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 java.util.concurrent; |
8 |
|
9 |
import java.lang.Thread.UncaughtExceptionHandler; |
10 |
import java.lang.invoke.MethodHandles; |
11 |
import java.lang.invoke.VarHandle; |
12 |
import java.security.AccessController; |
13 |
import java.security.AccessControlContext; |
14 |
import java.security.Permission; |
15 |
import java.security.Permissions; |
16 |
import java.security.PrivilegedAction; |
17 |
import java.security.ProtectionDomain; |
18 |
import java.util.ArrayList; |
19 |
import java.util.Collection; |
20 |
import java.util.Collections; |
21 |
import java.util.List; |
22 |
import java.util.function.Predicate; |
23 |
import java.util.concurrent.atomic.AtomicInteger; |
24 |
import java.util.concurrent.locks.LockSupport; |
25 |
import java.util.concurrent.locks.ReentrantLock; |
26 |
import java.util.concurrent.locks.Condition; |
27 |
|
28 |
/** |
29 |
* An {@link ExecutorService} for running {@link ForkJoinTask}s. |
30 |
* A {@code ForkJoinPool} provides the entry point for submissions |
31 |
* from non-{@code ForkJoinTask} clients, as well as management and |
32 |
* monitoring operations. |
33 |
* |
34 |
* <p>A {@code ForkJoinPool} differs from other kinds of {@link |
35 |
* ExecutorService} mainly by virtue of employing |
36 |
* <em>work-stealing</em>: all threads in the pool attempt to find and |
37 |
* execute tasks submitted to the pool and/or created by other active |
38 |
* tasks (eventually blocking waiting for work if none exist). This |
39 |
* enables efficient processing when most tasks spawn other subtasks |
40 |
* (as do most {@code ForkJoinTask}s), as well as when many small |
41 |
* tasks are submitted to the pool from external clients. Especially |
42 |
* when setting <em>asyncMode</em> to true in constructors, {@code |
43 |
* ForkJoinPool}s may also be appropriate for use with event-style |
44 |
* tasks that are never joined. All worker threads are initialized |
45 |
* with {@link Thread#isDaemon} set {@code true}. |
46 |
* |
47 |
* <p>A static {@link #commonPool()} is available and appropriate for |
48 |
* most applications. The common pool is used by any ForkJoinTask that |
49 |
* is not explicitly submitted to a specified pool. Using the common |
50 |
* pool normally reduces resource usage (its threads are slowly |
51 |
* reclaimed during periods of non-use, and reinstated upon subsequent |
52 |
* use). |
53 |
* |
54 |
* <p>For applications that require separate or custom pools, a {@code |
55 |
* ForkJoinPool} may be constructed with a given target parallelism |
56 |
* level; by default, equal to the number of available processors. |
57 |
* The pool attempts to maintain enough active (or available) threads |
58 |
* by dynamically adding, suspending, or resuming internal worker |
59 |
* threads, even if some tasks are stalled waiting to join others. |
60 |
* However, no such adjustments are guaranteed in the face of blocked |
61 |
* I/O or other unmanaged synchronization. The nested {@link |
62 |
* ManagedBlocker} interface enables extension of the kinds of |
63 |
* synchronization accommodated. The default policies may be |
64 |
* overridden using a constructor with parameters corresponding to |
65 |
* those documented in class {@link ThreadPoolExecutor}. |
66 |
* |
67 |
* <p>In addition to execution and lifecycle control methods, this |
68 |
* class provides status check methods (for example |
69 |
* {@link #getStealCount}) that are intended to aid in developing, |
70 |
* tuning, and monitoring fork/join applications. Also, method |
71 |
* {@link #toString} returns indications of pool state in a |
72 |
* convenient form for informal monitoring. |
73 |
* |
74 |
* <p>As is the case with other ExecutorServices, there are three |
75 |
* main task execution methods summarized in the following table. |
76 |
* These are designed to be used primarily by clients not already |
77 |
* engaged in fork/join computations in the current pool. The main |
78 |
* forms of these methods accept instances of {@code ForkJoinTask}, |
79 |
* but overloaded forms also allow mixed execution of plain {@code |
80 |
* Runnable}- or {@code Callable}- based activities as well. However, |
81 |
* tasks that are already executing in a pool should normally instead |
82 |
* use the within-computation forms listed in the table unless using |
83 |
* async event-style tasks that are not usually joined, in which case |
84 |
* there is little difference among choice of methods. |
85 |
* |
86 |
* <table class="plain"> |
87 |
* <caption>Summary of task execution methods</caption> |
88 |
* <tr> |
89 |
* <td></td> |
90 |
* <th scope="col"> Call from non-fork/join clients</th> |
91 |
* <th scope="col"> Call from within fork/join computations</th> |
92 |
* </tr> |
93 |
* <tr> |
94 |
* <th scope="row" style="text-align:left"> Arrange async execution</th> |
95 |
* <td> {@link #execute(ForkJoinTask)}</td> |
96 |
* <td> {@link ForkJoinTask#fork}</td> |
97 |
* </tr> |
98 |
* <tr> |
99 |
* <th scope="row" style="text-align:left"> Await and obtain result</th> |
100 |
* <td> {@link #invoke(ForkJoinTask)}</td> |
101 |
* <td> {@link ForkJoinTask#invoke}</td> |
102 |
* </tr> |
103 |
* <tr> |
104 |
* <th scope="row" style="text-align:left"> Arrange exec and obtain Future</th> |
105 |
* <td> {@link #submit(ForkJoinTask)}</td> |
106 |
* <td> {@link ForkJoinTask#fork} (ForkJoinTasks <em>are</em> Futures)</td> |
107 |
* </tr> |
108 |
* </table> |
109 |
* |
110 |
* <p>The parameters used to construct the common pool may be controlled by |
111 |
* setting the following {@linkplain System#getProperty system properties}: |
112 |
* <ul> |
113 |
* <li>{@systemProperty java.util.concurrent.ForkJoinPool.common.parallelism} |
114 |
* - the parallelism level, a non-negative integer |
115 |
* <li>{@systemProperty java.util.concurrent.ForkJoinPool.common.threadFactory} |
116 |
* - the class name of a {@link ForkJoinWorkerThreadFactory}. |
117 |
* The {@linkplain ClassLoader#getSystemClassLoader() system class loader} |
118 |
* is used to load this class. |
119 |
* <li>{@systemProperty java.util.concurrent.ForkJoinPool.common.exceptionHandler} |
120 |
* - the class name of a {@link UncaughtExceptionHandler}. |
121 |
* The {@linkplain ClassLoader#getSystemClassLoader() system class loader} |
122 |
* is used to load this class. |
123 |
* <li>{@systemProperty java.util.concurrent.ForkJoinPool.common.maximumSpares} |
124 |
* - the maximum number of allowed extra threads to maintain target |
125 |
* parallelism (default 256). |
126 |
* </ul> |
127 |
* If no thread factory is supplied via a system property, then the |
128 |
* common pool uses a factory that uses the system class loader as the |
129 |
* {@linkplain Thread#getContextClassLoader() thread context class loader}. |
130 |
* In addition, if a {@link SecurityManager} is present, then |
131 |
* the common pool uses a factory supplying threads that have no |
132 |
* {@link Permissions} enabled. |
133 |
* |
134 |
* Upon any error in establishing these settings, default parameters |
135 |
* are used. It is possible to disable or limit the use of threads in |
136 |
* the common pool by setting the parallelism property to zero, and/or |
137 |
* using a factory that may return {@code null}. However doing so may |
138 |
* cause unjoined tasks to never be executed. |
139 |
* |
140 |
* <p><b>Implementation notes:</b> This implementation restricts the |
141 |
* maximum number of running threads to 32767. Attempts to create |
142 |
* pools with greater than the maximum number result in |
143 |
* {@code IllegalArgumentException}. |
144 |
* |
145 |
* <p>This implementation rejects submitted tasks (that is, by throwing |
146 |
* {@link RejectedExecutionException}) only when the pool is shut down |
147 |
* or internal resources have been exhausted. |
148 |
* |
149 |
* @since 1.7 |
150 |
* @author Doug Lea |
151 |
*/ |
152 |
public class ForkJoinPool extends AbstractExecutorService { |
153 |
|
154 |
/* |
155 |
* Implementation Overview |
156 |
* |
157 |
* This class and its nested classes provide the main |
158 |
* functionality and control for a set of worker threads: |
159 |
* Submissions from non-FJ threads enter into submission queues. |
160 |
* Workers take these tasks and typically split them into subtasks |
161 |
* that may be stolen by other workers. Work-stealing based on |
162 |
* randomized scans generally leads to better throughput than |
163 |
* "work dealing" in which producers assign tasks to idle threads, |
164 |
* in part because threads that have finished other tasks before |
165 |
* the signalled thread wakes up (which can be a long time) can |
166 |
* take the task instead. Preference rules give first priority to |
167 |
* processing tasks from their own queues (LIFO or FIFO, depending |
168 |
* on mode), then to randomized FIFO steals of tasks in other |
169 |
* queues. This framework began as vehicle for supporting |
170 |
* tree-structured parallelism using work-stealing. Over time, |
171 |
* its scalability advantages led to extensions and changes to |
172 |
* better support more diverse usage contexts. Because most |
173 |
* internal methods and nested classes are interrelated, their |
174 |
* main rationale and descriptions are presented here; individual |
175 |
* methods and nested classes contain only brief comments about |
176 |
* details. |
177 |
* |
178 |
* WorkQueues |
179 |
* ========== |
180 |
* |
181 |
* Most operations occur within work-stealing queues (in nested |
182 |
* class WorkQueue). These are special forms of Deques that |
183 |
* support only three of the four possible end-operations -- push, |
184 |
* pop, and poll (aka steal), under the further constraints that |
185 |
* push and pop are called only from the owning thread (or, as |
186 |
* extended here, under a lock), while poll may be called from |
187 |
* other threads. (If you are unfamiliar with them, you probably |
188 |
* want to read Herlihy and Shavit's book "The Art of |
189 |
* Multiprocessor programming", chapter 16 describing these in |
190 |
* more detail before proceeding.) The main work-stealing queue |
191 |
* design is roughly similar to those in the papers "Dynamic |
192 |
* Circular Work-Stealing Deque" by Chase and Lev, SPAA 2005 |
193 |
* (http://research.sun.com/scalable/pubs/index.html) and |
194 |
* "Idempotent work stealing" by Michael, Saraswat, and Vechev, |
195 |
* PPoPP 2009 (http://portal.acm.org/citation.cfm?id=1504186). |
196 |
* The main differences ultimately stem from GC requirements that |
197 |
* we null out taken slots as soon as we can, to maintain as small |
198 |
* a footprint as possible even in programs generating huge |
199 |
* numbers of tasks. To accomplish this, we shift the CAS |
200 |
* arbitrating pop vs poll (steal) from being on the indices |
201 |
* ("base" and "top") to the slots themselves. |
202 |
* |
203 |
* Adding tasks then takes the form of a classic array push(task) |
204 |
* in a circular buffer: |
205 |
* q.array[q.top++ % length] = task; |
206 |
* |
207 |
* The actual code needs to null-check and size-check the array, |
208 |
* uses masking, not mod, for indexing a power-of-two-sized array, |
209 |
* enforces memory ordering, supports resizing, and possibly |
210 |
* signals waiting workers to start scanning -- see below. |
211 |
* |
212 |
* The pop operation (always performed by owner) is of the form: |
213 |
* if ((task = getAndSet(q.array, (q.top-1) % length, null)) != null) |
214 |
* decrement top and return task; |
215 |
* If this fails, the queue is empty. |
216 |
* |
217 |
* The poll operation by another stealer thread is, basically: |
218 |
* if (CAS nonnull task at q.array[q.base % length] to null) |
219 |
* increment base and return task; |
220 |
* |
221 |
* This may fail due to contention, and may be retried. |
222 |
* Implementations must ensure a consistent snapshot of the base |
223 |
* index and the task (by looping or trying elsewhere) before |
224 |
* trying CAS. There isn't actually a method of this form, |
225 |
* because failure due to inconsistency or contention is handled |
226 |
* in different ways in different contexts, normally by first |
227 |
* trying other queues. (For the most straightforward example, see |
228 |
* method pollScan.) There are further variants for cases |
229 |
* requiring inspection of elements before extracting them, so |
230 |
* must interleave these with variants of this code. Also, a more |
231 |
* efficient version (nextLocalTask) is used for polls by owners. |
232 |
* It avoids some overhead because the queue cannot be growing |
233 |
* during call. |
234 |
* |
235 |
* Memory ordering. See "Correct and Efficient Work-Stealing for |
236 |
* Weak Memory Models" by Le, Pop, Cohen, and Nardelli, PPoPP 2013 |
237 |
* (http://www.di.ens.fr/~zappa/readings/ppopp13.pdf) for an |
238 |
* analysis of memory ordering requirements in work-stealing |
239 |
* algorithms similar to the one used here. Inserting and |
240 |
* extracting tasks in array slots via volatile or atomic accesses |
241 |
* or explicit fences provides primary synchronization. |
242 |
* |
243 |
* Operations on deque elements require reads and writes of both |
244 |
* indices and slots. When possible, we allow these to occur in |
245 |
* any order. Because the base and top indices (along with other |
246 |
* pool or array fields accessed in many methods) only imprecisely |
247 |
* guide where to extract from, we let accesses other than the |
248 |
* element getAndSet/CAS/setVolatile appear in any order, using |
249 |
* plain mode. But we must still preface some methods (mainly |
250 |
* those that may be accessed externally) with an acquireFence to |
251 |
* avoid unbounded staleness. This is equivalent to acting as if |
252 |
* callers use an acquiring read of the reference to the pool or |
253 |
* queue when invoking the method, even when they do not. We use |
254 |
* explicit acquiring reads (getSlot) rather than plain array |
255 |
* access when acquire mode is required but not otherwise ensured |
256 |
* by context. To reduce stalls by other stealers, we encourage |
257 |
* timely writes to the base index by immediately following |
258 |
* updates with a write of a volatile field that must be updated |
259 |
* anyway, or an Opaque-mode write if there is no such |
260 |
* opportunity. |
261 |
* |
262 |
* Because indices and slot contents cannot always be consistent, |
263 |
* the emptiness check base == top is only quiescently accurate |
264 |
* (and so used where this suffices). Otherwise, it may err on the |
265 |
* side of possibly making the queue appear nonempty when a push, |
266 |
* pop, or poll have not fully committed, or making it appear |
267 |
* empty when an update of top or base has not yet been seen. |
268 |
* Similarly, the check in push for the queue array being full may |
269 |
* trigger when not completely full, causing a resize earlier than |
270 |
* required. |
271 |
* |
272 |
* Mainly because of these potential inconsistencies among slots |
273 |
* vs indices, the poll operation, considered individually, is not |
274 |
* wait-free. One thief cannot successfully continue until another |
275 |
* in-progress one (or, if previously empty, a push) visibly |
276 |
* completes. This can stall threads when required to consume |
277 |
* from a given queue (which may spin). However, in the |
278 |
* aggregate, we ensure probabilistic non-blockingness at least |
279 |
* until checking quiescence (which is intrinsically blocking): |
280 |
* If an attempted steal fails, a scanning thief chooses a |
281 |
* different victim target to try next. So, in order for one thief |
282 |
* to progress, it suffices for any in-progress poll or new push |
283 |
* on any empty queue to complete. The worst cases occur when many |
284 |
* threads are looking for tasks being produced by a stalled |
285 |
* producer. |
286 |
* |
287 |
* This approach also enables support of a user mode in which |
288 |
* local task processing is in FIFO, not LIFO order, simply by |
289 |
* using poll rather than pop. This can be useful in |
290 |
* message-passing frameworks in which tasks are never joined, |
291 |
* although with increased contention among task producers and |
292 |
* consumers. |
293 |
* |
294 |
* WorkQueues are also used in a similar way for tasks submitted |
295 |
* to the pool. We cannot mix these tasks in the same queues used |
296 |
* by workers. Instead, we randomly associate submission queues |
297 |
* with submitting threads, using a form of hashing. The |
298 |
* ThreadLocalRandom probe value serves as a hash code for |
299 |
* choosing existing queues, and may be randomly repositioned upon |
300 |
* contention with other submitters. In essence, submitters act |
301 |
* like workers except that they are restricted to executing local |
302 |
* tasks that they submitted (or when known, subtasks thereof). |
303 |
* Insertion of tasks in shared mode requires a lock. We use only |
304 |
* a simple spinlock (using field "source"), because submitters |
305 |
* encountering a busy queue move to a different position to use |
306 |
* or create other queues. They block only when registering new |
307 |
* queues. |
308 |
* |
309 |
* Management |
310 |
* ========== |
311 |
* |
312 |
* The main throughput advantages of work-stealing stem from |
313 |
* decentralized control -- workers mostly take tasks from |
314 |
* themselves or each other, at rates that can exceed a billion |
315 |
* per second. Most non-atomic control is performed by some form |
316 |
* of scanning across or within queues. The pool itself creates, |
317 |
* activates (enables scanning for and running tasks), |
318 |
* deactivates, blocks, and terminates threads, all with minimal |
319 |
* central information. There are only a few properties that we |
320 |
* can globally track or maintain, so we pack them into a small |
321 |
* number of variables, often maintaining atomicity without |
322 |
* blocking or locking. Nearly all essentially atomic control |
323 |
* state is held in a few volatile variables that are by far most |
324 |
* often read (not written) as status and consistency checks. We |
325 |
* pack as much information into them as we can. |
326 |
* |
327 |
* Field "ctl" contains 64 bits holding information needed to |
328 |
* atomically decide to add, enqueue (on an event queue), and |
329 |
* dequeue and release workers. To enable this packing, we |
330 |
* restrict maximum parallelism to (1<<15)-1 (which is far in |
331 |
* excess of normal operating range) to allow ids, counts, and |
332 |
* their negations (used for thresholding) to fit into 16bit |
333 |
* subfields. |
334 |
* |
335 |
* Field "mode" holds configuration parameters as well as lifetime |
336 |
* status, atomically and monotonically setting SHUTDOWN, STOP, |
337 |
* and finally TERMINATED bits. It is updated only via bitwise |
338 |
* atomics (getAndBitwiseOr). |
339 |
* |
340 |
* Array "queues" holds references to WorkQueues. It is updated |
341 |
* (only during worker creation and termination) under the |
342 |
* registrationLock, but is otherwise concurrently readable, and |
343 |
* accessed directly (although always prefaced by acquireFences or |
344 |
* other acquiring reads). To simplify index-based operations, the |
345 |
* array size is always a power of two, and all readers must |
346 |
* tolerate null slots. Worker queues are at odd indices. Worker |
347 |
* ids masked with SMASK match their index. Shared (submission) |
348 |
* queues are at even indices. Grouping them together in this way |
349 |
* simplifies and speeds up task scanning. |
350 |
* |
351 |
* All worker thread creation is on-demand, triggered by task |
352 |
* submissions, replacement of terminated workers, and/or |
353 |
* compensation for blocked workers. However, all other support |
354 |
* code is set up to work with other policies. To ensure that we |
355 |
* do not hold on to worker or task references that would prevent |
356 |
* GC, all accesses to workQueues are via indices into the |
357 |
* queues array (which is one source of some of the messy code |
358 |
* constructions here). In essence, the queues array serves as |
359 |
* a weak reference mechanism. Thus for example the stack top |
360 |
* subfield of ctl stores indices, not references. |
361 |
* |
362 |
* Queuing Idle Workers. Unlike HPC work-stealing frameworks, we |
363 |
* cannot let workers spin indefinitely scanning for tasks when |
364 |
* none can be found immediately, and we cannot start/resume |
365 |
* workers unless there appear to be tasks available. On the |
366 |
* other hand, we must quickly prod them into action when new |
367 |
* tasks are submitted or generated. These latencies are mainly a |
368 |
* function of JVM park/unpark (and underlying OS) performance, |
369 |
* which can be slow and variable. In many usages, ramp-up time |
370 |
* is the main limiting factor in overall performance, which is |
371 |
* compounded at program start-up by JIT compilation and |
372 |
* allocation. On the other hand, throughput degrades when too |
373 |
* many threads poll for too few tasks. |
374 |
* |
375 |
* The "ctl" field atomically maintains total and "released" |
376 |
* worker counts, plus the head of the available worker queue |
377 |
* (actually stack, represented by the lower 32bit subfield of |
378 |
* ctl). Released workers are those known to be scanning for |
379 |
* and/or running tasks. Unreleased ("available") workers are |
380 |
* recorded in the ctl stack. These workers are made available for |
381 |
* signalling by enqueuing in ctl (see method awaitWork). The |
382 |
* "queue" is a form of Treiber stack. This is ideal for |
383 |
* activating threads in most-recently used order, and improves |
384 |
* performance and locality, outweighing the disadvantages of |
385 |
* being prone to contention and inability to release a worker |
386 |
* unless it is topmost on stack. The top stack state holds the |
387 |
* value of the "phase" field of the worker: its index and status, |
388 |
* plus a version counter that, in addition to the count subfields |
389 |
* (also serving as version stamps) provide protection against |
390 |
* Treiber stack ABA effects. |
391 |
* |
392 |
* Creating workers. To create a worker, we pre-increment counts |
393 |
* (serving as a reservation), and attempt to construct a |
394 |
* ForkJoinWorkerThread via its factory. On starting, the new |
395 |
* thread first invokes registerWorker, where it constructs a |
396 |
* WorkQueue and is assigned an index in the queues array |
397 |
* (expanding the array if necessary). Upon any exception across |
398 |
* these steps, or null return from factory, deregisterWorker |
399 |
* adjusts counts and records accordingly. If a null return, the |
400 |
* pool continues running with fewer than the target number |
401 |
* workers. If exceptional, the exception is propagated, generally |
402 |
* to some external caller. |
403 |
* |
404 |
* WorkQueue field "phase" is used by both workers and the pool to |
405 |
* manage and track whether a worker is UNSIGNALLED (possibly |
406 |
* blocked waiting for a signal). When a worker is enqueued its |
407 |
* phase field is set negative. Note that phase field updates lag |
408 |
* queue CAS releases; seeing a negative phase does not guarantee |
409 |
* that the worker is available. When queued, the lower 16 bits of |
410 |
* its phase must hold its pool index. So we place the index there |
411 |
* upon initialization and never modify these bits. |
412 |
* |
413 |
* The ctl field also serves as the basis for memory |
414 |
* synchronization surrounding activation. This uses a more |
415 |
* efficient version of a Dekker-like rule that task producers and |
416 |
* consumers sync with each other by both writing/CASing ctl (even |
417 |
* if to its current value). However, rather than CASing ctl to |
418 |
* its current value in the common case where no action is |
419 |
* required, we reduce write contention by ensuring that |
420 |
* signalWork invocations are prefaced with a full-volatile memory |
421 |
* access (which is usually needed anyway). |
422 |
* |
423 |
* Signalling. Signals (in signalWork) cause new or reactivated |
424 |
* workers to scan for tasks. Method signalWork and its callers |
425 |
* try to approximate the unattainable goal of having the right |
426 |
* number of workers activated for the tasks at hand, but must err |
427 |
* on the side of too many workers vs too few to avoid stalls. If |
428 |
* computations are purely tree structured, it suffices for every |
429 |
* worker to activate another when it pushes a task into an empty |
430 |
* queue, resulting in O(log(#threads)) steps to full activation. |
431 |
* If instead, tasks come in serially from only a single producer, |
432 |
* each worker taking its first (since the last quiescence) task |
433 |
* from a queue should signal another if there are more tasks in |
434 |
* that queue. This is equivalent to, but generally faster than, |
435 |
* arranging the stealer take two tasks, re-pushing one on its own |
436 |
* queue, and signalling (because its queue is empty), also |
437 |
* resulting in logarithmic full activation time. Because we don't |
438 |
* know about usage patterns (or most commonly, mixtures), we use |
439 |
* both approaches. We approximate the second rule by arranging |
440 |
* that workers in scan() do not repeat signals when repeatedly |
441 |
* taking tasks from any given queue, by remembering the previous |
442 |
* one. There are narrow windows in which both rules may apply, |
443 |
* leading to duplicate or unnecessary signals. Despite such |
444 |
* limitations, these rules usually avoid slowdowns that otherwise |
445 |
* occur when too many workers contend to take too few tasks, or |
446 |
* when producers waste most of their time resignalling. However, |
447 |
* contention and overhead effects may still occur during ramp-up, |
448 |
* ramp-down, and small computations involving only a few workers. |
449 |
* |
450 |
* Scanning. Method scan performs top-level scanning for (and |
451 |
* execution of) tasks. Scans by different workers and/or at |
452 |
* different times are unlikely to poll queues in the same |
453 |
* order. Each scan traverses and tries to poll from each queue in |
454 |
* a pseudorandom permutation order by starting at a random index, |
455 |
* and using a constant cyclically exhaustive stride; restarting |
456 |
* upon contention. (Non-top-level scans; for example in |
457 |
* helpJoin, use simpler linear probes because they do not |
458 |
* systematically contend with top-level scans.) The pseudorandom |
459 |
* generator need not have high-quality statistical properties in |
460 |
* the long term. We use Marsaglia XorShifts, seeded with the Weyl |
461 |
* sequence from ThreadLocalRandom probes, which are cheap and |
462 |
* suffice. Scans do not otherwise explicitly take into account |
463 |
* core affinities, loads, cache localities, etc, However, they do |
464 |
* exploit temporal locality (which usually approximates these) by |
465 |
* preferring to re-poll from the same queue after a successful |
466 |
* poll before trying others (see method topLevelExec). This |
467 |
* reduces fairness, which is partially counteracted by using a |
468 |
* one-shot form of poll (tryPoll) that may lose to other workers. |
469 |
* |
470 |
* Deactivation. Method scan returns a sentinel when no tasks are |
471 |
* found, leading to deactivation (see awaitWork). The count |
472 |
* fields in ctl allow accurate discovery of quiescent states |
473 |
* (i.e., when all workers are idle) after deactivation. However, |
474 |
* this may also race with new (external) submissions, so a |
475 |
* recheck is also needed to determine quiescence. Upon apparently |
476 |
* triggering quiescence, awaitWork re-scans and self-signals if |
477 |
* it may have missed a signal. In other cases, a missed signal |
478 |
* may transiently lower parallelism because deactivation does not |
479 |
* necessarily mean that there is no more work, only that that |
480 |
* there were no tasks not taken by other workers. But more |
481 |
* signals are generated (see above) to eventually reactivate if |
482 |
* needed. |
483 |
* |
484 |
* Trimming workers. To release resources after periods of lack of |
485 |
* use, a worker starting to wait when the pool is quiescent will |
486 |
* time out and terminate if the pool has remained quiescent for |
487 |
* period given by field keepAlive. |
488 |
* |
489 |
* Shutdown and Termination. A call to shutdownNow invokes |
490 |
* tryTerminate to atomically set a mode bit. The calling thread, |
491 |
* as well as every other worker thereafter terminating, helps |
492 |
* terminate others by cancelling their unprocessed tasks, and |
493 |
* waking them up. Calls to non-abrupt shutdown() preface this by |
494 |
* checking isQuiescent before triggering the "STOP" phase of |
495 |
* termination. |
496 |
* |
497 |
* Joining Tasks |
498 |
* ============= |
499 |
* |
500 |
* Normally, the first option when joining a task that is not done |
501 |
* is to try to unfork it from local queue and run it. Otherwise, |
502 |
* any of several actions may be taken when one worker is waiting |
503 |
* to join a task stolen (or always held) by another. Because we |
504 |
* are multiplexing many tasks on to a pool of workers, we can't |
505 |
* always just let them block (as in Thread.join). We also cannot |
506 |
* just reassign the joiner's run-time stack with another and |
507 |
* replace it later, which would be a form of "continuation", that |
508 |
* even if possible is not necessarily a good idea since we may |
509 |
* need both an unblocked task and its continuation to progress. |
510 |
* Instead we combine two tactics: |
511 |
* |
512 |
* Helping: Arranging for the joiner to execute some task that it |
513 |
* could be running if the steal had not occurred. |
514 |
* |
515 |
* Compensating: Unless there are already enough live threads, |
516 |
* method tryCompensate() may create or re-activate a spare |
517 |
* thread to compensate for blocked joiners until they unblock. |
518 |
* |
519 |
* A third form (implemented via tryRemove) amounts to helping a |
520 |
* hypothetical compensator: If we can readily tell that a |
521 |
* possible action of a compensator is to steal and execute the |
522 |
* task being joined, the joining thread can do so directly, |
523 |
* without the need for a compensation thread; although with a |
524 |
* (rare) possibility of reduced parallelism because of a |
525 |
* transient gap in the queue array. |
526 |
* |
527 |
* Other intermediate forms available for specific task types (for |
528 |
* example helpAsyncBlocker) often avoid or postpone the need for |
529 |
* blocking or compensation. |
530 |
* |
531 |
* The ManagedBlocker extension API can't use helping so relies |
532 |
* only on compensation in method awaitBlocker. |
533 |
* |
534 |
* The algorithm in helpJoin entails a form of "linear helping". |
535 |
* Each worker records (in field "source") the id of the queue |
536 |
* from which it last stole a task. The scan in method helpJoin |
537 |
* uses these markers to try to find a worker to help (i.e., steal |
538 |
* back a task from and execute it) that could hasten completion |
539 |
* of the actively joined task. Thus, the joiner executes a task |
540 |
* that would be on its own local deque if the to-be-joined task |
541 |
* had not been stolen. This is a conservative variant of the |
542 |
* approach described in Wagner & Calder "Leapfrogging: a portable |
543 |
* technique for implementing efficient futures" SIGPLAN Notices, |
544 |
* 1993 (http://portal.acm.org/citation.cfm?id=155354). It differs |
545 |
* mainly in that we only record queue ids, not full dependency |
546 |
* links. This requires a linear scan of the queues array to |
547 |
* locate stealers, but isolates cost to when it is needed, rather |
548 |
* than adding to per-task overhead. Also, searches are limited to |
549 |
* direct and at most two levels of indirect stealers, after which |
550 |
* there are rapidly diminishing returns on increased overhead. |
551 |
* Searches can fail to locate stealers when stalls delay |
552 |
* recording sources. Further, even when accurately identified, |
553 |
* stealers might not ever produce a task that the joiner can in |
554 |
* turn help with. So, compensation is tried upon failure to find |
555 |
* tasks to run. |
556 |
* |
557 |
* Joining CountedCompleters (see helpComplete) differs from (and |
558 |
* is generally more efficient than) other cases because task |
559 |
* eligibility is determined by checking completion chains rather |
560 |
* than tracking stealers. |
561 |
* |
562 |
* Joining under timeouts (ForkJoinTask timed get) uses a |
563 |
* constrained mixture of helping and compensating in part because |
564 |
* pools (actually, only the common pool) may not have any |
565 |
* available threads: If the pool is saturated (all available |
566 |
* workers are busy), the caller tries to remove and otherwise |
567 |
* help; else it blocks under compensation so that it may time out |
568 |
* independently of any tasks. |
569 |
* |
570 |
* Compensation does not by default aim to keep exactly the target |
571 |
* parallelism number of unblocked threads running at any given |
572 |
* time. Some previous versions of this class employed immediate |
573 |
* compensations for any blocked join. However, in practice, the |
574 |
* vast majority of blockages are transient byproducts of GC and |
575 |
* other JVM or OS activities that are made worse by replacement |
576 |
* when they cause longer-term oversubscription. Rather than |
577 |
* impose arbitrary policies, we allow users to override the |
578 |
* default of only adding threads upon apparent starvation. The |
579 |
* compensation mechanism may also be bounded. Bounds for the |
580 |
* commonPool (see COMMON_MAX_SPARES) better enable JVMs to cope |
581 |
* with programming errors and abuse before running out of |
582 |
* resources to do so. |
583 |
* |
584 |
* Common Pool |
585 |
* =========== |
586 |
* |
587 |
* The static common pool always exists after static |
588 |
* initialization. Since it (or any other created pool) need |
589 |
* never be used, we minimize initial construction overhead and |
590 |
* footprint to the setup of about a dozen fields. |
591 |
* |
592 |
* When external threads submit to the common pool, they can |
593 |
* perform subtask processing (see helpComplete and related |
594 |
* methods) upon joins. This caller-helps policy makes it |
595 |
* sensible to set common pool parallelism level to one (or more) |
596 |
* less than the total number of available cores, or even zero for |
597 |
* pure caller-runs. We do not need to record whether external |
598 |
* submissions are to the common pool -- if not, external help |
599 |
* methods return quickly. These submitters would otherwise be |
600 |
* blocked waiting for completion, so the extra effort (with |
601 |
* liberally sprinkled task status checks) in inapplicable cases |
602 |
* amounts to an odd form of limited spin-wait before blocking in |
603 |
* ForkJoinTask.join. |
604 |
* |
605 |
* As a more appropriate default in managed environments, unless |
606 |
* overridden by system properties, we use workers of subclass |
607 |
* InnocuousForkJoinWorkerThread when there is a SecurityManager |
608 |
* present. These workers have no permissions set, do not belong |
609 |
* to any user-defined ThreadGroup, and erase all ThreadLocals |
610 |
* after executing any top-level task. The associated mechanics |
611 |
* may be JVM-dependent and must access particular Thread class |
612 |
* fields to achieve this effect. |
613 |
* |
614 |
* Interrupt handling |
615 |
* ================== |
616 |
* |
617 |
* The framework is designed to manage task cancellation |
618 |
* (ForkJoinTask.cancel) independently from the interrupt status |
619 |
* of threads running tasks. (See the public ForkJoinTask |
620 |
* documentation for rationale.) Interrupts are issued only in |
621 |
* tryTerminate, when workers should be terminating and tasks |
622 |
* should be cancelled anyway. Interrupts are cleared only when |
623 |
* necessary to ensure that calls to LockSupport.park do not loop |
624 |
* indefinitely (park returns immediately if the current thread is |
625 |
* interrupted). If so, interruption is reinstated after blocking |
626 |
* if status could be visible during the scope of any task. For |
627 |
* cases in which task bodies are specified or desired to |
628 |
* interrupt upon cancellation, ForkJoinTask.cancel can be |
629 |
* overridden to do so (as is done for invoke{Any,All}). |
630 |
* |
631 |
* Memory placement |
632 |
* ================ |
633 |
* |
634 |
* Performance can be very sensitive to placement of instances of |
635 |
* ForkJoinPool and WorkQueues and their queue arrays. To reduce |
636 |
* false-sharing impact, the @Contended annotation isolates the |
637 |
* ForkJoinPool.ctl field as well as the most heavily written |
638 |
* WorkQueue fields. These mainly reduce cache traffic by scanners. |
639 |
* WorkQueue arrays are presized large enough to avoid resizing |
640 |
* (which transiently reduces throughput) in most tree-like |
641 |
* computations, although not in some streaming usages. Initial |
642 |
* sizes are not large enough to avoid secondary contention |
643 |
* effects (especially for GC cardmarks) when queues are placed |
644 |
* near each other in memory. This is common, but has different |
645 |
* impact in different collectors and remains incompletely |
646 |
* addressed. |
647 |
* |
648 |
* Style notes |
649 |
* =========== |
650 |
* |
651 |
* Memory ordering relies mainly on atomic operations (CAS, |
652 |
* getAndSet, getAndAdd) along with explicit fences. This can be |
653 |
* awkward and ugly, but also reflects the need to control |
654 |
* outcomes across the unusual cases that arise in very racy code |
655 |
* with very few invariants. All fields are read into locals |
656 |
* before use, and null-checked if they are references, even if |
657 |
* they can never be null under current usages. Array accesses |
658 |
* using masked indices include checks (that are always true) that |
659 |
* the array length is non-zero to avoid compilers inserting more |
660 |
* expensive traps. This is usually done in a "C"-like style of |
661 |
* listing declarations at the heads of methods or blocks, and |
662 |
* using inline assignments on first encounter. Nearly all |
663 |
* explicit checks lead to bypass/return, not exception throws, |
664 |
* because they may legitimately arise during shutdown. |
665 |
* |
666 |
* There is a lot of representation-level coupling among classes |
667 |
* ForkJoinPool, ForkJoinWorkerThread, and ForkJoinTask. The |
668 |
* fields of WorkQueue maintain data structures managed by |
669 |
* ForkJoinPool, so are directly accessed. There is little point |
670 |
* trying to reduce this, since any associated future changes in |
671 |
* representations will need to be accompanied by algorithmic |
672 |
* changes anyway. Several methods intrinsically sprawl because |
673 |
* they must accumulate sets of consistent reads of fields held in |
674 |
* local variables. Some others are artificially broken up to |
675 |
* reduce producer/consumer imbalances due to dynamic compilation. |
676 |
* There are also other coding oddities (including several |
677 |
* unnecessary-looking hoisted null checks) that help some methods |
678 |
* perform reasonably even when interpreted (not compiled). |
679 |
* |
680 |
* The order of declarations in this file is (with a few exceptions): |
681 |
* (1) Static utility functions |
682 |
* (2) Nested (static) classes |
683 |
* (3) Static fields |
684 |
* (4) Fields, along with constants used when unpacking some of them |
685 |
* (5) Internal control methods |
686 |
* (6) Callbacks and other support for ForkJoinTask methods |
687 |
* (7) Exported methods |
688 |
* (8) Static block initializing statics in minimally dependent order |
689 |
* |
690 |
* Revision notes |
691 |
* ============== |
692 |
* |
693 |
* The main sources of differences of January 2020 ForkJoin |
694 |
* classes from previous version are: |
695 |
* |
696 |
* * ForkJoinTask now uses field "aux" to support blocking joins |
697 |
* and/or record exceptions, replacing reliance on builtin |
698 |
* monitors and side tables. |
699 |
* * Scans probe slots (vs compare indices), along with related |
700 |
* changes that reduce performance differences across most |
701 |
* garbage collectors, and reduce contention. |
702 |
* * Refactoring for better integration of special task types and |
703 |
* other capabilities that had been incrementally tacked on. Plus |
704 |
* many minor reworkings to improve consistency. |
705 |
*/ |
706 |
|
707 |
// Static utilities |
708 |
|
709 |
/** |
710 |
* If there is a security manager, makes sure caller has |
711 |
* permission to modify threads. |
712 |
*/ |
713 |
private static void checkPermission() { |
714 |
SecurityManager security = System.getSecurityManager(); |
715 |
if (security != null) |
716 |
security.checkPermission(modifyThreadPermission); |
717 |
} |
718 |
|
719 |
static AccessControlContext contextWithPermissions(Permission ... perms) { |
720 |
Permissions permissions = new Permissions(); |
721 |
for (Permission perm : perms) |
722 |
permissions.add(perm); |
723 |
return new AccessControlContext( |
724 |
new ProtectionDomain[] { new ProtectionDomain(null, permissions) }); |
725 |
} |
726 |
|
727 |
// Nested classes |
728 |
|
729 |
/** |
730 |
* Factory for creating new {@link ForkJoinWorkerThread}s. |
731 |
* A {@code ForkJoinWorkerThreadFactory} must be defined and used |
732 |
* for {@code ForkJoinWorkerThread} subclasses that extend base |
733 |
* functionality or initialize threads with different contexts. |
734 |
*/ |
735 |
public static interface ForkJoinWorkerThreadFactory { |
736 |
/** |
737 |
* Returns a new worker thread operating in the given pool. |
738 |
* Returning null or throwing an exception may result in tasks |
739 |
* never being executed. If this method throws an exception, |
740 |
* it is relayed to the caller of the method (for example |
741 |
* {@code execute}) causing attempted thread creation. If this |
742 |
* method returns null or throws an exception, it is not |
743 |
* retried until the next attempted creation (for example |
744 |
* another call to {@code execute}). |
745 |
* |
746 |
* @param pool the pool this thread works in |
747 |
* @return the new worker thread, or {@code null} if the request |
748 |
* to create a thread is rejected |
749 |
* @throws NullPointerException if the pool is null |
750 |
*/ |
751 |
public ForkJoinWorkerThread newThread(ForkJoinPool pool); |
752 |
} |
753 |
|
754 |
/** |
755 |
* Default ForkJoinWorkerThreadFactory implementation; creates a |
756 |
* new ForkJoinWorkerThread using the system class loader as the |
757 |
* thread context class loader. |
758 |
*/ |
759 |
static final class DefaultForkJoinWorkerThreadFactory |
760 |
implements ForkJoinWorkerThreadFactory { |
761 |
// ACC for access to the factory |
762 |
private static final AccessControlContext ACC = contextWithPermissions( |
763 |
new RuntimePermission("getClassLoader"), |
764 |
new RuntimePermission("setContextClassLoader")); |
765 |
public final ForkJoinWorkerThread newThread(ForkJoinPool pool) { |
766 |
return AccessController.doPrivileged( |
767 |
new PrivilegedAction<>() { |
768 |
public ForkJoinWorkerThread run() { |
769 |
return new ForkJoinWorkerThread(null, pool, true, false); |
770 |
}}, |
771 |
ACC); |
772 |
} |
773 |
} |
774 |
|
775 |
/** |
776 |
* Factory for CommonPool unless overridden by System property. |
777 |
* Creates InnocuousForkJoinWorkerThreads if a security manager is |
778 |
* present at time of invocation. Support requires that we break |
779 |
* quite a lot of encapsulation (some via helper methods in |
780 |
* ThreadLocalRandom) to access and set Thread fields. |
781 |
*/ |
782 |
static final class DefaultCommonPoolForkJoinWorkerThreadFactory |
783 |
implements ForkJoinWorkerThreadFactory { |
784 |
private static final AccessControlContext ACC = contextWithPermissions( |
785 |
modifyThreadPermission, |
786 |
new RuntimePermission("enableContextClassLoaderOverride"), |
787 |
new RuntimePermission("modifyThreadGroup"), |
788 |
new RuntimePermission("getClassLoader"), |
789 |
new RuntimePermission("setContextClassLoader")); |
790 |
|
791 |
public final ForkJoinWorkerThread newThread(ForkJoinPool pool) { |
792 |
return AccessController.doPrivileged( |
793 |
new PrivilegedAction<>() { |
794 |
public ForkJoinWorkerThread run() { |
795 |
return System.getSecurityManager() == null ? |
796 |
new ForkJoinWorkerThread(null, pool, true, true): |
797 |
new ForkJoinWorkerThread. |
798 |
InnocuousForkJoinWorkerThread(pool); }}, |
799 |
ACC); |
800 |
} |
801 |
} |
802 |
|
803 |
// Constants shared across ForkJoinPool and WorkQueue |
804 |
|
805 |
// Bounds |
806 |
static final int SWIDTH = 16; // width of short |
807 |
static final int SMASK = 0xffff; // short bits == max index |
808 |
static final int MAX_CAP = 0x7fff; // max #workers - 1 |
809 |
|
810 |
// Masks and units for WorkQueue.phase and ctl sp subfield |
811 |
static final int UNSIGNALLED = 1 << 31; // must be negative |
812 |
static final int SS_SEQ = 1 << 16; // version count |
813 |
|
814 |
// Mode bits and sentinels, some also used in WorkQueue fields |
815 |
static final int FIFO = 1 << 16; // fifo queue or access mode |
816 |
static final int SRC = 1 << 17; // set for valid queue ids |
817 |
static final int INNOCUOUS = 1 << 18; // set for Innocuous workers |
818 |
static final int QUIET = 1 << 19; // quiescing phase or source |
819 |
static final int SHUTDOWN = 1 << 24; |
820 |
static final int TERMINATED = 1 << 25; |
821 |
static final int STOP = 1 << 31; // must be negative |
822 |
static final int UNCOMPENSATE = 1 << 16; // tryCompensate return |
823 |
|
824 |
/** |
825 |
* Initial capacity of work-stealing queue array. Must be a power |
826 |
* of two, at least 2. See above. |
827 |
*/ |
828 |
static final int INITIAL_QUEUE_CAPACITY = 1 << 8; |
829 |
|
830 |
/** |
831 |
* Queues supporting work-stealing as well as external task |
832 |
* submission. See above for descriptions and algorithms. |
833 |
*/ |
834 |
static final class WorkQueue { |
835 |
volatile int phase; // versioned, negative if inactive |
836 |
int stackPred; // pool stack (ctl) predecessor link |
837 |
int config; // index, mode, ORed with SRC after init |
838 |
int base; // index of next slot for poll |
839 |
ForkJoinTask<?>[] array; // the queued tasks; power of 2 size |
840 |
final ForkJoinWorkerThread owner; // owning thread or null if shared |
841 |
|
842 |
// segregate fields frequently updated but not read by scans or steals |
843 |
@jdk.internal.vm.annotation.Contended("w") |
844 |
int top; // index of next slot for push |
845 |
@jdk.internal.vm.annotation.Contended("w") |
846 |
volatile int source; // source queue id, lock, or sentinel |
847 |
@jdk.internal.vm.annotation.Contended("w") |
848 |
int nsteals; // number of steals from other queues |
849 |
|
850 |
// Support for atomic operations |
851 |
private static final VarHandle QA; // for array slots |
852 |
private static final VarHandle SOURCE; |
853 |
private static final VarHandle BASE; |
854 |
static final ForkJoinTask<?> getSlot(ForkJoinTask<?>[] a, int i) { |
855 |
return (ForkJoinTask<?>)QA.getAcquire(a, i); |
856 |
} |
857 |
static final ForkJoinTask<?> getAndClearSlot(ForkJoinTask<?>[] a, |
858 |
int i) { |
859 |
return (ForkJoinTask<?>)QA.getAndSet(a, i, null); |
860 |
} |
861 |
static final void setSlotVolatile(ForkJoinTask<?>[] a, int i, |
862 |
ForkJoinTask<?> v) { |
863 |
QA.setVolatile(a, i, v); |
864 |
} |
865 |
static final boolean casSlotToNull(ForkJoinTask<?>[] a, int i, |
866 |
ForkJoinTask<?> c) { |
867 |
return QA.weakCompareAndSet(a, i, c, null); |
868 |
} |
869 |
final boolean tryLock() { |
870 |
return SOURCE.compareAndSet(this, 0, 1); |
871 |
} |
872 |
final void setBaseOpaque(int b) { |
873 |
BASE.setOpaque(this, b); |
874 |
} |
875 |
|
876 |
/** |
877 |
* Constructor used by ForkJoinWorkerThreads. Most fields |
878 |
* are initialized upon thread start, in pool.registerWorker. |
879 |
*/ |
880 |
WorkQueue(ForkJoinWorkerThread owner, boolean isInnocuous) { |
881 |
this.config = (isInnocuous) ? INNOCUOUS : 0; |
882 |
this.owner = owner; |
883 |
} |
884 |
|
885 |
/** |
886 |
* Constructor used for external queues. |
887 |
*/ |
888 |
WorkQueue(int config) { |
889 |
array = new ForkJoinTask<?>[INITIAL_QUEUE_CAPACITY]; |
890 |
this.config = config; |
891 |
owner = null; |
892 |
phase = -1; |
893 |
} |
894 |
|
895 |
/** |
896 |
* Returns an exportable index (used by ForkJoinWorkerThread). |
897 |
*/ |
898 |
final int getPoolIndex() { |
899 |
return (config & 0xffff) >>> 1; // ignore odd/even tag bit |
900 |
} |
901 |
|
902 |
/** |
903 |
* Returns the approximate number of tasks in the queue. |
904 |
*/ |
905 |
final int queueSize() { |
906 |
VarHandle.acquireFence(); // ensure fresh reads by external callers |
907 |
int n = top - base; |
908 |
return (n < 0) ? 0 : n; // ignore transient negative |
909 |
} |
910 |
|
911 |
/** |
912 |
* Provides a more conservative estimate of whether this queue |
913 |
* has any tasks than does queueSize. |
914 |
*/ |
915 |
final boolean isEmpty() { |
916 |
return !((source != 0 && owner == null) || top - base > 0); |
917 |
} |
918 |
|
919 |
/** |
920 |
* Pushes a task. Call only by owner in unshared queues. |
921 |
* |
922 |
* @param task the task. Caller must ensure non-null. |
923 |
* @param pool (no-op if null) |
924 |
* @throws RejectedExecutionException if array cannot be resized |
925 |
*/ |
926 |
final void push(ForkJoinTask<?> task, ForkJoinPool pool) { |
927 |
ForkJoinTask<?>[] a = array; |
928 |
int s = top++, d = s - base, cap, m; // skip insert if disabled |
929 |
if (a != null && pool != null && (cap = a.length) > 0) { |
930 |
setSlotVolatile(a, (m = cap - 1) & s, task); |
931 |
if (d == m) |
932 |
growArray(); |
933 |
if (d == m || a[m & (s - 1)] == null) |
934 |
pool.signalWork(); // signal if was empty or resized |
935 |
} |
936 |
} |
937 |
|
938 |
/** |
939 |
* Pushes task to a shared queue with lock already held, and unlocks. |
940 |
* |
941 |
* @return true if caller should signal work |
942 |
*/ |
943 |
final boolean lockedPush(ForkJoinTask<?> task) { |
944 |
ForkJoinTask<?>[] a = array; |
945 |
int s = top++, d = s - base, cap, m; |
946 |
if (a != null && (cap = a.length) > 0) { |
947 |
a[(m = cap - 1) & s] = task; |
948 |
if (d == m) |
949 |
growArray(); |
950 |
source = 0; // unlock |
951 |
if (d == m || a[m & (s - 1)] == null) |
952 |
return true; |
953 |
} |
954 |
return false; |
955 |
} |
956 |
|
957 |
/** |
958 |
* Doubles the capacity of array. Called by owner or with lock |
959 |
* held after pre-incrementing top, which is reverted on |
960 |
* allocation failure. |
961 |
*/ |
962 |
final void growArray() { |
963 |
ForkJoinTask<?>[] oldArray = array, newArray; |
964 |
int s = top - 1, oldCap, newCap; |
965 |
if (oldArray != null && (oldCap = oldArray.length) > 0 && |
966 |
(newCap = oldCap << 1) > 0) { // skip if disabled |
967 |
try { |
968 |
newArray = new ForkJoinTask<?>[newCap]; |
969 |
} catch (Throwable ex) { |
970 |
top = s; |
971 |
if (owner == null) |
972 |
source = 0; // unlock |
973 |
throw new RejectedExecutionException( |
974 |
"Queue capacity exceeded"); |
975 |
} |
976 |
int newMask = newCap - 1, oldMask = oldCap - 1; |
977 |
for (int k = oldCap; k > 0; --k, --s) { |
978 |
ForkJoinTask<?> x; // poll old, push to new |
979 |
if ((x = getAndClearSlot(oldArray, s & oldMask)) == null) |
980 |
break; // others already taken |
981 |
newArray[s & newMask] = x; |
982 |
} |
983 |
VarHandle.releaseFence(); // fill before publish |
984 |
array = newArray; |
985 |
} |
986 |
} |
987 |
|
988 |
// Variants of pop |
989 |
|
990 |
/** |
991 |
* Pops and returns task, or null if empty. Called only by owner. |
992 |
*/ |
993 |
private ForkJoinTask<?> pop() { |
994 |
ForkJoinTask<?> t = null; |
995 |
int s = top, cap; ForkJoinTask<?>[] a; |
996 |
if ((a = array) != null && (cap = a.length) > 0 && base != s-- && |
997 |
(t = getAndClearSlot(a, (cap - 1) & s)) != null) |
998 |
top = s; |
999 |
return t; |
1000 |
} |
1001 |
|
1002 |
/** |
1003 |
* Pops the given task for owner only if it is at the current top. |
1004 |
*/ |
1005 |
final boolean tryUnpush(ForkJoinTask<?> task) { |
1006 |
int s = top, cap; ForkJoinTask<?>[] a; |
1007 |
if ((a = array) != null && (cap = a.length) > 0 && base != s-- && |
1008 |
casSlotToNull(a, (cap - 1) & s, task)) { |
1009 |
top = s; |
1010 |
return true; |
1011 |
} |
1012 |
return false; |
1013 |
} |
1014 |
|
1015 |
/** |
1016 |
* Locking version of tryUnpush. |
1017 |
*/ |
1018 |
final boolean externalTryUnpush(ForkJoinTask<?> task) { |
1019 |
boolean taken = false; |
1020 |
for (;;) { |
1021 |
int s = top, cap, k; ForkJoinTask<?>[] a; |
1022 |
if ((a = array) == null || (cap = a.length) <= 0 || |
1023 |
a[k = (cap - 1) & (s - 1)] != task) |
1024 |
break; |
1025 |
if (tryLock()) { |
1026 |
if (top == s && array == a) { |
1027 |
if (taken = casSlotToNull(a, k, task)) { |
1028 |
top = s - 1; |
1029 |
source = 0; |
1030 |
break; |
1031 |
} |
1032 |
} |
1033 |
source = 0; // release lock for retry |
1034 |
} |
1035 |
Thread.yield(); // trylock failure |
1036 |
} |
1037 |
return taken; |
1038 |
} |
1039 |
|
1040 |
/** |
1041 |
* Deep form of tryUnpush: Traverses from top and removes task if |
1042 |
* present, shifting others to fill gap. |
1043 |
*/ |
1044 |
final boolean tryRemove(ForkJoinTask<?> task, boolean owned) { |
1045 |
boolean taken = false; |
1046 |
int p = top, cap; ForkJoinTask<?>[] a; ForkJoinTask<?> t; |
1047 |
if ((a = array) != null && task != null && (cap = a.length) > 0) { |
1048 |
int m = cap - 1, s = p - 1, d = p - base; |
1049 |
for (int i = s, k; d > 0; --i, --d) { |
1050 |
if ((t = a[k = i & m]) == task) { |
1051 |
if (owned || tryLock()) { |
1052 |
if ((owned || (array == a && top == p)) && |
1053 |
(taken = casSlotToNull(a, k, t))) { |
1054 |
for (int j = i; j != s; ) // shift down |
1055 |
a[j & m] = getAndClearSlot(a, ++j & m); |
1056 |
top = s; |
1057 |
} |
1058 |
if (!owned) |
1059 |
source = 0; |
1060 |
} |
1061 |
break; |
1062 |
} |
1063 |
} |
1064 |
} |
1065 |
return taken; |
1066 |
} |
1067 |
|
1068 |
// variants of poll |
1069 |
|
1070 |
/** |
1071 |
* Tries once to poll next task in FIFO order, failing on |
1072 |
* inconsistency or contention. |
1073 |
*/ |
1074 |
final ForkJoinTask<?> tryPoll() { |
1075 |
int cap, b, k; ForkJoinTask<?>[] a; |
1076 |
if ((a = array) != null && (cap = a.length) > 0) { |
1077 |
ForkJoinTask<?> t = getSlot(a, k = (cap - 1) & (b = base)); |
1078 |
if (base == b++ && t != null && casSlotToNull(a, k, t)) { |
1079 |
setBaseOpaque(b); |
1080 |
return t; |
1081 |
} |
1082 |
} |
1083 |
return null; |
1084 |
} |
1085 |
|
1086 |
/** |
1087 |
* Takes next task, if one exists, in order specified by mode. |
1088 |
*/ |
1089 |
final ForkJoinTask<?> nextLocalTask(int cfg) { |
1090 |
ForkJoinTask<?> t = null; |
1091 |
int s = top, cap; ForkJoinTask<?>[] a; |
1092 |
if ((a = array) != null && (cap = a.length) > 0) { |
1093 |
for (int b, d;;) { |
1094 |
if ((d = s - (b = base)) <= 0) |
1095 |
break; |
1096 |
if (d == 1 || (cfg & FIFO) == 0) { |
1097 |
if ((t = getAndClearSlot(a, --s & (cap - 1))) != null) |
1098 |
top = s; |
1099 |
break; |
1100 |
} |
1101 |
if ((t = getAndClearSlot(a, b++ & (cap - 1))) != null) { |
1102 |
setBaseOpaque(b); |
1103 |
break; |
1104 |
} |
1105 |
} |
1106 |
} |
1107 |
return t; |
1108 |
} |
1109 |
|
1110 |
/** |
1111 |
* Takes next task, if one exists, using configured mode. |
1112 |
*/ |
1113 |
final ForkJoinTask<?> nextLocalTask() { |
1114 |
return nextLocalTask(config); |
1115 |
} |
1116 |
|
1117 |
/** |
1118 |
* Returns next task, if one exists, in order specified by mode. |
1119 |
*/ |
1120 |
final ForkJoinTask<?> peek() { |
1121 |
VarHandle.acquireFence(); |
1122 |
int cap; ForkJoinTask<?>[] a; |
1123 |
return ((a = array) != null && (cap = a.length) > 0) ? |
1124 |
a[(cap - 1) & ((config & FIFO) != 0 ? base : top - 1)] : null; |
1125 |
} |
1126 |
|
1127 |
// specialized execution methods |
1128 |
|
1129 |
/** |
1130 |
* Runs the given (stolen) task if nonnull, as well as |
1131 |
* remaining local tasks and/or others available from the |
1132 |
* given queue. |
1133 |
*/ |
1134 |
final void topLevelExec(ForkJoinTask<?> task, WorkQueue q) { |
1135 |
int cfg = config, nstolen = 1; |
1136 |
while (task != null) { |
1137 |
task.doExec(); |
1138 |
if ((task = nextLocalTask(cfg)) == null && |
1139 |
q != null && (task = q.tryPoll()) != null) |
1140 |
++nstolen; |
1141 |
} |
1142 |
nsteals += nstolen; |
1143 |
source = 0; |
1144 |
if ((cfg & INNOCUOUS) != 0) |
1145 |
ThreadLocalRandom.eraseThreadLocals(Thread.currentThread()); |
1146 |
} |
1147 |
|
1148 |
/** |
1149 |
* Tries to pop and run tasks within the target's computation |
1150 |
* until done, not found, or limit exceeded. |
1151 |
* |
1152 |
* @param task root of CountedCompleter computation |
1153 |
* @param owned true if owned by a ForkJoinWorkerThread |
1154 |
* @param limit max runs, or zero for no limit |
1155 |
* @return task status on exit |
1156 |
*/ |
1157 |
final int helpComplete(ForkJoinTask<?> task, boolean owned, int limit) { |
1158 |
int status = 0, cap, k, p, s; ForkJoinTask<?>[] a; ForkJoinTask<?> t; |
1159 |
while (task != null && (status = task.status) >= 0 && |
1160 |
(a = array) != null && (cap = a.length) > 0 && |
1161 |
(t = a[k = (cap - 1) & (s = (p = top) - 1)]) |
1162 |
instanceof CountedCompleter) { |
1163 |
CountedCompleter<?> f = (CountedCompleter<?>)t; |
1164 |
boolean taken = false; |
1165 |
for (;;) { // exec if root task is a completer of t |
1166 |
if (f == task) { |
1167 |
if (owned) { |
1168 |
if ((taken = casSlotToNull(a, k, t))) |
1169 |
top = s; |
1170 |
} |
1171 |
else if (tryLock()) { |
1172 |
if (top == p && array == a && |
1173 |
(taken = casSlotToNull(a, k, t))) |
1174 |
top = s; |
1175 |
source = 0; |
1176 |
} |
1177 |
if (taken) |
1178 |
t.doExec(); |
1179 |
else if (!owned) |
1180 |
Thread.yield(); // tryLock failure |
1181 |
break; |
1182 |
} |
1183 |
else if ((f = f.completer) == null) |
1184 |
break; |
1185 |
} |
1186 |
if (taken && limit != 0 && --limit == 0) |
1187 |
break; |
1188 |
} |
1189 |
return status; |
1190 |
} |
1191 |
|
1192 |
/** |
1193 |
* Tries to poll and run AsynchronousCompletionTasks until |
1194 |
* none found or blocker is released. |
1195 |
* |
1196 |
* @param blocker the blocker |
1197 |
*/ |
1198 |
final void helpAsyncBlocker(ManagedBlocker blocker) { |
1199 |
int cap, b, d, k; ForkJoinTask<?>[] a; ForkJoinTask<?> t; |
1200 |
while (blocker != null && (d = top - (b = base)) > 0 && |
1201 |
(a = array) != null && (cap = a.length) > 0 && |
1202 |
(((t = getSlot(a, k = (cap - 1) & b)) == null && d > 1) || |
1203 |
t instanceof |
1204 |
CompletableFuture.AsynchronousCompletionTask) && |
1205 |
!blocker.isReleasable()) { |
1206 |
if (t != null && base == b++ && casSlotToNull(a, k, t)) { |
1207 |
setBaseOpaque(b); |
1208 |
t.doExec(); |
1209 |
} |
1210 |
} |
1211 |
} |
1212 |
|
1213 |
// misc |
1214 |
|
1215 |
/** AccessControlContext for innocuous workers, created on 1st use. */ |
1216 |
private static AccessControlContext INNOCUOUS_ACC; |
1217 |
|
1218 |
/** |
1219 |
* Initializes (upon registration) InnocuousForkJoinWorkerThreads. |
1220 |
*/ |
1221 |
final void initializeInnocuousWorker() { |
1222 |
AccessControlContext acc; // racy construction OK |
1223 |
if ((acc = INNOCUOUS_ACC) == null) |
1224 |
INNOCUOUS_ACC = acc = new AccessControlContext( |
1225 |
new ProtectionDomain[] { new ProtectionDomain(null, null) }); |
1226 |
Thread t = Thread.currentThread(); |
1227 |
ThreadLocalRandom.setInheritedAccessControlContext(t, acc); |
1228 |
ThreadLocalRandom.eraseThreadLocals(t); |
1229 |
} |
1230 |
|
1231 |
/** |
1232 |
* Returns true if owned by a worker thread and not known to be blocked. |
1233 |
*/ |
1234 |
final boolean isApparentlyUnblocked() { |
1235 |
Thread wt; Thread.State s; |
1236 |
return ((wt = owner) != null && |
1237 |
(s = wt.getState()) != Thread.State.BLOCKED && |
1238 |
s != Thread.State.WAITING && |
1239 |
s != Thread.State.TIMED_WAITING); |
1240 |
} |
1241 |
|
1242 |
static { |
1243 |
try { |
1244 |
QA = MethodHandles.arrayElementVarHandle(ForkJoinTask[].class); |
1245 |
MethodHandles.Lookup l = MethodHandles.lookup(); |
1246 |
SOURCE = l.findVarHandle(WorkQueue.class, "source", int.class); |
1247 |
BASE = l.findVarHandle(WorkQueue.class, "base", int.class); |
1248 |
} catch (ReflectiveOperationException e) { |
1249 |
throw new ExceptionInInitializerError(e); |
1250 |
} |
1251 |
} |
1252 |
} |
1253 |
|
1254 |
// static fields (initialized in static initializer below) |
1255 |
|
1256 |
/** |
1257 |
* Creates a new ForkJoinWorkerThread. This factory is used unless |
1258 |
* overridden in ForkJoinPool constructors. |
1259 |
*/ |
1260 |
public static final ForkJoinWorkerThreadFactory |
1261 |
defaultForkJoinWorkerThreadFactory; |
1262 |
|
1263 |
/** |
1264 |
* Permission required for callers of methods that may start or |
1265 |
* kill threads. |
1266 |
*/ |
1267 |
static final RuntimePermission modifyThreadPermission; |
1268 |
|
1269 |
/** |
1270 |
* Common (static) pool. Non-null for public use unless a static |
1271 |
* construction exception, but internal usages null-check on use |
1272 |
* to paranoically avoid potential initialization circularities |
1273 |
* as well as to simplify generated code. |
1274 |
*/ |
1275 |
static final ForkJoinPool common; |
1276 |
|
1277 |
/** |
1278 |
* Common pool parallelism. To allow simpler use and management |
1279 |
* when common pool threads are disabled, we allow the underlying |
1280 |
* common.parallelism field to be zero, but in that case still report |
1281 |
* parallelism as 1 to reflect resulting caller-runs mechanics. |
1282 |
*/ |
1283 |
static final int COMMON_PARALLELISM; |
1284 |
|
1285 |
/** |
1286 |
* Limit on spare thread construction in tryCompensate. |
1287 |
*/ |
1288 |
private static final int COMMON_MAX_SPARES; |
1289 |
|
1290 |
/** |
1291 |
* Sequence number for creating worker names |
1292 |
*/ |
1293 |
private static volatile int poolIds; |
1294 |
|
1295 |
// static configuration constants |
1296 |
|
1297 |
/** |
1298 |
* Default idle timeout value (in milliseconds) for the thread |
1299 |
* triggering quiescence to park waiting for new work |
1300 |
*/ |
1301 |
private static final long DEFAULT_KEEPALIVE = 60_000L; |
1302 |
|
1303 |
/** |
1304 |
* Undershoot tolerance for idle timeouts |
1305 |
*/ |
1306 |
private static final long TIMEOUT_SLOP = 20L; |
1307 |
|
1308 |
/** |
1309 |
* The default value for COMMON_MAX_SPARES. Overridable using the |
1310 |
* "java.util.concurrent.ForkJoinPool.common.maximumSpares" system |
1311 |
* property. The default value is far in excess of normal |
1312 |
* requirements, but also far short of MAX_CAP and typical OS |
1313 |
* thread limits, so allows JVMs to catch misuse/abuse before |
1314 |
* running out of resources needed to do so. |
1315 |
*/ |
1316 |
private static final int DEFAULT_COMMON_MAX_SPARES = 256; |
1317 |
|
1318 |
/* |
1319 |
* Bits and masks for field ctl, packed with 4 16 bit subfields: |
1320 |
* RC: Number of released (unqueued) workers minus target parallelism |
1321 |
* TC: Number of total workers minus target parallelism |
1322 |
* SS: version count and status of top waiting thread |
1323 |
* ID: poolIndex of top of Treiber stack of waiters |
1324 |
* |
1325 |
* When convenient, we can extract the lower 32 stack top bits |
1326 |
* (including version bits) as sp=(int)ctl. The offsets of counts |
1327 |
* by the target parallelism and the positionings of fields makes |
1328 |
* it possible to perform the most common checks via sign tests of |
1329 |
* fields: When ac is negative, there are not enough unqueued |
1330 |
* workers, when tc is negative, there are not enough total |
1331 |
* workers. When sp is non-zero, there are waiting workers. To |
1332 |
* deal with possibly negative fields, we use casts in and out of |
1333 |
* "short" and/or signed shifts to maintain signedness. |
1334 |
* |
1335 |
* Because it occupies uppermost bits, we can add one release |
1336 |
* count using getAndAdd of RC_UNIT, rather than CAS, when |
1337 |
* returning from a blocked join. Other updates entail multiple |
1338 |
* subfields and masking, requiring CAS. |
1339 |
* |
1340 |
* The limits packed in field "bounds" are also offset by the |
1341 |
* parallelism level to make them comparable to the ctl rc and tc |
1342 |
* fields. |
1343 |
*/ |
1344 |
|
1345 |
// Lower and upper word masks |
1346 |
private static final long SP_MASK = 0xffffffffL; |
1347 |
private static final long UC_MASK = ~SP_MASK; |
1348 |
|
1349 |
// Release counts |
1350 |
private static final int RC_SHIFT = 48; |
1351 |
private static final long RC_UNIT = 0x0001L << RC_SHIFT; |
1352 |
private static final long RC_MASK = 0xffffL << RC_SHIFT; |
1353 |
|
1354 |
// Total counts |
1355 |
private static final int TC_SHIFT = 32; |
1356 |
private static final long TC_UNIT = 0x0001L << TC_SHIFT; |
1357 |
private static final long TC_MASK = 0xffffL << TC_SHIFT; |
1358 |
private static final long ADD_WORKER = 0x0001L << (TC_SHIFT + 15); // sign |
1359 |
|
1360 |
// Instance fields |
1361 |
|
1362 |
final long keepAlive; // milliseconds before dropping if idle |
1363 |
volatile long stealCount; // collects worker nsteals |
1364 |
int scanRover; // advances across pollScan calls |
1365 |
volatile int threadIds; // for worker thread names |
1366 |
final int bounds; // min, max threads packed as shorts |
1367 |
volatile int mode; // parallelism, runstate, queue mode |
1368 |
WorkQueue[] queues; // main registry |
1369 |
final ReentrantLock registrationLock; |
1370 |
Condition termination; // lazily constructed |
1371 |
final String workerNamePrefix; // null for common pool |
1372 |
final ForkJoinWorkerThreadFactory factory; |
1373 |
final UncaughtExceptionHandler ueh; // per-worker UEH |
1374 |
final Predicate<? super ForkJoinPool> saturate; |
1375 |
|
1376 |
@jdk.internal.vm.annotation.Contended("fjpctl") // segregate |
1377 |
volatile long ctl; // main pool control |
1378 |
|
1379 |
// Support for atomic operations |
1380 |
private static final VarHandle CTL; |
1381 |
private static final VarHandle MODE; |
1382 |
private static final VarHandle THREADIDS; |
1383 |
private static final VarHandle POOLIDS; |
1384 |
private boolean compareAndSetCtl(long c, long v) { |
1385 |
return CTL.compareAndSet(this, c, v); |
1386 |
} |
1387 |
private long compareAndExchangeCtl(long c, long v) { |
1388 |
return (long)CTL.compareAndExchange(this, c, v); |
1389 |
} |
1390 |
private long getAndAddCtl(long v) { |
1391 |
return (long)CTL.getAndAdd(this, v); |
1392 |
} |
1393 |
private int getAndBitwiseOrMode(int v) { |
1394 |
return (int)MODE.getAndBitwiseOr(this, v); |
1395 |
} |
1396 |
private int getAndAddThreadIds(int x) { |
1397 |
return (int)THREADIDS.getAndAdd(this, x); |
1398 |
} |
1399 |
private static int getAndAddPoolIds(int x) { |
1400 |
return (int)POOLIDS.getAndAdd(x); |
1401 |
} |
1402 |
|
1403 |
// Creating, registering and deregistering workers |
1404 |
|
1405 |
/** |
1406 |
* Tries to construct and start one worker. Assumes that total |
1407 |
* count has already been incremented as a reservation. Invokes |
1408 |
* deregisterWorker on any failure. |
1409 |
* |
1410 |
* @return true if successful |
1411 |
*/ |
1412 |
private boolean createWorker() { |
1413 |
ForkJoinWorkerThreadFactory fac = factory; |
1414 |
Throwable ex = null; |
1415 |
ForkJoinWorkerThread wt = null; |
1416 |
try { |
1417 |
if (fac != null && (wt = fac.newThread(this)) != null) { |
1418 |
wt.start(); |
1419 |
return true; |
1420 |
} |
1421 |
} catch (Throwable rex) { |
1422 |
ex = rex; |
1423 |
} |
1424 |
deregisterWorker(wt, ex); |
1425 |
return false; |
1426 |
} |
1427 |
|
1428 |
/** |
1429 |
* Provides a name for ForkJoinWorkerThread constructor. |
1430 |
*/ |
1431 |
final String nextWorkerThreadName() { |
1432 |
String prefix = workerNamePrefix; |
1433 |
int tid = getAndAddThreadIds(1) + 1; |
1434 |
if (prefix == null) // commonPool has no prefix |
1435 |
prefix = "ForkJoinPool.commonPool-worker-"; |
1436 |
return prefix.concat(Integer.toString(tid)); |
1437 |
} |
1438 |
|
1439 |
/** |
1440 |
* Finishes initializing and records owned queue. |
1441 |
* |
1442 |
* @param w caller's WorkQueue |
1443 |
*/ |
1444 |
final void registerWorker(WorkQueue w) { |
1445 |
ReentrantLock lock = registrationLock; |
1446 |
ThreadLocalRandom.localInit(); |
1447 |
int seed = ThreadLocalRandom.getProbe(); |
1448 |
if (w != null && lock != null) { |
1449 |
int modebits = (mode & FIFO) | w.config; |
1450 |
w.array = new ForkJoinTask<?>[INITIAL_QUEUE_CAPACITY]; |
1451 |
w.stackPred = seed; // stash for runWorker |
1452 |
if ((modebits & INNOCUOUS) != 0) |
1453 |
w.initializeInnocuousWorker(); |
1454 |
int id = (seed << 1) | 1; // initial index guess |
1455 |
lock.lock(); |
1456 |
try { |
1457 |
WorkQueue[] qs; int n; // find queue index |
1458 |
if ((qs = queues) != null && (n = qs.length) > 0) { |
1459 |
int k = n, m = n - 1; |
1460 |
for (; qs[id &= m] != null && k > 0; id -= 2, k -= 2); |
1461 |
if (k == 0) |
1462 |
id = n | 1; // resize below |
1463 |
w.phase = w.config = id | modebits; // now publishable |
1464 |
|
1465 |
if (id < n) |
1466 |
qs[id] = w; |
1467 |
else { // expand array |
1468 |
int an = n << 1, am = an - 1; |
1469 |
WorkQueue[] as = new WorkQueue[an]; |
1470 |
as[id & am] = w; |
1471 |
for (int j = 1; j < n; j += 2) |
1472 |
as[j] = qs[j]; |
1473 |
for (int j = 0; j < n; j += 2) { |
1474 |
WorkQueue q; |
1475 |
if ((q = qs[j]) != null) // shared queues may move |
1476 |
as[q.config & am] = q; |
1477 |
} |
1478 |
VarHandle.releaseFence(); // fill before publish |
1479 |
queues = as; |
1480 |
} |
1481 |
} |
1482 |
} finally { |
1483 |
lock.unlock(); |
1484 |
} |
1485 |
} |
1486 |
} |
1487 |
|
1488 |
/** |
1489 |
* Final callback from terminating worker, as well as upon failure |
1490 |
* to construct or start a worker. Removes record of worker from |
1491 |
* array, and adjusts counts. If pool is shutting down, tries to |
1492 |
* complete termination. |
1493 |
* |
1494 |
* @param wt the worker thread, or null if construction failed |
1495 |
* @param ex the exception causing failure, or null if none |
1496 |
*/ |
1497 |
final void deregisterWorker(ForkJoinWorkerThread wt, Throwable ex) { |
1498 |
ReentrantLock lock = registrationLock; |
1499 |
WorkQueue w = null; |
1500 |
int cfg = 0; |
1501 |
if (wt != null && (w = wt.workQueue) != null && lock != null) { |
1502 |
WorkQueue[] qs; int n, i; |
1503 |
cfg = w.config; |
1504 |
long ns = w.nsteals & 0xffffffffL; |
1505 |
lock.lock(); // remove index from array |
1506 |
if ((qs = queues) != null && (n = qs.length) > 0 && |
1507 |
qs[i = cfg & (n - 1)] == w) |
1508 |
qs[i] = null; |
1509 |
stealCount += ns; // accumulate steals |
1510 |
lock.unlock(); |
1511 |
long c = ctl; |
1512 |
if ((cfg & QUIET) == 0) // unless self-signalled, decrement counts |
1513 |
do {} while (c != (c = compareAndExchangeCtl( |
1514 |
c, ((RC_MASK & (c - RC_UNIT)) | |
1515 |
(TC_MASK & (c - TC_UNIT)) | |
1516 |
(SP_MASK & c))))); |
1517 |
else if ((int)c == 0) // was dropped on timeout |
1518 |
cfg = 0; // suppress signal if last |
1519 |
for (ForkJoinTask<?> t; (t = w.pop()) != null; ) |
1520 |
ForkJoinTask.cancelIgnoringExceptions(t); // cancel tasks |
1521 |
} |
1522 |
|
1523 |
if (!tryTerminate(false, false) && w != null && (cfg & SRC) != 0) |
1524 |
signalWork(); // possibly replace worker |
1525 |
if (ex != null) |
1526 |
ForkJoinTask.rethrow(ex); |
1527 |
} |
1528 |
|
1529 |
/* |
1530 |
* Tries to create or release a worker if too few are running. |
1531 |
*/ |
1532 |
final void signalWork() { |
1533 |
for (long c = ctl; c < 0L;) { |
1534 |
int sp, i; WorkQueue[] qs; WorkQueue v; |
1535 |
if ((sp = (int)c & ~UNSIGNALLED) == 0) { // no idle workers |
1536 |
if ((c & ADD_WORKER) == 0L) // enough total workers |
1537 |
break; |
1538 |
if (c == (c = compareAndExchangeCtl( |
1539 |
c, ((RC_MASK & (c + RC_UNIT)) | |
1540 |
(TC_MASK & (c + TC_UNIT)))))) { |
1541 |
createWorker(); |
1542 |
break; |
1543 |
} |
1544 |
} |
1545 |
else if ((qs = queues) == null) |
1546 |
break; // unstarted/terminated |
1547 |
else if (qs.length <= (i = sp & SMASK)) |
1548 |
break; // terminated |
1549 |
else if ((v = qs[i]) == null) |
1550 |
break; // terminating |
1551 |
else { |
1552 |
long nc = (v.stackPred & SP_MASK) | (UC_MASK & (c + RC_UNIT)); |
1553 |
Thread vt = v.owner; |
1554 |
if (c == (c = compareAndExchangeCtl(c, nc))) { |
1555 |
v.phase = sp; |
1556 |
LockSupport.unpark(vt); // release idle worker |
1557 |
break; |
1558 |
} |
1559 |
} |
1560 |
} |
1561 |
} |
1562 |
|
1563 |
/** |
1564 |
* Top-level runloop for workers, called by ForkJoinWorkerThread.run. |
1565 |
* See above for explanation. |
1566 |
* |
1567 |
* @param w caller's WorkQueue (may be null on failed initialization) |
1568 |
*/ |
1569 |
final void runWorker(WorkQueue w) { |
1570 |
if (mode >= 0 && w != null) { // skip on failed init |
1571 |
w.config |= SRC; // mark as valid source |
1572 |
int r = w.stackPred, src = 0; // use seed from registerWorker |
1573 |
do { |
1574 |
r ^= r << 13; r ^= r >>> 17; r ^= r << 5; // xorshift |
1575 |
} while ((src = scan(w, src, r)) >= 0 || |
1576 |
(src = awaitWork(w)) == 0); |
1577 |
} |
1578 |
} |
1579 |
|
1580 |
/** |
1581 |
* Scans for and if found executes top-level tasks: Tries to poll |
1582 |
* each queue starting at a random index with random stride, |
1583 |
* returning source id or retry indicator if contended or |
1584 |
* inconsistent. |
1585 |
* |
1586 |
* @param w caller's WorkQueue |
1587 |
* @param prevSrc the previous queue stolen from in current phase, or 0 |
1588 |
* @param r random seed |
1589 |
* @return id of queue if taken, negative if none found, prevSrc for retry |
1590 |
*/ |
1591 |
private int scan(WorkQueue w, int prevSrc, int r) { |
1592 |
WorkQueue[] qs = queues; |
1593 |
int n = (w == null || qs == null) ? 0 : qs.length; |
1594 |
for (int step = (r >>> 16) | 1, i = n; i > 0; --i, r += step) { |
1595 |
int j, cap, b; WorkQueue q; ForkJoinTask<?>[] a; |
1596 |
if ((q = qs[j = r & (n - 1)]) != null && // poll at qs[j].array[k] |
1597 |
(a = q.array) != null && (cap = a.length) > 0) { |
1598 |
int k = (cap - 1) & (b = q.base), nextBase = b + 1; |
1599 |
int nextIndex = (cap - 1) & nextBase, src = j | SRC; |
1600 |
ForkJoinTask<?> t = WorkQueue.getSlot(a, k); |
1601 |
if (q.base != b) // inconsistent |
1602 |
return prevSrc; |
1603 |
else if (t != null && WorkQueue.casSlotToNull(a, k, t)) { |
1604 |
q.base = nextBase; |
1605 |
ForkJoinTask<?> next = a[nextIndex]; |
1606 |
if ((w.source = src) != prevSrc && next != null) |
1607 |
signalWork(); // propagate |
1608 |
w.topLevelExec(t, q); |
1609 |
return src; |
1610 |
} |
1611 |
else if (a[nextIndex] != null) // revisit |
1612 |
return prevSrc; |
1613 |
} |
1614 |
} |
1615 |
return (queues != qs) ? prevSrc: -1; // possibly resized |
1616 |
} |
1617 |
|
1618 |
/** |
1619 |
* Advances worker phase, pushes onto ctl stack, and awaits signal |
1620 |
* or reports termination. |
1621 |
* |
1622 |
* @return negative if terminated, else 0 |
1623 |
*/ |
1624 |
private int awaitWork(WorkQueue w) { |
1625 |
if (w == null) |
1626 |
return -1; // already terminated |
1627 |
int phase = (w.phase + SS_SEQ) & ~UNSIGNALLED; |
1628 |
w.phase = phase | UNSIGNALLED; // advance phase |
1629 |
long prevCtl = ctl, c; // enqueue |
1630 |
do { |
1631 |
w.stackPred = (int)prevCtl; |
1632 |
c = ((prevCtl - RC_UNIT) & UC_MASK) | (phase & SP_MASK); |
1633 |
} while (prevCtl != (prevCtl = compareAndExchangeCtl(prevCtl, c))); |
1634 |
|
1635 |
Thread.interrupted(); // clear status |
1636 |
LockSupport.setCurrentBlocker(this); // prepare to block (exit also OK) |
1637 |
long deadline = 0L; // nonzero if possibly quiescent |
1638 |
int ac = (int)(c >> RC_SHIFT), md; |
1639 |
if ((md = mode) < 0) // pool is terminating |
1640 |
return -1; |
1641 |
else if ((md & SMASK) + ac <= 0) { |
1642 |
boolean checkTermination = (md & SHUTDOWN) != 0; |
1643 |
if ((deadline = System.currentTimeMillis() + keepAlive) == 0L) |
1644 |
deadline = 1L; // avoid zero |
1645 |
WorkQueue[] qs = queues; // check for racing submission |
1646 |
int n = (qs == null) ? 0 : qs.length; |
1647 |
for (int i = 0; i < n; i += 2) { |
1648 |
WorkQueue q; ForkJoinTask<?>[] a; int cap, b; |
1649 |
if (ctl != c) { // already signalled |
1650 |
checkTermination = false; |
1651 |
break; |
1652 |
} |
1653 |
else if ((q = qs[i]) != null && |
1654 |
(a = q.array) != null && (cap = a.length) > 0 && |
1655 |
((b = q.base) != q.top || a[(cap - 1) & b] != null || |
1656 |
q.source != 0)) { |
1657 |
if (compareAndSetCtl(c, prevCtl)) |
1658 |
w.phase = phase; // self-signal |
1659 |
checkTermination = false; |
1660 |
break; |
1661 |
} |
1662 |
} |
1663 |
if (checkTermination && tryTerminate(false, false)) |
1664 |
return -1; // trigger quiescent termination |
1665 |
} |
1666 |
|
1667 |
for (boolean alt = false;;) { // await activation or termination |
1668 |
if (w.phase >= 0) |
1669 |
break; |
1670 |
else if (mode < 0) |
1671 |
return -1; |
1672 |
else if ((c = ctl) == prevCtl) |
1673 |
Thread.onSpinWait(); // signal in progress |
1674 |
else if (!(alt = !alt)) // check between park calls |
1675 |
Thread.interrupted(); |
1676 |
else if (deadline == 0L) |
1677 |
LockSupport.park(); |
1678 |
else if (deadline - System.currentTimeMillis() > TIMEOUT_SLOP) |
1679 |
LockSupport.parkUntil(deadline); |
1680 |
else if (((int)c & SMASK) == (w.config & SMASK) && |
1681 |
compareAndSetCtl(c, ((UC_MASK & (c - TC_UNIT)) | |
1682 |
(prevCtl & SP_MASK)))) { |
1683 |
w.config |= QUIET; // sentinel for deregisterWorker |
1684 |
return -1; // drop on timeout |
1685 |
} |
1686 |
else if ((deadline += keepAlive) == 0L) |
1687 |
deadline = 1L; // not at head; restart timer |
1688 |
} |
1689 |
return 0; |
1690 |
} |
1691 |
|
1692 |
// Utilities used by ForkJoinTask |
1693 |
|
1694 |
/** |
1695 |
* Returns true if all workers are busy, possibly creating one if allowed |
1696 |
*/ |
1697 |
final boolean isSaturated() { |
1698 |
int par = mode & SMASK, maxTotal = bounds >>> SWIDTH; |
1699 |
for (long c;;) { |
1700 |
if (((int)(c = ctl) & ~UNSIGNALLED) != 0) |
1701 |
return false; |
1702 |
if ((short)(c >>> TC_SHIFT) >= maxTotal || par == 0) |
1703 |
return true; // cannot create |
1704 |
long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK); |
1705 |
if (compareAndSetCtl(c, nc)) |
1706 |
return !createWorker(); |
1707 |
} |
1708 |
} |
1709 |
|
1710 |
/** |
1711 |
* Returns true if can start terminating if enabled, or already terminated |
1712 |
*/ |
1713 |
final boolean canStop() { |
1714 |
outer: for (long oldSum = 0L;;) { // repeat until stable |
1715 |
int md; WorkQueue[] qs; long c; |
1716 |
if ((qs = queues) == null || ((md = mode) & STOP) != 0) |
1717 |
return true; |
1718 |
if ((md & SMASK) + (int)((c = ctl) >> RC_SHIFT) > 0) |
1719 |
break; |
1720 |
long checkSum = c; |
1721 |
for (int i = 1; i < qs.length; i += 2) { // scan submitters |
1722 |
WorkQueue q; ForkJoinTask<?>[] a; int s = 0, cap; |
1723 |
if ((q = qs[i]) != null && (a = q.array) != null && |
1724 |
(cap = a.length) > 0 && |
1725 |
((s = q.top) != q.base || a[(cap - 1) & s] != null || |
1726 |
q.source != 0)) |
1727 |
break outer; |
1728 |
checkSum += (((long)i) << 32) ^ s; |
1729 |
} |
1730 |
if (oldSum == (oldSum = checkSum) && queues == qs) |
1731 |
return true; |
1732 |
} |
1733 |
return (mode & STOP) != 0; // recheck mode on false return |
1734 |
} |
1735 |
|
1736 |
/** |
1737 |
* Tries to decrement counts (sometimes implicitly) and possibly |
1738 |
* arrange for a compensating worker in preparation for |
1739 |
* blocking. May fail due to interference, in which case -1 is |
1740 |
* returned so caller may retry. A zero return value indicates |
1741 |
* that the caller doesn't need to re-adjust counts when later |
1742 |
* unblocked. |
1743 |
* |
1744 |
* @param c incoming ctl value |
1745 |
* @return UNCOMPENSATE: block then adjust, 0: block, -1 : retry |
1746 |
*/ |
1747 |
private int tryCompensate(long c) { |
1748 |
Predicate<? super ForkJoinPool> sat; |
1749 |
int md = mode, b = bounds; |
1750 |
// counts are signed; centered at parallelism level == 0 |
1751 |
int minActive = (short)(b & SMASK), |
1752 |
maxTotal = b >>> SWIDTH, |
1753 |
active = (int)(c >> RC_SHIFT), |
1754 |
total = (short)(c >>> TC_SHIFT), |
1755 |
sp = (int)c & ~UNSIGNALLED; |
1756 |
if ((md & SMASK) == 0) |
1757 |
return 0; // cannot compensate if parallelism zero |
1758 |
else if (total >= 0) { |
1759 |
if (sp != 0) { // activate idle worker |
1760 |
WorkQueue[] qs; int n; WorkQueue v; |
1761 |
if ((qs = queues) != null && (n = qs.length) > 0 && |
1762 |
(v = qs[sp & (n - 1)]) != null) { |
1763 |
Thread vt = v.owner; |
1764 |
long nc = ((long)v.stackPred & SP_MASK) | (UC_MASK & c); |
1765 |
if (compareAndSetCtl(c, nc)) { |
1766 |
v.phase = sp; |
1767 |
LockSupport.unpark(vt); |
1768 |
return UNCOMPENSATE; |
1769 |
} |
1770 |
} |
1771 |
return -1; // retry |
1772 |
} |
1773 |
else if (active > minActive) { // reduce parallelism |
1774 |
long nc = ((RC_MASK & (c - RC_UNIT)) | (~RC_MASK & c)); |
1775 |
return compareAndSetCtl(c, nc) ? UNCOMPENSATE : -1; |
1776 |
} |
1777 |
} |
1778 |
if (total < maxTotal) { // expand pool |
1779 |
long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK); |
1780 |
return (!compareAndSetCtl(c, nc) ? -1 : |
1781 |
!createWorker() ? 0 : UNCOMPENSATE); |
1782 |
} |
1783 |
else if (!compareAndSetCtl(c, c)) // validate |
1784 |
return -1; |
1785 |
else if ((sat = saturate) != null && sat.test(this)) |
1786 |
return 0; |
1787 |
else |
1788 |
throw new RejectedExecutionException( |
1789 |
"Thread limit exceeded replacing blocked worker"); |
1790 |
} |
1791 |
|
1792 |
/** |
1793 |
* Readjusts RC count; called from ForkJoinTask after blocking. |
1794 |
*/ |
1795 |
final void uncompensate() { |
1796 |
getAndAddCtl(RC_UNIT); |
1797 |
} |
1798 |
|
1799 |
/** |
1800 |
* Helps if possible until the given task is done. Scans other |
1801 |
* queues for a task produced by one of w's stealers; returning |
1802 |
* compensated blocking sentinel if none are found. |
1803 |
* |
1804 |
* @param task the task |
1805 |
* @param w caller's WorkQueue |
1806 |
* @return task status on exit, or UNCOMPENSATE for compensated blocking |
1807 |
*/ |
1808 |
final int helpJoin(ForkJoinTask<?> task, WorkQueue w) { |
1809 |
int s = 0; |
1810 |
if (task != null && w != null) { |
1811 |
int wsrc = w.source, wid = w.config & SMASK, r = wid + 2; |
1812 |
boolean scan = true; |
1813 |
long c = 0L; // track ctl stability |
1814 |
outer: for (;;) { |
1815 |
if ((s = task.status) < 0) |
1816 |
break; |
1817 |
else if (scan = !scan) { // previous scan was empty |
1818 |
if (mode < 0) |
1819 |
ForkJoinTask.cancelIgnoringExceptions(task); |
1820 |
else if (c == (c = ctl) && (s = tryCompensate(c)) >= 0) |
1821 |
break; // block |
1822 |
} |
1823 |
else { // scan for subtasks |
1824 |
WorkQueue[] qs = queues; |
1825 |
int n = (qs == null) ? 0 : qs.length, m = n - 1; |
1826 |
for (int i = n; i > 0; i -= 2, r += 2) { |
1827 |
int j; WorkQueue q, x, y; ForkJoinTask<?>[] a; |
1828 |
if ((q = qs[j = r & m]) != null) { |
1829 |
int sq = q.source & SMASK, cap, b; |
1830 |
if ((a = q.array) != null && (cap = a.length) > 0) { |
1831 |
int k = (cap - 1) & (b = q.base); |
1832 |
int nextBase = b + 1, src = j | SRC, sx; |
1833 |
ForkJoinTask<?> t = WorkQueue.getSlot(a, k); |
1834 |
boolean eligible = sq == wid || |
1835 |
((x = qs[sq & m]) != null && // indirect |
1836 |
((sx = (x.source & SMASK)) == wid || |
1837 |
((y = qs[sx & m]) != null && // 2-indirect |
1838 |
(y.source & SMASK) == wid))); |
1839 |
if ((s = task.status) < 0) |
1840 |
break outer; |
1841 |
else if ((q.source & SMASK) != sq || |
1842 |
q.base != b) |
1843 |
scan = true; // inconsistent |
1844 |
else if (t == null) |
1845 |
scan |= (a[nextBase & (cap - 1)] != null || |
1846 |
q.top != b); // lagging |
1847 |
else if (eligible) { |
1848 |
if (WorkQueue.casSlotToNull(a, k, t)) { |
1849 |
q.base = nextBase; |
1850 |
w.source = src; |
1851 |
t.doExec(); |
1852 |
w.source = wsrc; |
1853 |
} |
1854 |
scan = true; |
1855 |
break; |
1856 |
} |
1857 |
} |
1858 |
} |
1859 |
} |
1860 |
} |
1861 |
} |
1862 |
} |
1863 |
return s; |
1864 |
} |
1865 |
|
1866 |
/** |
1867 |
* Extra helpJoin steps for CountedCompleters. Scans for and runs |
1868 |
* subtasks of the given root task, returning if none are found. |
1869 |
* |
1870 |
* @param task root of CountedCompleter computation |
1871 |
* @param w caller's WorkQueue |
1872 |
* @param owned true if owned by a ForkJoinWorkerThread |
1873 |
* @return task status on exit |
1874 |
*/ |
1875 |
final int helpComplete(ForkJoinTask<?> task, WorkQueue w, boolean owned) { |
1876 |
int s = 0; |
1877 |
if (task != null && w != null) { |
1878 |
int r = w.config; |
1879 |
boolean scan = true, locals = true; |
1880 |
long c = 0L; |
1881 |
outer: for (;;) { |
1882 |
if (locals) { // try locals before scanning |
1883 |
if ((s = w.helpComplete(task, owned, 0)) < 0) |
1884 |
break; |
1885 |
locals = false; |
1886 |
} |
1887 |
else if ((s = task.status) < 0) |
1888 |
break; |
1889 |
else if (scan = !scan) { |
1890 |
if (c == (c = ctl)) |
1891 |
break; |
1892 |
} |
1893 |
else { // scan for subtasks |
1894 |
WorkQueue[] qs = queues; |
1895 |
int n = (qs == null) ? 0 : qs.length; |
1896 |
for (int i = n; i > 0; --i, ++r) { |
1897 |
int j, cap, b; WorkQueue q; ForkJoinTask<?>[] a; |
1898 |
boolean eligible = false; |
1899 |
if ((q = qs[j = r & (n - 1)]) != null && |
1900 |
(a = q.array) != null && (cap = a.length) > 0) { |
1901 |
int k = (cap - 1) & (b = q.base), nextBase = b + 1; |
1902 |
ForkJoinTask<?> t = WorkQueue.getSlot(a, k); |
1903 |
if (t instanceof CountedCompleter) { |
1904 |
CountedCompleter<?> f = (CountedCompleter<?>)t; |
1905 |
do {} while (!(eligible = (f == task)) && |
1906 |
(f = f.completer) != null); |
1907 |
} |
1908 |
if ((s = task.status) < 0) |
1909 |
break outer; |
1910 |
else if (q.base != b) |
1911 |
scan = true; // inconsistent |
1912 |
else if (t == null) |
1913 |
scan |= (a[nextBase & (cap - 1)] != null || |
1914 |
q.top != b); |
1915 |
else if (eligible) { |
1916 |
if (WorkQueue.casSlotToNull(a, k, t)) { |
1917 |
q.setBaseOpaque(nextBase); |
1918 |
t.doExec(); |
1919 |
locals = true; |
1920 |
} |
1921 |
scan = true; |
1922 |
break; |
1923 |
} |
1924 |
} |
1925 |
} |
1926 |
} |
1927 |
} |
1928 |
} |
1929 |
return s; |
1930 |
} |
1931 |
|
1932 |
/** |
1933 |
* Scans for and returns a polled task, if available. Used only |
1934 |
* for untracked polls. Begins scan at an index (scanRover) |
1935 |
* advanced on each call, to avoid systematic unfairness. |
1936 |
* |
1937 |
* @param submissionsOnly if true, only scan submission queues |
1938 |
*/ |
1939 |
private ForkJoinTask<?> pollScan(boolean submissionsOnly) { |
1940 |
VarHandle.acquireFence(); |
1941 |
int r = scanRover += 0x61c88647; // Weyl increment; raciness OK |
1942 |
if (submissionsOnly) // even indices only |
1943 |
r &= ~1; |
1944 |
int step = (submissionsOnly) ? 2 : 1; |
1945 |
WorkQueue[] qs; int n; |
1946 |
while ((qs = queues) != null && (n = qs.length) > 0) { |
1947 |
boolean scan = false; |
1948 |
for (int i = 0; i < n; i += step) { |
1949 |
int j, cap, b; WorkQueue q; ForkJoinTask<?>[] a; |
1950 |
if ((q = qs[j = (n - 1) & (r + i)]) != null && |
1951 |
(a = q.array) != null && (cap = a.length) > 0) { |
1952 |
int k = (cap - 1) & (b = q.base), nextBase = b + 1; |
1953 |
ForkJoinTask<?> t = WorkQueue.getSlot(a, k); |
1954 |
if (q.base != b) |
1955 |
scan = true; |
1956 |
else if (t == null) |
1957 |
scan |= (q.top != b || a[nextBase & (cap - 1)] != null); |
1958 |
else if (!WorkQueue.casSlotToNull(a, k, t)) |
1959 |
scan = true; |
1960 |
else { |
1961 |
q.setBaseOpaque(nextBase); |
1962 |
return t; |
1963 |
} |
1964 |
} |
1965 |
} |
1966 |
if (!scan && queues == qs) |
1967 |
break; |
1968 |
} |
1969 |
return null; |
1970 |
} |
1971 |
|
1972 |
/** |
1973 |
* Runs tasks until {@code isQuiescent()}. Rather than blocking |
1974 |
* when tasks cannot be found, rescans until all others cannot |
1975 |
* find tasks either. |
1976 |
* |
1977 |
* @param nanos max wait time (Long.MAX_VALUE if effectively untimed) |
1978 |
* @param interruptible true if return on interrupt |
1979 |
* @return positive if quiescent, negative if interrupted, else 0 |
1980 |
*/ |
1981 |
final int helpQuiescePool(WorkQueue w, long nanos, boolean interruptible) { |
1982 |
if (w == null) |
1983 |
return 0; |
1984 |
long startTime = System.nanoTime(), parkTime = 0L; |
1985 |
int prevSrc = w.source, wsrc = prevSrc, cfg = w.config, r = cfg + 1; |
1986 |
for (boolean active = true, locals = true;;) { |
1987 |
boolean busy = false, scan = false; |
1988 |
if (locals) { // run local tasks before (re)polling |
1989 |
locals = false; |
1990 |
for (ForkJoinTask<?> u; (u = w.nextLocalTask(cfg)) != null;) |
1991 |
u.doExec(); |
1992 |
} |
1993 |
WorkQueue[] qs = queues; |
1994 |
int n = (qs == null) ? 0 : qs.length; |
1995 |
for (int i = n; i > 0; --i, ++r) { |
1996 |
int j, b, cap; WorkQueue q; ForkJoinTask<?>[] a; |
1997 |
if ((q = qs[j = (n - 1) & r]) != null && q != w && |
1998 |
(a = q.array) != null && (cap = a.length) > 0) { |
1999 |
int k = (cap - 1) & (b = q.base); |
2000 |
int nextBase = b + 1, src = j | SRC; |
2001 |
ForkJoinTask<?> t = WorkQueue.getSlot(a, k); |
2002 |
if (q.base != b) |
2003 |
busy = scan = true; |
2004 |
else if (t != null) { |
2005 |
busy = scan = true; |
2006 |
if (!active) { // increment before taking |
2007 |
active = true; |
2008 |
getAndAddCtl(RC_UNIT); |
2009 |
} |
2010 |
if (WorkQueue.casSlotToNull(a, k, t)) { |
2011 |
q.base = nextBase; |
2012 |
w.source = src; |
2013 |
t.doExec(); |
2014 |
w.source = wsrc = prevSrc; |
2015 |
locals = true; |
2016 |
} |
2017 |
break; |
2018 |
} |
2019 |
else if (!busy) { |
2020 |
if (q.top != b || a[nextBase & (cap - 1)] != null) |
2021 |
busy = scan = true; |
2022 |
else if (q.source != QUIET && q.phase >= 0) |
2023 |
busy = true; |
2024 |
} |
2025 |
} |
2026 |
} |
2027 |
VarHandle.acquireFence(); |
2028 |
if (!scan && queues == qs) { |
2029 |
boolean interrupted; |
2030 |
if (!busy) { |
2031 |
w.source = prevSrc; |
2032 |
if (!active) |
2033 |
getAndAddCtl(RC_UNIT); |
2034 |
return 1; |
2035 |
} |
2036 |
if (wsrc != QUIET) |
2037 |
w.source = wsrc = QUIET; |
2038 |
if (active) { // decrement |
2039 |
active = false; |
2040 |
parkTime = 0L; |
2041 |
getAndAddCtl(RC_MASK & -RC_UNIT); |
2042 |
} |
2043 |
else if (parkTime == 0L) { |
2044 |
parkTime = 1L << 10; // initially about 1 usec |
2045 |
Thread.yield(); |
2046 |
} |
2047 |
else if ((interrupted = interruptible && Thread.interrupted()) || |
2048 |
System.nanoTime() - startTime > nanos) { |
2049 |
getAndAddCtl(RC_UNIT); |
2050 |
return interrupted ? -1 : 0; |
2051 |
} |
2052 |
else { |
2053 |
LockSupport.parkNanos(this, parkTime); |
2054 |
if (parkTime < nanos >>> 8 && parkTime < 1L << 20) |
2055 |
parkTime <<= 1; // max sleep approx 1 sec or 1% nanos |
2056 |
} |
2057 |
} |
2058 |
} |
2059 |
} |
2060 |
|
2061 |
/** |
2062 |
* Helps quiesce from external caller until done, interrupted, or timeout |
2063 |
* |
2064 |
* @param nanos max wait time (Long.MAX_VALUE if effectively untimed) |
2065 |
* @param interruptible true if return on interrupt |
2066 |
* @return positive if quiescent, negative if interrupted, else 0 |
2067 |
*/ |
2068 |
final int externalHelpQuiescePool(long nanos, boolean interruptible) { |
2069 |
for (long startTime = System.nanoTime(), parkTime = 0L;;) { |
2070 |
ForkJoinTask<?> t; |
2071 |
if ((t = pollScan(false)) != null) { |
2072 |
t.doExec(); |
2073 |
parkTime = 0L; |
2074 |
} |
2075 |
else if (canStop()) |
2076 |
return 1; |
2077 |
else if (parkTime == 0L) { |
2078 |
parkTime = 1L << 10; |
2079 |
Thread.yield(); |
2080 |
} |
2081 |
else if ((System.nanoTime() - startTime) > nanos) |
2082 |
return 0; |
2083 |
else if (interruptible && Thread.interrupted()) |
2084 |
return -1; |
2085 |
else { |
2086 |
LockSupport.parkNanos(this, parkTime); |
2087 |
if (parkTime < nanos >>> 8 && parkTime < 1L << 20) |
2088 |
parkTime <<= 1; |
2089 |
} |
2090 |
} |
2091 |
} |
2092 |
|
2093 |
/** |
2094 |
* Gets and removes a local or stolen task for the given worker. |
2095 |
* |
2096 |
* @return a task, if available |
2097 |
*/ |
2098 |
final ForkJoinTask<?> nextTaskFor(WorkQueue w) { |
2099 |
ForkJoinTask<?> t; |
2100 |
if (w == null || (t = w.nextLocalTask(w.config)) == null) |
2101 |
t = pollScan(false); |
2102 |
return t; |
2103 |
} |
2104 |
|
2105 |
// External operations |
2106 |
|
2107 |
/** |
2108 |
* Finds and locks a WorkQueue for an external submitter, or |
2109 |
* returns null if shutdown or terminating. |
2110 |
*/ |
2111 |
final WorkQueue submissionQueue() { |
2112 |
int r; |
2113 |
if ((r = ThreadLocalRandom.getProbe()) == 0) { |
2114 |
ThreadLocalRandom.localInit(); // initialize caller's probe |
2115 |
r = ThreadLocalRandom.getProbe(); |
2116 |
} |
2117 |
for (int id = r << 1;;) { // even indices only |
2118 |
int md = mode, n, i; WorkQueue q; ReentrantLock lock; |
2119 |
WorkQueue[] qs = queues; |
2120 |
if ((md & SHUTDOWN) != 0 || qs == null || (n = qs.length) <= 0) |
2121 |
return null; |
2122 |
else if ((q = qs[i = (n - 1) & id]) == null) { |
2123 |
if ((lock = registrationLock) != null) { |
2124 |
WorkQueue w = new WorkQueue(id | SRC); |
2125 |
lock.lock(); // install under lock |
2126 |
if (qs[i] == null) |
2127 |
qs[i] = w; // else lost race; discard |
2128 |
lock.unlock(); |
2129 |
} |
2130 |
} |
2131 |
else if (!q.tryLock()) // move and restart |
2132 |
id = (r = ThreadLocalRandom.advanceProbe(r)) << 1; |
2133 |
else |
2134 |
return q; |
2135 |
} |
2136 |
} |
2137 |
|
2138 |
/** |
2139 |
* Adds the given task to an external submission queue, or throws |
2140 |
* exception if shutdown or terminating. |
2141 |
* |
2142 |
* @param task the task. Caller must ensure non-null. |
2143 |
*/ |
2144 |
final void externalPush(ForkJoinTask<?> task) { |
2145 |
WorkQueue q; |
2146 |
if ((q = submissionQueue()) == null) |
2147 |
throw new RejectedExecutionException(); // shutdown or disabled |
2148 |
else if (q.lockedPush(task)) |
2149 |
signalWork(); |
2150 |
} |
2151 |
|
2152 |
/** |
2153 |
* Pushes a possibly-external submission. |
2154 |
*/ |
2155 |
private <T> ForkJoinTask<T> externalSubmit(ForkJoinTask<T> task) { |
2156 |
Thread t; ForkJoinWorkerThread wt; WorkQueue q; |
2157 |
if (task == null) |
2158 |
throw new NullPointerException(); |
2159 |
if (((t = Thread.currentThread()) instanceof ForkJoinWorkerThread) && |
2160 |
(q = (wt = (ForkJoinWorkerThread)t).workQueue) != null && |
2161 |
wt.pool == this) |
2162 |
q.push(task, this); |
2163 |
else |
2164 |
externalPush(task); |
2165 |
return task; |
2166 |
} |
2167 |
|
2168 |
/** |
2169 |
* Returns common pool queue for an external thread that has |
2170 |
* possibly ever submitted a common pool task (nonzero probe), or |
2171 |
* null if none. |
2172 |
*/ |
2173 |
static WorkQueue commonQueue() { |
2174 |
ForkJoinPool p; WorkQueue[] qs; |
2175 |
int r = ThreadLocalRandom.getProbe(), n; |
2176 |
return ((p = common) != null && (qs = p.queues) != null && |
2177 |
(n = qs.length) > 0 && r != 0) ? |
2178 |
qs[(n - 1) & (r << 1)] : null; |
2179 |
} |
2180 |
|
2181 |
/** |
2182 |
* If the given executor is a ForkJoinPool, poll and execute |
2183 |
* AsynchronousCompletionTasks from worker's queue until none are |
2184 |
* available or blocker is released. |
2185 |
*/ |
2186 |
static void helpAsyncBlocker(Executor e, ManagedBlocker blocker) { |
2187 |
WorkQueue w = null; Thread t; ForkJoinWorkerThread wt; |
2188 |
if ((t = Thread.currentThread()) instanceof ForkJoinWorkerThread) { |
2189 |
if ((wt = (ForkJoinWorkerThread)t).pool == e) |
2190 |
w = wt.workQueue; |
2191 |
} |
2192 |
else if (e == common) |
2193 |
w = commonQueue(); |
2194 |
if (w != null) |
2195 |
w.helpAsyncBlocker(blocker); |
2196 |
} |
2197 |
|
2198 |
/** |
2199 |
* Returns a cheap heuristic guide for task partitioning when |
2200 |
* programmers, frameworks, tools, or languages have little or no |
2201 |
* idea about task granularity. In essence, by offering this |
2202 |
* method, we ask users only about tradeoffs in overhead vs |
2203 |
* expected throughput and its variance, rather than how finely to |
2204 |
* partition tasks. |
2205 |
* |
2206 |
* In a steady state strict (tree-structured) computation, each |
2207 |
* thread makes available for stealing enough tasks for other |
2208 |
* threads to remain active. Inductively, if all threads play by |
2209 |
* the same rules, each thread should make available only a |
2210 |
* constant number of tasks. |
2211 |
* |
2212 |
* The minimum useful constant is just 1. But using a value of 1 |
2213 |
* would require immediate replenishment upon each steal to |
2214 |
* maintain enough tasks, which is infeasible. Further, |
2215 |
* partitionings/granularities of offered tasks should minimize |
2216 |
* steal rates, which in general means that threads nearer the top |
2217 |
* of computation tree should generate more than those nearer the |
2218 |
* bottom. In perfect steady state, each thread is at |
2219 |
* approximately the same level of computation tree. However, |
2220 |
* producing extra tasks amortizes the uncertainty of progress and |
2221 |
* diffusion assumptions. |
2222 |
* |
2223 |
* So, users will want to use values larger (but not much larger) |
2224 |
* than 1 to both smooth over transient shortages and hedge |
2225 |
* against uneven progress; as traded off against the cost of |
2226 |
* extra task overhead. We leave the user to pick a threshold |
2227 |
* value to compare with the results of this call to guide |
2228 |
* decisions, but recommend values such as 3. |
2229 |
* |
2230 |
* When all threads are active, it is on average OK to estimate |
2231 |
* surplus strictly locally. In steady-state, if one thread is |
2232 |
* maintaining say 2 surplus tasks, then so are others. So we can |
2233 |
* just use estimated queue length. However, this strategy alone |
2234 |
* leads to serious mis-estimates in some non-steady-state |
2235 |
* conditions (ramp-up, ramp-down, other stalls). We can detect |
2236 |
* many of these by further considering the number of "idle" |
2237 |
* threads, that are known to have zero queued tasks, so |
2238 |
* compensate by a factor of (#idle/#active) threads. |
2239 |
*/ |
2240 |
static int getSurplusQueuedTaskCount() { |
2241 |
Thread t; ForkJoinWorkerThread wt; ForkJoinPool pool; WorkQueue q; |
2242 |
if (((t = Thread.currentThread()) instanceof ForkJoinWorkerThread) && |
2243 |
(pool = (wt = (ForkJoinWorkerThread)t).pool) != null && |
2244 |
(q = wt.workQueue) != null) { |
2245 |
int p = pool.mode & SMASK; |
2246 |
int a = p + (int)(pool.ctl >> RC_SHIFT); |
2247 |
int n = q.top - q.base; |
2248 |
return n - (a > (p >>>= 1) ? 0 : |
2249 |
a > (p >>>= 1) ? 1 : |
2250 |
a > (p >>>= 1) ? 2 : |
2251 |
a > (p >>>= 1) ? 4 : |
2252 |
8); |
2253 |
} |
2254 |
return 0; |
2255 |
} |
2256 |
|
2257 |
// Termination |
2258 |
|
2259 |
/** |
2260 |
* Possibly initiates and/or completes termination. |
2261 |
* |
2262 |
* @param now if true, unconditionally terminate, else only |
2263 |
* if no work and no active workers |
2264 |
* @param enable if true, terminate when next possible |
2265 |
* @return true if terminating or terminated |
2266 |
*/ |
2267 |
private boolean tryTerminate(boolean now, boolean enable) { |
2268 |
int md; // try to set SHUTDOWN, then STOP, then help terminate |
2269 |
if (((md = mode) & SHUTDOWN) == 0) { |
2270 |
if (!enable) |
2271 |
return false; |
2272 |
md = getAndBitwiseOrMode(SHUTDOWN); |
2273 |
} |
2274 |
if ((md & STOP) == 0) { |
2275 |
if (!now && !canStop()) |
2276 |
return false; |
2277 |
md = getAndBitwiseOrMode(STOP); |
2278 |
} |
2279 |
for (boolean rescan = true;;) { // repeat until no changes |
2280 |
boolean changed = false; |
2281 |
for (ForkJoinTask<?> t; (t = pollScan(false)) != null; ) { |
2282 |
changed = true; |
2283 |
ForkJoinTask.cancelIgnoringExceptions(t); // help cancel |
2284 |
} |
2285 |
WorkQueue[] qs; int n; WorkQueue q; Thread thread; |
2286 |
if ((qs = queues) != null && (n = qs.length) > 0) { |
2287 |
for (int j = 1; j < n; j += 2) { // unblock other workers |
2288 |
if ((q = qs[j]) != null && (thread = q.owner) != null && |
2289 |
!thread.isInterrupted()) { |
2290 |
changed = true; |
2291 |
try { |
2292 |
thread.interrupt(); |
2293 |
} catch (Throwable ignore) { |
2294 |
} |
2295 |
} |
2296 |
} |
2297 |
} |
2298 |
ReentrantLock lock; Condition cond; // signal when no workers |
2299 |
if (((md = mode) & TERMINATED) == 0 && |
2300 |
(md & SMASK) + (short)(ctl >>> TC_SHIFT) <= 0 && |
2301 |
(getAndBitwiseOrMode(TERMINATED) & TERMINATED) == 0 && |
2302 |
(lock = registrationLock) != null) { |
2303 |
lock.lock(); |
2304 |
if ((cond = termination) != null) |
2305 |
cond.signalAll(); |
2306 |
lock.unlock(); |
2307 |
} |
2308 |
if (changed) |
2309 |
rescan = true; |
2310 |
else if (rescan) |
2311 |
rescan = false; |
2312 |
else |
2313 |
break; |
2314 |
} |
2315 |
return true; |
2316 |
} |
2317 |
|
2318 |
// Exported methods |
2319 |
|
2320 |
// Constructors |
2321 |
|
2322 |
/** |
2323 |
* Creates a {@code ForkJoinPool} with parallelism equal to {@link |
2324 |
* java.lang.Runtime#availableProcessors}, using defaults for all |
2325 |
* other parameters (see {@link #ForkJoinPool(int, |
2326 |
* ForkJoinWorkerThreadFactory, UncaughtExceptionHandler, boolean, |
2327 |
* int, int, int, Predicate, long, TimeUnit)}). |
2328 |
* |
2329 |
* @throws SecurityException if a security manager exists and |
2330 |
* the caller is not permitted to modify threads |
2331 |
* because it does not hold {@link |
2332 |
* java.lang.RuntimePermission}{@code ("modifyThread")} |
2333 |
*/ |
2334 |
public ForkJoinPool() { |
2335 |
this(Math.min(MAX_CAP, Runtime.getRuntime().availableProcessors()), |
2336 |
defaultForkJoinWorkerThreadFactory, null, false, |
2337 |
0, MAX_CAP, 1, null, DEFAULT_KEEPALIVE, TimeUnit.MILLISECONDS); |
2338 |
} |
2339 |
|
2340 |
/** |
2341 |
* Creates a {@code ForkJoinPool} with the indicated parallelism |
2342 |
* level, using defaults for all other parameters (see {@link |
2343 |
* #ForkJoinPool(int, ForkJoinWorkerThreadFactory, |
2344 |
* UncaughtExceptionHandler, boolean, int, int, int, Predicate, |
2345 |
* long, TimeUnit)}). |
2346 |
* |
2347 |
* @param parallelism the parallelism level |
2348 |
* @throws IllegalArgumentException if parallelism less than or |
2349 |
* equal to zero, or greater than implementation limit |
2350 |
* @throws SecurityException if a security manager exists and |
2351 |
* the caller is not permitted to modify threads |
2352 |
* because it does not hold {@link |
2353 |
* java.lang.RuntimePermission}{@code ("modifyThread")} |
2354 |
*/ |
2355 |
public ForkJoinPool(int parallelism) { |
2356 |
this(parallelism, defaultForkJoinWorkerThreadFactory, null, false, |
2357 |
0, MAX_CAP, 1, null, DEFAULT_KEEPALIVE, TimeUnit.MILLISECONDS); |
2358 |
} |
2359 |
|
2360 |
/** |
2361 |
* Creates a {@code ForkJoinPool} with the given parameters (using |
2362 |
* defaults for others -- see {@link #ForkJoinPool(int, |
2363 |
* ForkJoinWorkerThreadFactory, UncaughtExceptionHandler, boolean, |
2364 |
* int, int, int, Predicate, long, TimeUnit)}). |
2365 |
* |
2366 |
* @param parallelism the parallelism level. For default value, |
2367 |
* use {@link java.lang.Runtime#availableProcessors}. |
2368 |
* @param factory the factory for creating new threads. For default value, |
2369 |
* use {@link #defaultForkJoinWorkerThreadFactory}. |
2370 |
* @param handler the handler for internal worker threads that |
2371 |
* terminate due to unrecoverable errors encountered while executing |
2372 |
* tasks. For default value, use {@code null}. |
2373 |
* @param asyncMode if true, |
2374 |
* establishes local first-in-first-out scheduling mode for forked |
2375 |
* tasks that are never joined. This mode may be more appropriate |
2376 |
* than default locally stack-based mode in applications in which |
2377 |
* worker threads only process event-style asynchronous tasks. |
2378 |
* For default value, use {@code false}. |
2379 |
* @throws IllegalArgumentException if parallelism less than or |
2380 |
* equal to zero, or greater than implementation limit |
2381 |
* @throws NullPointerException if the factory is null |
2382 |
* @throws SecurityException if a security manager exists and |
2383 |
* the caller is not permitted to modify threads |
2384 |
* because it does not hold {@link |
2385 |
* java.lang.RuntimePermission}{@code ("modifyThread")} |
2386 |
*/ |
2387 |
public ForkJoinPool(int parallelism, |
2388 |
ForkJoinWorkerThreadFactory factory, |
2389 |
UncaughtExceptionHandler handler, |
2390 |
boolean asyncMode) { |
2391 |
this(parallelism, factory, handler, asyncMode, |
2392 |
0, MAX_CAP, 1, null, DEFAULT_KEEPALIVE, TimeUnit.MILLISECONDS); |
2393 |
} |
2394 |
|
2395 |
/** |
2396 |
* Creates a {@code ForkJoinPool} with the given parameters. |
2397 |
* |
2398 |
* @param parallelism the parallelism level. For default value, |
2399 |
* use {@link java.lang.Runtime#availableProcessors}. |
2400 |
* |
2401 |
* @param factory the factory for creating new threads. For |
2402 |
* default value, use {@link #defaultForkJoinWorkerThreadFactory}. |
2403 |
* |
2404 |
* @param handler the handler for internal worker threads that |
2405 |
* terminate due to unrecoverable errors encountered while |
2406 |
* executing tasks. For default value, use {@code null}. |
2407 |
* |
2408 |
* @param asyncMode if true, establishes local first-in-first-out |
2409 |
* scheduling mode for forked tasks that are never joined. This |
2410 |
* mode may be more appropriate than default locally stack-based |
2411 |
* mode in applications in which worker threads only process |
2412 |
* event-style asynchronous tasks. For default value, use {@code |
2413 |
* false}. |
2414 |
* |
2415 |
* @param corePoolSize the number of threads to keep in the pool |
2416 |
* (unless timed out after an elapsed keep-alive). Normally (and |
2417 |
* by default) this is the same value as the parallelism level, |
2418 |
* but may be set to a larger value to reduce dynamic overhead if |
2419 |
* tasks regularly block. Using a smaller value (for example |
2420 |
* {@code 0}) has the same effect as the default. |
2421 |
* |
2422 |
* @param maximumPoolSize the maximum number of threads allowed. |
2423 |
* When the maximum is reached, attempts to replace blocked |
2424 |
* threads fail. (However, because creation and termination of |
2425 |
* different threads may overlap, and may be managed by the given |
2426 |
* thread factory, this value may be transiently exceeded.) To |
2427 |
* arrange the same value as is used by default for the common |
2428 |
* pool, use {@code 256} plus the {@code parallelism} level. (By |
2429 |
* default, the common pool allows a maximum of 256 spare |
2430 |
* threads.) Using a value (for example {@code |
2431 |
* Integer.MAX_VALUE}) larger than the implementation's total |
2432 |
* thread limit has the same effect as using this limit (which is |
2433 |
* the default). |
2434 |
* |
2435 |
* @param minimumRunnable the minimum allowed number of core |
2436 |
* threads not blocked by a join or {@link ManagedBlocker}. To |
2437 |
* ensure progress, when too few unblocked threads exist and |
2438 |
* unexecuted tasks may exist, new threads are constructed, up to |
2439 |
* the given maximumPoolSize. For the default value, use {@code |
2440 |
* 1}, that ensures liveness. A larger value might improve |
2441 |
* throughput in the presence of blocked activities, but might |
2442 |
* not, due to increased overhead. A value of zero may be |
2443 |
* acceptable when submitted tasks cannot have dependencies |
2444 |
* requiring additional threads. |
2445 |
* |
2446 |
* @param saturate if non-null, a predicate invoked upon attempts |
2447 |
* to create more than the maximum total allowed threads. By |
2448 |
* default, when a thread is about to block on a join or {@link |
2449 |
* ManagedBlocker}, but cannot be replaced because the |
2450 |
* maximumPoolSize would be exceeded, a {@link |
2451 |
* RejectedExecutionException} is thrown. But if this predicate |
2452 |
* returns {@code true}, then no exception is thrown, so the pool |
2453 |
* continues to operate with fewer than the target number of |
2454 |
* runnable threads, which might not ensure progress. |
2455 |
* |
2456 |
* @param keepAliveTime the elapsed time since last use before |
2457 |
* a thread is terminated (and then later replaced if needed). |
2458 |
* For the default value, use {@code 60, TimeUnit.SECONDS}. |
2459 |
* |
2460 |
* @param unit the time unit for the {@code keepAliveTime} argument |
2461 |
* |
2462 |
* @throws IllegalArgumentException if parallelism is less than or |
2463 |
* equal to zero, or is greater than implementation limit, |
2464 |
* or if maximumPoolSize is less than parallelism, |
2465 |
* of if the keepAliveTime is less than or equal to zero. |
2466 |
* @throws NullPointerException if the factory is null |
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 |
* @since 9 |
2472 |
*/ |
2473 |
public ForkJoinPool(int parallelism, |
2474 |
ForkJoinWorkerThreadFactory factory, |
2475 |
UncaughtExceptionHandler handler, |
2476 |
boolean asyncMode, |
2477 |
int corePoolSize, |
2478 |
int maximumPoolSize, |
2479 |
int minimumRunnable, |
2480 |
Predicate<? super ForkJoinPool> saturate, |
2481 |
long keepAliveTime, |
2482 |
TimeUnit unit) { |
2483 |
checkPermission(); |
2484 |
int p = parallelism; |
2485 |
if (p <= 0 || p > MAX_CAP || p > maximumPoolSize || keepAliveTime <= 0L) |
2486 |
throw new IllegalArgumentException(); |
2487 |
if (factory == null || unit == null) |
2488 |
throw new NullPointerException(); |
2489 |
this.factory = factory; |
2490 |
this.ueh = handler; |
2491 |
this.saturate = saturate; |
2492 |
this.keepAlive = Math.max(unit.toMillis(keepAliveTime), TIMEOUT_SLOP); |
2493 |
int size = 1 << (33 - Integer.numberOfLeadingZeros(p - 1)); |
2494 |
int corep = Math.min(Math.max(corePoolSize, p), MAX_CAP); |
2495 |
int maxSpares = Math.min(maximumPoolSize, MAX_CAP) - p; |
2496 |
int minAvail = Math.min(Math.max(minimumRunnable, 0), MAX_CAP); |
2497 |
this.bounds = ((minAvail - p) & SMASK) | (maxSpares << SWIDTH); |
2498 |
this.mode = p | (asyncMode ? FIFO : 0); |
2499 |
this.ctl = ((((long)(-corep) << TC_SHIFT) & TC_MASK) | |
2500 |
(((long)(-p) << RC_SHIFT) & RC_MASK)); |
2501 |
this.registrationLock = new ReentrantLock(); |
2502 |
this.queues = new WorkQueue[size]; |
2503 |
String pid = Integer.toString(getAndAddPoolIds(1) + 1); |
2504 |
this.workerNamePrefix = "ForkJoinPool-" + pid + "-worker-"; |
2505 |
} |
2506 |
|
2507 |
// helper method for commonPool constructor |
2508 |
private static Object newInstanceFromSystemProperty(String property) |
2509 |
throws ReflectiveOperationException { |
2510 |
String className = System.getProperty(property); |
2511 |
return (className == null) |
2512 |
? null |
2513 |
: ClassLoader.getSystemClassLoader().loadClass(className) |
2514 |
.getConstructor().newInstance(); |
2515 |
} |
2516 |
|
2517 |
/** |
2518 |
* Constructor for common pool using parameters possibly |
2519 |
* overridden by system properties |
2520 |
*/ |
2521 |
private ForkJoinPool(byte forCommonPoolOnly) { |
2522 |
int parallelism = Runtime.getRuntime().availableProcessors() - 1; |
2523 |
ForkJoinWorkerThreadFactory fac = null; |
2524 |
UncaughtExceptionHandler handler = null; |
2525 |
try { // ignore exceptions in accessing/parsing properties |
2526 |
fac = (ForkJoinWorkerThreadFactory) newInstanceFromSystemProperty( |
2527 |
"java.util.concurrent.ForkJoinPool.common.threadFactory"); |
2528 |
handler = (UncaughtExceptionHandler) newInstanceFromSystemProperty( |
2529 |
"java.util.concurrent.ForkJoinPool.common.exceptionHandler"); |
2530 |
String pp = System.getProperty |
2531 |
("java.util.concurrent.ForkJoinPool.common.parallelism"); |
2532 |
if (pp != null) |
2533 |
parallelism = Integer.parseInt(pp); |
2534 |
} catch (Exception ignore) { |
2535 |
} |
2536 |
int p = this.mode = Math.min(Math.max(parallelism, 0), MAX_CAP); |
2537 |
int maxSpares = (p == 0) ? 0 : COMMON_MAX_SPARES; |
2538 |
int bnds = ((1 - p) & SMASK) | (maxSpares << SWIDTH); |
2539 |
int size = 1 << (33 - Integer.numberOfLeadingZeros(p > 0 ? p - 1 : 1)); |
2540 |
this.factory = (fac != null) ? fac : |
2541 |
new DefaultCommonPoolForkJoinWorkerThreadFactory(); |
2542 |
this.ueh = handler; |
2543 |
this.keepAlive = DEFAULT_KEEPALIVE; |
2544 |
this.saturate = null; |
2545 |
this.workerNamePrefix = null; |
2546 |
this.bounds = bnds; |
2547 |
this.ctl = ((((long)(-p) << TC_SHIFT) & TC_MASK) | |
2548 |
(((long)(-p) << RC_SHIFT) & RC_MASK)); |
2549 |
this.queues = new WorkQueue[size]; |
2550 |
this.registrationLock = new ReentrantLock(); |
2551 |
} |
2552 |
|
2553 |
/** |
2554 |
* Returns the common pool instance. This pool is statically |
2555 |
* constructed; its run state is unaffected by attempts to {@link |
2556 |
* #shutdown} or {@link #shutdownNow}. However this pool and any |
2557 |
* ongoing processing are automatically terminated upon program |
2558 |
* {@link System#exit}. Any program that relies on asynchronous |
2559 |
* task processing to complete before program termination should |
2560 |
* invoke {@code commonPool().}{@link #awaitQuiescence awaitQuiescence}, |
2561 |
* before exit. |
2562 |
* |
2563 |
* @return the common pool instance |
2564 |
* @since 1.8 |
2565 |
*/ |
2566 |
public static ForkJoinPool commonPool() { |
2567 |
// assert common != null : "static init error"; |
2568 |
return common; |
2569 |
} |
2570 |
|
2571 |
// Execution methods |
2572 |
|
2573 |
/** |
2574 |
* Performs the given task, returning its result upon completion. |
2575 |
* If the computation encounters an unchecked Exception or Error, |
2576 |
* it is rethrown as the outcome of this invocation. Rethrown |
2577 |
* exceptions behave in the same way as regular exceptions, but, |
2578 |
* when possible, contain stack traces (as displayed for example |
2579 |
* using {@code ex.printStackTrace()}) of both the current thread |
2580 |
* as well as the thread actually encountering the exception; |
2581 |
* minimally only the latter. |
2582 |
* |
2583 |
* @param task the task |
2584 |
* @param <T> the type of the task's result |
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 |
externalSubmit(task); |
2592 |
return task.joinForPoolInvoke(this); |
2593 |
} |
2594 |
|
2595 |
/** |
2596 |
* Arranges for (asynchronous) execution of the given task. |
2597 |
* |
2598 |
* @param task the task |
2599 |
* @throws NullPointerException if the task is null |
2600 |
* @throws RejectedExecutionException if the task cannot be |
2601 |
* scheduled for execution |
2602 |
*/ |
2603 |
public void execute(ForkJoinTask<?> task) { |
2604 |
externalSubmit(task); |
2605 |
} |
2606 |
|
2607 |
// AbstractExecutorService methods |
2608 |
|
2609 |
/** |
2610 |
* @throws NullPointerException if the task is null |
2611 |
* @throws RejectedExecutionException if the task cannot be |
2612 |
* scheduled for execution |
2613 |
*/ |
2614 |
@Override |
2615 |
@SuppressWarnings("unchecked") |
2616 |
public void execute(Runnable task) { |
2617 |
externalSubmit((task instanceof ForkJoinTask<?>) |
2618 |
? (ForkJoinTask<Void>) task // avoid re-wrap |
2619 |
: new ForkJoinTask.RunnableExecuteAction(task)); |
2620 |
} |
2621 |
|
2622 |
/** |
2623 |
* Submits a ForkJoinTask for execution. |
2624 |
* |
2625 |
* @param task the task to submit |
2626 |
* @param <T> the type of the task's result |
2627 |
* @return the task |
2628 |
* @throws NullPointerException if the task is null |
2629 |
* @throws RejectedExecutionException if the task cannot be |
2630 |
* scheduled for execution |
2631 |
*/ |
2632 |
public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) { |
2633 |
return externalSubmit(task); |
2634 |
} |
2635 |
|
2636 |
/** |
2637 |
* @throws NullPointerException if the task is null |
2638 |
* @throws RejectedExecutionException if the task cannot be |
2639 |
* scheduled for execution |
2640 |
*/ |
2641 |
@Override |
2642 |
public <T> ForkJoinTask<T> submit(Callable<T> task) { |
2643 |
return externalSubmit(new ForkJoinTask.AdaptedCallable<T>(task)); |
2644 |
} |
2645 |
|
2646 |
/** |
2647 |
* @throws NullPointerException if the task is null |
2648 |
* @throws RejectedExecutionException if the task cannot be |
2649 |
* scheduled for execution |
2650 |
*/ |
2651 |
@Override |
2652 |
public <T> ForkJoinTask<T> submit(Runnable task, T result) { |
2653 |
return externalSubmit(new ForkJoinTask.AdaptedRunnable<T>(task, result)); |
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 |
@Override |
2662 |
@SuppressWarnings("unchecked") |
2663 |
public ForkJoinTask<?> submit(Runnable task) { |
2664 |
return externalSubmit((task instanceof ForkJoinTask<?>) |
2665 |
? (ForkJoinTask<Void>) task // avoid re-wrap |
2666 |
: new ForkJoinTask.AdaptedRunnableAction(task)); |
2667 |
} |
2668 |
|
2669 |
/** |
2670 |
* @throws NullPointerException {@inheritDoc} |
2671 |
* @throws RejectedExecutionException {@inheritDoc} |
2672 |
*/ |
2673 |
@Override |
2674 |
public <T> List<Future<T>> invokeAll(Collection<? extends Callable<T>> tasks) { |
2675 |
ArrayList<Future<T>> futures = new ArrayList<>(tasks.size()); |
2676 |
try { |
2677 |
for (Callable<T> t : tasks) { |
2678 |
ForkJoinTask<T> f = |
2679 |
new ForkJoinTask.AdaptedInterruptibleCallable<T>(t); |
2680 |
futures.add(f); |
2681 |
externalSubmit(f); |
2682 |
} |
2683 |
for (int i = futures.size() - 1; i >= 0; --i) |
2684 |
((ForkJoinTask<?>)futures.get(i)).tryJoinForPoolInvoke(this); |
2685 |
return futures; |
2686 |
} catch (Throwable t) { |
2687 |
for (Future<T> e : futures) |
2688 |
ForkJoinTask.cancelIgnoringExceptions(e); |
2689 |
throw t; |
2690 |
} |
2691 |
} |
2692 |
|
2693 |
@Override |
2694 |
public <T> List<Future<T>> invokeAll(Collection<? extends Callable<T>> tasks, |
2695 |
long timeout, TimeUnit unit) |
2696 |
throws InterruptedException { |
2697 |
long nanos = unit.toNanos(timeout); |
2698 |
ArrayList<Future<T>> futures = new ArrayList<>(tasks.size()); |
2699 |
try { |
2700 |
for (Callable<T> t : tasks) { |
2701 |
ForkJoinTask<T> f = |
2702 |
new ForkJoinTask.AdaptedInterruptibleCallable<T>(t); |
2703 |
futures.add(f); |
2704 |
externalSubmit(f); |
2705 |
} |
2706 |
long startTime = System.nanoTime(), ns = nanos; |
2707 |
boolean timedOut = (ns < 0L); |
2708 |
for (int i = futures.size() - 1; i >= 0; --i) { |
2709 |
Future<T> f = futures.get(i); |
2710 |
if (!f.isDone()) { |
2711 |
if (timedOut) |
2712 |
ForkJoinTask.cancelIgnoringExceptions(f); |
2713 |
else { |
2714 |
try { |
2715 |
((ForkJoinTask<T>)f).getForPoolInvoke(this, ns); |
2716 |
} catch (CancellationException | TimeoutException | |
2717 |
ExecutionException ok) { |
2718 |
} |
2719 |
if ((ns = nanos - (System.nanoTime() - startTime)) < 0L) |
2720 |
timedOut = true; |
2721 |
} |
2722 |
} |
2723 |
} |
2724 |
return futures; |
2725 |
} catch (Throwable t) { |
2726 |
for (Future<T> e : futures) |
2727 |
ForkJoinTask.cancelIgnoringExceptions(e); |
2728 |
throw t; |
2729 |
} |
2730 |
} |
2731 |
|
2732 |
// Task to hold results from InvokeAnyTasks |
2733 |
static final class InvokeAnyRoot<E> extends ForkJoinTask<E> { |
2734 |
private static final long serialVersionUID = 2838392045355241008L; |
2735 |
@SuppressWarnings("serial") // Conditionally serializable |
2736 |
volatile E result; |
2737 |
final AtomicInteger count; // in case all throw |
2738 |
final ForkJoinPool pool; // to check shutdown while collecting |
2739 |
InvokeAnyRoot(int n, ForkJoinPool p) { |
2740 |
pool = p; |
2741 |
count = new AtomicInteger(n); |
2742 |
} |
2743 |
final void tryComplete(Callable<E> c) { // called by InvokeAnyTasks |
2744 |
Throwable ex = null; |
2745 |
boolean failed; |
2746 |
if (c == null || Thread.interrupted() || |
2747 |
(pool != null && pool.mode < 0)) |
2748 |
failed = true; |
2749 |
else if (isDone()) |
2750 |
failed = false; |
2751 |
else { |
2752 |
try { |
2753 |
complete(c.call()); |
2754 |
failed = false; |
2755 |
} catch (Throwable tx) { |
2756 |
ex = tx; |
2757 |
failed = true; |
2758 |
} |
2759 |
} |
2760 |
if ((pool != null && pool.mode < 0) || |
2761 |
(failed && count.getAndDecrement() <= 1)) |
2762 |
trySetThrown(ex != null ? ex : new CancellationException()); |
2763 |
} |
2764 |
public final boolean exec() { return false; } // never forked |
2765 |
public final E getRawResult() { return result; } |
2766 |
public final void setRawResult(E v) { result = v; } |
2767 |
} |
2768 |
|
2769 |
// Variant of AdaptedInterruptibleCallable with results in InvokeAnyRoot |
2770 |
static final class InvokeAnyTask<E> extends ForkJoinTask<E> { |
2771 |
private static final long serialVersionUID = 2838392045355241008L; |
2772 |
final InvokeAnyRoot<E> root; |
2773 |
@SuppressWarnings("serial") // Conditionally serializable |
2774 |
final Callable<E> callable; |
2775 |
transient volatile Thread runner; |
2776 |
InvokeAnyTask(InvokeAnyRoot<E> root, Callable<E> callable) { |
2777 |
this.root = root; |
2778 |
this.callable = callable; |
2779 |
} |
2780 |
public final boolean exec() { |
2781 |
Thread.interrupted(); |
2782 |
runner = Thread.currentThread(); |
2783 |
root.tryComplete(callable); |
2784 |
runner = null; |
2785 |
Thread.interrupted(); |
2786 |
return true; |
2787 |
} |
2788 |
public final boolean cancel(boolean mayInterruptIfRunning) { |
2789 |
Thread t; |
2790 |
boolean stat = super.cancel(false); |
2791 |
if (mayInterruptIfRunning && (t = runner) != null) { |
2792 |
try { |
2793 |
t.interrupt(); |
2794 |
} catch (Throwable ignore) { |
2795 |
} |
2796 |
} |
2797 |
return stat; |
2798 |
} |
2799 |
public final void setRawResult(E v) {} // unused |
2800 |
public final E getRawResult() { return null; } |
2801 |
} |
2802 |
|
2803 |
@Override |
2804 |
public <T> T invokeAny(Collection<? extends Callable<T>> tasks) |
2805 |
throws InterruptedException, ExecutionException { |
2806 |
int n = tasks.size(); |
2807 |
if (n <= 0) |
2808 |
throw new IllegalArgumentException(); |
2809 |
InvokeAnyRoot<T> root = new InvokeAnyRoot<T>(n, this); |
2810 |
ArrayList<InvokeAnyTask<T>> fs = new ArrayList<>(n); |
2811 |
try { |
2812 |
for (Callable<T> c : tasks) { |
2813 |
if (c == null) |
2814 |
throw new NullPointerException(); |
2815 |
InvokeAnyTask<T> f = new InvokeAnyTask<T>(root, c); |
2816 |
fs.add(f); |
2817 |
externalSubmit(f); |
2818 |
if (root.isDone()) |
2819 |
break; |
2820 |
} |
2821 |
return root.getForPoolInvoke(this); |
2822 |
} finally { |
2823 |
for (InvokeAnyTask<T> f : fs) |
2824 |
ForkJoinTask.cancelIgnoringExceptions(f); |
2825 |
} |
2826 |
} |
2827 |
|
2828 |
@Override |
2829 |
public <T> T invokeAny(Collection<? extends Callable<T>> tasks, |
2830 |
long timeout, TimeUnit unit) |
2831 |
throws InterruptedException, ExecutionException, TimeoutException { |
2832 |
long nanos = unit.toNanos(timeout); |
2833 |
int n = tasks.size(); |
2834 |
if (n <= 0) |
2835 |
throw new IllegalArgumentException(); |
2836 |
InvokeAnyRoot<T> root = new InvokeAnyRoot<T>(n, this); |
2837 |
ArrayList<InvokeAnyTask<T>> fs = new ArrayList<>(n); |
2838 |
try { |
2839 |
for (Callable<T> c : tasks) { |
2840 |
if (c == null) |
2841 |
throw new NullPointerException(); |
2842 |
InvokeAnyTask<T> f = new InvokeAnyTask<T>(root, c); |
2843 |
fs.add(f); |
2844 |
externalSubmit(f); |
2845 |
if (root.isDone()) |
2846 |
break; |
2847 |
} |
2848 |
return root.getForPoolInvoke(this, nanos); |
2849 |
} finally { |
2850 |
for (InvokeAnyTask<T> f : fs) |
2851 |
ForkJoinTask.cancelIgnoringExceptions(f); |
2852 |
} |
2853 |
} |
2854 |
|
2855 |
/** |
2856 |
* Returns the factory used for constructing new workers. |
2857 |
* |
2858 |
* @return the factory used for constructing new workers |
2859 |
*/ |
2860 |
public ForkJoinWorkerThreadFactory getFactory() { |
2861 |
return factory; |
2862 |
} |
2863 |
|
2864 |
/** |
2865 |
* Returns the handler for internal worker threads that terminate |
2866 |
* due to unrecoverable errors encountered while executing tasks. |
2867 |
* |
2868 |
* @return the handler, or {@code null} if none |
2869 |
*/ |
2870 |
public UncaughtExceptionHandler getUncaughtExceptionHandler() { |
2871 |
return ueh; |
2872 |
} |
2873 |
|
2874 |
/** |
2875 |
* Returns the targeted parallelism level of this pool. |
2876 |
* |
2877 |
* @return the targeted parallelism level of this pool |
2878 |
*/ |
2879 |
public int getParallelism() { |
2880 |
int par = mode & SMASK; |
2881 |
return (par > 0) ? par : 1; |
2882 |
} |
2883 |
|
2884 |
/** |
2885 |
* Returns the targeted parallelism level of the common pool. |
2886 |
* |
2887 |
* @return the targeted parallelism level of the common pool |
2888 |
* @since 1.8 |
2889 |
*/ |
2890 |
public static int getCommonPoolParallelism() { |
2891 |
return COMMON_PARALLELISM; |
2892 |
} |
2893 |
|
2894 |
/** |
2895 |
* Returns the number of worker threads that have started but not |
2896 |
* yet terminated. The result returned by this method may differ |
2897 |
* from {@link #getParallelism} when threads are created to |
2898 |
* maintain parallelism when others are cooperatively blocked. |
2899 |
* |
2900 |
* @return the number of worker threads |
2901 |
*/ |
2902 |
public int getPoolSize() { |
2903 |
return ((mode & SMASK) + (short)(ctl >>> TC_SHIFT)); |
2904 |
} |
2905 |
|
2906 |
/** |
2907 |
* Returns {@code true} if this pool uses local first-in-first-out |
2908 |
* scheduling mode for forked tasks that are never joined. |
2909 |
* |
2910 |
* @return {@code true} if this pool uses async mode |
2911 |
*/ |
2912 |
public boolean getAsyncMode() { |
2913 |
return (mode & FIFO) != 0; |
2914 |
} |
2915 |
|
2916 |
/** |
2917 |
* Returns an estimate of the number of worker threads that are |
2918 |
* not blocked waiting to join tasks or for other managed |
2919 |
* synchronization. This method may overestimate the |
2920 |
* number of running threads. |
2921 |
* |
2922 |
* @return the number of worker threads |
2923 |
*/ |
2924 |
public int getRunningThreadCount() { |
2925 |
VarHandle.acquireFence(); |
2926 |
WorkQueue[] qs; WorkQueue q; |
2927 |
int rc = 0; |
2928 |
if ((qs = queues) != null) { |
2929 |
for (int i = 1; i < qs.length; i += 2) { |
2930 |
if ((q = qs[i]) != null && q.isApparentlyUnblocked()) |
2931 |
++rc; |
2932 |
} |
2933 |
} |
2934 |
return rc; |
2935 |
} |
2936 |
|
2937 |
/** |
2938 |
* Returns an estimate of the number of threads that are currently |
2939 |
* stealing or executing tasks. This method may overestimate the |
2940 |
* number of active threads. |
2941 |
* |
2942 |
* @return the number of active threads |
2943 |
*/ |
2944 |
public int getActiveThreadCount() { |
2945 |
int r = (mode & SMASK) + (int)(ctl >> RC_SHIFT); |
2946 |
return (r <= 0) ? 0 : r; // suppress momentarily negative values |
2947 |
} |
2948 |
|
2949 |
/** |
2950 |
* Returns {@code true} if all worker threads are currently idle. |
2951 |
* An idle worker is one that cannot obtain a task to execute |
2952 |
* because none are available to steal from other threads, and |
2953 |
* there are no pending submissions to the pool. This method is |
2954 |
* conservative; it might not return {@code true} immediately upon |
2955 |
* idleness of all threads, but will eventually become true if |
2956 |
* threads remain inactive. |
2957 |
* |
2958 |
* @return {@code true} if all threads are currently idle |
2959 |
*/ |
2960 |
public boolean isQuiescent() { |
2961 |
return canStop(); |
2962 |
} |
2963 |
|
2964 |
/** |
2965 |
* Returns an estimate of the total number of completed tasks that |
2966 |
* were executed by a thread other than their submitter. The |
2967 |
* reported value underestimates the actual total number of steals |
2968 |
* when the pool is not quiescent. This value may be useful for |
2969 |
* monitoring and tuning fork/join programs: in general, steal |
2970 |
* counts should be high enough to keep threads busy, but low |
2971 |
* enough to avoid overhead and contention across threads. |
2972 |
* |
2973 |
* @return the number of steals |
2974 |
*/ |
2975 |
public long getStealCount() { |
2976 |
long count = stealCount; |
2977 |
WorkQueue[] qs; WorkQueue q; |
2978 |
if ((qs = queues) != null) { |
2979 |
for (int i = 1; i < qs.length; i += 2) { |
2980 |
if ((q = qs[i]) != null) |
2981 |
count += (long)q.nsteals & 0xffffffffL; |
2982 |
} |
2983 |
} |
2984 |
return count; |
2985 |
} |
2986 |
|
2987 |
/** |
2988 |
* Returns an estimate of the total number of tasks currently held |
2989 |
* in queues by worker threads (but not including tasks submitted |
2990 |
* to the pool that have not begun executing). This value is only |
2991 |
* an approximation, obtained by iterating across all threads in |
2992 |
* the pool. This method may be useful for tuning task |
2993 |
* granularities. |
2994 |
* |
2995 |
* @return the number of queued tasks |
2996 |
*/ |
2997 |
public long getQueuedTaskCount() { |
2998 |
VarHandle.acquireFence(); |
2999 |
WorkQueue[] qs; WorkQueue q; |
3000 |
int count = 0; |
3001 |
if ((qs = queues) != null) { |
3002 |
for (int i = 1; i < qs.length; i += 2) { |
3003 |
if ((q = qs[i]) != null) |
3004 |
count += q.queueSize(); |
3005 |
} |
3006 |
} |
3007 |
return count; |
3008 |
} |
3009 |
|
3010 |
/** |
3011 |
* Returns an estimate of the number of tasks submitted to this |
3012 |
* pool that have not yet begun executing. This method may take |
3013 |
* time proportional to the number of submissions. |
3014 |
* |
3015 |
* @return the number of queued submissions |
3016 |
*/ |
3017 |
public int getQueuedSubmissionCount() { |
3018 |
VarHandle.acquireFence(); |
3019 |
WorkQueue[] qs; WorkQueue q; |
3020 |
int count = 0; |
3021 |
if ((qs = queues) != null) { |
3022 |
for (int i = 0; i < qs.length; i += 2) { |
3023 |
if ((q = qs[i]) != null) |
3024 |
count += q.queueSize(); |
3025 |
} |
3026 |
} |
3027 |
return count; |
3028 |
} |
3029 |
|
3030 |
/** |
3031 |
* Returns {@code true} if there are any tasks submitted to this |
3032 |
* pool that have not yet begun executing. |
3033 |
* |
3034 |
* @return {@code true} if there are any queued submissions |
3035 |
*/ |
3036 |
public boolean hasQueuedSubmissions() { |
3037 |
VarHandle.acquireFence(); |
3038 |
WorkQueue[] qs; WorkQueue q; |
3039 |
if ((qs = queues) != null) { |
3040 |
for (int i = 0; i < qs.length; i += 2) { |
3041 |
if ((q = qs[i]) != null && !q.isEmpty()) |
3042 |
return true; |
3043 |
} |
3044 |
} |
3045 |
return false; |
3046 |
} |
3047 |
|
3048 |
/** |
3049 |
* Removes and returns the next unexecuted submission if one is |
3050 |
* available. This method may be useful in extensions to this |
3051 |
* class that re-assign work in systems with multiple pools. |
3052 |
* |
3053 |
* @return the next submission, or {@code null} if none |
3054 |
*/ |
3055 |
protected ForkJoinTask<?> pollSubmission() { |
3056 |
return pollScan(true); |
3057 |
} |
3058 |
|
3059 |
/** |
3060 |
* Removes all available unexecuted submitted and forked tasks |
3061 |
* from scheduling queues and adds them to the given collection, |
3062 |
* without altering their execution status. These may include |
3063 |
* artificially generated or wrapped tasks. This method is |
3064 |
* designed to be invoked only when the pool is known to be |
3065 |
* quiescent. Invocations at other times may not remove all |
3066 |
* tasks. A failure encountered while attempting to add elements |
3067 |
* to collection {@code c} may result in elements being in |
3068 |
* neither, either or both collections when the associated |
3069 |
* exception is thrown. The behavior of this operation is |
3070 |
* undefined if the specified collection is modified while the |
3071 |
* operation is in progress. |
3072 |
* |
3073 |
* @param c the collection to transfer elements into |
3074 |
* @return the number of elements transferred |
3075 |
*/ |
3076 |
protected int drainTasksTo(Collection<? super ForkJoinTask<?>> c) { |
3077 |
int count = 0; |
3078 |
for (ForkJoinTask<?> t; (t = pollScan(false)) != null; ) { |
3079 |
c.add(t); |
3080 |
++count; |
3081 |
} |
3082 |
return count; |
3083 |
} |
3084 |
|
3085 |
/** |
3086 |
* Returns a string identifying this pool, as well as its state, |
3087 |
* including indications of run state, parallelism level, and |
3088 |
* worker and task counts. |
3089 |
* |
3090 |
* @return a string identifying this pool, as well as its state |
3091 |
*/ |
3092 |
public String toString() { |
3093 |
// Use a single pass through queues to collect counts |
3094 |
int md = mode; // read volatile fields first |
3095 |
long c = ctl; |
3096 |
long st = stealCount; |
3097 |
long qt = 0L, ss = 0L; int rc = 0; |
3098 |
WorkQueue[] qs; WorkQueue q; |
3099 |
if ((qs = queues) != null) { |
3100 |
for (int i = 0; i < qs.length; ++i) { |
3101 |
if ((q = qs[i]) != null) { |
3102 |
int size = q.queueSize(); |
3103 |
if ((i & 1) == 0) |
3104 |
ss += size; |
3105 |
else { |
3106 |
qt += size; |
3107 |
st += (long)q.nsteals & 0xffffffffL; |
3108 |
if (q.isApparentlyUnblocked()) |
3109 |
++rc; |
3110 |
} |
3111 |
} |
3112 |
} |
3113 |
} |
3114 |
|
3115 |
int pc = (md & SMASK); |
3116 |
int tc = pc + (short)(c >>> TC_SHIFT); |
3117 |
int ac = pc + (int)(c >> RC_SHIFT); |
3118 |
if (ac < 0) // ignore transient negative |
3119 |
ac = 0; |
3120 |
String level = ((md & TERMINATED) != 0 ? "Terminated" : |
3121 |
(md & STOP) != 0 ? "Terminating" : |
3122 |
(md & SHUTDOWN) != 0 ? "Shutting down" : |
3123 |
"Running"); |
3124 |
return super.toString() + |
3125 |
"[" + level + |
3126 |
", parallelism = " + pc + |
3127 |
", size = " + tc + |
3128 |
", active = " + ac + |
3129 |
", running = " + rc + |
3130 |
", steals = " + st + |
3131 |
", tasks = " + qt + |
3132 |
", submissions = " + ss + |
3133 |
"]"; |
3134 |
} |
3135 |
|
3136 |
/** |
3137 |
* Possibly initiates an orderly shutdown in which previously |
3138 |
* submitted tasks are executed, but no new tasks will be |
3139 |
* accepted. Invocation has no effect on execution state if this |
3140 |
* is the {@link #commonPool()}, and no additional effect if |
3141 |
* already shut down. Tasks that are in the process of being |
3142 |
* submitted concurrently during the course of this method may or |
3143 |
* may not be rejected. |
3144 |
* |
3145 |
* @throws SecurityException if a security manager exists and |
3146 |
* the caller is not permitted to modify threads |
3147 |
* because it does not hold {@link |
3148 |
* java.lang.RuntimePermission}{@code ("modifyThread")} |
3149 |
*/ |
3150 |
public void shutdown() { |
3151 |
checkPermission(); |
3152 |
if (this != common) |
3153 |
tryTerminate(false, true); |
3154 |
} |
3155 |
|
3156 |
/** |
3157 |
* Possibly attempts to cancel and/or stop all tasks, and reject |
3158 |
* all subsequently submitted tasks. Invocation has no effect on |
3159 |
* execution state if this is the {@link #commonPool()}, and no |
3160 |
* additional effect if already shut down. Otherwise, tasks that |
3161 |
* are in the process of being submitted or executed concurrently |
3162 |
* during the course of this method may or may not be |
3163 |
* rejected. This method cancels both existing and unexecuted |
3164 |
* tasks, in order to permit termination in the presence of task |
3165 |
* dependencies. So the method always returns an empty list |
3166 |
* (unlike the case for some other Executors). |
3167 |
* |
3168 |
* @return an empty list |
3169 |
* @throws SecurityException if a security manager exists and |
3170 |
* the caller is not permitted to modify threads |
3171 |
* because it does not hold {@link |
3172 |
* java.lang.RuntimePermission}{@code ("modifyThread")} |
3173 |
*/ |
3174 |
public List<Runnable> shutdownNow() { |
3175 |
checkPermission(); |
3176 |
if (this != common) |
3177 |
tryTerminate(true, true); |
3178 |
return Collections.emptyList(); |
3179 |
} |
3180 |
|
3181 |
/** |
3182 |
* Returns {@code true} if all tasks have completed following shut down. |
3183 |
* |
3184 |
* @return {@code true} if all tasks have completed following shut down |
3185 |
*/ |
3186 |
public boolean isTerminated() { |
3187 |
return (mode & TERMINATED) != 0; |
3188 |
} |
3189 |
|
3190 |
/** |
3191 |
* Returns {@code true} if the process of termination has |
3192 |
* commenced but not yet completed. This method may be useful for |
3193 |
* debugging. A return of {@code true} reported a sufficient |
3194 |
* period after shutdown may indicate that submitted tasks have |
3195 |
* ignored or suppressed interruption, or are waiting for I/O, |
3196 |
* causing this executor not to properly terminate. (See the |
3197 |
* advisory notes for class {@link ForkJoinTask} stating that |
3198 |
* tasks should not normally entail blocking operations. But if |
3199 |
* they do, they must abort them on interrupt.) |
3200 |
* |
3201 |
* @return {@code true} if terminating but not yet terminated |
3202 |
*/ |
3203 |
public boolean isTerminating() { |
3204 |
return (mode & (STOP | TERMINATED)) == STOP; |
3205 |
} |
3206 |
|
3207 |
/** |
3208 |
* Returns {@code true} if this pool has been shut down. |
3209 |
* |
3210 |
* @return {@code true} if this pool has been shut down |
3211 |
*/ |
3212 |
public boolean isShutdown() { |
3213 |
return (mode & SHUTDOWN) != 0; |
3214 |
} |
3215 |
|
3216 |
/** |
3217 |
* Blocks until all tasks have completed execution after a |
3218 |
* shutdown request, or the timeout occurs, or the current thread |
3219 |
* is interrupted, whichever happens first. Because the {@link |
3220 |
* #commonPool()} never terminates until program shutdown, when |
3221 |
* applied to the common pool, this method is equivalent to {@link |
3222 |
* #awaitQuiescence(long, TimeUnit)} but always returns {@code false}. |
3223 |
* |
3224 |
* @param timeout the maximum time to wait |
3225 |
* @param unit the time unit of the timeout argument |
3226 |
* @return {@code true} if this executor terminated and |
3227 |
* {@code false} if the timeout elapsed before termination |
3228 |
* @throws InterruptedException if interrupted while waiting |
3229 |
*/ |
3230 |
public boolean awaitTermination(long timeout, TimeUnit unit) |
3231 |
throws InterruptedException { |
3232 |
ReentrantLock lock; Condition cond; |
3233 |
long nanos = unit.toNanos(timeout); |
3234 |
boolean terminated = false; |
3235 |
if (this == common) { |
3236 |
Thread t; ForkJoinWorkerThread wt; int q; |
3237 |
if ((t = Thread.currentThread()) instanceof ForkJoinWorkerThread && |
3238 |
(wt = (ForkJoinWorkerThread)t).pool == this) |
3239 |
q = helpQuiescePool(wt.workQueue, nanos, true); |
3240 |
else |
3241 |
q = externalHelpQuiescePool(nanos, true); |
3242 |
if (q < 0) |
3243 |
throw new InterruptedException(); |
3244 |
} |
3245 |
else if (!(terminated = ((mode & TERMINATED) != 0)) && |
3246 |
(lock = registrationLock) != null) { |
3247 |
lock.lock(); |
3248 |
try { |
3249 |
if ((cond = termination) == null) |
3250 |
termination = cond = lock.newCondition(); |
3251 |
while (!(terminated = ((mode & TERMINATED) != 0)) && nanos > 0L) |
3252 |
nanos = cond.awaitNanos(nanos); |
3253 |
} finally { |
3254 |
lock.unlock(); |
3255 |
} |
3256 |
} |
3257 |
return terminated; |
3258 |
} |
3259 |
|
3260 |
/** |
3261 |
* If called by a ForkJoinTask operating in this pool, equivalent |
3262 |
* in effect to {@link ForkJoinTask#helpQuiesce}. Otherwise, |
3263 |
* waits and/or attempts to assist performing tasks until this |
3264 |
* pool {@link #isQuiescent} or the indicated timeout elapses. |
3265 |
* |
3266 |
* @param timeout the maximum time to wait |
3267 |
* @param unit the time unit of the timeout argument |
3268 |
* @return {@code true} if quiescent; {@code false} if the |
3269 |
* timeout elapsed. |
3270 |
*/ |
3271 |
public boolean awaitQuiescence(long timeout, TimeUnit unit) { |
3272 |
Thread t; ForkJoinWorkerThread wt; int q; |
3273 |
long nanos = unit.toNanos(timeout); |
3274 |
if ((t = Thread.currentThread()) instanceof ForkJoinWorkerThread && |
3275 |
(wt = (ForkJoinWorkerThread)t).pool == this) |
3276 |
q = helpQuiescePool(wt.workQueue, nanos, false); |
3277 |
else |
3278 |
q = externalHelpQuiescePool(nanos, false); |
3279 |
return (q > 0); |
3280 |
} |
3281 |
|
3282 |
/** |
3283 |
* Interface for extending managed parallelism for tasks running |
3284 |
* in {@link ForkJoinPool}s. |
3285 |
* |
3286 |
* <p>A {@code ManagedBlocker} provides two methods. Method |
3287 |
* {@link #isReleasable} must return {@code true} if blocking is |
3288 |
* not necessary. Method {@link #block} blocks the current thread |
3289 |
* if necessary (perhaps internally invoking {@code isReleasable} |
3290 |
* before actually blocking). These actions are performed by any |
3291 |
* thread invoking {@link |
3292 |
* ForkJoinPool#managedBlock(ManagedBlocker)}. The unusual |
3293 |
* methods in this API accommodate synchronizers that may, but |
3294 |
* don't usually, block for long periods. Similarly, they allow |
3295 |
* more efficient internal handling of cases in which additional |
3296 |
* workers may be, but usually are not, needed to ensure |
3297 |
* sufficient parallelism. Toward this end, implementations of |
3298 |
* method {@code isReleasable} must be amenable to repeated |
3299 |
* invocation. Neither method is invoked after a prior invocation |
3300 |
* of {@code isReleasable} or {@code block} returns {@code true}. |
3301 |
* |
3302 |
* <p>For example, here is a ManagedBlocker based on a |
3303 |
* ReentrantLock: |
3304 |
* <pre> {@code |
3305 |
* class ManagedLocker implements ManagedBlocker { |
3306 |
* final ReentrantLock lock; |
3307 |
* boolean hasLock = false; |
3308 |
* ManagedLocker(ReentrantLock lock) { this.lock = lock; } |
3309 |
* public boolean block() { |
3310 |
* if (!hasLock) |
3311 |
* lock.lock(); |
3312 |
* return true; |
3313 |
* } |
3314 |
* public boolean isReleasable() { |
3315 |
* return hasLock || (hasLock = lock.tryLock()); |
3316 |
* } |
3317 |
* }}</pre> |
3318 |
* |
3319 |
* <p>Here is a class that possibly blocks waiting for an |
3320 |
* item on a given queue: |
3321 |
* <pre> {@code |
3322 |
* class QueueTaker<E> implements ManagedBlocker { |
3323 |
* final BlockingQueue<E> queue; |
3324 |
* volatile E item = null; |
3325 |
* QueueTaker(BlockingQueue<E> q) { this.queue = q; } |
3326 |
* public boolean block() throws InterruptedException { |
3327 |
* if (item == null) |
3328 |
* item = queue.take(); |
3329 |
* return true; |
3330 |
* } |
3331 |
* public boolean isReleasable() { |
3332 |
* return item != null || (item = queue.poll()) != null; |
3333 |
* } |
3334 |
* public E getItem() { // call after pool.managedBlock completes |
3335 |
* return item; |
3336 |
* } |
3337 |
* }}</pre> |
3338 |
*/ |
3339 |
public static interface ManagedBlocker { |
3340 |
/** |
3341 |
* Possibly blocks the current thread, for example waiting for |
3342 |
* a lock or condition. |
3343 |
* |
3344 |
* @return {@code true} if no additional blocking is necessary |
3345 |
* (i.e., if isReleasable would return true) |
3346 |
* @throws InterruptedException if interrupted while waiting |
3347 |
* (the method is not required to do so, but is allowed to) |
3348 |
*/ |
3349 |
boolean block() throws InterruptedException; |
3350 |
|
3351 |
/** |
3352 |
* Returns {@code true} if blocking is unnecessary. |
3353 |
* @return {@code true} if blocking is unnecessary |
3354 |
*/ |
3355 |
boolean isReleasable(); |
3356 |
} |
3357 |
|
3358 |
/** |
3359 |
* Runs the given possibly blocking task. When {@linkplain |
3360 |
* ForkJoinTask#inForkJoinPool() running in a ForkJoinPool}, this |
3361 |
* method possibly arranges for a spare thread to be activated if |
3362 |
* necessary to ensure sufficient parallelism while the current |
3363 |
* thread is blocked in {@link ManagedBlocker#block blocker.block()}. |
3364 |
* |
3365 |
* <p>This method repeatedly calls {@code blocker.isReleasable()} and |
3366 |
* {@code blocker.block()} until either method returns {@code true}. |
3367 |
* Every call to {@code blocker.block()} is preceded by a call to |
3368 |
* {@code blocker.isReleasable()} that returned {@code false}. |
3369 |
* |
3370 |
* <p>If not running in a ForkJoinPool, this method is |
3371 |
* behaviorally equivalent to |
3372 |
* <pre> {@code |
3373 |
* while (!blocker.isReleasable()) |
3374 |
* if (blocker.block()) |
3375 |
* break;}</pre> |
3376 |
* |
3377 |
* If running in a ForkJoinPool, the pool may first be expanded to |
3378 |
* ensure sufficient parallelism available during the call to |
3379 |
* {@code blocker.block()}. |
3380 |
* |
3381 |
* @param blocker the blocker task |
3382 |
* @throws InterruptedException if {@code blocker.block()} did so |
3383 |
*/ |
3384 |
public static void managedBlock(ManagedBlocker blocker) |
3385 |
throws InterruptedException { |
3386 |
Thread t; ForkJoinPool p; |
3387 |
if ((t = Thread.currentThread()) instanceof ForkJoinWorkerThread && |
3388 |
(p = ((ForkJoinWorkerThread)t).pool) != null) |
3389 |
p.compensatedBlock(blocker); |
3390 |
else |
3391 |
unmanagedBlock(blocker); |
3392 |
} |
3393 |
|
3394 |
/** ManagedBlock for ForkJoinWorkerThreads */ |
3395 |
private void compensatedBlock(ManagedBlocker blocker) |
3396 |
throws InterruptedException { |
3397 |
if (blocker == null) throw new NullPointerException(); |
3398 |
for (;;) { |
3399 |
int comp; boolean done; |
3400 |
long c = ctl; |
3401 |
if (blocker.isReleasable()) |
3402 |
break; |
3403 |
if ((comp = tryCompensate(c)) >= 0) { |
3404 |
long post = (comp == 0) ? 0L : RC_UNIT; |
3405 |
try { |
3406 |
done = blocker.block(); |
3407 |
} finally { |
3408 |
getAndAddCtl(post); |
3409 |
} |
3410 |
if (done) |
3411 |
break; |
3412 |
} |
3413 |
} |
3414 |
} |
3415 |
|
3416 |
/** ManagedBlock for external threads */ |
3417 |
private static void unmanagedBlock(ManagedBlocker blocker) |
3418 |
throws InterruptedException { |
3419 |
if (blocker == null) throw new NullPointerException(); |
3420 |
do {} while (!blocker.isReleasable() && !blocker.block()); |
3421 |
} |
3422 |
|
3423 |
// AbstractExecutorService.newTaskFor overrides rely on |
3424 |
// undocumented fact that ForkJoinTask.adapt returns ForkJoinTasks |
3425 |
// that also implement RunnableFuture. |
3426 |
|
3427 |
@Override |
3428 |
protected <T> RunnableFuture<T> newTaskFor(Runnable runnable, T value) { |
3429 |
return new ForkJoinTask.AdaptedRunnable<T>(runnable, value); |
3430 |
} |
3431 |
|
3432 |
@Override |
3433 |
protected <T> RunnableFuture<T> newTaskFor(Callable<T> callable) { |
3434 |
return new ForkJoinTask.AdaptedCallable<T>(callable); |
3435 |
} |
3436 |
|
3437 |
static { |
3438 |
try { |
3439 |
MethodHandles.Lookup l = MethodHandles.lookup(); |
3440 |
CTL = l.findVarHandle(ForkJoinPool.class, "ctl", long.class); |
3441 |
MODE = l.findVarHandle(ForkJoinPool.class, "mode", int.class); |
3442 |
THREADIDS = l.findVarHandle(ForkJoinPool.class, "threadIds", int.class); |
3443 |
POOLIDS = l.findStaticVarHandle(ForkJoinPool.class, "poolIds", int.class); |
3444 |
} catch (ReflectiveOperationException e) { |
3445 |
throw new ExceptionInInitializerError(e); |
3446 |
} |
3447 |
|
3448 |
// Reduce the risk of rare disastrous classloading in first call to |
3449 |
// LockSupport.park: https://bugs.openjdk.java.net/browse/JDK-8074773 |
3450 |
Class<?> ensureLoaded = LockSupport.class; |
3451 |
|
3452 |
int commonMaxSpares = DEFAULT_COMMON_MAX_SPARES; |
3453 |
try { |
3454 |
String p = System.getProperty |
3455 |
("java.util.concurrent.ForkJoinPool.common.maximumSpares"); |
3456 |
if (p != null) |
3457 |
commonMaxSpares = Integer.parseInt(p); |
3458 |
} catch (Exception ignore) {} |
3459 |
COMMON_MAX_SPARES = commonMaxSpares; |
3460 |
|
3461 |
defaultForkJoinWorkerThreadFactory = |
3462 |
new DefaultForkJoinWorkerThreadFactory(); |
3463 |
modifyThreadPermission = new RuntimePermission("modifyThread"); |
3464 |
common = AccessController.doPrivileged(new PrivilegedAction<>() { |
3465 |
public ForkJoinPool run() { |
3466 |
return new ForkJoinPool((byte)0); }}); |
3467 |
|
3468 |
COMMON_PARALLELISM = Math.max(common.mode & SMASK, 1); |
3469 |
} |
3470 |
} |