6 |
|
|
7 |
|
package jsr166y; |
8 |
|
|
9 |
– |
import java.util.Collection; |
10 |
– |
import java.util.concurrent.RejectedExecutionException; |
11 |
– |
|
9 |
|
/** |
10 |
|
* A thread managed by a {@link ForkJoinPool}, which executes |
11 |
|
* {@link ForkJoinTask}s. |
22 |
|
*/ |
23 |
|
public class ForkJoinWorkerThread extends Thread { |
24 |
|
/* |
28 |
– |
* Overview: |
29 |
– |
* |
25 |
|
* ForkJoinWorkerThreads are managed by ForkJoinPools and perform |
26 |
< |
* ForkJoinTasks. This class includes bookkeeping in support of |
27 |
< |
* 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 |
< |
* 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 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 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. |
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.) |
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 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 4, but is |
152 |
< |
* padded to minimize cache effects. |
153 |
< |
*/ |
154 |
< |
private static final int INITIAL_QUEUE_CAPACITY = 1 << 13; |
155 |
< |
|
156 |
< |
/** |
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 << 24; // 16M |
163 |
< |
|
164 |
< |
/** |
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 |
< |
ForkJoinTask<?>[] queue; |
170 |
< |
|
171 |
< |
/** |
172 |
< |
* The pool this thread works in. Accessed directly by ForkJoinTask. |
173 |
< |
*/ |
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, 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 |
< |
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 |
< |
volatile int queueBase; |
190 |
< |
|
191 |
< |
/** |
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 |
< |
int stealHint; |
199 |
< |
|
200 |
< |
/** |
201 |
< |
* Index of this worker in pool array. Set once by pool before |
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 |
< |
volatile int eventCount; |
218 |
< |
|
219 |
< |
/** |
220 |
< |
* Seed for random number generator for choosing steal victims. |
221 |
< |
* Uses Marsaglia xorshift. Must be initialized as nonzero. |
222 |
< |
*/ |
223 |
< |
int seed; |
224 |
< |
|
225 |
< |
/** |
226 |
< |
* Number of steals. Directly accessed (and reset) by pool when |
227 |
< |
* idle. |
228 |
< |
*/ |
229 |
< |
int stealCount; |
230 |
< |
|
231 |
< |
/** |
232 |
< |
* True if this worker should or did terminate |
233 |
< |
*/ |
234 |
< |
volatile boolean terminate; |
235 |
< |
|
236 |
< |
/** |
237 |
< |
* Set to true before LockSupport.park; false on return |
238 |
< |
*/ |
239 |
< |
volatile boolean parked; |
240 |
< |
|
241 |
< |
/** |
242 |
< |
* True if use local fifo, not default lifo, for local polling. |
243 |
< |
* Shadows value from ForkJoinPool. |
244 |
< |
*/ |
245 |
< |
final boolean locallyFifo; |
246 |
< |
|
247 |
< |
/** |
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. |
26 |
> |
* ForkJoinTasks. For explanation, see the internal documentation |
27 |
> |
* of class ForkJoinPool. |
28 |
|
*/ |
252 |
– |
ForkJoinTask<?> currentSteal; |
29 |
|
|
30 |
< |
/** |
31 |
< |
* 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 |
< |
ForkJoinTask<?> currentJoin; |
30 |
> |
final ForkJoinPool.WorkQueue workQueue; // Work-stealing mechanics |
31 |
> |
final ForkJoinPool pool; // the pool this thread works in |
32 |
|
|
33 |
|
/** |
34 |
|
* Creates a ForkJoinWorkerThread operating in the given pool. |
38 |
|
*/ |
39 |
|
protected ForkJoinWorkerThread(ForkJoinPool pool) { |
40 |
|
super(pool.nextWorkerName()); |
41 |
< |
this.pool = pool; |
270 |
< |
int k = pool.registerWorker(this); |
271 |
< |
poolIndex = k; |
272 |
< |
eventCount = ~k & SMASK; // clear wait count |
273 |
< |
locallyFifo = pool.locallyFifo; |
41 |
> |
setDaemon(true); |
42 |
|
Thread.UncaughtExceptionHandler ueh = pool.ueh; |
43 |
|
if (ueh != null) |
44 |
|
setUncaughtExceptionHandler(ueh); |
45 |
< |
setDaemon(true); |
45 |
> |
this.pool = pool; |
46 |
> |
this.workQueue = new ForkJoinPool.WorkQueue(this, pool.localMode); |
47 |
> |
pool.registerWorker(this); |
48 |
|
} |
49 |
|
|
280 |
– |
// Public methods |
281 |
– |
|
50 |
|
/** |
51 |
|
* Returns the pool hosting this thread. |
52 |
|
* |
66 |
|
* @return the index number |
67 |
|
*/ |
68 |
|
public int getPoolIndex() { |
69 |
< |
return poolIndex; |
69 |
> |
return workQueue.poolIndex; |
70 |
|
} |
71 |
|
|
304 |
– |
// Randomization |
305 |
– |
|
306 |
– |
/** |
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 |
– |
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 |
– |
// Run State management |
322 |
– |
|
72 |
|
/** |
73 |
|
* Initializes internal state after construction but before |
74 |
|
* processing any tasks. If you override this method, you must |
79 |
|
* processing tasks. |
80 |
|
*/ |
81 |
|
protected void onStart() { |
333 |
– |
queue = new ForkJoinTask<?>[INITIAL_QUEUE_CAPACITY]; |
334 |
– |
int r = pool.workerSeedGenerator.nextInt(); |
335 |
– |
seed = (r == 0)? 1 : r; // must be nonzero |
82 |
|
} |
83 |
|
|
84 |
|
/** |
90 |
|
* to an unrecoverable error, or {@code null} if completed normally |
91 |
|
*/ |
92 |
|
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 |
– |
} |
93 |
|
} |
94 |
|
|
95 |
|
/** |
101 |
|
Throwable exception = null; |
102 |
|
try { |
103 |
|
onStart(); |
104 |
< |
pool.work(this); |
104 |
> |
pool.runWorker(this); |
105 |
|
} catch (Throwable ex) { |
106 |
|
exception = ex; |
107 |
|
} finally { |
373 |
– |
onTermination(exception); |
374 |
– |
} |
375 |
– |
} |
376 |
– |
|
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 |
– |
*/ |
403 |
– |
|
404 |
– |
/** |
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 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 |
– |
* 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 |
– |
private static final void writeSlot(ForkJoinTask<?>[] q, int i, |
419 |
– |
ForkJoinTask<?> t) { |
420 |
– |
UNSAFE.putObjectVolatile(q, (i << ASHIFT) + ABASE, t); |
421 |
– |
} |
422 |
– |
|
423 |
– |
// queue methods |
424 |
– |
|
425 |
– |
/** |
426 |
– |
* Pushes a task. Call only from this thread. |
427 |
– |
* |
428 |
– |
* @param t the task. Caller must ensure non-null. |
429 |
– |
*/ |
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 |
– |
* 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 size = oldQ != null ? oldQ.length << 1 : INITIAL_QUEUE_CAPACITY; |
451 |
– |
if (size > MAXIMUM_QUEUE_CAPACITY) |
452 |
– |
throw new RejectedExecutionException("Queue capacity exceeded"); |
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 |
– |
* 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 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 |
– |
* 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 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 |
– |
|
513 |
– |
/** |
514 |
– |
* Returns a popped task, or null if empty. |
515 |
– |
* Called only by this thread. |
516 |
– |
*/ |
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 |
– |
* Specialized version of popTask to pop only if topmost element |
538 |
– |
* is the given task. Called only by this thread. |
539 |
– |
* |
540 |
– |
* @param t the task. Caller must ensure non-null. |
541 |
– |
*/ |
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 |
– |
* Returns next task, or null if empty or contended. |
556 |
– |
*/ |
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 |
– |
* Runs the given task, plus any local tasks until queue is empty |
570 |
– |
*/ |
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 |
– |
/** |
585 |
– |
* Removes and cancels all tasks in queue. Can be called from any |
586 |
– |
* thread. |
587 |
– |
*/ |
588 |
– |
final void cancelTasks() { |
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(); |
599 |
– |
} |
600 |
– |
} |
601 |
– |
|
602 |
– |
/** |
603 |
– |
* Drains tasks to given collection c. |
604 |
– |
* |
605 |
– |
* @return the number of tasks drained |
606 |
– |
*/ |
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 |
– |
* Returns an estimate of the number of tasks in the queue. |
623 |
– |
*/ |
624 |
– |
final int getQueueSize() { |
625 |
– |
return queueTop - queueBase; |
626 |
– |
} |
627 |
– |
|
628 |
– |
/** |
629 |
– |
* Gets and removes a local task. |
630 |
– |
* |
631 |
– |
* @return a task, if available |
632 |
– |
*/ |
633 |
– |
final ForkJoinTask<?> pollLocalTask() { |
634 |
– |
return locallyFifo ? locallyDeqTask() : popTask(); |
635 |
– |
} |
636 |
– |
|
637 |
– |
/** |
638 |
– |
* Gets and removes a local or stolen task. |
639 |
– |
* |
640 |
– |
* @return a task, if available |
641 |
– |
*/ |
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 |
– |
i = 0; |
657 |
– |
} |
658 |
– |
} |
659 |
– |
return null; |
660 |
– |
} |
661 |
– |
|
662 |
– |
/** |
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 |
– |
private static final int MAX_HELP = 16; |
672 |
– |
|
673 |
– |
/** |
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 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 |
– |
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 |
– |
* 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 |
– |
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 |
– |
/** |
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 helped; |
800 |
– |
} |
801 |
– |
|
802 |
– |
/** |
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 |
– |
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 |
– |
* 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 |
– |
final int getEstimatedSurplusTaskCount() { |
887 |
– |
return queueTop - queueBase - pool.idlePerActive(); |
888 |
– |
} |
889 |
– |
|
890 |
– |
/** |
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 |
– |
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 |
– |
// 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 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 |
– |
private static sun.misc.Unsafe getUnsafe() { |
977 |
– |
try { |
978 |
– |
return sun.misc.Unsafe.getUnsafe(); |
979 |
– |
} catch (SecurityException se) { |
108 |
|
try { |
109 |
< |
return java.security.AccessController.doPrivileged |
110 |
< |
(new java.security |
111 |
< |
.PrivilegedExceptionAction<sun.misc.Unsafe>() { |
112 |
< |
public sun.misc.Unsafe run() throws Exception { |
113 |
< |
java.lang.reflect.Field f = sun.misc |
114 |
< |
.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()); |
109 |
> |
onTermination(exception); |
110 |
> |
} catch (Throwable ex) { |
111 |
> |
if (exception == null) |
112 |
> |
exception = ex; |
113 |
> |
} finally { |
114 |
> |
pool.deregisterWorker(this, exception); |
115 |
|
} |
116 |
|
} |
117 |
|
} |
118 |
|
} |
119 |
+ |
|