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 |
5 |
*/ |
6 |
|
7 |
package jsr166y; |
8 |
|
9 |
import java.util.concurrent.*; |
10 |
|
11 |
import java.util.Random; |
12 |
import java.util.Collection; |
13 |
import java.util.concurrent.locks.LockSupport; |
14 |
|
15 |
/** |
16 |
* A thread managed by a {@link ForkJoinPool}. This class is |
17 |
* subclassable solely for the sake of adding functionality -- there |
18 |
* are no overridable methods dealing with scheduling or 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 |
* 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 |
* Work-stealing queues are special forms of Deques that support |
41 |
* only three of the four possible end-operations -- push, pop, |
42 |
* and deq (aka steal), under the further constraints that push |
43 |
* and pop are called only from the owning thread, while deq may |
44 |
* be called from other threads. (If you are unfamiliar with |
45 |
* them, you probably want to read Herlihy and Shavit's book "The |
46 |
* Art of Multiprocessor programming", chapter 16 describing these |
47 |
* in more detail before proceeding.) The main work-stealing |
48 |
* queue design is roughly similar to those in the papers "Dynamic |
49 |
* Circular Work-Stealing Deque" by Chase and Lev, SPAA 2005 |
50 |
* (http://research.sun.com/scalable/pubs/index.html) and |
51 |
* "Idempotent work stealing" by Michael, Saraswat, and Vechev, |
52 |
* PPoPP 2009 (http://portal.acm.org/citation.cfm?id=1504186). |
53 |
* The main differences ultimately stem from gc requirements that |
54 |
* we null out taken slots as soon as we can, to maintain as small |
55 |
* a footprint as possible even in programs generating huge |
56 |
* numbers of tasks. To accomplish this, we shift the CAS |
57 |
* arbitrating pop vs deq (steal) from being on the indices |
58 |
* ("base" and "sp") to the slots themselves (mainly via method |
59 |
* "casSlotNull()"). So, both a successful pop and deq mainly |
60 |
* entail a CAS of a slot from non-null to null. Because we rely |
61 |
* on CASes of references, we do not need tag bits on base or sp. |
62 |
* They are simple ints as used in any circular array-based queue |
63 |
* (see for example ArrayDeque). Updates to the indices must |
64 |
* still be ordered in a way that guarantees that sp == base means |
65 |
* the queue is empty, but otherwise may err on the side of |
66 |
* possibly making the queue appear nonempty when a push, pop, or |
67 |
* deq have not fully committed. Note that this means that the deq |
68 |
* operation, considered individually, is not wait-free. One thief |
69 |
* cannot successfully continue until another in-progress one (or, |
70 |
* if previously empty, a push) completes. However, in the |
71 |
* aggregate, we ensure at least probabilistic non-blockingness. |
72 |
* If an attempted steal fails, a thief always chooses a different |
73 |
* random victim target to try next. So, in order for one thief to |
74 |
* progress, it suffices for any in-progress deq or new push on |
75 |
* any empty queue to complete. One reason this works well here is |
76 |
* that apparently-nonempty often means soon-to-be-stealable, |
77 |
* which gives threads a chance to set activation status if |
78 |
* necessary before stealing. |
79 |
* |
80 |
* This approach also enables support for "async mode" where local |
81 |
* task processing is in FIFO, not LIFO order; simply by using a |
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 |
* Efficient implementation of this approach currently relies on |
87 |
* an uncomfortable amount of "Unsafe" mechanics. To maintain |
88 |
* correct orderings, reads and writes of variable base require |
89 |
* volatile ordering. Variable sp does not require volatile |
90 |
* writes but still needs store-ordering, which we accomplish by |
91 |
* pre-incrementing sp before filling the slot with an ordered |
92 |
* store. (Pre-incrementing also enables backouts used in |
93 |
* scanWhileJoining.) Because they are protected by volatile base |
94 |
* reads, reads of the queue array and its slots by other threads |
95 |
* do not need volatile load semantics, but writes (in push) |
96 |
* require store order and CASes (in pop and deq) require |
97 |
* (volatile) CAS semantics. (Michael, Saraswat, and Vechev's |
98 |
* algorithm has similar properties, but without support for |
99 |
* nulling slots.) Since these combinations aren't supported |
100 |
* using ordinary volatiles, the only way to accomplish these |
101 |
* efficiently is to use direct Unsafe calls. (Using external |
102 |
* AtomicIntegers and AtomicReferenceArrays for the indices and |
103 |
* array is significantly slower because of memory locality and |
104 |
* indirection effects.) |
105 |
* |
106 |
* Further, performance on most platforms is very sensitive to |
107 |
* placement and sizing of the (resizable) queue array. Even |
108 |
* though these queues don't usually become all that big, the |
109 |
* initial size must be large enough to counteract cache |
110 |
* contention effects across multiple queues (especially in the |
111 |
* presence of GC cardmarking). Also, to improve thread-locality, |
112 |
* queues are initialized after starting. All together, these |
113 |
* low-level implementation choices produce as much as a factor of |
114 |
* 4 performance improvement compared to naive implementations, |
115 |
* and enable the processing of billions of tasks per second, |
116 |
* sometimes at the expense of ugliness. |
117 |
*/ |
118 |
|
119 |
/** |
120 |
* Generator for initial random seeds for random victim |
121 |
* selection. This is used only to create initial seeds. Random |
122 |
* steals use a cheaper xorshift generator per steal attempt. We |
123 |
* expect only rare contention on seedGenerator, so just use a |
124 |
* plain Random. |
125 |
*/ |
126 |
private static final Random seedGenerator = new Random(); |
127 |
|
128 |
/** |
129 |
* The timeout value for suspending spares. Spare workers that |
130 |
* remain unsignalled for more than this time may be trimmed |
131 |
* (killed and removed from pool). Since our goal is to avoid |
132 |
* long-term thread buildup, the exact value of timeout does not |
133 |
* matter too much so long as it avoids most false-alarm timeouts |
134 |
* under GC stalls or momentarily high system load. |
135 |
*/ |
136 |
private static final long SPARE_KEEPALIVE_NANOS = |
137 |
5L * 1000L * 1000L * 1000L; // 5 secs |
138 |
|
139 |
/** |
140 |
* Capacity of work-stealing queue array upon initialization. |
141 |
* Must be a power of two. Initial size must be at least 2, but is |
142 |
* padded to minimize cache effects. |
143 |
*/ |
144 |
private static final int INITIAL_QUEUE_CAPACITY = 1 << 13; |
145 |
|
146 |
/** |
147 |
* Maximum work-stealing queue array size. Must be less than or |
148 |
* equal to 1 << 28 to ensure lack of index wraparound. (This |
149 |
* is less than usual bounds, because we need leftshift by 3 |
150 |
* to be in int range). |
151 |
*/ |
152 |
private static final int MAXIMUM_QUEUE_CAPACITY = 1 << 28; |
153 |
|
154 |
/** |
155 |
* The pool this thread works in. Accessed directly by ForkJoinTask. |
156 |
*/ |
157 |
final ForkJoinPool pool; |
158 |
|
159 |
/** |
160 |
* The work-stealing queue array. Size must be a power of two. |
161 |
* Initialized in onStart, to improve memory locality. |
162 |
*/ |
163 |
private ForkJoinTask<?>[] queue; |
164 |
|
165 |
/** |
166 |
* Index (mod queue.length) of least valid queue slot, which is |
167 |
* always the next position to steal from if nonempty. |
168 |
*/ |
169 |
private volatile int base; |
170 |
|
171 |
/** |
172 |
* Index (mod queue.length) of next queue slot to push to or pop |
173 |
* from. It is written only by owner thread, and accessed by other |
174 |
* threads only after reading (volatile) base. Both sp and base |
175 |
* are allowed to wrap around on overflow, but (sp - base) still |
176 |
* estimates size. |
177 |
*/ |
178 |
private int sp; |
179 |
|
180 |
/** |
181 |
* Run state of this worker. In addition to the usual run levels, |
182 |
* tracks if this worker is suspended as a spare, and if it was |
183 |
* killed (trimmed) while suspended. However, "active" status is |
184 |
* maintained separately. |
185 |
*/ |
186 |
private volatile int runState; |
187 |
|
188 |
private static final int TERMINATING = 0x01; |
189 |
private static final int TERMINATED = 0x02; |
190 |
private static final int SUSPENDED = 0x04; // inactive spare |
191 |
private static final int TRIMMED = 0x08; // killed while suspended |
192 |
|
193 |
/** |
194 |
* Number of LockSupport.park calls to block this thread for |
195 |
* suspension or event waits. Used for internal instrumention; |
196 |
* currently not exported but included because volatile write upon |
197 |
* park also provides a workaround for a JVM bug. |
198 |
*/ |
199 |
private volatile int parkCount; |
200 |
|
201 |
/** |
202 |
* Number of steals, transferred and reset in pool callbacks pool |
203 |
* when idle Accessed directly by pool. |
204 |
*/ |
205 |
int stealCount; |
206 |
|
207 |
/** |
208 |
* Seed for random number generator for choosing steal victims. |
209 |
* Uses Marsaglia xorshift. Must be initialized as nonzero. |
210 |
*/ |
211 |
private int seed; |
212 |
|
213 |
/** |
214 |
* Activity status. When true, this worker is considered active. |
215 |
* Accessed directly by pool. Must be false upon construction. |
216 |
*/ |
217 |
boolean active; |
218 |
|
219 |
/** |
220 |
* True if use local fifo, not default lifo, for local polling. |
221 |
* Shadows value from ForkJoinPool, which resets it if changed |
222 |
* pool-wide. |
223 |
*/ |
224 |
private boolean locallyFifo; |
225 |
|
226 |
/** |
227 |
* Index of this worker in pool array. Set once by pool before |
228 |
* running, and accessed directly by pool to locate this worker in |
229 |
* its workers array. |
230 |
*/ |
231 |
int poolIndex; |
232 |
|
233 |
/** |
234 |
* The last pool event waited for. Accessed only by pool in |
235 |
* callback methods invoked within this thread. |
236 |
*/ |
237 |
int lastEventCount; |
238 |
|
239 |
/** |
240 |
* Encoded index and event count of next event waiter. Used only |
241 |
* by ForkJoinPool for managing event waiters. |
242 |
*/ |
243 |
volatile long nextWaiter; |
244 |
|
245 |
/** |
246 |
* Creates a ForkJoinWorkerThread operating in the given pool. |
247 |
* |
248 |
* @param pool the pool this thread works in |
249 |
* @throws NullPointerException if pool is null |
250 |
*/ |
251 |
protected ForkJoinWorkerThread(ForkJoinPool pool) { |
252 |
if (pool == null) throw new NullPointerException(); |
253 |
this.pool = pool; |
254 |
// To avoid exposing construction details to subclasses, |
255 |
// remaining initialization is in start() and onStart() |
256 |
} |
257 |
|
258 |
/** |
259 |
* Performs additional initialization and starts this thread |
260 |
*/ |
261 |
final void start(int poolIndex, boolean locallyFifo, |
262 |
UncaughtExceptionHandler ueh) { |
263 |
this.poolIndex = poolIndex; |
264 |
this.locallyFifo = locallyFifo; |
265 |
if (ueh != null) |
266 |
setUncaughtExceptionHandler(ueh); |
267 |
setDaemon(true); |
268 |
start(); |
269 |
} |
270 |
|
271 |
// Public/protected methods |
272 |
|
273 |
/** |
274 |
* Returns the pool hosting this thread. |
275 |
* |
276 |
* @return the pool |
277 |
*/ |
278 |
public ForkJoinPool getPool() { |
279 |
return pool; |
280 |
} |
281 |
|
282 |
/** |
283 |
* Returns the index number of this thread in its pool. The |
284 |
* returned value ranges from zero to the maximum number of |
285 |
* threads (minus one) that have ever been created in the pool. |
286 |
* This method may be useful for applications that track status or |
287 |
* collect results per-worker rather than per-task. |
288 |
* |
289 |
* @return the index number |
290 |
*/ |
291 |
public int getPoolIndex() { |
292 |
return poolIndex; |
293 |
} |
294 |
|
295 |
/** |
296 |
* Initializes internal state after construction but before |
297 |
* processing any tasks. If you override this method, you must |
298 |
* invoke super.onStart() at the beginning of the method. |
299 |
* Initialization requires care: Most fields must have legal |
300 |
* default values, to ensure that attempted accesses from other |
301 |
* threads work correctly even before this thread starts |
302 |
* processing tasks. |
303 |
*/ |
304 |
protected void onStart() { |
305 |
int rs = seedGenerator.nextInt(); |
306 |
seed = rs == 0? 1 : rs; // seed must be nonzero |
307 |
|
308 |
// Allocate name string and queue array in this thread |
309 |
String pid = Integer.toString(pool.getPoolNumber()); |
310 |
String wid = Integer.toString(poolIndex); |
311 |
setName("ForkJoinPool-" + pid + "-worker-" + wid); |
312 |
|
313 |
queue = new ForkJoinTask<?>[INITIAL_QUEUE_CAPACITY]; |
314 |
} |
315 |
|
316 |
/** |
317 |
* Performs cleanup associated with termination of this worker |
318 |
* thread. If you override this method, you must invoke |
319 |
* {@code super.onTermination} at the end of the overridden method. |
320 |
* |
321 |
* @param exception the exception causing this thread to abort due |
322 |
* to an unrecoverable error, or {@code null} if completed normally |
323 |
*/ |
324 |
protected void onTermination(Throwable exception) { |
325 |
try { |
326 |
cancelTasks(); |
327 |
setTerminated(); |
328 |
pool.workerTerminated(this); |
329 |
} catch (Throwable ex) { // Shouldn't ever happen |
330 |
if (exception == null) // but if so, at least rethrown |
331 |
exception = ex; |
332 |
} finally { |
333 |
if (exception != null) |
334 |
UNSAFE.throwException(exception); |
335 |
} |
336 |
} |
337 |
|
338 |
/** |
339 |
* This method is required to be public, but should never be |
340 |
* called explicitly. It performs the main run loop to execute |
341 |
* ForkJoinTasks. |
342 |
*/ |
343 |
public void run() { |
344 |
Throwable exception = null; |
345 |
try { |
346 |
onStart(); |
347 |
mainLoop(); |
348 |
} catch (Throwable ex) { |
349 |
exception = ex; |
350 |
} finally { |
351 |
onTermination(exception); |
352 |
} |
353 |
} |
354 |
|
355 |
// helpers for run() |
356 |
|
357 |
/** |
358 |
* Find and execute tasks and check status while running |
359 |
*/ |
360 |
private void mainLoop() { |
361 |
boolean ran = false; // true if ran task on previous step |
362 |
ForkJoinPool p = pool; |
363 |
for (;;) { |
364 |
p.preStep(this, ran); |
365 |
if (runState != 0) |
366 |
return; |
367 |
ForkJoinTask<?> t; // try to get and run stolen or submitted task |
368 |
if (ran = (t = scan()) != null || (t = pollSubmission()) != null) { |
369 |
t.tryExec(); |
370 |
if (base != sp) |
371 |
runLocalTasks(); |
372 |
} |
373 |
} |
374 |
} |
375 |
|
376 |
/** |
377 |
* Runs local tasks until queue is empty or shut down. Call only |
378 |
* while active. |
379 |
*/ |
380 |
private void runLocalTasks() { |
381 |
while (runState == 0) { |
382 |
ForkJoinTask<?> t = locallyFifo? locallyDeqTask() : popTask(); |
383 |
if (t != null) |
384 |
t.tryExec(); |
385 |
else if (base == sp) |
386 |
break; |
387 |
} |
388 |
} |
389 |
|
390 |
/** |
391 |
* If a submission exists, try to activate and take it |
392 |
* |
393 |
* @return a task, if available |
394 |
*/ |
395 |
private ForkJoinTask<?> pollSubmission() { |
396 |
ForkJoinPool p = pool; |
397 |
while (p.hasQueuedSubmissions()) { |
398 |
if (active || (active = p.tryIncrementActiveCount())) { |
399 |
ForkJoinTask<?> t = p.pollSubmission(); |
400 |
return t != null ? t : scan(); // if missed, rescan |
401 |
} |
402 |
} |
403 |
return null; |
404 |
} |
405 |
|
406 |
/* |
407 |
* Intrinsics-based atomic writes for queue slots. These are |
408 |
* basically the same as methods in AtomicObjectArray, but |
409 |
* specialized for (1) ForkJoinTask elements (2) requirement that |
410 |
* nullness and bounds checks have already been performed by |
411 |
* callers and (3) effective offsets are known not to overflow |
412 |
* from int to long (because of MAXIMUM_QUEUE_CAPACITY). We don't |
413 |
* need corresponding version for reads: plain array reads are OK |
414 |
* because they protected by other volatile reads and are |
415 |
* confirmed by CASes. |
416 |
* |
417 |
* Most uses don't actually call these methods, but instead contain |
418 |
* inlined forms that enable more predictable optimization. We |
419 |
* don't define the version of write used in pushTask at all, but |
420 |
* instead inline there a store-fenced array slot write. |
421 |
*/ |
422 |
|
423 |
/** |
424 |
* CASes slot i of array q from t to null. Caller must ensure q is |
425 |
* non-null and index is in range. |
426 |
*/ |
427 |
private static final boolean casSlotNull(ForkJoinTask<?>[] q, int i, |
428 |
ForkJoinTask<?> t) { |
429 |
return UNSAFE.compareAndSwapObject(q, (i << qShift) + qBase, t, null); |
430 |
} |
431 |
|
432 |
/** |
433 |
* Performs a volatile write of the given task at given slot of |
434 |
* array q. Caller must ensure q is non-null and index is in |
435 |
* range. This method is used only during resets and backouts. |
436 |
*/ |
437 |
private static final void writeSlot(ForkJoinTask<?>[] q, int i, |
438 |
ForkJoinTask<?> t) { |
439 |
UNSAFE.putObjectVolatile(q, (i << qShift) + qBase, t); |
440 |
} |
441 |
|
442 |
// queue methods |
443 |
|
444 |
/** |
445 |
* Pushes a task. Call only from this thread. |
446 |
* |
447 |
* @param t the task. Caller must ensure non-null. |
448 |
*/ |
449 |
final void pushTask(ForkJoinTask<?> t) { |
450 |
int s; |
451 |
ForkJoinTask<?>[] q = queue; |
452 |
int mask = q.length - 1; // implicit assert q != null |
453 |
UNSAFE.putOrderedObject(q, (((s = sp++) & mask) << qShift) + qBase, t); |
454 |
if ((s -= base) <= 0) |
455 |
pool.signalWork(); |
456 |
else if (s + 1 >= mask) |
457 |
growQueue(); |
458 |
} |
459 |
|
460 |
/** |
461 |
* Tries to take a task from the base of the queue, failing if |
462 |
* empty or contended. Note: Specializations of this code appear |
463 |
* in scan and scanWhileJoining. |
464 |
* |
465 |
* @return a task, or null if none or contended |
466 |
*/ |
467 |
final ForkJoinTask<?> deqTask() { |
468 |
ForkJoinTask<?> t; |
469 |
ForkJoinTask<?>[] q; |
470 |
int b, i; |
471 |
if ((b = base) != sp && |
472 |
(q = queue) != null && // must read q after b |
473 |
(t = q[i = (q.length - 1) & b]) != null && |
474 |
UNSAFE.compareAndSwapObject(q, (i << qShift) + qBase, t, null)) { |
475 |
base = b + 1; |
476 |
return t; |
477 |
} |
478 |
return null; |
479 |
} |
480 |
|
481 |
/** |
482 |
* Tries to take a task from the base of own queue. Assumes active |
483 |
* status. Called only by current thread. |
484 |
* |
485 |
* @return a task, or null if none |
486 |
*/ |
487 |
final ForkJoinTask<?> locallyDeqTask() { |
488 |
ForkJoinTask<?>[] q = queue; |
489 |
if (q != null) { |
490 |
ForkJoinTask<?> t; |
491 |
int b, i; |
492 |
while (sp != (b = base)) { |
493 |
if ((t = q[i = (q.length - 1) & b]) != null && |
494 |
UNSAFE.compareAndSwapObject(q, (i << qShift) + qBase, |
495 |
t, null)) { |
496 |
base = b + 1; |
497 |
return t; |
498 |
} |
499 |
} |
500 |
} |
501 |
return null; |
502 |
} |
503 |
|
504 |
/** |
505 |
* Returns a popped task, or null if empty. Assumes active status. |
506 |
* Called only by current thread. (Note: a specialization of this |
507 |
* code appears in scanWhileJoining.) |
508 |
*/ |
509 |
final ForkJoinTask<?> popTask() { |
510 |
int s; |
511 |
ForkJoinTask<?>[] q = queue; |
512 |
if (q != null && (s = sp) != base) { |
513 |
int i = (q.length - 1) & --s; |
514 |
ForkJoinTask<?> t = q[i]; |
515 |
if (t != null && UNSAFE.compareAndSwapObject |
516 |
(q, (i << qShift) + qBase, t, null)) { |
517 |
sp = s; |
518 |
return t; |
519 |
} |
520 |
} |
521 |
return null; |
522 |
} |
523 |
|
524 |
/** |
525 |
* Specialized version of popTask to pop only if |
526 |
* topmost element is the given task. Called only |
527 |
* by current thread while active. |
528 |
* |
529 |
* @param t the task. Caller must ensure non-null. |
530 |
*/ |
531 |
final boolean unpushTask(ForkJoinTask<?> t) { |
532 |
int s; |
533 |
ForkJoinTask<?>[] q = queue; |
534 |
if (q != null && UNSAFE.compareAndSwapObject |
535 |
(q, (((q.length - 1) & (s = sp - 1)) << qShift) + qBase, t, null)){ |
536 |
sp = s; |
537 |
return true; |
538 |
} |
539 |
return false; |
540 |
} |
541 |
|
542 |
/** |
543 |
* Returns next task or null if empty or contended |
544 |
*/ |
545 |
final ForkJoinTask<?> peekTask() { |
546 |
ForkJoinTask<?>[] q = queue; |
547 |
if (q == null) |
548 |
return null; |
549 |
int mask = q.length - 1; |
550 |
int i = locallyFifo ? base : (sp - 1); |
551 |
return q[i & mask]; |
552 |
} |
553 |
|
554 |
/** |
555 |
* Doubles queue array size. Transfers elements by emulating |
556 |
* steals (deqs) from old array and placing, oldest first, into |
557 |
* new array. |
558 |
*/ |
559 |
private void growQueue() { |
560 |
ForkJoinTask<?>[] oldQ = queue; |
561 |
int oldSize = oldQ.length; |
562 |
int newSize = oldSize << 1; |
563 |
if (newSize > MAXIMUM_QUEUE_CAPACITY) |
564 |
throw new RejectedExecutionException("Queue capacity exceeded"); |
565 |
ForkJoinTask<?>[] newQ = queue = new ForkJoinTask<?>[newSize]; |
566 |
|
567 |
int b = base; |
568 |
int bf = b + oldSize; |
569 |
int oldMask = oldSize - 1; |
570 |
int newMask = newSize - 1; |
571 |
do { |
572 |
int oldIndex = b & oldMask; |
573 |
ForkJoinTask<?> t = oldQ[oldIndex]; |
574 |
if (t != null && !casSlotNull(oldQ, oldIndex, t)) |
575 |
t = null; |
576 |
writeSlot(newQ, b & newMask, t); |
577 |
} while (++b != bf); |
578 |
pool.signalWork(); |
579 |
} |
580 |
|
581 |
/** |
582 |
* Computes next value for random victim probe in scan(). Scans |
583 |
* don't require a very high quality generator, but also not a |
584 |
* crummy one. Marsaglia xor-shift is cheap and works well enough. |
585 |
* Note: This is manually inlined in scan() |
586 |
*/ |
587 |
private static final int xorShift(int r) { |
588 |
r ^= r << 13; |
589 |
r ^= r >>> 17; |
590 |
return r ^ (r << 5); |
591 |
} |
592 |
|
593 |
/** |
594 |
* Tries to steal a task from another worker. Starts at a random |
595 |
* index of workers array, and probes workers until finding one |
596 |
* with non-empty queue or finding that all are empty. It |
597 |
* randomly selects the first n probes. If these are empty, it |
598 |
* resorts to a circular sweep, which is necessary to accurately |
599 |
* set active status. (The circular sweep uses steps of |
600 |
* approximately half the array size plus 1, to avoid bias |
601 |
* stemming from leftmost packing of the array in ForkJoinPool.) |
602 |
* |
603 |
* This method must be both fast and quiet -- usually avoiding |
604 |
* memory accesses that could disrupt cache sharing etc other than |
605 |
* those needed to check for and take tasks (or to activate if not |
606 |
* already active). This accounts for, among other things, |
607 |
* updating random seed in place without storing it until exit. |
608 |
* |
609 |
* @return a task, or null if none found |
610 |
*/ |
611 |
private ForkJoinTask<?> scan() { |
612 |
ForkJoinPool p = pool; |
613 |
ForkJoinWorkerThread[] ws = p.workers; |
614 |
int n = ws.length; // upper bound of #workers |
615 |
boolean canSteal = active; // shadow active status |
616 |
int r = seed; // extract seed once |
617 |
int k = r; // index: random if j<0 else step |
618 |
for (int j = -n; j < n; ++j) { |
619 |
ForkJoinWorkerThread v = ws[k & (n - 1)]; |
620 |
r ^= r << 13; r ^= r >>> 17; r ^= r << 5; // xorshift |
621 |
if (v != null && v.base != v.sp) { |
622 |
if (canSteal || // ensure active status |
623 |
(canSteal = active = p.tryIncrementActiveCount())) { |
624 |
int b, i; // inlined specialization of deqTask |
625 |
ForkJoinTask<?> t; |
626 |
ForkJoinTask<?>[] q; |
627 |
if ((b = v.base) != v.sp && // recheck |
628 |
(q = v.queue) != null && |
629 |
(t = q[i = (q.length - 1) & b]) != null && |
630 |
UNSAFE.compareAndSwapObject |
631 |
(q, (i << qShift) + qBase, t, null)) { |
632 |
v.base = b + 1; |
633 |
seed = r; |
634 |
++stealCount; |
635 |
return t; |
636 |
} |
637 |
} |
638 |
j = -n; // reset on contention |
639 |
} |
640 |
k = j >= 0? k + ((n >>> 1) | 1) : r; |
641 |
} |
642 |
return null; |
643 |
} |
644 |
|
645 |
// Run State management |
646 |
|
647 |
// status check methods used mainly by ForkJoinPool |
648 |
final boolean isTerminating() { return (runState & TERMINATING) != 0; } |
649 |
final boolean isTerminated() { return (runState & TERMINATED) != 0; } |
650 |
final boolean isSuspended() { return (runState & SUSPENDED) != 0; } |
651 |
final boolean isTrimmed() { return (runState & TRIMMED) != 0; } |
652 |
|
653 |
/** |
654 |
* Sets state to TERMINATING, also resuming if suspended. |
655 |
*/ |
656 |
final void shutdown() { |
657 |
for (;;) { |
658 |
int s = runState; |
659 |
if ((s & SUSPENDED) != 0) { // kill and wakeup if suspended |
660 |
if (UNSAFE.compareAndSwapInt(this, runStateOffset, s, |
661 |
(s & ~SUSPENDED) | |
662 |
(TRIMMED|TERMINATING))) { |
663 |
LockSupport.unpark(this); |
664 |
break; |
665 |
} |
666 |
} |
667 |
else if (UNSAFE.compareAndSwapInt(this, runStateOffset, s, |
668 |
s | TERMINATING)) |
669 |
break; |
670 |
} |
671 |
} |
672 |
|
673 |
/** |
674 |
* Sets state to TERMINATED. Called only by this thread. |
675 |
*/ |
676 |
private void setTerminated() { |
677 |
int s; |
678 |
do {} while (!UNSAFE.compareAndSwapInt(this, runStateOffset, |
679 |
s = runState, |
680 |
s | (TERMINATING|TERMINATED))); |
681 |
} |
682 |
|
683 |
/** |
684 |
* Instrumented version of park. Also used by ForkJoinPool.awaitEvent |
685 |
*/ |
686 |
final void doPark() { |
687 |
++parkCount; |
688 |
LockSupport.park(this); |
689 |
} |
690 |
|
691 |
/** |
692 |
* If suspended, tries to set status to unsuspended. |
693 |
* Caller must unpark to actually resume |
694 |
* |
695 |
* @return true if successful |
696 |
*/ |
697 |
final boolean tryUnsuspend() { |
698 |
int s; |
699 |
return (((s = runState) & SUSPENDED) != 0 && |
700 |
UNSAFE.compareAndSwapInt(this, runStateOffset, s, |
701 |
s & ~SUSPENDED)); |
702 |
} |
703 |
|
704 |
/** |
705 |
* Sets suspended status and blocks as spare until resumed, |
706 |
* shutdown, or timed out. |
707 |
* |
708 |
* @return false if trimmed |
709 |
*/ |
710 |
final boolean suspendAsSpare() { |
711 |
for (;;) { // set suspended unless terminating |
712 |
int s = runState; |
713 |
if ((s & TERMINATING) != 0) { // must kill |
714 |
if (UNSAFE.compareAndSwapInt(this, runStateOffset, s, |
715 |
s | (TRIMMED | TERMINATING))) |
716 |
return false; |
717 |
} |
718 |
else if (UNSAFE.compareAndSwapInt(this, runStateOffset, s, |
719 |
s | SUSPENDED)) |
720 |
break; |
721 |
} |
722 |
lastEventCount = 0; // reset upon resume |
723 |
ForkJoinPool p = pool; |
724 |
p.releaseWaiters(); // help others progress |
725 |
p.accumulateStealCount(this); |
726 |
interrupted(); // clear/ignore interrupts |
727 |
if (poolIndex < p.getParallelism()) { // untimed wait |
728 |
while ((runState & SUSPENDED) != 0) |
729 |
doPark(); |
730 |
return true; |
731 |
} |
732 |
return timedSuspend(); // timed wait if apparently non-core |
733 |
} |
734 |
|
735 |
/** |
736 |
* Blocks as spare until resumed or timed out |
737 |
* @return false if trimmed |
738 |
*/ |
739 |
private boolean timedSuspend() { |
740 |
long nanos = SPARE_KEEPALIVE_NANOS; |
741 |
long startTime = System.nanoTime(); |
742 |
while ((runState & SUSPENDED) != 0) { |
743 |
++parkCount; |
744 |
if ((nanos -= (System.nanoTime() - startTime)) > 0) |
745 |
LockSupport.parkNanos(this, nanos); |
746 |
else { // try to trim on timeout |
747 |
int s = runState; |
748 |
if (UNSAFE.compareAndSwapInt(this, runStateOffset, s, |
749 |
(s & ~SUSPENDED) | |
750 |
(TRIMMED|TERMINATING))) |
751 |
return false; |
752 |
} |
753 |
} |
754 |
return true; |
755 |
} |
756 |
|
757 |
// Misc support methods for ForkJoinPool |
758 |
|
759 |
/** |
760 |
* Returns an estimate of the number of tasks in the queue. Also |
761 |
* used by ForkJoinTask. |
762 |
*/ |
763 |
final int getQueueSize() { |
764 |
return -base + sp; |
765 |
} |
766 |
|
767 |
/** |
768 |
* Set locallyFifo mode. Called only by ForkJoinPool |
769 |
*/ |
770 |
final void setAsyncMode(boolean async) { |
771 |
locallyFifo = async; |
772 |
} |
773 |
|
774 |
/** |
775 |
* Removes and cancels all tasks in queue. Can be called from any |
776 |
* thread. |
777 |
*/ |
778 |
final void cancelTasks() { |
779 |
while (base != sp) { |
780 |
ForkJoinTask<?> t = deqTask(); |
781 |
if (t != null) |
782 |
t.cancelIgnoringExceptions(); |
783 |
} |
784 |
} |
785 |
|
786 |
/** |
787 |
* Drains tasks to given collection c. |
788 |
* |
789 |
* @return the number of tasks drained |
790 |
*/ |
791 |
final int drainTasksTo(Collection<? super ForkJoinTask<?>> c) { |
792 |
int n = 0; |
793 |
while (base != sp) { |
794 |
ForkJoinTask<?> t = deqTask(); |
795 |
if (t != null) { |
796 |
c.add(t); |
797 |
++n; |
798 |
} |
799 |
} |
800 |
return n; |
801 |
} |
802 |
|
803 |
// Support methods for ForkJoinTask |
804 |
|
805 |
/** |
806 |
* Returns an estimate of the number of tasks, offset by a |
807 |
* function of number of idle workers. |
808 |
* |
809 |
* This method provides a cheap heuristic guide for task |
810 |
* partitioning when programmers, frameworks, tools, or languages |
811 |
* have little or no idea about task granularity. In essence by |
812 |
* offering this method, we ask users only about tradeoffs in |
813 |
* overhead vs expected throughput and its variance, rather than |
814 |
* how finely to partition tasks. |
815 |
* |
816 |
* In a steady state strict (tree-structured) computation, each |
817 |
* thread makes available for stealing enough tasks for other |
818 |
* threads to remain active. Inductively, if all threads play by |
819 |
* the same rules, each thread should make available only a |
820 |
* constant number of tasks. |
821 |
* |
822 |
* The minimum useful constant is just 1. But using a value of 1 |
823 |
* would require immediate replenishment upon each steal to |
824 |
* maintain enough tasks, which is infeasible. Further, |
825 |
* partitionings/granularities of offered tasks should minimize |
826 |
* steal rates, which in general means that threads nearer the top |
827 |
* of computation tree should generate more than those nearer the |
828 |
* bottom. In perfect steady state, each thread is at |
829 |
* approximately the same level of computation tree. However, |
830 |
* producing extra tasks amortizes the uncertainty of progress and |
831 |
* diffusion assumptions. |
832 |
* |
833 |
* So, users will want to use values larger, but not much larger |
834 |
* than 1 to both smooth over transient shortages and hedge |
835 |
* against uneven progress; as traded off against the cost of |
836 |
* extra task overhead. We leave the user to pick a threshold |
837 |
* value to compare with the results of this call to guide |
838 |
* decisions, but recommend values such as 3. |
839 |
* |
840 |
* When all threads are active, it is on average OK to estimate |
841 |
* surplus strictly locally. In steady-state, if one thread is |
842 |
* maintaining say 2 surplus tasks, then so are others. So we can |
843 |
* just use estimated queue length (although note that (sp - base) |
844 |
* can be an overestimate because of stealers lagging increments |
845 |
* of base). However, this strategy alone leads to serious |
846 |
* mis-estimates in some non-steady-state conditions (ramp-up, |
847 |
* ramp-down, other stalls). We can detect many of these by |
848 |
* further considering the number of "idle" threads, that are |
849 |
* known to have zero queued tasks, so compensate by a factor of |
850 |
* (#idle/#active) threads. |
851 |
*/ |
852 |
final int getEstimatedSurplusTaskCount() { |
853 |
return sp - base - pool.idlePerActive(); |
854 |
} |
855 |
|
856 |
/** |
857 |
* Gets and removes a local task. |
858 |
* |
859 |
* @return a task, if available |
860 |
*/ |
861 |
final ForkJoinTask<?> pollLocalTask() { |
862 |
while (base != sp) { |
863 |
if (active || (active = pool.tryIncrementActiveCount())) |
864 |
return locallyFifo? locallyDeqTask() : popTask(); |
865 |
} |
866 |
return null; |
867 |
} |
868 |
|
869 |
/** |
870 |
* Gets and removes a local or stolen task. |
871 |
* |
872 |
* @return a task, if available |
873 |
*/ |
874 |
final ForkJoinTask<?> pollTask() { |
875 |
ForkJoinTask<?> t; |
876 |
return (t = pollLocalTask()) != null ? t : scan(); |
877 |
} |
878 |
|
879 |
/** |
880 |
* Returns a stolen task, if available, unless joinMe is done |
881 |
* |
882 |
* This method is intrinsically nonmodular. To maintain the |
883 |
* property that tasks are never stolen if the awaited task is |
884 |
* ready, we must interleave mechanics of scan with status |
885 |
* checks. We rely here on the commit points of deq that allow us |
886 |
* to cancel a steal even after CASing slot to null, but before |
887 |
* adjusting base index: If, after the CAS, we see that joinMe is |
888 |
* ready, we can back out by placing the task back into the slot, |
889 |
* without adjusting index. The scan loop is otherwise the same as |
890 |
* in scan. |
891 |
* |
892 |
* The outer loop cannot be allowed to run forever, because it |
893 |
* could lead to a form of deadlock if all threads are executing |
894 |
* this method. However, we must also be patient before giving up, |
895 |
* to cope with GC stalls, transient high loads, etc. The loop |
896 |
* terminates (causing caller to possibly block this thread and |
897 |
* create a replacement) only after #workers clean sweeps during |
898 |
* which all running threads are active. |
899 |
*/ |
900 |
final ForkJoinTask<?> scanWhileJoining(ForkJoinTask<?> joinMe) { |
901 |
int sweeps = 0; |
902 |
int r = seed; |
903 |
ForkJoinPool p = pool; |
904 |
p.releaseWaiters(); // help other threads progress |
905 |
while (joinMe.status >= 0) { |
906 |
ForkJoinWorkerThread[] ws = p.workers; |
907 |
int n = ws.length; |
908 |
int k = r; |
909 |
for (int j = -n; j < n; ++j) { |
910 |
ForkJoinWorkerThread v = ws[k & (n - 1)]; |
911 |
r ^= r << 13; r ^= r >>> 17; r ^= r << 5; // xorshift |
912 |
if (v != null) { |
913 |
int b = v.base; |
914 |
ForkJoinTask<?>[] q; |
915 |
if (b != v.sp && (q = v.queue) != null) { |
916 |
int i = (q.length - 1) & b; |
917 |
ForkJoinTask<?> t = q[i]; |
918 |
if (t != null) { |
919 |
if (joinMe.status < 0) |
920 |
return null; |
921 |
if (UNSAFE.compareAndSwapObject |
922 |
(q, (i << qShift) + qBase, t, null)) { |
923 |
if (joinMe.status < 0) { |
924 |
writeSlot(q, i, t); // back out |
925 |
return null; |
926 |
} |
927 |
v.base = b + 1; |
928 |
seed = r; |
929 |
++stealCount; |
930 |
return t; |
931 |
} |
932 |
} |
933 |
sweeps = 0; // ensure rescan on contention |
934 |
} |
935 |
} |
936 |
k = j >= 0? k + ((n >>> 1) | 1) : r; |
937 |
if ((j & 7) == 0 && joinMe.status < 0) // periodically recheck |
938 |
return null; |
939 |
} |
940 |
if ((sweeps = p.inactiveCount() == 0 ? sweeps + 1 : 0) > n) |
941 |
return null; |
942 |
} |
943 |
return null; |
944 |
} |
945 |
|
946 |
/** |
947 |
* Runs tasks until {@code pool.isQuiescent()}. |
948 |
*/ |
949 |
final void helpQuiescePool() { |
950 |
for (;;) { |
951 |
ForkJoinTask<?> t = pollLocalTask(); |
952 |
if (t != null || (t = scan()) != null) |
953 |
t.tryExec(); |
954 |
else { |
955 |
ForkJoinPool p = pool; |
956 |
if (active) { |
957 |
active = false; // inactivate |
958 |
do {} while (!p.tryDecrementActiveCount()); |
959 |
} |
960 |
if (p.isQuiescent()) { |
961 |
active = true; // re-activate |
962 |
do {} while (!p.tryIncrementActiveCount()); |
963 |
return; |
964 |
} |
965 |
} |
966 |
} |
967 |
} |
968 |
|
969 |
// Unsafe mechanics |
970 |
|
971 |
private static final sun.misc.Unsafe UNSAFE = getUnsafe(); |
972 |
private static final long runStateOffset = |
973 |
objectFieldOffset("runState", ForkJoinWorkerThread.class); |
974 |
private static final long qBase = |
975 |
UNSAFE.arrayBaseOffset(ForkJoinTask[].class); |
976 |
private static final int qShift; |
977 |
|
978 |
static { |
979 |
int s = UNSAFE.arrayIndexScale(ForkJoinTask[].class); |
980 |
if ((s & (s-1)) != 0) |
981 |
throw new Error("data type scale not a power of two"); |
982 |
qShift = 31 - Integer.numberOfLeadingZeros(s); |
983 |
} |
984 |
|
985 |
private static long objectFieldOffset(String field, Class<?> klazz) { |
986 |
try { |
987 |
return UNSAFE.objectFieldOffset(klazz.getDeclaredField(field)); |
988 |
} catch (NoSuchFieldException e) { |
989 |
// Convert Exception to corresponding Error |
990 |
NoSuchFieldError error = new NoSuchFieldError(field); |
991 |
error.initCause(e); |
992 |
throw error; |
993 |
} |
994 |
} |
995 |
|
996 |
/** |
997 |
* Returns a sun.misc.Unsafe. Suitable for use in a 3rd party package. |
998 |
* Replace with a simple call to Unsafe.getUnsafe when integrating |
999 |
* into a jdk. |
1000 |
* |
1001 |
* @return a sun.misc.Unsafe |
1002 |
*/ |
1003 |
private static sun.misc.Unsafe getUnsafe() { |
1004 |
try { |
1005 |
return sun.misc.Unsafe.getUnsafe(); |
1006 |
} catch (SecurityException se) { |
1007 |
try { |
1008 |
return java.security.AccessController.doPrivileged |
1009 |
(new java.security |
1010 |
.PrivilegedExceptionAction<sun.misc.Unsafe>() { |
1011 |
public sun.misc.Unsafe run() throws Exception { |
1012 |
java.lang.reflect.Field f = sun.misc |
1013 |
.Unsafe.class.getDeclaredField("theUnsafe"); |
1014 |
f.setAccessible(true); |
1015 |
return (sun.misc.Unsafe) f.get(null); |
1016 |
}}); |
1017 |
} catch (java.security.PrivilegedActionException e) { |
1018 |
throw new RuntimeException("Could not initialize intrinsics", |
1019 |
e.getCause()); |
1020 |
} |
1021 |
} |
1022 |
} |
1023 |
} |