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 1.3 by dl, Wed Jan 7 19:12:36 2009 UTC vs.
Revision 1.66 by jsr166, Wed Jun 1 21:04:30 2011 UTC

#	Line 1 | Line 1
1		/*
2		* Written by Doug Lea with assistance from members of JCP JSR-166
3		* Expert Group and released to the public domain, as explained at
4	<	* http://creativecommons.org/licenses/publicdomain
4	>	* http://creativecommons.org/publicdomain/zero/1.0/
5		*/
6
7		package jsr166y;
8	<	import java.util.*;
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.*;
8	>
9	>	import java.util.Collection;
10	>	import java.util.concurrent.RejectedExecutionException;
11
12		/**
13	<	* A thread managed by a {@link ForkJoinPool}.  This class is
14	<	* subclassable solely for the sake of adding functionality -- there
15	<	* are no overridable methods dealing with scheduling or
16	<	* execution. However, you can override initialization and termination
17	<	* cleanup methods surrounding the main task processing loop.  If you
18	<	* do create such a subclass, you will also need to supply a custom
19	<	* ForkJoinWorkerThreadFactory to use it in a ForkJoinPool.
20	<	*
21	<	* <p>This class also provides methods for generating per-thread
22	<	* random numbers, with the same properties as {@link
23	<	* java.util.Random} but with each generator isolated from those of
24	<	* other threads.
13	>	* A thread managed by a {@link ForkJoinPool}, which executes
14	>	* {@link ForkJoinTask}s.
15	>	* This class is subclassable solely for the sake of adding
16	>	* functionality -- there are no overridable methods dealing with
17	>	* scheduling or execution.  However, you can override initialization
18	>	* and termination methods surrounding the main task processing loop.
19	>	* If you do create such a subclass, you will also need to supply a
20	>	* custom {@link ForkJoinPool.ForkJoinWorkerThreadFactory} to use it
21	>	* in a {@code ForkJoinPool}.
22	>	*
23	>	* @since 1.7
24	>	* @author Doug Lea
25		*/
26		public class ForkJoinWorkerThread extends Thread {
27		/*
28	<	* Algorithm overview:
28	>	* Overview:
29	>	*
30	>	* ForkJoinWorkerThreads are managed by ForkJoinPools and perform
31	>	* ForkJoinTasks. This class includes bookkeeping in support of
32	>	* worker activation, suspension, and lifecycle control described
33	>	* in more detail in the internal documentation of class
34	>	* ForkJoinPool. And as described further below, this class also
35	>	* includes special-cased support for some ForkJoinTask
36	>	* methods. But the main mechanics involve work-stealing:
37		*
38	<	* 1. Work-Stealing: Work-stealing queues are special forms of
39	<	* Deques that support only three of the four possible
40	<	* end-operations -- push, pop, and deq (aka steal), and only do
41	<	* so under the constraints that push and pop are called only from
42	<	* the owning thread, while deq may be called from other threads.
43	<	* (If you are unfamiliar with them, you probably want to read
44	<	* Herlihy and Shavit's book "The Art of Multiprocessor
45	<	* programming", chapter 16 describing these in more detail before
46	<	* proceeding.)  The main work-stealing queue design is roughly
47	<	* similar to "Dynamic Circular Work-Stealing Deque" by David
48	<	* Chase and Yossi Lev, SPAA 2005
49	<	* (http://research.sun.com/scalable/pubs/index.html).  The main
50	<	* difference ultimately stems from gc requirements that we null
51	<	* out taken slots as soon as we can, to maintain as small a
52	<	* footprint as possible even in programs generating huge numbers
53	<	* of tasks. To accomplish this, we shift the CAS arbitrating pop
54	<	* vs deq (steal) from being on the indices ("base" and "sp") to
55	<	* the slots themselves (mainly via method "casSlotNull()"). So,
56	<	* both a successful pop and deq mainly entail CAS'ing a nonnull
57	<	* slot to null.  Because we rely on CASes of references, we do
58	<	* not need tag bits on base or sp.  They are simple ints as used
59	<	* in any circular array-based queue (see for example ArrayDeque).
38	>	* Work-stealing queues are special forms of Deques that support
39	>	* only three of the four possible end-operations -- push, pop,
40	>	* and deq (aka steal), under the further constraints that push
41	>	* and pop are called only from the owning thread, while deq may
42	>	* be called from other threads.  (If you are unfamiliar with
43	>	* them, you probably want to read Herlihy and Shavit's book "The
44	>	* Art of Multiprocessor programming", chapter 16 describing these
45	>	* in more detail before proceeding.)  The main work-stealing
46	>	* queue design is roughly similar to those in the papers "Dynamic
47	>	* Circular Work-Stealing Deque" by Chase and Lev, SPAA 2005
48	>	* (http://research.sun.com/scalable/pubs/index.html) and
49	>	* "Idempotent work stealing" by Michael, Saraswat, and Vechev,
50	>	* PPoPP 2009 (http://portal.acm.org/citation.cfm?id=1504186).
51	>	* The main differences ultimately stem from gc requirements that
52	>	* we null out taken slots as soon as we can, to maintain as small
53	>	* a footprint as possible even in programs generating huge
54	>	* numbers of tasks. To accomplish this, we shift the CAS
55	>	* arbitrating pop vs deq (steal) from being on the indices
56	>	* ("queueBase" and "queueTop") to the slots themselves (mainly
57	>	* via method "casSlotNull()"). So, both a successful pop and deq
58	>	* mainly entail a CAS of a slot from non-null to null.  Because
59	>	* we rely on CASes of references, we do not need tag bits on
60	>	* queueBase or queueTop.  They are simple ints as used in any
61	>	* circular array-based queue (see for example ArrayDeque).
62		* Updates to the indices must still be ordered in a way that
63	<	* guarantees that (sp - base) > 0 means the queue is empty, but
64	<	* otherwise may err on the side of possibly making the queue
63	>	* guarantees that queueTop == queueBase means the queue is empty,
64	>	* but otherwise may err on the side of possibly making the queue
65		* appear nonempty when a push, pop, or deq have not fully
66		* committed. Note that this means that the deq operation,
67		* considered individually, is not wait-free. One thief cannot
68		* successfully continue until another in-progress one (or, if
69		* previously empty, a push) completes.  However, in the
70	<	* aggregate, we ensure at least probablistic non-blockingness. If
71	<	* an attempted steal fails, a thief always chooses a different
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).
78	<	*
79	<	* Efficient implementation of this approach currently relies on
80	<	* an uncomfortable amount of "Unsafe" mechanics. To maintain
81	<	* correct orderings, reads and writes of variable base require
82	<	* volatile ordering.  Variable sp does not require volatile write
83	<	* but needs cheaper store-ordering on writes.  Because they are
84	<	* protected by volatile base reads, reads of the queue array and
85	<	* its slots do not need volatile load semantics, but writes (in
86	<	* push) require store order and CASes (in pop and deq) require
87	<	* (volatile) CAS semantics. Since these combinations aren't
88	<	* supported using ordinary volatiles, the only way to accomplish
89	<	* these effciently is to use direct Unsafe calls. (Using external
74	>	* any empty queue to complete.
75	>	*
76	>	* This approach also enables support for "async mode" where local
77	>	* task processing is in FIFO, not LIFO order; simply by using a
78	>	* version of deq rather than pop when locallyFifo is true (as set
79	>	* by the ForkJoinPool).  This allows use in message-passing
80	>	* frameworks in which tasks are never joined.  However neither
81	>	* mode considers affinities, loads, cache localities, etc, so
82	>	* rarely provide the best possible performance on a given
83	>	* machine, but portably provide good throughput by averaging over
84	>	* these factors.  (Further, even if we did try to use such
85	>	* information, we do not usually have a basis for exploiting
86	>	* it. For example, some sets of tasks profit from cache
87	>	* affinities, but others are harmed by cache pollution effects.)
88	>	*
89	>	* When a worker would otherwise be blocked waiting to join a
90	>	* task, it first tries a form of linear helping: Each worker
91	>	* records (in field currentSteal) the most recent task it stole
92	>	* from some other worker. Plus, it records (in field currentJoin)
93	>	* the task it is currently actively joining. Method joinTask uses
94	>	* these markers to try to find a worker to help (i.e., steal back
95	>	* a task from and execute it) that could hasten completion of the
96	>	* actively joined task. In essence, the joiner executes a task
97	>	* that would be on its own local deque had the to-be-joined task
98	>	* not been stolen. This may be seen as a conservative variant of
99	>	* the approach in Wagner & Calder "Leapfrogging: a portable
100	>	* technique for implementing efficient futures" SIGPLAN Notices,
101	>	* 1993 (http://portal.acm.org/citation.cfm?id=155354). It differs
102	>	* in that: (1) We only maintain dependency links across workers
103	>	* upon steals, rather than use per-task bookkeeping.  This may
104	>	* require a linear scan of workers array to locate stealers, but
105	>	* usually doesn't because stealers leave hints (that may become
106	>	* stale/wrong) of where to locate them. This isolates cost to
107	>	* when it is needed, rather than adding to per-task overhead.
108	>	* (2) It is "shallow", ignoring nesting and potentially cyclic
109	>	* mutual steals.  (3) It is intentionally racy: field currentJoin
110	>	* is updated only while actively joining, which means that we
111	>	* miss links in the chain during long-lived tasks, GC stalls etc
112	>	* (which is OK since blocking in such cases is usually a good
113	>	* idea).  (4) We bound the number of attempts to find work (see
114	>	* MAX_HELP) and fall back to suspending the worker and if
115	>	* necessary replacing it with another.
116	>	*
117	>	* Efficient implementation of these algorithms currently relies
118	>	* on an uncomfortable amount of "Unsafe" mechanics. To maintain
119	>	* correct orderings, reads and writes of variable queueBase
120	>	* require volatile ordering.  Variable queueTop need not be
121	>	* volatile because non-local reads always follow those of
122	>	* queueBase.  Similarly, because they are protected by volatile
123	>	* queueBase reads, reads of the queue array and its slots by
124	>	* other threads do not need volatile load semantics, but writes
125	>	* (in push) require store order and CASes (in pop and deq)
126	>	* require (volatile) CAS semantics.  (Michael, Saraswat, and
127	>	* Vechev's algorithm has similar properties, but without support
128	>	* for nulling slots.)  Since these combinations aren't supported
129	>	* using ordinary volatiles, the only way to accomplish these
130	>	* efficiently is to use direct Unsafe calls. (Using external
131		* AtomicIntegers and AtomicReferenceArrays for the indices and
132		* array is significantly slower because of memory locality and
133	<	* indirection effects.) Further, performance on most platforms is
134	<	* very sensitive to placement and sizing of the (resizable) queue
135	<	* array.  Even though these queues don't usually become all that
136	<	* big, the initial size must be large enough to counteract cache
133	>	* indirection 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
96	<	* choices produce as much as a factor of 4 performance
97	<	* improvement compared to naive implementations, and enable the
98	<	* processing of billions of tasks per second, sometimes at the
99	<	* expense of ugliness.
100	<	*
101	<	* 2. Run control: The primary run control is based on a global
102	<	* counter (activeCount) held by the pool. It uses an algorithm
103	<	* similar to that in Herlihy and Shavit section 17.6 to cause
104	<	* threads to eventually block when all threads declare they are
105	<	* inactive. (See variable "scans".)  For this to work, threads
106	<	* must be declared active when executing tasks, and before
107	<	* stealing a task. They must be inactive before blocking on the
108	<	* Pool Barrier (awaiting a new submission or other Pool
109	<	* event). In between, there is some free play which we take
110	<	* advantage of to avoid contention and rapid flickering of the
111	<	* global activeCount: If inactive, we activate only if a victim
112	<	* queue appears to be nonempty (see above).  Similarly, a thread
113	<	* tries to inactivate only after a full scan of other threads.
114	<	* The net effect is that contention on activeCount is rarely a
115	<	* measurable performance issue. (There are also a few other cases
116	<	* where we scan for work rather than retry/block upon
117	<	* contention.)
118	<	*
119	<	* 3. Selection control. We maintain policy of always choosing to
120	<	* run local tasks rather than stealing, and always trying to
121	<	* steal tasks before trying to run a new submission. All steals
122	<	* are currently performed in randomly-chosen deq-order. It may be
123	<	* 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.
142	>	*/
143	>
144	>	/**
145	>	* Mask for pool indices encoded as shorts
146		*/
147	+	private static final int  SMASK  = 0xffff;
148
149		/**
150		* Capacity of work-stealing queue array upon initialization.
151	<	* Must be a power of two. Initial size must be at least 2, but is
151	>	* Must be a power of two. Initial size must be at least 4, but is
152		* padded to minimize cache effects.
153		*/
154		private static final int INITIAL_QUEUE_CAPACITY = 1 << 13;
155
156		/**
157	<	* Maximum work-stealing queue array size.  Must be less than or
158	<	* equal to 1 << 30 to ensure lack of index wraparound.
157	>	* Maximum size for queue array. Must be a power of two
158	>	* less than or equal to 1 << (31 - width of array entry) to
159	>	* ensure lack of index wraparound, but is capped at a lower
160	>	* value to help users trap runaway computations.
161		*/
162	<	private static final int MAXIMUM_QUEUE_CAPACITY = 1 << 30;
162	>	private static final int MAXIMUM_QUEUE_CAPACITY = 1 << 24; // 16M
163
164		/**
165	<	* Generator of seeds for per-thread random numbers.
165	>	* The work-stealing queue array. Size must be a power of two.
166	>	* Initialized when started (as oposed to when constructed), to
167	>	* improve memory locality.
168		*/
169	<	private static final Random randomSeedGenerator = new Random();
169	>	ForkJoinTask<?>[] queue;
170
171		/**
172	<	* The work-stealing queue array. Size must be a power of two.
172	>	* The pool this thread works in. Accessed directly by ForkJoinTask.
173		*/
174	<	private ForkJoinTask<?>[] queue;
174	>	final ForkJoinPool pool;
175
176		/**
177		* Index (mod queue.length) of next queue slot to push to or pop
178	<	* from. It is written only by owner thread, via ordered store.
179	<	* Both sp and base are allowed to wrap around on overflow, but
180	<	* (sp - base) still estimates size.
178	>	* from. It is written only by owner thread, and accessed by other
179	>	* threads only after reading (volatile) queueBase.  Both queueTop
180	>	* and queueBase are allowed to wrap around on overflow, but
181	>	* (queueTop - queueBase) still estimates size.
182		*/
183	<	private volatile int sp;
183	>	int queueTop;
184
185		/**
186		* Index (mod queue.length) of least valid queue slot, which is
187		* always the next position to steal from if nonempty.
188		*/
189	<	private volatile int base;
189	>	volatile int queueBase;
190
191		/**
192	<	* The pool this thread works in.
192	>	* The index of most recent stealer, used as a hint to avoid
193	>	* traversal in method helpJoinTask. This is only a hint because a
194	>	* worker might have had multiple steals and this only holds one
195	>	* of them (usually the most current). Declared non-volatile,
196	>	* relying on other prevailing sync to keep reasonably current.
197		*/
198	<	final ForkJoinPool pool;
198	>	int stealHint;
199
200		/**
201		* Index of this worker in pool array. Set once by pool before
202	<	* running, and accessed directly by pool during cleanup etc
202	>	* running, and accessed directly by pool to locate this worker in
203	>	* its workers array.
204	>	*/
205	>	final int poolIndex;
206	>
207	>	/**
208	>	* Encoded record for pool task waits. Usages are always
209	>	* surrounded by volatile reads/writes
210	>	*/
211	>	int nextWait;
212	>
213	>	/**
214	>	* Complement of poolIndex, offset by count of entries of task
215	>	* waits. Accessed by ForkJoinPool to manage event waiters.
216		*/
217	<	int poolIndex;
217	>	volatile int eventCount;
218
219		/**
220	<	* Run state of this worker. Supports simple versions of the usual
221	<	* shutdown/shutdownNow control.
220	>	* Seed for random number generator for choosing steal victims.
221	>	* Uses Marsaglia xorshift. Must be initialized as nonzero.
222		*/
223	<	private volatile int runState;
223	>	int seed;
224
225	<	// Runstate values. Order matters
226	<	private static final int RUNNING     = 0;
227	<	private static final int SHUTDOWN    = 1;
228	<	private static final int TERMINATING = 2;
229	<	private static final int TERMINATED  = 3;
225	>	/**
226	>	* Number of steals. Directly accessed (and reset) by pool when
227	>	* idle.
228	>	*/
229	>	int stealCount;
230
231		/**
232	<	* Activity status. When true, this worker is considered active.
190	<	* 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.
232	>	* True if this worker should or did terminate
233		*/
234	<	private boolean active;
234	>	volatile boolean terminate;
235
236		/**
237	<	* Number of steals, transferred to pool when idle
237	>	* Set to true before LockSupport.park; false on return
238		*/
239	<	private int stealCount;
239	>	volatile boolean parked;
240
241		/**
242	<	* Seed for random number generator for choosing steal victims
242	>	* True if use local fifo, not default lifo, for local polling.
243	>	* Shadows value from ForkJoinPool.
244		*/
245	<	private int randomVictimSeed;
245	>	final boolean locallyFifo;
246
247		/**
248	<	* Seed for embedded Jurandom
248	>	* The task most recently stolen from another worker (or
249	>	* submission queue).  All uses are surrounded by enough volatile
250	>	* reads/writes to maintain as non-volatile.
251		*/
252	<	private long juRandomSeed;
252	>	ForkJoinTask<?> currentSteal;
253
254		/**
255	<	* The last barrier event waited for
255	>	* The task currently being joined, set only when actively trying
256	>	* to help other stealers in helpJoinTask. All uses are surrounded
257	>	* by enough volatile reads/writes to maintain as non-volatile.
258		*/
259	<	private long eventCount;
259	>	ForkJoinTask<?> currentJoin;
260
261		/**
262		* Creates a ForkJoinWorkerThread operating in the given pool.
263	+	*
264		* @param pool the pool this thread works in
265		* @throws NullPointerException if pool is null
266		*/
267		protected ForkJoinWorkerThread(ForkJoinPool pool) {
268	<	if (pool == null) throw new NullPointerException();
268	>	super(pool.nextWorkerName());
269		this.pool = pool;
270	<	// remaining initialization deferred to onStart
270	>	int k = pool.registerWorker(this);
271	>	poolIndex = k;
272	>	eventCount = ~k & SMASK; // clear wait count
273	>	locallyFifo = pool.locallyFifo;
274	>	Thread.UncaughtExceptionHandler ueh = pool.ueh;
275	>	if (ueh != null)
276	>	setUncaughtExceptionHandler(ueh);
277	>	setDaemon(true);
278		}
279
280	<	// public access methods
280	>	// Public methods
281
282		/**
283	<	* Returns the pool hosting the current task execution.
283	>	* Returns the pool hosting this thread.
284	>	*
285		* @return the pool
286		*/
287	<	public static ForkJoinPool getPool() {
288	<	return ((ForkJoinWorkerThread)(Thread.currentThread())).pool;
287	>	public ForkJoinPool getPool() {
288	>	return pool;
289		}
290
291		/**
292	<	* Returns the index number of the current worker thread in its
293	<	* pool.  The returned value ranges from zero to the maximum
294	<	* number of threads (minus one) that have ever been created in
295	<	* the pool.  This method may be useful for applications that
296	<	* track status or collect results on a per-worker basis.
297	<	* @return the index number.
292	>	* Returns the index number of this thread in its pool.  The
293	>	* returned value ranges from zero to the maximum number of
294	>	* threads (minus one) that have ever been created in the pool.
295	>	* This method may be useful for applications that track status or
296	>	* collect results per-worker rather than per-task.
297	>	*
298	>	* @return the index number
299		*/
300	<	public static int getPoolIndex() {
301	<	return ((ForkJoinWorkerThread)(Thread.currentThread())).poolIndex;
300	>	public int getPoolIndex() {
301	>	return poolIndex;
302		}
303
304	<	//  Access methods used by Pool
304	>	// Randomization
305
306		/**
307	<	* Get and clear steal count for accumulation by pool.  Called
308	<	* only when known to be idle (in pool.sync and termination).
307	>	* Computes next value for random victim probes and backoffs.
308	>	* Scans don't require a very high quality generator, but also not
309	>	* a crummy one.  Marsaglia xor-shift is cheap and works well
310	>	* enough.  Note: This is manually inlined in FJP.scan() to avoid
311	>	* writes inside busy loops.
312		*/
313	<	final int getAndClearStealCount() {
314	<	int sc = stealCount;
315	<	stealCount = 0;
316	<	return sc;
313	>	private int nextSeed() {
314	>	int r = seed;
315	>	r ^= r << 13;
316	>	r ^= r >>> 17;
317	>	r ^= r << 5;
318	>	return seed = r;
319		}
320
321	<	/**
262	<	* Returns estimate of the number of tasks in the queue, without
263	<	* correcting for transient negative values
264	<	*/
265	<	final int getRawQueueSize() {
266	<	return sp - base;
267	<	}
268	<
269	<	// Intrinsics-based support for queue operations.
270	<	// Currently these three (setSp, setSlot, casSlotNull) are
271	<	// usually manually inlined to improve performance
321	>	// Run State management
322
323		/**
324	<	* Sets sp in store-order.
324	>	* Initializes internal state after construction but before
325	>	* processing any tasks. If you override this method, you must
326	>	* invoke {@code super.onStart()} at the beginning of the method.
327	>	* Initialization requires care: Most fields must have legal
328	>	* default values, to ensure that attempted accesses from other
329	>	* threads work correctly even before this thread starts
330	>	* processing tasks.
331		*/
332	<	private void setSp(int s) {
333	<	_unsafe.putOrderedInt(this, spOffset, s);
332	>	protected void onStart() {
333	>	queue = new ForkJoinTask<?>[INITIAL_QUEUE_CAPACITY];
334	>	int r = ForkJoinPool.workerSeedGenerator.nextInt();
335	>	seed = (r == 0) ? 1 : r; //  must be nonzero
336		}
337
338		/**
339	<	* Add in store-order the given task at given slot of q to
340	<	* null. Caller must ensure q is nonnull and index is in range.
339	>	* Performs cleanup associated with termination of this worker
340	>	* thread.  If you override this method, you must invoke
341	>	* {@code super.onTermination} at the end of the overridden method.
342	>	*
343	>	* @param exception the exception causing this thread to abort due
344	>	* to an unrecoverable error, or {@code null} if completed normally
345		*/
346	<	private static void setSlot(ForkJoinTask<?>[] q, int i,
347	<	ForkJoinTask<?> t){
348	<	_unsafe.putOrderedObject(q, (i << qShift) + qBase, t);
346	>	protected void onTermination(Throwable exception) {
347	>	try {
348	>	terminate = true;
349	>	cancelTasks();
350	>	pool.deregisterWorker(this, exception);
351	>	} catch (Throwable ex) {        // Shouldn't ever happen
352	>	if (exception == null)      // but if so, at least rethrown
353	>	exception = ex;
354	>	} finally {
355	>	if (exception != null)
356	>	UNSAFE.throwException(exception);
357	>	}
358		}
359
360		/**
361	<	* CAS given slot of q to null. Caller must ensure q is nonnull
362	<	* and index is in range.
361	>	* This method is required to be public, but should never be
362	>	* called explicitly. It performs the main run loop to execute
363	>	* {@link ForkJoinTask}s.
364		*/
365	<	private static boolean casSlotNull(ForkJoinTask<?>[] q, int i,
366	<	ForkJoinTask<?> t) {
367	<	return _unsafe.compareAndSwapObject(q, (i << qShift) + qBase, t, null);
365	>	public void run() {
366	>	Throwable exception = null;
367	>	try {
368	>	onStart();
369	>	pool.work(this);
370	>	} catch (Throwable ex) {
371	>	exception = ex;
372	>	} finally {
373	>	onTermination(exception);
374	>	}
375		}
376
377	<	// Main queue methods
378	<
379	<	/**
380	<	* Pushes a task. Called only by current thread.
381	<	* @param t the task. Caller must ensure nonnull
377	>	/*
378	>	* Intrinsics-based atomic writes for queue slots. These are
379	>	* basically the same as methods in AtomicReferenceArray, but
380	>	* specialized for (1) ForkJoinTask elements (2) requirement that
381	>	* nullness and bounds checks have already been performed by
382	>	* callers and (3) effective offsets are known not to overflow
383	>	* from int to long (because of MAXIMUM_QUEUE_CAPACITY). We don't
384	>	* need corresponding version for reads: plain array reads are OK
385	>	* because they are protected by other volatile reads and are
386	>	* confirmed by CASes.
387	>	*
388	>	* Most uses don't actually call these methods, but instead
389	>	* contain inlined forms that enable more predictable
390	>	* optimization.  We don't define the version of write used in
391	>	* pushTask at all, but instead inline there a store-fenced array
392	>	* slot write.
393	>	*
394	>	* Also in most methods, as a performance (not correctness) issue,
395	>	* we'd like to encourage compilers not to arbitrarily postpone
396	>	* setting queueTop after writing slot.  Currently there is no
397	>	* intrinsic for arranging this, but using Unsafe putOrderedInt
398	>	* may be a preferable strategy on some compilers even though its
399	>	* main effect is a pre-, not post- fence. To simplify possible
400	>	* changes, the option is left in comments next to the associated
401	>	* assignments.
402		*/
304	–	final void pushTask(ForkJoinTask<?> t) {
305	–	ForkJoinTask<?>[] q = queue;
306	–	int mask = q.length - 1;
307	–	int s = sp;
308	–	_unsafe.putOrderedObject(q, ((s & mask) << qShift) + qBase, t);
309	–	_unsafe.putOrderedInt(this, spOffset, ++s);
310	–	if ((s -= base) == 1)
311	–	pool.signalNonEmptyWorkerQueue();
312	–	else if (s >= mask)
313	–	growQueue();
314	–	}
403
404		/**
405	<	* Tries to take a task from the base of the queue, failing if
406	<	* either empty or contended.
319	<	* @return a task, or null if none or contended.
405	>	* CASes slot i of array q from t to null. Caller must ensure q is
406	>	* non-null and index is in range.
407		*/
408	<	private ForkJoinTask<?> deqTask() {
409	<	ForkJoinTask<?>[] q;
410	<	ForkJoinTask<?> t;
324	<	int i;
325	<	int b;
326	<	if (sp != (b = base) &&
327	<	(q = queue) != null && // must read q after b
328	<	(t = q[i = (q.length - 1) & b]) != null &&
329	<	_unsafe.compareAndSwapObject(q, (i << qShift) + qBase, t, null)) {
330	<	base = b + 1;
331	<	return t;
332	<	}
333	<	return null;
408	>	private static final boolean casSlotNull(ForkJoinTask<?>[] q, int i,
409	>	ForkJoinTask<?> t) {
410	>	return UNSAFE.compareAndSwapObject(q, (i << ASHIFT) + ABASE, t, null);
411		}
412
413		/**
414	<	* Returns a popped task, or null if empty.  Called only by
415	<	* current thread.
414	>	* Performs a volatile write of the given task at given slot of
415	>	* array q.  Caller must ensure q is non-null and index is in
416	>	* range. This method is used only during resets and backouts.
417		*/
418	<	final ForkJoinTask<?> popTask() {
419	<	ForkJoinTask<?> t;
420	<	int i;
343	<	ForkJoinTask<?>[] q = queue;
344	<	int mask = q.length - 1;
345	<	int s = sp;
346	<	if (s != base &&
347	<	(t = q[i = (s - 1) & mask]) != null &&
348	<	_unsafe.compareAndSwapObject(q, (i << qShift) + qBase, t, null)) {
349	<	_unsafe.putOrderedInt(this, spOffset, s - 1);
350	<	return t;
351	<	}
352	<	return null;
418	>	private static final void writeSlot(ForkJoinTask<?>[] q, int i,
419	>	ForkJoinTask<?> t) {
420	>	UNSAFE.putObjectVolatile(q, (i << ASHIFT) + ABASE, t);
421		}
422
423	<	/**
356	<	* Specialized version of popTask to pop only if
357	<	* topmost element is the given task. Called only
358	<	* by current thread.
359	<	* @param t the task. Caller must ensure nonnull
360	<	*/
361	<	final boolean unpushTask(ForkJoinTask<?> t) {
362	<	ForkJoinTask<?>[] q = queue;
363	<	int mask = q.length - 1;
364	<	int s = sp - 1;
365	<	if (_unsafe.compareAndSwapObject(q, ((s & mask) << qShift) + qBase,
366	<	t, null)) {
367	<	_unsafe.putOrderedInt(this, spOffset, s);
368	<	return true;
369	<	}
370	<	return false;
371	<	}
423	>	// queue methods
424
425		/**
426	<	* Returns next task to pop.
426	>	* Pushes a task. Call only from this thread.
427	>	*
428	>	* @param t the task. Caller must ensure non-null.
429		*/
430	<	final ForkJoinTask<?> peekTask() {
431	<	ForkJoinTask<?>[] q = queue;
432	<	return q == null? null : q[(sp - 1) & (q.length - 1)];
430	>	final void pushTask(ForkJoinTask<?> t) {
431	>	ForkJoinTask<?>[] q; int s, m;
432	>	if ((q = queue) != null) {    // ignore if queue removed
433	>	long u = (((s = queueTop) & (m = q.length - 1)) << ASHIFT) + ABASE;
434	>	UNSAFE.putOrderedObject(q, u, t);
435	>	queueTop = s + 1;         // or use putOrderedInt
436	>	if ((s -= queueBase) <= 2)
437	>	pool.signalWork();
438	>	else if (s == m)
439	>	growQueue();
440	>	}
441		}
442
443		/**
444	<	* Doubles queue array size. Transfers elements by emulating
445	<	* steals (deqs) from old array and placing, oldest first, into
446	<	* new array.
444	>	* Creates or doubles queue array.  Transfers elements by
445	>	* emulating steals (deqs) from old array and placing, oldest
446	>	* first, into new array.
447		*/
448		private void growQueue() {
449		ForkJoinTask<?>[] oldQ = queue;
450	<	int oldSize = oldQ.length;
451	<	int newSize = oldSize << 1;
390	<	if (newSize > MAXIMUM_QUEUE_CAPACITY)
450	>	int size = oldQ != null ? oldQ.length << 1 : INITIAL_QUEUE_CAPACITY;
451	>	if (size > MAXIMUM_QUEUE_CAPACITY)
452		throw new RejectedExecutionException("Queue capacity exceeded");
453	<	ForkJoinTask<?>[] newQ = queue = new ForkJoinTask<?>[newSize];
454	<
455	<	int b = base;
456	<	int bf = b + oldSize;
457	<	int oldMask = oldSize - 1;
458	<	int newMask = newSize - 1;
459	<	do {
460	<	int oldIndex = b & oldMask;
461	<	ForkJoinTask<?> t = oldQ[oldIndex];
462	<	if (t != null && !casSlotNull(oldQ, oldIndex, t))
463	<	t = null;
464	<	setSlot(newQ, b & newMask, t);
465	<	} while (++b != bf);
466	<	pool.signalIdleWorkers(false);
406	<	}
407	<
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	<
416	<	/**
417	<	* Transition to at least the given state. Return true if not
418	<	* already at least given state.
419	<	*/
420	<	private boolean transitionRunStateTo(int state) {
421	<	for (;;) {
422	<	int s = runState;
423	<	if (s >= state)
424	<	return false;
425	<	if (_unsafe.compareAndSwapInt(this, runStateOffset, s, state))
426	<	return true;
453	>	if (size < INITIAL_QUEUE_CAPACITY)
454	>	size = INITIAL_QUEUE_CAPACITY;
455	>	ForkJoinTask<?>[] q = queue = new ForkJoinTask<?>[size];
456	>	int mask = size - 1;
457	>	int top = queueTop;
458	>	int oldMask;
459	>	if (oldQ != null && (oldMask = oldQ.length - 1) >= 0) {
460	>	for (int b = queueBase; b != top; ++b) {
461	>	long u = ((b & oldMask) << ASHIFT) + ABASE;
462	>	Object x = UNSAFE.getObjectVolatile(oldQ, u);
463	>	if (x != null && UNSAFE.compareAndSwapObject(oldQ, u, x, null))
464	>	UNSAFE.putObjectVolatile
465	>	(q, ((b & mask) << ASHIFT) + ABASE, x);
466	>	}
467		}
468		}
469
470		/**
471	<	* Ensure status is active and if necessary adjust pool active count
471	>	* Tries to take a task from the base of the queue, failing if
472	>	* empty or contended. Note: Specializations of this code appear
473	>	* in locallyDeqTask and elsewhere.
474	>	*
475	>	* @return a task, or null if none or contended
476		*/
477	<	final void activate() {
478	<	if (!active) {
479	<	active = true;
480	<	pool.incrementActiveCount();
477	>	final ForkJoinTask<?> deqTask() {
478	>	ForkJoinTask<?> t; ForkJoinTask<?>[] q; int b, i;
479	>	if (queueTop != (b = queueBase) &&
480	>	(q = queue) != null && // must read q after b
481	>	(i = (q.length - 1) & b) >= 0 &&
482	>	(t = q[i]) != null && queueBase == b &&
483	>	UNSAFE.compareAndSwapObject(q, (i << ASHIFT) + ABASE, t, null)) {
484	>	queueBase = b + 1;
485	>	return t;
486		}
487	+	return null;
488		}
489
490		/**
491	<	* Ensure status is inactive and if necessary adjust pool active count
491	>	* Tries to take a task from the base of own queue.  Called only
492	>	* by this thread.
493	>	*
494	>	* @return a task, or null if none
495		*/
496	<	final void inactivate() {
497	<	if (active) {
498	<	active = false;
499	<	pool.decrementActiveCount();
496	>	final ForkJoinTask<?> locallyDeqTask() {
497	>	ForkJoinTask<?> t; int m, b, i;
498	>	ForkJoinTask<?>[] q = queue;
499	>	if (q != null && (m = q.length - 1) >= 0) {
500	>	while (queueTop != (b = queueBase)) {
501	>	if ((t = q[i = m & b]) != null &&
502	>	queueBase == b &&
503	>	UNSAFE.compareAndSwapObject(q, (i << ASHIFT) + ABASE,
504	>	t, null)) {
505	>	queueBase = b + 1;
506	>	return t;
507	>	}
508	>	}
509		}
510	+	return null;
511		}
512
450	–	// Lifecycle methods
451	–
513		/**
514	<	* Initializes internal state after construction but before
515	<	* 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.
514	>	* Returns a popped task, or null if empty.
515	>	* Called only by this thread.
516		*/
517	<	protected void onStart() {
518	<	juRandomSeed = randomSeedGenerator.nextLong();
519	<	do;while((randomVictimSeed = nextRandomInt()) == 0); // must be nonzero
520	<	if (queue == null)
521	<	queue = new ForkJoinTask<?>[INITIAL_QUEUE_CAPACITY];
522	<
523	<	// Heuristically allow one initial thread to warm up; others wait
524	<	if (poolIndex < pool.getParallelism() - 1) {
525	<	eventCount = pool.sync(this, 0);
526	<	activate();
517	>	private ForkJoinTask<?> popTask() {
518	>	int m;
519	>	ForkJoinTask<?>[] q = queue;
520	>	if (q != null && (m = q.length - 1) >= 0) {
521	>	for (int s; (s = queueTop) != queueBase;) {
522	>	int i = m & --s;
523	>	long u = (i << ASHIFT) + ABASE; // raw offset
524	>	ForkJoinTask<?> t = q[i];
525	>	if (t == null)   // lost to stealer
526	>	break;
527	>	if (UNSAFE.compareAndSwapObject(q, u, t, null)) {
528	>	queueTop = s; // or putOrderedInt
529	>	return t;
530	>	}
531	>	}
532		}
533	+	return null;
534		}
535
536		/**
537	<	* Perform cleanup associated with termination of this worker
538	<	* thread.  If you override this method, you must invoke
477	<	* super.onTermination at the end of the overridden method.
537	>	* Specialized version of popTask to pop only if topmost element
538	>	* is the given task. Called only by this thread.
539		*
540	<	* @param exception the exception causing this thread to abort due
480	<	* to an unrecoverable error, or null if completed normally.
540	>	* @param t the task. Caller must ensure non-null.
541		*/
542	<	protected void onTermination(Throwable exception) {
543	<	try {
544	<	clearLocalTasks();
545	<	inactivate();
546	<	cancelTasks();
547	<	} finally {
548	<	terminate(exception);
542	>	final boolean unpushTask(ForkJoinTask<?> t) {
543	>	ForkJoinTask<?>[] q;
544	>	int s;
545	>	if ((q = queue) != null && (s = queueTop) != queueBase &&
546	>	UNSAFE.compareAndSwapObject
547	>	(q, (((q.length - 1) & --s) << ASHIFT) + ABASE, t, null)) {
548	>	queueTop = s; // or putOrderedInt
549	>	return true;
550		}
551	+	return false;
552		}
553
554		/**
555	<	* Notify pool of termination and, if exception is nonnull,
494	<	* rethrow it to trigger this thread's uncaughtExceptionHandler
555	>	* Returns next task, or null if empty or contended.
556		*/
557	<	private void terminate(Throwable exception) {
558	<	transitionRunStateTo(TERMINATED);
559	<	try {
560	<	pool.workerTerminated(this);
561	<	} finally {
562	<	if (exception != null)
563	<	ForkJoinTask.rethrowException(exception);
503	<	}
557	>	final ForkJoinTask<?> peekTask() {
558	>	int m;
559	>	ForkJoinTask<?>[] q = queue;
560	>	if (q == null || (m = q.length - 1) < 0)
561	>	return null;
562	>	int i = locallyFifo ? queueBase : (queueTop - 1);
563	>	return q[i & m];
564		}
565
566	+	// Support methods for ForkJoinPool
567	+
568		/**
569	<	* Run local tasks on exit from main.
569	>	* Runs the given task, plus any local tasks until queue is empty
570		*/
571	<	private void clearLocalTasks() {
572	<	while (base != sp && !pool.isTerminating()) {
573	<	ForkJoinTask<?> t = popTask();
574	<	if (t != null) {
575	<	activate(); // ensure active status
576	<	t.quietlyExec();
577	<	}
571	>	final void execTask(ForkJoinTask<?> t) {
572	>	currentSteal = t;
573	>	for (;;) {
574	>	if (t != null)
575	>	t.doExec();
576	>	if (queueTop == queueBase)
577	>	break;
578	>	t = locallyFifo ? locallyDeqTask() : popTask();
579		}
580	+	++stealCount;
581	+	currentSteal = null;
582		}
583
584		/**
#	Line 521 | Line 586 | public class ForkJoinWorkerThread extend
586		* thread.
587		*/
588		final void cancelTasks() {
589	<	while (base != sp) {
589	>	ForkJoinTask<?> cj = currentJoin; // try to cancel ongoing tasks
590	>	if (cj != null && cj.status >= 0)
591	>	cj.cancelIgnoringExceptions();
592	>	ForkJoinTask<?> cs = currentSteal;
593	>	if (cs != null && cs.status >= 0)
594	>	cs.cancelIgnoringExceptions();
595	>	while (queueBase != queueTop) {
596		ForkJoinTask<?> t = deqTask();
597		if (t != null)
598		t.cancelIgnoringExceptions();
#	Line 529 | Line 600 | public class ForkJoinWorkerThread extend
600		}
601
602		/**
603	<	* This method is required to be public, but should never be
604	<	* called explicitly. It performs the main run loop to execute
605	<	* ForkJoinTasks.
603	>	* Drains tasks to given collection c.
604	>	*
605	>	* @return the number of tasks drained
606		*/
607	<	public void run() {
608	<	Throwable exception = null;
609	<	try {
610	<	onStart();
611	<	while (!isShutdown())
612	<	step();
613	<	} catch (Throwable ex) {
614	<	exception = ex;
544	<	} finally {
545	<	onTermination(exception);
607	>	final int drainTasksTo(Collection<? super ForkJoinTask<?>> c) {
608	>	int n = 0;
609	>	while (queueBase != queueTop) {
610	>	ForkJoinTask<?> t = deqTask();
611	>	if (t != null) {
612	>	c.add(t);
613	>	++n;
614	>	}
615		}
616	+	return n;
617		}
618
619	+	// Support methods for ForkJoinTask
620	+
621		/**
622	<	* Main top-level action.
622	>	* Returns an estimate of the number of tasks in the queue.
623		*/
624	<	private void step() {
625	<	ForkJoinTask<?> t = sp != base? popTask() : null;
554	<	if (t != null || (t = scan(null, true)) != null) {
555	<	activate();
556	<	t.quietlyExec();
557	<	}
558	<	else {
559	<	inactivate();
560	<	eventCount = pool.sync(this, eventCount);
561	<	}
624	>	final int getQueueSize() {
625	>	return queueTop - queueBase;
626		}
627
564	–	// scanning for and stealing tasks
565	–
628		/**
629	<	* 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.
629	>	* Gets and removes a local task.
630		*
631	<	* This is currently unused, and manually inlined
631	>	* @return a task, if available
632		*/
633	<	private static int xorShift(int r) {
634	<	r ^= r << 1;
575	<	r ^= r >>> 3;
576	<	r ^= r << 10;
577	<	return r;
633	>	final ForkJoinTask<?> pollLocalTask() {
634	>	return locallyFifo ? locallyDeqTask() : popTask();
635		}
636
637		/**
638	<	* Tries to steal a task from another worker and/or, if enabled,
582	<	* submission queue. Starts at a random index of workers array,
583	<	* and probes workers until finding one with non-empty queue or
584	<	* finding that all are empty.  It randomly selects the first n-1
585	<	* probes. If these are empty, it resorts to full circular
586	<	* 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.)
638	>	* Gets and removes a local or stolen task.
639		*
640	<	* @param joinMe if non null; exit early if done
600	<	* @param checkSubmissions true if OK to take submissions
601	<	* @return a task, or null if none found
640	>	* @return a task, if available
641		*/
642	<	private ForkJoinTask<?> scan(ForkJoinTask<?> joinMe,
643	<	boolean checkSubmissions) {
644	<	ForkJoinPool p = pool;
645	<	if (p == null)                    // Never null, but avoids
646	<	return null;                  //   implicit nullchecks below
647	<	int r = randomVictimSeed;         // extract once to keep scan quiet
648	<	restart:                          // outer loop refreshes ws array
649	<	while (joinMe == null || joinMe.status >= 0) {
650	<	int mask;
651	<	ForkJoinWorkerThread[] ws = p.workers;
652	<	if (ws != null && (mask = ws.length - 1) > 0) {
653	<	int probes = -mask;       // use random index while negative
654	<	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)
642	>	final ForkJoinTask<?> pollTask() {
643	>	ForkJoinWorkerThread[] ws;
644	>	ForkJoinTask<?> t = pollLocalTask();
645	>	if (t != null || (ws = pool.workers) == null)
646	>	return t;
647	>	int n = ws.length; // cheap version of FJP.scan
648	>	int steps = n << 1;
649	>	int r = nextSeed();
650	>	int i = 0;
651	>	while (i < steps) {
652	>	ForkJoinWorkerThread w = ws[(i++ + r) & (n - 1)];
653	>	if (w != null && w.queueBase != w.queueTop && w.queue != null) {
654	>	if ((t = w.deqTask()) != null)
655		return t;
656	<	}
643	<	else {
644	<	long ec = eventCount;     // restart on pool event
645	<	if ((eventCount = p.getEventCount()) == ec)
646	<	break;
656	>	i = 0;
657		}
658		}
659		return null;
660		}
661
662		/**
663	<	* Callback from pool.sync to rescan before blocking.  If a
664	<	* task is found, it is pushed so it can be executed upon return.
665	<	* @return true if found and pushed a task
666	<	*/
667	<	final boolean prescan() {
668	<	ForkJoinTask<?> t = scan(null, true);
669	<	if (t != null) {
660	<	pushTask(t);
661	<	return true;
662	<	}
663	<	else {
664	<	inactivate();
665	<	return false;
666	<	}
667	<	}
668	<
669	<	// Support for ForkJoinTask methods
670	<
671	<	/**
672	<	* Scan, returning early if joinMe done
673	<	*/
674	<	final ForkJoinTask<?> scanWhileJoining(ForkJoinTask<?> joinMe) {
675	<	ForkJoinTask<?> t = scan(joinMe, false);
676	<	if (t != null && joinMe.status < 0 && sp == base) {
677	<	pushTask(t); // unsteal if done and this task would be stealable
678	<	t = null;
679	<	}
680	<	return t;
681	<	}
682	<
683	<	/**
684	<	* Pops or steals a task
685	<	* @return task, or null if none available
663	>	* The maximum stolen->joining link depth allowed in helpJoinTask,
664	>	* as well as the maximum number of retries (allowing on average
665	>	* one staleness retry per level) per attempt to instead try
666	>	* compensation.  Depths for legitimate chains are unbounded, but
667	>	* we use a fixed constant to avoid (otherwise unchecked) cycles
668	>	* and bound staleness of traversal parameters at the expense of
669	>	* sometimes blocking when we could be helping.
670		*/
671	<	final ForkJoinTask<?> pollLocalOrStolenTask() {
688	<	ForkJoinTask<?> t;
689	<	return (t = popTask()) == null? scan(null, false) : t;
690	<	}
671	>	private static final int MAX_HELP = 16;
672
673		/**
674	<	* Runs tasks until pool isQuiescent
674	>	* Possibly runs some tasks and/or blocks, until joinMe is done.
675	>	*
676	>	* @param joinMe the task to join
677	>	* @return completion status on exit
678		*/
679	<	final void helpQuiescePool() {
680	<	for (;;) {
681	<	ForkJoinTask<?> t = pollLocalOrStolenTask();
682	<	if (t != null) {
683	<	activate();
684	<	t.quietlyExec();
679	>	final int joinTask(ForkJoinTask<?> joinMe) {
680	>	ForkJoinTask<?> prevJoin = currentJoin;
681	>	currentJoin = joinMe;
682	>	for (int s, retries = MAX_HELP;;) {
683	>	if ((s = joinMe.status) < 0) {
684	>	currentJoin = prevJoin;
685	>	return s;
686		}
687	<	else {
688	<	inactivate();
689	<	if (pool.isQuiescent()) {
690	<	activate(); // re-activate on exit
691	<	break;
687	>	if (retries > 0) {
688	>	if (queueTop != queueBase) {
689	>	if (!localHelpJoinTask(joinMe))
690	>	retries = 0;           // cannot help
691	>	}
692	>	else if (retries == MAX_HELP >>> 1) {
693	>	--retries;                 // check uncommon case
694	>	if (tryDeqAndExec(joinMe) >= 0)
695	>	Thread.yield();        // for politeness
696		}
697	+	else
698	+	retries = helpJoinTask(joinMe) ? MAX_HELP : retries - 1;
699	+	}
700	+	else {
701	+	retries = MAX_HELP;           // restart if not done
702	+	pool.tryAwaitJoin(joinMe);
703		}
704		}
705		}
706
707		/**
708	<	* Returns an estimate of the number of tasks in the queue.
709	<	*/
710	<	final int getQueueSize() {
711	<	int n = sp - base;
717	<	return n <= 0? 0 : n; // suppress momentarily negative values
718	<	}
719	<
720	<	/**
721	<	* Returns an estimate of the number of tasks, offset by a
722	<	* function of number of idle workers.
708	>	* If present, pops and executes the given task, or any other
709	>	* cancelled task
710	>	*
711	>	* @return false if any other non-cancelled task exists in local queue
712		*/
713	<	final int getEstimatedSurplusTaskCount() {
714	<	// The halving approximates weighting idle vs non-idle workers
715	<	return (sp - base) - (pool.getIdleThreadCount() >>> 1);
716	<	}
717	<
718	<	// Per-worker exported random numbers
719	<
720	<	// Same constants as java.util.Random
721	<	final static long JURandomMultiplier = 0x5DEECE66DL;
722	<	final static long JURandomAddend = 0xBL;
723	<	final static long JURandomMask = (1L << 48) - 1;
724	<
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);
749	<
750	<	for (;;) {
751	<	int val = bits % n;
752	<	if (bits - val + (n-1) >= 0)
753	<	return val;
754	<	bits = nextJURandom(31);
713	>	private boolean localHelpJoinTask(ForkJoinTask<?> joinMe) {
714	>	int s, i; ForkJoinTask<?>[] q; ForkJoinTask<?> t;
715	>	if ((s = queueTop) != queueBase && (q = queue) != null &&
716	>	(i = (q.length - 1) & --s) >= 0 &&
717	>	(t = q[i]) != null) {
718	>	if (t != joinMe && t.status >= 0)
719	>	return false;
720	>	if (UNSAFE.compareAndSwapObject
721	>	(q, (i << ASHIFT) + ABASE, t, null)) {
722	>	queueTop = s;           // or putOrderedInt
723	>	t.doExec();
724	>	}
725		}
726	+	return true;
727		}
728
729	<	private final long nextJURandomLong() {
730	<	return ((long)(nextJURandom(32)) << 32) + nextJURandom(32);
731	<	}
732	<
733	<	private final long nextJURandomLong(long n) {
734	<	if (n <= 0)
735	<	throw new IllegalArgumentException("n must be positive");
736	<	long offset = 0;
737	<	while (n >= Integer.MAX_VALUE) { // randomly pick half range
738	<	int bits = nextJURandom(2); // 2nd bit for odd vs even split
739	<	long half = n >>> 1;
740	<	long nextn = ((bits & 2) == 0)? half : n - half;
741	<	if ((bits & 1) == 0)
742	<	offset += n - nextn;
743	<	n = nextn;
729	>	/**
730	>	* Tries to locate and execute tasks for a stealer of the given
731	>	* task, or in turn one of its stealers, Traces
732	>	* currentSteal->currentJoin links looking for a thread working on
733	>	* a descendant of the given task and with a non-empty queue to
734	>	* steal back and execute tasks from.  The implementation is very
735	>	* branchy to cope with potential inconsistencies or loops
736	>	* encountering chains that are stale, unknown, or of length
737	>	* greater than MAX_HELP links.  All of these cases are dealt with
738	>	* by just retrying by caller.
739	>	*
740	>	* @param joinMe the task to join
741	>	* @param canSteal true if local queue is empty
742	>	* @return true if ran a task
743	>	*/
744	>	private boolean helpJoinTask(ForkJoinTask<?> joinMe) {
745	>	boolean helped = false;
746	>	int m = pool.scanGuard & SMASK;
747	>	ForkJoinWorkerThread[] ws = pool.workers;
748	>	if (ws != null && ws.length > m && joinMe.status >= 0) {
749	>	int levels = MAX_HELP;              // remaining chain length
750	>	ForkJoinTask<?> task = joinMe;      // base of chain
751	>	outer:for (ForkJoinWorkerThread thread = this;;) {
752	>	// Try to find v, the stealer of task, by first using hint
753	>	ForkJoinWorkerThread v = ws[thread.stealHint & m];
754	>	if (v == null || v.currentSteal != task) {
755	>	for (int j = 0; ;) {        // search array
756	>	if ((v = ws[j]) != null && v.currentSteal == task) {
757	>	thread.stealHint = j;
758	>	break;              // save hint for next time
759	>	}
760	>	if (++j > m)
761	>	break outer;        // can't find stealer
762	>	}
763	>	}
764	>	// Try to help v, using specialized form of deqTask
765	>	for (;;) {
766	>	ForkJoinTask<?>[] q; int b, i;
767	>	if (joinMe.status < 0)
768	>	break outer;
769	>	if ((b = v.queueBase) == v.queueTop ||
770	>	(q = v.queue) == null ||
771	>	(i = (q.length-1) & b) < 0)
772	>	break;                  // empty
773	>	long u = (i << ASHIFT) + ABASE;
774	>	ForkJoinTask<?> t = q[i];
775	>	if (task.status < 0)
776	>	break outer;            // stale
777	>	if (t != null && v.queueBase == b &&
778	>	UNSAFE.compareAndSwapObject(q, u, t, null)) {
779	>	v.queueBase = b + 1;
780	>	v.stealHint = poolIndex;
781	>	ForkJoinTask<?> ps = currentSteal;
782	>	currentSteal = t;
783	>	t.doExec();
784	>	currentSteal = ps;
785	>	helped = true;
786	>	}
787	>	}
788	>	// Try to descend to find v's stealer
789	>	ForkJoinTask<?> next = v.currentJoin;
790	>	if (--levels > 0 && task.status >= 0 &&
791	>	next != null && next != task) {
792	>	task = next;
793	>	thread = v;
794	>	}
795	>	else
796	>	break;  // max levels, stale, dead-end, or cyclic
797	>	}
798		}
799	<	return offset + nextJURandomInt((int)n);
775	<	}
776	<
777	<	private final double nextJURandomDouble() {
778	<	return (((long)(nextJURandom(26)) << 27) + nextJURandom(27))
779	<	/ (double)(1L << 53);
799	>	return helped;
800		}
801
802		/**
803	<	* Returns a random integer using a per-worker random
804	<	* number generator with the same properties as
805	<	* {@link java.util.Random#nextInt}
806	<	* @return the next pseudorandom, uniformly distributed {@code int}
807	<	*         value from this worker's random number generator's sequence
803	>	* Performs an uncommon case for joinTask: If task t is at base of
804	>	* some workers queue, steals and executes it.
805	>	*
806	>	* @param t the task
807	>	* @return t's status
808		*/
809	<	public static int nextRandomInt() {
810	<	return ((ForkJoinWorkerThread)(Thread.currentThread())).
811	<	nextJURandom(32);
809	>	private int tryDeqAndExec(ForkJoinTask<?> t) {
810	>	int m = pool.scanGuard & SMASK;
811	>	ForkJoinWorkerThread[] ws = pool.workers;
812	>	if (ws != null && ws.length > m && t.status >= 0) {
813	>	for (int j = 0; j <= m; ++j) {
814	>	ForkJoinTask<?>[] q; int b, i;
815	>	ForkJoinWorkerThread v = ws[j];
816	>	if (v != null &&
817	>	(b = v.queueBase) != v.queueTop &&
818	>	(q = v.queue) != null &&
819	>	(i = (q.length - 1) & b) >= 0 &&
820	>	q[i] ==  t) {
821	>	long u = (i << ASHIFT) + ABASE;
822	>	if (v.queueBase == b &&
823	>	UNSAFE.compareAndSwapObject(q, u, t, null)) {
824	>	v.queueBase = b + 1;
825	>	v.stealHint = poolIndex;
826	>	ForkJoinTask<?> ps = currentSteal;
827	>	currentSteal = t;
828	>	t.doExec();
829	>	currentSteal = ps;
830	>	}
831	>	break;
832	>	}
833	>	}
834	>	}
835	>	return t.status;
836		}
837
838		/**
839	<	* Returns a random integer using a per-worker random
840	<	* number generator with the same properties as
841	<	* {@link java.util.Random#nextInt(int)}
842	<	* @param n the bound on the random number to be returned.  Must be
843	<	*        positive.
844	<	* @return the next pseudorandom, uniformly distributed {@code int}
845	<	*         value between {@code 0} (inclusive) and {@code n} (exclusive)
846	<	*         from this worker's random number generator's sequence
847	<	* @throws IllegalArgumentException if n is not positive
839	>	* Implements ForkJoinTask.getSurplusQueuedTaskCount().  Returns
840	>	* an estimate of the number of tasks, offset by a function of
841	>	* number of idle workers.
842	>	*
843	>	* This method provides a cheap heuristic guide for task
844	>	* partitioning when programmers, frameworks, tools, or languages
845	>	* have little or no idea about task granularity.  In essence by
846	>	* offering this method, we ask users only about tradeoffs in
847	>	* overhead vs expected throughput and its variance, rather than
848	>	* how finely to partition tasks.
849	>	*
850	>	* In a steady state strict (tree-structured) computation, each
851	>	* thread makes available for stealing enough tasks for other
852	>	* threads to remain active. Inductively, if all threads play by
853	>	* the same rules, each thread should make available only a
854	>	* constant number of tasks.
855	>	*
856	>	* The minimum useful constant is just 1. But using a value of 1
857	>	* would require immediate replenishment upon each steal to
858	>	* maintain enough tasks, which is infeasible.  Further,
859	>	* partitionings/granularities of offered tasks should minimize
860	>	* steal rates, which in general means that threads nearer the top
861	>	* of computation tree should generate more than those nearer the
862	>	* bottom. In perfect steady state, each thread is at
863	>	* approximately the same level of computation tree. However,
864	>	* producing extra tasks amortizes the uncertainty of progress and
865	>	* diffusion assumptions.
866	>	*
867	>	* So, users will want to use values larger, but not much larger
868	>	* than 1 to both smooth over transient shortages and hedge
869	>	* against uneven progress; as traded off against the cost of
870	>	* extra task overhead. We leave the user to pick a threshold
871	>	* value to compare with the results of this call to guide
872	>	* decisions, but recommend values such as 3.
873	>	*
874	>	* When all threads are active, it is on average OK to estimate
875	>	* surplus strictly locally. In steady-state, if one thread is
876	>	* maintaining say 2 surplus tasks, then so are others. So we can
877	>	* just use estimated queue length (although note that (queueTop -
878	>	* queueBase) can be an overestimate because of stealers lagging
879	>	* increments of queueBase).  However, this strategy alone leads
880	>	* to serious mis-estimates in some non-steady-state conditions
881	>	* (ramp-up, ramp-down, other stalls). We can detect many of these
882	>	* by further considering the number of "idle" threads, that are
883	>	* known to have zero queued tasks, so compensate by a factor of
884	>	* (#idle/#active) threads.
885		*/
886	<	public static int nextRandomInt(int n) {
887	<	return ((ForkJoinWorkerThread)(Thread.currentThread())).
807	<	nextJURandomInt(n);
886	>	final int getEstimatedSurplusTaskCount() {
887	>	return queueTop - queueBase - pool.idlePerActive();
888		}
889
890		/**
891	<	* Returns a random long using a per-worker random
892	<	* number generator with the same properties as
893	<	* {@link java.util.Random#nextLong}
894	<	* @return the next pseudorandom, uniformly distributed {@code long}
895	<	*         value from this worker's random number generator's sequence
891	>	* Runs tasks until {@code pool.isQuiescent()}. We piggyback on
892	>	* pool's active count ctl maintenance, but rather than blocking
893	>	* when tasks cannot be found, we rescan until all others cannot
894	>	* find tasks either. The bracketing by pool quiescerCounts
895	>	* updates suppresses pool auto-shutdown mechanics that could
896	>	* otherwise prematurely terminate the pool because all threads
897	>	* appear to be inactive.
898		*/
899	<	public static long nextRandomLong() {
900	<	return ((ForkJoinWorkerThread)(Thread.currentThread())).
901	<	nextJURandomLong();
899	>	final void helpQuiescePool() {
900	>	boolean active = true;
901	>	ForkJoinTask<?> ps = currentSteal; // to restore below
902	>	ForkJoinPool p = pool;
903	>	p.addQuiescerCount(1);
904	>	for (;;) {
905	>	ForkJoinWorkerThread[] ws = p.workers;
906	>	ForkJoinWorkerThread v = null;
907	>	int n;
908	>	if (queueTop != queueBase)
909	>	v = this;
910	>	else if (ws != null && (n = ws.length) > 1) {
911	>	ForkJoinWorkerThread w;
912	>	int r = nextSeed(); // cheap version of FJP.scan
913	>	int steps = n << 1;
914	>	for (int i = 0; i < steps; ++i) {
915	>	if ((w = ws[(i + r) & (n - 1)]) != null &&
916	>	w.queueBase != w.queueTop) {
917	>	v = w;
918	>	break;
919	>	}
920	>	}
921	>	}
922	>	if (v != null) {
923	>	ForkJoinTask<?> t;
924	>	if (!active) {
925	>	active = true;
926	>	p.addActiveCount(1);
927	>	}
928	>	if ((t = (v != this) ? v.deqTask() :
929	>	locallyFifo ? locallyDeqTask() : popTask()) != null) {
930	>	currentSteal = t;
931	>	t.doExec();
932	>	currentSteal = ps;
933	>	}
934	>	}
935	>	else {
936	>	if (active) {
937	>	active = false;
938	>	p.addActiveCount(-1);
939	>	}
940	>	if (p.isQuiescent()) {
941	>	p.addActiveCount(1);
942	>	p.addQuiescerCount(-1);
943	>	break;
944	>	}
945	>	}
946	>	}
947		}
948
949	<	/**
950	<	* Returns a random integer using a per-worker random
951	<	* number generator with the same properties as
952	<	* {@link java.util.Random#nextInt(int)}
953	<	* @param n the bound on the random number to be returned.  Must be
954	<	*        positive.
955	<	* @return the next pseudorandom, uniformly distributed {@code int}
956	<	*         value between {@code 0} (inclusive) and {@code n} (exclusive)
957	<	*         from this worker's random number generator's sequence
958	<	* @throws IllegalArgumentException if n is not positive
959	<	*/
960	<	public static long nextRandomLong(long n) {
961	<	return ((ForkJoinWorkerThread)(Thread.currentThread())).
962	<	nextJURandomLong(n);
949	>	// Unsafe mechanics
950	>	private static final sun.misc.Unsafe UNSAFE;
951	>	private static final long ABASE;
952	>	private static final int ASHIFT;
953	>
954	>	static {
955	>	int s;
956	>	try {
957	>	UNSAFE = getUnsafe();
958	>	Class<?> a = ForkJoinTask[].class;
959	>	ABASE = UNSAFE.arrayBaseOffset(a);
960	>	s = UNSAFE.arrayIndexScale(a);
961	>	} catch (Exception e) {
962	>	throw new Error(e);
963	>	}
964	>	if ((s & (s-1)) != 0)
965	>	throw new Error("data type scale not a power of two");
966	>	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
967		}
968
969		/**
970	<	* Returns a random double using a per-worker random
971	<	* number generator with the same properties as
972	<	* {@link java.util.Random#nextDouble}
973	<	* @return the next pseudorandom, uniformly distributed {@code double}
974	<	*         value between {@code 0.0} and {@code 1.0} from this
844	<	*         worker's random number generator's sequence
970	>	* Returns a sun.misc.Unsafe.  Suitable for use in a 3rd party package.
971	>	* Replace with a simple call to Unsafe.getUnsafe when integrating
972	>	* into a jdk.
973	>	*
974	>	* @return a sun.misc.Unsafe
975		*/
976	<	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 {
976	>	private static sun.misc.Unsafe getUnsafe() {
977		try {
978	<	if (ForkJoinWorkerThread.class.getClassLoader() != null) {
979	<	Field f = Unsafe.class.getDeclaredField("theUnsafe");
980	<	f.setAccessible(true);
981	<	_unsafe = (Unsafe)f.get(null);
982	<	}
983	<	else
984	<	_unsafe = Unsafe.getUnsafe();
985	<	baseOffset = _unsafe.objectFieldOffset
986	<	(ForkJoinWorkerThread.class.getDeclaredField("base"));
987	<	spOffset = _unsafe.objectFieldOffset
988	<	(ForkJoinWorkerThread.class.getDeclaredField("sp"));
989	<	runStateOffset = _unsafe.objectFieldOffset
990	<	(ForkJoinWorkerThread.class.getDeclaredField("runState"));
991	<	qBase = _unsafe.arrayBaseOffset(ForkJoinTask[].class);
992	<	int s = _unsafe.arrayIndexScale(ForkJoinTask[].class);
993	<	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);
978	>	return sun.misc.Unsafe.getUnsafe();
979	>	} catch (SecurityException se) {
980	>	try {
981	>	return java.security.AccessController.doPrivileged
982	>	(new java.security
983	>	.PrivilegedExceptionAction<sun.misc.Unsafe>() {
984	>	public sun.misc.Unsafe run() throws Exception {
985	>	java.lang.reflect.Field f = sun.misc
986	>	.Unsafe.class.getDeclaredField("theUnsafe");
987	>	f.setAccessible(true);
988	>	return (sun.misc.Unsafe) f.get(null);
989	>	}});
990	>	} catch (java.security.PrivilegedActionException e) {
991	>	throw new RuntimeException("Could not initialize intrinsics",
992	>	e.getCause());
993	>	}
994		}
995		}
996		}

Diff Legend

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