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 1.35 by dl, Wed Jul 7 19:52:32 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
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 stolen) the most recent task it stole
89	>	* from some other worker. Plus, it records (in field joining) the
90	>	* 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 maintain per-task bookkeeping.  This
101	>	* requires a linear scan of workers array to locate stealers,
102	>	* which isolates cost to when it is needed, rather than adding to
103	>	* per-task overhead.  (2) It is "shallow", ignoring nesting and
104	>	* potentially cyclic mutual steals.  (3) It is intentionally
105	>	* racy: field joining is updated only while actively joining,
106	>	* which means that we could miss links in the chain during
107	>	* long-lived tasks, GC stalls etc.  (4) We fall back to
108	>	* suspending the worker and if necessary replacing it with a
109	>	* spare (see ForkJoinPool.tryAwaitJoin).
110	>	*
111	>	* Efficient implementation of these algorithms currently relies on
112		* an uncomfortable amount of "Unsafe" mechanics. To maintain
113		* correct orderings, reads and writes of variable base require
114	<	* volatile ordering.  Variable sp does not require volatile write
115	<	* but needs cheaper store-ordering on writes.  Because they are
116	<	* protected by volatile base reads, reads of the queue array and
117	<	* its slots do not need volatile load semantics, but writes (in
118	<	* push) require store order and CASes (in pop and deq) require
119	<	* (volatile) CAS semantics. Since these combinations aren't
120	<	* supported using ordinary volatiles, the only way to accomplish
121	<	* these efficiently is to use direct Unsafe calls. (Using external
114	>	* volatile ordering.  Variable sp does not require volatile
115	>	* writes but still needs store-ordering, which we accomplish by
116	>	* pre-incrementing sp before filling the slot with an ordered
117	>	* store.  (Pre-incrementing also enables backouts used in
118	>	* scanWhileJoining.)  Because they are protected by volatile base
119	>	* reads, reads of the queue array and its slots by other threads
120	>	* do not need volatile load semantics, but writes (in push)
121	>	* require store order and CASes (in pop and deq) require
122	>	* (volatile) CAS semantics.  (Michael, Saraswat, and Vechev's
123	>	* algorithm has similar properties, but without support for
124	>	* nulling slots.)  Since these combinations aren't supported
125	>	* using ordinary volatiles, the only way to accomplish these
126	>	* efficiently is to use direct Unsafe calls. (Using external
127		* AtomicIntegers and AtomicReferenceArrays for the indices and
128		* array is significantly slower because of memory locality and
129	<	* indirection effects.) Further, performance on most platforms is
130	<	* very sensitive to placement and sizing of the (resizable) queue
131	<	* array.  Even though these queues don't usually become all that
132	<	* big, the initial size must be large enough to counteract cache
129	>	* indirection effects.)
130	>	*
131	>	* Further, performance on most platforms is very sensitive to
132	>	* placement and sizing of the (resizable) queue array.  Even
133	>	* though these queues don't usually become all that big, the
134	>	* initial size must be large enough to counteract cache
135		* contention effects across multiple queues (especially in the
136		* presence of GC cardmarking). Also, to improve thread-locality,
137	<	* queues are currently initialized immediately after the thread
138	<	* gets the initial signal to start processing tasks.  However,
139	<	* all queue-related methods except pushTask are written in a way
140	<	* that allows them to instead be lazily allocated and/or disposed
141	<	* of when empty. All together, these low-level implementation
142	<	* choices produce as much as a factor of 4 performance
143	<	* improvement compared to naive implementations, and enable the
144	<	* processing of billions of tasks per second, sometimes at the
145	<	* expense of ugliness.
146	<	*
147	<	* 2. Run control: The primary run control is based on a global
148	<	* counter (activeCount) held by the pool. It uses an algorithm
149	<	* similar to that in Herlihy and Shavit section 17.6 to cause
150	<	* threads to eventually block when all threads declare they are
151	<	* inactive. (See variable "scans".)  For this to work, threads
152	<	* must be declared active when executing tasks, and before
153	<	* stealing a task. They must be inactive before blocking on the
154	<	* Pool Barrier (awaiting a new submission or other Pool
155	<	* event). In between, there is some free play which we take
156	<	* advantage of to avoid contention and rapid flickering of the
157	<	* global activeCount: If inactive, we activate only if a victim
158	<	* queue appears to be nonempty (see above).  Similarly, a thread
159	<	* tries to inactivate only after a full scan of other threads.
117	<	* The net effect is that contention on activeCount is rarely a
118	<	* measurable performance issue. (There are also a few other cases
119	<	* where we scan for work rather than retry/block upon
120	<	* contention.)
121	<	*
122	<	* 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.
137	>	* queues are initialized after starting.  All together, these
138	>	* low-level implementation choices produce as much as a factor of
139	>	* 4 performance improvement compared to naive implementations,
140	>	* and enable the processing of billions of tasks per second,
141	>	* sometimes at the expense of ugliness.
142	>	*/
143	>
144	>	/**
145	>	* Generator for initial random seeds for random victim
146	>	* selection. This is used only to create initial seeds. Random
147	>	* steals use a cheaper xorshift generator per steal attempt. We
148	>	* expect only rare contention on seedGenerator, so just use a
149	>	* plain Random.
150	>	*/
151	>	private static final Random seedGenerator = new Random();
152	>
153	>	/**
154	>	* The timeout value for suspending spares. Spare workers that
155	>	* remain unsignalled for more than this time may be trimmed
156	>	* (killed and removed from pool).  Since our goal is to avoid
157	>	* long-term thread buildup, the exact value of timeout does not
158	>	* matter too much so long as it avoids most false-alarm timeouts
159	>	* under GC stalls or momentarily high system load.
160		*/
161	+	private static final long SPARE_KEEPALIVE_NANOS =
162	+	5L * 1000L * 1000L * 1000L; // 5 secs
163
164		/**
165		* Capacity of work-stealing queue array upon initialization.
166	<	* Must be a power of two. Initial size must be at least 2, but is
166	>	* Must be a power of two. Initial size must be at least 4, but is
167		* padded to minimize cache effects.
168		*/
169		private static final int INITIAL_QUEUE_CAPACITY = 1 << 13;
#	Line 149 | Line 182 | public class ForkJoinWorkerThread extend
182		final ForkJoinPool pool;
183
184		/**
185	<	* The work-stealing queue array. Size must be a power of two.
153	<	* Initialized when thread starts, to improve memory locality.
185	>	* The task most recently stolen from another worker
186		*/
187	<	private ForkJoinTask<?>[] queue;
187	>	private volatile ForkJoinTask<?> stolen;
188
189		/**
190	<	* Index (mod queue.length) of next queue slot to push to or pop
191	<	* 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.
190	>	* The task currently being joined, set only when actively
191	>	* trying to helpStealer.
192		*/
193	<	private volatile int sp;
193	>	private volatile ForkJoinTask<?> joining;
194	>
195	>	/**
196	>	* The work-stealing queue array. Size must be a power of two.
197	>	* Initialized in onStart, to improve memory locality.
198	>	*/
199	>	private ForkJoinTask<?>[] queue;
200
201		/**
202		* Index (mod queue.length) of least valid queue slot, which is
#	Line 169 | Line 205 | public class ForkJoinWorkerThread extend
205		private volatile int base;
206
207		/**
208	<	* Activity status. When true, this worker is considered active.
209	<	* Must be false upon construction. It must be true when executing
210	<	* tasks, and BEFORE stealing a task. It must be false before
211	<	* calling pool.sync.
212	<	*/
213	<	private boolean active;
208	>	* Index (mod queue.length) of next queue slot to push to or pop
209	>	* from. It is written only by owner thread, and accessed by other
210	>	* threads only after reading (volatile) base.  Both sp and base
211	>	* are allowed to wrap around on overflow, but (sp - base) still
212	>	* estimates size.
213	>	*/
214	>	private int sp;
215
216		/**
217	<	* Run state of this worker. Supports simple versions of the usual
218	<	* shutdown/shutdownNow control.
217	>	* Run state of this worker. In addition to the usual run levels,
218	>	* tracks if this worker is suspended as a spare, and if it was
219	>	* killed (trimmed) while suspended. However, "active" status is
220	>	* maintained separately.
221		*/
222		private volatile int runState;
223
224	+	private static final int TERMINATING = 0x01;
225	+	private static final int TERMINATED  = 0x02;
226	+	private static final int SUSPENDED   = 0x04; // inactive spare
227	+	private static final int TRIMMED     = 0x08; // killed while suspended
228	+
229	+	/**
230	+	* Number of LockSupport.park calls to block this thread for
231	+	* suspension or event waits. Used for internal instrumention;
232	+	* currently not exported but included because volatile write upon
233	+	* park also provides a workaround for a JVM bug.
234	+	*/
235	+	volatile int parkCount;
236	+
237	+	/**
238	+	* Number of steals, transferred and reset in pool callbacks pool
239	+	* when idle Accessed directly by pool.
240	+	*/
241	+	int stealCount;
242	+
243		/**
244		* Seed for random number generator for choosing steal victims.
245	<	* Uses Marsaglia xorshift. Must be nonzero upon initialization.
245	>	* Uses Marsaglia xorshift. Must be initialized as nonzero.
246		*/
247		private int seed;
248
249		/**
250	<	* Number of steals, transferred to pool when idle
250	>	* Activity status. When true, this worker is considered active.
251	>	* Accessed directly by pool.  Must be false upon construction.
252		*/
253	<	private int stealCount;
253	>	boolean active;
254
255		/**
256	+	* True if use local fifo, not default lifo, for local polling.
257	+	* Shadows value from ForkJoinPool, which resets it if changed
258	+	* pool-wide.
259	+	*/
260	+	private final boolean locallyFifo;
261	+
262	+	/**
263		* Index of this worker in pool array. Set once by pool before
264	<	* running, and accessed directly by pool during cleanup etc.
264	>	* running, and accessed directly by pool to locate this worker in
265	>	* its workers array.
266		*/
267		int poolIndex;
268
269		/**
270	<	* The last barrier event waited for. Accessed in pool callback
271	<	* methods, but only by current thread.
270	>	* The last pool event waited for. Accessed only by pool in
271	>	* callback methods invoked within this thread.
272		*/
273	<	long lastEventCount;
273	>	int lastEventCount;
274
275		/**
276	<	* True if use local fifo, not default lifo, for local polling
276	>	* Encoded index and event count of next event waiter. Used only
277	>	* by ForkJoinPool for managing event waiters.
278		*/
279	<	private boolean locallyFifo;
279	>	volatile long nextWaiter;
280
281		/**
282		* Creates a ForkJoinWorkerThread operating in the given pool.
#	Line 217 | Line 285 | public class ForkJoinWorkerThread extend
285		* @throws NullPointerException if pool is null
286		*/
287		protected ForkJoinWorkerThread(ForkJoinPool pool) {
220	–	if (pool == null) throw new NullPointerException();
288		this.pool = pool;
289	<	// Note: poolIndex is set by pool during construction
290	<	// Remaining initialization is deferred to onStart
289	>	this.locallyFifo = pool.locallyFifo;
290	>	// To avoid exposing construction details to subclasses,
291	>	// remaining initialization is in start() and onStart()
292		}
293
294	<	// Public access methods
294	>	/**
295	>	* Performs additional initialization and starts this thread
296	>	*/
297	>	final void start(int poolIndex, UncaughtExceptionHandler ueh) {
298	>	this.poolIndex = poolIndex;
299	>	if (ueh != null)
300	>	setUncaughtExceptionHandler(ueh);
301	>	setDaemon(true);
302	>	start();
303	>	}
304	>
305	>	// Public/protected methods
306
307		/**
308		* Returns the pool hosting this thread.
#	Line 248 | Line 327 | public class ForkJoinWorkerThread extend
327		}
328
329		/**
330	<	* Establishes local first-in-first-out scheduling mode for forked
331	<	* tasks that are never joined.
332	<	*
333	<	* @param async if true, use locally FIFO scheduling
330	>	* Initializes internal state after construction but before
331	>	* processing any tasks. If you override this method, you must
332	>	* invoke super.onStart() at the beginning of the method.
333	>	* Initialization requires care: Most fields must have legal
334	>	* default values, to ensure that attempted accesses from other
335	>	* threads work correctly even before this thread starts
336	>	* processing tasks.
337		*/
338	<	void setAsyncMode(boolean async) {
339	<	locallyFifo = async;
340	<	}
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); }
338	>	protected void onStart() {
339	>	int rs = seedGenerator.nextInt();
340	>	seed = rs == 0? 1 : rs; // seed must be nonzero
341
342	<	/**
343	<	* Transitions to at least the given state.
344	<	*
345	<	* @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	<	}
342	>	// Allocate name string and arrays in this thread
343	>	String pid = Integer.toString(pool.getPoolNumber());
344	>	String wid = Integer.toString(poolIndex);
345	>	setName("ForkJoinPool-" + pid + "-worker-" + wid);
346
347	<	/**
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;
347	>	queue = new ForkJoinTask<?>[INITIAL_QUEUE_CAPACITY];
348		}
349
350		/**
351	<	* Tries to set status to inactive; fails on contention.
351	>	* Performs cleanup associated with termination of this worker
352	>	* thread.  If you override this method, you must invoke
353	>	* {@code super.onTermination} at the end of the overridden method.
354	>	*
355	>	* @param exception the exception causing this thread to abort due
356	>	* to an unrecoverable error, or {@code null} if completed normally
357		*/
358	<	private boolean tryInactivate() {
359	<	if (active) {
360	<	if (!pool.tryDecrementActiveCount())
361	<	return false;
362	<	active = false;
358	>	protected void onTermination(Throwable exception) {
359	>	try {
360	>	stolen = null;
361	>	joining = null;
362	>	cancelTasks();
363	>	setTerminated();
364	>	pool.workerTerminated(this);
365	>	} catch (Throwable ex) {        // Shouldn't ever happen
366	>	if (exception == null)      // but if so, at least rethrown
367	>	exception = ex;
368	>	} finally {
369	>	if (exception != null)
370	>	UNSAFE.throwException(exception);
371		}
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);
372		}
373
324	–	// Lifecycle methods
325	–
374		/**
375		* This method is required to be public, but should never be
376		* called explicitly. It performs the main run loop to execute
#	Line 332 | Line 380 | public class ForkJoinWorkerThread extend
380		Throwable exception = null;
381		try {
382		onStart();
335	–	pool.sync(this); // await first pool event
383		mainLoop();
384		} catch (Throwable ex) {
385		exception = ex;
#	Line 341 | Line 388 | public class ForkJoinWorkerThread extend
388		}
389		}
390
391	+	// helpers for run()
392	+
393		/**
394	<	* Executes tasks until shut down.
394	>	* Find and execute tasks and check status while running
395		*/
396		private void mainLoop() {
397	<	while (!isShutdown()) {
398	<	ForkJoinTask<?> t = pollTask();
399	<	if (t != null || (t = pollSubmission()) != null)
400	<	t.quietlyExec();
401	<	else if (tryInactivate())
402	<	pool.sync(this);
397	>	boolean ran = false;      // true if ran task in last loop iter
398	>	boolean prevRan = false;  // true if ran on last or previous step
399	>	ForkJoinPool p = pool;
400	>	for (;;) {
401	>	p.preStep(this, prevRan);
402	>	if (runState != 0)
403	>	return;
404	>	ForkJoinTask<?> t; // try to get and run stolen or submitted task
405	>	if ((t = scan()) != null || (t = pollSubmission()) != null) {
406	>	t.tryExec();
407	>	if (base != sp)
408	>	runLocalTasks();
409	>	stolen = null;
410	>	prevRan = ran = true;
411	>	}
412	>	else {
413	>	prevRan = ran;
414	>	ran = false;
415	>	}
416		}
417		}
418
419		/**
420	<	* Initializes internal state after construction but before
421	<	* processing any tasks. If you override this method, you must
360	<	* 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.
420	>	* Runs local tasks until queue is empty or shut down.  Call only
421	>	* while active.
422		*/
423	<	protected void onStart() {
424	<	// Allocate while starting to improve chances of thread-local
425	<	// isolation
426	<	queue = new ForkJoinTask<?>[INITIAL_QUEUE_CAPACITY];
427	<	// Initial value of seed need not be especially random but
428	<	// should differ across workers and must be nonzero
429	<	int p = poolIndex + 1;
430	<	seed = p + (p << 8) + (p << 16) + (p << 24); // spread bits
423	>	private void runLocalTasks() {
424	>	while (runState == 0) {
425	>	ForkJoinTask<?> t = locallyFifo? locallyDeqTask() : popTask();
426	>	if (t != null)
427	>	t.tryExec();
428	>	else if (base == sp)
429	>	break;
430	>	}
431		}
432
433		/**
434	<	* 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.
434	>	* If a submission exists, try to activate and take it
435		*
436	<	* @param exception the exception causing this thread to abort due
382	<	* to an unrecoverable error, or {@code null} if completed normally
436	>	* @return a task, if available
437		*/
438	<	protected void onTermination(Throwable exception) {
439	<	// Execute remaining local tasks unless aborting or terminating
440	<	while (exception == null &&  !pool.isTerminating() && base != sp) {
441	<	try {
442	<	ForkJoinTask<?> t = popTask();
443	<	if (t != null)
390	<	t.quietlyExec();
391	<	} catch (Throwable ex) {
392	<	exception = ex;
438	>	private ForkJoinTask<?> pollSubmission() {
439	>	ForkJoinPool p = pool;
440	>	while (p.hasQueuedSubmissions()) {
441	>	if (active || (active = p.tryIncrementActiveCount())) {
442	>	ForkJoinTask<?> t = p.pollSubmission();
443	>	return t != null ? t : scan(); // if missed, rescan
444		}
445		}
446	<	// 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	<	}
446	>	return null;
447		}
448
449	<	// Intrinsics-based support for queue operations.
450	<
451	<	/**
452	<	* Adds in store-order the given task at given slot of q to null.
453	<	* Caller must ensure q is non-null and index is in range.
449	>	/*
450	>	* Intrinsics-based atomic writes for queue slots. These are
451	>	* basically the same as methods in AtomicObjectArray, but
452	>	* specialized for (1) ForkJoinTask elements (2) requirement that
453	>	* nullness and bounds checks have already been performed by
454	>	* callers and (3) effective offsets are known not to overflow
455	>	* from int to long (because of MAXIMUM_QUEUE_CAPACITY). We don't
456	>	* need corresponding version for reads: plain array reads are OK
457	>	* because they protected by other volatile reads and are
458	>	* confirmed by CASes.
459	>	*
460	>	* Most uses don't actually call these methods, but instead contain
461	>	* inlined forms that enable more predictable optimization.  We
462	>	* don't define the version of write used in pushTask at all, but
463	>	* instead inline there a store-fenced array slot write.
464		*/
417	–	private static void setSlot(ForkJoinTask<?>[] q, int i,
418	–	ForkJoinTask<?> t) {
419	–	UNSAFE.putOrderedObject(q, (i << qShift) + qBase, t);
420	–	}
465
466		/**
467	<	* CAS given slot of q to null. Caller must ensure q is non-null
468	<	* and index is in range.
467	>	* CASes slot i of array q from t to null. Caller must ensure q is
468	>	* non-null and index is in range.
469		*/
470	<	private static boolean casSlotNull(ForkJoinTask<?>[] q, int i,
471	<	ForkJoinTask<?> t) {
470	>	private static final boolean casSlotNull(ForkJoinTask<?>[] q, int i,
471	>	ForkJoinTask<?> t) {
472		return UNSAFE.compareAndSwapObject(q, (i << qShift) + qBase, t, null);
473		}
474
475		/**
476	<	* Sets sp in store-order.
476	>	* Performs a volatile write of the given task at given slot of
477	>	* array q.  Caller must ensure q is non-null and index is in
478	>	* range. This method is used only during resets and backouts.
479		*/
480	<	private void storeSp(int s) {
481	<	UNSAFE.putOrderedInt(this, spOffset, s);
480	>	private static final void writeSlot(ForkJoinTask<?>[] q, int i,
481	>	ForkJoinTask<?> t) {
482	>	UNSAFE.putObjectVolatile(q, (i << qShift) + qBase, t);
483		}
484
485	<	// Main queue methods
485	>	// queue methods
486
487		/**
488	<	* Pushes a task. Called only by current thread.
488	>	* Pushes a task. Call only from this thread.
489		*
490		* @param t the task. Caller must ensure non-null.
491		*/
492		final void pushTask(ForkJoinTask<?> t) {
493		ForkJoinTask<?>[] q = queue;
494	<	int mask = q.length - 1;
495	<	int s = sp;
496	<	setSlot(q, s & mask, t);
497	<	storeSp(++s);
498	<	if ((s -= base) == 1)
499	<	pool.signalWork();
500	<	else if (s >= mask)
454	<	growQueue();
494	>	int mask = q.length - 1; // implicit assert q != null
495	>	int s = sp++;            // ok to increment sp before slot write
496	>	UNSAFE.putOrderedObject(q, ((s & mask) << qShift) + qBase, t);
497	>	if ((s -= base) == 0)
498	>	pool.signalWork();   // was empty
499	>	else if (s == mask)
500	>	growQueue();         // is full
501		}
502
503		/**
504		* Tries to take a task from the base of the queue, failing if
505	<	* either empty or contended.
505	>	* empty or contended. Note: Specializations of this code appear
506	>	* in locallyDeqTask and elsewhere.
507		*
508		* @return a task, or null if none or contended
509		*/
510		final ForkJoinTask<?> deqTask() {
511		ForkJoinTask<?> t;
512		ForkJoinTask<?>[] q;
513	<	int i;
514	<	int b;
468	<	if (sp != (b = base) &&
513	>	int b, i;
514	>	if ((b = base) != sp &&
515		(q = queue) != null && // must read q after b
516	<	(t = q[i = (q.length - 1) & b]) != null &&
517	<	casSlotNull(q, i, t)) {
516	>	(t = q[i = (q.length - 1) & b]) != null && base == b &&
517	>	UNSAFE.compareAndSwapObject(q, (i << qShift) + qBase, t, null)) {
518		base = b + 1;
519		return t;
520		}
#	Line 476 | Line 522 | public class ForkJoinWorkerThread extend
522		}
523
524		/**
525	<	* Tries to take a task from the base of own queue, activating if
526	<	* necessary, failing only if empty. Called only by current thread.
525	>	* Tries to take a task from the base of own queue. Assumes active
526	>	* status.  Called only by current thread.
527		*
528		* @return a task, or null if none
529		*/
530		final ForkJoinTask<?> locallyDeqTask() {
531	<	int b;
532	<	while (sp != (b = base)) {
533	<	if (tryActivate()) {
534	<	ForkJoinTask<?>[] q = queue;
535	<	int i = (q.length - 1) & b;
536	<	ForkJoinTask<?> t = q[i];
537	<	if (t != null && casSlotNull(q, i, t)) {
531	>	ForkJoinTask<?>[] q = queue;
532	>	if (q != null) {
533	>	ForkJoinTask<?> t;
534	>	int b, i;
535	>	while (sp != (b = base)) {
536	>	if ((t = q[i = (q.length - 1) & b]) != null && base == b &&
537	>	UNSAFE.compareAndSwapObject(q, (i << qShift) + qBase,
538	>	t, null)) {
539		base = b + 1;
540		return t;
541		}
#	Line 498 | Line 545 | public class ForkJoinWorkerThread extend
545		}
546
547		/**
548	<	* Returns a popped task, or null if empty. Ensures active status
549	<	* if non-null. Called only by current thread.
548	>	* Returns a popped task, or null if empty. Assumes active status.
549	>	* Called only by current thread. (Note: a specialization of this
550	>	* code appears in popWhileJoining.)
551		*/
552		final ForkJoinTask<?> popTask() {
553	<	int s = sp;
554	<	while (s != base) {
555	<	if (tryActivate()) {
556	<	ForkJoinTask<?>[] q = queue;
557	<	int mask = q.length - 1;
558	<	int i = (s - 1) & mask;
559	<	ForkJoinTask<?> t = q[i];
560	<	if (t == null || !casSlotNull(q, i, t))
513	<	break;
514	<	storeSp(s - 1);
553	>	int s;
554	>	ForkJoinTask<?>[] q;
555	>	if (base != (s = sp) && (q = queue) != null) {
556	>	int i = (q.length - 1) & --s;
557	>	ForkJoinTask<?> t = q[i];
558	>	if (t != null && UNSAFE.compareAndSwapObject
559	>	(q, (i << qShift) + qBase, t, null)) {
560	>	sp = s;
561		return t;
562		}
563		}
#	Line 519 | Line 565 | public class ForkJoinWorkerThread extend
565		}
566
567		/**
568	<	* Specialized version of popTask to pop only if
569	<	* topmost element is the given task. Called only
570	<	* by current thread while active.
568	>	* Specialized version of popTask to pop only if topmost element
569	>	* is the given task. Called only by current thread while
570	>	* active.
571		*
572		* @param t the task. Caller must ensure non-null.
573		*/
574		final boolean unpushTask(ForkJoinTask<?> t) {
575	<	ForkJoinTask<?>[] q = queue;
576	<	int mask = q.length - 1;
577	<	int s = sp - 1;
578	<	if (casSlotNull(q, s & mask, t)) {
579	<	storeSp(s);
575	>	int s;
576	>	ForkJoinTask<?>[] q;
577	>	if (base != (s = sp) && (q = queue) != null &&
578	>	UNSAFE.compareAndSwapObject
579	>	(q, (((q.length - 1) & --s) << qShift) + qBase, t, null)) {
580	>	sp = s;
581		return true;
582		}
583		return false;
#	Line 570 | Line 617 | public class ForkJoinWorkerThread extend
617		ForkJoinTask<?> t = oldQ[oldIndex];
618		if (t != null && !casSlotNull(oldQ, oldIndex, t))
619		t = null;
620	<	setSlot(newQ, b & newMask, t);
620	>	writeSlot(newQ, b & newMask, t);
621		} while (++b != bf);
622		pool.signalWork();
623		}
624
625		/**
626	+	* Computes next value for random victim probe in scan().  Scans
627	+	* don't require a very high quality generator, but also not a
628	+	* crummy one.  Marsaglia xor-shift is cheap and works well enough.
629	+	* Note: This is manually inlined in scan()
630	+	*/
631	+	private static final int xorShift(int r) {
632	+	r ^= r << 13;
633	+	r ^= r >>> 17;
634	+	return r ^ (r << 5);
635	+	}
636	+
637	+	/**
638		* Tries to steal a task from another worker. Starts at a random
639		* index of workers array, and probes workers until finding one
640		* with non-empty queue or finding that all are empty.  It
641		* randomly selects the first n probes. If these are empty, it
642	<	* resorts to a full circular traversal, which is necessary to
643	<	* accurately set active status by caller. Also restarts if pool
644	<	* events occurred since last scan, which forces refresh of
645	<	* workers array, in case barrier was associated with resize.
642	>	* resorts to a circular sweep, which is necessary to accurately
643	>	* set active status. (The circular sweep uses steps of
644	>	* approximately half the array size plus 1, to avoid bias
645	>	* stemming from leftmost packing of the array in ForkJoinPool.)
646		*
647		* This method must be both fast and quiet -- usually avoiding
648		* memory accesses that could disrupt cache sharing etc other than
649	<	* those needed to check for and take tasks. This accounts for,
650	<	* among other things, updating random seed in place without
651	<	* storing it until exit.
649	>	* those needed to check for and take tasks (or to activate if not
650	>	* already active). This accounts for, among other things,
651	>	* updating random seed in place without storing it until exit.
652		*
653		* @return a task, or null if none found
654		*/
655		private ForkJoinTask<?> scan() {
656	<	ForkJoinTask<?> t = null;
657	<	int r = seed;                    // extract once to keep scan quiet
658	<	ForkJoinWorkerThread[] ws;       // refreshed on outer loop
659	<	int mask;                        // must be power 2 minus 1 and > 0
660	<	outer:do {
661	<	if ((ws = pool.workers) != null && (mask = ws.length - 1) > 0) {
662	<	int idx = r;
663	<	int probes = ~mask;      // use random index while negative
664	<	for (;;) {
665	<	r = xorShift(r);     // update random seed
666	<	ForkJoinWorkerThread v = ws[mask & idx];
667	<	if (v == null || v.sp == v.base) {
668	<	if (probes <= mask)
669	<	idx = (probes++ < 0) ? r : (idx + 1);
670	<	else
671	<	break;
656	>	ForkJoinPool p = pool;
657	>	ForkJoinWorkerThread[] ws;        // worker array
658	>	int n;                            // upper bound of #workers
659	>	if ((ws = p.workers) != null && (n = ws.length) > 1) {
660	>	boolean canSteal = active;    // shadow active status
661	>	int r = seed;                 // extract seed once
662	>	int mask = n - 1;
663	>	int j = -n;                   // loop counter
664	>	int k = r;                    // worker index, random if j < 0
665	>	for (;;) {
666	>	ForkJoinWorkerThread v = ws[k & mask];
667	>	r ^= r << 13; r ^= r >>> 17; r ^= r << 5; // inline xorshift
668	>	if (v != null && v.base != v.sp) {
669	>	if (canSteal ||       // ensure active status
670	>	(canSteal = active = p.tryIncrementActiveCount())) {
671	>	int b = v.base;   // inline specialized deqTask
672	>	ForkJoinTask<?>[] q;
673	>	if (b != v.sp && (q = v.queue) != null) {
674	>	ForkJoinTask<?> t;
675	>	int i = (q.length - 1) & b;
676	>	long u = (i << qShift) + qBase; // raw offset
677	>	if ((t = q[i]) != null && v.base == b &&
678	>	UNSAFE.compareAndSwapObject(q, u, t, null)) {
679	>	stolen = t;
680	>	v.base = b + 1;
681	>	seed = r;
682	>	++stealCount;
683	>	return t;
684	>	}
685	>	}
686		}
687	<	else if (!tryActivate() || (t = v.deqTask()) == null)
688	<	continue outer;  // restart on contention
616	<	else
617	<	break outer;
687	>	j = -n;
688	>	k = r;                // restart on contention
689		}
690	+	else if (++j <= 0)
691	+	k = r;
692	+	else if (j <= n)
693	+	k += (n >>> 1) | 1;
694	+	else
695	+	break;
696		}
697	<	} while (pool.hasNewSyncEvent(this)); // retry on pool events
698	<	seed = r;
622	<	return t;
697	>	}
698	>	return null;
699		}
700
701	+	// Run State management
702	+
703	+	// status check methods used mainly by ForkJoinPool
704	+	final boolean isTerminating() { return (runState & TERMINATING) != 0; }
705	+	final boolean isTerminated()  { return (runState & TERMINATED) != 0; }
706	+	final boolean isSuspended()   { return (runState & SUSPENDED) != 0; }
707	+	final boolean isTrimmed()     { return (runState & TRIMMED) != 0; }
708	+
709		/**
710	<	* Gets and removes a local or stolen task.
627	<	*
628	<	* @return a task, if available
710	>	* Sets state to TERMINATING, also resuming if suspended.
711		*/
712	<	final ForkJoinTask<?> pollTask() {
713	<	ForkJoinTask<?> t = locallyFifo ? locallyDeqTask() : popTask();
714	<	if (t == null && (t = scan()) != null)
715	<	++stealCount;
716	<	return t;
712	>	final void shutdown() {
713	>	for (;;) {
714	>	int s = runState;
715	>	if ((s & SUSPENDED) != 0) { // kill and wakeup if suspended
716	>	if (UNSAFE.compareAndSwapInt(this, runStateOffset, s,
717	>	(s & ~SUSPENDED) |
718	>	(TRIMMED|TERMINATING))) {
719	>	LockSupport.unpark(this);
720	>	break;
721	>	}
722	>	}
723	>	else if (UNSAFE.compareAndSwapInt(this, runStateOffset, s,
724	>	s | TERMINATING))
725	>	break;
726	>	}
727	>	}
728	>
729	>	/**
730	>	* Sets state to TERMINATED. Called only by this thread.
731	>	*/
732	>	private void setTerminated() {
733	>	int s;
734	>	do {} while (!UNSAFE.compareAndSwapInt(this, runStateOffset,
735	>	s = runState,
736	>	s | (TERMINATING|TERMINATED)));
737	>	}
738	>
739	>	/**
740	>	* Instrumented version of park used by ForkJoinPool.awaitEvent
741	>	*/
742	>	final void doPark() {
743	>	++parkCount;
744	>	LockSupport.park(this);
745		}
746
747		/**
748	<	* Gets a local task.
748	>	* If suspended, tries to set status to unsuspended.
749	>	* Caller must unpark to actually resume
750		*
751	<	* @return a task, if available
751	>	* @return true if successful
752		*/
753	<	final ForkJoinTask<?> pollLocalTask() {
754	<	return locallyFifo ? locallyDeqTask() : popTask();
753	>	final boolean tryUnsuspend() {
754	>	int s = runState;
755	>	if ((s & SUSPENDED) != 0)
756	>	return UNSAFE.compareAndSwapInt(this, runStateOffset, s,
757	>	s & ~SUSPENDED);
758	>	return false;
759		}
760
761		/**
762	<	* Returns a pool submission, if one exists, activating first.
762	>	* Sets suspended status and blocks as spare until resumed,
763	>	* shutdown, or timed out.
764		*
765	<	* @return a submission, if available
765	>	* @return false if trimmed
766		*/
767	<	private ForkJoinTask<?> pollSubmission() {
768	<	ForkJoinPool p = pool;
769	<	while (p.hasQueuedSubmissions()) {
770	<	ForkJoinTask<?> t;
771	<	if (tryActivate() && (t = p.pollSubmission()) != null)
772	<	return t;
767	>	final boolean suspendAsSpare() {
768	>	for (;;) {               // set suspended unless terminating
769	>	int s = runState;
770	>	if ((s & TERMINATING) != 0) { // must kill
771	>	if (UNSAFE.compareAndSwapInt(this, runStateOffset, s,
772	>	s | (TRIMMED | TERMINATING)))
773	>	return false;
774	>	}
775	>	else if (UNSAFE.compareAndSwapInt(this, runStateOffset, s,
776	>	s | SUSPENDED))
777	>	break;
778		}
779	<	return null;
779	>	boolean timed;
780	>	long nanos;
781	>	long startTime;
782	>	if (poolIndex < pool.parallelism) {
783	>	timed = false;
784	>	nanos = 0L;
785	>	startTime = 0L;
786	>	}
787	>	else {
788	>	timed = true;
789	>	nanos = SPARE_KEEPALIVE_NANOS;
790	>	startTime = System.nanoTime();
791	>	}
792	>	pool.accumulateStealCount(this);
793	>	lastEventCount = 0;      // reset upon resume
794	>	interrupted();           // clear/ignore interrupts
795	>	while ((runState & SUSPENDED) != 0) {
796	>	++parkCount;
797	>	if (!timed)
798	>	LockSupport.park(this);
799	>	else if ((nanos -= (System.nanoTime() - startTime)) > 0)
800	>	LockSupport.parkNanos(this, nanos);
801	>	else { // try to trim on timeout
802	>	int s = runState;
803	>	if (UNSAFE.compareAndSwapInt(this, runStateOffset, s,
804	>	(s & ~SUSPENDED) |
805	>	(TRIMMED|TERMINATING)))
806	>	return false;
807	>	}
808	>	}
809	>	return true;
810		}
811
812	<	// Methods accessed only by Pool
812	>	// Misc support methods for ForkJoinPool
813	>
814	>	/**
815	>	* Returns an estimate of the number of tasks in the queue.  Also
816	>	* used by ForkJoinTask.
817	>	*/
818	>	final int getQueueSize() {
819	>	return -base + sp;
820	>	}
821
822		/**
823		* Removes and cancels all tasks in queue.  Can be called from any
824		* thread.
825		*/
826		final void cancelTasks() {
827	<	ForkJoinTask<?> t;
828	<	while (base != sp && (t = deqTask()) != null)
829	<	t.cancelIgnoringExceptions();
827	>	while (base != sp) {
828	>	ForkJoinTask<?> t = deqTask();
829	>	if (t != null)
830	>	t.cancelIgnoringExceptions();
831	>	}
832		}
833
834		/**
#	Line 677 | Line 838 | public class ForkJoinWorkerThread extend
838		*/
839		final int drainTasksTo(Collection<? super ForkJoinTask<?>> c) {
840		int n = 0;
841	<	ForkJoinTask<?> t;
842	<	while (base != sp && (t = deqTask()) != null) {
843	<	c.add(t);
844	<	++n;
841	>	while (base != sp) {
842	>	ForkJoinTask<?> t = deqTask();
843	>	if (t != null) {
844	>	c.add(t);
845	>	++n;
846	>	}
847		}
848		return n;
849		}
850
851	<	/**
689	<	* Gets and clears steal count for accumulation by pool.  Called
690	<	* only when known to be idle (in pool.sync and termination).
691	<	*/
692	<	final int getAndClearStealCount() {
693	<	int sc = stealCount;
694	<	stealCount = 0;
695	<	return sc;
696	<	}
851	>	// Support methods for ForkJoinTask
852
853		/**
854	<	* Returns {@code true} if at least one worker in the given array
700	<	* appears to have at least one queued task.
854	>	* Possibly runs some tasks and/or blocks, until task is done.
855		*
856	<	* @param ws array of workers
856	>	* @param joinMe the task to join
857		*/
858	<	static boolean hasQueuedTasks(ForkJoinWorkerThread[] ws) {
859	<	if (ws != null) {
860	<	int len = ws.length;
861	<	for (int j = 0; j < 2; ++j) { // need two passes for clean sweep
862	<	for (int i = 0; i < len; ++i) {
863	<	ForkJoinWorkerThread w = ws[i];
864	<	if (w != null && w.sp != w.base)
865	<	return true;
858	>	final void joinTask(ForkJoinTask<?> joinMe) {
859	>	ForkJoinTask<?> prevJoining = joining;
860	>	joining = joinMe;
861	>	while (joinMe.status >= 0) {
862	>	int s = sp;
863	>	if (s == base) {
864	>	nonlocalJoinTask(joinMe);
865	>	break;
866	>	}
867	>	// process local task
868	>	ForkJoinTask<?> t;
869	>	ForkJoinTask<?>[] q = queue;
870	>	int i = (q.length - 1) & --s;
871	>	long u = (i << qShift) + qBase; // raw offset
872	>	if ((t = q[i]) != null &&
873	>	UNSAFE.compareAndSwapObject(q, u, t, null)) {
874	>	/*
875	>	* This recheck (and similarly in nonlocalJoinTask)
876	>	* handles cases where joinMe is independently
877	>	* cancelled or forced even though there is other work
878	>	* available. Back out of the pop by putting t back
879	>	* into slot before we commit by setting sp.
880	>	*/
881	>	if (joinMe.status < 0) {
882	>	UNSAFE.putObjectVolatile(q, u, t);
883	>	break;
884		}
885	+	sp = s;
886	+	t.tryExec();
887		}
888		}
889	<	return false;
889	>	joining = prevJoining;
890		}
891
718	–	// Support methods for ForkJoinTask
719	–
892		/**
893	<	* Returns an estimate of the number of tasks in the queue.
893	>	* Tries to locate and help perform tasks for a stealer of the
894	>	* given task (or in turn one of its stealers), blocking (via
895	>	* pool.tryAwaitJoin) upon failure to find work.  Traces
896	>	* stolen->joining links looking for a thread working on
897	>	* a descendant of the given task and with a non-empty queue to
898	>	* steal back and execute tasks from. Inhibits mutual steal chains
899	>	* and scans on outer joins upon nesting to avoid unbounded
900	>	* growth.  Restarts search upon encountering inconsistencies.
901	>	* Tries to block if two passes agree that there are no remaining
902	>	* targets.
903	>	*
904	>	* @param joinMe the task to join
905		*/
906	<	final int getQueueSize() {
907	<	// suppress momentarily negative values
908	<	return Math.max(0, sp - base);
906	>	private void nonlocalJoinTask(ForkJoinTask<?> joinMe) {
907	>	ForkJoinPool p = pool;
908	>	int scans = p.parallelism;       // give up if too many retries
909	>	ForkJoinTask<?> bottom = null;   // target seen when can't descend
910	>	restart: while (joinMe.status >= 0) {
911	>	ForkJoinTask<?> target = null;
912	>	ForkJoinTask<?> next = joinMe;
913	>	while (scans >= 0 && next != null) {
914	>	--scans;
915	>	target = next;
916	>	next = null;
917	>	ForkJoinWorkerThread v = null;
918	>	ForkJoinWorkerThread[] ws = p.workers;
919	>	int n = ws.length;
920	>	for (int j = 0; j < n; ++j) {
921	>	ForkJoinWorkerThread w = ws[j];
922	>	if (w != null && w.stolen == target) {
923	>	v = w;
924	>	break;
925	>	}
926	>	}
927	>	if (v != null && v != this) {
928	>	ForkJoinTask<?> prevStolen = stolen;
929	>	int b;
930	>	ForkJoinTask<?>[] q;
931	>	while ((b = v.base) != v.sp && (q = v.queue) != null) {
932	>	int i = (q.length - 1) & b;
933	>	long u = (i << qShift) + qBase;
934	>	ForkJoinTask<?> t = q[i];
935	>	if (target.status < 0)
936	>	continue restart;
937	>	if (t != null && v.base == b &&
938	>	UNSAFE.compareAndSwapObject(q, u, t, null)) {
939	>	if (joinMe.status < 0) {
940	>	UNSAFE.putObjectVolatile(q, u, t);
941	>	return; // back out
942	>	}
943	>	stolen = t;
944	>	v.base = b + 1;
945	>	t.tryExec();
946	>	stolen = prevStolen;
947	>	}
948	>	if (joinMe.status < 0)
949	>	return;
950	>	}
951	>	next = v.joining;
952	>	}
953	>	if (target.status < 0)
954	>	continue restart;  // inconsistent
955	>	if (joinMe.status < 0)
956	>	return;
957	>	}
958	>
959	>	if (bottom != target)
960	>	bottom = target;    // recheck landing spot
961	>	else if (p.tryAwaitJoin(joinMe) < 0)
962	>	return;             // successfully blocked
963	>	Thread.yield();         // tame spin in case too many active
964	>	}
965		}
966
967		/**
968		* Returns an estimate of the number of tasks, offset by a
969		* function of number of idle workers.
970	+	*
971	+	* This method provides a cheap heuristic guide for task
972	+	* partitioning when programmers, frameworks, tools, or languages
973	+	* have little or no idea about task granularity.  In essence by
974	+	* offering this method, we ask users only about tradeoffs in
975	+	* overhead vs expected throughput and its variance, rather than
976	+	* how finely to partition tasks.
977	+	*
978	+	* In a steady state strict (tree-structured) computation, each
979	+	* thread makes available for stealing enough tasks for other
980	+	* threads to remain active. Inductively, if all threads play by
981	+	* the same rules, each thread should make available only a
982	+	* constant number of tasks.
983	+	*
984	+	* The minimum useful constant is just 1. But using a value of 1
985	+	* would require immediate replenishment upon each steal to
986	+	* maintain enough tasks, which is infeasible.  Further,
987	+	* partitionings/granularities of offered tasks should minimize
988	+	* steal rates, which in general means that threads nearer the top
989	+	* of computation tree should generate more than those nearer the
990	+	* bottom. In perfect steady state, each thread is at
991	+	* approximately the same level of computation tree. However,
992	+	* producing extra tasks amortizes the uncertainty of progress and
993	+	* diffusion assumptions.
994	+	*
995	+	* So, users will want to use values larger, but not much larger
996	+	* than 1 to both smooth over transient shortages and hedge
997	+	* against uneven progress; as traded off against the cost of
998	+	* extra task overhead. We leave the user to pick a threshold
999	+	* value to compare with the results of this call to guide
1000	+	* decisions, but recommend values such as 3.
1001	+	*
1002	+	* When all threads are active, it is on average OK to estimate
1003	+	* surplus strictly locally. In steady-state, if one thread is
1004	+	* maintaining say 2 surplus tasks, then so are others. So we can
1005	+	* just use estimated queue length (although note that (sp - base)
1006	+	* can be an overestimate because of stealers lagging increments
1007	+	* of base).  However, this strategy alone leads to serious
1008	+	* mis-estimates in some non-steady-state conditions (ramp-up,
1009	+	* ramp-down, other stalls). We can detect many of these by
1010	+	* further considering the number of "idle" threads, that are
1011	+	* known to have zero queued tasks, so compensate by a factor of
1012	+	* (#idle/#active) threads.
1013		*/
1014		final int getEstimatedSurplusTaskCount() {
1015	<	// The halving approximates weighting idle vs non-idle workers
734	<	return (sp - base) - (pool.getIdleThreadCount() >>> 1);
1015	>	return sp - base - pool.idlePerActive();
1016		}
1017
1018		/**
1019	<	* Scans, returning early if joinMe done.
1019	>	* Gets and removes a local task.
1020	>	*
1021	>	* @return a task, if available
1022		*/
1023	<	final ForkJoinTask<?> scanWhileJoining(ForkJoinTask<?> joinMe) {
1024	<	ForkJoinTask<?> t = pollTask();
1025	<	if (t != null && joinMe.status < 0 && sp == base) {
1026	<	pushTask(t); // unsteal if done and this task would be stealable
744	<	t = null;
1023	>	final ForkJoinTask<?> pollLocalTask() {
1024	>	while (sp != base) {
1025	>	if (active || (active = pool.tryIncrementActiveCount()))
1026	>	return locallyFifo? locallyDeqTask() : popTask();
1027		}
1028	<	return t;
1028	>	return null;
1029	>	}
1030	>
1031	>	/**
1032	>	* Gets and removes a local or stolen task.
1033	>	*
1034	>	* @return a task, if available
1035	>	*/
1036	>	final ForkJoinTask<?> pollTask() {
1037	>	ForkJoinTask<?> t;
1038	>	return (t = pollLocalTask()) != null ? t : scan();
1039		}
1040
1041		/**
#	Line 751 | Line 1043 | public class ForkJoinWorkerThread extend
1043		*/
1044		final void helpQuiescePool() {
1045		for (;;) {
1046	<	ForkJoinTask<?> t = pollTask();
1047	<	if (t != null)
1048	<	t.quietlyExec();
1049	<	else if (tryInactivate() && pool.isQuiescent())
1050	<	break;
1046	>	ForkJoinTask<?> t = pollLocalTask();
1047	>	if (t != null || (t = scan()) != null) {
1048	>	t.tryExec();
1049	>	stolen = null;
1050	>	}
1051	>	else {
1052	>	ForkJoinPool p = pool;
1053	>	if (active) {
1054	>	active = false; // inactivate
1055	>	do {} while (!p.tryDecrementActiveCount());
1056	>	}
1057	>	if (p.isQuiescent()) {
1058	>	active = true; // re-activate
1059	>	do {} while (!p.tryIncrementActiveCount());
1060	>	return;
1061	>	}
1062	>	}
1063		}
760	–	do {} while (!tryActivate()); // re-activate on exit
1064		}
1065
1066		// Unsafe mechanics
1067
1068		private static final sun.misc.Unsafe UNSAFE = getUnsafe();
766	–	private static final long spOffset =
767	–	objectFieldOffset("sp", ForkJoinWorkerThread.class);
1069		private static final long runStateOffset =
1070		objectFieldOffset("runState", ForkJoinWorkerThread.class);
1071	<	private static final long qBase;
1071	>	private static final long qBase =
1072	>	UNSAFE.arrayBaseOffset(ForkJoinTask[].class);
1073		private static final int qShift;
1074
1075		static {
774	–	qBase = UNSAFE.arrayBaseOffset(ForkJoinTask[].class);
1076		int s = UNSAFE.arrayIndexScale(ForkJoinTask[].class);
1077		if ((s & (s-1)) != 0)
1078		throw new Error("data type scale not a power of two");

Diff Legend

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