ViewVC Help
View File | Revision Log | Show Annotations | Download File | Root Listing
root/jsr166/jsr166/src/jsr166y/ForkJoinPool.java
Revision: 1.189
Committed: Sat Sep 12 19:16:45 2015 UTC (8 years, 6 months ago) by jsr166
Branch: MAIN
CVS Tags: HEAD
Changes since 1.188: +6 -2 lines
Log Message:
code readability

File Contents

# Content
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/publicdomain/zero/1.0/
5 */
6
7 package jsr166y;
8
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.ExecutorService;
17 import java.util.concurrent.Future;
18 import java.util.concurrent.RejectedExecutionException;
19 import java.util.concurrent.RunnableFuture;
20 import java.util.concurrent.TimeUnit;
21
22 /**
23 * An {@link ExecutorService} for running {@link ForkJoinTask}s.
24 * A {@code ForkJoinPool} provides the entry point for submissions
25 * from non-{@code ForkJoinTask} clients, as well as management and
26 * monitoring operations.
27 *
28 * <p>A {@code ForkJoinPool} differs from other kinds of {@link
29 * ExecutorService} mainly by virtue of employing
30 * <em>work-stealing</em>: all threads in the pool attempt to find and
31 * execute tasks submitted to the pool and/or created by other active
32 * tasks (eventually blocking waiting for work if none exist). This
33 * enables efficient processing when most tasks spawn other subtasks
34 * (as do most {@code ForkJoinTask}s), as well as when many small
35 * tasks are submitted to the pool from external clients. Especially
36 * when setting <em>asyncMode</em> to true in constructors, {@code
37 * ForkJoinPool}s may also be appropriate for use with event-style
38 * tasks that are never joined.
39 *
40 * <p>A static {@link #commonPool()} is available and appropriate for
41 * most applications. The common pool is used by any ForkJoinTask that
42 * is not explicitly submitted to a specified pool. Using the common
43 * pool normally reduces resource usage (its threads are slowly
44 * reclaimed during periods of non-use, and reinstated upon subsequent
45 * use).
46 *
47 * <p>For applications that require separate or custom pools, a {@code
48 * ForkJoinPool} may be constructed with a given target parallelism
49 * level; by default, equal to the number of available processors. The
50 * pool attempts to maintain enough active (or available) threads by
51 * dynamically adding, suspending, or resuming internal worker
52 * threads, even if some tasks are stalled waiting to join
53 * others. However, no such adjustments are guaranteed in the face of
54 * blocked I/O or other unmanaged synchronization. The nested {@link
55 * ManagedBlocker} interface enables extension of the kinds of
56 * synchronization accommodated.
57 *
58 * <p>In addition to execution and lifecycle control methods, this
59 * class provides status check methods (for example
60 * {@link #getStealCount}) that are intended to aid in developing,
61 * tuning, and monitoring fork/join applications. Also, method
62 * {@link #toString} returns indications of pool state in a
63 * convenient form for informal monitoring.
64 *
65 * <p>As is the case with other ExecutorServices, there are three
66 * main task execution methods summarized in the following table.
67 * These are designed to be used primarily by clients not already
68 * engaged in fork/join computations in the current pool. The main
69 * forms of these methods accept instances of {@code ForkJoinTask},
70 * but overloaded forms also allow mixed execution of plain {@code
71 * Runnable}- or {@code Callable}- based activities as well. However,
72 * tasks that are already executing in a pool should normally instead
73 * use the within-computation forms listed in the table unless using
74 * async event-style tasks that are not usually joined, in which case
75 * there is little difference among choice of methods.
76 *
77 * <table BORDER CELLPADDING=3 CELLSPACING=1>
78 * <tr>
79 * <td></td>
80 * <td ALIGN=CENTER> <b>Call from non-fork/join clients</b></td>
81 * <td ALIGN=CENTER> <b>Call from within fork/join computations</b></td>
82 * </tr>
83 * <tr>
84 * <td> <b>Arrange async execution</td>
85 * <td> {@link #execute(ForkJoinTask)}</td>
86 * <td> {@link ForkJoinTask#fork}</td>
87 * </tr>
88 * <tr>
89 * <td> <b>Await and obtain result</td>
90 * <td> {@link #invoke(ForkJoinTask)}</td>
91 * <td> {@link ForkJoinTask#invoke}</td>
92 * </tr>
93 * <tr>
94 * <td> <b>Arrange exec and obtain Future</td>
95 * <td> {@link #submit(ForkJoinTask)}</td>
96 * <td> {@link ForkJoinTask#fork} (ForkJoinTasks <em>are</em> Futures)</td>
97 * </tr>
98 * </table>
99 *
100 * <p>The common pool is by default constructed with default
101 * parameters, but these may be controlled by setting three {@link
102 * System#getProperty system properties} with prefix {@code
103 * java.util.concurrent.ForkJoinPool.common}: {@code parallelism} --
104 * an integer greater than zero, {@code threadFactory} -- the class
105 * name of a {@link ForkJoinWorkerThreadFactory}, and {@code
106 * exceptionHandler} -- the class name of a {@link
107 * java.lang.Thread.UncaughtExceptionHandler
108 * Thread.UncaughtExceptionHandler}. Upon any error in establishing
109 * these settings, default parameters are used.
110 *
111 * <p><b>Implementation notes</b>: This implementation restricts the
112 * maximum number of running threads to 32767. Attempts to create
113 * pools with greater than the maximum number result in
114 * {@code IllegalArgumentException}.
115 *
116 * <p>This implementation rejects submitted tasks (that is, by throwing
117 * {@link RejectedExecutionException}) only when the pool is shut down
118 * or internal resources have been exhausted.
119 *
120 * @since 1.7
121 * @author Doug Lea
122 */
123 public class ForkJoinPool extends AbstractExecutorService {
124
125 /*
126 * Implementation Overview
127 *
128 * This class and its nested classes provide the main
129 * functionality and control for a set of worker threads:
130 * Submissions from non-FJ threads enter into submission queues.
131 * Workers take these tasks and typically split them into subtasks
132 * that may be stolen by other workers. Preference rules give
133 * first priority to processing tasks from their own queues (LIFO
134 * or FIFO, depending on mode), then to randomized FIFO steals of
135 * tasks in other queues.
136 *
137 * WorkQueues
138 * ==========
139 *
140 * Most operations occur within work-stealing queues (in nested
141 * class WorkQueue). These are special forms of Deques that
142 * support only three of the four possible end-operations -- push,
143 * pop, and poll (aka steal), under the further constraints that
144 * push and pop are called only from the owning thread (or, as
145 * extended here, under a lock), while poll may be called from
146 * other threads. (If you are unfamiliar with them, you probably
147 * want to read Herlihy and Shavit's book "The Art of
148 * Multiprocessor programming", chapter 16 describing these in
149 * more detail before proceeding.) The main work-stealing queue
150 * design is roughly similar to those in the papers "Dynamic
151 * Circular Work-Stealing Deque" by Chase and Lev, SPAA 2005
152 * (http://research.sun.com/scalable/pubs/index.html) and
153 * "Idempotent work stealing" by Michael, Saraswat, and Vechev,
154 * PPoPP 2009 (http://portal.acm.org/citation.cfm?id=1504186).
155 * The main differences ultimately stem from GC requirements that
156 * we null out taken slots as soon as we can, to maintain as small
157 * a footprint as possible even in programs generating huge
158 * numbers of tasks. To accomplish this, we shift the CAS
159 * arbitrating pop vs poll (steal) from being on the indices
160 * ("base" and "top") to the slots themselves. So, both a
161 * successful pop and poll mainly entail a CAS of a slot from
162 * non-null to null. Because we rely on CASes of references, we
163 * do not need tag bits on base or top. They are simple ints as
164 * used in any circular array-based queue (see for example
165 * ArrayDeque). Updates to the indices must still be ordered in a
166 * way that guarantees that top == base means the queue is empty,
167 * but otherwise may err on the side of possibly making the queue
168 * appear nonempty when a push, pop, or poll have not fully
169 * committed. Note that this means that the poll operation,
170 * considered individually, is not wait-free. One thief cannot
171 * successfully continue until another in-progress one (or, if
172 * previously empty, a push) completes. However, in the
173 * aggregate, we ensure at least probabilistic non-blockingness.
174 * If an attempted steal fails, a thief always chooses a different
175 * random victim target to try next. So, in order for one thief to
176 * progress, it suffices for any in-progress poll or new push on
177 * any empty queue to complete. (This is why we normally use
178 * method pollAt and its variants that try once at the apparent
179 * base index, else consider alternative actions, rather than
180 * method poll.)
181 *
182 * This approach also enables support of a user mode in which local
183 * task processing is in FIFO, not LIFO order, simply by using
184 * poll rather than pop. This can be useful in message-passing
185 * frameworks in which tasks are never joined. However neither
186 * mode considers affinities, loads, cache localities, etc, so
187 * rarely provide the best possible performance on a given
188 * machine, but portably provide good throughput by averaging over
189 * these factors. (Further, even if we did try to use such
190 * information, we do not usually have a basis for exploiting it.
191 * For example, some sets of tasks profit from cache affinities,
192 * but others are harmed by cache pollution effects.)
193 *
194 * WorkQueues are also used in a similar way for tasks submitted
195 * to the pool. We cannot mix these tasks in the same queues used
196 * for work-stealing (this would contaminate lifo/fifo
197 * processing). Instead, we randomly associate submission queues
198 * with submitting threads, using a form of hashing. The
199 * ThreadLocal Submitter class contains a value initially used as
200 * a hash code for choosing existing queues, but may be randomly
201 * repositioned upon contention with other submitters. In
202 * essence, submitters act like workers except that they are
203 * restricted to executing local tasks that they submitted (or in
204 * the case of CountedCompleters, others with the same root task).
205 * However, because most shared/external queue operations are more
206 * expensive than internal, and because, at steady state, external
207 * submitters will compete for CPU with workers, ForkJoinTask.join
208 * and related methods disable them from repeatedly helping to
209 * process tasks if all workers are active. Insertion of tasks in
210 * shared mode requires a lock (mainly to protect in the case of
211 * resizing) but we use only a simple spinlock (using bits in
212 * field qlock), because submitters encountering a busy queue move
213 * on to try or create other queues -- they block only when
214 * creating and registering new queues.
215 *
216 * Management
217 * ==========
218 *
219 * The main throughput advantages of work-stealing stem from
220 * decentralized control -- workers mostly take tasks from
221 * themselves or each other. We cannot negate this in the
222 * implementation of other management responsibilities. The main
223 * tactic for avoiding bottlenecks is packing nearly all
224 * essentially atomic control state into two volatile variables
225 * that are by far most often read (not written) as status and
226 * consistency checks.
227 *
228 * Field "ctl" contains 64 bits holding all the information needed
229 * to atomically decide to add, inactivate, enqueue (on an event
230 * queue), dequeue, and/or re-activate workers. To enable this
231 * packing, we restrict maximum parallelism to (1<<15)-1 (which is
232 * far in excess of normal operating range) to allow ids, counts,
233 * and their negations (used for thresholding) to fit into 16bit
234 * fields.
235 *
236 * Field "plock" is a form of sequence lock with a saturating
237 * shutdown bit (similarly for per-queue "qlocks"), mainly
238 * protecting updates to the workQueues array, as well as to
239 * enable shutdown. When used as a lock, it is normally only very
240 * briefly held, so is nearly always available after at most a
241 * brief spin, but we use a monitor-based backup strategy to
242 * block when needed.
243 *
244 * Recording WorkQueues. WorkQueues are recorded in the
245 * "workQueues" array that is created upon first use and expanded
246 * if necessary. Updates to the array while recording new workers
247 * and unrecording terminated ones are protected from each other
248 * by a lock but the array is otherwise concurrently readable, and
249 * accessed directly. To simplify index-based operations, the
250 * array size is always a power of two, and all readers must
251 * tolerate null slots. Worker queues are at odd indices. Shared
252 * (submission) queues are at even indices, up to a maximum of 64
253 * slots, to limit growth even if array needs to expand to add
254 * more workers. Grouping them together in this way simplifies and
255 * speeds up task scanning.
256 *
257 * All worker thread creation is on-demand, triggered by task
258 * submissions, replacement of terminated workers, and/or
259 * compensation for blocked workers. However, all other support
260 * code is set up to work with other policies. To ensure that we
261 * do not hold on to worker references that would prevent GC, ALL
262 * accesses to workQueues are via indices into the workQueues
263 * array (which is one source of some of the messy code
264 * constructions here). In essence, the workQueues array serves as
265 * a weak reference mechanism. Thus for example the wait queue
266 * field of ctl stores indices, not references. Access to the
267 * workQueues in associated methods (for example signalWork) must
268 * both index-check and null-check the IDs. All such accesses
269 * ignore bad IDs by returning out early from what they are doing,
270 * since this can only be associated with termination, in which
271 * case it is OK to give up. All uses of the workQueues array
272 * also check that it is non-null (even if previously
273 * non-null). This allows nulling during termination, which is
274 * currently not necessary, but remains an option for
275 * resource-revocation-based shutdown schemes. It also helps
276 * reduce JIT issuance of uncommon-trap code, which tends to
277 * unnecessarily complicate control flow in some methods.
278 *
279 * Event Queuing. Unlike HPC work-stealing frameworks, we cannot
280 * let workers spin indefinitely scanning for tasks when none can
281 * be found immediately, and we cannot start/resume workers unless
282 * there appear to be tasks available. On the other hand, we must
283 * quickly prod them into action when new tasks are submitted or
284 * generated. In many usages, ramp-up time to activate workers is
285 * the main limiting factor in overall performance (this is
286 * compounded at program start-up by JIT compilation and
287 * allocation). So we try to streamline this as much as possible.
288 * We park/unpark workers after placing in an event wait queue
289 * when they cannot find work. This "queue" is actually a simple
290 * Treiber stack, headed by the "id" field of ctl, plus a 15bit
291 * counter value (that reflects the number of times a worker has
292 * been inactivated) to avoid ABA effects (we need only as many
293 * version numbers as worker threads). Successors are held in
294 * field WorkQueue.nextWait. Queuing deals with several intrinsic
295 * races, mainly that a task-producing thread can miss seeing (and
296 * signalling) another thread that gave up looking for work but
297 * has not yet entered the wait queue. We solve this by requiring
298 * a full sweep of all workers (via repeated calls to method
299 * scan()) both before and after a newly waiting worker is added
300 * to the wait queue. During a rescan, the worker might release
301 * some other queued worker rather than itself, which has the same
302 * net effect. Because enqueued workers may actually be rescanning
303 * rather than waiting, we set and clear the "parker" field of
304 * WorkQueues to reduce unnecessary calls to unpark. (This
305 * requires a secondary recheck to avoid missed signals.) Note
306 * the unusual conventions about Thread.interrupts surrounding
307 * parking and other blocking: Because interrupts are used solely
308 * to alert threads to check termination, which is checked anyway
309 * upon blocking, we clear status (using Thread.interrupted)
310 * before any call to park, so that park does not immediately
311 * return due to status being set via some other unrelated call to
312 * interrupt in user code.
313 *
314 * Signalling. We create or wake up workers only when there
315 * appears to be at least one task they might be able to find and
316 * execute. However, many other threads may notice the same task
317 * and each signal to wake up a thread that might take it. So in
318 * general, pools will be over-signalled. When a submission is
319 * added or another worker adds a task to a queue that has fewer
320 * than two tasks, they signal waiting workers (or trigger
321 * creation of new ones if fewer than the given parallelism level
322 * -- signalWork), and may leave a hint to the unparked worker to
323 * help signal others upon wakeup). These primary signals are
324 * buttressed by others (see method helpSignal) whenever other
325 * threads scan for work or do not have a task to process. On
326 * most platforms, signalling (unpark) overhead time is noticeably
327 * long, and the time between signalling a thread and it actually
328 * making progress can be very noticeably long, so it is worth
329 * offloading these delays from critical paths as much as
330 * possible.
331 *
332 * Trimming workers. To release resources after periods of lack of
333 * use, a worker starting to wait when the pool is quiescent will
334 * time out and terminate if the pool has remained quiescent for a
335 * given period -- a short period if there are more threads than
336 * parallelism, longer as the number of threads decreases. This
337 * will slowly propagate, eventually terminating all workers after
338 * periods of non-use.
339 *
340 * Shutdown and Termination. A call to shutdownNow atomically sets
341 * a plock bit and then (non-atomically) sets each worker's
342 * qlock status, cancels all unprocessed tasks, and wakes up
343 * all waiting workers. Detecting whether termination should
344 * commence after a non-abrupt shutdown() call requires more work
345 * and bookkeeping. We need consensus about quiescence (i.e., that
346 * there is no more work). The active count provides a primary
347 * indication but non-abrupt shutdown still requires a rechecking
348 * scan for any workers that are inactive but not queued.
349 *
350 * Joining Tasks
351 * =============
352 *
353 * Any of several actions may be taken when one worker is waiting
354 * to join a task stolen (or always held) by another. Because we
355 * are multiplexing many tasks on to a pool of workers, we can't
356 * just let them block (as in Thread.join). We also cannot just
357 * reassign the joiner's run-time stack with another and replace
358 * it later, which would be a form of "continuation", that even if
359 * possible is not necessarily a good idea since we sometimes need
360 * both an unblocked task and its continuation to progress.
361 * Instead we combine two tactics:
362 *
363 * Helping: Arranging for the joiner to execute some task that it
364 * would be running if the steal had not occurred.
365 *
366 * Compensating: Unless there are already enough live threads,
367 * method tryCompensate() may create or re-activate a spare
368 * thread to compensate for blocked joiners until they unblock.
369 *
370 * A third form (implemented in tryRemoveAndExec) amounts to
371 * helping a hypothetical compensator: If we can readily tell that
372 * a possible action of a compensator is to steal and execute the
373 * task being joined, the joining thread can do so directly,
374 * without the need for a compensation thread (although at the
375 * expense of larger run-time stacks, but the tradeoff is
376 * typically worthwhile).
377 *
378 * The ManagedBlocker extension API can't use helping so relies
379 * only on compensation in method awaitBlocker.
380 *
381 * The algorithm in tryHelpStealer entails a form of "linear"
382 * helping: Each worker records (in field currentSteal) the most
383 * recent task it stole from some other worker. Plus, it records
384 * (in field currentJoin) the task it is currently actively
385 * joining. Method tryHelpStealer uses these markers to try to
386 * find a worker to help (i.e., steal back a task from and execute
387 * it) that could hasten completion of the actively joined task.
388 * In essence, the joiner executes a task that would be on its own
389 * local deque had the to-be-joined task not been stolen. This may
390 * be seen as a conservative variant of the approach in Wagner &
391 * Calder "Leapfrogging: a portable technique for implementing
392 * efficient futures" SIGPLAN Notices, 1993
393 * (http://portal.acm.org/citation.cfm?id=155354). It differs in
394 * that: (1) We only maintain dependency links across workers upon
395 * steals, rather than use per-task bookkeeping. This sometimes
396 * requires a linear scan of workQueues array to locate stealers,
397 * but often doesn't because stealers leave hints (that may become
398 * stale/wrong) of where to locate them. It is only a hint
399 * because a worker might have had multiple steals and the hint
400 * records only one of them (usually the most current). Hinting
401 * isolates cost to when it is needed, rather than adding to
402 * per-task overhead. (2) It is "shallow", ignoring nesting and
403 * potentially cyclic mutual steals. (3) It is intentionally
404 * racy: field currentJoin is updated only while actively joining,
405 * which means that we miss links in the chain during long-lived
406 * tasks, GC stalls etc (which is OK since blocking in such cases
407 * is usually a good idea). (4) We bound the number of attempts
408 * to find work (see MAX_HELP) and fall back to suspending the
409 * worker and if necessary replacing it with another.
410 *
411 * Helping actions for CountedCompleters are much simpler: Method
412 * helpComplete can take and execute any task with the same root
413 * as the task being waited on. However, this still entails some
414 * traversal of completer chains, so is less efficient than using
415 * CountedCompleters without explicit joins.
416 *
417 * It is impossible to keep exactly the target parallelism number
418 * of threads running at any given time. Determining the
419 * existence of conservatively safe helping targets, the
420 * availability of already-created spares, and the apparent need
421 * to create new spares are all racy, so we rely on multiple
422 * retries of each. Compensation in the apparent absence of
423 * helping opportunities is challenging to control on JVMs, where
424 * GC and other activities can stall progress of tasks that in
425 * turn stall out many other dependent tasks, without us being
426 * able to determine whether they will ever require compensation.
427 * Even though work-stealing otherwise encounters little
428 * degradation in the presence of more threads than cores,
429 * aggressively adding new threads in such cases entails risk of
430 * unwanted positive feedback control loops in which more threads
431 * cause more dependent stalls (as well as delayed progress of
432 * unblocked threads to the point that we know they are available)
433 * leading to more situations requiring more threads, and so
434 * on. This aspect of control can be seen as an (analytically
435 * intractable) game with an opponent that may choose the worst
436 * (for us) active thread to stall at any time. We take several
437 * precautions to bound losses (and thus bound gains), mainly in
438 * methods tryCompensate and awaitJoin.
439 *
440 * Common Pool
441 * ===========
442 *
443 * The static common Pool always exists after static
444 * initialization. Since it (or any other created pool) need
445 * never be used, we minimize initial construction overhead and
446 * footprint to the setup of about a dozen fields, with no nested
447 * allocation. Most bootstrapping occurs within method
448 * fullExternalPush during the first submission to the pool.
449 *
450 * When external threads submit to the common pool, they can
451 * perform some subtask processing (see externalHelpJoin and
452 * related methods). We do not need to record whether these
453 * submissions are to the common pool -- if not, externalHelpJoin
454 * returns quickly (at the most helping to signal some common pool
455 * workers). These submitters would otherwise be blocked waiting
456 * for completion, so the extra effort (with liberally sprinkled
457 * task status checks) in inapplicable cases amounts to an odd
458 * form of limited spin-wait before blocking in ForkJoinTask.join.
459 *
460 * Style notes
461 * ===========
462 *
463 * There is a lot of representation-level coupling among classes
464 * ForkJoinPool, ForkJoinWorkerThread, and ForkJoinTask. The
465 * fields of WorkQueue maintain data structures managed by
466 * ForkJoinPool, so are directly accessed. There is little point
467 * trying to reduce this, since any associated future changes in
468 * representations will need to be accompanied by algorithmic
469 * changes anyway. Several methods intrinsically sprawl because
470 * they must accumulate sets of consistent reads of volatiles held
471 * in local variables. Methods signalWork() and scan() are the
472 * main bottlenecks, so are especially heavily
473 * micro-optimized/mangled. There are lots of inline assignments
474 * (of form "while ((local = field) != 0)") which are usually the
475 * simplest way to ensure the required read orderings (which are
476 * sometimes critical). This leads to a "C"-like style of listing
477 * declarations of these locals at the heads of methods or blocks.
478 * There are several occurrences of the unusual "do {} while
479 * (!cas...)" which is the simplest way to force an update of a
480 * CAS'ed variable. There are also other coding oddities (including
481 * several unnecessary-looking hoisted null checks) that help
482 * some methods perform reasonably even when interpreted (not
483 * compiled).
484 *
485 * The order of declarations in this file is:
486 * (1) Static utility functions
487 * (2) Nested (static) classes
488 * (3) Static fields
489 * (4) Fields, along with constants used when unpacking some of them
490 * (5) Internal control methods
491 * (6) Callbacks and other support for ForkJoinTask methods
492 * (7) Exported methods
493 * (8) Static block initializing statics in minimally dependent order
494 */
495
496 // Static utilities
497
498 /**
499 * If there is a security manager, makes sure caller has
500 * permission to modify threads.
501 */
502 private static void checkPermission() {
503 SecurityManager security = System.getSecurityManager();
504 if (security != null)
505 security.checkPermission(modifyThreadPermission);
506 }
507
508 // Nested classes
509
510 /**
511 * Factory for creating new {@link ForkJoinWorkerThread}s.
512 * A {@code ForkJoinWorkerThreadFactory} must be defined and used
513 * for {@code ForkJoinWorkerThread} subclasses that extend base
514 * functionality or initialize threads with different contexts.
515 */
516 public static interface ForkJoinWorkerThreadFactory {
517 /**
518 * Returns a new worker thread operating in the given pool.
519 *
520 * @param pool the pool this thread works in
521 * @throws NullPointerException if the pool is null
522 */
523 public ForkJoinWorkerThread newThread(ForkJoinPool pool);
524 }
525
526 /**
527 * Default ForkJoinWorkerThreadFactory implementation; creates a
528 * new ForkJoinWorkerThread.
529 */
530 static final class DefaultForkJoinWorkerThreadFactory
531 implements ForkJoinWorkerThreadFactory {
532 public final ForkJoinWorkerThread newThread(ForkJoinPool pool) {
533 return new ForkJoinWorkerThread(pool);
534 }
535 }
536
537 /**
538 * Per-thread records for threads that submit to pools. Currently
539 * holds only pseudo-random seed / index that is used to choose
540 * submission queues in method externalPush. In the future, this may
541 * also incorporate a means to implement different task rejection
542 * and resubmission policies.
543 *
544 * Seeds for submitters and workers/workQueues work in basically
545 * the same way but are initialized and updated using slightly
546 * different mechanics. Both are initialized using the same
547 * approach as in class ThreadLocal, where successive values are
548 * unlikely to collide with previous values. Seeds are then
549 * randomly modified upon collisions using xorshifts, which
550 * requires a non-zero seed.
551 */
552 static final class Submitter {
553 int seed;
554 Submitter(int s) { seed = s; }
555 }
556
557 /**
558 * Class for artificial tasks that are used to replace the target
559 * of local joins if they are removed from an interior queue slot
560 * in WorkQueue.tryRemoveAndExec. We don't need the proxy to
561 * actually do anything beyond having a unique identity.
562 */
563 static final class EmptyTask extends ForkJoinTask<Void> {
564 private static final long serialVersionUID = -7721805057305804111L;
565 EmptyTask() { status = ForkJoinTask.NORMAL; } // force done
566 public final Void getRawResult() { return null; }
567 public final void setRawResult(Void x) {}
568 public final boolean exec() { return true; }
569 }
570
571 /**
572 * Queues supporting work-stealing as well as external task
573 * submission. See above for main rationale and algorithms.
574 * Implementation relies heavily on "Unsafe" intrinsics
575 * and selective use of "volatile":
576 *
577 * Field "base" is the index (mod array.length) of the least valid
578 * queue slot, which is always the next position to steal (poll)
579 * from if nonempty. Reads and writes require volatile orderings
580 * but not CAS, because updates are only performed after slot
581 * CASes.
582 *
583 * Field "top" is the index (mod array.length) of the next queue
584 * slot to push to or pop from. It is written only by owner thread
585 * for push, or under lock for external/shared push, and accessed
586 * by other threads only after reading (volatile) base. Both top
587 * and base are allowed to wrap around on overflow, but (top -
588 * base) (or more commonly -(base - top) to force volatile read of
589 * base before top) still estimates size. The lock ("qlock") is
590 * forced to -1 on termination, causing all further lock attempts
591 * to fail. (Note: we don't need CAS for termination state because
592 * upon pool shutdown, all shared-queues will stop being used
593 * anyway.) Nearly all lock bodies are set up so that exceptions
594 * within lock bodies are "impossible" (modulo JVM errors that
595 * would cause failure anyway.)
596 *
597 * The array slots are read and written using the emulation of
598 * volatiles/atomics provided by Unsafe. Insertions must in
599 * general use putOrderedObject as a form of releasing store to
600 * ensure that all writes to the task object are ordered before
601 * its publication in the queue. All removals entail a CAS to
602 * null. The array is always a power of two. To ensure safety of
603 * Unsafe array operations, all accesses perform explicit null
604 * checks and implicit bounds checks via power-of-two masking.
605 *
606 * In addition to basic queuing support, this class contains
607 * fields described elsewhere to control execution. It turns out
608 * to work better memory-layout-wise to include them in this class
609 * rather than a separate class.
610 *
611 * Performance on most platforms is very sensitive to placement of
612 * instances of both WorkQueues and their arrays -- we absolutely
613 * do not want multiple WorkQueue instances or multiple queue
614 * arrays sharing cache lines. (It would be best for queue objects
615 * and their arrays to share, but there is nothing available to
616 * help arrange that). Unfortunately, because they are recorded
617 * in a common array, WorkQueue instances are often moved to be
618 * adjacent by garbage collectors. To reduce impact, we use field
619 * padding that works OK on common platforms; this effectively
620 * trades off slightly slower average field access for the sake of
621 * avoiding really bad worst-case access. (Until better JVM
622 * support is in place, this padding is dependent on transient
623 * properties of JVM field layout rules.) We also take care in
624 * allocating, sizing and resizing the array. Non-shared queue
625 * arrays are initialized by workers before use. Others are
626 * allocated on first use.
627 */
628 static final class WorkQueue {
629 /**
630 * Capacity of work-stealing queue array upon initialization.
631 * Must be a power of two; at least 4, but should be larger to
632 * reduce or eliminate cacheline sharing among queues.
633 * Currently, it is much larger, as a partial workaround for
634 * the fact that JVMs often place arrays in locations that
635 * share GC bookkeeping (especially cardmarks) such that
636 * per-write accesses encounter serious memory contention.
637 */
638 static final int INITIAL_QUEUE_CAPACITY = 1 << 13;
639
640 /**
641 * Maximum size for queue arrays. Must be a power of two less
642 * than or equal to 1 << (31 - width of array entry) to ensure
643 * lack of wraparound of index calculations, but defined to a
644 * value a bit less than this to help users trap runaway
645 * programs before saturating systems.
646 */
647 static final int MAXIMUM_QUEUE_CAPACITY = 1 << 26; // 64M
648
649 // Heuristic padding to ameliorate unfortunate memory placements
650 volatile long pad00, pad01, pad02, pad03, pad04, pad05, pad06;
651
652 int seed; // for random scanning; initialize nonzero
653 volatile int eventCount; // encoded inactivation count; < 0 if inactive
654 int nextWait; // encoded record of next event waiter
655 int hint; // steal or signal hint (index)
656 int poolIndex; // index of this queue in pool (or 0)
657 final int mode; // 0: lifo, > 0: fifo, < 0: shared
658 int nsteals; // number of steals
659 volatile int qlock; // 1: locked, -1: terminate; else 0
660 volatile int base; // index of next slot for poll
661 int top; // index of next slot for push
662 ForkJoinTask<?>[] array; // the elements (initially unallocated)
663 final ForkJoinPool pool; // the containing pool (may be null)
664 final ForkJoinWorkerThread owner; // owning thread or null if shared
665 volatile Thread parker; // == owner during call to park; else null
666 volatile ForkJoinTask<?> currentJoin; // task being joined in awaitJoin
667 ForkJoinTask<?> currentSteal; // current non-local task being executed
668
669 volatile Object pad10, pad11, pad12, pad13, pad14, pad15, pad16, pad17;
670 volatile Object pad18, pad19, pad1a, pad1b, pad1c, pad1d;
671
672 WorkQueue(ForkJoinPool pool, ForkJoinWorkerThread owner, int mode,
673 int seed) {
674 this.pool = pool;
675 this.owner = owner;
676 this.mode = mode;
677 this.seed = seed;
678 // Place indices in the center of array (that is not yet allocated)
679 base = top = INITIAL_QUEUE_CAPACITY >>> 1;
680 }
681
682 /**
683 * Returns the approximate number of tasks in the queue.
684 */
685 final int queueSize() {
686 int n = base - top; // non-owner callers must read base first
687 return (n >= 0) ? 0 : -n; // ignore transient negative
688 }
689
690 /**
691 * Provides a more accurate estimate of whether this queue has
692 * any tasks than does queueSize, by checking whether a
693 * near-empty queue has at least one unclaimed task.
694 */
695 final boolean isEmpty() {
696 ForkJoinTask<?>[] a; int m, s;
697 int n = base - (s = top);
698 return (n >= 0 ||
699 (n == -1 &&
700 ((a = array) == null ||
701 (m = a.length - 1) < 0 ||
702 U.getObject
703 (a, (long)((m & (s - 1)) << ASHIFT) + ABASE) == null)));
704 }
705
706 /**
707 * Pushes a task. Call only by owner in unshared queues. (The
708 * shared-queue version is embedded in method externalPush.)
709 *
710 * @param task the task. Caller must ensure non-null.
711 * @throws RejectedExecutionException if array cannot be resized
712 */
713 final void push(ForkJoinTask<?> task) {
714 ForkJoinTask<?>[] a; ForkJoinPool p;
715 int s = top, m, n;
716 if ((a = array) != null) { // ignore if queue removed
717 int j = (((m = a.length - 1) & s) << ASHIFT) + ABASE;
718 U.putOrderedObject(a, j, task);
719 if ((n = (top = s + 1) - base) <= 2) {
720 if ((p = pool) != null)
721 p.signalWork(this);
722 }
723 else if (n >= m)
724 growArray();
725 }
726 }
727
728 /**
729 * Initializes or doubles the capacity of array. Call either
730 * by owner or with lock held -- it is OK for base, but not
731 * top, to move while resizings are in progress.
732 */
733 final ForkJoinTask<?>[] growArray() {
734 ForkJoinTask<?>[] oldA = array;
735 int size = oldA != null ? oldA.length << 1 : INITIAL_QUEUE_CAPACITY;
736 if (size > MAXIMUM_QUEUE_CAPACITY)
737 throw new RejectedExecutionException("Queue capacity exceeded");
738 int oldMask, t, b;
739 ForkJoinTask<?>[] a = array = new ForkJoinTask<?>[size];
740 if (oldA != null && (oldMask = oldA.length - 1) >= 0 &&
741 (t = top) - (b = base) > 0) {
742 int mask = size - 1;
743 do {
744 ForkJoinTask<?> x;
745 int oldj = ((b & oldMask) << ASHIFT) + ABASE;
746 int j = ((b & mask) << ASHIFT) + ABASE;
747 x = (ForkJoinTask<?>)U.getObjectVolatile(oldA, oldj);
748 if (x != null &&
749 U.compareAndSwapObject(oldA, oldj, x, null))
750 U.putObjectVolatile(a, j, x);
751 } while (++b != t);
752 }
753 return a;
754 }
755
756 /**
757 * Takes next task, if one exists, in LIFO order. Call only
758 * by owner in unshared queues.
759 */
760 final ForkJoinTask<?> pop() {
761 ForkJoinTask<?>[] a; ForkJoinTask<?> t; int m;
762 if ((a = array) != null && (m = a.length - 1) >= 0) {
763 for (int s; (s = top - 1) - base >= 0;) {
764 long j = ((m & s) << ASHIFT) + ABASE;
765 if ((t = (ForkJoinTask<?>)U.getObject(a, j)) == null)
766 break;
767 if (U.compareAndSwapObject(a, j, t, null)) {
768 top = s;
769 return t;
770 }
771 }
772 }
773 return null;
774 }
775
776 /**
777 * Takes a task in FIFO order if b is base of queue and a task
778 * can be claimed without contention. Specialized versions
779 * appear in ForkJoinPool methods scan and tryHelpStealer.
780 */
781 final ForkJoinTask<?> pollAt(int b) {
782 ForkJoinTask<?> t; ForkJoinTask<?>[] a;
783 if ((a = array) != null) {
784 int j = (((a.length - 1) & b) << ASHIFT) + ABASE;
785 if ((t = (ForkJoinTask<?>)U.getObjectVolatile(a, j)) != null &&
786 base == b &&
787 U.compareAndSwapObject(a, j, t, null)) {
788 base = b + 1;
789 return t;
790 }
791 }
792 return null;
793 }
794
795 /**
796 * Takes next task, if one exists, in FIFO order.
797 */
798 final ForkJoinTask<?> poll() {
799 ForkJoinTask<?>[] a; int b; ForkJoinTask<?> t;
800 while ((b = base) - top < 0 && (a = array) != null) {
801 int j = (((a.length - 1) & b) << ASHIFT) + ABASE;
802 t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
803 if (t != null) {
804 if (base == b &&
805 U.compareAndSwapObject(a, j, t, null)) {
806 base = b + 1;
807 return t;
808 }
809 }
810 else if (base == b) {
811 if (b + 1 == top)
812 break;
813 Thread.yield(); // wait for lagging update (very rare)
814 }
815 }
816 return null;
817 }
818
819 /**
820 * Takes next task, if one exists, in order specified by mode.
821 */
822 final ForkJoinTask<?> nextLocalTask() {
823 return mode == 0 ? pop() : poll();
824 }
825
826 /**
827 * Returns next task, if one exists, in order specified by mode.
828 */
829 final ForkJoinTask<?> peek() {
830 ForkJoinTask<?>[] a = array; int m;
831 if (a == null || (m = a.length - 1) < 0)
832 return null;
833 int i = mode == 0 ? top - 1 : base;
834 int j = ((i & m) << ASHIFT) + ABASE;
835 return (ForkJoinTask<?>)U.getObjectVolatile(a, j);
836 }
837
838 /**
839 * Pops the given task only if it is at the current top.
840 * (A shared version is available only via FJP.tryExternalUnpush)
841 */
842 final boolean tryUnpush(ForkJoinTask<?> t) {
843 ForkJoinTask<?>[] a; int s;
844 if ((a = array) != null && (s = top) != base &&
845 U.compareAndSwapObject
846 (a, (((a.length - 1) & --s) << ASHIFT) + ABASE, t, null)) {
847 top = s;
848 return true;
849 }
850 return false;
851 }
852
853 /**
854 * Removes and cancels all known tasks, ignoring any exceptions.
855 */
856 final void cancelAll() {
857 ForkJoinTask.cancelIgnoringExceptions(currentJoin);
858 ForkJoinTask.cancelIgnoringExceptions(currentSteal);
859 for (ForkJoinTask<?> t; (t = poll()) != null; )
860 ForkJoinTask.cancelIgnoringExceptions(t);
861 }
862
863 /**
864 * Computes next value for random probes. Scans don't require
865 * a very high quality generator, but also not a crummy one.
866 * Marsaglia xor-shift is cheap and works well enough. Note:
867 * This is manually inlined in its usages in ForkJoinPool to
868 * avoid writes inside busy scan loops.
869 */
870 final int nextSeed() {
871 int r = seed;
872 r ^= r << 13;
873 r ^= r >>> 17;
874 return seed = r ^= r << 5;
875 }
876
877 // Specialized execution methods
878
879 /**
880 * Pops and runs tasks until empty.
881 */
882 private void popAndExecAll() {
883 // A bit faster than repeated pop calls
884 ForkJoinTask<?>[] a; int m, s; long j; ForkJoinTask<?> t;
885 while ((a = array) != null && (m = a.length - 1) >= 0 &&
886 (s = top - 1) - base >= 0 &&
887 (t = ((ForkJoinTask<?>)
888 U.getObject(a, j = ((m & s) << ASHIFT) + ABASE)))
889 != null) {
890 if (U.compareAndSwapObject(a, j, t, null)) {
891 top = s;
892 t.doExec();
893 }
894 }
895 }
896
897 /**
898 * Polls and runs tasks until empty.
899 */
900 private void pollAndExecAll() {
901 for (ForkJoinTask<?> t; (t = poll()) != null;)
902 t.doExec();
903 }
904
905 /**
906 * If present, removes from queue and executes the given task,
907 * or any other cancelled task. Returns (true) on any CAS
908 * or consistency check failure so caller can retry.
909 *
910 * @return false if no progress can be made, else true
911 */
912 final boolean tryRemoveAndExec(ForkJoinTask<?> task) {
913 boolean stat = true, removed = false, empty = true;
914 ForkJoinTask<?>[] a; int m, s, b, n;
915 if ((a = array) != null && (m = a.length - 1) >= 0 &&
916 (n = (s = top) - (b = base)) > 0) {
917 for (ForkJoinTask<?> t;;) { // traverse from s to b
918 int j = ((--s & m) << ASHIFT) + ABASE;
919 t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
920 if (t == null) // inconsistent length
921 break;
922 else if (t == task) {
923 if (s + 1 == top) { // pop
924 if (!U.compareAndSwapObject(a, j, task, null))
925 break;
926 top = s;
927 removed = true;
928 }
929 else if (base == b) // replace with proxy
930 removed = U.compareAndSwapObject(a, j, task,
931 new EmptyTask());
932 break;
933 }
934 else if (t.status >= 0)
935 empty = false;
936 else if (s + 1 == top) { // pop and throw away
937 if (U.compareAndSwapObject(a, j, t, null))
938 top = s;
939 break;
940 }
941 if (--n == 0) {
942 if (!empty && base == b)
943 stat = false;
944 break;
945 }
946 }
947 }
948 if (removed)
949 task.doExec();
950 return stat;
951 }
952
953 /**
954 * Polls for and executes the given task or any other task in
955 * its CountedCompleter computation.
956 */
957 final boolean pollAndExecCC(ForkJoinTask<?> root) {
958 ForkJoinTask<?>[] a; int b; Object o;
959 outer: while ((b = base) - top < 0 && (a = array) != null) {
960 long j = (((a.length - 1) & b) << ASHIFT) + ABASE;
961 if ((o = U.getObject(a, j)) == null ||
962 !(o instanceof CountedCompleter))
963 break;
964 for (CountedCompleter<?> t = (CountedCompleter<?>)o, r = t;;) {
965 if (r == root) {
966 if (base == b &&
967 U.compareAndSwapObject(a, j, t, null)) {
968 base = b + 1;
969 t.doExec();
970 return true;
971 }
972 else
973 break; // restart
974 }
975 if ((r = r.completer) == null)
976 break outer; // not part of root computation
977 }
978 }
979 return false;
980 }
981
982 /**
983 * Executes a top-level task and any local tasks remaining
984 * after execution.
985 */
986 final void runTask(ForkJoinTask<?> t) {
987 if (t != null) {
988 (currentSteal = t).doExec();
989 currentSteal = null;
990 ++nsteals;
991 if (base - top < 0) { // process remaining local tasks
992 if (mode == 0)
993 popAndExecAll();
994 else
995 pollAndExecAll();
996 }
997 }
998 }
999
1000 /**
1001 * Executes a non-top-level (stolen) task.
1002 */
1003 final void runSubtask(ForkJoinTask<?> t) {
1004 if (t != null) {
1005 ForkJoinTask<?> ps = currentSteal;
1006 (currentSteal = t).doExec();
1007 currentSteal = ps;
1008 }
1009 }
1010
1011 /**
1012 * Returns true if owned and not known to be blocked.
1013 */
1014 final boolean isApparentlyUnblocked() {
1015 Thread wt; Thread.State s;
1016 return (eventCount >= 0 &&
1017 (wt = owner) != null &&
1018 (s = wt.getState()) != Thread.State.BLOCKED &&
1019 s != Thread.State.WAITING &&
1020 s != Thread.State.TIMED_WAITING);
1021 }
1022
1023 // Unsafe mechanics
1024 private static final sun.misc.Unsafe U;
1025 private static final long QLOCK;
1026 private static final int ABASE;
1027 private static final int ASHIFT;
1028 static {
1029 try {
1030 U = getUnsafe();
1031 Class<?> k = WorkQueue.class;
1032 Class<?> ak = ForkJoinTask[].class;
1033 QLOCK = U.objectFieldOffset
1034 (k.getDeclaredField("qlock"));
1035 ABASE = U.arrayBaseOffset(ak);
1036 int scale = U.arrayIndexScale(ak);
1037 if ((scale & (scale - 1)) != 0)
1038 throw new Error("data type scale not a power of two");
1039 ASHIFT = 31 - Integer.numberOfLeadingZeros(scale);
1040 } catch (Exception e) {
1041 throw new Error(e);
1042 }
1043 }
1044 }
1045
1046 // static fields (initialized in static initializer below)
1047
1048 /**
1049 * Creates a new ForkJoinWorkerThread. This factory is used unless
1050 * overridden in ForkJoinPool constructors.
1051 */
1052 public static final ForkJoinWorkerThreadFactory
1053 defaultForkJoinWorkerThreadFactory;
1054
1055 /**
1056 * Per-thread submission bookkeeping. Shared across all pools
1057 * to reduce ThreadLocal pollution and because random motion
1058 * to avoid contention in one pool is likely to hold for others.
1059 * Lazily initialized on first submission (but null-checked
1060 * in other contexts to avoid unnecessary initialization).
1061 */
1062 static final ThreadLocal<Submitter> submitters;
1063
1064 /**
1065 * Permission required for callers of methods that may start or
1066 * kill threads.
1067 */
1068 private static final RuntimePermission modifyThreadPermission;
1069
1070 /**
1071 * Common (static) pool. Non-null for public use unless a static
1072 * construction exception, but internal usages null-check on use
1073 * to paranoically avoid potential initialization circularities
1074 * as well as to simplify generated code.
1075 */
1076 static final ForkJoinPool common;
1077
1078 /**
1079 * Common pool parallelism. Must equal common.parallelism.
1080 */
1081 static final int commonParallelism;
1082
1083 /**
1084 * Sequence number for creating workerNamePrefix.
1085 */
1086 private static int poolNumberSequence;
1087
1088 /**
1089 * Returns the next sequence number. We don't expect this to
1090 * ever contend, so use simple builtin sync.
1091 */
1092 private static final synchronized int nextPoolId() {
1093 return ++poolNumberSequence;
1094 }
1095
1096 // static constants
1097
1098 /**
1099 * Initial timeout value (in nanoseconds) for the thread
1100 * triggering quiescence to park waiting for new work. On timeout,
1101 * the thread will instead try to shrink the number of
1102 * workers. The value should be large enough to avoid overly
1103 * aggressive shrinkage during most transient stalls (long GCs
1104 * etc).
1105 */
1106 private static final long IDLE_TIMEOUT = 2000L * 1000L * 1000L; // 2sec
1107
1108 /**
1109 * Timeout value when there are more threads than parallelism level
1110 */
1111 private static final long FAST_IDLE_TIMEOUT = 200L * 1000L * 1000L;
1112
1113 /**
1114 * Tolerance for idle timeouts, to cope with timer undershoots
1115 */
1116 private static final long TIMEOUT_SLOP = 2000000L;
1117
1118 /**
1119 * The maximum stolen->joining link depth allowed in method
1120 * tryHelpStealer. Must be a power of two. Depths for legitimate
1121 * chains are unbounded, but we use a fixed constant to avoid
1122 * (otherwise unchecked) cycles and to bound staleness of
1123 * traversal parameters at the expense of sometimes blocking when
1124 * we could be helping.
1125 */
1126 private static final int MAX_HELP = 64;
1127
1128 /**
1129 * Increment for seed generators. See class ThreadLocal for
1130 * explanation.
1131 */
1132 private static final int SEED_INCREMENT = 0x61c88647;
1133
1134 /*
1135 * Bits and masks for control variables
1136 *
1137 * Field ctl is a long packed with:
1138 * AC: Number of active running workers minus target parallelism (16 bits)
1139 * TC: Number of total workers minus target parallelism (16 bits)
1140 * ST: true if pool is terminating (1 bit)
1141 * EC: the wait count of top waiting thread (15 bits)
1142 * ID: poolIndex of top of Treiber stack of waiters (16 bits)
1143 *
1144 * When convenient, we can extract the upper 32 bits of counts and
1145 * the lower 32 bits of queue state, u = (int)(ctl >>> 32) and e =
1146 * (int)ctl. The ec field is never accessed alone, but always
1147 * together with id and st. The offsets of counts by the target
1148 * parallelism and the positionings of fields makes it possible to
1149 * perform the most common checks via sign tests of fields: When
1150 * ac is negative, there are not enough active workers, when tc is
1151 * negative, there are not enough total workers, and when e is
1152 * negative, the pool is terminating. To deal with these possibly
1153 * negative fields, we use casts in and out of "short" and/or
1154 * signed shifts to maintain signedness.
1155 *
1156 * When a thread is queued (inactivated), its eventCount field is
1157 * set negative, which is the only way to tell if a worker is
1158 * prevented from executing tasks, even though it must continue to
1159 * scan for them to avoid queuing races. Note however that
1160 * eventCount updates lag releases so usage requires care.
1161 *
1162 * Field plock is an int packed with:
1163 * SHUTDOWN: true if shutdown is enabled (1 bit)
1164 * SEQ: a sequence lock, with PL_LOCK bit set if locked (30 bits)
1165 * SIGNAL: set when threads may be waiting on the lock (1 bit)
1166 *
1167 * The sequence number enables simple consistency checks:
1168 * Staleness of read-only operations on the workQueues array can
1169 * be checked by comparing plock before vs after the reads.
1170 */
1171
1172 // bit positions/shifts for fields
1173 private static final int AC_SHIFT = 48;
1174 private static final int TC_SHIFT = 32;
1175 private static final int ST_SHIFT = 31;
1176 private static final int EC_SHIFT = 16;
1177
1178 // bounds
1179 private static final int SMASK = 0xffff; // short bits
1180 private static final int MAX_CAP = 0x7fff; // max #workers - 1
1181 private static final int EVENMASK = 0xfffe; // even short bits
1182 private static final int SQMASK = 0x007e; // max 64 (even) slots
1183 private static final int SHORT_SIGN = 1 << 15;
1184 private static final int INT_SIGN = 1 << 31;
1185
1186 // masks
1187 private static final long STOP_BIT = 0x0001L << ST_SHIFT;
1188 private static final long AC_MASK = ((long)SMASK) << AC_SHIFT;
1189 private static final long TC_MASK = ((long)SMASK) << TC_SHIFT;
1190
1191 // units for incrementing and decrementing
1192 private static final long TC_UNIT = 1L << TC_SHIFT;
1193 private static final long AC_UNIT = 1L << AC_SHIFT;
1194
1195 // masks and units for dealing with u = (int)(ctl >>> 32)
1196 private static final int UAC_SHIFT = AC_SHIFT - 32;
1197 private static final int UTC_SHIFT = TC_SHIFT - 32;
1198 private static final int UAC_MASK = SMASK << UAC_SHIFT;
1199 private static final int UTC_MASK = SMASK << UTC_SHIFT;
1200 private static final int UAC_UNIT = 1 << UAC_SHIFT;
1201 private static final int UTC_UNIT = 1 << UTC_SHIFT;
1202
1203 // masks and units for dealing with e = (int)ctl
1204 private static final int E_MASK = 0x7fffffff; // no STOP_BIT
1205 private static final int E_SEQ = 1 << EC_SHIFT;
1206
1207 // plock bits
1208 private static final int SHUTDOWN = 1 << 31;
1209 private static final int PL_LOCK = 2;
1210 private static final int PL_SIGNAL = 1;
1211 private static final int PL_SPINS = 1 << 8;
1212
1213 // access mode for WorkQueue
1214 static final int LIFO_QUEUE = 0;
1215 static final int FIFO_QUEUE = 1;
1216 static final int SHARED_QUEUE = -1;
1217
1218 // bounds for #steps in scan loop -- must be power 2 minus 1
1219 private static final int MIN_SCAN = 0x1ff; // cover estimation slop
1220 private static final int MAX_SCAN = 0x1ffff; // 4 * max workers
1221
1222 // Instance fields
1223
1224 /*
1225 * Field layout of this class tends to matter more than one would
1226 * like. Runtime layout order is only loosely related to
1227 * declaration order and may differ across JVMs, but the following
1228 * empirically works OK on current JVMs.
1229 */
1230
1231 // Heuristic padding to ameliorate unfortunate memory placements
1232 volatile long pad00, pad01, pad02, pad03, pad04, pad05, pad06;
1233
1234 volatile long stealCount; // collects worker counts
1235 volatile long ctl; // main pool control
1236 volatile int plock; // shutdown status and seqLock
1237 volatile int indexSeed; // worker/submitter index seed
1238 final int config; // mode and parallelism level
1239 WorkQueue[] workQueues; // main registry
1240 final ForkJoinWorkerThreadFactory factory;
1241 final Thread.UncaughtExceptionHandler ueh; // per-worker UEH
1242 final String workerNamePrefix; // to create worker name string
1243
1244 volatile Object pad10, pad11, pad12, pad13, pad14, pad15, pad16, pad17;
1245 volatile Object pad18, pad19, pad1a, pad1b;
1246
1247 /**
1248 * Acquires the plock lock to protect worker array and related
1249 * updates. This method is called only if an initial CAS on plock
1250 * fails. This acts as a spinlock for normal cases, but falls back
1251 * to builtin monitor to block when (rarely) needed. This would be
1252 * a terrible idea for a highly contended lock, but works fine as
1253 * a more conservative alternative to a pure spinlock.
1254 */
1255 private int acquirePlock() {
1256 int spins = PL_SPINS, r = 0, ps, nps;
1257 for (;;) {
1258 if (((ps = plock) & PL_LOCK) == 0 &&
1259 U.compareAndSwapInt(this, PLOCK, ps, nps = ps + PL_LOCK))
1260 return nps;
1261 else if (r == 0) { // randomize spins if possible
1262 Thread t = Thread.currentThread(); WorkQueue w; Submitter z;
1263 if ((t instanceof ForkJoinWorkerThread) &&
1264 (w = ((ForkJoinWorkerThread)t).workQueue) != null)
1265 r = w.seed;
1266 else if ((z = submitters.get()) != null)
1267 r = z.seed;
1268 else
1269 r = 1;
1270 }
1271 else if (spins >= 0) {
1272 r ^= r << 1; r ^= r >>> 3; r ^= r << 10; // xorshift
1273 if (r >= 0)
1274 --spins;
1275 }
1276 else if (U.compareAndSwapInt(this, PLOCK, ps, ps | PL_SIGNAL)) {
1277 synchronized (this) {
1278 if ((plock & PL_SIGNAL) != 0) {
1279 try {
1280 wait();
1281 } catch (InterruptedException ie) {
1282 try {
1283 Thread.currentThread().interrupt();
1284 } catch (SecurityException ignore) {
1285 }
1286 }
1287 }
1288 else
1289 notifyAll();
1290 }
1291 }
1292 }
1293 }
1294
1295 /**
1296 * Unlocks and signals any thread waiting for plock. Called only
1297 * when CAS of seq value for unlock fails.
1298 */
1299 private void releasePlock(int ps) {
1300 plock = ps;
1301 synchronized (this) { notifyAll(); }
1302 }
1303
1304 /**
1305 * Tries to create and start one worker if fewer than target
1306 * parallelism level exist. Adjusts counts etc on failure.
1307 */
1308 private void tryAddWorker() {
1309 long c; int u;
1310 while ((u = (int)((c = ctl) >>> 32)) < 0 &&
1311 (u & SHORT_SIGN) != 0 && (int)c == 0) {
1312 long nc = (long)(((u + UTC_UNIT) & UTC_MASK) |
1313 ((u + UAC_UNIT) & UAC_MASK)) << 32;
1314 if (U.compareAndSwapLong(this, CTL, c, nc)) {
1315 ForkJoinWorkerThreadFactory fac;
1316 Throwable ex = null;
1317 ForkJoinWorkerThread wt = null;
1318 try {
1319 if ((fac = factory) != null &&
1320 (wt = fac.newThread(this)) != null) {
1321 wt.start();
1322 break;
1323 }
1324 } catch (Throwable e) {
1325 ex = e;
1326 }
1327 deregisterWorker(wt, ex);
1328 break;
1329 }
1330 }
1331 }
1332
1333 // Registering and deregistering workers
1334
1335 /**
1336 * Callback from ForkJoinWorkerThread to establish and record its
1337 * WorkQueue. To avoid scanning bias due to packing entries in
1338 * front of the workQueues array, we treat the array as a simple
1339 * power-of-two hash table using per-thread seed as hash,
1340 * expanding as needed.
1341 *
1342 * @param wt the worker thread
1343 * @return the worker's queue
1344 */
1345 final WorkQueue registerWorker(ForkJoinWorkerThread wt) {
1346 Thread.UncaughtExceptionHandler handler; WorkQueue[] ws; int s, ps;
1347 wt.setDaemon(true);
1348 if ((handler = ueh) != null)
1349 wt.setUncaughtExceptionHandler(handler);
1350 do {} while (!U.compareAndSwapInt(this, INDEXSEED, s = indexSeed,
1351 s += SEED_INCREMENT) ||
1352 s == 0); // skip 0
1353 WorkQueue w = new WorkQueue(this, wt, config >>> 16, s);
1354 if (((ps = plock) & PL_LOCK) != 0 ||
1355 !U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1356 ps = acquirePlock();
1357 int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1358 try {
1359 if ((ws = workQueues) != null) { // skip if shutting down
1360 int n = ws.length, m = n - 1;
1361 int r = (s << 1) | 1; // use odd-numbered indices
1362 if (ws[r &= m] != null) { // collision
1363 int probes = 0; // step by approx half size
1364 int step = (n <= 4) ? 2 : ((n >>> 1) & EVENMASK) + 2;
1365 while (ws[r = (r + step) & m] != null) {
1366 if (++probes >= n) {
1367 workQueues = ws = Arrays.copyOf(ws, n <<= 1);
1368 m = n - 1;
1369 probes = 0;
1370 }
1371 }
1372 }
1373 w.eventCount = w.poolIndex = r; // volatile write orders
1374 ws[r] = w;
1375 }
1376 } finally {
1377 if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1378 releasePlock(nps);
1379 }
1380 wt.setName(workerNamePrefix.concat(Integer.toString(w.poolIndex)));
1381 return w;
1382 }
1383
1384 /**
1385 * Final callback from terminating worker, as well as upon failure
1386 * to construct or start a worker. Removes record of worker from
1387 * array, and adjusts counts. If pool is shutting down, tries to
1388 * complete termination.
1389 *
1390 * @param wt the worker thread or null if construction failed
1391 * @param ex the exception causing failure, or null if none
1392 */
1393 final void deregisterWorker(ForkJoinWorkerThread wt, Throwable ex) {
1394 WorkQueue w = null;
1395 if (wt != null && (w = wt.workQueue) != null) {
1396 int ps;
1397 w.qlock = -1; // ensure set
1398 long ns = w.nsteals, sc; // collect steal count
1399 do {} while (!U.compareAndSwapLong(this, STEALCOUNT,
1400 sc = stealCount, sc + ns));
1401 if (((ps = plock) & PL_LOCK) != 0 ||
1402 !U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1403 ps = acquirePlock();
1404 int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1405 try {
1406 int idx = w.poolIndex;
1407 WorkQueue[] ws = workQueues;
1408 if (ws != null && idx >= 0 && idx < ws.length && ws[idx] == w)
1409 ws[idx] = null;
1410 } finally {
1411 if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1412 releasePlock(nps);
1413 }
1414 }
1415
1416 long c; // adjust ctl counts
1417 do {} while (!U.compareAndSwapLong
1418 (this, CTL, c = ctl, (((c - AC_UNIT) & AC_MASK) |
1419 ((c - TC_UNIT) & TC_MASK) |
1420 (c & ~(AC_MASK|TC_MASK)))));
1421
1422 if (!tryTerminate(false, false) && w != null && w.array != null) {
1423 w.cancelAll(); // cancel remaining tasks
1424 WorkQueue[] ws; WorkQueue v; Thread p; int u, i, e;
1425 while ((u = (int)((c = ctl) >>> 32)) < 0 && (e = (int)c) >= 0) {
1426 if (e > 0) { // activate or create replacement
1427 if ((ws = workQueues) == null ||
1428 (i = e & SMASK) >= ws.length ||
1429 (v = ws[i]) == null)
1430 break;
1431 long nc = (((long)(v.nextWait & E_MASK)) |
1432 ((long)(u + UAC_UNIT) << 32));
1433 if (v.eventCount != (e | INT_SIGN))
1434 break;
1435 if (U.compareAndSwapLong(this, CTL, c, nc)) {
1436 v.eventCount = (e + E_SEQ) & E_MASK;
1437 if ((p = v.parker) != null)
1438 U.unpark(p);
1439 break;
1440 }
1441 }
1442 else {
1443 if ((short)u < 0)
1444 tryAddWorker();
1445 break;
1446 }
1447 }
1448 }
1449 if (ex == null) // help clean refs on way out
1450 ForkJoinTask.helpExpungeStaleExceptions();
1451 else // rethrow
1452 ForkJoinTask.rethrow(ex);
1453 }
1454
1455 // Submissions
1456
1457 /**
1458 * Unless shutting down, adds the given task to a submission queue
1459 * at submitter's current queue index (modulo submission
1460 * range). Only the most common path is directly handled in this
1461 * method. All others are relayed to fullExternalPush.
1462 *
1463 * @param task the task. Caller must ensure non-null.
1464 */
1465 final void externalPush(ForkJoinTask<?> task) {
1466 WorkQueue[] ws; WorkQueue q; Submitter z; int m; ForkJoinTask<?>[] a;
1467 if ((z = submitters.get()) != null && plock > 0 &&
1468 (ws = workQueues) != null && (m = (ws.length - 1)) >= 0 &&
1469 (q = ws[m & z.seed & SQMASK]) != null &&
1470 U.compareAndSwapInt(q, QLOCK, 0, 1)) { // lock
1471 int b = q.base, s = q.top, n, an;
1472 if ((a = q.array) != null && (an = a.length) > (n = s + 1 - b)) {
1473 int j = (((an - 1) & s) << ASHIFT) + ABASE;
1474 U.putOrderedObject(a, j, task);
1475 q.top = s + 1; // push on to deque
1476 q.qlock = 0;
1477 if (n <= 2)
1478 signalWork(q);
1479 return;
1480 }
1481 q.qlock = 0;
1482 }
1483 fullExternalPush(task);
1484 }
1485
1486 /**
1487 * Full version of externalPush. This method is called, among
1488 * other times, upon the first submission of the first task to the
1489 * pool, so must perform secondary initialization. It also
1490 * detects first submission by an external thread by looking up
1491 * its ThreadLocal, and creates a new shared queue if the one at
1492 * index if empty or contended. The plock lock body must be
1493 * exception-free (so no try/finally) so we optimistically
1494 * allocate new queues outside the lock and throw them away if
1495 * (very rarely) not needed.
1496 *
1497 * Secondary initialization occurs when plock is zero, to create
1498 * workQueue array and set plock to a valid value. This lock body
1499 * must also be exception-free. Because the plock seq value can
1500 * eventually wrap around zero, this method harmlessly fails to
1501 * reinitialize if workQueues exists, while still advancing plock.
1502 */
1503 private void fullExternalPush(ForkJoinTask<?> task) {
1504 int r = 0; // random index seed
1505 for (Submitter z = submitters.get();;) {
1506 WorkQueue[] ws; WorkQueue q; int ps, m, k;
1507 if (z == null) {
1508 if (U.compareAndSwapInt(this, INDEXSEED, r = indexSeed,
1509 r += SEED_INCREMENT) && r != 0)
1510 submitters.set(z = new Submitter(r));
1511 }
1512 else if (r == 0) { // move to a different index
1513 r = z.seed;
1514 r ^= r << 13; // same xorshift as WorkQueues
1515 r ^= r >>> 17;
1516 z.seed = r ^ (r << 5);
1517 }
1518 else if ((ps = plock) < 0)
1519 throw new RejectedExecutionException();
1520 else if (ps == 0 || (ws = workQueues) == null ||
1521 (m = ws.length - 1) < 0) { // initialize workQueues
1522 int p = config & SMASK; // find power of two table size
1523 int n = (p > 1) ? p - 1 : 1; // ensure at least 2 slots
1524 n |= n >>> 1;
1525 n |= n >>> 2;
1526 n |= n >>> 4;
1527 n |= n >>> 8;
1528 n |= n >>> 16;
1529 n = (n + 1) << 1;
1530 WorkQueue[] nws = ((ws = workQueues) == null || ws.length == 0 ?
1531 new WorkQueue[n] : null);
1532 if (((ps = plock) & PL_LOCK) != 0 ||
1533 !U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1534 ps = acquirePlock();
1535 if (((ws = workQueues) == null || ws.length == 0) && nws != null)
1536 workQueues = nws;
1537 int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1538 if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1539 releasePlock(nps);
1540 }
1541 else if ((q = ws[k = r & m & SQMASK]) != null) {
1542 if (q.qlock == 0 && U.compareAndSwapInt(q, QLOCK, 0, 1)) {
1543 ForkJoinTask<?>[] a = q.array;
1544 int s = q.top;
1545 boolean submitted = false;
1546 try { // locked version of push
1547 if ((a != null && a.length > s + 1 - q.base) ||
1548 (a = q.growArray()) != null) { // must presize
1549 int j = (((a.length - 1) & s) << ASHIFT) + ABASE;
1550 U.putOrderedObject(a, j, task);
1551 q.top = s + 1;
1552 submitted = true;
1553 }
1554 } finally {
1555 q.qlock = 0; // unlock
1556 }
1557 if (submitted) {
1558 signalWork(q);
1559 return;
1560 }
1561 }
1562 r = 0; // move on failure
1563 }
1564 else if (((ps = plock) & PL_LOCK) == 0) { // create new queue
1565 q = new WorkQueue(this, null, SHARED_QUEUE, r);
1566 if (((ps = plock) & PL_LOCK) != 0 ||
1567 !U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1568 ps = acquirePlock();
1569 if ((ws = workQueues) != null && k < ws.length && ws[k] == null)
1570 ws[k] = q;
1571 int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN);
1572 if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1573 releasePlock(nps);
1574 }
1575 else
1576 r = 0; // try elsewhere while lock held
1577 }
1578 }
1579
1580 // Maintaining ctl counts
1581
1582 /**
1583 * Increments active count; mainly called upon return from blocking.
1584 */
1585 final void incrementActiveCount() {
1586 long c;
1587 do {} while (!U.compareAndSwapLong(this, CTL, c = ctl, c + AC_UNIT));
1588 }
1589
1590 /**
1591 * Tries to create or activate a worker if too few are active.
1592 *
1593 * @param q the (non-null) queue holding tasks to be signalled
1594 */
1595 final void signalWork(WorkQueue q) {
1596 int hint = q.poolIndex;
1597 long c; int e, u, i, n; WorkQueue[] ws; WorkQueue w; Thread p;
1598 while ((u = (int)((c = ctl) >>> 32)) < 0) {
1599 if ((e = (int)c) > 0) {
1600 if ((ws = workQueues) != null && ws.length > (i = e & SMASK) &&
1601 (w = ws[i]) != null && w.eventCount == (e | INT_SIGN)) {
1602 long nc = (((long)(w.nextWait & E_MASK)) |
1603 ((long)(u + UAC_UNIT) << 32));
1604 if (U.compareAndSwapLong(this, CTL, c, nc)) {
1605 w.hint = hint;
1606 w.eventCount = (e + E_SEQ) & E_MASK;
1607 if ((p = w.parker) != null)
1608 U.unpark(p);
1609 break;
1610 }
1611 if (q.top - q.base <= 0)
1612 break;
1613 }
1614 else
1615 break;
1616 }
1617 else {
1618 if ((short)u < 0)
1619 tryAddWorker();
1620 break;
1621 }
1622 }
1623 }
1624
1625 // Scanning for tasks
1626
1627 /**
1628 * Top-level runloop for workers, called by ForkJoinWorkerThread.run.
1629 */
1630 final void runWorker(WorkQueue w) {
1631 w.growArray(); // allocate queue
1632 do { w.runTask(scan(w)); } while (w.qlock >= 0);
1633 }
1634
1635 /**
1636 * Scans for and, if found, returns one task, else possibly
1637 * inactivates the worker. This method operates on single reads of
1638 * volatile state and is designed to be re-invoked continuously,
1639 * in part because it returns upon detecting inconsistencies,
1640 * contention, or state changes that indicate possible success on
1641 * re-invocation.
1642 *
1643 * The scan searches for tasks across queues (starting at a random
1644 * index, and relying on registerWorker to irregularly scatter
1645 * them within array to avoid bias), checking each at least twice.
1646 * The scan terminates upon either finding a non-empty queue, or
1647 * completing the sweep. If the worker is not inactivated, it
1648 * takes and returns a task from this queue. Otherwise, if not
1649 * activated, it signals workers (that may include itself) and
1650 * returns so caller can retry. Also returns for true if the
1651 * worker array may have changed during an empty scan. On failure
1652 * to find a task, we take one of the following actions, after
1653 * which the caller will retry calling this method unless
1654 * terminated.
1655 *
1656 * * If pool is terminating, terminate the worker.
1657 *
1658 * * If not already enqueued, try to inactivate and enqueue the
1659 * worker on wait queue. Or, if inactivating has caused the pool
1660 * to be quiescent, relay to idleAwaitWork to possibly shrink
1661 * pool.
1662 *
1663 * * If already enqueued and none of the above apply, possibly
1664 * park awaiting signal, else lingering to help scan and signal.
1665 *
1666 * * If a non-empty queue discovered or left as a hint,
1667 * help wake up other workers before return.
1668 *
1669 * @param w the worker (via its WorkQueue)
1670 * @return a task or null if none found
1671 */
1672 private final ForkJoinTask<?> scan(WorkQueue w) {
1673 WorkQueue[] ws; int m;
1674 int ps = plock; // read plock before ws
1675 if (w != null && (ws = workQueues) != null && (m = ws.length - 1) >= 0) {
1676 int ec = w.eventCount; // ec is negative if inactive
1677 int r = w.seed; r ^= r << 13; r ^= r >>> 17; w.seed = r ^= r << 5;
1678 w.hint = -1; // update seed and clear hint
1679 int j = ((m + m + 1) | MIN_SCAN) & MAX_SCAN;
1680 do {
1681 WorkQueue q; ForkJoinTask<?>[] a; int b;
1682 if ((q = ws[(r + j) & m]) != null && (b = q.base) - q.top < 0 &&
1683 (a = q.array) != null) { // probably nonempty
1684 int i = (((a.length - 1) & b) << ASHIFT) + ABASE;
1685 ForkJoinTask<?> t = (ForkJoinTask<?>)
1686 U.getObjectVolatile(a, i);
1687 if (q.base == b && ec >= 0 && t != null &&
1688 U.compareAndSwapObject(a, i, t, null)) {
1689 if ((q.base = b + 1) - q.top < 0)
1690 signalWork(q);
1691 return t; // taken
1692 }
1693 else if ((ec < 0 || j < m) && (int)(ctl >> AC_SHIFT) <= 0) {
1694 w.hint = (r + j) & m; // help signal below
1695 break; // cannot take
1696 }
1697 }
1698 } while (--j >= 0);
1699
1700 int h, e, ns; long c, sc; WorkQueue q;
1701 if ((ns = w.nsteals) != 0) {
1702 if (U.compareAndSwapLong(this, STEALCOUNT,
1703 sc = stealCount, sc + ns))
1704 w.nsteals = 0; // collect steals and rescan
1705 }
1706 else if (plock != ps) // consistency check
1707 ; // skip
1708 else if ((e = (int)(c = ctl)) < 0)
1709 w.qlock = -1; // pool is terminating
1710 else {
1711 if ((h = w.hint) < 0) {
1712 if (ec >= 0) { // try to enqueue/inactivate
1713 long nc = (((long)ec |
1714 ((c - AC_UNIT) & (AC_MASK|TC_MASK))));
1715 w.nextWait = e; // link and mark inactive
1716 w.eventCount = ec | INT_SIGN;
1717 if (ctl != c || !U.compareAndSwapLong(this, CTL, c, nc))
1718 w.eventCount = ec; // unmark on CAS failure
1719 else if ((int)(c >> AC_SHIFT) == 1 - (config & SMASK))
1720 idleAwaitWork(w, nc, c);
1721 }
1722 else if (w.eventCount < 0 && ctl == c) {
1723 Thread wt = Thread.currentThread();
1724 Thread.interrupted(); // clear status
1725 U.putObject(wt, PARKBLOCKER, this);
1726 w.parker = wt; // emulate LockSupport.park
1727 if (w.eventCount < 0) // recheck
1728 U.park(false, 0L); // block
1729 w.parker = null;
1730 U.putObject(wt, PARKBLOCKER, null);
1731 }
1732 }
1733 if ((h >= 0 || (h = w.hint) >= 0) &&
1734 (ws = workQueues) != null && h < ws.length &&
1735 (q = ws[h]) != null) { // signal others before retry
1736 WorkQueue v; Thread p; int u, i, s;
1737 for (int n = (config & SMASK) - 1;;) {
1738 int idleCount = (w.eventCount < 0) ? 0 : -1;
1739 if (((s = idleCount - q.base + q.top) <= n &&
1740 (n = s) <= 0) ||
1741 (u = (int)((c = ctl) >>> 32)) >= 0 ||
1742 (e = (int)c) <= 0 || m < (i = e & SMASK) ||
1743 (v = ws[i]) == null)
1744 break;
1745 long nc = (((long)(v.nextWait & E_MASK)) |
1746 ((long)(u + UAC_UNIT) << 32));
1747 if (v.eventCount != (e | INT_SIGN) ||
1748 !U.compareAndSwapLong(this, CTL, c, nc))
1749 break;
1750 v.hint = h;
1751 v.eventCount = (e + E_SEQ) & E_MASK;
1752 if ((p = v.parker) != null)
1753 U.unpark(p);
1754 if (--n <= 0)
1755 break;
1756 }
1757 }
1758 }
1759 }
1760 return null;
1761 }
1762
1763 /**
1764 * If inactivating worker w has caused the pool to become
1765 * quiescent, checks for pool termination, and, so long as this is
1766 * not the only worker, waits for event for up to a given
1767 * duration. On timeout, if ctl has not changed, terminates the
1768 * worker, which will in turn wake up another worker to possibly
1769 * repeat this process.
1770 *
1771 * @param w the calling worker
1772 * @param currentCtl the ctl value triggering possible quiescence
1773 * @param prevCtl the ctl value to restore if thread is terminated
1774 */
1775 private void idleAwaitWork(WorkQueue w, long currentCtl, long prevCtl) {
1776 if (w != null && w.eventCount < 0 &&
1777 !tryTerminate(false, false) && (int)prevCtl != 0 &&
1778 ctl == currentCtl) {
1779 int dc = -(short)(currentCtl >>> TC_SHIFT);
1780 long parkTime = dc < 0 ? FAST_IDLE_TIMEOUT: (dc + 1) * IDLE_TIMEOUT;
1781 long deadline = System.nanoTime() + parkTime - TIMEOUT_SLOP;
1782 Thread wt = Thread.currentThread();
1783 while (ctl == currentCtl) {
1784 Thread.interrupted(); // timed variant of version in scan()
1785 U.putObject(wt, PARKBLOCKER, this);
1786 w.parker = wt;
1787 if (ctl == currentCtl)
1788 U.park(false, parkTime);
1789 w.parker = null;
1790 U.putObject(wt, PARKBLOCKER, null);
1791 if (ctl != currentCtl)
1792 break;
1793 if (deadline - System.nanoTime() <= 0L &&
1794 U.compareAndSwapLong(this, CTL, currentCtl, prevCtl)) {
1795 w.eventCount = (w.eventCount + E_SEQ) | E_MASK;
1796 w.hint = -1;
1797 w.qlock = -1; // shrink
1798 break;
1799 }
1800 }
1801 }
1802 }
1803
1804 /**
1805 * Scans through queues looking for work while joining a task; if
1806 * any present, signals. May return early if more signalling is
1807 * detectably unneeded.
1808 *
1809 * @param task return early if done
1810 * @param origin an index to start scan
1811 */
1812 private void helpSignal(ForkJoinTask<?> task, int origin) {
1813 WorkQueue[] ws; WorkQueue w; Thread p; long c; int m, u, e, i, s;
1814 if (task != null && task.status >= 0 &&
1815 (u = (int)(ctl >>> 32)) < 0 && (u >> UAC_SHIFT) < 0 &&
1816 (ws = workQueues) != null && (m = ws.length - 1) >= 0) {
1817 outer: for (int k = origin, j = m; j >= 0; --j) {
1818 WorkQueue q = ws[k++ & m];
1819 for (int n = m;;) { // limit to at most m signals
1820 if (task.status < 0)
1821 break outer;
1822 if (q == null ||
1823 ((s = -q.base + q.top) <= n && (n = s) <= 0))
1824 break;
1825 if ((u = (int)((c = ctl) >>> 32)) >= 0 ||
1826 (e = (int)c) <= 0 || m < (i = e & SMASK) ||
1827 (w = ws[i]) == null)
1828 break outer;
1829 long nc = (((long)(w.nextWait & E_MASK)) |
1830 ((long)(u + UAC_UNIT) << 32));
1831 if (w.eventCount != (e | INT_SIGN))
1832 break outer;
1833 if (U.compareAndSwapLong(this, CTL, c, nc)) {
1834 w.eventCount = (e + E_SEQ) & E_MASK;
1835 if ((p = w.parker) != null)
1836 U.unpark(p);
1837 if (--n <= 0)
1838 break;
1839 }
1840 }
1841 }
1842 }
1843 }
1844
1845 /**
1846 * Tries to locate and execute tasks for a stealer of the given
1847 * task, or in turn one of its stealers, Traces currentSteal ->
1848 * currentJoin links looking for a thread working on a descendant
1849 * of the given task and with a non-empty queue to steal back and
1850 * execute tasks from. The first call to this method upon a
1851 * waiting join will often entail scanning/search, (which is OK
1852 * because the joiner has nothing better to do), but this method
1853 * leaves hints in workers to speed up subsequent calls. The
1854 * implementation is very branchy to cope with potential
1855 * inconsistencies or loops encountering chains that are stale,
1856 * unknown, or so long that they are likely cyclic.
1857 *
1858 * @param joiner the joining worker
1859 * @param task the task to join
1860 * @return 0 if no progress can be made, negative if task
1861 * known complete, else positive
1862 */
1863 private int tryHelpStealer(WorkQueue joiner, ForkJoinTask<?> task) {
1864 int stat = 0, steps = 0; // bound to avoid cycles
1865 if (joiner != null && task != null) { // hoist null checks
1866 restart: for (;;) {
1867 ForkJoinTask<?> subtask = task; // current target
1868 for (WorkQueue j = joiner, v;;) { // v is stealer of subtask
1869 WorkQueue[] ws; int m, s, h;
1870 if ((s = task.status) < 0) {
1871 stat = s;
1872 break restart;
1873 }
1874 if ((ws = workQueues) == null || (m = ws.length - 1) <= 0)
1875 break restart; // shutting down
1876 if ((v = ws[h = (j.hint | 1) & m]) == null ||
1877 v.currentSteal != subtask) {
1878 for (int origin = h;;) { // find stealer
1879 if (((h = (h + 2) & m) & 15) == 1 &&
1880 (subtask.status < 0 || j.currentJoin != subtask))
1881 continue restart; // occasional staleness check
1882 if ((v = ws[h]) != null &&
1883 v.currentSteal == subtask) {
1884 j.hint = h; // save hint
1885 break;
1886 }
1887 if (h == origin)
1888 break restart; // cannot find stealer
1889 }
1890 }
1891 for (;;) { // help stealer or descend to its stealer
1892 ForkJoinTask[] a; int b;
1893 if (subtask.status < 0) // surround probes with
1894 continue restart; // consistency checks
1895 if ((b = v.base) - v.top < 0 && (a = v.array) != null) {
1896 int i = (((a.length - 1) & b) << ASHIFT) + ABASE;
1897 ForkJoinTask<?> t =
1898 (ForkJoinTask<?>)U.getObjectVolatile(a, i);
1899 if (subtask.status < 0 || j.currentJoin != subtask ||
1900 v.currentSteal != subtask)
1901 continue restart; // stale
1902 stat = 1; // apparent progress
1903 if (t != null && v.base == b &&
1904 U.compareAndSwapObject(a, i, t, null)) {
1905 v.base = b + 1; // help stealer
1906 joiner.runSubtask(t);
1907 }
1908 else if (v.base == b && ++steps == MAX_HELP)
1909 break restart; // v apparently stalled
1910 }
1911 else { // empty -- try to descend
1912 ForkJoinTask<?> next = v.currentJoin;
1913 if (subtask.status < 0 || j.currentJoin != subtask ||
1914 v.currentSteal != subtask)
1915 continue restart; // stale
1916 else if (next == null || ++steps == MAX_HELP)
1917 break restart; // dead-end or maybe cyclic
1918 else {
1919 subtask = next;
1920 j = v;
1921 break;
1922 }
1923 }
1924 }
1925 }
1926 }
1927 }
1928 return stat;
1929 }
1930
1931 /**
1932 * Analog of tryHelpStealer for CountedCompleters. Tries to steal
1933 * and run tasks within the target's computation.
1934 *
1935 * @param task the task to join
1936 * @param mode if shared, exit upon completing any task
1937 * if all workers are active
1938 */
1939 private int helpComplete(ForkJoinTask<?> task, int mode) {
1940 WorkQueue[] ws; WorkQueue q; int m, n, s, u;
1941 if (task != null && (ws = workQueues) != null &&
1942 (m = ws.length - 1) >= 0) {
1943 for (int j = 1, origin = j;;) {
1944 if ((s = task.status) < 0)
1945 return s;
1946 if ((q = ws[j & m]) != null && q.pollAndExecCC(task)) {
1947 origin = j;
1948 if (mode == SHARED_QUEUE &&
1949 ((u = (int)(ctl >>> 32)) >= 0 || (u >> UAC_SHIFT) >= 0))
1950 break;
1951 }
1952 else if ((j = (j + 2) & m) == origin)
1953 break;
1954 }
1955 }
1956 return 0;
1957 }
1958
1959 /**
1960 * Tries to decrement active count (sometimes implicitly) and
1961 * possibly release or create a compensating worker in preparation
1962 * for blocking. Fails on contention or termination. Otherwise,
1963 * adds a new thread if no idle workers are available and pool
1964 * may become starved.
1965 */
1966 final boolean tryCompensate() {
1967 int pc = config & SMASK, e, i, tc; long c;
1968 WorkQueue[] ws; WorkQueue w; Thread p;
1969 if ((ws = workQueues) != null && (e = (int)(c = ctl)) >= 0) {
1970 if (e != 0 && (i = e & SMASK) < ws.length &&
1971 (w = ws[i]) != null && w.eventCount == (e | INT_SIGN)) {
1972 long nc = ((long)(w.nextWait & E_MASK) |
1973 (c & (AC_MASK|TC_MASK)));
1974 if (U.compareAndSwapLong(this, CTL, c, nc)) {
1975 w.eventCount = (e + E_SEQ) & E_MASK;
1976 if ((p = w.parker) != null)
1977 U.unpark(p);
1978 return true; // replace with idle worker
1979 }
1980 }
1981 else if ((tc = (short)(c >>> TC_SHIFT)) >= 0 &&
1982 (int)(c >> AC_SHIFT) + pc > 1) {
1983 long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK);
1984 if (U.compareAndSwapLong(this, CTL, c, nc))
1985 return true; // no compensation
1986 }
1987 else if (tc + pc < MAX_CAP) {
1988 long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK);
1989 if (U.compareAndSwapLong(this, CTL, c, nc)) {
1990 ForkJoinWorkerThreadFactory fac;
1991 Throwable ex = null;
1992 ForkJoinWorkerThread wt = null;
1993 try {
1994 if ((fac = factory) != null &&
1995 (wt = fac.newThread(this)) != null) {
1996 wt.start();
1997 return true;
1998 }
1999 } catch (Throwable rex) {
2000 ex = rex;
2001 }
2002 deregisterWorker(wt, ex); // clean up and return false
2003 }
2004 }
2005 }
2006 return false;
2007 }
2008
2009 /**
2010 * Helps and/or blocks until the given task is done.
2011 *
2012 * @param joiner the joining worker
2013 * @param task the task
2014 * @return task status on exit
2015 */
2016 final int awaitJoin(WorkQueue joiner, ForkJoinTask<?> task) {
2017 int s = 0;
2018 if (joiner != null && task != null && (s = task.status) >= 0) {
2019 ForkJoinTask<?> prevJoin = joiner.currentJoin;
2020 joiner.currentJoin = task;
2021 do {} while ((s = task.status) >= 0 && !joiner.isEmpty() &&
2022 joiner.tryRemoveAndExec(task)); // process local tasks
2023 if (s >= 0 && (s = task.status) >= 0) {
2024 helpSignal(task, joiner.poolIndex);
2025 if ((s = task.status) >= 0 &&
2026 (task instanceof CountedCompleter))
2027 s = helpComplete(task, LIFO_QUEUE);
2028 }
2029 while (s >= 0 && (s = task.status) >= 0) {
2030 if ((!joiner.isEmpty() || // try helping
2031 (s = tryHelpStealer(joiner, task)) == 0) &&
2032 (s = task.status) >= 0) {
2033 helpSignal(task, joiner.poolIndex);
2034 if ((s = task.status) >= 0 && tryCompensate()) {
2035 if (task.trySetSignal() && (s = task.status) >= 0) {
2036 synchronized (task) {
2037 if (task.status >= 0) {
2038 try { // see ForkJoinTask
2039 task.wait(); // for explanation
2040 } catch (InterruptedException ie) {
2041 }
2042 }
2043 else
2044 task.notifyAll();
2045 }
2046 }
2047 long c; // re-activate
2048 do {} while (!U.compareAndSwapLong
2049 (this, CTL, c = ctl, c + AC_UNIT));
2050 }
2051 }
2052 }
2053 joiner.currentJoin = prevJoin;
2054 }
2055 return s;
2056 }
2057
2058 /**
2059 * Stripped-down variant of awaitJoin used by timed joins. Tries
2060 * to help join only while there is continuous progress. (Caller
2061 * will then enter a timed wait.)
2062 *
2063 * @param joiner the joining worker
2064 * @param task the task
2065 */
2066 final void helpJoinOnce(WorkQueue joiner, ForkJoinTask<?> task) {
2067 int s;
2068 if (joiner != null && task != null && (s = task.status) >= 0) {
2069 ForkJoinTask<?> prevJoin = joiner.currentJoin;
2070 joiner.currentJoin = task;
2071 do {} while ((s = task.status) >= 0 && !joiner.isEmpty() &&
2072 joiner.tryRemoveAndExec(task));
2073 if (s >= 0 && (s = task.status) >= 0) {
2074 helpSignal(task, joiner.poolIndex);
2075 if ((s = task.status) >= 0 &&
2076 (task instanceof CountedCompleter))
2077 s = helpComplete(task, LIFO_QUEUE);
2078 }
2079 if (s >= 0 && joiner.isEmpty()) {
2080 do {} while (task.status >= 0 &&
2081 tryHelpStealer(joiner, task) > 0);
2082 }
2083 joiner.currentJoin = prevJoin;
2084 }
2085 }
2086
2087 /**
2088 * Returns a (probably) non-empty steal queue, if one is found
2089 * during a scan, else null. This method must be retried by
2090 * caller if, by the time it tries to use the queue, it is empty.
2091 * @param r a (random) seed for scanning
2092 */
2093 private WorkQueue findNonEmptyStealQueue(int r) {
2094 for (;;) {
2095 int ps = plock, m; WorkQueue[] ws; WorkQueue q;
2096 if ((ws = workQueues) != null && (m = ws.length - 1) >= 0) {
2097 for (int j = (m + 1) << 2; j >= 0; --j) {
2098 if ((q = ws[(((r + j) << 1) | 1) & m]) != null &&
2099 q.base - q.top < 0)
2100 return q;
2101 }
2102 }
2103 if (plock == ps)
2104 return null;
2105 }
2106 }
2107
2108 /**
2109 * Runs tasks until {@code isQuiescent()}. We piggyback on
2110 * active count ctl maintenance, but rather than blocking
2111 * when tasks cannot be found, we rescan until all others cannot
2112 * find tasks either.
2113 */
2114 final void helpQuiescePool(WorkQueue w) {
2115 for (boolean active = true;;) {
2116 long c; WorkQueue q; ForkJoinTask<?> t; int b;
2117 while ((t = w.nextLocalTask()) != null) {
2118 if (w.base - w.top < 0)
2119 signalWork(w);
2120 t.doExec();
2121 }
2122 if ((q = findNonEmptyStealQueue(w.nextSeed())) != null) {
2123 if (!active) { // re-establish active count
2124 active = true;
2125 do {} while (!U.compareAndSwapLong
2126 (this, CTL, c = ctl, c + AC_UNIT));
2127 }
2128 if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null) {
2129 if (q.base - q.top < 0)
2130 signalWork(q);
2131 w.runSubtask(t);
2132 }
2133 }
2134 else if (active) { // decrement active count without queuing
2135 long nc = (c = ctl) - AC_UNIT;
2136 if ((int)(nc >> AC_SHIFT) + (config & SMASK) == 0)
2137 return; // bypass decrement-then-increment
2138 if (U.compareAndSwapLong(this, CTL, c, nc))
2139 active = false;
2140 }
2141 else if ((int)((c = ctl) >> AC_SHIFT) + (config & SMASK) == 0 &&
2142 U.compareAndSwapLong(this, CTL, c, c + AC_UNIT))
2143 return;
2144 }
2145 }
2146
2147 /**
2148 * Gets and removes a local or stolen task for the given worker.
2149 *
2150 * @return a task, if available
2151 */
2152 final ForkJoinTask<?> nextTaskFor(WorkQueue w) {
2153 for (ForkJoinTask<?> t;;) {
2154 WorkQueue q; int b;
2155 if ((t = w.nextLocalTask()) != null)
2156 return t;
2157 if ((q = findNonEmptyStealQueue(w.nextSeed())) == null)
2158 return null;
2159 if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null) {
2160 if (q.base - q.top < 0)
2161 signalWork(q);
2162 return t;
2163 }
2164 }
2165 }
2166
2167 /**
2168 * Returns a cheap heuristic guide for task partitioning when
2169 * programmers, frameworks, tools, or languages have little or no
2170 * idea about task granularity. In essence by offering this
2171 * method, we ask users only about tradeoffs in overhead vs
2172 * expected throughput and its variance, rather than how finely to
2173 * partition tasks.
2174 *
2175 * In a steady state strict (tree-structured) computation, each
2176 * thread makes available for stealing enough tasks for other
2177 * threads to remain active. Inductively, if all threads play by
2178 * the same rules, each thread should make available only a
2179 * constant number of tasks.
2180 *
2181 * The minimum useful constant is just 1. But using a value of 1
2182 * would require immediate replenishment upon each steal to
2183 * maintain enough tasks, which is infeasible. Further,
2184 * partitionings/granularities of offered tasks should minimize
2185 * steal rates, which in general means that threads nearer the top
2186 * of computation tree should generate more than those nearer the
2187 * bottom. In perfect steady state, each thread is at
2188 * approximately the same level of computation tree. However,
2189 * producing extra tasks amortizes the uncertainty of progress and
2190 * diffusion assumptions.
2191 *
2192 * So, users will want to use values larger (but not much larger)
2193 * than 1 to both smooth over transient shortages and hedge
2194 * against uneven progress; as traded off against the cost of
2195 * extra task overhead. We leave the user to pick a threshold
2196 * value to compare with the results of this call to guide
2197 * decisions, but recommend values such as 3.
2198 *
2199 * When all threads are active, it is on average OK to estimate
2200 * surplus strictly locally. In steady-state, if one thread is
2201 * maintaining say 2 surplus tasks, then so are others. So we can
2202 * just use estimated queue length. However, this strategy alone
2203 * leads to serious mis-estimates in some non-steady-state
2204 * conditions (ramp-up, ramp-down, other stalls). We can detect
2205 * many of these by further considering the number of "idle"
2206 * threads, that are known to have zero queued tasks, so
2207 * compensate by a factor of (#idle/#active) threads.
2208 *
2209 * Note: The approximation of #busy workers as #active workers is
2210 * not very good under current signalling scheme, and should be
2211 * improved.
2212 */
2213 static int getSurplusQueuedTaskCount() {
2214 Thread t; ForkJoinWorkerThread wt; ForkJoinPool pool; WorkQueue q;
2215 if (((t = Thread.currentThread()) instanceof ForkJoinWorkerThread)) {
2216 int p = (pool = (wt = (ForkJoinWorkerThread)t).pool).config & SMASK;
2217 int n = (q = wt.workQueue).top - q.base;
2218 int a = (int)(pool.ctl >> AC_SHIFT) + p;
2219 return n - (a > (p >>>= 1) ? 0 :
2220 a > (p >>>= 1) ? 1 :
2221 a > (p >>>= 1) ? 2 :
2222 a > (p >>>= 1) ? 4 :
2223 8);
2224 }
2225 return 0;
2226 }
2227
2228 // Termination
2229
2230 /**
2231 * Possibly initiates and/or completes termination. The caller
2232 * triggering termination runs three passes through workQueues:
2233 * (0) Setting termination status, followed by wakeups of queued
2234 * workers; (1) cancelling all tasks; (2) interrupting lagging
2235 * threads (likely in external tasks, but possibly also blocked in
2236 * joins). Each pass repeats previous steps because of potential
2237 * lagging thread creation.
2238 *
2239 * @param now if true, unconditionally terminate, else only
2240 * if no work and no active workers
2241 * @param enable if true, enable shutdown when next possible
2242 * @return true if now terminating or terminated
2243 */
2244 private boolean tryTerminate(boolean now, boolean enable) {
2245 int ps;
2246 if (this == common) // cannot shut down
2247 return false;
2248 if ((ps = plock) >= 0) { // enable by setting plock
2249 if (!enable)
2250 return false;
2251 if ((ps & PL_LOCK) != 0 ||
2252 !U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
2253 ps = acquirePlock();
2254 int nps = ((ps + PL_LOCK) & ~SHUTDOWN) | SHUTDOWN;
2255 if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
2256 releasePlock(nps);
2257 }
2258 for (long c;;) {
2259 if (((c = ctl) & STOP_BIT) != 0) { // already terminating
2260 if ((short)(c >>> TC_SHIFT) == -(config & SMASK)) {
2261 synchronized (this) {
2262 notifyAll(); // signal when 0 workers
2263 }
2264 }
2265 return true;
2266 }
2267 if (!now) { // check if idle & no tasks
2268 WorkQueue[] ws; WorkQueue w;
2269 if ((int)(c >> AC_SHIFT) != -(config & SMASK))
2270 return false;
2271 if ((ws = workQueues) != null) {
2272 for (int i = 0; i < ws.length; ++i) {
2273 if ((w = ws[i]) != null) {
2274 if (!w.isEmpty()) { // signal unprocessed tasks
2275 signalWork(w);
2276 return false;
2277 }
2278 if ((i & 1) != 0 && w.eventCount >= 0)
2279 return false; // unqueued inactive worker
2280 }
2281 }
2282 }
2283 }
2284 if (U.compareAndSwapLong(this, CTL, c, c | STOP_BIT)) {
2285 for (int pass = 0; pass < 3; ++pass) {
2286 WorkQueue[] ws; WorkQueue w; Thread wt;
2287 if ((ws = workQueues) != null) {
2288 int n = ws.length;
2289 for (int i = 0; i < n; ++i) {
2290 if ((w = ws[i]) != null) {
2291 w.qlock = -1;
2292 if (pass > 0) {
2293 w.cancelAll();
2294 if (pass > 1 && (wt = w.owner) != null) {
2295 if (!wt.isInterrupted()) {
2296 try {
2297 wt.interrupt();
2298 } catch (Throwable ignore) {
2299 }
2300 }
2301 U.unpark(wt);
2302 }
2303 }
2304 }
2305 }
2306 // Wake up workers parked on event queue
2307 int i, e; long cc; Thread p;
2308 while ((e = (int)(cc = ctl) & E_MASK) != 0 &&
2309 (i = e & SMASK) < n && i >= 0 &&
2310 (w = ws[i]) != null) {
2311 long nc = ((long)(w.nextWait & E_MASK) |
2312 ((cc + AC_UNIT) & AC_MASK) |
2313 (cc & (TC_MASK|STOP_BIT)));
2314 if (w.eventCount == (e | INT_SIGN) &&
2315 U.compareAndSwapLong(this, CTL, cc, nc)) {
2316 w.eventCount = (e + E_SEQ) & E_MASK;
2317 w.qlock = -1;
2318 if ((p = w.parker) != null)
2319 U.unpark(p);
2320 }
2321 }
2322 }
2323 }
2324 }
2325 }
2326 }
2327
2328 // external operations on common pool
2329
2330 /**
2331 * Returns common pool queue for a thread that has submitted at
2332 * least one task.
2333 */
2334 static WorkQueue commonSubmitterQueue() {
2335 ForkJoinPool p; WorkQueue[] ws; int m; Submitter z;
2336 return ((z = submitters.get()) != null &&
2337 (p = common) != null &&
2338 (ws = p.workQueues) != null &&
2339 (m = ws.length - 1) >= 0) ?
2340 ws[m & z.seed & SQMASK] : null;
2341 }
2342
2343 /**
2344 * Tries to pop the given task from submitter's queue in common pool.
2345 */
2346 static boolean tryExternalUnpush(ForkJoinTask<?> t) {
2347 ForkJoinPool p; WorkQueue[] ws; WorkQueue q; Submitter z;
2348 ForkJoinTask<?>[] a; int m, s;
2349 if (t != null &&
2350 (z = submitters.get()) != null &&
2351 (p = common) != null &&
2352 (ws = p.workQueues) != null &&
2353 (m = ws.length - 1) >= 0 &&
2354 (q = ws[m & z.seed & SQMASK]) != null &&
2355 (s = q.top) != q.base &&
2356 (a = q.array) != null) {
2357 long j = (((a.length - 1) & (s - 1)) << ASHIFT) + ABASE;
2358 if (U.getObject(a, j) == t &&
2359 U.compareAndSwapInt(q, QLOCK, 0, 1)) {
2360 if (q.array == a && q.top == s && // recheck
2361 U.compareAndSwapObject(a, j, t, null)) {
2362 q.top = s - 1;
2363 q.qlock = 0;
2364 return true;
2365 }
2366 q.qlock = 0;
2367 }
2368 }
2369 return false;
2370 }
2371
2372 /**
2373 * Tries to pop and run local tasks within the same computation
2374 * as the given root. On failure, tries to help complete from
2375 * other queues via helpComplete.
2376 */
2377 private void externalHelpComplete(WorkQueue q, ForkJoinTask<?> root) {
2378 ForkJoinTask<?>[] a; int m;
2379 if (q != null && (a = q.array) != null && (m = (a.length - 1)) >= 0 &&
2380 root != null && root.status >= 0) {
2381 for (;;) {
2382 int s, u; Object o; CountedCompleter<?> task = null;
2383 if ((s = q.top) - q.base > 0) {
2384 long j = ((m & (s - 1)) << ASHIFT) + ABASE;
2385 if ((o = U.getObject(a, j)) != null &&
2386 (o instanceof CountedCompleter)) {
2387 CountedCompleter<?> t = (CountedCompleter<?>)o, r = t;
2388 do {
2389 if (r == root) {
2390 if (U.compareAndSwapInt(q, QLOCK, 0, 1)) {
2391 if (q.array == a && q.top == s &&
2392 U.compareAndSwapObject(a, j, t, null)) {
2393 q.top = s - 1;
2394 task = t;
2395 }
2396 q.qlock = 0;
2397 }
2398 break;
2399 }
2400 } while ((r = r.completer) != null);
2401 }
2402 }
2403 if (task != null)
2404 task.doExec();
2405 if (root.status < 0 ||
2406 (u = (int)(ctl >>> 32)) >= 0 || (u >> UAC_SHIFT) >= 0)
2407 break;
2408 if (task == null) {
2409 helpSignal(root, q.poolIndex);
2410 if (root.status >= 0)
2411 helpComplete(root, SHARED_QUEUE);
2412 break;
2413 }
2414 }
2415 }
2416 }
2417
2418 /**
2419 * Tries to help execute or signal availability of the given task
2420 * from submitter's queue in common pool.
2421 */
2422 static void externalHelpJoin(ForkJoinTask<?> t) {
2423 // Some hard-to-avoid overlap with tryExternalUnpush
2424 ForkJoinPool p; WorkQueue[] ws; WorkQueue q, w; Submitter z;
2425 ForkJoinTask<?>[] a; int m, s, n;
2426 if (t != null &&
2427 (z = submitters.get()) != null &&
2428 (p = common) != null &&
2429 (ws = p.workQueues) != null &&
2430 (m = ws.length - 1) >= 0 &&
2431 (q = ws[m & z.seed & SQMASK]) != null &&
2432 (a = q.array) != null) {
2433 int am = a.length - 1;
2434 if ((s = q.top) != q.base) {
2435 long j = ((am & (s - 1)) << ASHIFT) + ABASE;
2436 if (U.getObject(a, j) == t &&
2437 U.compareAndSwapInt(q, QLOCK, 0, 1)) {
2438 if (q.array == a && q.top == s &&
2439 U.compareAndSwapObject(a, j, t, null)) {
2440 q.top = s - 1;
2441 q.qlock = 0;
2442 t.doExec();
2443 }
2444 else
2445 q.qlock = 0;
2446 }
2447 }
2448 if (t.status >= 0) {
2449 if (t instanceof CountedCompleter)
2450 p.externalHelpComplete(q, t);
2451 else
2452 p.helpSignal(t, q.poolIndex);
2453 }
2454 }
2455 }
2456
2457 // Exported methods
2458
2459 // Constructors
2460
2461 /**
2462 * Creates a {@code ForkJoinPool} with parallelism equal to {@link
2463 * java.lang.Runtime#availableProcessors}, using the {@linkplain
2464 * #defaultForkJoinWorkerThreadFactory default thread factory},
2465 * no UncaughtExceptionHandler, and non-async LIFO processing mode.
2466 *
2467 * @throws SecurityException if a security manager exists and
2468 * the caller is not permitted to modify threads
2469 * because it does not hold {@link
2470 * java.lang.RuntimePermission}{@code ("modifyThread")}
2471 */
2472 public ForkJoinPool() {
2473 this(Math.min(MAX_CAP, Runtime.getRuntime().availableProcessors()),
2474 defaultForkJoinWorkerThreadFactory, null, false);
2475 }
2476
2477 /**
2478 * Creates a {@code ForkJoinPool} with the indicated parallelism
2479 * level, the {@linkplain
2480 * #defaultForkJoinWorkerThreadFactory default thread factory},
2481 * no UncaughtExceptionHandler, and non-async LIFO processing mode.
2482 *
2483 * @param parallelism the parallelism level
2484 * @throws IllegalArgumentException if parallelism less than or
2485 * equal to zero, or greater than implementation limit
2486 * @throws SecurityException if a security manager exists and
2487 * the caller is not permitted to modify threads
2488 * because it does not hold {@link
2489 * java.lang.RuntimePermission}{@code ("modifyThread")}
2490 */
2491 public ForkJoinPool(int parallelism) {
2492 this(parallelism, defaultForkJoinWorkerThreadFactory, null, false);
2493 }
2494
2495 /**
2496 * Creates a {@code ForkJoinPool} with the given parameters.
2497 *
2498 * @param parallelism the parallelism level. For default value,
2499 * use {@link java.lang.Runtime#availableProcessors}.
2500 * @param factory the factory for creating new threads. For default value,
2501 * use {@link #defaultForkJoinWorkerThreadFactory}.
2502 * @param handler the handler for internal worker threads that
2503 * terminate due to unrecoverable errors encountered while executing
2504 * tasks. For default value, use {@code null}.
2505 * @param asyncMode if true,
2506 * establishes local first-in-first-out scheduling mode for forked
2507 * tasks that are never joined. This mode may be more appropriate
2508 * than default locally stack-based mode in applications in which
2509 * worker threads only process event-style asynchronous tasks.
2510 * For default value, use {@code false}.
2511 * @throws IllegalArgumentException if parallelism less than or
2512 * equal to zero, or greater than implementation limit
2513 * @throws NullPointerException if the factory is null
2514 * @throws SecurityException if a security manager exists and
2515 * the caller is not permitted to modify threads
2516 * because it does not hold {@link
2517 * java.lang.RuntimePermission}{@code ("modifyThread")}
2518 */
2519 public ForkJoinPool(int parallelism,
2520 ForkJoinWorkerThreadFactory factory,
2521 Thread.UncaughtExceptionHandler handler,
2522 boolean asyncMode) {
2523 checkPermission();
2524 if (factory == null)
2525 throw new NullPointerException();
2526 if (parallelism <= 0 || parallelism > MAX_CAP)
2527 throw new IllegalArgumentException();
2528 this.factory = factory;
2529 this.ueh = handler;
2530 this.config = parallelism | (asyncMode ? (FIFO_QUEUE << 16) : 0);
2531 long np = (long)(-parallelism); // offset ctl counts
2532 this.ctl = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK);
2533 int pn = nextPoolId();
2534 StringBuilder sb = new StringBuilder("ForkJoinPool-");
2535 sb.append(Integer.toString(pn));
2536 sb.append("-worker-");
2537 this.workerNamePrefix = sb.toString();
2538 }
2539
2540 /**
2541 * Constructor for common pool, suitable only for static initialization.
2542 * Basically the same as above, but uses smallest possible initial footprint.
2543 */
2544 ForkJoinPool(int parallelism, long ctl,
2545 ForkJoinWorkerThreadFactory factory,
2546 Thread.UncaughtExceptionHandler handler) {
2547 this.config = parallelism;
2548 this.ctl = ctl;
2549 this.factory = factory;
2550 this.ueh = handler;
2551 this.workerNamePrefix = "ForkJoinPool.commonPool-worker-";
2552 }
2553
2554 /**
2555 * Returns the common pool instance. This pool is statically
2556 * constructed; its run state is unaffected by attempts to {@link
2557 * #shutdown} or {@link #shutdownNow}. However this pool and any
2558 * ongoing processing are automatically terminated upon program
2559 * {@link System#exit}. Any program that relies on asynchronous
2560 * task processing to complete before program termination should
2561 * invoke {@code commonPool().}{@link #awaitQuiescence}, before
2562 * exit.
2563 *
2564 * @return the common pool instance
2565 * @since 1.8
2566 */
2567 public static ForkJoinPool commonPool() {
2568 // assert common != null : "static init error";
2569 return common;
2570 }
2571
2572 // Execution methods
2573
2574 /**
2575 * Performs the given task, returning its result upon completion.
2576 * If the computation encounters an unchecked Exception or Error,
2577 * it is rethrown as the outcome of this invocation. Rethrown
2578 * exceptions behave in the same way as regular exceptions, but,
2579 * when possible, contain stack traces (as displayed for example
2580 * using {@code ex.printStackTrace()}) of both the current thread
2581 * as well as the thread actually encountering the exception;
2582 * minimally only the latter.
2583 *
2584 * @param task the task
2585 * @return the task's result
2586 * @throws NullPointerException if the task is null
2587 * @throws RejectedExecutionException if the task cannot be
2588 * scheduled for execution
2589 */
2590 public <T> T invoke(ForkJoinTask<T> task) {
2591 if (task == null)
2592 throw new NullPointerException();
2593 externalPush(task);
2594 return task.join();
2595 }
2596
2597 /**
2598 * Arranges for (asynchronous) execution of the given task.
2599 *
2600 * @param task the task
2601 * @throws NullPointerException if the task is null
2602 * @throws RejectedExecutionException if the task cannot be
2603 * scheduled for execution
2604 */
2605 public void execute(ForkJoinTask<?> task) {
2606 if (task == null)
2607 throw new NullPointerException();
2608 externalPush(task);
2609 }
2610
2611 // AbstractExecutorService methods
2612
2613 /**
2614 * @throws NullPointerException if the task is null
2615 * @throws RejectedExecutionException if the task cannot be
2616 * scheduled for execution
2617 */
2618 public void execute(Runnable task) {
2619 if (task == null)
2620 throw new NullPointerException();
2621 ForkJoinTask<?> job;
2622 if (task instanceof ForkJoinTask<?>) // avoid re-wrap
2623 job = (ForkJoinTask<?>) task;
2624 else
2625 job = new ForkJoinTask.AdaptedRunnableAction(task);
2626 externalPush(job);
2627 }
2628
2629 /**
2630 * Submits a ForkJoinTask for execution.
2631 *
2632 * @param task the task to submit
2633 * @return the task
2634 * @throws NullPointerException if the task is null
2635 * @throws RejectedExecutionException if the task cannot be
2636 * scheduled for execution
2637 */
2638 public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) {
2639 if (task == null)
2640 throw new NullPointerException();
2641 externalPush(task);
2642 return task;
2643 }
2644
2645 /**
2646 * @throws NullPointerException if the task is null
2647 * @throws RejectedExecutionException if the task cannot be
2648 * scheduled for execution
2649 */
2650 public <T> ForkJoinTask<T> submit(Callable<T> task) {
2651 ForkJoinTask<T> job = new ForkJoinTask.AdaptedCallable<T>(task);
2652 externalPush(job);
2653 return job;
2654 }
2655
2656 /**
2657 * @throws NullPointerException if the task is null
2658 * @throws RejectedExecutionException if the task cannot be
2659 * scheduled for execution
2660 */
2661 public <T> ForkJoinTask<T> submit(Runnable task, T result) {
2662 ForkJoinTask<T> job = new ForkJoinTask.AdaptedRunnable<T>(task, result);
2663 externalPush(job);
2664 return job;
2665 }
2666
2667 /**
2668 * @throws NullPointerException if the task is null
2669 * @throws RejectedExecutionException if the task cannot be
2670 * scheduled for execution
2671 */
2672 public ForkJoinTask<?> submit(Runnable task) {
2673 if (task == null)
2674 throw new NullPointerException();
2675 ForkJoinTask<?> job;
2676 if (task instanceof ForkJoinTask<?>) // avoid re-wrap
2677 job = (ForkJoinTask<?>) task;
2678 else
2679 job = new ForkJoinTask.AdaptedRunnableAction(task);
2680 externalPush(job);
2681 return job;
2682 }
2683
2684 /**
2685 * @throws NullPointerException {@inheritDoc}
2686 * @throws RejectedExecutionException {@inheritDoc}
2687 */
2688 public <T> List<Future<T>> invokeAll(Collection<? extends Callable<T>> tasks) {
2689 // In previous versions of this class, this method constructed
2690 // a task to run ForkJoinTask.invokeAll, but now external
2691 // invocation of multiple tasks is at least as efficient.
2692 ArrayList<Future<T>> futures = new ArrayList<Future<T>>(tasks.size());
2693
2694 boolean done = false;
2695 try {
2696 for (Callable<T> t : tasks) {
2697 ForkJoinTask<T> f = new ForkJoinTask.AdaptedCallable<T>(t);
2698 futures.add(f);
2699 externalPush(f);
2700 }
2701 for (int i = 0, size = futures.size(); i < size; i++)
2702 ((ForkJoinTask<?>)futures.get(i)).quietlyJoin();
2703 done = true;
2704 return futures;
2705 } finally {
2706 if (!done)
2707 for (int i = 0, size = futures.size(); i < size; i++)
2708 futures.get(i).cancel(false);
2709 }
2710 }
2711
2712 /**
2713 * Returns the factory used for constructing new workers.
2714 *
2715 * @return the factory used for constructing new workers
2716 */
2717 public ForkJoinWorkerThreadFactory getFactory() {
2718 return factory;
2719 }
2720
2721 /**
2722 * Returns the handler for internal worker threads that terminate
2723 * due to unrecoverable errors encountered while executing tasks.
2724 *
2725 * @return the handler, or {@code null} if none
2726 */
2727 public Thread.UncaughtExceptionHandler getUncaughtExceptionHandler() {
2728 return ueh;
2729 }
2730
2731 /**
2732 * Returns the targeted parallelism level of this pool.
2733 *
2734 * @return the targeted parallelism level of this pool
2735 */
2736 public int getParallelism() {
2737 return config & SMASK;
2738 }
2739
2740 /**
2741 * Returns the targeted parallelism level of the common pool.
2742 *
2743 * @return the targeted parallelism level of the common pool
2744 * @since 1.8
2745 */
2746 public static int getCommonPoolParallelism() {
2747 return commonParallelism;
2748 }
2749
2750 /**
2751 * Returns the number of worker threads that have started but not
2752 * yet terminated. The result returned by this method may differ
2753 * from {@link #getParallelism} when threads are created to
2754 * maintain parallelism when others are cooperatively blocked.
2755 *
2756 * @return the number of worker threads
2757 */
2758 public int getPoolSize() {
2759 return (config & SMASK) + (short)(ctl >>> TC_SHIFT);
2760 }
2761
2762 /**
2763 * Returns {@code true} if this pool uses local first-in-first-out
2764 * scheduling mode for forked tasks that are never joined.
2765 *
2766 * @return {@code true} if this pool uses async mode
2767 */
2768 public boolean getAsyncMode() {
2769 return (config >>> 16) == FIFO_QUEUE;
2770 }
2771
2772 /**
2773 * Returns an estimate of the number of worker threads that are
2774 * not blocked waiting to join tasks or for other managed
2775 * synchronization. This method may overestimate the
2776 * number of running threads.
2777 *
2778 * @return the number of worker threads
2779 */
2780 public int getRunningThreadCount() {
2781 int rc = 0;
2782 WorkQueue[] ws; WorkQueue w;
2783 if ((ws = workQueues) != null) {
2784 for (int i = 1; i < ws.length; i += 2) {
2785 if ((w = ws[i]) != null && w.isApparentlyUnblocked())
2786 ++rc;
2787 }
2788 }
2789 return rc;
2790 }
2791
2792 /**
2793 * Returns an estimate of the number of threads that are currently
2794 * stealing or executing tasks. This method may overestimate the
2795 * number of active threads.
2796 *
2797 * @return the number of active threads
2798 */
2799 public int getActiveThreadCount() {
2800 int r = (config & SMASK) + (int)(ctl >> AC_SHIFT);
2801 return (r <= 0) ? 0 : r; // suppress momentarily negative values
2802 }
2803
2804 /**
2805 * Returns {@code true} if all worker threads are currently idle.
2806 * An idle worker is one that cannot obtain a task to execute
2807 * because none are available to steal from other threads, and
2808 * there are no pending submissions to the pool. This method is
2809 * conservative; it might not return {@code true} immediately upon
2810 * idleness of all threads, but will eventually become true if
2811 * threads remain inactive.
2812 *
2813 * @return {@code true} if all threads are currently idle
2814 */
2815 public boolean isQuiescent() {
2816 return (int)(ctl >> AC_SHIFT) + (config & SMASK) == 0;
2817 }
2818
2819 /**
2820 * Returns an estimate of the total number of tasks stolen from
2821 * one thread's work queue by another. The reported value
2822 * underestimates the actual total number of steals when the pool
2823 * is not quiescent. This value may be useful for monitoring and
2824 * tuning fork/join programs: in general, steal counts should be
2825 * high enough to keep threads busy, but low enough to avoid
2826 * overhead and contention across threads.
2827 *
2828 * @return the number of steals
2829 */
2830 public long getStealCount() {
2831 long count = stealCount;
2832 WorkQueue[] ws; WorkQueue w;
2833 if ((ws = workQueues) != null) {
2834 for (int i = 1; i < ws.length; i += 2) {
2835 if ((w = ws[i]) != null)
2836 count += w.nsteals;
2837 }
2838 }
2839 return count;
2840 }
2841
2842 /**
2843 * Returns an estimate of the total number of tasks currently held
2844 * in queues by worker threads (but not including tasks submitted
2845 * to the pool that have not begun executing). This value is only
2846 * an approximation, obtained by iterating across all threads in
2847 * the pool. This method may be useful for tuning task
2848 * granularities.
2849 *
2850 * @return the number of queued tasks
2851 */
2852 public long getQueuedTaskCount() {
2853 long count = 0;
2854 WorkQueue[] ws; WorkQueue w;
2855 if ((ws = workQueues) != null) {
2856 for (int i = 1; i < ws.length; i += 2) {
2857 if ((w = ws[i]) != null)
2858 count += w.queueSize();
2859 }
2860 }
2861 return count;
2862 }
2863
2864 /**
2865 * Returns an estimate of the number of tasks submitted to this
2866 * pool that have not yet begun executing. This method may take
2867 * time proportional to the number of submissions.
2868 *
2869 * @return the number of queued submissions
2870 */
2871 public int getQueuedSubmissionCount() {
2872 int count = 0;
2873 WorkQueue[] ws; WorkQueue w;
2874 if ((ws = workQueues) != null) {
2875 for (int i = 0; i < ws.length; i += 2) {
2876 if ((w = ws[i]) != null)
2877 count += w.queueSize();
2878 }
2879 }
2880 return count;
2881 }
2882
2883 /**
2884 * Returns {@code true} if there are any tasks submitted to this
2885 * pool that have not yet begun executing.
2886 *
2887 * @return {@code true} if there are any queued submissions
2888 */
2889 public boolean hasQueuedSubmissions() {
2890 WorkQueue[] ws; WorkQueue w;
2891 if ((ws = workQueues) != null) {
2892 for (int i = 0; i < ws.length; i += 2) {
2893 if ((w = ws[i]) != null && !w.isEmpty())
2894 return true;
2895 }
2896 }
2897 return false;
2898 }
2899
2900 /**
2901 * Removes and returns the next unexecuted submission if one is
2902 * available. This method may be useful in extensions to this
2903 * class that re-assign work in systems with multiple pools.
2904 *
2905 * @return the next submission, or {@code null} if none
2906 */
2907 protected ForkJoinTask<?> pollSubmission() {
2908 WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2909 if ((ws = workQueues) != null) {
2910 for (int i = 0; i < ws.length; i += 2) {
2911 if ((w = ws[i]) != null && (t = w.poll()) != null)
2912 return t;
2913 }
2914 }
2915 return null;
2916 }
2917
2918 /**
2919 * Removes all available unexecuted submitted and forked tasks
2920 * from scheduling queues and adds them to the given collection,
2921 * without altering their execution status. These may include
2922 * artificially generated or wrapped tasks. This method is
2923 * designed to be invoked only when the pool is known to be
2924 * quiescent. Invocations at other times may not remove all
2925 * tasks. A failure encountered while attempting to add elements
2926 * to collection {@code c} may result in elements being in
2927 * neither, either or both collections when the associated
2928 * exception is thrown. The behavior of this operation is
2929 * undefined if the specified collection is modified while the
2930 * operation is in progress.
2931 *
2932 * @param c the collection to transfer elements into
2933 * @return the number of elements transferred
2934 */
2935 protected int drainTasksTo(Collection<? super ForkJoinTask<?>> c) {
2936 int count = 0;
2937 WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2938 if ((ws = workQueues) != null) {
2939 for (int i = 0; i < ws.length; ++i) {
2940 if ((w = ws[i]) != null) {
2941 while ((t = w.poll()) != null) {
2942 c.add(t);
2943 ++count;
2944 }
2945 }
2946 }
2947 }
2948 return count;
2949 }
2950
2951 /**
2952 * Returns a string identifying this pool, as well as its state,
2953 * including indications of run state, parallelism level, and
2954 * worker and task counts.
2955 *
2956 * @return a string identifying this pool, as well as its state
2957 */
2958 public String toString() {
2959 // Use a single pass through workQueues to collect counts
2960 long qt = 0L, qs = 0L; int rc = 0;
2961 long st = stealCount;
2962 long c = ctl;
2963 WorkQueue[] ws; WorkQueue w;
2964 if ((ws = workQueues) != null) {
2965 for (int i = 0; i < ws.length; ++i) {
2966 if ((w = ws[i]) != null) {
2967 int size = w.queueSize();
2968 if ((i & 1) == 0)
2969 qs += size;
2970 else {
2971 qt += size;
2972 st += w.nsteals;
2973 if (w.isApparentlyUnblocked())
2974 ++rc;
2975 }
2976 }
2977 }
2978 }
2979 int pc = (config & SMASK);
2980 int tc = pc + (short)(c >>> TC_SHIFT);
2981 int ac = pc + (int)(c >> AC_SHIFT);
2982 if (ac < 0) // ignore transient negative
2983 ac = 0;
2984 String level;
2985 if ((c & STOP_BIT) != 0)
2986 level = (tc == 0) ? "Terminated" : "Terminating";
2987 else
2988 level = plock < 0 ? "Shutting down" : "Running";
2989 return super.toString() +
2990 "[" + level +
2991 ", parallelism = " + pc +
2992 ", size = " + tc +
2993 ", active = " + ac +
2994 ", running = " + rc +
2995 ", steals = " + st +
2996 ", tasks = " + qt +
2997 ", submissions = " + qs +
2998 "]";
2999 }
3000
3001 /**
3002 * Possibly initiates an orderly shutdown in which previously
3003 * submitted tasks are executed, but no new tasks will be
3004 * accepted. Invocation has no effect on execution state if this
3005 * is the {@link #commonPool()}, and no additional effect if
3006 * already shut down. Tasks that are in the process of being
3007 * submitted concurrently during the course of this method may or
3008 * may not be rejected.
3009 *
3010 * @throws SecurityException if a security manager exists and
3011 * the caller is not permitted to modify threads
3012 * because it does not hold {@link
3013 * java.lang.RuntimePermission}{@code ("modifyThread")}
3014 */
3015 public void shutdown() {
3016 checkPermission();
3017 tryTerminate(false, true);
3018 }
3019
3020 /**
3021 * Possibly attempts to cancel and/or stop all tasks, and reject
3022 * all subsequently submitted tasks. Invocation has no effect on
3023 * execution state if this is the {@link #commonPool()}, and no
3024 * additional effect if already shut down. Otherwise, tasks that
3025 * are in the process of being submitted or executed concurrently
3026 * during the course of this method may or may not be
3027 * rejected. This method cancels both existing and unexecuted
3028 * tasks, in order to permit termination in the presence of task
3029 * dependencies. So the method always returns an empty list
3030 * (unlike the case for some other Executors).
3031 *
3032 * @return an empty list
3033 * @throws SecurityException if a security manager exists and
3034 * the caller is not permitted to modify threads
3035 * because it does not hold {@link
3036 * java.lang.RuntimePermission}{@code ("modifyThread")}
3037 */
3038 public List<Runnable> shutdownNow() {
3039 checkPermission();
3040 tryTerminate(true, true);
3041 return Collections.emptyList();
3042 }
3043
3044 /**
3045 * Returns {@code true} if all tasks have completed following shut down.
3046 *
3047 * @return {@code true} if all tasks have completed following shut down
3048 */
3049 public boolean isTerminated() {
3050 long c = ctl;
3051 return ((c & STOP_BIT) != 0L &&
3052 (short)(c >>> TC_SHIFT) == -(config & SMASK));
3053 }
3054
3055 /**
3056 * Returns {@code true} if the process of termination has
3057 * commenced but not yet completed. This method may be useful for
3058 * debugging. A return of {@code true} reported a sufficient
3059 * period after shutdown may indicate that submitted tasks have
3060 * ignored or suppressed interruption, or are waiting for I/O,
3061 * causing this executor not to properly terminate. (See the
3062 * advisory notes for class {@link ForkJoinTask} stating that
3063 * tasks should not normally entail blocking operations. But if
3064 * they do, they must abort them on interrupt.)
3065 *
3066 * @return {@code true} if terminating but not yet terminated
3067 */
3068 public boolean isTerminating() {
3069 long c = ctl;
3070 return ((c & STOP_BIT) != 0L &&
3071 (short)(c >>> TC_SHIFT) != -(config & SMASK));
3072 }
3073
3074 /**
3075 * Returns {@code true} if this pool has been shut down.
3076 *
3077 * @return {@code true} if this pool has been shut down
3078 */
3079 public boolean isShutdown() {
3080 return plock < 0;
3081 }
3082
3083 /**
3084 * Blocks until all tasks have completed execution after a
3085 * shutdown request, or the timeout occurs, or the current thread
3086 * is interrupted, whichever happens first. Because the {@link
3087 * #commonPool()} never terminates until program shutdown, when
3088 * applied to the common pool, this method is equivalent to {@link
3089 * #awaitQuiescence} but always returns {@code false}.
3090 *
3091 * @param timeout the maximum time to wait
3092 * @param unit the time unit of the timeout argument
3093 * @return {@code true} if this executor terminated and
3094 * {@code false} if the timeout elapsed before termination
3095 * @throws InterruptedException if interrupted while waiting
3096 */
3097 public boolean awaitTermination(long timeout, TimeUnit unit)
3098 throws InterruptedException {
3099 if (Thread.interrupted())
3100 throw new InterruptedException();
3101 if (this == common) {
3102 awaitQuiescence(timeout, unit);
3103 return false;
3104 }
3105 long nanos = unit.toNanos(timeout);
3106 if (isTerminated())
3107 return true;
3108 long startTime = System.nanoTime();
3109 boolean terminated = false;
3110 synchronized (this) {
3111 for (long waitTime = nanos, millis = 0L;;) {
3112 if (terminated = isTerminated() ||
3113 waitTime <= 0L ||
3114 (millis = unit.toMillis(waitTime)) <= 0L)
3115 break;
3116 wait(millis);
3117 waitTime = nanos - (System.nanoTime() - startTime);
3118 }
3119 }
3120 return terminated;
3121 }
3122
3123 /**
3124 * If called by a ForkJoinTask operating in this pool, equivalent
3125 * in effect to {@link ForkJoinTask#helpQuiesce}. Otherwise,
3126 * waits and/or attempts to assist performing tasks until this
3127 * pool {@link #isQuiescent} or the indicated timeout elapses.
3128 *
3129 * @param timeout the maximum time to wait
3130 * @param unit the time unit of the timeout argument
3131 * @return {@code true} if quiescent; {@code false} if the
3132 * timeout elapsed.
3133 */
3134 public boolean awaitQuiescence(long timeout, TimeUnit unit) {
3135 long nanos = unit.toNanos(timeout);
3136 ForkJoinWorkerThread wt;
3137 Thread thread = Thread.currentThread();
3138 if ((thread instanceof ForkJoinWorkerThread) &&
3139 (wt = (ForkJoinWorkerThread)thread).pool == this) {
3140 helpQuiescePool(wt.workQueue);
3141 return true;
3142 }
3143 long startTime = System.nanoTime();
3144 WorkQueue[] ws;
3145 int r = 0, m;
3146 boolean found = true;
3147 while (!isQuiescent() && (ws = workQueues) != null &&
3148 (m = ws.length - 1) >= 0) {
3149 if (!found) {
3150 if ((System.nanoTime() - startTime) > nanos)
3151 return false;
3152 Thread.yield(); // cannot block
3153 }
3154 found = false;
3155 for (int j = (m + 1) << 2; j >= 0; --j) {
3156 ForkJoinTask<?> t; WorkQueue q; int b;
3157 if ((q = ws[r++ & m]) != null && (b = q.base) - q.top < 0) {
3158 found = true;
3159 if ((t = q.pollAt(b)) != null) {
3160 if (q.base - q.top < 0)
3161 signalWork(q);
3162 t.doExec();
3163 }
3164 break;
3165 }
3166 }
3167 }
3168 return true;
3169 }
3170
3171 /**
3172 * Waits and/or attempts to assist performing tasks indefinitely
3173 * until the {@link #commonPool()} {@link #isQuiescent}.
3174 */
3175 static void quiesceCommonPool() {
3176 common.awaitQuiescence(Long.MAX_VALUE, TimeUnit.NANOSECONDS);
3177 }
3178
3179 /**
3180 * Interface for extending managed parallelism for tasks running
3181 * in {@link ForkJoinPool}s.
3182 *
3183 * <p>A {@code ManagedBlocker} provides two methods. Method
3184 * {@code isReleasable} must return {@code true} if blocking is
3185 * not necessary. Method {@code block} blocks the current thread
3186 * if necessary (perhaps internally invoking {@code isReleasable}
3187 * before actually blocking). These actions are performed by any
3188 * thread invoking {@link ForkJoinPool#managedBlock}. The
3189 * unusual methods in this API accommodate synchronizers that may,
3190 * but don't usually, block for long periods. Similarly, they
3191 * allow more efficient internal handling of cases in which
3192 * additional workers may be, but usually are not, needed to
3193 * ensure sufficient parallelism. Toward this end,
3194 * implementations of method {@code isReleasable} must be amenable
3195 * to repeated invocation.
3196 *
3197 * <p>For example, here is a ManagedBlocker based on a
3198 * ReentrantLock:
3199 * <pre> {@code
3200 * class ManagedLocker implements ManagedBlocker {
3201 * final ReentrantLock lock;
3202 * boolean hasLock = false;
3203 * ManagedLocker(ReentrantLock lock) { this.lock = lock; }
3204 * public boolean block() {
3205 * if (!hasLock)
3206 * lock.lock();
3207 * return true;
3208 * }
3209 * public boolean isReleasable() {
3210 * return hasLock || (hasLock = lock.tryLock());
3211 * }
3212 * }}</pre>
3213 *
3214 * <p>Here is a class that possibly blocks waiting for an
3215 * item on a given queue:
3216 * <pre> {@code
3217 * class QueueTaker<E> implements ManagedBlocker {
3218 * final BlockingQueue<E> queue;
3219 * volatile E item = null;
3220 * QueueTaker(BlockingQueue<E> q) { this.queue = q; }
3221 * public boolean block() throws InterruptedException {
3222 * if (item == null)
3223 * item = queue.take();
3224 * return true;
3225 * }
3226 * public boolean isReleasable() {
3227 * return item != null || (item = queue.poll()) != null;
3228 * }
3229 * public E getItem() { // call after pool.managedBlock completes
3230 * return item;
3231 * }
3232 * }}</pre>
3233 */
3234 public static interface ManagedBlocker {
3235 /**
3236 * Possibly blocks the current thread, for example waiting for
3237 * a lock or condition.
3238 *
3239 * @return {@code true} if no additional blocking is necessary
3240 * (i.e., if isReleasable would return true)
3241 * @throws InterruptedException if interrupted while waiting
3242 * (the method is not required to do so, but is allowed to)
3243 */
3244 boolean block() throws InterruptedException;
3245
3246 /**
3247 * Returns {@code true} if blocking is unnecessary.
3248 */
3249 boolean isReleasable();
3250 }
3251
3252 /**
3253 * Blocks in accord with the given blocker. If the current thread
3254 * is a {@link ForkJoinWorkerThread}, this method possibly
3255 * arranges for a spare thread to be activated if necessary to
3256 * ensure sufficient parallelism while the current thread is blocked.
3257 *
3258 * <p>If the caller is not a {@link ForkJoinTask}, this method is
3259 * behaviorally equivalent to
3260 * <pre> {@code
3261 * while (!blocker.isReleasable())
3262 * if (blocker.block())
3263 * return;
3264 * }</pre>
3265 *
3266 * If the caller is a {@code ForkJoinTask}, then the pool may
3267 * first be expanded to ensure parallelism, and later adjusted.
3268 *
3269 * @param blocker the blocker
3270 * @throws InterruptedException if blocker.block did so
3271 */
3272 public static void managedBlock(ManagedBlocker blocker)
3273 throws InterruptedException {
3274 Thread t = Thread.currentThread();
3275 if (t instanceof ForkJoinWorkerThread) {
3276 ForkJoinPool p = ((ForkJoinWorkerThread)t).pool;
3277 while (!blocker.isReleasable()) { // variant of helpSignal
3278 WorkQueue[] ws; WorkQueue q; int m, u;
3279 if ((ws = p.workQueues) != null && (m = ws.length - 1) >= 0) {
3280 for (int i = 0; i <= m; ++i) {
3281 if (blocker.isReleasable())
3282 return;
3283 if ((q = ws[i]) != null && q.base - q.top < 0) {
3284 p.signalWork(q);
3285 if ((u = (int)(p.ctl >>> 32)) >= 0 ||
3286 (u >> UAC_SHIFT) >= 0)
3287 break;
3288 }
3289 }
3290 }
3291 if (p.tryCompensate()) {
3292 try {
3293 do {} while (!blocker.isReleasable() &&
3294 !blocker.block());
3295 } finally {
3296 p.incrementActiveCount();
3297 }
3298 break;
3299 }
3300 }
3301 }
3302 else {
3303 do {} while (!blocker.isReleasable() &&
3304 !blocker.block());
3305 }
3306 }
3307
3308 // AbstractExecutorService overrides. These rely on undocumented
3309 // fact that ForkJoinTask.adapt returns ForkJoinTasks that also
3310 // implement RunnableFuture.
3311
3312 protected <T> RunnableFuture<T> newTaskFor(Runnable runnable, T value) {
3313 return new ForkJoinTask.AdaptedRunnable<T>(runnable, value);
3314 }
3315
3316 protected <T> RunnableFuture<T> newTaskFor(Callable<T> callable) {
3317 return new ForkJoinTask.AdaptedCallable<T>(callable);
3318 }
3319
3320 // Unsafe mechanics
3321 private static final sun.misc.Unsafe U;
3322 private static final long CTL;
3323 private static final long PARKBLOCKER;
3324 private static final int ABASE;
3325 private static final int ASHIFT;
3326 private static final long STEALCOUNT;
3327 private static final long PLOCK;
3328 private static final long INDEXSEED;
3329 private static final long QLOCK;
3330
3331 static {
3332 // initialize field offsets for CAS etc
3333 try {
3334 U = getUnsafe();
3335 Class<?> k = ForkJoinPool.class;
3336 CTL = U.objectFieldOffset
3337 (k.getDeclaredField("ctl"));
3338 STEALCOUNT = U.objectFieldOffset
3339 (k.getDeclaredField("stealCount"));
3340 PLOCK = U.objectFieldOffset
3341 (k.getDeclaredField("plock"));
3342 INDEXSEED = U.objectFieldOffset
3343 (k.getDeclaredField("indexSeed"));
3344 Class<?> tk = Thread.class;
3345 PARKBLOCKER = U.objectFieldOffset
3346 (tk.getDeclaredField("parkBlocker"));
3347 Class<?> wk = WorkQueue.class;
3348 QLOCK = U.objectFieldOffset
3349 (wk.getDeclaredField("qlock"));
3350 Class<?> ak = ForkJoinTask[].class;
3351 ABASE = U.arrayBaseOffset(ak);
3352 int scale = U.arrayIndexScale(ak);
3353 if ((scale & (scale - 1)) != 0)
3354 throw new Error("data type scale not a power of two");
3355 ASHIFT = 31 - Integer.numberOfLeadingZeros(scale);
3356 } catch (Exception e) {
3357 throw new Error(e);
3358 }
3359
3360 submitters = new ThreadLocal<Submitter>();
3361 ForkJoinWorkerThreadFactory fac = defaultForkJoinWorkerThreadFactory =
3362 new DefaultForkJoinWorkerThreadFactory();
3363 modifyThreadPermission = new RuntimePermission("modifyThread");
3364
3365 /*
3366 * Establish common pool parameters. For extra caution,
3367 * computations to set up common pool state are here; the
3368 * constructor just assigns these values to fields.
3369 */
3370
3371 int par = 0;
3372 Thread.UncaughtExceptionHandler handler = null;
3373 try { // TBD: limit or report ignored exceptions?
3374 String pp = System.getProperty
3375 ("java.util.concurrent.ForkJoinPool.common.parallelism");
3376 String hp = System.getProperty
3377 ("java.util.concurrent.ForkJoinPool.common.exceptionHandler");
3378 String fp = System.getProperty
3379 ("java.util.concurrent.ForkJoinPool.common.threadFactory");
3380 if (fp != null)
3381 fac = ((ForkJoinWorkerThreadFactory)ClassLoader.
3382 getSystemClassLoader().loadClass(fp).newInstance());
3383 if (hp != null)
3384 handler = ((Thread.UncaughtExceptionHandler)ClassLoader.
3385 getSystemClassLoader().loadClass(hp).newInstance());
3386 if (pp != null)
3387 par = Integer.parseInt(pp);
3388 } catch (Exception ignore) {
3389 }
3390
3391 if (par <= 0)
3392 par = Runtime.getRuntime().availableProcessors();
3393 if (par > MAX_CAP)
3394 par = MAX_CAP;
3395 commonParallelism = par;
3396 long np = (long)(-par); // precompute initial ctl value
3397 long ct = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK);
3398
3399 common = new ForkJoinPool(par, ct, fac, handler);
3400 }
3401
3402 /**
3403 * Returns a sun.misc.Unsafe. Suitable for use in a 3rd party package.
3404 * Replace with a simple call to Unsafe.getUnsafe when integrating
3405 * into a jdk.
3406 *
3407 * @return a sun.misc.Unsafe
3408 */
3409 private static sun.misc.Unsafe getUnsafe() {
3410 try {
3411 return sun.misc.Unsafe.getUnsafe();
3412 } catch (SecurityException tryReflectionInstead) {}
3413 try {
3414 return java.security.AccessController.doPrivileged
3415 (new java.security.PrivilegedExceptionAction<sun.misc.Unsafe>() {
3416 public sun.misc.Unsafe run() throws Exception {
3417 Class<sun.misc.Unsafe> k = sun.misc.Unsafe.class;
3418 for (java.lang.reflect.Field f : k.getDeclaredFields()) {
3419 f.setAccessible(true);
3420 Object x = f.get(null);
3421 if (k.isInstance(x))
3422 return k.cast(x);
3423 }
3424 throw new NoSuchFieldError("the Unsafe");
3425 }});
3426 } catch (java.security.PrivilegedActionException e) {
3427 throw new RuntimeException("Could not initialize intrinsics",
3428 e.getCause());
3429 }
3430 }
3431 }