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 1.4 by dl, Wed Jan 7 20:51:36 2009 UTC vs.
Revision 1.39 by dl, Sat Jul 24 20:28:18 2010 UTC

#	Line 5 | Line 5
5		*/
6
7		package jsr166y;
8	<	import java.util.*;
8	>
9		import java.util.concurrent.*;
10	<	import java.util.concurrent.atomic.*;
11	<	import java.util.concurrent.locks.*;
12	<	import sun.misc.Unsafe;
13	<	import java.lang.reflect.*;
10	>
11	>	import java.util.Random;
12	>	import java.util.Collection;
13	>	import java.util.concurrent.locks.LockSupport;
14
15		/**
16		* A thread managed by a {@link ForkJoinPool}.  This class is
17		* subclassable solely for the sake of adding functionality -- there
18	<	* are no overridable methods dealing with scheduling or
19	<	* execution. However, you can override initialization and termination
20	<	* cleanup methods surrounding the main task processing loop.  If you
21	<	* do create such a subclass, you will also need to supply a custom
22	<	* ForkJoinWorkerThreadFactory to use it in a ForkJoinPool.
23	<	*
24	<	* <p>This class also provides methods for generating per-thread
25	<	* random numbers, with the same properties as {@link
26	<	* java.util.Random} but with each generator isolated from those of
27	<	* other threads.
18	>	* are no overridable methods dealing with scheduling or execution.
19	>	* However, you can override initialization and termination methods
20	>	* surrounding the main task processing loop.  If you do create such a
21	>	* subclass, you will also need to supply a custom {@link
22	>	* ForkJoinPool.ForkJoinWorkerThreadFactory} to use it in a {@code
23	>	* ForkJoinPool}.
24	>	*
25	>	* @since 1.7
26	>	* @author Doug Lea
27		*/
28		public class ForkJoinWorkerThread extends Thread {
29		/*
30	<	* Algorithm overview:
30	>	* Overview:
31	>	*
32	>	* 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	<	* 1. Work-Stealing: Work-stealing queues are special forms of
41	<	* Deques that support only three of the four possible
42	<	* end-operations -- push, pop, and deq (aka steal), and only do
43	<	* so under the constraints that push and pop are called only from
44	<	* the owning thread, while deq may be called from other threads.
45	<	* (If you are unfamiliar with them, you probably want to read
46	<	* Herlihy and Shavit's book "The Art of Multiprocessor
47	<	* programming", chapter 16 describing these in more detail before
48	<	* proceeding.)  The main work-stealing queue design is roughly
49	<	* similar to "Dynamic Circular Work-Stealing Deque" by David
50	<	* Chase and Yossi Lev, SPAA 2005
51	<	* (http://research.sun.com/scalable/pubs/index.html).  The main
52	<	* difference ultimately stems from gc requirements that we null
53	<	* out taken slots as soon as we can, to maintain as small a
54	<	* footprint as possible even in programs generating huge numbers
55	<	* of tasks. To accomplish this, we shift the CAS arbitrating pop
56	<	* vs deq (steal) from being on the indices ("base" and "sp") to
57	<	* the slots themselves (mainly via method "casSlotNull()"). So,
58	<	* both a successful pop and deq mainly entail CAS'ing a nonnull
59	<	* slot to null.  Because we rely on CASes of references, we do
60	<	* not need tag bits on base or sp.  They are simple ints as used
61	<	* in any circular array-based queue (see for example ArrayDeque).
62	<	* Updates to the indices must still be ordered in a way that
63	<	* guarantees that (sp - base) > 0 means the queue is empty, but
64	<	* otherwise may err on the side of possibly making the queue
65	<	* appear nonempty when a push, pop, or deq have not fully
66	<	* committed. Note that this means that the deq operation,
67	<	* considered individually, is not wait-free. One thief cannot
68	<	* successfully continue until another in-progress one (or, if
69	<	* previously empty, a push) completes.  However, in the
70	<	* aggregate, we ensure at least probablistic non-blockingness. If
71	<	* an attempted steal fails, a thief always chooses a different
40	>	* Work-stealing queues are special forms of Deques that support
41	>	* only three of the four possible end-operations -- push, pop,
42	>	* and deq (aka steal), under the further constraints that push
43	>	* and pop are called only from the owning thread, while deq may
44	>	* be called from other threads.  (If you are unfamiliar with
45	>	* them, you probably want to read Herlihy and Shavit's book "The
46	>	* Art of Multiprocessor programming", chapter 16 describing these
47	>	* in more detail before proceeding.)  The main work-stealing
48	>	* queue design is roughly similar to those in the papers "Dynamic
49	>	* Circular Work-Stealing Deque" by Chase and Lev, SPAA 2005
50	>	* (http://research.sun.com/scalable/pubs/index.html) and
51	>	* "Idempotent work stealing" by Michael, Saraswat, and Vechev,
52	>	* PPoPP 2009 (http://portal.acm.org/citation.cfm?id=1504186).
53	>	* The main differences ultimately stem from gc requirements that
54	>	* we null out taken slots as soon as we can, to maintain as small
55	>	* a footprint as possible even in programs generating huge
56	>	* numbers of tasks. To accomplish this, we shift the CAS
57	>	* arbitrating pop vs deq (steal) from being on the indices
58	>	* ("base" and "sp") to the slots themselves (mainly via method
59	>	* "casSlotNull()"). So, both a successful pop and deq mainly
60	>	* entail a CAS of a slot from non-null to null.  Because we rely
61	>	* on CASes of references, we do not need tag bits on base or sp.
62	>	* They are simple ints as used in any circular array-based queue
63	>	* (see for example ArrayDeque).  Updates to the indices must
64	>	* still be ordered in a way that guarantees that sp == base means
65	>	* the queue is empty, but otherwise may err on the side of
66	>	* possibly making the queue appear nonempty when a push, pop, or
67	>	* deq have not fully committed. Note that this means that the deq
68	>	* operation, considered individually, is not wait-free. One thief
69	>	* cannot successfully continue until another in-progress one (or,
70	>	* if previously empty, a push) completes.  However, in the
71	>	* aggregate, we ensure at least probabilistic non-blockingness.
72	>	* If an attempted steal fails, a thief always chooses a different
73		* random victim target to try next. So, in order for one thief to
74		* progress, it suffices for any in-progress deq or new push on
75		* any empty queue to complete. One reason this works well here is
76		* that apparently-nonempty often means soon-to-be-stealable,
77	<	* which gives threads a chance to activate if necessary before
78	<	* stealing (see below).
77	>	* which gives threads a chance to set activation status if
78	>	* necessary before stealing.
79	>	*
80	>	* This approach also enables support for "async mode" where local
81	>	* task processing is in FIFO, not LIFO order; simply by using a
82	>	* version of deq rather than pop when locallyFifo is true (as set
83	>	* by the ForkJoinPool).  This allows use in message-passing
84	>	* frameworks in which tasks are never joined.
85	>	*
86	>	* When a worker would otherwise be blocked waiting to join a
87	>	* task, it first tries a form of linear helping: Each worker
88	>	* records (in field currentSteal) the most recent task it stole
89	>	* from some other worker. Plus, it records (in field currentJoin)
90	>	* the task it is currently actively joining. Method joinTask uses
91	>	* these markers to try to find a worker to help (i.e., steal back
92	>	* a task from and execute it) that could hasten completion of the
93	>	* actively joined task. In essence, the joiner executes a task
94	>	* that would be on its own local deque had the to-be-joined task
95	>	* not been stolen. This may be seen as a conservative variant of
96	>	* the approach in Wagner & Calder "Leapfrogging: a portable
97	>	* technique for implementing efficient futures" SIGPLAN Notices,
98	>	* 1993 (http://portal.acm.org/citation.cfm?id=155354). It differs
99	>	* in that: (1) We only maintain dependency links across workers
100	>	* upon steals, rather than use per-task bookkeeping.  This may
101	>	* require a linear scan of workers array to locate stealers, but
102	>	* usually doesn't because stealers leave hints (that may become
103	>	* stale/wrong) of where to locate them. This isolates cost to
104	>	* when it is needed, rather than adding to per-task overhead.
105	>	* (2) It is "shallow", ignoring nesting and potentially cyclic
106	>	* mutual steals.  (3) It is intentionally racy: field currentJoin
107	>	* is updated only while actively joining, which means that we
108	>	* miss links in the chain during long-lived tasks, GC stalls etc
109	>	* (which is OK since blocking in such cases is usually a good
110	>	* idea).  (4) We bound the number of attempts to find work (see
111	>	* MAX_HELP_DEPTH) and fall back to suspending the worker and if
112	>	* necessary replacing it with a spare (see
113	>	* ForkJoinPool.tryAwaitJoin).
114		*
115	<	* Efficient implementation of this approach currently relies on
116	<	* an uncomfortable amount of "Unsafe" mechanics. To maintain
115	>	* Efficient implementation of these algorithms currently relies
116	>	* on an uncomfortable amount of "Unsafe" mechanics. To maintain
117		* correct orderings, reads and writes of variable base require
118	<	* volatile ordering.  Variable sp does not require volatile write
119	<	* but needs cheaper store-ordering on writes.  Because they are
120	<	* protected by volatile base reads, reads of the queue array and
121	<	* its slots do not need volatile load semantics, but writes (in
122	<	* push) require store order and CASes (in pop and deq) require
123	<	* (volatile) CAS semantics. Since these combinations aren't
124	<	* supported using ordinary volatiles, the only way to accomplish
125	<	* these effciently is to use direct Unsafe calls. (Using external
126	<	* AtomicIntegers and AtomicReferenceArrays for the indices and
127	<	* array is significantly slower because of memory locality and
128	<	* indirection effects.) Further, performance on most platforms is
129	<	* very sensitive to placement and sizing of the (resizable) queue
130	<	* array.  Even though these queues don't usually become all that
131	<	* big, the initial size must be large enough to counteract cache
118	>	* volatile ordering.  Variable sp does not require volatile
119	>	* writes but still needs store-ordering, which we accomplish by
120	>	* pre-incrementing sp before filling the slot with an ordered
121	>	* store.  (Pre-incrementing also enables backouts used in
122	>	* joinTask.)  Because they are protected by volatile base reads,
123	>	* reads of the queue array and its slots by other threads do not
124	>	* need volatile load semantics, but writes (in push) require
125	>	* store order and CASes (in pop and deq) require (volatile) CAS
126	>	* semantics.  (Michael, Saraswat, and Vechev's algorithm has
127	>	* similar properties, but without support for nulling slots.)
128	>	* Since these combinations aren't supported using ordinary
129	>	* volatiles, the only way to accomplish these efficiently is to
130	>	* use direct Unsafe calls. (Using external AtomicIntegers and
131	>	* AtomicReferenceArrays for the indices and array is
132	>	* significantly slower because of memory locality and indirection
133	>	* effects.)
134	>	*
135	>	* Further, performance on most platforms is very sensitive to
136	>	* placement and sizing of the (resizable) queue array.  Even
137	>	* though these queues don't usually become all that big, the
138	>	* initial size must be large enough to counteract cache
139		* contention effects across multiple queues (especially in the
140		* presence of GC cardmarking). Also, to improve thread-locality,
141	<	* queues are currently initialized immediately after the thread
142	<	* gets the initial signal to start processing tasks.  However,
143	<	* all queue-related methods except pushTask are written in a way
144	<	* that allows them to instead be lazily allocated and/or disposed
145	<	* of when empty. All together, these low-level implementation
146	<	* choices produce as much as a factor of 4 performance
147	<	* improvement compared to naive implementations, and enable the
148	<	* processing of billions of tasks per second, sometimes at the
149	<	* expense of ugliness.
150	<	*
151	<	* 2. Run control: The primary run control is based on a global
152	<	* counter (activeCount) held by the pool. It uses an algorithm
153	<	* similar to that in Herlihy and Shavit section 17.6 to cause
154	<	* threads to eventually block when all threads declare they are
155	<	* inactive. (See variable "scans".)  For this to work, threads
156	<	* must be declared active when executing tasks, and before
157	<	* stealing a task. They must be inactive before blocking on the
158	<	* Pool Barrier (awaiting a new submission or other Pool
159	<	* event). In between, there is some free play which we take
160	<	* advantage of to avoid contention and rapid flickering of the
161	<	* global activeCount: If inactive, we activate only if a victim
162	<	* queue appears to be nonempty (see above).  Similarly, a thread
163	<	* tries to inactivate only after a full scan of other threads.
164	<	* The net effect is that contention on activeCount is rarely a
165	<	* measurable performance issue. (There are also a few other cases
166	<	* where we scan for work rather than retry/block upon
167	<	* contention.)
168	<	*
169	<	* 3. Selection control. We maintain policy of always choosing to
170	<	* run local tasks rather than stealing, and always trying to
171	<	* steal tasks before trying to run a new submission. All steals
172	<	* are currently performed in randomly-chosen deq-order. It may be
173	<	* worthwhile to bias these with locality / anti-locality
124	<	* information, but doing this well probably requires more
125	<	* 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 timeout value for suspending spares. Spare workers that
159	>	* remain unsignalled for more than this time may be trimmed
160	>	* (killed and removed from pool).  Since our goal is to avoid
161	>	* long-term thread buildup, the exact value of timeout does not
162	>	* matter too much so long as it avoids most false-alarm timeouts
163	>	* under GC stalls or momentarily high system load.
164	>	*/
165	>	private static final long SPARE_KEEPALIVE_NANOS =
166	>	5L * 1000L * 1000L * 1000L; // 5 secs
167	>
168	>	/**
169	>	* The maximum stolen->joining link depth allowed in helpJoinTask.
170	>	* Depths for legitimate chains are unbounded, but we use a fixed
171	>	* constant to avoid (otherwise unchecked) cycles and bound
172	>	* staleness of traversal parameters at the expense of sometimes
173	>	* blocking when we could be helping.
174		*/
175	+	private static final int MAX_HELP_DEPTH = 8;
176
177		/**
178		* Capacity of work-stealing queue array upon initialization.
179	<	* Must be a power of two. Initial size must be at least 2, but is
179	>	* Must be a power of two. Initial size must be at least 4, but is
180		* padded to minimize cache effects.
181		*/
182		private static final int INITIAL_QUEUE_CAPACITY = 1 << 13;
183
184		/**
185		* Maximum work-stealing queue array size.  Must be less than or
186	<	* equal to 1 << 30 to ensure lack of index wraparound.
186	>	* equal to 1 << 28 to ensure lack of index wraparound. (This
187	>	* is less than usual bounds, because we need leftshift by 3
188	>	* to be in int range).
189		*/
190	<	private static final int MAXIMUM_QUEUE_CAPACITY = 1 << 30;
190	>	private static final int MAXIMUM_QUEUE_CAPACITY = 1 << 28;
191
192		/**
193	<	* Generator of seeds for per-thread random numbers.
193	>	* The pool this thread works in. Accessed directly by ForkJoinTask.
194		*/
195	<	private static final Random randomSeedGenerator = new Random();
195	>	final ForkJoinPool pool;
196
197		/**
198		* The work-stealing queue array. Size must be a power of two.
199	+	* Initialized in onStart, to improve memory locality.
200		*/
201		private ForkJoinTask<?>[] queue;
202
203		/**
152	–	* Index (mod queue.length) of next queue slot to push to or pop
153	–	* from. It is written only by owner thread, via ordered store.
154	–	* Both sp and base are allowed to wrap around on overflow, but
155	–	* (sp - base) still estimates size.
156	–	*/
157	–	private volatile int sp;
158	–
159	–	/**
204		* Index (mod queue.length) of least valid queue slot, which is
205		* always the next position to steal from if nonempty.
206		*/
207		private volatile int base;
208
209		/**
210	<	* The pool this thread works in.
210	>	* Index (mod queue.length) of next queue slot to push to or pop
211	>	* from. It is written only by owner thread, and accessed by other
212	>	* threads only after reading (volatile) base.  Both sp and base
213	>	* are allowed to wrap around on overflow, but (sp - base) still
214	>	* estimates size.
215		*/
216	<	final ForkJoinPool pool;
216	>	private int sp;
217
218		/**
219	<	* Index of this worker in pool array. Set once by pool before
220	<	* running, and accessed directly by pool during cleanup etc
219	>	* The index of most recent stealer, used as a hint to avoid
220	>	* traversal in method helpJoinTask. This is only a hint because a
221	>	* worker might have had multiple steals and this only holds one
222	>	* of them (usually the most current). Declared non-volatile,
223	>	* relying on other prevailing sync to keep reasonably current.
224		*/
225	<	int poolIndex;
225	>	private int stealHint;
226
227		/**
228	<	* Run state of this worker. Supports simple versions of the usual
229	<	* shutdown/shutdownNow control.
228	>	* Run state of this worker. In addition to the usual run levels,
229	>	* tracks if this worker is suspended as a spare, and if it was
230	>	* killed (trimmed) while suspended. However, "active" status is
231	>	* maintained separately.
232		*/
233		private volatile int runState;
234
235	<	// Runstate values. Order matters
236	<	private static final int RUNNING     = 0;
237	<	private static final int SHUTDOWN    = 1;
238	<	private static final int TERMINATING = 2;
239	<	private static final int TERMINATED  = 3;
235	>	private static final int TERMINATING = 0x01;
236	>	private static final int TERMINATED  = 0x02;
237	>	private static final int SUSPENDED   = 0x04; // inactive spare
238	>	private static final int TRIMMED     = 0x08; // killed while suspended
239	>
240	>	/**
241	>	* Number of LockSupport.park calls to block this thread for
242	>	* suspension or event waits. Used for internal instrumention;
243	>	* currently not exported but included because volatile write upon
244	>	* park also provides a workaround for a JVM bug.
245	>	*/
246	>	volatile int parkCount;
247	>
248	>	/**
249	>	* Number of steals, transferred and reset in pool callbacks pool
250	>	* when idle Accessed directly by pool.
251	>	*/
252	>	int stealCount;
253	>
254	>	/**
255	>	* Seed for random number generator for choosing steal victims.
256	>	* Uses Marsaglia xorshift. Must be initialized as nonzero.
257	>	*/
258	>	private int seed;
259
260		/**
261		* Activity status. When true, this worker is considered active.
262	<	* Must be false upon construction. It must be true when executing
191	<	* tasks, and BEFORE stealing a task. It must be false before
192	<	* blocking on the Pool Barrier.
262	>	* Accessed directly by pool.  Must be false upon construction.
263		*/
264	<	private boolean active;
264	>	boolean active;
265
266		/**
267	<	* Number of steals, transferred to pool when idle
267	>	* True if use local fifo, not default lifo, for local polling.
268	>	* Shadows value from ForkJoinPool.
269		*/
270	<	private int stealCount;
270	>	private final boolean locallyFifo;
271
272		/**
273	<	* Seed for random number generator for choosing steal victims
273	>	* Index of this worker in pool array. Set once by pool before
274	>	* running, and accessed directly by pool to locate this worker in
275	>	* its workers array.
276		*/
277	<	private int randomVictimSeed;
277	>	int poolIndex;
278
279		/**
280	<	* Seed for embedded Jurandom
280	>	* The last pool event waited for. Accessed only by pool in
281	>	* callback methods invoked within this thread.
282		*/
283	<	private long juRandomSeed;
283	>	int lastEventCount;
284
285		/**
286	<	* The last barrier event waited for
286	>	* Encoded index and event count of next event waiter. Used only
287	>	* by ForkJoinPool for managing event waiters.
288		*/
289	<	private long eventCount;
289	>	volatile long nextWaiter;
290	>
291	>	/**
292	>	* The task currently being joined, set only when actively trying
293	>	* to helpStealer. Written only by current thread, but read by
294	>	* others.
295	>	*/
296	>	private volatile ForkJoinTask<?> currentJoin;
297	>
298	>	/**
299	>	* The task most recently stolen from another worker (or
300	>	* submission queue).  Not volatile because always read/written in
301	>	* presence of related volatiles in those cases where it matters.
302	>	*/
303	>	private ForkJoinTask<?> currentSteal;
304
305		/**
306		* Creates a ForkJoinWorkerThread operating in the given pool.
307	+	*
308		* @param pool the pool this thread works in
309		* @throws NullPointerException if pool is null
310		*/
311		protected ForkJoinWorkerThread(ForkJoinPool pool) {
222	–	if (pool == null) throw new NullPointerException();
312		this.pool = pool;
313	<	// remaining initialization deferred to onStart
313	>	this.locallyFifo = pool.locallyFifo;
314	>	// To avoid exposing construction details to subclasses,
315	>	// remaining initialization is in start() and onStart()
316	>	}
317	>
318	>	/**
319	>	* Performs additional initialization and starts this thread
320	>	*/
321	>	final void start(int poolIndex, UncaughtExceptionHandler ueh) {
322	>	this.poolIndex = poolIndex;
323	>	setDaemon(true);
324	>	if (ueh != null)
325	>	setUncaughtExceptionHandler(ueh);
326	>	start();
327		}
328
329	<	// public access methods
329	>	// Public/protected methods
330
331		/**
332	<	* Returns the pool hosting this thread
332	>	* Returns the pool hosting this thread.
333	>	*
334		* @return the pool
335		*/
336		public ForkJoinPool getPool() {
#	Line 239 | Line 342 | public class ForkJoinWorkerThread extend
342		* returned value ranges from zero to the maximum number of
343		* threads (minus one) that have ever been created in the pool.
344		* This method may be useful for applications that track status or
345	<	* collect results on a per-worker basis.
346	<	* @return the index number.
345	>	* collect results per-worker rather than per-task.
346	>	*
347	>	* @return the index number
348		*/
349		public int getPoolIndex() {
350		return poolIndex;
351		}
352
353	<	//  Access methods used by Pool
353	>	/**
354	>	* Initializes internal state after construction but before
355	>	* processing any tasks. If you override this method, you must
356	>	* invoke super.onStart() at the beginning of the method.
357	>	* Initialization requires care: Most fields must have legal
358	>	* default values, to ensure that attempted accesses from other
359	>	* threads work correctly even before this thread starts
360	>	* processing tasks.
361	>	*/
362	>	protected void onStart() {
363	>	int rs = seedGenerator.nextInt();
364	>	seed = rs == 0? 1 : rs; // seed must be nonzero
365	>
366	>	// Allocate name string and arrays in this thread
367	>	String pid = Integer.toString(pool.getPoolNumber());
368	>	String wid = Integer.toString(poolIndex);
369	>	setName("ForkJoinPool-" + pid + "-worker-" + wid);
370	>
371	>	queue = new ForkJoinTask<?>[INITIAL_QUEUE_CAPACITY];
372	>	}
373	>
374	>	/**
375	>	* Performs cleanup associated with termination of this worker
376	>	* thread.  If you override this method, you must invoke
377	>	* {@code super.onTermination} at the end of the overridden method.
378	>	*
379	>	* @param exception the exception causing this thread to abort due
380	>	* to an unrecoverable error, or {@code null} if completed normally
381	>	*/
382	>	protected void onTermination(Throwable exception) {
383	>	try {
384	>	cancelTasks();
385	>	setTerminated();
386	>	pool.workerTerminated(this);
387	>	} catch (Throwable ex) {        // Shouldn't ever happen
388	>	if (exception == null)      // but if so, at least rethrown
389	>	exception = ex;
390	>	} finally {
391	>	if (exception != null)
392	>	UNSAFE.throwException(exception);
393	>	}
394	>	}
395
396		/**
397	<	* Get and clear steal count for accumulation by pool.  Called
398	<	* only when known to be idle (in pool.sync and termination).
397	>	* This method is required to be public, but should never be
398	>	* called explicitly. It performs the main run loop to execute
399	>	* ForkJoinTasks.
400		*/
401	<	final int getAndClearStealCount() {
402	<	int sc = stealCount;
403	<	stealCount = 0;
404	<	return sc;
401	>	public void run() {
402	>	Throwable exception = null;
403	>	try {
404	>	onStart();
405	>	mainLoop();
406	>	} catch (Throwable ex) {
407	>	exception = ex;
408	>	} finally {
409	>	onTermination(exception);
410	>	}
411		}
412
413	+	// helpers for run()
414	+
415		/**
416	<	* Returns estimate of the number of tasks in the queue, without
263	<	* correcting for transient negative values
416	>	* Find and execute tasks and check status while running
417		*/
418	<	final int getRawQueueSize() {
419	<	return sp - base;
418	>	private void mainLoop() {
419	>	int emptyScans = 0; // consecutive times failed to find work
420	>	ForkJoinPool p = pool;
421	>	for (;;) {
422	>	p.preStep(this, emptyScans);
423	>	if (runState != 0)
424	>	return;
425	>	ForkJoinTask<?> t; // try to get and run stolen or submitted task
426	>	if ((t = scan()) != null || (t = pollSubmission()) != null) {
427	>	t.tryExec();
428	>	if (base != sp)
429	>	runLocalTasks();
430	>	currentSteal = null;
431	>	emptyScans = 0;
432	>	}
433	>	else
434	>	++emptyScans;
435	>	}
436		}
437
438	<	// Intrinsics-based support for queue operations.
439	<	// Currently these three (setSp, setSlot, casSlotNull) are
440	<	// usually manually inlined to improve performance
438	>	/**
439	>	* Runs local tasks until queue is empty or shut down.  Call only
440	>	* while active.
441	>	*/
442	>	private void runLocalTasks() {
443	>	while (runState == 0) {
444	>	ForkJoinTask<?> t = locallyFifo? locallyDeqTask() : popTask();
445	>	if (t != null)
446	>	t.tryExec();
447	>	else if (base == sp)
448	>	break;
449	>	}
450	>	}
451
452		/**
453	<	* Sets sp in store-order.
453	>	* If a submission exists, try to activate and take it
454	>	*
455	>	* @return a task, if available
456		*/
457	<	private void setSp(int s) {
458	<	_unsafe.putOrderedInt(this, spOffset, s);
457	>	private ForkJoinTask<?> pollSubmission() {
458	>	ForkJoinPool p = pool;
459	>	while (p.hasQueuedSubmissions()) {
460	>	if (active || (active = p.tryIncrementActiveCount())) {
461	>	ForkJoinTask<?> t = p.pollSubmission();
462	>	if (t != null) {
463	>	currentSteal = t;
464	>	return t;
465	>	}
466	>	return scan(); // if missed, rescan
467	>	}
468	>	}
469	>	return null;
470		}
471
472	+	/*
473	+	* Intrinsics-based atomic writes for queue slots. These are
474	+	* basically the same as methods in AtomicObjectArray, but
475	+	* specialized for (1) ForkJoinTask elements (2) requirement that
476	+	* nullness and bounds checks have already been performed by
477	+	* callers and (3) effective offsets are known not to overflow
478	+	* from int to long (because of MAXIMUM_QUEUE_CAPACITY). We don't
479	+	* need corresponding version for reads: plain array reads are OK
480	+	* because they protected by other volatile reads and are
481	+	* confirmed by CASes.
482	+	*
483	+	* Most uses don't actually call these methods, but instead contain
484	+	* inlined forms that enable more predictable optimization.  We
485	+	* don't define the version of write used in pushTask at all, but
486	+	* instead inline there a store-fenced array slot write.
487	+	*/
488	+
489		/**
490	<	* Add in store-order the given task at given slot of q to
491	<	* null. Caller must ensure q is nonnull and index is in range.
490	>	* CASes slot i of array q from t to null. Caller must ensure q is
491	>	* non-null and index is in range.
492		*/
493	<	private static void setSlot(ForkJoinTask<?>[] q, int i,
494	<	ForkJoinTask<?> t){
495	<	_unsafe.putOrderedObject(q, (i << qShift) + qBase, t);
493	>	private static final boolean casSlotNull(ForkJoinTask<?>[] q, int i,
494	>	ForkJoinTask<?> t) {
495	>	return UNSAFE.compareAndSwapObject(q, (i << qShift) + qBase, t, null);
496		}
497
498		/**
499	<	* CAS given slot of q to null. Caller must ensure q is nonnull
500	<	* and index is in range.
499	>	* Performs a volatile write of the given task at given slot of
500	>	* array q.  Caller must ensure q is non-null and index is in
501	>	* range. This method is used only during resets and backouts.
502		*/
503	<	private static boolean casSlotNull(ForkJoinTask<?>[] q, int i,
504	<	ForkJoinTask<?> t) {
505	<	return _unsafe.compareAndSwapObject(q, (i << qShift) + qBase, t, null);
503	>	private static final void writeSlot(ForkJoinTask<?>[] q, int i,
504	>	ForkJoinTask<?> t) {
505	>	UNSAFE.putObjectVolatile(q, (i << qShift) + qBase, t);
506		}
507
508	<	// Main queue methods
508	>	// queue methods
509
510		/**
511	<	* Pushes a task. Called only by current thread.
512	<	* @param t the task. Caller must ensure nonnull
511	>	* Pushes a task. Call only from this thread.
512	>	*
513	>	* @param t the task. Caller must ensure non-null.
514		*/
515		final void pushTask(ForkJoinTask<?> t) {
516		ForkJoinTask<?>[] q = queue;
517	<	int mask = q.length - 1;
518	<	int s = sp;
519	<	_unsafe.putOrderedObject(q, ((s & mask) << qShift) + qBase, t);
520	<	_unsafe.putOrderedInt(this, spOffset, ++s);
521	<	if ((s -= base) == 1)
522	<	pool.signalNonEmptyWorkerQueue();
523	<	else if (s >= mask)
313	<	growQueue();
517	>	int mask = q.length - 1; // implicit assert q != null
518	>	int s = sp++;            // ok to increment sp before slot write
519	>	UNSAFE.putOrderedObject(q, ((s & mask) << qShift) + qBase, t);
520	>	if ((s -= base) == 0)
521	>	pool.signalWork();   // was empty
522	>	else if (s == mask)
523	>	growQueue();         // is full
524		}
525
526		/**
527		* Tries to take a task from the base of the queue, failing if
528	<	* either empty or contended.
529	<	* @return a task, or null if none or contended.
528	>	* empty or contended. Note: Specializations of this code appear
529	>	* in locallyDeqTask and elsewhere.
530	>	*
531	>	* @return a task, or null if none or contended
532		*/
533	<	private ForkJoinTask<?> deqTask() {
322	<	ForkJoinTask<?>[] q;
533	>	final ForkJoinTask<?> deqTask() {
534		ForkJoinTask<?> t;
535	<	int i;
536	<	int b;
537	<	if (sp != (b = base) &&
535	>	ForkJoinTask<?>[] q;
536	>	int b, i;
537	>	if ((b = base) != sp &&
538		(q = queue) != null && // must read q after b
539	<	(t = q[i = (q.length - 1) & b]) != null &&
540	<	_unsafe.compareAndSwapObject(q, (i << qShift) + qBase, t, null)) {
539	>	(t = q[i = (q.length - 1) & b]) != null && base == b &&
540	>	UNSAFE.compareAndSwapObject(q, (i << qShift) + qBase, t, null)) {
541		base = b + 1;
542		return t;
543		}
#	Line 334 | Line 545 | public class ForkJoinWorkerThread extend
545		}
546
547		/**
548	<	* Returns a popped task, or null if empty.  Called only by
549	<	* current thread.
548	>	* Tries to take a task from the base of own queue. Assumes active
549	>	* status.  Called only by current thread.
550	>	*
551	>	* @return a task, or null if none
552		*/
553	<	final ForkJoinTask<?> popTask() {
341	<	ForkJoinTask<?> t;
342	<	int i;
553	>	final ForkJoinTask<?> locallyDeqTask() {
554		ForkJoinTask<?>[] q = queue;
555	<	int mask = q.length - 1;
556	<	int s = sp;
557	<	if (s != base &&
558	<	(t = q[i = (s - 1) & mask]) != null &&
559	<	_unsafe.compareAndSwapObject(q, (i << qShift) + qBase, t, null)) {
560	<	_unsafe.putOrderedInt(this, spOffset, s - 1);
561	<	return t;
555	>	if (q != null) {
556	>	ForkJoinTask<?> t;
557	>	int b, i;
558	>	while (sp != (b = base)) {
559	>	if ((t = q[i = (q.length - 1) & b]) != null && base == b &&
560	>	UNSAFE.compareAndSwapObject(q, (i << qShift) + qBase,
561	>	t, null)) {
562	>	base = b + 1;
563	>	return t;
564	>	}
565	>	}
566		}
567		return null;
568		}
569
570		/**
571	<	* Specialized version of popTask to pop only if
572	<	* topmost element is the given task. Called only
573	<	* by current thread.
574	<	* @param t the task. Caller must ensure nonnull
571	>	* Returns a popped task, or null if empty. Assumes active status.
572	>	* Called only by current thread.
573	>	*/
574	>	final ForkJoinTask<?> popTask() {
575	>	int s;
576	>	ForkJoinTask<?>[] q;
577	>	if (base != (s = sp) && (q = queue) != null) {
578	>	int i = (q.length - 1) & --s;
579	>	ForkJoinTask<?> t = q[i];
580	>	if (t != null && UNSAFE.compareAndSwapObject
581	>	(q, (i << qShift) + qBase, t, null)) {
582	>	sp = s;
583	>	return t;
584	>	}
585	>	}
586	>	return null;
587	>	}
588	>
589	>	/**
590	>	* Specialized version of popTask to pop only if topmost element
591	>	* is the given task. Called only by current thread while
592	>	* active.
593	>	*
594	>	* @param t the task. Caller must ensure non-null.
595		*/
596		final boolean unpushTask(ForkJoinTask<?> t) {
597	<	ForkJoinTask<?>[] q = queue;
598	<	int mask = q.length - 1;
599	<	int s = sp - 1;
600	<	if (_unsafe.compareAndSwapObject(q, ((s & mask) << qShift) + qBase,
601	<	t, null)) {
602	<	_unsafe.putOrderedInt(this, spOffset, s);
597	>	int s;
598	>	ForkJoinTask<?>[] q;
599	>	if (base != (s = sp) && (q = queue) != null &&
600	>	UNSAFE.compareAndSwapObject
601	>	(q, (((q.length - 1) & --s) << qShift) + qBase, t, null)) {
602	>	sp = s;
603		return true;
604		}
605		return false;
606		}
607
608		/**
609	<	* Returns next task to pop.
609	>	* Returns next task or null if empty or contended
610		*/
611		final ForkJoinTask<?> peekTask() {
612		ForkJoinTask<?>[] q = queue;
613	<	return q == null? null : q[(sp - 1) & (q.length - 1)];
613	>	if (q == null)
614	>	return null;
615	>	int mask = q.length - 1;
616	>	int i = locallyFifo ? base : (sp - 1);
617	>	return q[i & mask];
618		}
619
620		/**
#	Line 400 | Line 639 | public class ForkJoinWorkerThread extend
639		ForkJoinTask<?> t = oldQ[oldIndex];
640		if (t != null && !casSlotNull(oldQ, oldIndex, t))
641		t = null;
642	<	setSlot(newQ, b & newMask, t);
642	>	writeSlot(newQ, b & newMask, t);
643		} while (++b != bf);
644	<	pool.signalIdleWorkers(false);
644	>	pool.signalWork();
645		}
646
408	–	// Runstate management
409	–
410	–	final boolean isShutdown()    { return runState >= SHUTDOWN;  }
411	–	final boolean isTerminating() { return runState >= TERMINATING;  }
412	–	final boolean isTerminated()  { return runState == TERMINATED; }
413	–	final boolean shutdown()      { return transitionRunStateTo(SHUTDOWN); }
414	–	final boolean shutdownNow()   { return transitionRunStateTo(TERMINATING); }
415	–
647		/**
648	<	* Transition to at least the given state. Return true if not
649	<	* already at least given state.
648	>	* Computes next value for random victim probe in scan().  Scans
649	>	* don't require a very high quality generator, but also not a
650	>	* crummy one.  Marsaglia xor-shift is cheap and works well enough.
651	>	* Note: This is manually inlined in scan()
652		*/
653	<	private boolean transitionRunStateTo(int state) {
654	<	for (;;) {
655	<	int s = runState;
656	<	if (s >= state)
424	<	return false;
425	<	if (_unsafe.compareAndSwapInt(this, runStateOffset, s, state))
426	<	return true;
427	<	}
653	>	private static final int xorShift(int r) {
654	>	r ^= r << 13;
655	>	r ^= r >>> 17;
656	>	return r ^ (r << 5);
657		}
658
659		/**
660	<	* Ensure status is active and if necessary adjust pool active count
660	>	* Tries to steal a task from another worker. Starts at a random
661	>	* index of workers array, and probes workers until finding one
662	>	* with non-empty queue or finding that all are empty.  It
663	>	* randomly selects the first n probes. If these are empty, it
664	>	* resorts to a circular sweep, which is necessary to accurately
665	>	* set active status. (The circular sweep uses steps of
666	>	* approximately half the array size plus 1, to avoid bias
667	>	* stemming from leftmost packing of the array in ForkJoinPool.)
668	>	*
669	>	* This method must be both fast and quiet -- usually avoiding
670	>	* memory accesses that could disrupt cache sharing etc other than
671	>	* those needed to check for and take tasks (or to activate if not
672	>	* already active). This accounts for, among other things,
673	>	* updating random seed in place without storing it until exit.
674	>	*
675	>	* @return a task, or null if none found
676		*/
677	<	final void activate() {
678	<	if (!active) {
679	<	active = true;
680	<	pool.incrementActiveCount();
677	>	private ForkJoinTask<?> scan() {
678	>	ForkJoinPool p = pool;
679	>	ForkJoinWorkerThread[] ws;        // worker array
680	>	int n;                            // upper bound of #workers
681	>	if ((ws = p.workers) != null && (n = ws.length) > 1) {
682	>	boolean canSteal = active;    // shadow active status
683	>	int r = seed;                 // extract seed once
684	>	int mask = n - 1;
685	>	int j = -n;                   // loop counter
686	>	int k = r;                    // worker index, random if j < 0
687	>	for (;;) {
688	>	ForkJoinWorkerThread v = ws[k & mask];
689	>	r ^= r << 13; r ^= r >>> 17; r ^= r << 5; // inline xorshift
690	>	if (v != null && v.base != v.sp) {
691	>	if (canSteal ||       // ensure active status
692	>	(canSteal = active = p.tryIncrementActiveCount())) {
693	>	int b = v.base;   // inline specialized deqTask
694	>	ForkJoinTask<?>[] q;
695	>	if (b != v.sp && (q = v.queue) != null) {
696	>	ForkJoinTask<?> t;
697	>	int i = (q.length - 1) & b;
698	>	long u = (i << qShift) + qBase; // raw offset
699	>	if ((t = q[i]) != null && v.base == b &&
700	>	UNSAFE.compareAndSwapObject(q, u, t, null)) {
701	>	currentSteal = t;
702	>	v.stealHint = poolIndex;
703	>	v.base = b + 1;
704	>	seed = r;
705	>	++stealCount;
706	>	return t;
707	>	}
708	>	}
709	>	}
710	>	j = -n;
711	>	k = r;                // restart on contention
712	>	}
713	>	else if (++j <= 0)
714	>	k = r;
715	>	else if (j <= n)
716	>	k += (n >>> 1) | 1;
717	>	else
718	>	break;
719	>	}
720		}
721	+	return null;
722		}
723
724	+	// Run State management
725	+
726	+	// status check methods used mainly by ForkJoinPool
727	+	final boolean isTerminating() { return (runState & TERMINATING) != 0; }
728	+	final boolean isTerminated()  { return (runState & TERMINATED) != 0; }
729	+	final boolean isSuspended()   { return (runState & SUSPENDED) != 0; }
730	+	final boolean isTrimmed()     { return (runState & TRIMMED) != 0; }
731	+
732		/**
733	<	* Ensure status is inactive and if necessary adjust pool active count
733	>	* Sets state to TERMINATING, also resuming if suspended.
734		*/
735	<	final void inactivate() {
736	<	if (active) {
737	<	active = false;
738	<	pool.decrementActiveCount();
735	>	final void shutdown() {
736	>	for (;;) {
737	>	int s = runState;
738	>	if ((s & SUSPENDED) != 0) { // kill and wakeup if suspended
739	>	if (UNSAFE.compareAndSwapInt(this, runStateOffset, s,
740	>	(s & ~SUSPENDED) |
741	>	(TRIMMED|TERMINATING))) {
742	>	LockSupport.unpark(this);
743	>	break;
744	>	}
745	>	}
746	>	else if (UNSAFE.compareAndSwapInt(this, runStateOffset, s,
747	>	s | TERMINATING))
748	>	break;
749		}
750		}
751
450	–	// Lifecycle methods
451	–
752		/**
753	<	* Initializes internal state after construction but before
454	<	* processing any tasks. If you override this method, you must
455	<	* invoke super.onStart() at the beginning of the method.
456	<	* Initialization requires care: Most fields must have legal
457	<	* default values, to ensure that attempted accesses from other
458	<	* threads work correctly even before this thread starts
459	<	* processing tasks.
753	>	* Sets state to TERMINATED. Called only by this thread.
754		*/
755	<	protected void onStart() {
756	<	juRandomSeed = randomSeedGenerator.nextLong();
757	<	do;while((randomVictimSeed = nextRandomInt()) == 0); // must be nonzero
758	<	if (queue == null)
759	<	queue = new ForkJoinTask<?>[INITIAL_QUEUE_CAPACITY];
755	>	private void setTerminated() {
756	>	int s;
757	>	do {} while (!UNSAFE.compareAndSwapInt(this, runStateOffset,
758	>	s = runState,
759	>	s | (TERMINATING|TERMINATED)));
760	>	}
761
762	<	// Heuristically allow one initial thread to warm up; others wait
763	<	if (poolIndex < pool.getParallelism() - 1) {
764	<	eventCount = pool.sync(this, 0);
765	<	activate();
766	<	}
762	>	/**
763	>	* Instrumented version of park used by ForkJoinPool.eventSync
764	>	*/
765	>	final void doPark() {
766	>	++parkCount;
767	>	LockSupport.park(this);
768		}
769
770		/**
771	<	* Perform cleanup associated with termination of this worker
476	<	* thread.  If you override this method, you must invoke
477	<	* super.onTermination at the end of the overridden method.
771	>	* If suspended, tries to set status to unsuspended and unparks.
772		*
773	<	* @param exception the exception causing this thread to abort due
480	<	* to an unrecoverable error, or null if completed normally.
773	>	* @return true if successful
774		*/
775	<	protected void onTermination(Throwable exception) {
776	<	try {
777	<	clearLocalTasks();
778	<	inactivate();
779	<	cancelTasks();
780	<	} finally {
781	<	terminate(exception);
775	>	final boolean tryResumeSpare() {
776	>	int s = runState;
777	>	if ((s & SUSPENDED) != 0 &&
778	>	UNSAFE.compareAndSwapInt(this, runStateOffset, s,
779	>	s & ~SUSPENDED)) {
780	>	LockSupport.unpark(this);
781	>	return true;
782		}
783	+	return false;
784		}
785
786		/**
787	<	* Notify pool of termination and, if exception is nonnull,
788	<	* rethrow it to trigger this thread's uncaughtExceptionHandler
787	>	* Sets suspended status and blocks as spare until resumed,
788	>	* shutdown, or timed out.
789	>	*
790	>	* @return false if trimmed
791		*/
792	<	private void terminate(Throwable exception) {
793	<	transitionRunStateTo(TERMINATED);
794	<	try {
795	<	pool.workerTerminated(this);
796	<	} finally {
797	<	if (exception != null)
798	<	ForkJoinTask.rethrowException(exception);
792	>	final boolean suspendAsSpare() {
793	>	for (;;) {               // set suspended unless terminating
794	>	int s = runState;
795	>	if ((s & TERMINATING) != 0) { // must kill
796	>	if (UNSAFE.compareAndSwapInt(this, runStateOffset, s,
797	>	s | (TRIMMED | TERMINATING)))
798	>	return false;
799	>	}
800	>	else if (UNSAFE.compareAndSwapInt(this, runStateOffset, s,
801	>	s | SUSPENDED))
802	>	break;
803	>	}
804	>	int pc = pool.parallelism;
805	>	pool.accumulateStealCount(this);
806	>	boolean timed;
807	>	long nanos;
808	>	long startTime;
809	>	if (poolIndex < pc) { // untimed wait for core threads
810	>	timed = false;
811	>	nanos = 0L;
812	>	startTime = 0L;
813	>	}
814	>	else {                // timed wait for added threads
815	>	timed = true;
816	>	nanos = SPARE_KEEPALIVE_NANOS;
817	>	startTime = System.nanoTime();
818	>	}
819	>	lastEventCount = 0;      // reset upon resume
820	>	interrupted();           // clear/ignore interrupts
821	>	while ((runState & SUSPENDED) != 0) {
822	>	++parkCount;
823	>	if (!timed)
824	>	LockSupport.park(this);
825	>	else if ((nanos -= (System.nanoTime() - startTime)) > 0)
826	>	LockSupport.parkNanos(this, nanos);
827	>	else { // try to trim on timeout
828	>	int s = runState;
829	>	if (UNSAFE.compareAndSwapInt(this, runStateOffset, s,
830	>	(s & ~SUSPENDED) |
831	>	(TRIMMED|TERMINATING)))
832	>	return false;
833	>	}
834		}
835	+	return true;
836		}
837
838	+	// Misc support methods for ForkJoinPool
839	+
840		/**
841	<	* Run local tasks on exit from main.
841	>	* Returns an estimate of the number of tasks in the queue.  Also
842	>	* used by ForkJoinTask.
843		*/
844	<	private void clearLocalTasks() {
845	<	while (base != sp && !pool.isTerminating()) {
511	<	ForkJoinTask<?> t = popTask();
512	<	if (t != null) {
513	<	activate(); // ensure active status
514	<	t.quietlyExec();
515	<	}
516	<	}
844	>	final int getQueueSize() {
845	>	return -base + sp;
846		}
847
848		/**
#	Line 521 | Line 850 | public class ForkJoinWorkerThread extend
850		* thread.
851		*/
852		final void cancelTasks() {
853	+	ForkJoinTask<?> cj = currentJoin; // try to kill live tasks
854	+	if (cj != null) {
855	+	currentJoin = null;
856	+	cj.cancelIgnoringExceptions();
857	+	}
858	+	ForkJoinTask<?> cs = currentSteal;
859	+	if (cs != null) {
860	+	currentSteal = null;
861	+	cs.cancelIgnoringExceptions();
862	+	}
863		while (base != sp) {
864		ForkJoinTask<?> t = deqTask();
865		if (t != null)
#	Line 529 | Line 868 | public class ForkJoinWorkerThread extend
868		}
869
870		/**
871	<	* This method is required to be public, but should never be
872	<	* called explicitly. It performs the main run loop to execute
873	<	* ForkJoinTasks.
871	>	* Drains tasks to given collection c.
872	>	*
873	>	* @return the number of tasks drained
874		*/
875	<	public void run() {
876	<	Throwable exception = null;
877	<	try {
878	<	onStart();
879	<	while (!isShutdown())
880	<	step();
881	<	} catch (Throwable ex) {
882	<	exception = ex;
544	<	} finally {
545	<	onTermination(exception);
875	>	final int drainTasksTo(Collection<? super ForkJoinTask<?>> c) {
876	>	int n = 0;
877	>	while (base != sp) {
878	>	ForkJoinTask<?> t = deqTask();
879	>	if (t != null) {
880	>	c.add(t);
881	>	++n;
882	>	}
883		}
884	+	return n;
885		}
886
887	+	// Support methods for ForkJoinTask
888	+
889		/**
890	<	* Main top-level action.
890	>	* Gets and removes a local task.
891	>	*
892	>	* @return a task, if available
893		*/
894	<	private void step() {
895	<	ForkJoinTask<?> t = sp != base? popTask() : null;
896	<	if (t != null || (t = scan(null, true)) != null) {
897	<	activate();
556	<	t.quietlyExec();
557	<	}
558	<	else {
559	<	inactivate();
560	<	eventCount = pool.sync(this, eventCount);
894	>	final ForkJoinTask<?> pollLocalTask() {
895	>	while (sp != base) {
896	>	if (active || (active = pool.tryIncrementActiveCount()))
897	>	return locallyFifo? locallyDeqTask() : popTask();
898		}
899	+	return null;
900		}
901
564	–	// scanning for and stealing tasks
565	–
902		/**
903	<	* Computes next value for random victim probe. Scans don't
568	<	* require a very high quality generator, but also not a crummy
569	<	* one. Marsaglia xor-shift is cheap and works well.
903	>	* Gets and removes a local or stolen task.
904		*
905	<	* This is currently unused, and manually inlined
905	>	* @return a task, if available
906		*/
907	<	private static int xorShift(int r) {
908	<	r ^= r << 1;
909	<	r ^= r >>> 3;
910	<	r ^= r << 10;
911	<	return r;
907	>	final ForkJoinTask<?> pollTask() {
908	>	ForkJoinTask<?> t = pollLocalTask();
909	>	if (t == null) {
910	>	t = scan();
911	>	currentSteal = null; // cannot retain/track
912	>	}
913	>	return t;
914		}
915
916		/**
917	<	* Tries to steal a task from another worker and/or, if enabled,
918	<	* submission queue. Starts at a random index of workers array,
919	<	* and probes workers until finding one with non-empty queue or
920	<	* finding that all are empty.  It randomly selects the first n-1
921	<	* probes. If these are empty, it resorts to full circular
922	<	* traversal, which is necessary to accurately set active status
587	<	* by caller. Also restarts if pool barrier has tripped since last
588	<	* scan, which forces refresh of workers array, in case barrier
589	<	* was associated with resize.
590	<	*
591	<	* This method must be both fast and quiet -- usually avoiding
592	<	* memory accesses that could disrupt cache sharing etc other than
593	<	* those needed to check for and take tasks. This accounts for,
594	<	* among other things, updating random seed in place without
595	<	* storing it until exit. (Note that we only need to store it if
596	<	* we found a task; otherwise it doesn't matter if we start at the
597	<	* same place next time.)
917	>	* Possibly runs some tasks and/or blocks, until task is done.
918	>	* The main body is basically a big spinloop, alternating between
919	>	* calls to helpJoinTask and pool.tryAwaitJoin with increased
920	>	* patience parameters until either the task is done without
921	>	* waiting, or we have, if necessary, created or resumed a
922	>	* replacement for this thread while it blocks.
923		*
924	<	* @param joinMe if non null; exit early if done
925	<	* @param checkSubmissions true if OK to take submissions
601	<	* @return a task, or null if none found
924	>	* @param joinMe the task to join
925	>	* @return task status on exit
926		*/
927	<	private ForkJoinTask<?> scan(ForkJoinTask<?> joinMe,
928	<	boolean checkSubmissions) {
929	<	ForkJoinPool p = pool;
930	<	if (p == null)                    // Never null, but avoids
931	<	return null;                  //   implicit nullchecks below
932	<	int r = randomVictimSeed;         // extract once to keep scan quiet
933	<	restart:                          // outer loop refreshes ws array
934	<	while (joinMe == null || joinMe.status >= 0) {
935	<	int mask;
936	<	ForkJoinWorkerThread[] ws = p.workers;
937	<	if (ws != null && (mask = ws.length - 1) > 0) {
938	<	int probes = -mask;       // use random index while negative
939	<	int idx = r;
616	<	for (;;) {
617	<	ForkJoinWorkerThread v;
618	<	// inlined xorshift to update seed
619	<	r ^= r << 1;  r ^= r >>> 3; r ^= r << 10;
620	<	if ((v = ws[mask & idx]) != null && v.sp != v.base) {
621	<	ForkJoinTask<?> t;
622	<	activate();
623	<	if ((joinMe == null || joinMe.status >= 0) &&
624	<	(t = v.deqTask()) != null) {
625	<	randomVictimSeed = r;
626	<	++stealCount;
627	<	return t;
628	<	}
629	<	continue restart; // restart on contention
630	<	}
631	<	if ((probes >> 1) <= mask) // n-1 random then circular
632	<	idx = (probes++ < 0)? r : (idx + 1);
633	<	else
634	<	break;
635	<	}
636	<	}
637	<	if (checkSubmissions && p.hasQueuedSubmissions()) {
638	<	activate();
639	<	ForkJoinTask<?> t = p.pollSubmission();
640	<	if (t != null)
641	<	return t;
642	<	}
643	<	else {
644	<	long ec = eventCount;     // restart on pool event
645	<	if ((eventCount = p.getEventCount()) == ec)
927	>	final int joinTask(ForkJoinTask<?> joinMe) {
928	>	int stat;
929	>	ForkJoinTask<?> prevJoin = currentJoin;
930	>	// Only written by this thread; only need ordered store
931	>	UNSAFE.putOrderedObject(this, currentJoinOffset, joinMe);
932	>	if ((stat = joinMe.status) >= 0 &&
933	>	(sp == base || (stat = localHelpJoinTask(joinMe)) >= 0)) {
934	>	for (int retries = 0; ; ++retries) {
935	>	helpJoinTask(joinMe, retries);
936	>	if ((stat = joinMe.status) < 0)
937	>	break;
938	>	pool.tryAwaitJoin(joinMe, retries);
939	>	if ((stat = joinMe.status) < 0)
940		break;
941	+	Thread.yield(); // tame unbounded loop
942		}
943		}
944	<	return null;
944	>	UNSAFE.putOrderedObject(this, currentJoinOffset, prevJoin);
945	>	return stat;
946		}
947
948		/**
949	<	* Callback from pool.sync to rescan before blocking.  If a
950	<	* task is found, it is pushed so it can be executed upon return.
951	<	* @return true if found and pushed a task
952	<	*/
953	<	final boolean prescan() {
954	<	ForkJoinTask<?> t = scan(null, true);
955	<	if (t != null) {
956	<	pushTask(t);
957	<	return true;
958	<	}
959	<	else {
960	<	inactivate();
961	<	return false;
949	>	* Run tasks in local queue until given task is done.
950	>	*
951	>	* @param joinMe the task to join
952	>	* @return task status on exit
953	>	*/
954	>	private int localHelpJoinTask(ForkJoinTask<?> joinMe) {
955	>	int stat, s;
956	>	ForkJoinTask<?>[] q;
957	>	while ((stat = joinMe.status) >= 0 &&
958	>	base != (s = sp) && (q = queue) != null) {
959	>	ForkJoinTask<?> t;
960	>	int i = (q.length - 1) & --s;
961	>	long u = (i << qShift) + qBase; // raw offset
962	>	if ((t = q[i]) != null &&
963	>	UNSAFE.compareAndSwapObject(q, u, t, null)) {
964	>	/*
965	>	* This recheck (and similarly in helpJoinTask)
966	>	* handles cases where joinMe is independently
967	>	* cancelled or forced even though there is other work
968	>	* available. Back out of the pop by putting t back
969	>	* into slot before we commit by writing sp.
970	>	*/
971	>	if ((stat = joinMe.status) < 0) {
972	>	UNSAFE.putObjectVolatile(q, u, t);
973	>	break;
974	>	}
975	>	sp = s;
976	>	t.tryExec();
977	>	}
978		}
979	+	return stat;
980		}
981
669	–	// Support for ForkJoinTask methods
670	–
982		/**
983	<	* Scan, returning early if joinMe done
983	>	* Tries to locate and help perform tasks for a stealer of the
984	>	* given task, or in turn one of its stealers.  Traces
985	>	* currentSteal->currentJoin links looking for a thread working on
986	>	* a descendant of the given task and with a non-empty queue to
987	>	* steal back and execute tasks from. Restarts search upon
988	>	* encountering chains that are stale, unknown, or of length
989	>	* greater than MAX_HELP_DEPTH links, to avoid unbounded cycles.
990	>	*
991	>	* The implementation is very branchy to cope with the restart
992	>	* cases.  Returns void, not task status (which must be reread by
993	>	* caller anyway) to slightly simplify control paths.
994	>	*
995	>	* @param joinMe the task to join
996	>	* @param rescans the number of times to recheck for work
997		*/
998	<	final ForkJoinTask<?> scanWhileJoining(ForkJoinTask<?> joinMe) {
999	<	ForkJoinTask<?> t = scan(joinMe, false);
1000	<	if (t != null && joinMe.status < 0 && sp == base) {
1001	<	pushTask(t); // unsteal if done and this task would be stealable
1002	<	t = null;
998	>	private void helpJoinTask(ForkJoinTask<?> joinMe, int rescans) {
999	>	ForkJoinWorkerThread[] ws = pool.workers;
1000	>	int n;
1001	>	if (ws == null || (n = ws.length) <= 1)
1002	>	return;                   // need at least 2 workers
1003	>	restart:while (rescans-- >= 0 && joinMe.status >= 0) {
1004	>	ForkJoinTask<?> task = joinMe;        // base of chain
1005	>	ForkJoinWorkerThread thread = this;   // thread with stolen task
1006	>	for (int depth = 0; depth < MAX_HELP_DEPTH; ++depth) {
1007	>	// Try to find v, the stealer of task, by first using hint
1008	>	ForkJoinWorkerThread v = ws[thread.stealHint & (n - 1)];
1009	>	if (v == null || v.currentSteal != task) {
1010	>	for (int j = 0; ; ++j) {      // search array
1011	>	if (task.status < 0 || j == n)
1012	>	continue restart;     // stale or no stealer
1013	>	if ((v = ws[j]) != null && v.currentSteal == task) {
1014	>	thread.stealHint = j; // save for next time
1015	>	break;
1016	>	}
1017	>	}
1018	>	}
1019	>	// Try to help v, using specialized form of deqTask
1020	>	int b;
1021	>	ForkJoinTask<?>[] q;
1022	>	while ((b = v.base) != v.sp && (q = v.queue) != null) {
1023	>	int i = (q.length - 1) & b;
1024	>	long u = (i << qShift) + qBase;
1025	>	ForkJoinTask<?> t = q[i];
1026	>	if (task.status < 0)          // stale
1027	>	continue restart;
1028	>	if (t != null) {
1029	>	if (v.base == b &&
1030	>	UNSAFE.compareAndSwapObject(q, u, t, null)) {
1031	>	if (joinMe.status < 0) {
1032	>	UNSAFE.putObjectVolatile(q, u, t);
1033	>	return;           // back out on cancel
1034	>	}
1035	>	ForkJoinTask<?> prevSteal = currentSteal;
1036	>	currentSteal = t;
1037	>	v.stealHint = poolIndex;
1038	>	v.base = b + 1;
1039	>	t.tryExec();
1040	>	currentSteal = prevSteal;
1041	>	}
1042	>	}
1043	>	else if (v.base == b)          // producer stalled
1044	>	continue restart;          // retry via restart
1045	>	if (joinMe.status < 0)
1046	>	return;
1047	>	}
1048	>	// Try to descend to find v's stealer
1049	>	ForkJoinTask<?> next = v.currentJoin;
1050	>	if (next == null || next == task || task.status < 0)
1051	>	continue restart;             // no descendent or stale
1052	>	if (joinMe.status < 0)
1053	>	return;
1054	>	task = next;
1055	>	thread = v;
1056	>	}
1057		}
680	–	return t;
1058		}
1059	<
1059	>
1060		/**
1061	<	* Pops or steals a task
1062	<	* @return task, or null if none available
1061	>	* Returns an estimate of the number of tasks, offset by a
1062	>	* function of number of idle workers.
1063	>	*
1064	>	* This method provides a cheap heuristic guide for task
1065	>	* partitioning when programmers, frameworks, tools, or languages
1066	>	* have little or no idea about task granularity.  In essence by
1067	>	* offering this method, we ask users only about tradeoffs in
1068	>	* overhead vs expected throughput and its variance, rather than
1069	>	* how finely to partition tasks.
1070	>	*
1071	>	* In a steady state strict (tree-structured) computation, each
1072	>	* thread makes available for stealing enough tasks for other
1073	>	* threads to remain active. Inductively, if all threads play by
1074	>	* the same rules, each thread should make available only a
1075	>	* constant number of tasks.
1076	>	*
1077	>	* The minimum useful constant is just 1. But using a value of 1
1078	>	* would require immediate replenishment upon each steal to
1079	>	* maintain enough tasks, which is infeasible.  Further,
1080	>	* partitionings/granularities of offered tasks should minimize
1081	>	* steal rates, which in general means that threads nearer the top
1082	>	* of computation tree should generate more than those nearer the
1083	>	* bottom. In perfect steady state, each thread is at
1084	>	* approximately the same level of computation tree. However,
1085	>	* producing extra tasks amortizes the uncertainty of progress and
1086	>	* diffusion assumptions.
1087	>	*
1088	>	* So, users will want to use values larger, but not much larger
1089	>	* than 1 to both smooth over transient shortages and hedge
1090	>	* against uneven progress; as traded off against the cost of
1091	>	* extra task overhead. We leave the user to pick a threshold
1092	>	* value to compare with the results of this call to guide
1093	>	* decisions, but recommend values such as 3.
1094	>	*
1095	>	* When all threads are active, it is on average OK to estimate
1096	>	* surplus strictly locally. In steady-state, if one thread is
1097	>	* maintaining say 2 surplus tasks, then so are others. So we can
1098	>	* just use estimated queue length (although note that (sp - base)
1099	>	* can be an overestimate because of stealers lagging increments
1100	>	* of base).  However, this strategy alone leads to serious
1101	>	* mis-estimates in some non-steady-state conditions (ramp-up,
1102	>	* ramp-down, other stalls). We can detect many of these by
1103	>	* further considering the number of "idle" threads, that are
1104	>	* known to have zero queued tasks, so compensate by a factor of
1105	>	* (#idle/#active) threads.
1106		*/
1107	<	final ForkJoinTask<?> pollLocalOrStolenTask() {
1108	<	ForkJoinTask<?> t;
689	<	return (t = popTask()) == null? scan(null, false) : t;
1107	>	final int getEstimatedSurplusTaskCount() {
1108	>	return sp - base - pool.idlePerActive();
1109		}
1110
1111		/**
1112	<	* Runs tasks until pool isQuiescent
1112	>	* Runs tasks until {@code pool.isQuiescent()}.
1113		*/
1114		final void helpQuiescePool() {
1115		for (;;) {
1116	<	ForkJoinTask<?> t = pollLocalOrStolenTask();
1117	<	if (t != null) {
1118	<	activate();
1119	<	t.quietlyExec();
1116	>	ForkJoinTask<?> t = pollLocalTask();
1117	>	if (t != null || (t = scan()) != null) {
1118	>	t.tryExec();
1119	>	currentSteal = null;
1120		}
1121		else {
1122	<	inactivate();
1123	<	if (pool.isQuiescent()) {
1124	<	activate(); // re-activate on exit
1125	<	break;
1122	>	ForkJoinPool p = pool;
1123	>	if (active) {
1124	>	active = false; // inactivate
1125	>	do {} while (!p.tryDecrementActiveCount());
1126	>	}
1127	>	if (p.isQuiescent()) {
1128	>	active = true; // re-activate
1129	>	do {} while (!p.tryIncrementActiveCount());
1130	>	return;
1131		}
1132		}
1133		}
1134		}
1135
1136	<	/**
713	<	* Returns an estimate of the number of tasks in the queue.
714	<	*/
715	<	final int getQueueSize() {
716	<	int n = sp - base;
717	<	return n <= 0? 0 : n; // suppress momentarily negative values
718	<	}
1136	>	// Unsafe mechanics
1137
1138	<	/**
1139	<	* Returns an estimate of the number of tasks, offset by a
1140	<	* function of number of idle workers.
1141	<	*/
1142	<	final int getEstimatedSurplusTaskCount() {
1143	<	// The halving approximates weighting idle vs non-idle workers
1144	<	return (sp - base) - (pool.getIdleThreadCount() >>> 1);
1145	<	}
728	<
729	<	// Per-worker exported random numbers
730	<
731	<	// Same constants as java.util.Random
732	<	final static long JURandomMultiplier = 0x5DEECE66DL;
733	<	final static long JURandomAddend = 0xBL;
734	<	final static long JURandomMask = (1L << 48) - 1;
735	<
736	<	private final int nextJURandom(int bits) {
737	<	long next = (juRandomSeed * JURandomMultiplier + JURandomAddend) &
738	<	JURandomMask;
739	<	juRandomSeed = next;
740	<	return (int)(next >>> (48 - bits));
741	<	}
742	<
743	<	private final int nextJURandomInt(int n) {
744	<	if (n <= 0)
745	<	throw new IllegalArgumentException("n must be positive");
746	<	int bits = nextJURandom(31);
747	<	if ((n & -n) == n)
748	<	return (int)((n * (long)bits) >> 31);
1138	>	private static final sun.misc.Unsafe UNSAFE = getUnsafe();
1139	>	private static final long runStateOffset =
1140	>	objectFieldOffset("runState", ForkJoinWorkerThread.class);
1141	>	private static final long currentJoinOffset =
1142	>	objectFieldOffset("currentJoin", ForkJoinWorkerThread.class);
1143	>	private static final long qBase =
1144	>	UNSAFE.arrayBaseOffset(ForkJoinTask[].class);
1145	>	private static final int qShift;
1146
1147	<	for (;;) {
1148	<	int val = bits % n;
1149	<	if (bits - val + (n-1) >= 0)
1150	<	return val;
1151	<	bits = nextJURandom(31);
755	<	}
756	<	}
757	<
758	<	private final long nextJURandomLong() {
759	<	return ((long)(nextJURandom(32)) << 32) + nextJURandom(32);
1147	>	static {
1148	>	int s = UNSAFE.arrayIndexScale(ForkJoinTask[].class);
1149	>	if ((s & (s-1)) != 0)
1150	>	throw new Error("data type scale not a power of two");
1151	>	qShift = 31 - Integer.numberOfLeadingZeros(s);
1152		}
1153
1154	<	private final long nextJURandomLong(long n) {
1155	<	if (n <= 0)
1156	<	throw new IllegalArgumentException("n must be positive");
1157	<	long offset = 0;
1158	<	while (n >= Integer.MAX_VALUE) { // randomly pick half range
1159	<	int bits = nextJURandom(2); // 2nd bit for odd vs even split
1160	<	long half = n >>> 1;
1161	<	long nextn = ((bits & 2) == 0)? half : n - half;
770	<	if ((bits & 1) == 0)
771	<	offset += n - nextn;
772	<	n = nextn;
1154	>	private static long objectFieldOffset(String field, Class<?> klazz) {
1155	>	try {
1156	>	return UNSAFE.objectFieldOffset(klazz.getDeclaredField(field));
1157	>	} catch (NoSuchFieldException e) {
1158	>	// Convert Exception to corresponding Error
1159	>	NoSuchFieldError error = new NoSuchFieldError(field);
1160	>	error.initCause(e);
1161	>	throw error;
1162		}
774	–	return offset + nextJURandomInt((int)n);
775	–	}
776	–
777	–	private final double nextJURandomDouble() {
778	–	return (((long)(nextJURandom(26)) << 27) + nextJURandom(27))
779	–	/ (double)(1L << 53);
780	–	}
781	–
782	–	/**
783	–	* Returns a random integer using a per-worker random
784	–	* number generator with the same properties as
785	–	* {@link java.util.Random#nextInt}
786	–	* @return the next pseudorandom, uniformly distributed {@code int}
787	–	*         value from this worker's random number generator's sequence
788	–	*/
789	–	public static int nextRandomInt() {
790	–	return ((ForkJoinWorkerThread)(Thread.currentThread())).
791	–	nextJURandom(32);
792	–	}
793	–
794	–	/**
795	–	* Returns a random integer using a per-worker random
796	–	* number generator with the same properties as
797	–	* {@link java.util.Random#nextInt(int)}
798	–	* @param n the bound on the random number to be returned.  Must be
799	–	*        positive.
800	–	* @return the next pseudorandom, uniformly distributed {@code int}
801	–	*         value between {@code 0} (inclusive) and {@code n} (exclusive)
802	–	*         from this worker's random number generator's sequence
803	–	* @throws IllegalArgumentException if n is not positive
804	–	*/
805	–	public static int nextRandomInt(int n) {
806	–	return ((ForkJoinWorkerThread)(Thread.currentThread())).
807	–	nextJURandomInt(n);
1163		}
1164
1165		/**
1166	<	* Returns a random long using a per-worker random
1167	<	* number generator with the same properties as
1168	<	* {@link java.util.Random#nextLong}
1169	<	* @return the next pseudorandom, uniformly distributed {@code long}
1170	<	*         value from this worker's random number generator's sequence
816	<	*/
817	<	public static long nextRandomLong() {
818	<	return ((ForkJoinWorkerThread)(Thread.currentThread())).
819	<	nextJURandomLong();
820	<	}
821	<
822	<	/**
823	<	* Returns a random integer using a per-worker random
824	<	* number generator with the same properties as
825	<	* {@link java.util.Random#nextInt(int)}
826	<	* @param n the bound on the random number to be returned.  Must be
827	<	*        positive.
828	<	* @return the next pseudorandom, uniformly distributed {@code int}
829	<	*         value between {@code 0} (inclusive) and {@code n} (exclusive)
830	<	*         from this worker's random number generator's sequence
831	<	* @throws IllegalArgumentException if n is not positive
832	<	*/
833	<	public static long nextRandomLong(long n) {
834	<	return ((ForkJoinWorkerThread)(Thread.currentThread())).
835	<	nextJURandomLong(n);
836	<	}
837	<
838	<	/**
839	<	* Returns a random double using a per-worker random
840	<	* number generator with the same properties as
841	<	* {@link java.util.Random#nextDouble}
842	<	* @return the next pseudorandom, uniformly distributed {@code double}
843	<	*         value between {@code 0.0} and {@code 1.0} from this
844	<	*         worker's random number generator's sequence
1166	>	* Returns a sun.misc.Unsafe.  Suitable for use in a 3rd party package.
1167	>	* Replace with a simple call to Unsafe.getUnsafe when integrating
1168	>	* into a jdk.
1169	>	*
1170	>	* @return a sun.misc.Unsafe
1171		*/
1172	<	public static double nextRandomDouble() {
847	<	return ((ForkJoinWorkerThread)(Thread.currentThread())).
848	<	nextJURandomDouble();
849	<	}
850	<
851	<	// Temporary Unsafe mechanics for preliminary release
852	<
853	<	static final Unsafe _unsafe;
854	<	static final long baseOffset;
855	<	static final long spOffset;
856	<	static final long qBase;
857	<	static final int qShift;
858	<	static final long runStateOffset;
859	<	static {
1172	>	private static sun.misc.Unsafe getUnsafe() {
1173		try {
1174	<	if (ForkJoinWorkerThread.class.getClassLoader() != null) {
1175	<	Field f = Unsafe.class.getDeclaredField("theUnsafe");
1176	<	f.setAccessible(true);
1177	<	_unsafe = (Unsafe)f.get(null);
1174	>	return sun.misc.Unsafe.getUnsafe();
1175	>	} catch (SecurityException se) {
1176	>	try {
1177	>	return java.security.AccessController.doPrivileged
1178	>	(new java.security
1179	>	.PrivilegedExceptionAction<sun.misc.Unsafe>() {
1180	>	public sun.misc.Unsafe run() throws Exception {
1181	>	java.lang.reflect.Field f = sun.misc
1182	>	.Unsafe.class.getDeclaredField("theUnsafe");
1183	>	f.setAccessible(true);
1184	>	return (sun.misc.Unsafe) f.get(null);
1185	>	}});
1186	>	} catch (java.security.PrivilegedActionException e) {
1187	>	throw new RuntimeException("Could not initialize intrinsics",
1188	>	e.getCause());
1189		}
866	–	else
867	–	_unsafe = Unsafe.getUnsafe();
868	–	baseOffset = _unsafe.objectFieldOffset
869	–	(ForkJoinWorkerThread.class.getDeclaredField("base"));
870	–	spOffset = _unsafe.objectFieldOffset
871	–	(ForkJoinWorkerThread.class.getDeclaredField("sp"));
872	–	runStateOffset = _unsafe.objectFieldOffset
873	–	(ForkJoinWorkerThread.class.getDeclaredField("runState"));
874	–	qBase = _unsafe.arrayBaseOffset(ForkJoinTask[].class);
875	–	int s = _unsafe.arrayIndexScale(ForkJoinTask[].class);
876	–	if ((s & (s-1)) != 0)
877	–	throw new Error("data type scale not a power of two");
878	–	qShift = 31 - Integer.numberOfLeadingZeros(s);
879	–	} catch (Exception e) {
880	–	throw new RuntimeException("Could not initialize intrinsics", e);
1190		}
1191		}
1192		}

Diff Legend

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