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.AccessControlContext; |
11 |
import java.security.Permissions; |
12 |
import java.security.ProtectionDomain; |
13 |
import java.util.ArrayList; |
14 |
import java.util.Arrays; |
15 |
import java.util.Collection; |
16 |
import java.util.Collections; |
17 |
import java.util.List; |
18 |
import java.util.concurrent.locks.ReentrantLock; |
19 |
import java.util.concurrent.locks.LockSupport; |
20 |
|
21 |
/** |
22 |
* An {@link ExecutorService} for running {@link ForkJoinTask}s. |
23 |
* A {@code ForkJoinPool} provides the entry point for submissions |
24 |
* from non-{@code ForkJoinTask} clients, as well as management and |
25 |
* monitoring operations. |
26 |
* |
27 |
* <p>A {@code ForkJoinPool} differs from other kinds of {@link |
28 |
* ExecutorService} mainly by virtue of employing |
29 |
* <em>work-stealing</em>: all threads in the pool attempt to find and |
30 |
* execute tasks submitted to the pool and/or created by other active |
31 |
* tasks (eventually blocking waiting for work if none exist). This |
32 |
* enables efficient processing when most tasks spawn other subtasks |
33 |
* (as do most {@code ForkJoinTask}s), as well as when many small |
34 |
* tasks are submitted to the pool from external clients. Especially |
35 |
* when setting <em>asyncMode</em> to true in constructors, {@code |
36 |
* ForkJoinPool}s may also be appropriate for use with event-style |
37 |
* tasks that are never joined. |
38 |
* |
39 |
* <p>A static {@link #commonPool()} is available and appropriate for |
40 |
* most applications. The common pool is used by any ForkJoinTask that |
41 |
* is not explicitly submitted to a specified pool. Using the common |
42 |
* pool normally reduces resource usage (its threads are slowly |
43 |
* reclaimed during periods of non-use, and reinstated upon subsequent |
44 |
* use). |
45 |
* |
46 |
* <p>For applications that require separate or custom pools, a {@code |
47 |
* ForkJoinPool} may be constructed with a given target parallelism |
48 |
* level; by default, equal to the number of available processors. |
49 |
* The pool attempts to maintain enough active (or available) threads |
50 |
* by dynamically adding, suspending, or resuming internal worker |
51 |
* threads, even if some tasks are stalled waiting to join others. |
52 |
* However, no such adjustments are guaranteed in the face of blocked |
53 |
* I/O or other unmanaged synchronization. The nested {@link |
54 |
* ManagedBlocker} interface enables extension of the kinds of |
55 |
* synchronization accommodated. |
56 |
* |
57 |
* <p>In addition to execution and lifecycle control methods, this |
58 |
* class provides status check methods (for example |
59 |
* {@link #getStealCount}) that are intended to aid in developing, |
60 |
* tuning, and monitoring fork/join applications. Also, method |
61 |
* {@link #toString} returns indications of pool state in a |
62 |
* convenient form for informal monitoring. |
63 |
* |
64 |
* <p>As is the case with other ExecutorServices, there are three |
65 |
* main task execution methods summarized in the following table. |
66 |
* These are designed to be used primarily by clients not already |
67 |
* engaged in fork/join computations in the current pool. The main |
68 |
* forms of these methods accept instances of {@code ForkJoinTask}, |
69 |
* but overloaded forms also allow mixed execution of plain {@code |
70 |
* Runnable}- or {@code Callable}- based activities as well. However, |
71 |
* tasks that are already executing in a pool should normally instead |
72 |
* use the within-computation forms listed in the table unless using |
73 |
* async event-style tasks that are not usually joined, in which case |
74 |
* there is little difference among choice of methods. |
75 |
* |
76 |
* <table BORDER CELLPADDING=3 CELLSPACING=1> |
77 |
* <caption>Summary of task execution methods</caption> |
78 |
* <tr> |
79 |
* <td></td> |
80 |
* <td ALIGN=CENTER> <b>Call from non-fork/join clients</b></td> |
81 |
* <td ALIGN=CENTER> <b>Call from within fork/join computations</b></td> |
82 |
* </tr> |
83 |
* <tr> |
84 |
* <td> <b>Arrange async execution</b></td> |
85 |
* <td> {@link #execute(ForkJoinTask)}</td> |
86 |
* <td> {@link ForkJoinTask#fork}</td> |
87 |
* </tr> |
88 |
* <tr> |
89 |
* <td> <b>Await and obtain result</b></td> |
90 |
* <td> {@link #invoke(ForkJoinTask)}</td> |
91 |
* <td> {@link ForkJoinTask#invoke}</td> |
92 |
* </tr> |
93 |
* <tr> |
94 |
* <td> <b>Arrange exec and obtain Future</b></td> |
95 |
* <td> {@link #submit(ForkJoinTask)}</td> |
96 |
* <td> {@link ForkJoinTask#fork} (ForkJoinTasks <em>are</em> Futures)</td> |
97 |
* </tr> |
98 |
* </table> |
99 |
* |
100 |
* <p>The common pool is by default constructed with default |
101 |
* parameters, but these may be controlled by setting three |
102 |
* {@linkplain System#getProperty system properties}: |
103 |
* <ul> |
104 |
* <li>{@code java.util.concurrent.ForkJoinPool.common.parallelism} |
105 |
* - the parallelism level, a non-negative integer |
106 |
* <li>{@code java.util.concurrent.ForkJoinPool.common.threadFactory} |
107 |
* - the class name of a {@link ForkJoinWorkerThreadFactory} |
108 |
* <li>{@code java.util.concurrent.ForkJoinPool.common.exceptionHandler} |
109 |
* - the class name of a {@link UncaughtExceptionHandler} |
110 |
* <li>{@code java.util.concurrent.ForkJoinPool.common.maximumSpares} |
111 |
* - the maximum number of allowed extra threads to maintain target |
112 |
* parallelism (default 256). |
113 |
* </ul> |
114 |
* If a {@link SecurityManager} is present and no factory is |
115 |
* specified, then the default pool uses a factory supplying |
116 |
* threads that have no {@link Permissions} enabled. |
117 |
* The system class loader is used to load these classes. |
118 |
* Upon any error in establishing these settings, default parameters |
119 |
* are used. It is possible to disable or limit the use of threads in |
120 |
* the common pool by setting the parallelism property to zero, and/or |
121 |
* using a factory that may return {@code null}. However doing so may |
122 |
* cause unjoined tasks to never be executed. |
123 |
* |
124 |
* <p><b>Implementation notes</b>: This implementation restricts the |
125 |
* maximum number of running threads to 32767. Attempts to create |
126 |
* pools with greater than the maximum number result in |
127 |
* {@code IllegalArgumentException}. |
128 |
* |
129 |
* <p>This implementation rejects submitted tasks (that is, by throwing |
130 |
* {@link RejectedExecutionException}) only when the pool is shut down |
131 |
* or internal resources have been exhausted. |
132 |
* |
133 |
* @since 1.7 |
134 |
* @author Doug Lea |
135 |
*/ |
136 |
@sun.misc.Contended |
137 |
public class ForkJoinPool extends AbstractExecutorService { |
138 |
|
139 |
/* |
140 |
* Implementation Overview |
141 |
* |
142 |
* This class and its nested classes provide the main |
143 |
* functionality and control for a set of worker threads: |
144 |
* Submissions from non-FJ threads enter into submission queues. |
145 |
* Workers take these tasks and typically split them into subtasks |
146 |
* that may be stolen by other workers. Preference rules give |
147 |
* first priority to processing tasks from their own queues (LIFO |
148 |
* or FIFO, depending on mode), then to randomized FIFO steals of |
149 |
* tasks in other queues. This framework began as vehicle for |
150 |
* supporting tree-structured parallelism using work-stealing. |
151 |
* Over time, its scalability advantages led to extensions and |
152 |
* changes to better support more diverse usage contexts. Because |
153 |
* most internal methods and nested classes are interrelated, |
154 |
* their main rationale and descriptions are presented here; |
155 |
* individual methods and nested classes contain only brief |
156 |
* comments about details. |
157 |
* |
158 |
* WorkQueues |
159 |
* ========== |
160 |
* |
161 |
* Most operations occur within work-stealing queues (in nested |
162 |
* class WorkQueue). These are special forms of Deques that |
163 |
* support only three of the four possible end-operations -- push, |
164 |
* pop, and poll (aka steal), under the further constraints that |
165 |
* push and pop are called only from the owning thread (or, as |
166 |
* extended here, under a lock), while poll may be called from |
167 |
* other threads. (If you are unfamiliar with them, you probably |
168 |
* want to read Herlihy and Shavit's book "The Art of |
169 |
* Multiprocessor programming", chapter 16 describing these in |
170 |
* more detail before proceeding.) The main work-stealing queue |
171 |
* design is roughly similar to those in the papers "Dynamic |
172 |
* Circular Work-Stealing Deque" by Chase and Lev, SPAA 2005 |
173 |
* (http://research.sun.com/scalable/pubs/index.html) and |
174 |
* "Idempotent work stealing" by Michael, Saraswat, and Vechev, |
175 |
* PPoPP 2009 (http://portal.acm.org/citation.cfm?id=1504186). |
176 |
* The main differences ultimately stem from GC requirements that |
177 |
* we null out taken slots as soon as we can, to maintain as small |
178 |
* a footprint as possible even in programs generating huge |
179 |
* numbers of tasks. To accomplish this, we shift the CAS |
180 |
* arbitrating pop vs poll (steal) from being on the indices |
181 |
* ("base" and "top") to the slots themselves. |
182 |
* |
183 |
* Adding tasks then takes the form of a classic array push(task) |
184 |
* in a circular buffer: |
185 |
* q.array[q.top++ % length] = task; |
186 |
* |
187 |
* (The actual code needs to null-check and size-check the array, |
188 |
* uses masking, not mod, for indexing a power-of-two-sized array, |
189 |
* properly fences accesses, and possibly signals waiting workers |
190 |
* to start scanning -- see below.) Both a successful pop and |
191 |
* poll mainly entail a CAS of a slot from non-null to null. |
192 |
* |
193 |
* The pop operation (always performed by owner) is: |
194 |
* if ((the task at top slot is not null) and |
195 |
* (CAS slot to null)) |
196 |
* decrement top and return task; |
197 |
* |
198 |
* And the poll operation (usually by a stealer) is |
199 |
* if ((the task at base slot is not null) and |
200 |
* (CAS slot to null)) |
201 |
* increment base and return task; |
202 |
* |
203 |
* There are several variants of each of these; for example most |
204 |
* versions of poll pre-screen the CAS by rechecking that the base |
205 |
* has not changed since reading the slot, and most methods only |
206 |
* attempt the CAS if base appears not to be equal to top. |
207 |
* |
208 |
* Memory ordering. See "Correct and Efficient Work-Stealing for |
209 |
* Weak Memory Models" by Le, Pop, Cohen, and Nardelli, PPoPP 2013 |
210 |
* (http://www.di.ens.fr/~zappa/readings/ppopp13.pdf) for an |
211 |
* analysis of memory ordering requirements in work-stealing |
212 |
* algorithms similar to (but different than) the one used here. |
213 |
* Extracting tasks in array slots via (fully fenced) CAS provides |
214 |
* primary synchronization. The base and top indices imprecisely |
215 |
* guide where to extract from. We do not always require strict |
216 |
* orderings of array and index updates, so sometimes let them be |
217 |
* subject to compiler and processor reorderings. However, the |
218 |
* volatile "base" index also serves as a basis for memory |
219 |
* ordering: Slot accesses are preceded by a read of base, |
220 |
* ensuring happens-before ordering with respect to stealers (so |
221 |
* the slots themselves can be read via plain array reads.) The |
222 |
* only other memory orderings relied on are maintained in the |
223 |
* course of signalling and activation (see below). A check that |
224 |
* base == top indicates (momentary) emptiness, but otherwise may |
225 |
* err on the side of possibly making the queue appear nonempty |
226 |
* when a push, pop, or poll have not fully committed, or making |
227 |
* it appear empty when an update of top has not yet been visibly |
228 |
* written. (Method isEmpty() checks the case of a partially |
229 |
* completed removal of the last element.) Because of this, the |
230 |
* poll operation, considered individually, is not wait-free. One |
231 |
* thief cannot successfully continue until another in-progress |
232 |
* one (or, if previously empty, a push) visibly completes. |
233 |
* However, in the aggregate, we ensure at least probabilistic |
234 |
* non-blockingness. If an attempted steal fails, a scanning |
235 |
* thief chooses a different random victim target to try next. So, |
236 |
* in order for one thief to progress, it suffices for any |
237 |
* in-progress poll or new push on any empty queue to |
238 |
* complete. (This is why we normally use method pollAt and its |
239 |
* variants that try once at the apparent base index, else |
240 |
* consider alternative actions, rather than method poll, which |
241 |
* retries.) |
242 |
* |
243 |
* This approach also enables support of a user mode in which |
244 |
* local task processing is in FIFO, not LIFO order, simply by |
245 |
* using poll rather than pop. This can be useful in |
246 |
* message-passing frameworks in which tasks are never joined. |
247 |
* |
248 |
* WorkQueues are also used in a similar way for tasks submitted |
249 |
* to the pool. We cannot mix these tasks in the same queues used |
250 |
* by workers. Instead, we randomly associate submission queues |
251 |
* with submitting threads, using a form of hashing. The |
252 |
* ThreadLocalRandom probe value serves as a hash code for |
253 |
* choosing existing queues, and may be randomly repositioned upon |
254 |
* contention with other submitters. In essence, submitters act |
255 |
* like workers except that they are restricted to executing local |
256 |
* tasks that they submitted (or in the case of CountedCompleters, |
257 |
* others with the same root task). Insertion of tasks in shared |
258 |
* mode requires a lock but we use only a simple spinlock (using |
259 |
* field qlock), because submitters encountering a busy queue move |
260 |
* on to try or create other queues -- they block only when |
261 |
* creating and registering new queues. Because it is used only as |
262 |
* a spinlock, unlocking requires only a "releasing" store (using |
263 |
* putOrderedInt). The qlock is also used during termination |
264 |
* detection, in which case it is forced to a negative |
265 |
* nonlockable value. |
266 |
* |
267 |
* Management |
268 |
* ========== |
269 |
* |
270 |
* The main throughput advantages of work-stealing stem from |
271 |
* decentralized control -- workers mostly take tasks from |
272 |
* themselves or each other, at rates that can exceed a billion |
273 |
* per second. The pool itself creates, activates (enables |
274 |
* scanning for and running tasks), deactivates, blocks, and |
275 |
* terminates threads, all with minimal central information. |
276 |
* There are only a few properties that we can globally track or |
277 |
* maintain, so we pack them into a small number of variables, |
278 |
* often maintaining atomicity without blocking or locking. |
279 |
* Nearly all essentially atomic control state is held in two |
280 |
* volatile variables that are by far most often read (not |
281 |
* written) as status and consistency checks. (Also, field |
282 |
* "config" holds unchanging configuration state.) |
283 |
* |
284 |
* Field "ctl" contains 64 bits holding information needed to |
285 |
* atomically decide to add, inactivate, enqueue (on an event |
286 |
* queue), dequeue, and/or re-activate workers. To enable this |
287 |
* packing, we restrict maximum parallelism to (1<<15)-1 (which is |
288 |
* far in excess of normal operating range) to allow ids, counts, |
289 |
* and their negations (used for thresholding) to fit into 16bit |
290 |
* subfields. |
291 |
* |
292 |
* Field "runState" holds state bits (STARTED, STOP, etc). After |
293 |
* starting, the field is updated (only during shutdown) under the |
294 |
* auxState lock. |
295 |
* |
296 |
* Field "auxState" is a ReentrantLock subclass that also |
297 |
* opportunistically holds some other bookkeeping fields accessed |
298 |
* only when locked. It is mainly used to lock (infrequent) |
299 |
* updates to runState and workQueues. The auxState instance is |
300 |
* itself lazily constructed (see tryInitialize), requiring a |
301 |
* double-check-style bootstrapping use of field runState, and |
302 |
* locking a private static. |
303 |
* |
304 |
* Field "workQueues" holds references to WorkQueues. It is |
305 |
* updated (only during worker creation and termination) under the |
306 |
* lock, but is otherwise concurrently readable, and accessed |
307 |
* directly. We also ensure that reads of the array reference |
308 |
* itself never become too stale (for example, re-reading before |
309 |
* each scan). To simplify index-based operations, the array size |
310 |
* is always a power of two, and all readers must tolerate null |
311 |
* slots. Worker queues are at odd indices. Shared (submission) |
312 |
* queues are at even indices, up to a maximum of 64 slots, to |
313 |
* limit growth even if array needs to expand to add more |
314 |
* workers. Grouping them together in this way simplifies and |
315 |
* speeds up task scanning. |
316 |
* |
317 |
* All worker thread creation is on-demand, triggered by task |
318 |
* submissions, replacement of terminated workers, and/or |
319 |
* compensation for blocked workers. However, all other support |
320 |
* code is set up to work with other policies. To ensure that we |
321 |
* do not hold on to worker references that would prevent GC, All |
322 |
* accesses to workQueues are via indices into the workQueues |
323 |
* array (which is one source of some of the messy code |
324 |
* constructions here). In essence, the workQueues array serves as |
325 |
* a weak reference mechanism. Thus for example the stack top |
326 |
* subfield of ctl stores indices, not references. |
327 |
* |
328 |
* Queuing Idle Workers. Unlike HPC work-stealing frameworks, we |
329 |
* cannot let workers spin indefinitely scanning for tasks when |
330 |
* none can be found immediately, and we cannot start/resume |
331 |
* workers unless there appear to be tasks available. On the |
332 |
* other hand, we must quickly prod them into action when new |
333 |
* tasks are submitted or generated. In many usages, ramp-up time |
334 |
* to activate workers is the main limiting factor in overall |
335 |
* performance, which is compounded at program start-up by JIT |
336 |
* compilation and allocation. So we streamline this as much as |
337 |
* possible. |
338 |
* |
339 |
* The "ctl" field atomically maintains active and total worker |
340 |
* counts as well as a queue to place waiting threads so they can |
341 |
* be located for signalling. Active counts also play the role of |
342 |
* quiescence indicators, so are decremented when workers believe |
343 |
* that there are no more tasks to execute. The "queue" is |
344 |
* actually a form of Treiber stack. A stack is ideal for |
345 |
* activating threads in most-recently used order. This improves |
346 |
* performance and locality, outweighing the disadvantages of |
347 |
* being prone to contention and inability to release a worker |
348 |
* unless it is topmost on stack. We block/unblock workers after |
349 |
* pushing on the idle worker stack (represented by the lower |
350 |
* 32bit subfield of ctl) when they cannot find work. The top |
351 |
* stack state holds the value of the "scanState" field of the |
352 |
* worker: its index and status, plus a version counter that, in |
353 |
* addition to the count subfields (also serving as version |
354 |
* stamps) provide protection against Treiber stack ABA effects. |
355 |
* |
356 |
* Creating workers. To create a worker, we pre-increment total |
357 |
* count (serving as a reservation), and attempt to construct a |
358 |
* ForkJoinWorkerThread via its factory. Upon construction, the |
359 |
* new thread invokes registerWorker, where it constructs a |
360 |
* WorkQueue and is assigned an index in the workQueues array |
361 |
* (expanding the array if necessary). The thread is then |
362 |
* started. Upon any exception across these steps, or null return |
363 |
* from factory, deregisterWorker adjusts counts and records |
364 |
* accordingly. If a null return, the pool continues running with |
365 |
* fewer than the target number workers. If exceptional, the |
366 |
* exception is propagated, generally to some external caller. |
367 |
* Worker index assignment avoids the bias in scanning that would |
368 |
* occur if entries were sequentially packed starting at the front |
369 |
* of the workQueues array. We treat the array as a simple |
370 |
* power-of-two hash table, expanding as needed. The seedIndex |
371 |
* increment ensures no collisions until a resize is needed or a |
372 |
* worker is deregistered and replaced, and thereafter keeps |
373 |
* probability of collision low. We cannot use |
374 |
* ThreadLocalRandom.getProbe() for similar purposes here because |
375 |
* the thread has not started yet, but do so for creating |
376 |
* submission queues for existing external threads (see |
377 |
* externalPush). |
378 |
* |
379 |
* WorkQueue field scanState is used by both workers and the pool |
380 |
* to manage and track whether a worker is UNSIGNALLED (possibly |
381 |
* blocked waiting for a signal). When a worker is inactivated, |
382 |
* its scanState field is set, and is prevented from executing |
383 |
* tasks, even though it must scan once for them to avoid queuing |
384 |
* races. Note that scanState updates lag queue CAS releases so |
385 |
* usage requires care. When queued, the lower 16 bits of |
386 |
* scanState must hold its pool index. So we place the index there |
387 |
* upon initialization (see registerWorker) and otherwise keep it |
388 |
* there or restore it when necessary. |
389 |
* |
390 |
* The ctl field also serves as the basis for memory |
391 |
* synchronization surrounding activation. This uses a more |
392 |
* efficient version of a Dekker-like rule that task producers and |
393 |
* consumers sync with each other by both writing/CASing ctl (even |
394 |
* if to its current value). This would be extremely costly. So |
395 |
* we relax it in several ways: (1) Producers only signal when |
396 |
* their queue is empty. Other workers propagate this signal (in |
397 |
* method scan) when they find tasks. (2) Workers only enqueue |
398 |
* after scanning (see below) and not finding any tasks. (3) |
399 |
* Rather than CASing ctl to its current value in the common case |
400 |
* where no action is required, we reduce write contention by |
401 |
* equivalently prefacing signalWork when called by an external |
402 |
* task producer using a memory access with full-volatile |
403 |
* semantics or a "fullFence". (4) For internal task producers we |
404 |
* rely on the fact that even if no other workers awaken, the |
405 |
* producer itself will eventually see the task and execute it. |
406 |
* |
407 |
* Almost always, too many signals are issued. A task producer |
408 |
* cannot in general tell if some existing worker is in the midst |
409 |
* of finishing one task (or already scanning) and ready to take |
410 |
* another without being signalled. So the producer might instead |
411 |
* activate a different worker that does not find any work, and |
412 |
* then inactivates. This scarcely matters in steady-state |
413 |
* computations involving all workers, but can create contention |
414 |
* and bookkeeping bottlenecks during ramp-up, ramp-down, and small |
415 |
* computations involving only a few workers. |
416 |
* |
417 |
* Scanning. Method scan() performs top-level scanning for tasks. |
418 |
* Each scan traverses (and tries to poll from) each queue in |
419 |
* pseudorandom permutation order by randomly selecting an origin |
420 |
* index and a step value. (The pseudorandom generator need not |
421 |
* have high-quality statistical properties in the long term, but |
422 |
* just within computations; We use 64bit and 32bit Marsaglia |
423 |
* XorShifts, which are cheap and suffice here.) Scanning also |
424 |
* employs contention reduction: When scanning workers fail a CAS |
425 |
* polling for work, they soon restart with a different |
426 |
* pseudorandom scan order (thus likely retrying at different |
427 |
* intervals). This improves throughput when many threads are |
428 |
* trying to take tasks from few queues. Scans do not otherwise |
429 |
* explicitly take into account core affinities, loads, cache |
430 |
* localities, etc, However, they do exploit temporal locality |
431 |
* (which usually approximates these) by preferring to re-poll (up |
432 |
* to POLL_LIMIT times) from the same queue after a successful |
433 |
* poll before trying others. Restricted forms of scanning occur |
434 |
* in methods helpComplete and findNonEmptyStealQueue, and take |
435 |
* similar but simpler forms. |
436 |
* |
437 |
* Deactivation and waiting. Queuing encounters several intrinsic |
438 |
* races; most notably that a inactivating scanning worker can |
439 |
* miss seeing a task produced during a scan. So when a worker |
440 |
* cannot find a task to steal, it inactivates and enqueues, and |
441 |
* then rescans to ensure that it didn't miss one, reactivating |
442 |
* upon seeing one with probability approximately proportional to |
443 |
* probability of a miss. (In most cases, the worker will be |
444 |
* signalled before self-signalling, avoiding cascades of multiple |
445 |
* signals for the same task). |
446 |
* |
447 |
* Workers block (in method awaitWork) using park/unpark; |
448 |
* advertising the need for signallers to unpark by setting their |
449 |
* "parker" fields. |
450 |
* |
451 |
* Trimming workers. To release resources after periods of lack of |
452 |
* use, a worker starting to wait when the pool is quiescent will |
453 |
* time out and terminate (see awaitWork) if the pool has remained |
454 |
* quiescent for period IDLE_TIMEOUT, increasing the period as the |
455 |
* number of threads decreases, eventually removing all workers. |
456 |
* |
457 |
* Shutdown and Termination. A call to shutdownNow invokes |
458 |
* tryTerminate to atomically set a runState bit. The calling |
459 |
* thread, as well as every other worker thereafter terminating, |
460 |
* helps terminate others by setting their (qlock) status, |
461 |
* cancelling their unprocessed tasks, and waking them up, doing |
462 |
* so repeatedly until stable (but with a loop bounded by the |
463 |
* number of workers). Calls to non-abrupt shutdown() preface |
464 |
* this by checking whether termination should commence. This |
465 |
* relies primarily on the active count bits of "ctl" maintaining |
466 |
* consensus -- tryTerminate is called from awaitWork whenever |
467 |
* quiescent. However, external submitters do not take part in |
468 |
* this consensus. So, tryTerminate sweeps through queues (until |
469 |
* stable) to ensure lack of in-flight submissions and workers |
470 |
* about to process them before triggering the "STOP" phase of |
471 |
* termination. (Note: there is an intrinsic conflict if |
472 |
* helpQuiescePool is called when shutdown is enabled. Both wait |
473 |
* for quiescence, but tryTerminate is biased to not trigger until |
474 |
* helpQuiescePool completes.) |
475 |
* |
476 |
* Joining Tasks |
477 |
* ============= |
478 |
* |
479 |
* Any of several actions may be taken when one worker is waiting |
480 |
* to join a task stolen (or always held) by another. Because we |
481 |
* are multiplexing many tasks on to a pool of workers, we can't |
482 |
* just let them block (as in Thread.join). We also cannot just |
483 |
* reassign the joiner's run-time stack with another and replace |
484 |
* it later, which would be a form of "continuation", that even if |
485 |
* possible is not necessarily a good idea since we may need both |
486 |
* an unblocked task and its continuation to progress. Instead we |
487 |
* combine two tactics: |
488 |
* |
489 |
* Helping: Arranging for the joiner to execute some task that it |
490 |
* would be running if the steal had not occurred. |
491 |
* |
492 |
* Compensating: Unless there are already enough live threads, |
493 |
* method tryCompensate() may create or re-activate a spare |
494 |
* thread to compensate for blocked joiners until they unblock. |
495 |
* |
496 |
* A third form (implemented in tryRemoveAndExec) amounts to |
497 |
* helping a hypothetical compensator: If we can readily tell that |
498 |
* a possible action of a compensator is to steal and execute the |
499 |
* task being joined, the joining thread can do so directly, |
500 |
* without the need for a compensation thread (although at the |
501 |
* expense of larger run-time stacks, but the tradeoff is |
502 |
* typically worthwhile). |
503 |
* |
504 |
* The ManagedBlocker extension API can't use helping so relies |
505 |
* only on compensation in method awaitBlocker. |
506 |
* |
507 |
* The algorithm in helpStealer entails a form of "linear |
508 |
* helping". Each worker records (in field currentSteal) the most |
509 |
* recent task it stole from some other worker (or a submission). |
510 |
* It also records (in field currentJoin) the task it is currently |
511 |
* actively joining. Method helpStealer uses these markers to try |
512 |
* to find a worker to help (i.e., steal back a task from and |
513 |
* execute it) that could hasten completion of the actively joined |
514 |
* task. Thus, the joiner executes a task that would be on its |
515 |
* own local deque had the to-be-joined task not been stolen. This |
516 |
* is a conservative variant of the approach described in Wagner & |
517 |
* Calder "Leapfrogging: a portable technique for implementing |
518 |
* efficient futures" SIGPLAN Notices, 1993 |
519 |
* (http://portal.acm.org/citation.cfm?id=155354). It differs in |
520 |
* that: (1) We only maintain dependency links across workers upon |
521 |
* steals, rather than use per-task bookkeeping. This sometimes |
522 |
* requires a linear scan of workQueues array to locate stealers, |
523 |
* but often doesn't because stealers leave hints (that may become |
524 |
* stale/wrong) of where to locate them. It is only a hint |
525 |
* because a worker might have had multiple steals and the hint |
526 |
* records only one of them (usually the most current). Hinting |
527 |
* isolates cost to when it is needed, rather than adding to |
528 |
* per-task overhead. (2) It is "shallow", ignoring nesting and |
529 |
* potentially cyclic mutual steals. (3) It is intentionally |
530 |
* racy: field currentJoin is updated only while actively joining, |
531 |
* which means that we miss links in the chain during long-lived |
532 |
* tasks, GC stalls etc (which is OK since blocking in such cases |
533 |
* is usually a good idea). (4) We bound the number of attempts |
534 |
* to find work using checksums and fall back to suspending the |
535 |
* worker and if necessary replacing it with another. |
536 |
* |
537 |
* Helping actions for CountedCompleters do not require tracking |
538 |
* currentJoins: Method helpComplete takes and executes any task |
539 |
* with the same root as the task being waited on (preferring |
540 |
* local pops to non-local polls). However, this still entails |
541 |
* some traversal of completer chains, so is less efficient than |
542 |
* using CountedCompleters without explicit joins. |
543 |
* |
544 |
* Compensation does not aim to keep exactly the target |
545 |
* parallelism number of unblocked threads running at any given |
546 |
* time. Some previous versions of this class employed immediate |
547 |
* compensations for any blocked join. However, in practice, the |
548 |
* vast majority of blockages are transient byproducts of GC and |
549 |
* other JVM or OS activities that are made worse by replacement. |
550 |
* Currently, compensation is attempted only after validating that |
551 |
* all purportedly active threads are processing tasks by checking |
552 |
* field WorkQueue.scanState, which eliminates most false |
553 |
* positives. Also, compensation is bypassed (tolerating fewer |
554 |
* threads) in the most common case in which it is rarely |
555 |
* beneficial: when a worker with an empty queue (thus no |
556 |
* continuation tasks) blocks on a join and there still remain |
557 |
* enough threads to ensure liveness. |
558 |
* |
559 |
* Spare threads are removed as soon as they notice that the |
560 |
* target parallelism level has been exceeded, in method |
561 |
* tryDropSpare. (Method scan arranges returns for rechecks upon |
562 |
* each probe via the "bound" parameter.) |
563 |
* |
564 |
* The compensation mechanism may be bounded. Bounds for the |
565 |
* commonPool (see commonMaxSpares) better enable JVMs to cope |
566 |
* with programming errors and abuse before running out of |
567 |
* resources to do so. In other cases, users may supply factories |
568 |
* that limit thread construction. The effects of bounding in this |
569 |
* pool (like all others) is imprecise. Total worker counts are |
570 |
* decremented when threads deregister, not when they exit and |
571 |
* resources are reclaimed by the JVM and OS. So the number of |
572 |
* simultaneously live threads may transiently exceed bounds. |
573 |
* |
574 |
* Common Pool |
575 |
* =========== |
576 |
* |
577 |
* The static common pool always exists after static |
578 |
* initialization. Since it (or any other created pool) need |
579 |
* never be used, we minimize initial construction overhead and |
580 |
* footprint to the setup of about a dozen fields, with no nested |
581 |
* allocation. Most bootstrapping occurs within method |
582 |
* externalSubmit during the first submission to the pool. |
583 |
* |
584 |
* When external threads submit to the common pool, they can |
585 |
* perform subtask processing (see externalHelpComplete and |
586 |
* related methods) upon joins. This caller-helps policy makes it |
587 |
* sensible to set common pool parallelism level to one (or more) |
588 |
* less than the total number of available cores, or even zero for |
589 |
* pure caller-runs. We do not need to record whether external |
590 |
* submissions are to the common pool -- if not, external help |
591 |
* methods return quickly. These submitters would otherwise be |
592 |
* blocked waiting for completion, so the extra effort (with |
593 |
* liberally sprinkled task status checks) in inapplicable cases |
594 |
* amounts to an odd form of limited spin-wait before blocking in |
595 |
* ForkJoinTask.join. |
596 |
* |
597 |
* As a more appropriate default in managed environments, unless |
598 |
* overridden by system properties, we use workers of subclass |
599 |
* InnocuousForkJoinWorkerThread when there is a SecurityManager |
600 |
* present. These workers have no permissions set, do not belong |
601 |
* to any user-defined ThreadGroup, and erase all ThreadLocals |
602 |
* after executing any top-level task (see WorkQueue.runTask). |
603 |
* The associated mechanics (mainly in ForkJoinWorkerThread) may |
604 |
* be JVM-dependent and must access particular Thread class fields |
605 |
* to achieve this effect. |
606 |
* |
607 |
* Style notes |
608 |
* =========== |
609 |
* |
610 |
* Memory ordering relies mainly on Unsafe intrinsics that carry |
611 |
* the further responsibility of explicitly performing null- and |
612 |
* bounds- checks otherwise carried out implicitly by JVMs. This |
613 |
* can be awkward and ugly, but also reflects the need to control |
614 |
* outcomes across the unusual cases that arise in very racy code |
615 |
* with very few invariants. So these explicit checks would exist |
616 |
* in some form anyway. All fields are read into locals before |
617 |
* use, and null-checked if they are references. This is usually |
618 |
* done in a "C"-like style of listing declarations at the heads |
619 |
* of methods or blocks, and using inline assignments on first |
620 |
* encounter. Array bounds-checks are usually performed by |
621 |
* masking with array.length-1, which relies on the invariant that |
622 |
* these arrays are created with positive lengths, which is itself |
623 |
* paranoically checked. Nearly all explicit checks lead to |
624 |
* bypass/return, not exception throws, because they may |
625 |
* legitimately arise due to cancellation/revocation during |
626 |
* shutdown. |
627 |
* |
628 |
* There is a lot of representation-level coupling among classes |
629 |
* ForkJoinPool, ForkJoinWorkerThread, and ForkJoinTask. The |
630 |
* fields of WorkQueue maintain data structures managed by |
631 |
* ForkJoinPool, so are directly accessed. There is little point |
632 |
* trying to reduce this, since any associated future changes in |
633 |
* representations will need to be accompanied by algorithmic |
634 |
* changes anyway. Several methods intrinsically sprawl because |
635 |
* they must accumulate sets of consistent reads of fields held in |
636 |
* local variables. There are also other coding oddities |
637 |
* (including several unnecessary-looking hoisted null checks) |
638 |
* that help some methods perform reasonably even when interpreted |
639 |
* (not compiled). |
640 |
* |
641 |
* The order of declarations in this file is (with a few exceptions): |
642 |
* (1) Static utility functions |
643 |
* (2) Nested (static) classes |
644 |
* (3) Static fields |
645 |
* (4) Fields, along with constants used when unpacking some of them |
646 |
* (5) Internal control methods |
647 |
* (6) Callbacks and other support for ForkJoinTask methods |
648 |
* (7) Exported methods |
649 |
* (8) Static block initializing statics in minimally dependent order |
650 |
*/ |
651 |
|
652 |
// Static utilities |
653 |
|
654 |
/** |
655 |
* If there is a security manager, makes sure caller has |
656 |
* permission to modify threads. |
657 |
*/ |
658 |
private static void checkPermission() { |
659 |
SecurityManager security = System.getSecurityManager(); |
660 |
if (security != null) |
661 |
security.checkPermission(modifyThreadPermission); |
662 |
} |
663 |
|
664 |
// Nested classes |
665 |
|
666 |
/** |
667 |
* Factory for creating new {@link ForkJoinWorkerThread}s. |
668 |
* A {@code ForkJoinWorkerThreadFactory} must be defined and used |
669 |
* for {@code ForkJoinWorkerThread} subclasses that extend base |
670 |
* functionality or initialize threads with different contexts. |
671 |
*/ |
672 |
public static interface ForkJoinWorkerThreadFactory { |
673 |
/** |
674 |
* Returns a new worker thread operating in the given pool. |
675 |
* |
676 |
* @param pool the pool this thread works in |
677 |
* @return the new worker thread |
678 |
* @throws NullPointerException if the pool is null |
679 |
*/ |
680 |
public ForkJoinWorkerThread newThread(ForkJoinPool pool); |
681 |
} |
682 |
|
683 |
/** |
684 |
* Default ForkJoinWorkerThreadFactory implementation; creates a |
685 |
* new ForkJoinWorkerThread. |
686 |
*/ |
687 |
static final class DefaultForkJoinWorkerThreadFactory |
688 |
implements ForkJoinWorkerThreadFactory { |
689 |
public final ForkJoinWorkerThread newThread(ForkJoinPool pool) { |
690 |
return new ForkJoinWorkerThread(pool); |
691 |
} |
692 |
} |
693 |
|
694 |
/** |
695 |
* Class for artificial tasks that are used to replace the target |
696 |
* of local joins if they are removed from an interior queue slot |
697 |
* in WorkQueue.tryRemoveAndExec. We don't need the proxy to |
698 |
* actually do anything beyond having a unique identity. |
699 |
*/ |
700 |
static final class EmptyTask extends ForkJoinTask<Void> { |
701 |
private static final long serialVersionUID = -7721805057305804111L; |
702 |
EmptyTask() { status = ForkJoinTask.NORMAL; } // force done |
703 |
public final Void getRawResult() { return null; } |
704 |
public final void setRawResult(Void x) {} |
705 |
public final boolean exec() { return true; } |
706 |
} |
707 |
|
708 |
/** |
709 |
* Additional fields and lock created upon initialization. |
710 |
*/ |
711 |
static final class AuxState extends ReentrantLock { |
712 |
private static final long serialVersionUID = -6001602636862214147L; |
713 |
volatile long stealCount; // cumulative steal count |
714 |
long indexSeed; // index bits for registerWorker |
715 |
AuxState() {} |
716 |
} |
717 |
|
718 |
// Constants shared across ForkJoinPool and WorkQueue |
719 |
|
720 |
// Bounds |
721 |
static final int SMASK = 0xffff; // short bits == max index |
722 |
static final int MAX_CAP = 0x7fff; // max #workers - 1 |
723 |
static final int EVENMASK = 0xfffe; // even short bits |
724 |
static final int SQMASK = 0x007e; // max 64 (even) slots |
725 |
|
726 |
// Masks and units for WorkQueue.scanState and ctl sp subfield |
727 |
static final int UNSIGNALLED = 1 << 31; // must be negative |
728 |
static final int SS_SEQ = 1 << 16; // version count |
729 |
|
730 |
// Mode bits for ForkJoinPool.config and WorkQueue.config |
731 |
static final int MODE_MASK = 0xffff << 16; // top half of int |
732 |
static final int SPARE_WORKER = 1 << 17; // set if tc > 0 on creation |
733 |
static final int UNREGISTERED = 1 << 18; // to skip some of deregister |
734 |
static final int FIFO_QUEUE = 1 << 31; // must be negative |
735 |
static final int LIFO_QUEUE = 0; // for clarity |
736 |
static final int IS_OWNED = 1; // low bit 0 if shared |
737 |
|
738 |
/** |
739 |
* The maximum number of task executions from the same queue |
740 |
* before checking other queues, bounding unfairness and impact of |
741 |
* infinite user task recursion. Must be a power of two minus 1. |
742 |
*/ |
743 |
static final int POLL_LIMIT = (1 << 10) - 1; |
744 |
|
745 |
/** |
746 |
* Queues supporting work-stealing as well as external task |
747 |
* submission. See above for descriptions and algorithms. |
748 |
* Performance on most platforms is very sensitive to placement of |
749 |
* instances of both WorkQueues and their arrays -- we absolutely |
750 |
* do not want multiple WorkQueue instances or multiple queue |
751 |
* arrays sharing cache lines. The @Contended annotation alerts |
752 |
* JVMs to try to keep instances apart. |
753 |
*/ |
754 |
@sun.misc.Contended |
755 |
static final class WorkQueue { |
756 |
|
757 |
/** |
758 |
* Capacity of work-stealing queue array upon initialization. |
759 |
* Must be a power of two; at least 4, but should be larger to |
760 |
* reduce or eliminate cacheline sharing among queues. |
761 |
* Currently, it is much larger, as a partial workaround for |
762 |
* the fact that JVMs often place arrays in locations that |
763 |
* share GC bookkeeping (especially cardmarks) such that |
764 |
* per-write accesses encounter serious memory contention. |
765 |
*/ |
766 |
static final int INITIAL_QUEUE_CAPACITY = 1 << 13; |
767 |
|
768 |
/** |
769 |
* Maximum size for queue arrays. Must be a power of two less |
770 |
* than or equal to 1 << (31 - width of array entry) to ensure |
771 |
* lack of wraparound of index calculations, but defined to a |
772 |
* value a bit less than this to help users trap runaway |
773 |
* programs before saturating systems. |
774 |
*/ |
775 |
static final int MAXIMUM_QUEUE_CAPACITY = 1 << 26; // 64M |
776 |
|
777 |
// Instance fields |
778 |
|
779 |
volatile int scanState; // versioned, negative if inactive |
780 |
int stackPred; // pool stack (ctl) predecessor |
781 |
int nsteals; // number of steals |
782 |
int hint; // randomization and stealer index hint |
783 |
int config; // pool index and mode |
784 |
volatile int qlock; // 1: locked, < 0: terminate; else 0 |
785 |
volatile int base; // index of next slot for poll |
786 |
int top; // index of next slot for push |
787 |
ForkJoinTask<?>[] array; // the elements (initially unallocated) |
788 |
final ForkJoinPool pool; // the containing pool (may be null) |
789 |
final ForkJoinWorkerThread owner; // owning thread or null if shared |
790 |
volatile Thread parker; // == owner during call to park; else null |
791 |
volatile ForkJoinTask<?> currentJoin; // task being joined in awaitJoin |
792 |
@sun.misc.Contended("group2") // separate from other fields |
793 |
volatile ForkJoinTask<?> currentSteal; // nonnull when running some task |
794 |
|
795 |
WorkQueue(ForkJoinPool pool, ForkJoinWorkerThread owner) { |
796 |
this.pool = pool; |
797 |
this.owner = owner; |
798 |
// Place indices in the center of array (that is not yet allocated) |
799 |
base = top = INITIAL_QUEUE_CAPACITY >>> 1; |
800 |
} |
801 |
|
802 |
/** |
803 |
* Returns an exportable index (used by ForkJoinWorkerThread). |
804 |
*/ |
805 |
final int getPoolIndex() { |
806 |
return (config & 0xffff) >>> 1; // ignore odd/even tag bit |
807 |
} |
808 |
|
809 |
/** |
810 |
* Returns the approximate number of tasks in the queue. |
811 |
*/ |
812 |
final int queueSize() { |
813 |
int n = base - top; // read base first |
814 |
return (n >= 0) ? 0 : -n; // ignore transient negative |
815 |
} |
816 |
|
817 |
/** |
818 |
* Provides a more accurate estimate of whether this queue has |
819 |
* any tasks than does queueSize, by checking whether a |
820 |
* near-empty queue has at least one unclaimed task. |
821 |
*/ |
822 |
final boolean isEmpty() { |
823 |
ForkJoinTask<?>[] a; int n, al, s; |
824 |
return ((n = base - (s = top)) >= 0 || // possibly one task |
825 |
(n == -1 && ((a = array) == null || |
826 |
(al = a.length) == 0 || |
827 |
a[(al - 1) & (s - 1)] == null))); |
828 |
} |
829 |
|
830 |
/** |
831 |
* Pushes a task. Call only by owner in unshared queues. |
832 |
* |
833 |
* @param task the task. Caller must ensure non-null. |
834 |
* @throws RejectedExecutionException if array cannot be resized |
835 |
*/ |
836 |
final void push(ForkJoinTask<?> task) { |
837 |
int b = base, s = top, al, d; ForkJoinTask<?>[] a; |
838 |
if ((a = array) != null && (al = a.length) > 0) { |
839 |
a[(al - 1) & s] = task; // relaxed writes OK |
840 |
top = s + 1; |
841 |
ForkJoinPool p = pool; |
842 |
U.storeFence(); // ensure fields written before use |
843 |
if ((d = b - s) == 0 && p != null) |
844 |
p.signalWork(); |
845 |
else if (al + d == 1) |
846 |
growArray(); |
847 |
} |
848 |
} |
849 |
|
850 |
/** |
851 |
* Initializes or doubles the capacity of array. Call either |
852 |
* by owner or with lock held -- it is OK for base, but not |
853 |
* top, to move while resizings are in progress. |
854 |
*/ |
855 |
final ForkJoinTask<?>[] growArray() { |
856 |
ForkJoinTask<?>[] oldA = array; |
857 |
int size = oldA != null ? oldA.length << 1 : INITIAL_QUEUE_CAPACITY; |
858 |
if (size < INITIAL_QUEUE_CAPACITY || size > MAXIMUM_QUEUE_CAPACITY) |
859 |
throw new RejectedExecutionException("Queue capacity exceeded"); |
860 |
int oldMask, t, b; |
861 |
ForkJoinTask<?>[] a = array = new ForkJoinTask<?>[size]; |
862 |
if (oldA != null && (oldMask = oldA.length - 1) > 0 && |
863 |
(t = top) - (b = base) > 0) { |
864 |
int mask = size - 1; |
865 |
do { // emulate poll from old array, push to new array |
866 |
int index = b & oldMask; |
867 |
long offset = ((long)index << ASHIFT) + ABASE; |
868 |
ForkJoinTask<?> x = (ForkJoinTask<?>) |
869 |
U.getObjectVolatile(oldA, offset); |
870 |
if (x != null && |
871 |
U.compareAndSwapObject(oldA, offset, x, null)) |
872 |
a[b & mask] = x; |
873 |
} while (++b != t); |
874 |
U.storeFence(); |
875 |
} |
876 |
return a; |
877 |
} |
878 |
|
879 |
/** |
880 |
* Takes next task, if one exists, in LIFO order. Call only |
881 |
* by owner in unshared queues. |
882 |
*/ |
883 |
final ForkJoinTask<?> pop() { |
884 |
int b = base, s = top, al, i; ForkJoinTask<?>[] a; |
885 |
if ((a = array) != null && b != s && (al = a.length) > 0) { |
886 |
int index = (al - 1) & --s; |
887 |
long offset = ((long)index << ASHIFT) + ABASE; |
888 |
ForkJoinTask<?> t = (ForkJoinTask<?>) |
889 |
U.getObject(a, offset); |
890 |
if (t != null && |
891 |
U.compareAndSwapObject(a, offset, t, null)) { |
892 |
top = s; |
893 |
U.storeFence(); |
894 |
return t; |
895 |
} |
896 |
} |
897 |
return null; |
898 |
} |
899 |
|
900 |
/** |
901 |
* Takes a task in FIFO order if b is base of queue and a task |
902 |
* can be claimed without contention. Specialized versions |
903 |
* appear in ForkJoinPool methods scan and helpStealer. |
904 |
*/ |
905 |
final ForkJoinTask<?> pollAt(int b) { |
906 |
ForkJoinTask<?>[] a; int al; |
907 |
if ((a = array) != null && (al = a.length) > 0) { |
908 |
int index = (al - 1) & b; |
909 |
long offset = ((long)index << ASHIFT) + ABASE; |
910 |
ForkJoinTask<?> t = (ForkJoinTask<?>) |
911 |
U.getObjectVolatile(a, offset); |
912 |
if (t != null && b++ == base && |
913 |
U.compareAndSwapObject(a, offset, t, null)) { |
914 |
base = b; |
915 |
return t; |
916 |
} |
917 |
} |
918 |
return null; |
919 |
} |
920 |
|
921 |
/** |
922 |
* Takes next task, if one exists, in FIFO order. |
923 |
*/ |
924 |
final ForkJoinTask<?> poll() { |
925 |
for (;;) { |
926 |
int b = base, s = top, d, al; ForkJoinTask<?>[] a; |
927 |
if ((a = array) != null && (d = b - s) < 0 && |
928 |
(al = a.length) > 0) { |
929 |
int index = (al - 1) & b; |
930 |
long offset = ((long)index << ASHIFT) + ABASE; |
931 |
ForkJoinTask<?> t = (ForkJoinTask<?>) |
932 |
U.getObjectVolatile(a, offset); |
933 |
if (b++ == base) { |
934 |
if (t != null) { |
935 |
if (U.compareAndSwapObject(a, offset, t, null)) { |
936 |
base = b; |
937 |
return t; |
938 |
} |
939 |
} |
940 |
else if (d == -1) |
941 |
break; // now empty |
942 |
} |
943 |
} |
944 |
else |
945 |
break; |
946 |
} |
947 |
return null; |
948 |
} |
949 |
|
950 |
/** |
951 |
* Takes next task, if one exists, in order specified by mode. |
952 |
*/ |
953 |
final ForkJoinTask<?> nextLocalTask() { |
954 |
return (config < 0) ? poll() : pop(); |
955 |
} |
956 |
|
957 |
/** |
958 |
* Returns next task, if one exists, in order specified by mode. |
959 |
*/ |
960 |
final ForkJoinTask<?> peek() { |
961 |
int b = base, al; ForkJoinTask<?>[] a; |
962 |
return ((a = array) != null && (al = a.length) > 0) ? |
963 |
a[(al - 1) & (config < 0 ? b : top - 1)] : null; |
964 |
} |
965 |
|
966 |
/** |
967 |
* Pops the given task only if it is at the current top. |
968 |
*/ |
969 |
final boolean tryUnpush(ForkJoinTask<?> task) { |
970 |
int b = base, s = top, al; ForkJoinTask<?>[] a; |
971 |
if ((a = array) != null && b != s && (al = a.length) > 0) { |
972 |
int index = (al - 1) & --s; |
973 |
long offset = ((long)index << ASHIFT) + ABASE; |
974 |
if (U.compareAndSwapObject(a, offset, task, null)) { |
975 |
top = s; |
976 |
U.storeFence(); |
977 |
return true; |
978 |
} |
979 |
} |
980 |
return false; |
981 |
} |
982 |
|
983 |
/** |
984 |
* Shared version of push. Fails if already locked. |
985 |
* |
986 |
* @return status: > 0 locked, 0 was empty, < 0 was nonempty |
987 |
*/ |
988 |
final int sharedPush(ForkJoinTask<?> task) { |
989 |
int stat; |
990 |
if (U.compareAndSwapInt(this, QLOCK, 0, 1)) { |
991 |
int b = base, s = top, al, d; ForkJoinTask<?>[] a; |
992 |
if ((a = array) != null && (al = a.length) > 0 && |
993 |
al - 1 + (d = b - s) > 0) { |
994 |
a[(al - 1) & s] = task; |
995 |
top = s + 1; // relaxed writes OK here |
996 |
U.putOrderedInt(this, QLOCK, 0); |
997 |
if (d < 0) |
998 |
stat = d; |
999 |
else { |
1000 |
U.fullFence(); // sync with signallees |
1001 |
stat = 0; |
1002 |
} |
1003 |
} |
1004 |
else { |
1005 |
growAndSharedPush(task); |
1006 |
stat = 0; |
1007 |
} |
1008 |
} |
1009 |
else |
1010 |
stat = 1; |
1011 |
return stat; |
1012 |
} |
1013 |
|
1014 |
/* |
1015 |
* Helper for sharedPush; called only when locked and resize |
1016 |
* needed. |
1017 |
*/ |
1018 |
private void growAndSharedPush(ForkJoinTask<?> task) { |
1019 |
try { |
1020 |
growArray(); |
1021 |
int s = top, al; ForkJoinTask<?>[] a; |
1022 |
if ((a = array) != null && (al = a.length) > 0) { |
1023 |
a[(al - 1) & s] = task; |
1024 |
top = s + 1; |
1025 |
} |
1026 |
} finally { |
1027 |
qlock = 0; |
1028 |
} |
1029 |
} |
1030 |
|
1031 |
/** |
1032 |
* Shared version of pop |
1033 |
*/ |
1034 |
final boolean trySharedUnpush(ForkJoinTask<?> task) { |
1035 |
boolean popped = false; |
1036 |
int s = top - 1, al; ForkJoinTask<?>[] a; |
1037 |
if ((a = array) != null && (al = a.length) > 0) { |
1038 |
int index = (al - 1) & s; |
1039 |
long offset = ((long)index << ASHIFT) + ABASE; |
1040 |
ForkJoinTask<?> t = (ForkJoinTask<?>) |
1041 |
U.getObject(a, offset); |
1042 |
if (t == task && |
1043 |
U.compareAndSwapInt(this, QLOCK, 0, 1)) { |
1044 |
if (U.compareAndSwapObject(a, offset, task, null)) { |
1045 |
popped = true; |
1046 |
top = s; |
1047 |
} |
1048 |
U.putOrderedInt(this, QLOCK, 0); |
1049 |
} |
1050 |
} |
1051 |
return popped; |
1052 |
} |
1053 |
|
1054 |
/** |
1055 |
* Removes and cancels all known tasks, ignoring any exceptions. |
1056 |
*/ |
1057 |
final void cancelAll() { |
1058 |
ForkJoinTask<?> t; |
1059 |
if ((t = currentJoin) != null) { |
1060 |
currentJoin = null; |
1061 |
ForkJoinTask.cancelIgnoringExceptions(t); |
1062 |
} |
1063 |
if ((t = currentSteal) != null) { |
1064 |
currentSteal = null; |
1065 |
ForkJoinTask.cancelIgnoringExceptions(t); |
1066 |
} |
1067 |
while ((t = poll()) != null) |
1068 |
ForkJoinTask.cancelIgnoringExceptions(t); |
1069 |
} |
1070 |
|
1071 |
// Specialized execution methods |
1072 |
|
1073 |
/** |
1074 |
* Pops and executes up to POLL_LIMIT tasks or until empty. |
1075 |
*/ |
1076 |
final void localPopAndExec() { |
1077 |
for (int nexec = 0;;) { |
1078 |
int b = base, s = top, al; ForkJoinTask<?>[] a; |
1079 |
if ((a = array) != null && b != s && (al = a.length) > 0) { |
1080 |
int index = (al - 1) & --s; |
1081 |
long offset = ((long)index << ASHIFT) + ABASE; |
1082 |
ForkJoinTask<?> t = (ForkJoinTask<?>) |
1083 |
U.getAndSetObject(a, offset, null); |
1084 |
if (t != null) { |
1085 |
top = s; |
1086 |
(currentSteal = t).doExec(); |
1087 |
if (++nexec > POLL_LIMIT) |
1088 |
break; |
1089 |
} |
1090 |
else |
1091 |
break; |
1092 |
} |
1093 |
else |
1094 |
break; |
1095 |
} |
1096 |
} |
1097 |
|
1098 |
/** |
1099 |
* Polls and executes up to POLL_LIMIT tasks or until empty. |
1100 |
*/ |
1101 |
final void localPollAndExec() { |
1102 |
for (int nexec = 0;;) { |
1103 |
int b = base, s = top, al; ForkJoinTask<?>[] a; |
1104 |
if ((a = array) != null && b != s && (al = a.length) > 0) { |
1105 |
int index = (al - 1) & b++; |
1106 |
long offset = ((long)index << ASHIFT) + ABASE; |
1107 |
ForkJoinTask<?> t = (ForkJoinTask<?>) |
1108 |
U.getAndSetObject(a, offset, null); |
1109 |
if (t != null) { |
1110 |
base = b; |
1111 |
t.doExec(); |
1112 |
if (++nexec > POLL_LIMIT) |
1113 |
break; |
1114 |
} |
1115 |
} |
1116 |
else |
1117 |
break; |
1118 |
} |
1119 |
} |
1120 |
|
1121 |
/** |
1122 |
* Executes the given task and (some) remaining local tasks. |
1123 |
*/ |
1124 |
final void runTask(ForkJoinTask<?> task) { |
1125 |
if (task != null) { |
1126 |
task.doExec(); |
1127 |
if (config < 0) |
1128 |
localPollAndExec(); |
1129 |
else |
1130 |
localPopAndExec(); |
1131 |
int ns = ++nsteals; |
1132 |
ForkJoinWorkerThread thread = owner; |
1133 |
currentSteal = null; |
1134 |
if (ns < 0) // collect on overflow |
1135 |
transferStealCount(pool); |
1136 |
if (thread != null) |
1137 |
thread.afterTopLevelExec(); |
1138 |
} |
1139 |
} |
1140 |
|
1141 |
/** |
1142 |
* Adds steal count to pool steal count if it exists, and resets. |
1143 |
*/ |
1144 |
final void transferStealCount(ForkJoinPool p) { |
1145 |
AuxState aux; |
1146 |
if (p != null && (aux = p.auxState) != null) { |
1147 |
int s = nsteals; |
1148 |
nsteals = 0; // if negative, correct for overflow |
1149 |
long inc = (long)(s < 0 ? Integer.MAX_VALUE : s); |
1150 |
aux.lock(); |
1151 |
try { |
1152 |
aux.stealCount += inc; |
1153 |
} finally { |
1154 |
aux.unlock(); |
1155 |
} |
1156 |
} |
1157 |
} |
1158 |
|
1159 |
/** |
1160 |
* If present, removes from queue and executes the given task, |
1161 |
* or any other cancelled task. Used only by awaitJoin. |
1162 |
* |
1163 |
* @return true if queue empty and task not known to be done |
1164 |
*/ |
1165 |
final boolean tryRemoveAndExec(ForkJoinTask<?> task) { |
1166 |
if (task != null && task.status >= 0) { |
1167 |
int b, s, d, al; ForkJoinTask<?>[] a; |
1168 |
while ((d = (b = base) - (s = top)) < 0 && |
1169 |
(a = array) != null && (al = a.length) > 0) { |
1170 |
for (;;) { // traverse from s to b |
1171 |
int index = --s & (al - 1); |
1172 |
long offset = (index << ASHIFT) + ABASE; |
1173 |
ForkJoinTask<?> t = (ForkJoinTask<?>) |
1174 |
U.getObjectVolatile(a, offset); |
1175 |
if (t == null) |
1176 |
break; // restart |
1177 |
else if (t == task) { |
1178 |
boolean removed = false; |
1179 |
if (s + 1 == top) { // pop |
1180 |
if (U.compareAndSwapObject(a, offset, t, null)) { |
1181 |
top = s; |
1182 |
U.storeFence(); |
1183 |
removed = true; |
1184 |
} |
1185 |
} |
1186 |
else if (base == b) // replace with proxy |
1187 |
removed = U.compareAndSwapObject(a, offset, t, |
1188 |
new EmptyTask()); |
1189 |
if (removed) { |
1190 |
ForkJoinTask<?> ps = currentSteal; |
1191 |
(currentSteal = task).doExec(); |
1192 |
currentSteal = ps; |
1193 |
} |
1194 |
break; |
1195 |
} |
1196 |
else if (t.status < 0 && s + 1 == top) { |
1197 |
if (U.compareAndSwapObject(a, offset, t, null)) { |
1198 |
top = s; |
1199 |
U.storeFence(); |
1200 |
} |
1201 |
break; // was cancelled |
1202 |
} |
1203 |
else if (++d == 0) { |
1204 |
if (base != b) // rescan |
1205 |
break; |
1206 |
return false; |
1207 |
} |
1208 |
} |
1209 |
if (task.status < 0) |
1210 |
return false; |
1211 |
} |
1212 |
} |
1213 |
return true; |
1214 |
} |
1215 |
|
1216 |
/** |
1217 |
* Pops task if in the same CC computation as the given task, |
1218 |
* in either shared or owned mode. Used only by helpComplete. |
1219 |
*/ |
1220 |
final CountedCompleter<?> popCC(CountedCompleter<?> task, int mode) { |
1221 |
int b = base, s = top, al; ForkJoinTask<?>[] a; |
1222 |
if ((a = array) != null && b != s && (al = a.length) > 0) { |
1223 |
int index = (al - 1) & (s - 1); |
1224 |
long offset = ((long)index << ASHIFT) + ABASE; |
1225 |
ForkJoinTask<?> o = (ForkJoinTask<?>) |
1226 |
U.getObjectVolatile(a, offset); |
1227 |
if (o instanceof CountedCompleter) { |
1228 |
CountedCompleter<?> t = (CountedCompleter<?>)o; |
1229 |
for (CountedCompleter<?> r = t;;) { |
1230 |
if (r == task) { |
1231 |
if ((mode & IS_OWNED) == 0) { |
1232 |
boolean popped; |
1233 |
if (U.compareAndSwapInt(this, QLOCK, 0, 1)) { |
1234 |
if (popped = |
1235 |
U.compareAndSwapObject(a, offset, |
1236 |
t, null)) |
1237 |
top = s - 1; |
1238 |
U.putOrderedInt(this, QLOCK, 0); |
1239 |
if (popped) |
1240 |
return t; |
1241 |
} |
1242 |
} |
1243 |
else if (U.compareAndSwapObject(a, offset, |
1244 |
t, null)) { |
1245 |
top = s - 1; |
1246 |
U.storeFence(); |
1247 |
return t; |
1248 |
} |
1249 |
break; |
1250 |
} |
1251 |
else if ((r = r.completer) == null) // try parent |
1252 |
break; |
1253 |
} |
1254 |
} |
1255 |
} |
1256 |
return null; |
1257 |
} |
1258 |
|
1259 |
/** |
1260 |
* Steals and runs a task in the same CC computation as the |
1261 |
* given task if one exists and can be taken without |
1262 |
* contention. Otherwise returns a checksum/control value for |
1263 |
* use by method helpComplete. |
1264 |
* |
1265 |
* @return 1 if successful, 2 if retryable (lost to another |
1266 |
* stealer), -1 if non-empty but no matching task found, else |
1267 |
* the base index, forced negative. |
1268 |
*/ |
1269 |
final int pollAndExecCC(CountedCompleter<?> task) { |
1270 |
ForkJoinTask<?>[] a; |
1271 |
int b = base, s = top, al, h; |
1272 |
if ((a = array) != null && b != s && (al = a.length) > 0) { |
1273 |
int index = (al - 1) & b; |
1274 |
long offset = ((long)index << ASHIFT) + ABASE; |
1275 |
ForkJoinTask<?> o = (ForkJoinTask<?>) |
1276 |
U.getObjectVolatile(a, offset); |
1277 |
if (o == null) |
1278 |
h = 2; // retryable |
1279 |
else if (!(o instanceof CountedCompleter)) |
1280 |
h = -1; // unmatchable |
1281 |
else { |
1282 |
CountedCompleter<?> t = (CountedCompleter<?>)o; |
1283 |
for (CountedCompleter<?> r = t;;) { |
1284 |
if (r == task) { |
1285 |
if (b++ == base && |
1286 |
U.compareAndSwapObject(a, offset, t, null)) { |
1287 |
base = b; |
1288 |
t.doExec(); |
1289 |
h = 1; // success |
1290 |
} |
1291 |
else |
1292 |
h = 2; // lost CAS |
1293 |
break; |
1294 |
} |
1295 |
else if ((r = r.completer) == null) { |
1296 |
h = -1; // unmatched |
1297 |
break; |
1298 |
} |
1299 |
} |
1300 |
} |
1301 |
} |
1302 |
else |
1303 |
h = b | Integer.MIN_VALUE; // to sense movement on re-poll |
1304 |
return h; |
1305 |
} |
1306 |
|
1307 |
/** |
1308 |
* Returns true if owned and not known to be blocked. |
1309 |
*/ |
1310 |
final boolean isApparentlyUnblocked() { |
1311 |
Thread wt; Thread.State s; |
1312 |
return (scanState >= 0 && |
1313 |
(wt = owner) != null && |
1314 |
(s = wt.getState()) != Thread.State.BLOCKED && |
1315 |
s != Thread.State.WAITING && |
1316 |
s != Thread.State.TIMED_WAITING); |
1317 |
} |
1318 |
|
1319 |
// Unsafe mechanics. Note that some are (and must be) the same as in FJP |
1320 |
private static final sun.misc.Unsafe U = sun.misc.Unsafe.getUnsafe(); |
1321 |
private static final int ABASE; |
1322 |
private static final int ASHIFT; |
1323 |
private static final long QLOCK; |
1324 |
static { |
1325 |
try { |
1326 |
QLOCK = U.objectFieldOffset |
1327 |
(WorkQueue.class.getDeclaredField("qlock")); |
1328 |
ABASE = U.arrayBaseOffset(ForkJoinTask[].class); |
1329 |
int scale = U.arrayIndexScale(ForkJoinTask[].class); |
1330 |
if ((scale & (scale - 1)) != 0) |
1331 |
throw new Error("array index scale not a power of two"); |
1332 |
ASHIFT = 31 - Integer.numberOfLeadingZeros(scale); |
1333 |
} catch (ReflectiveOperationException e) { |
1334 |
throw new Error(e); |
1335 |
} |
1336 |
} |
1337 |
} |
1338 |
|
1339 |
// static fields (initialized in static initializer below) |
1340 |
|
1341 |
/** |
1342 |
* Creates a new ForkJoinWorkerThread. This factory is used unless |
1343 |
* overridden in ForkJoinPool constructors. |
1344 |
*/ |
1345 |
public static final ForkJoinWorkerThreadFactory |
1346 |
defaultForkJoinWorkerThreadFactory; |
1347 |
|
1348 |
/** |
1349 |
* Permission required for callers of methods that may start or |
1350 |
* kill threads. Also used as a private static lock in tryInitialize. |
1351 |
*/ |
1352 |
private static final RuntimePermission modifyThreadPermission; |
1353 |
|
1354 |
/** |
1355 |
* Common (static) pool. Non-null for public use unless a static |
1356 |
* construction exception, but internal usages null-check on use |
1357 |
* to paranoically avoid potential initialization circularities |
1358 |
* as well as to simplify generated code. |
1359 |
*/ |
1360 |
static final ForkJoinPool common; |
1361 |
|
1362 |
/** |
1363 |
* Common pool parallelism. To allow simpler use and management |
1364 |
* when common pool threads are disabled, we allow the underlying |
1365 |
* common.parallelism field to be zero, but in that case still report |
1366 |
* parallelism as 1 to reflect resulting caller-runs mechanics. |
1367 |
*/ |
1368 |
static final int commonParallelism; |
1369 |
|
1370 |
/** |
1371 |
* Limit on spare thread construction in tryCompensate. |
1372 |
*/ |
1373 |
private static int commonMaxSpares; |
1374 |
|
1375 |
/** |
1376 |
* Sequence number for creating workerNamePrefix. |
1377 |
*/ |
1378 |
private static int poolNumberSequence; |
1379 |
|
1380 |
/** |
1381 |
* Returns the next sequence number. We don't expect this to |
1382 |
* ever contend, so use simple builtin sync. |
1383 |
*/ |
1384 |
private static final synchronized int nextPoolId() { |
1385 |
return ++poolNumberSequence; |
1386 |
} |
1387 |
|
1388 |
// static configuration constants |
1389 |
|
1390 |
/** |
1391 |
* Initial timeout value (in milliseconds) for the thread |
1392 |
* triggering quiescence to park waiting for new work. On timeout, |
1393 |
* the thread will instead try to shrink the number of |
1394 |
* workers. The value should be large enough to avoid overly |
1395 |
* aggressive shrinkage during most transient stalls (long GCs |
1396 |
* etc). |
1397 |
*/ |
1398 |
private static final long IDLE_TIMEOUT_MS = 2000L; // 2sec |
1399 |
|
1400 |
/** |
1401 |
* Tolerance for idle timeouts, to cope with timer undershoots |
1402 |
*/ |
1403 |
private static final long TIMEOUT_SLOP_MS = 20L; // 20ms |
1404 |
|
1405 |
/** |
1406 |
* The initial value for commonMaxSpares during static |
1407 |
* initialization unless overridden using System property |
1408 |
* "java.util.concurrent.ForkJoinPool.common.maximumSpares". The |
1409 |
* default value is far in excess of normal requirements, but also |
1410 |
* far short of MAX_CAP and typical OS thread limits, so allows |
1411 |
* JVMs to catch misuse/abuse before running out of resources |
1412 |
* needed to do so. |
1413 |
*/ |
1414 |
private static final int DEFAULT_COMMON_MAX_SPARES = 256; |
1415 |
|
1416 |
/** |
1417 |
* Increment for seed generators. See class ThreadLocal for |
1418 |
* explanation. |
1419 |
*/ |
1420 |
private static final int SEED_INCREMENT = 0x9e3779b9; |
1421 |
|
1422 |
/* |
1423 |
* Bits and masks for field ctl, packed with 4 16 bit subfields: |
1424 |
* AC: Number of active running workers minus target parallelism |
1425 |
* TC: Number of total workers minus target parallelism |
1426 |
* SS: version count and status of top waiting thread |
1427 |
* ID: poolIndex of top of Treiber stack of waiters |
1428 |
* |
1429 |
* When convenient, we can extract the lower 32 stack top bits |
1430 |
* (including version bits) as sp=(int)ctl. The offsets of counts |
1431 |
* by the target parallelism and the positionings of fields makes |
1432 |
* it possible to perform the most common checks via sign tests of |
1433 |
* fields: When ac is negative, there are not enough active |
1434 |
* workers, when tc is negative, there are not enough total |
1435 |
* workers. When sp is non-zero, there are waiting workers. To |
1436 |
* deal with possibly negative fields, we use casts in and out of |
1437 |
* "short" and/or signed shifts to maintain signedness. |
1438 |
* |
1439 |
* Because it occupies uppermost bits, we can add one active count |
1440 |
* using getAndAddLong of AC_UNIT, rather than CAS, when returning |
1441 |
* from a blocked join. Other updates entail multiple subfields |
1442 |
* and masking, requiring CAS. |
1443 |
*/ |
1444 |
|
1445 |
// Lower and upper word masks |
1446 |
private static final long SP_MASK = 0xffffffffL; |
1447 |
private static final long UC_MASK = ~SP_MASK; |
1448 |
|
1449 |
// Active counts |
1450 |
private static final int AC_SHIFT = 48; |
1451 |
private static final long AC_UNIT = 0x0001L << AC_SHIFT; |
1452 |
private static final long AC_MASK = 0xffffL << AC_SHIFT; |
1453 |
|
1454 |
// Total counts |
1455 |
private static final int TC_SHIFT = 32; |
1456 |
private static final long TC_UNIT = 0x0001L << TC_SHIFT; |
1457 |
private static final long TC_MASK = 0xffffL << TC_SHIFT; |
1458 |
private static final long ADD_WORKER = 0x0001L << (TC_SHIFT + 15); // sign |
1459 |
|
1460 |
// runState bits: SHUTDOWN must be negative, others arbitrary powers of two |
1461 |
private static final int STARTED = 1; |
1462 |
private static final int STOP = 1 << 1; |
1463 |
private static final int TERMINATED = 1 << 2; |
1464 |
private static final int SHUTDOWN = 1 << 31; |
1465 |
|
1466 |
// Instance fields |
1467 |
volatile long ctl; // main pool control |
1468 |
volatile int runState; |
1469 |
final int config; // parallelism, mode |
1470 |
AuxState auxState; // lock, steal counts |
1471 |
volatile WorkQueue[] workQueues; // main registry |
1472 |
final String workerNamePrefix; // to create worker name string |
1473 |
final ForkJoinWorkerThreadFactory factory; |
1474 |
final UncaughtExceptionHandler ueh; // per-worker UEH |
1475 |
|
1476 |
/** |
1477 |
* Instantiates fields upon first submission, or upon shutdown if |
1478 |
* no submissions. If checkTermination true, also responds to |
1479 |
* termination by external calls submitting tasks. |
1480 |
*/ |
1481 |
private void tryInitialize(boolean checkTermination) { |
1482 |
if ((runState & STARTED) == 0) { // bootstrap by locking static field |
1483 |
int rs; // create workQueues array with size a power of two |
1484 |
int p = config & SMASK; // ensure at least 2 slots |
1485 |
int n = (p > 1) ? p - 1 : 1; |
1486 |
n |= n >>> 1; n |= n >>> 2; n |= n >>> 4; |
1487 |
n |= n >>> 8; n |= n >>> 16; n = ((n + 1) << 1) & SMASK; |
1488 |
AuxState aux = new AuxState(); |
1489 |
WorkQueue[] ws = new WorkQueue[n]; |
1490 |
synchronized (modifyThreadPermission) { // double-check |
1491 |
if (((rs = runState) & STARTED) == 0) { |
1492 |
rs |= STARTED; |
1493 |
workQueues = ws; |
1494 |
auxState = aux; |
1495 |
runState = rs; |
1496 |
} |
1497 |
} |
1498 |
} |
1499 |
if (checkTermination && runState < 0) { |
1500 |
tryTerminate(false, false); // help terminate |
1501 |
throw new RejectedExecutionException(); |
1502 |
} |
1503 |
} |
1504 |
|
1505 |
// Creating, registering and deregistering workers |
1506 |
|
1507 |
/** |
1508 |
* Tries to construct and start one worker. Assumes that total |
1509 |
* count has already been incremented as a reservation. Invokes |
1510 |
* deregisterWorker on any failure. |
1511 |
* |
1512 |
* @param isSpare true if this is a spare thread |
1513 |
* @return true if successful |
1514 |
*/ |
1515 |
private boolean createWorker(boolean isSpare) { |
1516 |
ForkJoinWorkerThreadFactory fac = factory; |
1517 |
Throwable ex = null; |
1518 |
ForkJoinWorkerThread wt = null; |
1519 |
WorkQueue q; |
1520 |
try { |
1521 |
if (fac != null && (wt = fac.newThread(this)) != null) { |
1522 |
if (isSpare && (q = wt.workQueue) != null) |
1523 |
q.config |= SPARE_WORKER; |
1524 |
wt.start(); |
1525 |
return true; |
1526 |
} |
1527 |
} catch (Throwable rex) { |
1528 |
ex = rex; |
1529 |
} |
1530 |
deregisterWorker(wt, ex); |
1531 |
return false; |
1532 |
} |
1533 |
|
1534 |
/** |
1535 |
* Tries to add one worker, incrementing ctl counts before doing |
1536 |
* so, relying on createWorker to back out on failure. |
1537 |
* |
1538 |
* @param c incoming ctl value, with total count negative and no |
1539 |
* idle workers. On CAS failure, c is refreshed and retried if |
1540 |
* this holds (otherwise, a new worker is not needed). |
1541 |
*/ |
1542 |
private void tryAddWorker(long c) { |
1543 |
do { |
1544 |
long nc = ((AC_MASK & (c + AC_UNIT)) | |
1545 |
(TC_MASK & (c + TC_UNIT))); |
1546 |
if (ctl == c && U.compareAndSwapLong(this, CTL, c, nc)) { |
1547 |
createWorker(false); |
1548 |
break; |
1549 |
} |
1550 |
} while (((c = ctl) & ADD_WORKER) != 0L && (int)c == 0); |
1551 |
} |
1552 |
|
1553 |
/** |
1554 |
* Callback from ForkJoinWorkerThread constructor to establish and |
1555 |
* record its WorkQueue. |
1556 |
* |
1557 |
* @param wt the worker thread |
1558 |
* @return the worker's queue |
1559 |
*/ |
1560 |
final WorkQueue registerWorker(ForkJoinWorkerThread wt) { |
1561 |
UncaughtExceptionHandler handler; |
1562 |
AuxState aux; |
1563 |
wt.setDaemon(true); // configure thread |
1564 |
if ((handler = ueh) != null) |
1565 |
wt.setUncaughtExceptionHandler(handler); |
1566 |
WorkQueue w = new WorkQueue(this, wt); |
1567 |
int i = 0; // assign a pool index |
1568 |
int mode = config & MODE_MASK; |
1569 |
int rs = runState; |
1570 |
if ((aux = auxState) != null) { |
1571 |
aux.lock(); |
1572 |
try { |
1573 |
int s = (int)(aux.indexSeed += SEED_INCREMENT), n, m; |
1574 |
WorkQueue[] ws = workQueues; |
1575 |
if (ws != null && (n = ws.length) > 0) { |
1576 |
i = (m = n - 1) & ((s << 1) | 1); // odd-numbered indices |
1577 |
if (ws[i] != null) { // collision |
1578 |
int probes = 0; // step by approx half n |
1579 |
int step = (n <= 4) ? 2 : ((n >>> 1) & EVENMASK) + 2; |
1580 |
while (ws[i = (i + step) & m] != null) { |
1581 |
if (++probes >= n) { |
1582 |
workQueues = ws = Arrays.copyOf(ws, n <<= 1); |
1583 |
m = n - 1; |
1584 |
probes = 0; |
1585 |
} |
1586 |
} |
1587 |
} |
1588 |
w.hint = s; // use as random seed |
1589 |
w.config = i | mode; |
1590 |
w.scanState = i | (s & 0x7fff0000); // random seq bits |
1591 |
ws[i] = w; |
1592 |
} |
1593 |
} finally { |
1594 |
aux.unlock(); |
1595 |
} |
1596 |
} |
1597 |
wt.setName(workerNamePrefix.concat(Integer.toString(i >>> 1))); |
1598 |
return w; |
1599 |
} |
1600 |
|
1601 |
/** |
1602 |
* Final callback from terminating worker, as well as upon failure |
1603 |
* to construct or start a worker. Removes record of worker from |
1604 |
* array, and adjusts counts. If pool is shutting down, tries to |
1605 |
* complete termination. |
1606 |
* |
1607 |
* @param wt the worker thread, or null if construction failed |
1608 |
* @param ex the exception causing failure, or null if none |
1609 |
*/ |
1610 |
final void deregisterWorker(ForkJoinWorkerThread wt, Throwable ex) { |
1611 |
WorkQueue w = null; |
1612 |
int rs = runState; |
1613 |
if (wt != null && (w = wt.workQueue) != null) { |
1614 |
AuxState aux; WorkQueue[] ws; // remove index from array |
1615 |
int idx = w.config & SMASK; |
1616 |
int ns = w.nsteals; |
1617 |
if ((aux = auxState) != null) { |
1618 |
aux.lock(); |
1619 |
try { |
1620 |
if ((ws = workQueues) != null && ws.length > idx && |
1621 |
ws[idx] == w) |
1622 |
ws[idx] = null; |
1623 |
aux.stealCount += ns; |
1624 |
} finally { |
1625 |
aux.unlock(); |
1626 |
} |
1627 |
} |
1628 |
} |
1629 |
if (w == null || (w.config & UNREGISTERED) == 0) { // else pre-adjusted |
1630 |
long c; // decrement counts |
1631 |
do {} while (!U.compareAndSwapLong |
1632 |
(this, CTL, c = ctl, ((AC_MASK & (c - AC_UNIT)) | |
1633 |
(TC_MASK & (c - TC_UNIT)) | |
1634 |
(SP_MASK & c)))); |
1635 |
} |
1636 |
if (w != null) { |
1637 |
w.qlock = -1; // ensure set |
1638 |
w.cancelAll(); // cancel remaining tasks |
1639 |
} |
1640 |
for (;;) { // possibly replace |
1641 |
WorkQueue[] ws; int wl, sp; long c; |
1642 |
if (tryTerminate(false, false) || w == null || w.array == null || |
1643 |
(runState & STOP) != 0 || (ws = workQueues) == null || |
1644 |
(wl = ws.length) <= 0) // already terminating |
1645 |
break; |
1646 |
else if ((sp = (int)(c = ctl)) != 0) { // wake up replacement |
1647 |
if (tryRelease(c, ws[(wl - 1) & sp], AC_UNIT)) |
1648 |
break; |
1649 |
} |
1650 |
else if (ex != null && (c & ADD_WORKER) != 0L) { |
1651 |
tryAddWorker(c); // create replacement |
1652 |
break; |
1653 |
} |
1654 |
else // don't need replacement |
1655 |
break; |
1656 |
} |
1657 |
if (ex == null) // help clean on way out |
1658 |
ForkJoinTask.helpExpungeStaleExceptions(); |
1659 |
else // rethrow |
1660 |
ForkJoinTask.rethrow(ex); |
1661 |
} |
1662 |
|
1663 |
// Signalling |
1664 |
|
1665 |
/** |
1666 |
* Tries to create or activate a worker if too few are active. |
1667 |
*/ |
1668 |
final void signalWork() { |
1669 |
for (;;) { |
1670 |
long c; int sp, i; WorkQueue v; WorkQueue[] ws; |
1671 |
if ((c = ctl) >= 0L) // enough workers |
1672 |
break; |
1673 |
else if ((sp = (int)c) == 0) { // no idle workers |
1674 |
if ((c & ADD_WORKER) != 0L) // too few workers |
1675 |
tryAddWorker(c); |
1676 |
break; |
1677 |
} |
1678 |
else if ((ws = workQueues) == null) |
1679 |
break; // unstarted/terminated |
1680 |
else if (ws.length <= (i = sp & SMASK)) |
1681 |
break; // terminated |
1682 |
else if ((v = ws[i]) == null) |
1683 |
break; // terminating |
1684 |
else { |
1685 |
int ns = sp & ~UNSIGNALLED; |
1686 |
int vs = v.scanState; |
1687 |
long nc = (v.stackPred & SP_MASK) | (UC_MASK & (c + AC_UNIT)); |
1688 |
if (sp == vs && U.compareAndSwapLong(this, CTL, c, nc)) { |
1689 |
v.scanState = ns; |
1690 |
LockSupport.unpark(v.parker); |
1691 |
break; |
1692 |
} |
1693 |
} |
1694 |
} |
1695 |
} |
1696 |
|
1697 |
/** |
1698 |
* Signals and releases worker v if it is top of idle worker |
1699 |
* stack. This performs a one-shot version of signalWork only if |
1700 |
* there is (apparently) at least one idle worker. |
1701 |
* |
1702 |
* @param c incoming ctl value |
1703 |
* @param v if non-null, a worker |
1704 |
* @param inc the increment to active count (zero when compensating) |
1705 |
* @return true if successful |
1706 |
*/ |
1707 |
private boolean tryRelease(long c, WorkQueue v, long inc) { |
1708 |
int sp = (int)c, ns = sp & ~UNSIGNALLED; |
1709 |
if (v != null) { |
1710 |
int vs = v.scanState; |
1711 |
long nc = (v.stackPred & SP_MASK) | (UC_MASK & (c + inc)); |
1712 |
if (sp == vs && U.compareAndSwapLong(this, CTL, c, nc)) { |
1713 |
v.scanState = ns; |
1714 |
LockSupport.unpark(v.parker); |
1715 |
return true; |
1716 |
} |
1717 |
} |
1718 |
return false; |
1719 |
} |
1720 |
|
1721 |
/** |
1722 |
* With approx probability of a missed signal, tries (once) to |
1723 |
* reactivate worker w (or some other worker), failing if stale or |
1724 |
* known to be already active. |
1725 |
* |
1726 |
* @param w the worker |
1727 |
* @param ws the workQueue array to use |
1728 |
* @param r random seed |
1729 |
*/ |
1730 |
private void tryReactivate(WorkQueue w, WorkQueue[] ws, int r) { |
1731 |
long c; int sp, wl, m; WorkQueue v; |
1732 |
if ((sp = (int)(c = ctl)) != 0 && w != null && |
1733 |
ws != null && (wl = ws.length) > 0 && |
1734 |
((m = wl - 1) & ((sp ^ r) >>> 16)) <= 1 && |
1735 |
(v = ws[m & sp]) != null) { |
1736 |
long nc = (v.stackPred & SP_MASK) | (UC_MASK & (c + AC_UNIT)); |
1737 |
int ns = sp & ~UNSIGNALLED; |
1738 |
if (w.scanState < 0 && |
1739 |
v.scanState == sp && |
1740 |
U.compareAndSwapLong(this, CTL, c, nc)) { |
1741 |
v.scanState = ns; |
1742 |
LockSupport.unpark(v.parker); |
1743 |
} |
1744 |
} |
1745 |
} |
1746 |
|
1747 |
/** |
1748 |
* If worker w exists and is active, enqueues and sets status to inactive. |
1749 |
* |
1750 |
* @param w the worker |
1751 |
* @param ss current (non-negative) scanState |
1752 |
*/ |
1753 |
private void inactivate(WorkQueue w, int ss) { |
1754 |
int ns = (ss + SS_SEQ) | UNSIGNALLED; |
1755 |
long lc = ns & SP_MASK, nc, c; |
1756 |
if (w != null) { |
1757 |
w.scanState = ns; |
1758 |
do { |
1759 |
nc = lc | (UC_MASK & ((c = ctl) - AC_UNIT)); |
1760 |
w.stackPred = (int)c; |
1761 |
} while (!U.compareAndSwapLong(this, CTL, c, nc)); |
1762 |
} |
1763 |
} |
1764 |
|
1765 |
/** |
1766 |
* Possibly blocks worker w waiting for signal, or returns |
1767 |
* negative status if the worker should terminate. May return |
1768 |
* without status change if multiple stale unparks and/or |
1769 |
* interrupts occur. |
1770 |
* |
1771 |
* @param w the calling worker |
1772 |
* @return negative if w should terminate |
1773 |
*/ |
1774 |
private int awaitWork(WorkQueue w) { |
1775 |
int stat = 0; |
1776 |
if (w != null && w.scanState < 0) { |
1777 |
long c = ctl; |
1778 |
if ((int)(c >> AC_SHIFT) + (config & SMASK) <= 0) |
1779 |
stat = timedAwaitWork(w, c); // possibly quiescent |
1780 |
else if ((runState & STOP) != 0) |
1781 |
stat = w.qlock = -1; // pool terminating |
1782 |
else if ((stat = w.qlock) >= 0 && w.scanState < 0) { |
1783 |
w.parker = Thread.currentThread(); |
1784 |
if (w.scanState < 0) { // recheck after write |
1785 |
LockSupport.park(this); |
1786 |
if ((stat = w.qlock) >= 0 && w.scanState < 0) { |
1787 |
Thread.interrupted(); // clear status |
1788 |
LockSupport.park(this); // retry once |
1789 |
} |
1790 |
} |
1791 |
w.parker = null; |
1792 |
} |
1793 |
} |
1794 |
return stat; |
1795 |
} |
1796 |
|
1797 |
/** |
1798 |
* Possibly triggers shutdown and tries (once) to block worker |
1799 |
* when pool is (or may be) quiescent. Waits up to a duration |
1800 |
* determined by number of workers. On timeout, if ctl has not |
1801 |
* changed, terminates the worker, which will in turn wake up |
1802 |
* another worker to possibly repeat this process. |
1803 |
* |
1804 |
* @param w the calling worker |
1805 |
* @return negative if w should terminate |
1806 |
*/ |
1807 |
private int timedAwaitWork(WorkQueue w, long c) { |
1808 |
int stat = 0; |
1809 |
int scale = 1 - (short)(c >>> TC_SHIFT); |
1810 |
long deadline = (((scale <= 0) ? 1 : scale) * IDLE_TIMEOUT_MS + |
1811 |
System.currentTimeMillis()); |
1812 |
if (w != null && w.scanState < 0) { |
1813 |
int ss; AuxState aux; |
1814 |
if (runState < 0 && tryTerminate(false, false)) |
1815 |
stat = w.qlock = -1; // help terminate |
1816 |
else if ((stat = w.qlock) >= 0 && w.scanState < 0) { |
1817 |
w.parker = Thread.currentThread(); |
1818 |
if (w.scanState < 0) |
1819 |
LockSupport.parkUntil(this, deadline); |
1820 |
w.parker = null; |
1821 |
if ((stat = w.qlock) >= 0 && (ss = w.scanState) < 0 && |
1822 |
!Thread.interrupted() && (int)c == ss && |
1823 |
(aux = auxState) != null && ctl == c && |
1824 |
deadline - System.currentTimeMillis() <= TIMEOUT_SLOP_MS) { |
1825 |
aux.lock(); |
1826 |
try { // pre-deregister |
1827 |
WorkQueue[] ws; |
1828 |
int cfg = w.config, idx = cfg & SMASK; |
1829 |
long nc = ((UC_MASK & (c - TC_UNIT)) | |
1830 |
(SP_MASK & w.stackPred)); |
1831 |
if ((ws = workQueues) != null && |
1832 |
idx < ws.length && idx >= 0 && ws[idx] == w && |
1833 |
U.compareAndSwapLong(this, CTL, c, nc)) { |
1834 |
ws[idx] = null; |
1835 |
w.config = cfg | UNREGISTERED; |
1836 |
stat = w.qlock = -1; |
1837 |
} |
1838 |
} finally { |
1839 |
aux.unlock(); |
1840 |
} |
1841 |
} |
1842 |
} |
1843 |
} |
1844 |
return stat; |
1845 |
} |
1846 |
|
1847 |
/** |
1848 |
* If the given worker is a spare with no queued tasks, and there |
1849 |
* are enough existing workers, drops it from ctl counts and sets |
1850 |
* its state to terminated. |
1851 |
* |
1852 |
* @param w the calling worker -- must be a spare |
1853 |
* @return true if dropped (in which case it must not process more tasks) |
1854 |
*/ |
1855 |
private boolean tryDropSpare(WorkQueue w) { |
1856 |
if (w != null && w.isEmpty()) { // no local tasks |
1857 |
long c; int sp, wl; WorkQueue[] ws; WorkQueue v; |
1858 |
while ((short)((c = ctl) >> TC_SHIFT) > 0 && |
1859 |
((sp = (int)c) != 0 || (int)(c >> AC_SHIFT) > 0) && |
1860 |
(ws = workQueues) != null && (wl = ws.length) > 0) { |
1861 |
boolean dropped, canDrop; |
1862 |
if (sp == 0) { // no queued workers |
1863 |
long nc = ((AC_MASK & (c - AC_UNIT)) | |
1864 |
(TC_MASK & (c - TC_UNIT)) | (SP_MASK & c)); |
1865 |
dropped = U.compareAndSwapLong(this, CTL, c, nc); |
1866 |
} |
1867 |
else if ( |
1868 |
(v = ws[(wl - 1) & sp]) == null || v.scanState != sp) |
1869 |
dropped = false; // stale; retry |
1870 |
else { |
1871 |
long nc = v.stackPred & SP_MASK; |
1872 |
if (w == v || w.scanState >= 0) { |
1873 |
canDrop = true; // w unqueued or topmost |
1874 |
nc |= ((AC_MASK & c) | // ensure replacement |
1875 |
(TC_MASK & (c - TC_UNIT))); |
1876 |
} |
1877 |
else { // w may be queued |
1878 |
canDrop = false; // help uncover |
1879 |
nc |= ((AC_MASK & (c + AC_UNIT)) | |
1880 |
(TC_MASK & c)); |
1881 |
} |
1882 |
if (U.compareAndSwapLong(this, CTL, c, nc)) { |
1883 |
v.scanState = sp & ~UNSIGNALLED; |
1884 |
LockSupport.unpark(v.parker); |
1885 |
dropped = canDrop; |
1886 |
} |
1887 |
else |
1888 |
dropped = false; |
1889 |
} |
1890 |
if (dropped) { // pre-deregister |
1891 |
int cfg = w.config, idx = cfg & SMASK; |
1892 |
if (idx >= 0 && idx < ws.length && ws[idx] == w) |
1893 |
ws[idx] = null; |
1894 |
w.config = cfg | UNREGISTERED; |
1895 |
w.qlock = -1; |
1896 |
return true; |
1897 |
} |
1898 |
} |
1899 |
} |
1900 |
return false; |
1901 |
} |
1902 |
|
1903 |
/** |
1904 |
* Top-level runloop for workers, called by ForkJoinWorkerThread.run. |
1905 |
*/ |
1906 |
final void runWorker(WorkQueue w) { |
1907 |
w.growArray(); // allocate queue |
1908 |
int bound = (w.config & SPARE_WORKER) != 0 ? 0 : POLL_LIMIT; |
1909 |
long seed = w.hint * 0xdaba0b6eb09322e3L; // initial random seed |
1910 |
for (long r = (seed == 0L) ? 1L : seed;;) { // ensure nonzero |
1911 |
if (bound == 0 && tryDropSpare(w)) |
1912 |
break; |
1913 |
// high bits of prev seed for step; current low bits for idx |
1914 |
int step = (int)(r >>> 48) | 1; |
1915 |
r ^= r >>> 12; r ^= r << 25; r ^= r >>> 27; // xorshift |
1916 |
if (scan(w, bound, step, (int)r) < 0 && awaitWork(w) < 0) |
1917 |
break; |
1918 |
} |
1919 |
} |
1920 |
|
1921 |
// Scanning for tasks |
1922 |
|
1923 |
/** |
1924 |
* Repeatedly scans for and tries to steal and execute (via |
1925 |
* workQueue.runTask) a queued task. Each scan traverses queues in |
1926 |
* pseudorandom permutation. Upon finding a non-empty queue, makes |
1927 |
* at most the given bound attempts to re-poll (fewer if |
1928 |
* contended) on the same queue before returning (impossible |
1929 |
* scanState value) 0 to restart scan. Else returns after at least |
1930 |
* 1 and at most 32 full scans. |
1931 |
* |
1932 |
* @param w the worker (via its WorkQueue) |
1933 |
* @param bound repoll bound as bitmask (0 if spare) |
1934 |
* @param step (circular) index increment per iteration (must be odd) |
1935 |
* @param r a random seed for origin index |
1936 |
* @return negative if should await signal |
1937 |
*/ |
1938 |
private int scan(WorkQueue w, int bound, int step, int r) { |
1939 |
int stat = 0, wl; WorkQueue[] ws; |
1940 |
if ((ws = workQueues) != null && w != null && (wl = ws.length) > 0) { |
1941 |
for (int m = wl - 1, |
1942 |
origin = m & r, idx = origin, |
1943 |
npolls = 0, |
1944 |
ss = w.scanState;;) { // negative if inactive |
1945 |
WorkQueue q; ForkJoinTask<?>[] a; int b, d, al; |
1946 |
if ((q = ws[idx]) != null && (d = (b = q.base) - q.top) < 0 && |
1947 |
(a = q.array) != null && (al = a.length) > 0) { |
1948 |
int index = (al - 1) & b; |
1949 |
long offset = ((long)index << ASHIFT) + ABASE; |
1950 |
ForkJoinTask<?> t = (ForkJoinTask<?>) |
1951 |
U.getObjectVolatile(a, offset); |
1952 |
if (t == null) |
1953 |
break; // empty or busy |
1954 |
else if (b++ != q.base) |
1955 |
break; // busy |
1956 |
else if (ss < 0) { |
1957 |
tryReactivate(w, ws, r); |
1958 |
break; // retry upon rescan |
1959 |
} |
1960 |
else if (!U.compareAndSwapObject(a, offset, t, null)) |
1961 |
break; // contended |
1962 |
else { |
1963 |
q.base = b; |
1964 |
w.currentSteal = t; |
1965 |
if (d != -1) // propagate signal |
1966 |
signalWork(); |
1967 |
w.runTask(t); |
1968 |
if (++npolls > bound) |
1969 |
break; |
1970 |
} |
1971 |
} |
1972 |
else if (npolls != 0) // rescan |
1973 |
break; |
1974 |
else if ((idx = (idx + step) & m) == origin) { |
1975 |
if (ss < 0) { // await signal |
1976 |
stat = ss; |
1977 |
break; |
1978 |
} |
1979 |
else if (r >= 0) { |
1980 |
inactivate(w, ss); |
1981 |
break; |
1982 |
} |
1983 |
else |
1984 |
r <<= 1; // at most 31 rescans |
1985 |
} |
1986 |
} |
1987 |
} |
1988 |
return stat; |
1989 |
} |
1990 |
|
1991 |
// Joining tasks |
1992 |
|
1993 |
/** |
1994 |
* Tries to steal and run tasks within the target's computation. |
1995 |
* Uses a variant of the top-level algorithm, restricted to tasks |
1996 |
* with the given task as ancestor: It prefers taking and running |
1997 |
* eligible tasks popped from the worker's own queue (via |
1998 |
* popCC). Otherwise it scans others, randomly moving on |
1999 |
* contention or execution, deciding to give up based on a |
2000 |
* checksum (via return codes from pollAndExecCC). The maxTasks |
2001 |
* argument supports external usages; internal calls use zero, |
2002 |
* allowing unbounded steps (external calls trap non-positive |
2003 |
* values). |
2004 |
* |
2005 |
* @param w caller |
2006 |
* @param maxTasks if non-zero, the maximum number of other tasks to run |
2007 |
* @return task status on exit |
2008 |
*/ |
2009 |
final int helpComplete(WorkQueue w, CountedCompleter<?> task, |
2010 |
int maxTasks) { |
2011 |
WorkQueue[] ws; int s = 0, wl; |
2012 |
if ((ws = workQueues) != null && (wl = ws.length) > 1 && |
2013 |
task != null && w != null) { |
2014 |
for (int m = wl - 1, |
2015 |
mode = w.config, |
2016 |
r = ~mode, // scanning seed |
2017 |
origin = r & m, k = origin, // first queue to scan |
2018 |
step = 3, // first scan step |
2019 |
h = 1, // 1:ran, >1:contended, <0:hash |
2020 |
oldSum = 0, checkSum = 0;;) { |
2021 |
CountedCompleter<?> p; WorkQueue q; int i; |
2022 |
if ((s = task.status) < 0) |
2023 |
break; |
2024 |
if (h == 1 && (p = w.popCC(task, mode)) != null) { |
2025 |
p.doExec(); // run local task |
2026 |
if (maxTasks != 0 && --maxTasks == 0) |
2027 |
break; |
2028 |
origin = k; // reset |
2029 |
oldSum = checkSum = 0; |
2030 |
} |
2031 |
else { // poll other worker queues |
2032 |
if ((i = k | 1) < 0 || i > m || (q = ws[i]) == null) |
2033 |
h = 0; |
2034 |
else if ((h = q.pollAndExecCC(task)) < 0) |
2035 |
checkSum += h; |
2036 |
if (h > 0) { |
2037 |
if (h == 1 && maxTasks != 0 && --maxTasks == 0) |
2038 |
break; |
2039 |
step = (r >>> 16) | 3; |
2040 |
r ^= r << 13; r ^= r >>> 17; r ^= r << 5; // xorshift |
2041 |
k = origin = r & m; // move and restart |
2042 |
oldSum = checkSum = 0; |
2043 |
} |
2044 |
else if ((k = (k + step) & m) == origin) { |
2045 |
if (oldSum == (oldSum = checkSum)) |
2046 |
break; |
2047 |
checkSum = 0; |
2048 |
} |
2049 |
} |
2050 |
} |
2051 |
} |
2052 |
return s; |
2053 |
} |
2054 |
|
2055 |
/** |
2056 |
* Tries to locate and execute tasks for a stealer of the given |
2057 |
* task, or in turn one of its stealers, Traces currentSteal -> |
2058 |
* currentJoin links looking for a thread working on a descendant |
2059 |
* of the given task and with a non-empty queue to steal back and |
2060 |
* execute tasks from. The first call to this method upon a |
2061 |
* waiting join will often entail scanning/search, (which is OK |
2062 |
* because the joiner has nothing better to do), but this method |
2063 |
* leaves hints in workers to speed up subsequent calls. |
2064 |
* |
2065 |
* @param w caller |
2066 |
* @param task the task to join |
2067 |
*/ |
2068 |
private void helpStealer(WorkQueue w, ForkJoinTask<?> task) { |
2069 |
if (task != null && w != null) { |
2070 |
ForkJoinTask<?> ps = w.currentSteal; |
2071 |
WorkQueue[] ws; int wl, oldSum = 0; |
2072 |
outer: while (w.tryRemoveAndExec(task) && task.status >= 0 && |
2073 |
(ws = workQueues) != null && (wl = ws.length) > 0) { |
2074 |
ForkJoinTask<?> subtask; |
2075 |
int m = wl - 1, checkSum = 0; // for stability check |
2076 |
WorkQueue j = w, v; // v is subtask stealer |
2077 |
descent: for (subtask = task; subtask.status >= 0; ) { |
2078 |
for (int h = j.hint | 1, k = 0, i;;) { |
2079 |
if ((v = ws[i = (h + (k << 1)) & m]) != null) { |
2080 |
if (v.currentSteal == subtask) { |
2081 |
j.hint = i; |
2082 |
break; |
2083 |
} |
2084 |
checkSum += v.base; |
2085 |
} |
2086 |
if (++k > m) // can't find stealer |
2087 |
break outer; |
2088 |
} |
2089 |
|
2090 |
for (;;) { // help v or descend |
2091 |
ForkJoinTask<?>[] a; int b, al; |
2092 |
if (subtask.status < 0) // too late to help |
2093 |
break descent; |
2094 |
checkSum += (b = v.base); |
2095 |
ForkJoinTask<?> next = v.currentJoin; |
2096 |
ForkJoinTask<?> t = null; |
2097 |
if ((a = v.array) != null && (al = a.length) > 0) { |
2098 |
int index = (al - 1) & b; |
2099 |
long offset = ((long)index << ASHIFT) + ABASE; |
2100 |
t = (ForkJoinTask<?>) |
2101 |
U.getObjectVolatile(a, offset); |
2102 |
if (t != null && b++ == v.base) { |
2103 |
if (j.currentJoin != subtask || |
2104 |
v.currentSteal != subtask || |
2105 |
subtask.status < 0) |
2106 |
break descent; // stale |
2107 |
if (U.compareAndSwapObject(a, offset, t, null)) { |
2108 |
v.base = b; |
2109 |
w.currentSteal = t; |
2110 |
for (int top = w.top;;) { |
2111 |
t.doExec(); // help |
2112 |
w.currentSteal = ps; |
2113 |
if (task.status < 0) |
2114 |
break outer; |
2115 |
if (w.top == top) |
2116 |
break; // run local tasks |
2117 |
if ((t = w.pop()) == null) |
2118 |
break descent; |
2119 |
w.currentSteal = t; |
2120 |
} |
2121 |
} |
2122 |
} |
2123 |
} |
2124 |
if (t == null && b == v.base && b - v.top >= 0) { |
2125 |
if ((subtask = next) == null) { // try to descend |
2126 |
if (next == v.currentJoin && |
2127 |
oldSum == (oldSum = checkSum)) |
2128 |
break outer; |
2129 |
break descent; |
2130 |
} |
2131 |
j = v; |
2132 |
break; |
2133 |
} |
2134 |
} |
2135 |
} |
2136 |
} |
2137 |
} |
2138 |
} |
2139 |
|
2140 |
/** |
2141 |
* Tries to decrement active count (sometimes implicitly) and |
2142 |
* possibly release or create a compensating worker in preparation |
2143 |
* for blocking. Returns false (retryable by caller), on |
2144 |
* contention, detected staleness, instability, or termination. |
2145 |
* |
2146 |
* @param w caller |
2147 |
*/ |
2148 |
private boolean tryCompensate(WorkQueue w) { |
2149 |
boolean canBlock; int wl; |
2150 |
long c = ctl; |
2151 |
WorkQueue[] ws = workQueues; |
2152 |
int pc = config & SMASK; |
2153 |
int ac = pc + (int)(c >> AC_SHIFT); |
2154 |
int tc = pc + (short)(c >> TC_SHIFT); |
2155 |
if (w == null || w.qlock < 0 || pc == 0 || // terminating or disabled |
2156 |
ws == null || (wl = ws.length) <= 0) |
2157 |
canBlock = false; |
2158 |
else { |
2159 |
int m = wl - 1, sp; |
2160 |
boolean busy = true; // validate ac |
2161 |
for (int i = 0; i <= m; ++i) { |
2162 |
int k; WorkQueue v; |
2163 |
if ((k = (i << 1) | 1) <= m && k >= 0 && (v = ws[k]) != null && |
2164 |
v.scanState >= 0 && v.currentSteal == null) { |
2165 |
busy = false; |
2166 |
break; |
2167 |
} |
2168 |
} |
2169 |
if (!busy || ctl != c) |
2170 |
canBlock = false; // unstable or stale |
2171 |
else if ((sp = (int)c) != 0) // release idle worker |
2172 |
canBlock = tryRelease(c, ws[m & sp], 0L); |
2173 |
else if (tc >= pc && ac > 1 && w.isEmpty()) { |
2174 |
long nc = ((AC_MASK & (c - AC_UNIT)) | |
2175 |
(~AC_MASK & c)); // uncompensated |
2176 |
canBlock = U.compareAndSwapLong(this, CTL, c, nc); |
2177 |
} |
2178 |
else if (tc >= MAX_CAP || |
2179 |
(this == common && tc >= pc + commonMaxSpares)) |
2180 |
throw new RejectedExecutionException( |
2181 |
"Thread limit exceeded replacing blocked worker"); |
2182 |
else { // similar to tryAddWorker |
2183 |
boolean isSpare = (tc >= pc); |
2184 |
long nc = (AC_MASK & c) | (TC_MASK & (c + TC_UNIT)); |
2185 |
canBlock = (U.compareAndSwapLong(this, CTL, c, nc) && |
2186 |
createWorker(isSpare)); // throws on exception |
2187 |
} |
2188 |
} |
2189 |
return canBlock; |
2190 |
} |
2191 |
|
2192 |
/** |
2193 |
* Helps and/or blocks until the given task is done or timeout. |
2194 |
* |
2195 |
* @param w caller |
2196 |
* @param task the task |
2197 |
* @param deadline for timed waits, if nonzero |
2198 |
* @return task status on exit |
2199 |
*/ |
2200 |
final int awaitJoin(WorkQueue w, ForkJoinTask<?> task, long deadline) { |
2201 |
int s = 0; |
2202 |
if (w != null) { |
2203 |
ForkJoinTask<?> prevJoin = w.currentJoin; |
2204 |
if (task != null && (s = task.status) >= 0) { |
2205 |
w.currentJoin = task; |
2206 |
CountedCompleter<?> cc = (task instanceof CountedCompleter) ? |
2207 |
(CountedCompleter<?>)task : null; |
2208 |
for (;;) { |
2209 |
if (cc != null) |
2210 |
helpComplete(w, cc, 0); |
2211 |
else |
2212 |
helpStealer(w, task); |
2213 |
if ((s = task.status) < 0) |
2214 |
break; |
2215 |
long ms, ns; |
2216 |
if (deadline == 0L) |
2217 |
ms = 0L; |
2218 |
else if ((ns = deadline - System.nanoTime()) <= 0L) |
2219 |
break; |
2220 |
else if ((ms = TimeUnit.NANOSECONDS.toMillis(ns)) <= 0L) |
2221 |
ms = 1L; |
2222 |
if (tryCompensate(w)) { |
2223 |
task.internalWait(ms); |
2224 |
U.getAndAddLong(this, CTL, AC_UNIT); |
2225 |
} |
2226 |
if ((s = task.status) < 0) |
2227 |
break; |
2228 |
} |
2229 |
w.currentJoin = prevJoin; |
2230 |
} |
2231 |
} |
2232 |
return s; |
2233 |
} |
2234 |
|
2235 |
// Specialized scanning |
2236 |
|
2237 |
/** |
2238 |
* Returns a (probably) non-empty steal queue, if one is found |
2239 |
* during a scan, else null. This method must be retried by |
2240 |
* caller if, by the time it tries to use the queue, it is empty. |
2241 |
*/ |
2242 |
private WorkQueue findNonEmptyStealQueue() { |
2243 |
WorkQueue[] ws; int wl; // one-shot version of scan loop |
2244 |
int r = ThreadLocalRandom.nextSecondarySeed(); |
2245 |
if ((ws = workQueues) != null && (wl = ws.length) > 0) { |
2246 |
int m = wl - 1, origin = r & m; |
2247 |
for (int k = origin, oldSum = 0, checkSum = 0;;) { |
2248 |
WorkQueue q; int b; |
2249 |
if ((q = ws[k]) != null) { |
2250 |
if ((b = q.base) - q.top < 0) |
2251 |
return q; |
2252 |
checkSum += b; |
2253 |
} |
2254 |
if ((k = (k + 1) & m) == origin) { |
2255 |
if (oldSum == (oldSum = checkSum)) |
2256 |
break; |
2257 |
checkSum = 0; |
2258 |
} |
2259 |
} |
2260 |
} |
2261 |
return null; |
2262 |
} |
2263 |
|
2264 |
/** |
2265 |
* Runs tasks until {@code isQuiescent()}. We piggyback on |
2266 |
* active count ctl maintenance, but rather than blocking |
2267 |
* when tasks cannot be found, we rescan until all others cannot |
2268 |
* find tasks either. |
2269 |
*/ |
2270 |
final void helpQuiescePool(WorkQueue w) { |
2271 |
ForkJoinTask<?> ps = w.currentSteal; // save context |
2272 |
int wc = w.config; |
2273 |
for (boolean active = true;;) { |
2274 |
long c; WorkQueue q; ForkJoinTask<?> t; |
2275 |
if (wc >= 0 && (t = w.pop()) != null) { // run locals if LIFO |
2276 |
(w.currentSteal = t).doExec(); |
2277 |
w.currentSteal = ps; |
2278 |
} |
2279 |
else if ((q = findNonEmptyStealQueue()) != null) { |
2280 |
if (!active) { // re-establish active count |
2281 |
active = true; |
2282 |
U.getAndAddLong(this, CTL, AC_UNIT); |
2283 |
} |
2284 |
if ((t = q.pollAt(q.base)) != null) { |
2285 |
(w.currentSteal = t).doExec(); |
2286 |
w.currentSteal = ps; |
2287 |
if (++w.nsteals < 0) |
2288 |
w.transferStealCount(this); |
2289 |
} |
2290 |
} |
2291 |
else if (active) { // decrement active count without queuing |
2292 |
long nc = (AC_MASK & ((c = ctl) - AC_UNIT)) | (~AC_MASK & c); |
2293 |
if (U.compareAndSwapLong(this, CTL, c, nc)) |
2294 |
active = false; |
2295 |
} |
2296 |
else if ((int)((c = ctl) >> AC_SHIFT) + (config & SMASK) <= 0 && |
2297 |
U.compareAndSwapLong(this, CTL, c, c + AC_UNIT)) |
2298 |
break; |
2299 |
} |
2300 |
} |
2301 |
|
2302 |
/** |
2303 |
* Gets and removes a local or stolen task for the given worker. |
2304 |
* |
2305 |
* @return a task, if available |
2306 |
*/ |
2307 |
final ForkJoinTask<?> nextTaskFor(WorkQueue w) { |
2308 |
for (ForkJoinTask<?> t;;) { |
2309 |
WorkQueue q; |
2310 |
if ((t = w.nextLocalTask()) != null) |
2311 |
return t; |
2312 |
if ((q = findNonEmptyStealQueue()) == null) |
2313 |
return null; |
2314 |
if ((t = q.pollAt(q.base)) != null) |
2315 |
return t; |
2316 |
} |
2317 |
} |
2318 |
|
2319 |
/** |
2320 |
* Returns a cheap heuristic guide for task partitioning when |
2321 |
* programmers, frameworks, tools, or languages have little or no |
2322 |
* idea about task granularity. In essence, by offering this |
2323 |
* method, we ask users only about tradeoffs in overhead vs |
2324 |
* expected throughput and its variance, rather than how finely to |
2325 |
* partition tasks. |
2326 |
* |
2327 |
* In a steady state strict (tree-structured) computation, each |
2328 |
* thread makes available for stealing enough tasks for other |
2329 |
* threads to remain active. Inductively, if all threads play by |
2330 |
* the same rules, each thread should make available only a |
2331 |
* constant number of tasks. |
2332 |
* |
2333 |
* The minimum useful constant is just 1. But using a value of 1 |
2334 |
* would require immediate replenishment upon each steal to |
2335 |
* maintain enough tasks, which is infeasible. Further, |
2336 |
* partitionings/granularities of offered tasks should minimize |
2337 |
* steal rates, which in general means that threads nearer the top |
2338 |
* of computation tree should generate more than those nearer the |
2339 |
* bottom. In perfect steady state, each thread is at |
2340 |
* approximately the same level of computation tree. However, |
2341 |
* producing extra tasks amortizes the uncertainty of progress and |
2342 |
* diffusion assumptions. |
2343 |
* |
2344 |
* So, users will want to use values larger (but not much larger) |
2345 |
* than 1 to both smooth over transient shortages and hedge |
2346 |
* against uneven progress; as traded off against the cost of |
2347 |
* extra task overhead. We leave the user to pick a threshold |
2348 |
* value to compare with the results of this call to guide |
2349 |
* decisions, but recommend values such as 3. |
2350 |
* |
2351 |
* When all threads are active, it is on average OK to estimate |
2352 |
* surplus strictly locally. In steady-state, if one thread is |
2353 |
* maintaining say 2 surplus tasks, then so are others. So we can |
2354 |
* just use estimated queue length. However, this strategy alone |
2355 |
* leads to serious mis-estimates in some non-steady-state |
2356 |
* conditions (ramp-up, ramp-down, other stalls). We can detect |
2357 |
* many of these by further considering the number of "idle" |
2358 |
* threads, that are known to have zero queued tasks, so |
2359 |
* compensate by a factor of (#idle/#active) threads. |
2360 |
*/ |
2361 |
static int getSurplusQueuedTaskCount() { |
2362 |
Thread t; ForkJoinWorkerThread wt; ForkJoinPool pool; WorkQueue q; |
2363 |
if ((t = Thread.currentThread()) instanceof ForkJoinWorkerThread) { |
2364 |
int p = (pool = (wt = (ForkJoinWorkerThread)t).pool).config & SMASK; |
2365 |
int n = (q = wt.workQueue).top - q.base; |
2366 |
int a = (int)(pool.ctl >> AC_SHIFT) + p; |
2367 |
return n - (a > (p >>>= 1) ? 0 : |
2368 |
a > (p >>>= 1) ? 1 : |
2369 |
a > (p >>>= 1) ? 2 : |
2370 |
a > (p >>>= 1) ? 4 : |
2371 |
8); |
2372 |
} |
2373 |
return 0; |
2374 |
} |
2375 |
|
2376 |
// Termination |
2377 |
|
2378 |
/** |
2379 |
* Possibly initiates and/or completes termination. |
2380 |
* |
2381 |
* @param now if true, unconditionally terminate, else only |
2382 |
* if no work and no active workers |
2383 |
* @param enable if true, terminate when next possible |
2384 |
* @return true if now terminating or terminated |
2385 |
*/ |
2386 |
private boolean tryTerminate(boolean now, boolean enable) { |
2387 |
AuxState aux; int rs; |
2388 |
if ((rs = runState) >= 0 && (!enable || this == common)) |
2389 |
return false; |
2390 |
while ((aux = auxState) == null) |
2391 |
tryInitialize(false); |
2392 |
aux.lock(); |
2393 |
rs = runState = runState | SHUTDOWN; |
2394 |
aux.unlock(); |
2395 |
if ((rs & STOP) == 0) { |
2396 |
if (!now) { // check quiescence |
2397 |
for (long oldSum = 0L;;) { // repeat until stable |
2398 |
WorkQueue[] ws; WorkQueue w; int m, b, sp; long c; |
2399 |
long checkSum = ctl; |
2400 |
if ((int)(checkSum >> AC_SHIFT) + (config & SMASK) > 0) |
2401 |
return false; // still active workers |
2402 |
if ((ws = workQueues) == null || (m = ws.length - 1) < 0) |
2403 |
break; // check queues |
2404 |
for (int i = 0; i <= m; ++i) { |
2405 |
if ((w = ws[i]) != null) { |
2406 |
if ((b = w.base) != w.top || |
2407 |
w.currentSteal != null) { |
2408 |
while ((sp = (int)(c = ctl)) != 0 && |
2409 |
!tryRelease(c, ws[m & sp], AC_UNIT)) |
2410 |
; |
2411 |
return false; // arrange for recheck |
2412 |
} |
2413 |
checkSum += b; |
2414 |
if ((i & 1) == 0) |
2415 |
w.qlock = -1; // try to disable external |
2416 |
} |
2417 |
} |
2418 |
if (oldSum == (oldSum = checkSum)) |
2419 |
break; |
2420 |
} |
2421 |
} |
2422 |
aux.lock(); |
2423 |
rs = runState = runState | STOP; |
2424 |
aux.unlock(); |
2425 |
} |
2426 |
|
2427 |
int pass = 0; // 3 passes to help terminate |
2428 |
for (long oldSum = 0L;;) { // or until done or stable |
2429 |
WorkQueue[] ws; WorkQueue w; ForkJoinWorkerThread wt; int m; |
2430 |
long checkSum = ctl; |
2431 |
if ((short)(checkSum >>> TC_SHIFT) + (config & SMASK) <= 0 || |
2432 |
(ws = workQueues) == null || (m = ws.length - 1) < 0) { |
2433 |
aux.lock(); |
2434 |
runState |= TERMINATED; |
2435 |
aux.unlock(); |
2436 |
synchronized (this) { |
2437 |
notifyAll(); // for awaitTermination |
2438 |
} |
2439 |
break; |
2440 |
} |
2441 |
for (int i = 0; i <= m; ++i) { |
2442 |
if ((w = ws[i]) != null) { |
2443 |
checkSum += w.base; |
2444 |
w.qlock = -1; // try to disable |
2445 |
if (pass > 0) { |
2446 |
w.cancelAll(); // clear queue |
2447 |
if (pass > 1 && (wt = w.owner) != null) { |
2448 |
if (!wt.isInterrupted()) { |
2449 |
try { // unblock join |
2450 |
wt.interrupt(); |
2451 |
} catch (Throwable ignore) { |
2452 |
} |
2453 |
} |
2454 |
LockSupport.unpark(wt); |
2455 |
} |
2456 |
} |
2457 |
} |
2458 |
} |
2459 |
if (checkSum != oldSum) { // unstable |
2460 |
oldSum = checkSum; |
2461 |
pass = 0; |
2462 |
} |
2463 |
else if (pass > 3 && pass > m) // can't further help |
2464 |
break; |
2465 |
else if (++pass > 1) { // try to dequeue |
2466 |
long c; int j = 0, sp; // bound attempts |
2467 |
while (j++ <= m && (sp = (int)(c = ctl)) != 0) |
2468 |
tryRelease(c, ws[sp & m], AC_UNIT); |
2469 |
} |
2470 |
} |
2471 |
return true; |
2472 |
} |
2473 |
|
2474 |
// External operations |
2475 |
|
2476 |
/** |
2477 |
* Constructs and tries to install a new external queue, |
2478 |
* failing if the workQueues array already has a queue at |
2479 |
* the given index. |
2480 |
* |
2481 |
* @param index the index of the new queue |
2482 |
*/ |
2483 |
private void tryCreateExternalQueue(int index) { |
2484 |
AuxState aux; |
2485 |
if ((aux = auxState) != null && index >= 0) { |
2486 |
WorkQueue q = new WorkQueue(this, null); |
2487 |
q.config = index; |
2488 |
q.scanState = ~UNSIGNALLED; |
2489 |
q.qlock = 1; // lock queue |
2490 |
boolean installed = false; |
2491 |
aux.lock(); |
2492 |
try { // lock pool to install |
2493 |
WorkQueue[] ws; |
2494 |
if ((ws = workQueues) != null && index < ws.length && |
2495 |
ws[index] == null) { |
2496 |
ws[index] = q; // else throw away |
2497 |
installed = true; |
2498 |
} |
2499 |
} finally { |
2500 |
aux.unlock(); |
2501 |
} |
2502 |
if (installed) { |
2503 |
try { |
2504 |
q.growArray(); |
2505 |
} finally { |
2506 |
q.qlock = 0; |
2507 |
} |
2508 |
} |
2509 |
} |
2510 |
} |
2511 |
|
2512 |
/** |
2513 |
* Adds the given task to a submission queue at submitter's |
2514 |
* current queue. Also performs secondary initialization upon the |
2515 |
* first submission of the first task to the pool, and detects |
2516 |
* first submission by an external thread and creates a new shared |
2517 |
* queue if the one at index if empty or contended. |
2518 |
* |
2519 |
* @param task the task. Caller must ensure non-null. |
2520 |
*/ |
2521 |
final void externalPush(ForkJoinTask<?> task) { |
2522 |
int r; // initialize caller's probe |
2523 |
if ((r = ThreadLocalRandom.getProbe()) == 0) { |
2524 |
ThreadLocalRandom.localInit(); |
2525 |
r = ThreadLocalRandom.getProbe(); |
2526 |
} |
2527 |
for (;;) { |
2528 |
WorkQueue q; int wl, k, stat; |
2529 |
int rs = runState; |
2530 |
WorkQueue[] ws = workQueues; |
2531 |
if (rs <= 0 || ws == null || (wl = ws.length) <= 0) |
2532 |
tryInitialize(true); |
2533 |
else if ((q = ws[k = (wl - 1) & r & SQMASK]) == null) |
2534 |
tryCreateExternalQueue(k); |
2535 |
else if ((stat = q.sharedPush(task)) < 0) |
2536 |
break; |
2537 |
else if (stat == 0) { |
2538 |
signalWork(); |
2539 |
break; |
2540 |
} |
2541 |
else // move if busy |
2542 |
r = ThreadLocalRandom.advanceProbe(r); |
2543 |
} |
2544 |
} |
2545 |
|
2546 |
/** |
2547 |
* Pushes a possibly-external submission. |
2548 |
*/ |
2549 |
private <T> ForkJoinTask<T> externalSubmit(ForkJoinTask<T> task) { |
2550 |
Thread t; ForkJoinWorkerThread w; WorkQueue q; |
2551 |
if (task == null) |
2552 |
throw new NullPointerException(); |
2553 |
if (((t = Thread.currentThread()) instanceof ForkJoinWorkerThread) && |
2554 |
(w = (ForkJoinWorkerThread)t).pool == this && |
2555 |
(q = w.workQueue) != null) |
2556 |
q.push(task); |
2557 |
else |
2558 |
externalPush(task); |
2559 |
return task; |
2560 |
} |
2561 |
|
2562 |
/** |
2563 |
* Returns common pool queue for an external thread. |
2564 |
*/ |
2565 |
static WorkQueue commonSubmitterQueue() { |
2566 |
ForkJoinPool p = common; |
2567 |
int r = ThreadLocalRandom.getProbe(); |
2568 |
WorkQueue[] ws; int wl; |
2569 |
return (p != null && (ws = p.workQueues) != null && |
2570 |
(wl = ws.length) > 0) ? |
2571 |
ws[(wl - 1) & r & SQMASK] : null; |
2572 |
} |
2573 |
|
2574 |
/** |
2575 |
* Performs tryUnpush for an external submitter. |
2576 |
*/ |
2577 |
final boolean tryExternalUnpush(ForkJoinTask<?> task) { |
2578 |
int r = ThreadLocalRandom.getProbe(); |
2579 |
WorkQueue[] ws; WorkQueue w; int wl; |
2580 |
return ((ws = workQueues) != null && |
2581 |
(wl = ws.length) > 0 && |
2582 |
(w = ws[(wl - 1) & r & SQMASK]) != null && |
2583 |
w.trySharedUnpush(task)); |
2584 |
} |
2585 |
|
2586 |
/** |
2587 |
* Performs helpComplete for an external submitter. |
2588 |
*/ |
2589 |
final int externalHelpComplete(CountedCompleter<?> task, int maxTasks) { |
2590 |
WorkQueue[] ws; int wl; |
2591 |
int r = ThreadLocalRandom.getProbe(); |
2592 |
return ((ws = workQueues) != null && (wl = ws.length) > 0) ? |
2593 |
helpComplete(ws[(wl - 1) & r & SQMASK], task, maxTasks) : 0; |
2594 |
} |
2595 |
|
2596 |
// Exported methods |
2597 |
|
2598 |
// Constructors |
2599 |
|
2600 |
/** |
2601 |
* Creates a {@code ForkJoinPool} with parallelism equal to {@link |
2602 |
* java.lang.Runtime#availableProcessors}, using the {@linkplain |
2603 |
* #defaultForkJoinWorkerThreadFactory default thread factory}, |
2604 |
* no UncaughtExceptionHandler, and non-async LIFO processing mode. |
2605 |
* |
2606 |
* @throws SecurityException if a security manager exists and |
2607 |
* the caller is not permitted to modify threads |
2608 |
* because it does not hold {@link |
2609 |
* java.lang.RuntimePermission}{@code ("modifyThread")} |
2610 |
*/ |
2611 |
public ForkJoinPool() { |
2612 |
this(Math.min(MAX_CAP, Runtime.getRuntime().availableProcessors()), |
2613 |
defaultForkJoinWorkerThreadFactory, null, false); |
2614 |
} |
2615 |
|
2616 |
/** |
2617 |
* Creates a {@code ForkJoinPool} with the indicated parallelism |
2618 |
* level, the {@linkplain |
2619 |
* #defaultForkJoinWorkerThreadFactory default thread factory}, |
2620 |
* no UncaughtExceptionHandler, and non-async LIFO processing mode. |
2621 |
* |
2622 |
* @param parallelism the parallelism level |
2623 |
* @throws IllegalArgumentException if parallelism less than or |
2624 |
* equal to zero, or greater than implementation limit |
2625 |
* @throws SecurityException if a security manager exists and |
2626 |
* the caller is not permitted to modify threads |
2627 |
* because it does not hold {@link |
2628 |
* java.lang.RuntimePermission}{@code ("modifyThread")} |
2629 |
*/ |
2630 |
public ForkJoinPool(int parallelism) { |
2631 |
this(parallelism, defaultForkJoinWorkerThreadFactory, null, false); |
2632 |
} |
2633 |
|
2634 |
/** |
2635 |
* Creates a {@code ForkJoinPool} with the given parameters. |
2636 |
* |
2637 |
* @param parallelism the parallelism level. For default value, |
2638 |
* use {@link java.lang.Runtime#availableProcessors}. |
2639 |
* @param factory the factory for creating new threads. For default value, |
2640 |
* use {@link #defaultForkJoinWorkerThreadFactory}. |
2641 |
* @param handler the handler for internal worker threads that |
2642 |
* terminate due to unrecoverable errors encountered while executing |
2643 |
* tasks. For default value, use {@code null}. |
2644 |
* @param asyncMode if true, |
2645 |
* establishes local first-in-first-out scheduling mode for forked |
2646 |
* tasks that are never joined. This mode may be more appropriate |
2647 |
* than default locally stack-based mode in applications in which |
2648 |
* worker threads only process event-style asynchronous tasks. |
2649 |
* For default value, use {@code false}. |
2650 |
* @throws IllegalArgumentException if parallelism less than or |
2651 |
* equal to zero, or greater than implementation limit |
2652 |
* @throws NullPointerException if the factory is null |
2653 |
* @throws SecurityException if a security manager exists and |
2654 |
* the caller is not permitted to modify threads |
2655 |
* because it does not hold {@link |
2656 |
* java.lang.RuntimePermission}{@code ("modifyThread")} |
2657 |
*/ |
2658 |
public ForkJoinPool(int parallelism, |
2659 |
ForkJoinWorkerThreadFactory factory, |
2660 |
UncaughtExceptionHandler handler, |
2661 |
boolean asyncMode) { |
2662 |
this(checkParallelism(parallelism), |
2663 |
checkFactory(factory), |
2664 |
handler, |
2665 |
asyncMode ? FIFO_QUEUE : LIFO_QUEUE, |
2666 |
"ForkJoinPool-" + nextPoolId() + "-worker-"); |
2667 |
checkPermission(); |
2668 |
} |
2669 |
|
2670 |
private static int checkParallelism(int parallelism) { |
2671 |
if (parallelism <= 0 || parallelism > MAX_CAP) |
2672 |
throw new IllegalArgumentException(); |
2673 |
return parallelism; |
2674 |
} |
2675 |
|
2676 |
private static ForkJoinWorkerThreadFactory checkFactory |
2677 |
(ForkJoinWorkerThreadFactory factory) { |
2678 |
if (factory == null) |
2679 |
throw new NullPointerException(); |
2680 |
return factory; |
2681 |
} |
2682 |
|
2683 |
/** |
2684 |
* Creates a {@code ForkJoinPool} with the given parameters, without |
2685 |
* any security checks or parameter validation. Invoked directly by |
2686 |
* makeCommonPool. |
2687 |
*/ |
2688 |
private ForkJoinPool(int parallelism, |
2689 |
ForkJoinWorkerThreadFactory factory, |
2690 |
UncaughtExceptionHandler handler, |
2691 |
int mode, |
2692 |
String workerNamePrefix) { |
2693 |
this.workerNamePrefix = workerNamePrefix; |
2694 |
this.factory = factory; |
2695 |
this.ueh = handler; |
2696 |
this.config = (parallelism & SMASK) | mode; |
2697 |
long np = (long)(-parallelism); // offset ctl counts |
2698 |
this.ctl = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK); |
2699 |
} |
2700 |
|
2701 |
/** |
2702 |
* Returns the common pool instance. This pool is statically |
2703 |
* constructed; its run state is unaffected by attempts to {@link |
2704 |
* #shutdown} or {@link #shutdownNow}. However this pool and any |
2705 |
* ongoing processing are automatically terminated upon program |
2706 |
* {@link System#exit}. Any program that relies on asynchronous |
2707 |
* task processing to complete before program termination should |
2708 |
* invoke {@code commonPool().}{@link #awaitQuiescence awaitQuiescence}, |
2709 |
* before exit. |
2710 |
* |
2711 |
* @return the common pool instance |
2712 |
* @since 1.8 |
2713 |
*/ |
2714 |
public static ForkJoinPool commonPool() { |
2715 |
// assert common != null : "static init error"; |
2716 |
return common; |
2717 |
} |
2718 |
|
2719 |
// Execution methods |
2720 |
|
2721 |
/** |
2722 |
* Performs the given task, returning its result upon completion. |
2723 |
* If the computation encounters an unchecked Exception or Error, |
2724 |
* it is rethrown as the outcome of this invocation. Rethrown |
2725 |
* exceptions behave in the same way as regular exceptions, but, |
2726 |
* when possible, contain stack traces (as displayed for example |
2727 |
* using {@code ex.printStackTrace()}) of both the current thread |
2728 |
* as well as the thread actually encountering the exception; |
2729 |
* minimally only the latter. |
2730 |
* |
2731 |
* @param task the task |
2732 |
* @param <T> the type of the task's result |
2733 |
* @return the task's result |
2734 |
* @throws NullPointerException if the task is null |
2735 |
* @throws RejectedExecutionException if the task cannot be |
2736 |
* scheduled for execution |
2737 |
*/ |
2738 |
public <T> T invoke(ForkJoinTask<T> task) { |
2739 |
if (task == null) |
2740 |
throw new NullPointerException(); |
2741 |
externalSubmit(task); |
2742 |
return task.join(); |
2743 |
} |
2744 |
|
2745 |
/** |
2746 |
* Arranges for (asynchronous) execution of the given task. |
2747 |
* |
2748 |
* @param task the task |
2749 |
* @throws NullPointerException if the task is null |
2750 |
* @throws RejectedExecutionException if the task cannot be |
2751 |
* scheduled for execution |
2752 |
*/ |
2753 |
public void execute(ForkJoinTask<?> task) { |
2754 |
externalSubmit(task); |
2755 |
} |
2756 |
|
2757 |
// AbstractExecutorService methods |
2758 |
|
2759 |
/** |
2760 |
* @throws NullPointerException if the task is null |
2761 |
* @throws RejectedExecutionException if the task cannot be |
2762 |
* scheduled for execution |
2763 |
*/ |
2764 |
public void execute(Runnable task) { |
2765 |
if (task == null) |
2766 |
throw new NullPointerException(); |
2767 |
ForkJoinTask<?> job; |
2768 |
if (task instanceof ForkJoinTask<?>) // avoid re-wrap |
2769 |
job = (ForkJoinTask<?>) task; |
2770 |
else |
2771 |
job = new ForkJoinTask.RunnableExecuteAction(task); |
2772 |
externalSubmit(job); |
2773 |
} |
2774 |
|
2775 |
/** |
2776 |
* Submits a ForkJoinTask for execution. |
2777 |
* |
2778 |
* @param task the task to submit |
2779 |
* @param <T> the type of the task's result |
2780 |
* @return the task |
2781 |
* @throws NullPointerException if the task is null |
2782 |
* @throws RejectedExecutionException if the task cannot be |
2783 |
* scheduled for execution |
2784 |
*/ |
2785 |
public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) { |
2786 |
return externalSubmit(task); |
2787 |
} |
2788 |
|
2789 |
/** |
2790 |
* @throws NullPointerException if the task is null |
2791 |
* @throws RejectedExecutionException if the task cannot be |
2792 |
* scheduled for execution |
2793 |
*/ |
2794 |
public <T> ForkJoinTask<T> submit(Callable<T> task) { |
2795 |
return externalSubmit(new ForkJoinTask.AdaptedCallable<T>(task)); |
2796 |
} |
2797 |
|
2798 |
/** |
2799 |
* @throws NullPointerException if the task is null |
2800 |
* @throws RejectedExecutionException if the task cannot be |
2801 |
* scheduled for execution |
2802 |
*/ |
2803 |
public <T> ForkJoinTask<T> submit(Runnable task, T result) { |
2804 |
return externalSubmit(new ForkJoinTask.AdaptedRunnable<T>(task, result)); |
2805 |
} |
2806 |
|
2807 |
/** |
2808 |
* @throws NullPointerException if the task is null |
2809 |
* @throws RejectedExecutionException if the task cannot be |
2810 |
* scheduled for execution |
2811 |
*/ |
2812 |
public ForkJoinTask<?> submit(Runnable task) { |
2813 |
if (task == null) |
2814 |
throw new NullPointerException(); |
2815 |
ForkJoinTask<?> job; |
2816 |
if (task instanceof ForkJoinTask<?>) // avoid re-wrap |
2817 |
job = (ForkJoinTask<?>) task; |
2818 |
else |
2819 |
job = new ForkJoinTask.AdaptedRunnableAction(task); |
2820 |
return externalSubmit(job); |
2821 |
} |
2822 |
|
2823 |
/** |
2824 |
* @throws NullPointerException {@inheritDoc} |
2825 |
* @throws RejectedExecutionException {@inheritDoc} |
2826 |
*/ |
2827 |
public <T> List<Future<T>> invokeAll(Collection<? extends Callable<T>> tasks) { |
2828 |
// In previous versions of this class, this method constructed |
2829 |
// a task to run ForkJoinTask.invokeAll, but now external |
2830 |
// invocation of multiple tasks is at least as efficient. |
2831 |
ArrayList<Future<T>> futures = new ArrayList<>(tasks.size()); |
2832 |
|
2833 |
try { |
2834 |
for (Callable<T> t : tasks) { |
2835 |
ForkJoinTask<T> f = new ForkJoinTask.AdaptedCallable<T>(t); |
2836 |
futures.add(f); |
2837 |
externalSubmit(f); |
2838 |
} |
2839 |
for (int i = 0, size = futures.size(); i < size; i++) |
2840 |
((ForkJoinTask<?>)futures.get(i)).quietlyJoin(); |
2841 |
return futures; |
2842 |
} catch (Throwable t) { |
2843 |
for (int i = 0, size = futures.size(); i < size; i++) |
2844 |
futures.get(i).cancel(false); |
2845 |
throw t; |
2846 |
} |
2847 |
} |
2848 |
|
2849 |
/** |
2850 |
* Returns the factory used for constructing new workers. |
2851 |
* |
2852 |
* @return the factory used for constructing new workers |
2853 |
*/ |
2854 |
public ForkJoinWorkerThreadFactory getFactory() { |
2855 |
return factory; |
2856 |
} |
2857 |
|
2858 |
/** |
2859 |
* Returns the handler for internal worker threads that terminate |
2860 |
* due to unrecoverable errors encountered while executing tasks. |
2861 |
* |
2862 |
* @return the handler, or {@code null} if none |
2863 |
*/ |
2864 |
public UncaughtExceptionHandler getUncaughtExceptionHandler() { |
2865 |
return ueh; |
2866 |
} |
2867 |
|
2868 |
/** |
2869 |
* Returns the targeted parallelism level of this pool. |
2870 |
* |
2871 |
* @return the targeted parallelism level of this pool |
2872 |
*/ |
2873 |
public int getParallelism() { |
2874 |
int par; |
2875 |
return ((par = config & SMASK) > 0) ? par : 1; |
2876 |
} |
2877 |
|
2878 |
/** |
2879 |
* Returns the targeted parallelism level of the common pool. |
2880 |
* |
2881 |
* @return the targeted parallelism level of the common pool |
2882 |
* @since 1.8 |
2883 |
*/ |
2884 |
public static int getCommonPoolParallelism() { |
2885 |
return commonParallelism; |
2886 |
} |
2887 |
|
2888 |
/** |
2889 |
* Returns the number of worker threads that have started but not |
2890 |
* yet terminated. The result returned by this method may differ |
2891 |
* from {@link #getParallelism} when threads are created to |
2892 |
* maintain parallelism when others are cooperatively blocked. |
2893 |
* |
2894 |
* @return the number of worker threads |
2895 |
*/ |
2896 |
public int getPoolSize() { |
2897 |
return (config & SMASK) + (short)(ctl >>> TC_SHIFT); |
2898 |
} |
2899 |
|
2900 |
/** |
2901 |
* Returns {@code true} if this pool uses local first-in-first-out |
2902 |
* scheduling mode for forked tasks that are never joined. |
2903 |
* |
2904 |
* @return {@code true} if this pool uses async mode |
2905 |
*/ |
2906 |
public boolean getAsyncMode() { |
2907 |
return (config & FIFO_QUEUE) != 0; |
2908 |
} |
2909 |
|
2910 |
/** |
2911 |
* Returns an estimate of the number of worker threads that are |
2912 |
* not blocked waiting to join tasks or for other managed |
2913 |
* synchronization. This method may overestimate the |
2914 |
* number of running threads. |
2915 |
* |
2916 |
* @return the number of worker threads |
2917 |
*/ |
2918 |
public int getRunningThreadCount() { |
2919 |
int rc = 0; |
2920 |
WorkQueue[] ws; WorkQueue w; |
2921 |
if ((ws = workQueues) != null) { |
2922 |
for (int i = 1; i < ws.length; i += 2) { |
2923 |
if ((w = ws[i]) != null && w.isApparentlyUnblocked()) |
2924 |
++rc; |
2925 |
} |
2926 |
} |
2927 |
return rc; |
2928 |
} |
2929 |
|
2930 |
/** |
2931 |
* Returns an estimate of the number of threads that are currently |
2932 |
* stealing or executing tasks. This method may overestimate the |
2933 |
* number of active threads. |
2934 |
* |
2935 |
* @return the number of active threads |
2936 |
*/ |
2937 |
public int getActiveThreadCount() { |
2938 |
int r = (config & SMASK) + (int)(ctl >> AC_SHIFT); |
2939 |
return (r <= 0) ? 0 : r; // suppress momentarily negative values |
2940 |
} |
2941 |
|
2942 |
/** |
2943 |
* Returns {@code true} if all worker threads are currently idle. |
2944 |
* An idle worker is one that cannot obtain a task to execute |
2945 |
* because none are available to steal from other threads, and |
2946 |
* there are no pending submissions to the pool. This method is |
2947 |
* conservative; it might not return {@code true} immediately upon |
2948 |
* idleness of all threads, but will eventually become true if |
2949 |
* threads remain inactive. |
2950 |
* |
2951 |
* @return {@code true} if all threads are currently idle |
2952 |
*/ |
2953 |
public boolean isQuiescent() { |
2954 |
return (config & SMASK) + (int)(ctl >> AC_SHIFT) <= 0; |
2955 |
} |
2956 |
|
2957 |
/** |
2958 |
* Returns an estimate of the total number of tasks stolen from |
2959 |
* one thread's work queue by another. The reported value |
2960 |
* underestimates the actual total number of steals when the pool |
2961 |
* is not quiescent. This value may be useful for monitoring and |
2962 |
* tuning fork/join programs: in general, steal counts should be |
2963 |
* high enough to keep threads busy, but low enough to avoid |
2964 |
* overhead and contention across threads. |
2965 |
* |
2966 |
* @return the number of steals |
2967 |
*/ |
2968 |
public long getStealCount() { |
2969 |
AuxState sc = auxState; |
2970 |
long count = (sc == null) ? 0L : sc.stealCount; |
2971 |
WorkQueue[] ws; WorkQueue w; |
2972 |
if ((ws = workQueues) != null) { |
2973 |
for (int i = 1; i < ws.length; i += 2) { |
2974 |
if ((w = ws[i]) != null) |
2975 |
count += w.nsteals; |
2976 |
} |
2977 |
} |
2978 |
return count; |
2979 |
} |
2980 |
|
2981 |
/** |
2982 |
* Returns an estimate of the total number of tasks currently held |
2983 |
* in queues by worker threads (but not including tasks submitted |
2984 |
* to the pool that have not begun executing). This value is only |
2985 |
* an approximation, obtained by iterating across all threads in |
2986 |
* the pool. This method may be useful for tuning task |
2987 |
* granularities. |
2988 |
* |
2989 |
* @return the number of queued tasks |
2990 |
*/ |
2991 |
public long getQueuedTaskCount() { |
2992 |
long count = 0; |
2993 |
WorkQueue[] ws; WorkQueue w; |
2994 |
if ((ws = workQueues) != null) { |
2995 |
for (int i = 1; i < ws.length; i += 2) { |
2996 |
if ((w = ws[i]) != null) |
2997 |
count += w.queueSize(); |
2998 |
} |
2999 |
} |
3000 |
return count; |
3001 |
} |
3002 |
|
3003 |
/** |
3004 |
* Returns an estimate of the number of tasks submitted to this |
3005 |
* pool that have not yet begun executing. This method may take |
3006 |
* time proportional to the number of submissions. |
3007 |
* |
3008 |
* @return the number of queued submissions |
3009 |
*/ |
3010 |
public int getQueuedSubmissionCount() { |
3011 |
int count = 0; |
3012 |
WorkQueue[] ws; WorkQueue w; |
3013 |
if ((ws = workQueues) != null) { |
3014 |
for (int i = 0; i < ws.length; i += 2) { |
3015 |
if ((w = ws[i]) != null) |
3016 |
count += w.queueSize(); |
3017 |
} |
3018 |
} |
3019 |
return count; |
3020 |
} |
3021 |
|
3022 |
/** |
3023 |
* Returns {@code true} if there are any tasks submitted to this |
3024 |
* pool that have not yet begun executing. |
3025 |
* |
3026 |
* @return {@code true} if there are any queued submissions |
3027 |
*/ |
3028 |
public boolean hasQueuedSubmissions() { |
3029 |
WorkQueue[] ws; WorkQueue w; |
3030 |
if ((ws = workQueues) != null) { |
3031 |
for (int i = 0; i < ws.length; i += 2) { |
3032 |
if ((w = ws[i]) != null && !w.isEmpty()) |
3033 |
return true; |
3034 |
} |
3035 |
} |
3036 |
return false; |
3037 |
} |
3038 |
|
3039 |
/** |
3040 |
* Removes and returns the next unexecuted submission if one is |
3041 |
* available. This method may be useful in extensions to this |
3042 |
* class that re-assign work in systems with multiple pools. |
3043 |
* |
3044 |
* @return the next submission, or {@code null} if none |
3045 |
*/ |
3046 |
protected ForkJoinTask<?> pollSubmission() { |
3047 |
WorkQueue[] ws; int wl; WorkQueue w; ForkJoinTask<?> t; |
3048 |
int r = ThreadLocalRandom.nextSecondarySeed(); |
3049 |
if ((ws = workQueues) != null && (wl = ws.length) > 0) { |
3050 |
for (int m = wl - 1, i = 0; i < wl; ++i) { |
3051 |
if ((w = ws[(i << 1) & m]) != null && (t = w.poll()) != null) |
3052 |
return t; |
3053 |
} |
3054 |
} |
3055 |
return null; |
3056 |
} |
3057 |
|
3058 |
/** |
3059 |
* Removes all available unexecuted submitted and forked tasks |
3060 |
* from scheduling queues and adds them to the given collection, |
3061 |
* without altering their execution status. These may include |
3062 |
* artificially generated or wrapped tasks. This method is |
3063 |
* designed to be invoked only when the pool is known to be |
3064 |
* quiescent. Invocations at other times may not remove all |
3065 |
* tasks. A failure encountered while attempting to add elements |
3066 |
* to collection {@code c} may result in elements being in |
3067 |
* neither, either or both collections when the associated |
3068 |
* exception is thrown. The behavior of this operation is |
3069 |
* undefined if the specified collection is modified while the |
3070 |
* operation is in progress. |
3071 |
* |
3072 |
* @param c the collection to transfer elements into |
3073 |
* @return the number of elements transferred |
3074 |
*/ |
3075 |
protected int drainTasksTo(Collection<? super ForkJoinTask<?>> c) { |
3076 |
int count = 0; |
3077 |
WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t; |
3078 |
if ((ws = workQueues) != null) { |
3079 |
for (int i = 0; i < ws.length; ++i) { |
3080 |
if ((w = ws[i]) != null) { |
3081 |
while ((t = w.poll()) != null) { |
3082 |
c.add(t); |
3083 |
++count; |
3084 |
} |
3085 |
} |
3086 |
} |
3087 |
} |
3088 |
return count; |
3089 |
} |
3090 |
|
3091 |
/** |
3092 |
* Returns a string identifying this pool, as well as its state, |
3093 |
* including indications of run state, parallelism level, and |
3094 |
* worker and task counts. |
3095 |
* |
3096 |
* @return a string identifying this pool, as well as its state |
3097 |
*/ |
3098 |
public String toString() { |
3099 |
// Use a single pass through workQueues to collect counts |
3100 |
long qt = 0L, qs = 0L; int rc = 0; |
3101 |
AuxState sc = auxState; |
3102 |
long st = (sc == null) ? 0L : sc.stealCount; |
3103 |
long c = ctl; |
3104 |
WorkQueue[] ws; WorkQueue w; |
3105 |
if ((ws = workQueues) != null) { |
3106 |
for (int i = 0; i < ws.length; ++i) { |
3107 |
if ((w = ws[i]) != null) { |
3108 |
int size = w.queueSize(); |
3109 |
if ((i & 1) == 0) |
3110 |
qs += size; |
3111 |
else { |
3112 |
qt += size; |
3113 |
st += w.nsteals; |
3114 |
if (w.isApparentlyUnblocked()) |
3115 |
++rc; |
3116 |
} |
3117 |
} |
3118 |
} |
3119 |
} |
3120 |
int pc = (config & SMASK); |
3121 |
int tc = pc + (short)(c >>> TC_SHIFT); |
3122 |
int ac = pc + (int)(c >> AC_SHIFT); |
3123 |
if (ac < 0) // ignore transient negative |
3124 |
ac = 0; |
3125 |
int rs = runState; |
3126 |
String level = ((rs & TERMINATED) != 0 ? "Terminated" : |
3127 |
(rs & STOP) != 0 ? "Terminating" : |
3128 |
(rs & SHUTDOWN) != 0 ? "Shutting down" : |
3129 |
"Running"); |
3130 |
return super.toString() + |
3131 |
"[" + level + |
3132 |
", parallelism = " + pc + |
3133 |
", size = " + tc + |
3134 |
", active = " + ac + |
3135 |
", running = " + rc + |
3136 |
", steals = " + st + |
3137 |
", tasks = " + qt + |
3138 |
", submissions = " + qs + |
3139 |
"]"; |
3140 |
} |
3141 |
|
3142 |
/** |
3143 |
* Possibly initiates an orderly shutdown in which previously |
3144 |
* submitted tasks are executed, but no new tasks will be |
3145 |
* accepted. Invocation has no effect on execution state if this |
3146 |
* is the {@link #commonPool()}, and no additional effect if |
3147 |
* already shut down. Tasks that are in the process of being |
3148 |
* submitted concurrently during the course of this method may or |
3149 |
* may not be rejected. |
3150 |
* |
3151 |
* @throws SecurityException if a security manager exists and |
3152 |
* the caller is not permitted to modify threads |
3153 |
* because it does not hold {@link |
3154 |
* java.lang.RuntimePermission}{@code ("modifyThread")} |
3155 |
*/ |
3156 |
public void shutdown() { |
3157 |
checkPermission(); |
3158 |
tryTerminate(false, true); |
3159 |
} |
3160 |
|
3161 |
/** |
3162 |
* Possibly attempts to cancel and/or stop all tasks, and reject |
3163 |
* all subsequently submitted tasks. Invocation has no effect on |
3164 |
* execution state if this is the {@link #commonPool()}, and no |
3165 |
* additional effect if already shut down. Otherwise, tasks that |
3166 |
* are in the process of being submitted or executed concurrently |
3167 |
* during the course of this method may or may not be |
3168 |
* rejected. This method cancels both existing and unexecuted |
3169 |
* tasks, in order to permit termination in the presence of task |
3170 |
* dependencies. So the method always returns an empty list |
3171 |
* (unlike the case for some other Executors). |
3172 |
* |
3173 |
* @return an empty list |
3174 |
* @throws SecurityException if a security manager exists and |
3175 |
* the caller is not permitted to modify threads |
3176 |
* because it does not hold {@link |
3177 |
* java.lang.RuntimePermission}{@code ("modifyThread")} |
3178 |
*/ |
3179 |
public List<Runnable> shutdownNow() { |
3180 |
checkPermission(); |
3181 |
tryTerminate(true, true); |
3182 |
return Collections.emptyList(); |
3183 |
} |
3184 |
|
3185 |
/** |
3186 |
* Returns {@code true} if all tasks have completed following shut down. |
3187 |
* |
3188 |
* @return {@code true} if all tasks have completed following shut down |
3189 |
*/ |
3190 |
public boolean isTerminated() { |
3191 |
return (runState & TERMINATED) != 0; |
3192 |
} |
3193 |
|
3194 |
/** |
3195 |
* Returns {@code true} if the process of termination has |
3196 |
* commenced but not yet completed. This method may be useful for |
3197 |
* debugging. A return of {@code true} reported a sufficient |
3198 |
* period after shutdown may indicate that submitted tasks have |
3199 |
* ignored or suppressed interruption, or are waiting for I/O, |
3200 |
* causing this executor not to properly terminate. (See the |
3201 |
* advisory notes for class {@link ForkJoinTask} stating that |
3202 |
* tasks should not normally entail blocking operations. But if |
3203 |
* they do, they must abort them on interrupt.) |
3204 |
* |
3205 |
* @return {@code true} if terminating but not yet terminated |
3206 |
*/ |
3207 |
public boolean isTerminating() { |
3208 |
int rs = runState; |
3209 |
return (rs & STOP) != 0 && (rs & TERMINATED) == 0; |
3210 |
} |
3211 |
|
3212 |
/** |
3213 |
* Returns {@code true} if this pool has been shut down. |
3214 |
* |
3215 |
* @return {@code true} if this pool has been shut down |
3216 |
*/ |
3217 |
public boolean isShutdown() { |
3218 |
return (runState & SHUTDOWN) != 0; |
3219 |
} |
3220 |
|
3221 |
/** |
3222 |
* Blocks until all tasks have completed execution after a |
3223 |
* shutdown request, or the timeout occurs, or the current thread |
3224 |
* is interrupted, whichever happens first. Because the {@link |
3225 |
* #commonPool()} never terminates until program shutdown, when |
3226 |
* applied to the common pool, this method is equivalent to {@link |
3227 |
* #awaitQuiescence(long, TimeUnit)} but always returns {@code false}. |
3228 |
* |
3229 |
* @param timeout the maximum time to wait |
3230 |
* @param unit the time unit of the timeout argument |
3231 |
* @return {@code true} if this executor terminated and |
3232 |
* {@code false} if the timeout elapsed before termination |
3233 |
* @throws InterruptedException if interrupted while waiting |
3234 |
*/ |
3235 |
public boolean awaitTermination(long timeout, TimeUnit unit) |
3236 |
throws InterruptedException { |
3237 |
if (Thread.interrupted()) |
3238 |
throw new InterruptedException(); |
3239 |
if (this == common) { |
3240 |
awaitQuiescence(timeout, unit); |
3241 |
return false; |
3242 |
} |
3243 |
long nanos = unit.toNanos(timeout); |
3244 |
if (isTerminated()) |
3245 |
return true; |
3246 |
if (nanos <= 0L) |
3247 |
return false; |
3248 |
long deadline = System.nanoTime() + nanos; |
3249 |
synchronized (this) { |
3250 |
for (;;) { |
3251 |
if (isTerminated()) |
3252 |
return true; |
3253 |
if (nanos <= 0L) |
3254 |
return false; |
3255 |
long millis = TimeUnit.NANOSECONDS.toMillis(nanos); |
3256 |
wait(millis > 0L ? millis : 1L); |
3257 |
nanos = deadline - System.nanoTime(); |
3258 |
} |
3259 |
} |
3260 |
} |
3261 |
|
3262 |
/** |
3263 |
* If called by a ForkJoinTask operating in this pool, equivalent |
3264 |
* in effect to {@link ForkJoinTask#helpQuiesce}. Otherwise, |
3265 |
* waits and/or attempts to assist performing tasks until this |
3266 |
* pool {@link #isQuiescent} or the indicated timeout elapses. |
3267 |
* |
3268 |
* @param timeout the maximum time to wait |
3269 |
* @param unit the time unit of the timeout argument |
3270 |
* @return {@code true} if quiescent; {@code false} if the |
3271 |
* timeout elapsed. |
3272 |
*/ |
3273 |
public boolean awaitQuiescence(long timeout, TimeUnit unit) { |
3274 |
long nanos = unit.toNanos(timeout); |
3275 |
ForkJoinWorkerThread wt; |
3276 |
Thread thread = Thread.currentThread(); |
3277 |
if ((thread instanceof ForkJoinWorkerThread) && |
3278 |
(wt = (ForkJoinWorkerThread)thread).pool == this) { |
3279 |
helpQuiescePool(wt.workQueue); |
3280 |
return true; |
3281 |
} |
3282 |
long startTime = System.nanoTime(); |
3283 |
WorkQueue[] ws; |
3284 |
int r = 0, wl; |
3285 |
boolean found = true; |
3286 |
while (!isQuiescent() && (ws = workQueues) != null && |
3287 |
(wl = ws.length) > 0) { |
3288 |
if (!found) { |
3289 |
if ((System.nanoTime() - startTime) > nanos) |
3290 |
return false; |
3291 |
Thread.yield(); // cannot block |
3292 |
} |
3293 |
found = false; |
3294 |
for (int m = wl - 1, j = (m + 1) << 2; j >= 0; --j) { |
3295 |
ForkJoinTask<?> t; WorkQueue q; int b, k; |
3296 |
if ((k = r++ & m) <= m && k >= 0 && (q = ws[k]) != null && |
3297 |
(b = q.base) - q.top < 0) { |
3298 |
found = true; |
3299 |
if ((t = q.pollAt(b)) != null) |
3300 |
t.doExec(); |
3301 |
break; |
3302 |
} |
3303 |
} |
3304 |
} |
3305 |
return true; |
3306 |
} |
3307 |
|
3308 |
/** |
3309 |
* Waits and/or attempts to assist performing tasks indefinitely |
3310 |
* until the {@link #commonPool()} {@link #isQuiescent}. |
3311 |
*/ |
3312 |
static void quiesceCommonPool() { |
3313 |
common.awaitQuiescence(Long.MAX_VALUE, TimeUnit.NANOSECONDS); |
3314 |
} |
3315 |
|
3316 |
/** |
3317 |
* Interface for extending managed parallelism for tasks running |
3318 |
* in {@link ForkJoinPool}s. |
3319 |
* |
3320 |
* <p>A {@code ManagedBlocker} provides two methods. Method |
3321 |
* {@link #isReleasable} must return {@code true} if blocking is |
3322 |
* not necessary. Method {@link #block} blocks the current thread |
3323 |
* if necessary (perhaps internally invoking {@code isReleasable} |
3324 |
* before actually blocking). These actions are performed by any |
3325 |
* thread invoking {@link ForkJoinPool#managedBlock(ManagedBlocker)}. |
3326 |
* The unusual methods in this API accommodate synchronizers that |
3327 |
* may, but don't usually, block for long periods. Similarly, they |
3328 |
* allow more efficient internal handling of cases in which |
3329 |
* additional workers may be, but usually are not, needed to |
3330 |
* ensure sufficient parallelism. Toward this end, |
3331 |
* implementations of method {@code isReleasable} must be amenable |
3332 |
* to repeated invocation. |
3333 |
* |
3334 |
* <p>For example, here is a ManagedBlocker based on a |
3335 |
* ReentrantLock: |
3336 |
* <pre> {@code |
3337 |
* class ManagedLocker implements ManagedBlocker { |
3338 |
* final ReentrantLock lock; |
3339 |
* boolean hasLock = false; |
3340 |
* ManagedLocker(ReentrantLock lock) { this.lock = lock; } |
3341 |
* public boolean block() { |
3342 |
* if (!hasLock) |
3343 |
* lock.lock(); |
3344 |
* return true; |
3345 |
* } |
3346 |
* public boolean isReleasable() { |
3347 |
* return hasLock || (hasLock = lock.tryLock()); |
3348 |
* } |
3349 |
* }}</pre> |
3350 |
* |
3351 |
* <p>Here is a class that possibly blocks waiting for an |
3352 |
* item on a given queue: |
3353 |
* <pre> {@code |
3354 |
* class QueueTaker<E> implements ManagedBlocker { |
3355 |
* final BlockingQueue<E> queue; |
3356 |
* volatile E item = null; |
3357 |
* QueueTaker(BlockingQueue<E> q) { this.queue = q; } |
3358 |
* public boolean block() throws InterruptedException { |
3359 |
* if (item == null) |
3360 |
* item = queue.take(); |
3361 |
* return true; |
3362 |
* } |
3363 |
* public boolean isReleasable() { |
3364 |
* return item != null || (item = queue.poll()) != null; |
3365 |
* } |
3366 |
* public E getItem() { // call after pool.managedBlock completes |
3367 |
* return item; |
3368 |
* } |
3369 |
* }}</pre> |
3370 |
*/ |
3371 |
public static interface ManagedBlocker { |
3372 |
/** |
3373 |
* Possibly blocks the current thread, for example waiting for |
3374 |
* a lock or condition. |
3375 |
* |
3376 |
* @return {@code true} if no additional blocking is necessary |
3377 |
* (i.e., if isReleasable would return true) |
3378 |
* @throws InterruptedException if interrupted while waiting |
3379 |
* (the method is not required to do so, but is allowed to) |
3380 |
*/ |
3381 |
boolean block() throws InterruptedException; |
3382 |
|
3383 |
/** |
3384 |
* Returns {@code true} if blocking is unnecessary. |
3385 |
* @return {@code true} if blocking is unnecessary |
3386 |
*/ |
3387 |
boolean isReleasable(); |
3388 |
} |
3389 |
|
3390 |
/** |
3391 |
* Runs the given possibly blocking task. When {@linkplain |
3392 |
* ForkJoinTask#inForkJoinPool() running in a ForkJoinPool}, this |
3393 |
* method possibly arranges for a spare thread to be activated if |
3394 |
* necessary to ensure sufficient parallelism while the current |
3395 |
* thread is blocked in {@link ManagedBlocker#block blocker.block()}. |
3396 |
* |
3397 |
* <p>This method repeatedly calls {@code blocker.isReleasable()} and |
3398 |
* {@code blocker.block()} until either method returns {@code true}. |
3399 |
* Every call to {@code blocker.block()} is preceded by a call to |
3400 |
* {@code blocker.isReleasable()} that returned {@code false}. |
3401 |
* |
3402 |
* <p>If not running in a ForkJoinPool, this method is |
3403 |
* behaviorally equivalent to |
3404 |
* <pre> {@code |
3405 |
* while (!blocker.isReleasable()) |
3406 |
* if (blocker.block()) |
3407 |
* break;}</pre> |
3408 |
* |
3409 |
* If running in a ForkJoinPool, the pool may first be expanded to |
3410 |
* ensure sufficient parallelism available during the call to |
3411 |
* {@code blocker.block()}. |
3412 |
* |
3413 |
* @param blocker the blocker task |
3414 |
* @throws InterruptedException if {@code blocker.block()} did so |
3415 |
*/ |
3416 |
public static void managedBlock(ManagedBlocker blocker) |
3417 |
throws InterruptedException { |
3418 |
ForkJoinPool p; |
3419 |
ForkJoinWorkerThread wt; |
3420 |
Thread t = Thread.currentThread(); |
3421 |
if ((t instanceof ForkJoinWorkerThread) && |
3422 |
(p = (wt = (ForkJoinWorkerThread)t).pool) != null) { |
3423 |
WorkQueue w = wt.workQueue; |
3424 |
while (!blocker.isReleasable()) { |
3425 |
if (p.tryCompensate(w)) { |
3426 |
try { |
3427 |
do {} while (!blocker.isReleasable() && |
3428 |
!blocker.block()); |
3429 |
} finally { |
3430 |
U.getAndAddLong(p, CTL, AC_UNIT); |
3431 |
} |
3432 |
break; |
3433 |
} |
3434 |
} |
3435 |
} |
3436 |
else { |
3437 |
do {} while (!blocker.isReleasable() && |
3438 |
!blocker.block()); |
3439 |
} |
3440 |
} |
3441 |
|
3442 |
// AbstractExecutorService overrides. These rely on undocumented |
3443 |
// fact that ForkJoinTask.adapt returns ForkJoinTasks that also |
3444 |
// implement RunnableFuture. |
3445 |
|
3446 |
protected <T> RunnableFuture<T> newTaskFor(Runnable runnable, T value) { |
3447 |
return new ForkJoinTask.AdaptedRunnable<T>(runnable, value); |
3448 |
} |
3449 |
|
3450 |
protected <T> RunnableFuture<T> newTaskFor(Callable<T> callable) { |
3451 |
return new ForkJoinTask.AdaptedCallable<T>(callable); |
3452 |
} |
3453 |
|
3454 |
// Unsafe mechanics |
3455 |
private static final sun.misc.Unsafe U = sun.misc.Unsafe.getUnsafe(); |
3456 |
private static final long CTL; |
3457 |
private static final int ABASE; |
3458 |
private static final int ASHIFT; |
3459 |
|
3460 |
static { |
3461 |
try { |
3462 |
CTL = U.objectFieldOffset |
3463 |
(ForkJoinPool.class.getDeclaredField("ctl")); |
3464 |
ABASE = U.arrayBaseOffset(ForkJoinTask[].class); |
3465 |
int scale = U.arrayIndexScale(ForkJoinTask[].class); |
3466 |
if ((scale & (scale - 1)) != 0) |
3467 |
throw new Error("array index scale not a power of two"); |
3468 |
ASHIFT = 31 - Integer.numberOfLeadingZeros(scale); |
3469 |
} catch (ReflectiveOperationException e) { |
3470 |
throw new Error(e); |
3471 |
} |
3472 |
|
3473 |
// Reduce the risk of rare disastrous classloading in first call to |
3474 |
// LockSupport.park: https://bugs.openjdk.java.net/browse/JDK-8074773 |
3475 |
Class<?> ensureLoaded = LockSupport.class; |
3476 |
|
3477 |
commonMaxSpares = DEFAULT_COMMON_MAX_SPARES; |
3478 |
defaultForkJoinWorkerThreadFactory = |
3479 |
new DefaultForkJoinWorkerThreadFactory(); |
3480 |
modifyThreadPermission = new RuntimePermission("modifyThread"); |
3481 |
|
3482 |
common = java.security.AccessController.doPrivileged |
3483 |
(new java.security.PrivilegedAction<ForkJoinPool>() { |
3484 |
public ForkJoinPool run() { return makeCommonPool(); }}); |
3485 |
int par = common.config & SMASK; // report 1 even if threads disabled |
3486 |
commonParallelism = par > 0 ? par : 1; |
3487 |
} |
3488 |
|
3489 |
/** |
3490 |
* Creates and returns the common pool, respecting user settings |
3491 |
* specified via system properties. |
3492 |
*/ |
3493 |
private static ForkJoinPool makeCommonPool() { |
3494 |
int parallelism = -1; |
3495 |
ForkJoinWorkerThreadFactory factory = null; |
3496 |
UncaughtExceptionHandler handler = null; |
3497 |
try { // ignore exceptions in accessing/parsing properties |
3498 |
String pp = System.getProperty |
3499 |
("java.util.concurrent.ForkJoinPool.common.parallelism"); |
3500 |
String fp = System.getProperty |
3501 |
("java.util.concurrent.ForkJoinPool.common.threadFactory"); |
3502 |
String hp = System.getProperty |
3503 |
("java.util.concurrent.ForkJoinPool.common.exceptionHandler"); |
3504 |
String mp = System.getProperty |
3505 |
("java.util.concurrent.ForkJoinPool.common.maximumSpares"); |
3506 |
if (pp != null) |
3507 |
parallelism = Integer.parseInt(pp); |
3508 |
if (fp != null) |
3509 |
factory = ((ForkJoinWorkerThreadFactory)ClassLoader. |
3510 |
getSystemClassLoader().loadClass(fp).newInstance()); |
3511 |
if (hp != null) |
3512 |
handler = ((UncaughtExceptionHandler)ClassLoader. |
3513 |
getSystemClassLoader().loadClass(hp).newInstance()); |
3514 |
if (mp != null) |
3515 |
commonMaxSpares = Integer.parseInt(mp); |
3516 |
} catch (Exception ignore) { |
3517 |
} |
3518 |
if (factory == null) { |
3519 |
if (System.getSecurityManager() == null) |
3520 |
factory = defaultForkJoinWorkerThreadFactory; |
3521 |
else // use security-managed default |
3522 |
factory = new InnocuousForkJoinWorkerThreadFactory(); |
3523 |
} |
3524 |
if (parallelism < 0 && // default 1 less than #cores |
3525 |
(parallelism = Runtime.getRuntime().availableProcessors() - 1) <= 0) |
3526 |
parallelism = 1; |
3527 |
if (parallelism > MAX_CAP) |
3528 |
parallelism = MAX_CAP; |
3529 |
return new ForkJoinPool(parallelism, factory, handler, LIFO_QUEUE, |
3530 |
"ForkJoinPool.commonPool-worker-"); |
3531 |
} |
3532 |
|
3533 |
/** |
3534 |
* Factory for innocuous worker threads |
3535 |
*/ |
3536 |
static final class InnocuousForkJoinWorkerThreadFactory |
3537 |
implements ForkJoinWorkerThreadFactory { |
3538 |
|
3539 |
/** |
3540 |
* An ACC to restrict permissions for the factory itself. |
3541 |
* The constructed workers have no permissions set. |
3542 |
*/ |
3543 |
private static final AccessControlContext innocuousAcc; |
3544 |
static { |
3545 |
Permissions innocuousPerms = new Permissions(); |
3546 |
innocuousPerms.add(modifyThreadPermission); |
3547 |
innocuousPerms.add(new RuntimePermission( |
3548 |
"enableContextClassLoaderOverride")); |
3549 |
innocuousPerms.add(new RuntimePermission( |
3550 |
"modifyThreadGroup")); |
3551 |
innocuousAcc = new AccessControlContext(new ProtectionDomain[] { |
3552 |
new ProtectionDomain(null, innocuousPerms) |
3553 |
}); |
3554 |
} |
3555 |
|
3556 |
public final ForkJoinWorkerThread newThread(ForkJoinPool pool) { |
3557 |
return java.security.AccessController.doPrivileged( |
3558 |
new java.security.PrivilegedAction<ForkJoinWorkerThread>() { |
3559 |
public ForkJoinWorkerThread run() { |
3560 |
return new ForkJoinWorkerThread. |
3561 |
InnocuousForkJoinWorkerThread(pool); |
3562 |
}}, innocuousAcc); |
3563 |
} |
3564 |
} |
3565 |
|
3566 |
} |