6 |
|
|
7 |
|
package jsr166y; |
8 |
|
|
9 |
– |
import java.util.concurrent.*; |
10 |
– |
|
9 |
|
import java.util.ArrayList; |
10 |
|
import java.util.Arrays; |
11 |
|
import java.util.Collection; |
12 |
|
import java.util.Collections; |
13 |
|
import java.util.List; |
14 |
+ |
import java.util.concurrent.AbstractExecutorService; |
15 |
+ |
import java.util.concurrent.Callable; |
16 |
+ |
import java.util.concurrent.CountDownLatch; |
17 |
+ |
import java.util.concurrent.ExecutorService; |
18 |
+ |
import java.util.concurrent.Future; |
19 |
+ |
import java.util.concurrent.RejectedExecutionException; |
20 |
+ |
import java.util.concurrent.RunnableFuture; |
21 |
+ |
import java.util.concurrent.TimeUnit; |
22 |
+ |
import java.util.concurrent.TimeoutException; |
23 |
+ |
import java.util.concurrent.atomic.AtomicInteger; |
24 |
|
import java.util.concurrent.locks.LockSupport; |
25 |
|
import java.util.concurrent.locks.ReentrantLock; |
18 |
– |
import java.util.concurrent.atomic.AtomicInteger; |
19 |
– |
import java.util.concurrent.CountDownLatch; |
26 |
|
|
27 |
|
/** |
28 |
|
* An {@link ExecutorService} for running {@link ForkJoinTask}s. |
75 |
|
* <td ALIGN=CENTER> <b>Call from within fork/join computations</b></td> |
76 |
|
* </tr> |
77 |
|
* <tr> |
78 |
< |
* <td> <b>Arange async execution</td> |
78 |
> |
* <td> <b>Arrange async execution</td> |
79 |
|
* <td> {@link #execute(ForkJoinTask)}</td> |
80 |
|
* <td> {@link ForkJoinTask#fork}</td> |
81 |
|
* </tr> |
146 |
|
* Beyond work-stealing support and essential bookkeeping, the |
147 |
|
* main responsibility of this framework is to take actions when |
148 |
|
* one worker is waiting to join a task stolen (or always held by) |
149 |
< |
* another. Becauae we are multiplexing many tasks on to a pool |
149 |
> |
* another. Because we are multiplexing many tasks on to a pool |
150 |
|
* of workers, we can't just let them block (as in Thread.join). |
151 |
|
* We also cannot just reassign the joiner's run-time stack with |
152 |
|
* another and replace it later, which would be a form of |
163 |
|
* links to try to find such a task. |
164 |
|
* |
165 |
|
* Compensating: Unless there are already enough live threads, |
166 |
< |
* method helpMaintainParallelism() may create or or |
166 |
> |
* method helpMaintainParallelism() may create or |
167 |
|
* re-activate a spare thread to compensate for blocked |
168 |
|
* joiners until they unblock. |
169 |
|
* |
232 |
|
* ManagedBlocker), we may create or resume others to take their |
233 |
|
* place until they unblock (see below). Implementing this |
234 |
|
* requires counts of the number of "running" threads (i.e., those |
235 |
< |
* that are neither blocked nor artifically suspended) as well as |
235 |
> |
* that are neither blocked nor artificially suspended) as well as |
236 |
|
* the total number. These two values are packed into one field, |
237 |
|
* "workerCounts" because we need accurate snapshots when deciding |
238 |
|
* to create, resume or suspend. Note however that the |
239 |
< |
* correspondance of these counts to reality is not guaranteed. In |
239 |
> |
* correspondence of these counts to reality is not guaranteed. In |
240 |
|
* particular updates for unblocked threads may lag until they |
241 |
|
* actually wake up. |
242 |
|
* |
307 |
|
* about the same time as another is needlessly being created. We |
308 |
|
* counteract this and related slop in part by requiring resumed |
309 |
|
* spares to immediately recheck (in preStep) to see whether they |
310 |
< |
* they should re-suspend. |
310 |
> |
* should re-suspend. |
311 |
|
* |
312 |
|
* 6. Killing off unneeded workers. A timeout mechanism is used to |
313 |
|
* shed unused workers: The oldest (first) event queue waiter uses |
321 |
|
* 7. Deciding when to create new workers. The main dynamic |
322 |
|
* control in this class is deciding when to create extra threads |
323 |
|
* in method helpMaintainParallelism. We would like to keep |
324 |
< |
* exactly #parallelism threads running, which is an impossble |
324 |
> |
* exactly #parallelism threads running, which is an impossible |
325 |
|
* task. We always need to create one when the number of running |
326 |
|
* threads would become zero and all workers are busy. Beyond |
327 |
< |
* this, we must rely on heuristics that work well in the the |
328 |
< |
* presence of transients phenomena such as GC stalls, dynamic |
327 |
> |
* this, we must rely on heuristics that work well in the |
328 |
> |
* presence of transient phenomena such as GC stalls, dynamic |
329 |
|
* compilation, and wake-up lags. These transients are extremely |
330 |
|
* common -- we are normally trying to fully saturate the CPUs on |
331 |
|
* a machine, so almost any activity other than running tasks |
352 |
|
* "while ((local = field) != 0)") which are usually the simplest |
353 |
|
* way to ensure the required read orderings (which are sometimes |
354 |
|
* critical). Also several occurrences of the unusual "do {} |
355 |
< |
* while(!cas...)" which is the simplest way to force an update of |
355 |
> |
* while (!cas...)" which is the simplest way to force an update of |
356 |
|
* a CAS'ed variable. There are also other coding oddities that |
357 |
|
* help some methods perform reasonably even when interpreted (not |
358 |
|
* compiled), at the expense of some messy constructions that |
426 |
|
/** |
427 |
|
* The time to block in a join (see awaitJoin) before checking if |
428 |
|
* a new worker should be (re)started to maintain parallelism |
429 |
< |
* level. The value should be short enough to maintain gloabal |
429 |
> |
* level. The value should be short enough to maintain global |
430 |
|
* responsiveness and progress but long enough to avoid |
431 |
|
* counterproductive firings during GC stalls or unrelated system |
432 |
|
* activity, and to not bog down systems with continual re-firings |
436 |
|
|
437 |
|
/** |
438 |
|
* The wakeup interval (in nanoseconds) for the oldest worker |
439 |
< |
* worker waiting for an event invokes tryShutdownUnusedWorker to shrink |
440 |
< |
* the number of workers. The exact value does not matter too |
441 |
< |
* much, but should be long enough to slowly release resources |
442 |
< |
* during long periods without use without disrupting normal use. |
439 |
> |
* waiting for an event to invoke tryShutdownUnusedWorker to |
440 |
> |
* shrink the number of workers. The exact value does not matter |
441 |
> |
* too much. It must be short enough to release resources during |
442 |
> |
* sustained periods of idleness, but not so short that threads |
443 |
> |
* are continually re-created. |
444 |
|
*/ |
445 |
|
private static final long SHRINK_RATE_NANOS = |
446 |
|
30L * 1000L * 1000L * 1000L; // 2 per minute |
490 |
|
private volatile long stealCount; |
491 |
|
|
492 |
|
/** |
493 |
< |
* Encoded record of top of treiber stack of threads waiting for |
493 |
> |
* Encoded record of top of Treiber stack of threads waiting for |
494 |
|
* events. The top 32 bits contain the count being waited for. The |
495 |
|
* bottom 16 bits contains one plus the pool index of waiting |
496 |
|
* worker thread. (Bits 16-31 are unused.) |
509 |
|
private volatile int eventCount; |
510 |
|
|
511 |
|
/** |
512 |
< |
* Encoded record of top of treiber stack of spare threads waiting |
512 |
> |
* Encoded record of top of Treiber stack of spare threads waiting |
513 |
|
* for resumption. The top 16 bits contain an arbitrary count to |
514 |
|
* avoid ABA effects. The bottom 16bits contains one plus the pool |
515 |
|
* index of waiting worker thread. |
523 |
|
* Lifecycle control. The low word contains the number of workers |
524 |
|
* that are (probably) executing tasks. This value is atomically |
525 |
|
* incremented before a worker gets a task to run, and decremented |
526 |
< |
* when worker has no tasks and cannot find any. Bits 16-18 |
526 |
> |
* when a worker has no tasks and cannot find any. Bits 16-18 |
527 |
|
* contain runLevel value. When all are zero, the pool is |
528 |
|
* running. Level transitions are monotonic (running -> shutdown |
529 |
|
* -> terminating -> terminated) so each transition adds a bit. |
612 |
|
* (rarely) necessary when other count updates lag. |
613 |
|
* |
614 |
|
* @param dr -- either zero or ONE_RUNNING |
615 |
< |
* @param dt == either zero or ONE_TOTAL |
615 |
> |
* @param dt -- either zero or ONE_TOTAL |
616 |
|
*/ |
617 |
|
private void decrementWorkerCounts(int dr, int dt) { |
618 |
|
for (;;) { |
681 |
|
} |
682 |
|
|
683 |
|
/** |
684 |
< |
* Nulls out record of worker in workers array |
684 |
> |
* Nulls out record of worker in workers array. |
685 |
|
*/ |
686 |
|
private void forgetWorker(ForkJoinWorkerThread w) { |
687 |
|
int idx = w.poolIndex; |
688 |
< |
// Locking helps method recordWorker avoid unecessary expansion |
688 |
> |
// Locking helps method recordWorker avoid unnecessary expansion |
689 |
|
final ReentrantLock lock = this.workerLock; |
690 |
|
lock.lock(); |
691 |
|
try { |
700 |
|
/** |
701 |
|
* Final callback from terminating worker. Removes record of |
702 |
|
* worker from array, and adjusts counts. If pool is shutting |
703 |
< |
* down, tries to complete terminatation. |
703 |
> |
* down, tries to complete termination. |
704 |
|
* |
705 |
|
* @param w the worker |
706 |
|
*/ |
814 |
|
// Maintaining parallelism |
815 |
|
|
816 |
|
/** |
817 |
< |
* Pushes worker onto the spare stack |
817 |
> |
* Pushes worker onto the spare stack. |
818 |
|
*/ |
819 |
|
final void pushSpare(ForkJoinWorkerThread w) { |
820 |
|
int ns = (++w.spareCount << SPARE_COUNT_SHIFT) | (w.poolIndex + 1); |
839 |
|
UNSAFE.compareAndSwapInt(this, spareWaitersOffset, |
840 |
|
sw, w.nextSpare)) { |
841 |
|
int c; // increment running count before resume |
842 |
< |
do {} while(!UNSAFE.compareAndSwapInt |
843 |
< |
(this, workerCountsOffset, |
844 |
< |
c = workerCounts, c + ONE_RUNNING)); |
842 |
> |
do {} while (!UNSAFE.compareAndSwapInt |
843 |
> |
(this, workerCountsOffset, |
844 |
> |
c = workerCounts, c + ONE_RUNNING)); |
845 |
|
if (w.tryUnsuspend()) |
846 |
|
LockSupport.unpark(w); |
847 |
|
else // back out if w was shutdown |
853 |
|
* Tries to increase the number of running workers if below target |
854 |
|
* parallelism: If a spare exists tries to resume it via |
855 |
|
* tryResumeSpare. Otherwise, if not enough total workers or all |
856 |
< |
* existing workers are busy, adds a new worker. In all casses also |
856 |
> |
* existing workers are busy, adds a new worker. In all cases also |
857 |
|
* helps wake up releasable workers waiting for work. |
858 |
|
*/ |
859 |
|
private void helpMaintainParallelism() { |
992 |
|
w.lastEventCount = ec; // no need to wait |
993 |
|
break; |
994 |
|
} |
995 |
< |
else if (!(inactivate |= active)) |
995 |
> |
else if (!(inactivate |= active)) |
996 |
|
eventSync(w, wec); // must inactivate before sync |
997 |
|
} |
998 |
|
else |
1114 |
|
c = eventCount, c+1); |
1115 |
|
eventWaiters = 0L; // clobber lists |
1116 |
|
spareWaiters = 0; |
1117 |
< |
ForkJoinWorkerThread[] ws = workers; |
1111 |
< |
int n = ws.length; |
1112 |
< |
for (int i = 0; i < n; ++i) { |
1113 |
< |
ForkJoinWorkerThread w = ws[i]; |
1117 |
> |
for (ForkJoinWorkerThread w : workers) { |
1118 |
|
if (w != null) { |
1119 |
|
w.shutdown(); |
1120 |
|
if (passes > 0 && !w.isTerminated()) { |
1133 |
|
} |
1134 |
|
|
1135 |
|
/** |
1136 |
< |
* Clear out and cancel submissions, ignoring exceptions |
1136 |
> |
* Clears out and cancels submissions, ignoring exceptions. |
1137 |
|
*/ |
1138 |
|
private void cancelSubmissions() { |
1139 |
|
ForkJoinTask<?> task; |
1148 |
|
// misc support for ForkJoinWorkerThread |
1149 |
|
|
1150 |
|
/** |
1151 |
< |
* Returns pool number |
1151 |
> |
* Returns pool number. |
1152 |
|
*/ |
1153 |
|
final int getPoolNumber() { |
1154 |
|
return poolNumber; |
1155 |
|
} |
1156 |
|
|
1157 |
|
/** |
1158 |
< |
* Tries to accumulates steal count from a worker, clearing |
1159 |
< |
* the worker's value. |
1158 |
> |
* Tries to accumulate steal count from a worker, clearing |
1159 |
> |
* the worker's value if successful. |
1160 |
|
* |
1161 |
|
* @return true if worker steal count now zero |
1162 |
|
*/ |
1178 |
|
*/ |
1179 |
|
final int idlePerActive() { |
1180 |
|
int pc = parallelism; // use parallelism, not rc |
1181 |
< |
int ac = runState; // no mask -- artifically boosts during shutdown |
1181 |
> |
int ac = runState; // no mask -- artificially boosts during shutdown |
1182 |
|
// Use exact results for small values, saturate past 4 |
1183 |
< |
return pc <= ac? 0 : pc >>> 1 <= ac? 1 : pc >>> 2 <= ac? 3 : pc >>> 3; |
1183 |
> |
return ((pc <= ac) ? 0 : |
1184 |
> |
(pc >>> 1 <= ac) ? 1 : |
1185 |
> |
(pc >>> 2 <= ac) ? 3 : |
1186 |
> |
pc >>> 3); |
1187 |
|
} |
1188 |
|
|
1189 |
|
// Public and protected methods |
1233 |
|
* use {@link #defaultForkJoinWorkerThreadFactory}. |
1234 |
|
* @param handler the handler for internal worker threads that |
1235 |
|
* terminate due to unrecoverable errors encountered while executing |
1236 |
< |
* tasks. For default value, use <code>null</code>. |
1236 |
> |
* tasks. For default value, use {@code null}. |
1237 |
|
* @param asyncMode if true, |
1238 |
|
* establishes local first-in-first-out scheduling mode for forked |
1239 |
|
* tasks that are never joined. This mode may be more appropriate |
1240 |
|
* than default locally stack-based mode in applications in which |
1241 |
|
* worker threads only process event-style asynchronous tasks. |
1242 |
< |
* For default value, use <code>false</code>. |
1242 |
> |
* For default value, use {@code false}. |
1243 |
|
* @throws IllegalArgumentException if parallelism less than or |
1244 |
|
* equal to zero, or greater than implementation limit |
1245 |
|
* @throws NullPointerException if the factory is null |
1449 |
|
|
1450 |
|
/** |
1451 |
|
* Returns the number of worker threads that have started but not |
1452 |
< |
* yet terminated. This result returned by this method may differ |
1452 |
> |
* yet terminated. The result returned by this method may differ |
1453 |
|
* from {@link #getParallelism} when threads are created to |
1454 |
|
* maintain parallelism when others are cooperatively blocked. |
1455 |
|
* |
1534 |
|
*/ |
1535 |
|
public long getQueuedTaskCount() { |
1536 |
|
long count = 0; |
1537 |
< |
ForkJoinWorkerThread[] ws = workers; |
1531 |
< |
int n = ws.length; |
1532 |
< |
for (int i = 0; i < n; ++i) { |
1533 |
< |
ForkJoinWorkerThread w = ws[i]; |
1537 |
> |
for (ForkJoinWorkerThread w : workers) |
1538 |
|
if (w != null) |
1539 |
|
count += w.getQueueSize(); |
1536 |
– |
} |
1540 |
|
return count; |
1541 |
|
} |
1542 |
|
|
1591 |
|
*/ |
1592 |
|
protected int drainTasksTo(Collection<? super ForkJoinTask<?>> c) { |
1593 |
|
int count = submissionQueue.drainTo(c); |
1594 |
< |
ForkJoinWorkerThread[] ws = workers; |
1592 |
< |
int n = ws.length; |
1593 |
< |
for (int i = 0; i < n; ++i) { |
1594 |
< |
ForkJoinWorkerThread w = ws[i]; |
1594 |
> |
for (ForkJoinWorkerThread w : workers) |
1595 |
|
if (w != null) |
1596 |
|
count += w.drainTasksTo(c); |
1597 |
– |
} |
1597 |
|
return count; |
1598 |
|
} |
1599 |
|
|
1720 |
|
throws InterruptedException { |
1721 |
|
try { |
1722 |
|
return termination.awaitAdvanceInterruptibly(0, timeout, unit) > 0; |
1723 |
< |
} catch(TimeoutException ex) { |
1723 |
> |
} catch (TimeoutException ex) { |
1724 |
|
return false; |
1725 |
|
} |
1726 |
|
} |
1850 |
|
private static final long eventCountOffset = |
1851 |
|
objectFieldOffset("eventCount", ForkJoinPool.class); |
1852 |
|
private static final long eventWaitersOffset = |
1853 |
< |
objectFieldOffset("eventWaiters",ForkJoinPool.class); |
1853 |
> |
objectFieldOffset("eventWaiters", ForkJoinPool.class); |
1854 |
|
private static final long stealCountOffset = |
1855 |
< |
objectFieldOffset("stealCount",ForkJoinPool.class); |
1855 |
> |
objectFieldOffset("stealCount", ForkJoinPool.class); |
1856 |
|
private static final long spareWaitersOffset = |
1857 |
< |
objectFieldOffset("spareWaiters",ForkJoinPool.class); |
1857 |
> |
objectFieldOffset("spareWaiters", ForkJoinPool.class); |
1858 |
|
|
1859 |
|
private static long objectFieldOffset(String field, Class<?> klazz) { |
1860 |
|
try { |