1 |
/* |
2 |
* Written by Doug Lea with assistance from members of JCP JSR-166 |
3 |
* Expert Group and released to the public domain, as explained at |
4 |
* http://creativecommons.org/publicdomain/zero/1.0/ |
5 |
*/ |
6 |
|
7 |
package java.util.concurrent; |
8 |
|
9 |
import java.lang.Thread.UncaughtExceptionHandler; |
10 |
import java.security.AccessControlContext; |
11 |
import java.security.Permissions; |
12 |
import java.security.ProtectionDomain; |
13 |
import java.util.ArrayList; |
14 |
import java.util.Arrays; |
15 |
import java.util.Collection; |
16 |
import java.util.Collections; |
17 |
import java.util.List; |
18 |
import java.util.concurrent.atomic.AtomicLong; |
19 |
|
20 |
/** |
21 |
* An {@link ExecutorService} for running {@link ForkJoinTask}s. |
22 |
* A {@code ForkJoinPool} provides the entry point for submissions |
23 |
* from non-{@code ForkJoinTask} clients, as well as management and |
24 |
* monitoring operations. |
25 |
* |
26 |
* <p>A {@code ForkJoinPool} differs from other kinds of {@link |
27 |
* ExecutorService} mainly by virtue of employing |
28 |
* <em>work-stealing</em>: all threads in the pool attempt to find and |
29 |
* execute tasks submitted to the pool and/or created by other active |
30 |
* tasks (eventually blocking waiting for work if none exist). This |
31 |
* enables efficient processing when most tasks spawn other subtasks |
32 |
* (as do most {@code ForkJoinTask}s), as well as when many small |
33 |
* tasks are submitted to the pool from external clients. Especially |
34 |
* when setting <em>asyncMode</em> to true in constructors, {@code |
35 |
* ForkJoinPool}s may also be appropriate for use with event-style |
36 |
* tasks that are never joined. |
37 |
* |
38 |
* <p>A static {@link #commonPool()} is available and appropriate for |
39 |
* most applications. The common pool is used by any ForkJoinTask that |
40 |
* is not explicitly submitted to a specified pool. Using the common |
41 |
* pool normally reduces resource usage (its threads are slowly |
42 |
* reclaimed during periods of non-use, and reinstated upon subsequent |
43 |
* use). |
44 |
* |
45 |
* <p>For applications that require separate or custom pools, a {@code |
46 |
* ForkJoinPool} may be constructed with a given target parallelism |
47 |
* level; by default, equal to the number of available processors. |
48 |
* The pool attempts to maintain enough active (or available) threads |
49 |
* by dynamically adding, suspending, or resuming internal worker |
50 |
* threads, even if some tasks are stalled waiting to join others. |
51 |
* However, no such adjustments are guaranteed in the face of blocked |
52 |
* I/O or other unmanaged synchronization. The nested {@link |
53 |
* ManagedBlocker} interface enables extension of the kinds of |
54 |
* synchronization accommodated. |
55 |
* |
56 |
* <p>In addition to execution and lifecycle control methods, this |
57 |
* class provides status check methods (for example |
58 |
* {@link #getStealCount}) that are intended to aid in developing, |
59 |
* tuning, and monitoring fork/join applications. Also, method |
60 |
* {@link #toString} returns indications of pool state in a |
61 |
* convenient form for informal monitoring. |
62 |
* |
63 |
* <p>As is the case with other ExecutorServices, there are three |
64 |
* main task execution methods summarized in the following table. |
65 |
* These are designed to be used primarily by clients not already |
66 |
* engaged in fork/join computations in the current pool. The main |
67 |
* forms of these methods accept instances of {@code ForkJoinTask}, |
68 |
* but overloaded forms also allow mixed execution of plain {@code |
69 |
* Runnable}- or {@code Callable}- based activities as well. However, |
70 |
* tasks that are already executing in a pool should normally instead |
71 |
* use the within-computation forms listed in the table unless using |
72 |
* async event-style tasks that are not usually joined, in which case |
73 |
* there is little difference among choice of methods. |
74 |
* |
75 |
* <table BORDER CELLPADDING=3 CELLSPACING=1> |
76 |
* <caption>Summary of task execution methods</caption> |
77 |
* <tr> |
78 |
* <td></td> |
79 |
* <td ALIGN=CENTER> <b>Call from non-fork/join clients</b></td> |
80 |
* <td ALIGN=CENTER> <b>Call from within fork/join computations</b></td> |
81 |
* </tr> |
82 |
* <tr> |
83 |
* <td> <b>Arrange async execution</b></td> |
84 |
* <td> {@link #execute(ForkJoinTask)}</td> |
85 |
* <td> {@link ForkJoinTask#fork}</td> |
86 |
* </tr> |
87 |
* <tr> |
88 |
* <td> <b>Await and obtain result</b></td> |
89 |
* <td> {@link #invoke(ForkJoinTask)}</td> |
90 |
* <td> {@link ForkJoinTask#invoke}</td> |
91 |
* </tr> |
92 |
* <tr> |
93 |
* <td> <b>Arrange exec and obtain Future</b></td> |
94 |
* <td> {@link #submit(ForkJoinTask)}</td> |
95 |
* <td> {@link ForkJoinTask#fork} (ForkJoinTasks <em>are</em> Futures)</td> |
96 |
* </tr> |
97 |
* </table> |
98 |
* |
99 |
* <p>The common pool is by default constructed with default |
100 |
* parameters, but these may be controlled by setting three |
101 |
* {@linkplain System#getProperty system properties}: |
102 |
* <ul> |
103 |
* <li>{@code java.util.concurrent.ForkJoinPool.common.parallelism} |
104 |
* - the parallelism level, a non-negative integer |
105 |
* <li>{@code java.util.concurrent.ForkJoinPool.common.threadFactory} |
106 |
* - the class name of a {@link ForkJoinWorkerThreadFactory} |
107 |
* <li>{@code java.util.concurrent.ForkJoinPool.common.exceptionHandler} |
108 |
* - the class name of a {@link UncaughtExceptionHandler} |
109 |
* <li>{@code java.util.concurrent.ForkJoinPool.common.maximumSpares} |
110 |
* - the maximum number of allowed extra threads to maintain target |
111 |
* parallelism (default 256). |
112 |
* </ul> |
113 |
* If a {@link SecurityManager} is present and no factory is |
114 |
* specified, then the default pool uses a factory supplying |
115 |
* threads that have no {@link Permissions} enabled. |
116 |
* The system class loader is used to load these classes. |
117 |
* Upon any error in establishing these settings, default parameters |
118 |
* are used. It is possible to disable or limit the use of threads in |
119 |
* the common pool by setting the parallelism property to zero, and/or |
120 |
* using a factory that may return {@code null}. However doing so may |
121 |
* cause unjoined tasks to never be executed. |
122 |
* |
123 |
* <p><b>Implementation notes</b>: This implementation restricts the |
124 |
* maximum number of running threads to 32767. Attempts to create |
125 |
* pools with greater than the maximum number result in |
126 |
* {@code IllegalArgumentException}. |
127 |
* |
128 |
* <p>This implementation rejects submitted tasks (that is, by throwing |
129 |
* {@link RejectedExecutionException}) only when the pool is shut down |
130 |
* or internal resources have been exhausted. |
131 |
* |
132 |
* @since 1.7 |
133 |
* @author Doug Lea |
134 |
*/ |
135 |
@sun.misc.Contended |
136 |
public class ForkJoinPool extends AbstractExecutorService { |
137 |
|
138 |
/* |
139 |
* Implementation Overview |
140 |
* |
141 |
* This class and its nested classes provide the main |
142 |
* functionality and control for a set of worker threads: |
143 |
* Submissions from non-FJ threads enter into submission queues. |
144 |
* Workers take these tasks and typically split them into subtasks |
145 |
* that may be stolen by other workers. Preference rules give |
146 |
* first priority to processing tasks from their own queues (LIFO |
147 |
* or FIFO, depending on mode), then to randomized FIFO steals of |
148 |
* tasks in other queues. This framework began as vehicle for |
149 |
* supporting tree-structured parallelism using work-stealing. |
150 |
* Over time, its scalability advantages led to extensions and |
151 |
* changes to better support more diverse usage contexts. Because |
152 |
* most internal methods and nested classes are interrelated, |
153 |
* their main rationale and descriptions are presented here; |
154 |
* individual methods and nested classes contain only brief |
155 |
* comments about details. |
156 |
* |
157 |
* WorkQueues |
158 |
* ========== |
159 |
* |
160 |
* Most operations occur within work-stealing queues (in nested |
161 |
* class WorkQueue). These are special forms of Deques that |
162 |
* support only three of the four possible end-operations -- push, |
163 |
* pop, and poll (aka steal), under the further constraints that |
164 |
* push and pop are called only from the owning thread (or, as |
165 |
* extended here, under a lock), while poll may be called from |
166 |
* other threads. (If you are unfamiliar with them, you probably |
167 |
* want to read Herlihy and Shavit's book "The Art of |
168 |
* Multiprocessor programming", chapter 16 describing these in |
169 |
* more detail before proceeding.) The main work-stealing queue |
170 |
* design is roughly similar to those in the papers "Dynamic |
171 |
* Circular Work-Stealing Deque" by Chase and Lev, SPAA 2005 |
172 |
* (http://research.sun.com/scalable/pubs/index.html) and |
173 |
* "Idempotent work stealing" by Michael, Saraswat, and Vechev, |
174 |
* PPoPP 2009 (http://portal.acm.org/citation.cfm?id=1504186). |
175 |
* The main differences ultimately stem from GC requirements that |
176 |
* we null out taken slots as soon as we can, to maintain as small |
177 |
* a footprint as possible even in programs generating huge |
178 |
* numbers of tasks. To accomplish this, we shift the CAS |
179 |
* arbitrating pop vs poll (steal) from being on the indices |
180 |
* ("base" and "top") to the slots themselves. |
181 |
* |
182 |
* Adding tasks then takes the form of a classic array push(task): |
183 |
* q.array[q.top] = task; ++q.top; |
184 |
* |
185 |
* (The actual code needs to null-check and size-check the array, |
186 |
* properly fence the accesses, and possibly signal waiting |
187 |
* workers to start scanning -- see below.) Both a successful pop |
188 |
* and poll mainly entail a CAS of a slot from non-null to null. |
189 |
* |
190 |
* The pop operation (always performed by owner) is: |
191 |
* if ((base != top) and |
192 |
* (the task at top slot is not null) and |
193 |
* (CAS slot to null)) |
194 |
* decrement top and return task; |
195 |
* |
196 |
* And the poll operation (usually by a stealer) is |
197 |
* if ((base != top) and |
198 |
* (the task at base slot is not null) and |
199 |
* (base has not changed) and |
200 |
* (CAS slot to null)) |
201 |
* increment base and return task; |
202 |
* |
203 |
* Because we rely on CASes of references, we do not need tag bits |
204 |
* on base or top. They are simple ints as used in any circular |
205 |
* array-based queue (see for example ArrayDeque). Updates to the |
206 |
* indices guarantee that top == base means the queue is empty, |
207 |
* but otherwise may err on the side of possibly making the queue |
208 |
* appear nonempty when a push, pop, or poll have not fully |
209 |
* committed. (Method isEmpty() checks the case of a partially |
210 |
* completed removal of the last element.) Because of this, the |
211 |
* poll operation, considered individually, is not wait-free. One |
212 |
* thief cannot successfully continue until another in-progress |
213 |
* one (or, if previously empty, a push) completes. However, in |
214 |
* the aggregate, we ensure at least probabilistic |
215 |
* non-blockingness. If an attempted steal fails, a thief always |
216 |
* chooses a different random victim target to try next. So, in |
217 |
* order for one thief to progress, it suffices for any |
218 |
* in-progress poll or new push on any empty queue to |
219 |
* complete. (This is why we normally use method pollAt and its |
220 |
* variants that try once at the apparent base index, else |
221 |
* consider alternative actions, rather than method poll, which |
222 |
* retries.) |
223 |
* |
224 |
* This approach also enables support of a user mode in which |
225 |
* local task processing is in FIFO, not LIFO order, simply by |
226 |
* using poll rather than pop. This can be useful in |
227 |
* message-passing frameworks in which tasks are never joined. |
228 |
* However neither mode considers affinities, loads, cache |
229 |
* localities, etc, so rarely provide the best possible |
230 |
* performance on a given machine, but portably provide good |
231 |
* throughput by averaging over these factors. Further, even if |
232 |
* we did try to use such information, we do not usually have a |
233 |
* basis for exploiting it. For example, some sets of tasks |
234 |
* profit from cache affinities, but others are harmed by cache |
235 |
* pollution effects. Additionally, even though it requires |
236 |
* scanning, long-term throughput is often best using random |
237 |
* selection rather than directed selection policies, so cheap |
238 |
* randomization of sufficient quality is used whenever |
239 |
* applicable. Various Marsaglia XorShifts (some with different |
240 |
* shift constants) are inlined at use points. |
241 |
* |
242 |
* WorkQueues are also used in a similar way for tasks submitted |
243 |
* to the pool. We cannot mix these tasks in the same queues used |
244 |
* by workers. Instead, we randomly associate submission queues |
245 |
* with submitting threads, using a form of hashing. The |
246 |
* ThreadLocalRandom probe value serves as a hash code for |
247 |
* choosing existing queues, and may be randomly repositioned upon |
248 |
* contention with other submitters. In essence, submitters act |
249 |
* like workers except that they are restricted to executing local |
250 |
* tasks that they submitted (or in the case of CountedCompleters, |
251 |
* others with the same root task). Insertion of tasks in shared |
252 |
* mode requires a lock (mainly to protect in the case of |
253 |
* resizing) but we use only a simple spinlock (using field |
254 |
* qlock), because submitters encountering a busy queue move on to |
255 |
* try or create other queues -- they block only when creating and |
256 |
* registering new queues. Additionally, "qlock" saturates to an |
257 |
* unlockable value (-1) at shutdown. Unlocking still can be and |
258 |
* is performed by cheaper ordered writes of "qlock" in successful |
259 |
* cases, but uses CAS in unsuccessful cases. |
260 |
* |
261 |
* Management |
262 |
* ========== |
263 |
* |
264 |
* The main throughput advantages of work-stealing stem from |
265 |
* decentralized control -- workers mostly take tasks from |
266 |
* themselves or each other, at rates that can exceed a billion |
267 |
* per second. The pool itself creates, activates (enables |
268 |
* scanning for and running tasks), deactivates, blocks, and |
269 |
* terminates threads, all with minimal central information. |
270 |
* There are only a few properties that we can globally track or |
271 |
* maintain, so we pack them into a small number of variables, |
272 |
* often maintaining atomicity without blocking or locking. |
273 |
* Nearly all essentially atomic control state is held in two |
274 |
* volatile variables that are by far most often read (not |
275 |
* written) as status and consistency checks. (Also, field |
276 |
* "config" holds unchanging configuration state.) |
277 |
* |
278 |
* Field "ctl" contains 64 bits holding information needed to |
279 |
* atomically decide to add, inactivate, enqueue (on an event |
280 |
* queue), dequeue, and/or re-activate workers. To enable this |
281 |
* packing, we restrict maximum parallelism to (1<<15)-1 (which is |
282 |
* far in excess of normal operating range) to allow ids, counts, |
283 |
* and their negations (used for thresholding) to fit into 16bit |
284 |
* subfields. |
285 |
* |
286 |
* Field "runState" holds lockable state bits (STARTED, STOP, etc) |
287 |
* also protecting updates to the workQueues array. When used as |
288 |
* a lock, it is normally held only for a few instructions (the |
289 |
* only exceptions are one-time array initialization and uncommon |
290 |
* resizing), so is nearly always available after at most a brief |
291 |
* spin. But to be extra-cautious, after spinning, method |
292 |
* awaitRunStateLock (called only if an initial CAS fails), uses a |
293 |
* wait/notify mechanics on a builtin monitor to block when |
294 |
* (rarely) needed. This would be a terrible idea for a highly |
295 |
* contended lock, but most pools run without the lock ever |
296 |
* contending after the spin limit, so this works fine as a more |
297 |
* conservative alternative. Because we don't otherwise have an |
298 |
* internal Object to use as a monitor, the "stealCounter" (an |
299 |
* AtomicLong) is used when available (it too must be lazily |
300 |
* initialized; see externalSubmit). |
301 |
* |
302 |
* Usages of "runState" vs "ctl" interact in only one case: |
303 |
* deciding to add a worker thread (see tryAddWorker), in which |
304 |
* case the ctl CAS is performed while the lock is held. |
305 |
* |
306 |
* Recording WorkQueues. WorkQueues are recorded in the |
307 |
* "workQueues" array. The array is created upon first use (see |
308 |
* externalSubmit) and expanded if necessary. Updates to the |
309 |
* array while recording new workers and unrecording terminated |
310 |
* ones are protected from each other by the runState lock, but |
311 |
* the array is otherwise concurrently readable, and accessed |
312 |
* directly. We also ensure that reads of the array reference |
313 |
* itself never become too stale. To simplify index-based |
314 |
* operations, the array size is always a power of two, and all |
315 |
* readers must tolerate null slots. Worker queues are at odd |
316 |
* indices. Shared (submission) queues are at even indices, up to |
317 |
* a maximum of 64 slots, to limit growth even if array needs to |
318 |
* expand to add more workers. Grouping them together in this way |
319 |
* simplifies and speeds up task scanning. |
320 |
* |
321 |
* All worker thread creation is on-demand, triggered by task |
322 |
* submissions, replacement of terminated workers, and/or |
323 |
* compensation for blocked workers. However, all other support |
324 |
* code is set up to work with other policies. To ensure that we |
325 |
* do not hold on to worker references that would prevent GC, All |
326 |
* accesses to workQueues are via indices into the workQueues |
327 |
* array (which is one source of some of the messy code |
328 |
* constructions here). In essence, the workQueues array serves as |
329 |
* a weak reference mechanism. Thus for example the stack top |
330 |
* subfield of ctl stores indices, not references. |
331 |
* |
332 |
* Queuing Idle Workers. Unlike HPC work-stealing frameworks, we |
333 |
* cannot let workers spin indefinitely scanning for tasks when |
334 |
* none can be found immediately, and we cannot start/resume |
335 |
* workers unless there appear to be tasks available. On the |
336 |
* other hand, we must quickly prod them into action when new |
337 |
* tasks are submitted or generated. In many usages, ramp-up time |
338 |
* to activate workers is the main limiting factor in overall |
339 |
* performance, which is compounded at program start-up by JIT |
340 |
* compilation and allocation. So we streamline this as much as |
341 |
* possible. |
342 |
* |
343 |
* The "ctl" field atomically maintains active and total worker |
344 |
* counts as well as a queue to place waiting threads so they can |
345 |
* be located for signalling. Active counts also play the role of |
346 |
* quiescence indicators, so are decremented when workers believe |
347 |
* that there are no more tasks to execute. The "queue" is |
348 |
* actually a form of Treiber stack. A stack is ideal for |
349 |
* activating threads in most-recently used order. This improves |
350 |
* performance and locality, outweighing the disadvantages of |
351 |
* being prone to contention and inability to release a worker |
352 |
* unless it is topmost on stack. We park/unpark workers after |
353 |
* pushing on the idle worker stack (represented by the lower |
354 |
* 32bit subfield of ctl) when they cannot find work. The top |
355 |
* stack state holds the value of the "scanState" field of the |
356 |
* worker: its index and status, plus a version counter that, in |
357 |
* addition to the count subfields (also serving as version |
358 |
* stamps) provide protection against Treiber stack ABA effects. |
359 |
* |
360 |
* Field scanState is used by both workers and the pool to manage |
361 |
* and track whether a worker is INACTIVE (possibly blocked |
362 |
* waiting for a signal), or SCANNING for tasks (when neither hold |
363 |
* it is busy running tasks). When a worker is inactivated, its |
364 |
* scanState field is set, and is prevented from executing tasks, |
365 |
* even though it must scan once for them to avoid queuing |
366 |
* races. Note that scanState updates lag queue CAS releases so |
367 |
* usage requires care. When queued, the lower 16 bits of |
368 |
* scanState must hold its pool index. So we place the index there |
369 |
* upon initialization (see registerWorker) and otherwise keep it |
370 |
* there or restore it when necessary. |
371 |
* |
372 |
* Memory ordering. See "Correct and Efficient Work-Stealing for |
373 |
* Weak Memory Models" by Le, Pop, Cohen, and Nardelli, PPoPP 2013 |
374 |
* (http://www.di.ens.fr/~zappa/readings/ppopp13.pdf) for an |
375 |
* analysis of memory ordering requirements in work-stealing |
376 |
* algorithms similar to the one used here. We usually need |
377 |
* stronger than minimal ordering because we must sometimes signal |
378 |
* workers, requiring Dekker-like full-fences to avoid lost |
379 |
* signals. Arranging for enough ordering without expensive |
380 |
* over-fencing requires tradeoffs among the supported means of |
381 |
* expressing access constraints. The most central operations, |
382 |
* taking from queues and updating ctl state, require full-fence |
383 |
* CAS. Array slots are read using the emulation of volatiles |
384 |
* provided by Unsafe. Access from other threads to WorkQueue |
385 |
* base, top, and array requires a volatile load of the first of |
386 |
* any of these read. We use the convention of declaring the |
387 |
* "base" index volatile, and always read it before other fields. |
388 |
* The owner thread must ensure ordered updates, so writes use |
389 |
* ordered intrinsics unless they can piggyback on those for other |
390 |
* writes. Similar conventions and rationales hold for other |
391 |
* WorkQueue fields (such as "currentSteal") that are only written |
392 |
* by owners but observed by others. |
393 |
* |
394 |
* Creating workers. To create a worker, we pre-increment total |
395 |
* count (serving as a reservation), and attempt to construct a |
396 |
* ForkJoinWorkerThread via its factory. Upon construction, the |
397 |
* new thread invokes registerWorker, where it constructs a |
398 |
* WorkQueue and is assigned an index in the workQueues array |
399 |
* (expanding the array if necessary). The thread is then |
400 |
* started. Upon any exception across these steps, or null return |
401 |
* from factory, deregisterWorker adjusts counts and records |
402 |
* accordingly. If a null return, the pool continues running with |
403 |
* fewer than the target number workers. If exceptional, the |
404 |
* exception is propagated, generally to some external caller. |
405 |
* Worker index assignment avoids the bias in scanning that would |
406 |
* occur if entries were sequentially packed starting at the front |
407 |
* of the workQueues array. We treat the array as a simple |
408 |
* power-of-two hash table, expanding as needed. The seedIndex |
409 |
* increment ensures no collisions until a resize is needed or a |
410 |
* worker is deregistered and replaced, and thereafter keeps |
411 |
* probability of collision low. We cannot use |
412 |
* ThreadLocalRandom.getProbe() for similar purposes here because |
413 |
* the thread has not started yet, but do so for creating |
414 |
* submission queues for existing external threads. |
415 |
* |
416 |
* Deactivation and waiting. Queuing encounters several intrinsic |
417 |
* races; most notably that a task-producing thread can miss |
418 |
* seeing (and signalling) another thread that gave up looking for |
419 |
* work but has not yet entered the wait queue. When a worker |
420 |
* cannot find a task to steal, it deactivates and enqueues. Very |
421 |
* often, the lack of tasks is transient due to GC or OS |
422 |
* scheduling. To reduce false-alarm deactivation, scanners |
423 |
* compute checksums of queue states during sweeps. (The |
424 |
* stability checks used here and elsewhere are probabilistic |
425 |
* variants of snapshot techniques -- see Herlihy & Shavit.) |
426 |
* Workers give up and try to deactivate only after the sum is |
427 |
* stable across scans. Further, to avoid missed signals, they |
428 |
* repeat this scanning process after successful enqueuing until |
429 |
* again stable. In this state, the worker cannot take/run a task |
430 |
* it sees until it is released from the queue, so the worker |
431 |
* itself eventually tries to release itself or any successor (see |
432 |
* tryRelease). Otherwise, upon an empty scan, a deactivated |
433 |
* worker uses an adaptive local spin construction (see awaitWork) |
434 |
* before blocking (via park). Note the unusual conventions about |
435 |
* Thread.interrupts surrounding parking and other blocking: |
436 |
* Because interrupts are used solely to alert threads to check |
437 |
* termination, which is checked anyway upon blocking, we clear |
438 |
* status (using Thread.interrupted) before any call to park, so |
439 |
* that park does not immediately return due to status being set |
440 |
* via some other unrelated call to interrupt in user code. |
441 |
* |
442 |
* Signalling and activation. Workers are created or activated |
443 |
* only when there appears to be at least one task they might be |
444 |
* able to find and execute. Upon push (either by a worker or an |
445 |
* external submission) to a previously (possibly) empty queue, |
446 |
* workers are signalled if idle, or created if fewer exist than |
447 |
* the given parallelism level. These primary signals are |
448 |
* buttressed by others whenever other threads remove a task from |
449 |
* a queue and notice that there are other tasks there as well. |
450 |
* On most platforms, signalling (unpark) overhead time is |
451 |
* noticeably long, and the time between signalling a thread and |
452 |
* it actually making progress can be very noticeably long, so it |
453 |
* is worth offloading these delays from critical paths as much as |
454 |
* possible. Also, because inactive workers are often rescanning |
455 |
* or spinning rather than blocking, we set and clear the "parker" |
456 |
* field of WorkQueues to reduce unnecessary calls to unpark. |
457 |
* (This requires a secondary recheck to avoid missed signals.) |
458 |
* |
459 |
* Trimming workers. To release resources after periods of lack of |
460 |
* use, a worker starting to wait when the pool is quiescent will |
461 |
* time out and terminate (see awaitWork) if the pool has remained |
462 |
* quiescent for period IDLE_TIMEOUT, increasing the period as the |
463 |
* number of threads decreases, eventually removing all workers. |
464 |
* Also, when more than two spare threads exist, excess threads |
465 |
* are immediately terminated at the next quiescent point. |
466 |
* (Padding by two avoids hysteresis.) |
467 |
* |
468 |
* Shutdown and Termination. A call to shutdownNow invokes |
469 |
* tryTerminate to atomically set a runState bit. The calling |
470 |
* thread, as well as every other worker thereafter terminating, |
471 |
* helps terminate others by setting their (qlock) status, |
472 |
* cancelling their unprocessed tasks, and waking them up, doing |
473 |
* so repeatedly until stable (but with a loop bounded by the |
474 |
* number of workers). Calls to non-abrupt shutdown() preface |
475 |
* this by checking whether termination should commence. This |
476 |
* relies primarily on the active count bits of "ctl" maintaining |
477 |
* consensus -- tryTerminate is called from awaitWork whenever |
478 |
* quiescent. However, external submitters do not take part in |
479 |
* this consensus. So, tryTerminate sweeps through queues (until |
480 |
* stable) to ensure lack of in-flight submissions and workers |
481 |
* about to process them before triggering the "STOP" phase of |
482 |
* termination. (Note: there is an intrinsic conflict if |
483 |
* helpQuiescePool is called when shutdown is enabled. Both wait |
484 |
* for quiescence, but tryTerminate is biased to not trigger until |
485 |
* helpQuiescePool completes.) |
486 |
* |
487 |
* |
488 |
* Joining Tasks |
489 |
* ============= |
490 |
* |
491 |
* Any of several actions may be taken when one worker is waiting |
492 |
* to join a task stolen (or always held) by another. Because we |
493 |
* are multiplexing many tasks on to a pool of workers, we can't |
494 |
* just let them block (as in Thread.join). We also cannot just |
495 |
* reassign the joiner's run-time stack with another and replace |
496 |
* it later, which would be a form of "continuation", that even if |
497 |
* possible is not necessarily a good idea since we may need both |
498 |
* an unblocked task and its continuation to progress. Instead we |
499 |
* combine two tactics: |
500 |
* |
501 |
* Helping: Arranging for the joiner to execute some task that it |
502 |
* would be running if the steal had not occurred. |
503 |
* |
504 |
* Compensating: Unless there are already enough live threads, |
505 |
* method tryCompensate() may create or re-activate a spare |
506 |
* thread to compensate for blocked joiners until they unblock. |
507 |
* |
508 |
* A third form (implemented in tryRemoveAndExec) amounts to |
509 |
* helping a hypothetical compensator: If we can readily tell that |
510 |
* a possible action of a compensator is to steal and execute the |
511 |
* task being joined, the joining thread can do so directly, |
512 |
* without the need for a compensation thread (although at the |
513 |
* expense of larger run-time stacks, but the tradeoff is |
514 |
* typically worthwhile). |
515 |
* |
516 |
* The ManagedBlocker extension API can't use helping so relies |
517 |
* only on compensation in method awaitBlocker. |
518 |
* |
519 |
* The algorithm in helpStealer entails a form of "linear |
520 |
* helping". Each worker records (in field currentSteal) the most |
521 |
* recent task it stole from some other worker (or a submission). |
522 |
* It also records (in field currentJoin) the task it is currently |
523 |
* actively joining. Method helpStealer uses these markers to try |
524 |
* to find a worker to help (i.e., steal back a task from and |
525 |
* execute it) that could hasten completion of the actively joined |
526 |
* task. Thus, the joiner executes a task that would be on its |
527 |
* own local deque had the to-be-joined task not been stolen. This |
528 |
* is a conservative variant of the approach described in Wagner & |
529 |
* Calder "Leapfrogging: a portable technique for implementing |
530 |
* efficient futures" SIGPLAN Notices, 1993 |
531 |
* (http://portal.acm.org/citation.cfm?id=155354). It differs in |
532 |
* that: (1) We only maintain dependency links across workers upon |
533 |
* steals, rather than use per-task bookkeeping. This sometimes |
534 |
* requires a linear scan of workQueues array to locate stealers, |
535 |
* but often doesn't because stealers leave hints (that may become |
536 |
* stale/wrong) of where to locate them. It is only a hint |
537 |
* because a worker might have had multiple steals and the hint |
538 |
* records only one of them (usually the most current). Hinting |
539 |
* isolates cost to when it is needed, rather than adding to |
540 |
* per-task overhead. (2) It is "shallow", ignoring nesting and |
541 |
* potentially cyclic mutual steals. (3) It is intentionally |
542 |
* racy: field currentJoin is updated only while actively joining, |
543 |
* which means that we miss links in the chain during long-lived |
544 |
* tasks, GC stalls etc (which is OK since blocking in such cases |
545 |
* is usually a good idea). (4) We bound the number of attempts |
546 |
* to find work using checksums and fall back to suspending the |
547 |
* worker and if necessary replacing it with another. |
548 |
* |
549 |
* Helping actions for CountedCompleters do not require tracking |
550 |
* currentJoins: Method helpComplete takes and executes any task |
551 |
* with the same root as the task being waited on (preferring |
552 |
* local pops to non-local polls). However, this still entails |
553 |
* some traversal of completer chains, so is less efficient than |
554 |
* using CountedCompleters without explicit joins. |
555 |
* |
556 |
* Compensation does not aim to keep exactly the target |
557 |
* parallelism number of unblocked threads running at any given |
558 |
* time. Some previous versions of this class employed immediate |
559 |
* compensations for any blocked join. However, in practice, the |
560 |
* vast majority of blockages are transient byproducts of GC and |
561 |
* other JVM or OS activities that are made worse by replacement. |
562 |
* Currently, compensation is attempted only after validating that |
563 |
* all purportedly active threads are processing tasks by checking |
564 |
* field WorkQueue.scanState, which eliminates most false |
565 |
* positives. Also, compensation is bypassed (tolerating fewer |
566 |
* threads) in the most common case in which it is rarely |
567 |
* beneficial: when a worker with an empty queue (thus no |
568 |
* continuation tasks) blocks on a join and there still remain |
569 |
* enough threads to ensure liveness. |
570 |
* |
571 |
* The compensation mechanism may be bounded. Bounds for the |
572 |
* commonPool (see commonMaxSpares) better enable JVMs to cope |
573 |
* with programming errors and abuse before running out of |
574 |
* resources to do so. In other cases, users may supply factories |
575 |
* that limit thread construction. The effects of bounding in this |
576 |
* pool (like all others) is imprecise. Total worker counts are |
577 |
* decremented when threads deregister, not when they exit and |
578 |
* resources are reclaimed by the JVM and OS. So the number of |
579 |
* simultaneously live threads may transiently exceed bounds. |
580 |
* |
581 |
* Common Pool |
582 |
* =========== |
583 |
* |
584 |
* The static common pool always exists after static |
585 |
* initialization. Since it (or any other created pool) need |
586 |
* never be used, we minimize initial construction overhead and |
587 |
* footprint to the setup of about a dozen fields, with no nested |
588 |
* allocation. Most bootstrapping occurs within method |
589 |
* externalSubmit during the first submission to the pool. |
590 |
* |
591 |
* When external threads submit to the common pool, they can |
592 |
* perform subtask processing (see externalHelpComplete and |
593 |
* related methods) upon joins. This caller-helps policy makes it |
594 |
* sensible to set common pool parallelism level to one (or more) |
595 |
* less than the total number of available cores, or even zero for |
596 |
* pure caller-runs. We do not need to record whether external |
597 |
* submissions are to the common pool -- if not, external help |
598 |
* methods return quickly. These submitters would otherwise be |
599 |
* blocked waiting for completion, so the extra effort (with |
600 |
* liberally sprinkled task status checks) in inapplicable cases |
601 |
* amounts to an odd form of limited spin-wait before blocking in |
602 |
* ForkJoinTask.join. |
603 |
* |
604 |
* As a more appropriate default in managed environments, unless |
605 |
* overridden by system properties, we use workers of subclass |
606 |
* InnocuousForkJoinWorkerThread when there is a SecurityManager |
607 |
* present. These workers have no permissions set, do not belong |
608 |
* to any user-defined ThreadGroup, and erase all ThreadLocals |
609 |
* after executing any top-level task (see WorkQueue.runTask). |
610 |
* The associated mechanics (mainly in ForkJoinWorkerThread) may |
611 |
* be JVM-dependent and must access particular Thread class fields |
612 |
* to achieve this effect. |
613 |
* |
614 |
* Style notes |
615 |
* =========== |
616 |
* |
617 |
* Memory ordering relies mainly on Unsafe intrinsics that carry |
618 |
* the further responsibility of explicitly performing null- and |
619 |
* bounds- checks otherwise carried out implicitly by JVMs. This |
620 |
* can be awkward and ugly, but also reflects the need to control |
621 |
* outcomes across the unusual cases that arise in very racy code |
622 |
* with very few invariants. So these explicit checks would exist |
623 |
* in some form anyway. All fields are read into locals before |
624 |
* use, and null-checked if they are references. This is usually |
625 |
* done in a "C"-like style of listing declarations at the heads |
626 |
* of methods or blocks, and using inline assignments on first |
627 |
* encounter. Array bounds-checks are usually performed by |
628 |
* masking with array.length-1, which relies on the invariant that |
629 |
* these arrays are created with positive lengths, which is itself |
630 |
* paranoically checked. Nearly all explicit checks lead to |
631 |
* bypass/return, not exception throws, because they may |
632 |
* legitimately arise due to cancellation/revocation during |
633 |
* shutdown. |
634 |
* |
635 |
* There is a lot of representation-level coupling among classes |
636 |
* ForkJoinPool, ForkJoinWorkerThread, and ForkJoinTask. The |
637 |
* fields of WorkQueue maintain data structures managed by |
638 |
* ForkJoinPool, so are directly accessed. There is little point |
639 |
* trying to reduce this, since any associated future changes in |
640 |
* representations will need to be accompanied by algorithmic |
641 |
* changes anyway. Several methods intrinsically sprawl because |
642 |
* they must accumulate sets of consistent reads of fields held in |
643 |
* local variables. There are also other coding oddities |
644 |
* (including several unnecessary-looking hoisted null checks) |
645 |
* that help some methods perform reasonably even when interpreted |
646 |
* (not compiled). |
647 |
* |
648 |
* The order of declarations in this file is (with a few exceptions): |
649 |
* (1) Static utility functions |
650 |
* (2) Nested (static) classes |
651 |
* (3) Static fields |
652 |
* (4) Fields, along with constants used when unpacking some of them |
653 |
* (5) Internal control methods |
654 |
* (6) Callbacks and other support for ForkJoinTask methods |
655 |
* (7) Exported methods |
656 |
* (8) Static block initializing statics in minimally dependent order |
657 |
*/ |
658 |
|
659 |
// Static utilities |
660 |
|
661 |
/** |
662 |
* If there is a security manager, makes sure caller has |
663 |
* permission to modify threads. |
664 |
*/ |
665 |
private static void checkPermission() { |
666 |
SecurityManager security = System.getSecurityManager(); |
667 |
if (security != null) |
668 |
security.checkPermission(modifyThreadPermission); |
669 |
} |
670 |
|
671 |
// Nested classes |
672 |
|
673 |
/** |
674 |
* Factory for creating new {@link ForkJoinWorkerThread}s. |
675 |
* A {@code ForkJoinWorkerThreadFactory} must be defined and used |
676 |
* for {@code ForkJoinWorkerThread} subclasses that extend base |
677 |
* functionality or initialize threads with different contexts. |
678 |
*/ |
679 |
public static interface ForkJoinWorkerThreadFactory { |
680 |
/** |
681 |
* Returns a new worker thread operating in the given pool. |
682 |
* |
683 |
* @param pool the pool this thread works in |
684 |
* @return the new worker thread |
685 |
* @throws NullPointerException if the pool is null |
686 |
*/ |
687 |
public ForkJoinWorkerThread newThread(ForkJoinPool pool); |
688 |
} |
689 |
|
690 |
/** |
691 |
* Default ForkJoinWorkerThreadFactory implementation; creates a |
692 |
* new ForkJoinWorkerThread. |
693 |
*/ |
694 |
static final class DefaultForkJoinWorkerThreadFactory |
695 |
implements ForkJoinWorkerThreadFactory { |
696 |
public final ForkJoinWorkerThread newThread(ForkJoinPool pool) { |
697 |
return new ForkJoinWorkerThread(pool); |
698 |
} |
699 |
} |
700 |
|
701 |
/** |
702 |
* Class for artificial tasks that are used to replace the target |
703 |
* of local joins if they are removed from an interior queue slot |
704 |
* in WorkQueue.tryRemoveAndExec. We don't need the proxy to |
705 |
* actually do anything beyond having a unique identity. |
706 |
*/ |
707 |
static final class EmptyTask extends ForkJoinTask<Void> { |
708 |
private static final long serialVersionUID = -7721805057305804111L; |
709 |
EmptyTask() { status = ForkJoinTask.NORMAL; } // force done |
710 |
public final Void getRawResult() { return null; } |
711 |
public final void setRawResult(Void x) {} |
712 |
public final boolean exec() { return true; } |
713 |
} |
714 |
|
715 |
// Constants shared across ForkJoinPool and WorkQueue |
716 |
|
717 |
// Bounds |
718 |
static final int SMASK = 0xffff; // short bits == max index |
719 |
static final int MAX_CAP = 0x7fff; // max #workers - 1 |
720 |
static final int EVENMASK = 0xfffe; // even short bits |
721 |
static final int SQMASK = 0x007e; // max 64 (even) slots |
722 |
|
723 |
// Masks and units for WorkQueue.scanState and ctl sp subfield |
724 |
static final int SCANNING = 1; // false when running tasks |
725 |
static final int INACTIVE = 1 << 31; // must be negative |
726 |
static final int SS_SEQ = 1 << 16; // version count |
727 |
|
728 |
// Mode bits for ForkJoinPool.config and WorkQueue.config |
729 |
static final int MODE_MASK = 0xffff << 16; // top half of int |
730 |
static final int LIFO_QUEUE = 0; |
731 |
static final int FIFO_QUEUE = 1 << 16; |
732 |
static final int SHARED_QUEUE = 1 << 31; // must be negative |
733 |
|
734 |
// This version uses array access methods in anticipation of JDK9 support |
735 |
// that should eliminate their need |
736 |
|
737 |
static final ForkJoinTask<?> getAt(ForkJoinTask<?>[] a, int i) { |
738 |
return (ForkJoinTask<?>)U.getObjectVolatile( |
739 |
a, (long)((i << ASHIFT) + ABASE)); |
740 |
} |
741 |
|
742 |
static final void setAt(ForkJoinTask<?>[] a, int i, ForkJoinTask<?> x) { |
743 |
U.putOrderedObject(a, (long)((i << ASHIFT) + ABASE), x); |
744 |
} |
745 |
|
746 |
static final boolean casAt(ForkJoinTask<?>[] a, int i, |
747 |
ForkJoinTask<?> c, ForkJoinTask<?> v) { |
748 |
return U.compareAndSwapObject( |
749 |
a, (long)((i << ASHIFT) + ABASE), c, v); |
750 |
} |
751 |
|
752 |
static final ForkJoinTask<?> xchgAt(ForkJoinTask<?>[] a, int i, |
753 |
ForkJoinTask<?> x) { |
754 |
return (ForkJoinTask<?>)U.getAndSetObject( |
755 |
a, (long)((i << ASHIFT) + ABASE), x); |
756 |
} |
757 |
|
758 |
/** |
759 |
* Queues supporting work-stealing as well as external task |
760 |
* submission. See above for descriptions and algorithms. |
761 |
* Performance on most platforms is very sensitive to placement of |
762 |
* instances of both WorkQueues and their arrays -- we absolutely |
763 |
* do not want multiple WorkQueue instances or multiple queue |
764 |
* arrays sharing cache lines. The @Contended annotation alerts |
765 |
* JVMs to try to keep instances apart. |
766 |
*/ |
767 |
@sun.misc.Contended |
768 |
static final class WorkQueue { |
769 |
|
770 |
/** |
771 |
* Capacity of work-stealing queue array upon initialization. |
772 |
* Must be a power of two; at least 4, but should be larger to |
773 |
* reduce or eliminate cacheline sharing among queues. |
774 |
* Currently, it is much larger, as a partial workaround for |
775 |
* the fact that JVMs often place arrays in locations that |
776 |
* share GC bookkeeping (especially cardmarks) such that |
777 |
* per-write accesses encounter serious memory contention. |
778 |
*/ |
779 |
static final int INITIAL_QUEUE_CAPACITY = 1 << 13; |
780 |
|
781 |
/** |
782 |
* Maximum size for queue arrays. Must be a power of two less |
783 |
* than or equal to 1 << (31 - width of array entry) to ensure |
784 |
* lack of wraparound of index calculations, but defined to a |
785 |
* value a bit less than this to help users trap runaway |
786 |
* programs before saturating systems. |
787 |
*/ |
788 |
static final int MAXIMUM_QUEUE_CAPACITY = 1 << 26; // 64M |
789 |
|
790 |
// Instance fields |
791 |
volatile int scanState; // versioned, <0: inactive; odd:scanning |
792 |
int stackPred; // pool stack (ctl) predecessor |
793 |
int nsteals; // number of steals |
794 |
int hint; // randomization and stealer index hint |
795 |
int config; // pool index and mode |
796 |
volatile int qlock; // 1: locked, < 0: terminate; else 0 |
797 |
volatile int base; // index of next slot for poll |
798 |
int top; // index of next slot for push |
799 |
ForkJoinTask<?>[] array; // the elements (initially unallocated) |
800 |
final ForkJoinPool pool; // the containing pool (may be null) |
801 |
final ForkJoinWorkerThread owner; // owning thread or null if shared |
802 |
volatile Thread parker; // == owner during call to park; else null |
803 |
volatile ForkJoinTask<?> currentJoin; // task being joined in awaitJoin |
804 |
volatile ForkJoinTask<?> currentSteal; // mainly used by helpStealer |
805 |
|
806 |
// Temporary repeats of array access methods |
807 |
|
808 |
static final ForkJoinTask<?> getAt(ForkJoinTask<?>[] a, int i) { |
809 |
return (ForkJoinTask<?>)U.getObjectVolatile( |
810 |
a, (long)((i << ASHIFT) + ABASE)); |
811 |
} |
812 |
|
813 |
static final void setAt(ForkJoinTask<?>[] a, int i, ForkJoinTask<?> x) { |
814 |
U.putOrderedObject(a, (long)((i << ASHIFT) + ABASE), x); |
815 |
} |
816 |
|
817 |
static final boolean casAt(ForkJoinTask<?>[] a, int i, |
818 |
ForkJoinTask<?> c, ForkJoinTask<?> v) { |
819 |
return U.compareAndSwapObject( |
820 |
a, (long)((i << ASHIFT) + ABASE), c, v); |
821 |
} |
822 |
|
823 |
static final ForkJoinTask<?> xchgAt(ForkJoinTask<?>[] a, int i, |
824 |
ForkJoinTask<?> x) { |
825 |
return (ForkJoinTask<?>)U.getAndSetObject( |
826 |
a, (long)((i << ASHIFT) + ABASE), x); |
827 |
} |
828 |
|
829 |
WorkQueue(ForkJoinPool pool, ForkJoinWorkerThread owner) { |
830 |
this.pool = pool; |
831 |
this.owner = owner; |
832 |
// Place indices in the center of array (that is not yet allocated) |
833 |
base = top = INITIAL_QUEUE_CAPACITY >>> 1; |
834 |
} |
835 |
|
836 |
/** |
837 |
* Returns an exportable index (used by ForkJoinWorkerThread). |
838 |
*/ |
839 |
final int getPoolIndex() { |
840 |
return (config & 0xffff) >>> 1; // ignore odd/even tag bit |
841 |
} |
842 |
|
843 |
/** |
844 |
* Returns the approximate number of tasks in the queue. |
845 |
*/ |
846 |
final int queueSize() { |
847 |
int n = base - top; // non-owner callers must read base first |
848 |
return (n >= 0) ? 0 : -n; // ignore transient negative |
849 |
} |
850 |
|
851 |
/** |
852 |
* Provides a more accurate estimate of whether this queue has |
853 |
* any tasks than does queueSize, by checking whether a |
854 |
* near-empty queue has at least one unclaimed task. |
855 |
*/ |
856 |
final boolean isEmpty() { |
857 |
ForkJoinTask<?>[] a; int n, al, s; |
858 |
return ((n = base - (s = top)) >= 0 || |
859 |
(n == -1 && // possibly one task |
860 |
((a = array) == null || (al = a.length) == 0 || |
861 |
getAt(a, (al - 1) & (s - 1)) == null))); |
862 |
} |
863 |
|
864 |
/** |
865 |
* Pushes a task. Call only by owner in unshared queues. (The |
866 |
* shared-queue version is embedded in method externalPush.) |
867 |
* |
868 |
* @param task the task. Caller must ensure non-null. |
869 |
* @throws RejectedExecutionException if array cannot be resized |
870 |
*/ |
871 |
final void push(ForkJoinTask<?> task) { |
872 |
ForkJoinTask<?>[] a; ForkJoinPool p; |
873 |
if ((a = array) != null) { // ignore if queue removed |
874 |
int b = base, m = a.length - 1, s = top, n; |
875 |
if (m > 0) { // always true, but check required |
876 |
setAt(a, m & s, task); |
877 |
U.putOrderedInt(this, QTOP, s + 1); |
878 |
if ((n = s - b) <= 1) { |
879 |
if ((p = pool) != null) |
880 |
p.signalWork(p.workQueues, this); |
881 |
} |
882 |
else if (n >= m) |
883 |
growArray(); |
884 |
} |
885 |
} |
886 |
} |
887 |
|
888 |
/** |
889 |
* Initializes or doubles the capacity of array. Call either |
890 |
* by owner or with lock held -- it is OK for base, but not |
891 |
* top, to move while resizings are in progress. |
892 |
*/ |
893 |
final ForkJoinTask<?>[] growArray() { |
894 |
ForkJoinTask<?>[] oldA = array; |
895 |
int size = oldA != null ? oldA.length << 1 : INITIAL_QUEUE_CAPACITY; |
896 |
if (size < INITIAL_QUEUE_CAPACITY || size > MAXIMUM_QUEUE_CAPACITY) |
897 |
throw new RejectedExecutionException("Queue capacity exceeded"); |
898 |
int oldMask, t, b; |
899 |
ForkJoinTask<?>[] a = array = new ForkJoinTask<?>[size]; |
900 |
if (oldA != null && (oldMask = oldA.length - 1) > 0 && |
901 |
(t = top) - (b = base) > 0) { |
902 |
int mask = size - 1; |
903 |
do { // emulate poll from old array, push to new array |
904 |
ForkJoinTask<?> x; |
905 |
int oldj = b & oldMask, j = b & mask; |
906 |
if ((x = getAt(oldA, oldj)) != null && |
907 |
casAt(oldA, oldj, x, null)) |
908 |
setAt(a, j, x); |
909 |
} while (++b != t); |
910 |
} |
911 |
return a; |
912 |
} |
913 |
|
914 |
/** |
915 |
* Takes next task, if one exists, in LIFO order. Call only |
916 |
* by owner in unshared queues. |
917 |
*/ |
918 |
final ForkJoinTask<?> pop() { |
919 |
ForkJoinTask<?>[] a; ForkJoinTask<?> t; int al; |
920 |
if ((a = array) != null && (al = a.length) > 0) { |
921 |
for (int s; (s = top - 1) - base >= 0;) { |
922 |
int j = (al - 1) & s; |
923 |
if ((t = getAt(a, j)) == null) |
924 |
break; |
925 |
if (casAt(a, j, t, null)) { |
926 |
U.putOrderedInt(this, QTOP, s); |
927 |
return t; |
928 |
} |
929 |
} |
930 |
} |
931 |
return null; |
932 |
} |
933 |
|
934 |
/** |
935 |
* Takes a task in FIFO order if b is base of queue and a task |
936 |
* can be claimed without contention. Specialized versions |
937 |
* appear in ForkJoinPool methods scan and helpStealer. |
938 |
*/ |
939 |
final ForkJoinTask<?> pollAt(int b) { |
940 |
ForkJoinTask<?> t; ForkJoinTask<?>[] a; |
941 |
if ((a = array) != null) { |
942 |
int al = a.length, j = (al - 1) & b; |
943 |
if (al > 0 && (t = getAt(a, j)) != null && |
944 |
base == b && casAt(a, j, t, null)) { |
945 |
base = b + 1; |
946 |
return t; |
947 |
} |
948 |
} |
949 |
return null; |
950 |
} |
951 |
|
952 |
/** |
953 |
* Takes next task, if one exists, in FIFO order. |
954 |
*/ |
955 |
final ForkJoinTask<?> poll() { |
956 |
ForkJoinTask<?>[] a; int b, al, j; |
957 |
while ((b = base) - top < 0 && (a = array) != null && |
958 |
(al = a.length) > 0) { |
959 |
ForkJoinTask<?> t = getAt(a, j = (al - 1) & b); |
960 |
if (base == b) { |
961 |
if (t != null) { |
962 |
if (casAt(a, j, t, null)) { |
963 |
base = b + 1; |
964 |
return t; |
965 |
} |
966 |
} |
967 |
else if (b + 1 == top) // now empty |
968 |
break; |
969 |
} |
970 |
} |
971 |
return null; |
972 |
} |
973 |
|
974 |
/** |
975 |
* Takes next task, if one exists, in order specified by mode. |
976 |
*/ |
977 |
final ForkJoinTask<?> nextLocalTask() { |
978 |
return (config & FIFO_QUEUE) == 0 ? pop() : poll(); |
979 |
} |
980 |
|
981 |
/** |
982 |
* Returns next task, if one exists, in order specified by mode. |
983 |
*/ |
984 |
final ForkJoinTask<?> peek() { |
985 |
ForkJoinTask<?>[] a = array; int al; |
986 |
if (a != null && (al = a.length) > 0) { |
987 |
int i = (config & FIFO_QUEUE) == 0 ? top - 1 : base; |
988 |
return getAt(a, (al - 1) & i); |
989 |
} |
990 |
return null; |
991 |
} |
992 |
|
993 |
/** |
994 |
* Pops the given task only if it is at the current top. |
995 |
* (A shared version is available only via FJP.tryExternalUnpush) |
996 |
*/ |
997 |
final boolean tryUnpush(ForkJoinTask<?> t) { |
998 |
ForkJoinTask<?>[] a; |
999 |
if ((a = array) != null) { |
1000 |
int b = base, al = a.length, s = top; |
1001 |
if (s != b && al > 0 && |
1002 |
casAt(a, (al - 1) & (s - 1), t, null)) { |
1003 |
U.putOrderedInt(this, QTOP, s - 1); |
1004 |
return true; |
1005 |
} |
1006 |
} |
1007 |
return false; |
1008 |
} |
1009 |
|
1010 |
/** |
1011 |
* Removes and cancels all known tasks, ignoring any exceptions. |
1012 |
*/ |
1013 |
final void cancelAll() { |
1014 |
ForkJoinTask<?> t; |
1015 |
if ((t = currentJoin) != null) { |
1016 |
currentJoin = null; |
1017 |
ForkJoinTask.cancelIgnoringExceptions(t); |
1018 |
} |
1019 |
if ((t = currentSteal) != null) { |
1020 |
currentSteal = null; |
1021 |
ForkJoinTask.cancelIgnoringExceptions(t); |
1022 |
} |
1023 |
while ((t = poll()) != null) |
1024 |
ForkJoinTask.cancelIgnoringExceptions(t); |
1025 |
} |
1026 |
|
1027 |
// Specialized execution methods |
1028 |
|
1029 |
/** |
1030 |
* Polls and runs tasks until empty. |
1031 |
*/ |
1032 |
final void pollAndExecAll() { |
1033 |
for (ForkJoinTask<?> t; (t = poll()) != null;) |
1034 |
t.doExec(); |
1035 |
} |
1036 |
|
1037 |
/** |
1038 |
* Pops and runs tasks until empty. |
1039 |
*/ |
1040 |
final void popAndExecAll() { |
1041 |
ForkJoinTask<?>[] a; ForkJoinTask<?> t; |
1042 |
while ((a = array) != null) { |
1043 |
int b = base, al = a.length, s = top, i = (al - 1) & (s - 1); |
1044 |
if (b != s && al > 0 && |
1045 |
(t = xchgAt(a, i, null)) != null) { |
1046 |
U.putOrderedInt(this, QTOP, s - 1); |
1047 |
t.doExec(); |
1048 |
} |
1049 |
else |
1050 |
break; |
1051 |
} |
1052 |
} |
1053 |
|
1054 |
/** |
1055 |
* Executes the given task and any remaining local tasks. |
1056 |
*/ |
1057 |
final void runTask(ForkJoinTask<?> task) { |
1058 |
if (task != null) { |
1059 |
scanState &= ~SCANNING; // mark as busy |
1060 |
(currentSteal = task).doExec(); |
1061 |
U.putOrderedObject(this, QCURRENTSTEAL, null); // release for GC |
1062 |
if ((config & FIFO_QUEUE) != 0) |
1063 |
pollAndExecAll(); |
1064 |
else |
1065 |
popAndExecAll(); |
1066 |
ForkJoinWorkerThread thread = owner; |
1067 |
if (++nsteals < 0) // collect on overflow |
1068 |
transferStealCount(pool); |
1069 |
scanState |= SCANNING; |
1070 |
if (thread != null) |
1071 |
thread.afterTopLevelExec(); |
1072 |
} |
1073 |
} |
1074 |
|
1075 |
/** |
1076 |
* Adds steal count to pool stealCounter if it exists, and resets. |
1077 |
*/ |
1078 |
final void transferStealCount(ForkJoinPool p) { |
1079 |
AtomicLong sc; |
1080 |
if (p != null && (sc = p.stealCounter) != null) { |
1081 |
int s = nsteals; |
1082 |
nsteals = 0; // if negative, correct for overflow |
1083 |
sc.getAndAdd((long)(s < 0 ? Integer.MAX_VALUE : s)); |
1084 |
} |
1085 |
} |
1086 |
|
1087 |
/** |
1088 |
* If present, removes from queue and executes the given task, |
1089 |
* or any other cancelled task. Used only by awaitJoin. |
1090 |
* |
1091 |
* @return true if queue empty and task not known to be done |
1092 |
*/ |
1093 |
final boolean tryRemoveAndExec(ForkJoinTask<?> task) { |
1094 |
ForkJoinTask<?>[] a; int al, s, b, n; |
1095 |
if ((a = array) != null && (al = a.length) > 0 && task != null) { |
1096 |
while ((n = (s = top) - (b = base)) > 0) { |
1097 |
for (ForkJoinTask<?> t;;) { // traverse from s to b |
1098 |
int j = --s & (al - 1); |
1099 |
if ((t = getAt(a, j)) == null) |
1100 |
return s + 1 == top; // shorter than expected |
1101 |
else if (t == task) { |
1102 |
boolean removed = false; |
1103 |
if (s + 1 == top) { // pop |
1104 |
if (casAt(a, j, task, null)) { |
1105 |
U.putOrderedInt(this, QTOP, s); |
1106 |
removed = true; |
1107 |
} |
1108 |
} |
1109 |
else if (base == b) // replace with proxy |
1110 |
removed = casAt(a, j, task, new EmptyTask()); |
1111 |
if (removed) |
1112 |
task.doExec(); |
1113 |
break; |
1114 |
} |
1115 |
else if (t.status < 0 && s + 1 == top) { |
1116 |
if (casAt(a, j, t, null)) |
1117 |
U.putOrderedInt(this, QTOP, s); |
1118 |
break; // was cancelled |
1119 |
} |
1120 |
if (--n == 0) |
1121 |
return false; |
1122 |
} |
1123 |
if (task.status < 0) |
1124 |
return false; |
1125 |
} |
1126 |
} |
1127 |
return true; |
1128 |
} |
1129 |
|
1130 |
/** |
1131 |
* Pops task if in the same CC computation as the given task, |
1132 |
* in either shared or owned mode. Used only by helpComplete. |
1133 |
*/ |
1134 |
final CountedCompleter<?> popCC(CountedCompleter<?> task, int mode) { |
1135 |
ForkJoinTask<?>[] a; ForkJoinTask<?> o; |
1136 |
if ((a = array) != null) { |
1137 |
int b = base, al = a.length, s = top, i = (al - 1) & (s - 1); |
1138 |
if (b != s && al > 0 && |
1139 |
((o = a[i]) instanceof CountedCompleter)) { |
1140 |
CountedCompleter<?> t = (CountedCompleter<?>)o; |
1141 |
for (CountedCompleter<?> r = t;;) { |
1142 |
if (r == task) { |
1143 |
if (mode < 0) { // must lock |
1144 |
if (U.compareAndSwapInt(this, QLOCK, 0, 1)) { |
1145 |
if (top == s && array == a && |
1146 |
casAt(a, i, t, null)) { |
1147 |
U.putOrderedInt(this, QTOP, s - 1); |
1148 |
U.putOrderedInt(this, QLOCK, 0); |
1149 |
return t; |
1150 |
} |
1151 |
U.compareAndSwapInt(this, QLOCK, 1, 0); |
1152 |
} |
1153 |
} |
1154 |
else if (casAt(a, i, t, null)) { |
1155 |
U.putOrderedInt(this, QTOP, s - 1); |
1156 |
return t; |
1157 |
} |
1158 |
break; |
1159 |
} |
1160 |
else if ((r = r.completer) == null) // try parent |
1161 |
break; |
1162 |
} |
1163 |
} |
1164 |
} |
1165 |
return null; |
1166 |
} |
1167 |
|
1168 |
/** |
1169 |
* Steals and runs a task in the same CC computation as the |
1170 |
* given task if one exists and can be taken without |
1171 |
* contention. Otherwise returns a checksum/control value for |
1172 |
* use by method helpComplete. |
1173 |
* |
1174 |
* @return 1 if successful, 2 if retryable (lost to another |
1175 |
* stealer), -1 if non-empty but no matching task found, else |
1176 |
* the base index, forced negative. |
1177 |
*/ |
1178 |
final int pollAndExecCC(CountedCompleter<?> task) { |
1179 |
int b, h, j, al; ForkJoinTask<?>[] a; Object o; |
1180 |
if ((b = base) - top >= 0 || (a = array) == null || |
1181 |
(al = a.length) <= 0) |
1182 |
h = b | Integer.MIN_VALUE; // to sense movement on re-poll |
1183 |
else if ((o = getAt(a, j = (al - 1) & b)) == null) |
1184 |
h = 2; // retryable |
1185 |
else if (!(o instanceof CountedCompleter)) |
1186 |
h = -1; // unmatchable |
1187 |
else { |
1188 |
CountedCompleter<?> t = (CountedCompleter<?>)o; |
1189 |
for (CountedCompleter<?> r = t;;) { |
1190 |
if (r == task) { |
1191 |
if (base == b && casAt(a, j, t, null)) { |
1192 |
base = b + 1; |
1193 |
t.doExec(); |
1194 |
h = 1; // success |
1195 |
} |
1196 |
else |
1197 |
h = 2; // lost CAS |
1198 |
break; |
1199 |
} |
1200 |
else if ((r = r.completer) == null) { |
1201 |
h = -1; // unmatched |
1202 |
break; |
1203 |
} |
1204 |
} |
1205 |
} |
1206 |
return h; |
1207 |
} |
1208 |
|
1209 |
/** |
1210 |
* Returns true if owned and not known to be blocked. |
1211 |
*/ |
1212 |
final boolean isApparentlyUnblocked() { |
1213 |
Thread wt; Thread.State s; |
1214 |
return (scanState >= 0 && |
1215 |
(wt = owner) != null && |
1216 |
(s = wt.getState()) != Thread.State.BLOCKED && |
1217 |
s != Thread.State.WAITING && |
1218 |
s != Thread.State.TIMED_WAITING); |
1219 |
} |
1220 |
|
1221 |
// Unsafe mechanics. Note that some are (and must be) the same as in FJP |
1222 |
private static final sun.misc.Unsafe U = sun.misc.Unsafe.getUnsafe(); |
1223 |
private static final int ABASE; |
1224 |
private static final int ASHIFT; |
1225 |
private static final long QTOP; |
1226 |
private static final long QLOCK; |
1227 |
private static final long QCURRENTSTEAL; |
1228 |
static { |
1229 |
try { |
1230 |
QTOP = U.objectFieldOffset |
1231 |
(WorkQueue.class.getDeclaredField("top")); |
1232 |
QLOCK = U.objectFieldOffset |
1233 |
(WorkQueue.class.getDeclaredField("qlock")); |
1234 |
QCURRENTSTEAL = U.objectFieldOffset |
1235 |
(WorkQueue.class.getDeclaredField("currentSteal")); |
1236 |
|
1237 |
ABASE = U.arrayBaseOffset(ForkJoinTask[].class); |
1238 |
int scale = U.arrayIndexScale(ForkJoinTask[].class); |
1239 |
if ((scale & (scale - 1)) != 0) |
1240 |
throw new Error("array index scale not a power of two"); |
1241 |
ASHIFT = 31 - Integer.numberOfLeadingZeros(scale); |
1242 |
} catch (ReflectiveOperationException e) { |
1243 |
throw new Error(e); |
1244 |
} |
1245 |
} |
1246 |
} |
1247 |
|
1248 |
// static fields (initialized in static initializer below) |
1249 |
|
1250 |
/** |
1251 |
* Creates a new ForkJoinWorkerThread. This factory is used unless |
1252 |
* overridden in ForkJoinPool constructors. |
1253 |
*/ |
1254 |
public static final ForkJoinWorkerThreadFactory |
1255 |
defaultForkJoinWorkerThreadFactory; |
1256 |
|
1257 |
/** |
1258 |
* Permission required for callers of methods that may start or |
1259 |
* kill threads. |
1260 |
*/ |
1261 |
private static final RuntimePermission modifyThreadPermission; |
1262 |
|
1263 |
/** |
1264 |
* Common (static) pool. Non-null for public use unless a static |
1265 |
* construction exception, but internal usages null-check on use |
1266 |
* to paranoically avoid potential initialization circularities |
1267 |
* as well as to simplify generated code. |
1268 |
*/ |
1269 |
static final ForkJoinPool common; |
1270 |
|
1271 |
/** |
1272 |
* Common pool parallelism. To allow simpler use and management |
1273 |
* when common pool threads are disabled, we allow the underlying |
1274 |
* common.parallelism field to be zero, but in that case still report |
1275 |
* parallelism as 1 to reflect resulting caller-runs mechanics. |
1276 |
*/ |
1277 |
static final int commonParallelism; |
1278 |
|
1279 |
/** |
1280 |
* Limit on spare thread construction in tryCompensate. |
1281 |
*/ |
1282 |
private static int commonMaxSpares; |
1283 |
|
1284 |
/** |
1285 |
* Sequence number for creating workerNamePrefix. |
1286 |
*/ |
1287 |
private static int poolNumberSequence; |
1288 |
|
1289 |
/** |
1290 |
* Returns the next sequence number. We don't expect this to |
1291 |
* ever contend, so use simple builtin sync. |
1292 |
*/ |
1293 |
private static final synchronized int nextPoolId() { |
1294 |
return ++poolNumberSequence; |
1295 |
} |
1296 |
|
1297 |
// static configuration constants |
1298 |
|
1299 |
/** |
1300 |
* Initial timeout value (in nanoseconds) for the thread |
1301 |
* triggering quiescence to park waiting for new work. On timeout, |
1302 |
* the thread will instead try to shrink the number of |
1303 |
* workers. The value should be large enough to avoid overly |
1304 |
* aggressive shrinkage during most transient stalls (long GCs |
1305 |
* etc). |
1306 |
*/ |
1307 |
private static final long IDLE_TIMEOUT = 2000L * 1000L * 1000L; // 2sec |
1308 |
|
1309 |
/** |
1310 |
* Tolerance for idle timeouts, to cope with timer undershoots |
1311 |
*/ |
1312 |
private static final long TIMEOUT_SLOP = 20L * 1000L * 1000L; // 20ms |
1313 |
|
1314 |
/** |
1315 |
* The initial value for commonMaxSpares during static |
1316 |
* initialization unless overridden using System property |
1317 |
* "java.util.concurrent.ForkJoinPool.common.maximumSpares". The |
1318 |
* default value is far in excess of normal requirements, but also |
1319 |
* far short of MAX_CAP and typical OS thread limits, so allows |
1320 |
* JVMs to catch misuse/abuse before running out of resources |
1321 |
* needed to do so. |
1322 |
*/ |
1323 |
private static final int DEFAULT_COMMON_MAX_SPARES = 256; |
1324 |
|
1325 |
/** |
1326 |
* Number of times to spin-wait before blocking. The spins (in |
1327 |
* awaitRunStateLock and awaitWork) currently use randomized |
1328 |
* spins. If/when MWAIT-like intrinsics becomes available, they |
1329 |
* may allow quieter spinning. The value of SPINS must be a power |
1330 |
* of two, at least 4. The current value causes spinning for a |
1331 |
* small fraction of typical context-switch times, well worthwhile |
1332 |
* given the typical likelihoods that blocking is not necessary. |
1333 |
*/ |
1334 |
private static final int SPINS = 1 << 11; |
1335 |
|
1336 |
/** |
1337 |
* Increment for seed generators. See class ThreadLocal for |
1338 |
* explanation. |
1339 |
*/ |
1340 |
private static final int SEED_INCREMENT = 0x9e3779b9; |
1341 |
|
1342 |
/* |
1343 |
* Bits and masks for field ctl, packed with 4 16 bit subfields: |
1344 |
* AC: Number of active running workers minus target parallelism |
1345 |
* TC: Number of total workers minus target parallelism |
1346 |
* SS: version count and status of top waiting thread |
1347 |
* ID: poolIndex of top of Treiber stack of waiters |
1348 |
* |
1349 |
* When convenient, we can extract the lower 32 stack top bits |
1350 |
* (including version bits) as sp=(int)ctl. The offsets of counts |
1351 |
* by the target parallelism and the positionings of fields makes |
1352 |
* it possible to perform the most common checks via sign tests of |
1353 |
* fields: When ac is negative, there are not enough active |
1354 |
* workers, when tc is negative, there are not enough total |
1355 |
* workers. When sp is non-zero, there are waiting workers. To |
1356 |
* deal with possibly negative fields, we use casts in and out of |
1357 |
* "short" and/or signed shifts to maintain signedness. |
1358 |
* |
1359 |
* Because it occupies uppermost bits, we can add one active count |
1360 |
* using getAndAddLong of AC_UNIT, rather than CAS, when returning |
1361 |
* from a blocked join. Other updates entail multiple subfields |
1362 |
* and masking, requiring CAS. |
1363 |
*/ |
1364 |
|
1365 |
// Lower and upper word masks |
1366 |
private static final long SP_MASK = 0xffffffffL; |
1367 |
private static final long UC_MASK = ~SP_MASK; |
1368 |
|
1369 |
// Active counts |
1370 |
private static final int AC_SHIFT = 48; |
1371 |
private static final long AC_UNIT = 0x0001L << AC_SHIFT; |
1372 |
private static final long AC_MASK = 0xffffL << AC_SHIFT; |
1373 |
|
1374 |
// Total counts |
1375 |
private static final int TC_SHIFT = 32; |
1376 |
private static final long TC_UNIT = 0x0001L << TC_SHIFT; |
1377 |
private static final long TC_MASK = 0xffffL << TC_SHIFT; |
1378 |
private static final long ADD_WORKER = 0x0001L << (TC_SHIFT + 15); // sign |
1379 |
|
1380 |
// runState bits: SHUTDOWN must be negative, others arbitrary powers of two |
1381 |
private static final int RSLOCK = 1; |
1382 |
private static final int RSIGNAL = 1 << 1; |
1383 |
private static final int STARTED = 1 << 2; |
1384 |
private static final int STOP = 1 << 29; |
1385 |
private static final int TERMINATED = 1 << 30; |
1386 |
private static final int SHUTDOWN = 1 << 31; |
1387 |
|
1388 |
// Instance fields |
1389 |
volatile long ctl; // main pool control |
1390 |
volatile int runState; // lockable status |
1391 |
final int config; // parallelism, mode |
1392 |
int indexSeed; // to generate worker index |
1393 |
volatile WorkQueue[] workQueues; // main registry |
1394 |
final ForkJoinWorkerThreadFactory factory; |
1395 |
final UncaughtExceptionHandler ueh; // per-worker UEH |
1396 |
final String workerNamePrefix; // to create worker name string |
1397 |
volatile AtomicLong stealCounter; // also used as sync monitor |
1398 |
|
1399 |
/** |
1400 |
* Acquires the runState lock; returns current (locked) runState. |
1401 |
*/ |
1402 |
private int lockRunState() { |
1403 |
int rs; |
1404 |
return ((((rs = runState) & RSLOCK) != 0 || |
1405 |
!U.compareAndSwapInt(this, RUNSTATE, rs, rs |= RSLOCK)) ? |
1406 |
awaitRunStateLock() : rs); |
1407 |
} |
1408 |
|
1409 |
/** |
1410 |
* Spins and/or blocks until runstate lock is available. See |
1411 |
* above for explanation. |
1412 |
*/ |
1413 |
private int awaitRunStateLock() { |
1414 |
Object lock; |
1415 |
boolean wasInterrupted = false; |
1416 |
for (int spins = SPINS, r = 0, rs, ns;;) { |
1417 |
if (((rs = runState) & RSLOCK) == 0) { |
1418 |
if (U.compareAndSwapInt(this, RUNSTATE, rs, ns = rs | RSLOCK)) { |
1419 |
if (wasInterrupted) { |
1420 |
try { |
1421 |
Thread.currentThread().interrupt(); |
1422 |
} catch (SecurityException ignore) { |
1423 |
} |
1424 |
} |
1425 |
return ns; |
1426 |
} |
1427 |
} |
1428 |
else if (r == 0) |
1429 |
r = ThreadLocalRandom.nextSecondarySeed(); |
1430 |
else if (spins > 0) { |
1431 |
r ^= r << 6; r ^= r >>> 21; r ^= r << 7; // xorshift |
1432 |
if (r >= 0) |
1433 |
--spins; |
1434 |
} |
1435 |
else if ((rs & STARTED) == 0 || (lock = stealCounter) == null) |
1436 |
Thread.yield(); // initialization race |
1437 |
else if (U.compareAndSwapInt(this, RUNSTATE, rs, rs | RSIGNAL)) { |
1438 |
synchronized (lock) { |
1439 |
if ((runState & RSIGNAL) != 0) { |
1440 |
try { |
1441 |
lock.wait(); |
1442 |
} catch (InterruptedException ie) { |
1443 |
if (!(Thread.currentThread() instanceof |
1444 |
ForkJoinWorkerThread)) |
1445 |
wasInterrupted = true; |
1446 |
} |
1447 |
} |
1448 |
else |
1449 |
lock.notifyAll(); |
1450 |
} |
1451 |
} |
1452 |
} |
1453 |
} |
1454 |
|
1455 |
/** |
1456 |
* Unlocks and sets runState to newRunState. |
1457 |
* |
1458 |
* @param oldRunState a value returned from lockRunState |
1459 |
* @param newRunState the next value (must have lock bit clear). |
1460 |
*/ |
1461 |
private void unlockRunState(int oldRunState, int newRunState) { |
1462 |
if (!U.compareAndSwapInt(this, RUNSTATE, oldRunState, newRunState)) { |
1463 |
Object lock = stealCounter; |
1464 |
runState = newRunState; // clears RSIGNAL bit |
1465 |
if (lock != null) |
1466 |
synchronized (lock) { lock.notifyAll(); } |
1467 |
} |
1468 |
} |
1469 |
|
1470 |
// Creating, registering and deregistering workers |
1471 |
|
1472 |
/** |
1473 |
* Tries to construct and start one worker. Assumes that total |
1474 |
* count has already been incremented as a reservation. Invokes |
1475 |
* deregisterWorker on any failure. |
1476 |
* |
1477 |
* @return true if successful |
1478 |
*/ |
1479 |
private boolean createWorker() { |
1480 |
ForkJoinWorkerThreadFactory fac = factory; |
1481 |
Throwable ex = null; |
1482 |
ForkJoinWorkerThread wt = null; |
1483 |
try { |
1484 |
if (fac != null && (wt = fac.newThread(this)) != null) { |
1485 |
wt.start(); |
1486 |
return true; |
1487 |
} |
1488 |
} catch (Throwable rex) { |
1489 |
ex = rex; |
1490 |
} |
1491 |
deregisterWorker(wt, ex); |
1492 |
return false; |
1493 |
} |
1494 |
|
1495 |
/** |
1496 |
* Tries to add one worker, incrementing ctl counts before doing |
1497 |
* so, relying on createWorker to back out on failure. |
1498 |
* |
1499 |
* @param c incoming ctl value, with total count negative and no |
1500 |
* idle workers. On CAS failure, c is refreshed and retried if |
1501 |
* this holds (otherwise, a new worker is not needed). |
1502 |
*/ |
1503 |
private void tryAddWorker(long c) { |
1504 |
boolean add = false; |
1505 |
do { |
1506 |
long nc = ((AC_MASK & (c + AC_UNIT)) | |
1507 |
(TC_MASK & (c + TC_UNIT))); |
1508 |
if (ctl == c) { |
1509 |
int rs, stop; // check if terminating |
1510 |
if ((stop = (rs = lockRunState()) & STOP) == 0) |
1511 |
add = U.compareAndSwapLong(this, CTL, c, nc); |
1512 |
unlockRunState(rs, rs & ~RSLOCK); |
1513 |
if (stop != 0) |
1514 |
break; |
1515 |
if (add) { |
1516 |
createWorker(); |
1517 |
break; |
1518 |
} |
1519 |
} |
1520 |
} while (((c = ctl) & ADD_WORKER) != 0L && (int)c == 0); |
1521 |
} |
1522 |
|
1523 |
/** |
1524 |
* Callback from ForkJoinWorkerThread constructor to establish and |
1525 |
* record its WorkQueue. |
1526 |
* |
1527 |
* @param wt the worker thread |
1528 |
* @return the worker's queue |
1529 |
*/ |
1530 |
final WorkQueue registerWorker(ForkJoinWorkerThread wt) { |
1531 |
UncaughtExceptionHandler handler; |
1532 |
wt.setDaemon(true); // configure thread |
1533 |
if ((handler = ueh) != null) |
1534 |
wt.setUncaughtExceptionHandler(handler); |
1535 |
WorkQueue w = new WorkQueue(this, wt); |
1536 |
int i = 0; // assign a pool index |
1537 |
int mode = config & MODE_MASK; |
1538 |
int rs = lockRunState(); |
1539 |
try { |
1540 |
WorkQueue[] ws; int n; // skip if no array |
1541 |
if ((ws = workQueues) != null && (n = ws.length) > 0) { |
1542 |
int s = indexSeed += SEED_INCREMENT; // unlikely to collide |
1543 |
int m = n - 1; |
1544 |
i = ((s << 1) | 1) & m; // odd-numbered indices |
1545 |
if (ws[i] != null) { // collision |
1546 |
int probes = 0; // step by approx half n |
1547 |
int step = (n <= 4) ? 2 : ((n >>> 1) & EVENMASK) + 2; |
1548 |
while (ws[i = (i + step) & m] != null) { |
1549 |
if (++probes >= n) { |
1550 |
workQueues = ws = Arrays.copyOf(ws, n <<= 1); |
1551 |
m = n - 1; |
1552 |
probes = 0; |
1553 |
} |
1554 |
} |
1555 |
} |
1556 |
w.hint = s; // use as random seed |
1557 |
w.config = i | mode; |
1558 |
w.scanState = i; // publication fence |
1559 |
ws[i] = w; |
1560 |
} |
1561 |
} finally { |
1562 |
unlockRunState(rs, rs & ~RSLOCK); |
1563 |
} |
1564 |
wt.setName(workerNamePrefix.concat(Integer.toString(i >>> 1))); |
1565 |
return w; |
1566 |
} |
1567 |
|
1568 |
/** |
1569 |
* Final callback from terminating worker, as well as upon failure |
1570 |
* to construct or start a worker. Removes record of worker from |
1571 |
* array, and adjusts counts. If pool is shutting down, tries to |
1572 |
* complete termination. |
1573 |
* |
1574 |
* @param wt the worker thread, or null if construction failed |
1575 |
* @param ex the exception causing failure, or null if none |
1576 |
*/ |
1577 |
final void deregisterWorker(ForkJoinWorkerThread wt, Throwable ex) { |
1578 |
WorkQueue w = null; |
1579 |
if (wt != null && (w = wt.workQueue) != null) { |
1580 |
WorkQueue[] ws; // remove index from array |
1581 |
int idx = w.config & SMASK; |
1582 |
int rs = lockRunState(); |
1583 |
if ((ws = workQueues) != null && ws.length > idx && ws[idx] == w) |
1584 |
ws[idx] = null; |
1585 |
unlockRunState(rs, rs & ~RSLOCK); |
1586 |
} |
1587 |
long c; // decrement counts |
1588 |
do {} while (!U.compareAndSwapLong |
1589 |
(this, CTL, c = ctl, ((AC_MASK & (c - AC_UNIT)) | |
1590 |
(TC_MASK & (c - TC_UNIT)) | |
1591 |
(SP_MASK & c)))); |
1592 |
if (w != null) { |
1593 |
w.qlock = -1; // ensure set |
1594 |
w.transferStealCount(this); |
1595 |
w.cancelAll(); // cancel remaining tasks |
1596 |
} |
1597 |
for (;;) { // possibly replace |
1598 |
WorkQueue[] ws; int m, sp; |
1599 |
if (tryTerminate(false, false) || w == null || w.array == null || |
1600 |
(runState & STOP) != 0 || (ws = workQueues) == null || |
1601 |
(m = ws.length - 1) < 0) // already terminating |
1602 |
break; |
1603 |
if ((sp = (int)(c = ctl)) != 0) { // wake up replacement |
1604 |
if (tryRelease(c, ws[sp & m], AC_UNIT)) |
1605 |
break; |
1606 |
} |
1607 |
else if (ex != null && (c & ADD_WORKER) != 0L) { |
1608 |
tryAddWorker(c); // create replacement |
1609 |
break; |
1610 |
} |
1611 |
else // don't need replacement |
1612 |
break; |
1613 |
} |
1614 |
if (ex == null) // help clean on way out |
1615 |
ForkJoinTask.helpExpungeStaleExceptions(); |
1616 |
else // rethrow |
1617 |
ForkJoinTask.rethrow(ex); |
1618 |
} |
1619 |
|
1620 |
// Signalling |
1621 |
|
1622 |
/** |
1623 |
* Tries to create or activate a worker if too few are active. |
1624 |
* |
1625 |
* @param ws the worker array to use to find signallees |
1626 |
* @param q a WorkQueue --if non-null, don't retry if now empty |
1627 |
*/ |
1628 |
final void signalWork(WorkQueue[] ws, WorkQueue q) { |
1629 |
long c; int sp, i; WorkQueue v; Thread p; |
1630 |
while ((c = ctl) < 0L) { // too few active |
1631 |
if ((sp = (int)c) == 0) { // no idle workers |
1632 |
if ((c & ADD_WORKER) != 0L) // too few workers |
1633 |
tryAddWorker(c); |
1634 |
break; |
1635 |
} |
1636 |
if (ws == null) // unstarted/terminated |
1637 |
break; |
1638 |
if (ws.length <= (i = sp & SMASK)) // terminated |
1639 |
break; |
1640 |
if ((v = ws[i]) == null) // terminating |
1641 |
break; |
1642 |
int vs = (sp + SS_SEQ) & ~INACTIVE; // next scanState |
1643 |
int d = sp - v.scanState; // screen CAS |
1644 |
long nc = (UC_MASK & (c + AC_UNIT)) | (SP_MASK & v.stackPred); |
1645 |
if (d == 0 && U.compareAndSwapLong(this, CTL, c, nc)) { |
1646 |
v.scanState = vs; // activate v |
1647 |
if ((p = v.parker) != null) |
1648 |
U.unpark(p); |
1649 |
break; |
1650 |
} |
1651 |
if (q != null && q.base == q.top) // no more work |
1652 |
break; |
1653 |
} |
1654 |
} |
1655 |
|
1656 |
/** |
1657 |
* Signals and releases worker v if it is top of idle worker |
1658 |
* stack. This performs a one-shot version of signalWork only if |
1659 |
* there is (apparently) at least one idle worker. |
1660 |
* |
1661 |
* @param c incoming ctl value |
1662 |
* @param v if non-null, a worker |
1663 |
* @param inc the increment to active count (zero when compensating) |
1664 |
* @return true if successful |
1665 |
*/ |
1666 |
private boolean tryRelease(long c, WorkQueue v, long inc) { |
1667 |
int sp = (int)c, vs = (sp + SS_SEQ) & ~INACTIVE; Thread p; |
1668 |
if (v != null && v.scanState == sp) { // v is at top of stack |
1669 |
long nc = (UC_MASK & (c + inc)) | (SP_MASK & v.stackPred); |
1670 |
if (U.compareAndSwapLong(this, CTL, c, nc)) { |
1671 |
v.scanState = vs; |
1672 |
if ((p = v.parker) != null) |
1673 |
U.unpark(p); |
1674 |
return true; |
1675 |
} |
1676 |
} |
1677 |
return false; |
1678 |
} |
1679 |
|
1680 |
// Scanning for tasks |
1681 |
|
1682 |
/** |
1683 |
* Top-level runloop for workers, called by ForkJoinWorkerThread.run. |
1684 |
*/ |
1685 |
final void runWorker(WorkQueue w) { |
1686 |
w.growArray(); // allocate queue |
1687 |
int seed = w.hint; // initially holds randomization hint |
1688 |
int r = (seed == 0) ? 1 : seed; // avoid 0 for xorShift |
1689 |
for (ForkJoinTask<?> t;;) { |
1690 |
if ((t = scan(w, r)) != null) |
1691 |
w.runTask(t); |
1692 |
else if (!awaitWork(w, r)) |
1693 |
break; |
1694 |
r ^= r << 13; r ^= r >>> 17; r ^= r << 5; // xorshift |
1695 |
} |
1696 |
} |
1697 |
|
1698 |
/** |
1699 |
* Scans for and tries to steal a top-level task. Scans start at a |
1700 |
* random location, randomly moving on apparent contention, |
1701 |
* otherwise continuing linearly until reaching two consecutive |
1702 |
* empty passes over all queues with the same checksum (summing |
1703 |
* each base index of each queue, that moves on each steal), at |
1704 |
* which point the worker tries to inactivate and then re-scans, |
1705 |
* attempting to re-activate (itself or some other worker) if |
1706 |
* finding a task; otherwise returning null to await work. Scans |
1707 |
* otherwise touch as little memory as possible, to reduce |
1708 |
* disruption on other scanning threads. |
1709 |
* |
1710 |
* @param w the worker (via its WorkQueue) |
1711 |
* @param r a random seed |
1712 |
* @return a task, or null if none found |
1713 |
*/ |
1714 |
private ForkJoinTask<?> scan(WorkQueue w, int r) { |
1715 |
WorkQueue[] ws; int m; |
1716 |
if ((ws = workQueues) != null && (m = ws.length - 1) > 0 && w != null) { |
1717 |
int ss = w.scanState; // initially non-negative |
1718 |
for (int origin = r & m, k = origin, oldSum = 0, checkSum = 0;;) { |
1719 |
WorkQueue q; ForkJoinTask<?> t; int al, i, n; long c; |
1720 |
if ((q = ws[k]) != null) { |
1721 |
int b = q.base; ForkJoinTask<?>[] a = q.array; |
1722 |
if ((n = b - q.top) < 0 && a != null && |
1723 |
(al = a.length) > 0) { // non-empty |
1724 |
if ((t = getAt(a, i = (al - 1) & b)) != null && |
1725 |
q.base == b) { |
1726 |
if (ss >= 0) { |
1727 |
if (casAt(a, i, t, null)) { |
1728 |
q.base = b + 1; |
1729 |
if (n < -1) // signal others |
1730 |
signalWork(ws, q); |
1731 |
return t; |
1732 |
} |
1733 |
} |
1734 |
else if (oldSum == 0 && // try to activate |
1735 |
w.scanState < 0) |
1736 |
tryRelease(c = ctl, ws[m & (int)c], AC_UNIT); |
1737 |
} |
1738 |
if (ss < 0) // refresh |
1739 |
ss = w.scanState; |
1740 |
r ^= r << 1; r ^= r >>> 3; r ^= r << 10; |
1741 |
origin = k = r & m; // move and rescan |
1742 |
oldSum = checkSum = 0; |
1743 |
continue; |
1744 |
} |
1745 |
checkSum += b; |
1746 |
} |
1747 |
if ((k = (k + 1) & m) == origin) { // continue until stable |
1748 |
if ((ss >= 0 || (ss == (ss = w.scanState))) && |
1749 |
oldSum == (oldSum = checkSum)) { |
1750 |
if (ss < 0 || w.qlock < 0) // already inactive |
1751 |
break; |
1752 |
int ns = ss | INACTIVE; // try to inactivate |
1753 |
long nc = ((SP_MASK & ns) | |
1754 |
(UC_MASK & ((c = ctl) - AC_UNIT))); |
1755 |
w.stackPred = (int)c; // hold prev stack top |
1756 |
U.putInt(w, QSCANSTATE, ns); |
1757 |
if (U.compareAndSwapLong(this, CTL, c, nc)) |
1758 |
ss = ns; |
1759 |
else |
1760 |
w.scanState = ss; // back out |
1761 |
} |
1762 |
checkSum = 0; |
1763 |
} |
1764 |
} |
1765 |
} |
1766 |
return null; |
1767 |
} |
1768 |
|
1769 |
/** |
1770 |
* Possibly blocks worker w waiting for a task to steal, or |
1771 |
* returns false if the worker should terminate. If inactivating |
1772 |
* w has caused the pool to become quiescent, checks for pool |
1773 |
* termination, and, so long as this is not the only worker, waits |
1774 |
* for up to a given duration. On timeout, if ctl has not |
1775 |
* changed, terminates the worker, which will in turn wake up |
1776 |
* another worker to possibly repeat this process. |
1777 |
* |
1778 |
* @param w the calling worker |
1779 |
* @param r a random seed (for spins) |
1780 |
* @return false if the worker should terminate |
1781 |
*/ |
1782 |
private boolean awaitWork(WorkQueue w, int r) { |
1783 |
if (w == null || w.qlock < 0) // w is terminating |
1784 |
return false; |
1785 |
for (int pred = w.stackPred, spins = SPINS, ss;;) { |
1786 |
if ((ss = w.scanState) >= 0) |
1787 |
break; |
1788 |
else if (spins > 0) { |
1789 |
r ^= r << 6; r ^= r >>> 21; r ^= r << 7; |
1790 |
if (r >= 0 && --spins == 0) { // randomize spins |
1791 |
WorkQueue v; WorkQueue[] ws; int s, j; AtomicLong sc; |
1792 |
if (pred != 0 && (ws = workQueues) != null && |
1793 |
(j = pred & SMASK) < ws.length && |
1794 |
(v = ws[j]) != null && // see if pred parking |
1795 |
(v.parker == null || v.scanState >= 0)) |
1796 |
spins = SPINS; // continue spinning |
1797 |
} |
1798 |
} |
1799 |
else if (w.qlock < 0) // recheck after spins |
1800 |
return false; |
1801 |
else if (!Thread.interrupted()) { |
1802 |
long c, prevctl, parkTime, deadline; |
1803 |
int ac = (int)((c = ctl) >> AC_SHIFT) + (config & SMASK); |
1804 |
if ((ac <= 0 && tryTerminate(false, false)) || |
1805 |
(runState & STOP) != 0) // pool terminating |
1806 |
return false; |
1807 |
if (ac <= 0 && ss == (int)c) { // is last waiter |
1808 |
prevctl = (UC_MASK & (c + AC_UNIT)) | (SP_MASK & pred); |
1809 |
int t = (short)(c >>> TC_SHIFT); // shrink excess spares |
1810 |
if (t > 2 && U.compareAndSwapLong(this, CTL, c, prevctl)) |
1811 |
return false; // else use timed wait |
1812 |
parkTime = IDLE_TIMEOUT * ((t >= 0) ? 1 : 1 - t); |
1813 |
deadline = System.nanoTime() + parkTime - TIMEOUT_SLOP; |
1814 |
} |
1815 |
else |
1816 |
prevctl = parkTime = deadline = 0L; |
1817 |
Thread wt = Thread.currentThread(); |
1818 |
U.putObject(wt, PARKBLOCKER, this); // emulate LockSupport |
1819 |
w.parker = wt; |
1820 |
if (w.scanState < 0 && ctl == c) // recheck before park |
1821 |
U.park(false, parkTime); |
1822 |
U.putOrderedObject(w, QPARKER, null); |
1823 |
U.putObject(wt, PARKBLOCKER, null); |
1824 |
if (w.scanState >= 0) |
1825 |
break; |
1826 |
if (parkTime != 0L && ctl == c && |
1827 |
deadline - System.nanoTime() <= 0L && |
1828 |
U.compareAndSwapLong(this, CTL, c, prevctl)) |
1829 |
return false; // shrink pool |
1830 |
} |
1831 |
} |
1832 |
return true; |
1833 |
} |
1834 |
|
1835 |
// Joining tasks |
1836 |
|
1837 |
/** |
1838 |
* Tries to steal and run tasks within the target's computation. |
1839 |
* Uses a variant of the top-level algorithm, restricted to tasks |
1840 |
* with the given task as ancestor: It prefers taking and running |
1841 |
* eligible tasks popped from the worker's own queue (via |
1842 |
* popCC). Otherwise it scans others, randomly moving on |
1843 |
* contention or execution, deciding to give up based on a |
1844 |
* checksum (via return codes frob pollAndExecCC). The maxTasks |
1845 |
* argument supports external usages; internal calls use zero, |
1846 |
* allowing unbounded steps (external calls trap non-positive |
1847 |
* values). |
1848 |
* |
1849 |
* @param w caller |
1850 |
* @param maxTasks if non-zero, the maximum number of other tasks to run |
1851 |
* @return task status on exit |
1852 |
*/ |
1853 |
final int helpComplete(WorkQueue w, CountedCompleter<?> task, |
1854 |
int maxTasks) { |
1855 |
WorkQueue[] ws; int s = 0, m; |
1856 |
if ((ws = workQueues) != null && (m = ws.length - 1) > 0 && |
1857 |
task != null && w != null) { |
1858 |
int mode = w.config; // for popCC |
1859 |
int r = w.hint ^ w.top; // arbitrary seed for origin |
1860 |
int origin = r & m; // first queue to scan |
1861 |
int h = 1; // 1:ran, >1:contended, <0:hash |
1862 |
for (int k = origin, oldSum = 0, checkSum = 0;;) { |
1863 |
CountedCompleter<?> p; WorkQueue q; |
1864 |
if ((s = task.status) < 0) |
1865 |
break; |
1866 |
if (h == 1 && (p = w.popCC(task, mode)) != null) { |
1867 |
p.doExec(); // run local task |
1868 |
if (maxTasks != 0 && --maxTasks == 0) |
1869 |
break; |
1870 |
origin = k; // reset |
1871 |
oldSum = checkSum = 0; |
1872 |
} |
1873 |
else { // poll other queues |
1874 |
if ((q = ws[k]) == null) |
1875 |
h = 0; |
1876 |
else if ((h = q.pollAndExecCC(task)) < 0) |
1877 |
checkSum += h; |
1878 |
if (h > 0) { |
1879 |
if (h == 1 && maxTasks != 0 && --maxTasks == 0) |
1880 |
break; |
1881 |
r ^= r << 13; r ^= r >>> 17; r ^= r << 5; // xorshift |
1882 |
origin = k = r & m; // move and restart |
1883 |
oldSum = checkSum = 0; |
1884 |
} |
1885 |
else if ((k = (k + 1) & m) == origin) { |
1886 |
if (oldSum == (oldSum = checkSum)) |
1887 |
break; |
1888 |
checkSum = 0; |
1889 |
} |
1890 |
} |
1891 |
} |
1892 |
} |
1893 |
return s; |
1894 |
} |
1895 |
|
1896 |
/** |
1897 |
* Tries to locate and execute tasks for a stealer of the given |
1898 |
* task, or in turn one of its stealers, Traces currentSteal -> |
1899 |
* currentJoin links looking for a thread working on a descendant |
1900 |
* of the given task and with a non-empty queue to steal back and |
1901 |
* execute tasks from. The first call to this method upon a |
1902 |
* waiting join will often entail scanning/search, (which is OK |
1903 |
* because the joiner has nothing better to do), but this method |
1904 |
* leaves hints in workers to speed up subsequent calls. |
1905 |
* |
1906 |
* @param w caller |
1907 |
* @param task the task to join |
1908 |
*/ |
1909 |
private void helpStealer(WorkQueue w, ForkJoinTask<?> task) { |
1910 |
WorkQueue[] ws = workQueues; |
1911 |
int oldSum = 0, checkSum, m; |
1912 |
if (ws != null && (m = ws.length - 1) > 0 && w != null && |
1913 |
task != null) { |
1914 |
do { // restart point |
1915 |
checkSum = 0; // for stability check |
1916 |
ForkJoinTask<?> subtask; |
1917 |
WorkQueue j = w, v; // v is subtask stealer |
1918 |
descent: for (subtask = task; subtask.status >= 0; ) { |
1919 |
for (int h = j.hint | 1, k = 0, i; ; k += 2) { |
1920 |
if (k > m) // can't find stealer |
1921 |
break descent; |
1922 |
if ((v = ws[i = (h + k) & m]) != null) { |
1923 |
if (v.currentSteal == subtask) { |
1924 |
j.hint = i; |
1925 |
break; |
1926 |
} |
1927 |
checkSum += v.base; |
1928 |
} |
1929 |
} |
1930 |
for (;;) { // help v or descend |
1931 |
ForkJoinTask<?>[] a; int b, al, i; |
1932 |
checkSum += (b = v.base); |
1933 |
ForkJoinTask<?> next = v.currentJoin; |
1934 |
if (subtask.status < 0 || j.currentJoin != subtask || |
1935 |
v.currentSteal != subtask) // stale |
1936 |
break descent; |
1937 |
if (b - v.top >= 0 || (a = v.array) == null || |
1938 |
(al = a.length) <= 0) { |
1939 |
if ((subtask = next) == null) |
1940 |
break descent; |
1941 |
j = v; |
1942 |
break; |
1943 |
} |
1944 |
ForkJoinTask<?> t = getAt(a, i = (al - 1) & b); |
1945 |
if (v.base == b) { |
1946 |
if (t == null) // stale |
1947 |
break descent; |
1948 |
if (casAt(a, i, t, null)) { |
1949 |
v.base = b + 1; |
1950 |
ForkJoinTask<?> ps = w.currentSteal; |
1951 |
int top = w.top; |
1952 |
do { |
1953 |
U.putOrderedObject(w, QCURRENTSTEAL, t); |
1954 |
t.doExec(); // clear local tasks too |
1955 |
} while (task.status >= 0 && |
1956 |
w.top != top && |
1957 |
(t = w.pop()) != null); |
1958 |
U.putOrderedObject(w, QCURRENTSTEAL, ps); |
1959 |
if (w.base != w.top) |
1960 |
return; // can't further help |
1961 |
} |
1962 |
} |
1963 |
} |
1964 |
} |
1965 |
} while (task.status >= 0 && oldSum != (oldSum = checkSum)); |
1966 |
} |
1967 |
} |
1968 |
|
1969 |
/** |
1970 |
* Tries to decrement active count (sometimes implicitly) and |
1971 |
* possibly release or create a compensating worker in preparation |
1972 |
* for blocking. Returns false (retryable by caller), on |
1973 |
* contention, detected staleness, instability, or termination. |
1974 |
* |
1975 |
* @param w caller |
1976 |
*/ |
1977 |
private boolean tryCompensate(WorkQueue w) { |
1978 |
boolean canBlock; |
1979 |
WorkQueue[] ws; long c; int m, pc, sp; |
1980 |
if (w == null || w.qlock < 0 || // caller terminating |
1981 |
(ws = workQueues) == null || (m = ws.length - 1) <= 0 || |
1982 |
(pc = config & SMASK) == 0) // parallelism disabled |
1983 |
canBlock = false; |
1984 |
else if ((sp = (int)(c = ctl)) != 0) // release idle worker |
1985 |
canBlock = tryRelease(c, ws[sp & m], 0L); |
1986 |
else { |
1987 |
int ac = (int)(c >> AC_SHIFT) + pc; |
1988 |
int tc = (short)(c >> TC_SHIFT) + pc; |
1989 |
int nbusy = 0; // validate saturation |
1990 |
for (int i = 0; i <= m; ++i) { // two passes of odd indices |
1991 |
WorkQueue v; |
1992 |
if ((v = ws[((i << 1) | 1) & m]) != null) { |
1993 |
if ((v.scanState & SCANNING) != 0) |
1994 |
break; |
1995 |
++nbusy; |
1996 |
} |
1997 |
} |
1998 |
if (nbusy != (tc << 1) || ctl != c) |
1999 |
canBlock = false; // unstable or stale |
2000 |
else if (tc >= pc && ac > 1 && w.isEmpty()) { |
2001 |
long nc = ((AC_MASK & (c - AC_UNIT)) | |
2002 |
(~AC_MASK & c)); // uncompensated |
2003 |
canBlock = U.compareAndSwapLong(this, CTL, c, nc); |
2004 |
} |
2005 |
else if (tc >= MAX_CAP || |
2006 |
(this == common && tc >= pc + commonMaxSpares)) |
2007 |
throw new RejectedExecutionException( |
2008 |
"Thread limit exceeded replacing blocked worker"); |
2009 |
else { // similar to tryAddWorker |
2010 |
boolean add = false; int rs; // CAS within lock |
2011 |
long nc = ((AC_MASK & c) | |
2012 |
(TC_MASK & (c + TC_UNIT))); |
2013 |
if (((rs = lockRunState()) & STOP) == 0) |
2014 |
add = U.compareAndSwapLong(this, CTL, c, nc); |
2015 |
unlockRunState(rs, rs & ~RSLOCK); |
2016 |
canBlock = add && createWorker(); // throws on exception |
2017 |
} |
2018 |
} |
2019 |
return canBlock; |
2020 |
} |
2021 |
|
2022 |
/** |
2023 |
* Helps and/or blocks until the given task is done or timeout. |
2024 |
* |
2025 |
* @param w caller |
2026 |
* @param task the task |
2027 |
* @param deadline for timed waits, if nonzero |
2028 |
* @return task status on exit |
2029 |
*/ |
2030 |
final int awaitJoin(WorkQueue w, ForkJoinTask<?> task, long deadline) { |
2031 |
int s = 0; |
2032 |
if (task != null && w != null) { |
2033 |
ForkJoinTask<?> prevJoin = w.currentJoin; |
2034 |
U.putOrderedObject(w, QCURRENTJOIN, task); |
2035 |
CountedCompleter<?> cc = (task instanceof CountedCompleter) ? |
2036 |
(CountedCompleter<?>)task : null; |
2037 |
for (;;) { |
2038 |
if ((s = task.status) < 0) |
2039 |
break; |
2040 |
if (cc != null) |
2041 |
helpComplete(w, cc, 0); |
2042 |
else if (w.base == w.top || w.tryRemoveAndExec(task)) |
2043 |
helpStealer(w, task); |
2044 |
if ((s = task.status) < 0) |
2045 |
break; |
2046 |
long ms, ns; |
2047 |
if (deadline == 0L) |
2048 |
ms = 0L; |
2049 |
else if ((ns = deadline - System.nanoTime()) <= 0L) |
2050 |
break; |
2051 |
else if ((ms = TimeUnit.NANOSECONDS.toMillis(ns)) <= 0L) |
2052 |
ms = 1L; |
2053 |
if (tryCompensate(w)) { |
2054 |
task.internalWait(ms); |
2055 |
U.getAndAddLong(this, CTL, AC_UNIT); |
2056 |
} |
2057 |
} |
2058 |
U.putOrderedObject(w, QCURRENTJOIN, prevJoin); |
2059 |
} |
2060 |
return s; |
2061 |
} |
2062 |
|
2063 |
// Specialized scanning |
2064 |
|
2065 |
/** |
2066 |
* Returns a (probably) non-empty steal queue, if one is found |
2067 |
* during a scan, else null. This method must be retried by |
2068 |
* caller if, by the time it tries to use the queue, it is empty. |
2069 |
*/ |
2070 |
private WorkQueue findNonEmptyStealQueue() { |
2071 |
WorkQueue[] ws; int m; // one-shot version of scan loop |
2072 |
int r = ThreadLocalRandom.nextSecondarySeed(); |
2073 |
if ((ws = workQueues) != null && (m = ws.length - 1) > 0) { |
2074 |
for (int origin = r & m, k = origin, oldSum = 0, checkSum = 0;;) { |
2075 |
WorkQueue q; int b; |
2076 |
if ((q = ws[k]) != null) { |
2077 |
if ((b = q.base) - q.top < 0) |
2078 |
return q; |
2079 |
checkSum += b; |
2080 |
} |
2081 |
if ((k = (k + 1) & m) == origin) { |
2082 |
if (oldSum == (oldSum = checkSum)) |
2083 |
break; |
2084 |
checkSum = 0; |
2085 |
} |
2086 |
} |
2087 |
} |
2088 |
return null; |
2089 |
} |
2090 |
|
2091 |
/** |
2092 |
* Runs tasks until {@code isQuiescent()}. We piggyback on |
2093 |
* active count ctl maintenance, but rather than blocking |
2094 |
* when tasks cannot be found, we rescan until all others cannot |
2095 |
* find tasks either. |
2096 |
*/ |
2097 |
final void helpQuiescePool(WorkQueue w) { |
2098 |
ForkJoinTask<?> ps = w.currentSteal; // save context |
2099 |
for (boolean active = true;;) { |
2100 |
long c; WorkQueue q; ForkJoinTask<?> t; int b; |
2101 |
if ((w.config & FIFO_QUEUE) != 0) |
2102 |
w.pollAndExecAll(); // run locals before each scan |
2103 |
else |
2104 |
w.popAndExecAll(); |
2105 |
if ((q = findNonEmptyStealQueue()) != null) { |
2106 |
if (!active) { // re-establish active count |
2107 |
active = true; |
2108 |
U.getAndAddLong(this, CTL, AC_UNIT); |
2109 |
} |
2110 |
if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null) { |
2111 |
U.putOrderedObject(w, QCURRENTSTEAL, t); |
2112 |
t.doExec(); |
2113 |
if (++w.nsteals < 0) |
2114 |
w.transferStealCount(this); |
2115 |
} |
2116 |
} |
2117 |
else if (active) { // decrement active count without queuing |
2118 |
long nc = (AC_MASK & ((c = ctl) - AC_UNIT)) | (~AC_MASK & c); |
2119 |
if ((int)(nc >> AC_SHIFT) + (config & SMASK) <= 0) |
2120 |
break; // bypass decrement-then-increment |
2121 |
if (U.compareAndSwapLong(this, CTL, c, nc)) |
2122 |
active = false; |
2123 |
} |
2124 |
else if ((int)((c = ctl) >> AC_SHIFT) + (config & SMASK) <= 0 && |
2125 |
U.compareAndSwapLong(this, CTL, c, c + AC_UNIT)) |
2126 |
break; |
2127 |
} |
2128 |
U.putOrderedObject(w, QCURRENTSTEAL, ps); |
2129 |
} |
2130 |
|
2131 |
/** |
2132 |
* Gets and removes a local or stolen task for the given worker. |
2133 |
* |
2134 |
* @return a task, if available |
2135 |
*/ |
2136 |
final ForkJoinTask<?> nextTaskFor(WorkQueue w) { |
2137 |
for (ForkJoinTask<?> t;;) { |
2138 |
WorkQueue q; int b; |
2139 |
if ((t = w.nextLocalTask()) != null) |
2140 |
return t; |
2141 |
if ((q = findNonEmptyStealQueue()) == null) |
2142 |
return null; |
2143 |
if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null) |
2144 |
return t; |
2145 |
} |
2146 |
} |
2147 |
|
2148 |
/** |
2149 |
* Returns a cheap heuristic guide for task partitioning when |
2150 |
* programmers, frameworks, tools, or languages have little or no |
2151 |
* idea about task granularity. In essence, by offering this |
2152 |
* method, we ask users only about tradeoffs in overhead vs |
2153 |
* expected throughput and its variance, rather than how finely to |
2154 |
* partition tasks. |
2155 |
* |
2156 |
* In a steady state strict (tree-structured) computation, each |
2157 |
* thread makes available for stealing enough tasks for other |
2158 |
* threads to remain active. Inductively, if all threads play by |
2159 |
* the same rules, each thread should make available only a |
2160 |
* constant number of tasks. |
2161 |
* |
2162 |
* The minimum useful constant is just 1. But using a value of 1 |
2163 |
* would require immediate replenishment upon each steal to |
2164 |
* maintain enough tasks, which is infeasible. Further, |
2165 |
* partitionings/granularities of offered tasks should minimize |
2166 |
* steal rates, which in general means that threads nearer the top |
2167 |
* of computation tree should generate more than those nearer the |
2168 |
* bottom. In perfect steady state, each thread is at |
2169 |
* approximately the same level of computation tree. However, |
2170 |
* producing extra tasks amortizes the uncertainty of progress and |
2171 |
* diffusion assumptions. |
2172 |
* |
2173 |
* So, users will want to use values larger (but not much larger) |
2174 |
* than 1 to both smooth over transient shortages and hedge |
2175 |
* against uneven progress; as traded off against the cost of |
2176 |
* extra task overhead. We leave the user to pick a threshold |
2177 |
* value to compare with the results of this call to guide |
2178 |
* decisions, but recommend values such as 3. |
2179 |
* |
2180 |
* When all threads are active, it is on average OK to estimate |
2181 |
* surplus strictly locally. In steady-state, if one thread is |
2182 |
* maintaining say 2 surplus tasks, then so are others. So we can |
2183 |
* just use estimated queue length. However, this strategy alone |
2184 |
* leads to serious mis-estimates in some non-steady-state |
2185 |
* conditions (ramp-up, ramp-down, other stalls). We can detect |
2186 |
* many of these by further considering the number of "idle" |
2187 |
* threads, that are known to have zero queued tasks, so |
2188 |
* compensate by a factor of (#idle/#active) threads. |
2189 |
*/ |
2190 |
static int getSurplusQueuedTaskCount() { |
2191 |
Thread t; ForkJoinWorkerThread wt; ForkJoinPool pool; WorkQueue q; |
2192 |
if ((t = Thread.currentThread()) instanceof ForkJoinWorkerThread) { |
2193 |
int p = (pool = (wt = (ForkJoinWorkerThread)t).pool). |
2194 |
config & SMASK; |
2195 |
int n = (q = wt.workQueue).top - q.base; |
2196 |
int a = (int)(pool.ctl >> AC_SHIFT) + p; |
2197 |
return n - (a > (p >>>= 1) ? 0 : |
2198 |
a > (p >>>= 1) ? 1 : |
2199 |
a > (p >>>= 1) ? 2 : |
2200 |
a > (p >>>= 1) ? 4 : |
2201 |
8); |
2202 |
} |
2203 |
return 0; |
2204 |
} |
2205 |
|
2206 |
// Termination |
2207 |
|
2208 |
/** |
2209 |
* Possibly initiates and/or completes termination. |
2210 |
* |
2211 |
* @param now if true, unconditionally terminate, else only |
2212 |
* if no work and no active workers |
2213 |
* @param enable if true, enable shutdown when next possible |
2214 |
* @return true if now terminating or terminated |
2215 |
*/ |
2216 |
private boolean tryTerminate(boolean now, boolean enable) { |
2217 |
int rs; |
2218 |
if (this == common) // cannot shut down |
2219 |
return false; |
2220 |
if ((rs = runState) >= 0) { |
2221 |
if (!enable) |
2222 |
return false; |
2223 |
rs = lockRunState(); // enter SHUTDOWN phase |
2224 |
unlockRunState(rs, (rs & ~RSLOCK) | SHUTDOWN); |
2225 |
} |
2226 |
|
2227 |
if ((rs & STOP) == 0) { |
2228 |
if (!now) { // check quiescence |
2229 |
for (long oldSum = 0L;;) { // repeat until stable |
2230 |
WorkQueue[] ws; WorkQueue w; int m, b; long c; |
2231 |
long checkSum = ctl; |
2232 |
if ((int)(checkSum >> AC_SHIFT) + (config & SMASK) > 0) |
2233 |
return false; // still active workers |
2234 |
if ((ws = workQueues) == null || (m = ws.length - 1) <= 0) |
2235 |
break; // check queues |
2236 |
for (int i = 0; i <= m; ++i) { |
2237 |
if ((w = ws[i]) != null) { |
2238 |
if ((b = w.base) != w.top || w.scanState >= 0 || |
2239 |
w.currentSteal != null) { |
2240 |
tryRelease(c = ctl, ws[m & (int)c], AC_UNIT); |
2241 |
return false; // arrange for recheck |
2242 |
} |
2243 |
checkSum += b; |
2244 |
if ((i & 1) == 0) |
2245 |
w.qlock = -1; // try to disable external |
2246 |
} |
2247 |
} |
2248 |
if (oldSum == (oldSum = checkSum)) |
2249 |
break; |
2250 |
} |
2251 |
} |
2252 |
if ((runState & STOP) == 0) { |
2253 |
rs = lockRunState(); // enter STOP phase |
2254 |
unlockRunState(rs, (rs & ~RSLOCK) | STOP); |
2255 |
} |
2256 |
} |
2257 |
|
2258 |
int pass = 0; // 3 passes to help terminate |
2259 |
for (long oldSum = 0L;;) { // or until done or stable |
2260 |
WorkQueue[] ws; WorkQueue w; ForkJoinWorkerThread wt; int m; |
2261 |
long checkSum = ctl; |
2262 |
if ((short)(checkSum >>> TC_SHIFT) + (config & SMASK) <= 0 || |
2263 |
(ws = workQueues) == null || (m = ws.length - 1) <= 0) { |
2264 |
if ((runState & TERMINATED) == 0) { |
2265 |
rs = lockRunState(); // done |
2266 |
unlockRunState(rs, (rs & ~RSLOCK) | TERMINATED); |
2267 |
synchronized (this) { notifyAll(); } // for awaitTermination |
2268 |
} |
2269 |
break; |
2270 |
} |
2271 |
for (int i = 0; i <= m; ++i) { |
2272 |
if ((w = ws[i]) != null) { |
2273 |
checkSum += w.base; |
2274 |
w.qlock = -1; // try to disable |
2275 |
if (pass > 0) { |
2276 |
w.cancelAll(); // clear queue |
2277 |
if (pass > 1 && (wt = w.owner) != null) { |
2278 |
if (!wt.isInterrupted()) { |
2279 |
try { // unblock join |
2280 |
wt.interrupt(); |
2281 |
} catch (Throwable ignore) { |
2282 |
} |
2283 |
} |
2284 |
if (w.scanState < 0) |
2285 |
U.unpark(wt); // wake up |
2286 |
} |
2287 |
} |
2288 |
} |
2289 |
} |
2290 |
if (checkSum != oldSum) { // unstable |
2291 |
oldSum = checkSum; |
2292 |
pass = 0; |
2293 |
} |
2294 |
else if (pass > 3 && pass > m) // can't further help |
2295 |
break; |
2296 |
else if (++pass > 1) { // try to dequeue |
2297 |
long c; int j = 0, sp; // bound attempts |
2298 |
while (j++ <= m && (sp = (int)(c = ctl)) != 0) |
2299 |
tryRelease(c, ws[sp & m], AC_UNIT); |
2300 |
} |
2301 |
} |
2302 |
return true; |
2303 |
} |
2304 |
|
2305 |
// External operations |
2306 |
|
2307 |
/** |
2308 |
* Full version of externalPush, handling uncommon cases, as well |
2309 |
* as performing secondary initialization upon the first |
2310 |
* submission of the first task to the pool. It also detects |
2311 |
* first submission by an external thread and creates a new shared |
2312 |
* queue if the one at index if empty or contended. |
2313 |
* |
2314 |
* @param task the task. Caller must ensure non-null. |
2315 |
*/ |
2316 |
private void externalSubmit(ForkJoinTask<?> task) { |
2317 |
int r; // initialize caller's probe |
2318 |
if ((r = ThreadLocalRandom.getProbe()) == 0) { |
2319 |
ThreadLocalRandom.localInit(); |
2320 |
r = ThreadLocalRandom.getProbe(); |
2321 |
} |
2322 |
for (;;) { |
2323 |
WorkQueue[] ws; WorkQueue q; int rs, m, k; |
2324 |
boolean move = false; |
2325 |
if ((rs = runState) < 0) { |
2326 |
tryTerminate(false, false); // help terminate |
2327 |
throw new RejectedExecutionException(); |
2328 |
} |
2329 |
else if ((rs & STARTED) == 0 || // initialize |
2330 |
((ws = workQueues) == null || (m = ws.length - 1) <= 0)) { |
2331 |
int ns = 0; |
2332 |
rs = lockRunState(); |
2333 |
try { |
2334 |
if ((rs & STARTED) == 0) { |
2335 |
U.compareAndSwapObject(this, STEALCOUNTER, null, |
2336 |
new AtomicLong()); |
2337 |
// create workQueues array with size a power of two |
2338 |
int p = config & SMASK; // ensure at least 2 slots |
2339 |
int n = (p > 1) ? p - 1 : 1; |
2340 |
n |= n >>> 1; n |= n >>> 2; n |= n >>> 4; |
2341 |
n |= n >>> 8; n |= n >>> 16; n = (n + 1) << 1; |
2342 |
workQueues = new WorkQueue[n]; |
2343 |
ns = STARTED; |
2344 |
} |
2345 |
} finally { |
2346 |
unlockRunState(rs, (rs & ~RSLOCK) | ns); |
2347 |
} |
2348 |
} |
2349 |
else if ((q = ws[k = r & m & SQMASK]) != null) { |
2350 |
if (q.qlock == 0 && U.compareAndSwapInt(q, QLOCK, 0, 1)) { |
2351 |
ForkJoinTask<?>[] a = q.array; |
2352 |
int s = q.top; |
2353 |
boolean submitted = false; // initial submission or resizing |
2354 |
try { // locked version of push |
2355 |
if ((a != null && a.length > s + 1 - q.base) || |
2356 |
(a = q.growArray()) != null) { |
2357 |
int al = a.length, j = (al - 1) & s; |
2358 |
if (al > 0) { |
2359 |
setAt(a, j, task); |
2360 |
U.putOrderedInt(q, QTOP, s + 1); |
2361 |
submitted = true; |
2362 |
} |
2363 |
} |
2364 |
} finally { |
2365 |
U.compareAndSwapInt(q, QLOCK, 1, 0); |
2366 |
} |
2367 |
if (submitted) { |
2368 |
signalWork(ws, q); |
2369 |
return; |
2370 |
} |
2371 |
} |
2372 |
move = true; // move on failure |
2373 |
} |
2374 |
else if (((rs = runState) & RSLOCK) == 0) { // create new queue |
2375 |
q = new WorkQueue(this, null); |
2376 |
q.hint = r; |
2377 |
q.config = k | SHARED_QUEUE; |
2378 |
q.scanState = INACTIVE; |
2379 |
rs = lockRunState(); // publish index |
2380 |
if (rs > 0 && (ws = workQueues) != null && |
2381 |
k < ws.length && ws[k] == null) |
2382 |
ws[k] = q; // else terminated |
2383 |
unlockRunState(rs, rs & ~RSLOCK); |
2384 |
} |
2385 |
else |
2386 |
move = true; // move if busy |
2387 |
if (move) |
2388 |
r = ThreadLocalRandom.advanceProbe(r); |
2389 |
} |
2390 |
} |
2391 |
|
2392 |
/** |
2393 |
* Tries to add the given task to a submission queue at |
2394 |
* submitter's current queue. Only the (vastly) most common path |
2395 |
* is directly handled in this method, while screening for need |
2396 |
* for externalSubmit. |
2397 |
* |
2398 |
* @param task the task. Caller must ensure non-null. |
2399 |
*/ |
2400 |
final void externalPush(ForkJoinTask<?> task) { |
2401 |
WorkQueue[] ws; WorkQueue q; int m; |
2402 |
int r = ThreadLocalRandom.getProbe(); |
2403 |
int rs = runState; |
2404 |
if ((ws = workQueues) != null && (m = (ws.length - 1)) > 0 && |
2405 |
(q = ws[m & r & SQMASK]) != null && r != 0 && rs > 0 && |
2406 |
U.compareAndSwapInt(q, QLOCK, 0, 1)) { |
2407 |
ForkJoinTask<?>[] a; |
2408 |
if ((a = q.array) != null) { |
2409 |
int b = q.base, al = a.length, s = q.top; |
2410 |
if (al > 0) { |
2411 |
int am = al - 1, j = am & s, n; |
2412 |
if ((n = s - b) < am) { |
2413 |
setAt(a, j, task); |
2414 |
U.putOrderedInt(q, QTOP, s + 1); |
2415 |
U.putOrderedInt(q, QLOCK, 0); |
2416 |
if (n <= 1) |
2417 |
signalWork(ws, q); |
2418 |
return; |
2419 |
} |
2420 |
} |
2421 |
} |
2422 |
U.compareAndSwapInt(q, QLOCK, 1, 0); |
2423 |
} |
2424 |
externalSubmit(task); |
2425 |
} |
2426 |
|
2427 |
/** |
2428 |
* Pushes a possibly-external submission. |
2429 |
*/ |
2430 |
final <T> ForkJoinTask<T> doSubmit(ForkJoinTask<T> task) { |
2431 |
Thread t; ForkJoinWorkerThread w; |
2432 |
if (task == null) |
2433 |
throw new NullPointerException(); |
2434 |
if (((t = Thread.currentThread()) instanceof ForkJoinWorkerThread) && |
2435 |
(w = (ForkJoinWorkerThread)t).pool == this) |
2436 |
w.workQueue.push(task); |
2437 |
else |
2438 |
externalPush(task); |
2439 |
return task; |
2440 |
} |
2441 |
|
2442 |
/** |
2443 |
* Returns common pool queue for an external thread. |
2444 |
*/ |
2445 |
static WorkQueue commonSubmitterQueue() { |
2446 |
ForkJoinPool p = common; |
2447 |
int r = ThreadLocalRandom.getProbe(); |
2448 |
WorkQueue[] ws; int m; |
2449 |
return (p != null && (ws = p.workQueues) != null && |
2450 |
(m = ws.length - 1) > 0) ? |
2451 |
ws[m & r & SQMASK] : null; |
2452 |
} |
2453 |
|
2454 |
/** |
2455 |
* Performs tryUnpush for an external submitter: Finds queue, |
2456 |
* locks if apparently non-empty, validates upon locking, and |
2457 |
* adjusts top. Each check can fail but rarely does. |
2458 |
*/ |
2459 |
final boolean tryExternalUnpush(ForkJoinTask<?> task) { |
2460 |
WorkQueue[] ws; WorkQueue w; ForkJoinTask<?>[] a; int m; |
2461 |
int r = ThreadLocalRandom.getProbe(); |
2462 |
if ((ws = workQueues) != null && (m = ws.length - 1) > 0 && |
2463 |
(w = ws[m & r & SQMASK]) != null && |
2464 |
(a = w.array) != null) { |
2465 |
int b = w.base, al = a.length, s = w.top; |
2466 |
if (s != b && al > 0 && |
2467 |
U.compareAndSwapInt(w, QLOCK, 0, 1)) { |
2468 |
int i = (al - 1) & (s - 1); |
2469 |
if (w.top == s && w.array == a && casAt(a, i, task, null)) { |
2470 |
U.putOrderedInt(w, QTOP, s - 1); |
2471 |
U.putOrderedInt(w, QLOCK, 0); |
2472 |
return true; |
2473 |
} |
2474 |
U.compareAndSwapInt(w, QLOCK, 1, 0); |
2475 |
} |
2476 |
} |
2477 |
return false; |
2478 |
} |
2479 |
|
2480 |
/** |
2481 |
* Performs helpComplete for an external submitter. |
2482 |
*/ |
2483 |
final int externalHelpComplete(CountedCompleter<?> task, int maxTasks) { |
2484 |
WorkQueue[] ws; int n; |
2485 |
int r = ThreadLocalRandom.getProbe(); |
2486 |
return ((ws = workQueues) == null || (n = ws.length) == 0) ? 0 : |
2487 |
helpComplete(ws[(n - 1) & r & SQMASK], task, maxTasks); |
2488 |
} |
2489 |
|
2490 |
// Exported methods |
2491 |
|
2492 |
// Constructors |
2493 |
|
2494 |
/** |
2495 |
* Creates a {@code ForkJoinPool} with parallelism equal to {@link |
2496 |
* java.lang.Runtime#availableProcessors}, using the {@linkplain |
2497 |
* #defaultForkJoinWorkerThreadFactory default thread factory}, |
2498 |
* no UncaughtExceptionHandler, and non-async LIFO processing mode. |
2499 |
* |
2500 |
* @throws SecurityException if a security manager exists and |
2501 |
* the caller is not permitted to modify threads |
2502 |
* because it does not hold {@link |
2503 |
* java.lang.RuntimePermission}{@code ("modifyThread")} |
2504 |
*/ |
2505 |
public ForkJoinPool() { |
2506 |
this(Math.min(MAX_CAP, Runtime.getRuntime().availableProcessors()), |
2507 |
defaultForkJoinWorkerThreadFactory, null, false); |
2508 |
} |
2509 |
|
2510 |
/** |
2511 |
* Creates a {@code ForkJoinPool} with the indicated parallelism |
2512 |
* level, the {@linkplain |
2513 |
* #defaultForkJoinWorkerThreadFactory default thread factory}, |
2514 |
* no UncaughtExceptionHandler, and non-async LIFO processing mode. |
2515 |
* |
2516 |
* @param parallelism the parallelism level |
2517 |
* @throws IllegalArgumentException if parallelism less than or |
2518 |
* equal to zero, or greater than implementation limit |
2519 |
* @throws SecurityException if a security manager exists and |
2520 |
* the caller is not permitted to modify threads |
2521 |
* because it does not hold {@link |
2522 |
* java.lang.RuntimePermission}{@code ("modifyThread")} |
2523 |
*/ |
2524 |
public ForkJoinPool(int parallelism) { |
2525 |
this(parallelism, defaultForkJoinWorkerThreadFactory, null, false); |
2526 |
} |
2527 |
|
2528 |
/** |
2529 |
* Creates a {@code ForkJoinPool} with the given parameters. |
2530 |
* |
2531 |
* @param parallelism the parallelism level. For default value, |
2532 |
* use {@link java.lang.Runtime#availableProcessors}. |
2533 |
* @param factory the factory for creating new threads. For default value, |
2534 |
* use {@link #defaultForkJoinWorkerThreadFactory}. |
2535 |
* @param handler the handler for internal worker threads that |
2536 |
* terminate due to unrecoverable errors encountered while executing |
2537 |
* tasks. For default value, use {@code null}. |
2538 |
* @param asyncMode if true, |
2539 |
* establishes local first-in-first-out scheduling mode for forked |
2540 |
* tasks that are never joined. This mode may be more appropriate |
2541 |
* than default locally stack-based mode in applications in which |
2542 |
* worker threads only process event-style asynchronous tasks. |
2543 |
* For default value, use {@code false}. |
2544 |
* @throws IllegalArgumentException if parallelism less than or |
2545 |
* equal to zero, or greater than implementation limit |
2546 |
* @throws NullPointerException if the factory is null |
2547 |
* @throws SecurityException if a security manager exists and |
2548 |
* the caller is not permitted to modify threads |
2549 |
* because it does not hold {@link |
2550 |
* java.lang.RuntimePermission}{@code ("modifyThread")} |
2551 |
*/ |
2552 |
public ForkJoinPool(int parallelism, |
2553 |
ForkJoinWorkerThreadFactory factory, |
2554 |
UncaughtExceptionHandler handler, |
2555 |
boolean asyncMode) { |
2556 |
this(checkParallelism(parallelism), |
2557 |
checkFactory(factory), |
2558 |
handler, |
2559 |
asyncMode ? FIFO_QUEUE : LIFO_QUEUE, |
2560 |
"ForkJoinPool-" + nextPoolId() + "-worker-"); |
2561 |
checkPermission(); |
2562 |
} |
2563 |
|
2564 |
private static int checkParallelism(int parallelism) { |
2565 |
if (parallelism <= 0 || parallelism > MAX_CAP) |
2566 |
throw new IllegalArgumentException(); |
2567 |
return parallelism; |
2568 |
} |
2569 |
|
2570 |
private static ForkJoinWorkerThreadFactory checkFactory |
2571 |
(ForkJoinWorkerThreadFactory factory) { |
2572 |
if (factory == null) |
2573 |
throw new NullPointerException(); |
2574 |
return factory; |
2575 |
} |
2576 |
|
2577 |
/** |
2578 |
* Creates a {@code ForkJoinPool} with the given parameters, without |
2579 |
* any security checks or parameter validation. Invoked directly by |
2580 |
* makeCommonPool. |
2581 |
*/ |
2582 |
private ForkJoinPool(int parallelism, |
2583 |
ForkJoinWorkerThreadFactory factory, |
2584 |
UncaughtExceptionHandler handler, |
2585 |
int mode, |
2586 |
String workerNamePrefix) { |
2587 |
this.workerNamePrefix = workerNamePrefix; |
2588 |
this.factory = factory; |
2589 |
this.ueh = handler; |
2590 |
this.config = (parallelism & SMASK) | mode; |
2591 |
long np = (long)(-parallelism); // offset ctl counts |
2592 |
this.ctl = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK); |
2593 |
} |
2594 |
|
2595 |
/** |
2596 |
* Returns the common pool instance. This pool is statically |
2597 |
* constructed; its run state is unaffected by attempts to {@link |
2598 |
* #shutdown} or {@link #shutdownNow}. However this pool and any |
2599 |
* ongoing processing are automatically terminated upon program |
2600 |
* {@link System#exit}. Any program that relies on asynchronous |
2601 |
* task processing to complete before program termination should |
2602 |
* invoke {@code commonPool().}{@link #awaitQuiescence awaitQuiescence}, |
2603 |
* before exit. |
2604 |
* |
2605 |
* @return the common pool instance |
2606 |
* @since 1.8 |
2607 |
*/ |
2608 |
public static ForkJoinPool commonPool() { |
2609 |
// assert common != null : "static init error"; |
2610 |
return common; |
2611 |
} |
2612 |
|
2613 |
// Execution methods |
2614 |
|
2615 |
/** |
2616 |
* Performs the given task, returning its result upon completion. |
2617 |
* If the computation encounters an unchecked Exception or Error, |
2618 |
* it is rethrown as the outcome of this invocation. Rethrown |
2619 |
* exceptions behave in the same way as regular exceptions, but, |
2620 |
* when possible, contain stack traces (as displayed for example |
2621 |
* using {@code ex.printStackTrace()}) of both the current thread |
2622 |
* as well as the thread actually encountering the exception; |
2623 |
* minimally only the latter. |
2624 |
* |
2625 |
* @param task the task |
2626 |
* @param <T> the type of the task's result |
2627 |
* @return the task's result |
2628 |
* @throws NullPointerException if the task is null |
2629 |
* @throws RejectedExecutionException if the task cannot be |
2630 |
* scheduled for execution |
2631 |
*/ |
2632 |
public <T> T invoke(ForkJoinTask<T> task) { |
2633 |
if (task == null) |
2634 |
throw new NullPointerException(); |
2635 |
doSubmit(task); |
2636 |
return task.join(); |
2637 |
} |
2638 |
|
2639 |
/** |
2640 |
* Arranges for (asynchronous) execution of the given task. |
2641 |
* |
2642 |
* @param task the task |
2643 |
* @throws NullPointerException if the task is null |
2644 |
* @throws RejectedExecutionException if the task cannot be |
2645 |
* scheduled for execution |
2646 |
*/ |
2647 |
public void execute(ForkJoinTask<?> task) { |
2648 |
doSubmit(task); |
2649 |
} |
2650 |
|
2651 |
// AbstractExecutorService methods |
2652 |
|
2653 |
/** |
2654 |
* @throws NullPointerException if the task is null |
2655 |
* @throws RejectedExecutionException if the task cannot be |
2656 |
* scheduled for execution |
2657 |
*/ |
2658 |
public void execute(Runnable task) { |
2659 |
if (task == null) |
2660 |
throw new NullPointerException(); |
2661 |
ForkJoinTask<?> job; |
2662 |
if (task instanceof ForkJoinTask<?>) // avoid re-wrap |
2663 |
job = (ForkJoinTask<?>) task; |
2664 |
else |
2665 |
job = new ForkJoinTask.RunnableExecuteAction(task); |
2666 |
doSubmit(job); |
2667 |
} |
2668 |
|
2669 |
/** |
2670 |
* Submits a ForkJoinTask for execution. |
2671 |
* |
2672 |
* @param task the task to submit |
2673 |
* @param <T> the type of the task's result |
2674 |
* @return the task |
2675 |
* @throws NullPointerException if the task is null |
2676 |
* @throws RejectedExecutionException if the task cannot be |
2677 |
* scheduled for execution |
2678 |
*/ |
2679 |
public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) { |
2680 |
return doSubmit(task); |
2681 |
} |
2682 |
|
2683 |
/** |
2684 |
* @throws NullPointerException if the task is null |
2685 |
* @throws RejectedExecutionException if the task cannot be |
2686 |
* scheduled for execution |
2687 |
*/ |
2688 |
public <T> ForkJoinTask<T> submit(Callable<T> task) { |
2689 |
return doSubmit(new ForkJoinTask.AdaptedCallable<T>(task)); |
2690 |
} |
2691 |
|
2692 |
/** |
2693 |
* @throws NullPointerException if the task is null |
2694 |
* @throws RejectedExecutionException if the task cannot be |
2695 |
* scheduled for execution |
2696 |
*/ |
2697 |
public <T> ForkJoinTask<T> submit(Runnable task, T result) { |
2698 |
return doSubmit(new ForkJoinTask.AdaptedRunnable<T>(task, result)); |
2699 |
} |
2700 |
|
2701 |
/** |
2702 |
* @throws NullPointerException if the task is null |
2703 |
* @throws RejectedExecutionException if the task cannot be |
2704 |
* scheduled for execution |
2705 |
*/ |
2706 |
public ForkJoinTask<?> submit(Runnable task) { |
2707 |
if (task == null) |
2708 |
throw new NullPointerException(); |
2709 |
ForkJoinTask<?> job; |
2710 |
if (task instanceof ForkJoinTask<?>) // avoid re-wrap |
2711 |
job = (ForkJoinTask<?>) task; |
2712 |
else |
2713 |
job = new ForkJoinTask.AdaptedRunnableAction(task); |
2714 |
return doSubmit(job); |
2715 |
} |
2716 |
|
2717 |
/** |
2718 |
* @throws NullPointerException {@inheritDoc} |
2719 |
* @throws RejectedExecutionException {@inheritDoc} |
2720 |
*/ |
2721 |
public <T> List<Future<T>> invokeAll(Collection<? extends Callable<T>> tasks) { |
2722 |
// In previous versions of this class, this method constructed |
2723 |
// a task to run ForkJoinTask.invokeAll, but now external |
2724 |
// invocation of multiple tasks is at least as efficient. |
2725 |
ArrayList<Future<T>> futures = new ArrayList<>(tasks.size()); |
2726 |
|
2727 |
try { |
2728 |
for (Callable<T> t : tasks) { |
2729 |
ForkJoinTask<T> f = new ForkJoinTask.AdaptedCallable<T>(t); |
2730 |
futures.add(f); |
2731 |
doSubmit(f); |
2732 |
} |
2733 |
for (int i = 0, size = futures.size(); i < size; i++) |
2734 |
((ForkJoinTask<?>)futures.get(i)).quietlyJoin(); |
2735 |
return futures; |
2736 |
} catch (Throwable t) { |
2737 |
for (int i = 0, size = futures.size(); i < size; i++) |
2738 |
futures.get(i).cancel(false); |
2739 |
throw t; |
2740 |
} |
2741 |
} |
2742 |
|
2743 |
/** |
2744 |
* Returns the factory used for constructing new workers. |
2745 |
* |
2746 |
* @return the factory used for constructing new workers |
2747 |
*/ |
2748 |
public ForkJoinWorkerThreadFactory getFactory() { |
2749 |
return factory; |
2750 |
} |
2751 |
|
2752 |
/** |
2753 |
* Returns the handler for internal worker threads that terminate |
2754 |
* due to unrecoverable errors encountered while executing tasks. |
2755 |
* |
2756 |
* @return the handler, or {@code null} if none |
2757 |
*/ |
2758 |
public UncaughtExceptionHandler getUncaughtExceptionHandler() { |
2759 |
return ueh; |
2760 |
} |
2761 |
|
2762 |
/** |
2763 |
* Returns the targeted parallelism level of this pool. |
2764 |
* |
2765 |
* @return the targeted parallelism level of this pool |
2766 |
*/ |
2767 |
public int getParallelism() { |
2768 |
int par; |
2769 |
return ((par = config & SMASK) > 0) ? par : 1; |
2770 |
} |
2771 |
|
2772 |
/** |
2773 |
* Returns the targeted parallelism level of the common pool. |
2774 |
* |
2775 |
* @return the targeted parallelism level of the common pool |
2776 |
* @since 1.8 |
2777 |
*/ |
2778 |
public static int getCommonPoolParallelism() { |
2779 |
return commonParallelism; |
2780 |
} |
2781 |
|
2782 |
/** |
2783 |
* Returns the number of worker threads that have started but not |
2784 |
* yet terminated. The result returned by this method may differ |
2785 |
* from {@link #getParallelism} when threads are created to |
2786 |
* maintain parallelism when others are cooperatively blocked. |
2787 |
* |
2788 |
* @return the number of worker threads |
2789 |
*/ |
2790 |
public int getPoolSize() { |
2791 |
return (config & SMASK) + (short)(ctl >>> TC_SHIFT); |
2792 |
} |
2793 |
|
2794 |
/** |
2795 |
* Returns {@code true} if this pool uses local first-in-first-out |
2796 |
* scheduling mode for forked tasks that are never joined. |
2797 |
* |
2798 |
* @return {@code true} if this pool uses async mode |
2799 |
*/ |
2800 |
public boolean getAsyncMode() { |
2801 |
return (config & FIFO_QUEUE) != 0; |
2802 |
} |
2803 |
|
2804 |
/** |
2805 |
* Returns an estimate of the number of worker threads that are |
2806 |
* not blocked waiting to join tasks or for other managed |
2807 |
* synchronization. This method may overestimate the |
2808 |
* number of running threads. |
2809 |
* |
2810 |
* @return the number of worker threads |
2811 |
*/ |
2812 |
public int getRunningThreadCount() { |
2813 |
int rc = 0; |
2814 |
WorkQueue[] ws; WorkQueue w; |
2815 |
if ((ws = workQueues) != null) { |
2816 |
for (int i = 1; i < ws.length; i += 2) { |
2817 |
if ((w = ws[i]) != null && w.isApparentlyUnblocked()) |
2818 |
++rc; |
2819 |
} |
2820 |
} |
2821 |
return rc; |
2822 |
} |
2823 |
|
2824 |
/** |
2825 |
* Returns an estimate of the number of threads that are currently |
2826 |
* stealing or executing tasks. This method may overestimate the |
2827 |
* number of active threads. |
2828 |
* |
2829 |
* @return the number of active threads |
2830 |
*/ |
2831 |
public int getActiveThreadCount() { |
2832 |
int r = (config & SMASK) + (int)(ctl >> AC_SHIFT); |
2833 |
return (r <= 0) ? 0 : r; // suppress momentarily negative values |
2834 |
} |
2835 |
|
2836 |
/** |
2837 |
* Returns {@code true} if all worker threads are currently idle. |
2838 |
* An idle worker is one that cannot obtain a task to execute |
2839 |
* because none are available to steal from other threads, and |
2840 |
* there are no pending submissions to the pool. This method is |
2841 |
* conservative; it might not return {@code true} immediately upon |
2842 |
* idleness of all threads, but will eventually become true if |
2843 |
* threads remain inactive. |
2844 |
* |
2845 |
* @return {@code true} if all threads are currently idle |
2846 |
*/ |
2847 |
public boolean isQuiescent() { |
2848 |
return (config & SMASK) + (int)(ctl >> AC_SHIFT) <= 0; |
2849 |
} |
2850 |
|
2851 |
/** |
2852 |
* Returns an estimate of the total number of tasks stolen from |
2853 |
* one thread's work queue by another. The reported value |
2854 |
* underestimates the actual total number of steals when the pool |
2855 |
* is not quiescent. This value may be useful for monitoring and |
2856 |
* tuning fork/join programs: in general, steal counts should be |
2857 |
* high enough to keep threads busy, but low enough to avoid |
2858 |
* overhead and contention across threads. |
2859 |
* |
2860 |
* @return the number of steals |
2861 |
*/ |
2862 |
public long getStealCount() { |
2863 |
AtomicLong sc = stealCounter; |
2864 |
long count = (sc == null) ? 0L : sc.get(); |
2865 |
WorkQueue[] ws; WorkQueue w; |
2866 |
if ((ws = workQueues) != null) { |
2867 |
for (int i = 1; i < ws.length; i += 2) { |
2868 |
if ((w = ws[i]) != null) |
2869 |
count += w.nsteals; |
2870 |
} |
2871 |
} |
2872 |
return count; |
2873 |
} |
2874 |
|
2875 |
/** |
2876 |
* Returns an estimate of the total number of tasks currently held |
2877 |
* in queues by worker threads (but not including tasks submitted |
2878 |
* to the pool that have not begun executing). This value is only |
2879 |
* an approximation, obtained by iterating across all threads in |
2880 |
* the pool. This method may be useful for tuning task |
2881 |
* granularities. |
2882 |
* |
2883 |
* @return the number of queued tasks |
2884 |
*/ |
2885 |
public long getQueuedTaskCount() { |
2886 |
long count = 0; |
2887 |
WorkQueue[] ws; WorkQueue w; |
2888 |
if ((ws = workQueues) != null) { |
2889 |
for (int i = 1; i < ws.length; i += 2) { |
2890 |
if ((w = ws[i]) != null) |
2891 |
count += w.queueSize(); |
2892 |
} |
2893 |
} |
2894 |
return count; |
2895 |
} |
2896 |
|
2897 |
/** |
2898 |
* Returns an estimate of the number of tasks submitted to this |
2899 |
* pool that have not yet begun executing. This method may take |
2900 |
* time proportional to the number of submissions. |
2901 |
* |
2902 |
* @return the number of queued submissions |
2903 |
*/ |
2904 |
public int getQueuedSubmissionCount() { |
2905 |
int count = 0; |
2906 |
WorkQueue[] ws; WorkQueue w; |
2907 |
if ((ws = workQueues) != null) { |
2908 |
for (int i = 0; i < ws.length; i += 2) { |
2909 |
if ((w = ws[i]) != null) |
2910 |
count += w.queueSize(); |
2911 |
} |
2912 |
} |
2913 |
return count; |
2914 |
} |
2915 |
|
2916 |
/** |
2917 |
* Returns {@code true} if there are any tasks submitted to this |
2918 |
* pool that have not yet begun executing. |
2919 |
* |
2920 |
* @return {@code true} if there are any queued submissions |
2921 |
*/ |
2922 |
public boolean hasQueuedSubmissions() { |
2923 |
WorkQueue[] ws; WorkQueue w; |
2924 |
if ((ws = workQueues) != null) { |
2925 |
for (int i = 0; i < ws.length; i += 2) { |
2926 |
if ((w = ws[i]) != null && !w.isEmpty()) |
2927 |
return true; |
2928 |
} |
2929 |
} |
2930 |
return false; |
2931 |
} |
2932 |
|
2933 |
/** |
2934 |
* Removes and returns the next unexecuted submission if one is |
2935 |
* available. This method may be useful in extensions to this |
2936 |
* class that re-assign work in systems with multiple pools. |
2937 |
* |
2938 |
* @return the next submission, or {@code null} if none |
2939 |
*/ |
2940 |
protected ForkJoinTask<?> pollSubmission() { |
2941 |
WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t; |
2942 |
if ((ws = workQueues) != null) { |
2943 |
for (int i = 0; i < ws.length; i += 2) { |
2944 |
if ((w = ws[i]) != null && (t = w.poll()) != null) |
2945 |
return t; |
2946 |
} |
2947 |
} |
2948 |
return null; |
2949 |
} |
2950 |
|
2951 |
/** |
2952 |
* Removes all available unexecuted submitted and forked tasks |
2953 |
* from scheduling queues and adds them to the given collection, |
2954 |
* without altering their execution status. These may include |
2955 |
* artificially generated or wrapped tasks. This method is |
2956 |
* designed to be invoked only when the pool is known to be |
2957 |
* quiescent. Invocations at other times may not remove all |
2958 |
* tasks. A failure encountered while attempting to add elements |
2959 |
* to collection {@code c} may result in elements being in |
2960 |
* neither, either or both collections when the associated |
2961 |
* exception is thrown. The behavior of this operation is |
2962 |
* undefined if the specified collection is modified while the |
2963 |
* operation is in progress. |
2964 |
* |
2965 |
* @param c the collection to transfer elements into |
2966 |
* @return the number of elements transferred |
2967 |
*/ |
2968 |
protected int drainTasksTo(Collection<? super ForkJoinTask<?>> c) { |
2969 |
int count = 0; |
2970 |
WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t; |
2971 |
if ((ws = workQueues) != null) { |
2972 |
for (int i = 0; i < ws.length; ++i) { |
2973 |
if ((w = ws[i]) != null) { |
2974 |
while ((t = w.poll()) != null) { |
2975 |
c.add(t); |
2976 |
++count; |
2977 |
} |
2978 |
} |
2979 |
} |
2980 |
} |
2981 |
return count; |
2982 |
} |
2983 |
|
2984 |
/** |
2985 |
* Returns a string identifying this pool, as well as its state, |
2986 |
* including indications of run state, parallelism level, and |
2987 |
* worker and task counts. |
2988 |
* |
2989 |
* @return a string identifying this pool, as well as its state |
2990 |
*/ |
2991 |
public String toString() { |
2992 |
// Use a single pass through workQueues to collect counts |
2993 |
long qt = 0L, qs = 0L; int rc = 0; |
2994 |
AtomicLong sc = stealCounter; |
2995 |
long st = (sc == null) ? 0L : sc.get(); |
2996 |
long c = ctl; |
2997 |
WorkQueue[] ws; WorkQueue w; |
2998 |
if ((ws = workQueues) != null) { |
2999 |
for (int i = 0; i < ws.length; ++i) { |
3000 |
if ((w = ws[i]) != null) { |
3001 |
int size = w.queueSize(); |
3002 |
if ((i & 1) == 0) |
3003 |
qs += size; |
3004 |
else { |
3005 |
qt += size; |
3006 |
st += w.nsteals; |
3007 |
if (w.isApparentlyUnblocked()) |
3008 |
++rc; |
3009 |
} |
3010 |
} |
3011 |
} |
3012 |
} |
3013 |
int pc = (config & SMASK); |
3014 |
int tc = pc + (short)(c >>> TC_SHIFT); |
3015 |
int ac = pc + (int)(c >> AC_SHIFT); |
3016 |
if (ac < 0) // ignore transient negative |
3017 |
ac = 0; |
3018 |
int rs = runState; |
3019 |
String level = ((rs & TERMINATED) != 0 ? "Terminated" : |
3020 |
(rs & STOP) != 0 ? "Terminating" : |
3021 |
(rs & SHUTDOWN) != 0 ? "Shutting down" : |
3022 |
"Running"); |
3023 |
return super.toString() + |
3024 |
"[" + level + |
3025 |
", parallelism = " + pc + |
3026 |
", size = " + tc + |
3027 |
", active = " + ac + |
3028 |
", running = " + rc + |
3029 |
", steals = " + st + |
3030 |
", tasks = " + qt + |
3031 |
", submissions = " + qs + |
3032 |
"]"; |
3033 |
} |
3034 |
|
3035 |
/** |
3036 |
* Possibly initiates an orderly shutdown in which previously |
3037 |
* submitted tasks are executed, but no new tasks will be |
3038 |
* accepted. Invocation has no effect on execution state if this |
3039 |
* is the {@link #commonPool()}, and no additional effect if |
3040 |
* already shut down. Tasks that are in the process of being |
3041 |
* submitted concurrently during the course of this method may or |
3042 |
* may not be rejected. |
3043 |
* |
3044 |
* @throws SecurityException if a security manager exists and |
3045 |
* the caller is not permitted to modify threads |
3046 |
* because it does not hold {@link |
3047 |
* java.lang.RuntimePermission}{@code ("modifyThread")} |
3048 |
*/ |
3049 |
public void shutdown() { |
3050 |
checkPermission(); |
3051 |
tryTerminate(false, true); |
3052 |
} |
3053 |
|
3054 |
/** |
3055 |
* Possibly attempts to cancel and/or stop all tasks, and reject |
3056 |
* all subsequently submitted tasks. Invocation has no effect on |
3057 |
* execution state if this is the {@link #commonPool()}, and no |
3058 |
* additional effect if already shut down. Otherwise, tasks that |
3059 |
* are in the process of being submitted or executed concurrently |
3060 |
* during the course of this method may or may not be |
3061 |
* rejected. This method cancels both existing and unexecuted |
3062 |
* tasks, in order to permit termination in the presence of task |
3063 |
* dependencies. So the method always returns an empty list |
3064 |
* (unlike the case for some other Executors). |
3065 |
* |
3066 |
* @return an empty list |
3067 |
* @throws SecurityException if a security manager exists and |
3068 |
* the caller is not permitted to modify threads |
3069 |
* because it does not hold {@link |
3070 |
* java.lang.RuntimePermission}{@code ("modifyThread")} |
3071 |
*/ |
3072 |
public List<Runnable> shutdownNow() { |
3073 |
checkPermission(); |
3074 |
tryTerminate(true, true); |
3075 |
return Collections.emptyList(); |
3076 |
} |
3077 |
|
3078 |
/** |
3079 |
* Returns {@code true} if all tasks have completed following shut down. |
3080 |
* |
3081 |
* @return {@code true} if all tasks have completed following shut down |
3082 |
*/ |
3083 |
public boolean isTerminated() { |
3084 |
return (runState & TERMINATED) != 0; |
3085 |
} |
3086 |
|
3087 |
/** |
3088 |
* Returns {@code true} if the process of termination has |
3089 |
* commenced but not yet completed. This method may be useful for |
3090 |
* debugging. A return of {@code true} reported a sufficient |
3091 |
* period after shutdown may indicate that submitted tasks have |
3092 |
* ignored or suppressed interruption, or are waiting for I/O, |
3093 |
* causing this executor not to properly terminate. (See the |
3094 |
* advisory notes for class {@link ForkJoinTask} stating that |
3095 |
* tasks should not normally entail blocking operations. But if |
3096 |
* they do, they must abort them on interrupt.) |
3097 |
* |
3098 |
* @return {@code true} if terminating but not yet terminated |
3099 |
*/ |
3100 |
public boolean isTerminating() { |
3101 |
int rs = runState; |
3102 |
return (rs & STOP) != 0 && (rs & TERMINATED) == 0; |
3103 |
} |
3104 |
|
3105 |
/** |
3106 |
* Returns {@code true} if this pool has been shut down. |
3107 |
* |
3108 |
* @return {@code true} if this pool has been shut down |
3109 |
*/ |
3110 |
public boolean isShutdown() { |
3111 |
return (runState & SHUTDOWN) != 0; |
3112 |
} |
3113 |
|
3114 |
/** |
3115 |
* Blocks until all tasks have completed execution after a |
3116 |
* shutdown request, or the timeout occurs, or the current thread |
3117 |
* is interrupted, whichever happens first. Because the {@link |
3118 |
* #commonPool()} never terminates until program shutdown, when |
3119 |
* applied to the common pool, this method is equivalent to {@link |
3120 |
* #awaitQuiescence(long, TimeUnit)} but always returns {@code false}. |
3121 |
* |
3122 |
* @param timeout the maximum time to wait |
3123 |
* @param unit the time unit of the timeout argument |
3124 |
* @return {@code true} if this executor terminated and |
3125 |
* {@code false} if the timeout elapsed before termination |
3126 |
* @throws InterruptedException if interrupted while waiting |
3127 |
*/ |
3128 |
public boolean awaitTermination(long timeout, TimeUnit unit) |
3129 |
throws InterruptedException { |
3130 |
if (Thread.interrupted()) |
3131 |
throw new InterruptedException(); |
3132 |
if (this == common) { |
3133 |
awaitQuiescence(timeout, unit); |
3134 |
return false; |
3135 |
} |
3136 |
long nanos = unit.toNanos(timeout); |
3137 |
if (isTerminated()) |
3138 |
return true; |
3139 |
if (nanos <= 0L) |
3140 |
return false; |
3141 |
long deadline = System.nanoTime() + nanos; |
3142 |
synchronized (this) { |
3143 |
for (;;) { |
3144 |
if (isTerminated()) |
3145 |
return true; |
3146 |
if (nanos <= 0L) |
3147 |
return false; |
3148 |
long millis = TimeUnit.NANOSECONDS.toMillis(nanos); |
3149 |
wait(millis > 0L ? millis : 1L); |
3150 |
nanos = deadline - System.nanoTime(); |
3151 |
} |
3152 |
} |
3153 |
} |
3154 |
|
3155 |
/** |
3156 |
* If called by a ForkJoinTask operating in this pool, equivalent |
3157 |
* in effect to {@link ForkJoinTask#helpQuiesce}. Otherwise, |
3158 |
* waits and/or attempts to assist performing tasks until this |
3159 |
* pool {@link #isQuiescent} or the indicated timeout elapses. |
3160 |
* |
3161 |
* @param timeout the maximum time to wait |
3162 |
* @param unit the time unit of the timeout argument |
3163 |
* @return {@code true} if quiescent; {@code false} if the |
3164 |
* timeout elapsed. |
3165 |
*/ |
3166 |
public boolean awaitQuiescence(long timeout, TimeUnit unit) { |
3167 |
long nanos = unit.toNanos(timeout); |
3168 |
ForkJoinWorkerThread wt; |
3169 |
Thread thread = Thread.currentThread(); |
3170 |
if ((thread instanceof ForkJoinWorkerThread) && |
3171 |
(wt = (ForkJoinWorkerThread)thread).pool == this) { |
3172 |
helpQuiescePool(wt.workQueue); |
3173 |
return true; |
3174 |
} |
3175 |
long startTime = System.nanoTime(); |
3176 |
WorkQueue[] ws; |
3177 |
int r = 0, m; |
3178 |
boolean found = true; |
3179 |
while (!isQuiescent() && (ws = workQueues) != null && |
3180 |
(m = ws.length - 1) > 0) { |
3181 |
if (!found) { |
3182 |
if ((System.nanoTime() - startTime) > nanos) |
3183 |
return false; |
3184 |
Thread.yield(); // cannot block |
3185 |
} |
3186 |
found = false; |
3187 |
for (int j = (m + 1) << 2; j >= 0; --j) { |
3188 |
ForkJoinTask<?> t; WorkQueue q; int b, k; |
3189 |
if ((k = r++ & m) <= m && k >= 0 && (q = ws[k]) != null && |
3190 |
(b = q.base) - q.top < 0) { |
3191 |
found = true; |
3192 |
if ((t = q.pollAt(b)) != null) |
3193 |
t.doExec(); |
3194 |
break; |
3195 |
} |
3196 |
} |
3197 |
} |
3198 |
return true; |
3199 |
} |
3200 |
|
3201 |
/** |
3202 |
* Waits and/or attempts to assist performing tasks indefinitely |
3203 |
* until the {@link #commonPool()} {@link #isQuiescent}. |
3204 |
*/ |
3205 |
static void quiesceCommonPool() { |
3206 |
common.awaitQuiescence(Long.MAX_VALUE, TimeUnit.NANOSECONDS); |
3207 |
} |
3208 |
|
3209 |
/** |
3210 |
* Interface for extending managed parallelism for tasks running |
3211 |
* in {@link ForkJoinPool}s. |
3212 |
* |
3213 |
* <p>A {@code ManagedBlocker} provides two methods. Method |
3214 |
* {@link #isReleasable} must return {@code true} if blocking is |
3215 |
* not necessary. Method {@link #block} blocks the current thread |
3216 |
* if necessary (perhaps internally invoking {@code isReleasable} |
3217 |
* before actually blocking). These actions are performed by any |
3218 |
* thread invoking {@link ForkJoinPool#managedBlock(ManagedBlocker)}. |
3219 |
* The unusual methods in this API accommodate synchronizers that |
3220 |
* may, but don't usually, block for long periods. Similarly, they |
3221 |
* allow more efficient internal handling of cases in which |
3222 |
* additional workers may be, but usually are not, needed to |
3223 |
* ensure sufficient parallelism. Toward this end, |
3224 |
* implementations of method {@code isReleasable} must be amenable |
3225 |
* to repeated invocation. |
3226 |
* |
3227 |
* <p>For example, here is a ManagedBlocker based on a |
3228 |
* ReentrantLock: |
3229 |
* <pre> {@code |
3230 |
* class ManagedLocker implements ManagedBlocker { |
3231 |
* final ReentrantLock lock; |
3232 |
* boolean hasLock = false; |
3233 |
* ManagedLocker(ReentrantLock lock) { this.lock = lock; } |
3234 |
* public boolean block() { |
3235 |
* if (!hasLock) |
3236 |
* lock.lock(); |
3237 |
* return true; |
3238 |
* } |
3239 |
* public boolean isReleasable() { |
3240 |
* return hasLock || (hasLock = lock.tryLock()); |
3241 |
* } |
3242 |
* }}</pre> |
3243 |
* |
3244 |
* <p>Here is a class that possibly blocks waiting for an |
3245 |
* item on a given queue: |
3246 |
* <pre> {@code |
3247 |
* class QueueTaker<E> implements ManagedBlocker { |
3248 |
* final BlockingQueue<E> queue; |
3249 |
* volatile E item = null; |
3250 |
* QueueTaker(BlockingQueue<E> q) { this.queue = q; } |
3251 |
* public boolean block() throws InterruptedException { |
3252 |
* if (item == null) |
3253 |
* item = queue.take(); |
3254 |
* return true; |
3255 |
* } |
3256 |
* public boolean isReleasable() { |
3257 |
* return item != null || (item = queue.poll()) != null; |
3258 |
* } |
3259 |
* public E getItem() { // call after pool.managedBlock completes |
3260 |
* return item; |
3261 |
* } |
3262 |
* }}</pre> |
3263 |
*/ |
3264 |
public static interface ManagedBlocker { |
3265 |
/** |
3266 |
* Possibly blocks the current thread, for example waiting for |
3267 |
* a lock or condition. |
3268 |
* |
3269 |
* @return {@code true} if no additional blocking is necessary |
3270 |
* (i.e., if isReleasable would return true) |
3271 |
* @throws InterruptedException if interrupted while waiting |
3272 |
* (the method is not required to do so, but is allowed to) |
3273 |
*/ |
3274 |
boolean block() throws InterruptedException; |
3275 |
|
3276 |
/** |
3277 |
* Returns {@code true} if blocking is unnecessary. |
3278 |
* @return {@code true} if blocking is unnecessary |
3279 |
*/ |
3280 |
boolean isReleasable(); |
3281 |
} |
3282 |
|
3283 |
/** |
3284 |
* Runs the given possibly blocking task. When {@linkplain |
3285 |
* ForkJoinTask#inForkJoinPool() running in a ForkJoinPool}, this |
3286 |
* method possibly arranges for a spare thread to be activated if |
3287 |
* necessary to ensure sufficient parallelism while the current |
3288 |
* thread is blocked in {@link ManagedBlocker#block blocker.block()}. |
3289 |
* |
3290 |
* <p>This method repeatedly calls {@code blocker.isReleasable()} and |
3291 |
* {@code blocker.block()} until either method returns {@code true}. |
3292 |
* Every call to {@code blocker.block()} is preceded by a call to |
3293 |
* {@code blocker.isReleasable()} that returned {@code false}. |
3294 |
* |
3295 |
* <p>If not running in a ForkJoinPool, this method is |
3296 |
* behaviorally equivalent to |
3297 |
* <pre> {@code |
3298 |
* while (!blocker.isReleasable()) |
3299 |
* if (blocker.block()) |
3300 |
* break;}</pre> |
3301 |
* |
3302 |
* If running in a ForkJoinPool, the pool may first be expanded to |
3303 |
* ensure sufficient parallelism available during the call to |
3304 |
* {@code blocker.block()}. |
3305 |
* |
3306 |
* @param blocker the blocker task |
3307 |
* @throws InterruptedException if {@code blocker.block()} did so |
3308 |
*/ |
3309 |
public static void managedBlock(ManagedBlocker blocker) |
3310 |
throws InterruptedException { |
3311 |
ForkJoinPool p; |
3312 |
ForkJoinWorkerThread wt; |
3313 |
Thread t = Thread.currentThread(); |
3314 |
if ((t instanceof ForkJoinWorkerThread) && |
3315 |
(p = (wt = (ForkJoinWorkerThread)t).pool) != null) { |
3316 |
WorkQueue w = wt.workQueue; |
3317 |
while (!blocker.isReleasable()) { |
3318 |
if (p.tryCompensate(w)) { |
3319 |
try { |
3320 |
do {} while (!blocker.isReleasable() && |
3321 |
!blocker.block()); |
3322 |
} finally { |
3323 |
U.getAndAddLong(p, CTL, AC_UNIT); |
3324 |
} |
3325 |
break; |
3326 |
} |
3327 |
} |
3328 |
} |
3329 |
else { |
3330 |
do {} while (!blocker.isReleasable() && |
3331 |
!blocker.block()); |
3332 |
} |
3333 |
} |
3334 |
|
3335 |
// AbstractExecutorService overrides. These rely on undocumented |
3336 |
// fact that ForkJoinTask.adapt returns ForkJoinTasks that also |
3337 |
// implement RunnableFuture. |
3338 |
|
3339 |
protected <T> RunnableFuture<T> newTaskFor(Runnable runnable, T value) { |
3340 |
return new ForkJoinTask.AdaptedRunnable<T>(runnable, value); |
3341 |
} |
3342 |
|
3343 |
protected <T> RunnableFuture<T> newTaskFor(Callable<T> callable) { |
3344 |
return new ForkJoinTask.AdaptedCallable<T>(callable); |
3345 |
} |
3346 |
|
3347 |
// Unsafe mechanics |
3348 |
private static final sun.misc.Unsafe U = sun.misc.Unsafe.getUnsafe(); |
3349 |
private static final int ABASE; |
3350 |
private static final int ASHIFT; |
3351 |
private static final long CTL; |
3352 |
private static final long RUNSTATE; |
3353 |
private static final long STEALCOUNTER; |
3354 |
private static final long PARKBLOCKER; |
3355 |
private static final long QTOP; |
3356 |
private static final long QLOCK; |
3357 |
private static final long QSCANSTATE; |
3358 |
private static final long QPARKER; |
3359 |
private static final long QCURRENTSTEAL; |
3360 |
private static final long QCURRENTJOIN; |
3361 |
|
3362 |
static { |
3363 |
try { |
3364 |
CTL = U.objectFieldOffset |
3365 |
(ForkJoinPool.class.getDeclaredField("ctl")); |
3366 |
RUNSTATE = U.objectFieldOffset |
3367 |
(ForkJoinPool.class.getDeclaredField("runState")); |
3368 |
STEALCOUNTER = U.objectFieldOffset |
3369 |
(ForkJoinPool.class.getDeclaredField("stealCounter")); |
3370 |
|
3371 |
PARKBLOCKER = U.objectFieldOffset |
3372 |
(Thread.class.getDeclaredField("parkBlocker")); |
3373 |
|
3374 |
QTOP = U.objectFieldOffset |
3375 |
(WorkQueue.class.getDeclaredField("top")); |
3376 |
QLOCK = U.objectFieldOffset |
3377 |
(WorkQueue.class.getDeclaredField("qlock")); |
3378 |
QSCANSTATE = U.objectFieldOffset |
3379 |
(WorkQueue.class.getDeclaredField("scanState")); |
3380 |
QPARKER = U.objectFieldOffset |
3381 |
(WorkQueue.class.getDeclaredField("parker")); |
3382 |
QCURRENTSTEAL = U.objectFieldOffset |
3383 |
(WorkQueue.class.getDeclaredField("currentSteal")); |
3384 |
QCURRENTJOIN = U.objectFieldOffset |
3385 |
(WorkQueue.class.getDeclaredField("currentJoin")); |
3386 |
|
3387 |
ABASE = U.arrayBaseOffset(ForkJoinTask[].class); |
3388 |
int scale = U.arrayIndexScale(ForkJoinTask[].class); |
3389 |
if ((scale & (scale - 1)) != 0) |
3390 |
throw new Error("array index scale not a power of two"); |
3391 |
ASHIFT = 31 - Integer.numberOfLeadingZeros(scale); |
3392 |
} catch (ReflectiveOperationException e) { |
3393 |
throw new Error(e); |
3394 |
} |
3395 |
|
3396 |
commonMaxSpares = DEFAULT_COMMON_MAX_SPARES; |
3397 |
defaultForkJoinWorkerThreadFactory = |
3398 |
new DefaultForkJoinWorkerThreadFactory(); |
3399 |
modifyThreadPermission = new RuntimePermission("modifyThread"); |
3400 |
|
3401 |
common = java.security.AccessController.doPrivileged |
3402 |
(new java.security.PrivilegedAction<ForkJoinPool>() { |
3403 |
public ForkJoinPool run() { return makeCommonPool(); }}); |
3404 |
int par = common.config & SMASK; // report 1 even if threads disabled |
3405 |
commonParallelism = par > 0 ? par : 1; |
3406 |
} |
3407 |
|
3408 |
/** |
3409 |
* Creates and returns the common pool, respecting user settings |
3410 |
* specified via system properties. |
3411 |
*/ |
3412 |
private static ForkJoinPool makeCommonPool() { |
3413 |
int parallelism = -1; |
3414 |
ForkJoinWorkerThreadFactory factory = null; |
3415 |
UncaughtExceptionHandler handler = null; |
3416 |
try { // ignore exceptions in accessing/parsing properties |
3417 |
String pp = System.getProperty |
3418 |
("java.util.concurrent.ForkJoinPool.common.parallelism"); |
3419 |
String fp = System.getProperty |
3420 |
("java.util.concurrent.ForkJoinPool.common.threadFactory"); |
3421 |
String hp = System.getProperty |
3422 |
("java.util.concurrent.ForkJoinPool.common.exceptionHandler"); |
3423 |
String mp = System.getProperty |
3424 |
("java.util.concurrent.ForkJoinPool.common.maximumSpares"); |
3425 |
if (pp != null) |
3426 |
parallelism = Integer.parseInt(pp); |
3427 |
if (fp != null) |
3428 |
factory = ((ForkJoinWorkerThreadFactory)ClassLoader. |
3429 |
getSystemClassLoader().loadClass(fp).newInstance()); |
3430 |
if (hp != null) |
3431 |
handler = ((UncaughtExceptionHandler)ClassLoader. |
3432 |
getSystemClassLoader().loadClass(hp).newInstance()); |
3433 |
if (mp != null) |
3434 |
commonMaxSpares = Integer.parseInt(mp); |
3435 |
} catch (Exception ignore) { |
3436 |
} |
3437 |
if (factory == null) { |
3438 |
if (System.getSecurityManager() == null) |
3439 |
factory = defaultForkJoinWorkerThreadFactory; |
3440 |
else // use security-managed default |
3441 |
factory = new InnocuousForkJoinWorkerThreadFactory(); |
3442 |
} |
3443 |
if (parallelism < 0 && // default 1 less than #cores |
3444 |
(parallelism = Runtime.getRuntime().availableProcessors() - 1) <= 0) |
3445 |
parallelism = 1; |
3446 |
if (parallelism > MAX_CAP) |
3447 |
parallelism = MAX_CAP; |
3448 |
return new ForkJoinPool(parallelism, factory, handler, LIFO_QUEUE, |
3449 |
"ForkJoinPool.commonPool-worker-"); |
3450 |
} |
3451 |
|
3452 |
/** |
3453 |
* Factory for innocuous worker threads |
3454 |
*/ |
3455 |
static final class InnocuousForkJoinWorkerThreadFactory |
3456 |
implements ForkJoinWorkerThreadFactory { |
3457 |
|
3458 |
/** |
3459 |
* An ACC to restrict permissions for the factory itself. |
3460 |
* The constructed workers have no permissions set. |
3461 |
*/ |
3462 |
private static final AccessControlContext innocuousAcc; |
3463 |
static { |
3464 |
Permissions innocuousPerms = new Permissions(); |
3465 |
innocuousPerms.add(modifyThreadPermission); |
3466 |
innocuousPerms.add(new RuntimePermission( |
3467 |
"enableContextClassLoaderOverride")); |
3468 |
innocuousPerms.add(new RuntimePermission( |
3469 |
"modifyThreadGroup")); |
3470 |
innocuousAcc = new AccessControlContext(new ProtectionDomain[] { |
3471 |
new ProtectionDomain(null, innocuousPerms) |
3472 |
}); |
3473 |
} |
3474 |
|
3475 |
public final ForkJoinWorkerThread newThread(ForkJoinPool pool) { |
3476 |
return java.security.AccessController.doPrivileged( |
3477 |
new java.security.PrivilegedAction<ForkJoinWorkerThread>() { |
3478 |
public ForkJoinWorkerThread run() { |
3479 |
return new ForkJoinWorkerThread. |
3480 |
InnocuousForkJoinWorkerThread(pool); |
3481 |
}}, innocuousAcc); |
3482 |
} |
3483 |
} |
3484 |
|
3485 |
} |