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 1.22 by dl, Wed Jul 29 12:05:55 2009 UTC vs.
Revision 1.41 by dl, Tue Aug 17 18:30:33 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
17		* subclassable solely for the sake of adding functionality -- there
18	<	* are no overridable methods dealing with scheduling or
19	<	* execution. However, you can override initialization and termination
20	<	* methods surrounding the main task processing loop.  If you do
21	<	* create such a subclass, you will also need to supply a custom
22	<	* ForkJoinWorkerThreadFactory to use it in a ForkJoinPool.
18	>	* are no overridable methods dealing with scheduling or execution.
19	>	* However, you can override initialization and termination methods
20	>	* surrounding the main task processing loop.  If you do create such a
21	>	* subclass, you will also need to supply a custom {@link
22	>	* ForkJoinPool.ForkJoinWorkerThreadFactory} to use it in a {@code
23	>	* ForkJoinPool}.
24		*
25		* @since 1.7
26		* @author Doug Lea
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
82	>	* version of deq rather than pop when locallyFifo is true (as set
83	>	* by the ForkJoinPool).  This allows use in message-passing
84	>	* frameworks in which tasks are never joined.
85	>	*
86	>	* When a worker would otherwise be blocked waiting to join a
87	>	* task, it first tries a form of linear helping: Each worker
88	>	* records (in field currentSteal) the most recent task it stole
89	>	* from some other worker. Plus, it records (in field currentJoin)
90	>	* the task it is currently actively joining. Method joinTask uses
91	>	* these markers to try to find a worker to help (i.e., steal back
92	>	* a task from and execute it) that could hasten completion of the
93	>	* actively joined task. In essence, the joiner executes a task
94	>	* that would be on its own local deque had the to-be-joined task
95	>	* not been stolen. This may be seen as a conservative variant of
96	>	* the approach in Wagner & Calder "Leapfrogging: a portable
97	>	* technique for implementing efficient futures" SIGPLAN Notices,
98	>	* 1993 (http://portal.acm.org/citation.cfm?id=155354). It differs
99	>	* in that: (1) We only maintain dependency links across workers
100	>	* upon steals, rather than use per-task bookkeeping.  This may
101	>	* require a linear scan of workers array to locate stealers, but
102	>	* usually doesn't because stealers leave hints (that may become
103	>	* stale/wrong) of where to locate them. This isolates cost to
104	>	* when it is needed, rather than adding to per-task overhead.
105	>	* (2) It is "shallow", ignoring nesting and potentially cyclic
106	>	* mutual steals.  (3) It is intentionally racy: field currentJoin
107	>	* is updated only while actively joining, which means that we
108	>	* miss links in the chain during long-lived tasks, GC stalls etc
109	>	* (which is OK since blocking in such cases is usually a good
110	>	* idea).  (4) We bound the number of attempts to find work (see
111	>	* MAX_HELP_DEPTH) and fall back to suspending the worker and if
112	>	* necessary replacing it with a spare (see
113	>	* ForkJoinPool.tryAwaitJoin).
114		*
115	<	* Efficient implementation of this approach currently relies on
116	<	* an uncomfortable amount of "Unsafe" mechanics. To maintain
115	>	* Efficient implementation of these algorithms currently relies
116	>	* on an uncomfortable amount of "Unsafe" mechanics. To maintain
117		* correct orderings, reads and writes of variable base require
118	<	* volatile ordering.  Variable sp does not require volatile write
119	<	* but needs cheaper store-ordering on writes.  Because they are
120	<	* protected by volatile base reads, reads of the queue array and
121	<	* its slots do not need volatile load semantics, but writes (in
122	<	* push) require store order and CASes (in pop and deq) require
123	<	* (volatile) CAS semantics. Since these combinations aren't
124	<	* supported using ordinary volatiles, the only way to accomplish
125	<	* these efficiently is to use direct Unsafe calls. (Using external
126	<	* AtomicIntegers and AtomicReferenceArrays for the indices and
127	<	* array is significantly slower because of memory locality and
128	<	* indirection effects.) Further, performance on most platforms is
129	<	* very sensitive to placement and sizing of the (resizable) queue
130	<	* array.  Even though these queues don't usually become all that
131	<	* big, the initial size must be large enough to counteract cache
118	>	* volatile ordering.  Variable sp does not require volatile
119	>	* writes but still needs store-ordering, which we accomplish by
120	>	* pre-incrementing sp before filling the slot with an ordered
121	>	* store.  (Pre-incrementing also enables backouts used in
122	>	* joinTask.)  Because they are protected by volatile base reads,
123	>	* reads of the queue array and its slots by other threads do not
124	>	* need volatile load semantics, but writes (in push) require
125	>	* store order and CASes (in pop and deq) require (volatile) CAS
126	>	* semantics.  (Michael, Saraswat, and Vechev's algorithm has
127	>	* similar properties, but without support for nulling slots.)
128	>	* Since these combinations aren't supported using ordinary
129	>	* volatiles, the only way to accomplish these efficiently is to
130	>	* use direct Unsafe calls. (Using external AtomicIntegers and
131	>	* AtomicReferenceArrays for the indices and array is
132	>	* significantly slower because of memory locality and indirection
133	>	* effects.)
134	>	*
135	>	* Further, performance on most platforms is very sensitive to
136	>	* placement and sizing of the (resizable) queue array.  Even
137	>	* though these queues don't usually become all that big, the
138	>	* initial size must be large enough to counteract cache
139		* contention effects across multiple queues (especially in the
140		* presence of GC cardmarking). Also, to improve thread-locality,
141	<	* queues are currently initialized immediately after the thread
142	<	* gets the initial signal to start processing tasks.  However,
143	<	* all queue-related methods except pushTask are written in a way
144	<	* that allows them to instead be lazily allocated and/or disposed
145	<	* of when empty. All together, these low-level implementation
92	<	* choices produce as much as a factor of 4 performance
93	<	* improvement compared to naive implementations, and enable the
94	<	* processing of billions of tasks per second, sometimes at the
95	<	* expense of ugliness.
96	<	*
97	<	* 2. Run control: The primary run control is based on a global
98	<	* counter (activeCount) held by the pool. It uses an algorithm
99	<	* similar to that in Herlihy and Shavit section 17.6 to cause
100	<	* threads to eventually block when all threads declare they are
101	<	* inactive. (See variable "scans".)  For this to work, threads
102	<	* must be declared active when executing tasks, and before
103	<	* stealing a task. They must be inactive before blocking on the
104	<	* Pool Barrier (awaiting a new submission or other Pool
105	<	* event). In between, there is some free play which we take
106	<	* advantage of to avoid contention and rapid flickering of the
107	<	* global activeCount: If inactive, we activate only if a victim
108	<	* queue appears to be nonempty (see above).  Similarly, a thread
109	<	* tries to inactivate only after a full scan of other threads.
110	<	* The net effect is that contention on activeCount is rarely a
111	<	* measurable performance issue. (There are also a few other cases
112	<	* where we scan for work rather than retry/block upon
113	<	* contention.)
114	<	*
115	<	* 3. Selection control. We maintain policy of always choosing to
116	<	* run local tasks rather than stealing, and always trying to
117	<	* steal tasks before trying to run a new submission. All steals
118	<	* are currently performed in randomly-chosen deq-order. It may be
119	<	* worthwhile to bias these with locality / anti-locality
120	<	* information, but doing this well probably requires more
121	<	* lower-level information from JVMs than currently provided.
141	>	* queues are initialized after starting.  All together, these
142	>	* low-level implementation choices produce as much as a factor of
143	>	* 4 performance improvement compared to naive implementations,
144	>	* and enable the processing of billions of tasks per second,
145	>	* sometimes at the expense of ugliness.
146		*/
147
148		/**
149	+	* Generator for initial random seeds for random victim
150	+	* selection. This is used only to create initial seeds. Random
151	+	* steals use a cheaper xorshift generator per steal attempt. We
152	+	* expect only rare contention on seedGenerator, so just use a
153	+	* plain Random.
154	+	*/
155	+	private static final Random seedGenerator = new Random();
156	+
157	+	/**
158	+	* The maximum stolen->joining link depth allowed in helpJoinTask.
159	+	* Depths for legitimate chains are unbounded, but we use a fixed
160	+	* constant to avoid (otherwise unchecked) cycles and bound
161	+	* staleness of traversal parameters at the expense of sometimes
162	+	* blocking when we could be helping.
163	+	*/
164	+	private static final int MAX_HELP_DEPTH = 8;
165	+
166	+	/**
167	+	* The wakeup interval (in nanoseconds) for the oldest worker
168	+	* suspended as spare.  On each wakeup not signalled by a
169	+	* resumption, it may ask the pool to reduce the number of spares.
170	+	*/
171	+	private static final long TRIM_RATE_NANOS =
172	+	5L * 1000L * 1000L * 1000L; // 5sec
173	+
174	+	/**
175		* Capacity of work-stealing queue array upon initialization.
176	<	* Must be a power of two. Initial size must be at least 2, but is
176	>	* Must be a power of two. Initial size must be at least 4, but is
177		* padded to minimize cache effects.
178		*/
179		private static final int INITIAL_QUEUE_CAPACITY = 1 << 13;
#	Line 143 | Line 193 | public class ForkJoinWorkerThread extend
193
194		/**
195		* The work-stealing queue array. Size must be a power of two.
196	<	* Initialized when thread starts, to improve memory locality.
196	>	* Initialized in onStart, to improve memory locality.
197		*/
198		private ForkJoinTask<?>[] queue;
199
200		/**
151	–	* Index (mod queue.length) of next queue slot to push to or pop
152	–	* from. It is written only by owner thread, via ordered store.
153	–	* Both sp and base are allowed to wrap around on overflow, but
154	–	* (sp - base) still estimates size.
155	–	*/
156	–	private volatile int sp;
157	–
158	–	/**
201		* Index (mod queue.length) of least valid queue slot, which is
202		* always the next position to steal from if nonempty.
203		*/
204		private volatile int base;
205
206		/**
207	<	* Activity status. When true, this worker is considered active.
208	<	* Must be false upon construction. It must be true when executing
209	<	* tasks, and BEFORE stealing a task. It must be false before
210	<	* calling pool.sync.
207	>	* Index (mod queue.length) of next queue slot to push to or pop
208	>	* from. It is written only by owner thread, and accessed by other
209	>	* threads only after reading (volatile) base.  Both sp and base
210	>	* are allowed to wrap around on overflow, but (sp - base) still
211	>	* estimates size.
212	>	*/
213	>	private int sp;
214	>
215	>	/**
216	>	* The index of most recent stealer, used as a hint to avoid
217	>	* traversal in method helpJoinTask. This is only a hint because a
218	>	* worker might have had multiple steals and this only holds one
219	>	* of them (usually the most current). Declared non-volatile,
220	>	* relying on other prevailing sync to keep reasonably current.
221		*/
222	<	private boolean active;
222	>	private int stealHint;
223
224		/**
225	<	* Run state of this worker. Supports simple versions of the usual
226	<	* shutdown/shutdownNow control.
225	>	* Run state of this worker. In addition to the usual run levels,
226	>	* tracks if this worker is suspended as a spare, and if it was
227	>	* killed (trimmed) while suspended. However, "active" status is
228	>	* maintained separately and modified only in conjunction with
229	>	* CASes of the pool's runState (which are currently sadly manually
230	>	* inlined for performance.)
231		*/
232		private volatile int runState;
233
234	+	private static final int TERMINATING = 0x01;
235	+	private static final int TERMINATED  = 0x02;
236	+	private static final int SUSPENDED   = 0x04; // inactive spare
237	+	private static final int TRIMMED     = 0x08; // killed while suspended
238	+
239	+	/**
240	+	* Number of steals, transferred and reset in pool callbacks pool
241	+	* when idle Accessed directly by pool.
242	+	*/
243	+	int stealCount;
244	+
245		/**
246		* Seed for random number generator for choosing steal victims.
247	<	* Uses Marsaglia xorshift. Must be nonzero upon initialization.
247	>	* Uses Marsaglia xorshift. Must be initialized as nonzero.
248		*/
249		private int seed;
250
251		/**
252	<	* Number of steals, transferred to pool when idle
252	>	* Activity status. When true, this worker is considered active.
253	>	* Accessed directly by pool.  Must be false upon construction.
254	>	*/
255	>	boolean active;
256	>
257	>	/**
258	>	* True if use local fifo, not default lifo, for local polling.
259	>	* Shadows value from ForkJoinPool.
260		*/
261	<	private int stealCount;
261	>	private final boolean locallyFifo;
262
263		/**
264		* Index of this worker in pool array. Set once by pool before
265	<	* running, and accessed directly by pool during cleanup etc.
265	>	* running, and accessed directly by pool to locate this worker in
266	>	* its workers array.
267		*/
268		int poolIndex;
269
270		/**
271	<	* The last barrier event waited for. Accessed in pool callback
272	<	* methods, but only by current thread.
271	>	* The last pool event waited for. Accessed only by pool in
272	>	* callback methods invoked within this thread.
273	>	*/
274	>	int lastEventCount;
275	>
276	>	/**
277	>	* Encoded index and event count of next event waiter. Used only
278	>	* by ForkJoinPool for managing event waiters.
279	>	*/
280	>	volatile long nextWaiter;
281	>
282	>	/**
283	>	* Number of times this thread suspended as spare
284	>	*/
285	>	int spareCount;
286	>
287	>	/**
288	>	* Encoded index and count of next spare waiter. Used only
289	>	* by ForkJoinPool for managing spares.
290		*/
291	<	long lastEventCount;
291	>	volatile int nextSpare;
292
293		/**
294	<	* True if use local fifo, not default lifo, for local polling
294	>	* The task currently being joined, set only when actively trying
295	>	* to helpStealer. Written only by current thread, but read by
296	>	* others.
297		*/
298	<	private boolean locallyFifo;
298	>	private volatile ForkJoinTask<?> currentJoin;
299	>
300	>	/**
301	>	* The task most recently stolen from another worker (or
302	>	* submission queue).  Not volatile because always read/written in
303	>	* presence of related volatiles in those cases where it matters.
304	>	*/
305	>	private ForkJoinTask<?> currentSteal;
306
307		/**
308		* Creates a ForkJoinWorkerThread operating in the given pool.
#	Line 210 | Line 311 | public class ForkJoinWorkerThread extend
311		* @throws NullPointerException if pool is null
312		*/
313		protected ForkJoinWorkerThread(ForkJoinPool pool) {
213	–	if (pool == null) throw new NullPointerException();
314		this.pool = pool;
315	<	// Note: poolIndex is set by pool during construction
316	<	// Remaining initialization is deferred to onStart
315	>	this.locallyFifo = pool.locallyFifo;
316	>	setDaemon(true);
317	>	// To avoid exposing construction details to subclasses,
318	>	// remaining initialization is in start() and onStart()
319		}
320
321	<	// Public access methods
321	>	/**
322	>	* Performs additional initialization and starts this thread
323	>	*/
324	>	final void start(int poolIndex, UncaughtExceptionHandler ueh) {
325	>	this.poolIndex = poolIndex;
326	>	if (ueh != null)
327	>	setUncaughtExceptionHandler(ueh);
328	>	start();
329	>	}
330	>
331	>	// Public/protected methods
332
333		/**
334		* Returns the pool hosting this thread.
#	Line 241 | Line 353 | public class ForkJoinWorkerThread extend
353		}
354
355		/**
356	<	* Establishes local first-in-first-out scheduling mode for forked
357	<	* tasks that are never joined.
358	<	*
359	<	* @param async if true, use locally FIFO scheduling
356	>	* Initializes internal state after construction but before
357	>	* processing any tasks. If you override this method, you must
358	>	* invoke super.onStart() at the beginning of the method.
359	>	* Initialization requires care: Most fields must have legal
360	>	* default values, to ensure that attempted accesses from other
361	>	* threads work correctly even before this thread starts
362	>	* processing tasks.
363		*/
364	<	void setAsyncMode(boolean async) {
365	<	locallyFifo = async;
366	<	}
252	<
253	<	// Runstate management
254	<
255	<	// Runstate values. Order matters
256	<	private static final int RUNNING     = 0;
257	<	private static final int SHUTDOWN    = 1;
258	<	private static final int TERMINATING = 2;
259	<	private static final int TERMINATED  = 3;
364	>	protected void onStart() {
365	>	int rs = seedGenerator.nextInt();
366	>	seed = rs == 0? 1 : rs; // seed must be nonzero
367
368	<	final boolean isShutdown()    { return runState >= SHUTDOWN;  }
369	<	final boolean isTerminating() { return runState >= TERMINATING;  }
370	<	final boolean isTerminated()  { return runState == TERMINATED; }
371	<	final boolean shutdown()      { return transitionRunStateTo(SHUTDOWN); }
265	<	final boolean shutdownNow()   { return transitionRunStateTo(TERMINATING); }
368	>	// Allocate name string and arrays in this thread
369	>	String pid = Integer.toString(pool.getPoolNumber());
370	>	String wid = Integer.toString(poolIndex);
371	>	setName("ForkJoinPool-" + pid + "-worker-" + wid);
372
373	<	/**
268	<	* Transitions to at least the given state.
269	<	*
270	<	* @return {@code true} if not already at least at given state
271	<	*/
272	<	private boolean transitionRunStateTo(int state) {
273	<	for (;;) {
274	<	int s = runState;
275	<	if (s >= state)
276	<	return false;
277	<	if (UNSAFE.compareAndSwapInt(this, runStateOffset, s, state))
278	<	return true;
279	<	}
280	<	}
281	<
282	<	/**
283	<	* Tries to set status to active; fails on contention.
284	<	*/
285	<	private boolean tryActivate() {
286	<	if (!active) {
287	<	if (!pool.tryIncrementActiveCount())
288	<	return false;
289	<	active = true;
290	<	}
291	<	return true;
373	>	queue = new ForkJoinTask<?>[INITIAL_QUEUE_CAPACITY];
374		}
375
376		/**
377	<	* Tries to set status to inactive; fails on contention.
377	>	* 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.
380	>	*
381	>	* @param exception the exception causing this thread to abort due
382	>	* to an unrecoverable error, or {@code null} if completed normally
383		*/
384	<	private boolean tryInactivate() {
385	<	if (active) {
386	<	if (!pool.tryDecrementActiveCount())
387	<	return false;
388	<	active = false;
384	>	protected void onTermination(Throwable exception) {
385	>	try {
386	>	ForkJoinPool p = pool;
387	>	if (active) {
388	>	int a; // inline p.tryDecrementActiveCount
389	>	active = false;
390	>	do {} while(!UNSAFE.compareAndSwapInt
391	>	(p, poolRunStateOffset, a = p.runState, a - 1));
392	>	}
393	>	cancelTasks();
394	>	setTerminated();
395	>	p.workerTerminated(this);
396	>	} catch (Throwable ex) {        // Shouldn't ever happen
397	>	if (exception == null)      // but if so, at least rethrown
398	>	exception = ex;
399	>	} finally {
400	>	if (exception != null)
401	>	UNSAFE.throwException(exception);
402		}
303	–	return true;
403		}
404
405		/**
307	–	* Computes next value for random victim probe.  Scans don't
308	–	* require a very high quality generator, but also not a crummy
309	–	* one.  Marsaglia xor-shift is cheap and works well.
310	–	*/
311	–	private static int xorShift(int r) {
312	–	r ^= r << 1;
313	–	r ^= r >>> 3;
314	–	r ^= r << 10;
315	–	return r;
316	–	}
317	–
318	–	// Lifecycle methods
319	–
320	–	/**
406		* This method is required to be public, but should never be
407		* called explicitly. It performs the main run loop to execute
408		* ForkJoinTasks.
#	Line 326 | Line 411 | public class ForkJoinWorkerThread extend
411		Throwable exception = null;
412		try {
413		onStart();
329	–	pool.sync(this); // await first pool event
414		mainLoop();
415		} catch (Throwable ex) {
416		exception = ex;
#	Line 335 | Line 419 | public class ForkJoinWorkerThread extend
419		}
420		}
421
422	+	// helpers for run()
423	+
424		/**
425	<	* Executes tasks until shut down.
425	>	* Find and execute tasks and check status while running
426		*/
427		private void mainLoop() {
428	<	while (!isShutdown()) {
429	<	ForkJoinTask<?> t = pollTask();
430	<	if (t != null || (t = pollSubmission()) != null)
431	<	t.quietlyExec();
432	<	else if (tryInactivate())
433	<	pool.sync(this);
428	>	int misses = 0; // track consecutive times failed to find work; max 2
429	>	ForkJoinPool p = pool;
430	>	for (;;) {
431	>	p.preStep(this, misses);
432	>	if (runState != 0)
433	>	break;
434	>	misses = ((tryExecSteal() || tryExecSubmission()) ? 0 :
435	>	(misses < 2 ? misses + 1 : 2));
436		}
437		}
438
439		/**
440	<	* Initializes internal state after construction but before
441	<	* processing any tasks. If you override this method, you must
442	<	* invoke super.onStart() at the beginning of the method.
355	<	* Initialization requires care: Most fields must have legal
356	<	* default values, to ensure that attempted accesses from other
357	<	* threads work correctly even before this thread starts
358	<	* processing tasks.
440	>	* Try to steal a task and execute it
441	>	*
442	>	* @return true if ran a task
443		*/
444	<	protected void onStart() {
445	<	// Allocate while starting to improve chances of thread-local
446	<	// isolation
447	<	queue = new ForkJoinTask<?>[INITIAL_QUEUE_CAPACITY];
448	<	// Initial value of seed need not be especially random but
449	<	// should differ across workers and must be nonzero
450	<	int p = poolIndex + 1;
451	<	seed = p + (p << 8) + (p << 16) + (p << 24); // spread bits
444	>	private boolean tryExecSteal() {
445	>	ForkJoinTask<?> t;
446	>	if ((t  = scan()) != null) {
447	>	t.quietlyExec();
448	>	currentSteal = null;
449	>	if (sp != base)
450	>	execLocalTasks();
451	>	return true;
452	>	}
453	>	return false;
454		}
455
456		/**
457	<	* Performs cleanup associated with termination of this worker
372	<	* thread.  If you override this method, you must invoke
373	<	* {@code super.onTermination} at the end of the overridden method.
457	>	* If a submission exists, try to activate and run it;
458		*
459	<	* @param exception the exception causing this thread to abort due
376	<	* to an unrecoverable error, or {@code null} if completed normally
459	>	* @return true if ran a task
460		*/
461	<	protected void onTermination(Throwable exception) {
462	<	// Execute remaining local tasks unless aborting or terminating
463	<	while (exception == null &&  !pool.isTerminating() && base != sp) {
464	<	try {
465	<	ForkJoinTask<?> t = popTask();
466	<	if (t != null)
461	>	private boolean tryExecSubmission() {
462	>	ForkJoinPool p = pool;
463	>	while (p.hasQueuedSubmissions()) {
464	>	ForkJoinTask<?> t; int a;
465	>	if (active || // ugly/hacky: inline p.tryIncrementActiveCount
466	>	(active = UNSAFE.compareAndSwapInt(p, poolRunStateOffset,
467	>	a = p.runState, a + 1))) {
468	>	if ((t = p.pollSubmission()) != null) {
469	>	currentSteal = t;
470		t.quietlyExec();
471	<	} catch (Throwable ex) {
472	<	exception = ex;
471	>	currentSteal = null;
472	>	if (sp != base)
473	>	execLocalTasks();
474	>	return true;
475	>	}
476		}
477		}
478	<	// Cancel other tasks, transition status, notify pool, and
390	<	// propagate exception to uncaught exception handler
391	<	try {
392	<	do {} while (!tryInactivate()); // ensure inactive
393	<	cancelTasks();
394	<	runState = TERMINATED;
395	<	pool.workerTerminated(this);
396	<	} catch (Throwable ex) {        // Shouldn't ever happen
397	<	if (exception == null)      // but if so, at least rethrown
398	<	exception = ex;
399	<	} finally {
400	<	if (exception != null)
401	<	ForkJoinTask.rethrowException(exception);
402	<	}
478	>	return false;
479		}
480
405	–	// Intrinsics-based support for queue operations.
406	–
481		/**
482	<	* Adds in store-order the given task at given slot of q to null.
483	<	* Caller must ensure q is non-null and index is in range.
482	>	* Runs local tasks until queue is empty or shut down.  Call only
483	>	* while active.
484		*/
485	<	private static void setSlot(ForkJoinTask<?>[] q, int i,
486	<	ForkJoinTask<?> t) {
487	<	UNSAFE.putOrderedObject(q, (i << qShift) + qBase, t);
485	>	private void execLocalTasks() {
486	>	while (runState == 0) {
487	>	ForkJoinTask<?> t = locallyFifo? locallyDeqTask() : popTask();
488	>	if (t != null)
489	>	t.quietlyExec();
490	>	else if (sp == base)
491	>	break;
492	>	}
493		}
494
495	+	/*
496	+	* Intrinsics-based atomic writes for queue slots. These are
497	+	* basically the same as methods in AtomicObjectArray, but
498	+	* specialized for (1) ForkJoinTask elements (2) requirement that
499	+	* nullness and bounds checks have already been performed by
500	+	* callers and (3) effective offsets are known not to overflow
501	+	* from int to long (because of MAXIMUM_QUEUE_CAPACITY). We don't
502	+	* need corresponding version for reads: plain array reads are OK
503	+	* because they protected by other volatile reads and are
504	+	* confirmed by CASes.
505	+	*
506	+	* Most uses don't actually call these methods, but instead contain
507	+	* inlined forms that enable more predictable optimization.  We
508	+	* don't define the version of write used in pushTask at all, but
509	+	* instead inline there a store-fenced array slot write.
510	+	*/
511	+
512		/**
513	<	* CAS given slot of q to null. Caller must ensure q is non-null
514	<	* and index is in range.
513	>	* CASes slot i of array q from t to null. Caller must ensure q is
514	>	* non-null and index is in range.
515		*/
516	<	private static boolean casSlotNull(ForkJoinTask<?>[] q, int i,
517	<	ForkJoinTask<?> t) {
516	>	private static final boolean casSlotNull(ForkJoinTask<?>[] q, int i,
517	>	ForkJoinTask<?> t) {
518		return UNSAFE.compareAndSwapObject(q, (i << qShift) + qBase, t, null);
519		}
520
521		/**
522	<	* Sets sp in store-order.
522	>	* Performs a volatile write of the given task at given slot of
523	>	* array q.  Caller must ensure q is non-null and index is in
524	>	* range. This method is used only during resets and backouts.
525		*/
526	<	private void storeSp(int s) {
527	<	UNSAFE.putOrderedInt(this, spOffset, s);
526	>	private static final void writeSlot(ForkJoinTask<?>[] q, int i,
527	>	ForkJoinTask<?> t) {
528	>	UNSAFE.putObjectVolatile(q, (i << qShift) + qBase, t);
529		}
530
531	<	// Main queue methods
531	>	// queue methods
532
533		/**
534	<	* Pushes a task. Called only by current thread.
534	>	* Pushes a task. Call only from this thread.
535		*
536		* @param t the task. Caller must ensure non-null.
537		*/
538		final void pushTask(ForkJoinTask<?> t) {
539		ForkJoinTask<?>[] q = queue;
540	<	int mask = q.length - 1;
541	<	int s = sp;
542	<	setSlot(q, s & mask, t);
543	<	storeSp(++s);
544	<	if ((s -= base) == 1)
545	<	pool.signalWork();
546	<	else if (s >= mask)
448	<	growQueue();
540	>	int mask = q.length - 1; // implicit assert q != null
541	>	int s = sp++;            // ok to increment sp before slot write
542	>	UNSAFE.putOrderedObject(q, ((s & mask) << qShift) + qBase, t);
543	>	if ((s -= base) == 0)
544	>	pool.signalWork();   // was empty
545	>	else if (s == mask)
546	>	growQueue();         // is full
547		}
548
549		/**
550		* Tries to take a task from the base of the queue, failing if
551	<	* either empty or contended.
551	>	* empty or contended. Note: Specializations of this code appear
552	>	* in locallyDeqTask and elsewhere.
553		*
554		* @return a task, or null if none or contended
555		*/
556		final ForkJoinTask<?> deqTask() {
557		ForkJoinTask<?> t;
558		ForkJoinTask<?>[] q;
559	<	int i;
461	<	int b;
559	>	int b, i;
560		if (sp != (b = base) &&
561		(q = queue) != null && // must read q after b
562	<	(t = q[i = (q.length - 1) & b]) != null &&
563	<	casSlotNull(q, i, t)) {
562	>	(t = q[i = (q.length - 1) & b]) != null && base == b &&
563	>	UNSAFE.compareAndSwapObject(q, (i << qShift) + qBase, t, null)) {
564		base = b + 1;
565		return t;
566		}
#	Line 470 | Line 568 | public class ForkJoinWorkerThread extend
568		}
569
570		/**
571	<	* Returns a popped task, or null if empty. Ensures active status
572	<	* if non-null. Called only by current thread.
571	>	* Tries to take a task from the base of own queue. Assumes active
572	>	* status.  Called only by current thread.
573	>	*
574	>	* @return a task, or null if none
575		*/
576	<	final ForkJoinTask<?> popTask() {
577	<	int s = sp;
578	<	while (s != base) {
579	<	if (tryActivate()) {
580	<	ForkJoinTask<?>[] q = queue;
581	<	int mask = q.length - 1;
582	<	int i = (s - 1) & mask;
576	>	final ForkJoinTask<?> locallyDeqTask() {
577	>	ForkJoinTask<?>[] q = queue;
578	>	if (q != null) {
579	>	ForkJoinTask<?> t;
580	>	int b, i;
581	>	while (sp != (b = base)) {
582	>	if ((t = q[i = (q.length - 1) & b]) != null && base == b &&
583	>	UNSAFE.compareAndSwapObject(q, (i << qShift) + qBase,
584	>	t, null)) {
585	>	base = b + 1;
586	>	return t;
587	>	}
588	>	}
589	>	}
590	>	return null;
591	>	}
592	>
593	>	/**
594	>	* Returns a popped task, or null if empty. Assumes active status.
595	>	* Called only by current thread.
596	>	*/
597	>	private ForkJoinTask<?> popTask() {
598	>	ForkJoinTask<?>[] q = queue;
599	>	if (q != null) {
600	>	int s;
601	>	while ((s = sp) != base) {
602	>	int i = (q.length - 1) & --s;
603	>	long u = (i << qShift) + qBase; // raw offset
604		ForkJoinTask<?> t = q[i];
605	<	if (t == null || !casSlotNull(q, i, t))
605	>	if (t == null)   // lost to stealer
606		break;
607	<	storeSp(s - 1);
608	<	return t;
607	>	if (UNSAFE.compareAndSwapObject(q, u, t, null)) {
608	>	sp = s; // putOrderedInt may encourage more timely write
609	>	// UNSAFE.putOrderedInt(this, spOffset, s);
610	>	return t;
611	>	}
612		}
613		}
614		return null;
615		}
616
617		/**
618	<	* Specialized version of popTask to pop only if
619	<	* topmost element is the given task. Called only
620	<	* by current thread while active.
618	>	* Specialized version of popTask to pop only if topmost element
619	>	* is the given task. Called only by current thread while
620	>	* active.
621		*
622		* @param t the task. Caller must ensure non-null.
623		*/
624		final boolean unpushTask(ForkJoinTask<?> t) {
625	+	int s;
626		ForkJoinTask<?>[] q = queue;
627	<	int mask = q.length - 1;
628	<	int s = sp - 1;
629	<	if (casSlotNull(q, s & mask, t)) {
630	<	storeSp(s);
627	>	if ((s = sp) != base && q != null &&
628	>	UNSAFE.compareAndSwapObject
629	>	(q, (((q.length - 1) & --s) << qShift) + qBase, t, null)) {
630	>	sp = s;
631	>	// UNSAFE.putOrderedInt(this, spOffset, s);
632		return true;
633		}
634		return false;
635		}
636
637		/**
638	<	* Returns next task.
638	>	* Returns next task or null if empty or contended
639		*/
640		final ForkJoinTask<?> peekTask() {
641		ForkJoinTask<?>[] q = queue;
#	Line 542 | Line 668 | public class ForkJoinWorkerThread extend
668		ForkJoinTask<?> t = oldQ[oldIndex];
669		if (t != null && !casSlotNull(oldQ, oldIndex, t))
670		t = null;
671	<	setSlot(newQ, b & newMask, t);
671	>	writeSlot(newQ, b & newMask, t);
672		} while (++b != bf);
673		pool.signalWork();
674		}
675
676		/**
677	+	* Computes next value for random victim probe in scan().  Scans
678	+	* don't require a very high quality generator, but also not a
679	+	* crummy one.  Marsaglia xor-shift is cheap and works well enough.
680	+	* Note: This is manually inlined in scan()
681	+	*/
682	+	private static final int xorShift(int r) {
683	+	r ^= r << 13;
684	+	r ^= r >>> 17;
685	+	return r ^ (r << 5);
686	+	}
687	+
688	+	/**
689		* Tries to steal a task from another worker. Starts at a random
690		* index of workers array, and probes workers until finding one
691		* with non-empty queue or finding that all are empty.  It
692		* randomly selects the first n probes. If these are empty, it
693	<	* resorts to a full circular traversal, which is necessary to
694	<	* accurately set active status by caller. Also restarts if pool
695	<	* events occurred since last scan, which forces refresh of
696	<	* workers array, in case barrier was associated with resize.
693	>	* resorts to a circular sweep, which is necessary to accurately
694	>	* set active status. (The circular sweep uses steps of
695	>	* approximately half the array size plus 1, to avoid bias
696	>	* stemming from leftmost packing of the array in ForkJoinPool.)
697		*
698		* This method must be both fast and quiet -- usually avoiding
699		* memory accesses that could disrupt cache sharing etc other than
700	<	* those needed to check for and take tasks. This accounts for,
701	<	* among other things, updating random seed in place without
702	<	* storing it until exit.
700	>	* those needed to check for and take tasks (or to activate if not
701	>	* already active). This accounts for, among other things,
702	>	* updating random seed in place without storing it until exit.
703		*
704		* @return a task, or null if none found
705		*/
706		private ForkJoinTask<?> scan() {
707	<	ForkJoinTask<?> t = null;
708	<	int r = seed;                    // extract once to keep scan quiet
709	<	ForkJoinWorkerThread[] ws;       // refreshed on outer loop
710	<	int mask;                        // must be power 2 minus 1 and > 0
711	<	outer:do {
712	<	if ((ws = pool.workers) != null && (mask = ws.length - 1) > 0) {
713	<	int idx = r;
714	<	int probes = ~mask;      // use random index while negative
715	<	for (;;) {
716	<	r = xorShift(r);     // update random seed
717	<	ForkJoinWorkerThread v = ws[mask & idx];
718	<	if (v == null || v.sp == v.base) {
719	<	if (probes <= mask)
720	<	idx = (probes++ < 0) ? r : (idx + 1);
721	<	else
722	<	break;
707	>	ForkJoinPool p = pool;
708	>	ForkJoinWorkerThread[] ws;        // worker array
709	>	int n;                            // upper bound of #workers
710	>	if ((ws = p.workers) != null && (n = ws.length) > 1) {
711	>	boolean canSteal = active;    // shadow active status
712	>	int r = seed;                 // extract seed once
713	>	int mask = n - 1;
714	>	int j = -n;                   // loop counter
715	>	int k = r;                    // worker index, random if j < 0
716	>	for (;;) {
717	>	ForkJoinWorkerThread v = ws[k & mask];
718	>	r ^= r << 13; r ^= r >>> 17; r ^= r << 5; // inline xorshift
719	>	if (v != null && v.base != v.sp) {
720	>	ForkJoinTask<?>[] q; int b, a;
721	>	if ((canSteal ||      // Ugly/hacky: inline
722	>	(canSteal = active =  // p.tryIncrementActiveCount
723	>	UNSAFE.compareAndSwapInt(p, poolRunStateOffset,
724	>	a = p.runState, a + 1))) &&
725	>	(q = v.queue) != null && (b = v.base) != v.sp) {
726	>	int i = (q.length - 1) & b;
727	>	long u = (i << qShift) + qBase; // raw offset
728	>	ForkJoinTask<?> t = q[i];
729	>	if (v.base == b && t != null &&
730	>	UNSAFE.compareAndSwapObject(q, u, t, null)) {
731	>	int pid = poolIndex;
732	>	currentSteal = t;
733	>	v.stealHint = pid;
734	>	v.base = b + 1;
735	>	seed = r;
736	>	++stealCount;
737	>	return t;
738	>	}
739		}
740	<	else if (!tryActivate() || (t = v.deqTask()) == null)
741	<	continue outer;  // restart on contention
588	<	else
589	<	break outer;
740	>	j = -n;
741	>	k = r;                // restart on contention
742		}
743	+	else if (++j <= 0)
744	+	k = r;
745	+	else if (j <= n)
746	+	k += (n >>> 1) | 1;
747	+	else
748	+	break;
749		}
750	<	} while (pool.hasNewSyncEvent(this)); // retry on pool events
751	<	seed = r;
594	<	return t;
750	>	}
751	>	return null;
752		}
753
754	+	// Run State management
755	+
756	+	// status check methods used mainly by ForkJoinPool
757	+	final boolean isRunning()     { return runState == 0; }
758	+	final boolean isTerminating() { return (runState & TERMINATING) != 0; }
759	+	final boolean isTerminated()  { return (runState & TERMINATED) != 0; }
760	+	final boolean isSuspended()   { return (runState & SUSPENDED) != 0; }
761	+	final boolean isTrimmed()     { return (runState & TRIMMED) != 0; }
762	+
763		/**
764	<	* Gets and removes a local or stolen task.
765	<	*
600	<	* @return a task, if available
764	>	* Sets state to TERMINATING. Does NOT unpark or interrupt
765	>	* to wake up if currently blocked.
766		*/
767	<	final ForkJoinTask<?> pollTask() {
768	<	ForkJoinTask<?> t = locallyFifo ? deqTask() : popTask();
769	<	if (t == null && (t = scan()) != null)
770	<	++stealCount;
771	<	return t;
767	>	final void shutdown() {
768	>	for (;;) {
769	>	int s = runState;
770	>	if ((s & (TERMINATING|TERMINATED)) != 0)
771	>	break;
772	>	if ((s & SUSPENDED) != 0) { // kill and wakeup if suspended
773	>	if (UNSAFE.compareAndSwapInt(this, runStateOffset, s,
774	>	(s & ~SUSPENDED) |
775	>	(TRIMMED|TERMINATING)))
776	>	break;
777	>	}
778	>	else if (UNSAFE.compareAndSwapInt(this, runStateOffset, s,
779	>	s | TERMINATING))
780	>	break;
781	>	}
782		}
783
784		/**
785	<	* Gets a local task.
611	<	*
612	<	* @return a task, if available
785	>	* Sets state to TERMINATED. Called only by onTermination()
786		*/
787	<	final ForkJoinTask<?> pollLocalTask() {
788	<	return locallyFifo ? deqTask() : popTask();
787	>	private void setTerminated() {
788	>	int s;
789	>	do {} while (!UNSAFE.compareAndSwapInt(this, runStateOffset,
790	>	s = runState,
791	>	s | (TERMINATING|TERMINATED)));
792		}
793
794		/**
795	<	* Returns a pool submission, if one exists, activating first.
795	>	* If suspended, tries to set status to unsuspended.
796		*
797	<	* @return a submission, if available
797	>	* @return true if successful
798		*/
799	<	private ForkJoinTask<?> pollSubmission() {
799	>	final boolean tryUnsuspend() {
800	>	int s;
801	>	while (((s = runState) & SUSPENDED) != 0) {
802	>	if (UNSAFE.compareAndSwapInt(this, runStateOffset, s,
803	>	s & ~SUSPENDED))
804	>	return true;
805	>	}
806	>	return false;
807	>	}
808	>
809	>	/**
810	>	* Sets suspended status and blocks as spare until resumed
811	>	* or shutdown.
812	>	*/
813	>	final void suspendAsSpare() {
814	>	for (;;) {                  // set suspended unless terminating
815	>	int s = runState;
816	>	if ((s & TERMINATING) != 0) { // must kill
817	>	if (UNSAFE.compareAndSwapInt(this, runStateOffset, s,
818	>	s | (TRIMMED | TERMINATING)))
819	>	return;
820	>	}
821	>	else if (UNSAFE.compareAndSwapInt(this, runStateOffset, s,
822	>	s | SUSPENDED))
823	>	break;
824	>	}
825		ForkJoinPool p = pool;
826	<	while (p.hasQueuedSubmissions()) {
827	<	ForkJoinTask<?> t;
828	<	if (tryActivate() && (t = p.pollSubmission()) != null)
829	<	return t;
826	>	p.pushSpare(this);
827	>	lastEventCount = 0;         // reset upon resume
828	>	while ((runState & SUSPENDED) != 0) {
829	>	if (p.tryAccumulateStealCount(this)) {
830	>	boolean untimed = nextSpare != 0;
831	>	long startTime = untimed? 0 : System.nanoTime();
832	>	interrupted();          // clear/ignore interrupts
833	>	if ((runState & SUSPENDED) == 0)
834	>	break;
835	>	if (untimed)     // untimed
836	>	LockSupport.park(this);
837	>	else {
838	>	LockSupport.parkNanos(this, TRIM_RATE_NANOS);
839	>	if ((runState & SUSPENDED) == 0)
840	>	break;
841	>	if (System.nanoTime() - startTime >= TRIM_RATE_NANOS)
842	>	p.tryShutdownSpare();
843	>	}
844	>	}
845		}
630	–	return null;
846		}
847
848	<	// Methods accessed only by Pool
848	>	// Misc support methods for ForkJoinPool
849	>
850	>	/**
851	>	* Returns an estimate of the number of tasks in the queue.  Also
852	>	* used by ForkJoinTask.
853	>	*/
854	>	final int getQueueSize() {
855	>	int n; // external calls must read base first
856	>	return (n = -base + sp) <= 0 ? 0 : n;
857	>	}
858
859		/**
860		* Removes and cancels all tasks in queue.  Can be called from any
861		* thread.
862		*/
863		final void cancelTasks() {
864	<	ForkJoinTask<?> t;
865	<	while (base != sp && (t = deqTask()) != null)
866	<	t.cancelIgnoringExceptions();
864	>	ForkJoinTask<?> cj = currentJoin; // try to cancel ongoing tasks
865	>	if (cj != null) {
866	>	currentJoin = null;
867	>	cj.cancelIgnoringExceptions();
868	>	try {
869	>	this.interrupt(); // awaken wait
870	>	} catch (SecurityException ignore) {
871	>	}
872	>	}
873	>	ForkJoinTask<?> cs = currentSteal;
874	>	if (cs != null) {
875	>	currentSteal = null;
876	>	cs.cancelIgnoringExceptions();
877	>	}
878	>	while (base != sp) {
879	>	ForkJoinTask<?> t = deqTask();
880	>	if (t != null)
881	>	t.cancelIgnoringExceptions();
882	>	}
883		}
884
885		/**
#	Line 649 | Line 889 | public class ForkJoinWorkerThread extend
889		*/
890		final int drainTasksTo(Collection<? super ForkJoinTask<?>> c) {
891		int n = 0;
892	<	ForkJoinTask<?> t;
893	<	while (base != sp && (t = deqTask()) != null) {
894	<	c.add(t);
895	<	++n;
892	>	while (base != sp) {
893	>	ForkJoinTask<?> t = deqTask();
894	>	if (t != null) {
895	>	c.add(t);
896	>	++n;
897	>	}
898		}
899		return n;
900		}
901
902	+	// Support methods for ForkJoinTask
903	+
904		/**
905	<	* Gets and clears steal count for accumulation by pool.  Called
906	<	* only when known to be idle (in pool.sync and termination).
905	>	* Gets and removes a local task.
906	>	*
907	>	* @return a task, if available
908		*/
909	<	final int getAndClearStealCount() {
910	<	int sc = stealCount;
911	<	stealCount = 0;
912	<	return sc;
909	>	final ForkJoinTask<?> pollLocalTask() {
910	>	ForkJoinPool p = pool;
911	>	while (sp != base) {
912	>	int a; // inline p.tryIncrementActiveCount
913	>	if (active ||
914	>	(active = UNSAFE.compareAndSwapInt(p, poolRunStateOffset,
915	>	a = p.runState, a + 1)))
916	>	return locallyFifo? locallyDeqTask() : popTask();
917	>	}
918	>	return null;
919		}
920
921		/**
922	<	* Returns {@code true} if at least one worker in the given array
672	<	* appears to have at least one queued task.
922	>	* Gets and removes a local or stolen task.
923		*
924	<	* @param ws array of workers
924	>	* @return a task, if available
925		*/
926	<	static boolean hasQueuedTasks(ForkJoinWorkerThread[] ws) {
927	<	if (ws != null) {
928	<	int len = ws.length;
929	<	for (int j = 0; j < 2; ++j) { // need two passes for clean sweep
930	<	for (int i = 0; i < len; ++i) {
681	<	ForkJoinWorkerThread w = ws[i];
682	<	if (w != null && w.sp != w.base)
683	<	return true;
684	<	}
685	<	}
926	>	final ForkJoinTask<?> pollTask() {
927	>	ForkJoinTask<?> t = pollLocalTask();
928	>	if (t == null) {
929	>	t = scan();
930	>	currentSteal = null; // cannot retain/track/help
931		}
932	<	return false;
932	>	return t;
933		}
934
690	–	// Support methods for ForkJoinTask
691	–
935		/**
936	<	* Returns an estimate of the number of tasks in the queue.
936	>	* Possibly runs some tasks and/or blocks, until task is done.
937	>	*
938	>	* @param joinMe the task to join
939		*/
940	<	final int getQueueSize() {
941	<	// suppress momentarily negative values
942	<	return Math.max(0, sp - base);
940	>	final void joinTask(ForkJoinTask<?> joinMe) {
941	>	// currentJoin only written by this thread; only need ordered store
942	>	ForkJoinTask<?> prevJoin = currentJoin;
943	>	UNSAFE.putOrderedObject(this, currentJoinOffset, joinMe);
944	>	if (sp != base)
945	>	localHelpJoinTask(joinMe);
946	>	if (joinMe.status >= 0)
947	>	pool.awaitJoin(joinMe, this);
948	>	UNSAFE.putOrderedObject(this, currentJoinOffset, prevJoin);
949		}
950
951		/**
952	<	* Returns an estimate of the number of tasks, offset by a
953	<	* function of number of idle workers.
952	>	* Run tasks in local queue until given task is done.
953	>	*
954	>	* @param joinMe the task to join
955		*/
956	<	final int getEstimatedSurplusTaskCount() {
957	<	// The halving approximates weighting idle vs non-idle workers
958	<	return (sp - base) - (pool.getIdleThreadCount() >>> 1);
956	>	private void localHelpJoinTask(ForkJoinTask<?> joinMe) {
957	>	int s;
958	>	ForkJoinTask<?>[] q;
959	>	while (joinMe.status >= 0 && (s = sp) != base && (q = queue) != null) {
960	>	int i = (q.length - 1) & --s;
961	>	long u = (i << qShift) + qBase; // raw offset
962	>	ForkJoinTask<?> t = q[i];
963	>	if (t == null)  // lost to a stealer
964	>	break;
965	>	if (UNSAFE.compareAndSwapObject(q, u, t, null)) {
966	>	/*
967	>	* This recheck (and similarly in helpJoinTask)
968	>	* handles cases where joinMe is independently
969	>	* cancelled or forced even though there is other work
970	>	* available. Back out of the pop by putting t back
971	>	* into slot before we commit by writing sp.
972	>	*/
973	>	if (joinMe.status < 0) {
974	>	UNSAFE.putObjectVolatile(q, u, t);
975	>	break;
976	>	}
977	>	sp = s;
978	>	// UNSAFE.putOrderedInt(this, spOffset, s);
979	>	t.quietlyExec();
980	>	}
981	>	}
982		}
983
984		/**
985	<	* Scans, returning early if joinMe done.
985	>	* Tries to locate and help perform tasks for a stealer of the
986	>	* given task, or in turn one of its stealers.  Traces
987	>	* currentSteal->currentJoin links looking for a thread working on
988	>	* a descendant of the given task and with a non-empty queue to
989	>	* steal back and execute tasks from.
990	>	*
991	>	* The implementation is very branchy to cope with the potential
992	>	* inconsistencies or loops encountering chains that are stale,
993	>	* unknown, or of length greater than MAX_HELP_DEPTH links.  All
994	>	* of these cases are dealt with by just returning back to the
995	>	* caller, who is expected to retry if other join mechanisms also
996	>	* don't work out.
997	>	*
998	>	* @param joinMe the task to join
999		*/
1000	<	final ForkJoinTask<?> scanWhileJoining(ForkJoinTask<?> joinMe) {
1001	<	ForkJoinTask<?> t = pollTask();
1002	<	if (t != null && joinMe.status < 0 && sp == base) {
1003	<	pushTask(t); // unsteal if done and this task would be stealable
1004	<	t = null;
1000	>	final void helpJoinTask(ForkJoinTask<?> joinMe) {
1001	>	ForkJoinWorkerThread[] ws = pool.workers;
1002	>	int n; // need at least 2 workers
1003	>	if (ws != null && (n = ws.length) > 1 && joinMe.status >= 0) {
1004	>	ForkJoinTask<?> task = joinMe;        // base of chain
1005	>	ForkJoinWorkerThread thread = this;   // thread with stolen task
1006	>	for (int d = 0; d < MAX_HELP_DEPTH; ++d) { // chain length
1007	>	// Try to find v, the stealer of task, by first using hint
1008	>	ForkJoinWorkerThread v = ws[thread.stealHint & (n - 1)];
1009	>	if (v == null || v.currentSteal != task) {
1010	>	for (int j = 0; ; ++j) {      // search array
1011	>	if (j < n) {
1012	>	if ((v = ws[j]) != null) {
1013	>	if (task.status < 0)
1014	>	return;       // stale or done
1015	>	if (v.currentSteal == task) {
1016	>	thread.stealHint = j;
1017	>	break;        // save hint for next time
1018	>	}
1019	>	}
1020	>	}
1021	>	else
1022	>	return;               // no stealer
1023	>	}
1024	>	}
1025	>	// Try to help v, using specialized form of deqTask
1026	>	int b;
1027	>	ForkJoinTask<?>[] q;
1028	>	while ((b = v.base) != v.sp && (q = v.queue) != null) {
1029	>	int i = (q.length - 1) & b;
1030	>	long u = (i << qShift) + qBase;
1031	>	ForkJoinTask<?> t = q[i];
1032	>	if (task.status < 0)
1033	>	return;                   // stale or done
1034	>	if (v.base == b) {
1035	>	if (t == null)
1036	>	return;               // producer stalled
1037	>	if (UNSAFE.compareAndSwapObject(q, u, t, null)) {
1038	>	if (joinMe.status < 0) {
1039	>	UNSAFE.putObjectVolatile(q, u, t);
1040	>	return;           // back out on cancel
1041	>	}
1042	>	int pid = poolIndex;
1043	>	ForkJoinTask<?> prevSteal = currentSteal;
1044	>	currentSteal = t;
1045	>	v.stealHint = pid;
1046	>	v.base = b + 1;
1047	>	t.quietlyExec();
1048	>	currentSteal = prevSteal;
1049	>	}
1050	>	}
1051	>	if (joinMe.status < 0)
1052	>	return;
1053	>	}
1054	>	// Try to descend to find v's stealer
1055	>	ForkJoinTask<?> next = v.currentJoin;
1056	>	if (task.status < 0 || next == null || next == task ||
1057	>	joinMe.status < 0)
1058	>	return;
1059	>	task = next;
1060	>	thread = v;
1061	>	}
1062		}
1063	<	return t;
1063	>	}
1064	>
1065	>	/**
1066	>	* Returns an estimate of the number of tasks, offset by a
1067	>	* function of number of idle workers.
1068	>	*
1069	>	* This method provides a cheap heuristic guide for task
1070	>	* partitioning when programmers, frameworks, tools, or languages
1071	>	* have little or no idea about task granularity.  In essence by
1072	>	* offering this method, we ask users only about tradeoffs in
1073	>	* overhead vs expected throughput and its variance, rather than
1074	>	* how finely to partition tasks.
1075	>	*
1076	>	* In a steady state strict (tree-structured) computation, each
1077	>	* thread makes available for stealing enough tasks for other
1078	>	* threads to remain active. Inductively, if all threads play by
1079	>	* the same rules, each thread should make available only a
1080	>	* constant number of tasks.
1081	>	*
1082	>	* The minimum useful constant is just 1. But using a value of 1
1083	>	* would require immediate replenishment upon each steal to
1084	>	* maintain enough tasks, which is infeasible.  Further,
1085	>	* partitionings/granularities of offered tasks should minimize
1086	>	* steal rates, which in general means that threads nearer the top
1087	>	* of computation tree should generate more than those nearer the
1088	>	* bottom. In perfect steady state, each thread is at
1089	>	* approximately the same level of computation tree. However,
1090	>	* producing extra tasks amortizes the uncertainty of progress and
1091	>	* diffusion assumptions.
1092	>	*
1093	>	* So, users will want to use values larger, but not much larger
1094	>	* than 1 to both smooth over transient shortages and hedge
1095	>	* against uneven progress; as traded off against the cost of
1096	>	* extra task overhead. We leave the user to pick a threshold
1097	>	* value to compare with the results of this call to guide
1098	>	* decisions, but recommend values such as 3.
1099	>	*
1100	>	* When all threads are active, it is on average OK to estimate
1101	>	* surplus strictly locally. In steady-state, if one thread is
1102	>	* maintaining say 2 surplus tasks, then so are others. So we can
1103	>	* just use estimated queue length (although note that (sp - base)
1104	>	* can be an overestimate because of stealers lagging increments
1105	>	* of base).  However, this strategy alone leads to serious
1106	>	* mis-estimates in some non-steady-state conditions (ramp-up,
1107	>	* ramp-down, other stalls). We can detect many of these by
1108	>	* further considering the number of "idle" threads, that are
1109	>	* known to have zero queued tasks, so compensate by a factor of
1110	>	* (#idle/#active) threads.
1111	>	*/
1112	>	final int getEstimatedSurplusTaskCount() {
1113	>	return sp - base - pool.idlePerActive();
1114		}
1115
1116		/**
#	Line 723 | Line 1118 | public class ForkJoinWorkerThread extend
1118		*/
1119		final void helpQuiescePool() {
1120		for (;;) {
1121	<	ForkJoinTask<?> t = pollTask();
1122	<	if (t != null)
1121	>	ForkJoinTask<?> t = pollLocalTask();
1122	>	if (t != null || (t = scan()) != null) {
1123		t.quietlyExec();
1124	<	else if (tryInactivate() && pool.isQuiescent())
1125	<	break;
1124	>	currentSteal = null;
1125	>	}
1126	>	else {
1127	>	ForkJoinPool p = pool;
1128	>	int a; // to inline CASes
1129	>	if (active) {
1130	>	if (!UNSAFE.compareAndSwapInt
1131	>	(p, poolRunStateOffset, a = p.runState, a - 1))
1132	>	continue;   // retry later
1133	>	active = false; // inactivate
1134	>	}
1135	>	if (p.isQuiescent()) {
1136	>	active = true; // re-activate
1137	>	do {} while(!UNSAFE.compareAndSwapInt
1138	>	(p, poolRunStateOffset, a = p.runState, a+1));
1139	>	return;
1140	>	}
1141	>	}
1142		}
732	–	do {} while (!tryActivate()); // re-activate on exit
1143		}
1144
1145		// Unsafe mechanics
#	Line 739 | Line 1149 | public class ForkJoinWorkerThread extend
1149		objectFieldOffset("sp", ForkJoinWorkerThread.class);
1150		private static final long runStateOffset =
1151		objectFieldOffset("runState", ForkJoinWorkerThread.class);
1152	<	private static final long qBase;
1152	>	private static final long currentJoinOffset =
1153	>	objectFieldOffset("currentJoin", ForkJoinWorkerThread.class);
1154	>	private static final long currentStealOffset =
1155	>	objectFieldOffset("currentSteal", ForkJoinWorkerThread.class);
1156	>	private static final long qBase =
1157	>	UNSAFE.arrayBaseOffset(ForkJoinTask[].class);
1158	>	private static final long poolRunStateOffset = // to inline CAS
1159	>	objectFieldOffset("runState", ForkJoinPool.class);
1160	>
1161		private static final int qShift;
1162
1163		static {
746	–	qBase = UNSAFE.arrayBaseOffset(ForkJoinTask[].class);
1164		int s = UNSAFE.arrayIndexScale(ForkJoinTask[].class);
1165		if ((s & (s-1)) != 0)
1166		throw new Error("data type scale not a power of two");

Diff Legend

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