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 1.29 by jsr166, Thu Aug 6 15:13:54 2009 UTC vs.
Revision 1.51 by jsr166, Mon Sep 20 20:42:37 2010 UTC

#	Line 8 | Line 8 | package jsr166y;
8
9		import java.util.concurrent.*;
10
11	+	import java.util.Random;
12		import java.util.Collection;
13	+	import java.util.concurrent.locks.LockSupport;
14
15		/**
16		* A thread managed by a {@link ForkJoinPool}.  This class is
#	Line 25 | Line 27 | import java.util.Collection;
27		*/
28		public class ForkJoinWorkerThread extends Thread {
29		/*
30	<	* Algorithm overview:
30	>	* Overview:
31		*
32	<	* 1. Work-Stealing: Work-stealing queues are special forms of
33	<	* Deques that support only three of the four possible
34	<	* end-operations -- push, pop, and deq (aka steal), and only do
35	<	* so under the constraints that push and pop are called only from
36	<	* the owning thread, while deq may be called from other threads.
37	<	* (If you are unfamiliar with them, you probably want to read
38	<	* Herlihy and Shavit's book "The Art of Multiprocessor
39	<	* programming", chapter 16 describing these in more detail before
40	<	* proceeding.)  The main work-stealing queue design is roughly
41	<	* similar to "Dynamic Circular Work-Stealing Deque" by David
42	<	* Chase and Yossi Lev, SPAA 2005
43	<	* (http://research.sun.com/scalable/pubs/index.html).  The main
44	<	* difference ultimately stems from gc requirements that we null
45	<	* out taken slots as soon as we can, to maintain as small a
46	<	* footprint as possible even in programs generating huge numbers
47	<	* of tasks. To accomplish this, we shift the CAS arbitrating pop
48	<	* vs deq (steal) from being on the indices ("base" and "sp") to
49	<	* the slots themselves (mainly via method "casSlotNull()"). So,
50	<	* both a successful pop and deq mainly entail CAS'ing a non-null
51	<	* slot to null.  Because we rely on CASes of references, we do
52	<	* not need tag bits on base or sp.  They are simple ints as used
53	<	* in any circular array-based queue (see for example ArrayDeque).
54	<	* Updates to the indices must still be ordered in a way that
55	<	* guarantees that (sp - base) > 0 means the queue is empty, but
56	<	* otherwise may err on the side of possibly making the queue
57	<	* appear nonempty when a push, pop, or deq have not fully
58	<	* committed. Note that this means that the deq operation,
59	<	* considered individually, is not wait-free. One thief cannot
60	<	* successfully continue until another in-progress one (or, if
61	<	* previously empty, a push) completes.  However, in the
62	<	* aggregate, we ensure at least probabilistic
63	<	* non-blockingness. If an attempted steal fails, a thief always
64	<	* chooses a different random victim target to try next. So, in
65	<	* order for one thief to progress, it suffices for any
66	<	* in-progress deq or new push on any empty queue to complete. One
67	<	* reason this works well here is that apparently-nonempty often
68	<	* means soon-to-be-stealable, which gives threads a chance to
69	<	* activate if necessary before stealing (see below).
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 set activation status if
78	>	* necessary before stealing.
79		*
80		* This approach also enables support for "async mode" where local
81		* task processing is in FIFO, not LIFO order; simply by using a
#	Line 72 | Line 83 | public class ForkJoinWorkerThread extend
83		* by the ForkJoinPool).  This allows use in message-passing
84		* frameworks in which tasks are never joined.
85		*
86	<	* Efficient implementation of this approach currently relies on
87	<	* an uncomfortable amount of "Unsafe" mechanics. To maintain
86	>	* When a worker would otherwise be blocked waiting to join a
87	>	* task, it first tries a form of linear helping: Each worker
88	>	* records (in field currentSteal) the most recent task it stole
89	>	* from some other worker. Plus, it records (in field currentJoin)
90	>	* the task it is currently actively joining. Method joinTask uses
91	>	* these markers to try to find a worker to help (i.e., steal back
92	>	* a task from and execute it) that could hasten completion of the
93	>	* actively joined task. In essence, the joiner executes a task
94	>	* that would be on its own local deque had the to-be-joined task
95	>	* not been stolen. This may be seen as a conservative variant of
96	>	* the approach in Wagner & Calder "Leapfrogging: a portable
97	>	* technique for implementing efficient futures" SIGPLAN Notices,
98	>	* 1993 (http://portal.acm.org/citation.cfm?id=155354). It differs
99	>	* in that: (1) We only maintain dependency links across workers
100	>	* upon steals, rather than use per-task bookkeeping.  This may
101	>	* require a linear scan of workers array to locate stealers, but
102	>	* usually doesn't because stealers leave hints (that may become
103	>	* stale/wrong) of where to locate them. This isolates cost to
104	>	* when it is needed, rather than adding to per-task overhead.
105	>	* (2) It is "shallow", ignoring nesting and potentially cyclic
106	>	* mutual steals.  (3) It is intentionally racy: field currentJoin
107	>	* is updated only while actively joining, which means that we
108	>	* miss links in the chain during long-lived tasks, GC stalls etc
109	>	* (which is OK since blocking in such cases is usually a good
110	>	* idea).  (4) We bound the number of attempts to find work (see
111	>	* MAX_HELP_DEPTH) and fall back to suspending the worker and if
112	>	* necessary replacing it with a spare (see
113	>	* ForkJoinPool.awaitJoin).
114	>	*
115	>	* Efficient implementation of these algorithms currently relies
116	>	* on an uncomfortable amount of "Unsafe" mechanics. To maintain
117		* correct orderings, reads and writes of variable base require
118	<	* volatile ordering.  Variable sp does not require volatile write
119	<	* but needs cheaper store-ordering on writes.  Because they are
120	<	* protected by volatile base reads, reads of the queue array and
121	<	* its slots do not need volatile load semantics, but writes (in
122	<	* push) require store order and CASes (in pop and deq) require
123	<	* (volatile) CAS semantics.  (See "Idempotent work stealing" by
124	<	* Michael, Saraswat, and Vechev, PPoPP 2009
125	<	* http://portal.acm.org/citation.cfm?id=1504186 for an algorithm
126	<	* with similar properties, but without support for nulling
127	<	* slots.)  Since these combinations aren't supported using
128	<	* ordinary volatiles, the only way to accomplish these
129	<	* efficiently is to use direct Unsafe calls. (Using external
130	<	* AtomicIntegers and AtomicReferenceArrays for the indices and
131	<	* array is significantly slower because of memory locality and
132	<	* indirection effects.)
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
#	Line 97 | Line 138 | public class ForkJoinWorkerThread extend
138		* initial size must be large enough to counteract cache
139		* contention effects across multiple queues (especially in the
140		* presence of GC cardmarking). Also, to improve thread-locality,
141	<	* queues are 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
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.  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		* Capacity of work-stealing queue array upon initialization.
168	<	* Must be a power of two. Initial size must be at least 2, but is
168	>	* Must be a power of two. Initial size must be at least 4, but is
169		* padded to minimize cache effects.
170		*/
171		private static final int INITIAL_QUEUE_CAPACITY = 1 << 13;
172
173		/**
174		* Maximum work-stealing queue array size.  Must be less than or
175	<	* equal to 1 << 28 to ensure lack of index wraparound. (This
176	<	* is less than usual bounds, because we need leftshift by 3
177	<	* to be in int range).
175	>	* equal to 1 << (31 - width of array entry) to ensure lack of
176	>	* index wraparound. The value is set in the static block
177	>	* at the end of this file after obtaining width.
178		*/
179	<	private static final int MAXIMUM_QUEUE_CAPACITY = 1 << 28;
179	>	private static final int MAXIMUM_QUEUE_CAPACITY;
180
181		/**
182		* The pool this thread works in. Accessed directly by ForkJoinTask.
#	Line 155 | Line 185 | public class ForkJoinWorkerThread extend
185
186		/**
187		* The work-stealing queue array. Size must be a power of two.
188	<	* Initialized when thread starts, to improve memory locality.
188	>	* Initialized in onStart, to improve memory locality.
189		*/
190		private ForkJoinTask<?>[] queue;
191
192		/**
193	+	* Index (mod queue.length) of least valid queue slot, which is
194	+	* always the next position to steal from if nonempty.
195	+	*/
196	+	private volatile int base;
197	+
198	+	/**
199		* Index (mod queue.length) of next queue slot to push to or pop
200	<	* from. It is written only by owner thread, via ordered store.
201	<	* Both sp and base are allowed to wrap around on overflow, but
202	<	* (sp - base) still estimates size.
200	>	* from. It is written only by owner thread, and accessed by other
201	>	* threads only after reading (volatile) base.  Both sp and base
202	>	* are allowed to wrap around on overflow, but (sp - base) still
203	>	* estimates size.
204		*/
205	<	private volatile int sp;
205	>	private int sp;
206
207		/**
208	<	* Index (mod queue.length) of least valid queue slot, which is
209	<	* always the next position to steal from if nonempty.
208	>	* The index of most recent stealer, used as a hint to avoid
209	>	* traversal in method helpJoinTask. This is only a hint because a
210	>	* worker might have had multiple steals and this only holds one
211	>	* of them (usually the most current). Declared non-volatile,
212	>	* relying on other prevailing sync to keep reasonably current.
213		*/
214	<	private volatile int base;
214	>	private int stealHint;
215
216		/**
217	<	* Activity status. When true, this worker is considered active.
218	<	* Must be false upon construction. It must be true when executing
219	<	* tasks, and BEFORE stealing a task. It must be false before
220	<	* calling pool.sync.
217	>	* Run state of this worker. In addition to the usual run levels,
218	>	* tracks if this worker is suspended as a spare, and if it was
219	>	* killed (trimmed) while suspended. However, "active" status is
220	>	* maintained separately and modified only in conjunction with
221	>	* CASes of the pool's runState (which are currently sadly
222	>	* manually inlined for performance.)  Accessed directly by pool
223	>	* to simplify checks for normal (zero) status.
224		*/
225	<	private boolean active;
225	>	volatile int runState;
226	>
227	>	private static final int TERMINATING = 0x01;
228	>	private static final int TERMINATED  = 0x02;
229	>	private static final int SUSPENDED   = 0x04; // inactive spare
230	>	private static final int TRIMMED     = 0x08; // killed while suspended
231
232		/**
233	<	* Run state of this worker. Supports simple versions of the usual
234	<	* shutdown/shutdownNow control.
233	>	* Number of steals. Directly accessed (and reset) by
234	>	* pool.tryAccumulateStealCount when idle.
235		*/
236	<	private volatile int runState;
236	>	int stealCount;
237
238		/**
239		* Seed for random number generator for choosing steal victims.
240	<	* Uses Marsaglia xorshift. Must be nonzero upon initialization.
240	>	* Uses Marsaglia xorshift. Must be initialized as nonzero.
241		*/
242		private int seed;
243
244		/**
245	<	* Number of steals, transferred to pool when idle
245	>	* Activity status. When true, this worker is considered active.
246	>	* Accessed directly by pool.  Must be false upon construction.
247	>	*/
248	>	boolean active;
249	>
250	>	/**
251	>	* True if use local fifo, not default lifo, for local polling.
252	>	* Shadows value from ForkJoinPool.
253		*/
254	<	private int stealCount;
254	>	private final boolean locallyFifo;
255
256		/**
257		* Index of this worker in pool array. Set once by pool before
258	<	* running, and accessed directly by pool during cleanup etc.
258	>	* running, and accessed directly by pool to locate this worker in
259	>	* its workers array.
260		*/
261		int poolIndex;
262
263		/**
264	<	* The last barrier event waited for. Accessed in pool callback
265	<	* methods, but only by current thread.
264	>	* The last pool event waited for. Accessed only by pool in
265	>	* callback methods invoked within this thread.
266	>	*/
267	>	int lastEventCount;
268	>
269	>	/**
270	>	* Encoded index and event count of next event waiter. Accessed
271	>	* only by ForkJoinPool for managing event waiters.
272		*/
273	<	long lastEventCount;
273	>	volatile long nextWaiter;
274
275		/**
276	<	* True if use local fifo, not default lifo, for local polling
276	>	* Number of times this thread suspended as spare. Accessed only
277	>	* by pool.
278		*/
279	<	private boolean locallyFifo;
279	>	int spareCount;
280	>
281	>	/**
282	>	* Encoded index and count of next spare waiter. Accessed only
283	>	* by ForkJoinPool for managing spares.
284	>	*/
285	>	volatile int nextSpare;
286	>
287	>	/**
288	>	* The task currently being joined, set only when actively trying
289	>	* to help other stealers in helpJoinTask. Written only by this
290	>	* thread, but read by others.
291	>	*/
292	>	private volatile ForkJoinTask<?> currentJoin;
293	>
294	>	/**
295	>	* The task most recently stolen from another worker (or
296	>	* submission queue).  Written only by this thread, but read by
297	>	* others.
298	>	*/
299	>	private volatile ForkJoinTask<?> currentSteal;
300
301		/**
302		* Creates a ForkJoinWorkerThread operating in the given pool.
#	Line 222 | Line 305 | public class ForkJoinWorkerThread extend
305		* @throws NullPointerException if pool is null
306		*/
307		protected ForkJoinWorkerThread(ForkJoinPool pool) {
225	–	if (pool == null) throw new NullPointerException();
308		this.pool = pool;
309	<	// Note: poolIndex is set by pool during construction
310	<	// Remaining initialization is deferred to onStart
309	>	this.locallyFifo = pool.locallyFifo;
310	>	setDaemon(true);
311	>	// To avoid exposing construction details to subclasses,
312	>	// remaining initialization is in start() and onStart()
313	>	}
314	>
315	>	/**
316	>	* Performs additional initialization and starts this thread.
317	>	*/
318	>	final void start(int poolIndex, UncaughtExceptionHandler ueh) {
319	>	this.poolIndex = poolIndex;
320	>	if (ueh != null)
321	>	setUncaughtExceptionHandler(ueh);
322	>	start();
323		}
324
325	<	// Public access methods
325	>	// Public/protected methods
326
327		/**
328		* Returns the pool hosting this thread.
#	Line 253 | Line 347 | public class ForkJoinWorkerThread extend
347		}
348
349		/**
350	<	* Establishes local first-in-first-out scheduling mode for forked
351	<	* tasks that are never joined.
352	<	*
353	<	* @param async if true, use locally FIFO scheduling
350	>	* Initializes internal state after construction but before
351	>	* processing any tasks. If you override this method, you must
352	>	* invoke @code{super.onStart()} at the beginning of the method.
353	>	* Initialization requires care: Most fields must have legal
354	>	* default values, to ensure that attempted accesses from other
355	>	* threads work correctly even before this thread starts
356	>	* processing tasks.
357		*/
358	<	void setAsyncMode(boolean async) {
359	<	locallyFifo = async;
360	<	}
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); }
358	>	protected void onStart() {
359	>	int rs = seedGenerator.nextInt();
360	>	seed = rs == 0? 1 : rs; // seed must be nonzero
361
362	<	/**
363	<	* Transitions to at least the given state.
364	<	*
365	<	* @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	<	}
362	>	// Allocate name string and arrays in this thread
363	>	String pid = Integer.toString(pool.getPoolNumber());
364	>	String wid = Integer.toString(poolIndex);
365	>	setName("ForkJoinPool-" + pid + "-worker-" + wid);
366
367	<	/**
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;
367	>	queue = new ForkJoinTask<?>[INITIAL_QUEUE_CAPACITY];
368		}
369
370		/**
371	<	* Tries to set status to inactive; fails on contention.
371	>	* 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.
374	>	*
375	>	* @param exception the exception causing this thread to abort due
376	>	* to an unrecoverable error, or {@code null} if completed normally
377		*/
378	<	private boolean tryInactivate() {
379	<	if (active) {
380	<	if (!pool.tryDecrementActiveCount())
381	<	return false;
382	<	active = false;
378	>	protected void onTermination(Throwable exception) {
379	>	try {
380	>	ForkJoinPool p = pool;
381	>	if (active) {
382	>	int a; // inline p.tryDecrementActiveCount
383	>	active = false;
384	>	do {} while (!UNSAFE.compareAndSwapInt
385	>	(p, poolRunStateOffset, a = p.runState, a - 1));
386	>	}
387	>	cancelTasks();
388	>	setTerminated();
389	>	p.workerTerminated(this);
390	>	} catch (Throwable ex) {        // Shouldn't ever happen
391	>	if (exception == null)      // but if so, at least rethrown
392	>	exception = ex;
393	>	} finally {
394	>	if (exception != null)
395	>	UNSAFE.throwException(exception);
396		}
315	–	return true;
397		}
398
399		/**
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
330	–
331	–	/**
400		* This method is required to be public, but should never be
401		* called explicitly. It performs the main run loop to execute
402		* ForkJoinTasks.
#	Line 337 | Line 405 | public class ForkJoinWorkerThread extend
405		Throwable exception = null;
406		try {
407		onStart();
340	–	pool.sync(this); // await first pool event
408		mainLoop();
409		} catch (Throwable ex) {
410		exception = ex;
#	Line 346 | Line 413 | public class ForkJoinWorkerThread extend
413		}
414		}
415
416	+	// helpers for run()
417	+
418		/**
419	<	* Executes tasks until shut down.
419	>	* Finds and executes tasks, and checks status while running.
420		*/
421		private void mainLoop() {
422	<	while (!isShutdown()) {
423	<	ForkJoinTask<?> t = pollTask();
424	<	if (t != null || (t = pollSubmission()) != null)
425	<	t.quietlyExec();
426	<	else if (tryInactivate())
427	<	pool.sync(this);
422	>	boolean ran = false; // true if ran a task on last step
423	>	ForkJoinPool p = pool;
424	>	for (;;) {
425	>	p.preStep(this, ran);
426	>	if (runState != 0)
427	>	break;
428	>	ran = tryExecSteal() || tryExecSubmission();
429		}
430		}
431
432		/**
433	<	* Initializes internal state after construction but before
434	<	* processing any tasks. If you override this method, you must
435	<	* invoke super.onStart() at the beginning of the method.
366	<	* Initialization requires care: Most fields must have legal
367	<	* default values, to ensure that attempted accesses from other
368	<	* threads work correctly even before this thread starts
369	<	* processing tasks.
433	>	* Tries to steal a task and execute it.
434	>	*
435	>	* @return true if ran a task
436		*/
437	<	protected void onStart() {
438	<	// Allocate while starting to improve chances of thread-local
439	<	// isolation
440	<	queue = new ForkJoinTask<?>[INITIAL_QUEUE_CAPACITY];
441	<	// Initial value of seed need not be especially random but
442	<	// should differ across workers and must be nonzero
443	<	int p = poolIndex + 1;
444	<	seed = p + (p << 8) + (p << 16) + (p << 24); // spread bits
437	>	private boolean tryExecSteal() {
438	>	ForkJoinTask<?> t;
439	>	if ((t = scan()) != null) {
440	>	t.quietlyExec();
441	>	UNSAFE.putOrderedObject(this, currentStealOffset, null);
442	>	if (sp != base)
443	>	execLocalTasks();
444	>	return true;
445	>	}
446	>	return false;
447		}
448
449		/**
450	<	* Performs cleanup associated with termination of this worker
383	<	* thread.  If you override this method, you must invoke
384	<	* {@code super.onTermination} at the end of the overridden method.
450	>	* If a submission exists, try to activate and run it.
451		*
452	<	* @param exception the exception causing this thread to abort due
387	<	* to an unrecoverable error, or {@code null} if completed normally
452	>	* @return true if ran a task
453		*/
454	<	protected void onTermination(Throwable exception) {
455	<	// Execute remaining local tasks unless aborting or terminating
456	<	while (exception == null && pool.isProcessingTasks() && base != sp) {
457	<	try {
458	<	ForkJoinTask<?> t = popTask();
459	<	if (t != null)
454	>	private boolean tryExecSubmission() {
455	>	ForkJoinPool p = pool;
456	>	// This loop is needed in case attempt to activate fails, in
457	>	// which case we only retry if there still appears to be a
458	>	// submission.
459	>	while (p.hasQueuedSubmissions()) {
460	>	ForkJoinTask<?> t; int a;
461	>	if (active || // inline p.tryIncrementActiveCount
462	>	(active = UNSAFE.compareAndSwapInt(p, poolRunStateOffset,
463	>	a = p.runState, a + 1))) {
464	>	if ((t = p.pollSubmission()) != null) {
465	>	UNSAFE.putOrderedObject(this, currentStealOffset, t);
466		t.quietlyExec();
467	<	} catch (Throwable ex) {
468	<	exception = ex;
467	>	UNSAFE.putOrderedObject(this, currentStealOffset, null);
468	>	if (sp != base)
469	>	execLocalTasks();
470	>	return true;
471	>	}
472		}
473		}
474	<	// Cancel other tasks, transition status, notify pool, and
401	<	// propagate exception to uncaught exception handler
402	<	try {
403	<	do {} while (!tryInactivate()); // ensure inactive
404	<	cancelTasks();
405	<	runState = TERMINATED;
406	<	pool.workerTerminated(this);
407	<	} catch (Throwable ex) {        // Shouldn't ever happen
408	<	if (exception == null)      // but if so, at least rethrown
409	<	exception = ex;
410	<	} finally {
411	<	if (exception != null)
412	<	ForkJoinTask.rethrowException(exception);
413	<	}
474	>	return false;
475		}
476
416	–	// Intrinsics-based support for queue operations.
417	–
477		/**
478	<	* Adds in store-order the given task at given slot of q to null.
479	<	* Caller must ensure q is non-null and index is in range.
478	>	* Runs local tasks until queue is empty or shut down.  Call only
479	>	* while active.
480		*/
481	<	private static void setSlot(ForkJoinTask<?>[] q, int i,
482	<	ForkJoinTask<?> t) {
483	<	UNSAFE.putOrderedObject(q, (i << qShift) + qBase, t);
481	>	private void execLocalTasks() {
482	>	while (runState == 0) {
483	>	ForkJoinTask<?> t = locallyFifo ? locallyDeqTask() : popTask();
484	>	if (t != null)
485	>	t.quietlyExec();
486	>	else if (sp == base)
487	>	break;
488	>	}
489		}
490
491	+	/*
492	+	* Intrinsics-based atomic writes for queue slots. These are
493	+	* basically the same as methods in AtomicReferenceArray, but
494	+	* specialized for (1) ForkJoinTask elements (2) requirement that
495	+	* nullness and bounds checks have already been performed by
496	+	* callers and (3) effective offsets are known not to overflow
497	+	* from int to long (because of MAXIMUM_QUEUE_CAPACITY). We don't
498	+	* need corresponding version for reads: plain array reads are OK
499	+	* because they are protected by other volatile reads and are
500	+	* confirmed by CASes.
501	+	*
502	+	* Most uses don't actually call these methods, but instead contain
503	+	* inlined forms that enable more predictable optimization.  We
504	+	* don't define the version of write used in pushTask at all, but
505	+	* instead inline there a store-fenced array slot write.
506	+	*/
507	+
508		/**
509	<	* CAS given slot of q to null. Caller must ensure q is non-null
510	<	* and index is in range.
509	>	* CASes slot i of array q from t to null. Caller must ensure q is
510	>	* non-null and index is in range.
511		*/
512	<	private static boolean casSlotNull(ForkJoinTask<?>[] q, int i,
513	<	ForkJoinTask<?> t) {
512	>	private static final boolean casSlotNull(ForkJoinTask<?>[] q, int i,
513	>	ForkJoinTask<?> t) {
514		return UNSAFE.compareAndSwapObject(q, (i << qShift) + qBase, t, null);
515		}
516
517		/**
518	<	* Sets sp in store-order.
518	>	* Performs a volatile write of the given task at given slot of
519	>	* array q.  Caller must ensure q is non-null and index is in
520	>	* range. This method is used only during resets and backouts.
521		*/
522	<	private void storeSp(int s) {
523	<	UNSAFE.putOrderedInt(this, spOffset, s);
522	>	private static final void writeSlot(ForkJoinTask<?>[] q, int i,
523	>	ForkJoinTask<?> t) {
524	>	UNSAFE.putObjectVolatile(q, (i << qShift) + qBase, t);
525		}
526
527	<	// Main queue methods
527	>	// queue methods
528
529		/**
530	<	* Pushes a task. Called only by current thread.
530	>	* Pushes a task. Call only from this thread.
531		*
532		* @param t the task. Caller must ensure non-null.
533		*/
534		final void pushTask(ForkJoinTask<?> t) {
535		ForkJoinTask<?>[] q = queue;
536	<	int mask = q.length - 1;
537	<	int s = sp;
538	<	setSlot(q, s & mask, t);
539	<	storeSp(++s);
540	<	if ((s -= base) == 1)
541	<	pool.signalWork();
542	<	else if (s >= mask)
459	<	growQueue();
536	>	int mask = q.length - 1; // implicit assert q != null
537	>	int s = sp++;            // ok to increment sp before slot write
538	>	UNSAFE.putOrderedObject(q, ((s & mask) << qShift) + qBase, t);
539	>	if ((s -= base) == 0)
540	>	pool.signalWork();   // was empty
541	>	else if (s == mask)
542	>	growQueue();         // is full
543		}
544
545		/**
546		* Tries to take a task from the base of the queue, failing if
547	<	* either empty or contended.
547	>	* empty or contended. Note: Specializations of this code appear
548	>	* in locallyDeqTask and elsewhere.
549		*
550		* @return a task, or null if none or contended
551		*/
552		final ForkJoinTask<?> deqTask() {
553		ForkJoinTask<?> t;
554		ForkJoinTask<?>[] q;
555	<	int i;
472	<	int b;
555	>	int b, i;
556		if (sp != (b = base) &&
557		(q = queue) != null && // must read q after b
558	<	(t = q[i = (q.length - 1) & b]) != null &&
559	<	casSlotNull(q, i, t)) {
558	>	(t = q[i = (q.length - 1) & b]) != null && base == b &&
559	>	UNSAFE.compareAndSwapObject(q, (i << qShift) + qBase, t, null)) {
560		base = b + 1;
561		return t;
562		}
#	Line 481 | Line 564 | public class ForkJoinWorkerThread extend
564		}
565
566		/**
567	<	* Tries to take a task from the base of own queue, activating if
568	<	* necessary, failing only if empty. Called only by current thread.
567	>	* Tries to take a task from the base of own queue. Assumes active
568	>	* status.  Called only by this thread.
569		*
570		* @return a task, or null if none
571		*/
572		final ForkJoinTask<?> locallyDeqTask() {
573	<	int b;
574	<	while (sp != (b = base)) {
575	<	if (tryActivate()) {
576	<	ForkJoinTask<?>[] q = queue;
577	<	int i = (q.length - 1) & b;
578	<	ForkJoinTask<?> t = q[i];
579	<	if (t != null && casSlotNull(q, i, t)) {
573	>	ForkJoinTask<?>[] q = queue;
574	>	if (q != null) {
575	>	ForkJoinTask<?> t;
576	>	int b, i;
577	>	while (sp != (b = base)) {
578	>	if ((t = q[i = (q.length - 1) & b]) != null && base == b &&
579	>	UNSAFE.compareAndSwapObject(q, (i << qShift) + qBase,
580	>	t, null)) {
581		base = b + 1;
582		return t;
583		}
#	Line 503 | Line 587 | public class ForkJoinWorkerThread extend
587		}
588
589		/**
590	<	* Returns a popped task, or null if empty. Ensures active status
591	<	* if non-null. Called only by current thread.
590	>	* Returns a popped task, or null if empty. Assumes active status.
591	>	* Called only by this thread.
592		*/
593	<	final ForkJoinTask<?> popTask() {
594	<	int s = sp;
595	<	while (s != base) {
596	<	if (tryActivate()) {
597	<	ForkJoinTask<?>[] q = queue;
598	<	int mask = q.length - 1;
599	<	int i = (s - 1) & mask;
593	>	private ForkJoinTask<?> popTask() {
594	>	ForkJoinTask<?>[] q = queue;
595	>	if (q != null) {
596	>	int s;
597	>	while ((s = sp) != base) {
598	>	int i = (q.length - 1) & --s;
599	>	long u = (i << qShift) + qBase; // raw offset
600		ForkJoinTask<?> t = q[i];
601	<	if (t == null || !casSlotNull(q, i, t))
601	>	if (t == null)   // lost to stealer
602		break;
603	<	storeSp(s - 1);
604	<	return t;
603	>	if (UNSAFE.compareAndSwapObject(q, u, t, null)) {
604	>	sp = s; // putOrderedInt may encourage more timely write
605	>	// UNSAFE.putOrderedInt(this, spOffset, s);
606	>	return t;
607	>	}
608		}
609		}
610		return null;
611		}
612
613		/**
614	<	* Specialized version of popTask to pop only if
615	<	* topmost element is the given task. Called only
529	<	* by current thread while active.
614	>	* Specialized version of popTask to pop only if topmost element
615	>	* is the given task. Called only by this thread while active.
616		*
617		* @param t the task. Caller must ensure non-null.
618		*/
619		final boolean unpushTask(ForkJoinTask<?> t) {
620	+	int s;
621		ForkJoinTask<?>[] q = queue;
622	<	int mask = q.length - 1;
623	<	int s = sp - 1;
624	<	if (casSlotNull(q, s & mask, t)) {
625	<	storeSp(s);
622	>	if ((s = sp) != base && q != null &&
623	>	UNSAFE.compareAndSwapObject
624	>	(q, (((q.length - 1) & --s) << qShift) + qBase, t, null)) {
625	>	sp = s; // putOrderedInt may encourage more timely write
626	>	// UNSAFE.putOrderedInt(this, spOffset, s);
627		return true;
628		}
629		return false;
630		}
631
632		/**
633	<	* Returns next task or null if empty or contended
633	>	* Returns next task, or null if empty or contended.
634		*/
635		final ForkJoinTask<?> peekTask() {
636		ForkJoinTask<?>[] q = queue;
#	Line 575 | Line 663 | public class ForkJoinWorkerThread extend
663		ForkJoinTask<?> t = oldQ[oldIndex];
664		if (t != null && !casSlotNull(oldQ, oldIndex, t))
665		t = null;
666	<	setSlot(newQ, b & newMask, t);
666	>	writeSlot(newQ, b & newMask, t);
667		} while (++b != bf);
668		pool.signalWork();
669		}
670
671		/**
672	+	* Computes next value for random victim probe in scan().  Scans
673	+	* don't require a very high quality generator, but also not a
674	+	* crummy one.  Marsaglia xor-shift is cheap and works well enough.
675	+	* Note: This is manually inlined in scan().
676	+	*/
677	+	private static final int xorShift(int r) {
678	+	r ^= r << 13;
679	+	r ^= r >>> 17;
680	+	return r ^ (r << 5);
681	+	}
682	+
683	+	/**
684		* Tries to steal a task from another worker. Starts at a random
685		* index of workers array, and probes workers until finding one
686		* with non-empty queue or finding that all are empty.  It
687		* randomly selects the first n probes. If these are empty, it
688	<	* resorts to a full circular traversal, which is necessary to
689	<	* accurately set active status by caller. Also restarts if pool
690	<	* events occurred since last scan, which forces refresh of
691	<	* workers array, in case barrier was associated with resize.
688	>	* resorts to a circular sweep, which is necessary to accurately
689	>	* set active status. (The circular sweep uses steps of
690	>	* approximately half the array size plus 1, to avoid bias
691	>	* stemming from leftmost packing of the array in ForkJoinPool.)
692		*
693		* This method must be both fast and quiet -- usually avoiding
694		* memory accesses that could disrupt cache sharing etc other than
695	<	* those needed to check for and take tasks. This accounts for,
696	<	* among other things, updating random seed in place without
697	<	* storing it until exit.
695	>	* those needed to check for and take tasks (or to activate if not
696	>	* already active). This accounts for, among other things,
697	>	* updating random seed in place without storing it until exit.
698		*
699		* @return a task, or null if none found
700		*/
701		private ForkJoinTask<?> scan() {
702	<	ForkJoinTask<?> t = null;
703	<	int r = seed;                    // extract once to keep scan quiet
704	<	ForkJoinWorkerThread[] ws;       // refreshed on outer loop
705	<	int mask;                        // must be power 2 minus 1 and > 0
706	<	outer:do {
707	<	if ((ws = pool.workers) != null && (mask = ws.length - 1) > 0) {
708	<	int idx = r;
709	<	int probes = ~mask;      // use random index while negative
710	<	for (;;) {
711	<	r = xorShift(r);     // update random seed
712	<	ForkJoinWorkerThread v = ws[mask & idx];
713	<	if (v == null || v.sp == v.base) {
714	<	if (probes <= mask)
715	<	idx = (probes++ < 0) ? r : (idx + 1);
716	<	else
717	<	break;
702	>	ForkJoinPool p = pool;
703	>	ForkJoinWorkerThread[] ws;        // worker array
704	>	int n;                            // upper bound of #workers
705	>	if ((ws = p.workers) != null && (n = ws.length) > 1) {
706	>	boolean canSteal = active;    // shadow active status
707	>	int r = seed;                 // extract seed once
708	>	int mask = n - 1;
709	>	int j = -n;                   // loop counter
710	>	int k = r;                    // worker index, random if j < 0
711	>	for (;;) {
712	>	ForkJoinWorkerThread v = ws[k & mask];
713	>	r ^= r << 13; r ^= r >>> 17; r ^= r << 5; // inline xorshift
714	>	ForkJoinTask<?>[] q; ForkJoinTask<?> t; int b, a;
715	>	if (v != null && (b = v.base) != v.sp &&
716	>	(q = v.queue) != null) {
717	>	int i = (q.length - 1) & b;
718	>	long u = (i << qShift) + qBase; // raw offset
719	>	int pid = poolIndex;
720	>	if ((t = q[i]) != null) {
721	>	if (!canSteal &&  // inline p.tryIncrementActiveCount
722	>	UNSAFE.compareAndSwapInt(p, poolRunStateOffset,
723	>	a = p.runState, a + 1))
724	>	canSteal = active = true;
725	>	if (canSteal && v.base == b++ &&
726	>	UNSAFE.compareAndSwapObject(q, u, t, null)) {
727	>	v.base = b;
728	>	v.stealHint = pid;
729	>	UNSAFE.putOrderedObject(this,
730	>	currentStealOffset, t);
731	>	seed = r;
732	>	++stealCount;
733	>	return t;
734	>	}
735		}
736	<	else if (!tryActivate() || (t = v.deqTask()) == null)
737	<	continue outer;  // restart on contention
621	<	else
622	<	break outer;
736	>	j = -n;
737	>	k = r;                // restart on contention
738		}
739	+	else if (++j <= 0)
740	+	k = r;
741	+	else if (j <= n)
742	+	k += (n >>> 1) | 1;
743	+	else
744	+	break;
745		}
746	<	} while (pool.hasNewSyncEvent(this)); // retry on pool events
747	<	seed = r;
748	<	return t;
746	>	}
747	>	return null;
748	>	}
749	>
750	>	// Run State management
751	>
752	>	// status check methods used mainly by ForkJoinPool
753	>	final boolean isRunning()     { return runState == 0; }
754	>	final boolean isTerminated()  { return (runState & TERMINATED) != 0; }
755	>	final boolean isSuspended()   { return (runState & SUSPENDED) != 0; }
756	>	final boolean isTrimmed()     { return (runState & TRIMMED) != 0; }
757	>
758	>	final boolean isTerminating() {
759	>	if ((runState & TERMINATING) != 0)
760	>	return true;
761	>	if (pool.isAtLeastTerminating()) { // propagate pool state
762	>	shutdown();
763	>	return true;
764	>	}
765	>	return false;
766		}
767
768		/**
769	<	* Gets and removes a local or stolen task.
770	<	*
633	<	* @return a task, if available
769	>	* Sets state to TERMINATING. Does NOT unpark or interrupt
770	>	* to wake up if currently blocked. Callers must do so if desired.
771		*/
772	<	final ForkJoinTask<?> pollTask() {
773	<	ForkJoinTask<?> t = locallyFifo ? locallyDeqTask() : popTask();
774	<	if (t == null && (t = scan()) != null)
775	<	++stealCount;
776	<	return t;
772	>	final void shutdown() {
773	>	for (;;) {
774	>	int s = runState;
775	>	if ((s & (TERMINATING|TERMINATED)) != 0)
776	>	break;
777	>	if ((s & SUSPENDED) != 0) { // kill and wakeup if suspended
778	>	if (UNSAFE.compareAndSwapInt(this, runStateOffset, s,
779	>	(s & ~SUSPENDED) |
780	>	(TRIMMED|TERMINATING)))
781	>	break;
782	>	}
783	>	else if (UNSAFE.compareAndSwapInt(this, runStateOffset, s,
784	>	s | TERMINATING))
785	>	break;
786	>	}
787		}
788
789		/**
790	<	* Gets a local task.
644	<	*
645	<	* @return a task, if available
790	>	* Sets state to TERMINATED. Called only by onTermination().
791		*/
792	<	final ForkJoinTask<?> pollLocalTask() {
793	<	return locallyFifo ? locallyDeqTask() : popTask();
792	>	private void setTerminated() {
793	>	int s;
794	>	do {} while (!UNSAFE.compareAndSwapInt(this, runStateOffset,
795	>	s = runState,
796	>	s | (TERMINATING|TERMINATED)));
797		}
798
799		/**
800	<	* Returns a pool submission, if one exists, activating first.
800	>	* If suspended, tries to set status to unsuspended.
801	>	* Does NOT wake up if blocked.
802		*
803	<	* @return a submission, if available
803	>	* @return true if successful
804		*/
805	<	private ForkJoinTask<?> pollSubmission() {
805	>	final boolean tryUnsuspend() {
806	>	int s;
807	>	while (((s = runState) & SUSPENDED) != 0) {
808	>	if (UNSAFE.compareAndSwapInt(this, runStateOffset, s,
809	>	s & ~SUSPENDED))
810	>	return true;
811	>	}
812	>	return false;
813	>	}
814	>
815	>	/**
816	>	* Sets suspended status and blocks as spare until resumed
817	>	* or shutdown.
818	>	*/
819	>	final void suspendAsSpare() {
820	>	for (;;) {                  // set suspended unless terminating
821	>	int s = runState;
822	>	if ((s & TERMINATING) != 0) { // must kill
823	>	if (UNSAFE.compareAndSwapInt(this, runStateOffset, s,
824	>	s | (TRIMMED | TERMINATING)))
825	>	return;
826	>	}
827	>	else if (UNSAFE.compareAndSwapInt(this, runStateOffset, s,
828	>	s | SUSPENDED))
829	>	break;
830	>	}
831		ForkJoinPool p = pool;
832	<	while (p.hasQueuedSubmissions()) {
833	<	ForkJoinTask<?> t;
834	<	if (tryActivate() && (t = p.pollSubmission()) != null)
835	<	return t;
832	>	p.pushSpare(this);
833	>	while ((runState & SUSPENDED) != 0) {
834	>	if (p.tryAccumulateStealCount(this)) {
835	>	interrupted();          // clear/ignore interrupts
836	>	if ((runState & SUSPENDED) == 0)
837	>	break;
838	>	LockSupport.park(this);
839	>	}
840		}
663	–	return null;
841		}
842
843	<	// Methods accessed only by Pool
843	>	// Misc support methods for ForkJoinPool
844	>
845	>	/**
846	>	* Returns an estimate of the number of tasks in the queue.  Also
847	>	* used by ForkJoinTask.
848	>	*/
849	>	final int getQueueSize() {
850	>	int n; // external calls must read base first
851	>	return (n = -base + sp) <= 0 ? 0 : n;
852	>	}
853
854		/**
855		* Removes and cancels all tasks in queue.  Can be called from any
856		* thread.
857		*/
858		final void cancelTasks() {
859	<	ForkJoinTask<?> t;
860	<	while (base != sp && (t = deqTask()) != null)
861	<	t.cancelIgnoringExceptions();
859	>	ForkJoinTask<?> cj = currentJoin; // try to cancel ongoing tasks
860	>	if (cj != null) {
861	>	currentJoin = null;
862	>	cj.cancelIgnoringExceptions();
863	>	try {
864	>	this.interrupt(); // awaken wait
865	>	} catch (SecurityException ignore) {
866	>	}
867	>	}
868	>	ForkJoinTask<?> cs = currentSteal;
869	>	if (cs != null) {
870	>	currentSteal = null;
871	>	cs.cancelIgnoringExceptions();
872	>	}
873	>	while (base != sp) {
874	>	ForkJoinTask<?> t = deqTask();
875	>	if (t != null)
876	>	t.cancelIgnoringExceptions();
877	>	}
878		}
879
880		/**
#	Line 682 | Line 884 | public class ForkJoinWorkerThread extend
884		*/
885		final int drainTasksTo(Collection<? super ForkJoinTask<?>> c) {
886		int n = 0;
887	<	ForkJoinTask<?> t;
888	<	while (base != sp && (t = deqTask()) != null) {
889	<	c.add(t);
890	<	++n;
887	>	while (base != sp) {
888	>	ForkJoinTask<?> t = deqTask();
889	>	if (t != null) {
890	>	c.add(t);
891	>	++n;
892	>	}
893		}
894		return n;
895		}
896
897	+	// Support methods for ForkJoinTask
898	+
899		/**
900	<	* Gets and clears steal count for accumulation by pool.  Called
901	<	* only when known to be idle (in pool.sync and termination).
900	>	* Gets and removes a local task.
901	>	*
902	>	* @return a task, if available
903		*/
904	<	final int getAndClearStealCount() {
905	<	int sc = stealCount;
906	<	stealCount = 0;
907	<	return sc;
904	>	final ForkJoinTask<?> pollLocalTask() {
905	>	ForkJoinPool p = pool;
906	>	while (sp != base) {
907	>	int a; // inline p.tryIncrementActiveCount
908	>	if (active ||
909	>	(active = UNSAFE.compareAndSwapInt(p, poolRunStateOffset,
910	>	a = p.runState, a + 1)))
911	>	return locallyFifo ? locallyDeqTask() : popTask();
912	>	}
913	>	return null;
914		}
915
916		/**
917	<	* Returns {@code true} if at least one worker in the given array
705	<	* appears to have at least one queued task.
917	>	* Gets and removes a local or stolen task.
918		*
919	<	* @param ws array of workers
919	>	* @return a task, if available
920		*/
921	<	static boolean hasQueuedTasks(ForkJoinWorkerThread[] ws) {
922	<	if (ws != null) {
923	<	int len = ws.length;
924	<	for (int j = 0; j < 2; ++j) { // need two passes for clean sweep
925	<	for (int i = 0; i < len; ++i) {
926	<	ForkJoinWorkerThread w = ws[i];
715	<	if (w != null && w.sp != w.base)
716	<	return true;
717	<	}
718	<	}
921	>	final ForkJoinTask<?> pollTask() {
922	>	ForkJoinTask<?> t = pollLocalTask();
923	>	if (t == null) {
924	>	t = scan();
925	>	// cannot retain/track/help steal
926	>	UNSAFE.putOrderedObject(this, currentStealOffset, null);
927		}
928	<	return false;
928	>	return t;
929		}
930
723	–	// Support methods for ForkJoinTask
724	–
931		/**
932	<	* Returns an estimate of the number of tasks in the queue.
932	>	* Possibly runs some tasks and/or blocks, until task is done.
933	>	*
934	>	* @param joinMe the task to join
935		*/
936	<	final int getQueueSize() {
937	<	// suppress momentarily negative values
938	<	return Math.max(0, sp - base);
936	>	final void joinTask(ForkJoinTask<?> joinMe) {
937	>	// currentJoin only written by this thread; only need ordered store
938	>	ForkJoinTask<?> prevJoin = currentJoin;
939	>	UNSAFE.putOrderedObject(this, currentJoinOffset, joinMe);
940	>	if (sp != base)
941	>	localHelpJoinTask(joinMe);
942	>	if (joinMe.status >= 0)
943	>	pool.awaitJoin(joinMe, this);
944	>	UNSAFE.putOrderedObject(this, currentJoinOffset, prevJoin);
945		}
946
947		/**
948	<	* Returns an estimate of the number of tasks, offset by a
949	<	* function of number of idle workers.
948	>	* Run tasks in local queue until given task is done.
949	>	*
950	>	* @param joinMe the task to join
951		*/
952	<	final int getEstimatedSurplusTaskCount() {
953	<	// The halving approximates weighting idle vs non-idle workers
954	<	return (sp - base) - (pool.getIdleThreadCount() >>> 1);
952	>	private void localHelpJoinTask(ForkJoinTask<?> joinMe) {
953	>	int s;
954	>	ForkJoinTask<?>[] q;
955	>	while (joinMe.status >= 0 && (s = sp) != base && (q = queue) != null) {
956	>	int i = (q.length - 1) & --s;
957	>	long u = (i << qShift) + qBase; // raw offset
958	>	ForkJoinTask<?> t = q[i];
959	>	if (t == null)  // lost to a stealer
960	>	break;
961	>	if (UNSAFE.compareAndSwapObject(q, u, t, null)) {
962	>	/*
963	>	* This recheck (and similarly in helpJoinTask)
964	>	* handles cases where joinMe is independently
965	>	* cancelled or forced even though there is other work
966	>	* available. Back out of the pop by putting t back
967	>	* into slot before we commit by writing sp.
968	>	*/
969	>	if (joinMe.status < 0) {
970	>	UNSAFE.putObjectVolatile(q, u, t);
971	>	break;
972	>	}
973	>	sp = s;
974	>	// UNSAFE.putOrderedInt(this, spOffset, s);
975	>	t.quietlyExec();
976	>	}
977	>	}
978		}
979
980		/**
981	<	* Scans, returning early if joinMe done.
982	<	*/
983	<	final ForkJoinTask<?> scanWhileJoining(ForkJoinTask<?> joinMe) {
984	<	ForkJoinTask<?> t = pollTask();
985	<	if (t != null && joinMe.status < 0 && sp == base) {
986	<	pushTask(t); // unsteal if done and this task would be stealable
987	<	t = null;
981	>	* Unless terminating, tries to locate and help perform tasks for
982	>	* a stealer of the given task, or in turn one of its stealers.
983	>	* Traces currentSteal->currentJoin links looking for a thread
984	>	* working on a descendant of the given task and with a non-empty
985	>	* queue to steal back and execute tasks from.
986	>	*
987	>	* The implementation is very branchy to cope with potential
988	>	* inconsistencies or loops encountering chains that are stale,
989	>	* unknown, or of length greater than MAX_HELP_DEPTH links.  All
990	>	* of these cases are dealt with by just returning back to the
991	>	* caller, who is expected to retry if other join mechanisms also
992	>	* don't work out.
993	>	*
994	>	* @param joinMe the task to join
995	>	*/
996	>	final void helpJoinTask(ForkJoinTask<?> joinMe) {
997	>	ForkJoinWorkerThread[] ws;
998	>	int n;
999	>	if (joinMe.status < 0)                // already done
1000	>	return;
1001	>	if ((runState & TERMINATING) != 0) {  // cancel if shutting down
1002	>	joinMe.cancelIgnoringExceptions();
1003	>	return;
1004		}
1005	<	return t;
1005	>	if ((ws = pool.workers) == null || (n = ws.length) <= 1)
1006	>	return;                           // need at least 2 workers
1007	>
1008	>	ForkJoinTask<?> task = joinMe;        // base of chain
1009	>	ForkJoinWorkerThread thread = this;   // thread with stolen task
1010	>	for (int d = 0; d < MAX_HELP_DEPTH; ++d) { // chain length
1011	>	// Try to find v, the stealer of task, by first using hint
1012	>	ForkJoinWorkerThread v = ws[thread.stealHint & (n - 1)];
1013	>	if (v == null || v.currentSteal != task) {
1014	>	for (int j = 0; ; ++j) {      // search array
1015	>	if (j < n) {
1016	>	ForkJoinTask<?> vs;
1017	>	if ((v = ws[j]) != null &&
1018	>	(vs = v.currentSteal) != null) {
1019	>	if (joinMe.status < 0 || task.status < 0)
1020	>	return;       // stale or done
1021	>	if (vs == task) {
1022	>	thread.stealHint = j;
1023	>	break;        // save hint for next time
1024	>	}
1025	>	}
1026	>	}
1027	>	else
1028	>	return;               // no stealer
1029	>	}
1030	>	}
1031	>	for (;;) { // Try to help v, using specialized form of deqTask
1032	>	if (joinMe.status < 0)
1033	>	return;
1034	>	int b = v.base;
1035	>	ForkJoinTask<?>[] q = v.queue;
1036	>	if (b == v.sp || q == null)
1037	>	break;
1038	>	int i = (q.length - 1) & b;
1039	>	long u = (i << qShift) + qBase;
1040	>	ForkJoinTask<?> t = q[i];
1041	>	int pid = poolIndex;
1042	>	ForkJoinTask<?> ps = currentSteal;
1043	>	if (task.status < 0)
1044	>	return;                   // stale or done
1045	>	if (t != null && v.base == b++ &&
1046	>	UNSAFE.compareAndSwapObject(q, u, t, null)) {
1047	>	if (joinMe.status < 0) {
1048	>	UNSAFE.putObjectVolatile(q, u, t);
1049	>	return;               // back out on cancel
1050	>	}
1051	>	v.base = b;
1052	>	v.stealHint = pid;
1053	>	UNSAFE.putOrderedObject(this, currentStealOffset, t);
1054	>	t.quietlyExec();
1055	>	UNSAFE.putOrderedObject(this, currentStealOffset, ps);
1056	>	}
1057	>	}
1058	>	// Try to descend to find v's stealer
1059	>	ForkJoinTask<?> next = v.currentJoin;
1060	>	if (task.status < 0 || next == null || next == task ||
1061	>	joinMe.status < 0)
1062	>	return;
1063	>	task = next;
1064	>	thread = v;
1065	>	}
1066	>	}
1067	>
1068	>	/**
1069	>	* Implements ForkJoinTask.getSurplusQueuedTaskCount().
1070	>	* Returns an estimate of the number of tasks, offset by a
1071	>	* function of number of idle workers.
1072	>	*
1073	>	* This method provides a cheap heuristic guide for task
1074	>	* partitioning when programmers, frameworks, tools, or languages
1075	>	* have little or no idea about task granularity.  In essence by
1076	>	* offering this method, we ask users only about tradeoffs in
1077	>	* overhead vs expected throughput and its variance, rather than
1078	>	* how finely to partition tasks.
1079	>	*
1080	>	* In a steady state strict (tree-structured) computation, each
1081	>	* thread makes available for stealing enough tasks for other
1082	>	* threads to remain active. Inductively, if all threads play by
1083	>	* the same rules, each thread should make available only a
1084	>	* constant number of tasks.
1085	>	*
1086	>	* The minimum useful constant is just 1. But using a value of 1
1087	>	* would require immediate replenishment upon each steal to
1088	>	* maintain enough tasks, which is infeasible.  Further,
1089	>	* partitionings/granularities of offered tasks should minimize
1090	>	* steal rates, which in general means that threads nearer the top
1091	>	* of computation tree should generate more than those nearer the
1092	>	* bottom. In perfect steady state, each thread is at
1093	>	* approximately the same level of computation tree. However,
1094	>	* producing extra tasks amortizes the uncertainty of progress and
1095	>	* diffusion assumptions.
1096	>	*
1097	>	* So, users will want to use values larger, but not much larger
1098	>	* than 1 to both smooth over transient shortages and hedge
1099	>	* against uneven progress; as traded off against the cost of
1100	>	* extra task overhead. We leave the user to pick a threshold
1101	>	* value to compare with the results of this call to guide
1102	>	* decisions, but recommend values such as 3.
1103	>	*
1104	>	* When all threads are active, it is on average OK to estimate
1105	>	* surplus strictly locally. In steady-state, if one thread is
1106	>	* maintaining say 2 surplus tasks, then so are others. So we can
1107	>	* just use estimated queue length (although note that (sp - base)
1108	>	* can be an overestimate because of stealers lagging increments
1109	>	* of base).  However, this strategy alone leads to serious
1110	>	* mis-estimates in some non-steady-state conditions (ramp-up,
1111	>	* ramp-down, other stalls). We can detect many of these by
1112	>	* further considering the number of "idle" threads, that are
1113	>	* known to have zero queued tasks, so compensate by a factor of
1114	>	* (#idle/#active) threads.
1115	>	*/
1116	>	final int getEstimatedSurplusTaskCount() {
1117	>	return sp - base - pool.idlePerActive();
1118		}
1119
1120		/**
1121		* Runs tasks until {@code pool.isQuiescent()}.
1122		*/
1123		final void helpQuiescePool() {
1124	+	ForkJoinTask<?> ps = currentSteal; // to restore below
1125		for (;;) {
1126	<	ForkJoinTask<?> t = pollTask();
1127	<	if (t != null)
1126	>	ForkJoinTask<?> t = pollLocalTask();
1127	>	if (t != null || (t = scan()) != null)
1128		t.quietlyExec();
1129	<	else if (tryInactivate() && pool.isQuiescent())
1130	<	break;
1129	>	else {
1130	>	ForkJoinPool p = pool;
1131	>	int a; // to inline CASes
1132	>	if (active) {
1133	>	if (!UNSAFE.compareAndSwapInt
1134	>	(p, poolRunStateOffset, a = p.runState, a - 1))
1135	>	continue;   // retry later
1136	>	active = false; // inactivate
1137	>	UNSAFE.putOrderedObject(this, currentStealOffset, ps);
1138	>	}
1139	>	if (p.isQuiescent()) {
1140	>	active = true; // re-activate
1141	>	do {} while (!UNSAFE.compareAndSwapInt
1142	>	(p, poolRunStateOffset, a = p.runState, a+1));
1143	>	return;
1144	>	}
1145	>	}
1146		}
765	–	do {} while (!tryActivate()); // re-activate on exit
1147		}
1148
1149		// Unsafe mechanics
#	Line 772 | Line 1153 | public class ForkJoinWorkerThread extend
1153		objectFieldOffset("sp", ForkJoinWorkerThread.class);
1154		private static final long runStateOffset =
1155		objectFieldOffset("runState", ForkJoinWorkerThread.class);
1156	<	private static final long qBase;
1156	>	private static final long currentJoinOffset =
1157	>	objectFieldOffset("currentJoin", ForkJoinWorkerThread.class);
1158	>	private static final long currentStealOffset =
1159	>	objectFieldOffset("currentSteal", ForkJoinWorkerThread.class);
1160	>	private static final long qBase =
1161	>	UNSAFE.arrayBaseOffset(ForkJoinTask[].class);
1162	>	private static final long poolRunStateOffset = // to inline CAS
1163	>	objectFieldOffset("runState", ForkJoinPool.class);
1164	>
1165		private static final int qShift;
1166
1167		static {
779	–	qBase = UNSAFE.arrayBaseOffset(ForkJoinTask[].class);
1168		int s = UNSAFE.arrayIndexScale(ForkJoinTask[].class);
1169		if ((s & (s-1)) != 0)
1170		throw new Error("data type scale not a power of two");
1171		qShift = 31 - Integer.numberOfLeadingZeros(s);
1172	+	MAXIMUM_QUEUE_CAPACITY = 1 << (31 - qShift);
1173		}
1174
1175		private static long objectFieldOffset(String field, Class<?> klazz) {

Diff Legend

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