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.security.AccessController; |
11 |
import java.security.AccessControlContext; |
12 |
import java.security.Permission; |
13 |
import java.security.Permissions; |
14 |
import java.security.PrivilegedAction; |
15 |
import java.security.ProtectionDomain; |
16 |
import java.util.ArrayList; |
17 |
import java.util.Collection; |
18 |
import java.util.Collections; |
19 |
import java.util.List; |
20 |
import java.util.function.Predicate; |
21 |
import java.util.concurrent.atomic.AtomicInteger; |
22 |
import java.util.concurrent.locks.LockSupport; |
23 |
import java.util.concurrent.locks.ReentrantLock; |
24 |
import java.util.concurrent.locks.Condition; |
25 |
import jdk.internal.misc.Unsafe; |
26 |
//import jdk.internal.vm.SharedThreadContainer; // for loom |
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, and are not guaranteed to preserve |
133 |
* the values of {@link java.lang.ThreadLocal} variables across tasks. |
134 |
* |
135 |
* Upon any error in establishing these settings, default parameters |
136 |
* are used. It is possible to disable or limit the use of threads in |
137 |
* the common pool by setting the parallelism property to zero, and/or |
138 |
* using a factory that may return {@code null}. However doing so may |
139 |
* cause unjoined tasks to never be executed. |
140 |
* |
141 |
* @implNote This implementation restricts the maximum number of |
142 |
* running threads to 32767. Attempts to create pools with greater |
143 |
* than the maximum number result in {@code |
144 |
* IllegalArgumentException}. Also, this implementation rejects |
145 |
* submitted tasks (that is, by throwing {@link |
146 |
* RejectedExecutionException}) only when the pool is shut down or |
147 |
* 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. There are a fair number of odd code constructions and |
177 |
* design decisions for components that reside at the edge of Java |
178 |
* vs JVM functionality. |
179 |
* |
180 |
* WorkQueues |
181 |
* ========== |
182 |
* |
183 |
* Most operations occur within work-stealing queues (in nested |
184 |
* class WorkQueue). These are special forms of Deques that |
185 |
* support only three of the four possible end-operations -- push, |
186 |
* pop, and poll (aka steal), under the further constraints that |
187 |
* push and pop are called only from the owning thread (or, as |
188 |
* extended here, under a lock), while poll may be called from |
189 |
* other threads. (If you are unfamiliar with them, you probably |
190 |
* want to read Herlihy and Shavit's book "The Art of |
191 |
* Multiprocessor programming", chapter 16 describing these in |
192 |
* more detail before proceeding.) The main work-stealing queue |
193 |
* design is roughly similar to those in the papers "Dynamic |
194 |
* Circular Work-Stealing Deque" by Chase and Lev, SPAA 2005 |
195 |
* (http://research.sun.com/scalable/pubs/index.html) and |
196 |
* "Idempotent work stealing" by Michael, Saraswat, and Vechev, |
197 |
* PPoPP 2009 (http://portal.acm.org/citation.cfm?id=1504186). |
198 |
* The main differences ultimately stem from GC requirements that |
199 |
* we null out taken slots as soon as we can, to maintain as small |
200 |
* a footprint as possible even in programs generating huge |
201 |
* numbers of tasks. To accomplish this, we shift the CAS |
202 |
* arbitrating pop vs poll (steal) from being on the indices |
203 |
* ("base" and "top") to the slots themselves. These provide the |
204 |
* primary required memory ordering -- see "Correct and Efficient |
205 |
* Work-Stealing for Weak Memory Models" by Le, Pop, Cohen, and |
206 |
* Nardelli, PPoPP 2013 |
207 |
* (http://www.di.ens.fr/~zappa/readings/ppopp13.pdf) for an |
208 |
* analysis of memory ordering requirements in work-stealing |
209 |
* algorithms similar to the one used here. We also use ordered, |
210 |
* moded accesses and/or fences for other control, with modes |
211 |
* reflecting the presence or absence of other contextual sync |
212 |
* provided by atomic and/or volatile accesses. Some methods (or |
213 |
* their primary loops) begin with an acquire fence or |
214 |
* otherwise-unnecessary volatile read that amounts to an |
215 |
* acquiring read of "this" to cover all fields (which is |
216 |
* sometimes stronger than necessary, but less brittle). Some |
217 |
* constructions are intentionally racy because they use read |
218 |
* values as hints, not for correctness. |
219 |
* |
220 |
* We also support a user mode in which local task processing is |
221 |
* in FIFO, not LIFO order, simply by using a local version of |
222 |
* poll rather than pop. This can be useful in message-passing |
223 |
* frameworks in which tasks are never joined, although with |
224 |
* increased contention among task producers and consumers. Also, |
225 |
* the same data structure (and class) is used for "submission |
226 |
* queues" (described below) holding externally submitted tasks, |
227 |
* that differ only in that a lock (field "access"; see below) is |
228 |
* required by external callers to push and pop tasks. |
229 |
* |
230 |
* Adding tasks then takes the form of a classic array push(task) |
231 |
* in a circular buffer: |
232 |
* q.array[q.top++ % length] = task; |
233 |
* |
234 |
* The actual code needs to null-check and size-check the array, |
235 |
* uses masking, not mod, for indexing a power-of-two-sized array, |
236 |
* enforces memory ordering, supports resizing, and possibly |
237 |
* signals waiting workers to start scanning (described below), |
238 |
* which requires even internal usages to strictly order accesses |
239 |
* (using a form of lock release). |
240 |
* |
241 |
* The pop operation (always performed by owner) is of the form: |
242 |
* if ((task = getAndSet(q.array, (q.top-1) % length, null)) != null) |
243 |
* decrement top and return task; |
244 |
* If this fails, the queue is empty. This operation is one part |
245 |
* of the nextLocalTask method, that instead does a local-poll |
246 |
* in FIFO mode. |
247 |
* |
248 |
* The poll operation is, basically: |
249 |
* if (CAS nonnull task t = q.array[k = q.base % length] to null) |
250 |
* increment base and return task; |
251 |
* |
252 |
* However, there are several more cases that must be dealt with. |
253 |
* Some of them are just due to asynchrony; others reflect |
254 |
* contention and stealing policies. Stepping through them |
255 |
* illustrates some of the implementation decisions in this class. |
256 |
* |
257 |
* * Slot k must be read with an acquiring read, which it must |
258 |
* anyway to dereference and run the task if the (acquiring) |
259 |
* CAS succeeds, but uses an explicit acquire fence to support |
260 |
* the following rechecks even if the CAS is not attempted. |
261 |
* |
262 |
* * q.base may change between reading and using its value to |
263 |
* index the slot. To avoid trying to use the wrong t, the |
264 |
* index and slot must be reread (not necessarily immediately) |
265 |
* until consistent, unless this is a local poll by owner, in |
266 |
* which case this form of inconsistency can only appear as t |
267 |
* being null, below. |
268 |
* |
269 |
* * Similarly, q.array may change (due to a resize), unless this |
270 |
* is a local poll by owner. Otherwise, when t is present, this |
271 |
* only needs consideration on CAS failure (since a CAS |
272 |
* confirms the non-resized case.) |
273 |
* |
274 |
* * t may appear null because a previous poll operation has not |
275 |
* yet incremented q.base, so the read is from an already-taken |
276 |
* index. This form of stall reflects the non-lock-freedom of |
277 |
* the poll operation. Stalls can be detected by observing that |
278 |
* q.base doesn't change on repeated reads of null t and when |
279 |
* no other alternatives apply, spin-wait for it to settle. To |
280 |
* reduce producing these kinds of stalls by other stealers, we |
281 |
* encourage timely writes to indices using store fences when |
282 |
* memory ordering is not already constrained by context. |
283 |
* |
284 |
* * The CAS may fail, in which case we may want to retry unless |
285 |
* there is too much contention. One goal is to balance and |
286 |
* spread out the many forms of contention that may be |
287 |
* encountered across polling and other operations to avoid |
288 |
* sustained performance degradations. Across all cases where |
289 |
* alternatives exist, a bounded number of CAS misses or stalls |
290 |
* are tolerated (for slots, ctl, and elsewhere described |
291 |
* below) before taking alternative action. These may move |
292 |
* contention or retries elsewhere, which is still preferable |
293 |
* to single-point bottlenecks. |
294 |
* |
295 |
* * Even though the check "top == base" is quiescently accurate |
296 |
* to determine whether a queue is empty, it is not of much use |
297 |
* when deciding whether to try to poll or repoll after a |
298 |
* failure. Both top and base may move independently, and both |
299 |
* lag updates to the underlying array. To reduce memory |
300 |
* contention, when possible, non-owners avoid reading the |
301 |
* "top" index at all, and instead use array reads, including |
302 |
* one-ahead reads to check whether to repoll, relying on the |
303 |
* fact that a non-empty queue does not have two null slots in |
304 |
* a row, except in cases (resizes and shifts) that can be |
305 |
* detected with a secondary recheck. |
306 |
* |
307 |
* The poll operations in q.poll(), scan(), helpJoin(), and |
308 |
* elsewhere differ with respect to whether other queues are |
309 |
* available to try, and the presence or nature of screening steps |
310 |
* when only some kinds of tasks can be taken. When alternatives |
311 |
* (or failing) is an option, they uniformly give up after |
312 |
* bounded numbers of stalls and/or CAS failures, which reduces |
313 |
* contention when too many workers are polling too few tasks. |
314 |
* Overall, in the aggregate, we ensure probabilistic |
315 |
* non-blockingness of work-stealing at least until checking |
316 |
* quiescence (which is intrinsically blocking): If an attempted |
317 |
* steal fails in these ways, a scanning thief chooses a different |
318 |
* target to try next. In contexts where alternatives aren't |
319 |
* available, and when progress conditions can be isolated to |
320 |
* values of a single variable, simple spinloops (using |
321 |
* Thread.onSpinWait) are used to reduce memory traffic. |
322 |
* |
323 |
* WorkQueues are also used in a similar way for tasks submitted |
324 |
* to the pool. We cannot mix these tasks in the same queues used |
325 |
* by workers. Instead, we randomly associate submission queues |
326 |
* with submitting threads, using a form of hashing. The |
327 |
* ThreadLocalRandom probe value serves as a hash code for |
328 |
* choosing existing queues, and may be randomly repositioned upon |
329 |
* contention with other submitters. In essence, submitters act |
330 |
* like workers except that they are restricted to executing local |
331 |
* tasks that they submitted (or when known, subtasks thereof). |
332 |
* Insertion of tasks in shared mode requires a lock. We use only |
333 |
* a simple spinlock because submitters encountering a busy queue |
334 |
* move to a different position to use or create other queues. |
335 |
* They (spin) block when registering new queues, and less |
336 |
* often in tryRemove and helpComplete. The lock needed for |
337 |
* external queues is generalized (as field "access") for |
338 |
* operations on owned queues that require a fully-fenced write |
339 |
* (including push, parking status, and termination) in order to |
340 |
* deal with Dekker-like signalling constructs described below. |
341 |
* |
342 |
* Management |
343 |
* ========== |
344 |
* |
345 |
* The main throughput advantages of work-stealing stem from |
346 |
* decentralized control -- workers mostly take tasks from |
347 |
* themselves or each other, at rates that can exceed a billion |
348 |
* per second. Most non-atomic control is performed by some form |
349 |
* of scanning across or within queues. The pool itself creates, |
350 |
* activates (enables scanning for and running tasks), |
351 |
* deactivates, blocks, and terminates threads, all with minimal |
352 |
* central information. There are only a few properties that we |
353 |
* can globally track or maintain, so we pack them into a small |
354 |
* number of variables, often maintaining atomicity without |
355 |
* blocking or locking. Nearly all essentially atomic control |
356 |
* state is held in a few variables that are by far most often |
357 |
* read (not written) as status and consistency checks. We pack as |
358 |
* much information into them as we can. |
359 |
* |
360 |
* Field "ctl" contains 64 bits holding information needed to |
361 |
* atomically decide to add, enqueue (on an event queue), and |
362 |
* dequeue and release workers. To enable this packing, we |
363 |
* restrict maximum parallelism to (1<<15)-1 (which is far in |
364 |
* excess of normal operating range) to allow ids, counts, and |
365 |
* their negations (used for thresholding) to fit into 16bit |
366 |
* subfields. Field "parallelism" holds the target parallelism |
367 |
* (normally corresponding to pool size). It is needed (nearly) |
368 |
* only in methods updating ctl, so is packed nearby. As of the |
369 |
* current release, users can dynamically reset target |
370 |
* parallelism, which is read once per update, so only slowly has |
371 |
* an effect in creating threads or letting them time out and |
372 |
* terminate when idle. |
373 |
* |
374 |
* Field "runState" holds lifetime status, atomically and |
375 |
* monotonically setting SHUTDOWN, STOP, and finally TERMINATED |
376 |
* bits. It is updated only via bitwise atomics (getAndBitwiseOr). |
377 |
* |
378 |
* Array "queues" holds references to WorkQueues. It is updated |
379 |
* (only during worker creation and termination) under the |
380 |
* registrationLock, but is otherwise concurrently readable (often |
381 |
* prefaced by a volatile read of mode to check termination, that |
382 |
* is required anyway, and serves as an acquire fence). To |
383 |
* simplify index-based operations, the array size is always a |
384 |
* power of two, and all readers must tolerate null slots. Worker |
385 |
* queues are at odd indices. Worker ids masked with SMASK match |
386 |
* their index. Shared (submission) queues are at even |
387 |
* indices. Grouping them together in this way simplifies and |
388 |
* speeds up task scanning. |
389 |
* |
390 |
* All worker thread creation is on-demand, triggered by task |
391 |
* submissions, replacement of terminated workers, and/or |
392 |
* compensation for blocked workers. However, all other support |
393 |
* code is set up to work with other policies. To ensure that we |
394 |
* do not hold on to worker or task references that would prevent |
395 |
* GC, all accesses to workQueues in waiting, signalling, and |
396 |
* control methods are via indices into the queues array (which is |
397 |
* one source of some of the messy code constructions here). In |
398 |
* essence, the queues array serves as a weak reference |
399 |
* mechanism. In particular, the stack top subfield of ctl stores |
400 |
* indices, not references. |
401 |
* |
402 |
* Queuing Idle Workers. Unlike HPC work-stealing frameworks, we |
403 |
* cannot let workers spin indefinitely scanning for tasks when |
404 |
* none can be found immediately, and we cannot start/resume |
405 |
* workers unless there appear to be tasks available. On the |
406 |
* other hand, we must quickly prod them into action when new |
407 |
* tasks are submitted or generated. These latencies are mainly a |
408 |
* function of JVM park/unpark (and underlying OS) performance, |
409 |
* which can be slow and variable. In many usages, ramp-up time |
410 |
* is the main limiting factor in overall performance, which is |
411 |
* compounded at program start-up by JIT compilation and |
412 |
* allocation. On the other hand, throughput degrades when too |
413 |
* many threads poll for too few tasks. |
414 |
* |
415 |
* The "ctl" field atomically maintains total and "released" |
416 |
* worker counts, plus the head of the available worker queue |
417 |
* (actually stack, represented by the lower 32bit subfield of |
418 |
* ctl). Released workers are those known to be scanning for |
419 |
* and/or running tasks. Unreleased ("available") workers are |
420 |
* recorded in the ctl stack. These workers are made eligible for |
421 |
* signalling by enqueuing in ctl (see method awaitWork). The |
422 |
* "queue" is a form of Treiber stack. This is ideal for |
423 |
* activating threads in most-recently used order, and improves |
424 |
* performance and locality, outweighing the disadvantages of |
425 |
* being prone to contention and inability to release a worker |
426 |
* unless it is topmost on stack. The top stack state holds the |
427 |
* value of the "phase" field of the worker: its index and status, |
428 |
* plus a version counter that, in addition to the count subfields |
429 |
* (also serving as version stamps) provide protection against |
430 |
* Treiber stack ABA effects. |
431 |
* |
432 |
* Creating workers. To create a worker, we pre-increment counts |
433 |
* (serving as a reservation), and attempt to construct a |
434 |
* ForkJoinWorkerThread via its factory. On starting, the new |
435 |
* thread first invokes registerWorker, where it constructs a |
436 |
* WorkQueue and is assigned an index in the queues array |
437 |
* (expanding the array if necessary). Upon any exception across |
438 |
* these steps, or null return from factory, deregisterWorker |
439 |
* adjusts counts and records accordingly. If a null return, the |
440 |
* pool continues running with fewer than the target number |
441 |
* workers. If exceptional, the exception is propagated, generally |
442 |
* to some external caller. |
443 |
* |
444 |
* WorkQueue field "phase" is used by both workers and the pool to |
445 |
* manage and track whether a worker is unsignalled (possibly |
446 |
* blocked waiting for a signal), conveniently using the sign bit |
447 |
* to check. When a worker is enqueued its phase field is set |
448 |
* negative. Note that phase field updates lag queue CAS releases; |
449 |
* seeing a negative phase does not guarantee that the worker is |
450 |
* available (and so is never checked in this way). When queued, |
451 |
* the lower 16 bits of its phase must hold its pool index. So we |
452 |
* place the index there upon initialization and never modify |
453 |
* these bits. |
454 |
* |
455 |
* The ctl field also serves as the basis for memory |
456 |
* synchronization surrounding activation. This uses a more |
457 |
* efficient version of a Dekker-like rule that task producers and |
458 |
* consumers sync with each other by both writing/CASing ctl (even |
459 |
* if to its current value). However, rather than CASing ctl to |
460 |
* its current value in the common case where no action is |
461 |
* required, we reduce write contention by ensuring that |
462 |
* signalWork invocations are prefaced with a fully fenced memory |
463 |
* access (which is usually needed anyway). |
464 |
* |
465 |
* Signalling. Signals (in signalWork) cause new or reactivated |
466 |
* workers to scan for tasks. Method signalWork and its callers |
467 |
* try to approximate the unattainable goal of having the right |
468 |
* number of workers activated for the tasks at hand, but must err |
469 |
* on the side of too many workers vs too few to avoid stalls. If |
470 |
* computations are purely tree structured, it suffices for every |
471 |
* worker to activate another when it pushes a task into an empty |
472 |
* queue, resulting in O(log(#threads)) steps to full activation. |
473 |
* (To reduce resource usages in some cases, at the expense of |
474 |
* slower startup in others, activation of an idle thread is |
475 |
* preferred over creating a new one, here and elsewhere.) If |
476 |
* instead, tasks come in serially from only a single producer, |
477 |
* each worker taking its first (since the last activation) task |
478 |
* from a queue should signal another if there are more tasks in |
479 |
* that queue. This is equivalent to, but generally faster than, |
480 |
* arranging the stealer take two tasks, re-pushing one on its own |
481 |
* queue, and signalling (because its queue is empty), also |
482 |
* resulting in logarithmic full activation time. Because we don't |
483 |
* know about usage patterns (or most commonly, mixtures), we use |
484 |
* both approaches. Together these are minimally necessary for |
485 |
* maintaining liveness. However, they do not account for the fact |
486 |
* that when tasks are short-lived, signals are unnecessary |
487 |
* because workers will already be scanning for new tasks without |
488 |
* the need of new signals. We track these cases (variable |
489 |
* "prevSrc" in scan() and related methods) to avoid some |
490 |
* unnecessary signals and scans. However, signal contention and |
491 |
* overhead effects may still occur during ramp-up, ramp-down, and |
492 |
* small computations involving only a few workers. |
493 |
* |
494 |
* Scanning. Method scan performs top-level scanning for (and |
495 |
* execution of) tasks by polling a pseudo-random permutation of |
496 |
* the array (by starting at a random index, and using a constant |
497 |
* cyclically exhaustive stride.) It uses the same basic polling |
498 |
* method as WorkQueue.poll(), but restarts with a different |
499 |
* permutation on each invocation. (Non-top-level scans; for |
500 |
* example in helpJoin, use simpler and faster linear probes |
501 |
* because they do not systematically contend with top-level |
502 |
* scans.) The pseudorandom generator need not have high-quality |
503 |
* statistical properties in the long term. We use Marsaglia |
504 |
* XorShifts, seeded with the Weyl sequence from ThreadLocalRandom |
505 |
* probes, which are cheap and suffice. Scans do not otherwise |
506 |
* explicitly take into account core affinities, loads, cache |
507 |
* localities, etc, However, they do exploit temporal locality |
508 |
* (which usually approximates these) by preferring to re-poll |
509 |
* from the same queue (using method tryPoll()) after a successful |
510 |
* poll before trying others (see method topLevelExec), which also |
511 |
* reduces bookkeeping and scanning overhead. This also reduces |
512 |
* fairness, which is partially counteracted by giving up on |
513 |
* contention. |
514 |
* |
515 |
* Deactivation. When method scan indicates that no tasks are |
516 |
* found by a worker, it deactivates (see awaitWork). Note that |
517 |
* not finding tasks doesn't mean that there won't soon be |
518 |
* some. Further, a scan may give up under contention, returning |
519 |
* even without knowing whether any tasks are still present, which |
520 |
* is OK, given the above signalling rules that will eventually |
521 |
* maintain progress. Blocking and unblocking via park/unpark can |
522 |
* cause serious slowdowns when tasks are rapidly but irregularly |
523 |
* generated (which is often due to garbage collectors and other |
524 |
* activities). One way to ameliorate is for workers to rescan |
525 |
* multiple times, even when there are unlikely to be tasks. But |
526 |
* this causes enough memory and CAS contention to prefer using |
527 |
* quieter spinwaits in awaitWork; currently set to small values |
528 |
* that only cover near-miss scenarios for deactivate vs activate |
529 |
* races. Because idle workers are often not yet blocked (via |
530 |
* LockSupport.park), we use the WorkQueue access field to |
531 |
* advertise that a waiter actually needs unparking upon signal. |
532 |
* |
533 |
* When idle workers are not continually woken up, the count |
534 |
* fields in ctl allow efficient and accurate discovery of |
535 |
* quiescent states (i.e., when all workers are idle) after |
536 |
* deactivation. However, this voting mechanism alone does not |
537 |
* guarantee that a pool can become dormant (quiesced or |
538 |
* terminated), because external racing producers do not vote, and |
539 |
* can asynchronously submit new tasks. To deal with this, the |
540 |
* final unparked thread (in awaitWork) scans external queues to |
541 |
* check for tasks that could have been added during a race window |
542 |
* that would not be accompanied by a signal, in which case |
543 |
* re-activating itself (or any other worker) to recheck. The same |
544 |
* sets of checks are used in tryTerminate, to correctly trigger |
545 |
* delayed termination (shutDown, followed by quiescence) in the |
546 |
* presence of racing submissions. In all cases, the notion of the |
547 |
* "final" unparked thread is an approximation, because new |
548 |
* workers could be in the process of being constructed, which |
549 |
* occasionally adds some extra unnecessary processing. |
550 |
* |
551 |
* Shutdown and Termination. A call to shutdownNow invokes |
552 |
* tryTerminate to atomically set a mode bit. The calling thread, |
553 |
* as well as every other worker thereafter terminating, helps |
554 |
* terminate others by cancelling their unprocessed tasks, and |
555 |
* interrupting other workers. Calls to non-abrupt shutdown() |
556 |
* preface this by checking isQuiescent before triggering the |
557 |
* "STOP" phase of termination. During termination, workers are |
558 |
* stopped using all three of (often in parallel): releasing via |
559 |
* ctl (method reactivate), interrupts, and cancelling tasks that |
560 |
* will cause workers to not find work and exit. To support this, |
561 |
* worker references not removed from the queues array during |
562 |
* termination. It is possible for late thread creations to still |
563 |
* be in progress after a quiescent termination reports terminated |
564 |
* status, but they will also immediately terminate. To conform to |
565 |
* ExecutorService invoke, invokeAll, and invokeAny specs, we must |
566 |
* track pool status while waiting in ForkJoinTask.awaitDone, and |
567 |
* interrupt interruptible callers on termination, while also |
568 |
* avoiding cancelling other tasks that are normally completing |
569 |
* during quiescent termination. This is tracked by recording |
570 |
* ForkJoinTask.POOLSUBMIT in task status and/or as a bit flag |
571 |
* argument to joining methods. |
572 |
* |
573 |
* Trimming workers. To release resources after periods of lack of |
574 |
* use, a worker starting to wait when the pool is quiescent will |
575 |
* time out and terminate if the pool has remained quiescent for |
576 |
* period given by field keepAlive. |
577 |
* |
578 |
* Joining Tasks |
579 |
* ============= |
580 |
* |
581 |
* Normally, the first option when joining a task that is not done |
582 |
* is to try to take it from local queue and run it. Otherwise, |
583 |
* any of several actions may be taken when one worker is waiting |
584 |
* to join a task stolen (or always held) by another. Because we |
585 |
* are multiplexing many tasks on to a pool of workers, we can't |
586 |
* always just let them block (as in Thread.join). We also cannot |
587 |
* just reassign the joiner's run-time stack with another and |
588 |
* replace it later, which would be a form of "continuation", that |
589 |
* even if possible is not necessarily a good idea since we may |
590 |
* need both an unblocked task and its continuation to progress. |
591 |
* Instead we combine two tactics: |
592 |
* |
593 |
* Helping: Arranging for the joiner to execute some task that it |
594 |
* could be running if the steal had not occurred. |
595 |
* |
596 |
* Compensating: Unless there are already enough live threads, |
597 |
* method tryCompensate() may create or re-activate a spare |
598 |
* thread to compensate for blocked joiners until they unblock. |
599 |
* |
600 |
* A third form (implemented via tryRemove) amounts to helping a |
601 |
* hypothetical compensator: If we can readily tell that a |
602 |
* possible action of a compensator is to steal and execute the |
603 |
* task being joined, the joining thread can do so directly, |
604 |
* without the need for a compensation thread; although with a |
605 |
* possibility of reduced parallelism because of a transient gap |
606 |
* in the queue array that stalls stealers. |
607 |
* |
608 |
* Other intermediate forms available for specific task types (for |
609 |
* example helpAsyncBlocker) often avoid or postpone the need for |
610 |
* blocking or compensation. |
611 |
* |
612 |
* The ManagedBlocker extension API can't use helping so relies |
613 |
* only on compensation in method awaitBlocker. |
614 |
* |
615 |
* The algorithm in helpJoin entails a form of "linear helping". |
616 |
* Each worker records (in field "source") a reference to the |
617 |
* queue from which it last stole a task. The scan in method |
618 |
* helpJoin uses these markers to try to find a worker to help |
619 |
* (i.e., steal back a task from and execute it) that could hasten |
620 |
* completion of the actively joined task. Thus, the joiner |
621 |
* executes a task that would be on its own local deque if the |
622 |
* to-be-joined task had not been stolen. This is a conservative |
623 |
* variant of the approach described in Wagner & Calder |
624 |
* "Leapfrogging: a portable technique for implementing efficient |
625 |
* futures" SIGPLAN Notices, 1993 |
626 |
* (http://portal.acm.org/citation.cfm?id=155354). It differs |
627 |
* mainly in that we only record queues, not full dependency |
628 |
* links. This requires a linear scan of the queues array to |
629 |
* locate stealers, but isolates cost to when it is needed, rather |
630 |
* than adding to per-task overhead. For CountedCompleters, the |
631 |
* analogous method helpComplete doesn't need stealer-tracking, |
632 |
* but requires a similar check of completion chains. |
633 |
* |
634 |
* In either case, searches can fail to locate stealers when |
635 |
* stalls delay recording sources. We avoid some of these cases by |
636 |
* using snapshotted values of ctl as a check that the numbers of |
637 |
* workers are not changing. But even when accurately identified, |
638 |
* stealers might not ever produce a task that the joiner can in |
639 |
* turn help with. So, compensation is tried upon failure to find |
640 |
* tasks to run. |
641 |
* |
642 |
* Compensation does not by default aim to keep exactly the target |
643 |
* parallelism number of unblocked threads running at any given |
644 |
* time. Some previous versions of this class employed immediate |
645 |
* compensations for any blocked join. However, in practice, the |
646 |
* vast majority of blockages are transient byproducts of GC and |
647 |
* other JVM or OS activities that are made worse by replacement |
648 |
* when they cause longer-term oversubscription. Rather than |
649 |
* impose arbitrary policies, we allow users to override the |
650 |
* default of only adding threads upon apparent starvation. The |
651 |
* compensation mechanism may also be bounded. Bounds for the |
652 |
* commonPool better enable JVMs to cope with programming errors |
653 |
* and abuse before running out of resources to do so. |
654 |
* |
655 |
* Common Pool |
656 |
* =========== |
657 |
* |
658 |
* The static common pool always exists after static |
659 |
* initialization. Since it (or any other created pool) need |
660 |
* never be used, we minimize initial construction overhead and |
661 |
* footprint to the setup of about a dozen fields, although with |
662 |
* some System property parsing and with security processing that |
663 |
* takes far longer than the actual construction when |
664 |
* SecurityManagers are used or properties are set. The common |
665 |
* pool is distinguished internally by having both a null |
666 |
* workerNamePrefix and ISCOMMON config bit set, along with |
667 |
* PRESET_SIZE set if parallelism was configured by system |
668 |
* property. |
669 |
* |
670 |
* When external threads use ForkJoinTask.fork for the common |
671 |
* pool, they can perform subtask processing (see helpComplete and |
672 |
* related methods) upon joins. This caller-helps policy makes it |
673 |
* sensible to set common pool parallelism level to one (or more) |
674 |
* less than the total number of available cores, or even zero for |
675 |
* pure caller-runs. For the sake of ExecutorService specs, we can |
676 |
* only do this for tasks entered via fork, not submit. We track |
677 |
* this using a task status bit (markPoolSubmission). In all |
678 |
* other cases, external threads waiting for joins first check the |
679 |
* common pool for their task, which fails quickly if the caller |
680 |
* did not fork to common pool. |
681 |
* |
682 |
* Guarantees for common pool parallelism zero are limited to |
683 |
* tasks that are joined by their callers in a tree-structured |
684 |
* fashion or use CountedCompleters (as is true for jdk |
685 |
* parallelStreams). Support infiltrates several methods, |
686 |
* including those that retry helping steps or spin until we are |
687 |
* sure that none apply if there are no workers. |
688 |
* |
689 |
* As a more appropriate default in managed environments, unless |
690 |
* overridden by system properties, we use workers of subclass |
691 |
* InnocuousForkJoinWorkerThread when there is a SecurityManager |
692 |
* present. These workers have no permissions set, do not belong |
693 |
* to any user-defined ThreadGroup, and clear all ThreadLocals |
694 |
* after executing any top-level task. The associated mechanics |
695 |
* may be JVM-dependent and must access particular Thread class |
696 |
* fields to achieve this effect. |
697 |
* |
698 |
* Interrupt handling |
699 |
* ================== |
700 |
* |
701 |
* The framework is designed to manage task cancellation |
702 |
* (ForkJoinTask.cancel) independently from the interrupt status |
703 |
* of threads running tasks. (See the public ForkJoinTask |
704 |
* documentation for rationale.) Interrupts are issued only in |
705 |
* tryTerminate, when workers should be terminating and tasks |
706 |
* should be cancelled anyway. Interrupts are cleared only when |
707 |
* necessary to ensure that calls to LockSupport.park do not loop |
708 |
* indefinitely (park returns immediately if the current thread is |
709 |
* interrupted). For cases in which task bodies are specified or |
710 |
* desired to interrupt upon cancellation, ForkJoinTask.cancel can |
711 |
* be overridden to do so (as is done for invoke{Any,All}). |
712 |
* |
713 |
* Memory placement |
714 |
* ================ |
715 |
* |
716 |
* Performance is very sensitive to placement of instances of |
717 |
* ForkJoinPool and WorkQueues and their queue arrays, as well the |
718 |
* placement of their fields. Caches misses and contention due to |
719 |
* false-sharing have been observed to slow down some programs by |
720 |
* more than a factor of four. There is no perfect solution, in |
721 |
* part because isolating more fields also generates more cache |
722 |
* misses in more common cases (because some fields snd slots are |
723 |
* usually read at the same time), and the main means of placing |
724 |
* memory, the @Contended annotation provides only rough control |
725 |
* (for good reason). We isolate the ForkJoinPool.ctl field as |
726 |
* well the set of WorkQueue fields that otherwise cause the most |
727 |
* false-sharing misses with respect to other fields. Also, |
728 |
* ForkJoinPool fields are ordered such that fields less prone to |
729 |
* contention effects are first, offsetting those that otherwise |
730 |
* would be, while also reducing total footprint vs using |
731 |
* multiple @Contended regions, which tends to slow down |
732 |
* less-contended applications. These arrangements mainly reduce |
733 |
* cache traffic by scanners, which speeds up finding tasks to |
734 |
* run. Initial sizing and resizing of WorkQueue arrays is an |
735 |
* even more delicate tradeoff because the best strategy may vary |
736 |
* across garbage collectors. Small arrays are better for locality |
737 |
* and reduce GC scan time, but large arrays reduce both direct |
738 |
* false-sharing and indirect cases due to GC bookkeeping |
739 |
* (cardmarks etc), and reduce the number of resizes, which are |
740 |
* not especially fast because they require atomic transfers, and |
741 |
* may cause other scanning workers to stall or give up. |
742 |
* Currently, arrays are initialized to be fairly small but early |
743 |
* resizes rapidly increase size by more than a factor of two |
744 |
* until very large. (Maintenance note: any changes in fields, |
745 |
* queues, or their uses must be accompanied by re-evaluation of |
746 |
* these placement and sizing decisions.) |
747 |
* |
748 |
* Style notes |
749 |
* =========== |
750 |
* |
751 |
* Memory ordering relies mainly on atomic operations (CAS, |
752 |
* getAndSet, getAndAdd) along with moded accesses. These use |
753 |
* jdk-internal Unsafe for atomics and special memory modes, |
754 |
* rather than VarHandles, to avoid initialization dependencies in |
755 |
* other jdk components that require early parallelism. This can |
756 |
* be awkward and ugly, but also reflects the need to control |
757 |
* outcomes across the unusual cases that arise in very racy code |
758 |
* with very few invariants. All fields are read into locals |
759 |
* before use, and null-checked if they are references, even if |
760 |
* they can never be null under current usages. Usually, |
761 |
* computations (held in local variables) are defined as soon as |
762 |
* logically enabled, sometimes to convince compilers that they |
763 |
* may be performed despite memory ordering constraints. Array |
764 |
* accesses using masked indices include checks (that are always |
765 |
* true) that the array length is non-zero to avoid compilers |
766 |
* inserting more expensive traps. This is usually done in a |
767 |
* "C"-like style of listing declarations at the heads of methods |
768 |
* or blocks, and using inline assignments on first encounter. |
769 |
* Nearly all explicit checks lead to bypass/return, not exception |
770 |
* throws, because they may legitimately arise during shutdown. A |
771 |
* few unusual loop constructions encourage (with varying |
772 |
* effectiveness) JVMs about where (not) to place safepoints. |
773 |
* |
774 |
* There is a lot of representation-level coupling among classes |
775 |
* ForkJoinPool, ForkJoinWorkerThread, and ForkJoinTask. The |
776 |
* fields of WorkQueue maintain data structures managed by |
777 |
* ForkJoinPool, so are directly accessed. There is little point |
778 |
* trying to reduce this, since any associated future changes in |
779 |
* representations will need to be accompanied by algorithmic |
780 |
* changes anyway. Several methods intrinsically sprawl because |
781 |
* they must accumulate sets of consistent reads of fields held in |
782 |
* local variables. Some others are artificially broken up to |
783 |
* reduce producer/consumer imbalances due to dynamic compilation. |
784 |
* There are also other coding oddities (including several |
785 |
* unnecessary-looking hoisted null checks) that help some methods |
786 |
* perform reasonably even when interpreted (not compiled). |
787 |
* |
788 |
* The order of declarations in this file is (with a few exceptions): |
789 |
* (1) Static constants |
790 |
* (2) Static utility functions |
791 |
* (3) Nested (static) classes |
792 |
* (4) Fields, along with constants used when unpacking some of them |
793 |
* (5) Internal control methods |
794 |
* (6) Callbacks and other support for ForkJoinTask methods |
795 |
* (7) Exported methods |
796 |
* (8) Static block initializing statics in minimally dependent order |
797 |
* |
798 |
* Revision notes |
799 |
* ============== |
800 |
* |
801 |
* The main sources of differences from previous version are: |
802 |
* |
803 |
* * Use of Unsafe vs VarHandle, including re-instatement of some |
804 |
* constructions from pre-VarHandle versions. |
805 |
* * Reduced memory and signal contention, mainly by distinguishing |
806 |
* failure cases. |
807 |
* * Improved initialization, in part by preparing for possible |
808 |
* removal of SecurityManager |
809 |
* * Enable resizing (includes refactoring quiescence/termination) |
810 |
* * Unification of most internal vs external operations; some made |
811 |
* possible via use of WorkQueue.access, and POOLSUBMIT status in tasks. |
812 |
*/ |
813 |
|
814 |
// static configuration constants |
815 |
|
816 |
/** |
817 |
* Default idle timeout value (in milliseconds) for idle threads |
818 |
* to park waiting for new work before terminating. |
819 |
*/ |
820 |
static final long DEFAULT_KEEPALIVE = 60_000L; |
821 |
|
822 |
/** |
823 |
* Undershoot tolerance for idle timeouts |
824 |
*/ |
825 |
static final long TIMEOUT_SLOP = 20L; |
826 |
|
827 |
/** |
828 |
* The default value for common pool maxSpares. Overridable using |
829 |
* the "java.util.concurrent.ForkJoinPool.common.maximumSpares" |
830 |
* system property. The default value is far in excess of normal |
831 |
* requirements, but also far short of MAX_CAP and typical OS |
832 |
* thread limits, so allows JVMs to catch misuse/abuse before |
833 |
* running out of resources needed to do so. |
834 |
*/ |
835 |
static final int DEFAULT_COMMON_MAX_SPARES = 256; |
836 |
|
837 |
/** |
838 |
* Initial capacity of work-stealing queue array. Must be a power |
839 |
* of two, at least 2. See above. |
840 |
*/ |
841 |
static final int INITIAL_QUEUE_CAPACITY = 1 << 6; |
842 |
|
843 |
// Bounds |
844 |
static final int SWIDTH = 16; // width of short |
845 |
static final int SMASK = 0xffff; // short bits == max index |
846 |
static final int MAX_CAP = 0x7fff; // max #workers - 1 |
847 |
|
848 |
// pool.runState and workQueue.access bits and sentinels |
849 |
static final int STOP = 1 << 31; // must be negative |
850 |
static final int SHUTDOWN = 1; |
851 |
static final int TERMINATED = 2; |
852 |
static final int PARKED = -1; // access value when parked |
853 |
|
854 |
// {pool, workQueue}.config bits |
855 |
static final int FIFO = 1 << 16; // fifo queue or access mode |
856 |
static final int SRC = 1 << 17; // set when stealable |
857 |
static final int CLEAR_TLS = 1 << 18; // set for Innocuous workers |
858 |
static final int TRIMMED = 1 << 19; // timed out while idle |
859 |
static final int ISCOMMON = 1 << 20; // set for common pool |
860 |
static final int PRESET_SIZE = 1 << 21; // size was set by property |
861 |
|
862 |
static final int UNCOMPENSATE = 1 << 16; // tryCompensate return |
863 |
|
864 |
/* |
865 |
* Bits and masks for ctl and bounds are packed with 4 16 bit subfields: |
866 |
* RC: Number of released (unqueued) workers |
867 |
* TC: Number of total workers |
868 |
* SS: version count and status of top waiting thread |
869 |
* ID: poolIndex of top of Treiber stack of waiters |
870 |
* |
871 |
* When convenient, we can extract the lower 32 stack top bits |
872 |
* (including version bits) as sp=(int)ctl. When sp is non-zero, |
873 |
* there are waiting workers. Count fields may be transiently |
874 |
* negative during termination because of out-of-order updates. |
875 |
* To deal with this, we use casts in and out of "short" and/or |
876 |
* signed shifts to maintain signedness. Because it occupies |
877 |
* uppermost bits, we can add one release count using getAndAdd of |
878 |
* RC_UNIT, rather than CAS, when returning from a blocked join. |
879 |
* Other updates of multiple subfields require CAS. |
880 |
*/ |
881 |
|
882 |
// Lower and upper word masks |
883 |
static final long SP_MASK = 0xffffffffL; |
884 |
static final long UC_MASK = ~SP_MASK; |
885 |
// Release counts |
886 |
static final int RC_SHIFT = 48; |
887 |
static final long RC_UNIT = 0x0001L << RC_SHIFT; |
888 |
static final long RC_MASK = 0xffffL << RC_SHIFT; |
889 |
// Total counts |
890 |
static final int TC_SHIFT = 32; |
891 |
static final long TC_UNIT = 0x0001L << TC_SHIFT; |
892 |
static final long TC_MASK = 0xffffL << TC_SHIFT; |
893 |
// sp bits |
894 |
static final int SS_SEQ = 1 << 16; // version count |
895 |
static final int INACTIVE = 1 << 31; // phase bit when idle |
896 |
|
897 |
// Static utilities |
898 |
|
899 |
/** |
900 |
* If there is a security manager, makes sure caller has |
901 |
* permission to modify threads. |
902 |
*/ |
903 |
@SuppressWarnings("removal") |
904 |
private static void checkPermission() { |
905 |
SecurityManager security; RuntimePermission perm; |
906 |
if ((security = System.getSecurityManager()) != null) { |
907 |
if ((perm = modifyThreadPermission) == null) |
908 |
modifyThreadPermission = perm = // races OK |
909 |
new RuntimePermission("modifyThread"); |
910 |
security.checkPermission(perm); |
911 |
} |
912 |
} |
913 |
|
914 |
// Nested classes |
915 |
|
916 |
/** |
917 |
* Factory for creating new {@link ForkJoinWorkerThread}s. |
918 |
* A {@code ForkJoinWorkerThreadFactory} must be defined and used |
919 |
* for {@code ForkJoinWorkerThread} subclasses that extend base |
920 |
* functionality or initialize threads with different contexts. |
921 |
*/ |
922 |
public static interface ForkJoinWorkerThreadFactory { |
923 |
/** |
924 |
* Returns a new worker thread operating in the given pool. |
925 |
* Returning null or throwing an exception may result in tasks |
926 |
* never being executed. If this method throws an exception, |
927 |
* it is relayed to the caller of the method (for example |
928 |
* {@code execute}) causing attempted thread creation. If this |
929 |
* method returns null or throws an exception, it is not |
930 |
* retried until the next attempted creation (for example |
931 |
* another call to {@code execute}). |
932 |
* |
933 |
* @param pool the pool this thread works in |
934 |
* @return the new worker thread, or {@code null} if the request |
935 |
* to create a thread is rejected |
936 |
* @throws NullPointerException if the pool is null |
937 |
*/ |
938 |
public ForkJoinWorkerThread newThread(ForkJoinPool pool); |
939 |
} |
940 |
|
941 |
/** |
942 |
* Default ForkJoinWorkerThreadFactory implementation; creates a |
943 |
* new ForkJoinWorkerThread using the system class loader as the |
944 |
* thread context class loader. |
945 |
*/ |
946 |
static final class DefaultForkJoinWorkerThreadFactory |
947 |
implements ForkJoinWorkerThreadFactory { |
948 |
public final ForkJoinWorkerThread newThread(ForkJoinPool pool) { |
949 |
boolean isCommon = (pool.workerNamePrefix == null); |
950 |
@SuppressWarnings("removal") |
951 |
SecurityManager sm = System.getSecurityManager(); |
952 |
if (sm == null) |
953 |
return new ForkJoinWorkerThread(null, pool, true, false); |
954 |
else if (isCommon) |
955 |
return newCommonWithACC(pool); |
956 |
else |
957 |
return newRegularWithACC(pool); |
958 |
} |
959 |
|
960 |
/* |
961 |
* Create and use static AccessControlContexts only if there |
962 |
* is a SecurityManager. (These can be removed if/when |
963 |
* SecurityManagers are removed from platform.) The ACCs are |
964 |
* immutable and equivalent even when racily initialized, so |
965 |
* they don't require locking, although with the chance of |
966 |
* needlessly duplicate construction. |
967 |
*/ |
968 |
@SuppressWarnings("removal") |
969 |
static volatile AccessControlContext regularACC, commonACC; |
970 |
|
971 |
@SuppressWarnings("removal") |
972 |
static ForkJoinWorkerThread newRegularWithACC(ForkJoinPool pool) { |
973 |
AccessControlContext acc = regularACC; |
974 |
if (acc == null) { |
975 |
Permissions ps = new Permissions(); |
976 |
ps.add(new RuntimePermission("getClassLoader")); |
977 |
ps.add(new RuntimePermission("setContextClassLoader")); |
978 |
regularACC = acc = |
979 |
new AccessControlContext(new ProtectionDomain[] { |
980 |
new ProtectionDomain(null, ps) }); |
981 |
} |
982 |
return AccessController.doPrivileged( |
983 |
new PrivilegedAction<>() { |
984 |
public ForkJoinWorkerThread run() { |
985 |
return new ForkJoinWorkerThread(null, pool, true, false); |
986 |
}}, acc); |
987 |
} |
988 |
|
989 |
@SuppressWarnings("removal") |
990 |
static ForkJoinWorkerThread newCommonWithACC(ForkJoinPool pool) { |
991 |
AccessControlContext acc = commonACC; |
992 |
if (acc == null) { |
993 |
Permissions ps = new Permissions(); |
994 |
ps.add(new RuntimePermission("getClassLoader")); |
995 |
ps.add(new RuntimePermission("setContextClassLoader")); |
996 |
ps.add(new RuntimePermission("modifyThread")); |
997 |
ps.add(new RuntimePermission("enableContextClassLoaderOverride")); |
998 |
ps.add(new RuntimePermission("modifyThreadGroup")); |
999 |
commonACC = acc = |
1000 |
new AccessControlContext(new ProtectionDomain[] { |
1001 |
new ProtectionDomain(null, ps) }); |
1002 |
} |
1003 |
return AccessController.doPrivileged( |
1004 |
new PrivilegedAction<>() { |
1005 |
public ForkJoinWorkerThread run() { |
1006 |
return new ForkJoinWorkerThread. |
1007 |
InnocuousForkJoinWorkerThread(pool); |
1008 |
}}, acc); |
1009 |
} |
1010 |
} |
1011 |
|
1012 |
/** |
1013 |
* Queues supporting work-stealing as well as external task |
1014 |
* submission. See above for descriptions and algorithms. |
1015 |
*/ |
1016 |
static final class WorkQueue { |
1017 |
int stackPred; // pool stack (ctl) predecessor link |
1018 |
int config; // index, mode, ORed with SRC after init |
1019 |
int base; // index of next slot for poll |
1020 |
ForkJoinTask<?>[] array; // the queued tasks; power of 2 size |
1021 |
final ForkJoinWorkerThread owner; // owning thread or null if shared |
1022 |
|
1023 |
// fields otherwise causing more unnecessary false-sharing cache misses |
1024 |
@jdk.internal.vm.annotation.Contended("w") |
1025 |
int top; // index of next slot for push |
1026 |
@jdk.internal.vm.annotation.Contended("w") |
1027 |
volatile int access; // values 0, 1 (locked), PARKED, STOP |
1028 |
@jdk.internal.vm.annotation.Contended("w") |
1029 |
volatile int phase; // versioned, negative if inactive |
1030 |
@jdk.internal.vm.annotation.Contended("w") |
1031 |
volatile int source; // source queue id in topLevelExec |
1032 |
@jdk.internal.vm.annotation.Contended("w") |
1033 |
int nsteals; // number of steals from other queues |
1034 |
|
1035 |
// Support for atomic operations |
1036 |
private static final Unsafe U; |
1037 |
private static final long ACCESS; |
1038 |
private static final long PHASE; |
1039 |
private static final long ABASE; |
1040 |
private static final int ASHIFT; |
1041 |
|
1042 |
static ForkJoinTask<?> getAndClearSlot(ForkJoinTask<?>[] a, int i) { |
1043 |
return (ForkJoinTask<?>) |
1044 |
U.getAndSetReference(a, ((long)i << ASHIFT) + ABASE, null); |
1045 |
} |
1046 |
static boolean casSlotToNull(ForkJoinTask<?>[] a, int i, |
1047 |
ForkJoinTask<?> c) { |
1048 |
return U.compareAndSetReference(a, ((long)i << ASHIFT) + ABASE, |
1049 |
c, null); |
1050 |
} |
1051 |
final void forcePhaseActive() { // clear sign bit |
1052 |
U.getAndBitwiseAndInt(this, PHASE, 0x7fffffff); |
1053 |
} |
1054 |
final int getAndSetAccess(int v) { |
1055 |
return U.getAndSetInt(this, ACCESS, v); |
1056 |
} |
1057 |
final void releaseAccess() { |
1058 |
U.putIntRelease(this, ACCESS, 0); |
1059 |
} |
1060 |
|
1061 |
/** |
1062 |
* Constructor. For owned queues, most fields are initialized |
1063 |
* upon thread start in pool.registerWorker. |
1064 |
*/ |
1065 |
WorkQueue(ForkJoinWorkerThread owner, int config) { |
1066 |
this.owner = owner; |
1067 |
this.config = config; |
1068 |
base = top = 1; |
1069 |
} |
1070 |
|
1071 |
/** |
1072 |
* Returns an exportable index (used by ForkJoinWorkerThread). |
1073 |
*/ |
1074 |
final int getPoolIndex() { |
1075 |
return (config & 0xffff) >>> 1; // ignore odd/even tag bit |
1076 |
} |
1077 |
|
1078 |
/** |
1079 |
* Returns the approximate number of tasks in the queue. |
1080 |
*/ |
1081 |
final int queueSize() { |
1082 |
int unused = access; // for ordering effect |
1083 |
return Math.max(top - base, 0); // ignore transient negative |
1084 |
} |
1085 |
|
1086 |
/** |
1087 |
* Pushes a task. Called only by owner or if already locked |
1088 |
* |
1089 |
* @param task the task. Caller must ensure non-null. |
1090 |
* @param pool the pool. Must be non-null unless terminating. |
1091 |
* @param signalIfEmpty true if signal when pushing to empty queue |
1092 |
* @throws RejectedExecutionException if array cannot be resized |
1093 |
*/ |
1094 |
final void push(ForkJoinTask<?> task, ForkJoinPool pool, |
1095 |
boolean signalIfEmpty) { |
1096 |
boolean resize = false; |
1097 |
int s = top++, b = base, cap, m; ForkJoinTask<?>[] a; |
1098 |
if ((a = array) != null && (cap = a.length) > 0) { |
1099 |
if ((m = (cap - 1)) == s - b) { |
1100 |
resize = true; // rapidly grow until large |
1101 |
int newCap = (cap < 1 << 24) ? cap << 2 : cap << 1; |
1102 |
ForkJoinTask<?>[] newArray; |
1103 |
try { |
1104 |
newArray = new ForkJoinTask<?>[newCap]; |
1105 |
} catch (Throwable ex) { |
1106 |
top = s; |
1107 |
access = 0; |
1108 |
throw new RejectedExecutionException( |
1109 |
"Queue capacity exceeded"); |
1110 |
} |
1111 |
if (newCap > 0) { // always true |
1112 |
int newMask = newCap - 1, k = s; |
1113 |
do { // poll old, push to new |
1114 |
newArray[k-- & newMask] = task; |
1115 |
} while ((task = getAndClearSlot(a, k & m)) != null); |
1116 |
} |
1117 |
array = newArray; |
1118 |
} |
1119 |
else |
1120 |
a[m & s] = task; |
1121 |
getAndSetAccess(0); // for memory effects if owned |
1122 |
if ((resize || (a[m & (s - 1)] == null && signalIfEmpty)) && |
1123 |
pool != null) |
1124 |
pool.signalWork(); |
1125 |
} |
1126 |
} |
1127 |
|
1128 |
/** |
1129 |
* Takes next task, if one exists, in order specified by mode, |
1130 |
* so acts as either local-pop or local-poll. Called only by owner. |
1131 |
* @param fifo nonzero if FIFO mode |
1132 |
*/ |
1133 |
final ForkJoinTask<?> nextLocalTask(int fifo) { |
1134 |
ForkJoinTask<?> t = null; |
1135 |
ForkJoinTask<?>[] a = array; |
1136 |
int p = top, s = p - 1, b = base, nb, cap; |
1137 |
if (p - b > 0 && a != null && (cap = a.length) > 0) { |
1138 |
do { |
1139 |
if (fifo == 0 || (nb = b + 1) == p) { |
1140 |
if ((t = getAndClearSlot(a, (cap - 1) & s)) != null) |
1141 |
top = s; |
1142 |
break; // lost race for only task |
1143 |
} |
1144 |
else if ((t = getAndClearSlot(a, (cap - 1) & b)) != null) { |
1145 |
base = nb; |
1146 |
break; |
1147 |
} |
1148 |
else { |
1149 |
while (b == (b = base)) { |
1150 |
U.loadFence(); |
1151 |
Thread.onSpinWait(); // spin to reduce memory traffic |
1152 |
} |
1153 |
} |
1154 |
} while (p - b > 0); |
1155 |
U.storeStoreFence(); // for timely index updates |
1156 |
} |
1157 |
return t; |
1158 |
} |
1159 |
|
1160 |
/** |
1161 |
* Takes next task, if one exists, using configured mode. |
1162 |
* (Always owned, never called for Common pool.) |
1163 |
*/ |
1164 |
final ForkJoinTask<?> nextLocalTask() { |
1165 |
return nextLocalTask(config & FIFO); |
1166 |
} |
1167 |
|
1168 |
/** |
1169 |
* Pops the given task only if it is at the current top. |
1170 |
*/ |
1171 |
final boolean tryUnpush(ForkJoinTask<?> task, boolean owned) { |
1172 |
ForkJoinTask<?>[] a = array; |
1173 |
int p = top, s, cap, k; |
1174 |
if (task != null && base != p && a != null && (cap = a.length) > 0 && |
1175 |
a[k = (cap - 1) & (s = p - 1)] == task) { |
1176 |
if (owned || getAndSetAccess(1) == 0) { |
1177 |
if (top != p || a[k] != task || |
1178 |
getAndClearSlot(a, k) == null) |
1179 |
access = 0; |
1180 |
else { |
1181 |
top = s; |
1182 |
access = 0; |
1183 |
return true; |
1184 |
} |
1185 |
} |
1186 |
} |
1187 |
return false; |
1188 |
} |
1189 |
|
1190 |
/** |
1191 |
* Returns next task, if one exists, in order specified by mode. |
1192 |
*/ |
1193 |
final ForkJoinTask<?> peek() { |
1194 |
ForkJoinTask<?>[] a = array; |
1195 |
int cfg = config, p = top, b = base, cap; |
1196 |
if (p != b && a != null && (cap = a.length) > 0) { |
1197 |
if ((cfg & FIFO) == 0) |
1198 |
return a[(cap - 1) & (p - 1)]; |
1199 |
else { // skip over in-progress removals |
1200 |
ForkJoinTask<?> t; |
1201 |
for ( ; p - b > 0; ++b) { |
1202 |
if ((t = a[(cap - 1) & b]) != null) |
1203 |
return t; |
1204 |
} |
1205 |
} |
1206 |
} |
1207 |
return null; |
1208 |
} |
1209 |
|
1210 |
/** |
1211 |
* Polls for a task. Used only by non-owners in usually |
1212 |
* uncontended contexts. |
1213 |
* |
1214 |
* @param pool if nonnull, pool to signal if more tasks exist |
1215 |
*/ |
1216 |
final ForkJoinTask<?> poll(ForkJoinPool pool) { |
1217 |
for (int b = base;;) { |
1218 |
int cap; ForkJoinTask<?>[] a; |
1219 |
if ((a = array) == null || (cap = a.length) <= 0) |
1220 |
break; // currently impossible |
1221 |
int k = (cap - 1) & b, nb = b + 1, nk = (cap - 1) & nb; |
1222 |
ForkJoinTask<?> t = a[k]; |
1223 |
U.loadFence(); // for re-reads |
1224 |
if (b != (b = base)) // inconsistent |
1225 |
; |
1226 |
else if (t != null && casSlotToNull(a, k, t)) { |
1227 |
base = nb; |
1228 |
U.storeFence(); |
1229 |
if (pool != null && a[nk] != null) |
1230 |
pool.signalWork(); // propagate |
1231 |
return t; |
1232 |
} |
1233 |
else if (array != a || a[k] != null) |
1234 |
; // stale |
1235 |
else if (a[nk] == null && top - b <= 0) |
1236 |
break; // empty |
1237 |
} |
1238 |
return null; |
1239 |
} |
1240 |
|
1241 |
/** |
1242 |
* Tries to poll next task in FIFO order, failing on |
1243 |
* contention or stalls. Used only by topLevelExec to repoll |
1244 |
* from the queue obtained from pool.scan. |
1245 |
*/ |
1246 |
final ForkJoinTask<?> tryPoll() { |
1247 |
int b = base, cap; ForkJoinTask<?>[] a; |
1248 |
if ((a = array) != null && (cap = a.length) > 0) { |
1249 |
for (;;) { |
1250 |
int k = (cap - 1) & b, nb = b + 1; |
1251 |
ForkJoinTask<?> t = a[k]; |
1252 |
U.loadFence(); // for re-reads |
1253 |
if (b != (b = base)) |
1254 |
; // inconsistent |
1255 |
else if (t != null) { |
1256 |
if (casSlotToNull(a, k, t)) { |
1257 |
base = nb; |
1258 |
U.storeStoreFence(); |
1259 |
return t; |
1260 |
} |
1261 |
break; // contended |
1262 |
} |
1263 |
else if (a[k] == null) |
1264 |
break; // empty or stalled |
1265 |
} |
1266 |
} |
1267 |
return null; |
1268 |
} |
1269 |
|
1270 |
// specialized execution methods |
1271 |
|
1272 |
/** |
1273 |
* Runs the given (stolen) task if nonnull, as well as |
1274 |
* remaining local tasks and/or others available from its |
1275 |
* source queue, if any. |
1276 |
*/ |
1277 |
final void topLevelExec(ForkJoinTask<?> task, WorkQueue src) { |
1278 |
int cfg = config, fifo = cfg & FIFO, nstolen = 1; |
1279 |
while (task != null) { |
1280 |
task.doExec(); |
1281 |
if ((task = nextLocalTask(fifo)) == null && |
1282 |
src != null && (task = src.tryPoll()) != null) |
1283 |
++nstolen; |
1284 |
} |
1285 |
nsteals += nstolen; |
1286 |
source = 0; |
1287 |
if ((cfg & CLEAR_TLS) != 0) |
1288 |
ThreadLocalRandom.eraseThreadLocals(Thread.currentThread()); |
1289 |
} |
1290 |
|
1291 |
/** |
1292 |
* Deep form of tryUnpush: Traverses from top and removes and |
1293 |
* runs task if present, shifting others to fill gap. |
1294 |
* @return task status if removed, else 0 |
1295 |
*/ |
1296 |
final int tryRemoveAndExec(ForkJoinTask<?> task, boolean owned) { |
1297 |
ForkJoinTask<?>[] a = array; |
1298 |
int p = top, s = p - 1, d = p - base, cap; |
1299 |
if (task != null && d > 0 && a != null && (cap = a.length) > 0) { |
1300 |
for (int m = cap - 1, i = s; ; --i) { |
1301 |
ForkJoinTask<?> t; int k; |
1302 |
if ((t = a[k = i & m]) == task) { |
1303 |
if (!owned && getAndSetAccess(1) != 0) |
1304 |
break; // fail if locked |
1305 |
else if (top != p || a[k] != task || |
1306 |
getAndClearSlot(a, k) == null) { |
1307 |
access = 0; |
1308 |
break; // missed |
1309 |
} |
1310 |
else { |
1311 |
if (i != s && i == base) |
1312 |
base = i + 1; // avoid shift |
1313 |
else { |
1314 |
for (int j = i; j != s;) // shift down |
1315 |
a[j & m] = getAndClearSlot(a, ++j & m); |
1316 |
top = s; |
1317 |
} |
1318 |
releaseAccess(); |
1319 |
return task.doExec(); |
1320 |
} |
1321 |
} |
1322 |
else if (t == null || --d == 0) |
1323 |
break; |
1324 |
} |
1325 |
} |
1326 |
return 0; |
1327 |
} |
1328 |
|
1329 |
/** |
1330 |
* Tries to pop and run tasks within the target's computation |
1331 |
* until done, not found, or limit exceeded. |
1332 |
* |
1333 |
* @param task root of computation |
1334 |
* @param limit max runs, or zero for no limit |
1335 |
* @return task status on exit |
1336 |
*/ |
1337 |
final int helpComplete(ForkJoinTask<?> task, boolean owned, int limit) { |
1338 |
int status = 0; |
1339 |
if (task != null) { |
1340 |
outer: for (;;) { |
1341 |
ForkJoinTask<?>[] a; ForkJoinTask<?> t; |
1342 |
int p, s, cap, k; |
1343 |
if ((status = task.status) < 0) |
1344 |
return status; |
1345 |
if ((a = array) == null || (cap = a.length) <= 0 || |
1346 |
(t = a[k = (cap - 1) & (s = (p = top) - 1)]) == null || |
1347 |
!(t instanceof CountedCompleter)) |
1348 |
break; |
1349 |
for (CountedCompleter<?> f = (CountedCompleter<?>)t;;) { |
1350 |
if (f == task) |
1351 |
break; |
1352 |
else if ((f = f.completer) == null) |
1353 |
break outer; // ineligible |
1354 |
} |
1355 |
if (!owned && getAndSetAccess(1) != 0) |
1356 |
break; // fail if locked |
1357 |
if (top != p || a[k] != t || getAndClearSlot(a, k) == null) { |
1358 |
access = 0; |
1359 |
break; // missed |
1360 |
} |
1361 |
top = s; |
1362 |
releaseAccess(); |
1363 |
t.doExec(); |
1364 |
if (limit != 0 && --limit == 0) |
1365 |
break; |
1366 |
} |
1367 |
status = task.status; |
1368 |
} |
1369 |
return status; |
1370 |
} |
1371 |
|
1372 |
/** |
1373 |
* Tries to poll and run AsynchronousCompletionTasks until |
1374 |
* none found or blocker is released |
1375 |
* |
1376 |
* @param blocker the blocker |
1377 |
*/ |
1378 |
final void helpAsyncBlocker(ManagedBlocker blocker) { |
1379 |
if (blocker != null) { |
1380 |
for (;;) { |
1381 |
int b = base, cap; ForkJoinTask<?>[] a; |
1382 |
if ((a = array) == null || (cap = a.length) <= 0 || b == top) |
1383 |
break; |
1384 |
int k = (cap - 1) & b, nb = b + 1, nk = (cap - 1) & nb; |
1385 |
ForkJoinTask<?> t = a[k]; |
1386 |
U.loadFence(); // for re-reads |
1387 |
if (base != b) |
1388 |
; |
1389 |
else if (blocker.isReleasable()) |
1390 |
break; |
1391 |
else if (a[k] != t) |
1392 |
; |
1393 |
else if (t != null) { |
1394 |
if (!(t instanceof CompletableFuture |
1395 |
.AsynchronousCompletionTask)) |
1396 |
break; |
1397 |
else if (casSlotToNull(a, k, t)) { |
1398 |
base = nb; |
1399 |
U.storeStoreFence(); |
1400 |
t.doExec(); |
1401 |
} |
1402 |
} |
1403 |
else if (a[nk] == null) |
1404 |
break; |
1405 |
} |
1406 |
} |
1407 |
} |
1408 |
|
1409 |
// misc |
1410 |
|
1411 |
/** |
1412 |
* Returns true if owned by a worker thread and not known to be blocked. |
1413 |
*/ |
1414 |
final boolean isApparentlyUnblocked() { |
1415 |
Thread wt; Thread.State s; |
1416 |
return (access != STOP && (wt = owner) != null && |
1417 |
(s = wt.getState()) != Thread.State.BLOCKED && |
1418 |
s != Thread.State.WAITING && |
1419 |
s != Thread.State.TIMED_WAITING); |
1420 |
} |
1421 |
|
1422 |
/** |
1423 |
* Called in constructors if ThreadLocals not preserved |
1424 |
*/ |
1425 |
final void setClearThreadLocals() { |
1426 |
config |= CLEAR_TLS; |
1427 |
} |
1428 |
|
1429 |
static { |
1430 |
U = Unsafe.getUnsafe(); |
1431 |
Class<WorkQueue> klass = WorkQueue.class; |
1432 |
ACCESS = U.objectFieldOffset(klass, "access"); |
1433 |
PHASE = U.objectFieldOffset(klass, "phase"); |
1434 |
Class<ForkJoinTask[]> aklass = ForkJoinTask[].class; |
1435 |
ABASE = U.arrayBaseOffset(aklass); |
1436 |
int scale = U.arrayIndexScale(aklass); |
1437 |
ASHIFT = 31 - Integer.numberOfLeadingZeros(scale); |
1438 |
if ((scale & (scale - 1)) != 0) |
1439 |
throw new Error("array index scale not a power of two"); |
1440 |
} |
1441 |
} |
1442 |
|
1443 |
// static fields (initialized in static initializer below) |
1444 |
|
1445 |
/** |
1446 |
* Creates a new ForkJoinWorkerThread. This factory is used unless |
1447 |
* overridden in ForkJoinPool constructors. |
1448 |
*/ |
1449 |
public static final ForkJoinWorkerThreadFactory |
1450 |
defaultForkJoinWorkerThreadFactory; |
1451 |
|
1452 |
/** |
1453 |
* Common (static) pool. Non-null for public use unless a static |
1454 |
* construction exception, but internal usages null-check on use |
1455 |
* to paranoically avoid potential initialization circularities |
1456 |
* as well as to simplify generated code. |
1457 |
*/ |
1458 |
static final ForkJoinPool common; |
1459 |
|
1460 |
/** |
1461 |
* Sequence number for creating worker names |
1462 |
*/ |
1463 |
private static volatile int poolIds; |
1464 |
|
1465 |
/** |
1466 |
* Permission required for callers of methods that may start or |
1467 |
* kill threads. Lazily constructed. |
1468 |
*/ |
1469 |
static volatile RuntimePermission modifyThreadPermission; |
1470 |
|
1471 |
|
1472 |
// Instance fields |
1473 |
volatile long stealCount; // collects worker nsteals |
1474 |
volatile long threadIds; // for worker thread names |
1475 |
final long keepAlive; // milliseconds before dropping if idle |
1476 |
final long bounds; // min, max threads packed as shorts |
1477 |
final int config; // static configuration bits |
1478 |
volatile int runState; // SHUTDOWN, STOP, TERMINATED bits |
1479 |
WorkQueue[] queues; // main registry |
1480 |
final ReentrantLock registrationLock; |
1481 |
Condition termination; // lazily constructed |
1482 |
final String workerNamePrefix; // null for common pool |
1483 |
final ForkJoinWorkerThreadFactory factory; |
1484 |
final UncaughtExceptionHandler ueh; // per-worker UEH |
1485 |
final Predicate<? super ForkJoinPool> saturate; |
1486 |
// final SharedThreadContainer container; // for loom |
1487 |
|
1488 |
@jdk.internal.vm.annotation.Contended("fjpctl") // segregate |
1489 |
volatile long ctl; // main pool control |
1490 |
@jdk.internal.vm.annotation.Contended("fjpctl") // colocate |
1491 |
int parallelism; // target number of workers |
1492 |
|
1493 |
// Support for atomic operations |
1494 |
private static final Unsafe U; |
1495 |
private static final long CTL; |
1496 |
private static final long RUNSTATE; |
1497 |
private static final long PARALLELISM; |
1498 |
private static final long THREADIDS; |
1499 |
private static final long POOLIDS; |
1500 |
|
1501 |
private boolean compareAndSetCtl(long c, long v) { |
1502 |
return U.compareAndSetLong(this, CTL, c, v); |
1503 |
} |
1504 |
private long compareAndExchangeCtl(long c, long v) { |
1505 |
return U.compareAndExchangeLong(this, CTL, c, v); |
1506 |
} |
1507 |
private long getAndAddCtl(long v) { |
1508 |
return U.getAndAddLong(this, CTL, v); |
1509 |
} |
1510 |
private int getAndBitwiseOrRunState(int v) { |
1511 |
return U.getAndBitwiseOrInt(this, RUNSTATE, v); |
1512 |
} |
1513 |
private long incrementThreadIds() { |
1514 |
return U.getAndAddLong(this, THREADIDS, 1L); |
1515 |
} |
1516 |
private static int getAndAddPoolIds(int x) { |
1517 |
return U.getAndAddInt(ForkJoinPool.class, POOLIDS, x); |
1518 |
} |
1519 |
private int getAndSetParallelism(int v) { |
1520 |
return U.getAndSetInt(this, PARALLELISM, v); |
1521 |
} |
1522 |
private int getParallelismOpaque() { |
1523 |
return U.getIntOpaque(this, PARALLELISM); |
1524 |
} |
1525 |
|
1526 |
// Creating, registering, and deregistering workers |
1527 |
|
1528 |
/** |
1529 |
* Tries to construct and start one worker. Assumes that total |
1530 |
* count has already been incremented as a reservation. Invokes |
1531 |
* deregisterWorker on any failure. |
1532 |
* |
1533 |
* @return true if successful |
1534 |
*/ |
1535 |
private boolean createWorker() { |
1536 |
ForkJoinWorkerThreadFactory fac = factory; |
1537 |
Throwable ex = null; |
1538 |
ForkJoinWorkerThread wt = null; |
1539 |
try { |
1540 |
if (runState >= 0 && // avoid construction if terminating |
1541 |
fac != null && (wt = fac.newThread(this)) != null) { |
1542 |
wt.start(); // replace with following line for loom |
1543 |
// container.start(wt); |
1544 |
return true; |
1545 |
} |
1546 |
} catch (Throwable rex) { |
1547 |
ex = rex; |
1548 |
} |
1549 |
deregisterWorker(wt, ex); |
1550 |
return false; |
1551 |
} |
1552 |
|
1553 |
/** |
1554 |
* Provides a name for ForkJoinWorkerThread constructor. |
1555 |
*/ |
1556 |
final String nextWorkerThreadName() { |
1557 |
String prefix = workerNamePrefix; |
1558 |
long tid = incrementThreadIds() + 1L; |
1559 |
if (prefix == null) // commonPool has no prefix |
1560 |
prefix = "ForkJoinPool.commonPool-worker-"; |
1561 |
return prefix.concat(Long.toString(tid)); |
1562 |
} |
1563 |
|
1564 |
/** |
1565 |
* Finishes initializing and records owned queue. |
1566 |
* |
1567 |
* @param w caller's WorkQueue |
1568 |
*/ |
1569 |
final void registerWorker(WorkQueue w) { |
1570 |
ThreadLocalRandom.localInit(); |
1571 |
int seed = ThreadLocalRandom.getProbe(); |
1572 |
ReentrantLock lock = registrationLock; |
1573 |
int cfg = config & FIFO; |
1574 |
if (w != null && lock != null) { |
1575 |
w.array = new ForkJoinTask<?>[INITIAL_QUEUE_CAPACITY]; |
1576 |
cfg |= w.config | SRC; |
1577 |
w.stackPred = seed; |
1578 |
int id = (seed << 1) | 1; // initial index guess |
1579 |
lock.lock(); |
1580 |
try { |
1581 |
WorkQueue[] qs; int n; // find queue index |
1582 |
if ((qs = queues) != null && (n = qs.length) > 0) { |
1583 |
int k = n, m = n - 1; |
1584 |
for (; qs[id &= m] != null && k > 0; id -= 2, k -= 2); |
1585 |
if (k == 0) |
1586 |
id = n | 1; // resize below |
1587 |
w.phase = w.config = id | cfg; // now publishable |
1588 |
|
1589 |
if (id < n) |
1590 |
qs[id] = w; |
1591 |
else { // expand array |
1592 |
int an = n << 1, am = an - 1; |
1593 |
WorkQueue[] as = new WorkQueue[an]; |
1594 |
as[id & am] = w; |
1595 |
for (int j = 1; j < n; j += 2) |
1596 |
as[j] = qs[j]; |
1597 |
for (int j = 0; j < n; j += 2) { |
1598 |
WorkQueue q; |
1599 |
if ((q = qs[j]) != null) // shared queues may move |
1600 |
as[q.config & am] = q; |
1601 |
} |
1602 |
U.storeFence(); // fill before publish |
1603 |
queues = as; |
1604 |
} |
1605 |
} |
1606 |
} finally { |
1607 |
lock.unlock(); |
1608 |
} |
1609 |
} |
1610 |
} |
1611 |
|
1612 |
/** |
1613 |
* Final callback from terminating worker, as well as upon failure |
1614 |
* to construct or start a worker. Removes record of worker from |
1615 |
* array, and adjusts counts. If pool is shutting down, tries to |
1616 |
* complete termination. |
1617 |
* |
1618 |
* @param wt the worker thread, or null if construction failed |
1619 |
* @param ex the exception causing failure, or null if none |
1620 |
*/ |
1621 |
final void deregisterWorker(ForkJoinWorkerThread wt, Throwable ex) { |
1622 |
WorkQueue w = (wt == null) ? null : wt.workQueue; |
1623 |
int cfg = (w == null) ? 0 : w.config; |
1624 |
long c = ctl; |
1625 |
if ((cfg & TRIMMED) == 0) // decrement counts |
1626 |
do {} while (c != (c = compareAndExchangeCtl( |
1627 |
c, ((RC_MASK & (c - RC_UNIT)) | |
1628 |
(TC_MASK & (c - TC_UNIT)) | |
1629 |
(SP_MASK & c))))); |
1630 |
else if ((int)c == 0) // was dropped on timeout |
1631 |
cfg &= ~SRC; // suppress signal if last |
1632 |
if (!tryTerminate(false, false) && w != null) { |
1633 |
ReentrantLock lock; WorkQueue[] qs; int n, i; |
1634 |
long ns = w.nsteals & 0xffffffffL; |
1635 |
if ((lock = registrationLock) != null) { |
1636 |
lock.lock(); // remove index unless terminating |
1637 |
if ((qs = queues) != null && (n = qs.length) > 0 && |
1638 |
qs[i = cfg & (n - 1)] == w) |
1639 |
qs[i] = null; |
1640 |
stealCount += ns; // accumulate steals |
1641 |
lock.unlock(); |
1642 |
} |
1643 |
if ((cfg & SRC) != 0) |
1644 |
signalWork(); // possibly replace worker |
1645 |
} |
1646 |
if (ex != null) { |
1647 |
if (w != null) { |
1648 |
w.access = STOP; // cancel tasks |
1649 |
for (ForkJoinTask<?> t; (t = w.nextLocalTask(0)) != null; ) |
1650 |
ForkJoinTask.cancelIgnoringExceptions(t); |
1651 |
} |
1652 |
ForkJoinTask.rethrow(ex); |
1653 |
} |
1654 |
} |
1655 |
|
1656 |
/* |
1657 |
* Releases an idle worker, or creates one if not enough exist. |
1658 |
*/ |
1659 |
final void signalWork() { |
1660 |
int pc = parallelism, n; |
1661 |
long c = ctl; |
1662 |
WorkQueue[] qs = queues; |
1663 |
if ((short)(c >>> RC_SHIFT) < pc && qs != null && (n = qs.length) > 0) { |
1664 |
for (;;) { |
1665 |
boolean create = false; |
1666 |
int sp = (int)c & ~INACTIVE; |
1667 |
WorkQueue v = qs[sp & (n - 1)]; |
1668 |
int deficit = pc - (short)(c >>> TC_SHIFT); |
1669 |
long ac = (c + RC_UNIT) & RC_MASK, nc; |
1670 |
if (sp != 0 && v != null) |
1671 |
nc = (v.stackPred & SP_MASK) | (c & TC_MASK); |
1672 |
else if (deficit <= 0) |
1673 |
break; |
1674 |
else { |
1675 |
create = true; |
1676 |
nc = ((c + TC_UNIT) & TC_MASK); |
1677 |
} |
1678 |
if (c == (c = compareAndExchangeCtl(c, nc | ac))) { |
1679 |
if (create) |
1680 |
createWorker(); |
1681 |
else { |
1682 |
Thread owner = v.owner; |
1683 |
v.phase = sp; |
1684 |
if (v.access == PARKED) |
1685 |
LockSupport.unpark(owner); |
1686 |
} |
1687 |
break; |
1688 |
} |
1689 |
} |
1690 |
} |
1691 |
} |
1692 |
|
1693 |
/** |
1694 |
* Reactivates any idle worker, if one exists. |
1695 |
* |
1696 |
* @return the signalled worker, or null if none |
1697 |
*/ |
1698 |
private WorkQueue reactivate() { |
1699 |
WorkQueue[] qs; int n; |
1700 |
long c = ctl; |
1701 |
if ((qs = queues) != null && (n = qs.length) > 0) { |
1702 |
for (;;) { |
1703 |
int sp = (int)c & ~INACTIVE; |
1704 |
WorkQueue v = qs[sp & (n - 1)]; |
1705 |
long ac = UC_MASK & (c + RC_UNIT); |
1706 |
if (sp == 0 || v == null) |
1707 |
break; |
1708 |
if (c == (c = compareAndExchangeCtl( |
1709 |
c, (v.stackPred & SP_MASK) | ac))) { |
1710 |
Thread owner = v.owner; |
1711 |
v.phase = sp; |
1712 |
if (v.access == PARKED) |
1713 |
LockSupport.unpark(owner); |
1714 |
return v; |
1715 |
} |
1716 |
} |
1717 |
} |
1718 |
return null; |
1719 |
} |
1720 |
|
1721 |
/** |
1722 |
* Tries to deactivate worker w; called only on idle timeout. |
1723 |
*/ |
1724 |
private boolean tryTrim(WorkQueue w) { |
1725 |
if (w != null) { |
1726 |
int pred = w.stackPred, cfg = w.config | TRIMMED; |
1727 |
long c = ctl; |
1728 |
int sp = (int)c & ~INACTIVE; |
1729 |
if ((sp & SMASK) == (cfg & SMASK) && |
1730 |
compareAndSetCtl(c, ((pred & SP_MASK) | |
1731 |
(UC_MASK & (c - TC_UNIT))))) { |
1732 |
w.config = cfg; // add sentinel for deregisterWorker |
1733 |
w.phase = sp; |
1734 |
return true; |
1735 |
} |
1736 |
} |
1737 |
return false; |
1738 |
} |
1739 |
|
1740 |
/** |
1741 |
* Returns true if any queue is detectably nonempty. Accurate |
1742 |
* only when workers are quiescent; else conservatively |
1743 |
* approximate. |
1744 |
* @param submissionsOnly if true, only check submission queues |
1745 |
*/ |
1746 |
private boolean hasTasks(boolean submissionsOnly) { |
1747 |
int step = submissionsOnly ? 2 : 1; |
1748 |
for (int checkSum = 0;;) { // repeat until stable (normally twice) |
1749 |
U.loadFence(); |
1750 |
WorkQueue[] qs = queues; |
1751 |
int n = (qs == null) ? 0 : qs.length, sum = 0; |
1752 |
for (int i = 0; i < n; i += step) { |
1753 |
WorkQueue q; int s; |
1754 |
if ((q = qs[i]) != null) { |
1755 |
if (q.access > 0 || (s = q.top) != q.base) |
1756 |
return true; |
1757 |
sum += (s << 16) + i + 1; |
1758 |
} |
1759 |
} |
1760 |
if (checkSum == (checkSum = sum)) |
1761 |
return false; |
1762 |
} |
1763 |
} |
1764 |
|
1765 |
/** |
1766 |
* Top-level runloop for workers, called by ForkJoinWorkerThread.run. |
1767 |
* See above for explanation. |
1768 |
* |
1769 |
* @param w caller's WorkQueue (may be null on failed initialization) |
1770 |
*/ |
1771 |
final void runWorker(WorkQueue w) { |
1772 |
if (w != null) { // skip on failed init |
1773 |
int r = w.stackPred, src = 0; // use seed from registerWorker |
1774 |
do { |
1775 |
r ^= r << 13; r ^= r >>> 17; r ^= r << 5; // xorshift |
1776 |
} while ((src = scan(w, src, r)) >= 0 || |
1777 |
(src = awaitWork(w)) == 0); |
1778 |
w.access = STOP; // record normal termination |
1779 |
} |
1780 |
} |
1781 |
|
1782 |
/** |
1783 |
* Scans for and if found executes top-level tasks: Tries to poll |
1784 |
* each queue starting at a random index with random stride, |
1785 |
* returning source id or retry indicator. |
1786 |
* |
1787 |
* @param w caller's WorkQueue |
1788 |
* @param prevSrc the previous queue stolen from in current phase, or 0 |
1789 |
* @param r random seed |
1790 |
* @return id of queue if taken, negative if none found, prevSrc for retry |
1791 |
*/ |
1792 |
private int scan(WorkQueue w, int prevSrc, int r) { |
1793 |
WorkQueue[] qs = queues; |
1794 |
int n = (w == null || qs == null) ? 0 : qs.length; |
1795 |
for (int step = (r >>> 16) | 1, i = n; i > 0; --i, r += step) { |
1796 |
int j, cap; WorkQueue q; ForkJoinTask<?>[] a; |
1797 |
if ((q = qs[j = r & (n - 1)]) != null && |
1798 |
(a = q.array) != null && (cap = a.length) > 0) { |
1799 |
int src = j | SRC, b = q.base; |
1800 |
int k = (cap - 1) & b, nb = b + 1, nk = (cap - 1) & nb; |
1801 |
ForkJoinTask<?> t = a[k]; |
1802 |
U.loadFence(); // for re-reads |
1803 |
if (q.base != b) // inconsistent |
1804 |
return prevSrc; |
1805 |
else if (t != null && WorkQueue.casSlotToNull(a, k, t)) { |
1806 |
q.base = nb; |
1807 |
w.source = src; |
1808 |
if (prevSrc == 0 && q.base == nb && a[nk] != null) |
1809 |
signalWork(); // propagate |
1810 |
w.topLevelExec(t, q); |
1811 |
return src; |
1812 |
} |
1813 |
else if (q.array != a || a[k] != null || a[nk] != null) |
1814 |
return prevSrc; // revisit |
1815 |
} |
1816 |
} |
1817 |
return -1; |
1818 |
} |
1819 |
|
1820 |
/** |
1821 |
* Advances phase, enqueues, and awaits signal or termination. |
1822 |
* |
1823 |
* @return negative if terminated, else 0 |
1824 |
*/ |
1825 |
private int awaitWork(WorkQueue w) { |
1826 |
if (w == null) |
1827 |
return -1; // currently impossible |
1828 |
int p = (w.phase + SS_SEQ) & ~INACTIVE; // advance phase |
1829 |
boolean idle = false; // true if possibly quiescent |
1830 |
if (runState < 0) |
1831 |
return -1; // terminating |
1832 |
long sp = p & SP_MASK, pc = ctl, qc; |
1833 |
w.phase = p | INACTIVE; |
1834 |
do { // enqueue |
1835 |
w.stackPred = (int)pc; // set ctl stack link |
1836 |
} while (pc != (pc = compareAndExchangeCtl( |
1837 |
pc, qc = ((pc - RC_UNIT) & UC_MASK) | sp))); |
1838 |
if ((qc & RC_MASK) <= 0L) { |
1839 |
if (hasTasks(true) && (w.phase >= 0 || reactivate() == w)) |
1840 |
return 0; // check for stragglers |
1841 |
if (runState != 0 && tryTerminate(false, false)) |
1842 |
return -1; // quiescent termination |
1843 |
idle = true; |
1844 |
} |
1845 |
WorkQueue[] qs = queues; // spin for expected #accesses in scan+signal |
1846 |
int spins = ((qs == null) ? 0 : ((qs.length & SMASK) << 1)) | 0xf; |
1847 |
while ((p = w.phase) < 0 && --spins > 0) |
1848 |
Thread.onSpinWait(); |
1849 |
if (p < 0) { |
1850 |
long deadline = idle ? keepAlive + System.currentTimeMillis() : 0L; |
1851 |
LockSupport.setCurrentBlocker(this); |
1852 |
for (;;) { // await signal or termination |
1853 |
if (runState < 0) |
1854 |
return -1; |
1855 |
w.access = PARKED; // enable unpark |
1856 |
if (w.phase < 0) { |
1857 |
if (idle) |
1858 |
LockSupport.parkUntil(deadline); |
1859 |
else |
1860 |
LockSupport.park(); |
1861 |
} |
1862 |
w.access = 0; // disable unpark |
1863 |
if (w.phase >= 0) { |
1864 |
LockSupport.setCurrentBlocker(null); |
1865 |
break; |
1866 |
} |
1867 |
Thread.interrupted(); // clear status for next park |
1868 |
if (idle) { // check for idle timeout |
1869 |
if (deadline - System.currentTimeMillis() < TIMEOUT_SLOP) { |
1870 |
if (tryTrim(w)) |
1871 |
return -1; |
1872 |
else // not at head; restart timer |
1873 |
deadline += keepAlive; |
1874 |
} |
1875 |
} |
1876 |
} |
1877 |
} |
1878 |
return 0; |
1879 |
} |
1880 |
|
1881 |
/** |
1882 |
* Non-overridable version of isQuiescent. Returns true if |
1883 |
* quiescent or already terminating. |
1884 |
*/ |
1885 |
private boolean canStop() { |
1886 |
long c = ctl; |
1887 |
do { |
1888 |
if (runState < 0) |
1889 |
break; |
1890 |
if ((c & RC_MASK) > 0L || hasTasks(false)) |
1891 |
return false; |
1892 |
} while (c != (c = ctl)); // validate |
1893 |
return true; |
1894 |
} |
1895 |
|
1896 |
/** |
1897 |
* Scans for and returns a polled task, if available. Used only |
1898 |
* for untracked polls. Begins scan at a random index to avoid |
1899 |
* systematic unfairness. |
1900 |
* |
1901 |
* @param submissionsOnly if true, only scan submission queues |
1902 |
*/ |
1903 |
private ForkJoinTask<?> pollScan(boolean submissionsOnly) { |
1904 |
int r = ThreadLocalRandom.nextSecondarySeed(); |
1905 |
if (submissionsOnly) // even indices only |
1906 |
r &= ~1; |
1907 |
int step = (submissionsOnly) ? 2 : 1; |
1908 |
WorkQueue[] qs; int n; WorkQueue q; ForkJoinTask<?> t; |
1909 |
if (runState >= 0 && (qs = queues) != null && (n = qs.length) > 0) { |
1910 |
for (int i = n; i > 0; i -= step, r += step) { |
1911 |
if ((q = qs[r & (n - 1)]) != null && |
1912 |
(t = q.poll(this)) != null) |
1913 |
return t; |
1914 |
} |
1915 |
} |
1916 |
return null; |
1917 |
} |
1918 |
|
1919 |
/** |
1920 |
* Tries to decrement counts (sometimes implicitly) and possibly |
1921 |
* arrange for a compensating worker in preparation for |
1922 |
* blocking. May fail due to interference, in which case -1 is |
1923 |
* returned so caller may retry. A zero return value indicates |
1924 |
* that the caller doesn't need to re-adjust counts when later |
1925 |
* unblocked. |
1926 |
* |
1927 |
* @param c incoming ctl value |
1928 |
* @param canSaturate to override saturate predicate |
1929 |
* @return UNCOMPENSATE: block then adjust, 0: block, -1 : retry |
1930 |
*/ |
1931 |
private int tryCompensate(long c, boolean canSaturate) { |
1932 |
Predicate<? super ForkJoinPool> sat; |
1933 |
long b = bounds; // unpack fields |
1934 |
int pc = parallelism; |
1935 |
int minActive = (short)(b & SMASK), |
1936 |
maxTotal = (short)(b >>> SWIDTH) + pc, |
1937 |
active = (short)(c >>> RC_SHIFT), |
1938 |
total = (short)(c >>> TC_SHIFT), |
1939 |
sp = (int)c & ~INACTIVE; |
1940 |
if (sp != 0 && active <= pc) { // activate idle worker |
1941 |
WorkQueue[] qs; WorkQueue v; int i; |
1942 |
if (ctl == c && (qs = queues) != null && |
1943 |
qs.length > (i = sp & SMASK) && (v = qs[i]) != null) { |
1944 |
long nc = (v.stackPred & SP_MASK) | (UC_MASK & c); |
1945 |
if (compareAndSetCtl(c, nc)) { |
1946 |
v.phase = sp; |
1947 |
LockSupport.unpark(v.owner); |
1948 |
return UNCOMPENSATE; |
1949 |
} |
1950 |
} |
1951 |
return -1; // retry |
1952 |
} |
1953 |
else if (active > minActive && total >= pc) { // reduce active workers |
1954 |
long nc = ((RC_MASK & (c - RC_UNIT)) | (~RC_MASK & c)); |
1955 |
return compareAndSetCtl(c, nc) ? UNCOMPENSATE : -1; |
1956 |
} |
1957 |
else if (total < maxTotal && total < MAX_CAP) { // expand pool |
1958 |
long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK); |
1959 |
return (!compareAndSetCtl(c, nc) ? -1 : |
1960 |
!createWorker() ? 0 : UNCOMPENSATE); |
1961 |
} |
1962 |
else if (!compareAndSetCtl(c, c)) // validate |
1963 |
return -1; |
1964 |
else if (canSaturate || ((sat = saturate) != null && sat.test(this))) |
1965 |
return 0; |
1966 |
else |
1967 |
throw new RejectedExecutionException( |
1968 |
"Thread limit exceeded replacing blocked worker"); |
1969 |
} |
1970 |
|
1971 |
/** |
1972 |
* Readjusts RC count; called from ForkJoinTask after blocking. |
1973 |
*/ |
1974 |
final void uncompensate() { |
1975 |
getAndAddCtl(RC_UNIT); |
1976 |
} |
1977 |
|
1978 |
/** |
1979 |
* Helps if possible until the given task is done. Processes |
1980 |
* compatible local tasks and scans other queues for task produced |
1981 |
* by w's stealers; returning compensated blocking sentinel if |
1982 |
* none are found. |
1983 |
* |
1984 |
* @param task the task |
1985 |
* @param w caller's WorkQueue |
1986 |
* @param timed true if this is a timed join |
1987 |
* @return task status on exit, or UNCOMPENSATE for compensated blocking |
1988 |
*/ |
1989 |
final int helpJoin(ForkJoinTask<?> task, WorkQueue w, boolean timed) { |
1990 |
if (w == null || task == null) |
1991 |
return 0; |
1992 |
int wsrc = w.source, wid = (w.config & SMASK) | SRC, r = wid + 2; |
1993 |
long sctl = 0L; // track stability |
1994 |
for (boolean rescan = true;;) { |
1995 |
int s; WorkQueue[] qs; |
1996 |
if ((s = task.status) < 0) |
1997 |
return s; |
1998 |
if (!rescan && sctl == (sctl = ctl)) { |
1999 |
if (runState < 0) |
2000 |
return 0; |
2001 |
if ((s = tryCompensate(sctl, timed)) >= 0) |
2002 |
return s; // block |
2003 |
} |
2004 |
rescan = false; |
2005 |
int n = ((qs = queues) == null) ? 0 : qs.length, m = n - 1; |
2006 |
scan: for (int i = n >>> 1; i > 0; --i, r += 2) { |
2007 |
int j, cap; WorkQueue q; ForkJoinTask<?>[] a; |
2008 |
if ((q = qs[j = r & m]) != null && (a = q.array) != null && |
2009 |
(cap = a.length) > 0) { |
2010 |
for (int src = j | SRC;;) { |
2011 |
int sq = q.source, b = q.base; |
2012 |
int k = (cap - 1) & b, nb = b + 1; |
2013 |
ForkJoinTask<?> t = a[k]; |
2014 |
U.loadFence(); // for re-reads |
2015 |
boolean eligible = true; // check steal chain |
2016 |
for (int d = n, v = sq;;) { // may be cyclic; bound |
2017 |
WorkQueue p; |
2018 |
if (v == wid) |
2019 |
break; |
2020 |
if (v == 0 || --d == 0 || (p = qs[v & m]) == null) { |
2021 |
eligible = false; |
2022 |
break; |
2023 |
} |
2024 |
v = p.source; |
2025 |
} |
2026 |
if (q.source != sq || q.base != b) |
2027 |
; // stale |
2028 |
else if ((s = task.status) < 0) |
2029 |
return s; // recheck before taking |
2030 |
else if (t == null) { |
2031 |
if (a[k] == null) { |
2032 |
if (!rescan && eligible && |
2033 |
(q.array != a || q.top != b)) |
2034 |
rescan = true; // resized or stalled |
2035 |
break; |
2036 |
} |
2037 |
} |
2038 |
else if (t != task && !eligible) |
2039 |
break; |
2040 |
else if (WorkQueue.casSlotToNull(a, k, t)) { |
2041 |
q.base = nb; |
2042 |
w.source = src; |
2043 |
t.doExec(); |
2044 |
w.source = wsrc; |
2045 |
rescan = true; |
2046 |
break scan; |
2047 |
} |
2048 |
} |
2049 |
} |
2050 |
} |
2051 |
} |
2052 |
} |
2053 |
|
2054 |
/** |
2055 |
* Version of helpJoin for CountedCompleters. |
2056 |
* |
2057 |
* @param task the task |
2058 |
* @param w caller's WorkQueue |
2059 |
* @param owned true if w is owned by a ForkJoinWorkerThread |
2060 |
* @param timed true if this is a timed join |
2061 |
* @return task status on exit, or UNCOMPENSATE for compensated blocking |
2062 |
*/ |
2063 |
final int helpComplete(ForkJoinTask<?> task, WorkQueue w, boolean owned, |
2064 |
boolean timed) { |
2065 |
if (w == null || task == null) |
2066 |
return 0; |
2067 |
int wsrc = w.source, r = w.config; |
2068 |
long sctl = 0L; // track stability |
2069 |
for (boolean rescan = true;;) { |
2070 |
int s; WorkQueue[] qs; |
2071 |
if ((s = w.helpComplete(task, owned, 0)) < 0) |
2072 |
return s; |
2073 |
if (!rescan && sctl == (sctl = ctl)) { |
2074 |
if (!owned || runState < 0) |
2075 |
return 0; |
2076 |
if ((s = tryCompensate(sctl, timed)) >= 0) |
2077 |
return s; |
2078 |
} |
2079 |
rescan = false; |
2080 |
int n = ((qs = queues) == null) ? 0 : qs.length, m = n - 1; |
2081 |
scan: for (int i = n; i > 0; --i, ++r) { |
2082 |
int j, cap; WorkQueue q; ForkJoinTask<?>[] a; |
2083 |
if ((q = qs[j = r & m]) != null && (a = q.array) != null && |
2084 |
(cap = a.length) > 0) { |
2085 |
poll: for (int src = j | SRC, b = q.base;;) { |
2086 |
int k = (cap - 1) & b, nb = b + 1; |
2087 |
ForkJoinTask<?> t = a[k]; |
2088 |
U.loadFence(); // for re-reads |
2089 |
if (b != (b = q.base)) |
2090 |
; // stale |
2091 |
else if ((s = task.status) < 0) |
2092 |
return s; // recheck before taking |
2093 |
else if (t == null) { |
2094 |
if (a[k] == null) { |
2095 |
if (!rescan && // resized or stalled |
2096 |
(q.array != a || q.top != b)) |
2097 |
rescan = true; |
2098 |
break; |
2099 |
} |
2100 |
} |
2101 |
else if (t instanceof CountedCompleter) { |
2102 |
CountedCompleter<?> f; |
2103 |
for (f = (CountedCompleter<?>)t;;) { |
2104 |
if (f == task) |
2105 |
break; |
2106 |
else if ((f = f.completer) == null) |
2107 |
break poll; // ineligible |
2108 |
} |
2109 |
if (WorkQueue.casSlotToNull(a, k, t)) { |
2110 |
q.base = nb; |
2111 |
w.source = src; |
2112 |
t.doExec(); |
2113 |
w.source = wsrc; |
2114 |
rescan = true; |
2115 |
break scan; |
2116 |
} |
2117 |
} |
2118 |
else |
2119 |
break; |
2120 |
} |
2121 |
} |
2122 |
} |
2123 |
} |
2124 |
} |
2125 |
|
2126 |
/** |
2127 |
* Runs tasks until {@code isQuiescent()}. Rather than blocking |
2128 |
* when tasks cannot be found, rescans until all others cannot |
2129 |
* find tasks either. |
2130 |
* |
2131 |
* @param nanos max wait time (Long.MAX_VALUE if effectively untimed) |
2132 |
* @param interruptible true if return on interrupt |
2133 |
* @return positive if quiescent, negative if interrupted, else 0 |
2134 |
*/ |
2135 |
private int helpQuiesce(WorkQueue w, long nanos, boolean interruptible) { |
2136 |
long startTime = System.nanoTime(), parkTime = 0L; |
2137 |
int phase; // w.phase set negative when temporarily quiescent |
2138 |
if (w == null || (phase = w.phase) < 0) |
2139 |
return 0; |
2140 |
int activePhase = phase, inactivePhase = phase | INACTIVE; |
2141 |
int wsrc = w.source, r = 0; |
2142 |
for (boolean locals = true;;) { |
2143 |
WorkQueue[] qs; WorkQueue q; |
2144 |
if (runState < 0) { // terminating |
2145 |
w.phase = activePhase; |
2146 |
return 1; |
2147 |
} |
2148 |
if (locals) { // run local tasks before (re)polling |
2149 |
for (ForkJoinTask<?> u; (u = w.nextLocalTask()) != null;) |
2150 |
u.doExec(); |
2151 |
} |
2152 |
boolean rescan = false, busy = locals = false, interrupted; |
2153 |
int n = ((qs = queues) == null) ? 0 : qs.length, m = n - 1; |
2154 |
scan: for (int i = n, j; i > 0; --i, ++r) { |
2155 |
if ((q = qs[j = m & r]) != null && q != w) { |
2156 |
for (int src = j | SRC;;) { |
2157 |
ForkJoinTask<?>[] a = q.array; |
2158 |
int b = q.base, cap; |
2159 |
if (a == null || (cap = a.length) <= 0) |
2160 |
break; |
2161 |
int k = (cap - 1) & b, nb = b + 1, nk = (cap - 1) & nb; |
2162 |
ForkJoinTask<?> t = a[k]; |
2163 |
U.loadFence(); // for re-reads |
2164 |
if (q.base != b || q.array != a || a[k] != t) |
2165 |
; |
2166 |
else if (t == null) { |
2167 |
if (!rescan) { |
2168 |
if (a[nk] != null || q.top - b > 0) |
2169 |
rescan = true; |
2170 |
else if (!busy && |
2171 |
q.owner != null && q.phase >= 0) |
2172 |
busy = true; |
2173 |
} |
2174 |
break; |
2175 |
} |
2176 |
else if (phase < 0) // reactivate before taking |
2177 |
w.phase = phase = activePhase; |
2178 |
else if (WorkQueue.casSlotToNull(a, k, t)) { |
2179 |
q.base = nb; |
2180 |
w.source = src; |
2181 |
t.doExec(); |
2182 |
w.source = wsrc; |
2183 |
rescan = locals = true; |
2184 |
break scan; |
2185 |
} |
2186 |
} |
2187 |
} |
2188 |
} |
2189 |
if (rescan) |
2190 |
; // retry |
2191 |
else if (phase >= 0) { |
2192 |
parkTime = 0L; |
2193 |
w.phase = phase = inactivePhase; |
2194 |
} |
2195 |
else if (!busy) { |
2196 |
w.phase = activePhase; |
2197 |
return 1; |
2198 |
} |
2199 |
else if (parkTime == 0L) { |
2200 |
parkTime = 1L << 10; // initially about 1 usec |
2201 |
Thread.yield(); |
2202 |
} |
2203 |
else if ((interrupted = interruptible && Thread.interrupted()) || |
2204 |
System.nanoTime() - startTime > nanos) { |
2205 |
w.phase = activePhase; |
2206 |
return interrupted ? -1 : 0; |
2207 |
} |
2208 |
else { |
2209 |
LockSupport.parkNanos(this, parkTime); |
2210 |
if (parkTime < nanos >>> 8 && parkTime < 1L << 20) |
2211 |
parkTime <<= 1; // max sleep approx 1 sec or 1% nanos |
2212 |
} |
2213 |
} |
2214 |
} |
2215 |
|
2216 |
/** |
2217 |
* Helps quiesce from external caller until done, interrupted, or timeout |
2218 |
* |
2219 |
* @param nanos max wait time (Long.MAX_VALUE if effectively untimed) |
2220 |
* @param interruptible true if return on interrupt |
2221 |
* @return positive if quiescent, negative if interrupted, else 0 |
2222 |
*/ |
2223 |
private int externalHelpQuiesce(long nanos, boolean interruptible) { |
2224 |
for (long startTime = System.nanoTime(), parkTime = 0L;;) { |
2225 |
ForkJoinTask<?> t; |
2226 |
if ((t = pollScan(false)) != null) { |
2227 |
t.doExec(); |
2228 |
parkTime = 0L; |
2229 |
} |
2230 |
else if (canStop()) |
2231 |
return 1; |
2232 |
else if (parkTime == 0L) { |
2233 |
parkTime = 1L << 10; |
2234 |
Thread.yield(); |
2235 |
} |
2236 |
else if ((System.nanoTime() - startTime) > nanos) |
2237 |
return 0; |
2238 |
else if (interruptible && Thread.interrupted()) |
2239 |
return -1; |
2240 |
else { |
2241 |
LockSupport.parkNanos(this, parkTime); |
2242 |
if (parkTime < nanos >>> 8 && parkTime < 1L << 20) |
2243 |
parkTime <<= 1; |
2244 |
} |
2245 |
} |
2246 |
} |
2247 |
|
2248 |
/** |
2249 |
* Helps quiesce from either internal or external caller |
2250 |
* |
2251 |
* @param pool the pool to use, or null if any |
2252 |
* @param nanos max wait time (Long.MAX_VALUE if effectively untimed) |
2253 |
* @param interruptible true if return on interrupt |
2254 |
* @return positive if quiescent, negative if interrupted, else 0 |
2255 |
*/ |
2256 |
static final int helpQuiescePool(ForkJoinPool pool, long nanos, |
2257 |
boolean interruptible) { |
2258 |
Thread t; ForkJoinPool p; ForkJoinWorkerThread wt; |
2259 |
if ((t = Thread.currentThread()) instanceof ForkJoinWorkerThread && |
2260 |
(p = (wt = (ForkJoinWorkerThread)t).pool) != null && |
2261 |
(p == pool || pool == null)) |
2262 |
return p.helpQuiesce(wt.workQueue, nanos, interruptible); |
2263 |
else if ((p = pool) != null || (p = common) != null) |
2264 |
return p.externalHelpQuiesce(nanos, interruptible); |
2265 |
else |
2266 |
return 0; |
2267 |
} |
2268 |
|
2269 |
/** |
2270 |
* Gets and removes a local or stolen task for the given worker. |
2271 |
* |
2272 |
* @return a task, if available |
2273 |
*/ |
2274 |
final ForkJoinTask<?> nextTaskFor(WorkQueue w) { |
2275 |
ForkJoinTask<?> t; |
2276 |
if (w == null || (t = w.nextLocalTask()) == null) |
2277 |
t = pollScan(false); |
2278 |
return t; |
2279 |
} |
2280 |
|
2281 |
// External operations |
2282 |
|
2283 |
/** |
2284 |
* Finds and locks a WorkQueue for an external submitter, or |
2285 |
* throws RejectedExecutionException if shutdown or terminating. |
2286 |
* @param isSubmit false if this is for a common pool fork |
2287 |
*/ |
2288 |
final WorkQueue submissionQueue(boolean isSubmit) { |
2289 |
int r; |
2290 |
ReentrantLock lock = registrationLock; |
2291 |
if ((r = ThreadLocalRandom.getProbe()) == 0) { |
2292 |
ThreadLocalRandom.localInit(); // initialize caller's probe |
2293 |
r = ThreadLocalRandom.getProbe(); |
2294 |
} |
2295 |
if (lock != null) { // else init error |
2296 |
for (int id = r << 1;;) { // even indices only |
2297 |
int n, i; WorkQueue[] qs; WorkQueue q; |
2298 |
if ((qs = queues) == null || (n = qs.length) <= 0) |
2299 |
break; |
2300 |
else if ((q = qs[i = (n - 1) & id]) == null) { |
2301 |
WorkQueue w = new WorkQueue(null, id | SRC); |
2302 |
w.array = new ForkJoinTask<?>[INITIAL_QUEUE_CAPACITY]; |
2303 |
lock.lock(); // install under lock |
2304 |
if (queues == qs && qs[i] == null) |
2305 |
qs[i] = w; // else lost race; discard |
2306 |
lock.unlock(); |
2307 |
} |
2308 |
else if (q.getAndSetAccess(1) != 0) // move and restart |
2309 |
id = (r = ThreadLocalRandom.advanceProbe(r)) << 1; |
2310 |
else if (isSubmit && runState != 0) { |
2311 |
q.access = 0; // check while lock held |
2312 |
break; |
2313 |
} |
2314 |
else |
2315 |
return q; |
2316 |
} |
2317 |
} |
2318 |
throw new RejectedExecutionException(); |
2319 |
} |
2320 |
|
2321 |
/** |
2322 |
* Pushes a submission to the pool, using internal queue if called |
2323 |
* from ForkJoinWorkerThread, else external queue. |
2324 |
*/ |
2325 |
private <T> ForkJoinTask<T> poolSubmit(boolean signalIfEmpty, |
2326 |
ForkJoinTask<T> task) { |
2327 |
WorkQueue q; Thread t; ForkJoinWorkerThread wt; |
2328 |
U.storeStoreFence(); // ensure safely publishable |
2329 |
if (task == null) throw new NullPointerException(); |
2330 |
if (((t = Thread.currentThread()) instanceof ForkJoinWorkerThread) && |
2331 |
(wt = (ForkJoinWorkerThread)t).pool == this) |
2332 |
q = wt.workQueue; |
2333 |
else { |
2334 |
task.markPoolSubmission(); |
2335 |
q = submissionQueue(true); |
2336 |
} |
2337 |
q.push(task, this, signalIfEmpty); |
2338 |
return task; |
2339 |
} |
2340 |
|
2341 |
/** |
2342 |
* Returns queue for an external thread, if one exists that has |
2343 |
* possibly ever submitted to the given pool (nonzero probe), or |
2344 |
* null if none. |
2345 |
*/ |
2346 |
private static WorkQueue externalQueue(ForkJoinPool p) { |
2347 |
WorkQueue[] qs; |
2348 |
int r = ThreadLocalRandom.getProbe(), n; |
2349 |
return (p != null && (qs = p.queues) != null && |
2350 |
(n = qs.length) > 0 && r != 0) ? |
2351 |
qs[(n - 1) & (r << 1)] : null; |
2352 |
} |
2353 |
|
2354 |
/** |
2355 |
* Returns external queue for common pool. |
2356 |
*/ |
2357 |
static WorkQueue commonQueue() { |
2358 |
return externalQueue(common); |
2359 |
} |
2360 |
|
2361 |
/** |
2362 |
* Returns queue for an external thread, if one exists |
2363 |
*/ |
2364 |
final WorkQueue externalQueue() { |
2365 |
return externalQueue(this); |
2366 |
} |
2367 |
|
2368 |
/** |
2369 |
* If the given executor is a ForkJoinPool, poll and execute |
2370 |
* AsynchronousCompletionTasks from worker's queue until none are |
2371 |
* available or blocker is released. |
2372 |
*/ |
2373 |
static void helpAsyncBlocker(Executor e, ManagedBlocker blocker) { |
2374 |
WorkQueue w = null; Thread t; ForkJoinWorkerThread wt; |
2375 |
if ((t = Thread.currentThread()) instanceof ForkJoinWorkerThread) { |
2376 |
if ((wt = (ForkJoinWorkerThread)t).pool == e) |
2377 |
w = wt.workQueue; |
2378 |
} |
2379 |
else if (e instanceof ForkJoinPool) |
2380 |
w = ((ForkJoinPool)e).externalQueue(); |
2381 |
if (w != null) |
2382 |
w.helpAsyncBlocker(blocker); |
2383 |
} |
2384 |
|
2385 |
/** |
2386 |
* Returns a cheap heuristic guide for task partitioning when |
2387 |
* programmers, frameworks, tools, or languages have little or no |
2388 |
* idea about task granularity. In essence, by offering this |
2389 |
* method, we ask users only about tradeoffs in overhead vs |
2390 |
* expected throughput and its variance, rather than how finely to |
2391 |
* partition tasks. |
2392 |
* |
2393 |
* In a steady state strict (tree-structured) computation, each |
2394 |
* thread makes available for stealing enough tasks for other |
2395 |
* threads to remain active. Inductively, if all threads play by |
2396 |
* the same rules, each thread should make available only a |
2397 |
* constant number of tasks. |
2398 |
* |
2399 |
* The minimum useful constant is just 1. But using a value of 1 |
2400 |
* would require immediate replenishment upon each steal to |
2401 |
* maintain enough tasks, which is infeasible. Further, |
2402 |
* partitionings/granularities of offered tasks should minimize |
2403 |
* steal rates, which in general means that threads nearer the top |
2404 |
* of computation tree should generate more than those nearer the |
2405 |
* bottom. In perfect steady state, each thread is at |
2406 |
* approximately the same level of computation tree. However, |
2407 |
* producing extra tasks amortizes the uncertainty of progress and |
2408 |
* diffusion assumptions. |
2409 |
* |
2410 |
* So, users will want to use values larger (but not much larger) |
2411 |
* than 1 to both smooth over transient shortages and hedge |
2412 |
* against uneven progress; as traded off against the cost of |
2413 |
* extra task overhead. We leave the user to pick a threshold |
2414 |
* value to compare with the results of this call to guide |
2415 |
* decisions, but recommend values such as 3. |
2416 |
* |
2417 |
* When all threads are active, it is on average OK to estimate |
2418 |
* surplus strictly locally. In steady-state, if one thread is |
2419 |
* maintaining say 2 surplus tasks, then so are others. So we can |
2420 |
* just use estimated queue length. However, this strategy alone |
2421 |
* leads to serious mis-estimates in some non-steady-state |
2422 |
* conditions (ramp-up, ramp-down, other stalls). We can detect |
2423 |
* many of these by further considering the number of "idle" |
2424 |
* threads, that are known to have zero queued tasks, so |
2425 |
* compensate by a factor of (#idle/#active) threads. |
2426 |
*/ |
2427 |
static int getSurplusQueuedTaskCount() { |
2428 |
Thread t; ForkJoinWorkerThread wt; ForkJoinPool pool; WorkQueue q; |
2429 |
if (((t = Thread.currentThread()) instanceof ForkJoinWorkerThread) && |
2430 |
(pool = (wt = (ForkJoinWorkerThread)t).pool) != null && |
2431 |
(q = wt.workQueue) != null) { |
2432 |
int n = q.top - q.base; |
2433 |
int p = pool.parallelism; |
2434 |
int a = (short)(pool.ctl >>> RC_SHIFT); |
2435 |
return n - (a > (p >>>= 1) ? 0 : |
2436 |
a > (p >>>= 1) ? 1 : |
2437 |
a > (p >>>= 1) ? 2 : |
2438 |
a > (p >>>= 1) ? 4 : |
2439 |
8); |
2440 |
} |
2441 |
return 0; |
2442 |
} |
2443 |
|
2444 |
// Termination |
2445 |
|
2446 |
/** |
2447 |
* Possibly initiates and/or completes pool termination. |
2448 |
* |
2449 |
* @param now if true, unconditionally terminate, else only |
2450 |
* if no work and no active workers |
2451 |
* @param enable if true, terminate when next possible |
2452 |
* @return true if terminating or terminated |
2453 |
*/ |
2454 |
private boolean tryTerminate(boolean now, boolean enable) { |
2455 |
int rs; ReentrantLock lock; Condition cond; |
2456 |
if ((rs = runState) >= 0) { // set SHUTDOWN and/or STOP |
2457 |
if ((config & ISCOMMON) != 0) |
2458 |
return false; // cannot shutdown |
2459 |
if (!now) { |
2460 |
if ((rs & SHUTDOWN) == 0) { |
2461 |
if (!enable) |
2462 |
return false; |
2463 |
getAndBitwiseOrRunState(SHUTDOWN); |
2464 |
} |
2465 |
if (!canStop()) |
2466 |
return false; |
2467 |
} |
2468 |
getAndBitwiseOrRunState(SHUTDOWN | STOP); |
2469 |
} |
2470 |
WorkQueue released = reactivate(); // try signalling waiter |
2471 |
int tc = (short)(ctl >>> TC_SHIFT); |
2472 |
if (released == null && tc > 0) { // help unblock and cancel |
2473 |
Thread current = Thread.currentThread(); |
2474 |
WorkQueue w = ((current instanceof ForkJoinWorkerThread) ? |
2475 |
((ForkJoinWorkerThread)current).workQueue : null); |
2476 |
int r = (w == null) ? 0 : w.config + 1; // stagger traversals |
2477 |
WorkQueue[] qs = queues; |
2478 |
int n = (qs == null) ? 0 : qs.length; |
2479 |
for (int i = 0; i < n; ++i) { |
2480 |
WorkQueue q; Thread thread; |
2481 |
if ((q = qs[(r + i) & (n - 1)]) != null && |
2482 |
(thread = q.owner) != current && q.access != STOP) { |
2483 |
for (ForkJoinTask<?> t; (t = q.poll(null)) != null; ) |
2484 |
ForkJoinTask.cancelIgnoringExceptions(t); |
2485 |
if (thread != null && !thread.isInterrupted()) { |
2486 |
q.forcePhaseActive(); // for awaitWork |
2487 |
try { |
2488 |
thread.interrupt(); |
2489 |
} catch (Throwable ignore) { |
2490 |
} |
2491 |
} |
2492 |
} |
2493 |
} |
2494 |
} |
2495 |
if ((tc <= 0 || (short)(ctl >>> TC_SHIFT) <= 0) && |
2496 |
(getAndBitwiseOrRunState(TERMINATED) & TERMINATED) == 0 && |
2497 |
(lock = registrationLock) != null) { |
2498 |
lock.lock(); // signal when no workers |
2499 |
if ((cond = termination) != null) |
2500 |
cond.signalAll(); |
2501 |
lock.unlock(); |
2502 |
// container.close(); // for loom |
2503 |
} |
2504 |
return true; |
2505 |
} |
2506 |
|
2507 |
// Exported methods |
2508 |
|
2509 |
// Constructors |
2510 |
|
2511 |
/** |
2512 |
* Creates a {@code ForkJoinPool} with parallelism equal to {@link |
2513 |
* java.lang.Runtime#availableProcessors}, using defaults for all |
2514 |
* other parameters (see {@link #ForkJoinPool(int, |
2515 |
* ForkJoinWorkerThreadFactory, UncaughtExceptionHandler, boolean, |
2516 |
* int, int, int, Predicate, long, TimeUnit)}). |
2517 |
* |
2518 |
* @throws SecurityException if a security manager exists and |
2519 |
* the caller is not permitted to modify threads |
2520 |
* because it does not hold {@link |
2521 |
* java.lang.RuntimePermission}{@code ("modifyThread")} |
2522 |
*/ |
2523 |
public ForkJoinPool() { |
2524 |
this(Math.min(MAX_CAP, Runtime.getRuntime().availableProcessors()), |
2525 |
defaultForkJoinWorkerThreadFactory, null, false, |
2526 |
0, MAX_CAP, 1, null, DEFAULT_KEEPALIVE, TimeUnit.MILLISECONDS); |
2527 |
} |
2528 |
|
2529 |
/** |
2530 |
* Creates a {@code ForkJoinPool} with the indicated parallelism |
2531 |
* level, using defaults for all other parameters (see {@link |
2532 |
* #ForkJoinPool(int, ForkJoinWorkerThreadFactory, |
2533 |
* UncaughtExceptionHandler, boolean, int, int, int, Predicate, |
2534 |
* long, TimeUnit)}). |
2535 |
* |
2536 |
* @param parallelism the parallelism level |
2537 |
* @throws IllegalArgumentException if parallelism less than or |
2538 |
* equal to zero, or greater than implementation limit |
2539 |
* @throws SecurityException if a security manager exists and |
2540 |
* the caller is not permitted to modify threads |
2541 |
* because it does not hold {@link |
2542 |
* java.lang.RuntimePermission}{@code ("modifyThread")} |
2543 |
*/ |
2544 |
public ForkJoinPool(int parallelism) { |
2545 |
this(parallelism, defaultForkJoinWorkerThreadFactory, null, false, |
2546 |
0, MAX_CAP, 1, null, DEFAULT_KEEPALIVE, TimeUnit.MILLISECONDS); |
2547 |
} |
2548 |
|
2549 |
/** |
2550 |
* Creates a {@code ForkJoinPool} with the given parameters (using |
2551 |
* defaults for others -- see {@link #ForkJoinPool(int, |
2552 |
* ForkJoinWorkerThreadFactory, UncaughtExceptionHandler, boolean, |
2553 |
* int, int, int, Predicate, long, TimeUnit)}). |
2554 |
* |
2555 |
* @param parallelism the parallelism level. For default value, |
2556 |
* use {@link java.lang.Runtime#availableProcessors}. |
2557 |
* @param factory the factory for creating new threads. For default value, |
2558 |
* use {@link #defaultForkJoinWorkerThreadFactory}. |
2559 |
* @param handler the handler for internal worker threads that |
2560 |
* terminate due to unrecoverable errors encountered while executing |
2561 |
* tasks. For default value, use {@code null}. |
2562 |
* @param asyncMode if true, |
2563 |
* establishes local first-in-first-out scheduling mode for forked |
2564 |
* tasks that are never joined. This mode may be more appropriate |
2565 |
* than default locally stack-based mode in applications in which |
2566 |
* worker threads only process event-style asynchronous tasks. |
2567 |
* For default value, use {@code false}. |
2568 |
* @throws IllegalArgumentException if parallelism less than or |
2569 |
* equal to zero, or greater than implementation limit |
2570 |
* @throws NullPointerException if the factory is null |
2571 |
* @throws SecurityException if a security manager exists and |
2572 |
* the caller is not permitted to modify threads |
2573 |
* because it does not hold {@link |
2574 |
* java.lang.RuntimePermission}{@code ("modifyThread")} |
2575 |
*/ |
2576 |
public ForkJoinPool(int parallelism, |
2577 |
ForkJoinWorkerThreadFactory factory, |
2578 |
UncaughtExceptionHandler handler, |
2579 |
boolean asyncMode) { |
2580 |
this(parallelism, factory, handler, asyncMode, |
2581 |
0, MAX_CAP, 1, null, DEFAULT_KEEPALIVE, TimeUnit.MILLISECONDS); |
2582 |
} |
2583 |
|
2584 |
/** |
2585 |
* Creates a {@code ForkJoinPool} with the given parameters. |
2586 |
* |
2587 |
* @param parallelism the parallelism level. For default value, |
2588 |
* use {@link java.lang.Runtime#availableProcessors}. |
2589 |
* |
2590 |
* @param factory the factory for creating new threads. For |
2591 |
* default value, use {@link #defaultForkJoinWorkerThreadFactory}. |
2592 |
* |
2593 |
* @param handler the handler for internal worker threads that |
2594 |
* terminate due to unrecoverable errors encountered while |
2595 |
* executing tasks. For default value, use {@code null}. |
2596 |
* |
2597 |
* @param asyncMode if true, establishes local first-in-first-out |
2598 |
* scheduling mode for forked tasks that are never joined. This |
2599 |
* mode may be more appropriate than default locally stack-based |
2600 |
* mode in applications in which worker threads only process |
2601 |
* event-style asynchronous tasks. For default value, use {@code |
2602 |
* false}. |
2603 |
* |
2604 |
* @param corePoolSize the number of threads to keep in the pool |
2605 |
* (unless timed out after an elapsed keep-alive). Normally (and |
2606 |
* by default) this is the same value as the parallelism level, |
2607 |
* but may be set to a larger value to reduce dynamic overhead if |
2608 |
* tasks regularly block. Using a smaller value (for example |
2609 |
* {@code 0}) has the same effect as the default. |
2610 |
* |
2611 |
* @param maximumPoolSize the maximum number of threads allowed. |
2612 |
* When the maximum is reached, attempts to replace blocked |
2613 |
* threads fail. (However, because creation and termination of |
2614 |
* different threads may overlap, and may be managed by the given |
2615 |
* thread factory, this value may be transiently exceeded.) To |
2616 |
* arrange the same value as is used by default for the common |
2617 |
* pool, use {@code 256} plus the {@code parallelism} level. (By |
2618 |
* default, the common pool allows a maximum of 256 spare |
2619 |
* threads.) Using a value (for example {@code |
2620 |
* Integer.MAX_VALUE}) larger than the implementation's total |
2621 |
* thread limit has the same effect as using this limit (which is |
2622 |
* the default). |
2623 |
* |
2624 |
* @param minimumRunnable the minimum allowed number of core |
2625 |
* threads not blocked by a join or {@link ManagedBlocker}. To |
2626 |
* ensure progress, when too few unblocked threads exist and |
2627 |
* unexecuted tasks may exist, new threads are constructed, up to |
2628 |
* the given maximumPoolSize. For the default value, use {@code |
2629 |
* 1}, that ensures liveness. A larger value might improve |
2630 |
* throughput in the presence of blocked activities, but might |
2631 |
* not, due to increased overhead. A value of zero may be |
2632 |
* acceptable when submitted tasks cannot have dependencies |
2633 |
* requiring additional threads. |
2634 |
* |
2635 |
* @param saturate if non-null, a predicate invoked upon attempts |
2636 |
* to create more than the maximum total allowed threads. By |
2637 |
* default, when a thread is about to block on a join or {@link |
2638 |
* ManagedBlocker}, but cannot be replaced because the |
2639 |
* maximumPoolSize would be exceeded, a {@link |
2640 |
* RejectedExecutionException} is thrown. But if this predicate |
2641 |
* returns {@code true}, then no exception is thrown, so the pool |
2642 |
* continues to operate with fewer than the target number of |
2643 |
* runnable threads, which might not ensure progress. |
2644 |
* |
2645 |
* @param keepAliveTime the elapsed time since last use before |
2646 |
* a thread is terminated (and then later replaced if needed). |
2647 |
* For the default value, use {@code 60, TimeUnit.SECONDS}. |
2648 |
* |
2649 |
* @param unit the time unit for the {@code keepAliveTime} argument |
2650 |
* |
2651 |
* @throws IllegalArgumentException if parallelism is less than or |
2652 |
* equal to zero, or is greater than implementation limit, |
2653 |
* or if maximumPoolSize is less than parallelism, |
2654 |
* of if the keepAliveTime is less than or equal to zero. |
2655 |
* @throws NullPointerException if the factory is null |
2656 |
* @throws SecurityException if a security manager exists and |
2657 |
* the caller is not permitted to modify threads |
2658 |
* because it does not hold {@link |
2659 |
* java.lang.RuntimePermission}{@code ("modifyThread")} |
2660 |
* @since 9 |
2661 |
*/ |
2662 |
public ForkJoinPool(int parallelism, |
2663 |
ForkJoinWorkerThreadFactory factory, |
2664 |
UncaughtExceptionHandler handler, |
2665 |
boolean asyncMode, |
2666 |
int corePoolSize, |
2667 |
int maximumPoolSize, |
2668 |
int minimumRunnable, |
2669 |
Predicate<? super ForkJoinPool> saturate, |
2670 |
long keepAliveTime, |
2671 |
TimeUnit unit) { |
2672 |
checkPermission(); |
2673 |
int p = parallelism; |
2674 |
if (p <= 0 || p > MAX_CAP || p > maximumPoolSize || keepAliveTime <= 0L) |
2675 |
throw new IllegalArgumentException(); |
2676 |
if (factory == null || unit == null) |
2677 |
throw new NullPointerException(); |
2678 |
this.parallelism = p; |
2679 |
this.factory = factory; |
2680 |
this.ueh = handler; |
2681 |
this.saturate = saturate; |
2682 |
this.config = asyncMode ? FIFO : 0; |
2683 |
this.keepAlive = Math.max(unit.toMillis(keepAliveTime), TIMEOUT_SLOP); |
2684 |
int corep = Math.min(Math.max(corePoolSize, p), MAX_CAP); |
2685 |
int maxSpares = Math.max(0, Math.min(maximumPoolSize - p, MAX_CAP)); |
2686 |
int minAvail = Math.max(0, Math.min(minimumRunnable, MAX_CAP)); |
2687 |
this.bounds = (long)(minAvail & SMASK) | (long)(maxSpares << SWIDTH) | |
2688 |
((long)corep << 32); |
2689 |
int size = 1 << (33 - Integer.numberOfLeadingZeros(p - 1)); |
2690 |
this.registrationLock = new ReentrantLock(); |
2691 |
this.queues = new WorkQueue[size]; |
2692 |
String pid = Integer.toString(getAndAddPoolIds(1) + 1); |
2693 |
String name = "ForkJoinPool-" + pid; |
2694 |
this.workerNamePrefix = name + "-worker-"; |
2695 |
// this.container = SharedThreadContainer.create(name); // for loom |
2696 |
} |
2697 |
|
2698 |
/** |
2699 |
* Constructor for common pool using parameters possibly |
2700 |
* overridden by system properties |
2701 |
*/ |
2702 |
private ForkJoinPool(byte forCommonPoolOnly) { |
2703 |
ForkJoinWorkerThreadFactory fac = defaultForkJoinWorkerThreadFactory; |
2704 |
UncaughtExceptionHandler handler = null; |
2705 |
int maxSpares = DEFAULT_COMMON_MAX_SPARES; |
2706 |
int pc = 0, preset = 0; // nonzero if size set as property |
2707 |
try { // ignore exceptions in accessing/parsing properties |
2708 |
String pp = System.getProperty |
2709 |
("java.util.concurrent.ForkJoinPool.common.parallelism"); |
2710 |
if (pp != null) { |
2711 |
pc = Math.max(0, Integer.parseInt(pp)); |
2712 |
preset = PRESET_SIZE; |
2713 |
} |
2714 |
String ms = System.getProperty |
2715 |
("java.util.concurrent.ForkJoinPool.common.maximumSpares"); |
2716 |
if (ms != null) |
2717 |
maxSpares = Math.max(0, Math.min(MAX_CAP, Integer.parseInt(ms))); |
2718 |
String sf = System.getProperty |
2719 |
("java.util.concurrent.ForkJoinPool.common.threadFactory"); |
2720 |
String sh = System.getProperty |
2721 |
("java.util.concurrent.ForkJoinPool.common.exceptionHandler"); |
2722 |
if (sf != null || sh != null) { |
2723 |
ClassLoader ldr = ClassLoader.getSystemClassLoader(); |
2724 |
if (sf != null) |
2725 |
fac = (ForkJoinWorkerThreadFactory) |
2726 |
ldr.loadClass(sf).getConstructor().newInstance(); |
2727 |
if (sh != null) |
2728 |
handler = (UncaughtExceptionHandler) |
2729 |
ldr.loadClass(sh).getConstructor().newInstance(); |
2730 |
} |
2731 |
} catch (Exception ignore) { |
2732 |
} |
2733 |
if (preset == 0) |
2734 |
pc = Math.max(1, Runtime.getRuntime().availableProcessors() - 1); |
2735 |
int p = Math.min(pc, MAX_CAP); |
2736 |
int size = (p == 0) ? 1 : 1 << (33 - Integer.numberOfLeadingZeros(p-1)); |
2737 |
this.parallelism = p; |
2738 |
this.config = ISCOMMON | preset; |
2739 |
this.bounds = (long)(1 | (maxSpares << SWIDTH)); |
2740 |
this.factory = fac; |
2741 |
this.ueh = handler; |
2742 |
this.keepAlive = DEFAULT_KEEPALIVE; |
2743 |
this.saturate = null; |
2744 |
this.workerNamePrefix = null; |
2745 |
this.registrationLock = new ReentrantLock(); |
2746 |
this.queues = new WorkQueue[size]; |
2747 |
// this.container = SharedThreadContainer.create("ForkJoinPool.commonPool"); // for loom |
2748 |
} |
2749 |
|
2750 |
/** |
2751 |
* Returns the common pool instance. This pool is statically |
2752 |
* constructed; its run state is unaffected by attempts to {@link |
2753 |
* #shutdown} or {@link #shutdownNow}. However this pool and any |
2754 |
* ongoing processing are automatically terminated upon program |
2755 |
* {@link System#exit}. Any program that relies on asynchronous |
2756 |
* task processing to complete before program termination should |
2757 |
* invoke {@code commonPool().}{@link #awaitQuiescence awaitQuiescence}, |
2758 |
* before exit. |
2759 |
* |
2760 |
* @return the common pool instance |
2761 |
* @since 1.8 |
2762 |
*/ |
2763 |
public static ForkJoinPool commonPool() { |
2764 |
// assert common != null : "static init error"; |
2765 |
return common; |
2766 |
} |
2767 |
|
2768 |
// Execution methods |
2769 |
|
2770 |
/** |
2771 |
* Performs the given task, returning its result upon completion. |
2772 |
* If the computation encounters an unchecked Exception or Error, |
2773 |
* it is rethrown as the outcome of this invocation. Rethrown |
2774 |
* exceptions behave in the same way as regular exceptions, but, |
2775 |
* when possible, contain stack traces (as displayed for example |
2776 |
* using {@code ex.printStackTrace()}) of both the current thread |
2777 |
* as well as the thread actually encountering the exception; |
2778 |
* minimally only the latter. |
2779 |
* |
2780 |
* @param task the task |
2781 |
* @param <T> the type of the task's result |
2782 |
* @return the task's result |
2783 |
* @throws NullPointerException if the task is null |
2784 |
* @throws RejectedExecutionException if the task cannot be |
2785 |
* scheduled for execution |
2786 |
*/ |
2787 |
public <T> T invoke(ForkJoinTask<T> task) { |
2788 |
poolSubmit(true, task); |
2789 |
return task.join(); |
2790 |
} |
2791 |
|
2792 |
/** |
2793 |
* Arranges for (asynchronous) execution of the given task. |
2794 |
* |
2795 |
* @param task the task |
2796 |
* @throws NullPointerException if the task is null |
2797 |
* @throws RejectedExecutionException if the task cannot be |
2798 |
* scheduled for execution |
2799 |
*/ |
2800 |
public void execute(ForkJoinTask<?> task) { |
2801 |
poolSubmit(true, task); |
2802 |
} |
2803 |
|
2804 |
// AbstractExecutorService methods |
2805 |
|
2806 |
/** |
2807 |
* @throws NullPointerException if the task is null |
2808 |
* @throws RejectedExecutionException if the task cannot be |
2809 |
* scheduled for execution |
2810 |
*/ |
2811 |
@Override |
2812 |
@SuppressWarnings("unchecked") |
2813 |
public void execute(Runnable task) { |
2814 |
poolSubmit(true, (task instanceof ForkJoinTask<?>) |
2815 |
? (ForkJoinTask<Void>) task // avoid re-wrap |
2816 |
: new ForkJoinTask.RunnableExecuteAction(task)); |
2817 |
} |
2818 |
|
2819 |
/** |
2820 |
* Submits a ForkJoinTask for execution. |
2821 |
* |
2822 |
* @param task the task to submit |
2823 |
* @param <T> the type of the task's result |
2824 |
* @return the task |
2825 |
* @throws NullPointerException if the task is null |
2826 |
* @throws RejectedExecutionException if the task cannot be |
2827 |
* scheduled for execution |
2828 |
*/ |
2829 |
public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) { |
2830 |
return poolSubmit(true, task); |
2831 |
} |
2832 |
|
2833 |
/** |
2834 |
* @throws NullPointerException if the task is null |
2835 |
* @throws RejectedExecutionException if the task cannot be |
2836 |
* scheduled for execution |
2837 |
*/ |
2838 |
@Override |
2839 |
public <T> ForkJoinTask<T> submit(Callable<T> task) { |
2840 |
return poolSubmit(true, new ForkJoinTask.AdaptedCallable<T>(task)); |
2841 |
} |
2842 |
|
2843 |
/** |
2844 |
* @throws NullPointerException if the task is null |
2845 |
* @throws RejectedExecutionException if the task cannot be |
2846 |
* scheduled for execution |
2847 |
*/ |
2848 |
@Override |
2849 |
public <T> ForkJoinTask<T> submit(Runnable task, T result) { |
2850 |
return poolSubmit(true, new ForkJoinTask.AdaptedRunnable<T>(task, result)); |
2851 |
} |
2852 |
|
2853 |
/** |
2854 |
* @throws NullPointerException if the task is null |
2855 |
* @throws RejectedExecutionException if the task cannot be |
2856 |
* scheduled for execution |
2857 |
*/ |
2858 |
@Override |
2859 |
@SuppressWarnings("unchecked") |
2860 |
public ForkJoinTask<?> submit(Runnable task) { |
2861 |
return poolSubmit(true, (task instanceof ForkJoinTask<?>) |
2862 |
? (ForkJoinTask<Void>) task // avoid re-wrap |
2863 |
: new ForkJoinTask.AdaptedRunnableAction(task)); |
2864 |
} |
2865 |
|
2866 |
// Added mainly for possible use in Loom |
2867 |
|
2868 |
/** |
2869 |
* Submits the given task without guaranteeing that it will |
2870 |
* eventually execute in the absence of available active threads. |
2871 |
* In some contexts, this method may reduce contention and |
2872 |
* overhead by relying on context-specific knowledge that existing |
2873 |
* threads (possibly including the calling thread if operating in |
2874 |
* this pool) will eventually be available to execute the task. |
2875 |
* |
2876 |
* @param task the task |
2877 |
* @param <T> the type of the task's result |
2878 |
* @return the task |
2879 |
* @since 19 |
2880 |
*/ |
2881 |
public <T> ForkJoinTask<T> lazySubmit(ForkJoinTask<T> task) { |
2882 |
return poolSubmit(false, task); |
2883 |
} |
2884 |
|
2885 |
/** |
2886 |
* Changes the target parallelism of this pool, controlling the |
2887 |
* future creation, use, and termination of worker threads. |
2888 |
* Applications include contexts in which the number of available |
2889 |
* processors changes over time. |
2890 |
* |
2891 |
* @implNote This implementation restricts the maximum number of |
2892 |
* running threads to 32767 |
2893 |
* |
2894 |
* @param size the target parallelism level |
2895 |
* @return the previous parallelism level. |
2896 |
* @throws IllegalArgumentException if size is less than 1 or |
2897 |
* greater than the maximum supported by this pool. |
2898 |
* @throws UnsupportedOperationException this is the{@link |
2899 |
* #commonPool()} and parallelism level was set by System |
2900 |
* property {@systemProperty |
2901 |
* java.util.concurrent.ForkJoinPool.common.parallelism}. |
2902 |
* @throws SecurityException if a security manager exists and |
2903 |
* the caller is not permitted to modify threads |
2904 |
* because it does not hold {@link |
2905 |
* java.lang.RuntimePermission}{@code ("modifyThread")} |
2906 |
* @since 19 |
2907 |
*/ |
2908 |
public int setParallelism(int size) { |
2909 |
if (size < 1 || size > MAX_CAP) |
2910 |
throw new IllegalArgumentException(); |
2911 |
if ((config & PRESET_SIZE) != 0) |
2912 |
throw new UnsupportedOperationException("Cannot override System property"); |
2913 |
checkPermission(); |
2914 |
return getAndSetParallelism(size); |
2915 |
} |
2916 |
|
2917 |
/** |
2918 |
* @throws NullPointerException {@inheritDoc} |
2919 |
* @throws RejectedExecutionException {@inheritDoc} |
2920 |
*/ |
2921 |
@Override |
2922 |
public <T> List<Future<T>> invokeAll(Collection<? extends Callable<T>> tasks) { |
2923 |
ArrayList<Future<T>> futures = new ArrayList<>(tasks.size()); |
2924 |
try { |
2925 |
for (Callable<T> t : tasks) { |
2926 |
ForkJoinTask<T> f = |
2927 |
new ForkJoinTask.AdaptedInterruptibleCallable<T>(t); |
2928 |
futures.add(f); |
2929 |
poolSubmit(true, f); |
2930 |
} |
2931 |
for (int i = futures.size() - 1; i >= 0; --i) |
2932 |
((ForkJoinTask<?>)futures.get(i)).quietlyJoin(); |
2933 |
return futures; |
2934 |
} catch (Throwable t) { |
2935 |
for (Future<T> e : futures) |
2936 |
ForkJoinTask.cancelIgnoringExceptions(e); |
2937 |
throw t; |
2938 |
} |
2939 |
} |
2940 |
|
2941 |
@Override |
2942 |
public <T> List<Future<T>> invokeAll(Collection<? extends Callable<T>> tasks, |
2943 |
long timeout, TimeUnit unit) |
2944 |
throws InterruptedException { |
2945 |
long nanos = unit.toNanos(timeout); |
2946 |
ArrayList<Future<T>> futures = new ArrayList<>(tasks.size()); |
2947 |
try { |
2948 |
for (Callable<T> t : tasks) { |
2949 |
ForkJoinTask<T> f = |
2950 |
new ForkJoinTask.AdaptedInterruptibleCallable<T>(t); |
2951 |
futures.add(f); |
2952 |
poolSubmit(true, f); |
2953 |
} |
2954 |
long startTime = System.nanoTime(), ns = nanos; |
2955 |
boolean timedOut = (ns < 0L); |
2956 |
for (int i = futures.size() - 1; i >= 0; --i) { |
2957 |
ForkJoinTask<T> f = (ForkJoinTask<T>)futures.get(i); |
2958 |
if (!f.isDone()) { |
2959 |
if (!timedOut) |
2960 |
timedOut = !f.quietlyJoin(ns, TimeUnit.NANOSECONDS); |
2961 |
if (timedOut) |
2962 |
ForkJoinTask.cancelIgnoringExceptions(f); |
2963 |
else |
2964 |
ns = nanos - (System.nanoTime() - startTime); |
2965 |
} |
2966 |
} |
2967 |
return futures; |
2968 |
} catch (Throwable t) { |
2969 |
for (Future<T> e : futures) |
2970 |
ForkJoinTask.cancelIgnoringExceptions(e); |
2971 |
throw t; |
2972 |
} |
2973 |
} |
2974 |
|
2975 |
// Task to hold results from InvokeAnyTasks |
2976 |
static final class InvokeAnyRoot<E> extends ForkJoinTask<E> { |
2977 |
private static final long serialVersionUID = 2838392045355241008L; |
2978 |
@SuppressWarnings("serial") // Conditionally serializable |
2979 |
volatile E result; |
2980 |
final AtomicInteger count; // in case all throw |
2981 |
@SuppressWarnings("serial") |
2982 |
final ForkJoinPool pool; // to check shutdown while collecting |
2983 |
InvokeAnyRoot(int n, ForkJoinPool p) { |
2984 |
pool = p; |
2985 |
count = new AtomicInteger(n); |
2986 |
} |
2987 |
final void tryComplete(Callable<E> c) { // called by InvokeAnyTasks |
2988 |
Throwable ex = null; |
2989 |
boolean failed; |
2990 |
if (c == null || Thread.interrupted() || |
2991 |
(pool != null && pool.runState < 0)) |
2992 |
failed = true; |
2993 |
else if (isDone()) |
2994 |
failed = false; |
2995 |
else { |
2996 |
try { |
2997 |
complete(c.call()); |
2998 |
failed = false; |
2999 |
} catch (Throwable tx) { |
3000 |
ex = tx; |
3001 |
failed = true; |
3002 |
} |
3003 |
} |
3004 |
if ((pool != null && pool.runState < 0) || |
3005 |
(failed && count.getAndDecrement() <= 1)) |
3006 |
trySetThrown(ex != null ? ex : new CancellationException()); |
3007 |
} |
3008 |
public final boolean exec() { return false; } // never forked |
3009 |
public final E getRawResult() { return result; } |
3010 |
public final void setRawResult(E v) { result = v; } |
3011 |
} |
3012 |
|
3013 |
// Variant of AdaptedInterruptibleCallable with results in InvokeAnyRoot |
3014 |
static final class InvokeAnyTask<E> extends ForkJoinTask<E> { |
3015 |
private static final long serialVersionUID = 2838392045355241008L; |
3016 |
final InvokeAnyRoot<E> root; |
3017 |
@SuppressWarnings("serial") // Conditionally serializable |
3018 |
final Callable<E> callable; |
3019 |
transient volatile Thread runner; |
3020 |
InvokeAnyTask(InvokeAnyRoot<E> root, Callable<E> callable) { |
3021 |
this.root = root; |
3022 |
this.callable = callable; |
3023 |
} |
3024 |
public final boolean exec() { |
3025 |
Thread.interrupted(); |
3026 |
runner = Thread.currentThread(); |
3027 |
root.tryComplete(callable); |
3028 |
runner = null; |
3029 |
Thread.interrupted(); |
3030 |
return true; |
3031 |
} |
3032 |
public final boolean cancel(boolean mayInterruptIfRunning) { |
3033 |
Thread t; |
3034 |
boolean stat = super.cancel(false); |
3035 |
if (mayInterruptIfRunning && (t = runner) != null) { |
3036 |
try { |
3037 |
t.interrupt(); |
3038 |
} catch (Throwable ignore) { |
3039 |
} |
3040 |
} |
3041 |
return stat; |
3042 |
} |
3043 |
public final void setRawResult(E v) {} // unused |
3044 |
public final E getRawResult() { return null; } |
3045 |
} |
3046 |
|
3047 |
@Override |
3048 |
public <T> T invokeAny(Collection<? extends Callable<T>> tasks) |
3049 |
throws InterruptedException, ExecutionException { |
3050 |
int n = tasks.size(); |
3051 |
if (n <= 0) |
3052 |
throw new IllegalArgumentException(); |
3053 |
InvokeAnyRoot<T> root = new InvokeAnyRoot<T>(n, this); |
3054 |
ArrayList<InvokeAnyTask<T>> fs = new ArrayList<>(n); |
3055 |
try { |
3056 |
for (Callable<T> c : tasks) { |
3057 |
if (c == null) |
3058 |
throw new NullPointerException(); |
3059 |
InvokeAnyTask<T> f = new InvokeAnyTask<T>(root, c); |
3060 |
fs.add(f); |
3061 |
poolSubmit(true, f); |
3062 |
if (root.isDone()) |
3063 |
break; |
3064 |
} |
3065 |
return root.get(); |
3066 |
} finally { |
3067 |
for (InvokeAnyTask<T> f : fs) |
3068 |
ForkJoinTask.cancelIgnoringExceptions(f); |
3069 |
} |
3070 |
} |
3071 |
|
3072 |
@Override |
3073 |
public <T> T invokeAny(Collection<? extends Callable<T>> tasks, |
3074 |
long timeout, TimeUnit unit) |
3075 |
throws InterruptedException, ExecutionException, TimeoutException { |
3076 |
long nanos = unit.toNanos(timeout); |
3077 |
int n = tasks.size(); |
3078 |
if (n <= 0) |
3079 |
throw new IllegalArgumentException(); |
3080 |
InvokeAnyRoot<T> root = new InvokeAnyRoot<T>(n, this); |
3081 |
ArrayList<InvokeAnyTask<T>> fs = new ArrayList<>(n); |
3082 |
try { |
3083 |
for (Callable<T> c : tasks) { |
3084 |
if (c == null) |
3085 |
throw new NullPointerException(); |
3086 |
InvokeAnyTask<T> f = new InvokeAnyTask<T>(root, c); |
3087 |
fs.add(f); |
3088 |
poolSubmit(true, f); |
3089 |
if (root.isDone()) |
3090 |
break; |
3091 |
} |
3092 |
return root.get(nanos, TimeUnit.NANOSECONDS); |
3093 |
} finally { |
3094 |
for (InvokeAnyTask<T> f : fs) |
3095 |
ForkJoinTask.cancelIgnoringExceptions(f); |
3096 |
} |
3097 |
} |
3098 |
|
3099 |
/** |
3100 |
* Returns the factory used for constructing new workers. |
3101 |
* |
3102 |
* @return the factory used for constructing new workers |
3103 |
*/ |
3104 |
public ForkJoinWorkerThreadFactory getFactory() { |
3105 |
return factory; |
3106 |
} |
3107 |
|
3108 |
/** |
3109 |
* Returns the handler for internal worker threads that terminate |
3110 |
* due to unrecoverable errors encountered while executing tasks. |
3111 |
* |
3112 |
* @return the handler, or {@code null} if none |
3113 |
*/ |
3114 |
public UncaughtExceptionHandler getUncaughtExceptionHandler() { |
3115 |
return ueh; |
3116 |
} |
3117 |
|
3118 |
/** |
3119 |
* Returns the targeted parallelism level of this pool. |
3120 |
* |
3121 |
* @return the targeted parallelism level of this pool |
3122 |
*/ |
3123 |
public int getParallelism() { |
3124 |
return Math.max(getParallelismOpaque(), 1); |
3125 |
} |
3126 |
|
3127 |
/** |
3128 |
* Returns the targeted parallelism level of the common pool. |
3129 |
* |
3130 |
* @return the targeted parallelism level of the common pool |
3131 |
* @since 1.8 |
3132 |
*/ |
3133 |
public static int getCommonPoolParallelism() { |
3134 |
return common.getParallelism(); |
3135 |
} |
3136 |
|
3137 |
/** |
3138 |
* Returns the number of worker threads that have started but not |
3139 |
* yet terminated. The result returned by this method may differ |
3140 |
* from {@link #getParallelism} when threads are created to |
3141 |
* maintain parallelism when others are cooperatively blocked. |
3142 |
* |
3143 |
* @return the number of worker threads |
3144 |
*/ |
3145 |
public int getPoolSize() { |
3146 |
return (short)(ctl >>> TC_SHIFT); |
3147 |
} |
3148 |
|
3149 |
/** |
3150 |
* Returns {@code true} if this pool uses local first-in-first-out |
3151 |
* scheduling mode for forked tasks that are never joined. |
3152 |
* |
3153 |
* @return {@code true} if this pool uses async mode |
3154 |
*/ |
3155 |
public boolean getAsyncMode() { |
3156 |
return (config & FIFO) != 0; |
3157 |
} |
3158 |
|
3159 |
/** |
3160 |
* Returns an estimate of the number of worker threads that are |
3161 |
* not blocked waiting to join tasks or for other managed |
3162 |
* synchronization. This method may overestimate the |
3163 |
* number of running threads. |
3164 |
* |
3165 |
* @return the number of worker threads |
3166 |
*/ |
3167 |
public int getRunningThreadCount() { |
3168 |
WorkQueue[] qs; WorkQueue q; |
3169 |
int rc = 0; |
3170 |
if ((runState & TERMINATED) == 0 && (qs = queues) != null) { |
3171 |
for (int i = 1; i < qs.length; i += 2) { |
3172 |
if ((q = qs[i]) != null && q.isApparentlyUnblocked()) |
3173 |
++rc; |
3174 |
} |
3175 |
} |
3176 |
return rc; |
3177 |
} |
3178 |
|
3179 |
/** |
3180 |
* Returns an estimate of the number of threads that are currently |
3181 |
* stealing or executing tasks. This method may overestimate the |
3182 |
* number of active threads. |
3183 |
* |
3184 |
* @return the number of active threads |
3185 |
*/ |
3186 |
public int getActiveThreadCount() { |
3187 |
return Math.max((short)(ctl >>> RC_SHIFT), 0); |
3188 |
} |
3189 |
|
3190 |
/** |
3191 |
* Returns {@code true} if all worker threads are currently idle. |
3192 |
* An idle worker is one that cannot obtain a task to execute |
3193 |
* because none are available to steal from other threads, and |
3194 |
* there are no pending submissions to the pool. This method is |
3195 |
* conservative; it might not return {@code true} immediately upon |
3196 |
* idleness of all threads, but will eventually become true if |
3197 |
* threads remain inactive. |
3198 |
* |
3199 |
* @return {@code true} if all threads are currently idle |
3200 |
*/ |
3201 |
public boolean isQuiescent() { |
3202 |
return canStop(); |
3203 |
} |
3204 |
|
3205 |
/** |
3206 |
* Returns an estimate of the total number of completed tasks that |
3207 |
* were executed by a thread other than their submitter. The |
3208 |
* reported value underestimates the actual total number of steals |
3209 |
* when the pool is not quiescent. This value may be useful for |
3210 |
* monitoring and tuning fork/join programs: in general, steal |
3211 |
* counts should be high enough to keep threads busy, but low |
3212 |
* enough to avoid overhead and contention across threads. |
3213 |
* |
3214 |
* @return the number of steals |
3215 |
*/ |
3216 |
public long getStealCount() { |
3217 |
long count = stealCount; |
3218 |
WorkQueue[] qs; WorkQueue q; |
3219 |
if ((qs = queues) != null) { |
3220 |
for (int i = 1; i < qs.length; i += 2) { |
3221 |
if ((q = qs[i]) != null) |
3222 |
count += (long)q.nsteals & 0xffffffffL; |
3223 |
} |
3224 |
} |
3225 |
return count; |
3226 |
} |
3227 |
|
3228 |
/** |
3229 |
* Returns an estimate of the total number of tasks currently held |
3230 |
* in queues by worker threads (but not including tasks submitted |
3231 |
* to the pool that have not begun executing). This value is only |
3232 |
* an approximation, obtained by iterating across all threads in |
3233 |
* the pool. This method may be useful for tuning task |
3234 |
* granularities. |
3235 |
* |
3236 |
* @return the number of queued tasks |
3237 |
*/ |
3238 |
public long getQueuedTaskCount() { |
3239 |
WorkQueue[] qs; WorkQueue q; |
3240 |
int count = 0; |
3241 |
if ((runState & TERMINATED) == 0 && (qs = queues) != null) { |
3242 |
for (int i = 1; i < qs.length; i += 2) { |
3243 |
if ((q = qs[i]) != null) |
3244 |
count += q.queueSize(); |
3245 |
} |
3246 |
} |
3247 |
return count; |
3248 |
} |
3249 |
|
3250 |
/** |
3251 |
* Returns an estimate of the number of tasks submitted to this |
3252 |
* pool that have not yet begun executing. This method may take |
3253 |
* time proportional to the number of submissions. |
3254 |
* |
3255 |
* @return the number of queued submissions |
3256 |
*/ |
3257 |
public int getQueuedSubmissionCount() { |
3258 |
WorkQueue[] qs; WorkQueue q; |
3259 |
int count = 0; |
3260 |
if ((runState & TERMINATED) == 0 && (qs = queues) != null) { |
3261 |
for (int i = 0; i < qs.length; i += 2) { |
3262 |
if ((q = qs[i]) != null) |
3263 |
count += q.queueSize(); |
3264 |
} |
3265 |
} |
3266 |
return count; |
3267 |
} |
3268 |
|
3269 |
/** |
3270 |
* Returns {@code true} if there are any tasks submitted to this |
3271 |
* pool that have not yet begun executing. |
3272 |
* |
3273 |
* @return {@code true} if there are any queued submissions |
3274 |
*/ |
3275 |
public boolean hasQueuedSubmissions() { |
3276 |
return hasTasks(true); |
3277 |
} |
3278 |
|
3279 |
/** |
3280 |
* Removes and returns the next unexecuted submission if one is |
3281 |
* available. This method may be useful in extensions to this |
3282 |
* class that re-assign work in systems with multiple pools. |
3283 |
* |
3284 |
* @return the next submission, or {@code null} if none |
3285 |
*/ |
3286 |
protected ForkJoinTask<?> pollSubmission() { |
3287 |
return pollScan(true); |
3288 |
} |
3289 |
|
3290 |
/** |
3291 |
* Removes all available unexecuted submitted and forked tasks |
3292 |
* from scheduling queues and adds them to the given collection, |
3293 |
* without altering their execution status. These may include |
3294 |
* artificially generated or wrapped tasks. This method is |
3295 |
* designed to be invoked only when the pool is known to be |
3296 |
* quiescent. Invocations at other times may not remove all |
3297 |
* tasks. A failure encountered while attempting to add elements |
3298 |
* to collection {@code c} may result in elements being in |
3299 |
* neither, either or both collections when the associated |
3300 |
* exception is thrown. The behavior of this operation is |
3301 |
* undefined if the specified collection is modified while the |
3302 |
* operation is in progress. |
3303 |
* |
3304 |
* @param c the collection to transfer elements into |
3305 |
* @return the number of elements transferred |
3306 |
*/ |
3307 |
protected int drainTasksTo(Collection<? super ForkJoinTask<?>> c) { |
3308 |
int count = 0; |
3309 |
for (ForkJoinTask<?> t; (t = pollScan(false)) != null; ) { |
3310 |
c.add(t); |
3311 |
++count; |
3312 |
} |
3313 |
return count; |
3314 |
} |
3315 |
|
3316 |
/** |
3317 |
* Returns a string identifying this pool, as well as its state, |
3318 |
* including indications of run state, parallelism level, and |
3319 |
* worker and task counts. |
3320 |
* |
3321 |
* @return a string identifying this pool, as well as its state |
3322 |
*/ |
3323 |
public String toString() { |
3324 |
// Use a single pass through queues to collect counts |
3325 |
long st = stealCount; |
3326 |
long qt = 0L, ss = 0L; int rc = 0; |
3327 |
WorkQueue[] qs; WorkQueue q; |
3328 |
if ((qs = queues) != null) { |
3329 |
for (int i = 0; i < qs.length; ++i) { |
3330 |
if ((q = qs[i]) != null) { |
3331 |
int size = q.queueSize(); |
3332 |
if ((i & 1) == 0) |
3333 |
ss += size; |
3334 |
else { |
3335 |
qt += size; |
3336 |
st += (long)q.nsteals & 0xffffffffL; |
3337 |
if (q.isApparentlyUnblocked()) |
3338 |
++rc; |
3339 |
} |
3340 |
} |
3341 |
} |
3342 |
} |
3343 |
|
3344 |
int pc = parallelism; |
3345 |
long c = ctl; |
3346 |
int tc = (short)(c >>> TC_SHIFT); |
3347 |
int ac = (short)(c >>> RC_SHIFT); |
3348 |
if (ac < 0) // ignore transient negative |
3349 |
ac = 0; |
3350 |
int rs = runState; |
3351 |
String level = ((rs & TERMINATED) != 0 ? "Terminated" : |
3352 |
(rs & STOP) != 0 ? "Terminating" : |
3353 |
(rs & SHUTDOWN) != 0 ? "Shutting down" : |
3354 |
"Running"); |
3355 |
return super.toString() + |
3356 |
"[" + level + |
3357 |
", parallelism = " + pc + |
3358 |
", size = " + tc + |
3359 |
", active = " + ac + |
3360 |
", running = " + rc + |
3361 |
", steals = " + st + |
3362 |
", tasks = " + qt + |
3363 |
", submissions = " + ss + |
3364 |
"]"; |
3365 |
} |
3366 |
|
3367 |
/** |
3368 |
* Possibly initiates an orderly shutdown in which previously |
3369 |
* submitted tasks are executed, but no new tasks will be |
3370 |
* accepted. Invocation has no effect on execution state if this |
3371 |
* is the {@link #commonPool()}, and no additional effect if |
3372 |
* already shut down. Tasks that are in the process of being |
3373 |
* submitted concurrently during the course of this method may or |
3374 |
* may not be rejected. |
3375 |
* |
3376 |
* @throws SecurityException if a security manager exists and |
3377 |
* the caller is not permitted to modify threads |
3378 |
* because it does not hold {@link |
3379 |
* java.lang.RuntimePermission}{@code ("modifyThread")} |
3380 |
*/ |
3381 |
public void shutdown() { |
3382 |
checkPermission(); |
3383 |
tryTerminate(false, true); |
3384 |
} |
3385 |
|
3386 |
/** |
3387 |
* Possibly attempts to cancel and/or stop all tasks, and reject |
3388 |
* all subsequently submitted tasks. Invocation has no effect on |
3389 |
* execution state if this is the {@link #commonPool()}, and no |
3390 |
* additional effect if already shut down. Otherwise, tasks that |
3391 |
* are in the process of being submitted or executed concurrently |
3392 |
* during the course of this method may or may not be |
3393 |
* rejected. This method cancels both existing and unexecuted |
3394 |
* tasks, in order to permit termination in the presence of task |
3395 |
* dependencies. So the method always returns an empty list |
3396 |
* (unlike the case for some other Executors). |
3397 |
* |
3398 |
* @return an empty list |
3399 |
* @throws SecurityException if a security manager exists and |
3400 |
* the caller is not permitted to modify threads |
3401 |
* because it does not hold {@link |
3402 |
* java.lang.RuntimePermission}{@code ("modifyThread")} |
3403 |
*/ |
3404 |
public List<Runnable> shutdownNow() { |
3405 |
checkPermission(); |
3406 |
tryTerminate(true, true); |
3407 |
return Collections.emptyList(); |
3408 |
} |
3409 |
|
3410 |
/** |
3411 |
* Returns {@code true} if all tasks have completed following shut down. |
3412 |
* |
3413 |
* @return {@code true} if all tasks have completed following shut down |
3414 |
*/ |
3415 |
public boolean isTerminated() { |
3416 |
return (runState & TERMINATED) != 0; |
3417 |
} |
3418 |
|
3419 |
/** |
3420 |
* Returns {@code true} if the process of termination has |
3421 |
* commenced but not yet completed. This method may be useful for |
3422 |
* debugging. A return of {@code true} reported a sufficient |
3423 |
* period after shutdown may indicate that submitted tasks have |
3424 |
* ignored or suppressed interruption, or are waiting for I/O, |
3425 |
* causing this executor not to properly terminate. (See the |
3426 |
* advisory notes for class {@link ForkJoinTask} stating that |
3427 |
* tasks should not normally entail blocking operations. But if |
3428 |
* they do, they must abort them on interrupt.) |
3429 |
* |
3430 |
* @return {@code true} if terminating but not yet terminated |
3431 |
*/ |
3432 |
public boolean isTerminating() { |
3433 |
return (runState & (STOP | TERMINATED)) == STOP; |
3434 |
} |
3435 |
|
3436 |
/** |
3437 |
* Returns {@code true} if this pool has been shut down. |
3438 |
* |
3439 |
* @return {@code true} if this pool has been shut down |
3440 |
*/ |
3441 |
public boolean isShutdown() { |
3442 |
return runState != 0; |
3443 |
} |
3444 |
|
3445 |
/** |
3446 |
* Blocks until all tasks have completed execution after a |
3447 |
* shutdown request, or the timeout occurs, or the current thread |
3448 |
* is interrupted, whichever happens first. Because the {@link |
3449 |
* #commonPool()} never terminates until program shutdown, when |
3450 |
* applied to the common pool, this method is equivalent to {@link |
3451 |
* #awaitQuiescence(long, TimeUnit)} but always returns {@code false}. |
3452 |
* |
3453 |
* @param timeout the maximum time to wait |
3454 |
* @param unit the time unit of the timeout argument |
3455 |
* @return {@code true} if this executor terminated and |
3456 |
* {@code false} if the timeout elapsed before termination |
3457 |
* @throws InterruptedException if interrupted while waiting |
3458 |
*/ |
3459 |
public boolean awaitTermination(long timeout, TimeUnit unit) |
3460 |
throws InterruptedException { |
3461 |
ReentrantLock lock; Condition cond; boolean terminated; |
3462 |
long nanos = unit.toNanos(timeout); |
3463 |
if ((config & ISCOMMON) != 0) { |
3464 |
if (helpQuiescePool(this, nanos, true) < 0) |
3465 |
throw new InterruptedException(); |
3466 |
terminated = false; |
3467 |
} |
3468 |
else if (!(terminated = ((runState & TERMINATED) != 0))) { |
3469 |
tryTerminate(false, false); // reduce transient blocking |
3470 |
if ((lock = registrationLock) != null && |
3471 |
!(terminated = (((runState & TERMINATED) != 0)))) { |
3472 |
lock.lock(); |
3473 |
try { |
3474 |
if ((cond = termination) == null) |
3475 |
termination = cond = lock.newCondition(); |
3476 |
while (!(terminated = ((runState & TERMINATED) != 0)) && |
3477 |
nanos > 0L) |
3478 |
nanos = cond.awaitNanos(nanos); |
3479 |
} finally { |
3480 |
lock.unlock(); |
3481 |
} |
3482 |
} |
3483 |
} |
3484 |
return terminated; |
3485 |
} |
3486 |
|
3487 |
/** |
3488 |
* If called by a ForkJoinTask operating in this pool, equivalent |
3489 |
* in effect to {@link ForkJoinTask#helpQuiesce}. Otherwise, |
3490 |
* waits and/or attempts to assist performing tasks until this |
3491 |
* pool {@link #isQuiescent} or the indicated timeout elapses. |
3492 |
* |
3493 |
* @param timeout the maximum time to wait |
3494 |
* @param unit the time unit of the timeout argument |
3495 |
* @return {@code true} if quiescent; {@code false} if the |
3496 |
* timeout elapsed. |
3497 |
*/ |
3498 |
public boolean awaitQuiescence(long timeout, TimeUnit unit) { |
3499 |
return (helpQuiescePool(this, unit.toNanos(timeout), false) > 0); |
3500 |
} |
3501 |
|
3502 |
/** |
3503 |
* Unless this is the {@link #commonPool()}, initiates an orderly |
3504 |
* shutdown in which previously submitted tasks are executed, but |
3505 |
* no new tasks will be accepted, and waits until all tasks have |
3506 |
* completed execution and the executor has terminated. |
3507 |
* |
3508 |
* <p>If already terminated, or this is the {@link |
3509 |
* #commonPool()}, this method has no effect on execution, and |
3510 |
* does not wait. Otherwise, if interrupted while waiting, this |
3511 |
* method stops all executing tasks as if by invoking {@link |
3512 |
* #shutdownNow()}. It then continues to wait until all actively |
3513 |
* executing tasks have completed. Tasks that were awaiting |
3514 |
* execution are not executed. The interrupt status will be |
3515 |
* re-asserted before this method returns. |
3516 |
* |
3517 |
* @throws SecurityException if a security manager exists and |
3518 |
* shutting down this ExecutorService may manipulate |
3519 |
* threads that the caller is not permitted to modify |
3520 |
* because it does not hold {@link |
3521 |
* java.lang.RuntimePermission}{@code ("modifyThread")}, |
3522 |
* or the security manager's {@code checkAccess} method |
3523 |
* denies access. |
3524 |
* @since 19 |
3525 |
*/ |
3526 |
@Override |
3527 |
public void close() { |
3528 |
if ((config & ISCOMMON) == 0) { |
3529 |
boolean terminated = tryTerminate(false, false); |
3530 |
if (!terminated) { |
3531 |
shutdown(); |
3532 |
boolean interrupted = false; |
3533 |
while (!terminated) { |
3534 |
try { |
3535 |
terminated = awaitTermination(1L, TimeUnit.DAYS); |
3536 |
} catch (InterruptedException e) { |
3537 |
if (!interrupted) { |
3538 |
shutdownNow(); |
3539 |
interrupted = true; |
3540 |
} |
3541 |
} |
3542 |
} |
3543 |
if (interrupted) { |
3544 |
Thread.currentThread().interrupt(); |
3545 |
} |
3546 |
} |
3547 |
} |
3548 |
} |
3549 |
|
3550 |
/** |
3551 |
* Interface for extending managed parallelism for tasks running |
3552 |
* in {@link ForkJoinPool}s. |
3553 |
* |
3554 |
* <p>A {@code ManagedBlocker} provides two methods. Method |
3555 |
* {@link #isReleasable} must return {@code true} if blocking is |
3556 |
* not necessary. Method {@link #block} blocks the current thread |
3557 |
* if necessary (perhaps internally invoking {@code isReleasable} |
3558 |
* before actually blocking). These actions are performed by any |
3559 |
* thread invoking {@link |
3560 |
* ForkJoinPool#managedBlock(ManagedBlocker)}. The unusual |
3561 |
* methods in this API accommodate synchronizers that may, but |
3562 |
* don't usually, block for long periods. Similarly, they allow |
3563 |
* more efficient internal handling of cases in which additional |
3564 |
* workers may be, but usually are not, needed to ensure |
3565 |
* sufficient parallelism. Toward this end, implementations of |
3566 |
* method {@code isReleasable} must be amenable to repeated |
3567 |
* invocation. Neither method is invoked after a prior invocation |
3568 |
* of {@code isReleasable} or {@code block} returns {@code true}. |
3569 |
* |
3570 |
* <p>For example, here is a ManagedBlocker based on a |
3571 |
* ReentrantLock: |
3572 |
* <pre> {@code |
3573 |
* class ManagedLocker implements ManagedBlocker { |
3574 |
* final ReentrantLock lock; |
3575 |
* boolean hasLock = false; |
3576 |
* ManagedLocker(ReentrantLock lock) { this.lock = lock; } |
3577 |
* public boolean block() { |
3578 |
* if (!hasLock) |
3579 |
* lock.lock(); |
3580 |
* return true; |
3581 |
* } |
3582 |
* public boolean isReleasable() { |
3583 |
* return hasLock || (hasLock = lock.tryLock()); |
3584 |
* } |
3585 |
* }}</pre> |
3586 |
* |
3587 |
* <p>Here is a class that possibly blocks waiting for an |
3588 |
* item on a given queue: |
3589 |
* <pre> {@code |
3590 |
* class QueueTaker<E> implements ManagedBlocker { |
3591 |
* final BlockingQueue<E> queue; |
3592 |
* volatile E item = null; |
3593 |
* QueueTaker(BlockingQueue<E> q) { this.queue = q; } |
3594 |
* public boolean block() throws InterruptedException { |
3595 |
* if (item == null) |
3596 |
* item = queue.take(); |
3597 |
* return true; |
3598 |
* } |
3599 |
* public boolean isReleasable() { |
3600 |
* return item != null || (item = queue.poll()) != null; |
3601 |
* } |
3602 |
* public E getItem() { // call after pool.managedBlock completes |
3603 |
* return item; |
3604 |
* } |
3605 |
* }}</pre> |
3606 |
*/ |
3607 |
public static interface ManagedBlocker { |
3608 |
/** |
3609 |
* Possibly blocks the current thread, for example waiting for |
3610 |
* a lock or condition. |
3611 |
* |
3612 |
* @return {@code true} if no additional blocking is necessary |
3613 |
* (i.e., if isReleasable would return true) |
3614 |
* @throws InterruptedException if interrupted while waiting |
3615 |
* (the method is not required to do so, but is allowed to) |
3616 |
*/ |
3617 |
boolean block() throws InterruptedException; |
3618 |
|
3619 |
/** |
3620 |
* Returns {@code true} if blocking is unnecessary. |
3621 |
* @return {@code true} if blocking is unnecessary |
3622 |
*/ |
3623 |
boolean isReleasable(); |
3624 |
} |
3625 |
|
3626 |
/** |
3627 |
* Runs the given possibly blocking task. When {@linkplain |
3628 |
* ForkJoinTask#inForkJoinPool() running in a ForkJoinPool}, this |
3629 |
* method possibly arranges for a spare thread to be activated if |
3630 |
* necessary to ensure sufficient parallelism while the current |
3631 |
* thread is blocked in {@link ManagedBlocker#block blocker.block()}. |
3632 |
* |
3633 |
* <p>This method repeatedly calls {@code blocker.isReleasable()} and |
3634 |
* {@code blocker.block()} until either method returns {@code true}. |
3635 |
* Every call to {@code blocker.block()} is preceded by a call to |
3636 |
* {@code blocker.isReleasable()} that returned {@code false}. |
3637 |
* |
3638 |
* <p>If not running in a ForkJoinPool, this method is |
3639 |
* behaviorally equivalent to |
3640 |
* <pre> {@code |
3641 |
* while (!blocker.isReleasable()) |
3642 |
* if (blocker.block()) |
3643 |
* break;}</pre> |
3644 |
* |
3645 |
* If running in a ForkJoinPool, the pool may first be expanded to |
3646 |
* ensure sufficient parallelism available during the call to |
3647 |
* {@code blocker.block()}. |
3648 |
* |
3649 |
* @param blocker the blocker task |
3650 |
* @throws InterruptedException if {@code blocker.block()} did so |
3651 |
*/ |
3652 |
public static void managedBlock(ManagedBlocker blocker) |
3653 |
throws InterruptedException { |
3654 |
Thread t; ForkJoinPool p; |
3655 |
if ((t = Thread.currentThread()) instanceof ForkJoinWorkerThread && |
3656 |
(p = ((ForkJoinWorkerThread)t).pool) != null) |
3657 |
p.compensatedBlock(blocker); |
3658 |
else |
3659 |
unmanagedBlock(blocker); |
3660 |
} |
3661 |
|
3662 |
/** ManagedBlock for ForkJoinWorkerThreads */ |
3663 |
private void compensatedBlock(ManagedBlocker blocker) |
3664 |
throws InterruptedException { |
3665 |
if (blocker == null) throw new NullPointerException(); |
3666 |
for (;;) { |
3667 |
int comp; boolean done; |
3668 |
long c = ctl; |
3669 |
if (blocker.isReleasable()) |
3670 |
break; |
3671 |
if ((comp = tryCompensate(c, false)) >= 0) { |
3672 |
long post = (comp == 0) ? 0L : RC_UNIT; |
3673 |
try { |
3674 |
done = blocker.block(); |
3675 |
} finally { |
3676 |
getAndAddCtl(post); |
3677 |
} |
3678 |
if (done) |
3679 |
break; |
3680 |
} |
3681 |
} |
3682 |
} |
3683 |
|
3684 |
/** ManagedBlock for external threads */ |
3685 |
private static void unmanagedBlock(ManagedBlocker blocker) |
3686 |
throws InterruptedException { |
3687 |
if (blocker == null) throw new NullPointerException(); |
3688 |
do {} while (!blocker.isReleasable() && !blocker.block()); |
3689 |
} |
3690 |
|
3691 |
// AbstractExecutorService.newTaskFor overrides rely on |
3692 |
// undocumented fact that ForkJoinTask.adapt returns ForkJoinTasks |
3693 |
// that also implement RunnableFuture. |
3694 |
|
3695 |
@Override |
3696 |
protected <T> RunnableFuture<T> newTaskFor(Runnable runnable, T value) { |
3697 |
return new ForkJoinTask.AdaptedRunnable<T>(runnable, value); |
3698 |
} |
3699 |
|
3700 |
@Override |
3701 |
protected <T> RunnableFuture<T> newTaskFor(Callable<T> callable) { |
3702 |
return new ForkJoinTask.AdaptedCallable<T>(callable); |
3703 |
} |
3704 |
|
3705 |
static { |
3706 |
U = Unsafe.getUnsafe(); |
3707 |
Class<ForkJoinPool> klass = ForkJoinPool.class; |
3708 |
try { |
3709 |
POOLIDS = U.staticFieldOffset(klass.getDeclaredField("poolIds")); |
3710 |
} catch (NoSuchFieldException e) { |
3711 |
throw new ExceptionInInitializerError(e); |
3712 |
} |
3713 |
CTL = U.objectFieldOffset(klass, "ctl"); |
3714 |
RUNSTATE = U.objectFieldOffset(klass, "runState"); |
3715 |
PARALLELISM = U.objectFieldOffset(klass, "parallelism"); |
3716 |
THREADIDS = U.objectFieldOffset(klass, "threadIds"); |
3717 |
|
3718 |
defaultForkJoinWorkerThreadFactory = |
3719 |
new DefaultForkJoinWorkerThreadFactory(); |
3720 |
@SuppressWarnings("removal") |
3721 |
ForkJoinPool p = common = (System.getSecurityManager() == null) ? |
3722 |
new ForkJoinPool((byte)0) : |
3723 |
AccessController.doPrivileged(new PrivilegedAction<>() { |
3724 |
public ForkJoinPool run() { |
3725 |
return new ForkJoinPool((byte)0); }}); |
3726 |
Class<?> dep = LockSupport.class; // ensure loaded |
3727 |
} |
3728 |
} |