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 1.18 by jsr166, Sat Jul 25 00:34:00 2009 UTC vs.
Revision 1.58 by jsr166, Sun Nov 21 08:18:19 2010 UTC

#	Line 6 | Line 6
6
7		package jsr166y;
8
9	<	import java.util.concurrent.*;
10	<
9	>	import java.util.Random;
10		import java.util.Collection;
11	+	import java.util.concurrent.locks.LockSupport;
12	+	import java.util.concurrent.RejectedExecutionException;
13
14		/**
15		* A thread managed by a {@link ForkJoinPool}.  This class is
16		* subclassable solely for the sake of adding functionality -- there
17	<	* are no overridable methods dealing with scheduling or
18	<	* execution. However, you can override initialization and termination
19	<	* methods surrounding the main task processing loop.  If you do
20	<	* create such a subclass, you will also need to supply a custom
21	<	* ForkJoinWorkerThreadFactory to use it in a ForkJoinPool.
17	>	* are no overridable methods dealing with scheduling or execution.
18	>	* However, you can override initialization and termination methods
19	>	* surrounding the main task processing loop.  If you do create such a
20	>	* subclass, you will also need to supply a custom {@link
21	>	* ForkJoinPool.ForkJoinWorkerThreadFactory} to use it in a {@code
22	>	* ForkJoinPool}.
23		*
24		* @since 1.7
25		* @author Doug Lea
26		*/
27		public class ForkJoinWorkerThread extends Thread {
28		/*
29	<	* Algorithm overview:
29	>	* Overview:
30	>	*
31	>	* ForkJoinWorkerThreads are managed by ForkJoinPools and perform
32	>	* ForkJoinTasks. This class includes bookkeeping in support of
33	>	* worker activation, suspension, and lifecycle control described
34	>	* in more detail in the internal documentation of class
35	>	* ForkJoinPool. And as described further below, this class also
36	>	* includes special-cased support for some ForkJoinTask
37	>	* methods. But the main mechanics involve work-stealing:
38		*
39	<	* 1. Work-Stealing: Work-stealing queues are special forms of
40	<	* Deques that support only three of the four possible
41	<	* end-operations -- push, pop, and deq (aka steal), and only do
42	<	* so under the constraints that push and pop are called only from
43	<	* the owning thread, while deq may be called from other threads.
44	<	* (If you are unfamiliar with them, you probably want to read
45	<	* Herlihy and Shavit's book "The Art of Multiprocessor
46	<	* programming", chapter 16 describing these in more detail before
47	<	* proceeding.)  The main work-stealing queue design is roughly
48	<	* similar to "Dynamic Circular Work-Stealing Deque" by David
49	<	* Chase and Yossi Lev, SPAA 2005
50	<	* (http://research.sun.com/scalable/pubs/index.html).  The main
51	<	* difference ultimately stems from gc requirements that we null
52	<	* out taken slots as soon as we can, to maintain as small a
53	<	* footprint as possible even in programs generating huge numbers
54	<	* of tasks. To accomplish this, we shift the CAS arbitrating pop
55	<	* vs deq (steal) from being on the indices ("base" and "sp") to
56	<	* the slots themselves (mainly via method "casSlotNull()"). So,
57	<	* both a successful pop and deq mainly entail CAS'ing a non-null
58	<	* slot to null.  Because we rely on CASes of references, we do
59	<	* not need tag bits on base or sp.  They are simple ints as used
60	<	* in any circular array-based queue (see for example ArrayDeque).
61	<	* Updates to the indices must still be ordered in a way that
62	<	* guarantees that (sp - base) > 0 means the queue is empty, but
63	<	* otherwise may err on the side of possibly making the queue
64	<	* appear nonempty when a push, pop, or deq have not fully
65	<	* committed. Note that this means that the deq operation,
66	<	* considered individually, is not wait-free. One thief cannot
67	<	* successfully continue until another in-progress one (or, if
68	<	* previously empty, a push) completes.  However, in the
69	<	* aggregate, we ensure at least probabilistic non-blockingness. If
70	<	* an attempted steal fails, a thief always chooses a different
39	>	* Work-stealing queues are special forms of Deques that support
40	>	* only three of the four possible end-operations -- push, pop,
41	>	* and deq (aka steal), under the further constraints that push
42	>	* and pop are called only from the owning thread, while deq may
43	>	* be called from other threads.  (If you are unfamiliar with
44	>	* them, you probably want to read Herlihy and Shavit's book "The
45	>	* Art of Multiprocessor programming", chapter 16 describing these
46	>	* in more detail before proceeding.)  The main work-stealing
47	>	* queue design is roughly similar to those in the papers "Dynamic
48	>	* Circular Work-Stealing Deque" by Chase and Lev, SPAA 2005
49	>	* (http://research.sun.com/scalable/pubs/index.html) and
50	>	* "Idempotent work stealing" by Michael, Saraswat, and Vechev,
51	>	* PPoPP 2009 (http://portal.acm.org/citation.cfm?id=1504186).
52	>	* The main differences ultimately stem from gc requirements that
53	>	* we null out taken slots as soon as we can, to maintain as small
54	>	* a footprint as possible even in programs generating huge
55	>	* numbers of tasks. To accomplish this, we shift the CAS
56	>	* arbitrating pop vs deq (steal) from being on the indices
57	>	* ("base" and "sp") to the slots themselves (mainly via method
58	>	* "casSlotNull()"). So, both a successful pop and deq mainly
59	>	* entail a CAS of a slot from non-null to null.  Because we rely
60	>	* on CASes of references, we do not need tag bits on base or sp.
61	>	* They are simple ints as used in any circular array-based queue
62	>	* (see for example ArrayDeque).  Updates to the indices must
63	>	* still be ordered in a way that guarantees that sp == base means
64	>	* the queue is empty, but otherwise may err on the side of
65	>	* possibly making the queue appear nonempty when a push, pop, or
66	>	* deq have not fully committed. Note that this means that the deq
67	>	* operation, considered individually, is not wait-free. One thief
68	>	* cannot successfully continue until another in-progress one (or,
69	>	* if previously empty, a push) completes.  However, in the
70	>	* aggregate, we ensure at least probabilistic non-blockingness.
71	>	* If an attempted steal fails, a thief always chooses a different
72		* random victim target to try next. So, in order for one thief to
73		* progress, it suffices for any in-progress deq or new push on
74		* any empty queue to complete. One reason this works well here is
75		* that apparently-nonempty often means soon-to-be-stealable,
76	<	* which gives threads a chance to activate if necessary before
77	<	* stealing (see below).
76	>	* which gives threads a chance to set activation status if
77	>	* necessary before stealing.
78	>	*
79	>	* This approach also enables support for "async mode" where local
80	>	* task processing is in FIFO, not LIFO order; simply by using a
81	>	* version of deq rather than pop when locallyFifo is true (as set
82	>	* by the ForkJoinPool).  This allows use in message-passing
83	>	* frameworks in which tasks are never joined.
84		*
85	<	* Efficient implementation of this approach currently relies on
86	<	* an uncomfortable amount of "Unsafe" mechanics. To maintain
85	>	* When a worker would otherwise be blocked waiting to join a
86	>	* task, it first tries a form of linear helping: Each worker
87	>	* records (in field currentSteal) the most recent task it stole
88	>	* from some other worker. Plus, it records (in field currentJoin)
89	>	* the task it is currently actively joining. Method joinTask uses
90	>	* these markers to try to find a worker to help (i.e., steal back
91	>	* a task from and execute it) that could hasten completion of the
92	>	* actively joined task. In essence, the joiner executes a task
93	>	* that would be on its own local deque had the to-be-joined task
94	>	* not been stolen. This may be seen as a conservative variant of
95	>	* the approach in Wagner & Calder "Leapfrogging: a portable
96	>	* technique for implementing efficient futures" SIGPLAN Notices,
97	>	* 1993 (http://portal.acm.org/citation.cfm?id=155354). It differs
98	>	* in that: (1) We only maintain dependency links across workers
99	>	* upon steals, rather than use per-task bookkeeping.  This may
100	>	* require a linear scan of workers array to locate stealers, but
101	>	* usually doesn't because stealers leave hints (that may become
102	>	* stale/wrong) of where to locate them. This isolates cost to
103	>	* when it is needed, rather than adding to per-task overhead.
104	>	* (2) It is "shallow", ignoring nesting and potentially cyclic
105	>	* mutual steals.  (3) It is intentionally racy: field currentJoin
106	>	* is updated only while actively joining, which means that we
107	>	* miss links in the chain during long-lived tasks, GC stalls etc
108	>	* (which is OK since blocking in such cases is usually a good
109	>	* idea).  (4) We bound the number of attempts to find work (see
110	>	* MAX_HELP_DEPTH) and fall back to suspending the worker and if
111	>	* necessary replacing it with a spare (see
112	>	* ForkJoinPool.awaitJoin).
113	>	*
114	>	* Efficient implementation of these algorithms currently relies
115	>	* on an uncomfortable amount of "Unsafe" mechanics. To maintain
116		* correct orderings, reads and writes of variable base require
117	<	* volatile ordering.  Variable sp does not require volatile write
118	<	* but needs cheaper store-ordering on writes.  Because they are
119	<	* protected by volatile base reads, reads of the queue array and
120	<	* its slots do not need volatile load semantics, but writes (in
121	<	* push) require store order and CASes (in pop and deq) require
122	<	* (volatile) CAS semantics. Since these combinations aren't
123	<	* supported using ordinary volatiles, the only way to accomplish
124	<	* these efficiently is to use direct Unsafe calls. (Using external
125	<	* AtomicIntegers and AtomicReferenceArrays for the indices and
126	<	* array is significantly slower because of memory locality and
127	<	* indirection effects.) Further, performance on most platforms is
128	<	* very sensitive to placement and sizing of the (resizable) queue
129	<	* array.  Even though these queues don't usually become all that
130	<	* big, the initial size must be large enough to counteract cache
117	>	* volatile ordering.  Variable sp does not require volatile
118	>	* writes but still needs store-ordering, which we accomplish by
119	>	* pre-incrementing sp before filling the slot with an ordered
120	>	* store.  (Pre-incrementing also enables backouts used in
121	>	* joinTask.)  Because they are protected by volatile base reads,
122	>	* reads of the queue array and its slots by other threads do not
123	>	* need volatile load semantics, but writes (in push) require
124	>	* store order and CASes (in pop and deq) require (volatile) CAS
125	>	* semantics.  (Michael, Saraswat, and Vechev's algorithm has
126	>	* similar properties, but without support for nulling slots.)
127	>	* Since these combinations aren't supported using ordinary
128	>	* volatiles, the only way to accomplish these efficiently is to
129	>	* use direct Unsafe calls. (Using external AtomicIntegers and
130	>	* AtomicReferenceArrays for the indices and array is
131	>	* significantly slower because of memory locality and indirection
132	>	* effects.)
133	>	*
134	>	* Further, performance on most platforms is very sensitive to
135	>	* placement and sizing of the (resizable) queue array.  Even
136	>	* though these queues don't usually become all that big, the
137	>	* initial size must be large enough to counteract cache
138		* contention effects across multiple queues (especially in the
139		* presence of GC cardmarking). Also, to improve thread-locality,
140	<	* queues are currently initialized immediately after the thread
141	<	* gets the initial signal to start processing tasks.  However,
142	<	* all queue-related methods except pushTask are written in a way
143	<	* that allows them to instead be lazily allocated and/or disposed
144	<	* of when empty. All together, these low-level implementation
145	<	* choices produce as much as a factor of 4 performance
146	<	* improvement compared to naive implementations, and enable the
147	<	* processing of billions of tasks per second, sometimes at the
148	<	* expense of ugliness.
149	<	*
150	<	* 2. Run control: The primary run control is based on a global
151	<	* counter (activeCount) held by the pool. It uses an algorithm
152	<	* similar to that in Herlihy and Shavit section 17.6 to cause
153	<	* threads to eventually block when all threads declare they are
154	<	* inactive. (See variable "scans".)  For this to work, threads
155	<	* must be declared active when executing tasks, and before
156	<	* stealing a task. They must be inactive before blocking on the
157	<	* Pool Barrier (awaiting a new submission or other Pool
158	<	* event). In between, there is some free play which we take
159	<	* advantage of to avoid contention and rapid flickering of the
160	<	* global activeCount: If inactive, we activate only if a victim
161	<	* 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.
140	>	* queues are initialized after starting.  All together, these
141	>	* low-level implementation choices produce as much as a factor of
142	>	* 4 performance improvement compared to naive implementations,
143	>	* and enable the processing of billions of tasks per second,
144	>	* sometimes at the expense of ugliness.
145	>	*/
146	>
147	>	/**
148	>	* Generator for initial random seeds for random victim
149	>	* selection. This is used only to create initial seeds. Random
150	>	* steals use a cheaper xorshift generator per steal attempt. We
151	>	* expect only rare contention on seedGenerator, so just use a
152	>	* plain Random.
153	>	*/
154	>	private static final Random seedGenerator = new Random();
155	>
156	>	/**
157	>	* The maximum stolen->joining link depth allowed in helpJoinTask.
158	>	* Depths for legitimate chains are unbounded, but we use a fixed
159	>	* constant to avoid (otherwise unchecked) cycles and bound
160	>	* staleness of traversal parameters at the expense of sometimes
161	>	* blocking when we could be helping.
162		*/
163	+	private static final int MAX_HELP_DEPTH = 8;
164
165		/**
166		* Capacity of work-stealing queue array upon initialization.
167	<	* Must be a power of two. Initial size must be at least 2, but is
167	>	* Must be a power of two. Initial size must be at least 4, but is
168		* padded to minimize cache effects.
169		*/
170		private static final int INITIAL_QUEUE_CAPACITY = 1 << 13;
171
172		/**
173		* Maximum work-stealing queue array size.  Must be less than or
174	<	* equal to 1 << 28 to ensure lack of index wraparound. (This
175	<	* is less than usual bounds, because we need leftshift by 3
176	<	* to be in int range).
174	>	* equal to 1 << (31 - width of array entry) to ensure lack of
175	>	* index wraparound. The value is set in the static block
176	>	* at the end of this file after obtaining width.
177		*/
178	<	private static final int MAXIMUM_QUEUE_CAPACITY = 1 << 28;
178	>	private static final int MAXIMUM_QUEUE_CAPACITY;
179
180		/**
181		* The pool this thread works in. Accessed directly by ForkJoinTask.
#	Line 143 | Line 184 | public class ForkJoinWorkerThread extend
184
185		/**
186		* The work-stealing queue array. Size must be a power of two.
187	<	* Initialized when thread starts, to improve memory locality.
187	>	* Initialized in onStart, to improve memory locality.
188		*/
189		private ForkJoinTask<?>[] queue;
190
191		/**
192	+	* Index (mod queue.length) of least valid queue slot, which is
193	+	* always the next position to steal from if nonempty.
194	+	*/
195	+	private volatile int base;
196	+
197	+	/**
198		* Index (mod queue.length) of next queue slot to push to or pop
199	<	* from. It is written only by owner thread, via ordered store.
200	<	* Both sp and base are allowed to wrap around on overflow, but
201	<	* (sp - base) still estimates size.
199	>	* from. It is written only by owner thread, and accessed by other
200	>	* threads only after reading (volatile) base.  Both sp and base
201	>	* are allowed to wrap around on overflow, but (sp - base) still
202	>	* estimates size.
203		*/
204	<	private volatile int sp;
204	>	private int sp;
205
206		/**
207	<	* Index (mod queue.length) of least valid queue slot, which is
208	<	* always the next position to steal from if nonempty.
207	>	* The index of most recent stealer, used as a hint to avoid
208	>	* traversal in method helpJoinTask. This is only a hint because a
209	>	* worker might have had multiple steals and this only holds one
210	>	* of them (usually the most current). Declared non-volatile,
211	>	* relying on other prevailing sync to keep reasonably current.
212		*/
213	<	private volatile int base;
213	>	private int stealHint;
214
215		/**
216	<	* Activity status. When true, this worker is considered active.
217	<	* Must be false upon construction. It must be true when executing
218	<	* tasks, and BEFORE stealing a task. It must be false before
219	<	* calling pool.sync.
216	>	* Run state of this worker. In addition to the usual run levels,
217	>	* tracks if this worker is suspended as a spare, and if it was
218	>	* killed (trimmed) while suspended. However, "active" status is
219	>	* maintained separately and modified only in conjunction with
220	>	* CASes of the pool's runState (which are currently sadly
221	>	* manually inlined for performance.)  Accessed directly by pool
222	>	* to simplify checks for normal (zero) status.
223		*/
224	<	private boolean active;
224	>	volatile int runState;
225	>
226	>	private static final int TERMINATING = 0x01;
227	>	private static final int TERMINATED  = 0x02;
228	>	private static final int SUSPENDED   = 0x04; // inactive spare
229	>	private static final int TRIMMED     = 0x08; // killed while suspended
230
231		/**
232	<	* Run state of this worker. Supports simple versions of the usual
233	<	* shutdown/shutdownNow control.
232	>	* Number of steals. Directly accessed (and reset) by
233	>	* pool.tryAccumulateStealCount when idle.
234		*/
235	<	private volatile int runState;
235	>	int stealCount;
236
237		/**
238		* Seed for random number generator for choosing steal victims.
239	<	* Uses Marsaglia xorshift. Must be nonzero upon initialization.
239	>	* Uses Marsaglia xorshift. Must be initialized as nonzero.
240		*/
241		private int seed;
242
243		/**
244	<	* Number of steals, transferred to pool when idle
244	>	* Activity status. When true, this worker is considered active.
245	>	* Accessed directly by pool.  Must be false upon construction.
246	>	*/
247	>	boolean active;
248	>
249	>	/**
250	>	* True if use local fifo, not default lifo, for local polling.
251	>	* Shadows value from ForkJoinPool.
252		*/
253	<	private int stealCount;
253	>	private final boolean locallyFifo;
254
255		/**
256		* Index of this worker in pool array. Set once by pool before
257	<	* running, and accessed directly by pool during cleanup etc.
257	>	* running, and accessed directly by pool to locate this worker in
258	>	* its workers array.
259		*/
260		int poolIndex;
261
262		/**
263	<	* The last barrier event waited for. Accessed in pool callback
264	<	* methods, but only by current thread.
263	>	* The last pool event waited for. Accessed only by pool in
264	>	* callback methods invoked within this thread.
265		*/
266	<	long lastEventCount;
266	>	int lastEventCount;
267
268		/**
269	<	* True if use local fifo, not default lifo, for local polling
269	>	* Encoded index and event count of next event waiter. Accessed
270	>	* only by ForkJoinPool for managing event waiters.
271		*/
272	<	private boolean locallyFifo;
272	>	volatile long nextWaiter;
273	>
274	>	/**
275	>	* Number of times this thread suspended as spare. Accessed only
276	>	* by pool.
277	>	*/
278	>	int spareCount;
279	>
280	>	/**
281	>	* Encoded index and count of next spare waiter. Accessed only
282	>	* by ForkJoinPool for managing spares.
283	>	*/
284	>	volatile int nextSpare;
285	>
286	>	/**
287	>	* The task currently being joined, set only when actively trying
288	>	* to help other stealers in helpJoinTask. Written only by this
289	>	* thread, but read by others.
290	>	*/
291	>	private volatile ForkJoinTask<?> currentJoin;
292	>
293	>	/**
294	>	* The task most recently stolen from another worker (or
295	>	* submission queue).  Written only by this thread, but read by
296	>	* others.
297	>	*/
298	>	private volatile ForkJoinTask<?> currentSteal;
299
300		/**
301		* Creates a ForkJoinWorkerThread operating in the given pool.
#	Line 210 | Line 304 | public class ForkJoinWorkerThread extend
304		* @throws NullPointerException if pool is null
305		*/
306		protected ForkJoinWorkerThread(ForkJoinPool pool) {
213	–	if (pool == null) throw new NullPointerException();
307		this.pool = pool;
308	<	// Note: poolIndex is set by pool during construction
309	<	// Remaining initialization is deferred to onStart
308	>	this.locallyFifo = pool.locallyFifo;
309	>	setDaemon(true);
310	>	// To avoid exposing construction details to subclasses,
311	>	// remaining initialization is in start() and onStart()
312		}
313
314	<	// Public access methods
314	>	/**
315	>	* Performs additional initialization and starts this thread.
316	>	*/
317	>	final void start(int poolIndex, UncaughtExceptionHandler ueh) {
318	>	this.poolIndex = poolIndex;
319	>	if (ueh != null)
320	>	setUncaughtExceptionHandler(ueh);
321	>	start();
322	>	}
323	>
324	>	// Public/protected methods
325
326		/**
327		* Returns the pool hosting this thread.
#	Line 241 | Line 346 | public class ForkJoinWorkerThread extend
346		}
347
348		/**
349	<	* Establishes local first-in-first-out scheduling mode for forked
350	<	* tasks that are never joined.
351	<	*
352	<	* @param async if true, use locally FIFO scheduling
349	>	* Initializes internal state after construction but before
350	>	* processing any tasks. If you override this method, you must
351	>	* invoke {@code super.onStart()} at the beginning of the method.
352	>	* Initialization requires care: Most fields must have legal
353	>	* default values, to ensure that attempted accesses from other
354	>	* threads work correctly even before this thread starts
355	>	* processing tasks.
356		*/
357	<	void setAsyncMode(boolean async) {
358	<	locallyFifo = async;
359	<	}
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;
260	<
261	<	final boolean isShutdown()    { return runState >= SHUTDOWN;  }
262	<	final boolean isTerminating() { return runState >= TERMINATING;  }
263	<	final boolean isTerminated()  { return runState == TERMINATED; }
264	<	final boolean shutdown()      { return transitionRunStateTo(SHUTDOWN); }
265	<	final boolean shutdownNow()   { return transitionRunStateTo(TERMINATING); }
357	>	protected void onStart() {
358	>	int rs = seedGenerator.nextInt();
359	>	seed = rs == 0? 1 : rs; // seed must be nonzero
360
361	<	/**
362	<	* Transitions to at least the given state.  Returns true if not
363	<	* already at least at given state.
364	<	*/
271	<	private boolean transitionRunStateTo(int state) {
272	<	for (;;) {
273	<	int s = runState;
274	<	if (s >= state)
275	<	return false;
276	<	if (UNSAFE.compareAndSwapInt(this, runStateOffset, s, state))
277	<	return true;
278	<	}
279	<	}
361	>	// Allocate name string and arrays in this thread
362	>	String pid = Integer.toString(pool.getPoolNumber());
363	>	String wid = Integer.toString(poolIndex);
364	>	setName("ForkJoinPool-" + pid + "-worker-" + wid);
365
366	<	/**
282	<	* Tries to set status to active; fails on contention.
283	<	*/
284	<	private boolean tryActivate() {
285	<	if (!active) {
286	<	if (!pool.tryIncrementActiveCount())
287	<	return false;
288	<	active = true;
289	<	}
290	<	return true;
366	>	queue = new ForkJoinTask<?>[INITIAL_QUEUE_CAPACITY];
367		}
368
369		/**
370	<	* Tries to set status to inactive; fails on contention.
370	>	* Performs cleanup associated with termination of this worker
371	>	* thread.  If you override this method, you must invoke
372	>	* {@code super.onTermination} at the end of the overridden method.
373	>	*
374	>	* @param exception the exception causing this thread to abort due
375	>	* to an unrecoverable error, or {@code null} if completed normally
376		*/
377	<	private boolean tryInactivate() {
378	<	if (active) {
379	<	if (!pool.tryDecrementActiveCount())
380	<	return false;
381	<	active = false;
377	>	protected void onTermination(Throwable exception) {
378	>	try {
379	>	ForkJoinPool p = pool;
380	>	if (active) {
381	>	int a; // inline p.tryDecrementActiveCount
382	>	active = false;
383	>	do {} while (!UNSAFE.compareAndSwapInt
384	>	(p, poolRunStateOffset, a = p.runState, a - 1));
385	>	}
386	>	cancelTasks();
387	>	setTerminated();
388	>	p.workerTerminated(this);
389	>	} catch (Throwable ex) {        // Shouldn't ever happen
390	>	if (exception == null)      // but if so, at least rethrown
391	>	exception = ex;
392	>	} finally {
393	>	if (exception != null)
394	>	UNSAFE.throwException(exception);
395		}
302	–	return true;
303	–	}
304	–
305	–	/**
306	–	* Computes next value for random victim probe.  Scans don't
307	–	* require a very high quality generator, but also not a crummy
308	–	* one.  Marsaglia xor-shift is cheap and works well.
309	–	*/
310	–	private static int xorShift(int r) {
311	–	r ^= r << 1;
312	–	r ^= r >>> 3;
313	–	r ^= r << 10;
314	–	return r;
396		}
397
317	–	// Lifecycle methods
318	–
398		/**
399		* This method is required to be public, but should never be
400		* called explicitly. It performs the main run loop to execute
#	Line 325 | Line 404 | public class ForkJoinWorkerThread extend
404		Throwable exception = null;
405		try {
406		onStart();
328	–	pool.sync(this); // await first pool event
407		mainLoop();
408		} catch (Throwable ex) {
409		exception = ex;
#	Line 334 | Line 412 | public class ForkJoinWorkerThread extend
412		}
413		}
414
415	+	// helpers for run()
416	+
417		/**
418	<	* Executes tasks until shut down.
418	>	* Finds and executes tasks, and checks status while running.
419		*/
420		private void mainLoop() {
421	<	while (!isShutdown()) {
422	<	ForkJoinTask<?> t = pollTask();
423	<	if (t != null || (t = pollSubmission()) != null)
424	<	t.quietlyExec();
425	<	else if (tryInactivate())
426	<	pool.sync(this);
421	>	boolean ran = false; // true if ran a task on last step
422	>	ForkJoinPool p = pool;
423	>	for (;;) {
424	>	p.preStep(this, ran);
425	>	if (runState != 0)
426	>	break;
427	>	ran = tryExecSteal() || tryExecSubmission();
428		}
429		}
430
431		/**
432	<	* Initializes internal state after construction but before
433	<	* processing any tasks. If you override this method, you must
434	<	* invoke super.onStart() at the beginning of the method.
354	<	* Initialization requires care: Most fields must have legal
355	<	* default values, to ensure that attempted accesses from other
356	<	* threads work correctly even before this thread starts
357	<	* processing tasks.
432	>	* Tries to steal a task and execute it.
433	>	*
434	>	* @return true if ran a task
435		*/
436	<	protected void onStart() {
437	<	// Allocate while starting to improve chances of thread-local
438	<	// isolation
439	<	queue = new ForkJoinTask<?>[INITIAL_QUEUE_CAPACITY];
440	<	// Initial value of seed need not be especially random but
441	<	// should differ across workers and must be nonzero
442	<	int p = poolIndex + 1;
443	<	seed = p + (p << 8) + (p << 16) + (p << 24); // spread bits
436	>	private boolean tryExecSteal() {
437	>	ForkJoinTask<?> t;
438	>	if ((t = scan()) != null) {
439	>	t.quietlyExec();
440	>	UNSAFE.putOrderedObject(this, currentStealOffset, null);
441	>	if (sp != base)
442	>	execLocalTasks();
443	>	return true;
444	>	}
445	>	return false;
446		}
447
448		/**
449	<	* Performs cleanup associated with termination of this worker
371	<	* thread.  If you override this method, you must invoke
372	<	* {@code super.onTermination} at the end of the overridden method.
449	>	* If a submission exists, try to activate and run it.
450		*
451	<	* @param exception the exception causing this thread to abort due
375	<	* to an unrecoverable error, or null if completed normally
451	>	* @return true if ran a task
452		*/
453	<	protected void onTermination(Throwable exception) {
454	<	// Execute remaining local tasks unless aborting or terminating
455	<	while (exception == null &&  !pool.isTerminating() && base != sp) {
456	<	try {
457	<	ForkJoinTask<?> t = popTask();
458	<	if (t != null)
453	>	private boolean tryExecSubmission() {
454	>	ForkJoinPool p = pool;
455	>	// This loop is needed in case attempt to activate fails, in
456	>	// which case we only retry if there still appears to be a
457	>	// submission.
458	>	while (p.hasQueuedSubmissions()) {
459	>	ForkJoinTask<?> t; int a;
460	>	if (active || // inline p.tryIncrementActiveCount
461	>	(active = UNSAFE.compareAndSwapInt(p, poolRunStateOffset,
462	>	a = p.runState, a + 1))) {
463	>	if ((t = p.pollSubmission()) != null) {
464	>	UNSAFE.putOrderedObject(this, currentStealOffset, t);
465		t.quietlyExec();
466	<	} catch (Throwable ex) {
467	<	exception = ex;
466	>	UNSAFE.putOrderedObject(this, currentStealOffset, null);
467	>	if (sp != base)
468	>	execLocalTasks();
469	>	return true;
470	>	}
471		}
472		}
473	<	// Cancel other tasks, transition status, notify pool, and
389	<	// propagate exception to uncaught exception handler
390	<	try {
391	<	do {} while (!tryInactivate()); // ensure inactive
392	<	cancelTasks();
393	<	runState = TERMINATED;
394	<	pool.workerTerminated(this);
395	<	} catch (Throwable ex) {        // Shouldn't ever happen
396	<	if (exception == null)      // but if so, at least rethrown
397	<	exception = ex;
398	<	} finally {
399	<	if (exception != null)
400	<	ForkJoinTask.rethrowException(exception);
401	<	}
473	>	return false;
474		}
475
404	–	// Intrinsics-based support for queue operations.
405	–
476		/**
477	<	* Adds in store-order the given task at given slot of q to null.
478	<	* Caller must ensure q is non-null and index is in range.
477	>	* Runs local tasks until queue is empty or shut down.  Call only
478	>	* while active.
479		*/
480	<	private static void setSlot(ForkJoinTask<?>[] q, int i,
481	<	ForkJoinTask<?> t) {
482	<	UNSAFE.putOrderedObject(q, (i << qShift) + qBase, t);
480	>	private void execLocalTasks() {
481	>	while (runState == 0) {
482	>	ForkJoinTask<?> t = locallyFifo ? locallyDeqTask() : popTask();
483	>	if (t != null)
484	>	t.quietlyExec();
485	>	else if (sp == base)
486	>	break;
487	>	}
488		}
489
490	+	/*
491	+	* Intrinsics-based atomic writes for queue slots. These are
492	+	* basically the same as methods in AtomicReferenceArray, but
493	+	* specialized for (1) ForkJoinTask elements (2) requirement that
494	+	* nullness and bounds checks have already been performed by
495	+	* callers and (3) effective offsets are known not to overflow
496	+	* from int to long (because of MAXIMUM_QUEUE_CAPACITY). We don't
497	+	* need corresponding version for reads: plain array reads are OK
498	+	* because they are protected by other volatile reads and are
499	+	* confirmed by CASes.
500	+	*
501	+	* Most uses don't actually call these methods, but instead contain
502	+	* inlined forms that enable more predictable optimization.  We
503	+	* don't define the version of write used in pushTask at all, but
504	+	* instead inline there a store-fenced array slot write.
505	+	*/
506	+
507		/**
508	<	* CAS given slot of q to null. Caller must ensure q is non-null
509	<	* and index is in range.
508	>	* CASes slot i of array q from t to null. Caller must ensure q is
509	>	* non-null and index is in range.
510		*/
511	<	private static boolean casSlotNull(ForkJoinTask<?>[] q, int i,
512	<	ForkJoinTask<?> t) {
511	>	private static final boolean casSlotNull(ForkJoinTask<?>[] q, int i,
512	>	ForkJoinTask<?> t) {
513		return UNSAFE.compareAndSwapObject(q, (i << qShift) + qBase, t, null);
514		}
515
516		/**
517	<	* Sets sp in store-order.
517	>	* Performs a volatile write of the given task at given slot of
518	>	* array q.  Caller must ensure q is non-null and index is in
519	>	* range. This method is used only during resets and backouts.
520		*/
521	<	private void storeSp(int s) {
522	<	UNSAFE.putOrderedInt(this, spOffset, s);
521	>	private static final void writeSlot(ForkJoinTask<?>[] q, int i,
522	>	ForkJoinTask<?> t) {
523	>	UNSAFE.putObjectVolatile(q, (i << qShift) + qBase, t);
524		}
525
526	<	// Main queue methods
526	>	// queue methods
527
528		/**
529	<	* Pushes a task. Called only by current thread.
529	>	* Pushes a task. Call only from this thread.
530		*
531		* @param t the task. Caller must ensure non-null.
532		*/
533		final void pushTask(ForkJoinTask<?> t) {
534		ForkJoinTask<?>[] q = queue;
535	<	int mask = q.length - 1;
536	<	int s = sp;
537	<	setSlot(q, s & mask, t);
538	<	storeSp(++s);
539	<	if ((s -= base) == 1)
540	<	pool.signalWork();
541	<	else if (s >= mask)
447	<	growQueue();
535	>	int mask = q.length - 1; // implicit assert q != null
536	>	int s = sp++;            // ok to increment sp before slot write
537	>	UNSAFE.putOrderedObject(q, ((s & mask) << qShift) + qBase, t);
538	>	if ((s -= base) == 0)
539	>	pool.signalWork();   // was empty
540	>	else if (s == mask)
541	>	growQueue();         // is full
542		}
543
544		/**
545		* Tries to take a task from the base of the queue, failing if
546	<	* either empty or contended.
546	>	* empty or contended. Note: Specializations of this code appear
547	>	* in locallyDeqTask and elsewhere.
548		*
549		* @return a task, or null if none or contended
550		*/
551		final ForkJoinTask<?> deqTask() {
552		ForkJoinTask<?> t;
553		ForkJoinTask<?>[] q;
554	<	int i;
460	<	int b;
554	>	int b, i;
555		if (sp != (b = base) &&
556		(q = queue) != null && // must read q after b
557	<	(t = q[i = (q.length - 1) & b]) != null &&
558	<	casSlotNull(q, i, t)) {
557	>	(t = q[i = (q.length - 1) & b]) != null && base == b &&
558	>	UNSAFE.compareAndSwapObject(q, (i << qShift) + qBase, t, null)) {
559		base = b + 1;
560		return t;
561		}
#	Line 469 | Line 563 | public class ForkJoinWorkerThread extend
563		}
564
565		/**
566	<	* Returns a popped task, or null if empty. Ensures active status
567	<	* if non-null. Called only by current thread.
566	>	* Tries to take a task from the base of own queue. Assumes active
567	>	* status.  Called only by this thread.
568	>	*
569	>	* @return a task, or null if none
570	>	*/
571	>	final ForkJoinTask<?> locallyDeqTask() {
572	>	ForkJoinTask<?>[] q = queue;
573	>	if (q != null) {
574	>	ForkJoinTask<?> t;
575	>	int b, i;
576	>	while (sp != (b = base)) {
577	>	if ((t = q[i = (q.length - 1) & b]) != null && base == b &&
578	>	UNSAFE.compareAndSwapObject(q, (i << qShift) + qBase,
579	>	t, null)) {
580	>	base = b + 1;
581	>	return t;
582	>	}
583	>	}
584	>	}
585	>	return null;
586	>	}
587	>
588	>	/**
589	>	* Returns a popped task, or null if empty. Assumes active status.
590	>	* Called only by this thread.
591		*/
592	<	final ForkJoinTask<?> popTask() {
593	<	int s = sp;
594	<	while (s != base) {
595	<	if (tryActivate()) {
596	<	ForkJoinTask<?>[] q = queue;
597	<	int mask = q.length - 1;
598	<	int i = (s - 1) & mask;
592	>	private ForkJoinTask<?> popTask() {
593	>	ForkJoinTask<?>[] q = queue;
594	>	if (q != null) {
595	>	int s;
596	>	while ((s = sp) != base) {
597	>	int i = (q.length - 1) & --s;
598	>	long u = (i << qShift) + qBase; // raw offset
599		ForkJoinTask<?> t = q[i];
600	<	if (t == null || !casSlotNull(q, i, t))
600	>	if (t == null)   // lost to stealer
601		break;
602	<	storeSp(s - 1);
603	<	return t;
602	>	if (UNSAFE.compareAndSwapObject(q, u, t, null)) {
603	>	sp = s; // putOrderedInt may encourage more timely write
604	>	// UNSAFE.putOrderedInt(this, spOffset, s);
605	>	return t;
606	>	}
607		}
608		}
609		return null;
610		}
611
612		/**
613	<	* Specialized version of popTask to pop only if
614	<	* topmost element is the given task. Called only
495	<	* by current thread while active.
613	>	* Specialized version of popTask to pop only if topmost element
614	>	* is the given task. Called only by this thread while active.
615		*
616		* @param t the task. Caller must ensure non-null.
617		*/
618		final boolean unpushTask(ForkJoinTask<?> t) {
619	+	int s;
620		ForkJoinTask<?>[] q = queue;
621	<	int mask = q.length - 1;
622	<	int s = sp - 1;
623	<	if (casSlotNull(q, s & mask, t)) {
624	<	storeSp(s);
621	>	if ((s = sp) != base && q != null &&
622	>	UNSAFE.compareAndSwapObject
623	>	(q, (((q.length - 1) & --s) << qShift) + qBase, t, null)) {
624	>	sp = s; // putOrderedInt may encourage more timely write
625	>	// UNSAFE.putOrderedInt(this, spOffset, s);
626		return true;
627		}
628		return false;
629		}
630
631		/**
632	<	* Returns next task.
632	>	* Returns next task, or null if empty or contended.
633		*/
634		final ForkJoinTask<?> peekTask() {
635		ForkJoinTask<?>[] q = queue;
#	Line 541 | Line 662 | public class ForkJoinWorkerThread extend
662		ForkJoinTask<?> t = oldQ[oldIndex];
663		if (t != null && !casSlotNull(oldQ, oldIndex, t))
664		t = null;
665	<	setSlot(newQ, b & newMask, t);
665	>	writeSlot(newQ, b & newMask, t);
666		} while (++b != bf);
667		pool.signalWork();
668		}
669
670		/**
671	+	* Computes next value for random victim probe in scan().  Scans
672	+	* don't require a very high quality generator, but also not a
673	+	* crummy one.  Marsaglia xor-shift is cheap and works well enough.
674	+	* Note: This is manually inlined in scan().
675	+	*/
676	+	private static final int xorShift(int r) {
677	+	r ^= r << 13;
678	+	r ^= r >>> 17;
679	+	return r ^ (r << 5);
680	+	}
681	+
682	+	/**
683		* Tries to steal a task from another worker. Starts at a random
684		* index of workers array, and probes workers until finding one
685		* with non-empty queue or finding that all are empty.  It
686		* randomly selects the first n probes. If these are empty, it
687	<	* resorts to a full circular traversal, which is necessary to
688	<	* accurately set active status by caller. Also restarts if pool
689	<	* events occurred since last scan, which forces refresh of
690	<	* workers array, in case barrier was associated with resize.
687	>	* resorts to a circular sweep, which is necessary to accurately
688	>	* set active status. (The circular sweep uses steps of
689	>	* approximately half the array size plus 1, to avoid bias
690	>	* stemming from leftmost packing of the array in ForkJoinPool.)
691		*
692		* This method must be both fast and quiet -- usually avoiding
693		* memory accesses that could disrupt cache sharing etc other than
694	<	* those needed to check for and take tasks. This accounts for,
695	<	* among other things, updating random seed in place without
696	<	* storing it until exit.
694	>	* those needed to check for and take tasks (or to activate if not
695	>	* already active). This accounts for, among other things,
696	>	* updating random seed in place without storing it until exit.
697		*
698		* @return a task, or null if none found
699		*/
700		private ForkJoinTask<?> scan() {
701	<	ForkJoinTask<?> t = null;
702	<	int r = seed;                    // extract once to keep scan quiet
703	<	ForkJoinWorkerThread[] ws;       // refreshed on outer loop
704	<	int mask;                        // must be power 2 minus 1 and > 0
705	<	outer:do {
706	<	if ((ws = pool.workers) != null && (mask = ws.length - 1) > 0) {
707	<	int idx = r;
708	<	int probes = ~mask;      // use random index while negative
709	<	for (;;) {
710	<	r = xorShift(r);     // update random seed
711	<	ForkJoinWorkerThread v = ws[mask & idx];
712	<	if (v == null || v.sp == v.base) {
713	<	if (probes <= mask)
714	<	idx = (probes++ < 0) ? r : (idx + 1);
715	<	else
716	<	break;
701	>	ForkJoinPool p = pool;
702	>	ForkJoinWorkerThread[] ws;        // worker array
703	>	int n;                            // upper bound of #workers
704	>	if ((ws = p.workers) != null && (n = ws.length) > 1) {
705	>	boolean canSteal = active;    // shadow active status
706	>	int r = seed;                 // extract seed once
707	>	int mask = n - 1;
708	>	int j = -n;                   // loop counter
709	>	int k = r;                    // worker index, random if j < 0
710	>	for (;;) {
711	>	ForkJoinWorkerThread v = ws[k & mask];
712	>	r ^= r << 13; r ^= r >>> 17; r ^= r << 5; // inline xorshift
713	>	ForkJoinTask<?>[] q; ForkJoinTask<?> t; int b, a;
714	>	if (v != null && (b = v.base) != v.sp &&
715	>	(q = v.queue) != null) {
716	>	int i = (q.length - 1) & b;
717	>	long u = (i << qShift) + qBase; // raw offset
718	>	int pid = poolIndex;
719	>	if ((t = q[i]) != null) {
720	>	if (!canSteal &&  // inline p.tryIncrementActiveCount
721	>	UNSAFE.compareAndSwapInt(p, poolRunStateOffset,
722	>	a = p.runState, a + 1))
723	>	canSteal = active = true;
724	>	if (canSteal && v.base == b++ &&
725	>	UNSAFE.compareAndSwapObject(q, u, t, null)) {
726	>	v.base = b;
727	>	v.stealHint = pid;
728	>	UNSAFE.putOrderedObject(this,
729	>	currentStealOffset, t);
730	>	seed = r;
731	>	++stealCount;
732	>	return t;
733	>	}
734		}
735	<	else if (!tryActivate() || (t = v.deqTask()) == null)
736	<	continue outer;  // restart on contention
587	<	else
588	<	break outer;
735	>	j = -n;
736	>	k = r;                // restart on contention
737		}
738	+	else if (++j <= 0)
739	+	k = r;
740	+	else if (j <= n)
741	+	k += (n >>> 1) | 1;
742	+	else
743	+	break;
744		}
745	<	} while (pool.hasNewSyncEvent(this)); // retry on pool events
746	<	seed = r;
747	<	return t;
745	>	}
746	>	return null;
747	>	}
748	>
749	>	// Run State management
750	>
751	>	// status check methods used mainly by ForkJoinPool
752	>	final boolean isRunning()    { return runState == 0; }
753	>	final boolean isTerminated() { return (runState & TERMINATED) != 0; }
754	>	final boolean isSuspended()  { return (runState & SUSPENDED) != 0; }
755	>	final boolean isTrimmed()    { return (runState & TRIMMED) != 0; }
756	>
757	>	final boolean isTerminating() {
758	>	if ((runState & TERMINATING) != 0)
759	>	return true;
760	>	if (pool.isAtLeastTerminating()) { // propagate pool state
761	>	shutdown();
762	>	return true;
763	>	}
764	>	return false;
765		}
766
767		/**
768	<	* Gets and removes a local or stolen task.
769	<	*
599	<	* @return a task, if available
768	>	* Sets state to TERMINATING. Does NOT unpark or interrupt
769	>	* to wake up if currently blocked. Callers must do so if desired.
770		*/
771	<	final ForkJoinTask<?> pollTask() {
772	<	ForkJoinTask<?> t = locallyFifo ? deqTask() : popTask();
773	<	if (t == null && (t = scan()) != null)
774	<	++stealCount;
775	<	return t;
771	>	final void shutdown() {
772	>	for (;;) {
773	>	int s = runState;
774	>	if ((s & (TERMINATING|TERMINATED)) != 0)
775	>	break;
776	>	if ((s & SUSPENDED) != 0) { // kill and wakeup if suspended
777	>	if (UNSAFE.compareAndSwapInt(this, runStateOffset, s,
778	>	(s & ~SUSPENDED) |
779	>	(TRIMMED|TERMINATING)))
780	>	break;
781	>	}
782	>	else if (UNSAFE.compareAndSwapInt(this, runStateOffset, s,
783	>	s | TERMINATING))
784	>	break;
785	>	}
786		}
787
788		/**
789	<	* Gets a local task.
610	<	*
611	<	* @return a task, if available
789	>	* Sets state to TERMINATED. Called only by onTermination().
790		*/
791	<	final ForkJoinTask<?> pollLocalTask() {
792	<	return locallyFifo ? deqTask() : popTask();
791	>	private void setTerminated() {
792	>	int s;
793	>	do {} while (!UNSAFE.compareAndSwapInt(this, runStateOffset,
794	>	s = runState,
795	>	s | (TERMINATING|TERMINATED)));
796		}
797
798		/**
799	<	* Returns a pool submission, if one exists, activating first.
799	>	* If suspended, tries to set status to unsuspended.
800	>	* Does NOT wake up if blocked.
801		*
802	<	* @return a submission, if available
802	>	* @return true if successful
803		*/
804	<	private ForkJoinTask<?> pollSubmission() {
804	>	final boolean tryUnsuspend() {
805	>	int s;
806	>	while (((s = runState) & SUSPENDED) != 0) {
807	>	if (UNSAFE.compareAndSwapInt(this, runStateOffset, s,
808	>	s & ~SUSPENDED))
809	>	return true;
810	>	}
811	>	return false;
812	>	}
813	>
814	>	/**
815	>	* Sets suspended status and blocks as spare until resumed
816	>	* or shutdown.
817	>	*/
818	>	final void suspendAsSpare() {
819	>	for (;;) {                  // set suspended unless terminating
820	>	int s = runState;
821	>	if ((s & TERMINATING) != 0) { // must kill
822	>	if (UNSAFE.compareAndSwapInt(this, runStateOffset, s,
823	>	s | (TRIMMED | TERMINATING)))
824	>	return;
825	>	}
826	>	else if (UNSAFE.compareAndSwapInt(this, runStateOffset, s,
827	>	s | SUSPENDED))
828	>	break;
829	>	}
830		ForkJoinPool p = pool;
831	<	while (p.hasQueuedSubmissions()) {
832	<	ForkJoinTask<?> t;
833	<	if (tryActivate() && (t = p.pollSubmission()) != null)
834	<	return t;
831	>	p.pushSpare(this);
832	>	while ((runState & SUSPENDED) != 0) {
833	>	if (p.tryAccumulateStealCount(this)) {
834	>	interrupted();          // clear/ignore interrupts
835	>	if ((runState & SUSPENDED) == 0)
836	>	break;
837	>	LockSupport.park(this);
838	>	}
839		}
629	–	return null;
840		}
841
842	<	// Methods accessed only by Pool
842	>	// Misc support methods for ForkJoinPool
843	>
844	>	/**
845	>	* Returns an estimate of the number of tasks in the queue.  Also
846	>	* used by ForkJoinTask.
847	>	*/
848	>	final int getQueueSize() {
849	>	int n; // external calls must read base first
850	>	return (n = -base + sp) <= 0 ? 0 : n;
851	>	}
852
853		/**
854		* Removes and cancels all tasks in queue.  Can be called from any
855		* thread.
856		*/
857		final void cancelTasks() {
858	<	ForkJoinTask<?> t;
859	<	while (base != sp && (t = deqTask()) != null)
860	<	t.cancelIgnoringExceptions();
858	>	ForkJoinTask<?> cj = currentJoin; // try to cancel ongoing tasks
859	>	if (cj != null) {
860	>	currentJoin = null;
861	>	cj.cancelIgnoringExceptions();
862	>	try {
863	>	this.interrupt(); // awaken wait
864	>	} catch (SecurityException ignore) {
865	>	}
866	>	}
867	>	ForkJoinTask<?> cs = currentSteal;
868	>	if (cs != null) {
869	>	currentSteal = null;
870	>	cs.cancelIgnoringExceptions();
871	>	}
872	>	while (base != sp) {
873	>	ForkJoinTask<?> t = deqTask();
874	>	if (t != null)
875	>	t.cancelIgnoringExceptions();
876	>	}
877		}
878
879		/**
#	Line 646 | Line 881 | public class ForkJoinWorkerThread extend
881		*
882		* @return the number of tasks drained
883		*/
884	<	final int drainTasksTo(Collection<ForkJoinTask<?>> c) {
884	>	final int drainTasksTo(Collection<? super ForkJoinTask<?>> c) {
885		int n = 0;
886	<	ForkJoinTask<?> t;
887	<	while (base != sp && (t = deqTask()) != null) {
888	<	c.add(t);
889	<	++n;
886	>	while (base != sp) {
887	>	ForkJoinTask<?> t = deqTask();
888	>	if (t != null) {
889	>	c.add(t);
890	>	++n;
891	>	}
892		}
893		return n;
894		}
895
896	+	// Support methods for ForkJoinTask
897	+
898		/**
899	<	* Gets and clears steal count for accumulation by pool.  Called
900	<	* only when known to be idle (in pool.sync and termination).
899	>	* Gets and removes a local task.
900	>	*
901	>	* @return a task, if available
902		*/
903	<	final int getAndClearStealCount() {
904	<	int sc = stealCount;
905	<	stealCount = 0;
906	<	return sc;
903	>	final ForkJoinTask<?> pollLocalTask() {
904	>	ForkJoinPool p = pool;
905	>	while (sp != base) {
906	>	int a; // inline p.tryIncrementActiveCount
907	>	if (active ||
908	>	(active = UNSAFE.compareAndSwapInt(p, poolRunStateOffset,
909	>	a = p.runState, a + 1)))
910	>	return locallyFifo ? locallyDeqTask() : popTask();
911	>	}
912	>	return null;
913		}
914
915		/**
916	<	* Returns true if at least one worker in the given array appears
671	<	* to have at least one queued task.
916	>	* Gets and removes a local or stolen task.
917		*
918	<	* @param ws array of workers
918	>	* @return a task, if available
919		*/
920	<	static boolean hasQueuedTasks(ForkJoinWorkerThread[] ws) {
921	<	if (ws != null) {
922	<	int len = ws.length;
923	<	for (int j = 0; j < 2; ++j) { // need two passes for clean sweep
924	<	for (int i = 0; i < len; ++i) {
925	<	ForkJoinWorkerThread w = ws[i];
681	<	if (w != null && w.sp != w.base)
682	<	return true;
683	<	}
684	<	}
920	>	final ForkJoinTask<?> pollTask() {
921	>	ForkJoinTask<?> t = pollLocalTask();
922	>	if (t == null) {
923	>	t = scan();
924	>	// cannot retain/track/help steal
925	>	UNSAFE.putOrderedObject(this, currentStealOffset, null);
926		}
927	<	return false;
927	>	return t;
928		}
929
930	<	// Support methods for ForkJoinTask
930	>	/**
931	>	* Possibly runs some tasks and/or blocks, until task is done.
932	>	*
933	>	* @param joinMe the task to join
934	>	* @param timed true if use timed wait
935	>	* @param nanos wait time if timed
936	>	*/
937	>	final void joinTask(ForkJoinTask<?> joinMe, boolean timed, long nanos) {
938	>	// currentJoin only written by this thread; only need ordered store
939	>	ForkJoinTask<?> prevJoin = currentJoin;
940	>	UNSAFE.putOrderedObject(this, currentJoinOffset, joinMe);
941	>	if (isTerminating())                // cancel if shutting down
942	>	joinMe.cancelIgnoringExceptions();
943	>	else {
944	>	if (sp != base)
945	>	localHelpJoinTask(joinMe);
946	>	if (joinMe.status >= 0)
947	>	pool.awaitJoin(joinMe, this, timed, nanos);
948	>	}
949	>	UNSAFE.putOrderedObject(this, currentJoinOffset, prevJoin);
950	>	}
951
952		/**
953	<	* Returns an estimate of the number of tasks in the queue.
953	>	* Run tasks in local queue until given task is done.
954	>	* Not currently used because it complicates semantics.
955	>	*
956	>	* @param joinMe the task to join
957		*/
958	<	final int getQueueSize() {
959	<	// suppress momentarily negative values
960	<	return Math.max(0, sp - base);
958	>	private void localHelpJoinTask(ForkJoinTask<?> joinMe) {
959	>	int s;
960	>	ForkJoinTask<?>[] q;
961	>	while (joinMe.status >= 0 && (s = sp) != base && (q = queue) != null) {
962	>	int i = (q.length - 1) & --s;
963	>	long u = (i << qShift) + qBase; // raw offset
964	>	ForkJoinTask<?> t = q[i];
965	>	if (t == null)  // lost to a stealer
966	>	break;
967	>	if (UNSAFE.compareAndSwapObject(q, u, t, null)) {
968	>	/*
969	>	* This recheck (and similarly in helpJoinTask)
970	>	* handles cases where joinMe is independently
971	>	* cancelled or forced even though there is other work
972	>	* available. Back out of the pop by putting t back
973	>	* into slot before we commit by writing sp.
974	>	*/
975	>	if (joinMe.status < 0) {
976	>	UNSAFE.putObjectVolatile(q, u, t);
977	>	break;
978	>	}
979	>	sp = s;
980	>	// UNSAFE.putOrderedInt(this, spOffset, s);
981	>	t.quietlyExec();
982	>	}
983	>	}
984		}
985
986		/**
987	<	* Returns an estimate of the number of tasks, offset by a
988	<	* function of number of idle workers.
987	>	* Tries to locate and help perform tasks for a stealer of the
988	>	* given task, or in turn one of its stealers.  Traces
989	>	* currentSteal->currentJoin links looking for a thread working on
990	>	* a descendant of the given task and with a non-empty queue to
991	>	* steal back and execute tasks from.
992	>	*
993	>	* The implementation is very branchy to cope with potential
994	>	* inconsistencies or loops encountering chains that are stale,
995	>	* unknown, or of length greater than MAX_HELP_DEPTH links.  All
996	>	* of these cases are dealt with by just returning back to the
997	>	* caller, who is expected to retry if other join mechanisms also
998	>	* don't work out.
999	>	*
1000	>	* @param joinMe the task to join
1001		*/
1002	<	final int getEstimatedSurplusTaskCount() {
1003	<	// The halving approximates weighting idle vs non-idle workers
1004	<	return (sp - base) - (pool.getIdleThreadCount() >>> 1);
1002	>	final void helpJoinTask(ForkJoinTask<?> joinMe) {
1003	>	ForkJoinWorkerThread[] ws;
1004	>	int n;
1005	>	if (joinMe.status < 0)                // already done
1006	>	return;
1007	>	if ((ws = pool.workers) == null || (n = ws.length) <= 1)
1008	>	return;                           // need at least 2 workers
1009	>
1010	>	ForkJoinTask<?> task = joinMe;        // base of chain
1011	>	ForkJoinWorkerThread thread = this;   // thread with stolen task
1012	>	for (int d = 0; d < MAX_HELP_DEPTH; ++d) { // chain length
1013	>	// Try to find v, the stealer of task, by first using hint
1014	>	ForkJoinWorkerThread v = ws[thread.stealHint & (n - 1)];
1015	>	if (v == null || v.currentSteal != task) {
1016	>	for (int j = 0; ; ++j) {      // search array
1017	>	if (j < n) {
1018	>	ForkJoinTask<?> vs;
1019	>	if ((v = ws[j]) != null &&
1020	>	(v != this || base == sp) &&
1021	>	(vs = v.currentSteal) != null) {
1022	>	if (joinMe.status < 0 || task.status < 0)
1023	>	return;       // stale or done
1024	>	if (vs == task) {
1025	>	thread.stealHint = j;
1026	>	break;        // save hint for next time
1027	>	}
1028	>	}
1029	>	}
1030	>	else
1031	>	return;               // no stealer
1032	>	}
1033	>	}
1034	>	for (;;) { // Try to help v, using specialized form of deqTask
1035	>	if (joinMe.status < 0)
1036	>	return;
1037	>	int b = v.base;
1038	>	ForkJoinTask<?>[] q = v.queue;
1039	>	if (b == v.sp || q == null)
1040	>	break;
1041	>	int i = (q.length - 1) & b;
1042	>	long u = (i << qShift) + qBase;
1043	>	ForkJoinTask<?> t = q[i];
1044	>	int pid = poolIndex;
1045	>	ForkJoinTask<?> ps = currentSteal;
1046	>	if (task.status < 0)
1047	>	return;                   // stale or done
1048	>	if (t != null && v.base == b++ &&
1049	>	UNSAFE.compareAndSwapObject(q, u, t, null)) {
1050	>	if (joinMe.status < 0) {
1051	>	UNSAFE.putObjectVolatile(q, u, t);
1052	>	return;               // back out on cancel
1053	>	}
1054	>	v.base = b;
1055	>	v.stealHint = pid;
1056	>	UNSAFE.putOrderedObject(this, currentStealOffset, t);
1057	>	t.quietlyExec();
1058	>	UNSAFE.putOrderedObject(this, currentStealOffset, ps);
1059	>	}
1060	>	}
1061	>	// Try to descend to find v's stealer
1062	>	ForkJoinTask<?> next = v.currentJoin;
1063	>	if (task.status < 0 || next == null || next == task ||
1064	>	joinMe.status < 0)
1065	>	return;
1066	>	task = next;
1067	>	thread = v;
1068	>	}
1069		}
1070
1071		/**
1072	<	* Scans, returning early if joinMe done.
1072	>	* Implements ForkJoinTask.getSurplusQueuedTaskCount().
1073	>	* Returns an estimate of the number of tasks, offset by a
1074	>	* function of number of idle workers.
1075	>	*
1076	>	* This method provides a cheap heuristic guide for task
1077	>	* partitioning when programmers, frameworks, tools, or languages
1078	>	* have little or no idea about task granularity.  In essence by
1079	>	* offering this method, we ask users only about tradeoffs in
1080	>	* overhead vs expected throughput and its variance, rather than
1081	>	* how finely to partition tasks.
1082	>	*
1083	>	* In a steady state strict (tree-structured) computation, each
1084	>	* thread makes available for stealing enough tasks for other
1085	>	* threads to remain active. Inductively, if all threads play by
1086	>	* the same rules, each thread should make available only a
1087	>	* constant number of tasks.
1088	>	*
1089	>	* The minimum useful constant is just 1. But using a value of 1
1090	>	* would require immediate replenishment upon each steal to
1091	>	* maintain enough tasks, which is infeasible.  Further,
1092	>	* partitionings/granularities of offered tasks should minimize
1093	>	* steal rates, which in general means that threads nearer the top
1094	>	* of computation tree should generate more than those nearer the
1095	>	* bottom. In perfect steady state, each thread is at
1096	>	* approximately the same level of computation tree. However,
1097	>	* producing extra tasks amortizes the uncertainty of progress and
1098	>	* diffusion assumptions.
1099	>	*
1100	>	* So, users will want to use values larger, but not much larger
1101	>	* than 1 to both smooth over transient shortages and hedge
1102	>	* against uneven progress; as traded off against the cost of
1103	>	* extra task overhead. We leave the user to pick a threshold
1104	>	* value to compare with the results of this call to guide
1105	>	* decisions, but recommend values such as 3.
1106	>	*
1107	>	* When all threads are active, it is on average OK to estimate
1108	>	* surplus strictly locally. In steady-state, if one thread is
1109	>	* maintaining say 2 surplus tasks, then so are others. So we can
1110	>	* just use estimated queue length (although note that (sp - base)
1111	>	* can be an overestimate because of stealers lagging increments
1112	>	* of base).  However, this strategy alone leads to serious
1113	>	* mis-estimates in some non-steady-state conditions (ramp-up,
1114	>	* ramp-down, other stalls). We can detect many of these by
1115	>	* further considering the number of "idle" threads, that are
1116	>	* known to have zero queued tasks, so compensate by a factor of
1117	>	* (#idle/#active) threads.
1118		*/
1119	<	final ForkJoinTask<?> scanWhileJoining(ForkJoinTask<?> joinMe) {
1120	<	ForkJoinTask<?> t = pollTask();
713	<	if (t != null && joinMe.status < 0 && sp == base) {
714	<	pushTask(t); // unsteal if done and this task would be stealable
715	<	t = null;
716	<	}
717	<	return t;
1119	>	final int getEstimatedSurplusTaskCount() {
1120	>	return sp - base - pool.idlePerActive();
1121		}
1122
1123		/**
1124		* Runs tasks until {@code pool.isQuiescent()}.
1125		*/
1126		final void helpQuiescePool() {
1127	+	ForkJoinTask<?> ps = currentSteal; // to restore below
1128		for (;;) {
1129	<	ForkJoinTask<?> t = pollTask();
1130	<	if (t != null)
1129	>	ForkJoinTask<?> t = pollLocalTask();
1130	>	if (t != null || (t = scan()) != null)
1131		t.quietlyExec();
1132	<	else if (tryInactivate() && pool.isQuiescent())
1133	<	break;
1132	>	else {
1133	>	ForkJoinPool p = pool;
1134	>	int a; // to inline CASes
1135	>	if (active) {
1136	>	if (!UNSAFE.compareAndSwapInt
1137	>	(p, poolRunStateOffset, a = p.runState, a - 1))
1138	>	continue;   // retry later
1139	>	active = false; // inactivate
1140	>	UNSAFE.putOrderedObject(this, currentStealOffset, ps);
1141	>	}
1142	>	if (p.isQuiescent()) {
1143	>	active = true; // re-activate
1144	>	do {} while (!UNSAFE.compareAndSwapInt
1145	>	(p, poolRunStateOffset, a = p.runState, a+1));
1146	>	return;
1147	>	}
1148	>	}
1149	>	}
1150	>	}
1151	>
1152	>	// Unsafe mechanics
1153	>
1154	>	private static final sun.misc.Unsafe UNSAFE = getUnsafe();
1155	>	private static final long spOffset =
1156	>	objectFieldOffset("sp", ForkJoinWorkerThread.class);
1157	>	private static final long runStateOffset =
1158	>	objectFieldOffset("runState", ForkJoinWorkerThread.class);
1159	>	private static final long currentJoinOffset =
1160	>	objectFieldOffset("currentJoin", ForkJoinWorkerThread.class);
1161	>	private static final long currentStealOffset =
1162	>	objectFieldOffset("currentSteal", ForkJoinWorkerThread.class);
1163	>	private static final long qBase =
1164	>	UNSAFE.arrayBaseOffset(ForkJoinTask[].class);
1165	>	private static final long poolRunStateOffset = // to inline CAS
1166	>	objectFieldOffset("runState", ForkJoinPool.class);
1167	>
1168	>	private static final int qShift;
1169	>
1170	>	static {
1171	>	int s = UNSAFE.arrayIndexScale(ForkJoinTask[].class);
1172	>	if ((s & (s-1)) != 0)
1173	>	throw new Error("data type scale not a power of two");
1174	>	qShift = 31 - Integer.numberOfLeadingZeros(s);
1175	>	MAXIMUM_QUEUE_CAPACITY = 1 << (31 - qShift);
1176	>	}
1177	>
1178	>	private static long objectFieldOffset(String field, Class<?> klazz) {
1179	>	try {
1180	>	return UNSAFE.objectFieldOffset(klazz.getDeclaredField(field));
1181	>	} catch (NoSuchFieldException e) {
1182	>	// Convert Exception to corresponding Error
1183	>	NoSuchFieldError error = new NoSuchFieldError(field);
1184	>	error.initCause(e);
1185	>	throw error;
1186		}
731	–	do {} while (!tryActivate()); // re-activate on exit
1187		}
1188
1189	<	// Unsafe mechanics for jsr166y 3rd party package.
1189	>	/**
1190	>	* Returns a sun.misc.Unsafe.  Suitable for use in a 3rd party package.
1191	>	* Replace with a simple call to Unsafe.getUnsafe when integrating
1192	>	* into a jdk.
1193	>	*
1194	>	* @return a sun.misc.Unsafe
1195	>	*/
1196		private static sun.misc.Unsafe getUnsafe() {
1197		try {
1198		return sun.misc.Unsafe.getUnsafe();
1199		} catch (SecurityException se) {
1200		try {
1201		return java.security.AccessController.doPrivileged
1202	<	(new java.security.PrivilegedExceptionAction<sun.misc.Unsafe>() {
1202	>	(new java.security
1203	>	.PrivilegedExceptionAction<sun.misc.Unsafe>() {
1204		public sun.misc.Unsafe run() throws Exception {
1205	<	return getUnsafeByReflection();
1205	>	java.lang.reflect.Field f = sun.misc
1206	>	.Unsafe.class.getDeclaredField("theUnsafe");
1207	>	f.setAccessible(true);
1208	>	return (sun.misc.Unsafe) f.get(null);
1209		}});
1210		} catch (java.security.PrivilegedActionException e) {
1211		throw new RuntimeException("Could not initialize intrinsics",
#	Line 748 | Line 1213 | public class ForkJoinWorkerThread extend
1213		}
1214		}
1215		}
751	–
752	–	private static sun.misc.Unsafe getUnsafeByReflection()
753	–	throws NoSuchFieldException, IllegalAccessException {
754	–	java.lang.reflect.Field f =
755	–	sun.misc.Unsafe.class.getDeclaredField("theUnsafe");
756	–	f.setAccessible(true);
757	–	return (sun.misc.Unsafe) f.get(null);
758	–	}
759	–
760	–	private static long fieldOffset(String fieldName, Class<?> klazz) {
761	–	try {
762	–	return UNSAFE.objectFieldOffset(klazz.getDeclaredField(fieldName));
763	–	} catch (NoSuchFieldException e) {
764	–	// Convert Exception to Error
765	–	NoSuchFieldError error = new NoSuchFieldError(fieldName);
766	–	error.initCause(e);
767	–	throw error;
768	–	}
769	–	}
770	–
771	–	private static final sun.misc.Unsafe UNSAFE = getUnsafe();
772	–	static final long baseOffset =
773	–	fieldOffset("base", ForkJoinWorkerThread.class);
774	–	static final long spOffset =
775	–	fieldOffset("sp", ForkJoinWorkerThread.class);
776	–	static final long runStateOffset =
777	–	fieldOffset("runState", ForkJoinWorkerThread.class);
778	–	static final long qBase;
779	–	static final int qShift;
780	–
781	–	static {
782	–	qBase = UNSAFE.arrayBaseOffset(ForkJoinTask[].class);
783	–	int s = UNSAFE.arrayIndexScale(ForkJoinTask[].class);
784	–	if ((s & (s-1)) != 0)
785	–	throw new Error("data type scale not a power of two");
786	–	qShift = 31 - Integer.numberOfLeadingZeros(s);
787	–	}
1216		}

Diff Legend

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