1 |
dl |
1.1 |
/* |
2 |
|
|
* Written by Doug Lea with assistance from members of JCP JSR-166 |
3 |
|
|
* Expert Group and released to the public domain, as explained at |
4 |
|
|
* http://creativecommons.org/publicdomain/zero/1.0/ |
5 |
|
|
*/ |
6 |
|
|
|
7 |
|
|
package jsr166e; |
8 |
dl |
1.7 |
|
9 |
dl |
1.1 |
import java.util.ArrayList; |
10 |
|
|
import java.util.Arrays; |
11 |
|
|
import java.util.Collection; |
12 |
|
|
import java.util.Collections; |
13 |
|
|
import java.util.List; |
14 |
|
|
import java.util.concurrent.AbstractExecutorService; |
15 |
|
|
import java.util.concurrent.Callable; |
16 |
|
|
import java.util.concurrent.ExecutorService; |
17 |
|
|
import java.util.concurrent.Future; |
18 |
|
|
import java.util.concurrent.RejectedExecutionException; |
19 |
|
|
import java.util.concurrent.RunnableFuture; |
20 |
|
|
import java.util.concurrent.TimeUnit; |
21 |
|
|
|
22 |
|
|
/** |
23 |
|
|
* An {@link ExecutorService} for running {@link ForkJoinTask}s. |
24 |
|
|
* A {@code ForkJoinPool} provides the entry point for submissions |
25 |
|
|
* from non-{@code ForkJoinTask} clients, as well as management and |
26 |
|
|
* monitoring operations. |
27 |
|
|
* |
28 |
|
|
* <p>A {@code ForkJoinPool} differs from other kinds of {@link |
29 |
|
|
* ExecutorService} mainly by virtue of employing |
30 |
|
|
* <em>work-stealing</em>: all threads in the pool attempt to find and |
31 |
|
|
* execute tasks submitted to the pool and/or created by other active |
32 |
|
|
* tasks (eventually blocking waiting for work if none exist). This |
33 |
|
|
* enables efficient processing when most tasks spawn other subtasks |
34 |
|
|
* (as do most {@code ForkJoinTask}s), as well as when many small |
35 |
|
|
* tasks are submitted to the pool from external clients. Especially |
36 |
|
|
* when setting <em>asyncMode</em> to true in constructors, {@code |
37 |
|
|
* ForkJoinPool}s may also be appropriate for use with event-style |
38 |
|
|
* tasks that are never joined. |
39 |
|
|
* |
40 |
dl |
1.18 |
* <p>A static {@link #commonPool()} is available and appropriate for |
41 |
dl |
1.8 |
* most applications. The common pool is used by any ForkJoinTask that |
42 |
|
|
* is not explicitly submitted to a specified pool. Using the common |
43 |
|
|
* pool normally reduces resource usage (its threads are slowly |
44 |
|
|
* reclaimed during periods of non-use, and reinstated upon subsequent |
45 |
dl |
1.12 |
* use). |
46 |
dl |
1.7 |
* |
47 |
|
|
* <p>For applications that require separate or custom pools, a {@code |
48 |
|
|
* ForkJoinPool} may be constructed with a given target parallelism |
49 |
|
|
* level; by default, equal to the number of available processors. The |
50 |
|
|
* pool attempts to maintain enough active (or available) threads by |
51 |
|
|
* dynamically adding, suspending, or resuming internal worker |
52 |
|
|
* threads, even if some tasks are stalled waiting to join |
53 |
|
|
* others. However, no such adjustments are guaranteed in the face of |
54 |
jsr166 |
1.25 |
* blocked I/O or other unmanaged synchronization. The nested {@link |
55 |
dl |
1.7 |
* ManagedBlocker} interface enables extension of the kinds of |
56 |
dl |
1.1 |
* synchronization accommodated. |
57 |
|
|
* |
58 |
|
|
* <p>In addition to execution and lifecycle control methods, this |
59 |
|
|
* class provides status check methods (for example |
60 |
|
|
* {@link #getStealCount}) that are intended to aid in developing, |
61 |
|
|
* tuning, and monitoring fork/join applications. Also, method |
62 |
|
|
* {@link #toString} returns indications of pool state in a |
63 |
|
|
* convenient form for informal monitoring. |
64 |
|
|
* |
65 |
jsr166 |
1.16 |
* <p>As is the case with other ExecutorServices, there are three |
66 |
dl |
1.1 |
* main task execution methods summarized in the following table. |
67 |
|
|
* These are designed to be used primarily by clients not already |
68 |
|
|
* engaged in fork/join computations in the current pool. The main |
69 |
|
|
* forms of these methods accept instances of {@code ForkJoinTask}, |
70 |
|
|
* but overloaded forms also allow mixed execution of plain {@code |
71 |
|
|
* Runnable}- or {@code Callable}- based activities as well. However, |
72 |
|
|
* tasks that are already executing in a pool should normally instead |
73 |
|
|
* use the within-computation forms listed in the table unless using |
74 |
|
|
* async event-style tasks that are not usually joined, in which case |
75 |
|
|
* there is little difference among choice of methods. |
76 |
|
|
* |
77 |
|
|
* <table BORDER CELLPADDING=3 CELLSPACING=1> |
78 |
|
|
* <tr> |
79 |
|
|
* <td></td> |
80 |
|
|
* <td ALIGN=CENTER> <b>Call from non-fork/join clients</b></td> |
81 |
|
|
* <td ALIGN=CENTER> <b>Call from within fork/join computations</b></td> |
82 |
|
|
* </tr> |
83 |
|
|
* <tr> |
84 |
|
|
* <td> <b>Arrange async execution</td> |
85 |
|
|
* <td> {@link #execute(ForkJoinTask)}</td> |
86 |
|
|
* <td> {@link ForkJoinTask#fork}</td> |
87 |
|
|
* </tr> |
88 |
|
|
* <tr> |
89 |
|
|
* <td> <b>Await and obtain result</td> |
90 |
|
|
* <td> {@link #invoke(ForkJoinTask)}</td> |
91 |
|
|
* <td> {@link ForkJoinTask#invoke}</td> |
92 |
|
|
* </tr> |
93 |
|
|
* <tr> |
94 |
|
|
* <td> <b>Arrange exec and obtain Future</td> |
95 |
|
|
* <td> {@link #submit(ForkJoinTask)}</td> |
96 |
|
|
* <td> {@link ForkJoinTask#fork} (ForkJoinTasks <em>are</em> Futures)</td> |
97 |
|
|
* </tr> |
98 |
|
|
* </table> |
99 |
|
|
* |
100 |
dl |
1.12 |
* <p>The common pool is by default constructed with default |
101 |
|
|
* parameters, but these may be controlled by setting three {@link |
102 |
jsr166 |
1.23 |
* System#getProperty system properties} with prefix {@code |
103 |
dl |
1.12 |
* java.util.concurrent.ForkJoinPool.common}: {@code parallelism} -- |
104 |
|
|
* an integer greater than zero, {@code threadFactory} -- the class |
105 |
|
|
* name of a {@link ForkJoinWorkerThreadFactory}, and {@code |
106 |
|
|
* exceptionHandler} -- the class name of a {@link |
107 |
jsr166 |
1.15 |
* java.lang.Thread.UncaughtExceptionHandler |
108 |
dl |
1.12 |
* Thread.UncaughtExceptionHandler}. Upon any error in establishing |
109 |
|
|
* these settings, default parameters are used. |
110 |
|
|
* |
111 |
dl |
1.1 |
* <p><b>Implementation notes</b>: This implementation restricts the |
112 |
|
|
* maximum number of running threads to 32767. Attempts to create |
113 |
|
|
* pools with greater than the maximum number result in |
114 |
|
|
* {@code IllegalArgumentException}. |
115 |
|
|
* |
116 |
|
|
* <p>This implementation rejects submitted tasks (that is, by throwing |
117 |
|
|
* {@link RejectedExecutionException}) only when the pool is shut down |
118 |
|
|
* or internal resources have been exhausted. |
119 |
|
|
* |
120 |
|
|
* @since 1.7 |
121 |
|
|
* @author Doug Lea |
122 |
|
|
*/ |
123 |
|
|
public class ForkJoinPool extends AbstractExecutorService { |
124 |
|
|
|
125 |
|
|
/* |
126 |
|
|
* Implementation Overview |
127 |
|
|
* |
128 |
|
|
* This class and its nested classes provide the main |
129 |
|
|
* functionality and control for a set of worker threads: |
130 |
|
|
* Submissions from non-FJ threads enter into submission queues. |
131 |
|
|
* Workers take these tasks and typically split them into subtasks |
132 |
|
|
* that may be stolen by other workers. Preference rules give |
133 |
|
|
* first priority to processing tasks from their own queues (LIFO |
134 |
|
|
* or FIFO, depending on mode), then to randomized FIFO steals of |
135 |
|
|
* tasks in other queues. |
136 |
|
|
* |
137 |
|
|
* WorkQueues |
138 |
|
|
* ========== |
139 |
|
|
* |
140 |
|
|
* Most operations occur within work-stealing queues (in nested |
141 |
|
|
* class WorkQueue). These are special forms of Deques that |
142 |
|
|
* support only three of the four possible end-operations -- push, |
143 |
|
|
* pop, and poll (aka steal), under the further constraints that |
144 |
|
|
* push and pop are called only from the owning thread (or, as |
145 |
|
|
* extended here, under a lock), while poll may be called from |
146 |
|
|
* other threads. (If you are unfamiliar with them, you probably |
147 |
|
|
* want to read Herlihy and Shavit's book "The Art of |
148 |
|
|
* Multiprocessor programming", chapter 16 describing these in |
149 |
|
|
* more detail before proceeding.) The main work-stealing queue |
150 |
|
|
* design is roughly similar to those in the papers "Dynamic |
151 |
|
|
* Circular Work-Stealing Deque" by Chase and Lev, SPAA 2005 |
152 |
|
|
* (http://research.sun.com/scalable/pubs/index.html) and |
153 |
|
|
* "Idempotent work stealing" by Michael, Saraswat, and Vechev, |
154 |
|
|
* PPoPP 2009 (http://portal.acm.org/citation.cfm?id=1504186). |
155 |
|
|
* The main differences ultimately stem from GC requirements that |
156 |
|
|
* we null out taken slots as soon as we can, to maintain as small |
157 |
|
|
* a footprint as possible even in programs generating huge |
158 |
|
|
* numbers of tasks. To accomplish this, we shift the CAS |
159 |
|
|
* arbitrating pop vs poll (steal) from being on the indices |
160 |
|
|
* ("base" and "top") to the slots themselves. So, both a |
161 |
|
|
* successful pop and poll mainly entail a CAS of a slot from |
162 |
|
|
* non-null to null. Because we rely on CASes of references, we |
163 |
|
|
* do not need tag bits on base or top. They are simple ints as |
164 |
|
|
* used in any circular array-based queue (see for example |
165 |
|
|
* ArrayDeque). Updates to the indices must still be ordered in a |
166 |
|
|
* way that guarantees that top == base means the queue is empty, |
167 |
|
|
* but otherwise may err on the side of possibly making the queue |
168 |
|
|
* appear nonempty when a push, pop, or poll have not fully |
169 |
|
|
* committed. Note that this means that the poll operation, |
170 |
|
|
* considered individually, is not wait-free. One thief cannot |
171 |
|
|
* successfully continue until another in-progress one (or, if |
172 |
|
|
* previously empty, a push) completes. However, in the |
173 |
|
|
* aggregate, we ensure at least probabilistic non-blockingness. |
174 |
|
|
* If an attempted steal fails, a thief always chooses a different |
175 |
|
|
* random victim target to try next. So, in order for one thief to |
176 |
|
|
* progress, it suffices for any in-progress poll or new push on |
177 |
|
|
* any empty queue to complete. (This is why we normally use |
178 |
|
|
* method pollAt and its variants that try once at the apparent |
179 |
|
|
* base index, else consider alternative actions, rather than |
180 |
|
|
* method poll.) |
181 |
|
|
* |
182 |
|
|
* This approach also enables support of a user mode in which local |
183 |
|
|
* task processing is in FIFO, not LIFO order, simply by using |
184 |
|
|
* poll rather than pop. This can be useful in message-passing |
185 |
|
|
* frameworks in which tasks are never joined. However neither |
186 |
|
|
* mode considers affinities, loads, cache localities, etc, so |
187 |
|
|
* rarely provide the best possible performance on a given |
188 |
|
|
* machine, but portably provide good throughput by averaging over |
189 |
|
|
* these factors. (Further, even if we did try to use such |
190 |
|
|
* information, we do not usually have a basis for exploiting it. |
191 |
|
|
* For example, some sets of tasks profit from cache affinities, |
192 |
|
|
* but others are harmed by cache pollution effects.) |
193 |
|
|
* |
194 |
|
|
* WorkQueues are also used in a similar way for tasks submitted |
195 |
|
|
* to the pool. We cannot mix these tasks in the same queues used |
196 |
|
|
* for work-stealing (this would contaminate lifo/fifo |
197 |
dl |
1.12 |
* processing). Instead, we randomly associate submission queues |
198 |
dl |
1.1 |
* with submitting threads, using a form of hashing. The |
199 |
|
|
* ThreadLocal Submitter class contains a value initially used as |
200 |
|
|
* a hash code for choosing existing queues, but may be randomly |
201 |
|
|
* repositioned upon contention with other submitters. In |
202 |
dl |
1.12 |
* essence, submitters act like workers except that they are |
203 |
|
|
* restricted to executing local tasks that they submitted (or in |
204 |
|
|
* the case of CountedCompleters, others with the same root task). |
205 |
|
|
* However, because most shared/external queue operations are more |
206 |
|
|
* expensive than internal, and because, at steady state, external |
207 |
|
|
* submitters will compete for CPU with workers, ForkJoinTask.join |
208 |
|
|
* and related methods disable them from repeatedly helping to |
209 |
|
|
* process tasks if all workers are active. Insertion of tasks in |
210 |
|
|
* shared mode requires a lock (mainly to protect in the case of |
211 |
|
|
* resizing) but we use only a simple spinlock (using bits in |
212 |
|
|
* field qlock), because submitters encountering a busy queue move |
213 |
|
|
* on to try or create other queues -- they block only when |
214 |
|
|
* creating and registering new queues. |
215 |
dl |
1.1 |
* |
216 |
|
|
* Management |
217 |
|
|
* ========== |
218 |
|
|
* |
219 |
|
|
* The main throughput advantages of work-stealing stem from |
220 |
|
|
* decentralized control -- workers mostly take tasks from |
221 |
|
|
* themselves or each other. We cannot negate this in the |
222 |
|
|
* implementation of other management responsibilities. The main |
223 |
|
|
* tactic for avoiding bottlenecks is packing nearly all |
224 |
|
|
* essentially atomic control state into two volatile variables |
225 |
|
|
* that are by far most often read (not written) as status and |
226 |
|
|
* consistency checks. |
227 |
|
|
* |
228 |
|
|
* Field "ctl" contains 64 bits holding all the information needed |
229 |
|
|
* to atomically decide to add, inactivate, enqueue (on an event |
230 |
|
|
* queue), dequeue, and/or re-activate workers. To enable this |
231 |
|
|
* packing, we restrict maximum parallelism to (1<<15)-1 (which is |
232 |
|
|
* far in excess of normal operating range) to allow ids, counts, |
233 |
|
|
* and their negations (used for thresholding) to fit into 16bit |
234 |
|
|
* fields. |
235 |
|
|
* |
236 |
dl |
1.12 |
* Field "plock" is a form of sequence lock with a saturating |
237 |
|
|
* shutdown bit (similarly for per-queue "qlocks"), mainly |
238 |
|
|
* protecting updates to the workQueues array, as well as to |
239 |
|
|
* enable shutdown. When used as a lock, it is normally only very |
240 |
|
|
* briefly held, so is nearly always available after at most a |
241 |
|
|
* brief spin, but we use a monitor-based backup strategy to |
242 |
dl |
1.18 |
* block when needed. |
243 |
dl |
1.1 |
* |
244 |
|
|
* Recording WorkQueues. WorkQueues are recorded in the |
245 |
dl |
1.8 |
* "workQueues" array that is created upon first use and expanded |
246 |
|
|
* if necessary. Updates to the array while recording new workers |
247 |
|
|
* and unrecording terminated ones are protected from each other |
248 |
|
|
* by a lock but the array is otherwise concurrently readable, and |
249 |
|
|
* accessed directly. To simplify index-based operations, the |
250 |
|
|
* array size is always a power of two, and all readers must |
251 |
dl |
1.18 |
* tolerate null slots. Worker queues are at odd indices. Shared |
252 |
dl |
1.12 |
* (submission) queues are at even indices, up to a maximum of 64 |
253 |
|
|
* slots, to limit growth even if array needs to expand to add |
254 |
|
|
* more workers. Grouping them together in this way simplifies and |
255 |
|
|
* speeds up task scanning. |
256 |
dl |
1.1 |
* |
257 |
|
|
* All worker thread creation is on-demand, triggered by task |
258 |
|
|
* submissions, replacement of terminated workers, and/or |
259 |
|
|
* compensation for blocked workers. However, all other support |
260 |
|
|
* code is set up to work with other policies. To ensure that we |
261 |
|
|
* do not hold on to worker references that would prevent GC, ALL |
262 |
|
|
* accesses to workQueues are via indices into the workQueues |
263 |
|
|
* array (which is one source of some of the messy code |
264 |
|
|
* constructions here). In essence, the workQueues array serves as |
265 |
|
|
* a weak reference mechanism. Thus for example the wait queue |
266 |
|
|
* field of ctl stores indices, not references. Access to the |
267 |
|
|
* workQueues in associated methods (for example signalWork) must |
268 |
|
|
* both index-check and null-check the IDs. All such accesses |
269 |
|
|
* ignore bad IDs by returning out early from what they are doing, |
270 |
|
|
* since this can only be associated with termination, in which |
271 |
|
|
* case it is OK to give up. All uses of the workQueues array |
272 |
|
|
* also check that it is non-null (even if previously |
273 |
|
|
* non-null). This allows nulling during termination, which is |
274 |
|
|
* currently not necessary, but remains an option for |
275 |
|
|
* resource-revocation-based shutdown schemes. It also helps |
276 |
|
|
* reduce JIT issuance of uncommon-trap code, which tends to |
277 |
|
|
* unnecessarily complicate control flow in some methods. |
278 |
|
|
* |
279 |
|
|
* Event Queuing. Unlike HPC work-stealing frameworks, we cannot |
280 |
|
|
* let workers spin indefinitely scanning for tasks when none can |
281 |
|
|
* be found immediately, and we cannot start/resume workers unless |
282 |
|
|
* there appear to be tasks available. On the other hand, we must |
283 |
|
|
* quickly prod them into action when new tasks are submitted or |
284 |
|
|
* generated. In many usages, ramp-up time to activate workers is |
285 |
|
|
* the main limiting factor in overall performance (this is |
286 |
|
|
* compounded at program start-up by JIT compilation and |
287 |
|
|
* allocation). So we try to streamline this as much as possible. |
288 |
|
|
* We park/unpark workers after placing in an event wait queue |
289 |
|
|
* when they cannot find work. This "queue" is actually a simple |
290 |
|
|
* Treiber stack, headed by the "id" field of ctl, plus a 15bit |
291 |
|
|
* counter value (that reflects the number of times a worker has |
292 |
|
|
* been inactivated) to avoid ABA effects (we need only as many |
293 |
|
|
* version numbers as worker threads). Successors are held in |
294 |
|
|
* field WorkQueue.nextWait. Queuing deals with several intrinsic |
295 |
|
|
* races, mainly that a task-producing thread can miss seeing (and |
296 |
|
|
* signalling) another thread that gave up looking for work but |
297 |
|
|
* has not yet entered the wait queue. We solve this by requiring |
298 |
|
|
* a full sweep of all workers (via repeated calls to method |
299 |
|
|
* scan()) both before and after a newly waiting worker is added |
300 |
|
|
* to the wait queue. During a rescan, the worker might release |
301 |
|
|
* some other queued worker rather than itself, which has the same |
302 |
|
|
* net effect. Because enqueued workers may actually be rescanning |
303 |
|
|
* rather than waiting, we set and clear the "parker" field of |
304 |
|
|
* WorkQueues to reduce unnecessary calls to unpark. (This |
305 |
|
|
* requires a secondary recheck to avoid missed signals.) Note |
306 |
|
|
* the unusual conventions about Thread.interrupts surrounding |
307 |
|
|
* parking and other blocking: Because interrupts are used solely |
308 |
|
|
* to alert threads to check termination, which is checked anyway |
309 |
|
|
* upon blocking, we clear status (using Thread.interrupted) |
310 |
|
|
* before any call to park, so that park does not immediately |
311 |
|
|
* return due to status being set via some other unrelated call to |
312 |
|
|
* interrupt in user code. |
313 |
|
|
* |
314 |
|
|
* Signalling. We create or wake up workers only when there |
315 |
|
|
* appears to be at least one task they might be able to find and |
316 |
dl |
1.12 |
* execute. However, many other threads may notice the same task |
317 |
|
|
* and each signal to wake up a thread that might take it. So in |
318 |
|
|
* general, pools will be over-signalled. When a submission is |
319 |
dl |
1.21 |
* added or another worker adds a task to a queue that has fewer |
320 |
|
|
* than two tasks, they signal waiting workers (or trigger |
321 |
|
|
* creation of new ones if fewer than the given parallelism level |
322 |
|
|
* -- signalWork), and may leave a hint to the unparked worker to |
323 |
|
|
* help signal others upon wakeup). These primary signals are |
324 |
|
|
* buttressed by others (see method helpSignal) whenever other |
325 |
|
|
* threads scan for work or do not have a task to process. On |
326 |
|
|
* most platforms, signalling (unpark) overhead time is noticeably |
327 |
|
|
* long, and the time between signalling a thread and it actually |
328 |
|
|
* making progress can be very noticeably long, so it is worth |
329 |
|
|
* offloading these delays from critical paths as much as |
330 |
|
|
* possible. |
331 |
dl |
1.1 |
* |
332 |
|
|
* Trimming workers. To release resources after periods of lack of |
333 |
|
|
* use, a worker starting to wait when the pool is quiescent will |
334 |
dl |
1.7 |
* time out and terminate if the pool has remained quiescent for a |
335 |
|
|
* given period -- a short period if there are more threads than |
336 |
|
|
* parallelism, longer as the number of threads decreases. This |
337 |
|
|
* will slowly propagate, eventually terminating all workers after |
338 |
|
|
* periods of non-use. |
339 |
dl |
1.1 |
* |
340 |
|
|
* Shutdown and Termination. A call to shutdownNow atomically sets |
341 |
dl |
1.12 |
* a plock bit and then (non-atomically) sets each worker's |
342 |
|
|
* qlock status, cancels all unprocessed tasks, and wakes up |
343 |
dl |
1.1 |
* all waiting workers. Detecting whether termination should |
344 |
|
|
* commence after a non-abrupt shutdown() call requires more work |
345 |
|
|
* and bookkeeping. We need consensus about quiescence (i.e., that |
346 |
|
|
* there is no more work). The active count provides a primary |
347 |
|
|
* indication but non-abrupt shutdown still requires a rechecking |
348 |
|
|
* scan for any workers that are inactive but not queued. |
349 |
|
|
* |
350 |
|
|
* Joining Tasks |
351 |
|
|
* ============= |
352 |
|
|
* |
353 |
|
|
* Any of several actions may be taken when one worker is waiting |
354 |
|
|
* to join a task stolen (or always held) by another. Because we |
355 |
|
|
* are multiplexing many tasks on to a pool of workers, we can't |
356 |
|
|
* just let them block (as in Thread.join). We also cannot just |
357 |
|
|
* reassign the joiner's run-time stack with another and replace |
358 |
|
|
* it later, which would be a form of "continuation", that even if |
359 |
|
|
* possible is not necessarily a good idea since we sometimes need |
360 |
|
|
* both an unblocked task and its continuation to progress. |
361 |
|
|
* Instead we combine two tactics: |
362 |
|
|
* |
363 |
|
|
* Helping: Arranging for the joiner to execute some task that it |
364 |
|
|
* would be running if the steal had not occurred. |
365 |
|
|
* |
366 |
|
|
* Compensating: Unless there are already enough live threads, |
367 |
|
|
* method tryCompensate() may create or re-activate a spare |
368 |
|
|
* thread to compensate for blocked joiners until they unblock. |
369 |
|
|
* |
370 |
dl |
1.12 |
* A third form (implemented in tryRemoveAndExec) amounts to |
371 |
|
|
* helping a hypothetical compensator: If we can readily tell that |
372 |
|
|
* a possible action of a compensator is to steal and execute the |
373 |
|
|
* task being joined, the joining thread can do so directly, |
374 |
|
|
* without the need for a compensation thread (although at the |
375 |
|
|
* expense of larger run-time stacks, but the tradeoff is |
376 |
|
|
* typically worthwhile). |
377 |
dl |
1.1 |
* |
378 |
|
|
* The ManagedBlocker extension API can't use helping so relies |
379 |
|
|
* only on compensation in method awaitBlocker. |
380 |
|
|
* |
381 |
|
|
* The algorithm in tryHelpStealer entails a form of "linear" |
382 |
|
|
* helping: Each worker records (in field currentSteal) the most |
383 |
|
|
* recent task it stole from some other worker. Plus, it records |
384 |
|
|
* (in field currentJoin) the task it is currently actively |
385 |
|
|
* joining. Method tryHelpStealer uses these markers to try to |
386 |
|
|
* find a worker to help (i.e., steal back a task from and execute |
387 |
|
|
* it) that could hasten completion of the actively joined task. |
388 |
|
|
* In essence, the joiner executes a task that would be on its own |
389 |
|
|
* local deque had the to-be-joined task not been stolen. This may |
390 |
|
|
* be seen as a conservative variant of the approach in Wagner & |
391 |
|
|
* Calder "Leapfrogging: a portable technique for implementing |
392 |
|
|
* efficient futures" SIGPLAN Notices, 1993 |
393 |
|
|
* (http://portal.acm.org/citation.cfm?id=155354). It differs in |
394 |
|
|
* that: (1) We only maintain dependency links across workers upon |
395 |
|
|
* steals, rather than use per-task bookkeeping. This sometimes |
396 |
|
|
* requires a linear scan of workQueues array to locate stealers, |
397 |
|
|
* but often doesn't because stealers leave hints (that may become |
398 |
dl |
1.18 |
* stale/wrong) of where to locate them. It is only a hint |
399 |
|
|
* because a worker might have had multiple steals and the hint |
400 |
|
|
* records only one of them (usually the most current). Hinting |
401 |
|
|
* isolates cost to when it is needed, rather than adding to |
402 |
|
|
* per-task overhead. (2) It is "shallow", ignoring nesting and |
403 |
|
|
* potentially cyclic mutual steals. (3) It is intentionally |
404 |
dl |
1.1 |
* racy: field currentJoin is updated only while actively joining, |
405 |
|
|
* which means that we miss links in the chain during long-lived |
406 |
|
|
* tasks, GC stalls etc (which is OK since blocking in such cases |
407 |
|
|
* is usually a good idea). (4) We bound the number of attempts |
408 |
|
|
* to find work (see MAX_HELP) and fall back to suspending the |
409 |
|
|
* worker and if necessary replacing it with another. |
410 |
|
|
* |
411 |
dl |
1.12 |
* Helping actions for CountedCompleters are much simpler: Method |
412 |
|
|
* helpComplete can take and execute any task with the same root |
413 |
|
|
* as the task being waited on. However, this still entails some |
414 |
|
|
* traversal of completer chains, so is less efficient than using |
415 |
|
|
* CountedCompleters without explicit joins. |
416 |
|
|
* |
417 |
dl |
1.1 |
* It is impossible to keep exactly the target parallelism number |
418 |
|
|
* of threads running at any given time. Determining the |
419 |
|
|
* existence of conservatively safe helping targets, the |
420 |
|
|
* availability of already-created spares, and the apparent need |
421 |
|
|
* to create new spares are all racy, so we rely on multiple |
422 |
|
|
* retries of each. Compensation in the apparent absence of |
423 |
|
|
* helping opportunities is challenging to control on JVMs, where |
424 |
|
|
* GC and other activities can stall progress of tasks that in |
425 |
|
|
* turn stall out many other dependent tasks, without us being |
426 |
|
|
* able to determine whether they will ever require compensation. |
427 |
|
|
* Even though work-stealing otherwise encounters little |
428 |
|
|
* degradation in the presence of more threads than cores, |
429 |
|
|
* aggressively adding new threads in such cases entails risk of |
430 |
|
|
* unwanted positive feedback control loops in which more threads |
431 |
|
|
* cause more dependent stalls (as well as delayed progress of |
432 |
|
|
* unblocked threads to the point that we know they are available) |
433 |
|
|
* leading to more situations requiring more threads, and so |
434 |
|
|
* on. This aspect of control can be seen as an (analytically |
435 |
|
|
* intractable) game with an opponent that may choose the worst |
436 |
|
|
* (for us) active thread to stall at any time. We take several |
437 |
|
|
* precautions to bound losses (and thus bound gains), mainly in |
438 |
dl |
1.12 |
* methods tryCompensate and awaitJoin. |
439 |
|
|
* |
440 |
|
|
* Common Pool |
441 |
|
|
* =========== |
442 |
|
|
* |
443 |
dl |
1.40 |
* The static common Pool always exists after static |
444 |
dl |
1.12 |
* initialization. Since it (or any other created pool) need |
445 |
|
|
* never be used, we minimize initial construction overhead and |
446 |
|
|
* footprint to the setup of about a dozen fields, with no nested |
447 |
|
|
* allocation. Most bootstrapping occurs within method |
448 |
|
|
* fullExternalPush during the first submission to the pool. |
449 |
|
|
* |
450 |
|
|
* When external threads submit to the common pool, they can |
451 |
|
|
* perform some subtask processing (see externalHelpJoin and |
452 |
|
|
* related methods). We do not need to record whether these |
453 |
|
|
* submissions are to the common pool -- if not, externalHelpJoin |
454 |
jsr166 |
1.14 |
* returns quickly (at the most helping to signal some common pool |
455 |
dl |
1.12 |
* workers). These submitters would otherwise be blocked waiting |
456 |
|
|
* for completion, so the extra effort (with liberally sprinkled |
457 |
|
|
* task status checks) in inapplicable cases amounts to an odd |
458 |
|
|
* form of limited spin-wait before blocking in ForkJoinTask.join. |
459 |
|
|
* |
460 |
|
|
* Style notes |
461 |
|
|
* =========== |
462 |
|
|
* |
463 |
|
|
* There is a lot of representation-level coupling among classes |
464 |
|
|
* ForkJoinPool, ForkJoinWorkerThread, and ForkJoinTask. The |
465 |
|
|
* fields of WorkQueue maintain data structures managed by |
466 |
|
|
* ForkJoinPool, so are directly accessed. There is little point |
467 |
|
|
* trying to reduce this, since any associated future changes in |
468 |
|
|
* representations will need to be accompanied by algorithmic |
469 |
|
|
* changes anyway. Several methods intrinsically sprawl because |
470 |
|
|
* they must accumulate sets of consistent reads of volatiles held |
471 |
|
|
* in local variables. Methods signalWork() and scan() are the |
472 |
|
|
* main bottlenecks, so are especially heavily |
473 |
dl |
1.1 |
* micro-optimized/mangled. There are lots of inline assignments |
474 |
|
|
* (of form "while ((local = field) != 0)") which are usually the |
475 |
|
|
* simplest way to ensure the required read orderings (which are |
476 |
|
|
* sometimes critical). This leads to a "C"-like style of listing |
477 |
|
|
* declarations of these locals at the heads of methods or blocks. |
478 |
|
|
* There are several occurrences of the unusual "do {} while |
479 |
|
|
* (!cas...)" which is the simplest way to force an update of a |
480 |
dl |
1.12 |
* CAS'ed variable. There are also other coding oddities (including |
481 |
|
|
* several unnecessary-looking hoisted null checks) that help |
482 |
dl |
1.1 |
* some methods perform reasonably even when interpreted (not |
483 |
|
|
* compiled). |
484 |
|
|
* |
485 |
|
|
* The order of declarations in this file is: |
486 |
|
|
* (1) Static utility functions |
487 |
|
|
* (2) Nested (static) classes |
488 |
|
|
* (3) Static fields |
489 |
|
|
* (4) Fields, along with constants used when unpacking some of them |
490 |
|
|
* (5) Internal control methods |
491 |
|
|
* (6) Callbacks and other support for ForkJoinTask methods |
492 |
|
|
* (7) Exported methods |
493 |
|
|
* (8) Static block initializing statics in minimally dependent order |
494 |
|
|
*/ |
495 |
|
|
|
496 |
|
|
// Static utilities |
497 |
|
|
|
498 |
|
|
/** |
499 |
|
|
* If there is a security manager, makes sure caller has |
500 |
|
|
* permission to modify threads. |
501 |
|
|
*/ |
502 |
|
|
private static void checkPermission() { |
503 |
|
|
SecurityManager security = System.getSecurityManager(); |
504 |
|
|
if (security != null) |
505 |
|
|
security.checkPermission(modifyThreadPermission); |
506 |
|
|
} |
507 |
|
|
|
508 |
|
|
// Nested classes |
509 |
|
|
|
510 |
|
|
/** |
511 |
|
|
* Factory for creating new {@link ForkJoinWorkerThread}s. |
512 |
|
|
* A {@code ForkJoinWorkerThreadFactory} must be defined and used |
513 |
|
|
* for {@code ForkJoinWorkerThread} subclasses that extend base |
514 |
|
|
* functionality or initialize threads with different contexts. |
515 |
|
|
*/ |
516 |
|
|
public static interface ForkJoinWorkerThreadFactory { |
517 |
|
|
/** |
518 |
|
|
* Returns a new worker thread operating in the given pool. |
519 |
|
|
* |
520 |
|
|
* @param pool the pool this thread works in |
521 |
|
|
* @throws NullPointerException if the pool is null |
522 |
|
|
*/ |
523 |
|
|
public ForkJoinWorkerThread newThread(ForkJoinPool pool); |
524 |
|
|
} |
525 |
|
|
|
526 |
|
|
/** |
527 |
|
|
* Default ForkJoinWorkerThreadFactory implementation; creates a |
528 |
|
|
* new ForkJoinWorkerThread. |
529 |
|
|
*/ |
530 |
dl |
1.18 |
static final class DefaultForkJoinWorkerThreadFactory |
531 |
dl |
1.1 |
implements ForkJoinWorkerThreadFactory { |
532 |
dl |
1.18 |
public final ForkJoinWorkerThread newThread(ForkJoinPool pool) { |
533 |
dl |
1.1 |
return new ForkJoinWorkerThread(pool); |
534 |
|
|
} |
535 |
|
|
} |
536 |
|
|
|
537 |
|
|
/** |
538 |
dl |
1.21 |
* Per-thread records for threads that submit to pools. Currently |
539 |
|
|
* holds only pseudo-random seed / index that is used to choose |
540 |
|
|
* submission queues in method externalPush. In the future, this may |
541 |
|
|
* also incorporate a means to implement different task rejection |
542 |
|
|
* and resubmission policies. |
543 |
|
|
* |
544 |
|
|
* Seeds for submitters and workers/workQueues work in basically |
545 |
|
|
* the same way but are initialized and updated using slightly |
546 |
|
|
* different mechanics. Both are initialized using the same |
547 |
|
|
* approach as in class ThreadLocal, where successive values are |
548 |
|
|
* unlikely to collide with previous values. Seeds are then |
549 |
|
|
* randomly modified upon collisions using xorshifts, which |
550 |
|
|
* requires a non-zero seed. |
551 |
|
|
*/ |
552 |
|
|
static final class Submitter { |
553 |
|
|
int seed; |
554 |
|
|
Submitter(int s) { seed = s; } |
555 |
|
|
} |
556 |
|
|
|
557 |
|
|
/** |
558 |
dl |
1.1 |
* Class for artificial tasks that are used to replace the target |
559 |
|
|
* of local joins if they are removed from an interior queue slot |
560 |
|
|
* in WorkQueue.tryRemoveAndExec. We don't need the proxy to |
561 |
|
|
* actually do anything beyond having a unique identity. |
562 |
|
|
*/ |
563 |
|
|
static final class EmptyTask extends ForkJoinTask<Void> { |
564 |
dl |
1.12 |
private static final long serialVersionUID = -7721805057305804111L; |
565 |
dl |
1.1 |
EmptyTask() { status = ForkJoinTask.NORMAL; } // force done |
566 |
|
|
public final Void getRawResult() { return null; } |
567 |
|
|
public final void setRawResult(Void x) {} |
568 |
|
|
public final boolean exec() { return true; } |
569 |
|
|
} |
570 |
|
|
|
571 |
|
|
/** |
572 |
|
|
* Queues supporting work-stealing as well as external task |
573 |
|
|
* submission. See above for main rationale and algorithms. |
574 |
|
|
* Implementation relies heavily on "Unsafe" intrinsics |
575 |
|
|
* and selective use of "volatile": |
576 |
|
|
* |
577 |
|
|
* Field "base" is the index (mod array.length) of the least valid |
578 |
|
|
* queue slot, which is always the next position to steal (poll) |
579 |
|
|
* from if nonempty. Reads and writes require volatile orderings |
580 |
|
|
* but not CAS, because updates are only performed after slot |
581 |
|
|
* CASes. |
582 |
|
|
* |
583 |
|
|
* Field "top" is the index (mod array.length) of the next queue |
584 |
|
|
* slot to push to or pop from. It is written only by owner thread |
585 |
dl |
1.12 |
* for push, or under lock for external/shared push, and accessed |
586 |
|
|
* by other threads only after reading (volatile) base. Both top |
587 |
|
|
* and base are allowed to wrap around on overflow, but (top - |
588 |
|
|
* base) (or more commonly -(base - top) to force volatile read of |
589 |
|
|
* base before top) still estimates size. The lock ("qlock") is |
590 |
|
|
* forced to -1 on termination, causing all further lock attempts |
591 |
|
|
* to fail. (Note: we don't need CAS for termination state because |
592 |
|
|
* upon pool shutdown, all shared-queues will stop being used |
593 |
|
|
* anyway.) Nearly all lock bodies are set up so that exceptions |
594 |
|
|
* within lock bodies are "impossible" (modulo JVM errors that |
595 |
|
|
* would cause failure anyway.) |
596 |
dl |
1.1 |
* |
597 |
|
|
* The array slots are read and written using the emulation of |
598 |
|
|
* volatiles/atomics provided by Unsafe. Insertions must in |
599 |
|
|
* general use putOrderedObject as a form of releasing store to |
600 |
|
|
* ensure that all writes to the task object are ordered before |
601 |
dl |
1.12 |
* its publication in the queue. All removals entail a CAS to |
602 |
|
|
* null. The array is always a power of two. To ensure safety of |
603 |
|
|
* Unsafe array operations, all accesses perform explicit null |
604 |
|
|
* checks and implicit bounds checks via power-of-two masking. |
605 |
dl |
1.1 |
* |
606 |
|
|
* In addition to basic queuing support, this class contains |
607 |
|
|
* fields described elsewhere to control execution. It turns out |
608 |
dl |
1.12 |
* to work better memory-layout-wise to include them in this class |
609 |
|
|
* rather than a separate class. |
610 |
dl |
1.1 |
* |
611 |
|
|
* Performance on most platforms is very sensitive to placement of |
612 |
|
|
* instances of both WorkQueues and their arrays -- we absolutely |
613 |
|
|
* do not want multiple WorkQueue instances or multiple queue |
614 |
|
|
* arrays sharing cache lines. (It would be best for queue objects |
615 |
|
|
* and their arrays to share, but there is nothing available to |
616 |
|
|
* help arrange that). Unfortunately, because they are recorded |
617 |
|
|
* in a common array, WorkQueue instances are often moved to be |
618 |
|
|
* adjacent by garbage collectors. To reduce impact, we use field |
619 |
|
|
* padding that works OK on common platforms; this effectively |
620 |
|
|
* trades off slightly slower average field access for the sake of |
621 |
|
|
* avoiding really bad worst-case access. (Until better JVM |
622 |
|
|
* support is in place, this padding is dependent on transient |
623 |
dl |
1.21 |
* properties of JVM field layout rules.) We also take care in |
624 |
|
|
* allocating, sizing and resizing the array. Non-shared queue |
625 |
|
|
* arrays are initialized by workers before use. Others are |
626 |
|
|
* allocated on first use. |
627 |
dl |
1.1 |
*/ |
628 |
|
|
static final class WorkQueue { |
629 |
|
|
/** |
630 |
|
|
* Capacity of work-stealing queue array upon initialization. |
631 |
|
|
* Must be a power of two; at least 4, but should be larger to |
632 |
|
|
* reduce or eliminate cacheline sharing among queues. |
633 |
|
|
* Currently, it is much larger, as a partial workaround for |
634 |
|
|
* the fact that JVMs often place arrays in locations that |
635 |
|
|
* share GC bookkeeping (especially cardmarks) such that |
636 |
|
|
* per-write accesses encounter serious memory contention. |
637 |
|
|
*/ |
638 |
|
|
static final int INITIAL_QUEUE_CAPACITY = 1 << 13; |
639 |
|
|
|
640 |
|
|
/** |
641 |
|
|
* Maximum size for queue arrays. Must be a power of two less |
642 |
|
|
* than or equal to 1 << (31 - width of array entry) to ensure |
643 |
|
|
* lack of wraparound of index calculations, but defined to a |
644 |
|
|
* value a bit less than this to help users trap runaway |
645 |
|
|
* programs before saturating systems. |
646 |
|
|
*/ |
647 |
|
|
static final int MAXIMUM_QUEUE_CAPACITY = 1 << 26; // 64M |
648 |
|
|
|
649 |
dl |
1.21 |
// Heuristic padding to ameliorate unfortunate memory placements |
650 |
|
|
volatile long pad00, pad01, pad02, pad03, pad04, pad05, pad06; |
651 |
|
|
|
652 |
dl |
1.1 |
int seed; // for random scanning; initialize nonzero |
653 |
|
|
volatile int eventCount; // encoded inactivation count; < 0 if inactive |
654 |
|
|
int nextWait; // encoded record of next event waiter |
655 |
dl |
1.18 |
int hint; // steal or signal hint (index) |
656 |
dl |
1.1 |
int poolIndex; // index of this queue in pool (or 0) |
657 |
dl |
1.18 |
final int mode; // 0: lifo, > 0: fifo, < 0: shared |
658 |
|
|
int nsteals; // number of steals |
659 |
dl |
1.12 |
volatile int qlock; // 1: locked, -1: terminate; else 0 |
660 |
dl |
1.1 |
volatile int base; // index of next slot for poll |
661 |
|
|
int top; // index of next slot for push |
662 |
|
|
ForkJoinTask<?>[] array; // the elements (initially unallocated) |
663 |
|
|
final ForkJoinPool pool; // the containing pool (may be null) |
664 |
|
|
final ForkJoinWorkerThread owner; // owning thread or null if shared |
665 |
|
|
volatile Thread parker; // == owner during call to park; else null |
666 |
|
|
volatile ForkJoinTask<?> currentJoin; // task being joined in awaitJoin |
667 |
|
|
ForkJoinTask<?> currentSteal; // current non-local task being executed |
668 |
dl |
1.18 |
|
669 |
dl |
1.21 |
volatile Object pad10, pad11, pad12, pad13, pad14, pad15, pad16, pad17; |
670 |
|
|
volatile Object pad18, pad19, pad1a, pad1b, pad1c, pad1d; |
671 |
dl |
1.1 |
|
672 |
dl |
1.18 |
WorkQueue(ForkJoinPool pool, ForkJoinWorkerThread owner, int mode, |
673 |
|
|
int seed) { |
674 |
dl |
1.1 |
this.pool = pool; |
675 |
|
|
this.owner = owner; |
676 |
dl |
1.18 |
this.mode = mode; |
677 |
|
|
this.seed = seed; |
678 |
dl |
1.21 |
// Place indices in the center of array (that is not yet allocated) |
679 |
dl |
1.1 |
base = top = INITIAL_QUEUE_CAPACITY >>> 1; |
680 |
|
|
} |
681 |
|
|
|
682 |
|
|
/** |
683 |
dl |
1.21 |
* Returns the approximate number of tasks in the queue. |
684 |
|
|
*/ |
685 |
|
|
final int queueSize() { |
686 |
|
|
int n = base - top; // non-owner callers must read base first |
687 |
|
|
return (n >= 0) ? 0 : -n; // ignore transient negative |
688 |
|
|
} |
689 |
|
|
|
690 |
|
|
/** |
691 |
|
|
* Provides a more accurate estimate of whether this queue has |
692 |
|
|
* any tasks than does queueSize, by checking whether a |
693 |
|
|
* near-empty queue has at least one unclaimed task. |
694 |
|
|
*/ |
695 |
|
|
final boolean isEmpty() { |
696 |
|
|
ForkJoinTask<?>[] a; int m, s; |
697 |
|
|
int n = base - (s = top); |
698 |
|
|
return (n >= 0 || |
699 |
|
|
(n == -1 && |
700 |
|
|
((a = array) == null || |
701 |
|
|
(m = a.length - 1) < 0 || |
702 |
|
|
U.getObject |
703 |
|
|
(a, (long)((m & (s - 1)) << ASHIFT) + ABASE) == null))); |
704 |
|
|
} |
705 |
|
|
|
706 |
|
|
/** |
707 |
|
|
* Pushes a task. Call only by owner in unshared queues. (The |
708 |
|
|
* shared-queue version is embedded in method externalPush.) |
709 |
dl |
1.1 |
* |
710 |
|
|
* @param task the task. Caller must ensure non-null. |
711 |
|
|
* @throw RejectedExecutionException if array cannot be resized |
712 |
|
|
*/ |
713 |
|
|
final void push(ForkJoinTask<?> task) { |
714 |
dl |
1.18 |
ForkJoinTask<?>[] a; ForkJoinPool p; |
715 |
|
|
int s = top, m, n; |
716 |
|
|
if ((a = array) != null) { // ignore if queue removed |
717 |
dl |
1.21 |
int j = (((m = a.length - 1) & s) << ASHIFT) + ABASE; |
718 |
|
|
U.putOrderedObject(a, j, task); |
719 |
|
|
if ((n = (top = s + 1) - base) <= 2) { |
720 |
dl |
1.18 |
if ((p = pool) != null) |
721 |
dl |
1.21 |
p.signalWork(this); |
722 |
dl |
1.18 |
} |
723 |
|
|
else if (n >= m) |
724 |
|
|
growArray(); |
725 |
dl |
1.1 |
} |
726 |
|
|
} |
727 |
|
|
|
728 |
dl |
1.18 |
/** |
729 |
|
|
* Initializes or doubles the capacity of array. Call either |
730 |
|
|
* by owner or with lock held -- it is OK for base, but not |
731 |
|
|
* top, to move while resizings are in progress. |
732 |
|
|
*/ |
733 |
|
|
final ForkJoinTask<?>[] growArray() { |
734 |
|
|
ForkJoinTask<?>[] oldA = array; |
735 |
|
|
int size = oldA != null ? oldA.length << 1 : INITIAL_QUEUE_CAPACITY; |
736 |
|
|
if (size > MAXIMUM_QUEUE_CAPACITY) |
737 |
|
|
throw new RejectedExecutionException("Queue capacity exceeded"); |
738 |
|
|
int oldMask, t, b; |
739 |
|
|
ForkJoinTask<?>[] a = array = new ForkJoinTask<?>[size]; |
740 |
|
|
if (oldA != null && (oldMask = oldA.length - 1) >= 0 && |
741 |
|
|
(t = top) - (b = base) > 0) { |
742 |
|
|
int mask = size - 1; |
743 |
|
|
do { |
744 |
|
|
ForkJoinTask<?> x; |
745 |
|
|
int oldj = ((b & oldMask) << ASHIFT) + ABASE; |
746 |
|
|
int j = ((b & mask) << ASHIFT) + ABASE; |
747 |
|
|
x = (ForkJoinTask<?>)U.getObjectVolatile(oldA, oldj); |
748 |
|
|
if (x != null && |
749 |
|
|
U.compareAndSwapObject(oldA, oldj, x, null)) |
750 |
|
|
U.putObjectVolatile(a, j, x); |
751 |
|
|
} while (++b != t); |
752 |
dl |
1.1 |
} |
753 |
dl |
1.18 |
return a; |
754 |
dl |
1.1 |
} |
755 |
|
|
|
756 |
|
|
/** |
757 |
|
|
* Takes next task, if one exists, in LIFO order. Call only |
758 |
dl |
1.9 |
* by owner in unshared queues. |
759 |
dl |
1.1 |
*/ |
760 |
|
|
final ForkJoinTask<?> pop() { |
761 |
|
|
ForkJoinTask<?>[] a; ForkJoinTask<?> t; int m; |
762 |
|
|
if ((a = array) != null && (m = a.length - 1) >= 0) { |
763 |
|
|
for (int s; (s = top - 1) - base >= 0;) { |
764 |
|
|
long j = ((m & s) << ASHIFT) + ABASE; |
765 |
|
|
if ((t = (ForkJoinTask<?>)U.getObject(a, j)) == null) |
766 |
|
|
break; |
767 |
|
|
if (U.compareAndSwapObject(a, j, t, null)) { |
768 |
|
|
top = s; |
769 |
|
|
return t; |
770 |
|
|
} |
771 |
|
|
} |
772 |
|
|
} |
773 |
|
|
return null; |
774 |
|
|
} |
775 |
|
|
|
776 |
|
|
/** |
777 |
|
|
* Takes a task in FIFO order if b is base of queue and a task |
778 |
|
|
* can be claimed without contention. Specialized versions |
779 |
|
|
* appear in ForkJoinPool methods scan and tryHelpStealer. |
780 |
|
|
*/ |
781 |
|
|
final ForkJoinTask<?> pollAt(int b) { |
782 |
|
|
ForkJoinTask<?> t; ForkJoinTask<?>[] a; |
783 |
|
|
if ((a = array) != null) { |
784 |
|
|
int j = (((a.length - 1) & b) << ASHIFT) + ABASE; |
785 |
|
|
if ((t = (ForkJoinTask<?>)U.getObjectVolatile(a, j)) != null && |
786 |
|
|
base == b && |
787 |
|
|
U.compareAndSwapObject(a, j, t, null)) { |
788 |
|
|
base = b + 1; |
789 |
|
|
return t; |
790 |
|
|
} |
791 |
|
|
} |
792 |
|
|
return null; |
793 |
|
|
} |
794 |
|
|
|
795 |
|
|
/** |
796 |
|
|
* Takes next task, if one exists, in FIFO order. |
797 |
|
|
*/ |
798 |
|
|
final ForkJoinTask<?> poll() { |
799 |
|
|
ForkJoinTask<?>[] a; int b; ForkJoinTask<?> t; |
800 |
|
|
while ((b = base) - top < 0 && (a = array) != null) { |
801 |
|
|
int j = (((a.length - 1) & b) << ASHIFT) + ABASE; |
802 |
|
|
t = (ForkJoinTask<?>)U.getObjectVolatile(a, j); |
803 |
|
|
if (t != null) { |
804 |
|
|
if (base == b && |
805 |
|
|
U.compareAndSwapObject(a, j, t, null)) { |
806 |
|
|
base = b + 1; |
807 |
|
|
return t; |
808 |
|
|
} |
809 |
|
|
} |
810 |
|
|
else if (base == b) { |
811 |
|
|
if (b + 1 == top) |
812 |
|
|
break; |
813 |
dl |
1.12 |
Thread.yield(); // wait for lagging update (very rare) |
814 |
dl |
1.1 |
} |
815 |
|
|
} |
816 |
|
|
return null; |
817 |
|
|
} |
818 |
|
|
|
819 |
|
|
/** |
820 |
|
|
* Takes next task, if one exists, in order specified by mode. |
821 |
|
|
*/ |
822 |
|
|
final ForkJoinTask<?> nextLocalTask() { |
823 |
|
|
return mode == 0 ? pop() : poll(); |
824 |
|
|
} |
825 |
|
|
|
826 |
|
|
/** |
827 |
|
|
* Returns next task, if one exists, in order specified by mode. |
828 |
|
|
*/ |
829 |
|
|
final ForkJoinTask<?> peek() { |
830 |
|
|
ForkJoinTask<?>[] a = array; int m; |
831 |
|
|
if (a == null || (m = a.length - 1) < 0) |
832 |
|
|
return null; |
833 |
|
|
int i = mode == 0 ? top - 1 : base; |
834 |
|
|
int j = ((i & m) << ASHIFT) + ABASE; |
835 |
|
|
return (ForkJoinTask<?>)U.getObjectVolatile(a, j); |
836 |
|
|
} |
837 |
|
|
|
838 |
|
|
/** |
839 |
|
|
* Pops the given task only if it is at the current top. |
840 |
dl |
1.12 |
* (A shared version is available only via FJP.tryExternalUnpush) |
841 |
dl |
1.1 |
*/ |
842 |
|
|
final boolean tryUnpush(ForkJoinTask<?> t) { |
843 |
|
|
ForkJoinTask<?>[] a; int s; |
844 |
|
|
if ((a = array) != null && (s = top) != base && |
845 |
|
|
U.compareAndSwapObject |
846 |
|
|
(a, (((a.length - 1) & --s) << ASHIFT) + ABASE, t, null)) { |
847 |
|
|
top = s; |
848 |
|
|
return true; |
849 |
|
|
} |
850 |
|
|
return false; |
851 |
|
|
} |
852 |
|
|
|
853 |
|
|
/** |
854 |
|
|
* Removes and cancels all known tasks, ignoring any exceptions. |
855 |
|
|
*/ |
856 |
|
|
final void cancelAll() { |
857 |
|
|
ForkJoinTask.cancelIgnoringExceptions(currentJoin); |
858 |
|
|
ForkJoinTask.cancelIgnoringExceptions(currentSteal); |
859 |
|
|
for (ForkJoinTask<?> t; (t = poll()) != null; ) |
860 |
|
|
ForkJoinTask.cancelIgnoringExceptions(t); |
861 |
|
|
} |
862 |
|
|
|
863 |
|
|
/** |
864 |
|
|
* Computes next value for random probes. Scans don't require |
865 |
|
|
* a very high quality generator, but also not a crummy one. |
866 |
|
|
* Marsaglia xor-shift is cheap and works well enough. Note: |
867 |
|
|
* This is manually inlined in its usages in ForkJoinPool to |
868 |
|
|
* avoid writes inside busy scan loops. |
869 |
|
|
*/ |
870 |
|
|
final int nextSeed() { |
871 |
|
|
int r = seed; |
872 |
|
|
r ^= r << 13; |
873 |
|
|
r ^= r >>> 17; |
874 |
|
|
return seed = r ^= r << 5; |
875 |
|
|
} |
876 |
|
|
|
877 |
dl |
1.11 |
// Specialized execution methods |
878 |
dl |
1.1 |
|
879 |
|
|
/** |
880 |
|
|
* Pops and runs tasks until empty. |
881 |
|
|
*/ |
882 |
|
|
private void popAndExecAll() { |
883 |
|
|
// A bit faster than repeated pop calls |
884 |
|
|
ForkJoinTask<?>[] a; int m, s; long j; ForkJoinTask<?> t; |
885 |
|
|
while ((a = array) != null && (m = a.length - 1) >= 0 && |
886 |
|
|
(s = top - 1) - base >= 0 && |
887 |
|
|
(t = ((ForkJoinTask<?>) |
888 |
|
|
U.getObject(a, j = ((m & s) << ASHIFT) + ABASE))) |
889 |
|
|
!= null) { |
890 |
|
|
if (U.compareAndSwapObject(a, j, t, null)) { |
891 |
|
|
top = s; |
892 |
|
|
t.doExec(); |
893 |
|
|
} |
894 |
|
|
} |
895 |
|
|
} |
896 |
|
|
|
897 |
|
|
/** |
898 |
|
|
* Polls and runs tasks until empty. |
899 |
|
|
*/ |
900 |
|
|
private void pollAndExecAll() { |
901 |
|
|
for (ForkJoinTask<?> t; (t = poll()) != null;) |
902 |
|
|
t.doExec(); |
903 |
|
|
} |
904 |
|
|
|
905 |
|
|
/** |
906 |
dl |
1.12 |
* If present, removes from queue and executes the given task, |
907 |
|
|
* or any other cancelled task. Returns (true) on any CAS |
908 |
dl |
1.1 |
* or consistency check failure so caller can retry. |
909 |
|
|
* |
910 |
dl |
1.12 |
* @return false if no progress can be made, else true; |
911 |
dl |
1.1 |
*/ |
912 |
dl |
1.12 |
final boolean tryRemoveAndExec(ForkJoinTask<?> task) { |
913 |
|
|
boolean stat = true, removed = false, empty = true; |
914 |
dl |
1.1 |
ForkJoinTask<?>[] a; int m, s, b, n; |
915 |
|
|
if ((a = array) != null && (m = a.length - 1) >= 0 && |
916 |
|
|
(n = (s = top) - (b = base)) > 0) { |
917 |
|
|
for (ForkJoinTask<?> t;;) { // traverse from s to b |
918 |
|
|
int j = ((--s & m) << ASHIFT) + ABASE; |
919 |
|
|
t = (ForkJoinTask<?>)U.getObjectVolatile(a, j); |
920 |
|
|
if (t == null) // inconsistent length |
921 |
|
|
break; |
922 |
|
|
else if (t == task) { |
923 |
|
|
if (s + 1 == top) { // pop |
924 |
|
|
if (!U.compareAndSwapObject(a, j, task, null)) |
925 |
|
|
break; |
926 |
|
|
top = s; |
927 |
|
|
removed = true; |
928 |
|
|
} |
929 |
|
|
else if (base == b) // replace with proxy |
930 |
|
|
removed = U.compareAndSwapObject(a, j, task, |
931 |
|
|
new EmptyTask()); |
932 |
|
|
break; |
933 |
|
|
} |
934 |
|
|
else if (t.status >= 0) |
935 |
|
|
empty = false; |
936 |
|
|
else if (s + 1 == top) { // pop and throw away |
937 |
|
|
if (U.compareAndSwapObject(a, j, t, null)) |
938 |
|
|
top = s; |
939 |
|
|
break; |
940 |
|
|
} |
941 |
|
|
if (--n == 0) { |
942 |
|
|
if (!empty && base == b) |
943 |
dl |
1.12 |
stat = false; |
944 |
dl |
1.1 |
break; |
945 |
|
|
} |
946 |
|
|
} |
947 |
|
|
} |
948 |
|
|
if (removed) |
949 |
|
|
task.doExec(); |
950 |
|
|
return stat; |
951 |
|
|
} |
952 |
|
|
|
953 |
|
|
/** |
954 |
dl |
1.12 |
* Polls for and executes the given task or any other task in |
955 |
|
|
* its CountedCompleter computation |
956 |
dl |
1.11 |
*/ |
957 |
dl |
1.12 |
final boolean pollAndExecCC(ForkJoinTask<?> root) { |
958 |
|
|
ForkJoinTask<?>[] a; int b; Object o; |
959 |
|
|
outer: while ((b = base) - top < 0 && (a = array) != null) { |
960 |
|
|
long j = (((a.length - 1) & b) << ASHIFT) + ABASE; |
961 |
|
|
if ((o = U.getObject(a, j)) == null || |
962 |
|
|
!(o instanceof CountedCompleter)) |
963 |
|
|
break; |
964 |
|
|
for (CountedCompleter<?> t = (CountedCompleter<?>)o, r = t;;) { |
965 |
|
|
if (r == root) { |
966 |
|
|
if (base == b && |
967 |
|
|
U.compareAndSwapObject(a, j, t, null)) { |
968 |
|
|
base = b + 1; |
969 |
|
|
t.doExec(); |
970 |
|
|
return true; |
971 |
dl |
1.11 |
} |
972 |
dl |
1.12 |
else |
973 |
|
|
break; // restart |
974 |
dl |
1.11 |
} |
975 |
dl |
1.12 |
if ((r = r.completer) == null) |
976 |
|
|
break outer; // not part of root computation |
977 |
dl |
1.11 |
} |
978 |
|
|
} |
979 |
dl |
1.12 |
return false; |
980 |
dl |
1.11 |
} |
981 |
|
|
|
982 |
|
|
/** |
983 |
dl |
1.1 |
* Executes a top-level task and any local tasks remaining |
984 |
|
|
* after execution. |
985 |
|
|
*/ |
986 |
|
|
final void runTask(ForkJoinTask<?> t) { |
987 |
|
|
if (t != null) { |
988 |
dl |
1.12 |
(currentSteal = t).doExec(); |
989 |
|
|
currentSteal = null; |
990 |
dl |
1.32 |
++nsteals; |
991 |
dl |
1.21 |
if (base - top < 0) { // process remaining local tasks |
992 |
dl |
1.1 |
if (mode == 0) |
993 |
|
|
popAndExecAll(); |
994 |
|
|
else |
995 |
|
|
pollAndExecAll(); |
996 |
|
|
} |
997 |
|
|
} |
998 |
|
|
} |
999 |
|
|
|
1000 |
|
|
/** |
1001 |
|
|
* Executes a non-top-level (stolen) task. |
1002 |
|
|
*/ |
1003 |
|
|
final void runSubtask(ForkJoinTask<?> t) { |
1004 |
|
|
if (t != null) { |
1005 |
|
|
ForkJoinTask<?> ps = currentSteal; |
1006 |
dl |
1.12 |
(currentSteal = t).doExec(); |
1007 |
dl |
1.1 |
currentSteal = ps; |
1008 |
|
|
} |
1009 |
|
|
} |
1010 |
|
|
|
1011 |
|
|
/** |
1012 |
|
|
* Returns true if owned and not known to be blocked. |
1013 |
|
|
*/ |
1014 |
|
|
final boolean isApparentlyUnblocked() { |
1015 |
|
|
Thread wt; Thread.State s; |
1016 |
|
|
return (eventCount >= 0 && |
1017 |
|
|
(wt = owner) != null && |
1018 |
|
|
(s = wt.getState()) != Thread.State.BLOCKED && |
1019 |
|
|
s != Thread.State.WAITING && |
1020 |
|
|
s != Thread.State.TIMED_WAITING); |
1021 |
|
|
} |
1022 |
|
|
|
1023 |
|
|
// Unsafe mechanics |
1024 |
|
|
private static final sun.misc.Unsafe U; |
1025 |
dl |
1.12 |
private static final long QLOCK; |
1026 |
dl |
1.1 |
private static final int ABASE; |
1027 |
|
|
private static final int ASHIFT; |
1028 |
|
|
static { |
1029 |
|
|
int s; |
1030 |
|
|
try { |
1031 |
|
|
U = getUnsafe(); |
1032 |
|
|
Class<?> k = WorkQueue.class; |
1033 |
|
|
Class<?> ak = ForkJoinTask[].class; |
1034 |
dl |
1.12 |
QLOCK = U.objectFieldOffset |
1035 |
|
|
(k.getDeclaredField("qlock")); |
1036 |
dl |
1.1 |
ABASE = U.arrayBaseOffset(ak); |
1037 |
|
|
s = U.arrayIndexScale(ak); |
1038 |
|
|
} catch (Exception e) { |
1039 |
|
|
throw new Error(e); |
1040 |
|
|
} |
1041 |
|
|
if ((s & (s-1)) != 0) |
1042 |
|
|
throw new Error("data type scale not a power of two"); |
1043 |
|
|
ASHIFT = 31 - Integer.numberOfLeadingZeros(s); |
1044 |
|
|
} |
1045 |
|
|
} |
1046 |
jsr166 |
1.3 |
|
1047 |
dl |
1.18 |
// static fields (initialized in static initializer below) |
1048 |
|
|
|
1049 |
|
|
/** |
1050 |
|
|
* Creates a new ForkJoinWorkerThread. This factory is used unless |
1051 |
|
|
* overridden in ForkJoinPool constructors. |
1052 |
|
|
*/ |
1053 |
|
|
public static final ForkJoinWorkerThreadFactory |
1054 |
|
|
defaultForkJoinWorkerThreadFactory; |
1055 |
|
|
|
1056 |
dl |
1.1 |
/** |
1057 |
dl |
1.18 |
* Per-thread submission bookkeeping. Shared across all pools |
1058 |
|
|
* to reduce ThreadLocal pollution and because random motion |
1059 |
|
|
* to avoid contention in one pool is likely to hold for others. |
1060 |
|
|
* Lazily initialized on first submission (but null-checked |
1061 |
|
|
* in other contexts to avoid unnecessary initialization). |
1062 |
dl |
1.1 |
*/ |
1063 |
dl |
1.18 |
static final ThreadLocal<Submitter> submitters; |
1064 |
dl |
1.1 |
|
1065 |
dl |
1.8 |
/** |
1066 |
dl |
1.21 |
* Permission required for callers of methods that may start or |
1067 |
|
|
* kill threads. |
1068 |
|
|
*/ |
1069 |
|
|
private static final RuntimePermission modifyThreadPermission; |
1070 |
|
|
|
1071 |
|
|
/** |
1072 |
dl |
1.8 |
* Common (static) pool. Non-null for public use unless a static |
1073 |
dl |
1.12 |
* construction exception, but internal usages null-check on use |
1074 |
|
|
* to paranoically avoid potential initialization circularities |
1075 |
|
|
* as well as to simplify generated code. |
1076 |
dl |
1.8 |
*/ |
1077 |
dl |
1.40 |
static final ForkJoinPool common; |
1078 |
dl |
1.8 |
|
1079 |
|
|
/** |
1080 |
dl |
1.40 |
* Common pool parallelism. Must equal common.parallelism. |
1081 |
dl |
1.1 |
*/ |
1082 |
dl |
1.40 |
static final int commonParallelism; |
1083 |
dl |
1.1 |
|
1084 |
|
|
/** |
1085 |
dl |
1.12 |
* Sequence number for creating workerNamePrefix. |
1086 |
dl |
1.1 |
*/ |
1087 |
dl |
1.12 |
private static int poolNumberSequence; |
1088 |
dl |
1.1 |
|
1089 |
|
|
/** |
1090 |
jsr166 |
1.38 |
* Returns the next sequence number. We don't expect this to |
1091 |
|
|
* ever contend, so use simple builtin sync. |
1092 |
dl |
1.1 |
*/ |
1093 |
dl |
1.12 |
private static final synchronized int nextPoolId() { |
1094 |
|
|
return ++poolNumberSequence; |
1095 |
|
|
} |
1096 |
dl |
1.1 |
|
1097 |
|
|
// static constants |
1098 |
|
|
|
1099 |
|
|
/** |
1100 |
dl |
1.12 |
* Initial timeout value (in nanoseconds) for the thread |
1101 |
|
|
* triggering quiescence to park waiting for new work. On timeout, |
1102 |
|
|
* the thread will instead try to shrink the number of |
1103 |
|
|
* workers. The value should be large enough to avoid overly |
1104 |
|
|
* aggressive shrinkage during most transient stalls (long GCs |
1105 |
|
|
* etc). |
1106 |
dl |
1.1 |
*/ |
1107 |
dl |
1.12 |
private static final long IDLE_TIMEOUT = 2000L * 1000L * 1000L; // 2sec |
1108 |
dl |
1.1 |
|
1109 |
|
|
/** |
1110 |
dl |
1.7 |
* Timeout value when there are more threads than parallelism level |
1111 |
dl |
1.1 |
*/ |
1112 |
dl |
1.12 |
private static final long FAST_IDLE_TIMEOUT = 200L * 1000L * 1000L; |
1113 |
dl |
1.1 |
|
1114 |
|
|
/** |
1115 |
dl |
1.26 |
* Tolerance for idle timeouts, to cope with timer undershoots |
1116 |
|
|
*/ |
1117 |
dl |
1.33 |
private static final long TIMEOUT_SLOP = 2000000L; |
1118 |
dl |
1.26 |
|
1119 |
|
|
/** |
1120 |
dl |
1.1 |
* The maximum stolen->joining link depth allowed in method |
1121 |
dl |
1.12 |
* tryHelpStealer. Must be a power of two. Depths for legitimate |
1122 |
dl |
1.1 |
* chains are unbounded, but we use a fixed constant to avoid |
1123 |
|
|
* (otherwise unchecked) cycles and to bound staleness of |
1124 |
|
|
* traversal parameters at the expense of sometimes blocking when |
1125 |
|
|
* we could be helping. |
1126 |
|
|
*/ |
1127 |
|
|
private static final int MAX_HELP = 64; |
1128 |
|
|
|
1129 |
|
|
/** |
1130 |
|
|
* Increment for seed generators. See class ThreadLocal for |
1131 |
|
|
* explanation. |
1132 |
|
|
*/ |
1133 |
|
|
private static final int SEED_INCREMENT = 0x61c88647; |
1134 |
|
|
|
1135 |
|
|
/** |
1136 |
|
|
* Bits and masks for control variables |
1137 |
|
|
* |
1138 |
|
|
* Field ctl is a long packed with: |
1139 |
|
|
* AC: Number of active running workers minus target parallelism (16 bits) |
1140 |
|
|
* TC: Number of total workers minus target parallelism (16 bits) |
1141 |
|
|
* ST: true if pool is terminating (1 bit) |
1142 |
|
|
* EC: the wait count of top waiting thread (15 bits) |
1143 |
|
|
* ID: poolIndex of top of Treiber stack of waiters (16 bits) |
1144 |
|
|
* |
1145 |
|
|
* When convenient, we can extract the upper 32 bits of counts and |
1146 |
|
|
* the lower 32 bits of queue state, u = (int)(ctl >>> 32) and e = |
1147 |
|
|
* (int)ctl. The ec field is never accessed alone, but always |
1148 |
|
|
* together with id and st. The offsets of counts by the target |
1149 |
|
|
* parallelism and the positionings of fields makes it possible to |
1150 |
|
|
* perform the most common checks via sign tests of fields: When |
1151 |
|
|
* ac is negative, there are not enough active workers, when tc is |
1152 |
|
|
* negative, there are not enough total workers, and when e is |
1153 |
|
|
* negative, the pool is terminating. To deal with these possibly |
1154 |
|
|
* negative fields, we use casts in and out of "short" and/or |
1155 |
|
|
* signed shifts to maintain signedness. |
1156 |
|
|
* |
1157 |
|
|
* When a thread is queued (inactivated), its eventCount field is |
1158 |
|
|
* set negative, which is the only way to tell if a worker is |
1159 |
|
|
* prevented from executing tasks, even though it must continue to |
1160 |
|
|
* scan for them to avoid queuing races. Note however that |
1161 |
|
|
* eventCount updates lag releases so usage requires care. |
1162 |
|
|
* |
1163 |
dl |
1.12 |
* Field plock is an int packed with: |
1164 |
dl |
1.1 |
* SHUTDOWN: true if shutdown is enabled (1 bit) |
1165 |
dl |
1.12 |
* SEQ: a sequence lock, with PL_LOCK bit set if locked (30 bits) |
1166 |
|
|
* SIGNAL: set when threads may be waiting on the lock (1 bit) |
1167 |
dl |
1.1 |
* |
1168 |
|
|
* The sequence number enables simple consistency checks: |
1169 |
|
|
* Staleness of read-only operations on the workQueues array can |
1170 |
dl |
1.12 |
* be checked by comparing plock before vs after the reads. |
1171 |
dl |
1.1 |
*/ |
1172 |
|
|
|
1173 |
|
|
// bit positions/shifts for fields |
1174 |
|
|
private static final int AC_SHIFT = 48; |
1175 |
|
|
private static final int TC_SHIFT = 32; |
1176 |
|
|
private static final int ST_SHIFT = 31; |
1177 |
|
|
private static final int EC_SHIFT = 16; |
1178 |
|
|
|
1179 |
|
|
// bounds |
1180 |
|
|
private static final int SMASK = 0xffff; // short bits |
1181 |
|
|
private static final int MAX_CAP = 0x7fff; // max #workers - 1 |
1182 |
dl |
1.12 |
private static final int EVENMASK = 0xfffe; // even short bits |
1183 |
|
|
private static final int SQMASK = 0x007e; // max 64 (even) slots |
1184 |
dl |
1.1 |
private static final int SHORT_SIGN = 1 << 15; |
1185 |
|
|
private static final int INT_SIGN = 1 << 31; |
1186 |
|
|
|
1187 |
|
|
// masks |
1188 |
|
|
private static final long STOP_BIT = 0x0001L << ST_SHIFT; |
1189 |
|
|
private static final long AC_MASK = ((long)SMASK) << AC_SHIFT; |
1190 |
|
|
private static final long TC_MASK = ((long)SMASK) << TC_SHIFT; |
1191 |
|
|
|
1192 |
|
|
// units for incrementing and decrementing |
1193 |
|
|
private static final long TC_UNIT = 1L << TC_SHIFT; |
1194 |
|
|
private static final long AC_UNIT = 1L << AC_SHIFT; |
1195 |
|
|
|
1196 |
|
|
// masks and units for dealing with u = (int)(ctl >>> 32) |
1197 |
|
|
private static final int UAC_SHIFT = AC_SHIFT - 32; |
1198 |
|
|
private static final int UTC_SHIFT = TC_SHIFT - 32; |
1199 |
|
|
private static final int UAC_MASK = SMASK << UAC_SHIFT; |
1200 |
|
|
private static final int UTC_MASK = SMASK << UTC_SHIFT; |
1201 |
|
|
private static final int UAC_UNIT = 1 << UAC_SHIFT; |
1202 |
|
|
private static final int UTC_UNIT = 1 << UTC_SHIFT; |
1203 |
|
|
|
1204 |
|
|
// masks and units for dealing with e = (int)ctl |
1205 |
|
|
private static final int E_MASK = 0x7fffffff; // no STOP_BIT |
1206 |
|
|
private static final int E_SEQ = 1 << EC_SHIFT; |
1207 |
|
|
|
1208 |
dl |
1.12 |
// plock bits |
1209 |
dl |
1.1 |
private static final int SHUTDOWN = 1 << 31; |
1210 |
dl |
1.12 |
private static final int PL_LOCK = 2; |
1211 |
|
|
private static final int PL_SIGNAL = 1; |
1212 |
|
|
private static final int PL_SPINS = 1 << 8; |
1213 |
dl |
1.1 |
|
1214 |
|
|
// access mode for WorkQueue |
1215 |
|
|
static final int LIFO_QUEUE = 0; |
1216 |
|
|
static final int FIFO_QUEUE = 1; |
1217 |
|
|
static final int SHARED_QUEUE = -1; |
1218 |
|
|
|
1219 |
dl |
1.18 |
// bounds for #steps in scan loop -- must be power 2 minus 1 |
1220 |
|
|
private static final int MIN_SCAN = 0x1ff; // cover estimation slop |
1221 |
|
|
private static final int MAX_SCAN = 0x1ffff; // 4 * max workers |
1222 |
|
|
|
1223 |
dl |
1.1 |
// Instance fields |
1224 |
|
|
|
1225 |
|
|
/* |
1226 |
dl |
1.18 |
* Field layout of this class tends to matter more than one would |
1227 |
|
|
* like. Runtime layout order is only loosely related to |
1228 |
dl |
1.1 |
* declaration order and may differ across JVMs, but the following |
1229 |
|
|
* empirically works OK on current JVMs. |
1230 |
|
|
*/ |
1231 |
dl |
1.21 |
|
1232 |
|
|
// Heuristic padding to ameliorate unfortunate memory placements |
1233 |
|
|
volatile long pad00, pad01, pad02, pad03, pad04, pad05, pad06; |
1234 |
|
|
|
1235 |
dl |
1.8 |
volatile long stealCount; // collects worker counts |
1236 |
dl |
1.1 |
volatile long ctl; // main pool control |
1237 |
dl |
1.18 |
volatile int plock; // shutdown status and seqLock |
1238 |
dl |
1.12 |
volatile int indexSeed; // worker/submitter index seed |
1239 |
dl |
1.18 |
final int config; // mode and parallelism level |
1240 |
dl |
1.1 |
WorkQueue[] workQueues; // main registry |
1241 |
dl |
1.18 |
final ForkJoinWorkerThreadFactory factory; |
1242 |
dl |
1.1 |
final Thread.UncaughtExceptionHandler ueh; // per-worker UEH |
1243 |
dl |
1.8 |
final String workerNamePrefix; // to create worker name string |
1244 |
|
|
|
1245 |
dl |
1.21 |
volatile Object pad10, pad11, pad12, pad13, pad14, pad15, pad16, pad17; |
1246 |
|
|
volatile Object pad18, pad19, pad1a, pad1b; |
1247 |
|
|
|
1248 |
dl |
1.8 |
/* |
1249 |
dl |
1.12 |
* Acquires the plock lock to protect worker array and related |
1250 |
|
|
* updates. This method is called only if an initial CAS on plock |
1251 |
|
|
* fails. This acts as a spinLock for normal cases, but falls back |
1252 |
|
|
* to builtin monitor to block when (rarely) needed. This would be |
1253 |
|
|
* a terrible idea for a highly contended lock, but works fine as |
1254 |
dl |
1.32 |
* a more conservative alternative to a pure spinlock. |
1255 |
dl |
1.12 |
*/ |
1256 |
|
|
private int acquirePlock() { |
1257 |
|
|
int spins = PL_SPINS, r = 0, ps, nps; |
1258 |
|
|
for (;;) { |
1259 |
|
|
if (((ps = plock) & PL_LOCK) == 0 && |
1260 |
|
|
U.compareAndSwapInt(this, PLOCK, ps, nps = ps + PL_LOCK)) |
1261 |
|
|
return nps; |
1262 |
dl |
1.18 |
else if (r == 0) { // randomize spins if possible |
1263 |
|
|
Thread t = Thread.currentThread(); WorkQueue w; Submitter z; |
1264 |
|
|
if ((t instanceof ForkJoinWorkerThread) && |
1265 |
|
|
(w = ((ForkJoinWorkerThread)t).workQueue) != null) |
1266 |
|
|
r = w.seed; |
1267 |
|
|
else if ((z = submitters.get()) != null) |
1268 |
|
|
r = z.seed; |
1269 |
|
|
else |
1270 |
|
|
r = 1; |
1271 |
|
|
} |
1272 |
dl |
1.8 |
else if (spins >= 0) { |
1273 |
|
|
r ^= r << 1; r ^= r >>> 3; r ^= r << 10; // xorshift |
1274 |
|
|
if (r >= 0) |
1275 |
|
|
--spins; |
1276 |
|
|
} |
1277 |
dl |
1.12 |
else if (U.compareAndSwapInt(this, PLOCK, ps, ps | PL_SIGNAL)) { |
1278 |
jsr166 |
1.13 |
synchronized (this) { |
1279 |
dl |
1.12 |
if ((plock & PL_SIGNAL) != 0) { |
1280 |
dl |
1.8 |
try { |
1281 |
|
|
wait(); |
1282 |
|
|
} catch (InterruptedException ie) { |
1283 |
dl |
1.11 |
try { |
1284 |
|
|
Thread.currentThread().interrupt(); |
1285 |
|
|
} catch (SecurityException ignore) { |
1286 |
|
|
} |
1287 |
dl |
1.8 |
} |
1288 |
|
|
} |
1289 |
|
|
else |
1290 |
dl |
1.12 |
notifyAll(); |
1291 |
dl |
1.8 |
} |
1292 |
|
|
} |
1293 |
|
|
} |
1294 |
|
|
} |
1295 |
dl |
1.1 |
|
1296 |
|
|
/** |
1297 |
dl |
1.12 |
* Unlocks and signals any thread waiting for plock. Called only |
1298 |
|
|
* when CAS of seq value for unlock fails. |
1299 |
dl |
1.1 |
*/ |
1300 |
dl |
1.12 |
private void releasePlock(int ps) { |
1301 |
|
|
plock = ps; |
1302 |
jsr166 |
1.13 |
synchronized (this) { notifyAll(); } |
1303 |
dl |
1.1 |
} |
1304 |
|
|
|
1305 |
dl |
1.18 |
/** |
1306 |
dl |
1.26 |
* Tries to create and start one worker if fewer than target |
1307 |
|
|
* parallelism level exist. Adjusts counts etc on failure. |
1308 |
dl |
1.21 |
*/ |
1309 |
|
|
private void tryAddWorker() { |
1310 |
dl |
1.18 |
long c; int u; |
1311 |
dl |
1.21 |
while ((u = (int)((c = ctl) >>> 32)) < 0 && |
1312 |
|
|
(u & SHORT_SIGN) != 0 && (int)c == 0) { |
1313 |
dl |
1.18 |
long nc = (long)(((u + UTC_UNIT) & UTC_MASK) | |
1314 |
|
|
((u + UAC_UNIT) & UAC_MASK)) << 32; |
1315 |
dl |
1.21 |
if (U.compareAndSwapLong(this, CTL, c, nc)) { |
1316 |
|
|
ForkJoinWorkerThreadFactory fac; |
1317 |
|
|
Throwable ex = null; |
1318 |
|
|
ForkJoinWorkerThread wt = null; |
1319 |
|
|
try { |
1320 |
|
|
if ((fac = factory) != null && |
1321 |
|
|
(wt = fac.newThread(this)) != null) { |
1322 |
|
|
wt.start(); |
1323 |
|
|
break; |
1324 |
|
|
} |
1325 |
|
|
} catch (Throwable e) { |
1326 |
|
|
ex = e; |
1327 |
|
|
} |
1328 |
|
|
deregisterWorker(wt, ex); |
1329 |
|
|
break; |
1330 |
|
|
} |
1331 |
dl |
1.18 |
} |
1332 |
|
|
} |
1333 |
|
|
|
1334 |
|
|
// Registering and deregistering workers |
1335 |
|
|
|
1336 |
|
|
/** |
1337 |
|
|
* Callback from ForkJoinWorkerThread to establish and record its |
1338 |
|
|
* WorkQueue. To avoid scanning bias due to packing entries in |
1339 |
|
|
* front of the workQueues array, we treat the array as a simple |
1340 |
|
|
* power-of-two hash table using per-thread seed as hash, |
1341 |
|
|
* expanding as needed. |
1342 |
|
|
* |
1343 |
|
|
* @param wt the worker thread |
1344 |
dl |
1.21 |
* @return the worker's queue |
1345 |
dl |
1.18 |
*/ |
1346 |
dl |
1.21 |
final WorkQueue registerWorker(ForkJoinWorkerThread wt) { |
1347 |
|
|
Thread.UncaughtExceptionHandler handler; WorkQueue[] ws; int s, ps; |
1348 |
|
|
wt.setDaemon(true); |
1349 |
|
|
if ((handler = ueh) != null) |
1350 |
|
|
wt.setUncaughtExceptionHandler(handler); |
1351 |
|
|
do {} while (!U.compareAndSwapInt(this, INDEXSEED, s = indexSeed, |
1352 |
|
|
s += SEED_INCREMENT) || |
1353 |
|
|
s == 0); // skip 0 |
1354 |
|
|
WorkQueue w = new WorkQueue(this, wt, config >>> 16, s); |
1355 |
|
|
if (((ps = plock) & PL_LOCK) != 0 || |
1356 |
|
|
!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK)) |
1357 |
|
|
ps = acquirePlock(); |
1358 |
|
|
int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN); |
1359 |
|
|
try { |
1360 |
|
|
if ((ws = workQueues) != null) { // skip if shutting down |
1361 |
|
|
int n = ws.length, m = n - 1; |
1362 |
|
|
int r = (s << 1) | 1; // use odd-numbered indices |
1363 |
|
|
if (ws[r &= m] != null) { // collision |
1364 |
|
|
int probes = 0; // step by approx half size |
1365 |
|
|
int step = (n <= 4) ? 2 : ((n >>> 1) & EVENMASK) + 2; |
1366 |
|
|
while (ws[r = (r + step) & m] != null) { |
1367 |
|
|
if (++probes >= n) { |
1368 |
|
|
workQueues = ws = Arrays.copyOf(ws, n <<= 1); |
1369 |
|
|
m = n - 1; |
1370 |
|
|
probes = 0; |
1371 |
dl |
1.1 |
} |
1372 |
|
|
} |
1373 |
|
|
} |
1374 |
dl |
1.21 |
w.eventCount = w.poolIndex = r; // volatile write orders |
1375 |
|
|
ws[r] = w; |
1376 |
dl |
1.1 |
} |
1377 |
dl |
1.21 |
} finally { |
1378 |
|
|
if (!U.compareAndSwapInt(this, PLOCK, ps, nps)) |
1379 |
|
|
releasePlock(nps); |
1380 |
dl |
1.1 |
} |
1381 |
dl |
1.21 |
wt.setName(workerNamePrefix.concat(Integer.toString(w.poolIndex))); |
1382 |
|
|
return w; |
1383 |
dl |
1.1 |
} |
1384 |
|
|
|
1385 |
|
|
/** |
1386 |
|
|
* Final callback from terminating worker, as well as upon failure |
1387 |
dl |
1.12 |
* to construct or start a worker. Removes record of worker from |
1388 |
|
|
* array, and adjusts counts. If pool is shutting down, tries to |
1389 |
|
|
* complete termination. |
1390 |
dl |
1.1 |
* |
1391 |
dl |
1.12 |
* @param wt the worker thread or null if construction failed |
1392 |
dl |
1.1 |
* @param ex the exception causing failure, or null if none |
1393 |
|
|
*/ |
1394 |
|
|
final void deregisterWorker(ForkJoinWorkerThread wt, Throwable ex) { |
1395 |
|
|
WorkQueue w = null; |
1396 |
|
|
if (wt != null && (w = wt.workQueue) != null) { |
1397 |
dl |
1.12 |
int ps; |
1398 |
|
|
w.qlock = -1; // ensure set |
1399 |
dl |
1.18 |
long ns = w.nsteals, sc; // collect steal count |
1400 |
|
|
do {} while (!U.compareAndSwapLong(this, STEALCOUNT, |
1401 |
|
|
sc = stealCount, sc + ns)); |
1402 |
dl |
1.12 |
if (((ps = plock) & PL_LOCK) != 0 || |
1403 |
|
|
!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK)) |
1404 |
|
|
ps = acquirePlock(); |
1405 |
|
|
int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN); |
1406 |
dl |
1.8 |
try { |
1407 |
dl |
1.12 |
int idx = w.poolIndex; |
1408 |
dl |
1.1 |
WorkQueue[] ws = workQueues; |
1409 |
|
|
if (ws != null && idx >= 0 && idx < ws.length && ws[idx] == w) |
1410 |
|
|
ws[idx] = null; |
1411 |
|
|
} finally { |
1412 |
dl |
1.12 |
if (!U.compareAndSwapInt(this, PLOCK, ps, nps)) |
1413 |
|
|
releasePlock(nps); |
1414 |
dl |
1.1 |
} |
1415 |
|
|
} |
1416 |
|
|
|
1417 |
dl |
1.32 |
long c; // adjust ctl counts |
1418 |
dl |
1.1 |
do {} while (!U.compareAndSwapLong |
1419 |
|
|
(this, CTL, c = ctl, (((c - AC_UNIT) & AC_MASK) | |
1420 |
|
|
((c - TC_UNIT) & TC_MASK) | |
1421 |
|
|
(c & ~(AC_MASK|TC_MASK))))); |
1422 |
|
|
|
1423 |
dl |
1.26 |
if (!tryTerminate(false, false) && w != null && w.array != null) { |
1424 |
dl |
1.32 |
w.cancelAll(); // cancel remaining tasks |
1425 |
|
|
WorkQueue[] ws; WorkQueue v; Thread p; int u, i, e; |
1426 |
|
|
while ((u = (int)((c = ctl) >>> 32)) < 0 && (e = (int)c) >= 0) { |
1427 |
|
|
if (e > 0) { // activate or create replacement |
1428 |
|
|
if ((ws = workQueues) == null || |
1429 |
|
|
(i = e & SMASK) >= ws.length || |
1430 |
dl |
1.36 |
(v = ws[i]) == null) |
1431 |
dl |
1.32 |
break; |
1432 |
|
|
long nc = (((long)(v.nextWait & E_MASK)) | |
1433 |
|
|
((long)(u + UAC_UNIT) << 32)); |
1434 |
|
|
if (v.eventCount != (e | INT_SIGN)) |
1435 |
|
|
break; |
1436 |
|
|
if (U.compareAndSwapLong(this, CTL, c, nc)) { |
1437 |
|
|
v.eventCount = (e + E_SEQ) & E_MASK; |
1438 |
|
|
if ((p = v.parker) != null) |
1439 |
|
|
U.unpark(p); |
1440 |
|
|
break; |
1441 |
dl |
1.26 |
} |
1442 |
|
|
} |
1443 |
|
|
else { |
1444 |
|
|
if ((short)u < 0) |
1445 |
|
|
tryAddWorker(); |
1446 |
|
|
break; |
1447 |
|
|
} |
1448 |
|
|
} |
1449 |
dl |
1.1 |
} |
1450 |
dl |
1.26 |
if (ex == null) // help clean refs on way out |
1451 |
|
|
ForkJoinTask.helpExpungeStaleExceptions(); |
1452 |
|
|
else // rethrow |
1453 |
dl |
1.11 |
ForkJoinTask.rethrow(ex); |
1454 |
dl |
1.1 |
} |
1455 |
|
|
|
1456 |
|
|
// Submissions |
1457 |
|
|
|
1458 |
|
|
/** |
1459 |
|
|
* Unless shutting down, adds the given task to a submission queue |
1460 |
|
|
* at submitter's current queue index (modulo submission |
1461 |
dl |
1.12 |
* range). Only the most common path is directly handled in this |
1462 |
|
|
* method. All others are relayed to fullExternalPush. |
1463 |
dl |
1.1 |
* |
1464 |
|
|
* @param task the task. Caller must ensure non-null. |
1465 |
|
|
*/ |
1466 |
dl |
1.12 |
final void externalPush(ForkJoinTask<?> task) { |
1467 |
|
|
WorkQueue[] ws; WorkQueue q; Submitter z; int m; ForkJoinTask<?>[] a; |
1468 |
|
|
if ((z = submitters.get()) != null && plock > 0 && |
1469 |
|
|
(ws = workQueues) != null && (m = (ws.length - 1)) >= 0 && |
1470 |
|
|
(q = ws[m & z.seed & SQMASK]) != null && |
1471 |
|
|
U.compareAndSwapInt(q, QLOCK, 0, 1)) { // lock |
1472 |
dl |
1.18 |
int b = q.base, s = q.top, n, an; |
1473 |
|
|
if ((a = q.array) != null && (an = a.length) > (n = s + 1 - b)) { |
1474 |
dl |
1.21 |
int j = (((an - 1) & s) << ASHIFT) + ABASE; |
1475 |
|
|
U.putOrderedObject(a, j, task); |
1476 |
dl |
1.12 |
q.top = s + 1; // push on to deque |
1477 |
|
|
q.qlock = 0; |
1478 |
dl |
1.18 |
if (n <= 2) |
1479 |
dl |
1.21 |
signalWork(q); |
1480 |
dl |
1.1 |
return; |
1481 |
|
|
} |
1482 |
dl |
1.12 |
q.qlock = 0; |
1483 |
dl |
1.1 |
} |
1484 |
dl |
1.12 |
fullExternalPush(task); |
1485 |
dl |
1.1 |
} |
1486 |
|
|
|
1487 |
dl |
1.7 |
/** |
1488 |
dl |
1.12 |
* Full version of externalPush. This method is called, among |
1489 |
|
|
* other times, upon the first submission of the first task to the |
1490 |
dl |
1.37 |
* pool, so must perform secondary initialization. It also |
1491 |
|
|
* detects first submission by an external thread by looking up |
1492 |
|
|
* its ThreadLocal, and creates a new shared queue if the one at |
1493 |
|
|
* index if empty or contended. The plock lock body must be |
1494 |
|
|
* exception-free (so no try/finally) so we optimistically |
1495 |
|
|
* allocate new queues outside the lock and throw them away if |
1496 |
|
|
* (very rarely) not needed. |
1497 |
|
|
* |
1498 |
|
|
* Secondary initialization occurs when plock is zero, to create |
1499 |
|
|
* workQueue array and set plock to a valid value. This lock body |
1500 |
|
|
* must also be exception-free. Because the plock seq value can |
1501 |
|
|
* eventually wrap around zero, this method harmlessly fails to |
1502 |
|
|
* reinitialize if workQueues exists, while still advancing plock. |
1503 |
dl |
1.12 |
*/ |
1504 |
|
|
private void fullExternalPush(ForkJoinTask<?> task) { |
1505 |
dl |
1.21 |
int r = 0; // random index seed |
1506 |
dl |
1.18 |
for (Submitter z = submitters.get();;) { |
1507 |
|
|
WorkQueue[] ws; WorkQueue q; int ps, m, k; |
1508 |
|
|
if (z == null) { |
1509 |
|
|
if (U.compareAndSwapInt(this, INDEXSEED, r = indexSeed, |
1510 |
|
|
r += SEED_INCREMENT) && r != 0) |
1511 |
|
|
submitters.set(z = new Submitter(r)); |
1512 |
|
|
} |
1513 |
dl |
1.37 |
else if (r == 0) { // move to a different index |
1514 |
dl |
1.18 |
r = z.seed; |
1515 |
dl |
1.37 |
r ^= r << 13; // same xorshift as WorkQueues |
1516 |
dl |
1.18 |
r ^= r >>> 17; |
1517 |
|
|
z.seed = r ^ (r << 5); |
1518 |
|
|
} |
1519 |
|
|
else if ((ps = plock) < 0) |
1520 |
dl |
1.12 |
throw new RejectedExecutionException(); |
1521 |
dl |
1.18 |
else if (ps == 0 || (ws = workQueues) == null || |
1522 |
dl |
1.37 |
(m = ws.length - 1) < 0) { // initialize workQueues |
1523 |
|
|
int p = config & SMASK; // find power of two table size |
1524 |
|
|
int n = (p > 1) ? p - 1 : 1; // ensure at least 2 slots |
1525 |
|
|
n |= n >>> 1; n |= n >>> 2; n |= n >>> 4; |
1526 |
|
|
n |= n >>> 8; n |= n >>> 16; n = (n + 1) << 1; |
1527 |
|
|
WorkQueue[] nws = ((ws = workQueues) == null || ws.length == 0 ? |
1528 |
|
|
new WorkQueue[n] : null); |
1529 |
|
|
if (((ps = plock) & PL_LOCK) != 0 || |
1530 |
|
|
!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK)) |
1531 |
|
|
ps = acquirePlock(); |
1532 |
|
|
if (((ws = workQueues) == null || ws.length == 0) && nws != null) |
1533 |
|
|
workQueues = nws; |
1534 |
|
|
int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN); |
1535 |
|
|
if (!U.compareAndSwapInt(this, PLOCK, ps, nps)) |
1536 |
|
|
releasePlock(nps); |
1537 |
|
|
} |
1538 |
dl |
1.18 |
else if ((q = ws[k = r & m & SQMASK]) != null) { |
1539 |
dl |
1.21 |
if (q.qlock == 0 && U.compareAndSwapInt(q, QLOCK, 0, 1)) { |
1540 |
|
|
ForkJoinTask<?>[] a = q.array; |
1541 |
|
|
int s = q.top; |
1542 |
|
|
boolean submitted = false; |
1543 |
|
|
try { // locked version of push |
1544 |
|
|
if ((a != null && a.length > s + 1 - q.base) || |
1545 |
|
|
(a = q.growArray()) != null) { // must presize |
1546 |
|
|
int j = (((a.length - 1) & s) << ASHIFT) + ABASE; |
1547 |
|
|
U.putOrderedObject(a, j, task); |
1548 |
|
|
q.top = s + 1; |
1549 |
|
|
submitted = true; |
1550 |
|
|
} |
1551 |
|
|
} finally { |
1552 |
|
|
q.qlock = 0; // unlock |
1553 |
|
|
} |
1554 |
|
|
if (submitted) { |
1555 |
|
|
signalWork(q); |
1556 |
|
|
return; |
1557 |
|
|
} |
1558 |
|
|
} |
1559 |
|
|
r = 0; // move on failure |
1560 |
dl |
1.18 |
} |
1561 |
|
|
else if (((ps = plock) & PL_LOCK) == 0) { // create new queue |
1562 |
|
|
q = new WorkQueue(this, null, SHARED_QUEUE, r); |
1563 |
|
|
if (((ps = plock) & PL_LOCK) != 0 || |
1564 |
dl |
1.12 |
!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK)) |
1565 |
|
|
ps = acquirePlock(); |
1566 |
dl |
1.18 |
if ((ws = workQueues) != null && k < ws.length && ws[k] == null) |
1567 |
|
|
ws[k] = q; |
1568 |
dl |
1.12 |
int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN); |
1569 |
|
|
if (!U.compareAndSwapInt(this, PLOCK, ps, nps)) |
1570 |
|
|
releasePlock(nps); |
1571 |
|
|
} |
1572 |
dl |
1.18 |
else |
1573 |
|
|
r = 0; // try elsewhere while lock held |
1574 |
dl |
1.11 |
} |
1575 |
dl |
1.9 |
} |
1576 |
|
|
|
1577 |
dl |
1.1 |
// Maintaining ctl counts |
1578 |
|
|
|
1579 |
|
|
/** |
1580 |
|
|
* Increments active count; mainly called upon return from blocking. |
1581 |
|
|
*/ |
1582 |
|
|
final void incrementActiveCount() { |
1583 |
|
|
long c; |
1584 |
|
|
do {} while (!U.compareAndSwapLong(this, CTL, c = ctl, c + AC_UNIT)); |
1585 |
|
|
} |
1586 |
|
|
|
1587 |
|
|
/** |
1588 |
dl |
1.21 |
* Tries to create or activate a worker if too few are active. |
1589 |
|
|
* |
1590 |
|
|
* @param q the (non-null) queue holding tasks to be signalled |
1591 |
dl |
1.12 |
*/ |
1592 |
dl |
1.21 |
final void signalWork(WorkQueue q) { |
1593 |
|
|
int hint = q.poolIndex; |
1594 |
|
|
long c; int e, u, i, n; WorkQueue[] ws; WorkQueue w; Thread p; |
1595 |
dl |
1.12 |
while ((u = (int)((c = ctl) >>> 32)) < 0) { |
1596 |
|
|
if ((e = (int)c) > 0) { |
1597 |
|
|
if ((ws = workQueues) != null && ws.length > (i = e & SMASK) && |
1598 |
dl |
1.1 |
(w = ws[i]) != null && w.eventCount == (e | INT_SIGN)) { |
1599 |
|
|
long nc = (((long)(w.nextWait & E_MASK)) | |
1600 |
|
|
((long)(u + UAC_UNIT) << 32)); |
1601 |
|
|
if (U.compareAndSwapLong(this, CTL, c, nc)) { |
1602 |
dl |
1.21 |
w.hint = hint; |
1603 |
dl |
1.1 |
w.eventCount = (e + E_SEQ) & E_MASK; |
1604 |
dl |
1.21 |
if ((p = w.parker) != null) |
1605 |
dl |
1.12 |
U.unpark(p); |
1606 |
dl |
1.21 |
break; |
1607 |
dl |
1.12 |
} |
1608 |
dl |
1.21 |
if (q.top - q.base <= 0) |
1609 |
dl |
1.1 |
break; |
1610 |
|
|
} |
1611 |
|
|
else |
1612 |
|
|
break; |
1613 |
|
|
} |
1614 |
dl |
1.21 |
else { |
1615 |
|
|
if ((short)u < 0) |
1616 |
|
|
tryAddWorker(); |
1617 |
|
|
break; |
1618 |
dl |
1.1 |
} |
1619 |
|
|
} |
1620 |
|
|
} |
1621 |
|
|
|
1622 |
|
|
// Scanning for tasks |
1623 |
|
|
|
1624 |
|
|
/** |
1625 |
|
|
* Top-level runloop for workers, called by ForkJoinWorkerThread.run. |
1626 |
|
|
*/ |
1627 |
|
|
final void runWorker(WorkQueue w) { |
1628 |
dl |
1.21 |
w.growArray(); // allocate queue |
1629 |
|
|
do { w.runTask(scan(w)); } while (w.qlock >= 0); |
1630 |
dl |
1.1 |
} |
1631 |
|
|
|
1632 |
|
|
/** |
1633 |
|
|
* Scans for and, if found, returns one task, else possibly |
1634 |
|
|
* inactivates the worker. This method operates on single reads of |
1635 |
|
|
* volatile state and is designed to be re-invoked continuously, |
1636 |
|
|
* in part because it returns upon detecting inconsistencies, |
1637 |
|
|
* contention, or state changes that indicate possible success on |
1638 |
|
|
* re-invocation. |
1639 |
|
|
* |
1640 |
dl |
1.18 |
* The scan searches for tasks across queues (starting at a random |
1641 |
|
|
* index, and relying on registerWorker to irregularly scatter |
1642 |
|
|
* them within array to avoid bias), checking each at least twice. |
1643 |
|
|
* The scan terminates upon either finding a non-empty queue, or |
1644 |
|
|
* completing the sweep. If the worker is not inactivated, it |
1645 |
|
|
* takes and returns a task from this queue. Otherwise, if not |
1646 |
|
|
* activated, it signals workers (that may include itself) and |
1647 |
|
|
* returns so caller can retry. Also returns for true if the |
1648 |
|
|
* worker array may have changed during an empty scan. On failure |
1649 |
|
|
* to find a task, we take one of the following actions, after |
1650 |
|
|
* which the caller will retry calling this method unless |
1651 |
|
|
* terminated. |
1652 |
dl |
1.1 |
* |
1653 |
|
|
* * If pool is terminating, terminate the worker. |
1654 |
|
|
* |
1655 |
|
|
* * If not already enqueued, try to inactivate and enqueue the |
1656 |
|
|
* worker on wait queue. Or, if inactivating has caused the pool |
1657 |
dl |
1.32 |
* to be quiescent, relay to idleAwaitWork to possibly shrink |
1658 |
|
|
* pool. |
1659 |
dl |
1.1 |
* |
1660 |
dl |
1.12 |
* * If already enqueued and none of the above apply, possibly |
1661 |
dl |
1.32 |
* park awaiting signal, else lingering to help scan and signal. |
1662 |
|
|
* |
1663 |
|
|
* * If a non-empty queue discovered or left as a hint, |
1664 |
jsr166 |
1.39 |
* help wake up other workers before return. |
1665 |
dl |
1.1 |
* |
1666 |
|
|
* @param w the worker (via its WorkQueue) |
1667 |
jsr166 |
1.5 |
* @return a task or null if none found |
1668 |
dl |
1.1 |
*/ |
1669 |
|
|
private final ForkJoinTask<?> scan(WorkQueue w) { |
1670 |
dl |
1.21 |
WorkQueue[] ws; int m; |
1671 |
dl |
1.18 |
int ps = plock; // read plock before ws |
1672 |
|
|
if (w != null && (ws = workQueues) != null && (m = ws.length - 1) >= 0) { |
1673 |
|
|
int ec = w.eventCount; // ec is negative if inactive |
1674 |
|
|
int r = w.seed; r ^= r << 13; r ^= r >>> 17; w.seed = r ^= r << 5; |
1675 |
dl |
1.32 |
w.hint = -1; // update seed and clear hint |
1676 |
dl |
1.21 |
int j = ((m + m + 1) | MIN_SCAN) & MAX_SCAN; |
1677 |
|
|
do { |
1678 |
dl |
1.18 |
WorkQueue q; ForkJoinTask<?>[] a; int b; |
1679 |
|
|
if ((q = ws[(r + j) & m]) != null && (b = q.base) - q.top < 0 && |
1680 |
|
|
(a = q.array) != null) { // probably nonempty |
1681 |
dl |
1.1 |
int i = (((a.length - 1) & b) << ASHIFT) + ABASE; |
1682 |
dl |
1.18 |
ForkJoinTask<?> t = (ForkJoinTask<?>) |
1683 |
|
|
U.getObjectVolatile(a, i); |
1684 |
dl |
1.1 |
if (q.base == b && ec >= 0 && t != null && |
1685 |
|
|
U.compareAndSwapObject(a, i, t, null)) { |
1686 |
dl |
1.18 |
if ((q.base = b + 1) - q.top < 0) |
1687 |
dl |
1.21 |
signalWork(q); |
1688 |
dl |
1.18 |
return t; // taken |
1689 |
|
|
} |
1690 |
dl |
1.21 |
else if ((ec < 0 || j < m) && (int)(ctl >> AC_SHIFT) <= 0) { |
1691 |
|
|
w.hint = (r + j) & m; // help signal below |
1692 |
|
|
break; // cannot take |
1693 |
|
|
} |
1694 |
|
|
} |
1695 |
|
|
} while (--j >= 0); |
1696 |
|
|
|
1697 |
dl |
1.32 |
int h, e, ns; long c, sc; WorkQueue q; |
1698 |
|
|
if ((ns = w.nsteals) != 0) { |
1699 |
|
|
if (U.compareAndSwapLong(this, STEALCOUNT, |
1700 |
|
|
sc = stealCount, sc + ns)) |
1701 |
|
|
w.nsteals = 0; // collect steals and rescan |
1702 |
|
|
} |
1703 |
|
|
else if (plock != ps) // consistency check |
1704 |
|
|
; // skip |
1705 |
|
|
else if ((e = (int)(c = ctl)) < 0) |
1706 |
|
|
w.qlock = -1; // pool is terminating |
1707 |
|
|
else { |
1708 |
|
|
if ((h = w.hint) < 0) { |
1709 |
|
|
if (ec >= 0) { // try to enqueue/inactivate |
1710 |
|
|
long nc = (((long)ec | |
1711 |
|
|
((c - AC_UNIT) & (AC_MASK|TC_MASK)))); |
1712 |
|
|
w.nextWait = e; // link and mark inactive |
1713 |
|
|
w.eventCount = ec | INT_SIGN; |
1714 |
|
|
if (ctl != c || !U.compareAndSwapLong(this, CTL, c, nc)) |
1715 |
|
|
w.eventCount = ec; // unmark on CAS failure |
1716 |
|
|
else if ((int)(c >> AC_SHIFT) == 1 - (config & SMASK)) |
1717 |
|
|
idleAwaitWork(w, nc, c); |
1718 |
|
|
} |
1719 |
dl |
1.37 |
else if (w.eventCount < 0 && ctl == c) { |
1720 |
dl |
1.32 |
Thread wt = Thread.currentThread(); |
1721 |
|
|
Thread.interrupted(); // clear status |
1722 |
|
|
U.putObject(wt, PARKBLOCKER, this); |
1723 |
|
|
w.parker = wt; // emulate LockSupport.park |
1724 |
|
|
if (w.eventCount < 0) // recheck |
1725 |
dl |
1.37 |
U.park(false, 0L); // block |
1726 |
dl |
1.32 |
w.parker = null; |
1727 |
|
|
U.putObject(wt, PARKBLOCKER, null); |
1728 |
|
|
} |
1729 |
|
|
} |
1730 |
|
|
if ((h >= 0 || (h = w.hint) >= 0) && |
1731 |
|
|
(ws = workQueues) != null && h < ws.length && |
1732 |
|
|
(q = ws[h]) != null) { // signal others before retry |
1733 |
|
|
WorkQueue v; Thread p; int u, i, s; |
1734 |
dl |
1.37 |
for (int n = (config & SMASK) - 1;;) { |
1735 |
dl |
1.32 |
int idleCount = (w.eventCount < 0) ? 0 : -1; |
1736 |
|
|
if (((s = idleCount - q.base + q.top) <= n && |
1737 |
|
|
(n = s) <= 0) || |
1738 |
|
|
(u = (int)((c = ctl) >>> 32)) >= 0 || |
1739 |
|
|
(e = (int)c) <= 0 || m < (i = e & SMASK) || |
1740 |
|
|
(v = ws[i]) == null) |
1741 |
|
|
break; |
1742 |
|
|
long nc = (((long)(v.nextWait & E_MASK)) | |
1743 |
|
|
((long)(u + UAC_UNIT) << 32)); |
1744 |
|
|
if (v.eventCount != (e | INT_SIGN) || |
1745 |
|
|
!U.compareAndSwapLong(this, CTL, c, nc)) |
1746 |
|
|
break; |
1747 |
|
|
v.hint = h; |
1748 |
|
|
v.eventCount = (e + E_SEQ) & E_MASK; |
1749 |
|
|
if ((p = v.parker) != null) |
1750 |
|
|
U.unpark(p); |
1751 |
|
|
if (--n <= 0) |
1752 |
|
|
break; |
1753 |
|
|
} |
1754 |
dl |
1.21 |
} |
1755 |
|
|
} |
1756 |
dl |
1.1 |
} |
1757 |
|
|
return null; |
1758 |
|
|
} |
1759 |
|
|
|
1760 |
|
|
/** |
1761 |
|
|
* If inactivating worker w has caused the pool to become |
1762 |
|
|
* quiescent, checks for pool termination, and, so long as this is |
1763 |
dl |
1.7 |
* not the only worker, waits for event for up to a given |
1764 |
|
|
* duration. On timeout, if ctl has not changed, terminates the |
1765 |
dl |
1.1 |
* worker, which will in turn wake up another worker to possibly |
1766 |
|
|
* repeat this process. |
1767 |
|
|
* |
1768 |
|
|
* @param w the calling worker |
1769 |
|
|
* @param currentCtl the ctl value triggering possible quiescence |
1770 |
|
|
* @param prevCtl the ctl value to restore if thread is terminated |
1771 |
|
|
*/ |
1772 |
|
|
private void idleAwaitWork(WorkQueue w, long currentCtl, long prevCtl) { |
1773 |
dl |
1.18 |
if (w != null && w.eventCount < 0 && |
1774 |
dl |
1.37 |
!tryTerminate(false, false) && (int)prevCtl != 0 && |
1775 |
|
|
ctl == currentCtl) { |
1776 |
dl |
1.7 |
int dc = -(short)(currentCtl >>> TC_SHIFT); |
1777 |
|
|
long parkTime = dc < 0 ? FAST_IDLE_TIMEOUT: (dc + 1) * IDLE_TIMEOUT; |
1778 |
dl |
1.26 |
long deadline = System.nanoTime() + parkTime - TIMEOUT_SLOP; |
1779 |
dl |
1.1 |
Thread wt = Thread.currentThread(); |
1780 |
|
|
while (ctl == currentCtl) { |
1781 |
|
|
Thread.interrupted(); // timed variant of version in scan() |
1782 |
|
|
U.putObject(wt, PARKBLOCKER, this); |
1783 |
|
|
w.parker = wt; |
1784 |
|
|
if (ctl == currentCtl) |
1785 |
dl |
1.7 |
U.park(false, parkTime); |
1786 |
dl |
1.1 |
w.parker = null; |
1787 |
|
|
U.putObject(wt, PARKBLOCKER, null); |
1788 |
|
|
if (ctl != currentCtl) |
1789 |
|
|
break; |
1790 |
dl |
1.7 |
if (deadline - System.nanoTime() <= 0L && |
1791 |
dl |
1.1 |
U.compareAndSwapLong(this, CTL, currentCtl, prevCtl)) { |
1792 |
|
|
w.eventCount = (w.eventCount + E_SEQ) | E_MASK; |
1793 |
dl |
1.37 |
w.hint = -1; |
1794 |
dl |
1.12 |
w.qlock = -1; // shrink |
1795 |
dl |
1.1 |
break; |
1796 |
|
|
} |
1797 |
|
|
} |
1798 |
|
|
} |
1799 |
|
|
} |
1800 |
|
|
|
1801 |
|
|
/** |
1802 |
dl |
1.26 |
* Scans through queues looking for work while joining a task; if |
1803 |
|
|
* any present, signals. May return early if more signalling is |
1804 |
|
|
* detectably unneeded. |
1805 |
dl |
1.12 |
* |
1806 |
dl |
1.26 |
* @param task return early if done |
1807 |
dl |
1.12 |
* @param origin an index to start scan |
1808 |
|
|
*/ |
1809 |
dl |
1.26 |
private void helpSignal(ForkJoinTask<?> task, int origin) { |
1810 |
dl |
1.21 |
WorkQueue[] ws; WorkQueue w; Thread p; long c; int m, u, e, i, s; |
1811 |
dl |
1.26 |
if (task != null && task.status >= 0 && |
1812 |
|
|
(u = (int)(ctl >>> 32)) < 0 && (u >> UAC_SHIFT) < 0 && |
1813 |
dl |
1.21 |
(ws = workQueues) != null && (m = ws.length - 1) >= 0) { |
1814 |
dl |
1.26 |
outer: for (int k = origin, j = m; j >= 0; --j) { |
1815 |
dl |
1.21 |
WorkQueue q = ws[k++ & m]; |
1816 |
|
|
for (int n = m;;) { // limit to at most m signals |
1817 |
dl |
1.26 |
if (task.status < 0) |
1818 |
dl |
1.21 |
break outer; |
1819 |
|
|
if (q == null || |
1820 |
dl |
1.26 |
((s = -q.base + q.top) <= n && (n = s) <= 0)) |
1821 |
dl |
1.12 |
break; |
1822 |
dl |
1.21 |
if ((u = (int)((c = ctl) >>> 32)) >= 0 || |
1823 |
|
|
(e = (int)c) <= 0 || m < (i = e & SMASK) || |
1824 |
|
|
(w = ws[i]) == null) |
1825 |
|
|
break outer; |
1826 |
|
|
long nc = (((long)(w.nextWait & E_MASK)) | |
1827 |
|
|
((long)(u + UAC_UNIT) << 32)); |
1828 |
dl |
1.32 |
if (w.eventCount != (e | INT_SIGN)) |
1829 |
|
|
break outer; |
1830 |
|
|
if (U.compareAndSwapLong(this, CTL, c, nc)) { |
1831 |
dl |
1.28 |
w.eventCount = (e + E_SEQ) & E_MASK; |
1832 |
|
|
if ((p = w.parker) != null) |
1833 |
|
|
U.unpark(p); |
1834 |
|
|
if (--n <= 0) |
1835 |
|
|
break; |
1836 |
|
|
} |
1837 |
dl |
1.26 |
} |
1838 |
|
|
} |
1839 |
|
|
} |
1840 |
|
|
} |
1841 |
|
|
|
1842 |
|
|
/** |
1843 |
dl |
1.1 |
* Tries to locate and execute tasks for a stealer of the given |
1844 |
|
|
* task, or in turn one of its stealers, Traces currentSteal -> |
1845 |
|
|
* currentJoin links looking for a thread working on a descendant |
1846 |
|
|
* of the given task and with a non-empty queue to steal back and |
1847 |
|
|
* execute tasks from. The first call to this method upon a |
1848 |
|
|
* waiting join will often entail scanning/search, (which is OK |
1849 |
|
|
* because the joiner has nothing better to do), but this method |
1850 |
|
|
* leaves hints in workers to speed up subsequent calls. The |
1851 |
|
|
* implementation is very branchy to cope with potential |
1852 |
|
|
* inconsistencies or loops encountering chains that are stale, |
1853 |
|
|
* unknown, or so long that they are likely cyclic. |
1854 |
|
|
* |
1855 |
|
|
* @param joiner the joining worker |
1856 |
|
|
* @param task the task to join |
1857 |
|
|
* @return 0 if no progress can be made, negative if task |
1858 |
|
|
* known complete, else positive |
1859 |
|
|
*/ |
1860 |
|
|
private int tryHelpStealer(WorkQueue joiner, ForkJoinTask<?> task) { |
1861 |
|
|
int stat = 0, steps = 0; // bound to avoid cycles |
1862 |
|
|
if (joiner != null && task != null) { // hoist null checks |
1863 |
|
|
restart: for (;;) { |
1864 |
|
|
ForkJoinTask<?> subtask = task; // current target |
1865 |
|
|
for (WorkQueue j = joiner, v;;) { // v is stealer of subtask |
1866 |
|
|
WorkQueue[] ws; int m, s, h; |
1867 |
|
|
if ((s = task.status) < 0) { |
1868 |
|
|
stat = s; |
1869 |
|
|
break restart; |
1870 |
|
|
} |
1871 |
|
|
if ((ws = workQueues) == null || (m = ws.length - 1) <= 0) |
1872 |
|
|
break restart; // shutting down |
1873 |
dl |
1.18 |
if ((v = ws[h = (j.hint | 1) & m]) == null || |
1874 |
dl |
1.1 |
v.currentSteal != subtask) { |
1875 |
|
|
for (int origin = h;;) { // find stealer |
1876 |
|
|
if (((h = (h + 2) & m) & 15) == 1 && |
1877 |
|
|
(subtask.status < 0 || j.currentJoin != subtask)) |
1878 |
|
|
continue restart; // occasional staleness check |
1879 |
|
|
if ((v = ws[h]) != null && |
1880 |
|
|
v.currentSteal == subtask) { |
1881 |
dl |
1.18 |
j.hint = h; // save hint |
1882 |
dl |
1.1 |
break; |
1883 |
|
|
} |
1884 |
|
|
if (h == origin) |
1885 |
|
|
break restart; // cannot find stealer |
1886 |
|
|
} |
1887 |
|
|
} |
1888 |
|
|
for (;;) { // help stealer or descend to its stealer |
1889 |
|
|
ForkJoinTask[] a; int b; |
1890 |
|
|
if (subtask.status < 0) // surround probes with |
1891 |
|
|
continue restart; // consistency checks |
1892 |
|
|
if ((b = v.base) - v.top < 0 && (a = v.array) != null) { |
1893 |
|
|
int i = (((a.length - 1) & b) << ASHIFT) + ABASE; |
1894 |
|
|
ForkJoinTask<?> t = |
1895 |
|
|
(ForkJoinTask<?>)U.getObjectVolatile(a, i); |
1896 |
|
|
if (subtask.status < 0 || j.currentJoin != subtask || |
1897 |
|
|
v.currentSteal != subtask) |
1898 |
|
|
continue restart; // stale |
1899 |
|
|
stat = 1; // apparent progress |
1900 |
|
|
if (t != null && v.base == b && |
1901 |
|
|
U.compareAndSwapObject(a, i, t, null)) { |
1902 |
|
|
v.base = b + 1; // help stealer |
1903 |
|
|
joiner.runSubtask(t); |
1904 |
|
|
} |
1905 |
|
|
else if (v.base == b && ++steps == MAX_HELP) |
1906 |
|
|
break restart; // v apparently stalled |
1907 |
|
|
} |
1908 |
|
|
else { // empty -- try to descend |
1909 |
|
|
ForkJoinTask<?> next = v.currentJoin; |
1910 |
|
|
if (subtask.status < 0 || j.currentJoin != subtask || |
1911 |
|
|
v.currentSteal != subtask) |
1912 |
|
|
continue restart; // stale |
1913 |
|
|
else if (next == null || ++steps == MAX_HELP) |
1914 |
|
|
break restart; // dead-end or maybe cyclic |
1915 |
|
|
else { |
1916 |
|
|
subtask = next; |
1917 |
|
|
j = v; |
1918 |
|
|
break; |
1919 |
|
|
} |
1920 |
|
|
} |
1921 |
|
|
} |
1922 |
|
|
} |
1923 |
|
|
} |
1924 |
|
|
} |
1925 |
|
|
return stat; |
1926 |
|
|
} |
1927 |
|
|
|
1928 |
|
|
/** |
1929 |
dl |
1.12 |
* Analog of tryHelpStealer for CountedCompleters. Tries to steal |
1930 |
jsr166 |
1.17 |
* and run tasks within the target's computation. |
1931 |
dl |
1.12 |
* |
1932 |
|
|
* @param task the task to join |
1933 |
|
|
* @param mode if shared, exit upon completing any task |
1934 |
|
|
* if all workers are active |
1935 |
dl |
1.1 |
*/ |
1936 |
dl |
1.12 |
private int helpComplete(ForkJoinTask<?> task, int mode) { |
1937 |
dl |
1.18 |
WorkQueue[] ws; WorkQueue q; int m, n, s, u; |
1938 |
dl |
1.12 |
if (task != null && (ws = workQueues) != null && |
1939 |
|
|
(m = ws.length - 1) >= 0) { |
1940 |
|
|
for (int j = 1, origin = j;;) { |
1941 |
|
|
if ((s = task.status) < 0) |
1942 |
|
|
return s; |
1943 |
|
|
if ((q = ws[j & m]) != null && q.pollAndExecCC(task)) { |
1944 |
|
|
origin = j; |
1945 |
dl |
1.18 |
if (mode == SHARED_QUEUE && |
1946 |
|
|
((u = (int)(ctl >>> 32)) >= 0 || (u >> UAC_SHIFT) >= 0)) |
1947 |
dl |
1.12 |
break; |
1948 |
|
|
} |
1949 |
|
|
else if ((j = (j + 2) & m) == origin) |
1950 |
dl |
1.1 |
break; |
1951 |
|
|
} |
1952 |
|
|
} |
1953 |
dl |
1.12 |
return 0; |
1954 |
dl |
1.1 |
} |
1955 |
|
|
|
1956 |
|
|
/** |
1957 |
|
|
* Tries to decrement active count (sometimes implicitly) and |
1958 |
|
|
* possibly release or create a compensating worker in preparation |
1959 |
|
|
* for blocking. Fails on contention or termination. Otherwise, |
1960 |
dl |
1.12 |
* adds a new thread if no idle workers are available and pool |
1961 |
|
|
* may become starved. |
1962 |
dl |
1.1 |
*/ |
1963 |
dl |
1.12 |
final boolean tryCompensate() { |
1964 |
dl |
1.18 |
int pc = config & SMASK, e, i, tc; long c; |
1965 |
dl |
1.12 |
WorkQueue[] ws; WorkQueue w; Thread p; |
1966 |
dl |
1.18 |
if ((ws = workQueues) != null && (e = (int)(c = ctl)) >= 0) { |
1967 |
dl |
1.12 |
if (e != 0 && (i = e & SMASK) < ws.length && |
1968 |
|
|
(w = ws[i]) != null && w.eventCount == (e | INT_SIGN)) { |
1969 |
|
|
long nc = ((long)(w.nextWait & E_MASK) | |
1970 |
|
|
(c & (AC_MASK|TC_MASK))); |
1971 |
|
|
if (U.compareAndSwapLong(this, CTL, c, nc)) { |
1972 |
|
|
w.eventCount = (e + E_SEQ) & E_MASK; |
1973 |
|
|
if ((p = w.parker) != null) |
1974 |
|
|
U.unpark(p); |
1975 |
|
|
return true; // replace with idle worker |
1976 |
dl |
1.1 |
} |
1977 |
|
|
} |
1978 |
dl |
1.18 |
else if ((tc = (short)(c >>> TC_SHIFT)) >= 0 && |
1979 |
|
|
(int)(c >> AC_SHIFT) + pc > 1) { |
1980 |
dl |
1.12 |
long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK); |
1981 |
|
|
if (U.compareAndSwapLong(this, CTL, c, nc)) |
1982 |
dl |
1.18 |
return true; // no compensation |
1983 |
dl |
1.12 |
} |
1984 |
dl |
1.18 |
else if (tc + pc < MAX_CAP) { |
1985 |
dl |
1.12 |
long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK); |
1986 |
|
|
if (U.compareAndSwapLong(this, CTL, c, nc)) { |
1987 |
dl |
1.21 |
ForkJoinWorkerThreadFactory fac; |
1988 |
|
|
Throwable ex = null; |
1989 |
|
|
ForkJoinWorkerThread wt = null; |
1990 |
|
|
try { |
1991 |
|
|
if ((fac = factory) != null && |
1992 |
|
|
(wt = fac.newThread(this)) != null) { |
1993 |
|
|
wt.start(); |
1994 |
|
|
return true; |
1995 |
|
|
} |
1996 |
|
|
} catch (Throwable rex) { |
1997 |
|
|
ex = rex; |
1998 |
|
|
} |
1999 |
|
|
deregisterWorker(wt, ex); // clean up and return false |
2000 |
dl |
1.1 |
} |
2001 |
|
|
} |
2002 |
|
|
} |
2003 |
|
|
return false; |
2004 |
|
|
} |
2005 |
|
|
|
2006 |
|
|
/** |
2007 |
|
|
* Helps and/or blocks until the given task is done. |
2008 |
|
|
* |
2009 |
|
|
* @param joiner the joining worker |
2010 |
|
|
* @param task the task |
2011 |
|
|
* @return task status on exit |
2012 |
|
|
*/ |
2013 |
|
|
final int awaitJoin(WorkQueue joiner, ForkJoinTask<?> task) { |
2014 |
dl |
1.12 |
int s = 0; |
2015 |
|
|
if (joiner != null && task != null && (s = task.status) >= 0) { |
2016 |
dl |
1.1 |
ForkJoinTask<?> prevJoin = joiner.currentJoin; |
2017 |
|
|
joiner.currentJoin = task; |
2018 |
dl |
1.21 |
do {} while ((s = task.status) >= 0 && !joiner.isEmpty() && |
2019 |
dl |
1.12 |
joiner.tryRemoveAndExec(task)); // process local tasks |
2020 |
dl |
1.21 |
if (s >= 0 && (s = task.status) >= 0) { |
2021 |
dl |
1.26 |
helpSignal(task, joiner.poolIndex); |
2022 |
dl |
1.21 |
if ((s = task.status) >= 0 && |
2023 |
|
|
(task instanceof CountedCompleter)) |
2024 |
|
|
s = helpComplete(task, LIFO_QUEUE); |
2025 |
|
|
} |
2026 |
dl |
1.12 |
while (s >= 0 && (s = task.status) >= 0) { |
2027 |
dl |
1.21 |
if ((!joiner.isEmpty() || // try helping |
2028 |
dl |
1.12 |
(s = tryHelpStealer(joiner, task)) == 0) && |
2029 |
dl |
1.18 |
(s = task.status) >= 0) { |
2030 |
dl |
1.26 |
helpSignal(task, joiner.poolIndex); |
2031 |
dl |
1.21 |
if ((s = task.status) >= 0 && tryCompensate()) { |
2032 |
dl |
1.18 |
if (task.trySetSignal() && (s = task.status) >= 0) { |
2033 |
|
|
synchronized (task) { |
2034 |
|
|
if (task.status >= 0) { |
2035 |
|
|
try { // see ForkJoinTask |
2036 |
|
|
task.wait(); // for explanation |
2037 |
|
|
} catch (InterruptedException ie) { |
2038 |
|
|
} |
2039 |
dl |
1.1 |
} |
2040 |
dl |
1.18 |
else |
2041 |
|
|
task.notifyAll(); |
2042 |
dl |
1.1 |
} |
2043 |
|
|
} |
2044 |
dl |
1.18 |
long c; // re-activate |
2045 |
|
|
do {} while (!U.compareAndSwapLong |
2046 |
|
|
(this, CTL, c = ctl, c + AC_UNIT)); |
2047 |
dl |
1.1 |
} |
2048 |
|
|
} |
2049 |
|
|
} |
2050 |
dl |
1.12 |
joiner.currentJoin = prevJoin; |
2051 |
dl |
1.1 |
} |
2052 |
|
|
return s; |
2053 |
|
|
} |
2054 |
|
|
|
2055 |
|
|
/** |
2056 |
|
|
* Stripped-down variant of awaitJoin used by timed joins. Tries |
2057 |
|
|
* to help join only while there is continuous progress. (Caller |
2058 |
|
|
* will then enter a timed wait.) |
2059 |
|
|
* |
2060 |
|
|
* @param joiner the joining worker |
2061 |
|
|
* @param task the task |
2062 |
|
|
*/ |
2063 |
dl |
1.12 |
final void helpJoinOnce(WorkQueue joiner, ForkJoinTask<?> task) { |
2064 |
dl |
1.1 |
int s; |
2065 |
dl |
1.12 |
if (joiner != null && task != null && (s = task.status) >= 0) { |
2066 |
|
|
ForkJoinTask<?> prevJoin = joiner.currentJoin; |
2067 |
|
|
joiner.currentJoin = task; |
2068 |
dl |
1.21 |
do {} while ((s = task.status) >= 0 && !joiner.isEmpty() && |
2069 |
dl |
1.12 |
joiner.tryRemoveAndExec(task)); |
2070 |
dl |
1.21 |
if (s >= 0 && (s = task.status) >= 0) { |
2071 |
dl |
1.26 |
helpSignal(task, joiner.poolIndex); |
2072 |
dl |
1.21 |
if ((s = task.status) >= 0 && |
2073 |
|
|
(task instanceof CountedCompleter)) |
2074 |
|
|
s = helpComplete(task, LIFO_QUEUE); |
2075 |
|
|
} |
2076 |
|
|
if (s >= 0 && joiner.isEmpty()) { |
2077 |
dl |
1.12 |
do {} while (task.status >= 0 && |
2078 |
|
|
tryHelpStealer(joiner, task) > 0); |
2079 |
|
|
} |
2080 |
|
|
joiner.currentJoin = prevJoin; |
2081 |
|
|
} |
2082 |
dl |
1.1 |
} |
2083 |
|
|
|
2084 |
|
|
/** |
2085 |
|
|
* Returns a (probably) non-empty steal queue, if one is found |
2086 |
dl |
1.37 |
* during a scan, else null. This method must be retried by |
2087 |
|
|
* caller if, by the time it tries to use the queue, it is empty. |
2088 |
dl |
1.12 |
* @param r a (random) seed for scanning |
2089 |
dl |
1.1 |
*/ |
2090 |
dl |
1.12 |
private WorkQueue findNonEmptyStealQueue(int r) { |
2091 |
dl |
1.37 |
for (;;) { |
2092 |
|
|
int ps = plock, m; WorkQueue[] ws; WorkQueue q; |
2093 |
|
|
if ((ws = workQueues) != null && (m = ws.length - 1) >= 0) { |
2094 |
|
|
for (int j = (m + 1) << 2; j >= 0; --j) { |
2095 |
|
|
if ((q = ws[(((r + j) << 1) | 1) & m]) != null && |
2096 |
|
|
q.base - q.top < 0) |
2097 |
|
|
return q; |
2098 |
dl |
1.1 |
} |
2099 |
|
|
} |
2100 |
dl |
1.37 |
if (plock == ps) |
2101 |
|
|
return null; |
2102 |
dl |
1.1 |
} |
2103 |
|
|
} |
2104 |
|
|
|
2105 |
|
|
/** |
2106 |
|
|
* Runs tasks until {@code isQuiescent()}. We piggyback on |
2107 |
|
|
* active count ctl maintenance, but rather than blocking |
2108 |
|
|
* when tasks cannot be found, we rescan until all others cannot |
2109 |
|
|
* find tasks either. |
2110 |
|
|
*/ |
2111 |
|
|
final void helpQuiescePool(WorkQueue w) { |
2112 |
|
|
for (boolean active = true;;) { |
2113 |
dl |
1.37 |
long c; WorkQueue q; ForkJoinTask<?> t; int b; |
2114 |
|
|
while ((t = w.nextLocalTask()) != null) { |
2115 |
|
|
if (w.base - w.top < 0) |
2116 |
|
|
signalWork(w); |
2117 |
|
|
t.doExec(); |
2118 |
|
|
} |
2119 |
|
|
if ((q = findNonEmptyStealQueue(w.nextSeed())) != null) { |
2120 |
dl |
1.1 |
if (!active) { // re-establish active count |
2121 |
|
|
active = true; |
2122 |
|
|
do {} while (!U.compareAndSwapLong |
2123 |
|
|
(this, CTL, c = ctl, c + AC_UNIT)); |
2124 |
|
|
} |
2125 |
dl |
1.36 |
if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null) { |
2126 |
|
|
if (q.base - q.top < 0) |
2127 |
|
|
signalWork(q); |
2128 |
dl |
1.1 |
w.runSubtask(t); |
2129 |
dl |
1.36 |
} |
2130 |
dl |
1.1 |
} |
2131 |
dl |
1.37 |
else if (active) { // decrement active count without queuing |
2132 |
|
|
long nc = (c = ctl) - AC_UNIT; |
2133 |
|
|
if ((int)(nc >> AC_SHIFT) + (config & SMASK) == 0) |
2134 |
|
|
return; // bypass decrement-then-increment |
2135 |
|
|
if (U.compareAndSwapLong(this, CTL, c, nc)) |
2136 |
dl |
1.1 |
active = false; |
2137 |
|
|
} |
2138 |
dl |
1.37 |
else if ((int)((c = ctl) >> AC_SHIFT) + (config & SMASK) == 0 && |
2139 |
|
|
U.compareAndSwapLong(this, CTL, c, c + AC_UNIT)) |
2140 |
|
|
return; |
2141 |
dl |
1.1 |
} |
2142 |
|
|
} |
2143 |
|
|
|
2144 |
|
|
/** |
2145 |
|
|
* Gets and removes a local or stolen task for the given worker. |
2146 |
|
|
* |
2147 |
|
|
* @return a task, if available |
2148 |
|
|
*/ |
2149 |
|
|
final ForkJoinTask<?> nextTaskFor(WorkQueue w) { |
2150 |
|
|
for (ForkJoinTask<?> t;;) { |
2151 |
|
|
WorkQueue q; int b; |
2152 |
|
|
if ((t = w.nextLocalTask()) != null) |
2153 |
|
|
return t; |
2154 |
dl |
1.12 |
if ((q = findNonEmptyStealQueue(w.nextSeed())) == null) |
2155 |
dl |
1.1 |
return null; |
2156 |
dl |
1.36 |
if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null) { |
2157 |
|
|
if (q.base - q.top < 0) |
2158 |
|
|
signalWork(q); |
2159 |
dl |
1.1 |
return t; |
2160 |
dl |
1.36 |
} |
2161 |
dl |
1.1 |
} |
2162 |
|
|
} |
2163 |
|
|
|
2164 |
|
|
/** |
2165 |
dl |
1.12 |
* Returns a cheap heuristic guide for task partitioning when |
2166 |
|
|
* programmers, frameworks, tools, or languages have little or no |
2167 |
|
|
* idea about task granularity. In essence by offering this |
2168 |
|
|
* method, we ask users only about tradeoffs in overhead vs |
2169 |
|
|
* expected throughput and its variance, rather than how finely to |
2170 |
|
|
* partition tasks. |
2171 |
|
|
* |
2172 |
|
|
* In a steady state strict (tree-structured) computation, each |
2173 |
|
|
* thread makes available for stealing enough tasks for other |
2174 |
|
|
* threads to remain active. Inductively, if all threads play by |
2175 |
|
|
* the same rules, each thread should make available only a |
2176 |
|
|
* constant number of tasks. |
2177 |
|
|
* |
2178 |
|
|
* The minimum useful constant is just 1. But using a value of 1 |
2179 |
|
|
* would require immediate replenishment upon each steal to |
2180 |
|
|
* maintain enough tasks, which is infeasible. Further, |
2181 |
|
|
* partitionings/granularities of offered tasks should minimize |
2182 |
|
|
* steal rates, which in general means that threads nearer the top |
2183 |
|
|
* of computation tree should generate more than those nearer the |
2184 |
|
|
* bottom. In perfect steady state, each thread is at |
2185 |
|
|
* approximately the same level of computation tree. However, |
2186 |
|
|
* producing extra tasks amortizes the uncertainty of progress and |
2187 |
|
|
* diffusion assumptions. |
2188 |
|
|
* |
2189 |
|
|
* So, users will want to use values larger, but not much larger |
2190 |
|
|
* than 1 to both smooth over transient shortages and hedge |
2191 |
|
|
* against uneven progress; as traded off against the cost of |
2192 |
|
|
* extra task overhead. We leave the user to pick a threshold |
2193 |
|
|
* value to compare with the results of this call to guide |
2194 |
|
|
* decisions, but recommend values such as 3. |
2195 |
|
|
* |
2196 |
|
|
* When all threads are active, it is on average OK to estimate |
2197 |
|
|
* surplus strictly locally. In steady-state, if one thread is |
2198 |
|
|
* maintaining say 2 surplus tasks, then so are others. So we can |
2199 |
|
|
* just use estimated queue length. However, this strategy alone |
2200 |
|
|
* leads to serious mis-estimates in some non-steady-state |
2201 |
|
|
* conditions (ramp-up, ramp-down, other stalls). We can detect |
2202 |
|
|
* many of these by further considering the number of "idle" |
2203 |
|
|
* threads, that are known to have zero queued tasks, so |
2204 |
|
|
* compensate by a factor of (#idle/#active) threads. |
2205 |
|
|
* |
2206 |
|
|
* Note: The approximation of #busy workers as #active workers is |
2207 |
|
|
* not very good under current signalling scheme, and should be |
2208 |
|
|
* improved. |
2209 |
|
|
*/ |
2210 |
|
|
static int getSurplusQueuedTaskCount() { |
2211 |
|
|
Thread t; ForkJoinWorkerThread wt; ForkJoinPool pool; WorkQueue q; |
2212 |
|
|
if (((t = Thread.currentThread()) instanceof ForkJoinWorkerThread)) { |
2213 |
dl |
1.18 |
int p = (pool = (wt = (ForkJoinWorkerThread)t).pool).config & SMASK; |
2214 |
|
|
int n = (q = wt.workQueue).top - q.base; |
2215 |
dl |
1.12 |
int a = (int)(pool.ctl >> AC_SHIFT) + p; |
2216 |
dl |
1.18 |
return n - (a > (p >>>= 1) ? 0 : |
2217 |
|
|
a > (p >>>= 1) ? 1 : |
2218 |
|
|
a > (p >>>= 1) ? 2 : |
2219 |
|
|
a > (p >>>= 1) ? 4 : |
2220 |
|
|
8); |
2221 |
dl |
1.12 |
} |
2222 |
|
|
return 0; |
2223 |
dl |
1.7 |
} |
2224 |
|
|
|
2225 |
dl |
1.1 |
// Termination |
2226 |
|
|
|
2227 |
|
|
/** |
2228 |
|
|
* Possibly initiates and/or completes termination. The caller |
2229 |
|
|
* triggering termination runs three passes through workQueues: |
2230 |
|
|
* (0) Setting termination status, followed by wakeups of queued |
2231 |
|
|
* workers; (1) cancelling all tasks; (2) interrupting lagging |
2232 |
|
|
* threads (likely in external tasks, but possibly also blocked in |
2233 |
|
|
* joins). Each pass repeats previous steps because of potential |
2234 |
|
|
* lagging thread creation. |
2235 |
|
|
* |
2236 |
|
|
* @param now if true, unconditionally terminate, else only |
2237 |
|
|
* if no work and no active workers |
2238 |
|
|
* @param enable if true, enable shutdown when next possible |
2239 |
|
|
* @return true if now terminating or terminated |
2240 |
|
|
*/ |
2241 |
|
|
private boolean tryTerminate(boolean now, boolean enable) { |
2242 |
dl |
1.37 |
int ps; |
2243 |
dl |
1.40 |
if (this == common) // cannot shut down |
2244 |
dl |
1.12 |
return false; |
2245 |
dl |
1.37 |
if ((ps = plock) >= 0) { // enable by setting plock |
2246 |
|
|
if (!enable) |
2247 |
|
|
return false; |
2248 |
|
|
if ((ps & PL_LOCK) != 0 || |
2249 |
|
|
!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK)) |
2250 |
|
|
ps = acquirePlock(); |
2251 |
|
|
int nps = ((ps + PL_LOCK) & ~SHUTDOWN) | SHUTDOWN; |
2252 |
|
|
if (!U.compareAndSwapInt(this, PLOCK, ps, nps)) |
2253 |
|
|
releasePlock(nps); |
2254 |
|
|
} |
2255 |
dl |
1.1 |
for (long c;;) { |
2256 |
dl |
1.37 |
if (((c = ctl) & STOP_BIT) != 0) { // already terminating |
2257 |
dl |
1.18 |
if ((short)(c >>> TC_SHIFT) == -(config & SMASK)) { |
2258 |
jsr166 |
1.10 |
synchronized (this) { |
2259 |
dl |
1.37 |
notifyAll(); // signal when 0 workers |
2260 |
dl |
1.8 |
} |
2261 |
dl |
1.1 |
} |
2262 |
|
|
return true; |
2263 |
|
|
} |
2264 |
dl |
1.37 |
if (!now) { // check if idle & no tasks |
2265 |
|
|
WorkQueue[] ws; WorkQueue w; |
2266 |
|
|
if ((int)(c >> AC_SHIFT) != -(config & SMASK)) |
2267 |
dl |
1.1 |
return false; |
2268 |
dl |
1.37 |
if ((ws = workQueues) != null) { |
2269 |
|
|
for (int i = 0; i < ws.length; ++i) { |
2270 |
|
|
if ((w = ws[i]) != null) { |
2271 |
|
|
if (!w.isEmpty()) { // signal unprocessed tasks |
2272 |
|
|
signalWork(w); |
2273 |
|
|
return false; |
2274 |
|
|
} |
2275 |
|
|
if ((i & 1) != 0 && w.eventCount >= 0) |
2276 |
|
|
return false; // unqueued inactive worker |
2277 |
|
|
} |
2278 |
dl |
1.1 |
} |
2279 |
|
|
} |
2280 |
|
|
} |
2281 |
|
|
if (U.compareAndSwapLong(this, CTL, c, c | STOP_BIT)) { |
2282 |
|
|
for (int pass = 0; pass < 3; ++pass) { |
2283 |
dl |
1.37 |
WorkQueue[] ws; WorkQueue w; Thread wt; |
2284 |
|
|
if ((ws = workQueues) != null) { |
2285 |
dl |
1.1 |
int n = ws.length; |
2286 |
|
|
for (int i = 0; i < n; ++i) { |
2287 |
|
|
if ((w = ws[i]) != null) { |
2288 |
dl |
1.12 |
w.qlock = -1; |
2289 |
dl |
1.1 |
if (pass > 0) { |
2290 |
|
|
w.cancelAll(); |
2291 |
dl |
1.32 |
if (pass > 1 && (wt = w.owner) != null) { |
2292 |
|
|
if (!wt.isInterrupted()) { |
2293 |
|
|
try { |
2294 |
|
|
wt.interrupt(); |
2295 |
dl |
1.37 |
} catch (Throwable ignore) { |
2296 |
dl |
1.32 |
} |
2297 |
|
|
} |
2298 |
|
|
U.unpark(wt); |
2299 |
|
|
} |
2300 |
dl |
1.1 |
} |
2301 |
|
|
} |
2302 |
|
|
} |
2303 |
|
|
// Wake up workers parked on event queue |
2304 |
|
|
int i, e; long cc; Thread p; |
2305 |
|
|
while ((e = (int)(cc = ctl) & E_MASK) != 0 && |
2306 |
dl |
1.37 |
(i = e & SMASK) < n && i >= 0 && |
2307 |
dl |
1.1 |
(w = ws[i]) != null) { |
2308 |
|
|
long nc = ((long)(w.nextWait & E_MASK) | |
2309 |
|
|
((cc + AC_UNIT) & AC_MASK) | |
2310 |
|
|
(cc & (TC_MASK|STOP_BIT))); |
2311 |
|
|
if (w.eventCount == (e | INT_SIGN) && |
2312 |
|
|
U.compareAndSwapLong(this, CTL, cc, nc)) { |
2313 |
|
|
w.eventCount = (e + E_SEQ) & E_MASK; |
2314 |
dl |
1.12 |
w.qlock = -1; |
2315 |
dl |
1.1 |
if ((p = w.parker) != null) |
2316 |
|
|
U.unpark(p); |
2317 |
|
|
} |
2318 |
|
|
} |
2319 |
|
|
} |
2320 |
|
|
} |
2321 |
|
|
} |
2322 |
|
|
} |
2323 |
|
|
} |
2324 |
|
|
|
2325 |
dl |
1.12 |
// external operations on common pool |
2326 |
|
|
|
2327 |
|
|
/** |
2328 |
|
|
* Returns common pool queue for a thread that has submitted at |
2329 |
|
|
* least one task. |
2330 |
|
|
*/ |
2331 |
|
|
static WorkQueue commonSubmitterQueue() { |
2332 |
|
|
ForkJoinPool p; WorkQueue[] ws; int m; Submitter z; |
2333 |
|
|
return ((z = submitters.get()) != null && |
2334 |
dl |
1.40 |
(p = common) != null && |
2335 |
dl |
1.12 |
(ws = p.workQueues) != null && |
2336 |
|
|
(m = ws.length - 1) >= 0) ? |
2337 |
|
|
ws[m & z.seed & SQMASK] : null; |
2338 |
|
|
} |
2339 |
|
|
|
2340 |
|
|
/** |
2341 |
|
|
* Tries to pop the given task from submitter's queue in common pool. |
2342 |
|
|
*/ |
2343 |
|
|
static boolean tryExternalUnpush(ForkJoinTask<?> t) { |
2344 |
|
|
ForkJoinPool p; WorkQueue[] ws; WorkQueue q; Submitter z; |
2345 |
dl |
1.21 |
ForkJoinTask<?>[] a; int m, s; |
2346 |
|
|
if (t != null && |
2347 |
|
|
(z = submitters.get()) != null && |
2348 |
dl |
1.40 |
(p = common) != null && |
2349 |
dl |
1.12 |
(ws = p.workQueues) != null && |
2350 |
|
|
(m = ws.length - 1) >= 0 && |
2351 |
|
|
(q = ws[m & z.seed & SQMASK]) != null && |
2352 |
|
|
(s = q.top) != q.base && |
2353 |
dl |
1.21 |
(a = q.array) != null) { |
2354 |
|
|
long j = (((a.length - 1) & (s - 1)) << ASHIFT) + ABASE; |
2355 |
|
|
if (U.getObject(a, j) == t && |
2356 |
|
|
U.compareAndSwapInt(q, QLOCK, 0, 1)) { |
2357 |
|
|
if (q.array == a && q.top == s && // recheck |
2358 |
|
|
U.compareAndSwapObject(a, j, t, null)) { |
2359 |
|
|
q.top = s - 1; |
2360 |
|
|
q.qlock = 0; |
2361 |
|
|
return true; |
2362 |
|
|
} |
2363 |
dl |
1.12 |
q.qlock = 0; |
2364 |
|
|
} |
2365 |
|
|
} |
2366 |
|
|
return false; |
2367 |
|
|
} |
2368 |
|
|
|
2369 |
|
|
/** |
2370 |
|
|
* Tries to pop and run local tasks within the same computation |
2371 |
|
|
* as the given root. On failure, tries to help complete from |
2372 |
|
|
* other queues via helpComplete. |
2373 |
|
|
*/ |
2374 |
|
|
private void externalHelpComplete(WorkQueue q, ForkJoinTask<?> root) { |
2375 |
|
|
ForkJoinTask<?>[] a; int m; |
2376 |
|
|
if (q != null && (a = q.array) != null && (m = (a.length - 1)) >= 0 && |
2377 |
|
|
root != null && root.status >= 0) { |
2378 |
|
|
for (;;) { |
2379 |
dl |
1.18 |
int s, u; Object o; CountedCompleter<?> task = null; |
2380 |
dl |
1.12 |
if ((s = q.top) - q.base > 0) { |
2381 |
|
|
long j = ((m & (s - 1)) << ASHIFT) + ABASE; |
2382 |
|
|
if ((o = U.getObject(a, j)) != null && |
2383 |
|
|
(o instanceof CountedCompleter)) { |
2384 |
|
|
CountedCompleter<?> t = (CountedCompleter<?>)o, r = t; |
2385 |
|
|
do { |
2386 |
|
|
if (r == root) { |
2387 |
|
|
if (U.compareAndSwapInt(q, QLOCK, 0, 1)) { |
2388 |
|
|
if (q.array == a && q.top == s && |
2389 |
|
|
U.compareAndSwapObject(a, j, t, null)) { |
2390 |
|
|
q.top = s - 1; |
2391 |
|
|
task = t; |
2392 |
|
|
} |
2393 |
|
|
q.qlock = 0; |
2394 |
|
|
} |
2395 |
|
|
break; |
2396 |
|
|
} |
2397 |
jsr166 |
1.13 |
} while ((r = r.completer) != null); |
2398 |
dl |
1.12 |
} |
2399 |
|
|
} |
2400 |
|
|
if (task != null) |
2401 |
|
|
task.doExec(); |
2402 |
dl |
1.18 |
if (root.status < 0 || |
2403 |
|
|
(u = (int)(ctl >>> 32)) >= 0 || (u >> UAC_SHIFT) >= 0) |
2404 |
dl |
1.12 |
break; |
2405 |
|
|
if (task == null) { |
2406 |
dl |
1.26 |
helpSignal(root, q.poolIndex); |
2407 |
dl |
1.21 |
if (root.status >= 0) |
2408 |
dl |
1.12 |
helpComplete(root, SHARED_QUEUE); |
2409 |
|
|
break; |
2410 |
|
|
} |
2411 |
|
|
} |
2412 |
|
|
} |
2413 |
|
|
} |
2414 |
|
|
|
2415 |
|
|
/** |
2416 |
|
|
* Tries to help execute or signal availability of the given task |
2417 |
|
|
* from submitter's queue in common pool. |
2418 |
|
|
*/ |
2419 |
|
|
static void externalHelpJoin(ForkJoinTask<?> t) { |
2420 |
|
|
// Some hard-to-avoid overlap with tryExternalUnpush |
2421 |
|
|
ForkJoinPool p; WorkQueue[] ws; WorkQueue q, w; Submitter z; |
2422 |
dl |
1.21 |
ForkJoinTask<?>[] a; int m, s, n; |
2423 |
dl |
1.18 |
if (t != null && |
2424 |
dl |
1.12 |
(z = submitters.get()) != null && |
2425 |
dl |
1.40 |
(p = common) != null && |
2426 |
dl |
1.12 |
(ws = p.workQueues) != null && |
2427 |
|
|
(m = ws.length - 1) >= 0 && |
2428 |
|
|
(q = ws[m & z.seed & SQMASK]) != null && |
2429 |
dl |
1.21 |
(a = q.array) != null) { |
2430 |
|
|
int am = a.length - 1; |
2431 |
|
|
if ((s = q.top) != q.base) { |
2432 |
|
|
long j = ((am & (s - 1)) << ASHIFT) + ABASE; |
2433 |
|
|
if (U.getObject(a, j) == t && |
2434 |
|
|
U.compareAndSwapInt(q, QLOCK, 0, 1)) { |
2435 |
|
|
if (q.array == a && q.top == s && |
2436 |
|
|
U.compareAndSwapObject(a, j, t, null)) { |
2437 |
|
|
q.top = s - 1; |
2438 |
|
|
q.qlock = 0; |
2439 |
|
|
t.doExec(); |
2440 |
|
|
} |
2441 |
|
|
else |
2442 |
|
|
q.qlock = 0; |
2443 |
dl |
1.12 |
} |
2444 |
|
|
} |
2445 |
|
|
if (t.status >= 0) { |
2446 |
|
|
if (t instanceof CountedCompleter) |
2447 |
|
|
p.externalHelpComplete(q, t); |
2448 |
|
|
else |
2449 |
dl |
1.26 |
p.helpSignal(t, q.poolIndex); |
2450 |
dl |
1.12 |
} |
2451 |
|
|
} |
2452 |
|
|
} |
2453 |
|
|
|
2454 |
dl |
1.1 |
// Exported methods |
2455 |
|
|
|
2456 |
|
|
// Constructors |
2457 |
|
|
|
2458 |
|
|
/** |
2459 |
|
|
* Creates a {@code ForkJoinPool} with parallelism equal to {@link |
2460 |
|
|
* java.lang.Runtime#availableProcessors}, using the {@linkplain |
2461 |
|
|
* #defaultForkJoinWorkerThreadFactory default thread factory}, |
2462 |
|
|
* no UncaughtExceptionHandler, and non-async LIFO processing mode. |
2463 |
|
|
* |
2464 |
|
|
* @throws SecurityException if a security manager exists and |
2465 |
|
|
* the caller is not permitted to modify threads |
2466 |
|
|
* because it does not hold {@link |
2467 |
|
|
* java.lang.RuntimePermission}{@code ("modifyThread")} |
2468 |
|
|
*/ |
2469 |
|
|
public ForkJoinPool() { |
2470 |
|
|
this(Runtime.getRuntime().availableProcessors(), |
2471 |
|
|
defaultForkJoinWorkerThreadFactory, null, false); |
2472 |
|
|
} |
2473 |
|
|
|
2474 |
|
|
/** |
2475 |
|
|
* Creates a {@code ForkJoinPool} with the indicated parallelism |
2476 |
|
|
* level, the {@linkplain |
2477 |
|
|
* #defaultForkJoinWorkerThreadFactory default thread factory}, |
2478 |
|
|
* no UncaughtExceptionHandler, and non-async LIFO processing mode. |
2479 |
|
|
* |
2480 |
|
|
* @param parallelism the parallelism level |
2481 |
|
|
* @throws IllegalArgumentException if parallelism less than or |
2482 |
|
|
* equal to zero, or greater than implementation limit |
2483 |
|
|
* @throws SecurityException if a security manager exists and |
2484 |
|
|
* the caller is not permitted to modify threads |
2485 |
|
|
* because it does not hold {@link |
2486 |
|
|
* java.lang.RuntimePermission}{@code ("modifyThread")} |
2487 |
|
|
*/ |
2488 |
|
|
public ForkJoinPool(int parallelism) { |
2489 |
|
|
this(parallelism, defaultForkJoinWorkerThreadFactory, null, false); |
2490 |
|
|
} |
2491 |
|
|
|
2492 |
|
|
/** |
2493 |
|
|
* Creates a {@code ForkJoinPool} with the given parameters. |
2494 |
|
|
* |
2495 |
|
|
* @param parallelism the parallelism level. For default value, |
2496 |
|
|
* use {@link java.lang.Runtime#availableProcessors}. |
2497 |
|
|
* @param factory the factory for creating new threads. For default value, |
2498 |
|
|
* use {@link #defaultForkJoinWorkerThreadFactory}. |
2499 |
|
|
* @param handler the handler for internal worker threads that |
2500 |
|
|
* terminate due to unrecoverable errors encountered while executing |
2501 |
|
|
* tasks. For default value, use {@code null}. |
2502 |
|
|
* @param asyncMode if true, |
2503 |
|
|
* establishes local first-in-first-out scheduling mode for forked |
2504 |
|
|
* tasks that are never joined. This mode may be more appropriate |
2505 |
|
|
* than default locally stack-based mode in applications in which |
2506 |
|
|
* worker threads only process event-style asynchronous tasks. |
2507 |
|
|
* For default value, use {@code false}. |
2508 |
|
|
* @throws IllegalArgumentException if parallelism less than or |
2509 |
|
|
* equal to zero, or greater than implementation limit |
2510 |
|
|
* @throws NullPointerException if the factory is null |
2511 |
|
|
* @throws SecurityException if a security manager exists and |
2512 |
|
|
* the caller is not permitted to modify threads |
2513 |
|
|
* because it does not hold {@link |
2514 |
|
|
* java.lang.RuntimePermission}{@code ("modifyThread")} |
2515 |
|
|
*/ |
2516 |
|
|
public ForkJoinPool(int parallelism, |
2517 |
|
|
ForkJoinWorkerThreadFactory factory, |
2518 |
|
|
Thread.UncaughtExceptionHandler handler, |
2519 |
|
|
boolean asyncMode) { |
2520 |
|
|
checkPermission(); |
2521 |
|
|
if (factory == null) |
2522 |
|
|
throw new NullPointerException(); |
2523 |
|
|
if (parallelism <= 0 || parallelism > MAX_CAP) |
2524 |
|
|
throw new IllegalArgumentException(); |
2525 |
|
|
this.factory = factory; |
2526 |
|
|
this.ueh = handler; |
2527 |
jsr166 |
1.19 |
this.config = parallelism | (asyncMode ? (FIFO_QUEUE << 16) : 0); |
2528 |
dl |
1.1 |
long np = (long)(-parallelism); // offset ctl counts |
2529 |
|
|
this.ctl = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK); |
2530 |
dl |
1.12 |
int pn = nextPoolId(); |
2531 |
dl |
1.1 |
StringBuilder sb = new StringBuilder("ForkJoinPool-"); |
2532 |
|
|
sb.append(Integer.toString(pn)); |
2533 |
|
|
sb.append("-worker-"); |
2534 |
|
|
this.workerNamePrefix = sb.toString(); |
2535 |
|
|
} |
2536 |
|
|
|
2537 |
dl |
1.7 |
/** |
2538 |
dl |
1.8 |
* Constructor for common pool, suitable only for static initialization. |
2539 |
|
|
* Basically the same as above, but uses smallest possible initial footprint. |
2540 |
|
|
*/ |
2541 |
dl |
1.12 |
ForkJoinPool(int parallelism, long ctl, |
2542 |
dl |
1.8 |
ForkJoinWorkerThreadFactory factory, |
2543 |
|
|
Thread.UncaughtExceptionHandler handler) { |
2544 |
dl |
1.18 |
this.config = parallelism; |
2545 |
dl |
1.12 |
this.ctl = ctl; |
2546 |
dl |
1.8 |
this.factory = factory; |
2547 |
|
|
this.ueh = handler; |
2548 |
|
|
this.workerNamePrefix = "ForkJoinPool.commonPool-worker-"; |
2549 |
|
|
} |
2550 |
|
|
|
2551 |
|
|
/** |
2552 |
dl |
1.34 |
* Returns the common pool instance. This pool is statically |
2553 |
dl |
1.40 |
* constructed; its run state is unaffected by attempts to {@link |
2554 |
|
|
* #shutdown} or {@link #shutdownNow}. However this pool and any |
2555 |
|
|
* ongoing processing are automatically terminated upon program |
2556 |
|
|
* {@link System#exit}. Any program that relies on asynchronous |
2557 |
|
|
* task processing to complete before program termination should |
2558 |
dl |
1.42 |
* invoke {@code commonPool().}{@link #awaitQuiescence}, before |
2559 |
|
|
* exit. |
2560 |
dl |
1.7 |
* |
2561 |
|
|
* @return the common pool instance |
2562 |
|
|
*/ |
2563 |
|
|
public static ForkJoinPool commonPool() { |
2564 |
dl |
1.40 |
// assert common != null : "static init error"; |
2565 |
|
|
return common; |
2566 |
dl |
1.7 |
} |
2567 |
|
|
|
2568 |
dl |
1.1 |
// Execution methods |
2569 |
|
|
|
2570 |
|
|
/** |
2571 |
|
|
* Performs the given task, returning its result upon completion. |
2572 |
|
|
* If the computation encounters an unchecked Exception or Error, |
2573 |
|
|
* it is rethrown as the outcome of this invocation. Rethrown |
2574 |
|
|
* exceptions behave in the same way as regular exceptions, but, |
2575 |
|
|
* when possible, contain stack traces (as displayed for example |
2576 |
|
|
* using {@code ex.printStackTrace()}) of both the current thread |
2577 |
|
|
* as well as the thread actually encountering the exception; |
2578 |
|
|
* minimally only the latter. |
2579 |
|
|
* |
2580 |
|
|
* @param task the task |
2581 |
|
|
* @return the task's result |
2582 |
|
|
* @throws NullPointerException if the task is null |
2583 |
|
|
* @throws RejectedExecutionException if the task cannot be |
2584 |
|
|
* scheduled for execution |
2585 |
|
|
*/ |
2586 |
|
|
public <T> T invoke(ForkJoinTask<T> task) { |
2587 |
|
|
if (task == null) |
2588 |
|
|
throw new NullPointerException(); |
2589 |
dl |
1.12 |
externalPush(task); |
2590 |
dl |
1.1 |
return task.join(); |
2591 |
|
|
} |
2592 |
|
|
|
2593 |
|
|
/** |
2594 |
|
|
* Arranges for (asynchronous) execution of the given task. |
2595 |
|
|
* |
2596 |
|
|
* @param task the task |
2597 |
|
|
* @throws NullPointerException if the task is null |
2598 |
|
|
* @throws RejectedExecutionException if the task cannot be |
2599 |
|
|
* scheduled for execution |
2600 |
|
|
*/ |
2601 |
|
|
public void execute(ForkJoinTask<?> task) { |
2602 |
|
|
if (task == null) |
2603 |
|
|
throw new NullPointerException(); |
2604 |
dl |
1.12 |
externalPush(task); |
2605 |
dl |
1.1 |
} |
2606 |
|
|
|
2607 |
|
|
// AbstractExecutorService methods |
2608 |
|
|
|
2609 |
|
|
/** |
2610 |
|
|
* @throws NullPointerException if the task is null |
2611 |
|
|
* @throws RejectedExecutionException if the task cannot be |
2612 |
|
|
* scheduled for execution |
2613 |
|
|
*/ |
2614 |
|
|
public void execute(Runnable task) { |
2615 |
|
|
if (task == null) |
2616 |
|
|
throw new NullPointerException(); |
2617 |
|
|
ForkJoinTask<?> job; |
2618 |
|
|
if (task instanceof ForkJoinTask<?>) // avoid re-wrap |
2619 |
|
|
job = (ForkJoinTask<?>) task; |
2620 |
|
|
else |
2621 |
|
|
job = new ForkJoinTask.AdaptedRunnableAction(task); |
2622 |
dl |
1.12 |
externalPush(job); |
2623 |
dl |
1.1 |
} |
2624 |
|
|
|
2625 |
|
|
/** |
2626 |
|
|
* Submits a ForkJoinTask for execution. |
2627 |
|
|
* |
2628 |
|
|
* @param task the task to submit |
2629 |
|
|
* @return the task |
2630 |
|
|
* @throws NullPointerException if the task is null |
2631 |
|
|
* @throws RejectedExecutionException if the task cannot be |
2632 |
|
|
* scheduled for execution |
2633 |
|
|
*/ |
2634 |
|
|
public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) { |
2635 |
|
|
if (task == null) |
2636 |
|
|
throw new NullPointerException(); |
2637 |
dl |
1.12 |
externalPush(task); |
2638 |
dl |
1.1 |
return task; |
2639 |
|
|
} |
2640 |
|
|
|
2641 |
|
|
/** |
2642 |
|
|
* @throws NullPointerException if the task is null |
2643 |
|
|
* @throws RejectedExecutionException if the task cannot be |
2644 |
|
|
* scheduled for execution |
2645 |
|
|
*/ |
2646 |
|
|
public <T> ForkJoinTask<T> submit(Callable<T> task) { |
2647 |
|
|
ForkJoinTask<T> job = new ForkJoinTask.AdaptedCallable<T>(task); |
2648 |
dl |
1.12 |
externalPush(job); |
2649 |
dl |
1.1 |
return job; |
2650 |
|
|
} |
2651 |
|
|
|
2652 |
|
|
/** |
2653 |
|
|
* @throws NullPointerException if the task is null |
2654 |
|
|
* @throws RejectedExecutionException if the task cannot be |
2655 |
|
|
* scheduled for execution |
2656 |
|
|
*/ |
2657 |
|
|
public <T> ForkJoinTask<T> submit(Runnable task, T result) { |
2658 |
|
|
ForkJoinTask<T> job = new ForkJoinTask.AdaptedRunnable<T>(task, result); |
2659 |
dl |
1.12 |
externalPush(job); |
2660 |
dl |
1.1 |
return job; |
2661 |
|
|
} |
2662 |
|
|
|
2663 |
|
|
/** |
2664 |
|
|
* @throws NullPointerException if the task is null |
2665 |
|
|
* @throws RejectedExecutionException if the task cannot be |
2666 |
|
|
* scheduled for execution |
2667 |
|
|
*/ |
2668 |
|
|
public ForkJoinTask<?> submit(Runnable task) { |
2669 |
|
|
if (task == null) |
2670 |
|
|
throw new NullPointerException(); |
2671 |
|
|
ForkJoinTask<?> job; |
2672 |
|
|
if (task instanceof ForkJoinTask<?>) // avoid re-wrap |
2673 |
|
|
job = (ForkJoinTask<?>) task; |
2674 |
|
|
else |
2675 |
|
|
job = new ForkJoinTask.AdaptedRunnableAction(task); |
2676 |
dl |
1.12 |
externalPush(job); |
2677 |
dl |
1.1 |
return job; |
2678 |
|
|
} |
2679 |
|
|
|
2680 |
|
|
/** |
2681 |
|
|
* @throws NullPointerException {@inheritDoc} |
2682 |
|
|
* @throws RejectedExecutionException {@inheritDoc} |
2683 |
|
|
*/ |
2684 |
|
|
public <T> List<Future<T>> invokeAll(Collection<? extends Callable<T>> tasks) { |
2685 |
|
|
// In previous versions of this class, this method constructed |
2686 |
|
|
// a task to run ForkJoinTask.invokeAll, but now external |
2687 |
|
|
// invocation of multiple tasks is at least as efficient. |
2688 |
|
|
List<ForkJoinTask<T>> fs = new ArrayList<ForkJoinTask<T>>(tasks.size()); |
2689 |
|
|
// Workaround needed because method wasn't declared with |
2690 |
|
|
// wildcards in return type but should have been. |
2691 |
|
|
@SuppressWarnings({"unchecked", "rawtypes"}) |
2692 |
|
|
List<Future<T>> futures = (List<Future<T>>) (List) fs; |
2693 |
|
|
|
2694 |
|
|
boolean done = false; |
2695 |
|
|
try { |
2696 |
|
|
for (Callable<T> t : tasks) { |
2697 |
|
|
ForkJoinTask<T> f = new ForkJoinTask.AdaptedCallable<T>(t); |
2698 |
dl |
1.12 |
externalPush(f); |
2699 |
dl |
1.1 |
fs.add(f); |
2700 |
|
|
} |
2701 |
|
|
for (ForkJoinTask<T> f : fs) |
2702 |
|
|
f.quietlyJoin(); |
2703 |
|
|
done = true; |
2704 |
|
|
return futures; |
2705 |
|
|
} finally { |
2706 |
|
|
if (!done) |
2707 |
|
|
for (ForkJoinTask<T> f : fs) |
2708 |
|
|
f.cancel(false); |
2709 |
|
|
} |
2710 |
|
|
} |
2711 |
|
|
|
2712 |
|
|
/** |
2713 |
|
|
* Returns the factory used for constructing new workers. |
2714 |
|
|
* |
2715 |
|
|
* @return the factory used for constructing new workers |
2716 |
|
|
*/ |
2717 |
|
|
public ForkJoinWorkerThreadFactory getFactory() { |
2718 |
|
|
return factory; |
2719 |
|
|
} |
2720 |
|
|
|
2721 |
|
|
/** |
2722 |
|
|
* Returns the handler for internal worker threads that terminate |
2723 |
|
|
* due to unrecoverable errors encountered while executing tasks. |
2724 |
|
|
* |
2725 |
|
|
* @return the handler, or {@code null} if none |
2726 |
|
|
*/ |
2727 |
|
|
public Thread.UncaughtExceptionHandler getUncaughtExceptionHandler() { |
2728 |
|
|
return ueh; |
2729 |
|
|
} |
2730 |
|
|
|
2731 |
|
|
/** |
2732 |
|
|
* Returns the targeted parallelism level of this pool. |
2733 |
|
|
* |
2734 |
|
|
* @return the targeted parallelism level of this pool |
2735 |
|
|
*/ |
2736 |
|
|
public int getParallelism() { |
2737 |
dl |
1.18 |
return config & SMASK; |
2738 |
dl |
1.1 |
} |
2739 |
|
|
|
2740 |
|
|
/** |
2741 |
dl |
1.7 |
* Returns the targeted parallelism level of the common pool. |
2742 |
|
|
* |
2743 |
|
|
* @return the targeted parallelism level of the common pool |
2744 |
|
|
*/ |
2745 |
|
|
public static int getCommonPoolParallelism() { |
2746 |
dl |
1.40 |
return commonParallelism; |
2747 |
dl |
1.7 |
} |
2748 |
|
|
|
2749 |
|
|
/** |
2750 |
dl |
1.1 |
* Returns the number of worker threads that have started but not |
2751 |
|
|
* yet terminated. The result returned by this method may differ |
2752 |
|
|
* from {@link #getParallelism} when threads are created to |
2753 |
|
|
* maintain parallelism when others are cooperatively blocked. |
2754 |
|
|
* |
2755 |
|
|
* @return the number of worker threads |
2756 |
|
|
*/ |
2757 |
|
|
public int getPoolSize() { |
2758 |
dl |
1.18 |
return (config & SMASK) + (short)(ctl >>> TC_SHIFT); |
2759 |
dl |
1.1 |
} |
2760 |
|
|
|
2761 |
|
|
/** |
2762 |
|
|
* Returns {@code true} if this pool uses local first-in-first-out |
2763 |
|
|
* scheduling mode for forked tasks that are never joined. |
2764 |
|
|
* |
2765 |
|
|
* @return {@code true} if this pool uses async mode |
2766 |
|
|
*/ |
2767 |
|
|
public boolean getAsyncMode() { |
2768 |
dl |
1.18 |
return (config >>> 16) == FIFO_QUEUE; |
2769 |
dl |
1.1 |
} |
2770 |
|
|
|
2771 |
|
|
/** |
2772 |
|
|
* Returns an estimate of the number of worker threads that are |
2773 |
|
|
* not blocked waiting to join tasks or for other managed |
2774 |
|
|
* synchronization. This method may overestimate the |
2775 |
|
|
* number of running threads. |
2776 |
|
|
* |
2777 |
|
|
* @return the number of worker threads |
2778 |
|
|
*/ |
2779 |
|
|
public int getRunningThreadCount() { |
2780 |
|
|
int rc = 0; |
2781 |
|
|
WorkQueue[] ws; WorkQueue w; |
2782 |
|
|
if ((ws = workQueues) != null) { |
2783 |
|
|
for (int i = 1; i < ws.length; i += 2) { |
2784 |
|
|
if ((w = ws[i]) != null && w.isApparentlyUnblocked()) |
2785 |
|
|
++rc; |
2786 |
|
|
} |
2787 |
|
|
} |
2788 |
|
|
return rc; |
2789 |
|
|
} |
2790 |
|
|
|
2791 |
|
|
/** |
2792 |
|
|
* Returns an estimate of the number of threads that are currently |
2793 |
|
|
* stealing or executing tasks. This method may overestimate the |
2794 |
|
|
* number of active threads. |
2795 |
|
|
* |
2796 |
|
|
* @return the number of active threads |
2797 |
|
|
*/ |
2798 |
|
|
public int getActiveThreadCount() { |
2799 |
dl |
1.18 |
int r = (config & SMASK) + (int)(ctl >> AC_SHIFT); |
2800 |
dl |
1.1 |
return (r <= 0) ? 0 : r; // suppress momentarily negative values |
2801 |
|
|
} |
2802 |
|
|
|
2803 |
|
|
/** |
2804 |
|
|
* Returns {@code true} if all worker threads are currently idle. |
2805 |
|
|
* An idle worker is one that cannot obtain a task to execute |
2806 |
|
|
* because none are available to steal from other threads, and |
2807 |
|
|
* there are no pending submissions to the pool. This method is |
2808 |
|
|
* conservative; it might not return {@code true} immediately upon |
2809 |
|
|
* idleness of all threads, but will eventually become true if |
2810 |
|
|
* threads remain inactive. |
2811 |
|
|
* |
2812 |
|
|
* @return {@code true} if all threads are currently idle |
2813 |
|
|
*/ |
2814 |
|
|
public boolean isQuiescent() { |
2815 |
dl |
1.18 |
return (int)(ctl >> AC_SHIFT) + (config & SMASK) == 0; |
2816 |
dl |
1.1 |
} |
2817 |
|
|
|
2818 |
|
|
/** |
2819 |
|
|
* Returns an estimate of the total number of tasks stolen from |
2820 |
|
|
* one thread's work queue by another. The reported value |
2821 |
|
|
* underestimates the actual total number of steals when the pool |
2822 |
|
|
* is not quiescent. This value may be useful for monitoring and |
2823 |
|
|
* tuning fork/join programs: in general, steal counts should be |
2824 |
|
|
* high enough to keep threads busy, but low enough to avoid |
2825 |
|
|
* overhead and contention across threads. |
2826 |
|
|
* |
2827 |
|
|
* @return the number of steals |
2828 |
|
|
*/ |
2829 |
|
|
public long getStealCount() { |
2830 |
dl |
1.8 |
long count = stealCount; |
2831 |
dl |
1.1 |
WorkQueue[] ws; WorkQueue w; |
2832 |
|
|
if ((ws = workQueues) != null) { |
2833 |
|
|
for (int i = 1; i < ws.length; i += 2) { |
2834 |
|
|
if ((w = ws[i]) != null) |
2835 |
dl |
1.12 |
count += w.nsteals; |
2836 |
dl |
1.1 |
} |
2837 |
|
|
} |
2838 |
|
|
return count; |
2839 |
|
|
} |
2840 |
|
|
|
2841 |
|
|
/** |
2842 |
|
|
* Returns an estimate of the total number of tasks currently held |
2843 |
|
|
* in queues by worker threads (but not including tasks submitted |
2844 |
|
|
* to the pool that have not begun executing). This value is only |
2845 |
|
|
* an approximation, obtained by iterating across all threads in |
2846 |
|
|
* the pool. This method may be useful for tuning task |
2847 |
|
|
* granularities. |
2848 |
|
|
* |
2849 |
|
|
* @return the number of queued tasks |
2850 |
|
|
*/ |
2851 |
|
|
public long getQueuedTaskCount() { |
2852 |
|
|
long count = 0; |
2853 |
|
|
WorkQueue[] ws; WorkQueue w; |
2854 |
|
|
if ((ws = workQueues) != null) { |
2855 |
|
|
for (int i = 1; i < ws.length; i += 2) { |
2856 |
|
|
if ((w = ws[i]) != null) |
2857 |
|
|
count += w.queueSize(); |
2858 |
|
|
} |
2859 |
|
|
} |
2860 |
|
|
return count; |
2861 |
|
|
} |
2862 |
|
|
|
2863 |
|
|
/** |
2864 |
|
|
* Returns an estimate of the number of tasks submitted to this |
2865 |
|
|
* pool that have not yet begun executing. This method may take |
2866 |
|
|
* time proportional to the number of submissions. |
2867 |
|
|
* |
2868 |
|
|
* @return the number of queued submissions |
2869 |
|
|
*/ |
2870 |
|
|
public int getQueuedSubmissionCount() { |
2871 |
|
|
int count = 0; |
2872 |
|
|
WorkQueue[] ws; WorkQueue w; |
2873 |
|
|
if ((ws = workQueues) != null) { |
2874 |
|
|
for (int i = 0; i < ws.length; i += 2) { |
2875 |
|
|
if ((w = ws[i]) != null) |
2876 |
|
|
count += w.queueSize(); |
2877 |
|
|
} |
2878 |
|
|
} |
2879 |
|
|
return count; |
2880 |
|
|
} |
2881 |
|
|
|
2882 |
|
|
/** |
2883 |
|
|
* Returns {@code true} if there are any tasks submitted to this |
2884 |
|
|
* pool that have not yet begun executing. |
2885 |
|
|
* |
2886 |
|
|
* @return {@code true} if there are any queued submissions |
2887 |
|
|
*/ |
2888 |
|
|
public boolean hasQueuedSubmissions() { |
2889 |
|
|
WorkQueue[] ws; WorkQueue w; |
2890 |
|
|
if ((ws = workQueues) != null) { |
2891 |
|
|
for (int i = 0; i < ws.length; i += 2) { |
2892 |
dl |
1.21 |
if ((w = ws[i]) != null && !w.isEmpty()) |
2893 |
dl |
1.1 |
return true; |
2894 |
|
|
} |
2895 |
|
|
} |
2896 |
|
|
return false; |
2897 |
|
|
} |
2898 |
|
|
|
2899 |
|
|
/** |
2900 |
|
|
* Removes and returns the next unexecuted submission if one is |
2901 |
|
|
* available. This method may be useful in extensions to this |
2902 |
|
|
* class that re-assign work in systems with multiple pools. |
2903 |
|
|
* |
2904 |
|
|
* @return the next submission, or {@code null} if none |
2905 |
|
|
*/ |
2906 |
|
|
protected ForkJoinTask<?> pollSubmission() { |
2907 |
|
|
WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t; |
2908 |
|
|
if ((ws = workQueues) != null) { |
2909 |
|
|
for (int i = 0; i < ws.length; i += 2) { |
2910 |
|
|
if ((w = ws[i]) != null && (t = w.poll()) != null) |
2911 |
|
|
return t; |
2912 |
|
|
} |
2913 |
|
|
} |
2914 |
|
|
return null; |
2915 |
|
|
} |
2916 |
|
|
|
2917 |
|
|
/** |
2918 |
|
|
* Removes all available unexecuted submitted and forked tasks |
2919 |
|
|
* from scheduling queues and adds them to the given collection, |
2920 |
|
|
* without altering their execution status. These may include |
2921 |
|
|
* artificially generated or wrapped tasks. This method is |
2922 |
|
|
* designed to be invoked only when the pool is known to be |
2923 |
|
|
* quiescent. Invocations at other times may not remove all |
2924 |
|
|
* tasks. A failure encountered while attempting to add elements |
2925 |
|
|
* to collection {@code c} may result in elements being in |
2926 |
|
|
* neither, either or both collections when the associated |
2927 |
|
|
* exception is thrown. The behavior of this operation is |
2928 |
|
|
* undefined if the specified collection is modified while the |
2929 |
|
|
* operation is in progress. |
2930 |
|
|
* |
2931 |
|
|
* @param c the collection to transfer elements into |
2932 |
|
|
* @return the number of elements transferred |
2933 |
|
|
*/ |
2934 |
|
|
protected int drainTasksTo(Collection<? super ForkJoinTask<?>> c) { |
2935 |
|
|
int count = 0; |
2936 |
|
|
WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t; |
2937 |
|
|
if ((ws = workQueues) != null) { |
2938 |
|
|
for (int i = 0; i < ws.length; ++i) { |
2939 |
|
|
if ((w = ws[i]) != null) { |
2940 |
|
|
while ((t = w.poll()) != null) { |
2941 |
|
|
c.add(t); |
2942 |
|
|
++count; |
2943 |
|
|
} |
2944 |
|
|
} |
2945 |
|
|
} |
2946 |
|
|
} |
2947 |
|
|
return count; |
2948 |
|
|
} |
2949 |
|
|
|
2950 |
|
|
/** |
2951 |
|
|
* Returns a string identifying this pool, as well as its state, |
2952 |
|
|
* including indications of run state, parallelism level, and |
2953 |
|
|
* worker and task counts. |
2954 |
|
|
* |
2955 |
|
|
* @return a string identifying this pool, as well as its state |
2956 |
|
|
*/ |
2957 |
|
|
public String toString() { |
2958 |
|
|
// Use a single pass through workQueues to collect counts |
2959 |
|
|
long qt = 0L, qs = 0L; int rc = 0; |
2960 |
dl |
1.8 |
long st = stealCount; |
2961 |
dl |
1.1 |
long c = ctl; |
2962 |
|
|
WorkQueue[] ws; WorkQueue w; |
2963 |
|
|
if ((ws = workQueues) != null) { |
2964 |
|
|
for (int i = 0; i < ws.length; ++i) { |
2965 |
|
|
if ((w = ws[i]) != null) { |
2966 |
|
|
int size = w.queueSize(); |
2967 |
|
|
if ((i & 1) == 0) |
2968 |
|
|
qs += size; |
2969 |
|
|
else { |
2970 |
|
|
qt += size; |
2971 |
dl |
1.12 |
st += w.nsteals; |
2972 |
dl |
1.1 |
if (w.isApparentlyUnblocked()) |
2973 |
|
|
++rc; |
2974 |
|
|
} |
2975 |
|
|
} |
2976 |
|
|
} |
2977 |
|
|
} |
2978 |
dl |
1.18 |
int pc = (config & SMASK); |
2979 |
dl |
1.1 |
int tc = pc + (short)(c >>> TC_SHIFT); |
2980 |
|
|
int ac = pc + (int)(c >> AC_SHIFT); |
2981 |
|
|
if (ac < 0) // ignore transient negative |
2982 |
|
|
ac = 0; |
2983 |
|
|
String level; |
2984 |
|
|
if ((c & STOP_BIT) != 0) |
2985 |
|
|
level = (tc == 0) ? "Terminated" : "Terminating"; |
2986 |
|
|
else |
2987 |
dl |
1.12 |
level = plock < 0 ? "Shutting down" : "Running"; |
2988 |
dl |
1.1 |
return super.toString() + |
2989 |
|
|
"[" + level + |
2990 |
|
|
", parallelism = " + pc + |
2991 |
|
|
", size = " + tc + |
2992 |
|
|
", active = " + ac + |
2993 |
|
|
", running = " + rc + |
2994 |
|
|
", steals = " + st + |
2995 |
|
|
", tasks = " + qt + |
2996 |
|
|
", submissions = " + qs + |
2997 |
|
|
"]"; |
2998 |
|
|
} |
2999 |
|
|
|
3000 |
|
|
/** |
3001 |
dl |
1.7 |
* Possibly initiates an orderly shutdown in which previously |
3002 |
|
|
* submitted tasks are executed, but no new tasks will be |
3003 |
|
|
* accepted. Invocation has no effect on execution state if this |
3004 |
|
|
* is the {@link #commonPool}, and no additional effect if |
3005 |
|
|
* already shut down. Tasks that are in the process of being |
3006 |
|
|
* submitted concurrently during the course of this method may or |
3007 |
|
|
* may not be rejected. |
3008 |
dl |
1.1 |
* |
3009 |
|
|
* @throws SecurityException if a security manager exists and |
3010 |
|
|
* the caller is not permitted to modify threads |
3011 |
|
|
* because it does not hold {@link |
3012 |
|
|
* java.lang.RuntimePermission}{@code ("modifyThread")} |
3013 |
|
|
*/ |
3014 |
|
|
public void shutdown() { |
3015 |
|
|
checkPermission(); |
3016 |
dl |
1.12 |
tryTerminate(false, true); |
3017 |
dl |
1.1 |
} |
3018 |
|
|
|
3019 |
|
|
/** |
3020 |
dl |
1.7 |
* Possibly attempts to cancel and/or stop all tasks, and reject |
3021 |
|
|
* all subsequently submitted tasks. Invocation has no effect on |
3022 |
|
|
* execution state if this is the {@link #commonPool}, and no |
3023 |
|
|
* additional effect if already shut down. Otherwise, tasks that |
3024 |
|
|
* are in the process of being submitted or executed concurrently |
3025 |
|
|
* during the course of this method may or may not be |
3026 |
|
|
* rejected. This method cancels both existing and unexecuted |
3027 |
|
|
* tasks, in order to permit termination in the presence of task |
3028 |
|
|
* dependencies. So the method always returns an empty list |
3029 |
|
|
* (unlike the case for some other Executors). |
3030 |
dl |
1.1 |
* |
3031 |
|
|
* @return an empty list |
3032 |
|
|
* @throws SecurityException if a security manager exists and |
3033 |
|
|
* the caller is not permitted to modify threads |
3034 |
|
|
* because it does not hold {@link |
3035 |
|
|
* java.lang.RuntimePermission}{@code ("modifyThread")} |
3036 |
|
|
*/ |
3037 |
|
|
public List<Runnable> shutdownNow() { |
3038 |
|
|
checkPermission(); |
3039 |
dl |
1.12 |
tryTerminate(true, true); |
3040 |
dl |
1.1 |
return Collections.emptyList(); |
3041 |
|
|
} |
3042 |
|
|
|
3043 |
|
|
/** |
3044 |
|
|
* Returns {@code true} if all tasks have completed following shut down. |
3045 |
|
|
* |
3046 |
|
|
* @return {@code true} if all tasks have completed following shut down |
3047 |
|
|
*/ |
3048 |
|
|
public boolean isTerminated() { |
3049 |
|
|
long c = ctl; |
3050 |
|
|
return ((c & STOP_BIT) != 0L && |
3051 |
dl |
1.18 |
(short)(c >>> TC_SHIFT) == -(config & SMASK)); |
3052 |
dl |
1.1 |
} |
3053 |
|
|
|
3054 |
|
|
/** |
3055 |
|
|
* Returns {@code true} if the process of termination has |
3056 |
|
|
* commenced but not yet completed. This method may be useful for |
3057 |
|
|
* debugging. A return of {@code true} reported a sufficient |
3058 |
|
|
* period after shutdown may indicate that submitted tasks have |
3059 |
jsr166 |
1.25 |
* ignored or suppressed interruption, or are waiting for I/O, |
3060 |
dl |
1.1 |
* causing this executor not to properly terminate. (See the |
3061 |
|
|
* advisory notes for class {@link ForkJoinTask} stating that |
3062 |
|
|
* tasks should not normally entail blocking operations. But if |
3063 |
|
|
* they do, they must abort them on interrupt.) |
3064 |
|
|
* |
3065 |
|
|
* @return {@code true} if terminating but not yet terminated |
3066 |
|
|
*/ |
3067 |
|
|
public boolean isTerminating() { |
3068 |
|
|
long c = ctl; |
3069 |
|
|
return ((c & STOP_BIT) != 0L && |
3070 |
dl |
1.18 |
(short)(c >>> TC_SHIFT) != -(config & SMASK)); |
3071 |
dl |
1.1 |
} |
3072 |
|
|
|
3073 |
|
|
/** |
3074 |
|
|
* Returns {@code true} if this pool has been shut down. |
3075 |
|
|
* |
3076 |
|
|
* @return {@code true} if this pool has been shut down |
3077 |
|
|
*/ |
3078 |
|
|
public boolean isShutdown() { |
3079 |
dl |
1.12 |
return plock < 0; |
3080 |
dl |
1.1 |
} |
3081 |
|
|
|
3082 |
|
|
/** |
3083 |
dl |
1.12 |
* Blocks until all tasks have completed execution after a |
3084 |
|
|
* shutdown request, or the timeout occurs, or the current thread |
3085 |
dl |
1.40 |
* is interrupted, whichever happens first. Because the {@link |
3086 |
|
|
* #commonPool()} never terminates until program shutdown, when |
3087 |
|
|
* applied to the common pool, this method is equivalent to {@link |
3088 |
|
|
* #awaitQuiescence} but always returns {@code false}. |
3089 |
dl |
1.1 |
* |
3090 |
|
|
* @param timeout the maximum time to wait |
3091 |
|
|
* @param unit the time unit of the timeout argument |
3092 |
|
|
* @return {@code true} if this executor terminated and |
3093 |
|
|
* {@code false} if the timeout elapsed before termination |
3094 |
|
|
* @throws InterruptedException if interrupted while waiting |
3095 |
|
|
*/ |
3096 |
|
|
public boolean awaitTermination(long timeout, TimeUnit unit) |
3097 |
|
|
throws InterruptedException { |
3098 |
dl |
1.40 |
if (Thread.interrupted()) |
3099 |
|
|
throw new InterruptedException(); |
3100 |
|
|
if (this == common) { |
3101 |
|
|
awaitQuiescence(timeout, unit); |
3102 |
|
|
return false; |
3103 |
|
|
} |
3104 |
dl |
1.1 |
long nanos = unit.toNanos(timeout); |
3105 |
dl |
1.8 |
if (isTerminated()) |
3106 |
|
|
return true; |
3107 |
|
|
long startTime = System.nanoTime(); |
3108 |
|
|
boolean terminated = false; |
3109 |
jsr166 |
1.10 |
synchronized (this) { |
3110 |
dl |
1.8 |
for (long waitTime = nanos, millis = 0L;;) { |
3111 |
|
|
if (terminated = isTerminated() || |
3112 |
|
|
waitTime <= 0L || |
3113 |
|
|
(millis = unit.toMillis(waitTime)) <= 0L) |
3114 |
|
|
break; |
3115 |
|
|
wait(millis); |
3116 |
|
|
waitTime = nanos - (System.nanoTime() - startTime); |
3117 |
dl |
1.1 |
} |
3118 |
|
|
} |
3119 |
dl |
1.8 |
return terminated; |
3120 |
dl |
1.1 |
} |
3121 |
|
|
|
3122 |
|
|
/** |
3123 |
dl |
1.40 |
* If called by a ForkJoinTask operating in this pool, equivalent |
3124 |
|
|
* in effect to {@link ForkJoinTask#helpQuiesce}. Otherwise, |
3125 |
|
|
* waits and/or attempts to assist performing tasks until this |
3126 |
|
|
* pool {@link #isQuiescent} or the indicated timeout elapses. |
3127 |
|
|
* |
3128 |
|
|
* @param timeout the maximum time to wait |
3129 |
|
|
* @param unit the time unit of the timeout argument |
3130 |
|
|
* @return {@code true} if quiescent; {@code false} if the |
3131 |
|
|
* timeout elapsed. |
3132 |
|
|
*/ |
3133 |
|
|
public boolean awaitQuiescence(long timeout, TimeUnit unit) { |
3134 |
|
|
long nanos = unit.toNanos(timeout); |
3135 |
|
|
ForkJoinWorkerThread wt; |
3136 |
|
|
Thread thread = Thread.currentThread(); |
3137 |
|
|
if ((thread instanceof ForkJoinWorkerThread) && |
3138 |
|
|
(wt = (ForkJoinWorkerThread)thread).pool == this) { |
3139 |
|
|
helpQuiescePool(wt.workQueue); |
3140 |
|
|
return true; |
3141 |
|
|
} |
3142 |
|
|
long startTime = System.nanoTime(); |
3143 |
|
|
WorkQueue[] ws; |
3144 |
|
|
int r = 0, m; |
3145 |
|
|
boolean found = true; |
3146 |
|
|
while (!isQuiescent() && (ws = workQueues) != null && |
3147 |
|
|
(m = ws.length - 1) >= 0) { |
3148 |
|
|
if (!found) { |
3149 |
|
|
if ((System.nanoTime() - startTime) > nanos) |
3150 |
|
|
return false; |
3151 |
|
|
Thread.yield(); // cannot block |
3152 |
|
|
} |
3153 |
|
|
found = false; |
3154 |
|
|
for (int j = (m + 1) << 2; j >= 0; --j) { |
3155 |
|
|
ForkJoinTask<?> t; WorkQueue q; int b; |
3156 |
|
|
if ((q = ws[r++ & m]) != null && (b = q.base) - q.top < 0) { |
3157 |
|
|
found = true; |
3158 |
|
|
if ((t = q.pollAt(b)) != null) { |
3159 |
|
|
if (q.base - q.top < 0) |
3160 |
|
|
signalWork(q); |
3161 |
|
|
t.doExec(); |
3162 |
|
|
} |
3163 |
|
|
break; |
3164 |
|
|
} |
3165 |
|
|
} |
3166 |
|
|
} |
3167 |
|
|
return true; |
3168 |
|
|
} |
3169 |
|
|
|
3170 |
|
|
/** |
3171 |
|
|
* Waits and/or attempts to assist performing tasks indefinitely |
3172 |
|
|
* until the {@link #commonPool()} {@link #isQuiescent} |
3173 |
|
|
*/ |
3174 |
dl |
1.42 |
static void quiesceCommonPool() { |
3175 |
dl |
1.40 |
common.awaitQuiescence(Long.MAX_VALUE, TimeUnit.NANOSECONDS); |
3176 |
|
|
} |
3177 |
|
|
|
3178 |
|
|
/** |
3179 |
dl |
1.1 |
* Interface for extending managed parallelism for tasks running |
3180 |
|
|
* in {@link ForkJoinPool}s. |
3181 |
|
|
* |
3182 |
|
|
* <p>A {@code ManagedBlocker} provides two methods. Method |
3183 |
|
|
* {@code isReleasable} must return {@code true} if blocking is |
3184 |
|
|
* not necessary. Method {@code block} blocks the current thread |
3185 |
|
|
* if necessary (perhaps internally invoking {@code isReleasable} |
3186 |
|
|
* before actually blocking). These actions are performed by any |
3187 |
|
|
* thread invoking {@link ForkJoinPool#managedBlock}. The |
3188 |
|
|
* unusual methods in this API accommodate synchronizers that may, |
3189 |
|
|
* but don't usually, block for long periods. Similarly, they |
3190 |
|
|
* allow more efficient internal handling of cases in which |
3191 |
|
|
* additional workers may be, but usually are not, needed to |
3192 |
|
|
* ensure sufficient parallelism. Toward this end, |
3193 |
|
|
* implementations of method {@code isReleasable} must be amenable |
3194 |
|
|
* to repeated invocation. |
3195 |
|
|
* |
3196 |
|
|
* <p>For example, here is a ManagedBlocker based on a |
3197 |
|
|
* ReentrantLock: |
3198 |
|
|
* <pre> {@code |
3199 |
|
|
* class ManagedLocker implements ManagedBlocker { |
3200 |
|
|
* final ReentrantLock lock; |
3201 |
|
|
* boolean hasLock = false; |
3202 |
|
|
* ManagedLocker(ReentrantLock lock) { this.lock = lock; } |
3203 |
|
|
* public boolean block() { |
3204 |
|
|
* if (!hasLock) |
3205 |
|
|
* lock.lock(); |
3206 |
|
|
* return true; |
3207 |
|
|
* } |
3208 |
|
|
* public boolean isReleasable() { |
3209 |
|
|
* return hasLock || (hasLock = lock.tryLock()); |
3210 |
|
|
* } |
3211 |
|
|
* }}</pre> |
3212 |
|
|
* |
3213 |
|
|
* <p>Here is a class that possibly blocks waiting for an |
3214 |
|
|
* item on a given queue: |
3215 |
|
|
* <pre> {@code |
3216 |
|
|
* class QueueTaker<E> implements ManagedBlocker { |
3217 |
|
|
* final BlockingQueue<E> queue; |
3218 |
|
|
* volatile E item = null; |
3219 |
|
|
* QueueTaker(BlockingQueue<E> q) { this.queue = q; } |
3220 |
|
|
* public boolean block() throws InterruptedException { |
3221 |
|
|
* if (item == null) |
3222 |
|
|
* item = queue.take(); |
3223 |
|
|
* return true; |
3224 |
|
|
* } |
3225 |
|
|
* public boolean isReleasable() { |
3226 |
|
|
* return item != null || (item = queue.poll()) != null; |
3227 |
|
|
* } |
3228 |
|
|
* public E getItem() { // call after pool.managedBlock completes |
3229 |
|
|
* return item; |
3230 |
|
|
* } |
3231 |
|
|
* }}</pre> |
3232 |
|
|
*/ |
3233 |
|
|
public static interface ManagedBlocker { |
3234 |
|
|
/** |
3235 |
|
|
* Possibly blocks the current thread, for example waiting for |
3236 |
|
|
* a lock or condition. |
3237 |
|
|
* |
3238 |
|
|
* @return {@code true} if no additional blocking is necessary |
3239 |
|
|
* (i.e., if isReleasable would return true) |
3240 |
|
|
* @throws InterruptedException if interrupted while waiting |
3241 |
|
|
* (the method is not required to do so, but is allowed to) |
3242 |
|
|
*/ |
3243 |
|
|
boolean block() throws InterruptedException; |
3244 |
|
|
|
3245 |
|
|
/** |
3246 |
|
|
* Returns {@code true} if blocking is unnecessary. |
3247 |
|
|
*/ |
3248 |
|
|
boolean isReleasable(); |
3249 |
|
|
} |
3250 |
|
|
|
3251 |
|
|
/** |
3252 |
|
|
* Blocks in accord with the given blocker. If the current thread |
3253 |
|
|
* is a {@link ForkJoinWorkerThread}, this method possibly |
3254 |
|
|
* arranges for a spare thread to be activated if necessary to |
3255 |
|
|
* ensure sufficient parallelism while the current thread is blocked. |
3256 |
|
|
* |
3257 |
|
|
* <p>If the caller is not a {@link ForkJoinTask}, this method is |
3258 |
|
|
* behaviorally equivalent to |
3259 |
|
|
* <pre> {@code |
3260 |
|
|
* while (!blocker.isReleasable()) |
3261 |
|
|
* if (blocker.block()) |
3262 |
|
|
* return; |
3263 |
|
|
* }</pre> |
3264 |
|
|
* |
3265 |
|
|
* If the caller is a {@code ForkJoinTask}, then the pool may |
3266 |
|
|
* first be expanded to ensure parallelism, and later adjusted. |
3267 |
|
|
* |
3268 |
|
|
* @param blocker the blocker |
3269 |
|
|
* @throws InterruptedException if blocker.block did so |
3270 |
|
|
*/ |
3271 |
|
|
public static void managedBlock(ManagedBlocker blocker) |
3272 |
|
|
throws InterruptedException { |
3273 |
|
|
Thread t = Thread.currentThread(); |
3274 |
dl |
1.12 |
if (t instanceof ForkJoinWorkerThread) { |
3275 |
|
|
ForkJoinPool p = ((ForkJoinWorkerThread)t).pool; |
3276 |
|
|
while (!blocker.isReleasable()) { // variant of helpSignal |
3277 |
dl |
1.21 |
WorkQueue[] ws; WorkQueue q; int m, u; |
3278 |
dl |
1.12 |
if ((ws = p.workQueues) != null && (m = ws.length - 1) >= 0) { |
3279 |
|
|
for (int i = 0; i <= m; ++i) { |
3280 |
|
|
if (blocker.isReleasable()) |
3281 |
|
|
return; |
3282 |
dl |
1.21 |
if ((q = ws[i]) != null && q.base - q.top < 0) { |
3283 |
|
|
p.signalWork(q); |
3284 |
dl |
1.18 |
if ((u = (int)(p.ctl >>> 32)) >= 0 || |
3285 |
|
|
(u >> UAC_SHIFT) >= 0) |
3286 |
dl |
1.12 |
break; |
3287 |
|
|
} |
3288 |
|
|
} |
3289 |
|
|
} |
3290 |
|
|
if (p.tryCompensate()) { |
3291 |
|
|
try { |
3292 |
|
|
do {} while (!blocker.isReleasable() && |
3293 |
|
|
!blocker.block()); |
3294 |
|
|
} finally { |
3295 |
dl |
1.1 |
p.incrementActiveCount(); |
3296 |
dl |
1.12 |
} |
3297 |
|
|
break; |
3298 |
dl |
1.1 |
} |
3299 |
|
|
} |
3300 |
|
|
} |
3301 |
dl |
1.12 |
else { |
3302 |
|
|
do {} while (!blocker.isReleasable() && |
3303 |
|
|
!blocker.block()); |
3304 |
|
|
} |
3305 |
dl |
1.1 |
} |
3306 |
|
|
|
3307 |
|
|
// AbstractExecutorService overrides. These rely on undocumented |
3308 |
|
|
// fact that ForkJoinTask.adapt returns ForkJoinTasks that also |
3309 |
|
|
// implement RunnableFuture. |
3310 |
|
|
|
3311 |
|
|
protected <T> RunnableFuture<T> newTaskFor(Runnable runnable, T value) { |
3312 |
|
|
return new ForkJoinTask.AdaptedRunnable<T>(runnable, value); |
3313 |
|
|
} |
3314 |
|
|
|
3315 |
|
|
protected <T> RunnableFuture<T> newTaskFor(Callable<T> callable) { |
3316 |
|
|
return new ForkJoinTask.AdaptedCallable<T>(callable); |
3317 |
|
|
} |
3318 |
|
|
|
3319 |
|
|
// Unsafe mechanics |
3320 |
|
|
private static final sun.misc.Unsafe U; |
3321 |
|
|
private static final long CTL; |
3322 |
|
|
private static final long PARKBLOCKER; |
3323 |
|
|
private static final int ABASE; |
3324 |
|
|
private static final int ASHIFT; |
3325 |
dl |
1.8 |
private static final long STEALCOUNT; |
3326 |
dl |
1.12 |
private static final long PLOCK; |
3327 |
|
|
private static final long INDEXSEED; |
3328 |
|
|
private static final long QLOCK; |
3329 |
dl |
1.1 |
|
3330 |
|
|
static { |
3331 |
dl |
1.12 |
int s; // initialize field offsets for CAS etc |
3332 |
dl |
1.1 |
try { |
3333 |
|
|
U = getUnsafe(); |
3334 |
|
|
Class<?> k = ForkJoinPool.class; |
3335 |
|
|
CTL = U.objectFieldOffset |
3336 |
|
|
(k.getDeclaredField("ctl")); |
3337 |
dl |
1.8 |
STEALCOUNT = U.objectFieldOffset |
3338 |
|
|
(k.getDeclaredField("stealCount")); |
3339 |
dl |
1.12 |
PLOCK = U.objectFieldOffset |
3340 |
|
|
(k.getDeclaredField("plock")); |
3341 |
|
|
INDEXSEED = U.objectFieldOffset |
3342 |
|
|
(k.getDeclaredField("indexSeed")); |
3343 |
dl |
1.1 |
Class<?> tk = Thread.class; |
3344 |
|
|
PARKBLOCKER = U.objectFieldOffset |
3345 |
|
|
(tk.getDeclaredField("parkBlocker")); |
3346 |
dl |
1.12 |
Class<?> wk = WorkQueue.class; |
3347 |
|
|
QLOCK = U.objectFieldOffset |
3348 |
|
|
(wk.getDeclaredField("qlock")); |
3349 |
|
|
Class<?> ak = ForkJoinTask[].class; |
3350 |
dl |
1.1 |
ABASE = U.arrayBaseOffset(ak); |
3351 |
|
|
s = U.arrayIndexScale(ak); |
3352 |
dl |
1.8 |
ASHIFT = 31 - Integer.numberOfLeadingZeros(s); |
3353 |
dl |
1.1 |
} catch (Exception e) { |
3354 |
|
|
throw new Error(e); |
3355 |
|
|
} |
3356 |
|
|
if ((s & (s-1)) != 0) |
3357 |
|
|
throw new Error("data type scale not a power of two"); |
3358 |
dl |
1.12 |
|
3359 |
dl |
1.18 |
submitters = new ThreadLocal<Submitter>(); |
3360 |
|
|
ForkJoinWorkerThreadFactory fac = defaultForkJoinWorkerThreadFactory = |
3361 |
|
|
new DefaultForkJoinWorkerThreadFactory(); |
3362 |
dl |
1.21 |
modifyThreadPermission = new RuntimePermission("modifyThread"); |
3363 |
|
|
|
3364 |
dl |
1.12 |
/* |
3365 |
dl |
1.18 |
* Establish common pool parameters. For extra caution, |
3366 |
|
|
* computations to set up common pool state are here; the |
3367 |
|
|
* constructor just assigns these values to fields. |
3368 |
dl |
1.12 |
*/ |
3369 |
dl |
1.18 |
|
3370 |
|
|
int par = 0; |
3371 |
|
|
Thread.UncaughtExceptionHandler handler = null; |
3372 |
|
|
try { // TBD: limit or report ignored exceptions? |
3373 |
|
|
String pp = System.getProperty |
3374 |
|
|
("java.util.concurrent.ForkJoinPool.common.parallelism"); |
3375 |
|
|
String hp = System.getProperty |
3376 |
|
|
("java.util.concurrent.ForkJoinPool.common.exceptionHandler"); |
3377 |
|
|
String fp = System.getProperty |
3378 |
|
|
("java.util.concurrent.ForkJoinPool.common.threadFactory"); |
3379 |
|
|
if (fp != null) |
3380 |
|
|
fac = ((ForkJoinWorkerThreadFactory)ClassLoader. |
3381 |
|
|
getSystemClassLoader().loadClass(fp).newInstance()); |
3382 |
|
|
if (hp != null) |
3383 |
|
|
handler = ((Thread.UncaughtExceptionHandler)ClassLoader. |
3384 |
|
|
getSystemClassLoader().loadClass(hp).newInstance()); |
3385 |
|
|
if (pp != null) |
3386 |
|
|
par = Integer.parseInt(pp); |
3387 |
|
|
} catch (Exception ignore) { |
3388 |
|
|
} |
3389 |
|
|
|
3390 |
dl |
1.12 |
if (par <= 0) |
3391 |
|
|
par = Runtime.getRuntime().availableProcessors(); |
3392 |
|
|
if (par > MAX_CAP) |
3393 |
|
|
par = MAX_CAP; |
3394 |
dl |
1.40 |
commonParallelism = par; |
3395 |
dl |
1.12 |
long np = (long)(-par); // precompute initial ctl value |
3396 |
|
|
long ct = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK); |
3397 |
|
|
|
3398 |
dl |
1.40 |
common = new ForkJoinPool(par, ct, fac, handler); |
3399 |
dl |
1.1 |
} |
3400 |
|
|
|
3401 |
|
|
/** |
3402 |
|
|
* Returns a sun.misc.Unsafe. Suitable for use in a 3rd party package. |
3403 |
|
|
* Replace with a simple call to Unsafe.getUnsafe when integrating |
3404 |
|
|
* into a jdk. |
3405 |
|
|
* |
3406 |
|
|
* @return a sun.misc.Unsafe |
3407 |
|
|
*/ |
3408 |
|
|
private static sun.misc.Unsafe getUnsafe() { |
3409 |
|
|
try { |
3410 |
|
|
return sun.misc.Unsafe.getUnsafe(); |
3411 |
|
|
} catch (SecurityException se) { |
3412 |
|
|
try { |
3413 |
|
|
return java.security.AccessController.doPrivileged |
3414 |
|
|
(new java.security |
3415 |
|
|
.PrivilegedExceptionAction<sun.misc.Unsafe>() { |
3416 |
|
|
public sun.misc.Unsafe run() throws Exception { |
3417 |
|
|
java.lang.reflect.Field f = sun.misc |
3418 |
|
|
.Unsafe.class.getDeclaredField("theUnsafe"); |
3419 |
|
|
f.setAccessible(true); |
3420 |
|
|
return (sun.misc.Unsafe) f.get(null); |
3421 |
|
|
}}); |
3422 |
|
|
} catch (java.security.PrivilegedActionException e) { |
3423 |
|
|
throw new RuntimeException("Could not initialize intrinsics", |
3424 |
|
|
e.getCause()); |
3425 |
|
|
} |
3426 |
|
|
} |
3427 |
|
|
} |
3428 |
|
|
|
3429 |
|
|
} |