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 1.26 by dl, Sun Aug 2 11:54:31 2009 UTC vs.
Revision 1.40 by dl, Wed Aug 11 18:45:12 2010 UTC

#	Line 8 | Line 8 | 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
15		/**
16		* A thread managed by a {@link ForkJoinPool}.  This class is
#	Line 25 | Line 27 | import java.util.Collection;
27		*/
28		public class ForkJoinWorkerThread extends Thread {
29		/*
30	<	* Algorithm overview:
30	>	* Overview:
31		*
32	<	* 1. Work-Stealing: Work-stealing queues are special forms of
33	<	* Deques that support only three of the four possible
34	<	* end-operations -- push, pop, and deq (aka steal), and only do
35	<	* so under the constraints that push and pop are called only from
36	<	* the owning thread, while deq may be called from other threads.
37	<	* (If you are unfamiliar with them, you probably want to read
38	<	* Herlihy and Shavit's book "The Art of Multiprocessor
39	<	* programming", chapter 16 describing these in more detail before
40	<	* proceeding.)  The main work-stealing queue design is roughly
41	<	* similar to "Dynamic Circular Work-Stealing Deque" by David
42	<	* Chase and Yossi Lev, SPAA 2005
43	<	* (http://research.sun.com/scalable/pubs/index.html).  The main
44	<	* difference ultimately stems from gc requirements that we null
45	<	* out taken slots as soon as we can, to maintain as small a
46	<	* footprint as possible even in programs generating huge numbers
47	<	* of tasks. To accomplish this, we shift the CAS arbitrating pop
48	<	* vs deq (steal) from being on the indices ("base" and "sp") to
49	<	* the slots themselves (mainly via method "casSlotNull()"). So,
50	<	* both a successful pop and deq mainly entail CAS'ing a non-null
51	<	* slot to null.  Because we rely on CASes of references, we do
52	<	* not need tag bits on base or sp.  They are simple ints as used
53	<	* in any circular array-based queue (see for example ArrayDeque).
54	<	* Updates to the indices must still be ordered in a way that
55	<	* guarantees that (sp - base) > 0 means the queue is empty, but
56	<	* otherwise may err on the side of possibly making the queue
57	<	* appear nonempty when a push, pop, or deq have not fully
58	<	* committed. Note that this means that the deq operation,
59	<	* considered individually, is not wait-free. One thief cannot
60	<	* successfully continue until another in-progress one (or, if
61	<	* previously empty, a push) completes.  However, in the
62	<	* aggregate, we ensure at least probabilistic non-blockingness. If
63	<	* an attempted steal fails, a thief always chooses a different
32	>	* ForkJoinWorkerThreads are managed by ForkJoinPools and perform
33	>	* ForkJoinTasks. This class includes bookkeeping in support of
34	>	* 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 activate if necessary before
78	<	* stealing (see below).
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
#	Line 72 | Line 83 | public class ForkJoinWorkerThread extend
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
86	>	* When a worker would otherwise be blocked waiting to join a
87	>	* task, it first tries a form of linear helping: Each worker
88	>	* records (in field currentSteal) the most recent task it stole
89	>	* from some other worker. Plus, it records (in field currentJoin)
90	>	* the task it is currently actively joining. Method joinTask uses
91	>	* these markers to try to find a worker to help (i.e., steal back
92	>	* a task from and execute it) that could hasten completion of the
93	>	* actively joined task. In essence, the joiner executes a task
94	>	* that would be on its own local deque had the to-be-joined task
95	>	* not been stolen. This may be seen as a conservative variant of
96	>	* the approach in Wagner & Calder "Leapfrogging: a portable
97	>	* technique for implementing efficient futures" SIGPLAN Notices,
98	>	* 1993 (http://portal.acm.org/citation.cfm?id=155354). It differs
99	>	* in that: (1) We only maintain dependency links across workers
100	>	* upon steals, rather than use per-task bookkeeping.  This may
101	>	* require a linear scan of workers array to locate stealers, but
102	>	* usually doesn't because stealers leave hints (that may become
103	>	* stale/wrong) of where to locate them. This isolates cost to
104	>	* when it is needed, rather than adding to per-task overhead.
105	>	* (2) It is "shallow", ignoring nesting and potentially cyclic
106	>	* mutual steals.  (3) It is intentionally racy: field currentJoin
107	>	* is updated only while actively joining, which means that we
108	>	* miss links in the chain during long-lived tasks, GC stalls etc
109	>	* (which is OK since blocking in such cases is usually a good
110	>	* idea).  (4) We bound the number of attempts to find work (see
111	>	* MAX_HELP_DEPTH) and fall back to suspending the worker and if
112	>	* necessary replacing it with a spare (see
113	>	* ForkJoinPool.tryAwaitJoin).
114	>	*
115	>	* Efficient implementation of these algorithms currently relies
116	>	* on an uncomfortable amount of "Unsafe" mechanics. To maintain
117		* correct orderings, reads and writes of variable base require
118	<	* volatile ordering.  Variable sp does not require volatile write
119	<	* but needs cheaper store-ordering on writes.  Because they are
120	<	* protected by volatile base reads, reads of the queue array and
121	<	* its slots do not need volatile load semantics, but writes (in
122	<	* push) require store order and CASes (in pop and deq) require
123	<	* (volatile) CAS semantics. Since these combinations aren't
124	<	* supported using ordinary volatiles, the only way to accomplish
125	<	* these efficiently is to use direct Unsafe calls. (Using external
126	<	* AtomicIntegers and AtomicReferenceArrays for the indices and
127	<	* array is significantly slower because of memory locality and
128	<	* indirection effects.) Further, performance on most platforms is
129	<	* very sensitive to placement and sizing of the (resizable) queue
130	<	* array.  Even though these queues don't usually become all that
131	<	* big, the initial size must be large enough to counteract cache
118	>	* volatile ordering.  Variable sp does not require volatile
119	>	* writes but still needs store-ordering, which we accomplish by
120	>	* pre-incrementing sp before filling the slot with an ordered
121	>	* store.  (Pre-incrementing also enables backouts used in
122	>	* joinTask.)  Because they are protected by volatile base reads,
123	>	* reads of the queue array and its slots by other threads do not
124	>	* need volatile load semantics, but writes (in push) require
125	>	* store order and CASes (in pop and deq) require (volatile) CAS
126	>	* semantics.  (Michael, Saraswat, and Vechev's algorithm has
127	>	* similar properties, but without support for nulling slots.)
128	>	* Since these combinations aren't supported using ordinary
129	>	* volatiles, the only way to accomplish these efficiently is to
130	>	* use direct Unsafe calls. (Using external AtomicIntegers and
131	>	* AtomicReferenceArrays for the indices and array is
132	>	* significantly slower because of memory locality and indirection
133	>	* effects.)
134	>	*
135	>	* Further, performance on most platforms is very sensitive to
136	>	* placement and sizing of the (resizable) queue array.  Even
137	>	* though these queues don't usually become all that big, the
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 currently initialized immediately after the thread
142	<	* gets the initial signal to start processing tasks.  However,
143	<	* all queue-related methods except pushTask are written in a way
144	<	* that allows them to instead be lazily allocated and/or disposed
145	<	* of when empty. All together, these low-level implementation
146	<	* choices produce as much as a factor of 4 performance
147	<	* improvement compared to naive implementations, and enable the
148	<	* processing of billions of tasks per second, sometimes at the
149	<	* expense of ugliness.
150	<	*
151	<	* 2. Run control: The primary run control is based on a global
152	<	* counter (activeCount) held by the pool. It uses an algorithm
153	<	* similar to that in Herlihy and Shavit section 17.6 to cause
154	<	* threads to eventually block when all threads declare they are
155	<	* inactive. (See variable "scans".)  For this to work, threads
156	<	* must be declared active when executing tasks, and before
157	<	* stealing a task. They must be inactive before blocking on the
158	<	* Pool Barrier (awaiting a new submission or other Pool
159	<	* event). In between, there is some free play which we take
160	<	* advantage of to avoid contention and rapid flickering of the
161	<	* global activeCount: If inactive, we activate only if a victim
162	<	* queue appears to be nonempty (see above).  Similarly, a thread
163	<	* tries to inactivate only after a full scan of other threads.
164	<	* The net effect is that contention on activeCount is rarely a
165	<	* measurable performance issue. (There are also a few other cases
166	<	* where we scan for work rather than retry/block upon
167	<	* contention.)
168	<	*
169	<	* 3. Selection control. We maintain policy of always choosing to
123	<	* run local tasks rather than stealing, and always trying to
124	<	* steal tasks before trying to run a new submission. All steals
125	<	* are currently performed in randomly-chosen deq-order. It may be
126	<	* worthwhile to bias these with locality / anti-locality
127	<	* information, but doing this well probably requires more
128	<	* lower-level information from JVMs than currently provided.
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.
146	>	*/
147	>
148	>	/**
149	>	* 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.
154	>	*/
155	>	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;
165	>
166	>	/**
167	>	* The wakeup interval (in nanoseconds) for the first worker
168	>	* suspended as spare.  On each wakeup not signalled by a
169	>	* resumption, it may ask the pool to reduce the number of spares.
170		*/
171	+	private static final long TRIM_RATE_NANOS = 200L * 1000L * 1000L;
172
173		/**
174		* Capacity of work-stealing queue array upon initialization.
175	<	* Must be a power of two. Initial size must be at least 2, but is
175	>	* Must be a power of two. Initial size must be at least 4, but is
176		* padded to minimize cache effects.
177		*/
178		private static final int INITIAL_QUEUE_CAPACITY = 1 << 13;
#	Line 150 | Line 192 | public class ForkJoinWorkerThread extend
192
193		/**
194		* The work-stealing queue array. Size must be a power of two.
195	<	* Initialized when thread starts, to improve memory locality.
195	>	* Initialized in onStart, to improve memory locality.
196		*/
197		private ForkJoinTask<?>[] queue;
198
199		/**
158	–	* Index (mod queue.length) of next queue slot to push to or pop
159	–	* from. It is written only by owner thread, via ordered store.
160	–	* Both sp and base are allowed to wrap around on overflow, but
161	–	* (sp - base) still estimates size.
162	–	*/
163	–	private volatile int sp;
164	–
165	–	/**
200		* Index (mod queue.length) of least valid queue slot, which is
201		* always the next position to steal from if nonempty.
202		*/
203		private volatile int base;
204
205		/**
206	<	* Activity status. When true, this worker is considered active.
207	<	* Must be false upon construction. It must be true when executing
208	<	* tasks, and BEFORE stealing a task. It must be false before
209	<	* calling pool.sync.
206	>	* Index (mod queue.length) of next queue slot to push to or pop
207	>	* from. It is written only by owner thread, and accessed by other
208	>	* threads only after reading (volatile) base.  Both sp and base
209	>	* are allowed to wrap around on overflow, but (sp - base) still
210	>	* estimates size.
211		*/
212	<	private boolean active;
212	>	private int sp;
213
214		/**
215	<	* Run state of this worker. Supports simple versions of the usual
216	<	* shutdown/shutdownNow control.
215	>	* The index of most recent stealer, used as a hint to avoid
216	>	* traversal in method helpJoinTask. This is only a hint because a
217	>	* worker might have had multiple steals and this only holds one
218	>	* of them (usually the most current). Declared non-volatile,
219	>	* relying on other prevailing sync to keep reasonably current.
220	>	*/
221	>	private int stealHint;
222	>
223	>	/**
224	>	* Run state of this worker. In addition to the usual run levels,
225	>	* tracks if this worker is suspended as a spare, and if it was
226	>	* killed (trimmed) while suspended. However, "active" status is
227	>	* maintained separately.
228		*/
229		private volatile int runState;
230
231	+	private static final int TERMINATING = 0x01;
232	+	private static final int TERMINATED  = 0x02;
233	+	private static final int SUSPENDED   = 0x04; // inactive spare
234	+	private static final int TRIMMED     = 0x08; // killed while suspended
235	+
236	+	/**
237	+	* Number of steals, transferred and reset in pool callbacks pool
238	+	* when idle Accessed directly by pool.
239	+	*/
240	+	int stealCount;
241	+
242		/**
243		* Seed for random number generator for choosing steal victims.
244	<	* Uses Marsaglia xorshift. Must be nonzero upon initialization.
244	>	* Uses Marsaglia xorshift. Must be initialized as nonzero.
245		*/
246		private int seed;
247
248		/**
249	<	* Number of steals, transferred to pool when idle
249	>	* Activity status. When true, this worker is considered active.
250	>	* Accessed directly by pool.  Must be false upon construction.
251	>	*/
252	>	boolean active;
253	>
254	>	/**
255	>	* True if use local fifo, not default lifo, for local polling.
256	>	* Shadows value from ForkJoinPool.
257		*/
258	<	private int stealCount;
258	>	private final boolean locallyFifo;
259
260		/**
261		* Index of this worker in pool array. Set once by pool before
262	<	* running, and accessed directly by pool during cleanup etc.
262	>	* running, and accessed directly by pool to locate this worker in
263	>	* its workers array.
264		*/
265		int poolIndex;
266
267		/**
268	<	* The last barrier event waited for. Accessed in pool callback
269	<	* methods, but only by current thread.
268	>	* The last pool event waited for. Accessed only by pool in
269	>	* callback methods invoked within this thread.
270	>	*/
271	>	int lastEventCount;
272	>
273	>	/**
274	>	* Encoded index and event count of next event waiter. Used only
275	>	* by ForkJoinPool for managing event waiters.
276	>	*/
277	>	volatile long nextWaiter;
278	>
279	>	/**
280	>	* Number of times this thread suspended as spare
281	>	*/
282	>	int spareCount;
283	>
284	>	/**
285	>	* Encoded index and count of next spare waiter. Used only
286	>	* by ForkJoinPool for managing spares.
287		*/
288	<	long lastEventCount;
288	>	volatile int nextSpare;
289
290		/**
291	<	* True if use local fifo, not default lifo, for local polling
291	>	* The task currently being joined, set only when actively trying
292	>	* to helpStealer. Written only by current thread, but read by
293	>	* others.
294		*/
295	<	private boolean locallyFifo;
295	>	private volatile ForkJoinTask<?> currentJoin;
296	>
297	>	/**
298	>	* The task most recently stolen from another worker (or
299	>	* submission queue).  Not volatile because always read/written in
300	>	* presence of related volatiles in those cases where it matters.
301	>	*/
302	>	private ForkJoinTask<?> currentSteal;
303
304		/**
305		* Creates a ForkJoinWorkerThread operating in the given pool.
#	Line 217 | Line 308 | public class ForkJoinWorkerThread extend
308		* @throws NullPointerException if pool is null
309		*/
310		protected ForkJoinWorkerThread(ForkJoinPool pool) {
220	–	if (pool == null) throw new NullPointerException();
311		this.pool = pool;
312	<	// Note: poolIndex is set by pool during construction
313	<	// Remaining initialization is deferred to onStart
312	>	this.locallyFifo = pool.locallyFifo;
313	>	setDaemon(true);
314	>	// To avoid exposing construction details to subclasses,
315	>	// remaining initialization is in start() and onStart()
316	>	}
317	>
318	>	/**
319	>	* Performs additional initialization and starts this thread
320	>	*/
321	>	final void start(int poolIndex, UncaughtExceptionHandler ueh) {
322	>	this.poolIndex = poolIndex;
323	>	if (ueh != null)
324	>	setUncaughtExceptionHandler(ueh);
325	>	start();
326		}
327
328	<	// Public access methods
328	>	// Public/protected methods
329
330		/**
331		* Returns the pool hosting this thread.
#	Line 248 | Line 350 | public class ForkJoinWorkerThread extend
350		}
351
352		/**
353	<	* Establishes local first-in-first-out scheduling mode for forked
354	<	* tasks that are never joined.
355	<	*
356	<	* @param async if true, use locally FIFO scheduling
353	>	* Initializes internal state after construction but before
354	>	* processing any tasks. If you override this method, you must
355	>	* invoke super.onStart() at the beginning of the method.
356	>	* Initialization requires care: Most fields must have legal
357	>	* default values, to ensure that attempted accesses from other
358	>	* threads work correctly even before this thread starts
359	>	* processing tasks.
360		*/
361	<	void setAsyncMode(boolean async) {
362	<	locallyFifo = async;
363	<	}
259	<
260	<	// Runstate management
261	<
262	<	// Runstate values. Order matters
263	<	private static final int RUNNING     = 0;
264	<	private static final int SHUTDOWN    = 1;
265	<	private static final int TERMINATING = 2;
266	<	private static final int TERMINATED  = 3;
267	<
268	<	final boolean isShutdown()    { return runState >= SHUTDOWN;  }
269	<	final boolean isTerminating() { return runState >= TERMINATING;  }
270	<	final boolean isTerminated()  { return runState == TERMINATED; }
271	<	final boolean shutdown()      { return transitionRunStateTo(SHUTDOWN); }
272	<	final boolean shutdownNow()   { return transitionRunStateTo(TERMINATING); }
361	>	protected void onStart() {
362	>	int rs = seedGenerator.nextInt();
363	>	seed = rs == 0? 1 : rs; // seed must be nonzero
364
365	<	/**
366	<	* Transitions to at least the given state.
367	<	*
368	<	* @return {@code true} if not already at least at given state
278	<	*/
279	<	private boolean transitionRunStateTo(int state) {
280	<	for (;;) {
281	<	int s = runState;
282	<	if (s >= state)
283	<	return false;
284	<	if (UNSAFE.compareAndSwapInt(this, runStateOffset, s, state))
285	<	return true;
286	<	}
287	<	}
365	>	// Allocate name string and arrays in this thread
366	>	String pid = Integer.toString(pool.getPoolNumber());
367	>	String wid = Integer.toString(poolIndex);
368	>	setName("ForkJoinPool-" + pid + "-worker-" + wid);
369
370	<	/**
290	<	* Tries to set status to active; fails on contention.
291	<	*/
292	<	private boolean tryActivate() {
293	<	if (!active) {
294	<	if (!pool.tryIncrementActiveCount())
295	<	return false;
296	<	active = true;
297	<	}
298	<	return true;
370	>	queue = new ForkJoinTask<?>[INITIAL_QUEUE_CAPACITY];
371		}
372
373		/**
374	<	* Tries to set status to inactive; fails on contention.
374	>	* Performs cleanup associated with termination of this worker
375	>	* thread.  If you override this method, you must invoke
376	>	* {@code super.onTermination} at the end of the overridden method.
377	>	*
378	>	* @param exception the exception causing this thread to abort due
379	>	* to an unrecoverable error, or {@code null} if completed normally
380		*/
381	<	private boolean tryInactivate() {
382	<	if (active) {
383	<	if (!pool.tryDecrementActiveCount())
384	<	return false;
385	<	active = false;
381	>	protected void onTermination(Throwable exception) {
382	>	try {
383	>	cancelTasks();
384	>	while (active)              // force inactive
385	>	active = !pool.tryDecrementActiveCount();
386	>	setTerminated();
387	>	pool.workerTerminated(this);
388	>	} catch (Throwable ex) {        // Shouldn't ever happen
389	>	if (exception == null)      // but if so, at least rethrown
390	>	exception = ex;
391	>	} finally {
392	>	if (exception != null)
393	>	UNSAFE.throwException(exception);
394		}
310	–	return true;
311	–	}
312	–
313	–	/**
314	–	* Computes next value for random victim probe.  Scans don't
315	–	* require a very high quality generator, but also not a crummy
316	–	* one.  Marsaglia xor-shift is cheap and works well.
317	–	*/
318	–	private static int xorShift(int r) {
319	–	r ^= (r << 13);
320	–	r ^= (r >>> 17);
321	–	return r ^ (r << 5);
395		}
396
324	–	// Lifecycle methods
325	–
397		/**
398		* This method is required to be public, but should never be
399		* called explicitly. It performs the main run loop to execute
#	Line 332 | Line 403 | public class ForkJoinWorkerThread extend
403		Throwable exception = null;
404		try {
405		onStart();
335	–	pool.sync(this); // await first pool event
406		mainLoop();
407		} catch (Throwable ex) {
408		exception = ex;
#	Line 341 | Line 411 | public class ForkJoinWorkerThread extend
411		}
412		}
413
414	+	// helpers for run()
415	+
416		/**
417	<	* Executes tasks until shut down.
417	>	* Find and execute tasks and check status while running
418		*/
419		private void mainLoop() {
420	<	while (!isShutdown()) {
421	<	ForkJoinTask<?> t = pollTask();
422	<	if (t != null || (t = pollSubmission()) != null)
423	<	t.quietlyExec();
424	<	else if (tryInactivate())
425	<	pool.sync(this);
420	>	int misses = 0; // track consecutive times failed to find work; max 2
421	>	ForkJoinPool p = pool;
422	>	for (;;) {
423	>	p.preStep(this, misses);
424	>	if (runState != 0)
425	>	break;
426	>	misses = ((tryExecSteal() || tryExecSubmission()) ? 0 :
427	>	(misses < 2 ? misses + 1 : 2));
428		}
429		}
430
431		/**
432	<	* Initializes internal state after construction but before
433	<	* processing any tasks. If you override this method, you must
434	<	* invoke super.onStart() at the beginning of the method.
361	<	* Initialization requires care: Most fields must have legal
362	<	* default values, to ensure that attempted accesses from other
363	<	* threads work correctly even before this thread starts
364	<	* processing tasks.
432	>	* Try to steal a task and execute it
433	>	*
434	>	* @return true if ran a task
435		*/
436	<	protected void onStart() {
437	<	// Allocate while starting to improve chances of thread-local
438	<	// isolation
439	<	queue = new ForkJoinTask<?>[INITIAL_QUEUE_CAPACITY];
440	<	// Initial value of seed need not be especially random but
441	<	// should differ across workers and must be nonzero
442	<	int p = poolIndex + 1;
443	<	seed = p + (p << 8) + (p << 16) + (p << 24); // spread bits
436	>	private boolean tryExecSteal() {
437	>	ForkJoinTask<?> t;
438	>	if ((t  = scan()) != null) {
439	>	t.quietlyExec();
440	>	currentSteal = null;
441	>	if (sp != base)
442	>	execLocalTasks();
443	>	return true;
444	>	}
445	>	return false;
446		}
447
448		/**
449	<	* Performs cleanup associated with termination of this worker
378	<	* thread.  If you override this method, you must invoke
379	<	* {@code super.onTermination} at the end of the overridden method.
449	>	* If a submission exists, try to activate and run it;
450		*
451	<	* @param exception the exception causing this thread to abort due
382	<	* to an unrecoverable error, or {@code null} if completed normally
451	>	* @return true if ran a task
452		*/
453	<	protected void onTermination(Throwable exception) {
454	<	// Execute remaining local tasks unless aborting or terminating
455	<	while (exception == null &&  !pool.isTerminating() && base != sp) {
456	<	try {
457	<	ForkJoinTask<?> t = popTask();
458	<	if (t != null)
453	>	private boolean tryExecSubmission() {
454	>	ForkJoinPool p = pool;
455	>	while (p.hasQueuedSubmissions()) {
456	>	ForkJoinTask<?> t;
457	>	if (active || (active = p.tryIncrementActiveCount())) {
458	>	if ((t = p.pollSubmission()) != null) {
459	>	currentSteal = t;
460		t.quietlyExec();
461	<	} catch (Throwable ex) {
462	<	exception = ex;
461	>	currentSteal = null;
462	>	if (sp != base)
463	>	execLocalTasks();
464	>	return true;
465	>	}
466		}
467		}
468	<	// Cancel other tasks, transition status, notify pool, and
396	<	// propagate exception to uncaught exception handler
397	<	try {
398	<	do {} while (!tryInactivate()); // ensure inactive
399	<	cancelTasks();
400	<	runState = TERMINATED;
401	<	pool.workerTerminated(this);
402	<	} catch (Throwable ex) {        // Shouldn't ever happen
403	<	if (exception == null)      // but if so, at least rethrown
404	<	exception = ex;
405	<	} finally {
406	<	if (exception != null)
407	<	ForkJoinTask.rethrowException(exception);
408	<	}
468	>	return false;
469		}
470
411	–	// Intrinsics-based support for queue operations.
412	–
471		/**
472	<	* Adds in store-order the given task at given slot of q to null.
473	<	* Caller must ensure q is non-null and index is in range.
472	>	* Runs local tasks until queue is empty or shut down.  Call only
473	>	* while active.
474		*/
475	<	private static void setSlot(ForkJoinTask<?>[] q, int i,
476	<	ForkJoinTask<?> t) {
477	<	UNSAFE.putOrderedObject(q, (i << qShift) + qBase, t);
475	>	private void execLocalTasks() {
476	>	while (runState == 0) {
477	>	ForkJoinTask<?> t = locallyFifo? locallyDeqTask() : popTask();
478	>	if (t != null)
479	>	t.quietlyExec();
480	>	else if (sp == base)
481	>	break;
482	>	}
483		}
484
485	+	/*
486	+	* Intrinsics-based atomic writes for queue slots. These are
487	+	* basically the same as methods in AtomicObjectArray, but
488	+	* specialized for (1) ForkJoinTask elements (2) requirement that
489	+	* nullness and bounds checks have already been performed by
490	+	* callers and (3) effective offsets are known not to overflow
491	+	* from int to long (because of MAXIMUM_QUEUE_CAPACITY). We don't
492	+	* need corresponding version for reads: plain array reads are OK
493	+	* because they protected by other volatile reads and are
494	+	* confirmed by CASes.
495	+	*
496	+	* Most uses don't actually call these methods, but instead contain
497	+	* inlined forms that enable more predictable optimization.  We
498	+	* don't define the version of write used in pushTask at all, but
499	+	* instead inline there a store-fenced array slot write.
500	+	*/
501	+
502		/**
503	<	* CAS given slot of q to null. Caller must ensure q is non-null
504	<	* and index is in range.
503	>	* CASes slot i of array q from t to null. Caller must ensure q is
504	>	* non-null and index is in range.
505		*/
506	<	private static boolean casSlotNull(ForkJoinTask<?>[] q, int i,
507	<	ForkJoinTask<?> t) {
506	>	private static final boolean casSlotNull(ForkJoinTask<?>[] q, int i,
507	>	ForkJoinTask<?> t) {
508		return UNSAFE.compareAndSwapObject(q, (i << qShift) + qBase, t, null);
509		}
510
511		/**
512	<	* Sets sp in store-order.
512	>	* Performs a volatile write of the given task at given slot of
513	>	* array q.  Caller must ensure q is non-null and index is in
514	>	* range. This method is used only during resets and backouts.
515		*/
516	<	private void storeSp(int s) {
517	<	UNSAFE.putOrderedInt(this, spOffset, s);
516	>	private static final void writeSlot(ForkJoinTask<?>[] q, int i,
517	>	ForkJoinTask<?> t) {
518	>	UNSAFE.putObjectVolatile(q, (i << qShift) + qBase, t);
519		}
520
521	<	// Main queue methods
521	>	// queue methods
522
523		/**
524	<	* Pushes a task. Called only by current thread.
524	>	* Pushes a task. Call only from this thread.
525		*
526		* @param t the task. Caller must ensure non-null.
527		*/
528		final void pushTask(ForkJoinTask<?> t) {
529		ForkJoinTask<?>[] q = queue;
530	<	int mask = q.length - 1;
531	<	int s = sp;
532	<	setSlot(q, s & mask, t);
533	<	storeSp(++s);
534	<	if ((s -= base) == 1)
535	<	pool.signalWork();
536	<	else if (s >= mask)
454	<	growQueue();
530	>	int mask = q.length - 1; // implicit assert q != null
531	>	int s = sp++;            // ok to increment sp before slot write
532	>	UNSAFE.putOrderedObject(q, ((s & mask) << qShift) + qBase, t);
533	>	if ((s -= base) == 0)
534	>	pool.signalWork();   // was empty
535	>	else if (s == mask)
536	>	growQueue();         // is full
537		}
538
539		/**
540		* Tries to take a task from the base of the queue, failing if
541	<	* either empty or contended.
541	>	* empty or contended. Note: Specializations of this code appear
542	>	* in locallyDeqTask and elsewhere.
543		*
544		* @return a task, or null if none or contended
545		*/
546		final ForkJoinTask<?> deqTask() {
547		ForkJoinTask<?> t;
548		ForkJoinTask<?>[] q;
549	<	int i;
467	<	int b;
549	>	int b, i;
550		if (sp != (b = base) &&
551		(q = queue) != null && // must read q after b
552	<	(t = q[i = (q.length - 1) & b]) != null &&
553	<	casSlotNull(q, i, t)) {
552	>	(t = q[i = (q.length - 1) & b]) != null && base == b &&
553	>	UNSAFE.compareAndSwapObject(q, (i << qShift) + qBase, t, null)) {
554		base = b + 1;
555		return t;
556		}
#	Line 476 | Line 558 | public class ForkJoinWorkerThread extend
558		}
559
560		/**
561	<	* Tries to take a task from the base of own queue, activating if
562	<	* necessary, failing only if empty. Called only by current thread.
561	>	* Tries to take a task from the base of own queue. Assumes active
562	>	* status.  Called only by current thread.
563		*
564		* @return a task, or null if none
565		*/
566		final ForkJoinTask<?> locallyDeqTask() {
567	<	int b;
568	<	while (sp != (b = base)) {
569	<	if (tryActivate()) {
570	<	ForkJoinTask<?>[] q = queue;
571	<	int i = (q.length - 1) & b;
572	<	ForkJoinTask<?> t = q[i];
573	<	if (t != null && casSlotNull(q, i, t)) {
567	>	ForkJoinTask<?>[] q = queue;
568	>	if (q != null) {
569	>	ForkJoinTask<?> t;
570	>	int b, i;
571	>	while (sp != (b = base)) {
572	>	if ((t = q[i = (q.length - 1) & b]) != null && base == b &&
573	>	UNSAFE.compareAndSwapObject(q, (i << qShift) + qBase,
574	>	t, null)) {
575		base = b + 1;
576		return t;
577		}
#	Line 498 | Line 581 | public class ForkJoinWorkerThread extend
581		}
582
583		/**
584	<	* Returns a popped task, or null if empty. Ensures active status
585	<	* if non-null. Called only by current thread.
584	>	* Returns a popped task, or null if empty. Assumes active status.
585	>	* Called only by current thread.
586		*/
587	<	final ForkJoinTask<?> popTask() {
588	<	int s = sp;
589	<	while (s != base) {
590	<	if (tryActivate()) {
591	<	ForkJoinTask<?>[] q = queue;
592	<	int mask = q.length - 1;
593	<	int i = (s - 1) & mask;
587	>	private ForkJoinTask<?> popTask() {
588	>	ForkJoinTask<?>[] q = queue;
589	>	if (q != null) {
590	>	int s;
591	>	while ((s = sp) != base) {
592	>	int i = (q.length - 1) & --s;
593	>	long u = (i << qShift) + qBase; // raw offset
594		ForkJoinTask<?> t = q[i];
595	<	if (t == null || !casSlotNull(q, i, t))
595	>	if (t == null)   // lost to stealer
596		break;
597	<	storeSp(s - 1);
598	<	return t;
597	>	if (UNSAFE.compareAndSwapObject(q, u, t, null)) {
598	>	sp = s; // putOrderedInt may encourage more timely write
599	>	// UNSAFE.putOrderedInt(this, spOffset, s);
600	>	return t;
601	>	}
602		}
603		}
604		return null;
605		}
606
607		/**
608	<	* Specialized version of popTask to pop only if
609	<	* topmost element is the given task. Called only
610	<	* by current thread while active.
608	>	* Specialized version of popTask to pop only if topmost element
609	>	* is the given task. Called only by current thread while
610	>	* active.
611		*
612		* @param t the task. Caller must ensure non-null.
613		*/
614		final boolean unpushTask(ForkJoinTask<?> t) {
615	+	int s;
616		ForkJoinTask<?>[] q = queue;
617	<	int mask = q.length - 1;
618	<	int s = sp - 1;
619	<	if (casSlotNull(q, s & mask, t)) {
620	<	storeSp(s);
617	>	if ((s = sp) != base && q != null &&
618	>	UNSAFE.compareAndSwapObject
619	>	(q, (((q.length - 1) & --s) << qShift) + qBase, t, null)) {
620	>	sp = s;
621	>	// UNSAFE.putOrderedInt(this, spOffset, s);
622		return true;
623		}
624		return false;
#	Line 570 | Line 658 | public class ForkJoinWorkerThread extend
658		ForkJoinTask<?> t = oldQ[oldIndex];
659		if (t != null && !casSlotNull(oldQ, oldIndex, t))
660		t = null;
661	<	setSlot(newQ, b & newMask, t);
661	>	writeSlot(newQ, b & newMask, t);
662		} while (++b != bf);
663		pool.signalWork();
664		}
665
666		/**
667	+	* Computes next value for random victim probe in scan().  Scans
668	+	* don't require a very high quality generator, but also not a
669	+	* crummy one.  Marsaglia xor-shift is cheap and works well enough.
670	+	* Note: This is manually inlined in scan()
671	+	*/
672	+	private static final int xorShift(int r) {
673	+	r ^= r << 13;
674	+	r ^= r >>> 17;
675	+	return r ^ (r << 5);
676	+	}
677	+
678	+	/**
679		* Tries to steal a task from another worker. Starts at a random
680		* index of workers array, and probes workers until finding one
681		* with non-empty queue or finding that all are empty.  It
682		* randomly selects the first n probes. If these are empty, it
683	<	* resorts to a full circular traversal, which is necessary to
684	<	* accurately set active status by caller. Also restarts if pool
685	<	* events occurred since last scan, which forces refresh of
686	<	* workers array, in case barrier was associated with resize.
683	>	* resorts to a circular sweep, which is necessary to accurately
684	>	* set active status. (The circular sweep uses steps of
685	>	* approximately half the array size plus 1, to avoid bias
686	>	* stemming from leftmost packing of the array in ForkJoinPool.)
687		*
688		* This method must be both fast and quiet -- usually avoiding
689		* memory accesses that could disrupt cache sharing etc other than
690	<	* those needed to check for and take tasks. This accounts for,
691	<	* among other things, updating random seed in place without
692	<	* storing it until exit.
690	>	* those needed to check for and take tasks (or to activate if not
691	>	* already active). This accounts for, among other things,
692	>	* updating random seed in place without storing it until exit.
693		*
694		* @return a task, or null if none found
695		*/
696		private ForkJoinTask<?> scan() {
697	<	ForkJoinTask<?> t = null;
698	<	int r = seed;                    // extract once to keep scan quiet
699	<	ForkJoinWorkerThread[] ws;       // refreshed on outer loop
700	<	int mask;                        // must be power 2 minus 1 and > 0
701	<	outer:do {
702	<	if ((ws = pool.workers) != null && (mask = ws.length - 1) > 0) {
703	<	int idx = r;
704	<	int probes = ~mask;      // use random index while negative
705	<	for (;;) {
706	<	r = xorShift(r);     // update random seed
707	<	ForkJoinWorkerThread v = ws[mask & idx];
708	<	if (v == null || v.sp == v.base) {
709	<	if (probes <= mask)
710	<	idx = (probes++ < 0) ? r : (idx + 1);
711	<	else
712	<	break;
697	>	ForkJoinPool p = pool;
698	>	ForkJoinWorkerThread[] ws;        // worker array
699	>	int n;                            // upper bound of #workers
700	>	if ((ws = p.workers) != null && (n = ws.length) > 1) {
701	>	boolean canSteal = active;    // shadow active status
702	>	int r = seed;                 // extract seed once
703	>	int mask = n - 1;
704	>	int j = -n;                   // loop counter
705	>	int k = r;                    // worker index, random if j < 0
706	>	for (;;) {
707	>	ForkJoinWorkerThread v = ws[k & mask];
708	>	r ^= r << 13; r ^= r >>> 17; r ^= r << 5; // inline xorshift
709	>	if (v != null && v.base != v.sp) {
710	>	ForkJoinTask<?>[] q; int b;
711	>	if ((canSteal ||       // ensure active status
712	>	(canSteal = active = p.tryIncrementActiveCount())) &&
713	>	(q = v.queue) != null && (b = v.base) != v.sp) {
714	>	int i = (q.length - 1) & b;
715	>	long u = (i << qShift) + qBase; // raw offset
716	>	ForkJoinTask<?> t = q[i];
717	>	if (v.base == b && t != null &&
718	>	UNSAFE.compareAndSwapObject(q, u, t, null)) {
719	>	int pid = poolIndex;
720	>	currentSteal = t;
721	>	v.stealHint = pid;
722	>	v.base = b + 1;
723	>	seed = r;
724	>	++stealCount;
725	>	return t;
726	>	}
727		}
728	<	else if (!tryActivate() || (t = v.deqTask()) == null)
729	<	continue outer;  // restart on contention
616	<	else
617	<	break outer;
728	>	j = -n;
729	>	k = r;                // restart on contention
730		}
731	+	else if (++j <= 0)
732	+	k = r;
733	+	else if (j <= n)
734	+	k += (n >>> 1) | 1;
735	+	else
736	+	break;
737		}
738	<	} while (pool.hasNewSyncEvent(this)); // retry on pool events
739	<	seed = r;
622	<	return t;
738	>	}
739	>	return null;
740		}
741
742	+	// Run State management
743	+
744	+	// status check methods used mainly by ForkJoinPool
745	+	final boolean isRunning()     { return runState == 0; }
746	+	final boolean isTerminating() { return (runState & TERMINATING) != 0; }
747	+	final boolean isTerminated()  { return (runState & TERMINATED) != 0; }
748	+	final boolean isSuspended()   { return (runState & SUSPENDED) != 0; }
749	+	final boolean isTrimmed()     { return (runState & TRIMMED) != 0; }
750	+
751		/**
752	<	* Gets and removes a local or stolen task.
752	>	* Sets state to TERMINATING, also, unless "quiet", unparking if
753	>	* not already terminated
754		*
755	<	* @return a task, if available
755	>	* @param quiet don't unpark (used for faster status updates on
756	>	* pool termination)
757		*/
758	<	final ForkJoinTask<?> pollTask() {
759	<	ForkJoinTask<?> t = locallyFifo ? locallyDeqTask() : popTask();
760	<	if (t == null && (t = scan()) != null)
761	<	++stealCount;
762	<	return t;
758	>	final void shutdown(boolean quiet) {
759	>	for (;;) {
760	>	int s = runState;
761	>	if ((s & (TERMINATING|TERMINATED)) != 0)
762	>	break;
763	>	if ((s & SUSPENDED) != 0) { // kill and wakeup if suspended
764	>	if (UNSAFE.compareAndSwapInt(this, runStateOffset, s,
765	>	(s & ~SUSPENDED) |
766	>	(TRIMMED|TERMINATING)))
767	>	break;
768	>	}
769	>	else if (UNSAFE.compareAndSwapInt(this, runStateOffset, s,
770	>	s | TERMINATING))
771	>	break;
772	>	}
773	>	if (!quiet && (runState & TERMINATED) != 0)
774	>	LockSupport.unpark(this);
775		}
776
777		/**
778	<	* Gets a local task.
639	<	*
640	<	* @return a task, if available
778	>	* Sets state to TERMINATED. Called only by onTermination()
779		*/
780	<	final ForkJoinTask<?> pollLocalTask() {
781	<	return locallyFifo ? locallyDeqTask() : popTask();
780	>	private void setTerminated() {
781	>	int s;
782	>	do {} while (!UNSAFE.compareAndSwapInt(this, runStateOffset,
783	>	s = runState,
784	>	s | (TERMINATING|TERMINATED)));
785		}
786
787		/**
788	<	* Returns a pool submission, if one exists, activating first.
788	>	* If suspended, tries to set status to unsuspended and unparks.
789		*
790	<	* @return a submission, if available
790	>	* @return true if successful
791		*/
792	<	private ForkJoinTask<?> pollSubmission() {
792	>	final boolean tryUnsuspend() {
793	>	int s;
794	>	while (((s = runState) & SUSPENDED) != 0) {
795	>	if (UNSAFE.compareAndSwapInt(this, runStateOffset, s,
796	>	s & ~SUSPENDED))
797	>	return true;
798	>	}
799	>	return false;
800	>	}
801	>
802	>	/**
803	>	* Sets suspended status and blocks as spare until resumed
804	>	* or shutdown.
805	>	* @returns true if still running on exit
806	>	*/
807	>	final boolean suspendAsSpare() {
808	>	lastEventCount = 0;         // reset upon resume
809	>	for (;;) {                  // set suspended unless terminating
810	>	int s = runState;
811	>	if ((s & TERMINATING) != 0) { // must kill
812	>	if (UNSAFE.compareAndSwapInt(this, runStateOffset, s,
813	>	s | (TRIMMED | TERMINATING)))
814	>	return false;
815	>	}
816	>	else if (UNSAFE.compareAndSwapInt(this, runStateOffset, s,
817	>	s | SUSPENDED))
818	>	break;
819	>	}
820		ForkJoinPool p = pool;
821	<	while (p.hasQueuedSubmissions()) {
822	<	ForkJoinTask<?> t;
823	<	if (tryActivate() && (t = p.pollSubmission()) != null)
824	<	return t;
821	>	p.pushSpare(this);
822	>	while ((runState & SUSPENDED) != 0) {
823	>	if (!p.tryAccumulateStealCount(this))
824	>	continue;
825	>	interrupted();          // clear/ignore interrupts
826	>	if ((runState & SUSPENDED) == 0)
827	>	break;
828	>	if (nextSpare != 0)     // untimed
829	>	LockSupport.park(this);
830	>	else {
831	>	long startTime = System.nanoTime();
832	>	LockSupport.parkNanos(this, TRIM_RATE_NANOS);
833	>	if ((runState & SUSPENDED) == 0)
834	>	break;
835	>	long now = System.nanoTime();
836	>	if (now - startTime >= TRIM_RATE_NANOS)
837	>	pool.tryTrimSpare(now);
838	>	}
839		}
840	<	return null;
840	>	return runState == 0;
841		}
842
843	<	// Methods accessed only by Pool
843	>	// Misc support methods for ForkJoinPool
844	>
845	>	/**
846	>	* Returns an estimate of the number of tasks in the queue.  Also
847	>	* used by ForkJoinTask.
848	>	*/
849	>	final int getQueueSize() {
850	>	int n; // external calls must read base first
851	>	return (n = -base + sp) <= 0 ? 0 : n;
852	>	}
853
854		/**
855		* Removes and cancels all tasks in queue.  Can be called from any
856		* thread.
857		*/
858		final void cancelTasks() {
859	<	ForkJoinTask<?> t;
860	<	while (base != sp && (t = deqTask()) != null)
861	<	t.cancelIgnoringExceptions();
859	>	ForkJoinTask<?> cj = currentJoin; // try to cancel ongoing tasks
860	>	if (cj != null) {
861	>	currentJoin = null;
862	>	cj.cancelIgnoringExceptions();
863	>	try {
864	>	this.interrupt(); // awaken wait
865	>	} catch (SecurityException ignore) {
866	>	}
867	>	}
868	>	ForkJoinTask<?> cs = currentSteal;
869	>	if (cs != null) {
870	>	currentSteal = null;
871	>	cs.cancelIgnoringExceptions();
872	>	}
873	>	while (base != sp) {
874	>	ForkJoinTask<?> t = deqTask();
875	>	if (t != null)
876	>	t.cancelIgnoringExceptions();
877	>	}
878		}
879
880		/**
#	Line 677 | Line 884 | public class ForkJoinWorkerThread extend
884		*/
885		final int drainTasksTo(Collection<? super ForkJoinTask<?>> c) {
886		int n = 0;
887	<	ForkJoinTask<?> t;
888	<	while (base != sp && (t = deqTask()) != null) {
889	<	c.add(t);
890	<	++n;
887	>	while (base != sp) {
888	>	ForkJoinTask<?> t = deqTask();
889	>	if (t != null) {
890	>	c.add(t);
891	>	++n;
892	>	}
893		}
894		return n;
895		}
896
897	+	// Support methods for ForkJoinTask
898	+
899		/**
900	<	* Gets and clears steal count for accumulation by pool.  Called
901	<	* only when known to be idle (in pool.sync and termination).
900	>	* Gets and removes a local task.
901	>	*
902	>	* @return a task, if available
903		*/
904	<	final int getAndClearStealCount() {
905	<	int sc = stealCount;
906	<	stealCount = 0;
907	<	return sc;
904	>	final ForkJoinTask<?> pollLocalTask() {
905	>	while (sp != base) {
906	>	if (active || (active = pool.tryIncrementActiveCount()))
907	>	return locallyFifo? locallyDeqTask() : popTask();
908	>	}
909	>	return null;
910		}
911
912		/**
913	<	* Returns {@code true} if at least one worker in the given array
700	<	* appears to have at least one queued task.
913	>	* Gets and removes a local or stolen task.
914		*
915	<	* @param ws array of workers
915	>	* @return a task, if available
916		*/
917	<	static boolean hasQueuedTasks(ForkJoinWorkerThread[] ws) {
918	<	if (ws != null) {
919	<	int len = ws.length;
920	<	for (int j = 0; j < 2; ++j) { // need two passes for clean sweep
921	<	for (int i = 0; i < len; ++i) {
709	<	ForkJoinWorkerThread w = ws[i];
710	<	if (w != null && w.sp != w.base)
711	<	return true;
712	<	}
713	<	}
917	>	final ForkJoinTask<?> pollTask() {
918	>	ForkJoinTask<?> t = pollLocalTask();
919	>	if (t == null) {
920	>	t = scan();
921	>	currentSteal = null; // cannot retain/track/help
922		}
923	<	return false;
923	>	return t;
924		}
925
718	–	// Support methods for ForkJoinTask
719	–
926		/**
927	<	* Returns an estimate of the number of tasks in the queue.
927	>	* Possibly runs some tasks and/or blocks, until task is done.
928	>	*
929	>	* @param joinMe the task to join
930		*/
931	<	final int getQueueSize() {
932	<	// suppress momentarily negative values
933	<	return Math.max(0, sp - base);
931	>	final void joinTask(ForkJoinTask<?> joinMe) {
932	>	// currentJoin only written by this thread; only need ordered store
933	>	ForkJoinTask<?> prevJoin = currentJoin;
934	>	UNSAFE.putOrderedObject(this, currentJoinOffset, joinMe);
935	>	if (sp != base)
936	>	localHelpJoinTask(joinMe);
937	>	if (joinMe.status >= 0)
938	>	pool.awaitJoin(joinMe, this);
939	>	UNSAFE.putOrderedObject(this, currentJoinOffset, prevJoin);
940		}
941
942		/**
943	<	* Returns an estimate of the number of tasks, offset by a
944	<	* function of number of idle workers.
943	>	* Run tasks in local queue until given task is done.
944	>	*
945	>	* @param joinMe the task to join
946		*/
947	<	final int getEstimatedSurplusTaskCount() {
948	<	// The halving approximates weighting idle vs non-idle workers
949	<	return (sp - base) - (pool.getIdleThreadCount() >>> 1);
947	>	private void localHelpJoinTask(ForkJoinTask<?> joinMe) {
948	>	int s;
949	>	ForkJoinTask<?>[] q;
950	>	while (joinMe.status >= 0 && (s = sp) != base && (q = queue) != null) {
951	>	int i = (q.length - 1) & --s;
952	>	long u = (i << qShift) + qBase; // raw offset
953	>	ForkJoinTask<?> t = q[i];
954	>	if (t == null)  // lost to a stealer
955	>	break;
956	>	if (UNSAFE.compareAndSwapObject(q, u, t, null)) {
957	>	/*
958	>	* This recheck (and similarly in helpJoinTask)
959	>	* handles cases where joinMe is independently
960	>	* cancelled or forced even though there is other work
961	>	* available. Back out of the pop by putting t back
962	>	* into slot before we commit by writing sp.
963	>	*/
964	>	if (joinMe.status < 0) {
965	>	UNSAFE.putObjectVolatile(q, u, t);
966	>	break;
967	>	}
968	>	sp = s;
969	>	// UNSAFE.putOrderedInt(this, spOffset, s);
970	>	t.quietlyExec();
971	>	}
972	>	}
973		}
974
975		/**
976	<	* Scans, returning early if joinMe done.
977	<	*/
978	<	final ForkJoinTask<?> scanWhileJoining(ForkJoinTask<?> joinMe) {
979	<	ForkJoinTask<?> t = pollTask();
980	<	if (t != null && joinMe.status < 0 && sp == base) {
981	<	pushTask(t); // unsteal if done and this task would be stealable
982	<	t = null;
976	>	* Tries to locate and help perform tasks for a stealer of the
977	>	* given task, or in turn one of its stealers.  Traces
978	>	* currentSteal->currentJoin links looking for a thread working on
979	>	* a descendant of the given task and with a non-empty queue to
980	>	* steal back and execute tasks from.
981	>	*
982	>	* The implementation is very branchy to cope with the potential
983	>	* inconsistencies or loops encountering chains that are stale,
984	>	* unknown, or of length greater than MAX_HELP_DEPTH links.  All
985	>	* of these cases are dealt with by just returning back to the
986	>	* caller, who is expected to retry if other join mechanisms also
987	>	* don't work out.
988	>	*
989	>	* @param joinMe the task to join
990	>	*/
991	>	final void helpJoinTask(ForkJoinTask<?> joinMe) {
992	>	ForkJoinWorkerThread[] ws = pool.workers;
993	>	int n; // need at least 2 workers
994	>	if (ws != null && (n = ws.length) > 1 && joinMe.status >= 0) {
995	>	ForkJoinTask<?> task = joinMe;        // base of chain
996	>	ForkJoinWorkerThread thread = this;   // thread with stolen task
997	>	for (int d = 0; d < MAX_HELP_DEPTH; ++d) { // chain length
998	>	// Try to find v, the stealer of task, by first using hint
999	>	ForkJoinWorkerThread v = ws[thread.stealHint & (n - 1)];
1000	>	if (v == null || v.currentSteal != task) {
1001	>	for (int j = 0; ; ++j) {      // search array
1002	>	if (j < n) {
1003	>	if ((v = ws[j]) != null) {
1004	>	if (task.status < 0)
1005	>	return;       // stale or done
1006	>	if (v.currentSteal == task) {
1007	>	thread.stealHint = j;
1008	>	break;        // save hint for next time
1009	>	}
1010	>	}
1011	>	}
1012	>	else
1013	>	return;               // no stealer
1014	>	}
1015	>	}
1016	>	// Try to help v, using specialized form of deqTask
1017	>	int b;
1018	>	ForkJoinTask<?>[] q;
1019	>	while ((b = v.base) != v.sp && (q = v.queue) != null) {
1020	>	int i = (q.length - 1) & b;
1021	>	long u = (i << qShift) + qBase;
1022	>	ForkJoinTask<?> t = q[i];
1023	>	if (task.status < 0)
1024	>	return;                   // stale or done
1025	>	if (v.base == b) {
1026	>	if (t == null)
1027	>	return;               // producer stalled
1028	>	if (UNSAFE.compareAndSwapObject(q, u, t, null)) {
1029	>	if (joinMe.status < 0) {
1030	>	UNSAFE.putObjectVolatile(q, u, t);
1031	>	return;           // back out on cancel
1032	>	}
1033	>	int pid = poolIndex;
1034	>	ForkJoinTask<?> prevSteal = currentSteal;
1035	>	currentSteal = t;
1036	>	v.stealHint = pid;
1037	>	v.base = b + 1;
1038	>	t.quietlyExec();
1039	>	currentSteal = prevSteal;
1040	>	}
1041	>	}
1042	>	if (joinMe.status < 0)
1043	>	return;
1044	>	}
1045	>	// Try to descend to find v's stealer
1046	>	ForkJoinTask<?> next = v.currentJoin;
1047	>	if (task.status < 0 || next == null || next == task ||
1048	>	joinMe.status < 0)
1049	>	return;
1050	>	task = next;
1051	>	thread = v;
1052	>	}
1053		}
1054	<	return t;
1054	>	}
1055	>
1056	>	/**
1057	>	* Returns an estimate of the number of tasks, offset by a
1058	>	* function of number of idle workers.
1059	>	*
1060	>	* This method provides a cheap heuristic guide for task
1061	>	* partitioning when programmers, frameworks, tools, or languages
1062	>	* have little or no idea about task granularity.  In essence by
1063	>	* offering this method, we ask users only about tradeoffs in
1064	>	* overhead vs expected throughput and its variance, rather than
1065	>	* how finely to partition tasks.
1066	>	*
1067	>	* In a steady state strict (tree-structured) computation, each
1068	>	* thread makes available for stealing enough tasks for other
1069	>	* threads to remain active. Inductively, if all threads play by
1070	>	* the same rules, each thread should make available only a
1071	>	* constant number of tasks.
1072	>	*
1073	>	* The minimum useful constant is just 1. But using a value of 1
1074	>	* would require immediate replenishment upon each steal to
1075	>	* maintain enough tasks, which is infeasible.  Further,
1076	>	* partitionings/granularities of offered tasks should minimize
1077	>	* steal rates, which in general means that threads nearer the top
1078	>	* of computation tree should generate more than those nearer the
1079	>	* bottom. In perfect steady state, each thread is at
1080	>	* approximately the same level of computation tree. However,
1081	>	* producing extra tasks amortizes the uncertainty of progress and
1082	>	* diffusion assumptions.
1083	>	*
1084	>	* So, users will want to use values larger, but not much larger
1085	>	* than 1 to both smooth over transient shortages and hedge
1086	>	* against uneven progress; as traded off against the cost of
1087	>	* extra task overhead. We leave the user to pick a threshold
1088	>	* value to compare with the results of this call to guide
1089	>	* decisions, but recommend values such as 3.
1090	>	*
1091	>	* When all threads are active, it is on average OK to estimate
1092	>	* surplus strictly locally. In steady-state, if one thread is
1093	>	* maintaining say 2 surplus tasks, then so are others. So we can
1094	>	* just use estimated queue length (although note that (sp - base)
1095	>	* can be an overestimate because of stealers lagging increments
1096	>	* of base).  However, this strategy alone leads to serious
1097	>	* mis-estimates in some non-steady-state conditions (ramp-up,
1098	>	* ramp-down, other stalls). We can detect many of these by
1099	>	* further considering the number of "idle" threads, that are
1100	>	* known to have zero queued tasks, so compensate by a factor of
1101	>	* (#idle/#active) threads.
1102	>	*/
1103	>	final int getEstimatedSurplusTaskCount() {
1104	>	return sp - base - pool.idlePerActive();
1105		}
1106
1107		/**
#	Line 751 | Line 1109 | public class ForkJoinWorkerThread extend
1109		*/
1110		final void helpQuiescePool() {
1111		for (;;) {
1112	<	ForkJoinTask<?> t = pollTask();
1113	<	if (t != null)
1112	>	ForkJoinTask<?> t = pollLocalTask();
1113	>	if (t != null || (t = scan()) != null) {
1114		t.quietlyExec();
1115	<	else if (tryInactivate() && pool.isQuiescent())
1116	<	break;
1115	>	currentSteal = null;
1116	>	}
1117	>	else {
1118	>	ForkJoinPool p = pool;
1119	>	if (active) {
1120	>	if (!p.tryDecrementActiveCount())
1121	>	continue;   // retry later
1122	>	active = false; // inactivate
1123	>	}
1124	>	if (p.isQuiescent()) {
1125	>	active = true; // re-activate
1126	>	do {} while (!p.tryIncrementActiveCount());
1127	>	return;
1128	>	}
1129	>	}
1130		}
760	–	do {} while (!tryActivate()); // re-activate on exit
1131		}
1132
1133		// Unsafe mechanics
#	Line 767 | Line 1137 | public class ForkJoinWorkerThread extend
1137		objectFieldOffset("sp", ForkJoinWorkerThread.class);
1138		private static final long runStateOffset =
1139		objectFieldOffset("runState", ForkJoinWorkerThread.class);
1140	<	private static final long qBase;
1140	>	private static final long currentJoinOffset =
1141	>	objectFieldOffset("currentJoin", ForkJoinWorkerThread.class);
1142	>	private static final long currentStealOffset =
1143	>	objectFieldOffset("currentSteal", ForkJoinWorkerThread.class);
1144	>	private static final long qBase =
1145	>	UNSAFE.arrayBaseOffset(ForkJoinTask[].class);
1146	>
1147		private static final int qShift;
1148
1149		static {
774	–	qBase = UNSAFE.arrayBaseOffset(ForkJoinTask[].class);
1150		int s = UNSAFE.arrayIndexScale(ForkJoinTask[].class);
1151		if ((s & (s-1)) != 0)
1152		throw new Error("data type scale not a power of two");

Diff Legend

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