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 jsr166e; |
8 |
|
9 |
import java.lang.Thread.UncaughtExceptionHandler; |
10 |
import java.util.ArrayList; |
11 |
import java.util.Arrays; |
12 |
import java.util.Collection; |
13 |
import java.util.Collections; |
14 |
import java.util.List; |
15 |
import java.util.concurrent.AbstractExecutorService; |
16 |
import java.util.concurrent.Callable; |
17 |
import java.util.concurrent.ExecutorService; |
18 |
import java.util.concurrent.Future; |
19 |
import java.util.concurrent.RejectedExecutionException; |
20 |
import java.util.concurrent.RunnableFuture; |
21 |
import java.util.concurrent.ThreadLocalRandom; |
22 |
import java.util.concurrent.TimeUnit; |
23 |
|
24 |
/** |
25 |
* An {@link ExecutorService} for running {@link ForkJoinTask}s. |
26 |
* A {@code ForkJoinPool} provides the entry point for submissions |
27 |
* from non-{@code ForkJoinTask} clients, as well as management and |
28 |
* monitoring operations. |
29 |
* |
30 |
* <p>A {@code ForkJoinPool} differs from other kinds of {@link |
31 |
* ExecutorService} mainly by virtue of employing |
32 |
* <em>work-stealing</em>: all threads in the pool attempt to find and |
33 |
* execute tasks submitted to the pool and/or created by other active |
34 |
* tasks (eventually blocking waiting for work if none exist). This |
35 |
* enables efficient processing when most tasks spawn other subtasks |
36 |
* (as do most {@code ForkJoinTask}s), as well as when many small |
37 |
* tasks are submitted to the pool from external clients. Especially |
38 |
* when setting <em>asyncMode</em> to true in constructors, {@code |
39 |
* ForkJoinPool}s may also be appropriate for use with event-style |
40 |
* tasks that are never joined. |
41 |
* |
42 |
* <p>A static {@link #commonPool()} is available and appropriate for |
43 |
* most applications. The common pool is used by any ForkJoinTask that |
44 |
* is not explicitly submitted to a specified pool. Using the common |
45 |
* pool normally reduces resource usage (its threads are slowly |
46 |
* reclaimed during periods of non-use, and reinstated upon subsequent |
47 |
* use). |
48 |
* |
49 |
* <p>For applications that require separate or custom pools, a {@code |
50 |
* ForkJoinPool} may be constructed with a given target parallelism |
51 |
* level; by default, equal to the number of available processors. The |
52 |
* pool attempts to maintain enough active (or available) threads by |
53 |
* dynamically adding, suspending, or resuming internal worker |
54 |
* threads, even if some tasks are stalled waiting to join others. |
55 |
* However, no such adjustments are guaranteed in the face of blocked |
56 |
* I/O or other unmanaged synchronization. The nested {@link |
57 |
* ManagedBlocker} interface enables extension of the kinds of |
58 |
* synchronization accommodated. |
59 |
* |
60 |
* <p>In addition to execution and lifecycle control methods, this |
61 |
* class provides status check methods (for example |
62 |
* {@link #getStealCount}) that are intended to aid in developing, |
63 |
* tuning, and monitoring fork/join applications. Also, method |
64 |
* {@link #toString} returns indications of pool state in a |
65 |
* convenient form for informal monitoring. |
66 |
* |
67 |
* <p>As is the case with other ExecutorServices, there are three |
68 |
* main task execution methods summarized in the following table. |
69 |
* These are designed to be used primarily by clients not already |
70 |
* engaged in fork/join computations in the current pool. The main |
71 |
* forms of these methods accept instances of {@code ForkJoinTask}, |
72 |
* but overloaded forms also allow mixed execution of plain {@code |
73 |
* Runnable}- or {@code Callable}- based activities as well. However, |
74 |
* tasks that are already executing in a pool should normally instead |
75 |
* use the within-computation forms listed in the table unless using |
76 |
* async event-style tasks that are not usually joined, in which case |
77 |
* there is little difference among choice of methods. |
78 |
* |
79 |
* <table BORDER CELLPADDING=3 CELLSPACING=1> |
80 |
* <caption>Summary of task execution methods</caption> |
81 |
* <tr> |
82 |
* <td></td> |
83 |
* <td ALIGN=CENTER> <b>Call from non-fork/join clients</b></td> |
84 |
* <td ALIGN=CENTER> <b>Call from within fork/join computations</b></td> |
85 |
* </tr> |
86 |
* <tr> |
87 |
* <td> <b>Arrange async execution</b></td> |
88 |
* <td> {@link #execute(ForkJoinTask)}</td> |
89 |
* <td> {@link ForkJoinTask#fork}</td> |
90 |
* </tr> |
91 |
* <tr> |
92 |
* <td> <b>Await and obtain result</b></td> |
93 |
* <td> {@link #invoke(ForkJoinTask)}</td> |
94 |
* <td> {@link ForkJoinTask#invoke}</td> |
95 |
* </tr> |
96 |
* <tr> |
97 |
* <td> <b>Arrange exec and obtain Future</b></td> |
98 |
* <td> {@link #submit(ForkJoinTask)}</td> |
99 |
* <td> {@link ForkJoinTask#fork} (ForkJoinTasks <em>are</em> Futures)</td> |
100 |
* </tr> |
101 |
* </table> |
102 |
* |
103 |
* <p>The common pool is by default constructed with default |
104 |
* parameters, but these may be controlled by setting three |
105 |
* {@linkplain System#getProperty system properties}: |
106 |
* <ul> |
107 |
* <li>{@code java.util.concurrent.ForkJoinPool.common.parallelism} |
108 |
* - the parallelism level, a non-negative integer |
109 |
* <li>{@code java.util.concurrent.ForkJoinPool.common.threadFactory} |
110 |
* - the class name of a {@link ForkJoinWorkerThreadFactory} |
111 |
* <li>{@code java.util.concurrent.ForkJoinPool.common.exceptionHandler} |
112 |
* - the class name of a {@link UncaughtExceptionHandler} |
113 |
* </ul> |
114 |
* The system class loader is used to load these classes. |
115 |
* Upon any error in establishing these settings, default parameters |
116 |
* are used. It is possible to disable or limit the use of threads in |
117 |
* the common pool by setting the parallelism property to zero, and/or |
118 |
* using a factory that may return {@code null}. |
119 |
* |
120 |
* <p><b>Implementation notes</b>: This implementation restricts the |
121 |
* maximum number of running threads to 32767. Attempts to create |
122 |
* pools with greater than the maximum number result in |
123 |
* {@code IllegalArgumentException}. |
124 |
* |
125 |
* <p>This implementation rejects submitted tasks (that is, by throwing |
126 |
* {@link RejectedExecutionException}) only when the pool is shut down |
127 |
* or internal resources have been exhausted. |
128 |
* |
129 |
* @since 1.7 |
130 |
* @author Doug Lea |
131 |
*/ |
132 |
public class ForkJoinPool extends AbstractExecutorService { |
133 |
|
134 |
/* |
135 |
* Implementation Overview |
136 |
* |
137 |
* This class and its nested classes provide the main |
138 |
* functionality and control for a set of worker threads: |
139 |
* Submissions from non-FJ threads enter into submission queues. |
140 |
* Workers take these tasks and typically split them into subtasks |
141 |
* that may be stolen by other workers. Preference rules give |
142 |
* first priority to processing tasks from their own queues (LIFO |
143 |
* or FIFO, depending on mode), then to randomized FIFO steals of |
144 |
* tasks in other queues. |
145 |
* |
146 |
* WorkQueues |
147 |
* ========== |
148 |
* |
149 |
* Most operations occur within work-stealing queues (in nested |
150 |
* class WorkQueue). These are special forms of Deques that |
151 |
* support only three of the four possible end-operations -- push, |
152 |
* pop, and poll (aka steal), under the further constraints that |
153 |
* push and pop are called only from the owning thread (or, as |
154 |
* extended here, under a lock), while poll may be called from |
155 |
* other threads. (If you are unfamiliar with them, you probably |
156 |
* want to read Herlihy and Shavit's book "The Art of |
157 |
* Multiprocessor programming", chapter 16 describing these in |
158 |
* more detail before proceeding.) The main work-stealing queue |
159 |
* design is roughly similar to those in the papers "Dynamic |
160 |
* Circular Work-Stealing Deque" by Chase and Lev, SPAA 2005 |
161 |
* (http://research.sun.com/scalable/pubs/index.html) and |
162 |
* "Idempotent work stealing" by Michael, Saraswat, and Vechev, |
163 |
* PPoPP 2009 (http://portal.acm.org/citation.cfm?id=1504186). |
164 |
* See also "Correct and Efficient Work-Stealing for Weak Memory |
165 |
* Models" by Le, Pop, Cohen, and Nardelli, PPoPP 2013 |
166 |
* (http://www.di.ens.fr/~zappa/readings/ppopp13.pdf) for an |
167 |
* analysis of memory ordering (atomic, volatile etc) issues. The |
168 |
* main differences ultimately stem from GC requirements that we |
169 |
* null out taken slots as soon as we can, to maintain as small a |
170 |
* footprint as possible even in programs generating huge numbers |
171 |
* of tasks. To accomplish this, we shift the CAS arbitrating pop |
172 |
* vs poll (steal) from being on the indices ("base" and "top") to |
173 |
* the slots themselves. So, both a successful pop and poll |
174 |
* mainly entail a CAS of a slot from non-null to null. Because |
175 |
* we rely on CASes of references, we do not need tag bits on base |
176 |
* or top. They are simple ints as used in any circular |
177 |
* array-based queue (see for example ArrayDeque). Updates to the |
178 |
* indices must still be ordered in a way that guarantees that top |
179 |
* == base means the queue is empty, but otherwise may err on the |
180 |
* side of possibly making the queue appear nonempty when a push, |
181 |
* pop, or poll have not fully committed. Note that this means |
182 |
* that the poll operation, considered individually, is not |
183 |
* wait-free. One thief cannot successfully continue until another |
184 |
* in-progress one (or, if previously empty, a push) completes. |
185 |
* However, in the aggregate, we ensure at least probabilistic |
186 |
* non-blockingness. If an attempted steal fails, a thief always |
187 |
* chooses a different random victim target to try next. So, in |
188 |
* order for one thief to progress, it suffices for any |
189 |
* in-progress poll or new push on any empty queue to |
190 |
* complete. (This is why we normally use method pollAt and its |
191 |
* variants that try once at the apparent base index, else |
192 |
* consider alternative actions, rather than method poll.) |
193 |
* |
194 |
* This approach also enables support of a user mode in which local |
195 |
* task processing is in FIFO, not LIFO order, simply by using |
196 |
* poll rather than pop. This can be useful in message-passing |
197 |
* frameworks in which tasks are never joined. However neither |
198 |
* mode considers affinities, loads, cache localities, etc, so |
199 |
* rarely provide the best possible performance on a given |
200 |
* machine, but portably provide good throughput by averaging over |
201 |
* these factors. (Further, even if we did try to use such |
202 |
* information, we do not usually have a basis for exploiting it. |
203 |
* For example, some sets of tasks profit from cache affinities, |
204 |
* but others are harmed by cache pollution effects.) |
205 |
* |
206 |
* WorkQueues are also used in a similar way for tasks submitted |
207 |
* to the pool. We cannot mix these tasks in the same queues used |
208 |
* for work-stealing (this would contaminate lifo/fifo |
209 |
* processing). Instead, we randomly associate submission queues |
210 |
* with submitting threads, using a form of hashing. The |
211 |
* Submitter probe value serves as a hash code for |
212 |
* choosing existing queues, and may be randomly repositioned upon |
213 |
* contention with other submitters. In essence, submitters act |
214 |
* like workers except that they are restricted to executing local |
215 |
* tasks that they submitted (or in the case of CountedCompleters, |
216 |
* others with the same root task). However, because most |
217 |
* shared/external queue operations are more expensive than |
218 |
* internal, and because, at steady state, external submitters |
219 |
* will compete for CPU with workers, ForkJoinTask.join and |
220 |
* related methods disable them from repeatedly helping to process |
221 |
* tasks if all workers are active. Insertion of tasks in shared |
222 |
* mode requires a lock (mainly to protect in the case of |
223 |
* resizing) but we use only a simple spinlock (using bits in |
224 |
* field qlock), because submitters encountering a busy queue move |
225 |
* on to try or create other queues -- they block only when |
226 |
* creating and registering new queues. |
227 |
* |
228 |
* Management |
229 |
* ========== |
230 |
* |
231 |
* The main throughput advantages of work-stealing stem from |
232 |
* decentralized control -- workers mostly take tasks from |
233 |
* themselves or each other. We cannot negate this in the |
234 |
* implementation of other management responsibilities. The main |
235 |
* tactic for avoiding bottlenecks is packing nearly all |
236 |
* essentially atomic control state into two volatile variables |
237 |
* that are by far most often read (not written) as status and |
238 |
* consistency checks. |
239 |
* |
240 |
* Field "ctl" contains 64 bits holding all the information needed |
241 |
* to atomically decide to add, inactivate, enqueue (on an event |
242 |
* queue), dequeue, and/or re-activate workers. To enable this |
243 |
* packing, we restrict maximum parallelism to (1<<15)-1 (which is |
244 |
* far in excess of normal operating range) to allow ids, counts, |
245 |
* and their negations (used for thresholding) to fit into 16bit |
246 |
* fields. |
247 |
* |
248 |
* Field "plock" is a form of sequence lock with a saturating |
249 |
* shutdown bit (similarly for per-queue "qlocks"), mainly |
250 |
* protecting updates to the workQueues array, as well as to |
251 |
* enable shutdown. When used as a lock, it is normally only very |
252 |
* briefly held, so is nearly always available after at most a |
253 |
* brief spin, but we use a monitor-based backup strategy to |
254 |
* block when needed. |
255 |
* |
256 |
* Recording WorkQueues. WorkQueues are recorded in the |
257 |
* "workQueues" array that is created upon first use and expanded |
258 |
* if necessary. Updates to the array while recording new workers |
259 |
* and unrecording terminated ones are protected from each other |
260 |
* by a lock but the array is otherwise concurrently readable, and |
261 |
* accessed directly. To simplify index-based operations, the |
262 |
* array size is always a power of two, and all readers must |
263 |
* tolerate null slots. Worker queues are at odd indices. Shared |
264 |
* (submission) queues are at even indices, up to a maximum of 64 |
265 |
* slots, to limit growth even if array needs to expand to add |
266 |
* more workers. Grouping them together in this way simplifies and |
267 |
* speeds up task scanning. |
268 |
* |
269 |
* All worker thread creation is on-demand, triggered by task |
270 |
* submissions, replacement of terminated workers, and/or |
271 |
* compensation for blocked workers. However, all other support |
272 |
* code is set up to work with other policies. To ensure that we |
273 |
* do not hold on to worker references that would prevent GC, ALL |
274 |
* accesses to workQueues are via indices into the workQueues |
275 |
* array (which is one source of some of the messy code |
276 |
* constructions here). In essence, the workQueues array serves as |
277 |
* a weak reference mechanism. Thus for example the wait queue |
278 |
* field of ctl stores indices, not references. Access to the |
279 |
* workQueues in associated methods (for example signalWork) must |
280 |
* both index-check and null-check the IDs. All such accesses |
281 |
* ignore bad IDs by returning out early from what they are doing, |
282 |
* since this can only be associated with termination, in which |
283 |
* case it is OK to give up. All uses of the workQueues array |
284 |
* also check that it is non-null (even if previously |
285 |
* non-null). This allows nulling during termination, which is |
286 |
* currently not necessary, but remains an option for |
287 |
* resource-revocation-based shutdown schemes. It also helps |
288 |
* reduce JIT issuance of uncommon-trap code, which tends to |
289 |
* unnecessarily complicate control flow in some methods. |
290 |
* |
291 |
* Event Queuing. Unlike HPC work-stealing frameworks, we cannot |
292 |
* let workers spin indefinitely scanning for tasks when none can |
293 |
* be found immediately, and we cannot start/resume workers unless |
294 |
* there appear to be tasks available. On the other hand, we must |
295 |
* quickly prod them into action when new tasks are submitted or |
296 |
* generated. In many usages, ramp-up time to activate workers is |
297 |
* the main limiting factor in overall performance (this is |
298 |
* compounded at program start-up by JIT compilation and |
299 |
* allocation). So we try to streamline this as much as possible. |
300 |
* We park/unpark workers after placing in an event wait queue |
301 |
* when they cannot find work. This "queue" is actually a simple |
302 |
* Treiber stack, headed by the "id" field of ctl, plus a 15bit |
303 |
* counter value (that reflects the number of times a worker has |
304 |
* been inactivated) to avoid ABA effects (we need only as many |
305 |
* version numbers as worker threads). Successors are held in |
306 |
* field WorkQueue.nextWait. Queuing deals with several intrinsic |
307 |
* races, mainly that a task-producing thread can miss seeing (and |
308 |
* signalling) another thread that gave up looking for work but |
309 |
* has not yet entered the wait queue. We solve this by requiring |
310 |
* a full sweep of all workers (via repeated calls to method |
311 |
* scan()) both before and after a newly waiting worker is added |
312 |
* to the wait queue. Because enqueued workers may actually be |
313 |
* rescanning rather than waiting, we set and clear the "parker" |
314 |
* field of WorkQueues to reduce unnecessary calls to unpark. |
315 |
* (This requires a secondary recheck to avoid missed signals.) |
316 |
* Note the unusual conventions about Thread.interrupts |
317 |
* surrounding parking and other blocking: Because interrupts are |
318 |
* used solely to alert threads to check termination, which is |
319 |
* checked anyway upon blocking, we clear status (using |
320 |
* Thread.interrupted) before any call to park, so that park does |
321 |
* not immediately return due to status being set via some other |
322 |
* unrelated call to interrupt in user code. |
323 |
* |
324 |
* Signalling. We create or wake up workers only when there |
325 |
* appears to be at least one task they might be able to find and |
326 |
* execute. When a submission is added or another worker adds a |
327 |
* task to a queue that has fewer than two tasks, they signal |
328 |
* waiting workers (or trigger creation of new ones if fewer than |
329 |
* the given parallelism level -- signalWork). These primary |
330 |
* signals are buttressed by others whenever other threads remove |
331 |
* a task from a queue and notice that there are other tasks there |
332 |
* as well. So in general, pools will be over-signalled. On most |
333 |
* platforms, signalling (unpark) overhead time is noticeably |
334 |
* long, and the time between signalling a thread and it actually |
335 |
* making progress can be very noticeably long, so it is worth |
336 |
* offloading these delays from critical paths as much as |
337 |
* possible. Additionally, workers spin-down gradually, by staying |
338 |
* alive so long as they see the ctl state changing. Similar |
339 |
* stability-sensing techniques are also used before blocking in |
340 |
* awaitJoin and helpComplete. |
341 |
* |
342 |
* Trimming workers. To release resources after periods of lack of |
343 |
* use, a worker starting to wait when the pool is quiescent will |
344 |
* time out and terminate if the pool has remained quiescent for a |
345 |
* given period -- a short period if there are more threads than |
346 |
* parallelism, longer as the number of threads decreases. This |
347 |
* will slowly propagate, eventually terminating all workers after |
348 |
* periods of non-use. |
349 |
* |
350 |
* Shutdown and Termination. A call to shutdownNow atomically sets |
351 |
* a plock bit and then (non-atomically) sets each worker's |
352 |
* qlock status, cancels all unprocessed tasks, and wakes up |
353 |
* all waiting workers. Detecting whether termination should |
354 |
* commence after a non-abrupt shutdown() call requires more work |
355 |
* and bookkeeping. We need consensus about quiescence (i.e., that |
356 |
* there is no more work). The active count provides a primary |
357 |
* indication but non-abrupt shutdown still requires a rechecking |
358 |
* scan for any workers that are inactive but not queued. |
359 |
* |
360 |
* Joining Tasks |
361 |
* ============= |
362 |
* |
363 |
* Any of several actions may be taken when one worker is waiting |
364 |
* to join a task stolen (or always held) by another. Because we |
365 |
* are multiplexing many tasks on to a pool of workers, we can't |
366 |
* just let them block (as in Thread.join). We also cannot just |
367 |
* reassign the joiner's run-time stack with another and replace |
368 |
* it later, which would be a form of "continuation", that even if |
369 |
* possible is not necessarily a good idea since we sometimes need |
370 |
* both an unblocked task and its continuation to progress. |
371 |
* Instead we combine two tactics: |
372 |
* |
373 |
* Helping: Arranging for the joiner to execute some task that it |
374 |
* would be running if the steal had not occurred. |
375 |
* |
376 |
* Compensating: Unless there are already enough live threads, |
377 |
* method tryCompensate() may create or re-activate a spare |
378 |
* thread to compensate for blocked joiners until they unblock. |
379 |
* |
380 |
* A third form (implemented in tryRemoveAndExec) amounts to |
381 |
* helping a hypothetical compensator: If we can readily tell that |
382 |
* a possible action of a compensator is to steal and execute the |
383 |
* task being joined, the joining thread can do so directly, |
384 |
* without the need for a compensation thread (although at the |
385 |
* expense of larger run-time stacks, but the tradeoff is |
386 |
* typically worthwhile). |
387 |
* |
388 |
* The ManagedBlocker extension API can't use helping so relies |
389 |
* only on compensation in method awaitBlocker. |
390 |
* |
391 |
* The algorithm in tryHelpStealer entails a form of "linear" |
392 |
* helping: Each worker records (in field currentSteal) the most |
393 |
* recent task it stole from some other worker. Plus, it records |
394 |
* (in field currentJoin) the task it is currently actively |
395 |
* joining. Method tryHelpStealer uses these markers to try to |
396 |
* find a worker to help (i.e., steal back a task from and execute |
397 |
* it) that could hasten completion of the actively joined task. |
398 |
* In essence, the joiner executes a task that would be on its own |
399 |
* local deque had the to-be-joined task not been stolen. This may |
400 |
* be seen as a conservative variant of the approach in Wagner & |
401 |
* Calder "Leapfrogging: a portable technique for implementing |
402 |
* efficient futures" SIGPLAN Notices, 1993 |
403 |
* (http://portal.acm.org/citation.cfm?id=155354). It differs in |
404 |
* that: (1) We only maintain dependency links across workers upon |
405 |
* steals, rather than use per-task bookkeeping. This sometimes |
406 |
* requires a linear scan of workQueues array to locate stealers, |
407 |
* but often doesn't because stealers leave hints (that may become |
408 |
* stale/wrong) of where to locate them. It is only a hint |
409 |
* because a worker might have had multiple steals and the hint |
410 |
* records only one of them (usually the most current). Hinting |
411 |
* isolates cost to when it is needed, rather than adding to |
412 |
* per-task overhead. (2) It is "shallow", ignoring nesting and |
413 |
* potentially cyclic mutual steals. (3) It is intentionally |
414 |
* racy: field currentJoin is updated only while actively joining, |
415 |
* which means that we miss links in the chain during long-lived |
416 |
* tasks, GC stalls etc (which is OK since blocking in such cases |
417 |
* is usually a good idea). (4) We bound the number of attempts |
418 |
* to find work (see MAX_HELP) and fall back to suspending the |
419 |
* worker and if necessary replacing it with another. |
420 |
* |
421 |
* Helping actions for CountedCompleters are much simpler: Method |
422 |
* helpComplete can take and execute any task with the same root |
423 |
* as the task being waited on. However, this still entails some |
424 |
* traversal of completer chains, so is less efficient than using |
425 |
* CountedCompleters without explicit joins. |
426 |
* |
427 |
* It is impossible to keep exactly the target parallelism number |
428 |
* of threads running at any given time. Determining the |
429 |
* existence of conservatively safe helping targets, the |
430 |
* availability of already-created spares, and the apparent need |
431 |
* to create new spares are all racy, so we rely on multiple |
432 |
* retries of each. Compensation in the apparent absence of |
433 |
* helping opportunities is challenging to control on JVMs, where |
434 |
* GC and other activities can stall progress of tasks that in |
435 |
* turn stall out many other dependent tasks, without us being |
436 |
* able to determine whether they will ever require compensation. |
437 |
* Even though work-stealing otherwise encounters little |
438 |
* degradation in the presence of more threads than cores, |
439 |
* aggressively adding new threads in such cases entails risk of |
440 |
* unwanted positive feedback control loops in which more threads |
441 |
* cause more dependent stalls (as well as delayed progress of |
442 |
* unblocked threads to the point that we know they are available) |
443 |
* leading to more situations requiring more threads, and so |
444 |
* on. This aspect of control can be seen as an (analytically |
445 |
* intractable) game with an opponent that may choose the worst |
446 |
* (for us) active thread to stall at any time. We take several |
447 |
* precautions to bound losses (and thus bound gains), mainly in |
448 |
* methods tryCompensate and awaitJoin. |
449 |
* |
450 |
* Common Pool |
451 |
* =========== |
452 |
* |
453 |
* The static common pool always exists after static |
454 |
* initialization. Since it (or any other created pool) need |
455 |
* never be used, we minimize initial construction overhead and |
456 |
* footprint to the setup of about a dozen fields, with no nested |
457 |
* allocation. Most bootstrapping occurs within method |
458 |
* fullExternalPush during the first submission to the pool. |
459 |
* |
460 |
* When external threads submit to the common pool, they can |
461 |
* perform subtask processing (see externalHelpJoin and related |
462 |
* methods). This caller-helps policy makes it sensible to set |
463 |
* common pool parallelism level to one (or more) less than the |
464 |
* total number of available cores, or even zero for pure |
465 |
* caller-runs. We do not need to record whether external |
466 |
* submissions are to the common pool -- if not, externalHelpJoin |
467 |
* returns quickly (at the most helping to signal some common pool |
468 |
* workers). These submitters would otherwise be blocked waiting |
469 |
* for completion, so the extra effort (with liberally sprinkled |
470 |
* task status checks) in inapplicable cases amounts to an odd |
471 |
* form of limited spin-wait before blocking in ForkJoinTask.join. |
472 |
* |
473 |
* Style notes |
474 |
* =========== |
475 |
* |
476 |
* There is a lot of representation-level coupling among classes |
477 |
* ForkJoinPool, ForkJoinWorkerThread, and ForkJoinTask. The |
478 |
* fields of WorkQueue maintain data structures managed by |
479 |
* ForkJoinPool, so are directly accessed. There is little point |
480 |
* trying to reduce this, since any associated future changes in |
481 |
* representations will need to be accompanied by algorithmic |
482 |
* changes anyway. Several methods intrinsically sprawl because |
483 |
* they must accumulate sets of consistent reads of volatiles held |
484 |
* in local variables. Methods signalWork() and scan() are the |
485 |
* main bottlenecks, so are especially heavily |
486 |
* micro-optimized/mangled. There are lots of inline assignments |
487 |
* (of form "while ((local = field) != 0)") which are usually the |
488 |
* simplest way to ensure the required read orderings (which are |
489 |
* sometimes critical). This leads to a "C"-like style of listing |
490 |
* declarations of these locals at the heads of methods or blocks. |
491 |
* There are several occurrences of the unusual "do {} while |
492 |
* (!cas...)" which is the simplest way to force an update of a |
493 |
* CAS'ed variable. There are also other coding oddities (including |
494 |
* several unnecessary-looking hoisted null checks) that help |
495 |
* some methods perform reasonably even when interpreted (not |
496 |
* compiled). |
497 |
* |
498 |
* The order of declarations in this file is: |
499 |
* (1) Static utility functions |
500 |
* (2) Nested (static) classes |
501 |
* (3) Static fields |
502 |
* (4) Fields, along with constants used when unpacking some of them |
503 |
* (5) Internal control methods |
504 |
* (6) Callbacks and other support for ForkJoinTask methods |
505 |
* (7) Exported methods |
506 |
* (8) Static block initializing statics in minimally dependent order |
507 |
*/ |
508 |
|
509 |
// Static utilities |
510 |
|
511 |
/** |
512 |
* If there is a security manager, makes sure caller has |
513 |
* permission to modify threads. |
514 |
*/ |
515 |
private static void checkPermission() { |
516 |
SecurityManager security = System.getSecurityManager(); |
517 |
if (security != null) |
518 |
security.checkPermission(modifyThreadPermission); |
519 |
} |
520 |
|
521 |
// Nested classes |
522 |
|
523 |
/** |
524 |
* Factory for creating new {@link ForkJoinWorkerThread}s. |
525 |
* A {@code ForkJoinWorkerThreadFactory} must be defined and used |
526 |
* for {@code ForkJoinWorkerThread} subclasses that extend base |
527 |
* functionality or initialize threads with different contexts. |
528 |
*/ |
529 |
public static interface ForkJoinWorkerThreadFactory { |
530 |
/** |
531 |
* Returns a new worker thread operating in the given pool. |
532 |
* |
533 |
* @param pool the pool this thread works in |
534 |
* @throws NullPointerException if the pool is null |
535 |
* @return the new worker thread |
536 |
*/ |
537 |
public ForkJoinWorkerThread newThread(ForkJoinPool pool); |
538 |
} |
539 |
|
540 |
/** |
541 |
* Default ForkJoinWorkerThreadFactory implementation; creates a |
542 |
* new ForkJoinWorkerThread. |
543 |
*/ |
544 |
static final class DefaultForkJoinWorkerThreadFactory |
545 |
implements ForkJoinWorkerThreadFactory { |
546 |
public final ForkJoinWorkerThread newThread(ForkJoinPool pool) { |
547 |
return new ForkJoinWorkerThread(pool); |
548 |
} |
549 |
} |
550 |
|
551 |
/** |
552 |
* Class for artificial tasks that are used to replace the target |
553 |
* of local joins if they are removed from an interior queue slot |
554 |
* in WorkQueue.tryRemoveAndExec. We don't need the proxy to |
555 |
* actually do anything beyond having a unique identity. |
556 |
*/ |
557 |
static final class EmptyTask extends ForkJoinTask<Void> { |
558 |
private static final long serialVersionUID = -7721805057305804111L; |
559 |
EmptyTask() { status = ForkJoinTask.NORMAL; } // force done |
560 |
public final Void getRawResult() { return null; } |
561 |
public final void setRawResult(Void x) {} |
562 |
public final boolean exec() { return true; } |
563 |
} |
564 |
|
565 |
/** |
566 |
* Queues supporting work-stealing as well as external task |
567 |
* submission. See above for main rationale and algorithms. |
568 |
* Implementation relies heavily on "Unsafe" intrinsics |
569 |
* and selective use of "volatile": |
570 |
* |
571 |
* Field "base" is the index (mod array.length) of the least valid |
572 |
* queue slot, which is always the next position to steal (poll) |
573 |
* from if nonempty. Reads and writes require volatile orderings |
574 |
* but not CAS, because updates are only performed after slot |
575 |
* CASes. |
576 |
* |
577 |
* Field "top" is the index (mod array.length) of the next queue |
578 |
* slot to push to or pop from. It is written only by owner thread |
579 |
* for push, or under lock for external/shared push, and accessed |
580 |
* by other threads only after reading (volatile) base. Both top |
581 |
* and base are allowed to wrap around on overflow, but (top - |
582 |
* base) (or more commonly -(base - top) to force volatile read of |
583 |
* base before top) still estimates size. The lock ("qlock") is |
584 |
* forced to -1 on termination, causing all further lock attempts |
585 |
* to fail. (Note: we don't need CAS for termination state because |
586 |
* upon pool shutdown, all shared-queues will stop being used |
587 |
* anyway.) Nearly all lock bodies are set up so that exceptions |
588 |
* within lock bodies are "impossible" (modulo JVM errors that |
589 |
* would cause failure anyway.) |
590 |
* |
591 |
* The array slots are read and written using the emulation of |
592 |
* volatiles/atomics provided by Unsafe. Insertions must in |
593 |
* general use putOrderedObject as a form of releasing store to |
594 |
* ensure that all writes to the task object are ordered before |
595 |
* its publication in the queue. All removals entail a CAS to |
596 |
* null. The array is always a power of two. To ensure safety of |
597 |
* Unsafe array operations, all accesses perform explicit null |
598 |
* checks and implicit bounds checks via power-of-two masking. |
599 |
* |
600 |
* In addition to basic queuing support, this class contains |
601 |
* fields described elsewhere to control execution. It turns out |
602 |
* to work better memory-layout-wise to include them in this class |
603 |
* rather than a separate class. |
604 |
* |
605 |
* Performance on most platforms is very sensitive to placement of |
606 |
* instances of both WorkQueues and their arrays -- we absolutely |
607 |
* do not want multiple WorkQueue instances or multiple queue |
608 |
* arrays sharing cache lines. (It would be best for queue objects |
609 |
* and their arrays to share, but there is nothing available to |
610 |
* help arrange that). The @Contended annotation alerts JVMs to |
611 |
* try to keep instances apart. |
612 |
*/ |
613 |
static final class WorkQueue { |
614 |
/** |
615 |
* Capacity of work-stealing queue array upon initialization. |
616 |
* Must be a power of two; at least 4, but should be larger to |
617 |
* reduce or eliminate cacheline sharing among queues. |
618 |
* Currently, it is much larger, as a partial workaround for |
619 |
* the fact that JVMs often place arrays in locations that |
620 |
* share GC bookkeeping (especially cardmarks) such that |
621 |
* per-write accesses encounter serious memory contention. |
622 |
*/ |
623 |
static final int INITIAL_QUEUE_CAPACITY = 1 << 13; |
624 |
|
625 |
/** |
626 |
* Maximum size for queue arrays. Must be a power of two less |
627 |
* than or equal to 1 << (31 - width of array entry) to ensure |
628 |
* lack of wraparound of index calculations, but defined to a |
629 |
* value a bit less than this to help users trap runaway |
630 |
* programs before saturating systems. |
631 |
*/ |
632 |
static final int MAXIMUM_QUEUE_CAPACITY = 1 << 26; // 64M |
633 |
|
634 |
// Heuristic padding to ameliorate unfortunate memory placements |
635 |
volatile long pad00, pad01, pad02, pad03, pad04, pad05, pad06; |
636 |
|
637 |
volatile int eventCount; // encoded inactivation count; < 0 if inactive |
638 |
int nextWait; // encoded record of next event waiter |
639 |
int nsteals; // number of steals |
640 |
int hint; // steal index hint |
641 |
short poolIndex; // index of this queue in pool |
642 |
final short mode; // 0: lifo, > 0: fifo, < 0: shared |
643 |
volatile int qlock; // 1: locked, -1: terminate; else 0 |
644 |
volatile int base; // index of next slot for poll |
645 |
int top; // index of next slot for push |
646 |
ForkJoinTask<?>[] array; // the elements (initially unallocated) |
647 |
final ForkJoinPool pool; // the containing pool (may be null) |
648 |
final ForkJoinWorkerThread owner; // owning thread or null if shared |
649 |
volatile Thread parker; // == owner during call to park; else null |
650 |
volatile ForkJoinTask<?> currentJoin; // task being joined in awaitJoin |
651 |
ForkJoinTask<?> currentSteal; // current non-local task being executed |
652 |
|
653 |
volatile Object pad10, pad11, pad12, pad13, pad14, pad15, pad16, pad17; |
654 |
volatile Object pad18, pad19, pad1a, pad1b, pad1c, pad1d; |
655 |
|
656 |
WorkQueue(ForkJoinPool pool, ForkJoinWorkerThread owner, int mode, |
657 |
int seed) { |
658 |
this.pool = pool; |
659 |
this.owner = owner; |
660 |
this.mode = (short)mode; |
661 |
this.hint = seed; // store initial seed for runWorker |
662 |
// Place indices in the center of array (that is not yet allocated) |
663 |
base = top = INITIAL_QUEUE_CAPACITY >>> 1; |
664 |
} |
665 |
|
666 |
/** |
667 |
* Returns the approximate number of tasks in the queue. |
668 |
*/ |
669 |
final int queueSize() { |
670 |
int n = base - top; // non-owner callers must read base first |
671 |
return (n >= 0) ? 0 : -n; // ignore transient negative |
672 |
} |
673 |
|
674 |
/** |
675 |
* Provides a more accurate estimate of whether this queue has |
676 |
* any tasks than does queueSize, by checking whether a |
677 |
* near-empty queue has at least one unclaimed task. |
678 |
*/ |
679 |
final boolean isEmpty() { |
680 |
ForkJoinTask<?>[] a; int m, s; |
681 |
int n = base - (s = top); |
682 |
return (n >= 0 || |
683 |
(n == -1 && |
684 |
((a = array) == null || |
685 |
(m = a.length - 1) < 0 || |
686 |
U.getObject |
687 |
(a, (long)((m & (s - 1)) << ASHIFT) + ABASE) == null))); |
688 |
} |
689 |
|
690 |
/** |
691 |
* Pushes a task. Call only by owner in unshared queues. (The |
692 |
* shared-queue version is embedded in method externalPush.) |
693 |
* |
694 |
* @param task the task. Caller must ensure non-null. |
695 |
* @throws RejectedExecutionException if array cannot be resized |
696 |
*/ |
697 |
final void push(ForkJoinTask<?> task) { |
698 |
ForkJoinTask<?>[] a; ForkJoinPool p; |
699 |
int s = top, n; |
700 |
if ((a = array) != null) { // ignore if queue removed |
701 |
int m = a.length - 1; |
702 |
U.putOrderedObject(a, ((m & s) << ASHIFT) + ABASE, task); |
703 |
if ((n = (top = s + 1) - base) <= 2) |
704 |
(p = pool).signalWork(p.workQueues, this); |
705 |
else if (n >= m) |
706 |
growArray(); |
707 |
} |
708 |
} |
709 |
|
710 |
/** |
711 |
* Initializes or doubles the capacity of array. Call either |
712 |
* by owner or with lock held -- it is OK for base, but not |
713 |
* top, to move while resizings are in progress. |
714 |
*/ |
715 |
final ForkJoinTask<?>[] growArray() { |
716 |
ForkJoinTask<?>[] oldA = array; |
717 |
int size = oldA != null ? oldA.length << 1 : INITIAL_QUEUE_CAPACITY; |
718 |
if (size > MAXIMUM_QUEUE_CAPACITY) |
719 |
throw new RejectedExecutionException("Queue capacity exceeded"); |
720 |
int oldMask, t, b; |
721 |
ForkJoinTask<?>[] a = array = new ForkJoinTask<?>[size]; |
722 |
if (oldA != null && (oldMask = oldA.length - 1) >= 0 && |
723 |
(t = top) - (b = base) > 0) { |
724 |
int mask = size - 1; |
725 |
do { |
726 |
ForkJoinTask<?> x; |
727 |
int oldj = ((b & oldMask) << ASHIFT) + ABASE; |
728 |
int j = ((b & mask) << ASHIFT) + ABASE; |
729 |
x = (ForkJoinTask<?>)U.getObjectVolatile(oldA, oldj); |
730 |
if (x != null && |
731 |
U.compareAndSwapObject(oldA, oldj, x, null)) |
732 |
U.putObjectVolatile(a, j, x); |
733 |
} while (++b != t); |
734 |
} |
735 |
return a; |
736 |
} |
737 |
|
738 |
/** |
739 |
* Takes next task, if one exists, in LIFO order. Call only |
740 |
* by owner in unshared queues. |
741 |
*/ |
742 |
final ForkJoinTask<?> pop() { |
743 |
ForkJoinTask<?>[] a; ForkJoinTask<?> t; int m; |
744 |
if ((a = array) != null && (m = a.length - 1) >= 0) { |
745 |
for (int s; (s = top - 1) - base >= 0;) { |
746 |
long j = ((m & s) << ASHIFT) + ABASE; |
747 |
if ((t = (ForkJoinTask<?>)U.getObject(a, j)) == null) |
748 |
break; |
749 |
if (U.compareAndSwapObject(a, j, t, null)) { |
750 |
top = s; |
751 |
return t; |
752 |
} |
753 |
} |
754 |
} |
755 |
return null; |
756 |
} |
757 |
|
758 |
/** |
759 |
* Takes a task in FIFO order if b is base of queue and a task |
760 |
* can be claimed without contention. Specialized versions |
761 |
* appear in ForkJoinPool methods scan and tryHelpStealer. |
762 |
*/ |
763 |
final ForkJoinTask<?> pollAt(int b) { |
764 |
ForkJoinTask<?> t; ForkJoinTask<?>[] a; |
765 |
if ((a = array) != null) { |
766 |
int j = (((a.length - 1) & b) << ASHIFT) + ABASE; |
767 |
if ((t = (ForkJoinTask<?>)U.getObjectVolatile(a, j)) != null && |
768 |
base == b && U.compareAndSwapObject(a, j, t, null)) { |
769 |
U.putOrderedInt(this, QBASE, b + 1); |
770 |
return t; |
771 |
} |
772 |
} |
773 |
return null; |
774 |
} |
775 |
|
776 |
/** |
777 |
* Takes next task, if one exists, in FIFO order. |
778 |
*/ |
779 |
final ForkJoinTask<?> poll() { |
780 |
ForkJoinTask<?>[] a; int b; ForkJoinTask<?> t; |
781 |
while ((b = base) - top < 0 && (a = array) != null) { |
782 |
int j = (((a.length - 1) & b) << ASHIFT) + ABASE; |
783 |
t = (ForkJoinTask<?>)U.getObjectVolatile(a, j); |
784 |
if (t != null) { |
785 |
if (U.compareAndSwapObject(a, j, t, null)) { |
786 |
U.putOrderedInt(this, QBASE, b + 1); |
787 |
return t; |
788 |
} |
789 |
} |
790 |
else if (base == b) { |
791 |
if (b + 1 == top) |
792 |
break; |
793 |
Thread.yield(); // wait for lagging update (very rare) |
794 |
} |
795 |
} |
796 |
return null; |
797 |
} |
798 |
|
799 |
/** |
800 |
* Takes next task, if one exists, in order specified by mode. |
801 |
*/ |
802 |
final ForkJoinTask<?> nextLocalTask() { |
803 |
return mode == 0 ? pop() : poll(); |
804 |
} |
805 |
|
806 |
/** |
807 |
* Returns next task, if one exists, in order specified by mode. |
808 |
*/ |
809 |
final ForkJoinTask<?> peek() { |
810 |
ForkJoinTask<?>[] a = array; int m; |
811 |
if (a == null || (m = a.length - 1) < 0) |
812 |
return null; |
813 |
int i = mode == 0 ? top - 1 : base; |
814 |
int j = ((i & m) << ASHIFT) + ABASE; |
815 |
return (ForkJoinTask<?>)U.getObjectVolatile(a, j); |
816 |
} |
817 |
|
818 |
/** |
819 |
* Pops the given task only if it is at the current top. |
820 |
* (A shared version is available only via FJP.tryExternalUnpush) |
821 |
*/ |
822 |
final boolean tryUnpush(ForkJoinTask<?> t) { |
823 |
ForkJoinTask<?>[] a; int s; |
824 |
if ((a = array) != null && (s = top) != base && |
825 |
U.compareAndSwapObject |
826 |
(a, (((a.length - 1) & --s) << ASHIFT) + ABASE, t, null)) { |
827 |
top = s; |
828 |
return true; |
829 |
} |
830 |
return false; |
831 |
} |
832 |
|
833 |
/** |
834 |
* Removes and cancels all known tasks, ignoring any exceptions. |
835 |
*/ |
836 |
final void cancelAll() { |
837 |
ForkJoinTask.cancelIgnoringExceptions(currentJoin); |
838 |
ForkJoinTask.cancelIgnoringExceptions(currentSteal); |
839 |
for (ForkJoinTask<?> t; (t = poll()) != null; ) |
840 |
ForkJoinTask.cancelIgnoringExceptions(t); |
841 |
} |
842 |
|
843 |
// Specialized execution methods |
844 |
|
845 |
/** |
846 |
* Polls and runs tasks until empty. |
847 |
*/ |
848 |
final void pollAndExecAll() { |
849 |
for (ForkJoinTask<?> t; (t = poll()) != null;) |
850 |
t.doExec(); |
851 |
} |
852 |
|
853 |
/** |
854 |
* Executes a top-level task and any local tasks remaining |
855 |
* after execution. |
856 |
*/ |
857 |
final void runTask(ForkJoinTask<?> task) { |
858 |
if ((currentSteal = task) != null) { |
859 |
task.doExec(); |
860 |
ForkJoinTask<?>[] a = array; |
861 |
int md = mode; |
862 |
++nsteals; |
863 |
currentSteal = null; |
864 |
if (md != 0) |
865 |
pollAndExecAll(); |
866 |
else if (a != null) { |
867 |
int s, m = a.length - 1; |
868 |
while ((s = top - 1) - base >= 0) { |
869 |
long i = ((m & s) << ASHIFT) + ABASE; |
870 |
ForkJoinTask<?> t = (ForkJoinTask<?>)U.getObject(a, i); |
871 |
if (t == null) |
872 |
break; |
873 |
if (U.compareAndSwapObject(a, i, t, null)) { |
874 |
top = s; |
875 |
t.doExec(); |
876 |
} |
877 |
} |
878 |
} |
879 |
} |
880 |
} |
881 |
|
882 |
/** |
883 |
* If present, removes from queue and executes the given task, |
884 |
* or any other cancelled task. Returns (true) on any CAS |
885 |
* or consistency check failure so caller can retry. |
886 |
* |
887 |
* @return false if no progress can be made, else true |
888 |
*/ |
889 |
final boolean tryRemoveAndExec(ForkJoinTask<?> task) { |
890 |
boolean stat; |
891 |
ForkJoinTask<?>[] a; int m, s, b, n; |
892 |
if (task != null && (a = array) != null && (m = a.length - 1) >= 0 && |
893 |
(n = (s = top) - (b = base)) > 0) { |
894 |
boolean removed = false, empty = true; |
895 |
stat = true; |
896 |
for (ForkJoinTask<?> t;;) { // traverse from s to b |
897 |
long j = ((--s & m) << ASHIFT) + ABASE; |
898 |
t = (ForkJoinTask<?>)U.getObject(a, j); |
899 |
if (t == null) // inconsistent length |
900 |
break; |
901 |
else if (t == task) { |
902 |
if (s + 1 == top) { // pop |
903 |
if (!U.compareAndSwapObject(a, j, task, null)) |
904 |
break; |
905 |
top = s; |
906 |
removed = true; |
907 |
} |
908 |
else if (base == b) // replace with proxy |
909 |
removed = U.compareAndSwapObject(a, j, task, |
910 |
new EmptyTask()); |
911 |
break; |
912 |
} |
913 |
else if (t.status >= 0) |
914 |
empty = false; |
915 |
else if (s + 1 == top) { // pop and throw away |
916 |
if (U.compareAndSwapObject(a, j, t, null)) |
917 |
top = s; |
918 |
break; |
919 |
} |
920 |
if (--n == 0) { |
921 |
if (!empty && base == b) |
922 |
stat = false; |
923 |
break; |
924 |
} |
925 |
} |
926 |
if (removed) |
927 |
task.doExec(); |
928 |
} |
929 |
else |
930 |
stat = false; |
931 |
return stat; |
932 |
} |
933 |
|
934 |
/** |
935 |
* Tries to poll for and execute the given task or any other |
936 |
* task in its CountedCompleter computation. |
937 |
*/ |
938 |
final boolean pollAndExecCC(CountedCompleter<?> root) { |
939 |
ForkJoinTask<?>[] a; int b; Object o; CountedCompleter<?> t, r; |
940 |
if ((b = base) - top < 0 && (a = array) != null) { |
941 |
long j = (((a.length - 1) & b) << ASHIFT) + ABASE; |
942 |
if ((o = U.getObjectVolatile(a, j)) == null) |
943 |
return true; // retry |
944 |
if (o instanceof CountedCompleter) { |
945 |
for (t = (CountedCompleter<?>)o, r = t;;) { |
946 |
if (r == root) { |
947 |
if (base == b && |
948 |
U.compareAndSwapObject(a, j, t, null)) { |
949 |
U.putOrderedInt(this, QBASE, b + 1); |
950 |
t.doExec(); |
951 |
} |
952 |
return true; |
953 |
} |
954 |
else if ((r = r.completer) == null) |
955 |
break; // not part of root computation |
956 |
} |
957 |
} |
958 |
} |
959 |
return false; |
960 |
} |
961 |
|
962 |
/** |
963 |
* Tries to pop and execute the given task or any other task |
964 |
* in its CountedCompleter computation. |
965 |
*/ |
966 |
final boolean externalPopAndExecCC(CountedCompleter<?> root) { |
967 |
ForkJoinTask<?>[] a; int s; Object o; CountedCompleter<?> t, r; |
968 |
if (base - (s = top) < 0 && (a = array) != null) { |
969 |
long j = (((a.length - 1) & (s - 1)) << ASHIFT) + ABASE; |
970 |
if ((o = U.getObject(a, j)) instanceof CountedCompleter) { |
971 |
for (t = (CountedCompleter<?>)o, r = t;;) { |
972 |
if (r == root) { |
973 |
if (U.compareAndSwapInt(this, QLOCK, 0, 1)) { |
974 |
if (top == s && array == a && |
975 |
U.compareAndSwapObject(a, j, t, null)) { |
976 |
top = s - 1; |
977 |
qlock = 0; |
978 |
t.doExec(); |
979 |
} |
980 |
else |
981 |
qlock = 0; |
982 |
} |
983 |
return true; |
984 |
} |
985 |
else if ((r = r.completer) == null) |
986 |
break; |
987 |
} |
988 |
} |
989 |
} |
990 |
return false; |
991 |
} |
992 |
|
993 |
/** |
994 |
* Internal version |
995 |
*/ |
996 |
final boolean internalPopAndExecCC(CountedCompleter<?> root) { |
997 |
ForkJoinTask<?>[] a; int s; Object o; CountedCompleter<?> t, r; |
998 |
if (base - (s = top) < 0 && (a = array) != null) { |
999 |
long j = (((a.length - 1) & (s - 1)) << ASHIFT) + ABASE; |
1000 |
if ((o = U.getObject(a, j)) instanceof CountedCompleter) { |
1001 |
for (t = (CountedCompleter<?>)o, r = t;;) { |
1002 |
if (r == root) { |
1003 |
if (U.compareAndSwapObject(a, j, t, null)) { |
1004 |
top = s - 1; |
1005 |
t.doExec(); |
1006 |
} |
1007 |
return true; |
1008 |
} |
1009 |
else if ((r = r.completer) == null) |
1010 |
break; |
1011 |
} |
1012 |
} |
1013 |
} |
1014 |
return false; |
1015 |
} |
1016 |
|
1017 |
/** |
1018 |
* Returns true if owned and not known to be blocked. |
1019 |
*/ |
1020 |
final boolean isApparentlyUnblocked() { |
1021 |
Thread wt; Thread.State s; |
1022 |
return (eventCount >= 0 && |
1023 |
(wt = owner) != null && |
1024 |
(s = wt.getState()) != Thread.State.BLOCKED && |
1025 |
s != Thread.State.WAITING && |
1026 |
s != Thread.State.TIMED_WAITING); |
1027 |
} |
1028 |
|
1029 |
// Unsafe mechanics |
1030 |
private static final sun.misc.Unsafe U; |
1031 |
private static final long QBASE; |
1032 |
private static final long QLOCK; |
1033 |
private static final int ABASE; |
1034 |
private static final int ASHIFT; |
1035 |
static { |
1036 |
try { |
1037 |
U = getUnsafe(); |
1038 |
Class<?> k = WorkQueue.class; |
1039 |
Class<?> ak = ForkJoinTask[].class; |
1040 |
QBASE = U.objectFieldOffset |
1041 |
(k.getDeclaredField("base")); |
1042 |
QLOCK = U.objectFieldOffset |
1043 |
(k.getDeclaredField("qlock")); |
1044 |
ABASE = U.arrayBaseOffset(ak); |
1045 |
int scale = U.arrayIndexScale(ak); |
1046 |
if ((scale & (scale - 1)) != 0) |
1047 |
throw new Error("data type scale not a power of two"); |
1048 |
ASHIFT = 31 - Integer.numberOfLeadingZeros(scale); |
1049 |
} catch (Exception e) { |
1050 |
throw new Error(e); |
1051 |
} |
1052 |
} |
1053 |
} |
1054 |
|
1055 |
// static fields (initialized in static initializer below) |
1056 |
|
1057 |
/** |
1058 |
* Per-thread submission bookkeeping. Shared across all pools |
1059 |
* to reduce ThreadLocal pollution and because random motion |
1060 |
* to avoid contention in one pool is likely to hold for others. |
1061 |
* Lazily initialized on first submission (but null-checked |
1062 |
* in other contexts to avoid unnecessary initialization). |
1063 |
*/ |
1064 |
static final ThreadLocal<Submitter> submitters; |
1065 |
|
1066 |
/** |
1067 |
* Creates a new ForkJoinWorkerThread. This factory is used unless |
1068 |
* overridden in ForkJoinPool constructors. |
1069 |
*/ |
1070 |
public static final ForkJoinWorkerThreadFactory |
1071 |
defaultForkJoinWorkerThreadFactory; |
1072 |
|
1073 |
/** |
1074 |
* Permission required for callers of methods that may start or |
1075 |
* kill threads. |
1076 |
*/ |
1077 |
private static final RuntimePermission modifyThreadPermission; |
1078 |
|
1079 |
/** |
1080 |
* Common (static) pool. Non-null for public use unless a static |
1081 |
* construction exception, but internal usages null-check on use |
1082 |
* to paranoically avoid potential initialization circularities |
1083 |
* as well as to simplify generated code. |
1084 |
*/ |
1085 |
static final ForkJoinPool common; |
1086 |
|
1087 |
/** |
1088 |
* Common pool parallelism. To allow simpler use and management |
1089 |
* when common pool threads are disabled, we allow the underlying |
1090 |
* common.parallelism field to be zero, but in that case still report |
1091 |
* parallelism as 1 to reflect resulting caller-runs mechanics. |
1092 |
*/ |
1093 |
static final int commonParallelism; |
1094 |
|
1095 |
/** |
1096 |
* Sequence number for creating workerNamePrefix. |
1097 |
*/ |
1098 |
private static int poolNumberSequence; |
1099 |
|
1100 |
/** |
1101 |
* Returns the next sequence number. We don't expect this to |
1102 |
* ever contend, so use simple builtin sync. |
1103 |
*/ |
1104 |
private static final synchronized int nextPoolId() { |
1105 |
return ++poolNumberSequence; |
1106 |
} |
1107 |
|
1108 |
// static constants |
1109 |
|
1110 |
/** |
1111 |
* Initial timeout value (in nanoseconds) for the thread |
1112 |
* triggering quiescence to park waiting for new work. On timeout, |
1113 |
* the thread will instead try to shrink the number of |
1114 |
* workers. The value should be large enough to avoid overly |
1115 |
* aggressive shrinkage during most transient stalls (long GCs |
1116 |
* etc). |
1117 |
*/ |
1118 |
private static final long IDLE_TIMEOUT = 2000L * 1000L * 1000L; // 2sec |
1119 |
|
1120 |
/** |
1121 |
* Timeout value when there are more threads than parallelism level |
1122 |
*/ |
1123 |
private static final long FAST_IDLE_TIMEOUT = 200L * 1000L * 1000L; |
1124 |
|
1125 |
/** |
1126 |
* Tolerance for idle timeouts, to cope with timer undershoots |
1127 |
*/ |
1128 |
private static final long TIMEOUT_SLOP = 2000000L; |
1129 |
|
1130 |
/** |
1131 |
* The maximum stolen->joining link depth allowed in method |
1132 |
* tryHelpStealer. Must be a power of two. Depths for legitimate |
1133 |
* chains are unbounded, but we use a fixed constant to avoid |
1134 |
* (otherwise unchecked) cycles and to bound staleness of |
1135 |
* traversal parameters at the expense of sometimes blocking when |
1136 |
* we could be helping. |
1137 |
*/ |
1138 |
private static final int MAX_HELP = 64; |
1139 |
|
1140 |
/** |
1141 |
* Increment for seed generators. See class ThreadLocal for |
1142 |
* explanation. |
1143 |
*/ |
1144 |
private static final int SEED_INCREMENT = 0x61c88647; |
1145 |
|
1146 |
/* |
1147 |
* Bits and masks for control variables |
1148 |
* |
1149 |
* Field ctl is a long packed with: |
1150 |
* AC: Number of active running workers minus target parallelism (16 bits) |
1151 |
* TC: Number of total workers minus target parallelism (16 bits) |
1152 |
* ST: true if pool is terminating (1 bit) |
1153 |
* EC: the wait count of top waiting thread (15 bits) |
1154 |
* ID: poolIndex of top of Treiber stack of waiters (16 bits) |
1155 |
* |
1156 |
* When convenient, we can extract the upper 32 bits of counts and |
1157 |
* the lower 32 bits of queue state, u = (int)(ctl >>> 32) and e = |
1158 |
* (int)ctl. The ec field is never accessed alone, but always |
1159 |
* together with id and st. The offsets of counts by the target |
1160 |
* parallelism and the positionings of fields makes it possible to |
1161 |
* perform the most common checks via sign tests of fields: When |
1162 |
* ac is negative, there are not enough active workers, when tc is |
1163 |
* negative, there are not enough total workers, and when e is |
1164 |
* negative, the pool is terminating. To deal with these possibly |
1165 |
* negative fields, we use casts in and out of "short" and/or |
1166 |
* signed shifts to maintain signedness. |
1167 |
* |
1168 |
* When a thread is queued (inactivated), its eventCount field is |
1169 |
* set negative, which is the only way to tell if a worker is |
1170 |
* prevented from executing tasks, even though it must continue to |
1171 |
* scan for them to avoid queuing races. Note however that |
1172 |
* eventCount updates lag releases so usage requires care. |
1173 |
* |
1174 |
* Field plock is an int packed with: |
1175 |
* SHUTDOWN: true if shutdown is enabled (1 bit) |
1176 |
* SEQ: a sequence lock, with PL_LOCK bit set if locked (30 bits) |
1177 |
* SIGNAL: set when threads may be waiting on the lock (1 bit) |
1178 |
* |
1179 |
* The sequence number enables simple consistency checks: |
1180 |
* Staleness of read-only operations on the workQueues array can |
1181 |
* be checked by comparing plock before vs after the reads. |
1182 |
*/ |
1183 |
|
1184 |
// bit positions/shifts for fields |
1185 |
private static final int AC_SHIFT = 48; |
1186 |
private static final int TC_SHIFT = 32; |
1187 |
private static final int ST_SHIFT = 31; |
1188 |
private static final int EC_SHIFT = 16; |
1189 |
|
1190 |
// bounds |
1191 |
private static final int SMASK = 0xffff; // short bits |
1192 |
private static final int MAX_CAP = 0x7fff; // max #workers - 1 |
1193 |
private static final int EVENMASK = 0xfffe; // even short bits |
1194 |
private static final int SQMASK = 0x007e; // max 64 (even) slots |
1195 |
private static final int SHORT_SIGN = 1 << 15; |
1196 |
private static final int INT_SIGN = 1 << 31; |
1197 |
|
1198 |
// masks |
1199 |
private static final long STOP_BIT = 0x0001L << ST_SHIFT; |
1200 |
private static final long AC_MASK = ((long)SMASK) << AC_SHIFT; |
1201 |
private static final long TC_MASK = ((long)SMASK) << TC_SHIFT; |
1202 |
|
1203 |
// units for incrementing and decrementing |
1204 |
private static final long TC_UNIT = 1L << TC_SHIFT; |
1205 |
private static final long AC_UNIT = 1L << AC_SHIFT; |
1206 |
|
1207 |
// masks and units for dealing with u = (int)(ctl >>> 32) |
1208 |
private static final int UAC_SHIFT = AC_SHIFT - 32; |
1209 |
private static final int UTC_SHIFT = TC_SHIFT - 32; |
1210 |
private static final int UAC_MASK = SMASK << UAC_SHIFT; |
1211 |
private static final int UTC_MASK = SMASK << UTC_SHIFT; |
1212 |
private static final int UAC_UNIT = 1 << UAC_SHIFT; |
1213 |
private static final int UTC_UNIT = 1 << UTC_SHIFT; |
1214 |
|
1215 |
// masks and units for dealing with e = (int)ctl |
1216 |
private static final int E_MASK = 0x7fffffff; // no STOP_BIT |
1217 |
private static final int E_SEQ = 1 << EC_SHIFT; |
1218 |
|
1219 |
// plock bits |
1220 |
private static final int SHUTDOWN = 1 << 31; |
1221 |
private static final int PL_LOCK = 2; |
1222 |
private static final int PL_SIGNAL = 1; |
1223 |
private static final int PL_SPINS = 1 << 8; |
1224 |
|
1225 |
// access mode for WorkQueue |
1226 |
static final int LIFO_QUEUE = 0; |
1227 |
static final int FIFO_QUEUE = 1; |
1228 |
static final int SHARED_QUEUE = -1; |
1229 |
|
1230 |
// Heuristic padding to ameliorate unfortunate memory placements |
1231 |
volatile long pad00, pad01, pad02, pad03, pad04, pad05, pad06; |
1232 |
|
1233 |
// Instance fields |
1234 |
volatile long stealCount; // collects worker counts |
1235 |
volatile long ctl; // main pool control |
1236 |
volatile int plock; // shutdown status and seqLock |
1237 |
volatile int indexSeed; // worker/submitter index seed |
1238 |
final short parallelism; // parallelism level |
1239 |
final short mode; // LIFO/FIFO |
1240 |
WorkQueue[] workQueues; // main registry |
1241 |
final ForkJoinWorkerThreadFactory factory; |
1242 |
final UncaughtExceptionHandler ueh; // per-worker UEH |
1243 |
final String workerNamePrefix; // to create worker name string |
1244 |
|
1245 |
volatile Object pad10, pad11, pad12, pad13, pad14, pad15, pad16, pad17; |
1246 |
volatile Object pad18, pad19, pad1a, pad1b; |
1247 |
|
1248 |
/** |
1249 |
* Acquires the plock lock to protect worker array and related |
1250 |
* updates. This method is called only if an initial CAS on plock |
1251 |
* fails. This acts as a spinlock for normal cases, but falls back |
1252 |
* to builtin monitor to block when (rarely) needed. This would be |
1253 |
* a terrible idea for a highly contended lock, but works fine as |
1254 |
* a more conservative alternative to a pure spinlock. |
1255 |
*/ |
1256 |
private int acquirePlock() { |
1257 |
int spins = PL_SPINS, ps, nps; |
1258 |
for (;;) { |
1259 |
if (((ps = plock) & PL_LOCK) == 0 && |
1260 |
U.compareAndSwapInt(this, PLOCK, ps, nps = ps + PL_LOCK)) |
1261 |
return nps; |
1262 |
else if (spins >= 0) { |
1263 |
if (ThreadLocalRandom.current().nextInt() >= 0) |
1264 |
--spins; |
1265 |
} |
1266 |
else if (U.compareAndSwapInt(this, PLOCK, ps, ps | PL_SIGNAL)) { |
1267 |
synchronized (this) { |
1268 |
if ((plock & PL_SIGNAL) != 0) { |
1269 |
try { |
1270 |
wait(); |
1271 |
} catch (InterruptedException ie) { |
1272 |
try { |
1273 |
Thread.currentThread().interrupt(); |
1274 |
} catch (SecurityException ignore) { |
1275 |
} |
1276 |
} |
1277 |
} |
1278 |
else |
1279 |
notifyAll(); |
1280 |
} |
1281 |
} |
1282 |
} |
1283 |
} |
1284 |
|
1285 |
/** |
1286 |
* Unlocks and signals any thread waiting for plock. Called only |
1287 |
* when CAS of seq value for unlock fails. |
1288 |
*/ |
1289 |
private void releasePlock(int ps) { |
1290 |
plock = ps; |
1291 |
synchronized (this) { notifyAll(); } |
1292 |
} |
1293 |
|
1294 |
/** |
1295 |
* Tries to create and start one worker if fewer than target |
1296 |
* parallelism level exist. Adjusts counts etc on failure. |
1297 |
*/ |
1298 |
private void tryAddWorker() { |
1299 |
long c; int u, e; |
1300 |
while ((u = (int)((c = ctl) >>> 32)) < 0 && |
1301 |
(u & SHORT_SIGN) != 0 && (e = (int)c) >= 0) { |
1302 |
long nc = ((long)(((u + UTC_UNIT) & UTC_MASK) | |
1303 |
((u + UAC_UNIT) & UAC_MASK)) << 32) | (long)e; |
1304 |
if (U.compareAndSwapLong(this, CTL, c, nc)) { |
1305 |
ForkJoinWorkerThreadFactory fac; |
1306 |
Throwable ex = null; |
1307 |
ForkJoinWorkerThread wt = null; |
1308 |
try { |
1309 |
if ((fac = factory) != null && |
1310 |
(wt = fac.newThread(this)) != null) { |
1311 |
wt.start(); |
1312 |
break; |
1313 |
} |
1314 |
} catch (Throwable rex) { |
1315 |
ex = rex; |
1316 |
} |
1317 |
deregisterWorker(wt, ex); |
1318 |
break; |
1319 |
} |
1320 |
} |
1321 |
} |
1322 |
|
1323 |
// Registering and deregistering workers |
1324 |
|
1325 |
/** |
1326 |
* Callback from ForkJoinWorkerThread to establish and record its |
1327 |
* WorkQueue. To avoid scanning bias due to packing entries in |
1328 |
* front of the workQueues array, we treat the array as a simple |
1329 |
* power-of-two hash table using per-thread seed as hash, |
1330 |
* expanding as needed. |
1331 |
* |
1332 |
* @param wt the worker thread |
1333 |
* @return the worker's queue |
1334 |
*/ |
1335 |
final WorkQueue registerWorker(ForkJoinWorkerThread wt) { |
1336 |
UncaughtExceptionHandler handler; WorkQueue[] ws; int s, ps; |
1337 |
wt.setDaemon(true); |
1338 |
if ((handler = ueh) != null) |
1339 |
wt.setUncaughtExceptionHandler(handler); |
1340 |
do {} while (!U.compareAndSwapInt(this, INDEXSEED, s = indexSeed, |
1341 |
s += SEED_INCREMENT) || |
1342 |
s == 0); // skip 0 |
1343 |
WorkQueue w = new WorkQueue(this, wt, mode, s); |
1344 |
if (((ps = plock) & PL_LOCK) != 0 || |
1345 |
!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK)) |
1346 |
ps = acquirePlock(); |
1347 |
int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN); |
1348 |
try { |
1349 |
if ((ws = workQueues) != null) { // skip if shutting down |
1350 |
int n = ws.length, m = n - 1; |
1351 |
int r = (s << 1) | 1; // use odd-numbered indices |
1352 |
if (ws[r &= m] != null) { // collision |
1353 |
int probes = 0; // step by approx half size |
1354 |
int step = (n <= 4) ? 2 : ((n >>> 1) & EVENMASK) + 2; |
1355 |
while (ws[r = (r + step) & m] != null) { |
1356 |
if (++probes >= n) { |
1357 |
workQueues = ws = Arrays.copyOf(ws, n <<= 1); |
1358 |
m = n - 1; |
1359 |
probes = 0; |
1360 |
} |
1361 |
} |
1362 |
} |
1363 |
w.poolIndex = (short)r; |
1364 |
w.eventCount = r; // volatile write orders |
1365 |
ws[r] = w; |
1366 |
} |
1367 |
} finally { |
1368 |
if (!U.compareAndSwapInt(this, PLOCK, ps, nps)) |
1369 |
releasePlock(nps); |
1370 |
} |
1371 |
wt.setName(workerNamePrefix.concat(Integer.toString(w.poolIndex >>> 1))); |
1372 |
return w; |
1373 |
} |
1374 |
|
1375 |
/** |
1376 |
* Final callback from terminating worker, as well as upon failure |
1377 |
* to construct or start a worker. Removes record of worker from |
1378 |
* array, and adjusts counts. If pool is shutting down, tries to |
1379 |
* complete termination. |
1380 |
* |
1381 |
* @param wt the worker thread, or null if construction failed |
1382 |
* @param ex the exception causing failure, or null if none |
1383 |
*/ |
1384 |
final void deregisterWorker(ForkJoinWorkerThread wt, Throwable ex) { |
1385 |
WorkQueue w = null; |
1386 |
if (wt != null && (w = wt.workQueue) != null) { |
1387 |
int ps; long sc; |
1388 |
w.qlock = -1; // ensure set |
1389 |
do {} while (!U.compareAndSwapLong(this, STEALCOUNT, |
1390 |
sc = stealCount, |
1391 |
sc + w.nsteals)); |
1392 |
if (((ps = plock) & PL_LOCK) != 0 || |
1393 |
!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK)) |
1394 |
ps = acquirePlock(); |
1395 |
int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN); |
1396 |
try { |
1397 |
int idx = w.poolIndex; |
1398 |
WorkQueue[] ws = workQueues; |
1399 |
if (ws != null && idx >= 0 && idx < ws.length && ws[idx] == w) |
1400 |
ws[idx] = null; |
1401 |
} finally { |
1402 |
if (!U.compareAndSwapInt(this, PLOCK, ps, nps)) |
1403 |
releasePlock(nps); |
1404 |
} |
1405 |
} |
1406 |
|
1407 |
long c; // adjust ctl counts |
1408 |
do {} while (!U.compareAndSwapLong |
1409 |
(this, CTL, c = ctl, (((c - AC_UNIT) & AC_MASK) | |
1410 |
((c - TC_UNIT) & TC_MASK) | |
1411 |
(c & ~(AC_MASK|TC_MASK))))); |
1412 |
|
1413 |
if (!tryTerminate(false, false) && w != null && w.array != null) { |
1414 |
w.cancelAll(); // cancel remaining tasks |
1415 |
WorkQueue[] ws; WorkQueue v; Thread p; int u, i, e; |
1416 |
while ((u = (int)((c = ctl) >>> 32)) < 0 && (e = (int)c) >= 0) { |
1417 |
if (e > 0) { // activate or create replacement |
1418 |
if ((ws = workQueues) == null || |
1419 |
(i = e & SMASK) >= ws.length || |
1420 |
(v = ws[i]) == null) |
1421 |
break; |
1422 |
long nc = (((long)(v.nextWait & E_MASK)) | |
1423 |
((long)(u + UAC_UNIT) << 32)); |
1424 |
if (v.eventCount != (e | INT_SIGN)) |
1425 |
break; |
1426 |
if (U.compareAndSwapLong(this, CTL, c, nc)) { |
1427 |
v.eventCount = (e + E_SEQ) & E_MASK; |
1428 |
if ((p = v.parker) != null) |
1429 |
U.unpark(p); |
1430 |
break; |
1431 |
} |
1432 |
} |
1433 |
else { |
1434 |
if ((short)u < 0) |
1435 |
tryAddWorker(); |
1436 |
break; |
1437 |
} |
1438 |
} |
1439 |
} |
1440 |
if (ex == null) // help clean refs on way out |
1441 |
ForkJoinTask.helpExpungeStaleExceptions(); |
1442 |
else // rethrow |
1443 |
ForkJoinTask.rethrow(ex); |
1444 |
} |
1445 |
|
1446 |
// Submissions |
1447 |
|
1448 |
/** |
1449 |
* Per-thread records for threads that submit to pools. Currently |
1450 |
* holds only pseudo-random seed / index that is used to choose |
1451 |
* submission queues in method externalPush. In the future, this may |
1452 |
* also incorporate a means to implement different task rejection |
1453 |
* and resubmission policies. |
1454 |
* |
1455 |
* Seeds for submitters and workers/workQueues work in basically |
1456 |
* the same way but are initialized and updated using slightly |
1457 |
* different mechanics. Both are initialized using the same |
1458 |
* approach as in class ThreadLocal, where successive values are |
1459 |
* unlikely to collide with previous values. Seeds are then |
1460 |
* randomly modified upon collisions using xorshifts, which |
1461 |
* requires a non-zero seed. |
1462 |
*/ |
1463 |
static final class Submitter { |
1464 |
int seed; |
1465 |
Submitter(int s) { seed = s; } |
1466 |
} |
1467 |
|
1468 |
/** |
1469 |
* Unless shutting down, adds the given task to a submission queue |
1470 |
* at submitter's current queue index (modulo submission |
1471 |
* range). Only the most common path is directly handled in this |
1472 |
* method. All others are relayed to fullExternalPush. |
1473 |
* |
1474 |
* @param task the task. Caller must ensure non-null. |
1475 |
*/ |
1476 |
final void externalPush(ForkJoinTask<?> task) { |
1477 |
Submitter z = submitters.get(); |
1478 |
WorkQueue q; int r, m, s, n, am; ForkJoinTask<?>[] a; |
1479 |
int ps = plock; |
1480 |
WorkQueue[] ws = workQueues; |
1481 |
if (z != null && ps > 0 && ws != null && (m = (ws.length - 1)) >= 0 && |
1482 |
(q = ws[m & (r = z.seed) & SQMASK]) != null && r != 0 && |
1483 |
U.compareAndSwapInt(q, QLOCK, 0, 1)) { // lock |
1484 |
if ((a = q.array) != null && |
1485 |
(am = a.length - 1) > (n = (s = q.top) - q.base)) { |
1486 |
int j = ((am & s) << ASHIFT) + ABASE; |
1487 |
U.putOrderedObject(a, j, task); |
1488 |
q.top = s + 1; // push on to deque |
1489 |
q.qlock = 0; |
1490 |
if (n <= 1) |
1491 |
signalWork(ws, q); |
1492 |
return; |
1493 |
} |
1494 |
q.qlock = 0; |
1495 |
} |
1496 |
fullExternalPush(task); |
1497 |
} |
1498 |
|
1499 |
/** |
1500 |
* Full version of externalPush. This method is called, among |
1501 |
* other times, upon the first submission of the first task to the |
1502 |
* pool, so must perform secondary initialization. It also |
1503 |
* detects first submission by an external thread by looking up |
1504 |
* its ThreadLocal, and creates a new shared queue if the one at |
1505 |
* index if empty or contended. The plock lock body must be |
1506 |
* exception-free (so no try/finally) so we optimistically |
1507 |
* allocate new queues outside the lock and throw them away if |
1508 |
* (very rarely) not needed. |
1509 |
* |
1510 |
* Secondary initialization occurs when plock is zero, to create |
1511 |
* workQueue array and set plock to a valid value. This lock body |
1512 |
* must also be exception-free. Because the plock seq value can |
1513 |
* eventually wrap around zero, this method harmlessly fails to |
1514 |
* reinitialize if workQueues exists, while still advancing plock. |
1515 |
*/ |
1516 |
private void fullExternalPush(ForkJoinTask<?> task) { |
1517 |
int r = 0; // random index seed |
1518 |
for (Submitter z = submitters.get();;) { |
1519 |
WorkQueue[] ws; WorkQueue q; int ps, m, k; |
1520 |
if (z == null) { |
1521 |
if (U.compareAndSwapInt(this, INDEXSEED, r = indexSeed, |
1522 |
r += SEED_INCREMENT) && r != 0) |
1523 |
submitters.set(z = new Submitter(r)); |
1524 |
} |
1525 |
else if (r == 0) { // move to a different index |
1526 |
r = z.seed; |
1527 |
r ^= r << 13; // same xorshift as WorkQueues |
1528 |
r ^= r >>> 17; |
1529 |
z.seed = r ^= (r << 5); |
1530 |
} |
1531 |
if ((ps = plock) < 0) |
1532 |
throw new RejectedExecutionException(); |
1533 |
else if (ps == 0 || (ws = workQueues) == null || |
1534 |
(m = ws.length - 1) < 0) { // initialize workQueues |
1535 |
int p = parallelism; // find power of two table size |
1536 |
int n = (p > 1) ? p - 1 : 1; // ensure at least 2 slots |
1537 |
n |= n >>> 1; n |= n >>> 2; n |= n >>> 4; |
1538 |
n |= n >>> 8; n |= n >>> 16; n = (n + 1) << 1; |
1539 |
WorkQueue[] nws = ((ws = workQueues) == null || ws.length == 0 ? |
1540 |
new WorkQueue[n] : null); |
1541 |
if (((ps = plock) & PL_LOCK) != 0 || |
1542 |
!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK)) |
1543 |
ps = acquirePlock(); |
1544 |
if (((ws = workQueues) == null || ws.length == 0) && nws != null) |
1545 |
workQueues = nws; |
1546 |
int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN); |
1547 |
if (!U.compareAndSwapInt(this, PLOCK, ps, nps)) |
1548 |
releasePlock(nps); |
1549 |
} |
1550 |
else if ((q = ws[k = r & m & SQMASK]) != null) { |
1551 |
if (q.qlock == 0 && U.compareAndSwapInt(q, QLOCK, 0, 1)) { |
1552 |
ForkJoinTask<?>[] a = q.array; |
1553 |
int s = q.top; |
1554 |
boolean submitted = false; |
1555 |
try { // locked version of push |
1556 |
if ((a != null && a.length > s + 1 - q.base) || |
1557 |
(a = q.growArray()) != null) { // must presize |
1558 |
int j = (((a.length - 1) & s) << ASHIFT) + ABASE; |
1559 |
U.putOrderedObject(a, j, task); |
1560 |
q.top = s + 1; |
1561 |
submitted = true; |
1562 |
} |
1563 |
} finally { |
1564 |
q.qlock = 0; // unlock |
1565 |
} |
1566 |
if (submitted) { |
1567 |
signalWork(ws, q); |
1568 |
return; |
1569 |
} |
1570 |
} |
1571 |
r = 0; // move on failure |
1572 |
} |
1573 |
else if (((ps = plock) & PL_LOCK) == 0) { // create new queue |
1574 |
q = new WorkQueue(this, null, SHARED_QUEUE, r); |
1575 |
q.poolIndex = (short)k; |
1576 |
if (((ps = plock) & PL_LOCK) != 0 || |
1577 |
!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK)) |
1578 |
ps = acquirePlock(); |
1579 |
if ((ws = workQueues) != null && k < ws.length && ws[k] == null) |
1580 |
ws[k] = q; |
1581 |
int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN); |
1582 |
if (!U.compareAndSwapInt(this, PLOCK, ps, nps)) |
1583 |
releasePlock(nps); |
1584 |
} |
1585 |
else |
1586 |
r = 0; |
1587 |
} |
1588 |
} |
1589 |
|
1590 |
// Maintaining ctl counts |
1591 |
|
1592 |
/** |
1593 |
* Increments active count; mainly called upon return from blocking. |
1594 |
*/ |
1595 |
final void incrementActiveCount() { |
1596 |
long c; |
1597 |
do {} while (!U.compareAndSwapLong |
1598 |
(this, CTL, c = ctl, ((c & ~AC_MASK) | |
1599 |
((c & AC_MASK) + AC_UNIT)))); |
1600 |
} |
1601 |
|
1602 |
/** |
1603 |
* Tries to create or activate a worker if too few are active. |
1604 |
* |
1605 |
* @param ws the worker array to use to find signallees |
1606 |
* @param q if non-null, the queue holding tasks to be processed |
1607 |
*/ |
1608 |
final void signalWork(WorkQueue[] ws, WorkQueue q) { |
1609 |
for (;;) { |
1610 |
long c; int e, u, i; WorkQueue w; Thread p; |
1611 |
if ((u = (int)((c = ctl) >>> 32)) >= 0) |
1612 |
break; |
1613 |
if ((e = (int)c) <= 0) { |
1614 |
if ((short)u < 0) |
1615 |
tryAddWorker(); |
1616 |
break; |
1617 |
} |
1618 |
if (ws == null || ws.length <= (i = e & SMASK) || |
1619 |
(w = ws[i]) == null) |
1620 |
break; |
1621 |
long nc = (((long)(w.nextWait & E_MASK)) | |
1622 |
((long)(u + UAC_UNIT)) << 32); |
1623 |
int ne = (e + E_SEQ) & E_MASK; |
1624 |
if (w.eventCount == (e | INT_SIGN) && |
1625 |
U.compareAndSwapLong(this, CTL, c, nc)) { |
1626 |
w.eventCount = ne; |
1627 |
if ((p = w.parker) != null) |
1628 |
U.unpark(p); |
1629 |
break; |
1630 |
} |
1631 |
if (q != null && q.base >= q.top) |
1632 |
break; |
1633 |
} |
1634 |
} |
1635 |
|
1636 |
// Scanning for tasks |
1637 |
|
1638 |
/** |
1639 |
* Top-level runloop for workers, called by ForkJoinWorkerThread.run. |
1640 |
*/ |
1641 |
final void runWorker(WorkQueue w) { |
1642 |
w.growArray(); // allocate queue |
1643 |
for (int r = w.hint; scan(w, r) == 0; ) { |
1644 |
r ^= r << 13; r ^= r >>> 17; r ^= r << 5; // xorshift |
1645 |
} |
1646 |
} |
1647 |
|
1648 |
/** |
1649 |
* Scans for and, if found, runs one task, else possibly |
1650 |
* inactivates the worker. This method operates on single reads of |
1651 |
* volatile state and is designed to be re-invoked continuously, |
1652 |
* in part because it returns upon detecting inconsistencies, |
1653 |
* contention, or state changes that indicate possible success on |
1654 |
* re-invocation. |
1655 |
* |
1656 |
* The scan searches for tasks across queues starting at a random |
1657 |
* index, checking each at least twice. The scan terminates upon |
1658 |
* either finding a non-empty queue, or completing the sweep. If |
1659 |
* the worker is not inactivated, it takes and runs a task from |
1660 |
* this queue. Otherwise, if not activated, it tries to activate |
1661 |
* itself or some other worker by signalling. On failure to find a |
1662 |
* task, returns (for retry) if pool state may have changed during |
1663 |
* an empty scan, or tries to inactivate if active, else possibly |
1664 |
* blocks or terminates via method awaitWork. |
1665 |
* |
1666 |
* @param w the worker (via its WorkQueue) |
1667 |
* @param r a random seed |
1668 |
* @return worker qlock status if would have waited, else 0 |
1669 |
*/ |
1670 |
private final int scan(WorkQueue w, int r) { |
1671 |
WorkQueue[] ws; int m; |
1672 |
long c = ctl; // for consistency check |
1673 |
if ((ws = workQueues) != null && (m = ws.length - 1) >= 0 && w != null) { |
1674 |
for (int j = m + m + 1, ec = w.eventCount;;) { |
1675 |
WorkQueue q; int b, e; ForkJoinTask<?>[] a; ForkJoinTask<?> t; |
1676 |
if ((q = ws[(r - j) & m]) != null && |
1677 |
(b = q.base) - q.top < 0 && (a = q.array) != null) { |
1678 |
long i = (((a.length - 1) & b) << ASHIFT) + ABASE; |
1679 |
if ((t = ((ForkJoinTask<?>) |
1680 |
U.getObjectVolatile(a, i))) != null) { |
1681 |
if (ec < 0) |
1682 |
helpRelease(c, ws, w, q, b); |
1683 |
else if (q.base == b && |
1684 |
U.compareAndSwapObject(a, i, t, null)) { |
1685 |
U.putOrderedInt(q, QBASE, b + 1); |
1686 |
if ((b + 1) - q.top < 0) |
1687 |
signalWork(ws, q); |
1688 |
w.runTask(t); |
1689 |
} |
1690 |
} |
1691 |
break; |
1692 |
} |
1693 |
else if (--j < 0) { |
1694 |
if ((ec | (e = (int)c)) < 0) // inactive or terminating |
1695 |
return awaitWork(w, c, ec); |
1696 |
else if (ctl == c) { // try to inactivate and enqueue |
1697 |
long nc = (long)ec | ((c - AC_UNIT) & (AC_MASK|TC_MASK)); |
1698 |
w.nextWait = e; |
1699 |
w.eventCount = ec | INT_SIGN; |
1700 |
if (!U.compareAndSwapLong(this, CTL, c, nc)) |
1701 |
w.eventCount = ec; // back out |
1702 |
} |
1703 |
break; |
1704 |
} |
1705 |
} |
1706 |
} |
1707 |
return 0; |
1708 |
} |
1709 |
|
1710 |
/** |
1711 |
* A continuation of scan(), possibly blocking or terminating |
1712 |
* worker w. Returns without blocking if pool state has apparently |
1713 |
* changed since last invocation. Also, if inactivating w has |
1714 |
* caused the pool to become quiescent, checks for pool |
1715 |
* termination, and, so long as this is not the only worker, waits |
1716 |
* for event for up to a given duration. On timeout, if ctl has |
1717 |
* not changed, terminates the worker, which will in turn wake up |
1718 |
* another worker to possibly repeat this process. |
1719 |
* |
1720 |
* @param w the calling worker |
1721 |
* @param c the ctl value on entry to scan |
1722 |
* @param ec the worker's eventCount on entry to scan |
1723 |
*/ |
1724 |
private final int awaitWork(WorkQueue w, long c, int ec) { |
1725 |
int stat, ns; long parkTime, deadline; |
1726 |
if ((stat = w.qlock) >= 0 && w.eventCount == ec && ctl == c && |
1727 |
!Thread.interrupted()) { |
1728 |
int e = (int)c; |
1729 |
int u = (int)(c >>> 32); |
1730 |
int d = (u >> UAC_SHIFT) + parallelism; // active count |
1731 |
|
1732 |
if (e < 0 || (d <= 0 && tryTerminate(false, false))) |
1733 |
stat = w.qlock = -1; // pool is terminating |
1734 |
else if ((ns = w.nsteals) != 0) { // collect steals and retry |
1735 |
long sc; |
1736 |
w.nsteals = 0; |
1737 |
do {} while (!U.compareAndSwapLong(this, STEALCOUNT, |
1738 |
sc = stealCount, sc + ns)); |
1739 |
} |
1740 |
else { |
1741 |
long pc = ((d > 0 || ec != (e | INT_SIGN)) ? 0L : |
1742 |
((long)(w.nextWait & E_MASK)) | // ctl to restore |
1743 |
((long)(u + UAC_UNIT)) << 32); |
1744 |
if (pc != 0L) { // timed wait if last waiter |
1745 |
int dc = -(short)(c >>> TC_SHIFT); |
1746 |
parkTime = (dc < 0 ? FAST_IDLE_TIMEOUT: |
1747 |
(dc + 1) * IDLE_TIMEOUT); |
1748 |
deadline = System.nanoTime() + parkTime - TIMEOUT_SLOP; |
1749 |
} |
1750 |
else |
1751 |
parkTime = deadline = 0L; |
1752 |
if (w.eventCount == ec && ctl == c) { |
1753 |
Thread wt = Thread.currentThread(); |
1754 |
U.putObject(wt, PARKBLOCKER, this); |
1755 |
w.parker = wt; // emulate LockSupport.park |
1756 |
if (w.eventCount == ec && ctl == c) |
1757 |
U.park(false, parkTime); // must recheck before park |
1758 |
w.parker = null; |
1759 |
U.putObject(wt, PARKBLOCKER, null); |
1760 |
if (parkTime != 0L && ctl == c && |
1761 |
deadline - System.nanoTime() <= 0L && |
1762 |
U.compareAndSwapLong(this, CTL, c, pc)) |
1763 |
stat = w.qlock = -1; // shrink pool |
1764 |
} |
1765 |
} |
1766 |
} |
1767 |
return stat; |
1768 |
} |
1769 |
|
1770 |
/** |
1771 |
* Possibly releases (signals) a worker. Called only from scan() |
1772 |
* when a worker with apparently inactive status finds a non-empty |
1773 |
* queue. This requires revalidating all of the associated state |
1774 |
* from caller. |
1775 |
*/ |
1776 |
private final void helpRelease(long c, WorkQueue[] ws, WorkQueue w, |
1777 |
WorkQueue q, int b) { |
1778 |
WorkQueue v; int e, i; Thread p; |
1779 |
if (w != null && w.eventCount < 0 && (e = (int)c) > 0 && |
1780 |
ws != null && ws.length > (i = e & SMASK) && |
1781 |
(v = ws[i]) != null && ctl == c) { |
1782 |
long nc = (((long)(v.nextWait & E_MASK)) | |
1783 |
((long)((int)(c >>> 32) + UAC_UNIT)) << 32); |
1784 |
int ne = (e + E_SEQ) & E_MASK; |
1785 |
if (q != null && q.base == b && w.eventCount < 0 && |
1786 |
v.eventCount == (e | INT_SIGN) && |
1787 |
U.compareAndSwapLong(this, CTL, c, nc)) { |
1788 |
v.eventCount = ne; |
1789 |
if ((p = v.parker) != null) |
1790 |
U.unpark(p); |
1791 |
} |
1792 |
} |
1793 |
} |
1794 |
|
1795 |
/** |
1796 |
* Tries to locate and execute tasks for a stealer of the given |
1797 |
* task, or in turn one of its stealers, Traces currentSteal -> |
1798 |
* currentJoin links looking for a thread working on a descendant |
1799 |
* of the given task and with a non-empty queue to steal back and |
1800 |
* execute tasks from. The first call to this method upon a |
1801 |
* waiting join will often entail scanning/search, (which is OK |
1802 |
* because the joiner has nothing better to do), but this method |
1803 |
* leaves hints in workers to speed up subsequent calls. The |
1804 |
* implementation is very branchy to cope with potential |
1805 |
* inconsistencies or loops encountering chains that are stale, |
1806 |
* unknown, or so long that they are likely cyclic. |
1807 |
* |
1808 |
* @param joiner the joining worker |
1809 |
* @param task the task to join |
1810 |
* @return 0 if no progress can be made, negative if task |
1811 |
* known complete, else positive |
1812 |
*/ |
1813 |
private int tryHelpStealer(WorkQueue joiner, ForkJoinTask<?> task) { |
1814 |
int stat = 0, steps = 0; // bound to avoid cycles |
1815 |
if (task != null && joiner != null && |
1816 |
joiner.base - joiner.top >= 0) { // hoist checks |
1817 |
restart: for (;;) { |
1818 |
ForkJoinTask<?> subtask = task; // current target |
1819 |
for (WorkQueue j = joiner, v;;) { // v is stealer of subtask |
1820 |
WorkQueue[] ws; int m, s, h; |
1821 |
if ((s = task.status) < 0) { |
1822 |
stat = s; |
1823 |
break restart; |
1824 |
} |
1825 |
if ((ws = workQueues) == null || (m = ws.length - 1) <= 0) |
1826 |
break restart; // shutting down |
1827 |
if ((v = ws[h = (j.hint | 1) & m]) == null || |
1828 |
v.currentSteal != subtask) { |
1829 |
for (int origin = h;;) { // find stealer |
1830 |
if (((h = (h + 2) & m) & 15) == 1 && |
1831 |
(subtask.status < 0 || j.currentJoin != subtask)) |
1832 |
continue restart; // occasional staleness check |
1833 |
if ((v = ws[h]) != null && |
1834 |
v.currentSteal == subtask) { |
1835 |
j.hint = h; // save hint |
1836 |
break; |
1837 |
} |
1838 |
if (h == origin) |
1839 |
break restart; // cannot find stealer |
1840 |
} |
1841 |
} |
1842 |
for (;;) { // help stealer or descend to its stealer |
1843 |
ForkJoinTask[] a; int b; |
1844 |
if (subtask.status < 0) // surround probes with |
1845 |
continue restart; // consistency checks |
1846 |
if ((b = v.base) - v.top < 0 && (a = v.array) != null) { |
1847 |
int i = (((a.length - 1) & b) << ASHIFT) + ABASE; |
1848 |
ForkJoinTask<?> t = |
1849 |
(ForkJoinTask<?>)U.getObjectVolatile(a, i); |
1850 |
if (subtask.status < 0 || j.currentJoin != subtask || |
1851 |
v.currentSteal != subtask) |
1852 |
continue restart; // stale |
1853 |
stat = 1; // apparent progress |
1854 |
if (v.base == b) { |
1855 |
if (t == null) |
1856 |
break restart; |
1857 |
if (U.compareAndSwapObject(a, i, t, null)) { |
1858 |
U.putOrderedInt(v, QBASE, b + 1); |
1859 |
ForkJoinTask<?> ps = joiner.currentSteal; |
1860 |
int jt = joiner.top; |
1861 |
do { |
1862 |
joiner.currentSteal = t; |
1863 |
t.doExec(); // clear local tasks too |
1864 |
} while (task.status >= 0 && |
1865 |
joiner.top != jt && |
1866 |
(t = joiner.pop()) != null); |
1867 |
joiner.currentSteal = ps; |
1868 |
break restart; |
1869 |
} |
1870 |
} |
1871 |
} |
1872 |
else { // empty -- try to descend |
1873 |
ForkJoinTask<?> next = v.currentJoin; |
1874 |
if (subtask.status < 0 || j.currentJoin != subtask || |
1875 |
v.currentSteal != subtask) |
1876 |
continue restart; // stale |
1877 |
else if (next == null || ++steps == MAX_HELP) |
1878 |
break restart; // dead-end or maybe cyclic |
1879 |
else { |
1880 |
subtask = next; |
1881 |
j = v; |
1882 |
break; |
1883 |
} |
1884 |
} |
1885 |
} |
1886 |
} |
1887 |
} |
1888 |
} |
1889 |
return stat; |
1890 |
} |
1891 |
|
1892 |
/** |
1893 |
* Analog of tryHelpStealer for CountedCompleters. Tries to steal |
1894 |
* and run tasks within the target's computation. |
1895 |
* |
1896 |
* @param task the task to join |
1897 |
*/ |
1898 |
private int helpComplete(WorkQueue joiner, CountedCompleter<?> task) { |
1899 |
WorkQueue[] ws; int m; |
1900 |
int s = 0; |
1901 |
if ((ws = workQueues) != null && (m = ws.length - 1) >= 0 && |
1902 |
joiner != null && task != null) { |
1903 |
int j = joiner.poolIndex; |
1904 |
int scans = m + m + 1; |
1905 |
long c = 0L; // for stability check |
1906 |
for (int k = scans; ; j += 2) { |
1907 |
WorkQueue q; |
1908 |
if ((s = task.status) < 0) |
1909 |
break; |
1910 |
else if (joiner.internalPopAndExecCC(task)) |
1911 |
k = scans; |
1912 |
else if ((s = task.status) < 0) |
1913 |
break; |
1914 |
else if ((q = ws[j & m]) != null && q.pollAndExecCC(task)) |
1915 |
k = scans; |
1916 |
else if (--k < 0) { |
1917 |
if (c == (c = ctl)) |
1918 |
break; |
1919 |
k = scans; |
1920 |
} |
1921 |
} |
1922 |
} |
1923 |
return s; |
1924 |
} |
1925 |
|
1926 |
/** |
1927 |
* Tries to decrement active count (sometimes implicitly) and |
1928 |
* possibly release or create a compensating worker in preparation |
1929 |
* for blocking. Fails on contention or termination. Otherwise, |
1930 |
* adds a new thread if no idle workers are available and pool |
1931 |
* may become starved. |
1932 |
* |
1933 |
* @param c the assumed ctl value |
1934 |
*/ |
1935 |
final boolean tryCompensate(long c) { |
1936 |
WorkQueue[] ws = workQueues; |
1937 |
int pc = parallelism, e = (int)c, m, tc; |
1938 |
if (ws != null && (m = ws.length - 1) >= 0 && e >= 0 && ctl == c) { |
1939 |
WorkQueue w = ws[e & m]; |
1940 |
if (e != 0 && w != null) { |
1941 |
Thread p; |
1942 |
long nc = ((long)(w.nextWait & E_MASK) | |
1943 |
(c & (AC_MASK|TC_MASK))); |
1944 |
int ne = (e + E_SEQ) & E_MASK; |
1945 |
if (w.eventCount == (e | INT_SIGN) && |
1946 |
U.compareAndSwapLong(this, CTL, c, nc)) { |
1947 |
w.eventCount = ne; |
1948 |
if ((p = w.parker) != null) |
1949 |
U.unpark(p); |
1950 |
return true; // replace with idle worker |
1951 |
} |
1952 |
} |
1953 |
else if ((tc = (short)(c >>> TC_SHIFT)) >= 0 && |
1954 |
(int)(c >> AC_SHIFT) + pc > 1) { |
1955 |
long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK); |
1956 |
if (U.compareAndSwapLong(this, CTL, c, nc)) |
1957 |
return true; // no compensation |
1958 |
} |
1959 |
else if (tc + pc < MAX_CAP) { |
1960 |
long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK); |
1961 |
if (U.compareAndSwapLong(this, CTL, c, nc)) { |
1962 |
ForkJoinWorkerThreadFactory fac; |
1963 |
Throwable ex = null; |
1964 |
ForkJoinWorkerThread wt = null; |
1965 |
try { |
1966 |
if ((fac = factory) != null && |
1967 |
(wt = fac.newThread(this)) != null) { |
1968 |
wt.start(); |
1969 |
return true; |
1970 |
} |
1971 |
} catch (Throwable rex) { |
1972 |
ex = rex; |
1973 |
} |
1974 |
deregisterWorker(wt, ex); // clean up and return false |
1975 |
} |
1976 |
} |
1977 |
} |
1978 |
return false; |
1979 |
} |
1980 |
|
1981 |
/** |
1982 |
* Helps and/or blocks until the given task is done. |
1983 |
* |
1984 |
* @param joiner the joining worker |
1985 |
* @param task the task |
1986 |
* @return task status on exit |
1987 |
*/ |
1988 |
final int awaitJoin(WorkQueue joiner, ForkJoinTask<?> task) { |
1989 |
int s = 0; |
1990 |
if (task != null && (s = task.status) >= 0 && joiner != null) { |
1991 |
ForkJoinTask<?> prevJoin = joiner.currentJoin; |
1992 |
joiner.currentJoin = task; |
1993 |
do {} while (joiner.tryRemoveAndExec(task) && // process local tasks |
1994 |
(s = task.status) >= 0); |
1995 |
if (s >= 0 && (task instanceof CountedCompleter)) |
1996 |
s = helpComplete(joiner, (CountedCompleter<?>)task); |
1997 |
long cc = 0; // for stability checks |
1998 |
while (s >= 0 && (s = task.status) >= 0) { |
1999 |
if ((s = tryHelpStealer(joiner, task)) == 0 && |
2000 |
(s = task.status) >= 0) { |
2001 |
if (!tryCompensate(cc)) |
2002 |
cc = ctl; |
2003 |
else { |
2004 |
if (task.trySetSignal() && (s = task.status) >= 0) { |
2005 |
synchronized (task) { |
2006 |
if (task.status >= 0) { |
2007 |
try { // see ForkJoinTask |
2008 |
task.wait(); // for explanation |
2009 |
} catch (InterruptedException ie) { |
2010 |
} |
2011 |
} |
2012 |
else |
2013 |
task.notifyAll(); |
2014 |
} |
2015 |
} |
2016 |
long c; // reactivate |
2017 |
do {} while (!U.compareAndSwapLong |
2018 |
(this, CTL, c = ctl, |
2019 |
((c & ~AC_MASK) | |
2020 |
((c & AC_MASK) + AC_UNIT)))); |
2021 |
} |
2022 |
} |
2023 |
} |
2024 |
joiner.currentJoin = prevJoin; |
2025 |
} |
2026 |
return s; |
2027 |
} |
2028 |
|
2029 |
/** |
2030 |
* Stripped-down variant of awaitJoin used by timed joins. Tries |
2031 |
* to help join only while there is continuous progress. (Caller |
2032 |
* will then enter a timed wait.) |
2033 |
* |
2034 |
* @param joiner the joining worker |
2035 |
* @param task the task |
2036 |
*/ |
2037 |
final void helpJoinOnce(WorkQueue joiner, ForkJoinTask<?> task) { |
2038 |
int s; |
2039 |
if (joiner != null && task != null && (s = task.status) >= 0) { |
2040 |
ForkJoinTask<?> prevJoin = joiner.currentJoin; |
2041 |
joiner.currentJoin = task; |
2042 |
do {} while (joiner.tryRemoveAndExec(task) && // process local tasks |
2043 |
(s = task.status) >= 0); |
2044 |
if (s >= 0) { |
2045 |
if (task instanceof CountedCompleter) |
2046 |
helpComplete(joiner, (CountedCompleter<?>)task); |
2047 |
do {} while (task.status >= 0 && |
2048 |
tryHelpStealer(joiner, task) > 0); |
2049 |
} |
2050 |
joiner.currentJoin = prevJoin; |
2051 |
} |
2052 |
} |
2053 |
|
2054 |
/** |
2055 |
* Returns a (probably) non-empty steal queue, if one is found |
2056 |
* during a scan, else null. This method must be retried by |
2057 |
* caller if, by the time it tries to use the queue, it is empty. |
2058 |
*/ |
2059 |
private WorkQueue findNonEmptyStealQueue() { |
2060 |
int r = ThreadLocalRandom.current().nextInt(); |
2061 |
for (;;) { |
2062 |
int ps = plock, m; WorkQueue[] ws; WorkQueue q; |
2063 |
if ((ws = workQueues) != null && (m = ws.length - 1) >= 0) { |
2064 |
for (int j = (m + 1) << 2; j >= 0; --j) { |
2065 |
if ((q = ws[(((r - j) << 1) | 1) & m]) != null && |
2066 |
q.base - q.top < 0) |
2067 |
return q; |
2068 |
} |
2069 |
} |
2070 |
if (plock == ps) |
2071 |
return null; |
2072 |
} |
2073 |
} |
2074 |
|
2075 |
/** |
2076 |
* Runs tasks until {@code isQuiescent()}. We piggyback on |
2077 |
* active count ctl maintenance, but rather than blocking |
2078 |
* when tasks cannot be found, we rescan until all others cannot |
2079 |
* find tasks either. |
2080 |
*/ |
2081 |
final void helpQuiescePool(WorkQueue w) { |
2082 |
ForkJoinTask<?> ps = w.currentSteal; |
2083 |
for (boolean active = true;;) { |
2084 |
long c; WorkQueue q; ForkJoinTask<?> t; int b; |
2085 |
while ((t = w.nextLocalTask()) != null) |
2086 |
t.doExec(); |
2087 |
if ((q = findNonEmptyStealQueue()) != null) { |
2088 |
if (!active) { // re-establish active count |
2089 |
active = true; |
2090 |
do {} while (!U.compareAndSwapLong |
2091 |
(this, CTL, c = ctl, |
2092 |
((c & ~AC_MASK) | |
2093 |
((c & AC_MASK) + AC_UNIT)))); |
2094 |
} |
2095 |
if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null) { |
2096 |
(w.currentSteal = t).doExec(); |
2097 |
w.currentSteal = ps; |
2098 |
} |
2099 |
} |
2100 |
else if (active) { // decrement active count without queuing |
2101 |
long nc = ((c = ctl) & ~AC_MASK) | ((c & AC_MASK) - AC_UNIT); |
2102 |
if ((int)(nc >> AC_SHIFT) + parallelism == 0) |
2103 |
break; // bypass decrement-then-increment |
2104 |
if (U.compareAndSwapLong(this, CTL, c, nc)) |
2105 |
active = false; |
2106 |
} |
2107 |
else if ((int)((c = ctl) >> AC_SHIFT) + parallelism <= 0 && |
2108 |
U.compareAndSwapLong |
2109 |
(this, CTL, c, ((c & ~AC_MASK) | |
2110 |
((c & AC_MASK) + AC_UNIT)))) |
2111 |
break; |
2112 |
} |
2113 |
} |
2114 |
|
2115 |
/** |
2116 |
* Gets and removes a local or stolen task for the given worker. |
2117 |
* |
2118 |
* @return a task, if available |
2119 |
*/ |
2120 |
final ForkJoinTask<?> nextTaskFor(WorkQueue w) { |
2121 |
for (ForkJoinTask<?> t;;) { |
2122 |
WorkQueue q; int b; |
2123 |
if ((t = w.nextLocalTask()) != null) |
2124 |
return t; |
2125 |
if ((q = findNonEmptyStealQueue()) == null) |
2126 |
return null; |
2127 |
if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null) |
2128 |
return t; |
2129 |
} |
2130 |
} |
2131 |
|
2132 |
/** |
2133 |
* Returns a cheap heuristic guide for task partitioning when |
2134 |
* programmers, frameworks, tools, or languages have little or no |
2135 |
* idea about task granularity. In essence by offering this |
2136 |
* method, we ask users only about tradeoffs in overhead vs |
2137 |
* expected throughput and its variance, rather than how finely to |
2138 |
* partition tasks. |
2139 |
* |
2140 |
* In a steady state strict (tree-structured) computation, each |
2141 |
* thread makes available for stealing enough tasks for other |
2142 |
* threads to remain active. Inductively, if all threads play by |
2143 |
* the same rules, each thread should make available only a |
2144 |
* constant number of tasks. |
2145 |
* |
2146 |
* The minimum useful constant is just 1. But using a value of 1 |
2147 |
* would require immediate replenishment upon each steal to |
2148 |
* maintain enough tasks, which is infeasible. Further, |
2149 |
* partitionings/granularities of offered tasks should minimize |
2150 |
* steal rates, which in general means that threads nearer the top |
2151 |
* of computation tree should generate more than those nearer the |
2152 |
* bottom. In perfect steady state, each thread is at |
2153 |
* approximately the same level of computation tree. However, |
2154 |
* producing extra tasks amortizes the uncertainty of progress and |
2155 |
* diffusion assumptions. |
2156 |
* |
2157 |
* So, users will want to use values larger (but not much larger) |
2158 |
* than 1 to both smooth over transient shortages and hedge |
2159 |
* against uneven progress; as traded off against the cost of |
2160 |
* extra task overhead. We leave the user to pick a threshold |
2161 |
* value to compare with the results of this call to guide |
2162 |
* decisions, but recommend values such as 3. |
2163 |
* |
2164 |
* When all threads are active, it is on average OK to estimate |
2165 |
* surplus strictly locally. In steady-state, if one thread is |
2166 |
* maintaining say 2 surplus tasks, then so are others. So we can |
2167 |
* just use estimated queue length. However, this strategy alone |
2168 |
* leads to serious mis-estimates in some non-steady-state |
2169 |
* conditions (ramp-up, ramp-down, other stalls). We can detect |
2170 |
* many of these by further considering the number of "idle" |
2171 |
* threads, that are known to have zero queued tasks, so |
2172 |
* compensate by a factor of (#idle/#active) threads. |
2173 |
* |
2174 |
* Note: The approximation of #busy workers as #active workers is |
2175 |
* not very good under current signalling scheme, and should be |
2176 |
* improved. |
2177 |
*/ |
2178 |
static int getSurplusQueuedTaskCount() { |
2179 |
Thread t; ForkJoinWorkerThread wt; ForkJoinPool pool; WorkQueue q; |
2180 |
if (((t = Thread.currentThread()) instanceof ForkJoinWorkerThread)) { |
2181 |
int p = (pool = (wt = (ForkJoinWorkerThread)t).pool).parallelism; |
2182 |
int n = (q = wt.workQueue).top - q.base; |
2183 |
int a = (int)(pool.ctl >> AC_SHIFT) + p; |
2184 |
return n - (a > (p >>>= 1) ? 0 : |
2185 |
a > (p >>>= 1) ? 1 : |
2186 |
a > (p >>>= 1) ? 2 : |
2187 |
a > (p >>>= 1) ? 4 : |
2188 |
8); |
2189 |
} |
2190 |
return 0; |
2191 |
} |
2192 |
|
2193 |
// Termination |
2194 |
|
2195 |
/** |
2196 |
* Possibly initiates and/or completes termination. The caller |
2197 |
* triggering termination runs three passes through workQueues: |
2198 |
* (0) Setting termination status, followed by wakeups of queued |
2199 |
* workers; (1) cancelling all tasks; (2) interrupting lagging |
2200 |
* threads (likely in external tasks, but possibly also blocked in |
2201 |
* joins). Each pass repeats previous steps because of potential |
2202 |
* lagging thread creation. |
2203 |
* |
2204 |
* @param now if true, unconditionally terminate, else only |
2205 |
* if no work and no active workers |
2206 |
* @param enable if true, enable shutdown when next possible |
2207 |
* @return true if now terminating or terminated |
2208 |
*/ |
2209 |
private boolean tryTerminate(boolean now, boolean enable) { |
2210 |
int ps; |
2211 |
if (this == common) // cannot shut down |
2212 |
return false; |
2213 |
if ((ps = plock) >= 0) { // enable by setting plock |
2214 |
if (!enable) |
2215 |
return false; |
2216 |
if ((ps & PL_LOCK) != 0 || |
2217 |
!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK)) |
2218 |
ps = acquirePlock(); |
2219 |
int nps = ((ps + PL_LOCK) & ~SHUTDOWN) | SHUTDOWN; |
2220 |
if (!U.compareAndSwapInt(this, PLOCK, ps, nps)) |
2221 |
releasePlock(nps); |
2222 |
} |
2223 |
for (long c;;) { |
2224 |
if (((c = ctl) & STOP_BIT) != 0) { // already terminating |
2225 |
if ((short)(c >>> TC_SHIFT) + parallelism <= 0) { |
2226 |
synchronized (this) { |
2227 |
notifyAll(); // signal when 0 workers |
2228 |
} |
2229 |
} |
2230 |
return true; |
2231 |
} |
2232 |
if (!now) { // check if idle & no tasks |
2233 |
WorkQueue[] ws; WorkQueue w; |
2234 |
if ((int)(c >> AC_SHIFT) + parallelism > 0) |
2235 |
return false; |
2236 |
if ((ws = workQueues) != null) { |
2237 |
for (int i = 0; i < ws.length; ++i) { |
2238 |
if ((w = ws[i]) != null && |
2239 |
(!w.isEmpty() || |
2240 |
((i & 1) != 0 && w.eventCount >= 0))) { |
2241 |
signalWork(ws, w); |
2242 |
return false; |
2243 |
} |
2244 |
} |
2245 |
} |
2246 |
} |
2247 |
if (U.compareAndSwapLong(this, CTL, c, c | STOP_BIT)) { |
2248 |
for (int pass = 0; pass < 3; ++pass) { |
2249 |
WorkQueue[] ws; WorkQueue w; Thread wt; |
2250 |
if ((ws = workQueues) != null) { |
2251 |
int n = ws.length; |
2252 |
for (int i = 0; i < n; ++i) { |
2253 |
if ((w = ws[i]) != null) { |
2254 |
w.qlock = -1; |
2255 |
if (pass > 0) { |
2256 |
w.cancelAll(); |
2257 |
if (pass > 1 && (wt = w.owner) != null) { |
2258 |
if (!wt.isInterrupted()) { |
2259 |
try { |
2260 |
wt.interrupt(); |
2261 |
} catch (Throwable ignore) { |
2262 |
} |
2263 |
} |
2264 |
U.unpark(wt); |
2265 |
} |
2266 |
} |
2267 |
} |
2268 |
} |
2269 |
// Wake up workers parked on event queue |
2270 |
int i, e; long cc; Thread p; |
2271 |
while ((e = (int)(cc = ctl) & E_MASK) != 0 && |
2272 |
(i = e & SMASK) < n && i >= 0 && |
2273 |
(w = ws[i]) != null) { |
2274 |
long nc = ((long)(w.nextWait & E_MASK) | |
2275 |
((cc + AC_UNIT) & AC_MASK) | |
2276 |
(cc & (TC_MASK|STOP_BIT))); |
2277 |
if (w.eventCount == (e | INT_SIGN) && |
2278 |
U.compareAndSwapLong(this, CTL, cc, nc)) { |
2279 |
w.eventCount = (e + E_SEQ) & E_MASK; |
2280 |
w.qlock = -1; |
2281 |
if ((p = w.parker) != null) |
2282 |
U.unpark(p); |
2283 |
} |
2284 |
} |
2285 |
} |
2286 |
} |
2287 |
} |
2288 |
} |
2289 |
} |
2290 |
|
2291 |
// external operations on common pool |
2292 |
|
2293 |
/** |
2294 |
* Returns common pool queue for a thread that has submitted at |
2295 |
* least one task. |
2296 |
*/ |
2297 |
static WorkQueue commonSubmitterQueue() { |
2298 |
Submitter z; ForkJoinPool p; WorkQueue[] ws; int m, r; |
2299 |
return ((z = submitters.get()) != null && |
2300 |
(p = common) != null && |
2301 |
(ws = p.workQueues) != null && |
2302 |
(m = ws.length - 1) >= 0) ? |
2303 |
ws[m & z.seed & SQMASK] : null; |
2304 |
} |
2305 |
|
2306 |
/** |
2307 |
* Tries to pop the given task from submitter's queue in common pool. |
2308 |
*/ |
2309 |
final boolean tryExternalUnpush(ForkJoinTask<?> task) { |
2310 |
WorkQueue joiner; ForkJoinTask<?>[] a; int m, s; |
2311 |
Submitter z = submitters.get(); |
2312 |
WorkQueue[] ws = workQueues; |
2313 |
boolean popped = false; |
2314 |
if (z != null && ws != null && (m = ws.length - 1) >= 0 && |
2315 |
(joiner = ws[z.seed & m & SQMASK]) != null && |
2316 |
joiner.base != (s = joiner.top) && |
2317 |
(a = joiner.array) != null) { |
2318 |
long j = (((a.length - 1) & (s - 1)) << ASHIFT) + ABASE; |
2319 |
if (U.getObject(a, j) == task && |
2320 |
U.compareAndSwapInt(joiner, QLOCK, 0, 1)) { |
2321 |
if (joiner.top == s && joiner.array == a && |
2322 |
U.compareAndSwapObject(a, j, task, null)) { |
2323 |
joiner.top = s - 1; |
2324 |
popped = true; |
2325 |
} |
2326 |
joiner.qlock = 0; |
2327 |
} |
2328 |
} |
2329 |
return popped; |
2330 |
} |
2331 |
|
2332 |
final int externalHelpComplete(CountedCompleter<?> task) { |
2333 |
WorkQueue joiner; int m, j; |
2334 |
Submitter z = submitters.get(); |
2335 |
WorkQueue[] ws = workQueues; |
2336 |
int s = 0; |
2337 |
if (z != null && ws != null && (m = ws.length - 1) >= 0 && |
2338 |
(joiner = ws[(j = z.seed) & m & SQMASK]) != null && task != null) { |
2339 |
int scans = m + m + 1; |
2340 |
long c = 0L; // for stability check |
2341 |
j |= 1; // poll odd queues |
2342 |
for (int k = scans; ; j += 2) { |
2343 |
WorkQueue q; |
2344 |
if ((s = task.status) < 0) |
2345 |
break; |
2346 |
else if (joiner.externalPopAndExecCC(task)) |
2347 |
k = scans; |
2348 |
else if ((s = task.status) < 0) |
2349 |
break; |
2350 |
else if ((q = ws[j & m]) != null && q.pollAndExecCC(task)) |
2351 |
k = scans; |
2352 |
else if (--k < 0) { |
2353 |
if (c == (c = ctl)) |
2354 |
break; |
2355 |
k = scans; |
2356 |
} |
2357 |
} |
2358 |
} |
2359 |
return s; |
2360 |
} |
2361 |
|
2362 |
// Exported methods |
2363 |
|
2364 |
// Constructors |
2365 |
|
2366 |
/** |
2367 |
* Creates a {@code ForkJoinPool} with parallelism equal to {@link |
2368 |
* java.lang.Runtime#availableProcessors}, using the {@linkplain |
2369 |
* #defaultForkJoinWorkerThreadFactory default thread factory}, |
2370 |
* no UncaughtExceptionHandler, and non-async LIFO processing mode. |
2371 |
* |
2372 |
* @throws SecurityException if a security manager exists and |
2373 |
* the caller is not permitted to modify threads |
2374 |
* because it does not hold {@link |
2375 |
* java.lang.RuntimePermission}{@code ("modifyThread")} |
2376 |
*/ |
2377 |
public ForkJoinPool() { |
2378 |
this(Math.min(MAX_CAP, Runtime.getRuntime().availableProcessors()), |
2379 |
defaultForkJoinWorkerThreadFactory, null, false); |
2380 |
} |
2381 |
|
2382 |
/** |
2383 |
* Creates a {@code ForkJoinPool} with the indicated parallelism |
2384 |
* level, the {@linkplain |
2385 |
* #defaultForkJoinWorkerThreadFactory default thread factory}, |
2386 |
* no UncaughtExceptionHandler, and non-async LIFO processing mode. |
2387 |
* |
2388 |
* @param parallelism the parallelism level |
2389 |
* @throws IllegalArgumentException if parallelism less than or |
2390 |
* equal to zero, or greater than implementation limit |
2391 |
* @throws SecurityException if a security manager exists and |
2392 |
* the caller is not permitted to modify threads |
2393 |
* because it does not hold {@link |
2394 |
* java.lang.RuntimePermission}{@code ("modifyThread")} |
2395 |
*/ |
2396 |
public ForkJoinPool(int parallelism) { |
2397 |
this(parallelism, defaultForkJoinWorkerThreadFactory, null, false); |
2398 |
} |
2399 |
|
2400 |
/** |
2401 |
* Creates a {@code ForkJoinPool} with the given parameters. |
2402 |
* |
2403 |
* @param parallelism the parallelism level. For default value, |
2404 |
* use {@link java.lang.Runtime#availableProcessors}. |
2405 |
* @param factory the factory for creating new threads. For default value, |
2406 |
* use {@link #defaultForkJoinWorkerThreadFactory}. |
2407 |
* @param handler the handler for internal worker threads that |
2408 |
* terminate due to unrecoverable errors encountered while executing |
2409 |
* tasks. For default value, use {@code null}. |
2410 |
* @param asyncMode if true, |
2411 |
* establishes local first-in-first-out scheduling mode for forked |
2412 |
* tasks that are never joined. This mode may be more appropriate |
2413 |
* than default locally stack-based mode in applications in which |
2414 |
* worker threads only process event-style asynchronous tasks. |
2415 |
* For default value, use {@code false}. |
2416 |
* @throws IllegalArgumentException if parallelism less than or |
2417 |
* equal to zero, or greater than implementation limit |
2418 |
* @throws NullPointerException if the factory is null |
2419 |
* @throws SecurityException if a security manager exists and |
2420 |
* the caller is not permitted to modify threads |
2421 |
* because it does not hold {@link |
2422 |
* java.lang.RuntimePermission}{@code ("modifyThread")} |
2423 |
*/ |
2424 |
public ForkJoinPool(int parallelism, |
2425 |
ForkJoinWorkerThreadFactory factory, |
2426 |
UncaughtExceptionHandler handler, |
2427 |
boolean asyncMode) { |
2428 |
this(checkParallelism(parallelism), |
2429 |
checkFactory(factory), |
2430 |
handler, |
2431 |
(asyncMode ? FIFO_QUEUE : LIFO_QUEUE), |
2432 |
"ForkJoinPool-" + nextPoolId() + "-worker-"); |
2433 |
checkPermission(); |
2434 |
} |
2435 |
|
2436 |
private static int checkParallelism(int parallelism) { |
2437 |
if (parallelism <= 0 || parallelism > MAX_CAP) |
2438 |
throw new IllegalArgumentException(); |
2439 |
return parallelism; |
2440 |
} |
2441 |
|
2442 |
private static ForkJoinWorkerThreadFactory checkFactory |
2443 |
(ForkJoinWorkerThreadFactory factory) { |
2444 |
if (factory == null) |
2445 |
throw new NullPointerException(); |
2446 |
return factory; |
2447 |
} |
2448 |
|
2449 |
/** |
2450 |
* Creates a {@code ForkJoinPool} with the given parameters, without |
2451 |
* any security checks or parameter validation. Invoked directly by |
2452 |
* makeCommonPool. |
2453 |
*/ |
2454 |
private ForkJoinPool(int parallelism, |
2455 |
ForkJoinWorkerThreadFactory factory, |
2456 |
UncaughtExceptionHandler handler, |
2457 |
int mode, |
2458 |
String workerNamePrefix) { |
2459 |
this.workerNamePrefix = workerNamePrefix; |
2460 |
this.factory = factory; |
2461 |
this.ueh = handler; |
2462 |
this.mode = (short)mode; |
2463 |
this.parallelism = (short)parallelism; |
2464 |
long np = (long)(-parallelism); // offset ctl counts |
2465 |
this.ctl = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK); |
2466 |
} |
2467 |
|
2468 |
/** |
2469 |
* Returns the common pool instance. This pool is statically |
2470 |
* constructed; its run state is unaffected by attempts to {@link |
2471 |
* #shutdown} or {@link #shutdownNow}. However this pool and any |
2472 |
* ongoing processing are automatically terminated upon program |
2473 |
* {@link System#exit}. Any program that relies on asynchronous |
2474 |
* task processing to complete before program termination should |
2475 |
* invoke {@code commonPool().}{@link #awaitQuiescence awaitQuiescence}, |
2476 |
* before exit. |
2477 |
* |
2478 |
* @return the common pool instance |
2479 |
* @since 1.8 |
2480 |
*/ |
2481 |
public static ForkJoinPool commonPool() { |
2482 |
// assert common != null : "static init error"; |
2483 |
return common; |
2484 |
} |
2485 |
|
2486 |
// Execution methods |
2487 |
|
2488 |
/** |
2489 |
* Performs the given task, returning its result upon completion. |
2490 |
* If the computation encounters an unchecked Exception or Error, |
2491 |
* it is rethrown as the outcome of this invocation. Rethrown |
2492 |
* exceptions behave in the same way as regular exceptions, but, |
2493 |
* when possible, contain stack traces (as displayed for example |
2494 |
* using {@code ex.printStackTrace()}) of both the current thread |
2495 |
* as well as the thread actually encountering the exception; |
2496 |
* minimally only the latter. |
2497 |
* |
2498 |
* @param task the task |
2499 |
* @return the task's result |
2500 |
* @throws NullPointerException if the task is null |
2501 |
* @throws RejectedExecutionException if the task cannot be |
2502 |
* scheduled for execution |
2503 |
*/ |
2504 |
public <T> T invoke(ForkJoinTask<T> task) { |
2505 |
if (task == null) |
2506 |
throw new NullPointerException(); |
2507 |
externalPush(task); |
2508 |
return task.join(); |
2509 |
} |
2510 |
|
2511 |
/** |
2512 |
* Arranges for (asynchronous) execution of the given task. |
2513 |
* |
2514 |
* @param task the task |
2515 |
* @throws NullPointerException if the task is null |
2516 |
* @throws RejectedExecutionException if the task cannot be |
2517 |
* scheduled for execution |
2518 |
*/ |
2519 |
public void execute(ForkJoinTask<?> task) { |
2520 |
if (task == null) |
2521 |
throw new NullPointerException(); |
2522 |
externalPush(task); |
2523 |
} |
2524 |
|
2525 |
// AbstractExecutorService methods |
2526 |
|
2527 |
/** |
2528 |
* @throws NullPointerException if the task is null |
2529 |
* @throws RejectedExecutionException if the task cannot be |
2530 |
* scheduled for execution |
2531 |
*/ |
2532 |
public void execute(Runnable task) { |
2533 |
if (task == null) |
2534 |
throw new NullPointerException(); |
2535 |
ForkJoinTask<?> job; |
2536 |
if (task instanceof ForkJoinTask<?>) // avoid re-wrap |
2537 |
job = (ForkJoinTask<?>) task; |
2538 |
else |
2539 |
job = new ForkJoinTask.RunnableExecuteAction(task); |
2540 |
externalPush(job); |
2541 |
} |
2542 |
|
2543 |
/** |
2544 |
* Submits a ForkJoinTask for execution. |
2545 |
* |
2546 |
* @param task the task to submit |
2547 |
* @return the task |
2548 |
* @throws NullPointerException if the task is null |
2549 |
* @throws RejectedExecutionException if the task cannot be |
2550 |
* scheduled for execution |
2551 |
*/ |
2552 |
public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) { |
2553 |
if (task == null) |
2554 |
throw new NullPointerException(); |
2555 |
externalPush(task); |
2556 |
return task; |
2557 |
} |
2558 |
|
2559 |
/** |
2560 |
* @throws NullPointerException if the task is null |
2561 |
* @throws RejectedExecutionException if the task cannot be |
2562 |
* scheduled for execution |
2563 |
*/ |
2564 |
public <T> ForkJoinTask<T> submit(Callable<T> task) { |
2565 |
ForkJoinTask<T> job = new ForkJoinTask.AdaptedCallable<T>(task); |
2566 |
externalPush(job); |
2567 |
return job; |
2568 |
} |
2569 |
|
2570 |
/** |
2571 |
* @throws NullPointerException if the task is null |
2572 |
* @throws RejectedExecutionException if the task cannot be |
2573 |
* scheduled for execution |
2574 |
*/ |
2575 |
public <T> ForkJoinTask<T> submit(Runnable task, T result) { |
2576 |
ForkJoinTask<T> job = new ForkJoinTask.AdaptedRunnable<T>(task, result); |
2577 |
externalPush(job); |
2578 |
return job; |
2579 |
} |
2580 |
|
2581 |
/** |
2582 |
* @throws NullPointerException if the task is null |
2583 |
* @throws RejectedExecutionException if the task cannot be |
2584 |
* scheduled for execution |
2585 |
*/ |
2586 |
public ForkJoinTask<?> submit(Runnable task) { |
2587 |
if (task == null) |
2588 |
throw new NullPointerException(); |
2589 |
ForkJoinTask<?> job; |
2590 |
if (task instanceof ForkJoinTask<?>) // avoid re-wrap |
2591 |
job = (ForkJoinTask<?>) task; |
2592 |
else |
2593 |
job = new ForkJoinTask.AdaptedRunnableAction(task); |
2594 |
externalPush(job); |
2595 |
return job; |
2596 |
} |
2597 |
|
2598 |
/** |
2599 |
* @throws NullPointerException {@inheritDoc} |
2600 |
* @throws RejectedExecutionException {@inheritDoc} |
2601 |
*/ |
2602 |
public <T> List<Future<T>> invokeAll(Collection<? extends Callable<T>> tasks) { |
2603 |
// In previous versions of this class, this method constructed |
2604 |
// a task to run ForkJoinTask.invokeAll, but now external |
2605 |
// invocation of multiple tasks is at least as efficient. |
2606 |
ArrayList<Future<T>> futures = new ArrayList<Future<T>>(tasks.size()); |
2607 |
|
2608 |
boolean done = false; |
2609 |
try { |
2610 |
for (Callable<T> t : tasks) { |
2611 |
ForkJoinTask<T> f = new ForkJoinTask.AdaptedCallable<T>(t); |
2612 |
futures.add(f); |
2613 |
externalPush(f); |
2614 |
} |
2615 |
for (int i = 0, size = futures.size(); i < size; i++) |
2616 |
((ForkJoinTask<?>)futures.get(i)).quietlyJoin(); |
2617 |
done = true; |
2618 |
return futures; |
2619 |
} finally { |
2620 |
if (!done) |
2621 |
for (int i = 0, size = futures.size(); i < size; i++) |
2622 |
futures.get(i).cancel(false); |
2623 |
} |
2624 |
} |
2625 |
|
2626 |
/** |
2627 |
* Returns the factory used for constructing new workers. |
2628 |
* |
2629 |
* @return the factory used for constructing new workers |
2630 |
*/ |
2631 |
public ForkJoinWorkerThreadFactory getFactory() { |
2632 |
return factory; |
2633 |
} |
2634 |
|
2635 |
/** |
2636 |
* Returns the handler for internal worker threads that terminate |
2637 |
* due to unrecoverable errors encountered while executing tasks. |
2638 |
* |
2639 |
* @return the handler, or {@code null} if none |
2640 |
*/ |
2641 |
public UncaughtExceptionHandler getUncaughtExceptionHandler() { |
2642 |
return ueh; |
2643 |
} |
2644 |
|
2645 |
/** |
2646 |
* Returns the targeted parallelism level of this pool. |
2647 |
* |
2648 |
* @return the targeted parallelism level of this pool |
2649 |
*/ |
2650 |
public int getParallelism() { |
2651 |
int par; |
2652 |
return ((par = parallelism) > 0) ? par : 1; |
2653 |
} |
2654 |
|
2655 |
/** |
2656 |
* Returns the targeted parallelism level of the common pool. |
2657 |
* |
2658 |
* @return the targeted parallelism level of the common pool |
2659 |
* @since 1.8 |
2660 |
*/ |
2661 |
public static int getCommonPoolParallelism() { |
2662 |
return commonParallelism; |
2663 |
} |
2664 |
|
2665 |
/** |
2666 |
* Returns the number of worker threads that have started but not |
2667 |
* yet terminated. The result returned by this method may differ |
2668 |
* from {@link #getParallelism} when threads are created to |
2669 |
* maintain parallelism when others are cooperatively blocked. |
2670 |
* |
2671 |
* @return the number of worker threads |
2672 |
*/ |
2673 |
public int getPoolSize() { |
2674 |
return parallelism + (short)(ctl >>> TC_SHIFT); |
2675 |
} |
2676 |
|
2677 |
/** |
2678 |
* Returns {@code true} if this pool uses local first-in-first-out |
2679 |
* scheduling mode for forked tasks that are never joined. |
2680 |
* |
2681 |
* @return {@code true} if this pool uses async mode |
2682 |
*/ |
2683 |
public boolean getAsyncMode() { |
2684 |
return mode == FIFO_QUEUE; |
2685 |
} |
2686 |
|
2687 |
/** |
2688 |
* Returns an estimate of the number of worker threads that are |
2689 |
* not blocked waiting to join tasks or for other managed |
2690 |
* synchronization. This method may overestimate the |
2691 |
* number of running threads. |
2692 |
* |
2693 |
* @return the number of worker threads |
2694 |
*/ |
2695 |
public int getRunningThreadCount() { |
2696 |
int rc = 0; |
2697 |
WorkQueue[] ws; WorkQueue w; |
2698 |
if ((ws = workQueues) != null) { |
2699 |
for (int i = 1; i < ws.length; i += 2) { |
2700 |
if ((w = ws[i]) != null && w.isApparentlyUnblocked()) |
2701 |
++rc; |
2702 |
} |
2703 |
} |
2704 |
return rc; |
2705 |
} |
2706 |
|
2707 |
/** |
2708 |
* Returns an estimate of the number of threads that are currently |
2709 |
* stealing or executing tasks. This method may overestimate the |
2710 |
* number of active threads. |
2711 |
* |
2712 |
* @return the number of active threads |
2713 |
*/ |
2714 |
public int getActiveThreadCount() { |
2715 |
int r = parallelism + (int)(ctl >> AC_SHIFT); |
2716 |
return (r <= 0) ? 0 : r; // suppress momentarily negative values |
2717 |
} |
2718 |
|
2719 |
/** |
2720 |
* Returns {@code true} if all worker threads are currently idle. |
2721 |
* An idle worker is one that cannot obtain a task to execute |
2722 |
* because none are available to steal from other threads, and |
2723 |
* there are no pending submissions to the pool. This method is |
2724 |
* conservative; it might not return {@code true} immediately upon |
2725 |
* idleness of all threads, but will eventually become true if |
2726 |
* threads remain inactive. |
2727 |
* |
2728 |
* @return {@code true} if all threads are currently idle |
2729 |
*/ |
2730 |
public boolean isQuiescent() { |
2731 |
return parallelism + (int)(ctl >> AC_SHIFT) <= 0; |
2732 |
} |
2733 |
|
2734 |
/** |
2735 |
* Returns an estimate of the total number of tasks stolen from |
2736 |
* one thread's work queue by another. The reported value |
2737 |
* underestimates the actual total number of steals when the pool |
2738 |
* is not quiescent. This value may be useful for monitoring and |
2739 |
* tuning fork/join programs: in general, steal counts should be |
2740 |
* high enough to keep threads busy, but low enough to avoid |
2741 |
* overhead and contention across threads. |
2742 |
* |
2743 |
* @return the number of steals |
2744 |
*/ |
2745 |
public long getStealCount() { |
2746 |
long count = stealCount; |
2747 |
WorkQueue[] ws; WorkQueue w; |
2748 |
if ((ws = workQueues) != null) { |
2749 |
for (int i = 1; i < ws.length; i += 2) { |
2750 |
if ((w = ws[i]) != null) |
2751 |
count += w.nsteals; |
2752 |
} |
2753 |
} |
2754 |
return count; |
2755 |
} |
2756 |
|
2757 |
/** |
2758 |
* Returns an estimate of the total number of tasks currently held |
2759 |
* in queues by worker threads (but not including tasks submitted |
2760 |
* to the pool that have not begun executing). This value is only |
2761 |
* an approximation, obtained by iterating across all threads in |
2762 |
* the pool. This method may be useful for tuning task |
2763 |
* granularities. |
2764 |
* |
2765 |
* @return the number of queued tasks |
2766 |
*/ |
2767 |
public long getQueuedTaskCount() { |
2768 |
long count = 0; |
2769 |
WorkQueue[] ws; WorkQueue w; |
2770 |
if ((ws = workQueues) != null) { |
2771 |
for (int i = 1; i < ws.length; i += 2) { |
2772 |
if ((w = ws[i]) != null) |
2773 |
count += w.queueSize(); |
2774 |
} |
2775 |
} |
2776 |
return count; |
2777 |
} |
2778 |
|
2779 |
/** |
2780 |
* Returns an estimate of the number of tasks submitted to this |
2781 |
* pool that have not yet begun executing. This method may take |
2782 |
* time proportional to the number of submissions. |
2783 |
* |
2784 |
* @return the number of queued submissions |
2785 |
*/ |
2786 |
public int getQueuedSubmissionCount() { |
2787 |
int count = 0; |
2788 |
WorkQueue[] ws; WorkQueue w; |
2789 |
if ((ws = workQueues) != null) { |
2790 |
for (int i = 0; i < ws.length; i += 2) { |
2791 |
if ((w = ws[i]) != null) |
2792 |
count += w.queueSize(); |
2793 |
} |
2794 |
} |
2795 |
return count; |
2796 |
} |
2797 |
|
2798 |
/** |
2799 |
* Returns {@code true} if there are any tasks submitted to this |
2800 |
* pool that have not yet begun executing. |
2801 |
* |
2802 |
* @return {@code true} if there are any queued submissions |
2803 |
*/ |
2804 |
public boolean hasQueuedSubmissions() { |
2805 |
WorkQueue[] ws; WorkQueue w; |
2806 |
if ((ws = workQueues) != null) { |
2807 |
for (int i = 0; i < ws.length; i += 2) { |
2808 |
if ((w = ws[i]) != null && !w.isEmpty()) |
2809 |
return true; |
2810 |
} |
2811 |
} |
2812 |
return false; |
2813 |
} |
2814 |
|
2815 |
/** |
2816 |
* Removes and returns the next unexecuted submission if one is |
2817 |
* available. This method may be useful in extensions to this |
2818 |
* class that re-assign work in systems with multiple pools. |
2819 |
* |
2820 |
* @return the next submission, or {@code null} if none |
2821 |
*/ |
2822 |
protected ForkJoinTask<?> pollSubmission() { |
2823 |
WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t; |
2824 |
if ((ws = workQueues) != null) { |
2825 |
for (int i = 0; i < ws.length; i += 2) { |
2826 |
if ((w = ws[i]) != null && (t = w.poll()) != null) |
2827 |
return t; |
2828 |
} |
2829 |
} |
2830 |
return null; |
2831 |
} |
2832 |
|
2833 |
/** |
2834 |
* Removes all available unexecuted submitted and forked tasks |
2835 |
* from scheduling queues and adds them to the given collection, |
2836 |
* without altering their execution status. These may include |
2837 |
* artificially generated or wrapped tasks. This method is |
2838 |
* designed to be invoked only when the pool is known to be |
2839 |
* quiescent. Invocations at other times may not remove all |
2840 |
* tasks. A failure encountered while attempting to add elements |
2841 |
* to collection {@code c} may result in elements being in |
2842 |
* neither, either or both collections when the associated |
2843 |
* exception is thrown. The behavior of this operation is |
2844 |
* undefined if the specified collection is modified while the |
2845 |
* operation is in progress. |
2846 |
* |
2847 |
* @param c the collection to transfer elements into |
2848 |
* @return the number of elements transferred |
2849 |
*/ |
2850 |
protected int drainTasksTo(Collection<? super ForkJoinTask<?>> c) { |
2851 |
int count = 0; |
2852 |
WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t; |
2853 |
if ((ws = workQueues) != null) { |
2854 |
for (int i = 0; i < ws.length; ++i) { |
2855 |
if ((w = ws[i]) != null) { |
2856 |
while ((t = w.poll()) != null) { |
2857 |
c.add(t); |
2858 |
++count; |
2859 |
} |
2860 |
} |
2861 |
} |
2862 |
} |
2863 |
return count; |
2864 |
} |
2865 |
|
2866 |
/** |
2867 |
* Returns a string identifying this pool, as well as its state, |
2868 |
* including indications of run state, parallelism level, and |
2869 |
* worker and task counts. |
2870 |
* |
2871 |
* @return a string identifying this pool, as well as its state |
2872 |
*/ |
2873 |
public String toString() { |
2874 |
// Use a single pass through workQueues to collect counts |
2875 |
long qt = 0L, qs = 0L; int rc = 0; |
2876 |
long st = stealCount; |
2877 |
long c = ctl; |
2878 |
WorkQueue[] ws; WorkQueue w; |
2879 |
if ((ws = workQueues) != null) { |
2880 |
for (int i = 0; i < ws.length; ++i) { |
2881 |
if ((w = ws[i]) != null) { |
2882 |
int size = w.queueSize(); |
2883 |
if ((i & 1) == 0) |
2884 |
qs += size; |
2885 |
else { |
2886 |
qt += size; |
2887 |
st += w.nsteals; |
2888 |
if (w.isApparentlyUnblocked()) |
2889 |
++rc; |
2890 |
} |
2891 |
} |
2892 |
} |
2893 |
} |
2894 |
int pc = parallelism; |
2895 |
int tc = pc + (short)(c >>> TC_SHIFT); |
2896 |
int ac = pc + (int)(c >> AC_SHIFT); |
2897 |
if (ac < 0) // ignore transient negative |
2898 |
ac = 0; |
2899 |
String level; |
2900 |
if ((c & STOP_BIT) != 0) |
2901 |
level = (tc == 0) ? "Terminated" : "Terminating"; |
2902 |
else |
2903 |
level = plock < 0 ? "Shutting down" : "Running"; |
2904 |
return super.toString() + |
2905 |
"[" + level + |
2906 |
", parallelism = " + pc + |
2907 |
", size = " + tc + |
2908 |
", active = " + ac + |
2909 |
", running = " + rc + |
2910 |
", steals = " + st + |
2911 |
", tasks = " + qt + |
2912 |
", submissions = " + qs + |
2913 |
"]"; |
2914 |
} |
2915 |
|
2916 |
/** |
2917 |
* Possibly initiates an orderly shutdown in which previously |
2918 |
* submitted tasks are executed, but no new tasks will be |
2919 |
* accepted. Invocation has no effect on execution state if this |
2920 |
* is the {@link #commonPool()}, and no additional effect if |
2921 |
* already shut down. Tasks that are in the process of being |
2922 |
* submitted concurrently during the course of this method may or |
2923 |
* may not be rejected. |
2924 |
* |
2925 |
* @throws SecurityException if a security manager exists and |
2926 |
* the caller is not permitted to modify threads |
2927 |
* because it does not hold {@link |
2928 |
* java.lang.RuntimePermission}{@code ("modifyThread")} |
2929 |
*/ |
2930 |
public void shutdown() { |
2931 |
checkPermission(); |
2932 |
tryTerminate(false, true); |
2933 |
} |
2934 |
|
2935 |
/** |
2936 |
* Possibly attempts to cancel and/or stop all tasks, and reject |
2937 |
* all subsequently submitted tasks. Invocation has no effect on |
2938 |
* execution state if this is the {@link #commonPool()}, and no |
2939 |
* additional effect if already shut down. Otherwise, tasks that |
2940 |
* are in the process of being submitted or executed concurrently |
2941 |
* during the course of this method may or may not be |
2942 |
* rejected. This method cancels both existing and unexecuted |
2943 |
* tasks, in order to permit termination in the presence of task |
2944 |
* dependencies. So the method always returns an empty list |
2945 |
* (unlike the case for some other Executors). |
2946 |
* |
2947 |
* @return an empty list |
2948 |
* @throws SecurityException if a security manager exists and |
2949 |
* the caller is not permitted to modify threads |
2950 |
* because it does not hold {@link |
2951 |
* java.lang.RuntimePermission}{@code ("modifyThread")} |
2952 |
*/ |
2953 |
public List<Runnable> shutdownNow() { |
2954 |
checkPermission(); |
2955 |
tryTerminate(true, true); |
2956 |
return Collections.emptyList(); |
2957 |
} |
2958 |
|
2959 |
/** |
2960 |
* Returns {@code true} if all tasks have completed following shut down. |
2961 |
* |
2962 |
* @return {@code true} if all tasks have completed following shut down |
2963 |
*/ |
2964 |
public boolean isTerminated() { |
2965 |
long c = ctl; |
2966 |
return ((c & STOP_BIT) != 0L && |
2967 |
(short)(c >>> TC_SHIFT) + parallelism <= 0); |
2968 |
} |
2969 |
|
2970 |
/** |
2971 |
* Returns {@code true} if the process of termination has |
2972 |
* commenced but not yet completed. This method may be useful for |
2973 |
* debugging. A return of {@code true} reported a sufficient |
2974 |
* period after shutdown may indicate that submitted tasks have |
2975 |
* ignored or suppressed interruption, or are waiting for I/O, |
2976 |
* causing this executor not to properly terminate. (See the |
2977 |
* advisory notes for class {@link ForkJoinTask} stating that |
2978 |
* tasks should not normally entail blocking operations. But if |
2979 |
* they do, they must abort them on interrupt.) |
2980 |
* |
2981 |
* @return {@code true} if terminating but not yet terminated |
2982 |
*/ |
2983 |
public boolean isTerminating() { |
2984 |
long c = ctl; |
2985 |
return ((c & STOP_BIT) != 0L && |
2986 |
(short)(c >>> TC_SHIFT) + parallelism > 0); |
2987 |
} |
2988 |
|
2989 |
/** |
2990 |
* Returns {@code true} if this pool has been shut down. |
2991 |
* |
2992 |
* @return {@code true} if this pool has been shut down |
2993 |
*/ |
2994 |
public boolean isShutdown() { |
2995 |
return plock < 0; |
2996 |
} |
2997 |
|
2998 |
/** |
2999 |
* Blocks until all tasks have completed execution after a |
3000 |
* shutdown request, or the timeout occurs, or the current thread |
3001 |
* is interrupted, whichever happens first. Because the {@link |
3002 |
* #commonPool()} never terminates until program shutdown, when |
3003 |
* applied to the common pool, this method is equivalent to {@link |
3004 |
* #awaitQuiescence(long, TimeUnit)} but always returns {@code false}. |
3005 |
* |
3006 |
* @param timeout the maximum time to wait |
3007 |
* @param unit the time unit of the timeout argument |
3008 |
* @return {@code true} if this executor terminated and |
3009 |
* {@code false} if the timeout elapsed before termination |
3010 |
* @throws InterruptedException if interrupted while waiting |
3011 |
*/ |
3012 |
public boolean awaitTermination(long timeout, TimeUnit unit) |
3013 |
throws InterruptedException { |
3014 |
if (Thread.interrupted()) |
3015 |
throw new InterruptedException(); |
3016 |
if (this == common) { |
3017 |
awaitQuiescence(timeout, unit); |
3018 |
return false; |
3019 |
} |
3020 |
long nanos = unit.toNanos(timeout); |
3021 |
if (isTerminated()) |
3022 |
return true; |
3023 |
if (nanos <= 0L) |
3024 |
return false; |
3025 |
long deadline = System.nanoTime() + nanos; |
3026 |
synchronized (this) { |
3027 |
for (;;) { |
3028 |
if (isTerminated()) |
3029 |
return true; |
3030 |
if (nanos <= 0L) |
3031 |
return false; |
3032 |
long millis = TimeUnit.NANOSECONDS.toMillis(nanos); |
3033 |
wait(millis > 0L ? millis : 1L); |
3034 |
nanos = deadline - System.nanoTime(); |
3035 |
} |
3036 |
} |
3037 |
} |
3038 |
|
3039 |
/** |
3040 |
* If called by a ForkJoinTask operating in this pool, equivalent |
3041 |
* in effect to {@link ForkJoinTask#helpQuiesce}. Otherwise, |
3042 |
* waits and/or attempts to assist performing tasks until this |
3043 |
* pool {@link #isQuiescent} or the indicated timeout elapses. |
3044 |
* |
3045 |
* @param timeout the maximum time to wait |
3046 |
* @param unit the time unit of the timeout argument |
3047 |
* @return {@code true} if quiescent; {@code false} if the |
3048 |
* timeout elapsed. |
3049 |
*/ |
3050 |
public boolean awaitQuiescence(long timeout, TimeUnit unit) { |
3051 |
long nanos = unit.toNanos(timeout); |
3052 |
ForkJoinWorkerThread wt; |
3053 |
Thread thread = Thread.currentThread(); |
3054 |
if ((thread instanceof ForkJoinWorkerThread) && |
3055 |
(wt = (ForkJoinWorkerThread)thread).pool == this) { |
3056 |
helpQuiescePool(wt.workQueue); |
3057 |
return true; |
3058 |
} |
3059 |
long startTime = System.nanoTime(); |
3060 |
WorkQueue[] ws; |
3061 |
int r = 0, m; |
3062 |
boolean found = true; |
3063 |
while (!isQuiescent() && (ws = workQueues) != null && |
3064 |
(m = ws.length - 1) >= 0) { |
3065 |
if (!found) { |
3066 |
if ((System.nanoTime() - startTime) > nanos) |
3067 |
return false; |
3068 |
Thread.yield(); // cannot block |
3069 |
} |
3070 |
found = false; |
3071 |
for (int j = (m + 1) << 2; j >= 0; --j) { |
3072 |
ForkJoinTask<?> t; WorkQueue q; int b; |
3073 |
if ((q = ws[r++ & m]) != null && (b = q.base) - q.top < 0) { |
3074 |
found = true; |
3075 |
if ((t = q.pollAt(b)) != null) |
3076 |
t.doExec(); |
3077 |
break; |
3078 |
} |
3079 |
} |
3080 |
} |
3081 |
return true; |
3082 |
} |
3083 |
|
3084 |
/** |
3085 |
* Waits and/or attempts to assist performing tasks indefinitely |
3086 |
* until the {@link #commonPool()} {@link #isQuiescent}. |
3087 |
*/ |
3088 |
static void quiesceCommonPool() { |
3089 |
common.awaitQuiescence(Long.MAX_VALUE, TimeUnit.NANOSECONDS); |
3090 |
} |
3091 |
|
3092 |
/** |
3093 |
* Interface for extending managed parallelism for tasks running |
3094 |
* in {@link ForkJoinPool}s. |
3095 |
* |
3096 |
* <p>A {@code ManagedBlocker} provides two methods. Method |
3097 |
* {@code isReleasable} must return {@code true} if blocking is |
3098 |
* not necessary. Method {@code block} blocks the current thread |
3099 |
* if necessary (perhaps internally invoking {@code isReleasable} |
3100 |
* before actually blocking). These actions are performed by any |
3101 |
* thread invoking {@link ForkJoinPool#managedBlock(ManagedBlocker)}. |
3102 |
* The unusual methods in this API accommodate synchronizers that |
3103 |
* may, but don't usually, block for long periods. Similarly, they |
3104 |
* allow more efficient internal handling of cases in which |
3105 |
* additional workers may be, but usually are not, needed to |
3106 |
* ensure sufficient parallelism. Toward this end, |
3107 |
* implementations of method {@code isReleasable} must be amenable |
3108 |
* to repeated invocation. |
3109 |
* |
3110 |
* <p>For example, here is a ManagedBlocker based on a |
3111 |
* ReentrantLock: |
3112 |
* <pre> {@code |
3113 |
* class ManagedLocker implements ManagedBlocker { |
3114 |
* final ReentrantLock lock; |
3115 |
* boolean hasLock = false; |
3116 |
* ManagedLocker(ReentrantLock lock) { this.lock = lock; } |
3117 |
* public boolean block() { |
3118 |
* if (!hasLock) |
3119 |
* lock.lock(); |
3120 |
* return true; |
3121 |
* } |
3122 |
* public boolean isReleasable() { |
3123 |
* return hasLock || (hasLock = lock.tryLock()); |
3124 |
* } |
3125 |
* }}</pre> |
3126 |
* |
3127 |
* <p>Here is a class that possibly blocks waiting for an |
3128 |
* item on a given queue: |
3129 |
* <pre> {@code |
3130 |
* class QueueTaker<E> implements ManagedBlocker { |
3131 |
* final BlockingQueue<E> queue; |
3132 |
* volatile E item = null; |
3133 |
* QueueTaker(BlockingQueue<E> q) { this.queue = q; } |
3134 |
* public boolean block() throws InterruptedException { |
3135 |
* if (item == null) |
3136 |
* item = queue.take(); |
3137 |
* return true; |
3138 |
* } |
3139 |
* public boolean isReleasable() { |
3140 |
* return item != null || (item = queue.poll()) != null; |
3141 |
* } |
3142 |
* public E getItem() { // call after pool.managedBlock completes |
3143 |
* return item; |
3144 |
* } |
3145 |
* }}</pre> |
3146 |
*/ |
3147 |
public static interface ManagedBlocker { |
3148 |
/** |
3149 |
* Possibly blocks the current thread, for example waiting for |
3150 |
* a lock or condition. |
3151 |
* |
3152 |
* @return {@code true} if no additional blocking is necessary |
3153 |
* (i.e., if isReleasable would return true) |
3154 |
* @throws InterruptedException if interrupted while waiting |
3155 |
* (the method is not required to do so, but is allowed to) |
3156 |
*/ |
3157 |
boolean block() throws InterruptedException; |
3158 |
|
3159 |
/** |
3160 |
* Returns {@code true} if blocking is unnecessary. |
3161 |
* @return {@code true} if blocking is unnecessary |
3162 |
*/ |
3163 |
boolean isReleasable(); |
3164 |
} |
3165 |
|
3166 |
/** |
3167 |
* Blocks in accord with the given blocker. If the current thread |
3168 |
* is a {@link ForkJoinWorkerThread}, this method possibly |
3169 |
* arranges for a spare thread to be activated if necessary to |
3170 |
* ensure sufficient parallelism while the current thread is blocked. |
3171 |
* |
3172 |
* <p>If the caller is not a {@link ForkJoinTask}, this method is |
3173 |
* behaviorally equivalent to |
3174 |
* <pre> {@code |
3175 |
* while (!blocker.isReleasable()) |
3176 |
* if (blocker.block()) |
3177 |
* return; |
3178 |
* }</pre> |
3179 |
* |
3180 |
* If the caller is a {@code ForkJoinTask}, then the pool may |
3181 |
* first be expanded to ensure parallelism, and later adjusted. |
3182 |
* |
3183 |
* @param blocker the blocker |
3184 |
* @throws InterruptedException if blocker.block did so |
3185 |
*/ |
3186 |
public static void managedBlock(ManagedBlocker blocker) |
3187 |
throws InterruptedException { |
3188 |
Thread t = Thread.currentThread(); |
3189 |
if (t instanceof ForkJoinWorkerThread) { |
3190 |
ForkJoinPool p = ((ForkJoinWorkerThread)t).pool; |
3191 |
while (!blocker.isReleasable()) { |
3192 |
if (p.tryCompensate(p.ctl)) { |
3193 |
try { |
3194 |
do {} while (!blocker.isReleasable() && |
3195 |
!blocker.block()); |
3196 |
} finally { |
3197 |
p.incrementActiveCount(); |
3198 |
} |
3199 |
break; |
3200 |
} |
3201 |
} |
3202 |
} |
3203 |
else { |
3204 |
do {} while (!blocker.isReleasable() && |
3205 |
!blocker.block()); |
3206 |
} |
3207 |
} |
3208 |
|
3209 |
// AbstractExecutorService overrides. These rely on undocumented |
3210 |
// fact that ForkJoinTask.adapt returns ForkJoinTasks that also |
3211 |
// implement RunnableFuture. |
3212 |
|
3213 |
protected <T> RunnableFuture<T> newTaskFor(Runnable runnable, T value) { |
3214 |
return new ForkJoinTask.AdaptedRunnable<T>(runnable, value); |
3215 |
} |
3216 |
|
3217 |
protected <T> RunnableFuture<T> newTaskFor(Callable<T> callable) { |
3218 |
return new ForkJoinTask.AdaptedCallable<T>(callable); |
3219 |
} |
3220 |
|
3221 |
// Unsafe mechanics |
3222 |
private static final sun.misc.Unsafe U; |
3223 |
private static final long CTL; |
3224 |
private static final long PARKBLOCKER; |
3225 |
private static final int ABASE; |
3226 |
private static final int ASHIFT; |
3227 |
private static final long STEALCOUNT; |
3228 |
private static final long PLOCK; |
3229 |
private static final long INDEXSEED; |
3230 |
private static final long QBASE; |
3231 |
private static final long QLOCK; |
3232 |
|
3233 |
static { |
3234 |
// initialize field offsets for CAS etc |
3235 |
try { |
3236 |
U = getUnsafe(); |
3237 |
Class<?> k = ForkJoinPool.class; |
3238 |
CTL = U.objectFieldOffset |
3239 |
(k.getDeclaredField("ctl")); |
3240 |
STEALCOUNT = U.objectFieldOffset |
3241 |
(k.getDeclaredField("stealCount")); |
3242 |
PLOCK = U.objectFieldOffset |
3243 |
(k.getDeclaredField("plock")); |
3244 |
INDEXSEED = U.objectFieldOffset |
3245 |
(k.getDeclaredField("indexSeed")); |
3246 |
Class<?> tk = Thread.class; |
3247 |
PARKBLOCKER = U.objectFieldOffset |
3248 |
(tk.getDeclaredField("parkBlocker")); |
3249 |
Class<?> wk = WorkQueue.class; |
3250 |
QBASE = U.objectFieldOffset |
3251 |
(wk.getDeclaredField("base")); |
3252 |
QLOCK = U.objectFieldOffset |
3253 |
(wk.getDeclaredField("qlock")); |
3254 |
Class<?> ak = ForkJoinTask[].class; |
3255 |
ABASE = U.arrayBaseOffset(ak); |
3256 |
int scale = U.arrayIndexScale(ak); |
3257 |
if ((scale & (scale - 1)) != 0) |
3258 |
throw new Error("data type scale not a power of two"); |
3259 |
ASHIFT = 31 - Integer.numberOfLeadingZeros(scale); |
3260 |
} catch (Exception e) { |
3261 |
throw new Error(e); |
3262 |
} |
3263 |
|
3264 |
submitters = new ThreadLocal<Submitter>(); |
3265 |
defaultForkJoinWorkerThreadFactory = |
3266 |
new DefaultForkJoinWorkerThreadFactory(); |
3267 |
modifyThreadPermission = new RuntimePermission("modifyThread"); |
3268 |
|
3269 |
common = java.security.AccessController.doPrivileged |
3270 |
(new java.security.PrivilegedAction<ForkJoinPool>() { |
3271 |
public ForkJoinPool run() { return makeCommonPool(); }}); |
3272 |
int par = common.parallelism; // report 1 even if threads disabled |
3273 |
commonParallelism = par > 0 ? par : 1; |
3274 |
} |
3275 |
|
3276 |
/** |
3277 |
* Creates and returns the common pool, respecting user settings |
3278 |
* specified via system properties. |
3279 |
*/ |
3280 |
private static ForkJoinPool makeCommonPool() { |
3281 |
int parallelism = -1; |
3282 |
ForkJoinWorkerThreadFactory factory |
3283 |
= defaultForkJoinWorkerThreadFactory; |
3284 |
UncaughtExceptionHandler handler = null; |
3285 |
try { // ignore exceptions in accesing/parsing properties |
3286 |
String pp = System.getProperty |
3287 |
("java.util.concurrent.ForkJoinPool.common.parallelism"); |
3288 |
String fp = System.getProperty |
3289 |
("java.util.concurrent.ForkJoinPool.common.threadFactory"); |
3290 |
String hp = System.getProperty |
3291 |
("java.util.concurrent.ForkJoinPool.common.exceptionHandler"); |
3292 |
if (pp != null) |
3293 |
parallelism = Integer.parseInt(pp); |
3294 |
if (fp != null) |
3295 |
factory = ((ForkJoinWorkerThreadFactory)ClassLoader. |
3296 |
getSystemClassLoader().loadClass(fp).newInstance()); |
3297 |
if (hp != null) |
3298 |
handler = ((UncaughtExceptionHandler)ClassLoader. |
3299 |
getSystemClassLoader().loadClass(hp).newInstance()); |
3300 |
} catch (Exception ignore) { |
3301 |
} |
3302 |
|
3303 |
if (parallelism < 0 && // default 1 less than #cores |
3304 |
(parallelism = Runtime.getRuntime().availableProcessors() - 1) < 0) |
3305 |
parallelism = 0; |
3306 |
if (parallelism > MAX_CAP) |
3307 |
parallelism = MAX_CAP; |
3308 |
return new ForkJoinPool(parallelism, factory, handler, LIFO_QUEUE, |
3309 |
"ForkJoinPool.commonPool-worker-"); |
3310 |
} |
3311 |
|
3312 |
/** |
3313 |
* Returns a sun.misc.Unsafe. Suitable for use in a 3rd party package. |
3314 |
* Replace with a simple call to Unsafe.getUnsafe when integrating |
3315 |
* into a jdk. |
3316 |
* |
3317 |
* @return a sun.misc.Unsafe |
3318 |
*/ |
3319 |
private static sun.misc.Unsafe getUnsafe() { |
3320 |
try { |
3321 |
return sun.misc.Unsafe.getUnsafe(); |
3322 |
} catch (SecurityException tryReflectionInstead) {} |
3323 |
try { |
3324 |
return java.security.AccessController.doPrivileged |
3325 |
(new java.security.PrivilegedExceptionAction<sun.misc.Unsafe>() { |
3326 |
public sun.misc.Unsafe run() throws Exception { |
3327 |
Class<sun.misc.Unsafe> k = sun.misc.Unsafe.class; |
3328 |
for (java.lang.reflect.Field f : k.getDeclaredFields()) { |
3329 |
f.setAccessible(true); |
3330 |
Object x = f.get(null); |
3331 |
if (k.isInstance(x)) |
3332 |
return k.cast(x); |
3333 |
} |
3334 |
throw new NoSuchFieldError("the Unsafe"); |
3335 |
}}); |
3336 |
} catch (java.security.PrivilegedActionException e) { |
3337 |
throw new RuntimeException("Could not initialize intrinsics", |
3338 |
e.getCause()); |
3339 |
} |
3340 |
} |
3341 |
} |