1 |
dl |
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 |
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.*; |
14 |
|
|
|
15 |
|
|
/** |
16 |
dl |
1.2 |
* A thread managed by a {@link ForkJoinPool}. This class is |
17 |
|
|
* subclassable solely for the sake of adding functionality -- there |
18 |
|
|
* are no overridable methods dealing with scheduling or |
19 |
|
|
* execution. However, you can override initialization and termination |
20 |
dl |
1.7 |
* methods surrounding the main task processing loop. If you do |
21 |
|
|
* create such a subclass, you will also need to supply a custom |
22 |
dl |
1.2 |
* ForkJoinWorkerThreadFactory to use it in a ForkJoinPool. |
23 |
jsr166 |
1.6 |
* |
24 |
jsr166 |
1.13 |
* @since 1.7 |
25 |
|
|
* @author Doug Lea |
26 |
dl |
1.1 |
*/ |
27 |
|
|
public class ForkJoinWorkerThread extends Thread { |
28 |
|
|
/* |
29 |
|
|
* Algorithm overview: |
30 |
|
|
* |
31 |
|
|
* 1. Work-Stealing: Work-stealing queues are special forms of |
32 |
|
|
* Deques that support only three of the four possible |
33 |
|
|
* end-operations -- push, pop, and deq (aka steal), and only do |
34 |
|
|
* so under the constraints that push and pop are called only from |
35 |
|
|
* the owning thread, while deq may be called from other threads. |
36 |
|
|
* (If you are unfamiliar with them, you probably want to read |
37 |
|
|
* Herlihy and Shavit's book "The Art of Multiprocessor |
38 |
|
|
* programming", chapter 16 describing these in more detail before |
39 |
|
|
* proceeding.) The main work-stealing queue design is roughly |
40 |
|
|
* similar to "Dynamic Circular Work-Stealing Deque" by David |
41 |
|
|
* Chase and Yossi Lev, SPAA 2005 |
42 |
|
|
* (http://research.sun.com/scalable/pubs/index.html). The main |
43 |
|
|
* difference ultimately stems from gc requirements that we null |
44 |
|
|
* out taken slots as soon as we can, to maintain as small a |
45 |
|
|
* footprint as possible even in programs generating huge numbers |
46 |
|
|
* of tasks. To accomplish this, we shift the CAS arbitrating pop |
47 |
|
|
* vs deq (steal) from being on the indices ("base" and "sp") to |
48 |
|
|
* the slots themselves (mainly via method "casSlotNull()"). So, |
49 |
jsr166 |
1.10 |
* both a successful pop and deq mainly entail CAS'ing a non-null |
50 |
dl |
1.1 |
* slot to null. Because we rely on CASes of references, we do |
51 |
|
|
* not need tag bits on base or sp. They are simple ints as used |
52 |
|
|
* in any circular array-based queue (see for example ArrayDeque). |
53 |
|
|
* Updates to the indices must still be ordered in a way that |
54 |
|
|
* guarantees that (sp - base) > 0 means the queue is empty, but |
55 |
|
|
* otherwise may err on the side of possibly making the queue |
56 |
|
|
* appear nonempty when a push, pop, or deq have not fully |
57 |
|
|
* committed. Note that this means that the deq operation, |
58 |
|
|
* considered individually, is not wait-free. One thief cannot |
59 |
|
|
* successfully continue until another in-progress one (or, if |
60 |
|
|
* previously empty, a push) completes. However, in the |
61 |
jsr166 |
1.9 |
* aggregate, we ensure at least probabilistic non-blockingness. If |
62 |
dl |
1.1 |
* an attempted steal fails, a thief always chooses a different |
63 |
|
|
* random victim target to try next. So, in order for one thief to |
64 |
|
|
* progress, it suffices for any in-progress deq or new push on |
65 |
|
|
* any empty queue to complete. One reason this works well here is |
66 |
|
|
* that apparently-nonempty often means soon-to-be-stealable, |
67 |
|
|
* which gives threads a chance to activate if necessary before |
68 |
|
|
* stealing (see below). |
69 |
|
|
* |
70 |
|
|
* Efficient implementation of this approach currently relies on |
71 |
|
|
* an uncomfortable amount of "Unsafe" mechanics. To maintain |
72 |
|
|
* correct orderings, reads and writes of variable base require |
73 |
|
|
* volatile ordering. Variable sp does not require volatile write |
74 |
|
|
* but needs cheaper store-ordering on writes. Because they are |
75 |
|
|
* protected by volatile base reads, reads of the queue array and |
76 |
|
|
* its slots do not need volatile load semantics, but writes (in |
77 |
|
|
* push) require store order and CASes (in pop and deq) require |
78 |
|
|
* (volatile) CAS semantics. Since these combinations aren't |
79 |
|
|
* supported using ordinary volatiles, the only way to accomplish |
80 |
jsr166 |
1.9 |
* these efficiently is to use direct Unsafe calls. (Using external |
81 |
dl |
1.1 |
* AtomicIntegers and AtomicReferenceArrays for the indices and |
82 |
|
|
* array is significantly slower because of memory locality and |
83 |
|
|
* indirection effects.) Further, performance on most platforms is |
84 |
|
|
* very sensitive to placement and sizing of the (resizable) queue |
85 |
|
|
* array. Even though these queues don't usually become all that |
86 |
|
|
* big, the initial size must be large enough to counteract cache |
87 |
|
|
* contention effects across multiple queues (especially in the |
88 |
|
|
* presence of GC cardmarking). Also, to improve thread-locality, |
89 |
|
|
* queues are currently initialized immediately after the thread |
90 |
|
|
* gets the initial signal to start processing tasks. However, |
91 |
|
|
* all queue-related methods except pushTask are written in a way |
92 |
|
|
* that allows them to instead be lazily allocated and/or disposed |
93 |
|
|
* of when empty. All together, these low-level implementation |
94 |
|
|
* choices produce as much as a factor of 4 performance |
95 |
|
|
* improvement compared to naive implementations, and enable the |
96 |
|
|
* processing of billions of tasks per second, sometimes at the |
97 |
|
|
* expense of ugliness. |
98 |
|
|
* |
99 |
|
|
* 2. Run control: The primary run control is based on a global |
100 |
|
|
* counter (activeCount) held by the pool. It uses an algorithm |
101 |
|
|
* similar to that in Herlihy and Shavit section 17.6 to cause |
102 |
|
|
* threads to eventually block when all threads declare they are |
103 |
|
|
* inactive. (See variable "scans".) For this to work, threads |
104 |
|
|
* must be declared active when executing tasks, and before |
105 |
|
|
* stealing a task. They must be inactive before blocking on the |
106 |
|
|
* Pool Barrier (awaiting a new submission or other Pool |
107 |
|
|
* event). In between, there is some free play which we take |
108 |
|
|
* advantage of to avoid contention and rapid flickering of the |
109 |
|
|
* global activeCount: If inactive, we activate only if a victim |
110 |
|
|
* queue appears to be nonempty (see above). Similarly, a thread |
111 |
|
|
* tries to inactivate only after a full scan of other threads. |
112 |
|
|
* The net effect is that contention on activeCount is rarely a |
113 |
|
|
* measurable performance issue. (There are also a few other cases |
114 |
|
|
* where we scan for work rather than retry/block upon |
115 |
|
|
* contention.) |
116 |
|
|
* |
117 |
|
|
* 3. Selection control. We maintain policy of always choosing to |
118 |
|
|
* run local tasks rather than stealing, and always trying to |
119 |
|
|
* steal tasks before trying to run a new submission. All steals |
120 |
|
|
* are currently performed in randomly-chosen deq-order. It may be |
121 |
|
|
* worthwhile to bias these with locality / anti-locality |
122 |
|
|
* information, but doing this well probably requires more |
123 |
|
|
* lower-level information from JVMs than currently provided. |
124 |
|
|
*/ |
125 |
|
|
|
126 |
|
|
/** |
127 |
|
|
* Capacity of work-stealing queue array upon initialization. |
128 |
|
|
* Must be a power of two. Initial size must be at least 2, but is |
129 |
|
|
* padded to minimize cache effects. |
130 |
|
|
*/ |
131 |
|
|
private static final int INITIAL_QUEUE_CAPACITY = 1 << 13; |
132 |
|
|
|
133 |
|
|
/** |
134 |
|
|
* Maximum work-stealing queue array size. Must be less than or |
135 |
dl |
1.5 |
* equal to 1 << 28 to ensure lack of index wraparound. (This |
136 |
|
|
* is less than usual bounds, because we need leftshift by 3 |
137 |
|
|
* to be in int range). |
138 |
dl |
1.1 |
*/ |
139 |
dl |
1.5 |
private static final int MAXIMUM_QUEUE_CAPACITY = 1 << 28; |
140 |
dl |
1.1 |
|
141 |
|
|
/** |
142 |
dl |
1.5 |
* The pool this thread works in. Accessed directly by ForkJoinTask |
143 |
dl |
1.1 |
*/ |
144 |
dl |
1.5 |
final ForkJoinPool pool; |
145 |
dl |
1.1 |
|
146 |
|
|
/** |
147 |
|
|
* The work-stealing queue array. Size must be a power of two. |
148 |
dl |
1.5 |
* Initialized when thread starts, to improve memory locality. |
149 |
dl |
1.1 |
*/ |
150 |
|
|
private ForkJoinTask<?>[] queue; |
151 |
|
|
|
152 |
|
|
/** |
153 |
|
|
* Index (mod queue.length) of next queue slot to push to or pop |
154 |
|
|
* from. It is written only by owner thread, via ordered store. |
155 |
|
|
* Both sp and base are allowed to wrap around on overflow, but |
156 |
|
|
* (sp - base) still estimates size. |
157 |
|
|
*/ |
158 |
|
|
private volatile int sp; |
159 |
|
|
|
160 |
|
|
/** |
161 |
|
|
* Index (mod queue.length) of least valid queue slot, which is |
162 |
|
|
* always the next position to steal from if nonempty. |
163 |
|
|
*/ |
164 |
|
|
private volatile int base; |
165 |
|
|
|
166 |
|
|
/** |
167 |
dl |
1.5 |
* Activity status. When true, this worker is considered active. |
168 |
|
|
* Must be false upon construction. It must be true when executing |
169 |
|
|
* tasks, and BEFORE stealing a task. It must be false before |
170 |
|
|
* calling pool.sync |
171 |
dl |
1.1 |
*/ |
172 |
dl |
1.5 |
private boolean active; |
173 |
dl |
1.1 |
|
174 |
|
|
/** |
175 |
|
|
* Run state of this worker. Supports simple versions of the usual |
176 |
|
|
* shutdown/shutdownNow control. |
177 |
|
|
*/ |
178 |
|
|
private volatile int runState; |
179 |
|
|
|
180 |
|
|
/** |
181 |
dl |
1.5 |
* Seed for random number generator for choosing steal victims. |
182 |
|
|
* Uses Marsaglia xorshift. Must be nonzero upon initialization. |
183 |
dl |
1.1 |
*/ |
184 |
dl |
1.5 |
private int seed; |
185 |
dl |
1.1 |
|
186 |
|
|
/** |
187 |
|
|
* Number of steals, transferred to pool when idle |
188 |
|
|
*/ |
189 |
|
|
private int stealCount; |
190 |
|
|
|
191 |
|
|
/** |
192 |
dl |
1.5 |
* Index of this worker in pool array. Set once by pool before |
193 |
|
|
* running, and accessed directly by pool during cleanup etc |
194 |
dl |
1.1 |
*/ |
195 |
dl |
1.5 |
int poolIndex; |
196 |
dl |
1.1 |
|
197 |
|
|
/** |
198 |
dl |
1.5 |
* The last barrier event waited for. Accessed in pool callback |
199 |
|
|
* methods, but only by current thread. |
200 |
dl |
1.1 |
*/ |
201 |
dl |
1.5 |
long lastEventCount; |
202 |
dl |
1.1 |
|
203 |
|
|
/** |
204 |
dl |
1.7 |
* True if use local fifo, not default lifo, for local polling |
205 |
|
|
*/ |
206 |
|
|
private boolean locallyFifo; |
207 |
|
|
|
208 |
|
|
/** |
209 |
dl |
1.1 |
* Creates a ForkJoinWorkerThread operating in the given pool. |
210 |
jsr166 |
1.11 |
* |
211 |
dl |
1.1 |
* @param pool the pool this thread works in |
212 |
|
|
* @throws NullPointerException if pool is null |
213 |
|
|
*/ |
214 |
|
|
protected ForkJoinWorkerThread(ForkJoinPool pool) { |
215 |
|
|
if (pool == null) throw new NullPointerException(); |
216 |
|
|
this.pool = pool; |
217 |
dl |
1.5 |
// Note: poolIndex is set by pool during construction |
218 |
|
|
// Remaining initialization is deferred to onStart |
219 |
dl |
1.1 |
} |
220 |
|
|
|
221 |
jsr166 |
1.6 |
// Public access methods |
222 |
dl |
1.2 |
|
223 |
|
|
/** |
224 |
jsr166 |
1.11 |
* Returns the pool hosting this thread. |
225 |
|
|
* |
226 |
dl |
1.2 |
* @return the pool |
227 |
|
|
*/ |
228 |
dl |
1.4 |
public ForkJoinPool getPool() { |
229 |
|
|
return pool; |
230 |
dl |
1.2 |
} |
231 |
|
|
|
232 |
|
|
/** |
233 |
dl |
1.4 |
* Returns the index number of this thread in its pool. The |
234 |
|
|
* returned value ranges from zero to the maximum number of |
235 |
|
|
* threads (minus one) that have ever been created in the pool. |
236 |
|
|
* This method may be useful for applications that track status or |
237 |
dl |
1.5 |
* collect results per-worker rather than per-task. |
238 |
jsr166 |
1.11 |
* |
239 |
|
|
* @return the index number |
240 |
dl |
1.2 |
*/ |
241 |
dl |
1.4 |
public int getPoolIndex() { |
242 |
|
|
return poolIndex; |
243 |
dl |
1.2 |
} |
244 |
|
|
|
245 |
dl |
1.7 |
/** |
246 |
|
|
* Establishes local first-in-first-out scheduling mode for forked |
247 |
jsr166 |
1.8 |
* tasks that are never joined. |
248 |
jsr166 |
1.11 |
* |
249 |
dl |
1.7 |
* @param async if true, use locally FIFO scheduling |
250 |
|
|
*/ |
251 |
|
|
void setAsyncMode(boolean async) { |
252 |
|
|
locallyFifo = async; |
253 |
|
|
} |
254 |
dl |
1.5 |
|
255 |
|
|
// Runstate management |
256 |
|
|
|
257 |
|
|
// Runstate values. Order matters |
258 |
|
|
private static final int RUNNING = 0; |
259 |
|
|
private static final int SHUTDOWN = 1; |
260 |
|
|
private static final int TERMINATING = 2; |
261 |
|
|
private static final int TERMINATED = 3; |
262 |
|
|
|
263 |
|
|
final boolean isShutdown() { return runState >= SHUTDOWN; } |
264 |
|
|
final boolean isTerminating() { return runState >= TERMINATING; } |
265 |
|
|
final boolean isTerminated() { return runState == TERMINATED; } |
266 |
|
|
final boolean shutdown() { return transitionRunStateTo(SHUTDOWN); } |
267 |
|
|
final boolean shutdownNow() { return transitionRunStateTo(TERMINATING); } |
268 |
|
|
|
269 |
|
|
/** |
270 |
jsr166 |
1.11 |
* Transitions to at least the given state. Returns true if not |
271 |
|
|
* already at least at given state. |
272 |
dl |
1.5 |
*/ |
273 |
|
|
private boolean transitionRunStateTo(int state) { |
274 |
|
|
for (;;) { |
275 |
|
|
int s = runState; |
276 |
|
|
if (s >= state) |
277 |
|
|
return false; |
278 |
jsr166 |
1.12 |
if (UNSAFE.compareAndSwapInt(this, runStateOffset, s, state)) |
279 |
dl |
1.5 |
return true; |
280 |
|
|
} |
281 |
|
|
} |
282 |
|
|
|
283 |
|
|
/** |
284 |
jsr166 |
1.11 |
* Tries to set status to active; fails on contention. |
285 |
dl |
1.5 |
*/ |
286 |
|
|
private boolean tryActivate() { |
287 |
|
|
if (!active) { |
288 |
|
|
if (!pool.tryIncrementActiveCount()) |
289 |
|
|
return false; |
290 |
|
|
active = true; |
291 |
|
|
} |
292 |
|
|
return true; |
293 |
|
|
} |
294 |
|
|
|
295 |
|
|
/** |
296 |
jsr166 |
1.11 |
* Tries to set status to active; fails on contention. |
297 |
dl |
1.5 |
*/ |
298 |
|
|
private boolean tryInactivate() { |
299 |
|
|
if (active) { |
300 |
|
|
if (!pool.tryDecrementActiveCount()) |
301 |
|
|
return false; |
302 |
|
|
active = false; |
303 |
|
|
} |
304 |
|
|
return true; |
305 |
|
|
} |
306 |
|
|
|
307 |
|
|
/** |
308 |
jsr166 |
1.11 |
* Computes next value for random victim probe. Scans don't |
309 |
dl |
1.5 |
* require a very high quality generator, but also not a crummy |
310 |
jsr166 |
1.11 |
* one. Marsaglia xor-shift is cheap and works well. |
311 |
dl |
1.5 |
*/ |
312 |
|
|
private static int xorShift(int r) { |
313 |
|
|
r ^= r << 1; |
314 |
|
|
r ^= r >>> 3; |
315 |
|
|
r ^= r << 10; |
316 |
|
|
return r; |
317 |
|
|
} |
318 |
|
|
|
319 |
|
|
// Lifecycle methods |
320 |
dl |
1.1 |
|
321 |
|
|
/** |
322 |
dl |
1.5 |
* This method is required to be public, but should never be |
323 |
|
|
* called explicitly. It performs the main run loop to execute |
324 |
|
|
* ForkJoinTasks. |
325 |
dl |
1.1 |
*/ |
326 |
dl |
1.5 |
public void run() { |
327 |
|
|
Throwable exception = null; |
328 |
|
|
try { |
329 |
|
|
onStart(); |
330 |
|
|
pool.sync(this); // await first pool event |
331 |
|
|
mainLoop(); |
332 |
|
|
} catch (Throwable ex) { |
333 |
|
|
exception = ex; |
334 |
|
|
} finally { |
335 |
|
|
onTermination(exception); |
336 |
|
|
} |
337 |
dl |
1.1 |
} |
338 |
|
|
|
339 |
|
|
/** |
340 |
jsr166 |
1.11 |
* Executes tasks until shut down. |
341 |
dl |
1.1 |
*/ |
342 |
dl |
1.5 |
private void mainLoop() { |
343 |
|
|
while (!isShutdown()) { |
344 |
|
|
ForkJoinTask<?> t = pollTask(); |
345 |
jsr166 |
1.6 |
if (t != null || (t = pollSubmission()) != null) |
346 |
dl |
1.5 |
t.quietlyExec(); |
347 |
|
|
else if (tryInactivate()) |
348 |
|
|
pool.sync(this); |
349 |
|
|
} |
350 |
dl |
1.1 |
} |
351 |
|
|
|
352 |
dl |
1.5 |
/** |
353 |
|
|
* Initializes internal state after construction but before |
354 |
|
|
* processing any tasks. If you override this method, you must |
355 |
|
|
* invoke super.onStart() at the beginning of the method. |
356 |
|
|
* Initialization requires care: Most fields must have legal |
357 |
|
|
* default values, to ensure that attempted accesses from other |
358 |
|
|
* threads work correctly even before this thread starts |
359 |
|
|
* processing tasks. |
360 |
|
|
*/ |
361 |
|
|
protected void onStart() { |
362 |
|
|
// Allocate while starting to improve chances of thread-local |
363 |
|
|
// isolation |
364 |
|
|
queue = new ForkJoinTask<?>[INITIAL_QUEUE_CAPACITY]; |
365 |
|
|
// Initial value of seed need not be especially random but |
366 |
|
|
// should differ across workers and must be nonzero |
367 |
|
|
int p = poolIndex + 1; |
368 |
|
|
seed = p + (p << 8) + (p << 16) + (p << 24); // spread bits |
369 |
|
|
} |
370 |
dl |
1.1 |
|
371 |
|
|
/** |
372 |
jsr166 |
1.11 |
* Performs cleanup associated with termination of this worker |
373 |
dl |
1.5 |
* thread. If you override this method, you must invoke |
374 |
|
|
* super.onTermination at the end of the overridden method. |
375 |
|
|
* |
376 |
|
|
* @param exception the exception causing this thread to abort due |
377 |
jsr166 |
1.11 |
* to an unrecoverable error, or null if completed normally |
378 |
dl |
1.1 |
*/ |
379 |
dl |
1.5 |
protected void onTermination(Throwable exception) { |
380 |
|
|
// Execute remaining local tasks unless aborting or terminating |
381 |
|
|
while (exception == null && !pool.isTerminating() && base != sp) { |
382 |
|
|
try { |
383 |
|
|
ForkJoinTask<?> t = popTask(); |
384 |
|
|
if (t != null) |
385 |
|
|
t.quietlyExec(); |
386 |
|
|
} catch(Throwable ex) { |
387 |
|
|
exception = ex; |
388 |
|
|
} |
389 |
|
|
} |
390 |
|
|
// Cancel other tasks, transition status, notify pool, and |
391 |
|
|
// propagate exception to uncaught exception handler |
392 |
|
|
try { |
393 |
|
|
do;while (!tryInactivate()); // ensure inactive |
394 |
jsr166 |
1.6 |
cancelTasks(); |
395 |
dl |
1.5 |
runState = TERMINATED; |
396 |
|
|
pool.workerTerminated(this); |
397 |
|
|
} catch (Throwable ex) { // Shouldn't ever happen |
398 |
|
|
if (exception == null) // but if so, at least rethrown |
399 |
|
|
exception = ex; |
400 |
|
|
} finally { |
401 |
|
|
if (exception != null) |
402 |
|
|
ForkJoinTask.rethrowException(exception); |
403 |
|
|
} |
404 |
dl |
1.1 |
} |
405 |
|
|
|
406 |
jsr166 |
1.6 |
// Intrinsics-based support for queue operations. |
407 |
dl |
1.5 |
|
408 |
dl |
1.1 |
/** |
409 |
jsr166 |
1.10 |
* Adds in store-order the given task at given slot of q to null. |
410 |
|
|
* Caller must ensure q is non-null and index is in range. |
411 |
dl |
1.1 |
*/ |
412 |
|
|
private static void setSlot(ForkJoinTask<?>[] q, int i, |
413 |
jsr166 |
1.14 |
ForkJoinTask<?> t) { |
414 |
jsr166 |
1.12 |
UNSAFE.putOrderedObject(q, (i << qShift) + qBase, t); |
415 |
dl |
1.1 |
} |
416 |
|
|
|
417 |
|
|
/** |
418 |
jsr166 |
1.10 |
* CAS given slot of q to null. Caller must ensure q is non-null |
419 |
dl |
1.1 |
* and index is in range. |
420 |
|
|
*/ |
421 |
|
|
private static boolean casSlotNull(ForkJoinTask<?>[] q, int i, |
422 |
|
|
ForkJoinTask<?> t) { |
423 |
jsr166 |
1.12 |
return UNSAFE.compareAndSwapObject(q, (i << qShift) + qBase, t, null); |
424 |
dl |
1.1 |
} |
425 |
|
|
|
426 |
dl |
1.5 |
/** |
427 |
|
|
* Sets sp in store-order. |
428 |
|
|
*/ |
429 |
|
|
private void storeSp(int s) { |
430 |
jsr166 |
1.12 |
UNSAFE.putOrderedInt(this, spOffset, s); |
431 |
dl |
1.5 |
} |
432 |
|
|
|
433 |
dl |
1.1 |
// Main queue methods |
434 |
|
|
|
435 |
|
|
/** |
436 |
|
|
* Pushes a task. Called only by current thread. |
437 |
jsr166 |
1.11 |
* |
438 |
jsr166 |
1.10 |
* @param t the task. Caller must ensure non-null. |
439 |
dl |
1.1 |
*/ |
440 |
|
|
final void pushTask(ForkJoinTask<?> t) { |
441 |
|
|
ForkJoinTask<?>[] q = queue; |
442 |
|
|
int mask = q.length - 1; |
443 |
|
|
int s = sp; |
444 |
dl |
1.5 |
setSlot(q, s & mask, t); |
445 |
|
|
storeSp(++s); |
446 |
dl |
1.1 |
if ((s -= base) == 1) |
447 |
dl |
1.5 |
pool.signalWork(); |
448 |
dl |
1.1 |
else if (s >= mask) |
449 |
|
|
growQueue(); |
450 |
|
|
} |
451 |
|
|
|
452 |
|
|
/** |
453 |
|
|
* Tries to take a task from the base of the queue, failing if |
454 |
|
|
* either empty or contended. |
455 |
jsr166 |
1.11 |
* |
456 |
|
|
* @return a task, or null if none or contended |
457 |
dl |
1.1 |
*/ |
458 |
dl |
1.7 |
final ForkJoinTask<?> deqTask() { |
459 |
dl |
1.5 |
ForkJoinTask<?> t; |
460 |
dl |
1.1 |
ForkJoinTask<?>[] q; |
461 |
|
|
int i; |
462 |
|
|
int b; |
463 |
|
|
if (sp != (b = base) && |
464 |
|
|
(q = queue) != null && // must read q after b |
465 |
|
|
(t = q[i = (q.length - 1) & b]) != null && |
466 |
dl |
1.5 |
casSlotNull(q, i, t)) { |
467 |
dl |
1.1 |
base = b + 1; |
468 |
|
|
return t; |
469 |
|
|
} |
470 |
|
|
return null; |
471 |
|
|
} |
472 |
|
|
|
473 |
|
|
/** |
474 |
dl |
1.5 |
* Returns a popped task, or null if empty. Ensures active status |
475 |
jsr166 |
1.10 |
* if non-null. Called only by current thread. |
476 |
dl |
1.1 |
*/ |
477 |
|
|
final ForkJoinTask<?> popTask() { |
478 |
|
|
int s = sp; |
479 |
dl |
1.5 |
while (s != base) { |
480 |
|
|
if (tryActivate()) { |
481 |
|
|
ForkJoinTask<?>[] q = queue; |
482 |
|
|
int mask = q.length - 1; |
483 |
|
|
int i = (s - 1) & mask; |
484 |
|
|
ForkJoinTask<?> t = q[i]; |
485 |
|
|
if (t == null || !casSlotNull(q, i, t)) |
486 |
|
|
break; |
487 |
|
|
storeSp(s - 1); |
488 |
|
|
return t; |
489 |
|
|
} |
490 |
dl |
1.1 |
} |
491 |
|
|
return null; |
492 |
|
|
} |
493 |
|
|
|
494 |
|
|
/** |
495 |
|
|
* Specialized version of popTask to pop only if |
496 |
|
|
* topmost element is the given task. Called only |
497 |
dl |
1.5 |
* by current thread while active. |
498 |
jsr166 |
1.11 |
* |
499 |
|
|
* @param t the task. Caller must ensure non-null. |
500 |
dl |
1.1 |
*/ |
501 |
|
|
final boolean unpushTask(ForkJoinTask<?> t) { |
502 |
|
|
ForkJoinTask<?>[] q = queue; |
503 |
|
|
int mask = q.length - 1; |
504 |
|
|
int s = sp - 1; |
505 |
dl |
1.5 |
if (casSlotNull(q, s & mask, t)) { |
506 |
|
|
storeSp(s); |
507 |
dl |
1.1 |
return true; |
508 |
|
|
} |
509 |
|
|
return false; |
510 |
|
|
} |
511 |
|
|
|
512 |
|
|
/** |
513 |
dl |
1.7 |
* Returns next task. |
514 |
dl |
1.1 |
*/ |
515 |
dl |
1.2 |
final ForkJoinTask<?> peekTask() { |
516 |
dl |
1.1 |
ForkJoinTask<?>[] q = queue; |
517 |
dl |
1.7 |
if (q == null) |
518 |
|
|
return null; |
519 |
|
|
int mask = q.length - 1; |
520 |
|
|
int i = locallyFifo? base : (sp - 1); |
521 |
|
|
return q[i & mask]; |
522 |
dl |
1.1 |
} |
523 |
|
|
|
524 |
|
|
/** |
525 |
|
|
* Doubles queue array size. Transfers elements by emulating |
526 |
|
|
* steals (deqs) from old array and placing, oldest first, into |
527 |
|
|
* new array. |
528 |
|
|
*/ |
529 |
|
|
private void growQueue() { |
530 |
|
|
ForkJoinTask<?>[] oldQ = queue; |
531 |
|
|
int oldSize = oldQ.length; |
532 |
|
|
int newSize = oldSize << 1; |
533 |
|
|
if (newSize > MAXIMUM_QUEUE_CAPACITY) |
534 |
|
|
throw new RejectedExecutionException("Queue capacity exceeded"); |
535 |
|
|
ForkJoinTask<?>[] newQ = queue = new ForkJoinTask<?>[newSize]; |
536 |
|
|
|
537 |
|
|
int b = base; |
538 |
|
|
int bf = b + oldSize; |
539 |
|
|
int oldMask = oldSize - 1; |
540 |
|
|
int newMask = newSize - 1; |
541 |
|
|
do { |
542 |
|
|
int oldIndex = b & oldMask; |
543 |
|
|
ForkJoinTask<?> t = oldQ[oldIndex]; |
544 |
|
|
if (t != null && !casSlotNull(oldQ, oldIndex, t)) |
545 |
|
|
t = null; |
546 |
|
|
setSlot(newQ, b & newMask, t); |
547 |
|
|
} while (++b != bf); |
548 |
dl |
1.5 |
pool.signalWork(); |
549 |
dl |
1.1 |
} |
550 |
|
|
|
551 |
|
|
/** |
552 |
dl |
1.5 |
* Tries to steal a task from another worker. Starts at a random |
553 |
|
|
* index of workers array, and probes workers until finding one |
554 |
|
|
* with non-empty queue or finding that all are empty. It |
555 |
|
|
* randomly selects the first n probes. If these are empty, it |
556 |
|
|
* resorts to a full circular traversal, which is necessary to |
557 |
|
|
* accurately set active status by caller. Also restarts if pool |
558 |
|
|
* events occurred since last scan, which forces refresh of |
559 |
|
|
* workers array, in case barrier was associated with resize. |
560 |
dl |
1.1 |
* |
561 |
|
|
* This method must be both fast and quiet -- usually avoiding |
562 |
|
|
* memory accesses that could disrupt cache sharing etc other than |
563 |
|
|
* those needed to check for and take tasks. This accounts for, |
564 |
|
|
* among other things, updating random seed in place without |
565 |
dl |
1.5 |
* storing it until exit. |
566 |
dl |
1.1 |
* |
567 |
|
|
* @return a task, or null if none found |
568 |
|
|
*/ |
569 |
dl |
1.5 |
private ForkJoinTask<?> scan() { |
570 |
|
|
ForkJoinTask<?> t = null; |
571 |
|
|
int r = seed; // extract once to keep scan quiet |
572 |
|
|
ForkJoinWorkerThread[] ws; // refreshed on outer loop |
573 |
|
|
int mask; // must be power 2 minus 1 and > 0 |
574 |
|
|
outer:do { |
575 |
|
|
if ((ws = pool.workers) != null && (mask = ws.length - 1) > 0) { |
576 |
dl |
1.1 |
int idx = r; |
577 |
dl |
1.5 |
int probes = ~mask; // use random index while negative |
578 |
dl |
1.1 |
for (;;) { |
579 |
dl |
1.5 |
r = xorShift(r); // update random seed |
580 |
|
|
ForkJoinWorkerThread v = ws[mask & idx]; |
581 |
|
|
if (v == null || v.sp == v.base) { |
582 |
|
|
if (probes <= mask) |
583 |
|
|
idx = (probes++ < 0)? r : (idx + 1); |
584 |
|
|
else |
585 |
|
|
break; |
586 |
dl |
1.1 |
} |
587 |
dl |
1.5 |
else if (!tryActivate() || (t = v.deqTask()) == null) |
588 |
|
|
continue outer; // restart on contention |
589 |
dl |
1.1 |
else |
590 |
dl |
1.5 |
break outer; |
591 |
dl |
1.1 |
} |
592 |
|
|
} |
593 |
dl |
1.5 |
} while (pool.hasNewSyncEvent(this)); // retry on pool events |
594 |
|
|
seed = r; |
595 |
|
|
return t; |
596 |
dl |
1.1 |
} |
597 |
|
|
|
598 |
|
|
/** |
599 |
jsr166 |
1.11 |
* Gets and removes a local or stolen task. |
600 |
|
|
* |
601 |
dl |
1.5 |
* @return a task, if available |
602 |
|
|
*/ |
603 |
|
|
final ForkJoinTask<?> pollTask() { |
604 |
dl |
1.7 |
ForkJoinTask<?> t = locallyFifo? deqTask() : popTask(); |
605 |
dl |
1.5 |
if (t == null && (t = scan()) != null) |
606 |
|
|
++stealCount; |
607 |
|
|
return t; |
608 |
dl |
1.1 |
} |
609 |
|
|
|
610 |
|
|
/** |
611 |
jsr166 |
1.11 |
* Gets a local task. |
612 |
|
|
* |
613 |
dl |
1.7 |
* @return a task, if available |
614 |
|
|
*/ |
615 |
|
|
final ForkJoinTask<?> pollLocalTask() { |
616 |
|
|
return locallyFifo? deqTask() : popTask(); |
617 |
|
|
} |
618 |
|
|
|
619 |
|
|
/** |
620 |
dl |
1.5 |
* Returns a pool submission, if one exists, activating first. |
621 |
jsr166 |
1.11 |
* |
622 |
dl |
1.5 |
* @return a submission, if available |
623 |
dl |
1.1 |
*/ |
624 |
dl |
1.5 |
private ForkJoinTask<?> pollSubmission() { |
625 |
|
|
ForkJoinPool p = pool; |
626 |
|
|
while (p.hasQueuedSubmissions()) { |
627 |
|
|
ForkJoinTask<?> t; |
628 |
|
|
if (tryActivate() && (t = p.pollSubmission()) != null) |
629 |
|
|
return t; |
630 |
dl |
1.1 |
} |
631 |
dl |
1.5 |
return null; |
632 |
dl |
1.1 |
} |
633 |
dl |
1.5 |
|
634 |
|
|
// Methods accessed only by Pool |
635 |
|
|
|
636 |
dl |
1.1 |
/** |
637 |
dl |
1.5 |
* Removes and cancels all tasks in queue. Can be called from any |
638 |
|
|
* thread. |
639 |
dl |
1.1 |
*/ |
640 |
dl |
1.5 |
final void cancelTasks() { |
641 |
dl |
1.3 |
ForkJoinTask<?> t; |
642 |
dl |
1.5 |
while (base != sp && (t = deqTask()) != null) |
643 |
|
|
t.cancelIgnoringExceptions(); |
644 |
dl |
1.1 |
} |
645 |
|
|
|
646 |
|
|
/** |
647 |
jsr166 |
1.11 |
* Drains tasks to given collection c. |
648 |
|
|
* |
649 |
dl |
1.7 |
* @return the number of tasks drained |
650 |
|
|
*/ |
651 |
|
|
final int drainTasksTo(Collection<ForkJoinTask<?>> c) { |
652 |
|
|
int n = 0; |
653 |
|
|
ForkJoinTask<?> t; |
654 |
|
|
while (base != sp && (t = deqTask()) != null) { |
655 |
|
|
c.add(t); |
656 |
|
|
++n; |
657 |
|
|
} |
658 |
|
|
return n; |
659 |
|
|
} |
660 |
|
|
|
661 |
|
|
/** |
662 |
jsr166 |
1.11 |
* Gets and clears steal count for accumulation by pool. Called |
663 |
dl |
1.5 |
* only when known to be idle (in pool.sync and termination). |
664 |
dl |
1.1 |
*/ |
665 |
dl |
1.5 |
final int getAndClearStealCount() { |
666 |
|
|
int sc = stealCount; |
667 |
|
|
stealCount = 0; |
668 |
|
|
return sc; |
669 |
|
|
} |
670 |
|
|
|
671 |
|
|
/** |
672 |
|
|
* Returns true if at least one worker in the given array appears |
673 |
|
|
* to have at least one queued task. |
674 |
|
|
* @param ws array of workers |
675 |
|
|
*/ |
676 |
|
|
static boolean hasQueuedTasks(ForkJoinWorkerThread[] ws) { |
677 |
|
|
if (ws != null) { |
678 |
|
|
int len = ws.length; |
679 |
|
|
for (int j = 0; j < 2; ++j) { // need two passes for clean sweep |
680 |
|
|
for (int i = 0; i < len; ++i) { |
681 |
|
|
ForkJoinWorkerThread w = ws[i]; |
682 |
|
|
if (w != null && w.sp != w.base) |
683 |
|
|
return true; |
684 |
dl |
1.1 |
} |
685 |
|
|
} |
686 |
|
|
} |
687 |
dl |
1.5 |
return false; |
688 |
dl |
1.1 |
} |
689 |
|
|
|
690 |
dl |
1.5 |
// Support methods for ForkJoinTask |
691 |
|
|
|
692 |
dl |
1.1 |
/** |
693 |
dl |
1.2 |
* Returns an estimate of the number of tasks in the queue. |
694 |
|
|
*/ |
695 |
|
|
final int getQueueSize() { |
696 |
dl |
1.3 |
int n = sp - base; |
697 |
dl |
1.5 |
return n < 0? 0 : n; // suppress momentarily negative values |
698 |
dl |
1.2 |
} |
699 |
|
|
|
700 |
|
|
/** |
701 |
dl |
1.1 |
* Returns an estimate of the number of tasks, offset by a |
702 |
|
|
* function of number of idle workers. |
703 |
|
|
*/ |
704 |
|
|
final int getEstimatedSurplusTaskCount() { |
705 |
dl |
1.3 |
// The halving approximates weighting idle vs non-idle workers |
706 |
dl |
1.1 |
return (sp - base) - (pool.getIdleThreadCount() >>> 1); |
707 |
|
|
} |
708 |
|
|
|
709 |
|
|
/** |
710 |
jsr166 |
1.11 |
* Scans, returning early if joinMe done |
711 |
dl |
1.1 |
*/ |
712 |
dl |
1.5 |
final ForkJoinTask<?> scanWhileJoining(ForkJoinTask<?> joinMe) { |
713 |
|
|
ForkJoinTask<?> t = pollTask(); |
714 |
|
|
if (t != null && joinMe.status < 0 && sp == base) { |
715 |
|
|
pushTask(t); // unsteal if done and this task would be stealable |
716 |
|
|
t = null; |
717 |
|
|
} |
718 |
|
|
return t; |
719 |
dl |
1.1 |
} |
720 |
jsr166 |
1.6 |
|
721 |
dl |
1.1 |
/** |
722 |
jsr166 |
1.11 |
* Runs tasks until pool isQuiescent. |
723 |
dl |
1.1 |
*/ |
724 |
dl |
1.5 |
final void helpQuiescePool() { |
725 |
|
|
for (;;) { |
726 |
|
|
ForkJoinTask<?> t = pollTask(); |
727 |
jsr166 |
1.6 |
if (t != null) |
728 |
dl |
1.5 |
t.quietlyExec(); |
729 |
|
|
else if (tryInactivate() && pool.isQuiescent()) |
730 |
|
|
break; |
731 |
|
|
} |
732 |
|
|
do;while (!tryActivate()); // re-activate on exit |
733 |
dl |
1.1 |
} |
734 |
|
|
|
735 |
|
|
// Temporary Unsafe mechanics for preliminary release |
736 |
jsr166 |
1.6 |
private static Unsafe getUnsafe() throws Throwable { |
737 |
|
|
try { |
738 |
|
|
return Unsafe.getUnsafe(); |
739 |
|
|
} catch (SecurityException se) { |
740 |
|
|
try { |
741 |
|
|
return java.security.AccessController.doPrivileged |
742 |
|
|
(new java.security.PrivilegedExceptionAction<Unsafe>() { |
743 |
|
|
public Unsafe run() throws Exception { |
744 |
|
|
return getUnsafePrivileged(); |
745 |
|
|
}}); |
746 |
|
|
} catch (java.security.PrivilegedActionException e) { |
747 |
|
|
throw e.getCause(); |
748 |
|
|
} |
749 |
|
|
} |
750 |
|
|
} |
751 |
|
|
|
752 |
|
|
private static Unsafe getUnsafePrivileged() |
753 |
|
|
throws NoSuchFieldException, IllegalAccessException { |
754 |
|
|
Field f = Unsafe.class.getDeclaredField("theUnsafe"); |
755 |
|
|
f.setAccessible(true); |
756 |
|
|
return (Unsafe) f.get(null); |
757 |
|
|
} |
758 |
|
|
|
759 |
|
|
private static long fieldOffset(String fieldName) |
760 |
|
|
throws NoSuchFieldException { |
761 |
jsr166 |
1.12 |
return UNSAFE.objectFieldOffset |
762 |
jsr166 |
1.6 |
(ForkJoinWorkerThread.class.getDeclaredField(fieldName)); |
763 |
|
|
} |
764 |
dl |
1.1 |
|
765 |
jsr166 |
1.12 |
static final Unsafe UNSAFE; |
766 |
dl |
1.1 |
static final long baseOffset; |
767 |
|
|
static final long spOffset; |
768 |
dl |
1.5 |
static final long runStateOffset; |
769 |
dl |
1.1 |
static final long qBase; |
770 |
|
|
static final int qShift; |
771 |
|
|
static { |
772 |
|
|
try { |
773 |
jsr166 |
1.12 |
UNSAFE = getUnsafe(); |
774 |
jsr166 |
1.6 |
baseOffset = fieldOffset("base"); |
775 |
|
|
spOffset = fieldOffset("sp"); |
776 |
|
|
runStateOffset = fieldOffset("runState"); |
777 |
jsr166 |
1.12 |
qBase = UNSAFE.arrayBaseOffset(ForkJoinTask[].class); |
778 |
|
|
int s = UNSAFE.arrayIndexScale(ForkJoinTask[].class); |
779 |
dl |
1.1 |
if ((s & (s-1)) != 0) |
780 |
|
|
throw new Error("data type scale not a power of two"); |
781 |
|
|
qShift = 31 - Integer.numberOfLeadingZeros(s); |
782 |
jsr166 |
1.6 |
} catch (Throwable e) { |
783 |
dl |
1.1 |
throw new RuntimeException("Could not initialize intrinsics", e); |
784 |
|
|
} |
785 |
|
|
} |
786 |
|
|
} |