ForkJoinTask<?>[] q;
691	<	int s;
692	<	if (joinMe.status < 0)
693	<	return running;
694	<	else if ((runState & TERMINATING) != 0) {
695	<	joinMe.cancelIgnoringExceptions();
696	<	return running;
677	>	* @return completion status on exit
678	>	*/
679	>	final int joinTask(ForkJoinTask<?> joinMe) {
680	>	ForkJoinTask<?> prevJoin = currentJoin;
681	>	currentJoin = joinMe;
682	>	for (int s, retries = MAX_HELP;;) {
683	>	if ((s = joinMe.status) < 0) {
684	>	currentJoin = prevJoin;
685	>	return s;
686	>	}
687	>	if (retries > 0) {
688	>	if (queueTop != queueBase) {
689	>	if (!localHelpJoinTask(joinMe))
690	>	retries = 0;           // cannot help
691	>	}
692	>	else if (retries == MAX_HELP >>> 1) {
693	>	--retries;                 // check uncommon case
694	>	if (tryDeqAndExec(joinMe) >= 0)
695	>	Thread.yield();        // for politeness
696	>	}
697	>	else
698	>	retries = helpJoinTask(joinMe)? MAX_HELP : retries - 1;
699		}
992	–	else if ((s = sp) == base || (q = queue) == null)
993	–	break;                            // queue empty
700		else {
701	<	int i = (q.length - 1) & --s;
702	<	long u = (i << qShift) + qBase;   // raw offset
997	<	ForkJoinTask<?> t = q[i];
998	<	if (t == null)
999	<	break;                        // lost to a stealer
1000	<	else if (t != joinMe && t.status >= 0)
1001	<	return running;               // cannot safely help
1002	<	else if ((running ||
1003	<	(running = p.tryIncrementRunningCount())) &&
1004	<	UNSAFE.compareAndSwapObject(q, u, t, null)) {
1005	<	sp = s; // putOrderedInt may encourage more timely write
1006	<	// UNSAFE.putOrderedInt(this, spOffset, s);
1007	<	t.quietlyExec();
1008	<	}
701	>	retries = MAX_HELP;           // restart if not done
702	>	pool.tryAwaitJoin(joinMe);
703		}
704		}
705	+	}
706
707	<	int n;                                    // worker array size
708	<	ForkJoinWorkerThread[] ws = p.workers;
709	<	if (ws != null && (n = ws.length) > 1) {  // need at least 2 workers
710	<	ForkJoinTask<?> task = joinMe;        // base of chain
711	<	ForkJoinWorkerThread thread = this;   // thread with stolen task
707	>	/**
708	>	* If present, pops and executes the given task, or any other
709	>	* cancelled task
710	>	*
711	>	* @return false if any other non-cancelled task exists in local queue
712	>	*/
713	>	private boolean localHelpJoinTask(ForkJoinTask<?> joinMe) {
714	>	int s, i; ForkJoinTask<?>[] q; ForkJoinTask<?> t;
715	>	if ((s = queueTop) != queueBase && (q = queue) != null &&
716	>	(i = (q.length - 1) & --s) >= 0 &&
717	>	(t = q[i]) != null) {
718	>	if (t != joinMe && t.status >= 0)
719	>	return false;
720	>	if (UNSAFE.compareAndSwapObject
721	>	(q, (i << ASHIFT) + ABASE, t, null)) {
722	>	queueTop = s;           // or putOrderedInt
723	>	t.doExec();
724	>	}
725	>	}
726	>	return true;
727	>	}
728
729	<	outer:for (int d = 0; d < MAX_HELP_DEPTH; ++d) { // chain length
729	>	/**
730	>	* Tries to locate and execute tasks for a stealer of the given
731	>	* task, or in turn one of its stealers, Traces
732	>	* currentSteal->currentJoin links looking for a thread working on
733	>	* a descendant of the given task and with a non-empty queue to
734	>	* steal back and execute tasks from.  The implementation is very
735	>	* branchy to cope with potential inconsistencies or loops
736	>	* encountering chains that are stale, unknown, or of length
737	>	* greater than MAX_HELP links.  All of these cases are dealt with
738	>	* by just retrying by caller.
739	>	*
740	>	* @param joinMe the task to join
741	>	* @param canSteal true if local queue is empty
742	>	* @return true if ran a task
743	>	*/
744	>	private boolean helpJoinTask(ForkJoinTask<?> joinMe) {
745	>	boolean helped = false;
746	>	int m = pool.scanGuard & SMASK;
747	>	ForkJoinWorkerThread[] ws = pool.workers;
748	>	if (ws != null && ws.length > m && joinMe.status >= 0) {
749	>	int levels = MAX_HELP;              // remaining chain length
750	>	ForkJoinTask<?> task = joinMe;      // base of chain
751	>	outer:for (ForkJoinWorkerThread thread = this;;) {
752		// Try to find v, the stealer of task, by first using hint
753	<	ForkJoinWorkerThread v = ws[thread.stealHint & (n - 1)];
753	>	ForkJoinWorkerThread v = ws[thread.stealHint & m];
754		if (v == null || v.currentSteal != task) {
755	<	for (int j = 0; ; ++j) {      // search array
756	<	if (j < n) {
757	<	ForkJoinTask<?> vs;
758	<	if ((v = ws[j]) != null &&
1026	<	(vs = v.currentSteal) != null) {
1027	<	if (joinMe.status < 0)
1028	<	break outer;
1029	<	if (vs == task) {
1030	<	if (task.status < 0)
1031	<	break outer; // stale
1032	<	thread.stealHint = j;
1033	<	break;        // save hint for next time
1034	<	}
1035	<	}
755	>	for (int j = 0; ;) {        // search array
756	>	if ((v = ws[j]) != null && v.currentSteal == task) {
757	>	thread.stealHint = j;
758	>	break;              // save hint for next time
759		}
760	<	else
761	<	break outer;          // no stealer
760	>	if (++j > m)
761	>	break outer;        // can't find stealer
762		}
763		}
1041	–
764		// Try to help v, using specialized form of deqTask
765		for (;;) {
766	+	ForkJoinTask<?>[] q; int b, i;
767		if (joinMe.status < 0)
768		break outer;
769	<	int b = v.base;
770	<	ForkJoinTask<?>[] q = v.queue;
771	<	if (b == v.sp || q == null)
772	<	break;                    // empty
773	<	int i = (q.length - 1) & b;
1051	<	long u = (i << qShift) + qBase;
769	>	if ((b = v.queueBase) == v.queueTop ||
770	>	(q = v.queue) == null ||
771	>	(i = (q.length-1) & b) < 0)
772	>	break;                  // empty
773	>	long u = (i << ASHIFT) + ABASE;
774		ForkJoinTask<?> t = q[i];
775		if (task.status < 0)
776	<	break outer;              // stale
777	<	if (t != null &&
1056	<	(running ||
1057	<	(running = p.tryIncrementRunningCount())) &&
1058	<	v.base == b++ &&
776	>	break outer;            // stale
777	>	if (t != null && v.queueBase == b &&
778		UNSAFE.compareAndSwapObject(q, u, t, null)) {
779	<	if (t != joinMe && joinMe.status < 0) {
780	<	UNSAFE.putObjectVolatile(q, u, t);
781	<	break outer;          // joinMe cancelled; back out
782	<	}
783	<	v.base = b;
784	<	if (t.status >= 0) {
785	<	ForkJoinTask<?> ps = currentSteal;
1067	<	int pid = poolIndex;
1068	<	v.stealHint = pid;
1069	<	UNSAFE.putOrderedObject(this,
1070	<	currentStealOffset, t);
1071	<	t.quietlyExec();
1072	<	UNSAFE.putOrderedObject(this,
1073	<	currentStealOffset, ps);
1074	<	}
1075	<	}
1076	<	else if ((runState & TERMINATING) != 0) {
1077	<	joinMe.cancelIgnoringExceptions();
1078	<	break outer;
779	>	v.queueBase = b + 1;
780	>	v.stealHint = poolIndex;
781	>	ForkJoinTask<?> ps = currentSteal;
782	>	currentSteal = t;
783	>	t.doExec();
784	>	currentSteal = ps;
785	>	helped = true;
786		}
787		}
1081	–
788		// Try to descend to find v's stealer
789		ForkJoinTask<?> next = v.currentJoin;
790	<	if (task.status < 0 || next == null || next == task ||
791	<	joinMe.status < 0)
792	<	break;                 // done, stale, dead-end, or cyclic
793	<	task = next;
794	<	thread = v;
790	>	if (--levels > 0 && task.status >= 0 &&
791	>	next != null && next != task) {
792	>	task = next;
793	>	thread = v;
794	>	}
795	>	else
796	>	break;  // max levels, stale, dead-end, or cyclic
797	>	}
798	>	}
799	>	return helped;
800	>	}
801	>
802	>	/**
803	>	* Performs an uncommon case for joinTask: If task t is at base of
804	>	* some workers queue, steals and executes it.
805	>	*
806	>	* @param t the task
807	>	* @return t's status
808	>	*/
809	>	private int tryDeqAndExec(ForkJoinTask<?> t) {
810	>	int m = pool.scanGuard & SMASK;
811	>	ForkJoinWorkerThread[] ws = pool.workers;
812	>	if (ws != null && ws.length > m && t.status >= 0) {
813	>	for (int j = 0; j <= m; ++j) {
814	>	ForkJoinTask<?>[] q; int b, i;
815	>	ForkJoinWorkerThread v = ws[j];
816	>	if (v != null &&
817	>	(b = v.queueBase) != v.queueTop &&
818	>	(q = v.queue) != null &&
819	>	(i = (q.length - 1) & b) >= 0 &&
820	>	q[i] ==  t) {
821	>	long u = (i << ASHIFT) + ABASE;
822	>	if (v.queueBase == b &&
823	>	UNSAFE.compareAndSwapObject(q, u, t, null)) {
824	>	v.queueBase = b + 1;
825	>	v.stealHint = poolIndex;
826	>	ForkJoinTask<?> ps = currentSteal;
827	>	currentSteal = t;
828	>	t.doExec();
829	>	currentSteal = ps;
830	>	}
831	>	break;
832	>	}
833		}
834		}
835	<	return running;
835	>	return t.status;
836		}
837
838		/**
839	<	* Implements ForkJoinTask.getSurplusQueuedTaskCount().
840	<	* Returns an estimate of the number of tasks, offset by a
841	<	* function of number of idle workers.
839	>	* Implements ForkJoinTask.getSurplusQueuedTaskCount().  Returns
840	>	* an estimate of the number of tasks, offset by a function of
841	>	* number of idle workers.
842		*
843		* This method provides a cheap heuristic guide for task
844		* partitioning when programmers, frameworks, tools, or languages
#	Line 1130 | Line 874 | public class ForkJoinWorkerThread extend
874		* When all threads are active, it is on average OK to estimate
875		* surplus strictly locally. In steady-state, if one thread is
876		* maintaining say 2 surplus tasks, then so are others. So we can
877	<	* just use estimated queue length (although note that (sp - base)
878	<	* can be an overestimate because of stealers lagging increments
879	<	* of base).  However, this strategy alone leads to serious
880	<	* mis-estimates in some non-steady-state conditions (ramp-up,
881	<	* ramp-down, other stalls). We can detect many of these by
882	<	* further considering the number of "idle" threads, that are
877	>	* just use estimated queue length (although note that (queueTop -
878	>	* queueBase) can be an overestimate because of stealers lagging
879	>	* increments of queueBase).  However, this strategy alone leads
880	>	* to serious mis-estimates in some non-steady-state conditions
881	>	* (ramp-up, ramp-down, other stalls). We can detect many of these
882	>	* by further considering the number of "idle" threads, that are
883		* known to have zero queued tasks, so compensate by a factor of
884		* (#idle/#active) threads.
885		*/
886		final int getEstimatedSurplusTaskCount() {
887	<	return sp - base - pool.idlePerActive();
887	>	return queueTop - queueBase - pool.idlePerActive();
888		}
889
890		/**
891	<	* Runs tasks until {@code pool.isQuiescent()}.
891	>	* Runs tasks until {@code pool.isQuiescent()}. We piggyback on
892	>	* pool's active count ctl maintenance, but rather than blocking
893	>	* when tasks cannot be found, we rescan until all others cannot
894	>	* find tasks either. The bracketing by pool quiescerCounts
895	>	* updates suppresses pool auto-shutdown mechanics that could
896	>	* otherwise prematurely terminate the pool because all threads
897	>	* appear to be inactive.
898		*/
899		final void helpQuiescePool() {
900	+	boolean active = true;
901		ForkJoinTask<?> ps = currentSteal; // to restore below
902	+	ForkJoinPool p = pool;
903	+	p.addQuiescerCount(1);
904		for (;;) {
905	<	ForkJoinTask<?> t = pollLocalTask();
906	<	if (t != null || (t = scan()) != null)
907	<	t.quietlyExec();
905	>	ForkJoinWorkerThread[] ws = p.workers;
906	>	ForkJoinWorkerThread v = null;
907	>	int n;
908	>	if (queueTop != queueBase)
909	>	v = this;
910	>	else if (ws != null && (n = ws.length) > 1) {
911	>	ForkJoinWorkerThread w;
912	>	int r = nextSeed(); // cheap version of FJP.scan
913	>	int steps = n << 1;
914	>	for (int i = 0; i < steps; ++i) {
915	>	if ((w = ws[(i + r) & (n - 1)]) != null &&
916	>	w.queueBase != w.queueTop) {
917	>	v = w;
918	>	break;
919	>	}
920	>	}
921	>	}
922	>	if (v != null) {
923	>	ForkJoinTask<?> t;
924	>	if (!active) {
925	>	active = true;
926	>	p.addActiveCount(1);
927	>	}
928	>	if ((t = (v != this) ? v.deqTask() :
929	>	locallyFifo? locallyDeqTask() : popTask()) != null) {
930	>	currentSteal = t;
931	>	t.doExec();
932	>	currentSteal = ps;
933	>	}
934	>	}
935		else {
1156	–	ForkJoinPool p = pool;
1157	–	int a; // to inline CASes
936		if (active) {
937	<	if (!UNSAFE.compareAndSwapInt
938	<	(p, poolRunStateOffset, a = p.runState, a - 1))
1161	<	continue;   // retry later
1162	<	active = false; // inactivate
1163	<	UNSAFE.putOrderedObject(this, currentStealOffset, ps);
937	>	active = false;
938	>	p.addActiveCount(-1);
939		}
940		if (p.isQuiescent()) {
941	<	active = true; // re-activate
942	<	do {} while (!UNSAFE.compareAndSwapInt
943	<	(p, poolRunStateOffset, a = p.runState, a+1));
1169	<	return;
941	>	p.addActiveCount(1);
942	>	p.addQuiescerCount(-1);
943	>	break;
944		}
945		}
946		}
947		}
948
949		// Unsafe mechanics
950	<
951	<	private static final sun.misc.Unsafe UNSAFE = getUnsafe();
952	<	private static final long spOffset =
1179	<	objectFieldOffset("sp", ForkJoinWorkerThread.class);
1180	<	private static final long runStateOffset =
1181	<	objectFieldOffset("runState", ForkJoinWorkerThread.class);
1182	<	private static final long currentJoinOffset =
1183	<	objectFieldOffset("currentJoin", ForkJoinWorkerThread.class);
1184	<	private static final long currentStealOffset =
1185	<	objectFieldOffset("currentSteal", ForkJoinWorkerThread.class);
1186	<	private static final long qBase =
1187	<	UNSAFE.arrayBaseOffset(ForkJoinTask[].class);
1188	<	private static final long poolRunStateOffset = // to inline CAS
1189	<	objectFieldOffset("runState", ForkJoinPool.class);
1190	<
1191	<	private static final int qShift;
950	>	private static final sun.misc.Unsafe UNSAFE;
951	>	private static final long ABASE;
952	>	private static final int ASHIFT;
953
954		static {
955	<	int s = UNSAFE.arrayIndexScale(ForkJoinTask[].class);
1195	<	if ((s & (s-1)) != 0)
1196	<	throw new Error("data type scale not a power of two");
1197	<	qShift = 31 - Integer.numberOfLeadingZeros(s);
1198	<	MAXIMUM_QUEUE_CAPACITY = 1 << (31 - qShift);
1199	<	}
1200	<
1201	<	private static long objectFieldOffset(String field, Class<?> klazz) {
955	>	int s;
956		try {
957	<	return UNSAFE.objectFieldOffset(klazz.getDeclaredField(field));
958	<	} catch (NoSuchFieldException e) {
959	<	// Convert Exception to corresponding Error
960	<	NoSuchFieldError error = new NoSuchFieldError(field);
961	<	error.initCause(e);
962	<	throw error;
957	>	UNSAFE = getUnsafe();
958	>	Class a = ForkJoinTask[].class;
959	>	ABASE = UNSAFE.arrayBaseOffset(a);
960	>	s = UNSAFE.arrayIndexScale(a);
961	>	} catch (Exception e) {
962	>	throw new Error(e);
963		}
964	+	if ((s & (s-1)) != 0)
965	+	throw new Error("data type scale not a power of two");
966	+	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
967		}
968
969		/**

Diff Legend

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