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.util.ArrayList; |
10 |
import java.util.Arrays; |
11 |
import java.util.Collection; |
12 |
import java.util.Collections; |
13 |
import java.util.List; |
14 |
import java.util.Random; |
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 |
import java.util.concurrent.atomic.AtomicInteger; |
24 |
import java.util.concurrent.atomic.AtomicLong; |
25 |
import java.util.concurrent.locks.AbstractQueuedSynchronizer; |
26 |
import java.util.concurrent.locks.Condition; |
27 |
|
28 |
/** |
29 |
* An {@link ExecutorService} for running {@link ForkJoinTask}s. |
30 |
* A {@code ForkJoinPool} provides the entry point for submissions |
31 |
* from non-{@code ForkJoinTask} clients, as well as management and |
32 |
* monitoring operations. |
33 |
* |
34 |
* <p>A {@code ForkJoinPool} differs from other kinds of {@link |
35 |
* ExecutorService} mainly by virtue of employing |
36 |
* <em>work-stealing</em>: all threads in the pool attempt to find and |
37 |
* execute tasks submitted to the pool and/or created by other active |
38 |
* tasks (eventually blocking waiting for work if none exist). This |
39 |
* enables efficient processing when most tasks spawn other subtasks |
40 |
* (as do most {@code ForkJoinTask}s), as well as when many small |
41 |
* tasks are submitted to the pool from external clients. Especially |
42 |
* when setting <em>asyncMode</em> to true in constructors, {@code |
43 |
* ForkJoinPool}s may also be appropriate for use with event-style |
44 |
* tasks that are never joined. |
45 |
* |
46 |
* <p>A static {@link #commonPool} is available and appropriate for |
47 |
* most applications. The common pool is constructed upon first |
48 |
* access, or upon usage by any ForkJoinTask that is not explictly |
49 |
* submitted to a specified pool. Using the common pool normally |
50 |
* reduces resource usage (its threads are slowly reclaimed during |
51 |
* periods of non-use, and reinstated upon subsequent use). The |
52 |
* common pool is by default constructed with default parameters, but |
53 |
* these may be controlled by setting any or all of the three |
54 |
* properties {@code |
55 |
* java.util.concurrent.ForkJoinPool.common.{parallelism, |
56 |
* threadFactory, exceptionHandler}}. |
57 |
* |
58 |
* <p>For applications that require separate or custom pools, a {@code |
59 |
* ForkJoinPool} may be constructed with a given target parallelism |
60 |
* level; by default, equal to the number of available processors. The |
61 |
* pool attempts to maintain enough active (or available) threads by |
62 |
* dynamically adding, suspending, or resuming internal worker |
63 |
* threads, even if some tasks are stalled waiting to join |
64 |
* others. However, no such adjustments are guaranteed in the face of |
65 |
* blocked IO or other unmanaged synchronization. The nested {@link |
66 |
* ManagedBlocker} interface enables extension of the kinds of |
67 |
* synchronization accommodated. |
68 |
* |
69 |
* <p>In addition to execution and lifecycle control methods, this |
70 |
* class provides status check methods (for example |
71 |
* {@link #getStealCount}) that are intended to aid in developing, |
72 |
* tuning, and monitoring fork/join applications. Also, method |
73 |
* {@link #toString} returns indications of pool state in a |
74 |
* convenient form for informal monitoring. |
75 |
* |
76 |
* <p> As is the case with other ExecutorServices, there are three |
77 |
* main task execution methods summarized in the following table. |
78 |
* These are designed to be used primarily by clients not already |
79 |
* engaged in fork/join computations in the current pool. The main |
80 |
* forms of these methods accept instances of {@code ForkJoinTask}, |
81 |
* but overloaded forms also allow mixed execution of plain {@code |
82 |
* Runnable}- or {@code Callable}- based activities as well. However, |
83 |
* tasks that are already executing in a pool should normally instead |
84 |
* use the within-computation forms listed in the table unless using |
85 |
* async event-style tasks that are not usually joined, in which case |
86 |
* there is little difference among choice of methods. |
87 |
* |
88 |
* <table BORDER CELLPADDING=3 CELLSPACING=1> |
89 |
* <tr> |
90 |
* <td></td> |
91 |
* <td ALIGN=CENTER> <b>Call from non-fork/join clients</b></td> |
92 |
* <td ALIGN=CENTER> <b>Call from within fork/join computations</b></td> |
93 |
* </tr> |
94 |
* <tr> |
95 |
* <td> <b>Arrange async execution</td> |
96 |
* <td> {@link #execute(ForkJoinTask)}</td> |
97 |
* <td> {@link ForkJoinTask#fork}</td> |
98 |
* </tr> |
99 |
* <tr> |
100 |
* <td> <b>Await and obtain result</td> |
101 |
* <td> {@link #invoke(ForkJoinTask)}</td> |
102 |
* <td> {@link ForkJoinTask#invoke}</td> |
103 |
* </tr> |
104 |
* <tr> |
105 |
* <td> <b>Arrange exec and obtain Future</td> |
106 |
* <td> {@link #submit(ForkJoinTask)}</td> |
107 |
* <td> {@link ForkJoinTask#fork} (ForkJoinTasks <em>are</em> Futures)</td> |
108 |
* </tr> |
109 |
* </table> |
110 |
* |
111 |
* <p><b>Implementation notes</b>: This implementation restricts the |
112 |
* maximum number of running threads to 32767. Attempts to create |
113 |
* pools with greater than the maximum number result in |
114 |
* {@code IllegalArgumentException}. |
115 |
* |
116 |
* <p>This implementation rejects submitted tasks (that is, by throwing |
117 |
* {@link RejectedExecutionException}) only when the pool is shut down |
118 |
* or internal resources have been exhausted. |
119 |
* |
120 |
* @since 1.7 |
121 |
* @author Doug Lea |
122 |
*/ |
123 |
public class ForkJoinPool extends AbstractExecutorService { |
124 |
|
125 |
/* |
126 |
* Implementation Overview |
127 |
* |
128 |
* This class and its nested classes provide the main |
129 |
* functionality and control for a set of worker threads: |
130 |
* Submissions from non-FJ threads enter into submission queues. |
131 |
* Workers take these tasks and typically split them into subtasks |
132 |
* that may be stolen by other workers. Preference rules give |
133 |
* first priority to processing tasks from their own queues (LIFO |
134 |
* or FIFO, depending on mode), then to randomized FIFO steals of |
135 |
* tasks in other queues. |
136 |
* |
137 |
* WorkQueues |
138 |
* ========== |
139 |
* |
140 |
* Most operations occur within work-stealing queues (in nested |
141 |
* class WorkQueue). These are special forms of Deques that |
142 |
* support only three of the four possible end-operations -- push, |
143 |
* pop, and poll (aka steal), under the further constraints that |
144 |
* push and pop are called only from the owning thread (or, as |
145 |
* extended here, under a lock), while poll may be called from |
146 |
* other threads. (If you are unfamiliar with them, you probably |
147 |
* want to read Herlihy and Shavit's book "The Art of |
148 |
* Multiprocessor programming", chapter 16 describing these in |
149 |
* more detail before proceeding.) The main work-stealing queue |
150 |
* design is roughly similar to those in the papers "Dynamic |
151 |
* Circular Work-Stealing Deque" by Chase and Lev, SPAA 2005 |
152 |
* (http://research.sun.com/scalable/pubs/index.html) and |
153 |
* "Idempotent work stealing" by Michael, Saraswat, and Vechev, |
154 |
* PPoPP 2009 (http://portal.acm.org/citation.cfm?id=1504186). |
155 |
* The main differences ultimately stem from GC requirements that |
156 |
* we null out taken slots as soon as we can, to maintain as small |
157 |
* a footprint as possible even in programs generating huge |
158 |
* numbers of tasks. To accomplish this, we shift the CAS |
159 |
* arbitrating pop vs poll (steal) from being on the indices |
160 |
* ("base" and "top") to the slots themselves. So, both a |
161 |
* successful pop and poll mainly entail a CAS of a slot from |
162 |
* non-null to null. Because we rely on CASes of references, we |
163 |
* do not need tag bits on base or top. They are simple ints as |
164 |
* used in any circular array-based queue (see for example |
165 |
* ArrayDeque). Updates to the indices must still be ordered in a |
166 |
* way that guarantees that top == base means the queue is empty, |
167 |
* but otherwise may err on the side of possibly making the queue |
168 |
* appear nonempty when a push, pop, or poll have not fully |
169 |
* committed. Note that this means that the poll operation, |
170 |
* considered individually, is not wait-free. One thief cannot |
171 |
* successfully continue until another in-progress one (or, if |
172 |
* previously empty, a push) completes. However, in the |
173 |
* aggregate, we ensure at least probabilistic non-blockingness. |
174 |
* If an attempted steal fails, a thief always chooses a different |
175 |
* random victim target to try next. So, in order for one thief to |
176 |
* progress, it suffices for any in-progress poll or new push on |
177 |
* any empty queue to complete. (This is why we normally use |
178 |
* method pollAt and its variants that try once at the apparent |
179 |
* base index, else consider alternative actions, rather than |
180 |
* method poll.) |
181 |
* |
182 |
* This approach also enables support of a user mode in which local |
183 |
* task processing is in FIFO, not LIFO order, simply by using |
184 |
* poll rather than pop. This can be useful in message-passing |
185 |
* frameworks in which tasks are never joined. However neither |
186 |
* mode considers affinities, loads, cache localities, etc, so |
187 |
* rarely provide the best possible performance on a given |
188 |
* machine, but portably provide good throughput by averaging over |
189 |
* these factors. (Further, even if we did try to use such |
190 |
* information, we do not usually have a basis for exploiting it. |
191 |
* For example, some sets of tasks profit from cache affinities, |
192 |
* but others are harmed by cache pollution effects.) |
193 |
* |
194 |
* WorkQueues are also used in a similar way for tasks submitted |
195 |
* to the pool. We cannot mix these tasks in the same queues used |
196 |
* for work-stealing (this would contaminate lifo/fifo |
197 |
* processing). Instead, we loosely associate submission queues |
198 |
* with submitting threads, using a form of hashing. The |
199 |
* ThreadLocal Submitter class contains a value initially used as |
200 |
* a hash code for choosing existing queues, but may be randomly |
201 |
* repositioned upon contention with other submitters. In |
202 |
* essence, submitters act like workers except that they never |
203 |
* take tasks, and they are multiplexed on to a finite number of |
204 |
* shared work queues. However, classes are set up so that future |
205 |
* extensions could allow submitters to optionally help perform |
206 |
* tasks as well. Insertion of tasks in shared mode requires a |
207 |
* lock (mainly to protect in the case of resizing) but we use |
208 |
* only a simple spinlock (using bits in field runState), because |
209 |
* submitters encountering a busy queue move on to try or create |
210 |
* other queues -- they block only when creating and registering |
211 |
* new queues. |
212 |
* |
213 |
* Management |
214 |
* ========== |
215 |
* |
216 |
* The main throughput advantages of work-stealing stem from |
217 |
* decentralized control -- workers mostly take tasks from |
218 |
* themselves or each other. We cannot negate this in the |
219 |
* implementation of other management responsibilities. The main |
220 |
* tactic for avoiding bottlenecks is packing nearly all |
221 |
* essentially atomic control state into two volatile variables |
222 |
* that are by far most often read (not written) as status and |
223 |
* consistency checks. |
224 |
* |
225 |
* Field "ctl" contains 64 bits holding all the information needed |
226 |
* to atomically decide to add, inactivate, enqueue (on an event |
227 |
* queue), dequeue, and/or re-activate workers. To enable this |
228 |
* packing, we restrict maximum parallelism to (1<<15)-1 (which is |
229 |
* far in excess of normal operating range) to allow ids, counts, |
230 |
* and their negations (used for thresholding) to fit into 16bit |
231 |
* fields. |
232 |
* |
233 |
* Field "runState" contains 32 bits needed to register and |
234 |
* deregister WorkQueues, as well as to enable shutdown. It is |
235 |
* only modified under a lock (normally briefly held, but |
236 |
* occasionally protecting allocations and resizings) but even |
237 |
* when locked remains available to check consistency. |
238 |
* |
239 |
* Recording WorkQueues. WorkQueues are recorded in the |
240 |
* "workQueues" array that is created upon pool construction and |
241 |
* expanded if necessary. Updates to the array while recording |
242 |
* new workers and unrecording terminated ones are protected from |
243 |
* each other by a lock but the array is otherwise concurrently |
244 |
* readable, and accessed directly. To simplify index-based |
245 |
* operations, the array size is always a power of two, and all |
246 |
* readers must tolerate null slots. Shared (submission) queues |
247 |
* are at even indices, worker queues at odd indices. Grouping |
248 |
* them together in this way simplifies and speeds up task |
249 |
* scanning. |
250 |
* |
251 |
* All worker thread creation is on-demand, triggered by task |
252 |
* submissions, replacement of terminated workers, and/or |
253 |
* compensation for blocked workers. However, all other support |
254 |
* code is set up to work with other policies. To ensure that we |
255 |
* do not hold on to worker references that would prevent GC, ALL |
256 |
* accesses to workQueues are via indices into the workQueues |
257 |
* array (which is one source of some of the messy code |
258 |
* constructions here). In essence, the workQueues array serves as |
259 |
* a weak reference mechanism. Thus for example the wait queue |
260 |
* field of ctl stores indices, not references. Access to the |
261 |
* workQueues in associated methods (for example signalWork) must |
262 |
* both index-check and null-check the IDs. All such accesses |
263 |
* ignore bad IDs by returning out early from what they are doing, |
264 |
* since this can only be associated with termination, in which |
265 |
* case it is OK to give up. All uses of the workQueues array |
266 |
* also check that it is non-null (even if previously |
267 |
* non-null). This allows nulling during termination, which is |
268 |
* currently not necessary, but remains an option for |
269 |
* resource-revocation-based shutdown schemes. It also helps |
270 |
* reduce JIT issuance of uncommon-trap code, which tends to |
271 |
* unnecessarily complicate control flow in some methods. |
272 |
* |
273 |
* Event Queuing. Unlike HPC work-stealing frameworks, we cannot |
274 |
* let workers spin indefinitely scanning for tasks when none can |
275 |
* be found immediately, and we cannot start/resume workers unless |
276 |
* there appear to be tasks available. On the other hand, we must |
277 |
* quickly prod them into action when new tasks are submitted or |
278 |
* generated. In many usages, ramp-up time to activate workers is |
279 |
* the main limiting factor in overall performance (this is |
280 |
* compounded at program start-up by JIT compilation and |
281 |
* allocation). So we try to streamline this as much as possible. |
282 |
* We park/unpark workers after placing in an event wait queue |
283 |
* when they cannot find work. This "queue" is actually a simple |
284 |
* Treiber stack, headed by the "id" field of ctl, plus a 15bit |
285 |
* counter value (that reflects the number of times a worker has |
286 |
* been inactivated) to avoid ABA effects (we need only as many |
287 |
* version numbers as worker threads). Successors are held in |
288 |
* field WorkQueue.nextWait. Queuing deals with several intrinsic |
289 |
* races, mainly that a task-producing thread can miss seeing (and |
290 |
* signalling) another thread that gave up looking for work but |
291 |
* has not yet entered the wait queue. We solve this by requiring |
292 |
* a full sweep of all workers (via repeated calls to method |
293 |
* scan()) both before and after a newly waiting worker is added |
294 |
* to the wait queue. During a rescan, the worker might release |
295 |
* some other queued worker rather than itself, which has the same |
296 |
* net effect. Because enqueued workers may actually be rescanning |
297 |
* rather than waiting, we set and clear the "parker" field of |
298 |
* WorkQueues to reduce unnecessary calls to unpark. (This |
299 |
* requires a secondary recheck to avoid missed signals.) Note |
300 |
* the unusual conventions about Thread.interrupts surrounding |
301 |
* parking and other blocking: Because interrupts are used solely |
302 |
* to alert threads to check termination, which is checked anyway |
303 |
* upon blocking, we clear status (using Thread.interrupted) |
304 |
* before any call to park, so that park does not immediately |
305 |
* return due to status being set via some other unrelated call to |
306 |
* interrupt in user code. |
307 |
* |
308 |
* Signalling. We create or wake up workers only when there |
309 |
* appears to be at least one task they might be able to find and |
310 |
* execute. When a submission is added or another worker adds a |
311 |
* task to a queue that previously had fewer than two tasks, they |
312 |
* signal waiting workers (or trigger creation of new ones if |
313 |
* fewer than the given parallelism level -- see signalWork). |
314 |
* These primary signals are buttressed by signals during rescans; |
315 |
* together these cover the signals needed in cases when more |
316 |
* tasks are pushed but untaken, and improve performance compared |
317 |
* to having one thread wake up all workers. |
318 |
* |
319 |
* Trimming workers. To release resources after periods of lack of |
320 |
* use, a worker starting to wait when the pool is quiescent will |
321 |
* time out and terminate if the pool has remained quiescent for a |
322 |
* given period -- a short period if there are more threads than |
323 |
* parallelism, longer as the number of threads decreases. This |
324 |
* will slowly propagate, eventually terminating all workers after |
325 |
* periods of non-use. |
326 |
* |
327 |
* Shutdown and Termination. A call to shutdownNow atomically sets |
328 |
* a runState bit and then (non-atomically) sets each worker's |
329 |
* runState status, cancels all unprocessed tasks, and wakes up |
330 |
* all waiting workers. Detecting whether termination should |
331 |
* commence after a non-abrupt shutdown() call requires more work |
332 |
* and bookkeeping. We need consensus about quiescence (i.e., that |
333 |
* there is no more work). The active count provides a primary |
334 |
* indication but non-abrupt shutdown still requires a rechecking |
335 |
* scan for any workers that are inactive but not queued. |
336 |
* |
337 |
* Joining Tasks |
338 |
* ============= |
339 |
* |
340 |
* Any of several actions may be taken when one worker is waiting |
341 |
* to join a task stolen (or always held) by another. Because we |
342 |
* are multiplexing many tasks on to a pool of workers, we can't |
343 |
* just let them block (as in Thread.join). We also cannot just |
344 |
* reassign the joiner's run-time stack with another and replace |
345 |
* it later, which would be a form of "continuation", that even if |
346 |
* possible is not necessarily a good idea since we sometimes need |
347 |
* both an unblocked task and its continuation to progress. |
348 |
* Instead we combine two tactics: |
349 |
* |
350 |
* Helping: Arranging for the joiner to execute some task that it |
351 |
* would be running if the steal had not occurred. |
352 |
* |
353 |
* Compensating: Unless there are already enough live threads, |
354 |
* method tryCompensate() may create or re-activate a spare |
355 |
* thread to compensate for blocked joiners until they unblock. |
356 |
* |
357 |
* A third form (implemented in tryRemoveAndExec and |
358 |
* tryPollForAndExec) amounts to helping a hypothetical |
359 |
* compensator: If we can readily tell that a possible action of a |
360 |
* compensator is to steal and execute the task being joined, the |
361 |
* joining thread can do so directly, without the need for a |
362 |
* compensation thread (although at the expense of larger run-time |
363 |
* stacks, but the tradeoff is typically worthwhile). |
364 |
* |
365 |
* The ManagedBlocker extension API can't use helping so relies |
366 |
* only on compensation in method awaitBlocker. |
367 |
* |
368 |
* The algorithm in tryHelpStealer entails a form of "linear" |
369 |
* helping: Each worker records (in field currentSteal) the most |
370 |
* recent task it stole from some other worker. Plus, it records |
371 |
* (in field currentJoin) the task it is currently actively |
372 |
* joining. Method tryHelpStealer uses these markers to try to |
373 |
* find a worker to help (i.e., steal back a task from and execute |
374 |
* it) that could hasten completion of the actively joined task. |
375 |
* In essence, the joiner executes a task that would be on its own |
376 |
* local deque had the to-be-joined task not been stolen. This may |
377 |
* be seen as a conservative variant of the approach in Wagner & |
378 |
* Calder "Leapfrogging: a portable technique for implementing |
379 |
* efficient futures" SIGPLAN Notices, 1993 |
380 |
* (http://portal.acm.org/citation.cfm?id=155354). It differs in |
381 |
* that: (1) We only maintain dependency links across workers upon |
382 |
* steals, rather than use per-task bookkeeping. This sometimes |
383 |
* requires a linear scan of workQueues array to locate stealers, |
384 |
* but often doesn't because stealers leave hints (that may become |
385 |
* stale/wrong) of where to locate them. A stealHint is only a |
386 |
* hint because a worker might have had multiple steals and the |
387 |
* hint records only one of them (usually the most current). |
388 |
* Hinting isolates cost to when it is needed, rather than adding |
389 |
* to per-task overhead. (2) It is "shallow", ignoring nesting |
390 |
* and potentially cyclic mutual steals. (3) It is intentionally |
391 |
* racy: field currentJoin is updated only while actively joining, |
392 |
* which means that we miss links in the chain during long-lived |
393 |
* tasks, GC stalls etc (which is OK since blocking in such cases |
394 |
* is usually a good idea). (4) We bound the number of attempts |
395 |
* to find work (see MAX_HELP) and fall back to suspending the |
396 |
* worker and if necessary replacing it with another. |
397 |
* |
398 |
* It is impossible to keep exactly the target parallelism number |
399 |
* of threads running at any given time. Determining the |
400 |
* existence of conservatively safe helping targets, the |
401 |
* availability of already-created spares, and the apparent need |
402 |
* to create new spares are all racy, so we rely on multiple |
403 |
* retries of each. Compensation in the apparent absence of |
404 |
* helping opportunities is challenging to control on JVMs, where |
405 |
* GC and other activities can stall progress of tasks that in |
406 |
* turn stall out many other dependent tasks, without us being |
407 |
* able to determine whether they will ever require compensation. |
408 |
* Even though work-stealing otherwise encounters little |
409 |
* degradation in the presence of more threads than cores, |
410 |
* aggressively adding new threads in such cases entails risk of |
411 |
* unwanted positive feedback control loops in which more threads |
412 |
* cause more dependent stalls (as well as delayed progress of |
413 |
* unblocked threads to the point that we know they are available) |
414 |
* leading to more situations requiring more threads, and so |
415 |
* on. This aspect of control can be seen as an (analytically |
416 |
* intractable) game with an opponent that may choose the worst |
417 |
* (for us) active thread to stall at any time. We take several |
418 |
* precautions to bound losses (and thus bound gains), mainly in |
419 |
* methods tryCompensate and awaitJoin: (1) We only try |
420 |
* compensation after attempting enough helping steps (measured |
421 |
* via counting and timing) that we have already consumed the |
422 |
* estimated cost of creating and activating a new thread. (2) We |
423 |
* allow up to 50% of threads to be blocked before initially |
424 |
* adding any others, and unless completely saturated, check that |
425 |
* some work is available for a new worker before adding. Also, we |
426 |
* create up to only 50% more threads until entering a mode that |
427 |
* only adds a thread if all others are possibly blocked. All |
428 |
* together, this means that we might be half as fast to react, |
429 |
* and create half as many threads as possible in the ideal case, |
430 |
* but present vastly fewer anomalies in all other cases compared |
431 |
* to both more aggressive and more conservative alternatives. |
432 |
* |
433 |
* Style notes: There is a lot of representation-level coupling |
434 |
* among classes ForkJoinPool, ForkJoinWorkerThread, and |
435 |
* ForkJoinTask. The fields of WorkQueue maintain data structures |
436 |
* managed by ForkJoinPool, so are directly accessed. There is |
437 |
* little point trying to reduce this, since any associated future |
438 |
* changes in representations will need to be accompanied by |
439 |
* algorithmic changes anyway. Several methods intrinsically |
440 |
* sprawl because they must accumulate sets of consistent reads of |
441 |
* volatiles held in local variables. Methods signalWork() and |
442 |
* scan() are the main bottlenecks, so are especially heavily |
443 |
* micro-optimized/mangled. There are lots of inline assignments |
444 |
* (of form "while ((local = field) != 0)") which are usually the |
445 |
* simplest way to ensure the required read orderings (which are |
446 |
* sometimes critical). This leads to a "C"-like style of listing |
447 |
* declarations of these locals at the heads of methods or blocks. |
448 |
* There are several occurrences of the unusual "do {} while |
449 |
* (!cas...)" which is the simplest way to force an update of a |
450 |
* CAS'ed variable. There are also other coding oddities that help |
451 |
* some methods perform reasonably even when interpreted (not |
452 |
* compiled). |
453 |
* |
454 |
* The order of declarations in this file is: |
455 |
* (1) Static utility functions |
456 |
* (2) Nested (static) classes |
457 |
* (3) Static fields |
458 |
* (4) Fields, along with constants used when unpacking some of them |
459 |
* (5) Internal control methods |
460 |
* (6) Callbacks and other support for ForkJoinTask methods |
461 |
* (7) Exported methods |
462 |
* (8) Static block initializing statics in minimally dependent order |
463 |
*/ |
464 |
|
465 |
// Static utilities |
466 |
|
467 |
/** |
468 |
* If there is a security manager, makes sure caller has |
469 |
* permission to modify threads. |
470 |
*/ |
471 |
private static void checkPermission() { |
472 |
SecurityManager security = System.getSecurityManager(); |
473 |
if (security != null) |
474 |
security.checkPermission(modifyThreadPermission); |
475 |
} |
476 |
|
477 |
// Nested classes |
478 |
|
479 |
/** |
480 |
* Factory for creating new {@link ForkJoinWorkerThread}s. |
481 |
* A {@code ForkJoinWorkerThreadFactory} must be defined and used |
482 |
* for {@code ForkJoinWorkerThread} subclasses that extend base |
483 |
* functionality or initialize threads with different contexts. |
484 |
*/ |
485 |
public static interface ForkJoinWorkerThreadFactory { |
486 |
/** |
487 |
* Returns a new worker thread operating in the given pool. |
488 |
* |
489 |
* @param pool the pool this thread works in |
490 |
* @throws NullPointerException if the pool is null |
491 |
*/ |
492 |
public ForkJoinWorkerThread newThread(ForkJoinPool pool); |
493 |
} |
494 |
|
495 |
/** |
496 |
* Default ForkJoinWorkerThreadFactory implementation; creates a |
497 |
* new ForkJoinWorkerThread. |
498 |
*/ |
499 |
static class DefaultForkJoinWorkerThreadFactory |
500 |
implements ForkJoinWorkerThreadFactory { |
501 |
public ForkJoinWorkerThread newThread(ForkJoinPool pool) { |
502 |
return new ForkJoinWorkerThread(pool); |
503 |
} |
504 |
} |
505 |
|
506 |
/** |
507 |
* A simple non-reentrant lock used for exclusion when managing |
508 |
* queues and workers. We use a custom lock so that we can readily |
509 |
* probe lock state in constructions that check among alternative |
510 |
* actions. The lock is normally only very briefly held, and |
511 |
* sometimes treated as a spinlock, but other usages block to |
512 |
* reduce overall contention in those cases where locked code |
513 |
* bodies perform allocation/resizing. |
514 |
*/ |
515 |
static final class Mutex extends AbstractQueuedSynchronizer { |
516 |
public final boolean tryAcquire(int ignore) { |
517 |
return compareAndSetState(0, 1); |
518 |
} |
519 |
public final boolean tryRelease(int ignore) { |
520 |
setState(0); |
521 |
return true; |
522 |
} |
523 |
public final void lock() { acquire(0); } |
524 |
public final void unlock() { release(0); } |
525 |
public final boolean isHeldExclusively() { return getState() == 1; } |
526 |
public final Condition newCondition() { return new ConditionObject(); } |
527 |
} |
528 |
|
529 |
/** |
530 |
* Class for artificial tasks that are used to replace the target |
531 |
* of local joins if they are removed from an interior queue slot |
532 |
* in WorkQueue.tryRemoveAndExec. We don't need the proxy to |
533 |
* actually do anything beyond having a unique identity. |
534 |
*/ |
535 |
static final class EmptyTask extends ForkJoinTask<Void> { |
536 |
EmptyTask() { status = ForkJoinTask.NORMAL; } // force done |
537 |
public final Void getRawResult() { return null; } |
538 |
public final void setRawResult(Void x) {} |
539 |
public final boolean exec() { return true; } |
540 |
} |
541 |
|
542 |
/** |
543 |
* Queues supporting work-stealing as well as external task |
544 |
* submission. See above for main rationale and algorithms. |
545 |
* Implementation relies heavily on "Unsafe" intrinsics |
546 |
* and selective use of "volatile": |
547 |
* |
548 |
* Field "base" is the index (mod array.length) of the least valid |
549 |
* queue slot, which is always the next position to steal (poll) |
550 |
* from if nonempty. Reads and writes require volatile orderings |
551 |
* but not CAS, because updates are only performed after slot |
552 |
* CASes. |
553 |
* |
554 |
* Field "top" is the index (mod array.length) of the next queue |
555 |
* slot to push to or pop from. It is written only by owner thread |
556 |
* for push, or under lock for trySharedPush, and accessed by |
557 |
* other threads only after reading (volatile) base. Both top and |
558 |
* base are allowed to wrap around on overflow, but (top - base) |
559 |
* (or more commonly -(base - top) to force volatile read of base |
560 |
* before top) still estimates size. |
561 |
* |
562 |
* The array slots are read and written using the emulation of |
563 |
* volatiles/atomics provided by Unsafe. Insertions must in |
564 |
* general use putOrderedObject as a form of releasing store to |
565 |
* ensure that all writes to the task object are ordered before |
566 |
* its publication in the queue. (Although we can avoid one case |
567 |
* of this when locked in trySharedPush.) All removals entail a |
568 |
* CAS to null. The array is always a power of two. To ensure |
569 |
* safety of Unsafe array operations, all accesses perform |
570 |
* explicit null checks and implicit bounds checks via |
571 |
* power-of-two masking. |
572 |
* |
573 |
* In addition to basic queuing support, this class contains |
574 |
* fields described elsewhere to control execution. It turns out |
575 |
* to work better memory-layout-wise to include them in this |
576 |
* class rather than a separate class. |
577 |
* |
578 |
* Performance on most platforms is very sensitive to placement of |
579 |
* instances of both WorkQueues and their arrays -- we absolutely |
580 |
* do not want multiple WorkQueue instances or multiple queue |
581 |
* arrays sharing cache lines. (It would be best for queue objects |
582 |
* and their arrays to share, but there is nothing available to |
583 |
* help arrange that). Unfortunately, because they are recorded |
584 |
* in a common array, WorkQueue instances are often moved to be |
585 |
* adjacent by garbage collectors. To reduce impact, we use field |
586 |
* padding that works OK on common platforms; this effectively |
587 |
* trades off slightly slower average field access for the sake of |
588 |
* avoiding really bad worst-case access. (Until better JVM |
589 |
* support is in place, this padding is dependent on transient |
590 |
* properties of JVM field layout rules.) We also take care in |
591 |
* allocating, sizing and resizing the array. Non-shared queue |
592 |
* arrays are initialized (via method growArray) by workers before |
593 |
* use. Others are allocated on first use. |
594 |
*/ |
595 |
static final class WorkQueue { |
596 |
/** |
597 |
* Capacity of work-stealing queue array upon initialization. |
598 |
* Must be a power of two; at least 4, but should be larger to |
599 |
* reduce or eliminate cacheline sharing among queues. |
600 |
* Currently, it is much larger, as a partial workaround for |
601 |
* the fact that JVMs often place arrays in locations that |
602 |
* share GC bookkeeping (especially cardmarks) such that |
603 |
* per-write accesses encounter serious memory contention. |
604 |
*/ |
605 |
static final int INITIAL_QUEUE_CAPACITY = 1 << 13; |
606 |
|
607 |
/** |
608 |
* Maximum size for queue arrays. Must be a power of two less |
609 |
* than or equal to 1 << (31 - width of array entry) to ensure |
610 |
* lack of wraparound of index calculations, but defined to a |
611 |
* value a bit less than this to help users trap runaway |
612 |
* programs before saturating systems. |
613 |
*/ |
614 |
static final int MAXIMUM_QUEUE_CAPACITY = 1 << 26; // 64M |
615 |
|
616 |
volatile long totalSteals; // cumulative number of steals |
617 |
int seed; // for random scanning; initialize nonzero |
618 |
volatile int eventCount; // encoded inactivation count; < 0 if inactive |
619 |
int nextWait; // encoded record of next event waiter |
620 |
int rescans; // remaining scans until block |
621 |
int nsteals; // top-level task executions since last idle |
622 |
final int mode; // lifo, fifo, or shared |
623 |
int poolIndex; // index of this queue in pool (or 0) |
624 |
int stealHint; // index of most recent known stealer |
625 |
volatile int runState; // 1: locked, -1: terminate; else 0 |
626 |
volatile int base; // index of next slot for poll |
627 |
int top; // index of next slot for push |
628 |
ForkJoinTask<?>[] array; // the elements (initially unallocated) |
629 |
final ForkJoinPool pool; // the containing pool (may be null) |
630 |
final ForkJoinWorkerThread owner; // owning thread or null if shared |
631 |
volatile Thread parker; // == owner during call to park; else null |
632 |
volatile ForkJoinTask<?> currentJoin; // task being joined in awaitJoin |
633 |
ForkJoinTask<?> currentSteal; // current non-local task being executed |
634 |
// Heuristic padding to ameliorate unfortunate memory placements |
635 |
Object p00, p01, p02, p03, p04, p05, p06, p07; |
636 |
Object p08, p09, p0a, p0b, p0c, p0d, p0e; |
637 |
|
638 |
WorkQueue(ForkJoinPool pool, ForkJoinWorkerThread owner, int mode) { |
639 |
this.mode = mode; |
640 |
this.pool = pool; |
641 |
this.owner = owner; |
642 |
// Place indices in the center of array (that is not yet allocated) |
643 |
base = top = INITIAL_QUEUE_CAPACITY >>> 1; |
644 |
} |
645 |
|
646 |
/** |
647 |
* Returns the approximate number of tasks in the queue. |
648 |
*/ |
649 |
final int queueSize() { |
650 |
int n = base - top; // non-owner callers must read base first |
651 |
return (n >= 0) ? 0 : -n; // ignore transient negative |
652 |
} |
653 |
|
654 |
/** |
655 |
* Provides a more accurate estimate of whether this queue has |
656 |
* any tasks than does queueSize, by checking whether a |
657 |
* near-empty queue has at least one unclaimed task. |
658 |
*/ |
659 |
final boolean isEmpty() { |
660 |
ForkJoinTask<?>[] a; int m, s; |
661 |
int n = base - (s = top); |
662 |
return (n >= 0 || |
663 |
(n == -1 && |
664 |
((a = array) == null || |
665 |
(m = a.length - 1) < 0 || |
666 |
U.getObjectVolatile |
667 |
(a, ((m & (s - 1)) << ASHIFT) + ABASE) == null))); |
668 |
} |
669 |
|
670 |
/** |
671 |
* Pushes a task. Call only by owner in unshared queues. |
672 |
* |
673 |
* @param task the task. Caller must ensure non-null. |
674 |
* @throw RejectedExecutionException if array cannot be resized |
675 |
*/ |
676 |
final void push(ForkJoinTask<?> task) { |
677 |
ForkJoinTask<?>[] a; ForkJoinPool p; |
678 |
int s = top, m, n; |
679 |
if ((a = array) != null) { // ignore if queue removed |
680 |
U.putOrderedObject |
681 |
(a, (((m = a.length - 1) & s) << ASHIFT) + ABASE, task); |
682 |
if ((n = (top = s + 1) - base) <= 2) { |
683 |
if ((p = pool) != null) |
684 |
p.signalWork(); |
685 |
} |
686 |
else if (n >= m) |
687 |
growArray(true); |
688 |
} |
689 |
} |
690 |
|
691 |
/** |
692 |
* Pushes a task if lock is free and array is either big |
693 |
* enough or can be resized to be big enough. |
694 |
* |
695 |
* @param task the task. Caller must ensure non-null. |
696 |
* @return true if submitted |
697 |
*/ |
698 |
final boolean trySharedPush(ForkJoinTask<?> task) { |
699 |
boolean submitted = false; |
700 |
if (runState == 0 && U.compareAndSwapInt(this, RUNSTATE, 0, 1)) { |
701 |
ForkJoinTask<?>[] a = array; |
702 |
int s = top; |
703 |
try { |
704 |
if ((a != null && a.length > s + 1 - base) || |
705 |
(a = growArray(false)) != null) { // must presize |
706 |
int j = (((a.length - 1) & s) << ASHIFT) + ABASE; |
707 |
U.putObject(a, (long)j, task); // don't need "ordered" |
708 |
top = s + 1; |
709 |
submitted = true; |
710 |
} |
711 |
} finally { |
712 |
runState = 0; // unlock |
713 |
} |
714 |
} |
715 |
return submitted; |
716 |
} |
717 |
|
718 |
/** |
719 |
* Takes next task, if one exists, in LIFO order. Call only |
720 |
* by owner in unshared queues. (We do not have a shared |
721 |
* version of this method because it is never needed.) |
722 |
*/ |
723 |
final ForkJoinTask<?> pop() { |
724 |
ForkJoinTask<?>[] a; ForkJoinTask<?> t; int m; |
725 |
if ((a = array) != null && (m = a.length - 1) >= 0) { |
726 |
for (int s; (s = top - 1) - base >= 0;) { |
727 |
long j = ((m & s) << ASHIFT) + ABASE; |
728 |
if ((t = (ForkJoinTask<?>)U.getObject(a, j)) == null) |
729 |
break; |
730 |
if (U.compareAndSwapObject(a, j, t, null)) { |
731 |
top = s; |
732 |
return t; |
733 |
} |
734 |
} |
735 |
} |
736 |
return null; |
737 |
} |
738 |
|
739 |
/** |
740 |
* Takes a task in FIFO order if b is base of queue and a task |
741 |
* can be claimed without contention. Specialized versions |
742 |
* appear in ForkJoinPool methods scan and tryHelpStealer. |
743 |
*/ |
744 |
final ForkJoinTask<?> pollAt(int b) { |
745 |
ForkJoinTask<?> t; ForkJoinTask<?>[] a; |
746 |
if ((a = array) != null) { |
747 |
int j = (((a.length - 1) & b) << ASHIFT) + ABASE; |
748 |
if ((t = (ForkJoinTask<?>)U.getObjectVolatile(a, j)) != null && |
749 |
base == b && |
750 |
U.compareAndSwapObject(a, j, t, null)) { |
751 |
base = b + 1; |
752 |
return t; |
753 |
} |
754 |
} |
755 |
return null; |
756 |
} |
757 |
|
758 |
/** |
759 |
* Takes next task, if one exists, in FIFO order. |
760 |
*/ |
761 |
final ForkJoinTask<?> poll() { |
762 |
ForkJoinTask<?>[] a; int b; ForkJoinTask<?> t; |
763 |
while ((b = base) - top < 0 && (a = array) != null) { |
764 |
int j = (((a.length - 1) & b) << ASHIFT) + ABASE; |
765 |
t = (ForkJoinTask<?>)U.getObjectVolatile(a, j); |
766 |
if (t != null) { |
767 |
if (base == b && |
768 |
U.compareAndSwapObject(a, j, t, null)) { |
769 |
base = b + 1; |
770 |
return t; |
771 |
} |
772 |
} |
773 |
else if (base == b) { |
774 |
if (b + 1 == top) |
775 |
break; |
776 |
Thread.yield(); // wait for lagging update |
777 |
} |
778 |
} |
779 |
return null; |
780 |
} |
781 |
|
782 |
/** |
783 |
* Takes next task, if one exists, in order specified by mode. |
784 |
*/ |
785 |
final ForkJoinTask<?> nextLocalTask() { |
786 |
return mode == 0 ? pop() : poll(); |
787 |
} |
788 |
|
789 |
/** |
790 |
* Returns next task, if one exists, in order specified by mode. |
791 |
*/ |
792 |
final ForkJoinTask<?> peek() { |
793 |
ForkJoinTask<?>[] a = array; int m; |
794 |
if (a == null || (m = a.length - 1) < 0) |
795 |
return null; |
796 |
int i = mode == 0 ? top - 1 : base; |
797 |
int j = ((i & m) << ASHIFT) + ABASE; |
798 |
return (ForkJoinTask<?>)U.getObjectVolatile(a, j); |
799 |
} |
800 |
|
801 |
/** |
802 |
* Pops the given task only if it is at the current top. |
803 |
*/ |
804 |
final boolean tryUnpush(ForkJoinTask<?> t) { |
805 |
ForkJoinTask<?>[] a; int s; |
806 |
if ((a = array) != null && (s = top) != base && |
807 |
U.compareAndSwapObject |
808 |
(a, (((a.length - 1) & --s) << ASHIFT) + ABASE, t, null)) { |
809 |
top = s; |
810 |
return true; |
811 |
} |
812 |
return false; |
813 |
} |
814 |
|
815 |
/** |
816 |
* Version of tryUnpush for shared queues; called by non-FJ |
817 |
* submitters. Conservatively fails to unpush if all workers |
818 |
* are active unless there are multiple tasks in queue. |
819 |
*/ |
820 |
final boolean trySharedUnpush(ForkJoinTask<?> task, ForkJoinPool p) { |
821 |
boolean success = false; |
822 |
if (task != null && top != base && runState == 0 && |
823 |
U.compareAndSwapInt(this, RUNSTATE, 0, 1)) { |
824 |
try { |
825 |
ForkJoinTask<?>[] a; int n, s; |
826 |
if ((a = array) != null && (n = (s = top) - base) > 0 && |
827 |
(n > 1 || p == null || (int)(p.ctl >> AC_SHIFT) < 0)) { |
828 |
int j = (((a.length - 1) & --s) << ASHIFT) + ABASE; |
829 |
if (U.getObjectVolatile(a, j) == task && |
830 |
U.compareAndSwapObject(a, j, task, null)) { |
831 |
top = s; |
832 |
success = true; |
833 |
} |
834 |
} |
835 |
} finally { |
836 |
runState = 0; // unlock |
837 |
} |
838 |
} |
839 |
return success; |
840 |
} |
841 |
|
842 |
/** |
843 |
* Polls the given task only if it is at the current base. |
844 |
*/ |
845 |
final boolean pollFor(ForkJoinTask<?> task) { |
846 |
ForkJoinTask<?>[] a; int b; |
847 |
if ((b = base) - top < 0 && (a = array) != null) { |
848 |
int j = (((a.length - 1) & b) << ASHIFT) + ABASE; |
849 |
if (U.getObjectVolatile(a, j) == task && base == b && |
850 |
U.compareAndSwapObject(a, j, task, null)) { |
851 |
base = b + 1; |
852 |
return true; |
853 |
} |
854 |
} |
855 |
return false; |
856 |
} |
857 |
|
858 |
/** |
859 |
* Initializes or doubles the capacity of array. Call either |
860 |
* by owner or with lock held -- it is OK for base, but not |
861 |
* top, to move while resizings are in progress. |
862 |
* |
863 |
* @param rejectOnFailure if true, throw exception if capacity |
864 |
* exceeded (relayed ultimately to user); else return null. |
865 |
*/ |
866 |
final ForkJoinTask<?>[] growArray(boolean rejectOnFailure) { |
867 |
ForkJoinTask<?>[] oldA = array; |
868 |
int size = oldA != null ? oldA.length << 1 : INITIAL_QUEUE_CAPACITY; |
869 |
if (size <= MAXIMUM_QUEUE_CAPACITY) { |
870 |
int oldMask, t, b; |
871 |
ForkJoinTask<?>[] a = array = new ForkJoinTask<?>[size]; |
872 |
if (oldA != null && (oldMask = oldA.length - 1) >= 0 && |
873 |
(t = top) - (b = base) > 0) { |
874 |
int mask = size - 1; |
875 |
do { |
876 |
ForkJoinTask<?> x; |
877 |
int oldj = ((b & oldMask) << ASHIFT) + ABASE; |
878 |
int j = ((b & mask) << ASHIFT) + ABASE; |
879 |
x = (ForkJoinTask<?>)U.getObjectVolatile(oldA, oldj); |
880 |
if (x != null && |
881 |
U.compareAndSwapObject(oldA, oldj, x, null)) |
882 |
U.putObjectVolatile(a, j, x); |
883 |
} while (++b != t); |
884 |
} |
885 |
return a; |
886 |
} |
887 |
else if (!rejectOnFailure) |
888 |
return null; |
889 |
else |
890 |
throw new RejectedExecutionException("Queue capacity exceeded"); |
891 |
} |
892 |
|
893 |
/** |
894 |
* Removes and cancels all known tasks, ignoring any exceptions. |
895 |
*/ |
896 |
final void cancelAll() { |
897 |
ForkJoinTask.cancelIgnoringExceptions(currentJoin); |
898 |
ForkJoinTask.cancelIgnoringExceptions(currentSteal); |
899 |
for (ForkJoinTask<?> t; (t = poll()) != null; ) |
900 |
ForkJoinTask.cancelIgnoringExceptions(t); |
901 |
} |
902 |
|
903 |
/** |
904 |
* Computes next value for random probes. Scans don't require |
905 |
* a very high quality generator, but also not a crummy one. |
906 |
* Marsaglia xor-shift is cheap and works well enough. Note: |
907 |
* This is manually inlined in its usages in ForkJoinPool to |
908 |
* avoid writes inside busy scan loops. |
909 |
*/ |
910 |
final int nextSeed() { |
911 |
int r = seed; |
912 |
r ^= r << 13; |
913 |
r ^= r >>> 17; |
914 |
return seed = r ^= r << 5; |
915 |
} |
916 |
|
917 |
// Execution methods |
918 |
|
919 |
/** |
920 |
* Pops and runs tasks until empty. |
921 |
*/ |
922 |
private void popAndExecAll() { |
923 |
// A bit faster than repeated pop calls |
924 |
ForkJoinTask<?>[] a; int m, s; long j; ForkJoinTask<?> t; |
925 |
while ((a = array) != null && (m = a.length - 1) >= 0 && |
926 |
(s = top - 1) - base >= 0 && |
927 |
(t = ((ForkJoinTask<?>) |
928 |
U.getObject(a, j = ((m & s) << ASHIFT) + ABASE))) |
929 |
!= null) { |
930 |
if (U.compareAndSwapObject(a, j, t, null)) { |
931 |
top = s; |
932 |
t.doExec(); |
933 |
} |
934 |
} |
935 |
} |
936 |
|
937 |
/** |
938 |
* Polls and runs tasks until empty. |
939 |
*/ |
940 |
private void pollAndExecAll() { |
941 |
for (ForkJoinTask<?> t; (t = poll()) != null;) |
942 |
t.doExec(); |
943 |
} |
944 |
|
945 |
/** |
946 |
* If present, removes from queue and executes the given task, or |
947 |
* any other cancelled task. Returns (true) immediately on any CAS |
948 |
* or consistency check failure so caller can retry. |
949 |
* |
950 |
* @return 0 if no progress can be made, else positive |
951 |
* (this unusual convention simplifies use with tryHelpStealer.) |
952 |
*/ |
953 |
final int tryRemoveAndExec(ForkJoinTask<?> task) { |
954 |
int stat = 1; |
955 |
boolean removed = false, empty = true; |
956 |
ForkJoinTask<?>[] a; int m, s, b, n; |
957 |
if ((a = array) != null && (m = a.length - 1) >= 0 && |
958 |
(n = (s = top) - (b = base)) > 0) { |
959 |
for (ForkJoinTask<?> t;;) { // traverse from s to b |
960 |
int j = ((--s & m) << ASHIFT) + ABASE; |
961 |
t = (ForkJoinTask<?>)U.getObjectVolatile(a, j); |
962 |
if (t == null) // inconsistent length |
963 |
break; |
964 |
else if (t == task) { |
965 |
if (s + 1 == top) { // pop |
966 |
if (!U.compareAndSwapObject(a, j, task, null)) |
967 |
break; |
968 |
top = s; |
969 |
removed = true; |
970 |
} |
971 |
else if (base == b) // replace with proxy |
972 |
removed = U.compareAndSwapObject(a, j, task, |
973 |
new EmptyTask()); |
974 |
break; |
975 |
} |
976 |
else if (t.status >= 0) |
977 |
empty = false; |
978 |
else if (s + 1 == top) { // pop and throw away |
979 |
if (U.compareAndSwapObject(a, j, t, null)) |
980 |
top = s; |
981 |
break; |
982 |
} |
983 |
if (--n == 0) { |
984 |
if (!empty && base == b) |
985 |
stat = 0; |
986 |
break; |
987 |
} |
988 |
} |
989 |
} |
990 |
if (removed) |
991 |
task.doExec(); |
992 |
return stat; |
993 |
} |
994 |
|
995 |
/** |
996 |
* Executes a top-level task and any local tasks remaining |
997 |
* after execution. |
998 |
*/ |
999 |
final void runTask(ForkJoinTask<?> t) { |
1000 |
if (t != null) { |
1001 |
currentSteal = t; |
1002 |
t.doExec(); |
1003 |
if (top != base) { // process remaining local tasks |
1004 |
if (mode == 0) |
1005 |
popAndExecAll(); |
1006 |
else |
1007 |
pollAndExecAll(); |
1008 |
} |
1009 |
++nsteals; |
1010 |
currentSteal = null; |
1011 |
} |
1012 |
} |
1013 |
|
1014 |
/** |
1015 |
* Executes a non-top-level (stolen) task. |
1016 |
*/ |
1017 |
final void runSubtask(ForkJoinTask<?> t) { |
1018 |
if (t != null) { |
1019 |
ForkJoinTask<?> ps = currentSteal; |
1020 |
currentSteal = t; |
1021 |
t.doExec(); |
1022 |
currentSteal = ps; |
1023 |
} |
1024 |
} |
1025 |
|
1026 |
/** |
1027 |
* Returns true if owned and not known to be blocked. |
1028 |
*/ |
1029 |
final boolean isApparentlyUnblocked() { |
1030 |
Thread wt; Thread.State s; |
1031 |
return (eventCount >= 0 && |
1032 |
(wt = owner) != null && |
1033 |
(s = wt.getState()) != Thread.State.BLOCKED && |
1034 |
s != Thread.State.WAITING && |
1035 |
s != Thread.State.TIMED_WAITING); |
1036 |
} |
1037 |
|
1038 |
/** |
1039 |
* If this owned and is not already interrupted, try to |
1040 |
* interrupt and/or unpark, ignoring exceptions. |
1041 |
*/ |
1042 |
final void interruptOwner() { |
1043 |
Thread wt, p; |
1044 |
if ((wt = owner) != null && !wt.isInterrupted()) { |
1045 |
try { |
1046 |
wt.interrupt(); |
1047 |
} catch (SecurityException ignore) { |
1048 |
} |
1049 |
} |
1050 |
if ((p = parker) != null) |
1051 |
U.unpark(p); |
1052 |
} |
1053 |
|
1054 |
// Unsafe mechanics |
1055 |
private static final sun.misc.Unsafe U; |
1056 |
private static final long RUNSTATE; |
1057 |
private static final int ABASE; |
1058 |
private static final int ASHIFT; |
1059 |
static { |
1060 |
int s; |
1061 |
try { |
1062 |
U = getUnsafe(); |
1063 |
Class<?> k = WorkQueue.class; |
1064 |
Class<?> ak = ForkJoinTask[].class; |
1065 |
RUNSTATE = U.objectFieldOffset |
1066 |
(k.getDeclaredField("runState")); |
1067 |
ABASE = U.arrayBaseOffset(ak); |
1068 |
s = U.arrayIndexScale(ak); |
1069 |
} catch (Exception e) { |
1070 |
throw new Error(e); |
1071 |
} |
1072 |
if ((s & (s-1)) != 0) |
1073 |
throw new Error("data type scale not a power of two"); |
1074 |
ASHIFT = 31 - Integer.numberOfLeadingZeros(s); |
1075 |
} |
1076 |
} |
1077 |
|
1078 |
/** |
1079 |
* Per-thread records for threads that submit to pools. Currently |
1080 |
* holds only pseudo-random seed / index that is used to choose |
1081 |
* submission queues in method doSubmit. In the future, this may |
1082 |
* also incorporate a means to implement different task rejection |
1083 |
* and resubmission policies. |
1084 |
* |
1085 |
* Seeds for submitters and workers/workQueues work in basically |
1086 |
* the same way but are initialized and updated using slightly |
1087 |
* different mechanics. Both are initialized using the same |
1088 |
* approach as in class ThreadLocal, where successive values are |
1089 |
* unlikely to collide with previous values. This is done during |
1090 |
* registration for workers, but requires a separate AtomicInteger |
1091 |
* for submitters. Seeds are then randomly modified upon |
1092 |
* collisions using xorshifts, which requires a non-zero seed. |
1093 |
*/ |
1094 |
static final class Submitter { |
1095 |
int seed; |
1096 |
Submitter() { |
1097 |
int s = nextSubmitterSeed.getAndAdd(SEED_INCREMENT); |
1098 |
seed = (s == 0) ? 1 : s; // ensure non-zero |
1099 |
} |
1100 |
} |
1101 |
|
1102 |
/** ThreadLocal class for Submitters */ |
1103 |
static final class ThreadSubmitter extends ThreadLocal<Submitter> { |
1104 |
public Submitter initialValue() { return new Submitter(); } |
1105 |
} |
1106 |
|
1107 |
// static fields (initialized in static initializer below) |
1108 |
|
1109 |
/** |
1110 |
* Creates a new ForkJoinWorkerThread. This factory is used unless |
1111 |
* overridden in ForkJoinPool constructors. |
1112 |
*/ |
1113 |
public static final ForkJoinWorkerThreadFactory |
1114 |
defaultForkJoinWorkerThreadFactory; |
1115 |
|
1116 |
/** |
1117 |
* Generator for assigning sequence numbers as pool names. |
1118 |
*/ |
1119 |
private static final AtomicInteger poolNumberGenerator; |
1120 |
|
1121 |
/** |
1122 |
* Generator for initial hashes/seeds for submitters. Accessed by |
1123 |
* Submitter class constructor. |
1124 |
*/ |
1125 |
static final AtomicInteger nextSubmitterSeed; |
1126 |
|
1127 |
/** |
1128 |
* Permission required for callers of methods that may start or |
1129 |
* kill threads. |
1130 |
*/ |
1131 |
private static final RuntimePermission modifyThreadPermission; |
1132 |
|
1133 |
/** |
1134 |
* Per-thread submission bookkeeping. Shared across all pools |
1135 |
* to reduce ThreadLocal pollution and because random motion |
1136 |
* to avoid contention in one pool is likely to hold for others. |
1137 |
*/ |
1138 |
private static final ThreadSubmitter submitters; |
1139 |
|
1140 |
/** Common default pool */ |
1141 |
static volatile ForkJoinPool commonPool; |
1142 |
|
1143 |
// commonPool construction parameters |
1144 |
private static final String propPrefix = |
1145 |
"java.util.concurrent.ForkJoinPool.common."; |
1146 |
private static final Thread.UncaughtExceptionHandler commonPoolUEH; |
1147 |
private static final ForkJoinWorkerThreadFactory commonPoolFactory; |
1148 |
static final int commonPoolParallelism; |
1149 |
|
1150 |
/** Static initialization lock */ |
1151 |
private static final Mutex initializationLock; |
1152 |
|
1153 |
// static constants |
1154 |
|
1155 |
/** |
1156 |
* Initial timeout value (in nanoseconds) for the tread triggering |
1157 |
* quiescence to park waiting for new work. On timeout, the thread |
1158 |
* will instead try to shrink the number of workers. |
1159 |
*/ |
1160 |
private static final long IDLE_TIMEOUT = 1000L * 1000L * 1000L; // 1sec |
1161 |
|
1162 |
/** |
1163 |
* Timeout value when there are more threads than parallelism level |
1164 |
*/ |
1165 |
private static final long FAST_IDLE_TIMEOUT = 100L * 1000L * 1000L; |
1166 |
|
1167 |
/** |
1168 |
* The maximum stolen->joining link depth allowed in method |
1169 |
* tryHelpStealer. Must be a power of two. This value also |
1170 |
* controls the maximum number of times to try to help join a task |
1171 |
* without any apparent progress or change in pool state before |
1172 |
* giving up and blocking (see awaitJoin). Depths for legitimate |
1173 |
* chains are unbounded, but we use a fixed constant to avoid |
1174 |
* (otherwise unchecked) cycles and to bound staleness of |
1175 |
* traversal parameters at the expense of sometimes blocking when |
1176 |
* we could be helping. |
1177 |
*/ |
1178 |
private static final int MAX_HELP = 64; |
1179 |
|
1180 |
/** |
1181 |
* Secondary time-based bound (in nanosecs) for helping attempts |
1182 |
* before trying compensated blocking in awaitJoin. Used in |
1183 |
* conjunction with MAX_HELP to reduce variance due to different |
1184 |
* polling rates associated with different helping options. The |
1185 |
* value should roughly approximate the time required to create |
1186 |
* and/or activate a worker thread. |
1187 |
*/ |
1188 |
private static final long COMPENSATION_DELAY = 1L << 18; // ~0.25 millisec |
1189 |
|
1190 |
/** |
1191 |
* Increment for seed generators. See class ThreadLocal for |
1192 |
* explanation. |
1193 |
*/ |
1194 |
private static final int SEED_INCREMENT = 0x61c88647; |
1195 |
|
1196 |
/** |
1197 |
* Bits and masks for control variables |
1198 |
* |
1199 |
* Field ctl is a long packed with: |
1200 |
* AC: Number of active running workers minus target parallelism (16 bits) |
1201 |
* TC: Number of total workers minus target parallelism (16 bits) |
1202 |
* ST: true if pool is terminating (1 bit) |
1203 |
* EC: the wait count of top waiting thread (15 bits) |
1204 |
* ID: poolIndex of top of Treiber stack of waiters (16 bits) |
1205 |
* |
1206 |
* When convenient, we can extract the upper 32 bits of counts and |
1207 |
* the lower 32 bits of queue state, u = (int)(ctl >>> 32) and e = |
1208 |
* (int)ctl. The ec field is never accessed alone, but always |
1209 |
* together with id and st. The offsets of counts by the target |
1210 |
* parallelism and the positionings of fields makes it possible to |
1211 |
* perform the most common checks via sign tests of fields: When |
1212 |
* ac is negative, there are not enough active workers, when tc is |
1213 |
* negative, there are not enough total workers, and when e is |
1214 |
* negative, the pool is terminating. To deal with these possibly |
1215 |
* negative fields, we use casts in and out of "short" and/or |
1216 |
* signed shifts to maintain signedness. |
1217 |
* |
1218 |
* When a thread is queued (inactivated), its eventCount field is |
1219 |
* set negative, which is the only way to tell if a worker is |
1220 |
* prevented from executing tasks, even though it must continue to |
1221 |
* scan for them to avoid queuing races. Note however that |
1222 |
* eventCount updates lag releases so usage requires care. |
1223 |
* |
1224 |
* Field runState is an int packed with: |
1225 |
* SHUTDOWN: true if shutdown is enabled (1 bit) |
1226 |
* SEQ: a sequence number updated upon (de)registering workers (30 bits) |
1227 |
* INIT: set true after workQueues array construction (1 bit) |
1228 |
* |
1229 |
* The sequence number enables simple consistency checks: |
1230 |
* Staleness of read-only operations on the workQueues array can |
1231 |
* be checked by comparing runState before vs after the reads. |
1232 |
*/ |
1233 |
|
1234 |
// bit positions/shifts for fields |
1235 |
private static final int AC_SHIFT = 48; |
1236 |
private static final int TC_SHIFT = 32; |
1237 |
private static final int ST_SHIFT = 31; |
1238 |
private static final int EC_SHIFT = 16; |
1239 |
|
1240 |
// bounds |
1241 |
private static final int SMASK = 0xffff; // short bits |
1242 |
private static final int MAX_CAP = 0x7fff; // max #workers - 1 |
1243 |
private static final int SQMASK = 0xfffe; // even short bits |
1244 |
private static final int SHORT_SIGN = 1 << 15; |
1245 |
private static final int INT_SIGN = 1 << 31; |
1246 |
|
1247 |
// masks |
1248 |
private static final long STOP_BIT = 0x0001L << ST_SHIFT; |
1249 |
private static final long AC_MASK = ((long)SMASK) << AC_SHIFT; |
1250 |
private static final long TC_MASK = ((long)SMASK) << TC_SHIFT; |
1251 |
|
1252 |
// units for incrementing and decrementing |
1253 |
private static final long TC_UNIT = 1L << TC_SHIFT; |
1254 |
private static final long AC_UNIT = 1L << AC_SHIFT; |
1255 |
|
1256 |
// masks and units for dealing with u = (int)(ctl >>> 32) |
1257 |
private static final int UAC_SHIFT = AC_SHIFT - 32; |
1258 |
private static final int UTC_SHIFT = TC_SHIFT - 32; |
1259 |
private static final int UAC_MASK = SMASK << UAC_SHIFT; |
1260 |
private static final int UTC_MASK = SMASK << UTC_SHIFT; |
1261 |
private static final int UAC_UNIT = 1 << UAC_SHIFT; |
1262 |
private static final int UTC_UNIT = 1 << UTC_SHIFT; |
1263 |
|
1264 |
// masks and units for dealing with e = (int)ctl |
1265 |
private static final int E_MASK = 0x7fffffff; // no STOP_BIT |
1266 |
private static final int E_SEQ = 1 << EC_SHIFT; |
1267 |
|
1268 |
// runState bits |
1269 |
private static final int SHUTDOWN = 1 << 31; |
1270 |
|
1271 |
// access mode for WorkQueue |
1272 |
static final int LIFO_QUEUE = 0; |
1273 |
static final int FIFO_QUEUE = 1; |
1274 |
static final int SHARED_QUEUE = -1; |
1275 |
|
1276 |
// Instance fields |
1277 |
|
1278 |
/* |
1279 |
* Field layout order in this class tends to matter more than one |
1280 |
* would like. Runtime layout order is only loosely related to |
1281 |
* declaration order and may differ across JVMs, but the following |
1282 |
* empirically works OK on current JVMs. |
1283 |
*/ |
1284 |
|
1285 |
volatile long ctl; // main pool control |
1286 |
final int parallelism; // parallelism level |
1287 |
final int localMode; // per-worker scheduling mode |
1288 |
final int submitMask; // submit queue index bound |
1289 |
int nextSeed; // for initializing worker seeds |
1290 |
volatile int runState; // shutdown status and seq |
1291 |
WorkQueue[] workQueues; // main registry |
1292 |
final Mutex lock; // for registration |
1293 |
final Condition termination; // for awaitTermination |
1294 |
final ForkJoinWorkerThreadFactory factory; // factory for new workers |
1295 |
final Thread.UncaughtExceptionHandler ueh; // per-worker UEH |
1296 |
final AtomicLong stealCount; // collect counts when terminated |
1297 |
final AtomicInteger nextWorkerNumber; // to create worker name string |
1298 |
String workerNamePrefix; // to create worker name string |
1299 |
|
1300 |
// Creating, registering, and deregistering workers |
1301 |
|
1302 |
/** |
1303 |
* Tries to create and start a worker |
1304 |
*/ |
1305 |
private void addWorker() { |
1306 |
Throwable ex = null; |
1307 |
ForkJoinWorkerThread wt = null; |
1308 |
try { |
1309 |
if ((wt = factory.newThread(this)) != null) { |
1310 |
wt.start(); |
1311 |
return; |
1312 |
} |
1313 |
} catch (Throwable e) { |
1314 |
ex = e; |
1315 |
} |
1316 |
deregisterWorker(wt, ex); // adjust counts etc on failure |
1317 |
} |
1318 |
|
1319 |
/** |
1320 |
* Callback from ForkJoinWorkerThread constructor to assign a |
1321 |
* public name. This must be separate from registerWorker because |
1322 |
* it is called during the "super" constructor call in |
1323 |
* ForkJoinWorkerThread. |
1324 |
*/ |
1325 |
final String nextWorkerName() { |
1326 |
return workerNamePrefix.concat |
1327 |
(Integer.toString(nextWorkerNumber.addAndGet(1))); |
1328 |
} |
1329 |
|
1330 |
/** |
1331 |
* Callback from ForkJoinWorkerThread constructor to establish its |
1332 |
* poolIndex and record its WorkQueue. To avoid scanning bias due |
1333 |
* to packing entries in front of the workQueues array, we treat |
1334 |
* the array as a simple power-of-two hash table using per-thread |
1335 |
* seed as hash, expanding as needed. |
1336 |
* |
1337 |
* @param w the worker's queue |
1338 |
*/ |
1339 |
final void registerWorker(WorkQueue w) { |
1340 |
Mutex lock = this.lock; |
1341 |
lock.lock(); |
1342 |
try { |
1343 |
WorkQueue[] ws = workQueues; |
1344 |
if (w != null && ws != null) { // skip on shutdown/failure |
1345 |
int rs, n = ws.length, m = n - 1; |
1346 |
int s = nextSeed += SEED_INCREMENT; // rarely-colliding sequence |
1347 |
w.seed = (s == 0) ? 1 : s; // ensure non-zero seed |
1348 |
int r = (s << 1) | 1; // use odd-numbered indices |
1349 |
if (ws[r &= m] != null) { // collision |
1350 |
int probes = 0; // step by approx half size |
1351 |
int step = (n <= 4) ? 2 : ((n >>> 1) & SQMASK) + 2; |
1352 |
while (ws[r = (r + step) & m] != null) { |
1353 |
if (++probes >= n) { |
1354 |
workQueues = ws = Arrays.copyOf(ws, n <<= 1); |
1355 |
m = n - 1; |
1356 |
probes = 0; |
1357 |
} |
1358 |
} |
1359 |
} |
1360 |
w.eventCount = w.poolIndex = r; // establish before recording |
1361 |
ws[r] = w; // also update seq |
1362 |
runState = ((rs = runState) & SHUTDOWN) | ((rs + 2) & ~SHUTDOWN); |
1363 |
} |
1364 |
} finally { |
1365 |
lock.unlock(); |
1366 |
} |
1367 |
} |
1368 |
|
1369 |
/** |
1370 |
* Final callback from terminating worker, as well as upon failure |
1371 |
* to construct or start a worker in addWorker. Removes record of |
1372 |
* worker from array, and adjusts counts. If pool is shutting |
1373 |
* down, tries to complete termination. |
1374 |
* |
1375 |
* @param wt the worker thread or null if addWorker failed |
1376 |
* @param ex the exception causing failure, or null if none |
1377 |
*/ |
1378 |
final void deregisterWorker(ForkJoinWorkerThread wt, Throwable ex) { |
1379 |
Mutex lock = this.lock; |
1380 |
WorkQueue w = null; |
1381 |
if (wt != null && (w = wt.workQueue) != null) { |
1382 |
w.runState = -1; // ensure runState is set |
1383 |
stealCount.getAndAdd(w.totalSteals + w.nsteals); |
1384 |
int idx = w.poolIndex; |
1385 |
lock.lock(); |
1386 |
try { // remove record from array |
1387 |
WorkQueue[] ws = workQueues; |
1388 |
if (ws != null && idx >= 0 && idx < ws.length && ws[idx] == w) |
1389 |
ws[idx] = null; |
1390 |
} finally { |
1391 |
lock.unlock(); |
1392 |
} |
1393 |
} |
1394 |
|
1395 |
long c; // adjust ctl counts |
1396 |
do {} while (!U.compareAndSwapLong |
1397 |
(this, CTL, c = ctl, (((c - AC_UNIT) & AC_MASK) | |
1398 |
((c - TC_UNIT) & TC_MASK) | |
1399 |
(c & ~(AC_MASK|TC_MASK))))); |
1400 |
|
1401 |
if (!tryTerminate(false, false) && w != null) { |
1402 |
w.cancelAll(); // cancel remaining tasks |
1403 |
if (w.array != null) // suppress signal if never ran |
1404 |
signalWork(); // wake up or create replacement |
1405 |
if (ex == null) // help clean refs on way out |
1406 |
ForkJoinTask.helpExpungeStaleExceptions(); |
1407 |
} |
1408 |
|
1409 |
if (ex != null) // rethrow |
1410 |
U.throwException(ex); |
1411 |
} |
1412 |
|
1413 |
// Submissions |
1414 |
|
1415 |
/** |
1416 |
* Unless shutting down, adds the given task to a submission queue |
1417 |
* at submitter's current queue index (modulo submission |
1418 |
* range). If no queue exists at the index, one is created. If |
1419 |
* the queue is busy, another index is randomly chosen. The |
1420 |
* submitMask bounds the effective number of queues to the |
1421 |
* (nearest power of two for) parallelism level. |
1422 |
* |
1423 |
* @param task the task. Caller must ensure non-null. |
1424 |
*/ |
1425 |
private void doSubmit(ForkJoinTask<?> task) { |
1426 |
Submitter s = submitters.get(); |
1427 |
for (int r = s.seed, m = submitMask;;) { |
1428 |
WorkQueue[] ws; WorkQueue q; |
1429 |
int k = r & m & SQMASK; // use only even indices |
1430 |
if (runState < 0 || (ws = workQueues) == null || ws.length <= k) |
1431 |
throw new RejectedExecutionException(); // shutting down |
1432 |
else if ((q = ws[k]) == null) { // create new queue |
1433 |
WorkQueue nq = new WorkQueue(this, null, SHARED_QUEUE); |
1434 |
Mutex lock = this.lock; // construct outside lock |
1435 |
lock.lock(); |
1436 |
try { // recheck under lock |
1437 |
int rs = runState; // to update seq |
1438 |
if (ws == workQueues && ws[k] == null) { |
1439 |
ws[k] = nq; |
1440 |
runState = ((rs & SHUTDOWN) | ((rs + 2) & ~SHUTDOWN)); |
1441 |
} |
1442 |
} finally { |
1443 |
lock.unlock(); |
1444 |
} |
1445 |
} |
1446 |
else if (q.trySharedPush(task)) { |
1447 |
signalWork(); |
1448 |
return; |
1449 |
} |
1450 |
else if (m > 1) { // move to a different index |
1451 |
r ^= r << 13; // same xorshift as WorkQueues |
1452 |
r ^= r >>> 17; |
1453 |
s.seed = r ^= r << 5; |
1454 |
} |
1455 |
else |
1456 |
Thread.yield(); // yield if no alternatives |
1457 |
} |
1458 |
} |
1459 |
|
1460 |
/** |
1461 |
* Submits the given (non-null) task to the common pool, if possible. |
1462 |
*/ |
1463 |
static void submitToCommonPool(ForkJoinTask<?> task) { |
1464 |
ForkJoinPool p; |
1465 |
if ((p = commonPool) == null) |
1466 |
p = ensureCommonPool(); |
1467 |
p.doSubmit(task); |
1468 |
} |
1469 |
|
1470 |
/** |
1471 |
* Returns true if the given task was submitted to common pool |
1472 |
* and has not yet commenced execution, and is available for |
1473 |
* removal according to execution policies; if so removing the |
1474 |
* submission from the pool. |
1475 |
* |
1476 |
* @param task the task |
1477 |
* @return true if successful |
1478 |
*/ |
1479 |
static boolean tryUnsubmitFromCommonPool(ForkJoinTask<?> task) { |
1480 |
ForkJoinPool p; WorkQueue[] ws; WorkQueue q; |
1481 |
int k = submitters.get().seed & SQMASK; |
1482 |
return ((p = commonPool) != null && |
1483 |
(ws = p.workQueues) != null && |
1484 |
ws.length > (k &= p.submitMask) && |
1485 |
(q = ws[k]) != null && |
1486 |
q.trySharedUnpush(task, p)); |
1487 |
} |
1488 |
|
1489 |
// Maintaining ctl counts |
1490 |
|
1491 |
/** |
1492 |
* Increments active count; mainly called upon return from blocking. |
1493 |
*/ |
1494 |
final void incrementActiveCount() { |
1495 |
long c; |
1496 |
do {} while (!U.compareAndSwapLong(this, CTL, c = ctl, c + AC_UNIT)); |
1497 |
} |
1498 |
|
1499 |
/** |
1500 |
* Tries to create one or activate one or more workers if too few are active. |
1501 |
*/ |
1502 |
final void signalWork() { |
1503 |
long c; int u; |
1504 |
while ((u = (int)((c = ctl) >>> 32)) < 0) { // too few active |
1505 |
WorkQueue[] ws = workQueues; int e, i; WorkQueue w; Thread p; |
1506 |
if ((e = (int)c) > 0) { // at least one waiting |
1507 |
if (ws != null && (i = e & SMASK) < ws.length && |
1508 |
(w = ws[i]) != null && w.eventCount == (e | INT_SIGN)) { |
1509 |
long nc = (((long)(w.nextWait & E_MASK)) | |
1510 |
((long)(u + UAC_UNIT) << 32)); |
1511 |
if (U.compareAndSwapLong(this, CTL, c, nc)) { |
1512 |
w.eventCount = (e + E_SEQ) & E_MASK; |
1513 |
if ((p = w.parker) != null) |
1514 |
U.unpark(p); // activate and release |
1515 |
break; |
1516 |
} |
1517 |
} |
1518 |
else |
1519 |
break; |
1520 |
} |
1521 |
else if (e == 0 && (u & SHORT_SIGN) != 0) { // too few total |
1522 |
long nc = (long)(((u + UTC_UNIT) & UTC_MASK) | |
1523 |
((u + UAC_UNIT) & UAC_MASK)) << 32; |
1524 |
if (U.compareAndSwapLong(this, CTL, c, nc)) { |
1525 |
addWorker(); |
1526 |
break; |
1527 |
} |
1528 |
} |
1529 |
else |
1530 |
break; |
1531 |
} |
1532 |
} |
1533 |
|
1534 |
// Scanning for tasks |
1535 |
|
1536 |
/** |
1537 |
* Top-level runloop for workers, called by ForkJoinWorkerThread.run. |
1538 |
*/ |
1539 |
final void runWorker(WorkQueue w) { |
1540 |
w.growArray(false); // initialize queue array in this thread |
1541 |
do { w.runTask(scan(w)); } while (w.runState >= 0); |
1542 |
} |
1543 |
|
1544 |
/** |
1545 |
* Scans for and, if found, returns one task, else possibly |
1546 |
* inactivates the worker. This method operates on single reads of |
1547 |
* volatile state and is designed to be re-invoked continuously, |
1548 |
* in part because it returns upon detecting inconsistencies, |
1549 |
* contention, or state changes that indicate possible success on |
1550 |
* re-invocation. |
1551 |
* |
1552 |
* The scan searches for tasks across a random permutation of |
1553 |
* queues (starting at a random index and stepping by a random |
1554 |
* relative prime, checking each at least once). The scan |
1555 |
* terminates upon either finding a non-empty queue, or completing |
1556 |
* the sweep. If the worker is not inactivated, it takes and |
1557 |
* returns a task from this queue. On failure to find a task, we |
1558 |
* take one of the following actions, after which the caller will |
1559 |
* retry calling this method unless terminated. |
1560 |
* |
1561 |
* * If pool is terminating, terminate the worker. |
1562 |
* |
1563 |
* * If not a complete sweep, try to release a waiting worker. If |
1564 |
* the scan terminated because the worker is inactivated, then the |
1565 |
* released worker will often be the calling worker, and it can |
1566 |
* succeed obtaining a task on the next call. Or maybe it is |
1567 |
* another worker, but with same net effect. Releasing in other |
1568 |
* cases as well ensures that we have enough workers running. |
1569 |
* |
1570 |
* * If not already enqueued, try to inactivate and enqueue the |
1571 |
* worker on wait queue. Or, if inactivating has caused the pool |
1572 |
* to be quiescent, relay to idleAwaitWork to check for |
1573 |
* termination and possibly shrink pool. |
1574 |
* |
1575 |
* * If already inactive, and the caller has run a task since the |
1576 |
* last empty scan, return (to allow rescan) unless others are |
1577 |
* also inactivated. Field WorkQueue.rescans counts down on each |
1578 |
* scan to ensure eventual inactivation and blocking. |
1579 |
* |
1580 |
* * If already enqueued and none of the above apply, park |
1581 |
* awaiting signal, |
1582 |
* |
1583 |
* @param w the worker (via its WorkQueue) |
1584 |
* @return a task or null if none found |
1585 |
*/ |
1586 |
private final ForkJoinTask<?> scan(WorkQueue w) { |
1587 |
WorkQueue[] ws; // first update random seed |
1588 |
int r = w.seed; r ^= r << 13; r ^= r >>> 17; w.seed = r ^= r << 5; |
1589 |
int rs = runState, m; // volatile read order matters |
1590 |
if ((ws = workQueues) != null && (m = ws.length - 1) > 0) { |
1591 |
int ec = w.eventCount; // ec is negative if inactive |
1592 |
int step = (r >>> 16) | 1; // relative prime |
1593 |
for (int j = (m + 1) << 2; ; r += step) { |
1594 |
WorkQueue q; ForkJoinTask<?> t; ForkJoinTask<?>[] a; int b; |
1595 |
if ((q = ws[r & m]) != null && (b = q.base) - q.top < 0 && |
1596 |
(a = q.array) != null) { // probably nonempty |
1597 |
int i = (((a.length - 1) & b) << ASHIFT) + ABASE; |
1598 |
t = (ForkJoinTask<?>)U.getObjectVolatile(a, i); |
1599 |
if (q.base == b && ec >= 0 && t != null && |
1600 |
U.compareAndSwapObject(a, i, t, null)) { |
1601 |
if (q.top - (q.base = b + 1) > 0) |
1602 |
signalWork(); // help pushes signal |
1603 |
return t; |
1604 |
} |
1605 |
else if (ec < 0 || j <= m) { |
1606 |
rs = 0; // mark scan as imcomplete |
1607 |
break; // caller can retry after release |
1608 |
} |
1609 |
} |
1610 |
if (--j < 0) |
1611 |
break; |
1612 |
} |
1613 |
|
1614 |
long c = ctl; int e = (int)c, a = (int)(c >> AC_SHIFT), nr, ns; |
1615 |
if (e < 0) // decode ctl on empty scan |
1616 |
w.runState = -1; // pool is terminating |
1617 |
else if (rs == 0 || rs != runState) { // incomplete scan |
1618 |
WorkQueue v; Thread p; // try to release a waiter |
1619 |
if (e > 0 && a < 0 && w.eventCount == ec && |
1620 |
(v = ws[e & m]) != null && v.eventCount == (e | INT_SIGN)) { |
1621 |
long nc = ((long)(v.nextWait & E_MASK) | |
1622 |
((c + AC_UNIT) & (AC_MASK|TC_MASK))); |
1623 |
if (ctl == c && U.compareAndSwapLong(this, CTL, c, nc)) { |
1624 |
v.eventCount = (e + E_SEQ) & E_MASK; |
1625 |
if ((p = v.parker) != null) |
1626 |
U.unpark(p); |
1627 |
} |
1628 |
} |
1629 |
} |
1630 |
else if (ec >= 0) { // try to enqueue/inactivate |
1631 |
long nc = (long)ec | ((c - AC_UNIT) & (AC_MASK|TC_MASK)); |
1632 |
w.nextWait = e; |
1633 |
w.eventCount = ec | INT_SIGN; // mark as inactive |
1634 |
if (ctl != c || !U.compareAndSwapLong(this, CTL, c, nc)) |
1635 |
w.eventCount = ec; // unmark on CAS failure |
1636 |
else { |
1637 |
if ((ns = w.nsteals) != 0) { |
1638 |
w.nsteals = 0; // set rescans if ran task |
1639 |
w.rescans = (a > 0) ? 0 : a + parallelism; |
1640 |
w.totalSteals += ns; |
1641 |
} |
1642 |
if (a == 1 - parallelism) // quiescent |
1643 |
idleAwaitWork(w, nc, c); |
1644 |
} |
1645 |
} |
1646 |
else if (w.eventCount < 0) { // already queued |
1647 |
int ac = a + parallelism; |
1648 |
if ((nr = w.rescans) > 0) // continue rescanning |
1649 |
w.rescans = (ac < nr) ? ac : nr - 1; |
1650 |
else if (((w.seed >>> 16) & ac) == 0) { // randomize park |
1651 |
Thread.interrupted(); // clear status |
1652 |
Thread wt = Thread.currentThread(); |
1653 |
U.putObject(wt, PARKBLOCKER, this); |
1654 |
w.parker = wt; // emulate LockSupport.park |
1655 |
if (w.eventCount < 0) // recheck |
1656 |
U.park(false, 0L); |
1657 |
w.parker = null; |
1658 |
U.putObject(wt, PARKBLOCKER, null); |
1659 |
} |
1660 |
} |
1661 |
} |
1662 |
return null; |
1663 |
} |
1664 |
|
1665 |
/** |
1666 |
* If inactivating worker w has caused the pool to become |
1667 |
* quiescent, checks for pool termination, and, so long as this is |
1668 |
* not the only worker, waits for event for up to a given |
1669 |
* duration. On timeout, if ctl has not changed, terminates the |
1670 |
* worker, which will in turn wake up another worker to possibly |
1671 |
* repeat this process. |
1672 |
* |
1673 |
* @param w the calling worker |
1674 |
* @param currentCtl the ctl value triggering possible quiescence |
1675 |
* @param prevCtl the ctl value to restore if thread is terminated |
1676 |
*/ |
1677 |
private void idleAwaitWork(WorkQueue w, long currentCtl, long prevCtl) { |
1678 |
if (w.eventCount < 0 && !tryTerminate(false, false) && |
1679 |
(int)prevCtl != 0 && !hasQueuedSubmissions() && ctl == currentCtl) { |
1680 |
int dc = -(short)(currentCtl >>> TC_SHIFT); |
1681 |
long parkTime = dc < 0 ? FAST_IDLE_TIMEOUT: (dc + 1) * IDLE_TIMEOUT; |
1682 |
long deadline = System.nanoTime() + parkTime - 100000L; // 1ms slop |
1683 |
Thread wt = Thread.currentThread(); |
1684 |
while (ctl == currentCtl) { |
1685 |
Thread.interrupted(); // timed variant of version in scan() |
1686 |
U.putObject(wt, PARKBLOCKER, this); |
1687 |
w.parker = wt; |
1688 |
if (ctl == currentCtl) |
1689 |
U.park(false, parkTime); |
1690 |
w.parker = null; |
1691 |
U.putObject(wt, PARKBLOCKER, null); |
1692 |
if (ctl != currentCtl) |
1693 |
break; |
1694 |
if (deadline - System.nanoTime() <= 0L && |
1695 |
U.compareAndSwapLong(this, CTL, currentCtl, prevCtl)) { |
1696 |
w.eventCount = (w.eventCount + E_SEQ) | E_MASK; |
1697 |
w.runState = -1; // shrink |
1698 |
break; |
1699 |
} |
1700 |
} |
1701 |
} |
1702 |
} |
1703 |
|
1704 |
/** |
1705 |
* Tries to locate and execute tasks for a stealer of the given |
1706 |
* task, or in turn one of its stealers, Traces currentSteal -> |
1707 |
* currentJoin links looking for a thread working on a descendant |
1708 |
* of the given task and with a non-empty queue to steal back and |
1709 |
* execute tasks from. The first call to this method upon a |
1710 |
* waiting join will often entail scanning/search, (which is OK |
1711 |
* because the joiner has nothing better to do), but this method |
1712 |
* leaves hints in workers to speed up subsequent calls. The |
1713 |
* implementation is very branchy to cope with potential |
1714 |
* inconsistencies or loops encountering chains that are stale, |
1715 |
* unknown, or so long that they are likely cyclic. |
1716 |
* |
1717 |
* @param joiner the joining worker |
1718 |
* @param task the task to join |
1719 |
* @return 0 if no progress can be made, negative if task |
1720 |
* known complete, else positive |
1721 |
*/ |
1722 |
private int tryHelpStealer(WorkQueue joiner, ForkJoinTask<?> task) { |
1723 |
int stat = 0, steps = 0; // bound to avoid cycles |
1724 |
if (joiner != null && task != null) { // hoist null checks |
1725 |
restart: for (;;) { |
1726 |
ForkJoinTask<?> subtask = task; // current target |
1727 |
for (WorkQueue j = joiner, v;;) { // v is stealer of subtask |
1728 |
WorkQueue[] ws; int m, s, h; |
1729 |
if ((s = task.status) < 0) { |
1730 |
stat = s; |
1731 |
break restart; |
1732 |
} |
1733 |
if ((ws = workQueues) == null || (m = ws.length - 1) <= 0) |
1734 |
break restart; // shutting down |
1735 |
if ((v = ws[h = (j.stealHint | 1) & m]) == null || |
1736 |
v.currentSteal != subtask) { |
1737 |
for (int origin = h;;) { // find stealer |
1738 |
if (((h = (h + 2) & m) & 15) == 1 && |
1739 |
(subtask.status < 0 || j.currentJoin != subtask)) |
1740 |
continue restart; // occasional staleness check |
1741 |
if ((v = ws[h]) != null && |
1742 |
v.currentSteal == subtask) { |
1743 |
j.stealHint = h; // save hint |
1744 |
break; |
1745 |
} |
1746 |
if (h == origin) |
1747 |
break restart; // cannot find stealer |
1748 |
} |
1749 |
} |
1750 |
for (;;) { // help stealer or descend to its stealer |
1751 |
ForkJoinTask[] a; int b; |
1752 |
if (subtask.status < 0) // surround probes with |
1753 |
continue restart; // consistency checks |
1754 |
if ((b = v.base) - v.top < 0 && (a = v.array) != null) { |
1755 |
int i = (((a.length - 1) & b) << ASHIFT) + ABASE; |
1756 |
ForkJoinTask<?> t = |
1757 |
(ForkJoinTask<?>)U.getObjectVolatile(a, i); |
1758 |
if (subtask.status < 0 || j.currentJoin != subtask || |
1759 |
v.currentSteal != subtask) |
1760 |
continue restart; // stale |
1761 |
stat = 1; // apparent progress |
1762 |
if (t != null && v.base == b && |
1763 |
U.compareAndSwapObject(a, i, t, null)) { |
1764 |
v.base = b + 1; // help stealer |
1765 |
joiner.runSubtask(t); |
1766 |
} |
1767 |
else if (v.base == b && ++steps == MAX_HELP) |
1768 |
break restart; // v apparently stalled |
1769 |
} |
1770 |
else { // empty -- try to descend |
1771 |
ForkJoinTask<?> next = v.currentJoin; |
1772 |
if (subtask.status < 0 || j.currentJoin != subtask || |
1773 |
v.currentSteal != subtask) |
1774 |
continue restart; // stale |
1775 |
else if (next == null || ++steps == MAX_HELP) |
1776 |
break restart; // dead-end or maybe cyclic |
1777 |
else { |
1778 |
subtask = next; |
1779 |
j = v; |
1780 |
break; |
1781 |
} |
1782 |
} |
1783 |
} |
1784 |
} |
1785 |
} |
1786 |
} |
1787 |
return stat; |
1788 |
} |
1789 |
|
1790 |
/** |
1791 |
* If task is at base of some steal queue, steals and executes it. |
1792 |
* |
1793 |
* @param joiner the joining worker |
1794 |
* @param task the task |
1795 |
*/ |
1796 |
private void tryPollForAndExec(WorkQueue joiner, ForkJoinTask<?> task) { |
1797 |
WorkQueue[] ws; |
1798 |
if ((ws = workQueues) != null) { |
1799 |
for (int j = 1; j < ws.length && task.status >= 0; j += 2) { |
1800 |
WorkQueue q = ws[j]; |
1801 |
if (q != null && q.pollFor(task)) { |
1802 |
joiner.runSubtask(task); |
1803 |
break; |
1804 |
} |
1805 |
} |
1806 |
} |
1807 |
} |
1808 |
|
1809 |
/** |
1810 |
* Tries to decrement active count (sometimes implicitly) and |
1811 |
* possibly release or create a compensating worker in preparation |
1812 |
* for blocking. Fails on contention or termination. Otherwise, |
1813 |
* adds a new thread if no idle workers are available and either |
1814 |
* pool would become completely starved or: (at least half |
1815 |
* starved, and fewer than 50% spares exist, and there is at least |
1816 |
* one task apparently available). Even though the availability |
1817 |
* check requires a full scan, it is worthwhile in reducing false |
1818 |
* alarms. |
1819 |
* |
1820 |
* @param task if non-null, a task being waited for |
1821 |
* @param blocker if non-null, a blocker being waited for |
1822 |
* @return true if the caller can block, else should recheck and retry |
1823 |
*/ |
1824 |
final boolean tryCompensate(ForkJoinTask<?> task, ManagedBlocker blocker) { |
1825 |
int pc = parallelism, e; |
1826 |
long c = ctl; |
1827 |
WorkQueue[] ws = workQueues; |
1828 |
if ((e = (int)c) >= 0 && ws != null) { |
1829 |
int u, a, ac, hc; |
1830 |
int tc = (short)((u = (int)(c >>> 32)) >>> UTC_SHIFT) + pc; |
1831 |
boolean replace = false; |
1832 |
if ((a = u >> UAC_SHIFT) <= 0) { |
1833 |
if ((ac = a + pc) <= 1) |
1834 |
replace = true; |
1835 |
else if ((e > 0 || (task != null && |
1836 |
ac <= (hc = pc >>> 1) && tc < pc + hc))) { |
1837 |
WorkQueue w; |
1838 |
for (int j = 0; j < ws.length; ++j) { |
1839 |
if ((w = ws[j]) != null && !w.isEmpty()) { |
1840 |
replace = true; |
1841 |
break; // in compensation range and tasks available |
1842 |
} |
1843 |
} |
1844 |
} |
1845 |
} |
1846 |
if ((task == null || task.status >= 0) && // recheck need to block |
1847 |
(blocker == null || !blocker.isReleasable()) && ctl == c) { |
1848 |
if (!replace) { // no compensation |
1849 |
long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK); |
1850 |
if (U.compareAndSwapLong(this, CTL, c, nc)) |
1851 |
return true; |
1852 |
} |
1853 |
else if (e != 0) { // release an idle worker |
1854 |
WorkQueue w; Thread p; int i; |
1855 |
if ((i = e & SMASK) < ws.length && (w = ws[i]) != null) { |
1856 |
long nc = ((long)(w.nextWait & E_MASK) | |
1857 |
(c & (AC_MASK|TC_MASK))); |
1858 |
if (w.eventCount == (e | INT_SIGN) && |
1859 |
U.compareAndSwapLong(this, CTL, c, nc)) { |
1860 |
w.eventCount = (e + E_SEQ) & E_MASK; |
1861 |
if ((p = w.parker) != null) |
1862 |
U.unpark(p); |
1863 |
return true; |
1864 |
} |
1865 |
} |
1866 |
} |
1867 |
else if (tc < MAX_CAP) { // create replacement |
1868 |
long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK); |
1869 |
if (U.compareAndSwapLong(this, CTL, c, nc)) { |
1870 |
addWorker(); |
1871 |
return true; |
1872 |
} |
1873 |
} |
1874 |
} |
1875 |
} |
1876 |
return false; |
1877 |
} |
1878 |
|
1879 |
/** |
1880 |
* Helps and/or blocks until the given task is done. |
1881 |
* |
1882 |
* @param joiner the joining worker |
1883 |
* @param task the task |
1884 |
* @return task status on exit |
1885 |
*/ |
1886 |
final int awaitJoin(WorkQueue joiner, ForkJoinTask<?> task) { |
1887 |
int s; |
1888 |
if ((s = task.status) >= 0) { |
1889 |
ForkJoinTask<?> prevJoin = joiner.currentJoin; |
1890 |
joiner.currentJoin = task; |
1891 |
long startTime = 0L; |
1892 |
for (int k = 0;;) { |
1893 |
if ((s = (joiner.isEmpty() ? // try to help |
1894 |
tryHelpStealer(joiner, task) : |
1895 |
joiner.tryRemoveAndExec(task))) == 0 && |
1896 |
(s = task.status) >= 0) { |
1897 |
if (k == 0) { |
1898 |
startTime = System.nanoTime(); |
1899 |
tryPollForAndExec(joiner, task); // check uncommon case |
1900 |
} |
1901 |
else if ((k & (MAX_HELP - 1)) == 0 && |
1902 |
System.nanoTime() - startTime >= |
1903 |
COMPENSATION_DELAY && |
1904 |
tryCompensate(task, null)) { |
1905 |
if (task.trySetSignal()) { |
1906 |
synchronized (task) { |
1907 |
if (task.status >= 0) { |
1908 |
try { // see ForkJoinTask |
1909 |
task.wait(); // for explanation |
1910 |
} catch (InterruptedException ie) { |
1911 |
} |
1912 |
} |
1913 |
else |
1914 |
task.notifyAll(); |
1915 |
} |
1916 |
} |
1917 |
long c; // re-activate |
1918 |
do {} while (!U.compareAndSwapLong |
1919 |
(this, CTL, c = ctl, c + AC_UNIT)); |
1920 |
} |
1921 |
} |
1922 |
if (s < 0 || (s = task.status) < 0) { |
1923 |
joiner.currentJoin = prevJoin; |
1924 |
break; |
1925 |
} |
1926 |
else if ((k++ & (MAX_HELP - 1)) == MAX_HELP >>> 1) |
1927 |
Thread.yield(); // for politeness |
1928 |
} |
1929 |
} |
1930 |
return s; |
1931 |
} |
1932 |
|
1933 |
/** |
1934 |
* Stripped-down variant of awaitJoin used by timed joins. Tries |
1935 |
* to help join only while there is continuous progress. (Caller |
1936 |
* will then enter a timed wait.) |
1937 |
* |
1938 |
* @param joiner the joining worker |
1939 |
* @param task the task |
1940 |
* @return task status on exit |
1941 |
*/ |
1942 |
final int helpJoinOnce(WorkQueue joiner, ForkJoinTask<?> task) { |
1943 |
int s; |
1944 |
while ((s = task.status) >= 0 && |
1945 |
(joiner.isEmpty() ? |
1946 |
tryHelpStealer(joiner, task) : |
1947 |
joiner.tryRemoveAndExec(task)) != 0) |
1948 |
; |
1949 |
return s; |
1950 |
} |
1951 |
|
1952 |
/** |
1953 |
* Returns a (probably) non-empty steal queue, if one is found |
1954 |
* during a random, then cyclic scan, else null. This method must |
1955 |
* be retried by caller if, by the time it tries to use the queue, |
1956 |
* it is empty. |
1957 |
*/ |
1958 |
private WorkQueue findNonEmptyStealQueue(WorkQueue w) { |
1959 |
// Similar to loop in scan(), but ignoring submissions |
1960 |
int r; |
1961 |
if (w == null) // allow external callers |
1962 |
r = ThreadLocalRandom.current().nextInt(); |
1963 |
else { |
1964 |
r = w.seed; r ^= r << 13; r ^= r >>> 17; w.seed = r ^= r << 5; |
1965 |
} |
1966 |
int step = (r >>> 16) | 1; |
1967 |
for (WorkQueue[] ws;;) { |
1968 |
int rs = runState, m; |
1969 |
if ((ws = workQueues) == null || (m = ws.length - 1) < 1) |
1970 |
return null; |
1971 |
for (int j = (m + 1) << 2; ; r += step) { |
1972 |
WorkQueue q = ws[((r << 1) | 1) & m]; |
1973 |
if (q != null && !q.isEmpty()) |
1974 |
return q; |
1975 |
else if (--j < 0) { |
1976 |
if (runState == rs) |
1977 |
return null; |
1978 |
break; |
1979 |
} |
1980 |
} |
1981 |
} |
1982 |
} |
1983 |
|
1984 |
/** |
1985 |
* Runs tasks until {@code isQuiescent()}. We piggyback on |
1986 |
* active count ctl maintenance, but rather than blocking |
1987 |
* when tasks cannot be found, we rescan until all others cannot |
1988 |
* find tasks either. |
1989 |
*/ |
1990 |
final void helpQuiescePool(WorkQueue w) { |
1991 |
for (boolean active = true;;) { |
1992 |
ForkJoinTask<?> localTask; // exhaust local queue |
1993 |
while ((localTask = w.nextLocalTask()) != null) |
1994 |
localTask.doExec(); |
1995 |
WorkQueue q = findNonEmptyStealQueue(w); |
1996 |
if (q != null) { |
1997 |
ForkJoinTask<?> t; int b; |
1998 |
if (!active) { // re-establish active count |
1999 |
long c; |
2000 |
active = true; |
2001 |
do {} while (!U.compareAndSwapLong |
2002 |
(this, CTL, c = ctl, c + AC_UNIT)); |
2003 |
} |
2004 |
if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null) |
2005 |
w.runSubtask(t); |
2006 |
} |
2007 |
else { |
2008 |
long c; |
2009 |
if (active) { // decrement active count without queuing |
2010 |
active = false; |
2011 |
do {} while (!U.compareAndSwapLong |
2012 |
(this, CTL, c = ctl, c -= AC_UNIT)); |
2013 |
} |
2014 |
else |
2015 |
c = ctl; // re-increment on exit |
2016 |
if ((int)(c >> AC_SHIFT) + parallelism == 0) { |
2017 |
do {} while (!U.compareAndSwapLong |
2018 |
(this, CTL, c = ctl, c + AC_UNIT)); |
2019 |
break; |
2020 |
} |
2021 |
} |
2022 |
} |
2023 |
} |
2024 |
|
2025 |
/** |
2026 |
* Restricted version of helpQuiescePool for non-FJ callers |
2027 |
*/ |
2028 |
static void externalHelpQuiescePool() { |
2029 |
ForkJoinPool p; WorkQueue[] ws; WorkQueue w, q; |
2030 |
ForkJoinTask<?> t; int b; |
2031 |
int k = submitters.get().seed & SQMASK; |
2032 |
if ((p = commonPool) != null && |
2033 |
(ws = p.workQueues) != null && |
2034 |
ws.length > (k &= p.submitMask) && |
2035 |
(w = ws[k]) != null && |
2036 |
(q = p.findNonEmptyStealQueue(w)) != null && |
2037 |
(b = q.base) - q.top < 0 && |
2038 |
(t = q.pollAt(b)) != null) |
2039 |
t.doExec(); |
2040 |
} |
2041 |
|
2042 |
/** |
2043 |
* Gets and removes a local or stolen task for the given worker. |
2044 |
* |
2045 |
* @return a task, if available |
2046 |
*/ |
2047 |
final ForkJoinTask<?> nextTaskFor(WorkQueue w) { |
2048 |
for (ForkJoinTask<?> t;;) { |
2049 |
WorkQueue q; int b; |
2050 |
if ((t = w.nextLocalTask()) != null) |
2051 |
return t; |
2052 |
if ((q = findNonEmptyStealQueue(w)) == null) |
2053 |
return null; |
2054 |
if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null) |
2055 |
return t; |
2056 |
} |
2057 |
} |
2058 |
|
2059 |
/** |
2060 |
* Returns the approximate (non-atomic) number of idle threads per |
2061 |
* active thread to offset steal queue size for method |
2062 |
* ForkJoinTask.getSurplusQueuedTaskCount(). |
2063 |
*/ |
2064 |
final int idlePerActive() { |
2065 |
// Approximate at powers of two for small values, saturate past 4 |
2066 |
int p = parallelism; |
2067 |
int a = p + (int)(ctl >> AC_SHIFT); |
2068 |
return (a > (p >>>= 1) ? 0 : |
2069 |
a > (p >>>= 1) ? 1 : |
2070 |
a > (p >>>= 1) ? 2 : |
2071 |
a > (p >>>= 1) ? 4 : |
2072 |
8); |
2073 |
} |
2074 |
|
2075 |
/** |
2076 |
* Returns approximate submission queue length for the given caller |
2077 |
*/ |
2078 |
static int getEstimatedSubmitterQueueLength() { |
2079 |
ForkJoinPool p; WorkQueue[] ws; WorkQueue q; |
2080 |
int k = submitters.get().seed & SQMASK; |
2081 |
return ((p = commonPool) != null && |
2082 |
p.runState >= 0 && |
2083 |
(ws = p.workQueues) != null && |
2084 |
ws.length > (k &= p.submitMask) && |
2085 |
(q = ws[k]) != null) ? |
2086 |
q.queueSize() : 0; |
2087 |
} |
2088 |
|
2089 |
// Termination |
2090 |
|
2091 |
/** |
2092 |
* Possibly initiates and/or completes termination. The caller |
2093 |
* triggering termination runs three passes through workQueues: |
2094 |
* (0) Setting termination status, followed by wakeups of queued |
2095 |
* workers; (1) cancelling all tasks; (2) interrupting lagging |
2096 |
* threads (likely in external tasks, but possibly also blocked in |
2097 |
* joins). Each pass repeats previous steps because of potential |
2098 |
* lagging thread creation. |
2099 |
* |
2100 |
* @param now if true, unconditionally terminate, else only |
2101 |
* if no work and no active workers |
2102 |
* @param enable if true, enable shutdown when next possible |
2103 |
* @return true if now terminating or terminated |
2104 |
*/ |
2105 |
private boolean tryTerminate(boolean now, boolean enable) { |
2106 |
Mutex lock = this.lock; |
2107 |
for (long c;;) { |
2108 |
if (((c = ctl) & STOP_BIT) != 0) { // already terminating |
2109 |
if ((short)(c >>> TC_SHIFT) == -parallelism) { |
2110 |
lock.lock(); // don't need try/finally |
2111 |
termination.signalAll(); // signal when 0 workers |
2112 |
lock.unlock(); |
2113 |
} |
2114 |
return true; |
2115 |
} |
2116 |
if (runState >= 0) { // not yet enabled |
2117 |
if (!enable) |
2118 |
return false; |
2119 |
lock.lock(); |
2120 |
runState |= SHUTDOWN; |
2121 |
lock.unlock(); |
2122 |
} |
2123 |
if (!now) { // check if idle & no tasks |
2124 |
if ((int)(c >> AC_SHIFT) != -parallelism || |
2125 |
hasQueuedSubmissions()) |
2126 |
return false; |
2127 |
// Check for unqueued inactive workers. One pass suffices. |
2128 |
WorkQueue[] ws = workQueues; WorkQueue w; |
2129 |
if (ws != null) { |
2130 |
for (int i = 1; i < ws.length; i += 2) { |
2131 |
if ((w = ws[i]) != null && w.eventCount >= 0) |
2132 |
return false; |
2133 |
} |
2134 |
} |
2135 |
} |
2136 |
if (U.compareAndSwapLong(this, CTL, c, c | STOP_BIT)) { |
2137 |
for (int pass = 0; pass < 3; ++pass) { |
2138 |
WorkQueue[] ws = workQueues; |
2139 |
if (ws != null) { |
2140 |
WorkQueue w; |
2141 |
int n = ws.length; |
2142 |
for (int i = 0; i < n; ++i) { |
2143 |
if ((w = ws[i]) != null) { |
2144 |
w.runState = -1; |
2145 |
if (pass > 0) { |
2146 |
w.cancelAll(); |
2147 |
if (pass > 1) |
2148 |
w.interruptOwner(); |
2149 |
} |
2150 |
} |
2151 |
} |
2152 |
// Wake up workers parked on event queue |
2153 |
int i, e; long cc; Thread p; |
2154 |
while ((e = (int)(cc = ctl) & E_MASK) != 0 && |
2155 |
(i = e & SMASK) < n && |
2156 |
(w = ws[i]) != null) { |
2157 |
long nc = ((long)(w.nextWait & E_MASK) | |
2158 |
((cc + AC_UNIT) & AC_MASK) | |
2159 |
(cc & (TC_MASK|STOP_BIT))); |
2160 |
if (w.eventCount == (e | INT_SIGN) && |
2161 |
U.compareAndSwapLong(this, CTL, cc, nc)) { |
2162 |
w.eventCount = (e + E_SEQ) & E_MASK; |
2163 |
w.runState = -1; |
2164 |
if ((p = w.parker) != null) |
2165 |
U.unpark(p); |
2166 |
} |
2167 |
} |
2168 |
} |
2169 |
} |
2170 |
} |
2171 |
} |
2172 |
} |
2173 |
|
2174 |
// Exported methods |
2175 |
|
2176 |
// Constructors |
2177 |
|
2178 |
/** |
2179 |
* Creates a {@code ForkJoinPool} with parallelism equal to {@link |
2180 |
* java.lang.Runtime#availableProcessors}, using the {@linkplain |
2181 |
* #defaultForkJoinWorkerThreadFactory default thread factory}, |
2182 |
* no UncaughtExceptionHandler, and non-async LIFO processing mode. |
2183 |
* |
2184 |
* @throws SecurityException if a security manager exists and |
2185 |
* the caller is not permitted to modify threads |
2186 |
* because it does not hold {@link |
2187 |
* java.lang.RuntimePermission}{@code ("modifyThread")} |
2188 |
*/ |
2189 |
public ForkJoinPool() { |
2190 |
this(Runtime.getRuntime().availableProcessors(), |
2191 |
defaultForkJoinWorkerThreadFactory, null, false); |
2192 |
} |
2193 |
|
2194 |
/** |
2195 |
* Creates a {@code ForkJoinPool} with the indicated parallelism |
2196 |
* level, the {@linkplain |
2197 |
* #defaultForkJoinWorkerThreadFactory default thread factory}, |
2198 |
* no UncaughtExceptionHandler, and non-async LIFO processing mode. |
2199 |
* |
2200 |
* @param parallelism the parallelism level |
2201 |
* @throws IllegalArgumentException if parallelism less than or |
2202 |
* equal to zero, or greater than implementation limit |
2203 |
* @throws SecurityException if a security manager exists and |
2204 |
* the caller is not permitted to modify threads |
2205 |
* because it does not hold {@link |
2206 |
* java.lang.RuntimePermission}{@code ("modifyThread")} |
2207 |
*/ |
2208 |
public ForkJoinPool(int parallelism) { |
2209 |
this(parallelism, defaultForkJoinWorkerThreadFactory, null, false); |
2210 |
} |
2211 |
|
2212 |
/** |
2213 |
* Creates a {@code ForkJoinPool} with the given parameters. |
2214 |
* |
2215 |
* @param parallelism the parallelism level. For default value, |
2216 |
* use {@link java.lang.Runtime#availableProcessors}. |
2217 |
* @param factory the factory for creating new threads. For default value, |
2218 |
* use {@link #defaultForkJoinWorkerThreadFactory}. |
2219 |
* @param handler the handler for internal worker threads that |
2220 |
* terminate due to unrecoverable errors encountered while executing |
2221 |
* tasks. For default value, use {@code null}. |
2222 |
* @param asyncMode if true, |
2223 |
* establishes local first-in-first-out scheduling mode for forked |
2224 |
* tasks that are never joined. This mode may be more appropriate |
2225 |
* than default locally stack-based mode in applications in which |
2226 |
* worker threads only process event-style asynchronous tasks. |
2227 |
* For default value, use {@code false}. |
2228 |
* @throws IllegalArgumentException if parallelism less than or |
2229 |
* equal to zero, or greater than implementation limit |
2230 |
* @throws NullPointerException if the factory is null |
2231 |
* @throws SecurityException if a security manager exists and |
2232 |
* the caller is not permitted to modify threads |
2233 |
* because it does not hold {@link |
2234 |
* java.lang.RuntimePermission}{@code ("modifyThread")} |
2235 |
*/ |
2236 |
public ForkJoinPool(int parallelism, |
2237 |
ForkJoinWorkerThreadFactory factory, |
2238 |
Thread.UncaughtExceptionHandler handler, |
2239 |
boolean asyncMode) { |
2240 |
checkPermission(); |
2241 |
if (factory == null) |
2242 |
throw new NullPointerException(); |
2243 |
if (parallelism <= 0 || parallelism > MAX_CAP) |
2244 |
throw new IllegalArgumentException(); |
2245 |
this.parallelism = parallelism; |
2246 |
this.factory = factory; |
2247 |
this.ueh = handler; |
2248 |
this.localMode = asyncMode ? FIFO_QUEUE : LIFO_QUEUE; |
2249 |
long np = (long)(-parallelism); // offset ctl counts |
2250 |
this.ctl = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK); |
2251 |
// Use nearest power 2 for workQueues size. See Hackers Delight sec 3.2. |
2252 |
int n = parallelism - 1; |
2253 |
n |= n >>> 1; n |= n >>> 2; n |= n >>> 4; n |= n >>> 8; n |= n >>> 16; |
2254 |
int size = (n + 1) << 1; // #slots = 2*#workers |
2255 |
this.submitMask = size - 1; // room for max # of submit queues |
2256 |
this.workQueues = new WorkQueue[size]; |
2257 |
this.termination = (this.lock = new Mutex()).newCondition(); |
2258 |
this.stealCount = new AtomicLong(); |
2259 |
this.nextWorkerNumber = new AtomicInteger(); |
2260 |
int pn = poolNumberGenerator.incrementAndGet(); |
2261 |
StringBuilder sb = new StringBuilder("ForkJoinPool-"); |
2262 |
sb.append(Integer.toString(pn)); |
2263 |
sb.append("-worker-"); |
2264 |
this.workerNamePrefix = sb.toString(); |
2265 |
lock.lock(); |
2266 |
this.runState = 1; // set init flag |
2267 |
lock.unlock(); |
2268 |
} |
2269 |
|
2270 |
/** |
2271 |
* Returns the common pool instance |
2272 |
* |
2273 |
* @return the common pool instance |
2274 |
*/ |
2275 |
public static ForkJoinPool commonPool() { |
2276 |
ForkJoinPool p; |
2277 |
return (p = commonPool) != null? p : ensureCommonPool(); |
2278 |
} |
2279 |
|
2280 |
private static ForkJoinPool ensureCommonPool() { |
2281 |
ForkJoinPool p; |
2282 |
if ((p = commonPool) == null) { |
2283 |
final Mutex lock = initializationLock; |
2284 |
lock.lock(); |
2285 |
try { |
2286 |
if ((p = commonPool) == null) { |
2287 |
p = commonPool = new ForkJoinPool(commonPoolParallelism, |
2288 |
commonPoolFactory, |
2289 |
commonPoolUEH, false); |
2290 |
// use a more informative name string for workers |
2291 |
p.workerNamePrefix = "ForkJoinPool.commonPool-worker-"; |
2292 |
} |
2293 |
} finally { |
2294 |
lock.unlock(); |
2295 |
} |
2296 |
} |
2297 |
return p; |
2298 |
} |
2299 |
|
2300 |
// Execution methods |
2301 |
|
2302 |
/** |
2303 |
* Performs the given task, returning its result upon completion. |
2304 |
* If the computation encounters an unchecked Exception or Error, |
2305 |
* it is rethrown as the outcome of this invocation. Rethrown |
2306 |
* exceptions behave in the same way as regular exceptions, but, |
2307 |
* when possible, contain stack traces (as displayed for example |
2308 |
* using {@code ex.printStackTrace()}) of both the current thread |
2309 |
* as well as the thread actually encountering the exception; |
2310 |
* minimally only the latter. |
2311 |
* |
2312 |
* @param task the task |
2313 |
* @return the task's result |
2314 |
* @throws NullPointerException if the task is null |
2315 |
* @throws RejectedExecutionException if the task cannot be |
2316 |
* scheduled for execution |
2317 |
*/ |
2318 |
public <T> T invoke(ForkJoinTask<T> task) { |
2319 |
if (task == null) |
2320 |
throw new NullPointerException(); |
2321 |
doSubmit(task); |
2322 |
return task.join(); |
2323 |
} |
2324 |
|
2325 |
/** |
2326 |
* Arranges for (asynchronous) execution of the given task. |
2327 |
* |
2328 |
* @param task the task |
2329 |
* @throws NullPointerException if the task is null |
2330 |
* @throws RejectedExecutionException if the task cannot be |
2331 |
* scheduled for execution |
2332 |
*/ |
2333 |
public void execute(ForkJoinTask<?> task) { |
2334 |
if (task == null) |
2335 |
throw new NullPointerException(); |
2336 |
doSubmit(task); |
2337 |
} |
2338 |
|
2339 |
// AbstractExecutorService methods |
2340 |
|
2341 |
/** |
2342 |
* @throws NullPointerException if the task is null |
2343 |
* @throws RejectedExecutionException if the task cannot be |
2344 |
* scheduled for execution |
2345 |
*/ |
2346 |
public void execute(Runnable task) { |
2347 |
if (task == null) |
2348 |
throw new NullPointerException(); |
2349 |
ForkJoinTask<?> job; |
2350 |
if (task instanceof ForkJoinTask<?>) // avoid re-wrap |
2351 |
job = (ForkJoinTask<?>) task; |
2352 |
else |
2353 |
job = new ForkJoinTask.AdaptedRunnableAction(task); |
2354 |
doSubmit(job); |
2355 |
} |
2356 |
|
2357 |
/** |
2358 |
* Submits a ForkJoinTask for execution. |
2359 |
* |
2360 |
* @param task the task to submit |
2361 |
* @return the task |
2362 |
* @throws NullPointerException if the task is null |
2363 |
* @throws RejectedExecutionException if the task cannot be |
2364 |
* scheduled for execution |
2365 |
*/ |
2366 |
public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) { |
2367 |
if (task == null) |
2368 |
throw new NullPointerException(); |
2369 |
doSubmit(task); |
2370 |
return task; |
2371 |
} |
2372 |
|
2373 |
/** |
2374 |
* @throws NullPointerException if the task is null |
2375 |
* @throws RejectedExecutionException if the task cannot be |
2376 |
* scheduled for execution |
2377 |
*/ |
2378 |
public <T> ForkJoinTask<T> submit(Callable<T> task) { |
2379 |
ForkJoinTask<T> job = new ForkJoinTask.AdaptedCallable<T>(task); |
2380 |
doSubmit(job); |
2381 |
return job; |
2382 |
} |
2383 |
|
2384 |
/** |
2385 |
* @throws NullPointerException if the task is null |
2386 |
* @throws RejectedExecutionException if the task cannot be |
2387 |
* scheduled for execution |
2388 |
*/ |
2389 |
public <T> ForkJoinTask<T> submit(Runnable task, T result) { |
2390 |
ForkJoinTask<T> job = new ForkJoinTask.AdaptedRunnable<T>(task, result); |
2391 |
doSubmit(job); |
2392 |
return job; |
2393 |
} |
2394 |
|
2395 |
/** |
2396 |
* @throws NullPointerException if the task is null |
2397 |
* @throws RejectedExecutionException if the task cannot be |
2398 |
* scheduled for execution |
2399 |
*/ |
2400 |
public ForkJoinTask<?> submit(Runnable task) { |
2401 |
if (task == null) |
2402 |
throw new NullPointerException(); |
2403 |
ForkJoinTask<?> job; |
2404 |
if (task instanceof ForkJoinTask<?>) // avoid re-wrap |
2405 |
job = (ForkJoinTask<?>) task; |
2406 |
else |
2407 |
job = new ForkJoinTask.AdaptedRunnableAction(task); |
2408 |
doSubmit(job); |
2409 |
return job; |
2410 |
} |
2411 |
|
2412 |
/** |
2413 |
* @throws NullPointerException {@inheritDoc} |
2414 |
* @throws RejectedExecutionException {@inheritDoc} |
2415 |
*/ |
2416 |
public <T> List<Future<T>> invokeAll(Collection<? extends Callable<T>> tasks) { |
2417 |
// In previous versions of this class, this method constructed |
2418 |
// a task to run ForkJoinTask.invokeAll, but now external |
2419 |
// invocation of multiple tasks is at least as efficient. |
2420 |
List<ForkJoinTask<T>> fs = new ArrayList<ForkJoinTask<T>>(tasks.size()); |
2421 |
// Workaround needed because method wasn't declared with |
2422 |
// wildcards in return type but should have been. |
2423 |
@SuppressWarnings({"unchecked", "rawtypes"}) |
2424 |
List<Future<T>> futures = (List<Future<T>>) (List) fs; |
2425 |
|
2426 |
boolean done = false; |
2427 |
try { |
2428 |
for (Callable<T> t : tasks) { |
2429 |
ForkJoinTask<T> f = new ForkJoinTask.AdaptedCallable<T>(t); |
2430 |
doSubmit(f); |
2431 |
fs.add(f); |
2432 |
} |
2433 |
for (ForkJoinTask<T> f : fs) |
2434 |
f.quietlyJoin(); |
2435 |
done = true; |
2436 |
return futures; |
2437 |
} finally { |
2438 |
if (!done) |
2439 |
for (ForkJoinTask<T> f : fs) |
2440 |
f.cancel(false); |
2441 |
} |
2442 |
} |
2443 |
|
2444 |
/** |
2445 |
* Returns the factory used for constructing new workers. |
2446 |
* |
2447 |
* @return the factory used for constructing new workers |
2448 |
*/ |
2449 |
public ForkJoinWorkerThreadFactory getFactory() { |
2450 |
return factory; |
2451 |
} |
2452 |
|
2453 |
/** |
2454 |
* Returns the handler for internal worker threads that terminate |
2455 |
* due to unrecoverable errors encountered while executing tasks. |
2456 |
* |
2457 |
* @return the handler, or {@code null} if none |
2458 |
*/ |
2459 |
public Thread.UncaughtExceptionHandler getUncaughtExceptionHandler() { |
2460 |
return ueh; |
2461 |
} |
2462 |
|
2463 |
/** |
2464 |
* Returns the targeted parallelism level of this pool. |
2465 |
* |
2466 |
* @return the targeted parallelism level of this pool |
2467 |
*/ |
2468 |
public int getParallelism() { |
2469 |
return parallelism; |
2470 |
} |
2471 |
|
2472 |
/** |
2473 |
* Returns the targeted parallelism level of the common pool. |
2474 |
* |
2475 |
* @return the targeted parallelism level of the common pool |
2476 |
*/ |
2477 |
public static int getCommonPoolParallelism() { |
2478 |
return commonPoolParallelism; |
2479 |
} |
2480 |
|
2481 |
/** |
2482 |
* Returns the number of worker threads that have started but not |
2483 |
* yet terminated. The result returned by this method may differ |
2484 |
* from {@link #getParallelism} when threads are created to |
2485 |
* maintain parallelism when others are cooperatively blocked. |
2486 |
* |
2487 |
* @return the number of worker threads |
2488 |
*/ |
2489 |
public int getPoolSize() { |
2490 |
return parallelism + (short)(ctl >>> TC_SHIFT); |
2491 |
} |
2492 |
|
2493 |
/** |
2494 |
* Returns {@code true} if this pool uses local first-in-first-out |
2495 |
* scheduling mode for forked tasks that are never joined. |
2496 |
* |
2497 |
* @return {@code true} if this pool uses async mode |
2498 |
*/ |
2499 |
public boolean getAsyncMode() { |
2500 |
return localMode != 0; |
2501 |
} |
2502 |
|
2503 |
/** |
2504 |
* Returns an estimate of the number of worker threads that are |
2505 |
* not blocked waiting to join tasks or for other managed |
2506 |
* synchronization. This method may overestimate the |
2507 |
* number of running threads. |
2508 |
* |
2509 |
* @return the number of worker threads |
2510 |
*/ |
2511 |
public int getRunningThreadCount() { |
2512 |
int rc = 0; |
2513 |
WorkQueue[] ws; WorkQueue w; |
2514 |
if ((ws = workQueues) != null) { |
2515 |
for (int i = 1; i < ws.length; i += 2) { |
2516 |
if ((w = ws[i]) != null && w.isApparentlyUnblocked()) |
2517 |
++rc; |
2518 |
} |
2519 |
} |
2520 |
return rc; |
2521 |
} |
2522 |
|
2523 |
/** |
2524 |
* Returns an estimate of the number of threads that are currently |
2525 |
* stealing or executing tasks. This method may overestimate the |
2526 |
* number of active threads. |
2527 |
* |
2528 |
* @return the number of active threads |
2529 |
*/ |
2530 |
public int getActiveThreadCount() { |
2531 |
int r = parallelism + (int)(ctl >> AC_SHIFT); |
2532 |
return (r <= 0) ? 0 : r; // suppress momentarily negative values |
2533 |
} |
2534 |
|
2535 |
/** |
2536 |
* Returns {@code true} if all worker threads are currently idle. |
2537 |
* An idle worker is one that cannot obtain a task to execute |
2538 |
* because none are available to steal from other threads, and |
2539 |
* there are no pending submissions to the pool. This method is |
2540 |
* conservative; it might not return {@code true} immediately upon |
2541 |
* idleness of all threads, but will eventually become true if |
2542 |
* threads remain inactive. |
2543 |
* |
2544 |
* @return {@code true} if all threads are currently idle |
2545 |
*/ |
2546 |
public boolean isQuiescent() { |
2547 |
return (int)(ctl >> AC_SHIFT) + parallelism == 0; |
2548 |
} |
2549 |
|
2550 |
/** |
2551 |
* Returns an estimate of the total number of tasks stolen from |
2552 |
* one thread's work queue by another. The reported value |
2553 |
* underestimates the actual total number of steals when the pool |
2554 |
* is not quiescent. This value may be useful for monitoring and |
2555 |
* tuning fork/join programs: in general, steal counts should be |
2556 |
* high enough to keep threads busy, but low enough to avoid |
2557 |
* overhead and contention across threads. |
2558 |
* |
2559 |
* @return the number of steals |
2560 |
*/ |
2561 |
public long getStealCount() { |
2562 |
long count = stealCount.get(); |
2563 |
WorkQueue[] ws; WorkQueue w; |
2564 |
if ((ws = workQueues) != null) { |
2565 |
for (int i = 1; i < ws.length; i += 2) { |
2566 |
if ((w = ws[i]) != null) |
2567 |
count += w.totalSteals; |
2568 |
} |
2569 |
} |
2570 |
return count; |
2571 |
} |
2572 |
|
2573 |
/** |
2574 |
* Returns an estimate of the total number of tasks currently held |
2575 |
* in queues by worker threads (but not including tasks submitted |
2576 |
* to the pool that have not begun executing). This value is only |
2577 |
* an approximation, obtained by iterating across all threads in |
2578 |
* the pool. This method may be useful for tuning task |
2579 |
* granularities. |
2580 |
* |
2581 |
* @return the number of queued tasks |
2582 |
*/ |
2583 |
public long getQueuedTaskCount() { |
2584 |
long count = 0; |
2585 |
WorkQueue[] ws; WorkQueue w; |
2586 |
if ((ws = workQueues) != null) { |
2587 |
for (int i = 1; i < ws.length; i += 2) { |
2588 |
if ((w = ws[i]) != null) |
2589 |
count += w.queueSize(); |
2590 |
} |
2591 |
} |
2592 |
return count; |
2593 |
} |
2594 |
|
2595 |
/** |
2596 |
* Returns an estimate of the number of tasks submitted to this |
2597 |
* pool that have not yet begun executing. This method may take |
2598 |
* time proportional to the number of submissions. |
2599 |
* |
2600 |
* @return the number of queued submissions |
2601 |
*/ |
2602 |
public int getQueuedSubmissionCount() { |
2603 |
int count = 0; |
2604 |
WorkQueue[] ws; WorkQueue w; |
2605 |
if ((ws = workQueues) != null) { |
2606 |
for (int i = 0; i < ws.length; i += 2) { |
2607 |
if ((w = ws[i]) != null) |
2608 |
count += w.queueSize(); |
2609 |
} |
2610 |
} |
2611 |
return count; |
2612 |
} |
2613 |
|
2614 |
/** |
2615 |
* Returns {@code true} if there are any tasks submitted to this |
2616 |
* pool that have not yet begun executing. |
2617 |
* |
2618 |
* @return {@code true} if there are any queued submissions |
2619 |
*/ |
2620 |
public boolean hasQueuedSubmissions() { |
2621 |
WorkQueue[] ws; WorkQueue w; |
2622 |
if ((ws = workQueues) != null) { |
2623 |
for (int i = 0; i < ws.length; i += 2) { |
2624 |
if ((w = ws[i]) != null && !w.isEmpty()) |
2625 |
return true; |
2626 |
} |
2627 |
} |
2628 |
return false; |
2629 |
} |
2630 |
|
2631 |
/** |
2632 |
* Removes and returns the next unexecuted submission if one is |
2633 |
* available. This method may be useful in extensions to this |
2634 |
* class that re-assign work in systems with multiple pools. |
2635 |
* |
2636 |
* @return the next submission, or {@code null} if none |
2637 |
*/ |
2638 |
protected ForkJoinTask<?> pollSubmission() { |
2639 |
WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t; |
2640 |
if ((ws = workQueues) != null) { |
2641 |
for (int i = 0; i < ws.length; i += 2) { |
2642 |
if ((w = ws[i]) != null && (t = w.poll()) != null) |
2643 |
return t; |
2644 |
} |
2645 |
} |
2646 |
return null; |
2647 |
} |
2648 |
|
2649 |
/** |
2650 |
* Removes all available unexecuted submitted and forked tasks |
2651 |
* from scheduling queues and adds them to the given collection, |
2652 |
* without altering their execution status. These may include |
2653 |
* artificially generated or wrapped tasks. This method is |
2654 |
* designed to be invoked only when the pool is known to be |
2655 |
* quiescent. Invocations at other times may not remove all |
2656 |
* tasks. A failure encountered while attempting to add elements |
2657 |
* to collection {@code c} may result in elements being in |
2658 |
* neither, either or both collections when the associated |
2659 |
* exception is thrown. The behavior of this operation is |
2660 |
* undefined if the specified collection is modified while the |
2661 |
* operation is in progress. |
2662 |
* |
2663 |
* @param c the collection to transfer elements into |
2664 |
* @return the number of elements transferred |
2665 |
*/ |
2666 |
protected int drainTasksTo(Collection<? super ForkJoinTask<?>> c) { |
2667 |
int count = 0; |
2668 |
WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t; |
2669 |
if ((ws = workQueues) != null) { |
2670 |
for (int i = 0; i < ws.length; ++i) { |
2671 |
if ((w = ws[i]) != null) { |
2672 |
while ((t = w.poll()) != null) { |
2673 |
c.add(t); |
2674 |
++count; |
2675 |
} |
2676 |
} |
2677 |
} |
2678 |
} |
2679 |
return count; |
2680 |
} |
2681 |
|
2682 |
/** |
2683 |
* Returns a string identifying this pool, as well as its state, |
2684 |
* including indications of run state, parallelism level, and |
2685 |
* worker and task counts. |
2686 |
* |
2687 |
* @return a string identifying this pool, as well as its state |
2688 |
*/ |
2689 |
public String toString() { |
2690 |
// Use a single pass through workQueues to collect counts |
2691 |
long qt = 0L, qs = 0L; int rc = 0; |
2692 |
long st = stealCount.get(); |
2693 |
long c = ctl; |
2694 |
WorkQueue[] ws; WorkQueue w; |
2695 |
if ((ws = workQueues) != null) { |
2696 |
for (int i = 0; i < ws.length; ++i) { |
2697 |
if ((w = ws[i]) != null) { |
2698 |
int size = w.queueSize(); |
2699 |
if ((i & 1) == 0) |
2700 |
qs += size; |
2701 |
else { |
2702 |
qt += size; |
2703 |
st += w.totalSteals; |
2704 |
if (w.isApparentlyUnblocked()) |
2705 |
++rc; |
2706 |
} |
2707 |
} |
2708 |
} |
2709 |
} |
2710 |
int pc = parallelism; |
2711 |
int tc = pc + (short)(c >>> TC_SHIFT); |
2712 |
int ac = pc + (int)(c >> AC_SHIFT); |
2713 |
if (ac < 0) // ignore transient negative |
2714 |
ac = 0; |
2715 |
String level; |
2716 |
if ((c & STOP_BIT) != 0) |
2717 |
level = (tc == 0) ? "Terminated" : "Terminating"; |
2718 |
else |
2719 |
level = runState < 0 ? "Shutting down" : "Running"; |
2720 |
return super.toString() + |
2721 |
"[" + level + |
2722 |
", parallelism = " + pc + |
2723 |
", size = " + tc + |
2724 |
", active = " + ac + |
2725 |
", running = " + rc + |
2726 |
", steals = " + st + |
2727 |
", tasks = " + qt + |
2728 |
", submissions = " + qs + |
2729 |
"]"; |
2730 |
} |
2731 |
|
2732 |
/** |
2733 |
* Possibly initiates an orderly shutdown in which previously |
2734 |
* submitted tasks are executed, but no new tasks will be |
2735 |
* accepted. Invocation has no effect on execution state if this |
2736 |
* is the {@link #commonPool}, and no additional effect if |
2737 |
* already shut down. Tasks that are in the process of being |
2738 |
* submitted concurrently during the course of this method may or |
2739 |
* may not be rejected. |
2740 |
* |
2741 |
* @throws SecurityException if a security manager exists and |
2742 |
* the caller is not permitted to modify threads |
2743 |
* because it does not hold {@link |
2744 |
* java.lang.RuntimePermission}{@code ("modifyThread")} |
2745 |
*/ |
2746 |
public void shutdown() { |
2747 |
checkPermission(); |
2748 |
if (this != commonPool) |
2749 |
tryTerminate(false, true); |
2750 |
} |
2751 |
|
2752 |
/** |
2753 |
* Possibly attempts to cancel and/or stop all tasks, and reject |
2754 |
* all subsequently submitted tasks. Invocation has no effect on |
2755 |
* execution state if this is the {@link #commonPool}, and no |
2756 |
* additional effect if already shut down. Otherwise, tasks that |
2757 |
* are in the process of being submitted or executed concurrently |
2758 |
* during the course of this method may or may not be |
2759 |
* rejected. This method cancels both existing and unexecuted |
2760 |
* tasks, in order to permit termination in the presence of task |
2761 |
* dependencies. So the method always returns an empty list |
2762 |
* (unlike the case for some other Executors). |
2763 |
* |
2764 |
* @return an empty list |
2765 |
* @throws SecurityException if a security manager exists and |
2766 |
* the caller is not permitted to modify threads |
2767 |
* because it does not hold {@link |
2768 |
* java.lang.RuntimePermission}{@code ("modifyThread")} |
2769 |
*/ |
2770 |
public List<Runnable> shutdownNow() { |
2771 |
checkPermission(); |
2772 |
if (this != commonPool) |
2773 |
tryTerminate(true, true); |
2774 |
return Collections.emptyList(); |
2775 |
} |
2776 |
|
2777 |
/** |
2778 |
* Returns {@code true} if all tasks have completed following shut down. |
2779 |
* |
2780 |
* @return {@code true} if all tasks have completed following shut down |
2781 |
*/ |
2782 |
public boolean isTerminated() { |
2783 |
long c = ctl; |
2784 |
return ((c & STOP_BIT) != 0L && |
2785 |
(short)(c >>> TC_SHIFT) == -parallelism); |
2786 |
} |
2787 |
|
2788 |
/** |
2789 |
* Returns {@code true} if the process of termination has |
2790 |
* commenced but not yet completed. This method may be useful for |
2791 |
* debugging. A return of {@code true} reported a sufficient |
2792 |
* period after shutdown may indicate that submitted tasks have |
2793 |
* ignored or suppressed interruption, or are waiting for IO, |
2794 |
* causing this executor not to properly terminate. (See the |
2795 |
* advisory notes for class {@link ForkJoinTask} stating that |
2796 |
* tasks should not normally entail blocking operations. But if |
2797 |
* they do, they must abort them on interrupt.) |
2798 |
* |
2799 |
* @return {@code true} if terminating but not yet terminated |
2800 |
*/ |
2801 |
public boolean isTerminating() { |
2802 |
long c = ctl; |
2803 |
return ((c & STOP_BIT) != 0L && |
2804 |
(short)(c >>> TC_SHIFT) != -parallelism); |
2805 |
} |
2806 |
|
2807 |
/** |
2808 |
* Returns {@code true} if this pool has been shut down. |
2809 |
* |
2810 |
* @return {@code true} if this pool has been shut down |
2811 |
*/ |
2812 |
public boolean isShutdown() { |
2813 |
return runState < 0; |
2814 |
} |
2815 |
|
2816 |
/** |
2817 |
* Blocks until all tasks have completed execution after a shutdown |
2818 |
* request, or the timeout occurs, or the current thread is |
2819 |
* interrupted, whichever happens first. |
2820 |
* |
2821 |
* @param timeout the maximum time to wait |
2822 |
* @param unit the time unit of the timeout argument |
2823 |
* @return {@code true} if this executor terminated and |
2824 |
* {@code false} if the timeout elapsed before termination |
2825 |
* @throws InterruptedException if interrupted while waiting |
2826 |
*/ |
2827 |
public boolean awaitTermination(long timeout, TimeUnit unit) |
2828 |
throws InterruptedException { |
2829 |
long nanos = unit.toNanos(timeout); |
2830 |
final Mutex lock = this.lock; |
2831 |
lock.lock(); |
2832 |
try { |
2833 |
for (;;) { |
2834 |
if (isTerminated()) |
2835 |
return true; |
2836 |
if (nanos <= 0) |
2837 |
return false; |
2838 |
nanos = termination.awaitNanos(nanos); |
2839 |
} |
2840 |
} finally { |
2841 |
lock.unlock(); |
2842 |
} |
2843 |
} |
2844 |
|
2845 |
/** |
2846 |
* Interface for extending managed parallelism for tasks running |
2847 |
* in {@link ForkJoinPool}s. |
2848 |
* |
2849 |
* <p>A {@code ManagedBlocker} provides two methods. Method |
2850 |
* {@code isReleasable} must return {@code true} if blocking is |
2851 |
* not necessary. Method {@code block} blocks the current thread |
2852 |
* if necessary (perhaps internally invoking {@code isReleasable} |
2853 |
* before actually blocking). These actions are performed by any |
2854 |
* thread invoking {@link ForkJoinPool#managedBlock}. The |
2855 |
* unusual methods in this API accommodate synchronizers that may, |
2856 |
* but don't usually, block for long periods. Similarly, they |
2857 |
* allow more efficient internal handling of cases in which |
2858 |
* additional workers may be, but usually are not, needed to |
2859 |
* ensure sufficient parallelism. Toward this end, |
2860 |
* implementations of method {@code isReleasable} must be amenable |
2861 |
* to repeated invocation. |
2862 |
* |
2863 |
* <p>For example, here is a ManagedBlocker based on a |
2864 |
* ReentrantLock: |
2865 |
* <pre> {@code |
2866 |
* class ManagedLocker implements ManagedBlocker { |
2867 |
* final ReentrantLock lock; |
2868 |
* boolean hasLock = false; |
2869 |
* ManagedLocker(ReentrantLock lock) { this.lock = lock; } |
2870 |
* public boolean block() { |
2871 |
* if (!hasLock) |
2872 |
* lock.lock(); |
2873 |
* return true; |
2874 |
* } |
2875 |
* public boolean isReleasable() { |
2876 |
* return hasLock || (hasLock = lock.tryLock()); |
2877 |
* } |
2878 |
* }}</pre> |
2879 |
* |
2880 |
* <p>Here is a class that possibly blocks waiting for an |
2881 |
* item on a given queue: |
2882 |
* <pre> {@code |
2883 |
* class QueueTaker<E> implements ManagedBlocker { |
2884 |
* final BlockingQueue<E> queue; |
2885 |
* volatile E item = null; |
2886 |
* QueueTaker(BlockingQueue<E> q) { this.queue = q; } |
2887 |
* public boolean block() throws InterruptedException { |
2888 |
* if (item == null) |
2889 |
* item = queue.take(); |
2890 |
* return true; |
2891 |
* } |
2892 |
* public boolean isReleasable() { |
2893 |
* return item != null || (item = queue.poll()) != null; |
2894 |
* } |
2895 |
* public E getItem() { // call after pool.managedBlock completes |
2896 |
* return item; |
2897 |
* } |
2898 |
* }}</pre> |
2899 |
*/ |
2900 |
public static interface ManagedBlocker { |
2901 |
/** |
2902 |
* Possibly blocks the current thread, for example waiting for |
2903 |
* a lock or condition. |
2904 |
* |
2905 |
* @return {@code true} if no additional blocking is necessary |
2906 |
* (i.e., if isReleasable would return true) |
2907 |
* @throws InterruptedException if interrupted while waiting |
2908 |
* (the method is not required to do so, but is allowed to) |
2909 |
*/ |
2910 |
boolean block() throws InterruptedException; |
2911 |
|
2912 |
/** |
2913 |
* Returns {@code true} if blocking is unnecessary. |
2914 |
*/ |
2915 |
boolean isReleasable(); |
2916 |
} |
2917 |
|
2918 |
/** |
2919 |
* Blocks in accord with the given blocker. If the current thread |
2920 |
* is a {@link ForkJoinWorkerThread}, this method possibly |
2921 |
* arranges for a spare thread to be activated if necessary to |
2922 |
* ensure sufficient parallelism while the current thread is blocked. |
2923 |
* |
2924 |
* <p>If the caller is not a {@link ForkJoinTask}, this method is |
2925 |
* behaviorally equivalent to |
2926 |
* <pre> {@code |
2927 |
* while (!blocker.isReleasable()) |
2928 |
* if (blocker.block()) |
2929 |
* return; |
2930 |
* }</pre> |
2931 |
* |
2932 |
* If the caller is a {@code ForkJoinTask}, then the pool may |
2933 |
* first be expanded to ensure parallelism, and later adjusted. |
2934 |
* |
2935 |
* @param blocker the blocker |
2936 |
* @throws InterruptedException if blocker.block did so |
2937 |
*/ |
2938 |
public static void managedBlock(ManagedBlocker blocker) |
2939 |
throws InterruptedException { |
2940 |
Thread t = Thread.currentThread(); |
2941 |
ForkJoinPool p = ((t instanceof ForkJoinWorkerThread) ? |
2942 |
((ForkJoinWorkerThread)t).pool : null); |
2943 |
while (!blocker.isReleasable()) { |
2944 |
if (p == null || p.tryCompensate(null, blocker)) { |
2945 |
try { |
2946 |
do {} while (!blocker.isReleasable() && !blocker.block()); |
2947 |
} finally { |
2948 |
if (p != null) |
2949 |
p.incrementActiveCount(); |
2950 |
} |
2951 |
break; |
2952 |
} |
2953 |
} |
2954 |
} |
2955 |
|
2956 |
// AbstractExecutorService overrides. These rely on undocumented |
2957 |
// fact that ForkJoinTask.adapt returns ForkJoinTasks that also |
2958 |
// implement RunnableFuture. |
2959 |
|
2960 |
protected <T> RunnableFuture<T> newTaskFor(Runnable runnable, T value) { |
2961 |
return new ForkJoinTask.AdaptedRunnable<T>(runnable, value); |
2962 |
} |
2963 |
|
2964 |
protected <T> RunnableFuture<T> newTaskFor(Callable<T> callable) { |
2965 |
return new ForkJoinTask.AdaptedCallable<T>(callable); |
2966 |
} |
2967 |
|
2968 |
// Unsafe mechanics |
2969 |
private static final sun.misc.Unsafe U; |
2970 |
private static final long CTL; |
2971 |
private static final long PARKBLOCKER; |
2972 |
private static final int ABASE; |
2973 |
private static final int ASHIFT; |
2974 |
|
2975 |
static { |
2976 |
poolNumberGenerator = new AtomicInteger(); |
2977 |
nextSubmitterSeed = new AtomicInteger(0x55555555); |
2978 |
modifyThreadPermission = new RuntimePermission("modifyThread"); |
2979 |
defaultForkJoinWorkerThreadFactory = |
2980 |
new DefaultForkJoinWorkerThreadFactory(); |
2981 |
submitters = new ThreadSubmitter(); |
2982 |
initializationLock = new Mutex(); |
2983 |
int s; |
2984 |
try { |
2985 |
U = getUnsafe(); |
2986 |
Class<?> k = ForkJoinPool.class; |
2987 |
Class<?> ak = ForkJoinTask[].class; |
2988 |
CTL = U.objectFieldOffset |
2989 |
(k.getDeclaredField("ctl")); |
2990 |
Class<?> tk = Thread.class; |
2991 |
PARKBLOCKER = U.objectFieldOffset |
2992 |
(tk.getDeclaredField("parkBlocker")); |
2993 |
ABASE = U.arrayBaseOffset(ak); |
2994 |
s = U.arrayIndexScale(ak); |
2995 |
} catch (Exception e) { |
2996 |
throw new Error(e); |
2997 |
} |
2998 |
if ((s & (s-1)) != 0) |
2999 |
throw new Error("data type scale not a power of two"); |
3000 |
ASHIFT = 31 - Integer.numberOfLeadingZeros(s); |
3001 |
|
3002 |
// Establish configuration for default pool |
3003 |
try { |
3004 |
String pp = System.getProperty(propPrefix + "parallelism"); |
3005 |
String fp = System.getProperty(propPrefix + "threadFactory"); |
3006 |
String up = System.getProperty(propPrefix + "exceptionHandler"); |
3007 |
int par; |
3008 |
if ((pp == null || (par = Integer.parseInt(pp)) <= 0)) |
3009 |
par = Runtime.getRuntime().availableProcessors(); |
3010 |
commonPoolParallelism = par; |
3011 |
if (fp != null) |
3012 |
commonPoolFactory = (ForkJoinWorkerThreadFactory) |
3013 |
ClassLoader.getSystemClassLoader().loadClass(fp).newInstance(); |
3014 |
else |
3015 |
commonPoolFactory = defaultForkJoinWorkerThreadFactory; |
3016 |
if (up != null) |
3017 |
commonPoolUEH = (Thread.UncaughtExceptionHandler) |
3018 |
ClassLoader.getSystemClassLoader().loadClass(up).newInstance(); |
3019 |
else |
3020 |
commonPoolUEH = null; |
3021 |
} catch (Exception e) { |
3022 |
throw new Error(e); |
3023 |
} |
3024 |
} |
3025 |
|
3026 |
/** |
3027 |
* Returns a sun.misc.Unsafe. Suitable for use in a 3rd party package. |
3028 |
* Replace with a simple call to Unsafe.getUnsafe when integrating |
3029 |
* into a jdk. |
3030 |
* |
3031 |
* @return a sun.misc.Unsafe |
3032 |
*/ |
3033 |
private static sun.misc.Unsafe getUnsafe() { |
3034 |
try { |
3035 |
return sun.misc.Unsafe.getUnsafe(); |
3036 |
} catch (SecurityException se) { |
3037 |
try { |
3038 |
return java.security.AccessController.doPrivileged |
3039 |
(new java.security |
3040 |
.PrivilegedExceptionAction<sun.misc.Unsafe>() { |
3041 |
public sun.misc.Unsafe run() throws Exception { |
3042 |
java.lang.reflect.Field f = sun.misc |
3043 |
.Unsafe.class.getDeclaredField("theUnsafe"); |
3044 |
f.setAccessible(true); |
3045 |
return (sun.misc.Unsafe) f.get(null); |
3046 |
}}); |
3047 |
} catch (java.security.PrivilegedActionException e) { |
3048 |
throw new RuntimeException("Could not initialize intrinsics", |
3049 |
e.getCause()); |
3050 |
} |
3051 |
} |
3052 |
} |
3053 |
|
3054 |
} |