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 1.28 by dl, Mon Aug 3 13:40:07 2009 UTC vs.
Revision 1.66 by jsr166, Wed Jun 1 21:04:30 2011 UTC

#	Line 1 | Line 1
1		/*
2		* Written by Doug Lea with assistance from members of JCP JSR-166
3		* Expert Group and released to the public domain, as explained at
4	<	* http://creativecommons.org/licenses/publicdomain
4	>	* http://creativecommons.org/publicdomain/zero/1.0/
5		*/
6
7		package jsr166y;
8
9	–	import java.util.concurrent.*;
10	–
9		import java.util.Collection;
10	+	import java.util.concurrent.RejectedExecutionException;
11
12		/**
13	<	* A thread managed by a {@link ForkJoinPool}.  This class is
14	<	* subclassable solely for the sake of adding functionality -- there
15	<	* are no overridable methods dealing with scheduling or execution.
16	<	* However, you can override initialization and termination methods
17	<	* surrounding the main task processing loop.  If you do create such a
18	<	* subclass, you will also need to supply a custom {@link
19	<	* ForkJoinPool.ForkJoinWorkerThreadFactory} to use it in a {@code
20	<	* ForkJoinPool}.
13	>	* A thread managed by a {@link ForkJoinPool}, which executes
14	>	* {@link ForkJoinTask}s.
15	>	* This class is subclassable solely for the sake of adding
16	>	* functionality -- there are no overridable methods dealing with
17	>	* scheduling or execution.  However, you can override initialization
18	>	* and termination methods surrounding the main task processing loop.
19	>	* If you do create such a subclass, you will also need to supply a
20	>	* custom {@link ForkJoinPool.ForkJoinWorkerThreadFactory} to use it
21	>	* in a {@code ForkJoinPool}.
22		*
23		* @since 1.7
24		* @author Doug Lea
25		*/
26		public class ForkJoinWorkerThread extends Thread {
27		/*
28	<	* Algorithm overview:
28	>	* Overview:
29		*
30	<	* 1. Work-Stealing: Work-stealing queues are special forms of
31	<	* Deques that support only three of the four possible
32	<	* end-operations -- push, pop, and deq (aka steal), and only do
33	<	* so under the constraints that push and pop are called only from
34	<	* the owning thread, while deq may be called from other threads.
35	<	* (If you are unfamiliar with them, you probably want to read
36	<	* Herlihy and Shavit's book "The Art of Multiprocessor
37	<	* programming", chapter 16 describing these in more detail before
38	<	* proceeding.)  The main work-stealing queue design is roughly
39	<	* similar to "Dynamic Circular Work-Stealing Deque" by David
40	<	* Chase and Yossi Lev, SPAA 2005
41	<	* (http://research.sun.com/scalable/pubs/index.html).  The main
42	<	* difference ultimately stems from gc requirements that we null
43	<	* out taken slots as soon as we can, to maintain as small a
44	<	* footprint as possible even in programs generating huge numbers
45	<	* of tasks. To accomplish this, we shift the CAS arbitrating pop
46	<	* vs deq (steal) from being on the indices ("base" and "sp") to
47	<	* the slots themselves (mainly via method "casSlotNull()"). So,
48	<	* both a successful pop and deq mainly entail CAS'ing a non-null
49	<	* slot to null.  Because we rely on CASes of references, we do
50	<	* not need tag bits on base or sp.  They are simple ints as used
51	<	* in any circular array-based queue (see for example ArrayDeque).
30	>	* ForkJoinWorkerThreads are managed by ForkJoinPools and perform
31	>	* ForkJoinTasks. This class includes bookkeeping in support of
32	>	* worker activation, suspension, and lifecycle control described
33	>	* in more detail in the internal documentation of class
34	>	* ForkJoinPool. And as described further below, this class also
35	>	* includes special-cased support for some ForkJoinTask
36	>	* methods. But the main mechanics involve work-stealing:
37	>	*
38	>	* Work-stealing queues are special forms of Deques that support
39	>	* only three of the four possible end-operations -- push, pop,
40	>	* and deq (aka steal), under the further constraints that push
41	>	* and pop are called only from the owning thread, while deq may
42	>	* be called from other threads.  (If you are unfamiliar with
43	>	* them, you probably want to read Herlihy and Shavit's book "The
44	>	* Art of Multiprocessor programming", chapter 16 describing these
45	>	* in more detail before proceeding.)  The main work-stealing
46	>	* queue design is roughly similar to those in the papers "Dynamic
47	>	* Circular Work-Stealing Deque" by Chase and Lev, SPAA 2005
48	>	* (http://research.sun.com/scalable/pubs/index.html) and
49	>	* "Idempotent work stealing" by Michael, Saraswat, and Vechev,
50	>	* PPoPP 2009 (http://portal.acm.org/citation.cfm?id=1504186).
51	>	* The main differences ultimately stem from gc requirements that
52	>	* we null out taken slots as soon as we can, to maintain as small
53	>	* a footprint as possible even in programs generating huge
54	>	* numbers of tasks. To accomplish this, we shift the CAS
55	>	* arbitrating pop vs deq (steal) from being on the indices
56	>	* ("queueBase" and "queueTop") to the slots themselves (mainly
57	>	* via method "casSlotNull()"). So, both a successful pop and deq
58	>	* mainly entail a CAS of a slot from non-null to null.  Because
59	>	* we rely on CASes of references, we do not need tag bits on
60	>	* queueBase or queueTop.  They are simple ints as used in any
61	>	* circular array-based queue (see for example ArrayDeque).
62		* Updates to the indices must still be ordered in a way that
63	<	* guarantees that (sp - base) > 0 means the queue is empty, but
64	<	* otherwise may err on the side of possibly making the queue
63	>	* guarantees that queueTop == queueBase means the queue is empty,
64	>	* but otherwise may err on the side of possibly making the queue
65		* appear nonempty when a push, pop, or deq have not fully
66		* committed. Note that this means that the deq operation,
67		* considered individually, is not wait-free. One thief cannot
68		* successfully continue until another in-progress one (or, if
69		* previously empty, a push) completes.  However, in the
70	<	* aggregate, we ensure at least probabilistic
71	<	* non-blockingness. If an attempted steal fails, a thief always
72	<	* chooses a different random victim target to try next. So, in
73	<	* order for one thief to progress, it suffices for any
74	<	* in-progress deq or new push on any empty queue to complete. One
65	<	* reason this works well here is that apparently-nonempty often
66	<	* means soon-to-be-stealable, which gives threads a chance to
67	<	* activate if necessary before stealing (see below).
70	>	* aggregate, we ensure at least probabilistic non-blockingness.
71	>	* If an attempted steal fails, a thief always chooses a different
72	>	* random victim target to try next. So, in order for one thief to
73	>	* progress, it suffices for any in-progress deq or new push on
74	>	* any empty queue to complete.
75		*
76		* This approach also enables support for "async mode" where local
77		* task processing is in FIFO, not LIFO order; simply by using a
78		* version of deq rather than pop when locallyFifo is true (as set
79		* by the ForkJoinPool).  This allows use in message-passing
80	<	* frameworks in which tasks are never joined.
81	<	*
82	<	* Efficient implementation of this approach currently relies on
83	<	* an uncomfortable amount of "Unsafe" mechanics. To maintain
84	<	* correct orderings, reads and writes of variable base require
85	<	* volatile ordering.  Variable sp does not require volatile write
86	<	* but needs cheaper store-ordering on writes.  Because they are
87	<	* protected by volatile base reads, reads of the queue array and
88	<	* its slots do not need volatile load semantics, but writes (in
89	<	* push) require store order and CASes (in pop and deq) require
90	<	* (volatile) CAS semantics.  (See "Idempotent work stealing" by
91	<	* Michael, Saraswat, and Vechev, PPoPP 2009
92	<	* http://portal.acm.org/citation.cfm?id=1504186 for an algorithm
93	<	* with similar properties, but without support for nulling
94	<	* slots.)  Since these combinations aren't supported using
95	<	* ordinary volatiles, the only way to accomplish these
80	>	* frameworks in which tasks are never joined.  However neither
81	>	* mode considers affinities, loads, cache localities, etc, so
82	>	* rarely provide the best possible performance on a given
83	>	* machine, but portably provide good throughput by averaging over
84	>	* these factors.  (Further, even if we did try to use such
85	>	* information, we do not usually have a basis for exploiting
86	>	* it. For example, some sets of tasks profit from cache
87	>	* affinities, but others are harmed by cache pollution effects.)
88	>	*
89	>	* When a worker would otherwise be blocked waiting to join a
90	>	* task, it first tries a form of linear helping: Each worker
91	>	* records (in field currentSteal) the most recent task it stole
92	>	* from some other worker. Plus, it records (in field currentJoin)
93	>	* the task it is currently actively joining. Method joinTask uses
94	>	* these markers to try to find a worker to help (i.e., steal back
95	>	* a task from and execute it) that could hasten completion of the
96	>	* actively joined task. In essence, the joiner executes a task
97	>	* that would be on its own local deque had the to-be-joined task
98	>	* not been stolen. This may be seen as a conservative variant of
99	>	* the approach in Wagner & Calder "Leapfrogging: a portable
100	>	* technique for implementing efficient futures" SIGPLAN Notices,
101	>	* 1993 (http://portal.acm.org/citation.cfm?id=155354). It differs
102	>	* in that: (1) We only maintain dependency links across workers
103	>	* upon steals, rather than use per-task bookkeeping.  This may
104	>	* require a linear scan of workers array to locate stealers, but
105	>	* usually doesn't because stealers leave hints (that may become
106	>	* stale/wrong) of where to locate them. This isolates cost to
107	>	* when it is needed, rather than adding to per-task overhead.
108	>	* (2) It is "shallow", ignoring nesting and potentially cyclic
109	>	* mutual steals.  (3) It is intentionally racy: field currentJoin
110	>	* is updated only while actively joining, which means that we
111	>	* miss links in the chain during long-lived tasks, GC stalls etc
112	>	* (which is OK since blocking in such cases is usually a good
113	>	* idea).  (4) We bound the number of attempts to find work (see
114	>	* MAX_HELP) and fall back to suspending the worker and if
115	>	* necessary replacing it with another.
116	>	*
117	>	* Efficient implementation of these algorithms currently relies
118	>	* on an uncomfortable amount of "Unsafe" mechanics. To maintain
119	>	* correct orderings, reads and writes of variable queueBase
120	>	* require volatile ordering.  Variable queueTop need not be
121	>	* volatile because non-local reads always follow those of
122	>	* queueBase.  Similarly, because they are protected by volatile
123	>	* queueBase reads, reads of the queue array and its slots by
124	>	* other threads do not need volatile load semantics, but writes
125	>	* (in push) require store order and CASes (in pop and deq)
126	>	* require (volatile) CAS semantics.  (Michael, Saraswat, and
127	>	* Vechev's algorithm has similar properties, but without support
128	>	* for nulling slots.)  Since these combinations aren't supported
129	>	* using ordinary volatiles, the only way to accomplish these
130		* efficiently is to use direct Unsafe calls. (Using external
131		* AtomicIntegers and AtomicReferenceArrays for the indices and
132		* array is significantly slower because of memory locality and
133		* indirection effects.)
134	<	*
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
101	<	* gets the initial signal to start processing tasks.  However,
102	<	* all queue-related methods except pushTask are written in a way
103	<	* that allows them to instead be lazily allocated and/or disposed
104	<	* of when empty. All together, these low-level implementation
105	<	* choices produce as much as a factor of 4 performance
106	<	* improvement compared to naive implementations, and enable the
107	<	* processing of billions of tasks per second, sometimes at the
108	<	* expense of ugliness.
109	<	*
110	<	* 2. Run control: The primary run control is based on a global
111	<	* counter (activeCount) held by the pool. It uses an algorithm
112	<	* similar to that in Herlihy and Shavit section 17.6 to cause
113	<	* threads to eventually block when all threads declare they are
114	<	* inactive. For this to work, threads must be declared active
115	<	* when executing tasks, and before stealing a task. They must be
116	<	* inactive before blocking on the Pool Barrier (awaiting a new
117	<	* submission or other Pool event). In between, there is some free
118	<	* play which we take advantage of to avoid contention and rapid
119	<	* flickering of the global activeCount: If inactive, we activate
120	<	* only if a victim queue appears to be nonempty (see above).
121	<	* Similarly, a thread tries to inactivate only after a full scan
122	<	* of other threads.  The net effect is that contention on
123	<	* activeCount is rarely a measurable performance issue. (There
124	<	* are also a few other cases where we scan for work rather than
125	<	* retry/block upon contention.)
126	<	*
127	<	* 3. Selection control. We maintain policy of always choosing to
128	<	* run local tasks rather than stealing, and always trying to
129	<	* steal tasks before trying to run a new submission. All steals
130	<	* are currently performed in randomly-chosen deq-order. It may be
131	<	* worthwhile to bias these with locality / anti-locality
132	<	* information, but doing this well probably requires more
133	<	* lower-level information from JVMs than currently provided.
141	>	* queues are initialized after starting.
142		*/
143
144		/**
145	+	* Mask for pool indices encoded as shorts
146	+	*/
147	+	private static final int  SMASK  = 0xffff;
148	+
149	+	/**
150		* Capacity of work-stealing queue array upon initialization.
151	<	* Must be a power of two. Initial size must be at least 2, but is
151	>	* Must be a power of two. Initial size must be at least 4, but is
152		* padded to minimize cache effects.
153		*/
154		private static final int INITIAL_QUEUE_CAPACITY = 1 << 13;
155
156		/**
157	<	* Maximum work-stealing queue array size.  Must be less than or
158	<	* equal to 1 << 28 to ensure lack of index wraparound. (This
159	<	* is less than usual bounds, because we need leftshift by 3
160	<	* to be in int range).
157	>	* Maximum size for queue array. Must be a power of two
158	>	* less than or equal to 1 << (31 - width of array entry) to
159	>	* ensure lack of index wraparound, but is capped at a lower
160	>	* value to help users trap runaway computations.
161		*/
162	<	private static final int MAXIMUM_QUEUE_CAPACITY = 1 << 28;
162	>	private static final int MAXIMUM_QUEUE_CAPACITY = 1 << 24; // 16M
163
164		/**
165	<	* The pool this thread works in. Accessed directly by ForkJoinTask.
165	>	* The work-stealing queue array. Size must be a power of two.
166	>	* Initialized when started (as oposed to when constructed), to
167	>	* improve memory locality.
168		*/
169	<	final ForkJoinPool pool;
169	>	ForkJoinTask<?>[] queue;
170
171		/**
172	<	* The work-stealing queue array. Size must be a power of two.
158	<	* Initialized when thread starts, to improve memory locality.
172	>	* The pool this thread works in. Accessed directly by ForkJoinTask.
173		*/
174	<	private ForkJoinTask<?>[] queue;
174	>	final ForkJoinPool pool;
175
176		/**
177		* Index (mod queue.length) of next queue slot to push to or pop
178	<	* from. It is written only by owner thread, via ordered store.
179	<	* Both sp and base are allowed to wrap around on overflow, but
180	<	* (sp - base) still estimates size.
178	>	* from. It is written only by owner thread, and accessed by other
179	>	* threads only after reading (volatile) queueBase.  Both queueTop
180	>	* and queueBase are allowed to wrap around on overflow, but
181	>	* (queueTop - queueBase) still estimates size.
182		*/
183	<	private volatile int sp;
183	>	int queueTop;
184
185		/**
186		* Index (mod queue.length) of least valid queue slot, which is
187		* always the next position to steal from if nonempty.
188		*/
189	<	private volatile int base;
189	>	volatile int queueBase;
190	>
191	>	/**
192	>	* The index of most recent stealer, used as a hint to avoid
193	>	* traversal in method helpJoinTask. This is only a hint because a
194	>	* worker might have had multiple steals and this only holds one
195	>	* of them (usually the most current). Declared non-volatile,
196	>	* relying on other prevailing sync to keep reasonably current.
197	>	*/
198	>	int stealHint;
199
200		/**
201	<	* Activity status. When true, this worker is considered active.
202	<	* Must be false upon construction. It must be true when executing
203	<	* tasks, and BEFORE stealing a task. It must be false before
180	<	* calling pool.sync.
201	>	* Index of this worker in pool array. Set once by pool before
202	>	* running, and accessed directly by pool to locate this worker in
203	>	* its workers array.
204		*/
205	<	private boolean active;
205	>	final int poolIndex;
206
207		/**
208	<	* Run state of this worker. Supports simple versions of the usual
209	<	* shutdown/shutdownNow control.
208	>	* Encoded record for pool task waits. Usages are always
209	>	* surrounded by volatile reads/writes
210		*/
211	<	private volatile int runState;
211	>	int nextWait;
212	>
213	>	/**
214	>	* Complement of poolIndex, offset by count of entries of task
215	>	* waits. Accessed by ForkJoinPool to manage event waiters.
216	>	*/
217	>	volatile int eventCount;
218
219		/**
220		* Seed for random number generator for choosing steal victims.
221	<	* Uses Marsaglia xorshift. Must be nonzero upon initialization.
221	>	* Uses Marsaglia xorshift. Must be initialized as nonzero.
222		*/
223	<	private int seed;
223	>	int seed;
224
225		/**
226	<	* Number of steals, transferred to pool when idle
226	>	* Number of steals. Directly accessed (and reset) by pool when
227	>	* idle.
228		*/
229	<	private int stealCount;
229	>	int stealCount;
230
231		/**
232	<	* Index of this worker in pool array. Set once by pool before
233	<	* running, and accessed directly by pool during cleanup etc.
232	>	* True if this worker should or did terminate
233	>	*/
234	>	volatile boolean terminate;
235	>
236	>	/**
237	>	* Set to true before LockSupport.park; false on return
238		*/
239	<	int poolIndex;
239	>	volatile boolean parked;
240
241		/**
242	<	* The last barrier event waited for. Accessed in pool callback
243	<	* methods, but only by current thread.
242	>	* True if use local fifo, not default lifo, for local polling.
243	>	* Shadows value from ForkJoinPool.
244		*/
245	<	long lastEventCount;
245	>	final boolean locallyFifo;
246
247		/**
248	<	* True if use local fifo, not default lifo, for local polling
248	>	* The task most recently stolen from another worker (or
249	>	* submission queue).  All uses are surrounded by enough volatile
250	>	* reads/writes to maintain as non-volatile.
251		*/
252	<	private boolean locallyFifo;
252	>	ForkJoinTask<?> currentSteal;
253	>
254	>	/**
255	>	* The task currently being joined, set only when actively trying
256	>	* to help other stealers in helpJoinTask. All uses are surrounded
257	>	* by enough volatile reads/writes to maintain as non-volatile.
258	>	*/
259	>	ForkJoinTask<?> currentJoin;
260
261		/**
262		* Creates a ForkJoinWorkerThread operating in the given pool.
#	Line 222 | Line 265 | public class ForkJoinWorkerThread extend
265		* @throws NullPointerException if pool is null
266		*/
267		protected ForkJoinWorkerThread(ForkJoinPool pool) {
268	<	if (pool == null) throw new NullPointerException();
268	>	super(pool.nextWorkerName());
269		this.pool = pool;
270	<	// Note: poolIndex is set by pool during construction
271	<	// Remaining initialization is deferred to onStart
270	>	int k = pool.registerWorker(this);
271	>	poolIndex = k;
272	>	eventCount = ~k & SMASK; // clear wait count
273	>	locallyFifo = pool.locallyFifo;
274	>	Thread.UncaughtExceptionHandler ueh = pool.ueh;
275	>	if (ueh != null)
276	>	setUncaughtExceptionHandler(ueh);
277	>	setDaemon(true);
278		}
279
280	<	// Public access methods
280	>	// Public methods
281
282		/**
283		* Returns the pool hosting this thread.
#	Line 252 | Line 301 | public class ForkJoinWorkerThread extend
301		return poolIndex;
302		}
303
304	<	/**
256	<	* Establishes local first-in-first-out scheduling mode for forked
257	<	* tasks that are never joined.
258	<	*
259	<	* @param async if true, use locally FIFO scheduling
260	<	*/
261	<	void setAsyncMode(boolean async) {
262	<	locallyFifo = async;
263	<	}
264	<
265	<	// Runstate management
266	<
267	<	// Runstate values. Order matters
268	<	private static final int RUNNING     = 0;
269	<	private static final int SHUTDOWN    = 1;
270	<	private static final int TERMINATING = 2;
271	<	private static final int TERMINATED  = 3;
272	<
273	<	final boolean isShutdown()    { return runState >= SHUTDOWN;  }
274	<	final boolean isTerminating() { return runState >= TERMINATING;  }
275	<	final boolean isTerminated()  { return runState == TERMINATED; }
276	<	final boolean shutdown()      { return transitionRunStateTo(SHUTDOWN); }
277	<	final boolean shutdownNow()   { return transitionRunStateTo(TERMINATING); }
278	<
279	<	/**
280	<	* Transitions to at least the given state.
281	<	*
282	<	* @return {@code true} if not already at least at given state
283	<	*/
284	<	private boolean transitionRunStateTo(int state) {
285	<	for (;;) {
286	<	int s = runState;
287	<	if (s >= state)
288	<	return false;
289	<	if (UNSAFE.compareAndSwapInt(this, runStateOffset, s, state))
290	<	return true;
291	<	}
292	<	}
293	<
294	<	/**
295	<	* Tries to set status to active; fails on contention.
296	<	*/
297	<	private boolean tryActivate() {
298	<	if (!active) {
299	<	if (!pool.tryIncrementActiveCount())
300	<	return false;
301	<	active = true;
302	<	}
303	<	return true;
304	<	}
305	<
306	<	/**
307	<	* Tries to set status to inactive; fails on contention.
308	<	*/
309	<	private boolean tryInactivate() {
310	<	if (active) {
311	<	if (!pool.tryDecrementActiveCount())
312	<	return false;
313	<	active = false;
314	<	}
315	<	return true;
316	<	}
317	<
318	<	/**
319	<	* Computes next value for random victim probe.  Scans don't
320	<	* require a very high quality generator, but also not a crummy
321	<	* one.  Marsaglia xor-shift is cheap and works well.
322	<	*/
323	<	private static int xorShift(int r) {
324	<	r ^= (r << 13);
325	<	r ^= (r >>> 17);
326	<	return r ^ (r << 5);
327	<	}
328	<
329	<	// Lifecycle methods
304	>	// Randomization
305
306		/**
307	<	* This method is required to be public, but should never be
308	<	* called explicitly. It performs the main run loop to execute
309	<	* ForkJoinTasks.
307	>	* Computes next value for random victim probes and backoffs.
308	>	* Scans don't require a very high quality generator, but also not
309	>	* a crummy one.  Marsaglia xor-shift is cheap and works well
310	>	* enough.  Note: This is manually inlined in FJP.scan() to avoid
311	>	* writes inside busy loops.
312		*/
313	<	public void run() {
314	<	Throwable exception = null;
315	<	try {
316	<	onStart();
317	<	pool.sync(this); // await first pool event
318	<	mainLoop();
342	<	} catch (Throwable ex) {
343	<	exception = ex;
344	<	} finally {
345	<	onTermination(exception);
346	<	}
313	>	private int nextSeed() {
314	>	int r = seed;
315	>	r ^= r << 13;
316	>	r ^= r >>> 17;
317	>	r ^= r << 5;
318	>	return seed = r;
319		}
320
321	<	/**
350	<	* Executes tasks until shut down.
351	<	*/
352	<	private void mainLoop() {
353	<	while (!isShutdown()) {
354	<	ForkJoinTask<?> t = pollTask();
355	<	if (t != null || (t = pollSubmission()) != null)
356	<	t.quietlyExec();
357	<	else if (tryInactivate())
358	<	pool.sync(this);
359	<	}
360	<	}
321	>	// Run State management
322
323		/**
324		* Initializes internal state after construction but before
325		* processing any tasks. If you override this method, you must
326	<	* invoke super.onStart() at the beginning of the method.
326	>	* invoke {@code super.onStart()} at the beginning of the method.
327		* Initialization requires care: Most fields must have legal
328		* default values, to ensure that attempted accesses from other
329		* threads work correctly even before this thread starts
330		* processing tasks.
331		*/
332		protected void onStart() {
372	–	// Allocate while starting to improve chances of thread-local
373	–	// isolation
333		queue = new ForkJoinTask<?>[INITIAL_QUEUE_CAPACITY];
334	<	// Initial value of seed need not be especially random but
335	<	// should differ across workers and must be nonzero
377	<	int p = poolIndex + 1;
378	<	seed = p + (p << 8) + (p << 16) + (p << 24); // spread bits
334	>	int r = ForkJoinPool.workerSeedGenerator.nextInt();
335	>	seed = (r == 0) ? 1 : r; //  must be nonzero
336		}
337
338		/**
#	Line 387 | Line 344 | public class ForkJoinWorkerThread extend
344		* to an unrecoverable error, or {@code null} if completed normally
345		*/
346		protected void onTermination(Throwable exception) {
390	–	// Execute remaining local tasks unless aborting or terminating
391	–	while (exception == null && pool.isProcessingTasks() && base != sp) {
392	–	try {
393	–	ForkJoinTask<?> t = popTask();
394	–	if (t != null)
395	–	t.quietlyExec();
396	–	} catch (Throwable ex) {
397	–	exception = ex;
398	–	}
399	–	}
400	–	// Cancel other tasks, transition status, notify pool, and
401	–	// propagate exception to uncaught exception handler
347		try {
348	<	do {} while (!tryInactivate()); // ensure inactive
348	>	terminate = true;
349		cancelTasks();
350	<	runState = TERMINATED;
406	<	pool.workerTerminated(this);
350	>	pool.deregisterWorker(this, exception);
351		} catch (Throwable ex) {        // Shouldn't ever happen
352		if (exception == null)      // but if so, at least rethrown
353		exception = ex;
354		} finally {
355		if (exception != null)
356	<	ForkJoinTask.rethrowException(exception);
356	>	UNSAFE.throwException(exception);
357		}
358		}
359
416	–	// Intrinsics-based support for queue operations.
417	–
360		/**
361	<	* Adds in store-order the given task at given slot of q to null.
362	<	* Caller must ensure q is non-null and index is in range.
361	>	* This method is required to be public, but should never be
362	>	* called explicitly. It performs the main run loop to execute
363	>	* {@link ForkJoinTask}s.
364		*/
365	<	private static void setSlot(ForkJoinTask<?>[] q, int i,
366	<	ForkJoinTask<?> t) {
367	<	UNSAFE.putOrderedObject(q, (i << qShift) + qBase, t);
365	>	public void run() {
366	>	Throwable exception = null;
367	>	try {
368	>	onStart();
369	>	pool.work(this);
370	>	} catch (Throwable ex) {
371	>	exception = ex;
372	>	} finally {
373	>	onTermination(exception);
374	>	}
375		}
376
377	+	/*
378	+	* Intrinsics-based atomic writes for queue slots. These are
379	+	* basically the same as methods in AtomicReferenceArray, but
380	+	* specialized for (1) ForkJoinTask elements (2) requirement that
381	+	* nullness and bounds checks have already been performed by
382	+	* callers and (3) effective offsets are known not to overflow
383	+	* from int to long (because of MAXIMUM_QUEUE_CAPACITY). We don't
384	+	* need corresponding version for reads: plain array reads are OK
385	+	* because they are protected by other volatile reads and are
386	+	* confirmed by CASes.
387	+	*
388	+	* Most uses don't actually call these methods, but instead
389	+	* contain inlined forms that enable more predictable
390	+	* optimization.  We don't define the version of write used in
391	+	* pushTask at all, but instead inline there a store-fenced array
392	+	* slot write.
393	+	*
394	+	* Also in most methods, as a performance (not correctness) issue,
395	+	* we'd like to encourage compilers not to arbitrarily postpone
396	+	* setting queueTop after writing slot.  Currently there is no
397	+	* intrinsic for arranging this, but using Unsafe putOrderedInt
398	+	* may be a preferable strategy on some compilers even though its
399	+	* main effect is a pre-, not post- fence. To simplify possible
400	+	* changes, the option is left in comments next to the associated
401	+	* assignments.
402	+	*/
403	+
404		/**
405	<	* CAS given slot of q to null. Caller must ensure q is non-null
406	<	* and index is in range.
405	>	* CASes slot i of array q from t to null. Caller must ensure q is
406	>	* non-null and index is in range.
407		*/
408	<	private static boolean casSlotNull(ForkJoinTask<?>[] q, int i,
409	<	ForkJoinTask<?> t) {
410	<	return UNSAFE.compareAndSwapObject(q, (i << qShift) + qBase, t, null);
408	>	private static final boolean casSlotNull(ForkJoinTask<?>[] q, int i,
409	>	ForkJoinTask<?> t) {
410	>	return UNSAFE.compareAndSwapObject(q, (i << ASHIFT) + ABASE, t, null);
411		}
412
413		/**
414	<	* Sets sp in store-order.
414	>	* Performs a volatile write of the given task at given slot of
415	>	* array q.  Caller must ensure q is non-null and index is in
416	>	* range. This method is used only during resets and backouts.
417		*/
418	<	private void storeSp(int s) {
419	<	UNSAFE.putOrderedInt(this, spOffset, s);
418	>	private static final void writeSlot(ForkJoinTask<?>[] q, int i,
419	>	ForkJoinTask<?> t) {
420	>	UNSAFE.putObjectVolatile(q, (i << ASHIFT) + ABASE, t);
421		}
422
423	<	// Main queue methods
423	>	// queue methods
424
425		/**
426	<	* Pushes a task. Called only by current thread.
426	>	* Pushes a task. Call only from this thread.
427		*
428		* @param t the task. Caller must ensure non-null.
429		*/
430		final void pushTask(ForkJoinTask<?> t) {
431	<	ForkJoinTask<?>[] q = queue;
432	<	int mask = q.length - 1;
433	<	int s = sp;
434	<	setSlot(q, s & mask, t);
435	<	storeSp(++s);
436	<	if ((s -= base) == 1)
437	<	pool.signalWork();
438	<	else if (s >= mask)
439	<	growQueue();
431	>	ForkJoinTask<?>[] q; int s, m;
432	>	if ((q = queue) != null) {    // ignore if queue removed
433	>	long u = (((s = queueTop) & (m = q.length - 1)) << ASHIFT) + ABASE;
434	>	UNSAFE.putOrderedObject(q, u, t);
435	>	queueTop = s + 1;         // or use putOrderedInt
436	>	if ((s -= queueBase) <= 2)
437	>	pool.signalWork();
438	>	else if (s == m)
439	>	growQueue();
440	>	}
441	>	}
442	>
443	>	/**
444	>	* Creates or doubles queue array.  Transfers elements by
445	>	* emulating steals (deqs) from old array and placing, oldest
446	>	* first, into new array.
447	>	*/
448	>	private void growQueue() {
449	>	ForkJoinTask<?>[] oldQ = queue;
450	>	int size = oldQ != null ? oldQ.length << 1 : INITIAL_QUEUE_CAPACITY;
451	>	if (size > MAXIMUM_QUEUE_CAPACITY)
452	>	throw new RejectedExecutionException("Queue capacity exceeded");
453	>	if (size < INITIAL_QUEUE_CAPACITY)
454	>	size = INITIAL_QUEUE_CAPACITY;
455	>	ForkJoinTask<?>[] q = queue = new ForkJoinTask<?>[size];
456	>	int mask = size - 1;
457	>	int top = queueTop;
458	>	int oldMask;
459	>	if (oldQ != null && (oldMask = oldQ.length - 1) >= 0) {
460	>	for (int b = queueBase; b != top; ++b) {
461	>	long u = ((b & oldMask) << ASHIFT) + ABASE;
462	>	Object x = UNSAFE.getObjectVolatile(oldQ, u);
463	>	if (x != null && UNSAFE.compareAndSwapObject(oldQ, u, x, null))
464	>	UNSAFE.putObjectVolatile
465	>	(q, ((b & mask) << ASHIFT) + ABASE, x);
466	>	}
467	>	}
468		}
469
470		/**
471		* Tries to take a task from the base of the queue, failing if
472	<	* either empty or contended.
472	>	* empty or contended. Note: Specializations of this code appear
473	>	* in locallyDeqTask and elsewhere.
474		*
475		* @return a task, or null if none or contended
476		*/
477		final ForkJoinTask<?> deqTask() {
478	<	ForkJoinTask<?> t;
479	<	ForkJoinTask<?>[] q;
471	<	int i;
472	<	int b;
473	<	if (sp != (b = base) &&
478	>	ForkJoinTask<?> t; ForkJoinTask<?>[] q; int b, i;
479	>	if (queueTop != (b = queueBase) &&
480		(q = queue) != null && // must read q after b
481	<	(t = q[i = (q.length - 1) & b]) != null &&
482	<	casSlotNull(q, i, t)) {
483	<	base = b + 1;
481	>	(i = (q.length - 1) & b) >= 0 &&
482	>	(t = q[i]) != null && queueBase == b &&
483	>	UNSAFE.compareAndSwapObject(q, (i << ASHIFT) + ABASE, t, null)) {
484	>	queueBase = b + 1;
485		return t;
486		}
487		return null;
488		}
489
490		/**
491	<	* Tries to take a task from the base of own queue, activating if
492	<	* necessary, failing only if empty. Called only by current thread.
491	>	* Tries to take a task from the base of own queue.  Called only
492	>	* by this thread.
493		*
494		* @return a task, or null if none
495		*/
496		final ForkJoinTask<?> locallyDeqTask() {
497	<	int b;
498	<	while (sp != (b = base)) {
499	<	if (tryActivate()) {
500	<	ForkJoinTask<?>[] q = queue;
501	<	int i = (q.length - 1) & b;
502	<	ForkJoinTask<?> t = q[i];
503	<	if (t != null && casSlotNull(q, i, t)) {
504	<	base = b + 1;
497	>	ForkJoinTask<?> t; int m, b, i;
498	>	ForkJoinTask<?>[] q = queue;
499	>	if (q != null && (m = q.length - 1) >= 0) {
500	>	while (queueTop != (b = queueBase)) {
501	>	if ((t = q[i = m & b]) != null &&
502	>	queueBase == b &&
503	>	UNSAFE.compareAndSwapObject(q, (i << ASHIFT) + ABASE,
504	>	t, null)) {
505	>	queueBase = b + 1;
506		return t;
507		}
508		}
#	Line 503 | Line 511 | public class ForkJoinWorkerThread extend
511		}
512
513		/**
514	<	* Returns a popped task, or null if empty. Ensures active status
515	<	* if non-null. Called only by current thread.
514	>	* Returns a popped task, or null if empty.
515	>	* Called only by this thread.
516		*/
517	<	final ForkJoinTask<?> popTask() {
518	<	int s = sp;
519	<	while (s != base) {
520	<	if (tryActivate()) {
521	<	ForkJoinTask<?>[] q = queue;
522	<	int mask = q.length - 1;
523	<	int i = (s - 1) & mask;
517	>	private ForkJoinTask<?> popTask() {
518	>	int m;
519	>	ForkJoinTask<?>[] q = queue;
520	>	if (q != null && (m = q.length - 1) >= 0) {
521	>	for (int s; (s = queueTop) != queueBase;) {
522	>	int i = m & --s;
523	>	long u = (i << ASHIFT) + ABASE; // raw offset
524		ForkJoinTask<?> t = q[i];
525	<	if (t == null || !casSlotNull(q, i, t))
525	>	if (t == null)   // lost to stealer
526		break;
527	<	storeSp(s - 1);
528	<	return t;
527	>	if (UNSAFE.compareAndSwapObject(q, u, t, null)) {
528	>	queueTop = s; // or putOrderedInt
529	>	return t;
530	>	}
531		}
532		}
533		return null;
534		}
535
536		/**
537	<	* Specialized version of popTask to pop only if
538	<	* topmost element is the given task. Called only
529	<	* by current thread while active.
537	>	* Specialized version of popTask to pop only if topmost element
538	>	* is the given task. Called only by this thread.
539		*
540		* @param t the task. Caller must ensure non-null.
541		*/
542		final boolean unpushTask(ForkJoinTask<?> t) {
543	<	ForkJoinTask<?>[] q = queue;
544	<	int mask = q.length - 1;
545	<	int s = sp - 1;
546	<	if (casSlotNull(q, s & mask, t)) {
547	<	storeSp(s);
543	>	ForkJoinTask<?>[] q;
544	>	int s;
545	>	if ((q = queue) != null && (s = queueTop) != queueBase &&
546	>	UNSAFE.compareAndSwapObject
547	>	(q, (((q.length - 1) & --s) << ASHIFT) + ABASE, t, null)) {
548	>	queueTop = s; // or putOrderedInt
549		return true;
550		}
551		return false;
552		}
553
554		/**
555	<	* Returns next task or null if empty or contended
555	>	* Returns next task, or null if empty or contended.
556		*/
557		final ForkJoinTask<?> peekTask() {
558	+	int m;
559		ForkJoinTask<?>[] q = queue;
560	<	if (q == null)
560	>	if (q == null || (m = q.length - 1) < 0)
561		return null;
562	<	int mask = q.length - 1;
563	<	int i = locallyFifo ? base : (sp - 1);
553	<	return q[i & mask];
562	>	int i = locallyFifo ? queueBase : (queueTop - 1);
563	>	return q[i & m];
564		}
565
566	+	// Support methods for ForkJoinPool
567	+
568		/**
569	<	* Doubles queue array size. Transfers elements by emulating
558	<	* steals (deqs) from old array and placing, oldest first, into
559	<	* new array.
569	>	* Runs the given task, plus any local tasks until queue is empty
570		*/
571	<	private void growQueue() {
572	<	ForkJoinTask<?>[] oldQ = queue;
573	<	int oldSize = oldQ.length;
574	<	int newSize = oldSize << 1;
575	<	if (newSize > MAXIMUM_QUEUE_CAPACITY)
576	<	throw new RejectedExecutionException("Queue capacity exceeded");
577	<	ForkJoinTask<?>[] newQ = queue = new ForkJoinTask<?>[newSize];
571	>	final void execTask(ForkJoinTask<?> t) {
572	>	currentSteal = t;
573	>	for (;;) {
574	>	if (t != null)
575	>	t.doExec();
576	>	if (queueTop == queueBase)
577	>	break;
578	>	t = locallyFifo ? locallyDeqTask() : popTask();
579	>	}
580	>	++stealCount;
581	>	currentSteal = null;
582	>	}
583
584	<	int b = base;
585	<	int bf = b + oldSize;
586	<	int oldMask = oldSize - 1;
587	<	int newMask = newSize - 1;
588	<	do {
589	<	int oldIndex = b & oldMask;
590	<	ForkJoinTask<?> t = oldQ[oldIndex];
591	<	if (t != null && !casSlotNull(oldQ, oldIndex, t))
592	<	t = null;
593	<	setSlot(newQ, b & newMask, t);
594	<	} while (++b != bf);
595	<	pool.signalWork();
584	>	/**
585	>	* Removes and cancels all tasks in queue.  Can be called from any
586	>	* thread.
587	>	*/
588	>	final void cancelTasks() {
589	>	ForkJoinTask<?> cj = currentJoin; // try to cancel ongoing tasks
590	>	if (cj != null && cj.status >= 0)
591	>	cj.cancelIgnoringExceptions();
592	>	ForkJoinTask<?> cs = currentSteal;
593	>	if (cs != null && cs.status >= 0)
594	>	cs.cancelIgnoringExceptions();
595	>	while (queueBase != queueTop) {
596	>	ForkJoinTask<?> t = deqTask();
597	>	if (t != null)
598	>	t.cancelIgnoringExceptions();
599	>	}
600		}
601
602		/**
603	<	* Tries to steal a task from another worker. Starts at a random
604	<	* index of workers array, and probes workers until finding one
605	<	* with non-empty queue or finding that all are empty.  It
606	<	* randomly selects the first n probes. If these are empty, it
607	<	* resorts to a full circular traversal, which is necessary to
608	<	* accurately set active status by caller. Also restarts if pool
609	<	* events occurred since last scan, which forces refresh of
610	<	* workers array, in case barrier was associated with resize.
611	<	*
612	<	* This method must be both fast and quiet -- usually avoiding
613	<	* memory accesses that could disrupt cache sharing etc other than
595	<	* those needed to check for and take tasks. This accounts for,
596	<	* among other things, updating random seed in place without
597	<	* storing it until exit.
598	<	*
599	<	* @return a task, or null if none found
600	<	*/
601	<	private ForkJoinTask<?> scan() {
602	<	ForkJoinTask<?> t = null;
603	<	int r = seed;                    // extract once to keep scan quiet
604	<	ForkJoinWorkerThread[] ws;       // refreshed on outer loop
605	<	int mask;                        // must be power 2 minus 1 and > 0
606	<	outer:do {
607	<	if ((ws = pool.workers) != null && (mask = ws.length - 1) > 0) {
608	<	int idx = r;
609	<	int probes = ~mask;      // use random index while negative
610	<	for (;;) {
611	<	r = xorShift(r);     // update random seed
612	<	ForkJoinWorkerThread v = ws[mask & idx];
613	<	if (v == null || v.sp == v.base) {
614	<	if (probes <= mask)
615	<	idx = (probes++ < 0) ? r : (idx + 1);
616	<	else
617	<	break;
618	<	}
619	<	else if (!tryActivate() || (t = v.deqTask()) == null)
620	<	continue outer;  // restart on contention
621	<	else
622	<	break outer;
623	<	}
603	>	* Drains tasks to given collection c.
604	>	*
605	>	* @return the number of tasks drained
606	>	*/
607	>	final int drainTasksTo(Collection<? super ForkJoinTask<?>> c) {
608	>	int n = 0;
609	>	while (queueBase != queueTop) {
610	>	ForkJoinTask<?> t = deqTask();
611	>	if (t != null) {
612	>	c.add(t);
613	>	++n;
614		}
615	<	} while (pool.hasNewSyncEvent(this)); // retry on pool events
616	<	seed = r;
627	<	return t;
615	>	}
616	>	return n;
617		}
618
619	+	// Support methods for ForkJoinTask
620	+
621		/**
622	<	* Gets and removes a local or stolen task.
632	<	*
633	<	* @return a task, if available
622	>	* Returns an estimate of the number of tasks in the queue.
623		*/
624	<	final ForkJoinTask<?> pollTask() {
625	<	ForkJoinTask<?> t = locallyFifo ? locallyDeqTask() : popTask();
637	<	if (t == null && (t = scan()) != null)
638	<	++stealCount;
639	<	return t;
624	>	final int getQueueSize() {
625	>	return queueTop - queueBase;
626		}
627
628		/**
629	<	* Gets a local task.
629	>	* Gets and removes a local task.
630		*
631		* @return a task, if available
632		*/
#	Line 649 | Line 635 | public class ForkJoinWorkerThread extend
635		}
636
637		/**
638	<	* Returns a pool submission, if one exists, activating first.
638	>	* Gets and removes a local or stolen task.
639		*
640	<	* @return a submission, if available
640	>	* @return a task, if available
641		*/
642	<	private ForkJoinTask<?> pollSubmission() {
643	<	ForkJoinPool p = pool;
644	<	while (p.hasQueuedSubmissions()) {
645	<	ForkJoinTask<?> t;
646	<	if (tryActivate() && (t = p.pollSubmission()) != null)
647	<	return t;
642	>	final ForkJoinTask<?> pollTask() {
643	>	ForkJoinWorkerThread[] ws;
644	>	ForkJoinTask<?> t = pollLocalTask();
645	>	if (t != null || (ws = pool.workers) == null)
646	>	return t;
647	>	int n = ws.length; // cheap version of FJP.scan
648	>	int steps = n << 1;
649	>	int r = nextSeed();
650	>	int i = 0;
651	>	while (i < steps) {
652	>	ForkJoinWorkerThread w = ws[(i++ + r) & (n - 1)];
653	>	if (w != null && w.queueBase != w.queueTop && w.queue != null) {
654	>	if ((t = w.deqTask()) != null)
655	>	return t;
656	>	i = 0;
657	>	}
658		}
659		return null;
660		}
661
666	–	// Methods accessed only by Pool
667	–
662		/**
663	<	* Removes and cancels all tasks in queue.  Can be called from any
664	<	* thread.
665	<	*/
666	<	final void cancelTasks() {
667	<	ForkJoinTask<?> t;
668	<	while (base != sp && (t = deqTask()) != null)
669	<	t.cancelIgnoringExceptions();
663	>	* The maximum stolen->joining link depth allowed in helpJoinTask,
664	>	* as well as the maximum number of retries (allowing on average
665	>	* one staleness retry per level) per attempt to instead try
666	>	* compensation.  Depths for legitimate chains are unbounded, but
667	>	* we use a fixed constant to avoid (otherwise unchecked) cycles
668	>	* and bound staleness of traversal parameters at the expense of
669	>	* sometimes blocking when we could be helping.
670	>	*/
671	>	private static final int MAX_HELP = 16;
672	>
673	>	/**
674	>	* Possibly runs some tasks and/or blocks, until joinMe is done.
675	>	*
676	>	* @param joinMe the task to join
677	>	* @return completion status on exit
678	>	*/
679	>	final int joinTask(ForkJoinTask<?> joinMe) {
680	>	ForkJoinTask<?> prevJoin = currentJoin;
681	>	currentJoin = joinMe;
682	>	for (int s, retries = MAX_HELP;;) {
683	>	if ((s = joinMe.status) < 0) {
684	>	currentJoin = prevJoin;
685	>	return s;
686	>	}
687	>	if (retries > 0) {
688	>	if (queueTop != queueBase) {
689	>	if (!localHelpJoinTask(joinMe))
690	>	retries = 0;           // cannot help
691	>	}
692	>	else if (retries == MAX_HELP >>> 1) {
693	>	--retries;                 // check uncommon case
694	>	if (tryDeqAndExec(joinMe) >= 0)
695	>	Thread.yield();        // for politeness
696	>	}
697	>	else
698	>	retries = helpJoinTask(joinMe) ? MAX_HELP : retries - 1;
699	>	}
700	>	else {
701	>	retries = MAX_HELP;           // restart if not done
702	>	pool.tryAwaitJoin(joinMe);
703	>	}
704	>	}
705		}
706
707		/**
708	<	* Drains tasks to given collection c.
708	>	* If present, pops and executes the given task, or any other
709	>	* cancelled task
710		*
711	<	* @return the number of tasks drained
711	>	* @return false if any other non-cancelled task exists in local queue
712		*/
713	<	final int drainTasksTo(Collection<? super ForkJoinTask<?>> c) {
714	<	int n = 0;
715	<	ForkJoinTask<?> t;
716	<	while (base != sp && (t = deqTask()) != null) {
717	<	c.add(t);
718	<	++n;
713	>	private boolean localHelpJoinTask(ForkJoinTask<?> joinMe) {
714	>	int s, i; ForkJoinTask<?>[] q; ForkJoinTask<?> t;
715	>	if ((s = queueTop) != queueBase && (q = queue) != null &&
716	>	(i = (q.length - 1) & --s) >= 0 &&
717	>	(t = q[i]) != null) {
718	>	if (t != joinMe && t.status >= 0)
719	>	return false;
720	>	if (UNSAFE.compareAndSwapObject
721	>	(q, (i << ASHIFT) + ABASE, t, null)) {
722	>	queueTop = s;           // or putOrderedInt
723	>	t.doExec();
724	>	}
725		}
726	<	return n;
726	>	return true;
727		}
728
729		/**
730	<	* Gets and clears steal count for accumulation by pool.  Called
731	<	* only when known to be idle (in pool.sync and termination).
732	<	*/
733	<	final int getAndClearStealCount() {
734	<	int sc = stealCount;
735	<	stealCount = 0;
736	<	return sc;
730	>	* Tries to locate and execute tasks for a stealer of the given
731	>	* task, or in turn one of its stealers, Traces
732	>	* currentSteal->currentJoin links looking for a thread working on
733	>	* a descendant of the given task and with a non-empty queue to
734	>	* steal back and execute tasks from.  The implementation is very
735	>	* branchy to cope with potential inconsistencies or loops
736	>	* encountering chains that are stale, unknown, or of length
737	>	* greater than MAX_HELP links.  All of these cases are dealt with
738	>	* by just retrying by caller.
739	>	*
740	>	* @param joinMe the task to join
741	>	* @param canSteal true if local queue is empty
742	>	* @return true if ran a task
743	>	*/
744	>	private boolean helpJoinTask(ForkJoinTask<?> joinMe) {
745	>	boolean helped = false;
746	>	int m = pool.scanGuard & SMASK;
747	>	ForkJoinWorkerThread[] ws = pool.workers;
748	>	if (ws != null && ws.length > m && joinMe.status >= 0) {
749	>	int levels = MAX_HELP;              // remaining chain length
750	>	ForkJoinTask<?> task = joinMe;      // base of chain
751	>	outer:for (ForkJoinWorkerThread thread = this;;) {
752	>	// Try to find v, the stealer of task, by first using hint
753	>	ForkJoinWorkerThread v = ws[thread.stealHint & m];
754	>	if (v == null || v.currentSteal != task) {
755	>	for (int j = 0; ;) {        // search array
756	>	if ((v = ws[j]) != null && v.currentSteal == task) {
757	>	thread.stealHint = j;
758	>	break;              // save hint for next time
759	>	}
760	>	if (++j > m)
761	>	break outer;        // can't find stealer
762	>	}
763	>	}
764	>	// Try to help v, using specialized form of deqTask
765	>	for (;;) {
766	>	ForkJoinTask<?>[] q; int b, i;
767	>	if (joinMe.status < 0)
768	>	break outer;
769	>	if ((b = v.queueBase) == v.queueTop ||
770	>	(q = v.queue) == null ||
771	>	(i = (q.length-1) & b) < 0)
772	>	break;                  // empty
773	>	long u = (i << ASHIFT) + ABASE;
774	>	ForkJoinTask<?> t = q[i];
775	>	if (task.status < 0)
776	>	break outer;            // stale
777	>	if (t != null && v.queueBase == b &&
778	>	UNSAFE.compareAndSwapObject(q, u, t, null)) {
779	>	v.queueBase = b + 1;
780	>	v.stealHint = poolIndex;
781	>	ForkJoinTask<?> ps = currentSteal;
782	>	currentSteal = t;
783	>	t.doExec();
784	>	currentSteal = ps;
785	>	helped = true;
786	>	}
787	>	}
788	>	// Try to descend to find v's stealer
789	>	ForkJoinTask<?> next = v.currentJoin;
790	>	if (--levels > 0 && task.status >= 0 &&
791	>	next != null && next != task) {
792	>	task = next;
793	>	thread = v;
794	>	}
795	>	else
796	>	break;  // max levels, stale, dead-end, or cyclic
797	>	}
798	>	}
799	>	return helped;
800		}
801
802		/**
803	<	* Returns {@code true} if at least one worker in the given array
804	<	* appears to have at least one queued task.
803	>	* Performs an uncommon case for joinTask: If task t is at base of
804	>	* some workers queue, steals and executes it.
805		*
806	<	* @param ws array of workers
807	<	*/
808	<	static boolean hasQueuedTasks(ForkJoinWorkerThread[] ws) {
809	<	if (ws != null) {
810	<	int len = ws.length;
811	<	for (int j = 0; j < 2; ++j) { // need two passes for clean sweep
812	<	for (int i = 0; i < len; ++i) {
813	<	ForkJoinWorkerThread w = ws[i];
814	<	if (w != null && w.sp != w.base)
815	<	return true;
806	>	* @param t the task
807	>	* @return t's status
808	>	*/
809	>	private int tryDeqAndExec(ForkJoinTask<?> t) {
810	>	int m = pool.scanGuard & SMASK;
811	>	ForkJoinWorkerThread[] ws = pool.workers;
812	>	if (ws != null && ws.length > m && t.status >= 0) {
813	>	for (int j = 0; j <= m; ++j) {
814	>	ForkJoinTask<?>[] q; int b, i;
815	>	ForkJoinWorkerThread v = ws[j];
816	>	if (v != null &&
817	>	(b = v.queueBase) != v.queueTop &&
818	>	(q = v.queue) != null &&
819	>	(i = (q.length - 1) & b) >= 0 &&
820	>	q[i] ==  t) {
821	>	long u = (i << ASHIFT) + ABASE;
822	>	if (v.queueBase == b &&
823	>	UNSAFE.compareAndSwapObject(q, u, t, null)) {
824	>	v.queueBase = b + 1;
825	>	v.stealHint = poolIndex;
826	>	ForkJoinTask<?> ps = currentSteal;
827	>	currentSteal = t;
828	>	t.doExec();
829	>	currentSteal = ps;
830	>	}
831	>	break;
832		}
833		}
834		}
835	<	return false;
835	>	return t.status;
836		}
837
723	–	// Support methods for ForkJoinTask
724	–
838		/**
839	<	* Returns an estimate of the number of tasks in the queue.
840	<	*/
841	<	final int getQueueSize() {
842	<	// suppress momentarily negative values
843	<	return Math.max(0, sp - base);
844	<	}
845	<
846	<	/**
847	<	* Returns an estimate of the number of tasks, offset by a
848	<	* function of number of idle workers.
839	>	* Implements ForkJoinTask.getSurplusQueuedTaskCount().  Returns
840	>	* an estimate of the number of tasks, offset by a function of
841	>	* number of idle workers.
842	>	*
843	>	* This method provides a cheap heuristic guide for task
844	>	* partitioning when programmers, frameworks, tools, or languages
845	>	* have little or no idea about task granularity.  In essence by
846	>	* offering this method, we ask users only about tradeoffs in
847	>	* overhead vs expected throughput and its variance, rather than
848	>	* how finely to partition tasks.
849	>	*
850	>	* In a steady state strict (tree-structured) computation, each
851	>	* thread makes available for stealing enough tasks for other
852	>	* threads to remain active. Inductively, if all threads play by
853	>	* the same rules, each thread should make available only a
854	>	* constant number of tasks.
855	>	*
856	>	* The minimum useful constant is just 1. But using a value of 1
857	>	* would require immediate replenishment upon each steal to
858	>	* maintain enough tasks, which is infeasible.  Further,
859	>	* partitionings/granularities of offered tasks should minimize
860	>	* steal rates, which in general means that threads nearer the top
861	>	* of computation tree should generate more than those nearer the
862	>	* bottom. In perfect steady state, each thread is at
863	>	* approximately the same level of computation tree. However,
864	>	* producing extra tasks amortizes the uncertainty of progress and
865	>	* diffusion assumptions.
866	>	*
867	>	* So, users will want to use values larger, but not much larger
868	>	* than 1 to both smooth over transient shortages and hedge
869	>	* against uneven progress; as traded off against the cost of
870	>	* extra task overhead. We leave the user to pick a threshold
871	>	* value to compare with the results of this call to guide
872	>	* decisions, but recommend values such as 3.
873	>	*
874	>	* When all threads are active, it is on average OK to estimate
875	>	* surplus strictly locally. In steady-state, if one thread is
876	>	* maintaining say 2 surplus tasks, then so are others. So we can
877	>	* just use estimated queue length (although note that (queueTop -
878	>	* queueBase) can be an overestimate because of stealers lagging
879	>	* increments of queueBase).  However, this strategy alone leads
880	>	* to serious mis-estimates in some non-steady-state conditions
881	>	* (ramp-up, ramp-down, other stalls). We can detect many of these
882	>	* by further considering the number of "idle" threads, that are
883	>	* known to have zero queued tasks, so compensate by a factor of
884	>	* (#idle/#active) threads.
885		*/
886		final int getEstimatedSurplusTaskCount() {
887	<	// The halving approximates weighting idle vs non-idle workers
739	<	return (sp - base) - (pool.getIdleThreadCount() >>> 1);
740	<	}
741	<
742	<	/**
743	<	* Scans, returning early if joinMe done.
744	<	*/
745	<	final ForkJoinTask<?> scanWhileJoining(ForkJoinTask<?> joinMe) {
746	<	ForkJoinTask<?> t = pollTask();
747	<	if (t != null && joinMe.status < 0 && sp == base) {
748	<	pushTask(t); // unsteal if done and this task would be stealable
749	<	t = null;
750	<	}
751	<	return t;
887	>	return queueTop - queueBase - pool.idlePerActive();
888		}
889
890		/**
891	<	* Runs tasks until {@code pool.isQuiescent()}.
891	>	* Runs tasks until {@code pool.isQuiescent()}. We piggyback on
892	>	* pool's active count ctl maintenance, but rather than blocking
893	>	* when tasks cannot be found, we rescan until all others cannot
894	>	* find tasks either. The bracketing by pool quiescerCounts
895	>	* updates suppresses pool auto-shutdown mechanics that could
896	>	* otherwise prematurely terminate the pool because all threads
897	>	* appear to be inactive.
898		*/
899		final void helpQuiescePool() {
900	+	boolean active = true;
901	+	ForkJoinTask<?> ps = currentSteal; // to restore below
902	+	ForkJoinPool p = pool;
903	+	p.addQuiescerCount(1);
904		for (;;) {
905	<	ForkJoinTask<?> t = pollTask();
906	<	if (t != null)
907	<	t.quietlyExec();
908	<	else if (tryInactivate() && pool.isQuiescent())
909	<	break;
905	>	ForkJoinWorkerThread[] ws = p.workers;
906	>	ForkJoinWorkerThread v = null;
907	>	int n;
908	>	if (queueTop != queueBase)
909	>	v = this;
910	>	else if (ws != null && (n = ws.length) > 1) {
911	>	ForkJoinWorkerThread w;
912	>	int r = nextSeed(); // cheap version of FJP.scan
913	>	int steps = n << 1;
914	>	for (int i = 0; i < steps; ++i) {
915	>	if ((w = ws[(i + r) & (n - 1)]) != null &&
916	>	w.queueBase != w.queueTop) {
917	>	v = w;
918	>	break;
919	>	}
920	>	}
921	>	}
922	>	if (v != null) {
923	>	ForkJoinTask<?> t;
924	>	if (!active) {
925	>	active = true;
926	>	p.addActiveCount(1);
927	>	}
928	>	if ((t = (v != this) ? v.deqTask() :
929	>	locallyFifo ? locallyDeqTask() : popTask()) != null) {
930	>	currentSteal = t;
931	>	t.doExec();
932	>	currentSteal = ps;
933	>	}
934	>	}
935	>	else {
936	>	if (active) {
937	>	active = false;
938	>	p.addActiveCount(-1);
939	>	}
940	>	if (p.isQuiescent()) {
941	>	p.addActiveCount(1);
942	>	p.addQuiescerCount(-1);
943	>	break;
944	>	}
945	>	}
946		}
765	–	do {} while (!tryActivate()); // re-activate on exit
947		}
948
949		// Unsafe mechanics
950	<
951	<	private static final sun.misc.Unsafe UNSAFE = getUnsafe();
952	<	private static final long spOffset =
772	<	objectFieldOffset("sp", ForkJoinWorkerThread.class);
773	<	private static final long runStateOffset =
774	<	objectFieldOffset("runState", ForkJoinWorkerThread.class);
775	<	private static final long qBase;
776	<	private static final int qShift;
950	>	private static final sun.misc.Unsafe UNSAFE;
951	>	private static final long ABASE;
952	>	private static final int ASHIFT;
953
954		static {
955	<	qBase = UNSAFE.arrayBaseOffset(ForkJoinTask[].class);
780	<	int s = UNSAFE.arrayIndexScale(ForkJoinTask[].class);
781	<	if ((s & (s-1)) != 0)
782	<	throw new Error("data type scale not a power of two");
783	<	qShift = 31 - Integer.numberOfLeadingZeros(s);
784	<	}
785	<
786	<	private static long objectFieldOffset(String field, Class<?> klazz) {
955	>	int s;
956		try {
957	<	return UNSAFE.objectFieldOffset(klazz.getDeclaredField(field));
958	<	} catch (NoSuchFieldException e) {
959	<	// Convert Exception to corresponding Error
960	<	NoSuchFieldError error = new NoSuchFieldError(field);
961	<	error.initCause(e);
962	<	throw error;
957	>	UNSAFE = getUnsafe();
958	>	Class<?> a = ForkJoinTask[].class;
959	>	ABASE = UNSAFE.arrayBaseOffset(a);
960	>	s = UNSAFE.arrayIndexScale(a);
961	>	} catch (Exception e) {
962	>	throw new Error(e);
963		}
964	+	if ((s & (s-1)) != 0)
965	+	throw new Error("data type scale not a power of two");
966	+	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
967		}
968
969		/**

Diff Legend

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